2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
66 #include <asm/sections.h>
67 #include <asm/tlbflush.h>
68 #include <asm/div64.h>
71 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
72 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #define MIN_PERCPU_PAGELIST_FRACTION (8)
75 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
76 DEFINE_PER_CPU(int, numa_node);
77 EXPORT_PER_CPU_SYMBOL(numa_node);
80 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
82 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
83 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
84 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
85 * defined in <linux/topology.h>.
87 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
88 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
89 int _node_numa_mem_[MAX_NUMNODES];
93 * Array of node states.
95 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
96 [N_POSSIBLE] = NODE_MASK_ALL,
97 [N_ONLINE] = { { [0] = 1UL } },
99 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_HIGHMEM
101 [N_HIGH_MEMORY] = { { [0] = 1UL } },
103 #ifdef CONFIG_MOVABLE_NODE
104 [N_MEMORY] = { { [0] = 1UL } },
106 [N_CPU] = { { [0] = 1UL } },
109 EXPORT_SYMBOL(node_states);
111 /* Protect totalram_pages and zone->managed_pages */
112 static DEFINE_SPINLOCK(managed_page_count_lock);
114 unsigned long totalram_pages __read_mostly;
115 unsigned long totalreserve_pages __read_mostly;
116 unsigned long totalcma_pages __read_mostly;
118 * When calculating the number of globally allowed dirty pages, there
119 * is a certain number of per-zone reserves that should not be
120 * considered dirtyable memory. This is the sum of those reserves
121 * over all existing zones that contribute dirtyable memory.
123 unsigned long dirty_balance_reserve __read_mostly;
125 int percpu_pagelist_fraction;
126 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
136 static inline int get_pcppage_migratetype(struct page *page)
141 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
143 page->index = migratetype;
146 #ifdef CONFIG_PM_SLEEP
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
156 static gfp_t saved_gfp_mask;
158 void pm_restore_gfp_mask(void)
160 WARN_ON(!mutex_is_locked(&pm_mutex));
161 if (saved_gfp_mask) {
162 gfp_allowed_mask = saved_gfp_mask;
167 void pm_restrict_gfp_mask(void)
169 WARN_ON(!mutex_is_locked(&pm_mutex));
170 WARN_ON(saved_gfp_mask);
171 saved_gfp_mask = gfp_allowed_mask;
172 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
175 bool pm_suspended_storage(void)
177 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
181 #endif /* CONFIG_PM_SLEEP */
183 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184 int pageblock_order __read_mostly;
187 static void __free_pages_ok(struct page *page, unsigned int order);
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
200 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
201 #ifdef CONFIG_ZONE_DMA
204 #ifdef CONFIG_ZONE_DMA32
207 #ifdef CONFIG_HIGHMEM
213 EXPORT_SYMBOL(totalram_pages);
215 static char * const zone_names[MAX_NR_ZONES] = {
216 #ifdef CONFIG_ZONE_DMA
219 #ifdef CONFIG_ZONE_DMA32
223 #ifdef CONFIG_HIGHMEM
227 #ifdef CONFIG_ZONE_DEVICE
232 static void free_compound_page(struct page *page);
233 compound_page_dtor * const compound_page_dtors[] = {
236 #ifdef CONFIG_HUGETLB_PAGE
241 int min_free_kbytes = 1024;
242 int user_min_free_kbytes = -1;
244 static unsigned long __meminitdata nr_kernel_pages;
245 static unsigned long __meminitdata nr_all_pages;
246 static unsigned long __meminitdata dma_reserve;
248 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
249 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
250 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
251 static unsigned long __initdata required_kernelcore;
252 static unsigned long __initdata required_movablecore;
253 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
255 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
257 EXPORT_SYMBOL(movable_zone);
258 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
261 int nr_node_ids __read_mostly = MAX_NUMNODES;
262 int nr_online_nodes __read_mostly = 1;
263 EXPORT_SYMBOL(nr_node_ids);
264 EXPORT_SYMBOL(nr_online_nodes);
267 int page_group_by_mobility_disabled __read_mostly;
269 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
270 static inline void reset_deferred_meminit(pg_data_t *pgdat)
272 pgdat->first_deferred_pfn = ULONG_MAX;
275 /* Returns true if the struct page for the pfn is uninitialised */
276 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
278 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
284 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
286 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
293 * Returns false when the remaining initialisation should be deferred until
294 * later in the boot cycle when it can be parallelised.
296 static inline bool update_defer_init(pg_data_t *pgdat,
297 unsigned long pfn, unsigned long zone_end,
298 unsigned long *nr_initialised)
300 /* Always populate low zones for address-contrained allocations */
301 if (zone_end < pgdat_end_pfn(pgdat))
304 /* Initialise at least 2G of the highest zone */
306 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
307 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
308 pgdat->first_deferred_pfn = pfn;
315 static inline void reset_deferred_meminit(pg_data_t *pgdat)
319 static inline bool early_page_uninitialised(unsigned long pfn)
324 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
329 static inline bool update_defer_init(pg_data_t *pgdat,
330 unsigned long pfn, unsigned long zone_end,
331 unsigned long *nr_initialised)
338 void set_pageblock_migratetype(struct page *page, int migratetype)
340 if (unlikely(page_group_by_mobility_disabled &&
341 migratetype < MIGRATE_PCPTYPES))
342 migratetype = MIGRATE_UNMOVABLE;
344 set_pageblock_flags_group(page, (unsigned long)migratetype,
345 PB_migrate, PB_migrate_end);
348 #ifdef CONFIG_DEBUG_VM
349 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
353 unsigned long pfn = page_to_pfn(page);
354 unsigned long sp, start_pfn;
357 seq = zone_span_seqbegin(zone);
358 start_pfn = zone->zone_start_pfn;
359 sp = zone->spanned_pages;
360 if (!zone_spans_pfn(zone, pfn))
362 } while (zone_span_seqretry(zone, seq));
365 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
366 pfn, zone_to_nid(zone), zone->name,
367 start_pfn, start_pfn + sp);
372 static int page_is_consistent(struct zone *zone, struct page *page)
374 if (!pfn_valid_within(page_to_pfn(page)))
376 if (zone != page_zone(page))
382 * Temporary debugging check for pages not lying within a given zone.
384 static int bad_range(struct zone *zone, struct page *page)
386 if (page_outside_zone_boundaries(zone, page))
388 if (!page_is_consistent(zone, page))
394 static inline int bad_range(struct zone *zone, struct page *page)
400 static void bad_page(struct page *page, const char *reason,
401 unsigned long bad_flags)
403 static unsigned long resume;
404 static unsigned long nr_shown;
405 static unsigned long nr_unshown;
407 /* Don't complain about poisoned pages */
408 if (PageHWPoison(page)) {
409 page_mapcount_reset(page); /* remove PageBuddy */
414 * Allow a burst of 60 reports, then keep quiet for that minute;
415 * or allow a steady drip of one report per second.
417 if (nr_shown == 60) {
418 if (time_before(jiffies, resume)) {
424 "BUG: Bad page state: %lu messages suppressed\n",
431 resume = jiffies + 60 * HZ;
433 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
434 current->comm, page_to_pfn(page));
435 dump_page_badflags(page, reason, bad_flags);
440 /* Leave bad fields for debug, except PageBuddy could make trouble */
441 page_mapcount_reset(page); /* remove PageBuddy */
442 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
446 * Higher-order pages are called "compound pages". They are structured thusly:
448 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
450 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
451 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
453 * The first tail page's ->compound_dtor holds the offset in array of compound
454 * page destructors. See compound_page_dtors.
456 * The first tail page's ->compound_order holds the order of allocation.
457 * This usage means that zero-order pages may not be compound.
460 static void free_compound_page(struct page *page)
462 __free_pages_ok(page, compound_order(page));
465 void prep_compound_page(struct page *page, unsigned long order)
468 int nr_pages = 1 << order;
470 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
471 set_compound_order(page, order);
473 for (i = 1; i < nr_pages; i++) {
474 struct page *p = page + i;
475 set_page_count(p, 0);
476 set_compound_head(p, page);
480 #ifdef CONFIG_DEBUG_PAGEALLOC
481 unsigned int _debug_guardpage_minorder;
482 bool _debug_pagealloc_enabled __read_mostly;
483 bool _debug_guardpage_enabled __read_mostly;
485 static int __init early_debug_pagealloc(char *buf)
490 if (strcmp(buf, "on") == 0)
491 _debug_pagealloc_enabled = true;
495 early_param("debug_pagealloc", early_debug_pagealloc);
497 static bool need_debug_guardpage(void)
499 /* If we don't use debug_pagealloc, we don't need guard page */
500 if (!debug_pagealloc_enabled())
506 static void init_debug_guardpage(void)
508 if (!debug_pagealloc_enabled())
511 _debug_guardpage_enabled = true;
514 struct page_ext_operations debug_guardpage_ops = {
515 .need = need_debug_guardpage,
516 .init = init_debug_guardpage,
519 static int __init debug_guardpage_minorder_setup(char *buf)
523 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
524 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
527 _debug_guardpage_minorder = res;
528 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
531 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
533 static inline void set_page_guard(struct zone *zone, struct page *page,
534 unsigned int order, int migratetype)
536 struct page_ext *page_ext;
538 if (!debug_guardpage_enabled())
541 page_ext = lookup_page_ext(page);
542 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
544 INIT_LIST_HEAD(&page->lru);
545 set_page_private(page, order);
546 /* Guard pages are not available for any usage */
547 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
550 static inline void clear_page_guard(struct zone *zone, struct page *page,
551 unsigned int order, int migratetype)
553 struct page_ext *page_ext;
555 if (!debug_guardpage_enabled())
558 page_ext = lookup_page_ext(page);
559 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
561 set_page_private(page, 0);
562 if (!is_migrate_isolate(migratetype))
563 __mod_zone_freepage_state(zone, (1 << order), migratetype);
566 struct page_ext_operations debug_guardpage_ops = { NULL, };
567 static inline void set_page_guard(struct zone *zone, struct page *page,
568 unsigned int order, int migratetype) {}
569 static inline void clear_page_guard(struct zone *zone, struct page *page,
570 unsigned int order, int migratetype) {}
573 static inline void set_page_order(struct page *page, unsigned int order)
575 set_page_private(page, order);
576 __SetPageBuddy(page);
579 static inline void rmv_page_order(struct page *page)
581 __ClearPageBuddy(page);
582 set_page_private(page, 0);
586 * This function checks whether a page is free && is the buddy
587 * we can do coalesce a page and its buddy if
588 * (a) the buddy is not in a hole &&
589 * (b) the buddy is in the buddy system &&
590 * (c) a page and its buddy have the same order &&
591 * (d) a page and its buddy are in the same zone.
593 * For recording whether a page is in the buddy system, we set ->_mapcount
594 * PAGE_BUDDY_MAPCOUNT_VALUE.
595 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
596 * serialized by zone->lock.
598 * For recording page's order, we use page_private(page).
600 static inline int page_is_buddy(struct page *page, struct page *buddy,
603 if (!pfn_valid_within(page_to_pfn(buddy)))
606 if (page_is_guard(buddy) && page_order(buddy) == order) {
607 if (page_zone_id(page) != page_zone_id(buddy))
610 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
615 if (PageBuddy(buddy) && page_order(buddy) == order) {
617 * zone check is done late to avoid uselessly
618 * calculating zone/node ids for pages that could
621 if (page_zone_id(page) != page_zone_id(buddy))
624 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
632 * Freeing function for a buddy system allocator.
634 * The concept of a buddy system is to maintain direct-mapped table
635 * (containing bit values) for memory blocks of various "orders".
636 * The bottom level table contains the map for the smallest allocatable
637 * units of memory (here, pages), and each level above it describes
638 * pairs of units from the levels below, hence, "buddies".
639 * At a high level, all that happens here is marking the table entry
640 * at the bottom level available, and propagating the changes upward
641 * as necessary, plus some accounting needed to play nicely with other
642 * parts of the VM system.
643 * At each level, we keep a list of pages, which are heads of continuous
644 * free pages of length of (1 << order) and marked with _mapcount
645 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
647 * So when we are allocating or freeing one, we can derive the state of the
648 * other. That is, if we allocate a small block, and both were
649 * free, the remainder of the region must be split into blocks.
650 * If a block is freed, and its buddy is also free, then this
651 * triggers coalescing into a block of larger size.
656 static inline void __free_one_page(struct page *page,
658 struct zone *zone, unsigned int order,
661 unsigned long page_idx;
662 unsigned long combined_idx;
663 unsigned long uninitialized_var(buddy_idx);
665 int max_order = MAX_ORDER;
667 VM_BUG_ON(!zone_is_initialized(zone));
668 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
670 VM_BUG_ON(migratetype == -1);
671 if (is_migrate_isolate(migratetype)) {
673 * We restrict max order of merging to prevent merge
674 * between freepages on isolate pageblock and normal
675 * pageblock. Without this, pageblock isolation
676 * could cause incorrect freepage accounting.
678 max_order = min(MAX_ORDER, pageblock_order + 1);
680 __mod_zone_freepage_state(zone, 1 << order, migratetype);
683 page_idx = pfn & ((1 << max_order) - 1);
685 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
686 VM_BUG_ON_PAGE(bad_range(zone, page), page);
688 while (order < max_order - 1) {
689 buddy_idx = __find_buddy_index(page_idx, order);
690 buddy = page + (buddy_idx - page_idx);
691 if (!page_is_buddy(page, buddy, order))
694 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
695 * merge with it and move up one order.
697 if (page_is_guard(buddy)) {
698 clear_page_guard(zone, buddy, order, migratetype);
700 list_del(&buddy->lru);
701 zone->free_area[order].nr_free--;
702 rmv_page_order(buddy);
704 combined_idx = buddy_idx & page_idx;
705 page = page + (combined_idx - page_idx);
706 page_idx = combined_idx;
709 set_page_order(page, order);
712 * If this is not the largest possible page, check if the buddy
713 * of the next-highest order is free. If it is, it's possible
714 * that pages are being freed that will coalesce soon. In case,
715 * that is happening, add the free page to the tail of the list
716 * so it's less likely to be used soon and more likely to be merged
717 * as a higher order page
719 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
720 struct page *higher_page, *higher_buddy;
721 combined_idx = buddy_idx & page_idx;
722 higher_page = page + (combined_idx - page_idx);
723 buddy_idx = __find_buddy_index(combined_idx, order + 1);
724 higher_buddy = higher_page + (buddy_idx - combined_idx);
725 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
726 list_add_tail(&page->lru,
727 &zone->free_area[order].free_list[migratetype]);
732 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
734 zone->free_area[order].nr_free++;
737 static inline int free_pages_check(struct page *page)
739 const char *bad_reason = NULL;
740 unsigned long bad_flags = 0;
742 if (unlikely(page_mapcount(page)))
743 bad_reason = "nonzero mapcount";
744 if (unlikely(page->mapping != NULL))
745 bad_reason = "non-NULL mapping";
746 if (unlikely(atomic_read(&page->_count) != 0))
747 bad_reason = "nonzero _count";
748 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
749 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
750 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
753 if (unlikely(page->mem_cgroup))
754 bad_reason = "page still charged to cgroup";
756 if (unlikely(bad_reason)) {
757 bad_page(page, bad_reason, bad_flags);
760 page_cpupid_reset_last(page);
761 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
762 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
767 * Frees a number of pages from the PCP lists
768 * Assumes all pages on list are in same zone, and of same order.
769 * count is the number of pages to free.
771 * If the zone was previously in an "all pages pinned" state then look to
772 * see if this freeing clears that state.
774 * And clear the zone's pages_scanned counter, to hold off the "all pages are
775 * pinned" detection logic.
777 static void free_pcppages_bulk(struct zone *zone, int count,
778 struct per_cpu_pages *pcp)
783 unsigned long nr_scanned;
785 spin_lock(&zone->lock);
786 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
788 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
792 struct list_head *list;
795 * Remove pages from lists in a round-robin fashion. A
796 * batch_free count is maintained that is incremented when an
797 * empty list is encountered. This is so more pages are freed
798 * off fuller lists instead of spinning excessively around empty
803 if (++migratetype == MIGRATE_PCPTYPES)
805 list = &pcp->lists[migratetype];
806 } while (list_empty(list));
808 /* This is the only non-empty list. Free them all. */
809 if (batch_free == MIGRATE_PCPTYPES)
810 batch_free = to_free;
813 int mt; /* migratetype of the to-be-freed page */
815 page = list_entry(list->prev, struct page, lru);
816 /* must delete as __free_one_page list manipulates */
817 list_del(&page->lru);
819 mt = get_pcppage_migratetype(page);
820 /* MIGRATE_ISOLATE page should not go to pcplists */
821 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
822 /* Pageblock could have been isolated meanwhile */
823 if (unlikely(has_isolate_pageblock(zone)))
824 mt = get_pageblock_migratetype(page);
826 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
827 trace_mm_page_pcpu_drain(page, 0, mt);
828 } while (--to_free && --batch_free && !list_empty(list));
830 spin_unlock(&zone->lock);
833 static void free_one_page(struct zone *zone,
834 struct page *page, unsigned long pfn,
838 unsigned long nr_scanned;
839 spin_lock(&zone->lock);
840 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
842 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
844 if (unlikely(has_isolate_pageblock(zone) ||
845 is_migrate_isolate(migratetype))) {
846 migratetype = get_pfnblock_migratetype(page, pfn);
848 __free_one_page(page, pfn, zone, order, migratetype);
849 spin_unlock(&zone->lock);
852 static int free_tail_pages_check(struct page *head_page, struct page *page)
857 * We rely page->lru.next never has bit 0 set, unless the page
858 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
860 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
862 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
866 if (unlikely(!PageTail(page))) {
867 bad_page(page, "PageTail not set", 0);
870 if (unlikely(compound_head(page) != head_page)) {
871 bad_page(page, "compound_head not consistent", 0);
876 clear_compound_head(page);
880 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
881 unsigned long zone, int nid)
883 set_page_links(page, zone, nid, pfn);
884 init_page_count(page);
885 page_mapcount_reset(page);
886 page_cpupid_reset_last(page);
888 INIT_LIST_HEAD(&page->lru);
889 #ifdef WANT_PAGE_VIRTUAL
890 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
891 if (!is_highmem_idx(zone))
892 set_page_address(page, __va(pfn << PAGE_SHIFT));
896 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
899 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
902 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
903 static void init_reserved_page(unsigned long pfn)
908 if (!early_page_uninitialised(pfn))
911 nid = early_pfn_to_nid(pfn);
912 pgdat = NODE_DATA(nid);
914 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
915 struct zone *zone = &pgdat->node_zones[zid];
917 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
920 __init_single_pfn(pfn, zid, nid);
923 static inline void init_reserved_page(unsigned long pfn)
925 /* Avoid false-positive PageTail() */
926 INIT_LIST_HEAD(&pfn_to_page(pfn)->lru);
928 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
931 * Initialised pages do not have PageReserved set. This function is
932 * called for each range allocated by the bootmem allocator and
933 * marks the pages PageReserved. The remaining valid pages are later
934 * sent to the buddy page allocator.
936 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
938 unsigned long start_pfn = PFN_DOWN(start);
939 unsigned long end_pfn = PFN_UP(end);
941 for (; start_pfn < end_pfn; start_pfn++) {
942 if (pfn_valid(start_pfn)) {
943 struct page *page = pfn_to_page(start_pfn);
945 init_reserved_page(start_pfn);
946 SetPageReserved(page);
951 static bool free_pages_prepare(struct page *page, unsigned int order)
953 bool compound = PageCompound(page);
956 VM_BUG_ON_PAGE(PageTail(page), page);
957 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
959 trace_mm_page_free(page, order);
960 kmemcheck_free_shadow(page, order);
961 kasan_free_pages(page, order);
964 page->mapping = NULL;
965 bad += free_pages_check(page);
966 for (i = 1; i < (1 << order); i++) {
968 bad += free_tail_pages_check(page, page + i);
969 bad += free_pages_check(page + i);
974 reset_page_owner(page, order);
976 if (!PageHighMem(page)) {
977 debug_check_no_locks_freed(page_address(page),
979 debug_check_no_obj_freed(page_address(page),
982 arch_free_page(page, order);
983 kernel_map_pages(page, 1 << order, 0);
988 static void __free_pages_ok(struct page *page, unsigned int order)
992 unsigned long pfn = page_to_pfn(page);
994 if (!free_pages_prepare(page, order))
997 migratetype = get_pfnblock_migratetype(page, pfn);
998 local_irq_save(flags);
999 __count_vm_events(PGFREE, 1 << order);
1000 free_one_page(page_zone(page), page, pfn, order, migratetype);
1001 local_irq_restore(flags);
1004 static void __init __free_pages_boot_core(struct page *page,
1005 unsigned long pfn, unsigned int order)
1007 unsigned int nr_pages = 1 << order;
1008 struct page *p = page;
1012 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1014 __ClearPageReserved(p);
1015 set_page_count(p, 0);
1017 __ClearPageReserved(p);
1018 set_page_count(p, 0);
1020 page_zone(page)->managed_pages += nr_pages;
1021 set_page_refcounted(page);
1022 __free_pages(page, order);
1025 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1026 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1028 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1030 int __meminit early_pfn_to_nid(unsigned long pfn)
1032 static DEFINE_SPINLOCK(early_pfn_lock);
1035 spin_lock(&early_pfn_lock);
1036 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1039 spin_unlock(&early_pfn_lock);
1045 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1046 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1047 struct mminit_pfnnid_cache *state)
1051 nid = __early_pfn_to_nid(pfn, state);
1052 if (nid >= 0 && nid != node)
1057 /* Only safe to use early in boot when initialisation is single-threaded */
1058 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1060 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1065 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1069 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1070 struct mminit_pfnnid_cache *state)
1077 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1080 if (early_page_uninitialised(pfn))
1082 return __free_pages_boot_core(page, pfn, order);
1085 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1086 static void __init deferred_free_range(struct page *page,
1087 unsigned long pfn, int nr_pages)
1094 /* Free a large naturally-aligned chunk if possible */
1095 if (nr_pages == MAX_ORDER_NR_PAGES &&
1096 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1097 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1098 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1102 for (i = 0; i < nr_pages; i++, page++, pfn++)
1103 __free_pages_boot_core(page, pfn, 0);
1106 /* Completion tracking for deferred_init_memmap() threads */
1107 static atomic_t pgdat_init_n_undone __initdata;
1108 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1110 static inline void __init pgdat_init_report_one_done(void)
1112 if (atomic_dec_and_test(&pgdat_init_n_undone))
1113 complete(&pgdat_init_all_done_comp);
1116 /* Initialise remaining memory on a node */
1117 static int __init deferred_init_memmap(void *data)
1119 pg_data_t *pgdat = data;
1120 int nid = pgdat->node_id;
1121 struct mminit_pfnnid_cache nid_init_state = { };
1122 unsigned long start = jiffies;
1123 unsigned long nr_pages = 0;
1124 unsigned long walk_start, walk_end;
1127 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1128 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1130 if (first_init_pfn == ULONG_MAX) {
1131 pgdat_init_report_one_done();
1135 /* Bind memory initialisation thread to a local node if possible */
1136 if (!cpumask_empty(cpumask))
1137 set_cpus_allowed_ptr(current, cpumask);
1139 /* Sanity check boundaries */
1140 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1141 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1142 pgdat->first_deferred_pfn = ULONG_MAX;
1144 /* Only the highest zone is deferred so find it */
1145 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1146 zone = pgdat->node_zones + zid;
1147 if (first_init_pfn < zone_end_pfn(zone))
1151 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1152 unsigned long pfn, end_pfn;
1153 struct page *page = NULL;
1154 struct page *free_base_page = NULL;
1155 unsigned long free_base_pfn = 0;
1158 end_pfn = min(walk_end, zone_end_pfn(zone));
1159 pfn = first_init_pfn;
1160 if (pfn < walk_start)
1162 if (pfn < zone->zone_start_pfn)
1163 pfn = zone->zone_start_pfn;
1165 for (; pfn < end_pfn; pfn++) {
1166 if (!pfn_valid_within(pfn))
1170 * Ensure pfn_valid is checked every
1171 * MAX_ORDER_NR_PAGES for memory holes
1173 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1174 if (!pfn_valid(pfn)) {
1180 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1185 /* Minimise pfn page lookups and scheduler checks */
1186 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1189 nr_pages += nr_to_free;
1190 deferred_free_range(free_base_page,
1191 free_base_pfn, nr_to_free);
1192 free_base_page = NULL;
1193 free_base_pfn = nr_to_free = 0;
1195 page = pfn_to_page(pfn);
1200 VM_BUG_ON(page_zone(page) != zone);
1204 __init_single_page(page, pfn, zid, nid);
1205 if (!free_base_page) {
1206 free_base_page = page;
1207 free_base_pfn = pfn;
1212 /* Where possible, batch up pages for a single free */
1215 /* Free the current block of pages to allocator */
1216 nr_pages += nr_to_free;
1217 deferred_free_range(free_base_page, free_base_pfn,
1219 free_base_page = NULL;
1220 free_base_pfn = nr_to_free = 0;
1223 first_init_pfn = max(end_pfn, first_init_pfn);
1226 /* Sanity check that the next zone really is unpopulated */
1227 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1229 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1230 jiffies_to_msecs(jiffies - start));
1232 pgdat_init_report_one_done();
1236 void __init page_alloc_init_late(void)
1240 /* There will be num_node_state(N_MEMORY) threads */
1241 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1242 for_each_node_state(nid, N_MEMORY) {
1243 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1246 /* Block until all are initialised */
1247 wait_for_completion(&pgdat_init_all_done_comp);
1249 /* Reinit limits that are based on free pages after the kernel is up */
1250 files_maxfiles_init();
1252 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1255 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1256 void __init init_cma_reserved_pageblock(struct page *page)
1258 unsigned i = pageblock_nr_pages;
1259 struct page *p = page;
1262 __ClearPageReserved(p);
1263 set_page_count(p, 0);
1266 set_pageblock_migratetype(page, MIGRATE_CMA);
1268 if (pageblock_order >= MAX_ORDER) {
1269 i = pageblock_nr_pages;
1272 set_page_refcounted(p);
1273 __free_pages(p, MAX_ORDER - 1);
1274 p += MAX_ORDER_NR_PAGES;
1275 } while (i -= MAX_ORDER_NR_PAGES);
1277 set_page_refcounted(page);
1278 __free_pages(page, pageblock_order);
1281 adjust_managed_page_count(page, pageblock_nr_pages);
1286 * The order of subdivision here is critical for the IO subsystem.
1287 * Please do not alter this order without good reasons and regression
1288 * testing. Specifically, as large blocks of memory are subdivided,
1289 * the order in which smaller blocks are delivered depends on the order
1290 * they're subdivided in this function. This is the primary factor
1291 * influencing the order in which pages are delivered to the IO
1292 * subsystem according to empirical testing, and this is also justified
1293 * by considering the behavior of a buddy system containing a single
1294 * large block of memory acted on by a series of small allocations.
1295 * This behavior is a critical factor in sglist merging's success.
1299 static inline void expand(struct zone *zone, struct page *page,
1300 int low, int high, struct free_area *area,
1303 unsigned long size = 1 << high;
1305 while (high > low) {
1309 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1311 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1312 debug_guardpage_enabled() &&
1313 high < debug_guardpage_minorder()) {
1315 * Mark as guard pages (or page), that will allow to
1316 * merge back to allocator when buddy will be freed.
1317 * Corresponding page table entries will not be touched,
1318 * pages will stay not present in virtual address space
1320 set_page_guard(zone, &page[size], high, migratetype);
1323 list_add(&page[size].lru, &area->free_list[migratetype]);
1325 set_page_order(&page[size], high);
1330 * This page is about to be returned from the page allocator
1332 static inline int check_new_page(struct page *page)
1334 const char *bad_reason = NULL;
1335 unsigned long bad_flags = 0;
1337 if (unlikely(page_mapcount(page)))
1338 bad_reason = "nonzero mapcount";
1339 if (unlikely(page->mapping != NULL))
1340 bad_reason = "non-NULL mapping";
1341 if (unlikely(atomic_read(&page->_count) != 0))
1342 bad_reason = "nonzero _count";
1343 if (unlikely(page->flags & __PG_HWPOISON)) {
1344 bad_reason = "HWPoisoned (hardware-corrupted)";
1345 bad_flags = __PG_HWPOISON;
1347 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1348 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1349 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1352 if (unlikely(page->mem_cgroup))
1353 bad_reason = "page still charged to cgroup";
1355 if (unlikely(bad_reason)) {
1356 bad_page(page, bad_reason, bad_flags);
1362 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1367 for (i = 0; i < (1 << order); i++) {
1368 struct page *p = page + i;
1369 if (unlikely(check_new_page(p)))
1373 set_page_private(page, 0);
1374 set_page_refcounted(page);
1376 arch_alloc_page(page, order);
1377 kernel_map_pages(page, 1 << order, 1);
1378 kasan_alloc_pages(page, order);
1380 if (gfp_flags & __GFP_ZERO)
1381 for (i = 0; i < (1 << order); i++)
1382 clear_highpage(page + i);
1384 if (order && (gfp_flags & __GFP_COMP))
1385 prep_compound_page(page, order);
1387 set_page_owner(page, order, gfp_flags);
1390 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1391 * allocate the page. The expectation is that the caller is taking
1392 * steps that will free more memory. The caller should avoid the page
1393 * being used for !PFMEMALLOC purposes.
1395 if (alloc_flags & ALLOC_NO_WATERMARKS)
1396 set_page_pfmemalloc(page);
1398 clear_page_pfmemalloc(page);
1404 * Go through the free lists for the given migratetype and remove
1405 * the smallest available page from the freelists
1408 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1411 unsigned int current_order;
1412 struct free_area *area;
1415 /* Find a page of the appropriate size in the preferred list */
1416 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1417 area = &(zone->free_area[current_order]);
1418 if (list_empty(&area->free_list[migratetype]))
1421 page = list_entry(area->free_list[migratetype].next,
1423 list_del(&page->lru);
1424 rmv_page_order(page);
1426 expand(zone, page, order, current_order, area, migratetype);
1427 set_pcppage_migratetype(page, migratetype);
1436 * This array describes the order lists are fallen back to when
1437 * the free lists for the desirable migrate type are depleted
1439 static int fallbacks[MIGRATE_TYPES][4] = {
1440 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1441 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1442 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1444 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1446 #ifdef CONFIG_MEMORY_ISOLATION
1447 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1452 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1455 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1458 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1459 unsigned int order) { return NULL; }
1463 * Move the free pages in a range to the free lists of the requested type.
1464 * Note that start_page and end_pages are not aligned on a pageblock
1465 * boundary. If alignment is required, use move_freepages_block()
1467 int move_freepages(struct zone *zone,
1468 struct page *start_page, struct page *end_page,
1472 unsigned long order;
1473 int pages_moved = 0;
1475 #ifndef CONFIG_HOLES_IN_ZONE
1477 * page_zone is not safe to call in this context when
1478 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1479 * anyway as we check zone boundaries in move_freepages_block().
1480 * Remove at a later date when no bug reports exist related to
1481 * grouping pages by mobility
1483 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1486 for (page = start_page; page <= end_page;) {
1487 /* Make sure we are not inadvertently changing nodes */
1488 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1490 if (!pfn_valid_within(page_to_pfn(page))) {
1495 if (!PageBuddy(page)) {
1500 order = page_order(page);
1501 list_move(&page->lru,
1502 &zone->free_area[order].free_list[migratetype]);
1504 pages_moved += 1 << order;
1510 int move_freepages_block(struct zone *zone, struct page *page,
1513 unsigned long start_pfn, end_pfn;
1514 struct page *start_page, *end_page;
1516 start_pfn = page_to_pfn(page);
1517 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1518 start_page = pfn_to_page(start_pfn);
1519 end_page = start_page + pageblock_nr_pages - 1;
1520 end_pfn = start_pfn + pageblock_nr_pages - 1;
1522 /* Do not cross zone boundaries */
1523 if (!zone_spans_pfn(zone, start_pfn))
1525 if (!zone_spans_pfn(zone, end_pfn))
1528 return move_freepages(zone, start_page, end_page, migratetype);
1531 static void change_pageblock_range(struct page *pageblock_page,
1532 int start_order, int migratetype)
1534 int nr_pageblocks = 1 << (start_order - pageblock_order);
1536 while (nr_pageblocks--) {
1537 set_pageblock_migratetype(pageblock_page, migratetype);
1538 pageblock_page += pageblock_nr_pages;
1543 * When we are falling back to another migratetype during allocation, try to
1544 * steal extra free pages from the same pageblocks to satisfy further
1545 * allocations, instead of polluting multiple pageblocks.
1547 * If we are stealing a relatively large buddy page, it is likely there will
1548 * be more free pages in the pageblock, so try to steal them all. For
1549 * reclaimable and unmovable allocations, we steal regardless of page size,
1550 * as fragmentation caused by those allocations polluting movable pageblocks
1551 * is worse than movable allocations stealing from unmovable and reclaimable
1554 static bool can_steal_fallback(unsigned int order, int start_mt)
1557 * Leaving this order check is intended, although there is
1558 * relaxed order check in next check. The reason is that
1559 * we can actually steal whole pageblock if this condition met,
1560 * but, below check doesn't guarantee it and that is just heuristic
1561 * so could be changed anytime.
1563 if (order >= pageblock_order)
1566 if (order >= pageblock_order / 2 ||
1567 start_mt == MIGRATE_RECLAIMABLE ||
1568 start_mt == MIGRATE_UNMOVABLE ||
1569 page_group_by_mobility_disabled)
1576 * This function implements actual steal behaviour. If order is large enough,
1577 * we can steal whole pageblock. If not, we first move freepages in this
1578 * pageblock and check whether half of pages are moved or not. If half of
1579 * pages are moved, we can change migratetype of pageblock and permanently
1580 * use it's pages as requested migratetype in the future.
1582 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1585 int current_order = page_order(page);
1588 /* Take ownership for orders >= pageblock_order */
1589 if (current_order >= pageblock_order) {
1590 change_pageblock_range(page, current_order, start_type);
1594 pages = move_freepages_block(zone, page, start_type);
1596 /* Claim the whole block if over half of it is free */
1597 if (pages >= (1 << (pageblock_order-1)) ||
1598 page_group_by_mobility_disabled)
1599 set_pageblock_migratetype(page, start_type);
1603 * Check whether there is a suitable fallback freepage with requested order.
1604 * If only_stealable is true, this function returns fallback_mt only if
1605 * we can steal other freepages all together. This would help to reduce
1606 * fragmentation due to mixed migratetype pages in one pageblock.
1608 int find_suitable_fallback(struct free_area *area, unsigned int order,
1609 int migratetype, bool only_stealable, bool *can_steal)
1614 if (area->nr_free == 0)
1619 fallback_mt = fallbacks[migratetype][i];
1620 if (fallback_mt == MIGRATE_TYPES)
1623 if (list_empty(&area->free_list[fallback_mt]))
1626 if (can_steal_fallback(order, migratetype))
1629 if (!only_stealable)
1640 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1641 * there are no empty page blocks that contain a page with a suitable order
1643 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1644 unsigned int alloc_order)
1647 unsigned long max_managed, flags;
1650 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1651 * Check is race-prone but harmless.
1653 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1654 if (zone->nr_reserved_highatomic >= max_managed)
1657 spin_lock_irqsave(&zone->lock, flags);
1659 /* Recheck the nr_reserved_highatomic limit under the lock */
1660 if (zone->nr_reserved_highatomic >= max_managed)
1664 mt = get_pageblock_migratetype(page);
1665 if (mt != MIGRATE_HIGHATOMIC &&
1666 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1667 zone->nr_reserved_highatomic += pageblock_nr_pages;
1668 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1669 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1673 spin_unlock_irqrestore(&zone->lock, flags);
1677 * Used when an allocation is about to fail under memory pressure. This
1678 * potentially hurts the reliability of high-order allocations when under
1679 * intense memory pressure but failed atomic allocations should be easier
1680 * to recover from than an OOM.
1682 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1684 struct zonelist *zonelist = ac->zonelist;
1685 unsigned long flags;
1691 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1693 /* Preserve at least one pageblock */
1694 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1697 spin_lock_irqsave(&zone->lock, flags);
1698 for (order = 0; order < MAX_ORDER; order++) {
1699 struct free_area *area = &(zone->free_area[order]);
1701 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1704 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1708 * It should never happen but changes to locking could
1709 * inadvertently allow a per-cpu drain to add pages
1710 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1711 * and watch for underflows.
1713 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1714 zone->nr_reserved_highatomic);
1717 * Convert to ac->migratetype and avoid the normal
1718 * pageblock stealing heuristics. Minimally, the caller
1719 * is doing the work and needs the pages. More
1720 * importantly, if the block was always converted to
1721 * MIGRATE_UNMOVABLE or another type then the number
1722 * of pageblocks that cannot be completely freed
1725 set_pageblock_migratetype(page, ac->migratetype);
1726 move_freepages_block(zone, page, ac->migratetype);
1727 spin_unlock_irqrestore(&zone->lock, flags);
1730 spin_unlock_irqrestore(&zone->lock, flags);
1734 /* Remove an element from the buddy allocator from the fallback list */
1735 static inline struct page *
1736 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1738 struct free_area *area;
1739 unsigned int current_order;
1744 /* Find the largest possible block of pages in the other list */
1745 for (current_order = MAX_ORDER-1;
1746 current_order >= order && current_order <= MAX_ORDER-1;
1748 area = &(zone->free_area[current_order]);
1749 fallback_mt = find_suitable_fallback(area, current_order,
1750 start_migratetype, false, &can_steal);
1751 if (fallback_mt == -1)
1754 page = list_entry(area->free_list[fallback_mt].next,
1757 steal_suitable_fallback(zone, page, start_migratetype);
1759 /* Remove the page from the freelists */
1761 list_del(&page->lru);
1762 rmv_page_order(page);
1764 expand(zone, page, order, current_order, area,
1767 * The pcppage_migratetype may differ from pageblock's
1768 * migratetype depending on the decisions in
1769 * find_suitable_fallback(). This is OK as long as it does not
1770 * differ for MIGRATE_CMA pageblocks. Those can be used as
1771 * fallback only via special __rmqueue_cma_fallback() function
1773 set_pcppage_migratetype(page, start_migratetype);
1775 trace_mm_page_alloc_extfrag(page, order, current_order,
1776 start_migratetype, fallback_mt);
1785 * Do the hard work of removing an element from the buddy allocator.
1786 * Call me with the zone->lock already held.
1788 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1789 int migratetype, gfp_t gfp_flags)
1793 page = __rmqueue_smallest(zone, order, migratetype);
1794 if (unlikely(!page)) {
1795 if (migratetype == MIGRATE_MOVABLE)
1796 page = __rmqueue_cma_fallback(zone, order);
1799 page = __rmqueue_fallback(zone, order, migratetype);
1802 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1807 * Obtain a specified number of elements from the buddy allocator, all under
1808 * a single hold of the lock, for efficiency. Add them to the supplied list.
1809 * Returns the number of new pages which were placed at *list.
1811 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1812 unsigned long count, struct list_head *list,
1813 int migratetype, bool cold)
1817 spin_lock(&zone->lock);
1818 for (i = 0; i < count; ++i) {
1819 struct page *page = __rmqueue(zone, order, migratetype, 0);
1820 if (unlikely(page == NULL))
1824 * Split buddy pages returned by expand() are received here
1825 * in physical page order. The page is added to the callers and
1826 * list and the list head then moves forward. From the callers
1827 * perspective, the linked list is ordered by page number in
1828 * some conditions. This is useful for IO devices that can
1829 * merge IO requests if the physical pages are ordered
1833 list_add(&page->lru, list);
1835 list_add_tail(&page->lru, list);
1837 if (is_migrate_cma(get_pcppage_migratetype(page)))
1838 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1841 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1842 spin_unlock(&zone->lock);
1848 * Called from the vmstat counter updater to drain pagesets of this
1849 * currently executing processor on remote nodes after they have
1852 * Note that this function must be called with the thread pinned to
1853 * a single processor.
1855 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1857 unsigned long flags;
1858 int to_drain, batch;
1860 local_irq_save(flags);
1861 batch = READ_ONCE(pcp->batch);
1862 to_drain = min(pcp->count, batch);
1864 free_pcppages_bulk(zone, to_drain, pcp);
1865 pcp->count -= to_drain;
1867 local_irq_restore(flags);
1872 * Drain pcplists of the indicated processor and zone.
1874 * The processor must either be the current processor and the
1875 * thread pinned to the current processor or a processor that
1878 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1880 unsigned long flags;
1881 struct per_cpu_pageset *pset;
1882 struct per_cpu_pages *pcp;
1884 local_irq_save(flags);
1885 pset = per_cpu_ptr(zone->pageset, cpu);
1889 free_pcppages_bulk(zone, pcp->count, pcp);
1892 local_irq_restore(flags);
1896 * Drain pcplists of all zones on the indicated processor.
1898 * The processor must either be the current processor and the
1899 * thread pinned to the current processor or a processor that
1902 static void drain_pages(unsigned int cpu)
1906 for_each_populated_zone(zone) {
1907 drain_pages_zone(cpu, zone);
1912 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1914 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1915 * the single zone's pages.
1917 void drain_local_pages(struct zone *zone)
1919 int cpu = smp_processor_id();
1922 drain_pages_zone(cpu, zone);
1928 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1930 * When zone parameter is non-NULL, spill just the single zone's pages.
1932 * Note that this code is protected against sending an IPI to an offline
1933 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1934 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1935 * nothing keeps CPUs from showing up after we populated the cpumask and
1936 * before the call to on_each_cpu_mask().
1938 void drain_all_pages(struct zone *zone)
1943 * Allocate in the BSS so we wont require allocation in
1944 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1946 static cpumask_t cpus_with_pcps;
1949 * We don't care about racing with CPU hotplug event
1950 * as offline notification will cause the notified
1951 * cpu to drain that CPU pcps and on_each_cpu_mask
1952 * disables preemption as part of its processing
1954 for_each_online_cpu(cpu) {
1955 struct per_cpu_pageset *pcp;
1957 bool has_pcps = false;
1960 pcp = per_cpu_ptr(zone->pageset, cpu);
1964 for_each_populated_zone(z) {
1965 pcp = per_cpu_ptr(z->pageset, cpu);
1966 if (pcp->pcp.count) {
1974 cpumask_set_cpu(cpu, &cpus_with_pcps);
1976 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1978 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1982 #ifdef CONFIG_HIBERNATION
1984 void mark_free_pages(struct zone *zone)
1986 unsigned long pfn, max_zone_pfn;
1987 unsigned long flags;
1988 unsigned int order, t;
1989 struct list_head *curr;
1991 if (zone_is_empty(zone))
1994 spin_lock_irqsave(&zone->lock, flags);
1996 max_zone_pfn = zone_end_pfn(zone);
1997 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1998 if (pfn_valid(pfn)) {
1999 struct page *page = pfn_to_page(pfn);
2001 if (!swsusp_page_is_forbidden(page))
2002 swsusp_unset_page_free(page);
2005 for_each_migratetype_order(order, t) {
2006 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2009 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2010 for (i = 0; i < (1UL << order); i++)
2011 swsusp_set_page_free(pfn_to_page(pfn + i));
2014 spin_unlock_irqrestore(&zone->lock, flags);
2016 #endif /* CONFIG_PM */
2019 * Free a 0-order page
2020 * cold == true ? free a cold page : free a hot page
2022 void free_hot_cold_page(struct page *page, bool cold)
2024 struct zone *zone = page_zone(page);
2025 struct per_cpu_pages *pcp;
2026 unsigned long flags;
2027 unsigned long pfn = page_to_pfn(page);
2030 if (!free_pages_prepare(page, 0))
2033 migratetype = get_pfnblock_migratetype(page, pfn);
2034 set_pcppage_migratetype(page, migratetype);
2035 local_irq_save(flags);
2036 __count_vm_event(PGFREE);
2039 * We only track unmovable, reclaimable and movable on pcp lists.
2040 * Free ISOLATE pages back to the allocator because they are being
2041 * offlined but treat RESERVE as movable pages so we can get those
2042 * areas back if necessary. Otherwise, we may have to free
2043 * excessively into the page allocator
2045 if (migratetype >= MIGRATE_PCPTYPES) {
2046 if (unlikely(is_migrate_isolate(migratetype))) {
2047 free_one_page(zone, page, pfn, 0, migratetype);
2050 migratetype = MIGRATE_MOVABLE;
2053 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2055 list_add(&page->lru, &pcp->lists[migratetype]);
2057 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2059 if (pcp->count >= pcp->high) {
2060 unsigned long batch = READ_ONCE(pcp->batch);
2061 free_pcppages_bulk(zone, batch, pcp);
2062 pcp->count -= batch;
2066 local_irq_restore(flags);
2070 * Free a list of 0-order pages
2072 void free_hot_cold_page_list(struct list_head *list, bool cold)
2074 struct page *page, *next;
2076 list_for_each_entry_safe(page, next, list, lru) {
2077 trace_mm_page_free_batched(page, cold);
2078 free_hot_cold_page(page, cold);
2083 * split_page takes a non-compound higher-order page, and splits it into
2084 * n (1<<order) sub-pages: page[0..n]
2085 * Each sub-page must be freed individually.
2087 * Note: this is probably too low level an operation for use in drivers.
2088 * Please consult with lkml before using this in your driver.
2090 void split_page(struct page *page, unsigned int order)
2095 VM_BUG_ON_PAGE(PageCompound(page), page);
2096 VM_BUG_ON_PAGE(!page_count(page), page);
2098 #ifdef CONFIG_KMEMCHECK
2100 * Split shadow pages too, because free(page[0]) would
2101 * otherwise free the whole shadow.
2103 if (kmemcheck_page_is_tracked(page))
2104 split_page(virt_to_page(page[0].shadow), order);
2107 gfp_mask = get_page_owner_gfp(page);
2108 set_page_owner(page, 0, gfp_mask);
2109 for (i = 1; i < (1 << order); i++) {
2110 set_page_refcounted(page + i);
2111 set_page_owner(page + i, 0, gfp_mask);
2114 EXPORT_SYMBOL_GPL(split_page);
2116 int __isolate_free_page(struct page *page, unsigned int order)
2118 unsigned long watermark;
2122 BUG_ON(!PageBuddy(page));
2124 zone = page_zone(page);
2125 mt = get_pageblock_migratetype(page);
2127 if (!is_migrate_isolate(mt)) {
2128 /* Obey watermarks as if the page was being allocated */
2129 watermark = low_wmark_pages(zone) + (1 << order);
2130 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2133 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2136 /* Remove page from free list */
2137 list_del(&page->lru);
2138 zone->free_area[order].nr_free--;
2139 rmv_page_order(page);
2141 set_page_owner(page, order, __GFP_MOVABLE);
2143 /* Set the pageblock if the isolated page is at least a pageblock */
2144 if (order >= pageblock_order - 1) {
2145 struct page *endpage = page + (1 << order) - 1;
2146 for (; page < endpage; page += pageblock_nr_pages) {
2147 int mt = get_pageblock_migratetype(page);
2148 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2149 set_pageblock_migratetype(page,
2155 return 1UL << order;
2159 * Similar to split_page except the page is already free. As this is only
2160 * being used for migration, the migratetype of the block also changes.
2161 * As this is called with interrupts disabled, the caller is responsible
2162 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2165 * Note: this is probably too low level an operation for use in drivers.
2166 * Please consult with lkml before using this in your driver.
2168 int split_free_page(struct page *page)
2173 order = page_order(page);
2175 nr_pages = __isolate_free_page(page, order);
2179 /* Split into individual pages */
2180 set_page_refcounted(page);
2181 split_page(page, order);
2186 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2189 struct page *buffered_rmqueue(struct zone *preferred_zone,
2190 struct zone *zone, unsigned int order,
2191 gfp_t gfp_flags, int alloc_flags, int migratetype)
2193 unsigned long flags;
2195 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2197 if (likely(order == 0)) {
2198 struct per_cpu_pages *pcp;
2199 struct list_head *list;
2201 local_irq_save(flags);
2202 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2203 list = &pcp->lists[migratetype];
2204 if (list_empty(list)) {
2205 pcp->count += rmqueue_bulk(zone, 0,
2208 if (unlikely(list_empty(list)))
2213 page = list_entry(list->prev, struct page, lru);
2215 page = list_entry(list->next, struct page, lru);
2217 list_del(&page->lru);
2220 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2222 * __GFP_NOFAIL is not to be used in new code.
2224 * All __GFP_NOFAIL callers should be fixed so that they
2225 * properly detect and handle allocation failures.
2227 * We most definitely don't want callers attempting to
2228 * allocate greater than order-1 page units with
2231 WARN_ON_ONCE(order > 1);
2233 spin_lock_irqsave(&zone->lock, flags);
2236 if (alloc_flags & ALLOC_HARDER) {
2237 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2239 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2242 page = __rmqueue(zone, order, migratetype, gfp_flags);
2243 spin_unlock(&zone->lock);
2246 __mod_zone_freepage_state(zone, -(1 << order),
2247 get_pcppage_migratetype(page));
2250 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2251 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2252 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2253 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2255 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2256 zone_statistics(preferred_zone, zone, gfp_flags);
2257 local_irq_restore(flags);
2259 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2263 local_irq_restore(flags);
2267 #ifdef CONFIG_FAIL_PAGE_ALLOC
2270 struct fault_attr attr;
2272 u32 ignore_gfp_highmem;
2273 u32 ignore_gfp_reclaim;
2275 } fail_page_alloc = {
2276 .attr = FAULT_ATTR_INITIALIZER,
2277 .ignore_gfp_reclaim = 1,
2278 .ignore_gfp_highmem = 1,
2282 static int __init setup_fail_page_alloc(char *str)
2284 return setup_fault_attr(&fail_page_alloc.attr, str);
2286 __setup("fail_page_alloc=", setup_fail_page_alloc);
2288 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2290 if (order < fail_page_alloc.min_order)
2292 if (gfp_mask & __GFP_NOFAIL)
2294 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2296 if (fail_page_alloc.ignore_gfp_reclaim &&
2297 (gfp_mask & __GFP_DIRECT_RECLAIM))
2300 return should_fail(&fail_page_alloc.attr, 1 << order);
2303 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2305 static int __init fail_page_alloc_debugfs(void)
2307 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2310 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2311 &fail_page_alloc.attr);
2313 return PTR_ERR(dir);
2315 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2316 &fail_page_alloc.ignore_gfp_reclaim))
2318 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2319 &fail_page_alloc.ignore_gfp_highmem))
2321 if (!debugfs_create_u32("min-order", mode, dir,
2322 &fail_page_alloc.min_order))
2327 debugfs_remove_recursive(dir);
2332 late_initcall(fail_page_alloc_debugfs);
2334 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2336 #else /* CONFIG_FAIL_PAGE_ALLOC */
2338 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2343 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2346 * Return true if free base pages are above 'mark'. For high-order checks it
2347 * will return true of the order-0 watermark is reached and there is at least
2348 * one free page of a suitable size. Checking now avoids taking the zone lock
2349 * to check in the allocation paths if no pages are free.
2351 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2352 unsigned long mark, int classzone_idx, int alloc_flags,
2357 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2359 /* free_pages may go negative - that's OK */
2360 free_pages -= (1 << order) - 1;
2362 if (alloc_flags & ALLOC_HIGH)
2366 * If the caller does not have rights to ALLOC_HARDER then subtract
2367 * the high-atomic reserves. This will over-estimate the size of the
2368 * atomic reserve but it avoids a search.
2370 if (likely(!alloc_harder))
2371 free_pages -= z->nr_reserved_highatomic;
2376 /* If allocation can't use CMA areas don't use free CMA pages */
2377 if (!(alloc_flags & ALLOC_CMA))
2378 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2382 * Check watermarks for an order-0 allocation request. If these
2383 * are not met, then a high-order request also cannot go ahead
2384 * even if a suitable page happened to be free.
2386 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2389 /* If this is an order-0 request then the watermark is fine */
2393 /* For a high-order request, check at least one suitable page is free */
2394 for (o = order; o < MAX_ORDER; o++) {
2395 struct free_area *area = &z->free_area[o];
2404 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2405 if (!list_empty(&area->free_list[mt]))
2410 if ((alloc_flags & ALLOC_CMA) &&
2411 !list_empty(&area->free_list[MIGRATE_CMA])) {
2419 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2420 int classzone_idx, int alloc_flags)
2422 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2423 zone_page_state(z, NR_FREE_PAGES));
2426 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2427 unsigned long mark, int classzone_idx)
2429 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2431 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2432 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2434 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2439 static bool zone_local(struct zone *local_zone, struct zone *zone)
2441 return local_zone->node == zone->node;
2444 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2446 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2449 #else /* CONFIG_NUMA */
2450 static bool zone_local(struct zone *local_zone, struct zone *zone)
2455 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2459 #endif /* CONFIG_NUMA */
2461 static void reset_alloc_batches(struct zone *preferred_zone)
2463 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2466 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2467 high_wmark_pages(zone) - low_wmark_pages(zone) -
2468 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2469 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2470 } while (zone++ != preferred_zone);
2474 * get_page_from_freelist goes through the zonelist trying to allocate
2477 static struct page *
2478 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2479 const struct alloc_context *ac)
2481 struct zonelist *zonelist = ac->zonelist;
2483 struct page *page = NULL;
2485 int nr_fair_skipped = 0;
2486 bool zonelist_rescan;
2489 zonelist_rescan = false;
2492 * Scan zonelist, looking for a zone with enough free.
2493 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2495 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2499 if (cpusets_enabled() &&
2500 (alloc_flags & ALLOC_CPUSET) &&
2501 !cpuset_zone_allowed(zone, gfp_mask))
2504 * Distribute pages in proportion to the individual
2505 * zone size to ensure fair page aging. The zone a
2506 * page was allocated in should have no effect on the
2507 * time the page has in memory before being reclaimed.
2509 if (alloc_flags & ALLOC_FAIR) {
2510 if (!zone_local(ac->preferred_zone, zone))
2512 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2518 * When allocating a page cache page for writing, we
2519 * want to get it from a zone that is within its dirty
2520 * limit, such that no single zone holds more than its
2521 * proportional share of globally allowed dirty pages.
2522 * The dirty limits take into account the zone's
2523 * lowmem reserves and high watermark so that kswapd
2524 * should be able to balance it without having to
2525 * write pages from its LRU list.
2527 * This may look like it could increase pressure on
2528 * lower zones by failing allocations in higher zones
2529 * before they are full. But the pages that do spill
2530 * over are limited as the lower zones are protected
2531 * by this very same mechanism. It should not become
2532 * a practical burden to them.
2534 * XXX: For now, allow allocations to potentially
2535 * exceed the per-zone dirty limit in the slowpath
2536 * (spread_dirty_pages unset) before going into reclaim,
2537 * which is important when on a NUMA setup the allowed
2538 * zones are together not big enough to reach the
2539 * global limit. The proper fix for these situations
2540 * will require awareness of zones in the
2541 * dirty-throttling and the flusher threads.
2543 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2546 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2547 if (!zone_watermark_ok(zone, order, mark,
2548 ac->classzone_idx, alloc_flags)) {
2551 /* Checked here to keep the fast path fast */
2552 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2553 if (alloc_flags & ALLOC_NO_WATERMARKS)
2556 if (zone_reclaim_mode == 0 ||
2557 !zone_allows_reclaim(ac->preferred_zone, zone))
2560 ret = zone_reclaim(zone, gfp_mask, order);
2562 case ZONE_RECLAIM_NOSCAN:
2565 case ZONE_RECLAIM_FULL:
2566 /* scanned but unreclaimable */
2569 /* did we reclaim enough */
2570 if (zone_watermark_ok(zone, order, mark,
2571 ac->classzone_idx, alloc_flags))
2579 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2580 gfp_mask, alloc_flags, ac->migratetype);
2582 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2586 * If this is a high-order atomic allocation then check
2587 * if the pageblock should be reserved for the future
2589 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2590 reserve_highatomic_pageblock(page, zone, order);
2597 * The first pass makes sure allocations are spread fairly within the
2598 * local node. However, the local node might have free pages left
2599 * after the fairness batches are exhausted, and remote zones haven't
2600 * even been considered yet. Try once more without fairness, and
2601 * include remote zones now, before entering the slowpath and waking
2602 * kswapd: prefer spilling to a remote zone over swapping locally.
2604 if (alloc_flags & ALLOC_FAIR) {
2605 alloc_flags &= ~ALLOC_FAIR;
2606 if (nr_fair_skipped) {
2607 zonelist_rescan = true;
2608 reset_alloc_batches(ac->preferred_zone);
2610 if (nr_online_nodes > 1)
2611 zonelist_rescan = true;
2614 if (zonelist_rescan)
2621 * Large machines with many possible nodes should not always dump per-node
2622 * meminfo in irq context.
2624 static inline bool should_suppress_show_mem(void)
2629 ret = in_interrupt();
2634 static DEFINE_RATELIMIT_STATE(nopage_rs,
2635 DEFAULT_RATELIMIT_INTERVAL,
2636 DEFAULT_RATELIMIT_BURST);
2638 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2640 unsigned int filter = SHOW_MEM_FILTER_NODES;
2642 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2643 debug_guardpage_minorder() > 0)
2647 * This documents exceptions given to allocations in certain
2648 * contexts that are allowed to allocate outside current's set
2651 if (!(gfp_mask & __GFP_NOMEMALLOC))
2652 if (test_thread_flag(TIF_MEMDIE) ||
2653 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2654 filter &= ~SHOW_MEM_FILTER_NODES;
2655 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2656 filter &= ~SHOW_MEM_FILTER_NODES;
2659 struct va_format vaf;
2662 va_start(args, fmt);
2667 pr_warn("%pV", &vaf);
2672 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2673 current->comm, order, gfp_mask);
2676 if (!should_suppress_show_mem())
2680 static inline struct page *
2681 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2682 const struct alloc_context *ac, unsigned long *did_some_progress)
2684 struct oom_control oc = {
2685 .zonelist = ac->zonelist,
2686 .nodemask = ac->nodemask,
2687 .gfp_mask = gfp_mask,
2692 *did_some_progress = 0;
2695 * Acquire the oom lock. If that fails, somebody else is
2696 * making progress for us.
2698 if (!mutex_trylock(&oom_lock)) {
2699 *did_some_progress = 1;
2700 schedule_timeout_uninterruptible(1);
2705 * Go through the zonelist yet one more time, keep very high watermark
2706 * here, this is only to catch a parallel oom killing, we must fail if
2707 * we're still under heavy pressure.
2709 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2710 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2714 if (!(gfp_mask & __GFP_NOFAIL)) {
2715 /* Coredumps can quickly deplete all memory reserves */
2716 if (current->flags & PF_DUMPCORE)
2718 /* The OOM killer will not help higher order allocs */
2719 if (order > PAGE_ALLOC_COSTLY_ORDER)
2721 /* The OOM killer does not needlessly kill tasks for lowmem */
2722 if (ac->high_zoneidx < ZONE_NORMAL)
2724 /* The OOM killer does not compensate for IO-less reclaim */
2725 if (!(gfp_mask & __GFP_FS)) {
2727 * XXX: Page reclaim didn't yield anything,
2728 * and the OOM killer can't be invoked, but
2729 * keep looping as per tradition.
2731 *did_some_progress = 1;
2734 if (pm_suspended_storage())
2736 /* The OOM killer may not free memory on a specific node */
2737 if (gfp_mask & __GFP_THISNODE)
2740 /* Exhausted what can be done so it's blamo time */
2741 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2742 *did_some_progress = 1;
2744 mutex_unlock(&oom_lock);
2748 #ifdef CONFIG_COMPACTION
2749 /* Try memory compaction for high-order allocations before reclaim */
2750 static struct page *
2751 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2752 int alloc_flags, const struct alloc_context *ac,
2753 enum migrate_mode mode, int *contended_compaction,
2754 bool *deferred_compaction)
2756 unsigned long compact_result;
2762 current->flags |= PF_MEMALLOC;
2763 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2764 mode, contended_compaction);
2765 current->flags &= ~PF_MEMALLOC;
2767 switch (compact_result) {
2768 case COMPACT_DEFERRED:
2769 *deferred_compaction = true;
2771 case COMPACT_SKIPPED:
2778 * At least in one zone compaction wasn't deferred or skipped, so let's
2779 * count a compaction stall
2781 count_vm_event(COMPACTSTALL);
2783 page = get_page_from_freelist(gfp_mask, order,
2784 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2787 struct zone *zone = page_zone(page);
2789 zone->compact_blockskip_flush = false;
2790 compaction_defer_reset(zone, order, true);
2791 count_vm_event(COMPACTSUCCESS);
2796 * It's bad if compaction run occurs and fails. The most likely reason
2797 * is that pages exist, but not enough to satisfy watermarks.
2799 count_vm_event(COMPACTFAIL);
2806 static inline struct page *
2807 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2808 int alloc_flags, const struct alloc_context *ac,
2809 enum migrate_mode mode, int *contended_compaction,
2810 bool *deferred_compaction)
2814 #endif /* CONFIG_COMPACTION */
2816 /* Perform direct synchronous page reclaim */
2818 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2819 const struct alloc_context *ac)
2821 struct reclaim_state reclaim_state;
2826 /* We now go into synchronous reclaim */
2827 cpuset_memory_pressure_bump();
2828 current->flags |= PF_MEMALLOC;
2829 lockdep_set_current_reclaim_state(gfp_mask);
2830 reclaim_state.reclaimed_slab = 0;
2831 current->reclaim_state = &reclaim_state;
2833 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2836 current->reclaim_state = NULL;
2837 lockdep_clear_current_reclaim_state();
2838 current->flags &= ~PF_MEMALLOC;
2845 /* The really slow allocator path where we enter direct reclaim */
2846 static inline struct page *
2847 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2848 int alloc_flags, const struct alloc_context *ac,
2849 unsigned long *did_some_progress)
2851 struct page *page = NULL;
2852 bool drained = false;
2854 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2855 if (unlikely(!(*did_some_progress)))
2859 page = get_page_from_freelist(gfp_mask, order,
2860 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2863 * If an allocation failed after direct reclaim, it could be because
2864 * pages are pinned on the per-cpu lists or in high alloc reserves.
2865 * Shrink them them and try again
2867 if (!page && !drained) {
2868 unreserve_highatomic_pageblock(ac);
2869 drain_all_pages(NULL);
2878 * This is called in the allocator slow-path if the allocation request is of
2879 * sufficient urgency to ignore watermarks and take other desperate measures
2881 static inline struct page *
2882 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2883 const struct alloc_context *ac)
2888 page = get_page_from_freelist(gfp_mask, order,
2889 ALLOC_NO_WATERMARKS, ac);
2891 if (!page && gfp_mask & __GFP_NOFAIL)
2892 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2894 } while (!page && (gfp_mask & __GFP_NOFAIL));
2899 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2904 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2905 ac->high_zoneidx, ac->nodemask)
2906 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2910 gfp_to_alloc_flags(gfp_t gfp_mask)
2912 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2914 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2915 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2918 * The caller may dip into page reserves a bit more if the caller
2919 * cannot run direct reclaim, or if the caller has realtime scheduling
2920 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2921 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2923 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2925 if (gfp_mask & __GFP_ATOMIC) {
2927 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2928 * if it can't schedule.
2930 if (!(gfp_mask & __GFP_NOMEMALLOC))
2931 alloc_flags |= ALLOC_HARDER;
2933 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2934 * comment for __cpuset_node_allowed().
2936 alloc_flags &= ~ALLOC_CPUSET;
2937 } else if (unlikely(rt_task(current)) && !in_interrupt())
2938 alloc_flags |= ALLOC_HARDER;
2940 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2941 if (gfp_mask & __GFP_MEMALLOC)
2942 alloc_flags |= ALLOC_NO_WATERMARKS;
2943 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2944 alloc_flags |= ALLOC_NO_WATERMARKS;
2945 else if (!in_interrupt() &&
2946 ((current->flags & PF_MEMALLOC) ||
2947 unlikely(test_thread_flag(TIF_MEMDIE))))
2948 alloc_flags |= ALLOC_NO_WATERMARKS;
2951 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2952 alloc_flags |= ALLOC_CMA;
2957 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2959 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2962 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2964 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2967 static inline struct page *
2968 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2969 struct alloc_context *ac)
2971 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2972 struct page *page = NULL;
2974 unsigned long pages_reclaimed = 0;
2975 unsigned long did_some_progress;
2976 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2977 bool deferred_compaction = false;
2978 int contended_compaction = COMPACT_CONTENDED_NONE;
2981 * In the slowpath, we sanity check order to avoid ever trying to
2982 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2983 * be using allocators in order of preference for an area that is
2986 if (order >= MAX_ORDER) {
2987 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2992 * We also sanity check to catch abuse of atomic reserves being used by
2993 * callers that are not in atomic context.
2995 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
2996 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
2997 gfp_mask &= ~__GFP_ATOMIC;
3000 * If this allocation cannot block and it is for a specific node, then
3001 * fail early. There's no need to wakeup kswapd or retry for a
3002 * speculative node-specific allocation.
3004 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3008 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3009 wake_all_kswapds(order, ac);
3012 * OK, we're below the kswapd watermark and have kicked background
3013 * reclaim. Now things get more complex, so set up alloc_flags according
3014 * to how we want to proceed.
3016 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3019 * Find the true preferred zone if the allocation is unconstrained by
3022 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3023 struct zoneref *preferred_zoneref;
3024 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3025 ac->high_zoneidx, NULL, &ac->preferred_zone);
3026 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3029 /* This is the last chance, in general, before the goto nopage. */
3030 page = get_page_from_freelist(gfp_mask, order,
3031 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3035 /* Allocate without watermarks if the context allows */
3036 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3038 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3039 * the allocation is high priority and these type of
3040 * allocations are system rather than user orientated
3042 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3044 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3051 /* Caller is not willing to reclaim, we can't balance anything */
3052 if (!can_direct_reclaim) {
3054 * All existing users of the deprecated __GFP_NOFAIL are
3055 * blockable, so warn of any new users that actually allow this
3056 * type of allocation to fail.
3058 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3062 /* Avoid recursion of direct reclaim */
3063 if (current->flags & PF_MEMALLOC)
3066 /* Avoid allocations with no watermarks from looping endlessly */
3067 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3071 * Try direct compaction. The first pass is asynchronous. Subsequent
3072 * attempts after direct reclaim are synchronous
3074 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3076 &contended_compaction,
3077 &deferred_compaction);
3081 /* Checks for THP-specific high-order allocations */
3082 if (is_thp_gfp_mask(gfp_mask)) {
3084 * If compaction is deferred for high-order allocations, it is
3085 * because sync compaction recently failed. If this is the case
3086 * and the caller requested a THP allocation, we do not want
3087 * to heavily disrupt the system, so we fail the allocation
3088 * instead of entering direct reclaim.
3090 if (deferred_compaction)
3094 * In all zones where compaction was attempted (and not
3095 * deferred or skipped), lock contention has been detected.
3096 * For THP allocation we do not want to disrupt the others
3097 * so we fallback to base pages instead.
3099 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3103 * If compaction was aborted due to need_resched(), we do not
3104 * want to further increase allocation latency, unless it is
3105 * khugepaged trying to collapse.
3107 if (contended_compaction == COMPACT_CONTENDED_SCHED
3108 && !(current->flags & PF_KTHREAD))
3113 * It can become very expensive to allocate transparent hugepages at
3114 * fault, so use asynchronous memory compaction for THP unless it is
3115 * khugepaged trying to collapse.
3117 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3118 migration_mode = MIGRATE_SYNC_LIGHT;
3120 /* Try direct reclaim and then allocating */
3121 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3122 &did_some_progress);
3126 /* Do not loop if specifically requested */
3127 if (gfp_mask & __GFP_NORETRY)
3130 /* Keep reclaiming pages as long as there is reasonable progress */
3131 pages_reclaimed += did_some_progress;
3132 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3133 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3134 /* Wait for some write requests to complete then retry */
3135 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3139 /* Reclaim has failed us, start killing things */
3140 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3144 /* Retry as long as the OOM killer is making progress */
3145 if (did_some_progress)
3150 * High-order allocations do not necessarily loop after
3151 * direct reclaim and reclaim/compaction depends on compaction
3152 * being called after reclaim so call directly if necessary
3154 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3156 &contended_compaction,
3157 &deferred_compaction);
3161 warn_alloc_failed(gfp_mask, order, NULL);
3167 * This is the 'heart' of the zoned buddy allocator.
3170 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3171 struct zonelist *zonelist, nodemask_t *nodemask)
3173 struct zoneref *preferred_zoneref;
3174 struct page *page = NULL;
3175 unsigned int cpuset_mems_cookie;
3176 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3177 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3178 struct alloc_context ac = {
3179 .high_zoneidx = gfp_zone(gfp_mask),
3180 .nodemask = nodemask,
3181 .migratetype = gfpflags_to_migratetype(gfp_mask),
3184 gfp_mask &= gfp_allowed_mask;
3186 lockdep_trace_alloc(gfp_mask);
3188 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3190 if (should_fail_alloc_page(gfp_mask, order))
3194 * Check the zones suitable for the gfp_mask contain at least one
3195 * valid zone. It's possible to have an empty zonelist as a result
3196 * of __GFP_THISNODE and a memoryless node
3198 if (unlikely(!zonelist->_zonerefs->zone))
3201 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3202 alloc_flags |= ALLOC_CMA;
3205 cpuset_mems_cookie = read_mems_allowed_begin();
3207 /* We set it here, as __alloc_pages_slowpath might have changed it */
3208 ac.zonelist = zonelist;
3210 /* Dirty zone balancing only done in the fast path */
3211 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3213 /* The preferred zone is used for statistics later */
3214 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3215 ac.nodemask ? : &cpuset_current_mems_allowed,
3216 &ac.preferred_zone);
3217 if (!ac.preferred_zone)
3219 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3221 /* First allocation attempt */
3222 alloc_mask = gfp_mask|__GFP_HARDWALL;
3223 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3224 if (unlikely(!page)) {
3226 * Runtime PM, block IO and its error handling path
3227 * can deadlock because I/O on the device might not
3230 alloc_mask = memalloc_noio_flags(gfp_mask);
3231 ac.spread_dirty_pages = false;
3233 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3236 if (kmemcheck_enabled && page)
3237 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3239 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3243 * When updating a task's mems_allowed, it is possible to race with
3244 * parallel threads in such a way that an allocation can fail while
3245 * the mask is being updated. If a page allocation is about to fail,
3246 * check if the cpuset changed during allocation and if so, retry.
3248 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3253 EXPORT_SYMBOL(__alloc_pages_nodemask);
3256 * Common helper functions.
3258 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3263 * __get_free_pages() returns a 32-bit address, which cannot represent
3266 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3268 page = alloc_pages(gfp_mask, order);
3271 return (unsigned long) page_address(page);
3273 EXPORT_SYMBOL(__get_free_pages);
3275 unsigned long get_zeroed_page(gfp_t gfp_mask)
3277 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3279 EXPORT_SYMBOL(get_zeroed_page);
3281 void __free_pages(struct page *page, unsigned int order)
3283 if (put_page_testzero(page)) {
3285 free_hot_cold_page(page, false);
3287 __free_pages_ok(page, order);
3291 EXPORT_SYMBOL(__free_pages);
3293 void free_pages(unsigned long addr, unsigned int order)
3296 VM_BUG_ON(!virt_addr_valid((void *)addr));
3297 __free_pages(virt_to_page((void *)addr), order);
3301 EXPORT_SYMBOL(free_pages);
3305 * An arbitrary-length arbitrary-offset area of memory which resides
3306 * within a 0 or higher order page. Multiple fragments within that page
3307 * are individually refcounted, in the page's reference counter.
3309 * The page_frag functions below provide a simple allocation framework for
3310 * page fragments. This is used by the network stack and network device
3311 * drivers to provide a backing region of memory for use as either an
3312 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3314 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3317 struct page *page = NULL;
3318 gfp_t gfp = gfp_mask;
3320 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3321 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3323 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3324 PAGE_FRAG_CACHE_MAX_ORDER);
3325 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3327 if (unlikely(!page))
3328 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3330 nc->va = page ? page_address(page) : NULL;
3335 void *__alloc_page_frag(struct page_frag_cache *nc,
3336 unsigned int fragsz, gfp_t gfp_mask)
3338 unsigned int size = PAGE_SIZE;
3342 if (unlikely(!nc->va)) {
3344 page = __page_frag_refill(nc, gfp_mask);
3348 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3349 /* if size can vary use size else just use PAGE_SIZE */
3352 /* Even if we own the page, we do not use atomic_set().
3353 * This would break get_page_unless_zero() users.
3355 atomic_add(size - 1, &page->_count);
3357 /* reset page count bias and offset to start of new frag */
3358 nc->pfmemalloc = page_is_pfmemalloc(page);
3359 nc->pagecnt_bias = size;
3363 offset = nc->offset - fragsz;
3364 if (unlikely(offset < 0)) {
3365 page = virt_to_page(nc->va);
3367 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3370 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3371 /* if size can vary use size else just use PAGE_SIZE */
3374 /* OK, page count is 0, we can safely set it */
3375 atomic_set(&page->_count, size);
3377 /* reset page count bias and offset to start of new frag */
3378 nc->pagecnt_bias = size;
3379 offset = size - fragsz;
3383 nc->offset = offset;
3385 return nc->va + offset;
3387 EXPORT_SYMBOL(__alloc_page_frag);
3390 * Frees a page fragment allocated out of either a compound or order 0 page.
3392 void __free_page_frag(void *addr)
3394 struct page *page = virt_to_head_page(addr);
3396 if (unlikely(put_page_testzero(page)))
3397 __free_pages_ok(page, compound_order(page));
3399 EXPORT_SYMBOL(__free_page_frag);
3402 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3403 * of the current memory cgroup.
3405 * It should be used when the caller would like to use kmalloc, but since the
3406 * allocation is large, it has to fall back to the page allocator.
3408 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3412 page = alloc_pages(gfp_mask, order);
3413 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3414 __free_pages(page, order);
3420 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3424 page = alloc_pages_node(nid, gfp_mask, order);
3425 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3426 __free_pages(page, order);
3433 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3436 void __free_kmem_pages(struct page *page, unsigned int order)
3438 memcg_kmem_uncharge(page, order);
3439 __free_pages(page, order);
3442 void free_kmem_pages(unsigned long addr, unsigned int order)
3445 VM_BUG_ON(!virt_addr_valid((void *)addr));
3446 __free_kmem_pages(virt_to_page((void *)addr), order);
3450 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3453 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3454 unsigned long used = addr + PAGE_ALIGN(size);
3456 split_page(virt_to_page((void *)addr), order);
3457 while (used < alloc_end) {
3462 return (void *)addr;
3466 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3467 * @size: the number of bytes to allocate
3468 * @gfp_mask: GFP flags for the allocation
3470 * This function is similar to alloc_pages(), except that it allocates the
3471 * minimum number of pages to satisfy the request. alloc_pages() can only
3472 * allocate memory in power-of-two pages.
3474 * This function is also limited by MAX_ORDER.
3476 * Memory allocated by this function must be released by free_pages_exact().
3478 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3480 unsigned int order = get_order(size);
3483 addr = __get_free_pages(gfp_mask, order);
3484 return make_alloc_exact(addr, order, size);
3486 EXPORT_SYMBOL(alloc_pages_exact);
3489 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3491 * @nid: the preferred node ID where memory should be allocated
3492 * @size: the number of bytes to allocate
3493 * @gfp_mask: GFP flags for the allocation
3495 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3498 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3500 unsigned order = get_order(size);
3501 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3504 return make_alloc_exact((unsigned long)page_address(p), order, size);
3508 * free_pages_exact - release memory allocated via alloc_pages_exact()
3509 * @virt: the value returned by alloc_pages_exact.
3510 * @size: size of allocation, same value as passed to alloc_pages_exact().
3512 * Release the memory allocated by a previous call to alloc_pages_exact.
3514 void free_pages_exact(void *virt, size_t size)
3516 unsigned long addr = (unsigned long)virt;
3517 unsigned long end = addr + PAGE_ALIGN(size);
3519 while (addr < end) {
3524 EXPORT_SYMBOL(free_pages_exact);
3527 * nr_free_zone_pages - count number of pages beyond high watermark
3528 * @offset: The zone index of the highest zone
3530 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3531 * high watermark within all zones at or below a given zone index. For each
3532 * zone, the number of pages is calculated as:
3533 * managed_pages - high_pages
3535 static unsigned long nr_free_zone_pages(int offset)
3540 /* Just pick one node, since fallback list is circular */
3541 unsigned long sum = 0;
3543 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3545 for_each_zone_zonelist(zone, z, zonelist, offset) {
3546 unsigned long size = zone->managed_pages;
3547 unsigned long high = high_wmark_pages(zone);
3556 * nr_free_buffer_pages - count number of pages beyond high watermark
3558 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3559 * watermark within ZONE_DMA and ZONE_NORMAL.
3561 unsigned long nr_free_buffer_pages(void)
3563 return nr_free_zone_pages(gfp_zone(GFP_USER));
3565 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3568 * nr_free_pagecache_pages - count number of pages beyond high watermark
3570 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3571 * high watermark within all zones.
3573 unsigned long nr_free_pagecache_pages(void)
3575 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3578 static inline void show_node(struct zone *zone)
3580 if (IS_ENABLED(CONFIG_NUMA))
3581 printk("Node %d ", zone_to_nid(zone));
3584 void si_meminfo(struct sysinfo *val)
3586 val->totalram = totalram_pages;
3587 val->sharedram = global_page_state(NR_SHMEM);
3588 val->freeram = global_page_state(NR_FREE_PAGES);
3589 val->bufferram = nr_blockdev_pages();
3590 val->totalhigh = totalhigh_pages;
3591 val->freehigh = nr_free_highpages();
3592 val->mem_unit = PAGE_SIZE;
3595 EXPORT_SYMBOL(si_meminfo);
3598 void si_meminfo_node(struct sysinfo *val, int nid)
3600 int zone_type; /* needs to be signed */
3601 unsigned long managed_pages = 0;
3602 pg_data_t *pgdat = NODE_DATA(nid);
3604 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3605 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3606 val->totalram = managed_pages;
3607 val->sharedram = node_page_state(nid, NR_SHMEM);
3608 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3609 #ifdef CONFIG_HIGHMEM
3610 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3611 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3617 val->mem_unit = PAGE_SIZE;
3622 * Determine whether the node should be displayed or not, depending on whether
3623 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3625 bool skip_free_areas_node(unsigned int flags, int nid)
3628 unsigned int cpuset_mems_cookie;
3630 if (!(flags & SHOW_MEM_FILTER_NODES))
3634 cpuset_mems_cookie = read_mems_allowed_begin();
3635 ret = !node_isset(nid, cpuset_current_mems_allowed);
3636 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3641 #define K(x) ((x) << (PAGE_SHIFT-10))
3643 static void show_migration_types(unsigned char type)
3645 static const char types[MIGRATE_TYPES] = {
3646 [MIGRATE_UNMOVABLE] = 'U',
3647 [MIGRATE_RECLAIMABLE] = 'E',
3648 [MIGRATE_MOVABLE] = 'M',
3650 [MIGRATE_CMA] = 'C',
3652 #ifdef CONFIG_MEMORY_ISOLATION
3653 [MIGRATE_ISOLATE] = 'I',
3656 char tmp[MIGRATE_TYPES + 1];
3660 for (i = 0; i < MIGRATE_TYPES; i++) {
3661 if (type & (1 << i))
3666 printk("(%s) ", tmp);
3670 * Show free area list (used inside shift_scroll-lock stuff)
3671 * We also calculate the percentage fragmentation. We do this by counting the
3672 * memory on each free list with the exception of the first item on the list.
3675 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3678 void show_free_areas(unsigned int filter)
3680 unsigned long free_pcp = 0;
3684 for_each_populated_zone(zone) {
3685 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3688 for_each_online_cpu(cpu)
3689 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3692 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3693 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3694 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3695 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3696 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3697 " free:%lu free_pcp:%lu free_cma:%lu\n",
3698 global_page_state(NR_ACTIVE_ANON),
3699 global_page_state(NR_INACTIVE_ANON),
3700 global_page_state(NR_ISOLATED_ANON),
3701 global_page_state(NR_ACTIVE_FILE),
3702 global_page_state(NR_INACTIVE_FILE),
3703 global_page_state(NR_ISOLATED_FILE),
3704 global_page_state(NR_UNEVICTABLE),
3705 global_page_state(NR_FILE_DIRTY),
3706 global_page_state(NR_WRITEBACK),
3707 global_page_state(NR_UNSTABLE_NFS),
3708 global_page_state(NR_SLAB_RECLAIMABLE),
3709 global_page_state(NR_SLAB_UNRECLAIMABLE),
3710 global_page_state(NR_FILE_MAPPED),
3711 global_page_state(NR_SHMEM),
3712 global_page_state(NR_PAGETABLE),
3713 global_page_state(NR_BOUNCE),
3714 global_page_state(NR_FREE_PAGES),
3716 global_page_state(NR_FREE_CMA_PAGES));
3718 for_each_populated_zone(zone) {
3721 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3725 for_each_online_cpu(cpu)
3726 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3734 " active_anon:%lukB"
3735 " inactive_anon:%lukB"
3736 " active_file:%lukB"
3737 " inactive_file:%lukB"
3738 " unevictable:%lukB"
3739 " isolated(anon):%lukB"
3740 " isolated(file):%lukB"
3748 " slab_reclaimable:%lukB"
3749 " slab_unreclaimable:%lukB"
3750 " kernel_stack:%lukB"
3757 " writeback_tmp:%lukB"
3758 " pages_scanned:%lu"
3759 " all_unreclaimable? %s"
3762 K(zone_page_state(zone, NR_FREE_PAGES)),
3763 K(min_wmark_pages(zone)),
3764 K(low_wmark_pages(zone)),
3765 K(high_wmark_pages(zone)),
3766 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3767 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3768 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3769 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3770 K(zone_page_state(zone, NR_UNEVICTABLE)),
3771 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3772 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3773 K(zone->present_pages),
3774 K(zone->managed_pages),
3775 K(zone_page_state(zone, NR_MLOCK)),
3776 K(zone_page_state(zone, NR_FILE_DIRTY)),
3777 K(zone_page_state(zone, NR_WRITEBACK)),
3778 K(zone_page_state(zone, NR_FILE_MAPPED)),
3779 K(zone_page_state(zone, NR_SHMEM)),
3780 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3781 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3782 zone_page_state(zone, NR_KERNEL_STACK) *
3784 K(zone_page_state(zone, NR_PAGETABLE)),
3785 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3786 K(zone_page_state(zone, NR_BOUNCE)),
3788 K(this_cpu_read(zone->pageset->pcp.count)),
3789 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3790 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3791 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3792 (!zone_reclaimable(zone) ? "yes" : "no")
3794 printk("lowmem_reserve[]:");
3795 for (i = 0; i < MAX_NR_ZONES; i++)
3796 printk(" %ld", zone->lowmem_reserve[i]);
3800 for_each_populated_zone(zone) {
3801 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3802 unsigned char types[MAX_ORDER];
3804 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3807 printk("%s: ", zone->name);
3809 spin_lock_irqsave(&zone->lock, flags);
3810 for (order = 0; order < MAX_ORDER; order++) {
3811 struct free_area *area = &zone->free_area[order];
3814 nr[order] = area->nr_free;
3815 total += nr[order] << order;
3818 for (type = 0; type < MIGRATE_TYPES; type++) {
3819 if (!list_empty(&area->free_list[type]))
3820 types[order] |= 1 << type;
3823 spin_unlock_irqrestore(&zone->lock, flags);
3824 for (order = 0; order < MAX_ORDER; order++) {
3825 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3827 show_migration_types(types[order]);
3829 printk("= %lukB\n", K(total));
3832 hugetlb_show_meminfo();
3834 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3836 show_swap_cache_info();
3839 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3841 zoneref->zone = zone;
3842 zoneref->zone_idx = zone_idx(zone);
3846 * Builds allocation fallback zone lists.
3848 * Add all populated zones of a node to the zonelist.
3850 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3854 enum zone_type zone_type = MAX_NR_ZONES;
3858 zone = pgdat->node_zones + zone_type;
3859 if (populated_zone(zone)) {
3860 zoneref_set_zone(zone,
3861 &zonelist->_zonerefs[nr_zones++]);
3862 check_highest_zone(zone_type);
3864 } while (zone_type);
3872 * 0 = automatic detection of better ordering.
3873 * 1 = order by ([node] distance, -zonetype)
3874 * 2 = order by (-zonetype, [node] distance)
3876 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3877 * the same zonelist. So only NUMA can configure this param.
3879 #define ZONELIST_ORDER_DEFAULT 0
3880 #define ZONELIST_ORDER_NODE 1
3881 #define ZONELIST_ORDER_ZONE 2
3883 /* zonelist order in the kernel.
3884 * set_zonelist_order() will set this to NODE or ZONE.
3886 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3887 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3891 /* The value user specified ....changed by config */
3892 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3893 /* string for sysctl */
3894 #define NUMA_ZONELIST_ORDER_LEN 16
3895 char numa_zonelist_order[16] = "default";
3898 * interface for configure zonelist ordering.
3899 * command line option "numa_zonelist_order"
3900 * = "[dD]efault - default, automatic configuration.
3901 * = "[nN]ode - order by node locality, then by zone within node
3902 * = "[zZ]one - order by zone, then by locality within zone
3905 static int __parse_numa_zonelist_order(char *s)
3907 if (*s == 'd' || *s == 'D') {
3908 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3909 } else if (*s == 'n' || *s == 'N') {
3910 user_zonelist_order = ZONELIST_ORDER_NODE;
3911 } else if (*s == 'z' || *s == 'Z') {
3912 user_zonelist_order = ZONELIST_ORDER_ZONE;
3915 "Ignoring invalid numa_zonelist_order value: "
3922 static __init int setup_numa_zonelist_order(char *s)
3929 ret = __parse_numa_zonelist_order(s);
3931 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3935 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3938 * sysctl handler for numa_zonelist_order
3940 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3941 void __user *buffer, size_t *length,
3944 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3946 static DEFINE_MUTEX(zl_order_mutex);
3948 mutex_lock(&zl_order_mutex);
3950 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3954 strcpy(saved_string, (char *)table->data);
3956 ret = proc_dostring(table, write, buffer, length, ppos);
3960 int oldval = user_zonelist_order;
3962 ret = __parse_numa_zonelist_order((char *)table->data);
3965 * bogus value. restore saved string
3967 strncpy((char *)table->data, saved_string,
3968 NUMA_ZONELIST_ORDER_LEN);
3969 user_zonelist_order = oldval;
3970 } else if (oldval != user_zonelist_order) {
3971 mutex_lock(&zonelists_mutex);
3972 build_all_zonelists(NULL, NULL);
3973 mutex_unlock(&zonelists_mutex);
3977 mutex_unlock(&zl_order_mutex);
3982 #define MAX_NODE_LOAD (nr_online_nodes)
3983 static int node_load[MAX_NUMNODES];
3986 * find_next_best_node - find the next node that should appear in a given node's fallback list
3987 * @node: node whose fallback list we're appending
3988 * @used_node_mask: nodemask_t of already used nodes
3990 * We use a number of factors to determine which is the next node that should
3991 * appear on a given node's fallback list. The node should not have appeared
3992 * already in @node's fallback list, and it should be the next closest node
3993 * according to the distance array (which contains arbitrary distance values
3994 * from each node to each node in the system), and should also prefer nodes
3995 * with no CPUs, since presumably they'll have very little allocation pressure
3996 * on them otherwise.
3997 * It returns -1 if no node is found.
3999 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4002 int min_val = INT_MAX;
4003 int best_node = NUMA_NO_NODE;
4004 const struct cpumask *tmp = cpumask_of_node(0);
4006 /* Use the local node if we haven't already */
4007 if (!node_isset(node, *used_node_mask)) {
4008 node_set(node, *used_node_mask);
4012 for_each_node_state(n, N_MEMORY) {
4014 /* Don't want a node to appear more than once */
4015 if (node_isset(n, *used_node_mask))
4018 /* Use the distance array to find the distance */
4019 val = node_distance(node, n);
4021 /* Penalize nodes under us ("prefer the next node") */
4024 /* Give preference to headless and unused nodes */
4025 tmp = cpumask_of_node(n);
4026 if (!cpumask_empty(tmp))
4027 val += PENALTY_FOR_NODE_WITH_CPUS;
4029 /* Slight preference for less loaded node */
4030 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4031 val += node_load[n];
4033 if (val < min_val) {
4040 node_set(best_node, *used_node_mask);
4047 * Build zonelists ordered by node and zones within node.
4048 * This results in maximum locality--normal zone overflows into local
4049 * DMA zone, if any--but risks exhausting DMA zone.
4051 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4054 struct zonelist *zonelist;
4056 zonelist = &pgdat->node_zonelists[0];
4057 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4059 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4060 zonelist->_zonerefs[j].zone = NULL;
4061 zonelist->_zonerefs[j].zone_idx = 0;
4065 * Build gfp_thisnode zonelists
4067 static void build_thisnode_zonelists(pg_data_t *pgdat)
4070 struct zonelist *zonelist;
4072 zonelist = &pgdat->node_zonelists[1];
4073 j = build_zonelists_node(pgdat, zonelist, 0);
4074 zonelist->_zonerefs[j].zone = NULL;
4075 zonelist->_zonerefs[j].zone_idx = 0;
4079 * Build zonelists ordered by zone and nodes within zones.
4080 * This results in conserving DMA zone[s] until all Normal memory is
4081 * exhausted, but results in overflowing to remote node while memory
4082 * may still exist in local DMA zone.
4084 static int node_order[MAX_NUMNODES];
4086 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4089 int zone_type; /* needs to be signed */
4091 struct zonelist *zonelist;
4093 zonelist = &pgdat->node_zonelists[0];
4095 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4096 for (j = 0; j < nr_nodes; j++) {
4097 node = node_order[j];
4098 z = &NODE_DATA(node)->node_zones[zone_type];
4099 if (populated_zone(z)) {
4101 &zonelist->_zonerefs[pos++]);
4102 check_highest_zone(zone_type);
4106 zonelist->_zonerefs[pos].zone = NULL;
4107 zonelist->_zonerefs[pos].zone_idx = 0;
4110 #if defined(CONFIG_64BIT)
4112 * Devices that require DMA32/DMA are relatively rare and do not justify a
4113 * penalty to every machine in case the specialised case applies. Default
4114 * to Node-ordering on 64-bit NUMA machines
4116 static int default_zonelist_order(void)
4118 return ZONELIST_ORDER_NODE;
4122 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4123 * by the kernel. If processes running on node 0 deplete the low memory zone
4124 * then reclaim will occur more frequency increasing stalls and potentially
4125 * be easier to OOM if a large percentage of the zone is under writeback or
4126 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4127 * Hence, default to zone ordering on 32-bit.
4129 static int default_zonelist_order(void)
4131 return ZONELIST_ORDER_ZONE;
4133 #endif /* CONFIG_64BIT */
4135 static void set_zonelist_order(void)
4137 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4138 current_zonelist_order = default_zonelist_order();
4140 current_zonelist_order = user_zonelist_order;
4143 static void build_zonelists(pg_data_t *pgdat)
4147 nodemask_t used_mask;
4148 int local_node, prev_node;
4149 struct zonelist *zonelist;
4150 int order = current_zonelist_order;
4152 /* initialize zonelists */
4153 for (i = 0; i < MAX_ZONELISTS; i++) {
4154 zonelist = pgdat->node_zonelists + i;
4155 zonelist->_zonerefs[0].zone = NULL;
4156 zonelist->_zonerefs[0].zone_idx = 0;
4159 /* NUMA-aware ordering of nodes */
4160 local_node = pgdat->node_id;
4161 load = nr_online_nodes;
4162 prev_node = local_node;
4163 nodes_clear(used_mask);
4165 memset(node_order, 0, sizeof(node_order));
4168 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4170 * We don't want to pressure a particular node.
4171 * So adding penalty to the first node in same
4172 * distance group to make it round-robin.
4174 if (node_distance(local_node, node) !=
4175 node_distance(local_node, prev_node))
4176 node_load[node] = load;
4180 if (order == ZONELIST_ORDER_NODE)
4181 build_zonelists_in_node_order(pgdat, node);
4183 node_order[j++] = node; /* remember order */
4186 if (order == ZONELIST_ORDER_ZONE) {
4187 /* calculate node order -- i.e., DMA last! */
4188 build_zonelists_in_zone_order(pgdat, j);
4191 build_thisnode_zonelists(pgdat);
4194 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4196 * Return node id of node used for "local" allocations.
4197 * I.e., first node id of first zone in arg node's generic zonelist.
4198 * Used for initializing percpu 'numa_mem', which is used primarily
4199 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4201 int local_memory_node(int node)
4205 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4206 gfp_zone(GFP_KERNEL),
4213 #else /* CONFIG_NUMA */
4215 static void set_zonelist_order(void)
4217 current_zonelist_order = ZONELIST_ORDER_ZONE;
4220 static void build_zonelists(pg_data_t *pgdat)
4222 int node, local_node;
4224 struct zonelist *zonelist;
4226 local_node = pgdat->node_id;
4228 zonelist = &pgdat->node_zonelists[0];
4229 j = build_zonelists_node(pgdat, zonelist, 0);
4232 * Now we build the zonelist so that it contains the zones
4233 * of all the other nodes.
4234 * We don't want to pressure a particular node, so when
4235 * building the zones for node N, we make sure that the
4236 * zones coming right after the local ones are those from
4237 * node N+1 (modulo N)
4239 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4240 if (!node_online(node))
4242 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4244 for (node = 0; node < local_node; node++) {
4245 if (!node_online(node))
4247 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4250 zonelist->_zonerefs[j].zone = NULL;
4251 zonelist->_zonerefs[j].zone_idx = 0;
4254 #endif /* CONFIG_NUMA */
4257 * Boot pageset table. One per cpu which is going to be used for all
4258 * zones and all nodes. The parameters will be set in such a way
4259 * that an item put on a list will immediately be handed over to
4260 * the buddy list. This is safe since pageset manipulation is done
4261 * with interrupts disabled.
4263 * The boot_pagesets must be kept even after bootup is complete for
4264 * unused processors and/or zones. They do play a role for bootstrapping
4265 * hotplugged processors.
4267 * zoneinfo_show() and maybe other functions do
4268 * not check if the processor is online before following the pageset pointer.
4269 * Other parts of the kernel may not check if the zone is available.
4271 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4272 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4273 static void setup_zone_pageset(struct zone *zone);
4276 * Global mutex to protect against size modification of zonelists
4277 * as well as to serialize pageset setup for the new populated zone.
4279 DEFINE_MUTEX(zonelists_mutex);
4281 /* return values int ....just for stop_machine() */
4282 static int __build_all_zonelists(void *data)
4286 pg_data_t *self = data;
4289 memset(node_load, 0, sizeof(node_load));
4292 if (self && !node_online(self->node_id)) {
4293 build_zonelists(self);
4296 for_each_online_node(nid) {
4297 pg_data_t *pgdat = NODE_DATA(nid);
4299 build_zonelists(pgdat);
4303 * Initialize the boot_pagesets that are going to be used
4304 * for bootstrapping processors. The real pagesets for
4305 * each zone will be allocated later when the per cpu
4306 * allocator is available.
4308 * boot_pagesets are used also for bootstrapping offline
4309 * cpus if the system is already booted because the pagesets
4310 * are needed to initialize allocators on a specific cpu too.
4311 * F.e. the percpu allocator needs the page allocator which
4312 * needs the percpu allocator in order to allocate its pagesets
4313 * (a chicken-egg dilemma).
4315 for_each_possible_cpu(cpu) {
4316 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4318 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4320 * We now know the "local memory node" for each node--
4321 * i.e., the node of the first zone in the generic zonelist.
4322 * Set up numa_mem percpu variable for all possible cpus
4323 * if associated node has been onlined.
4325 if (node_online(cpu_to_node(cpu)))
4326 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4328 set_cpu_numa_mem(cpu, NUMA_NO_NODE);
4335 static noinline void __init
4336 build_all_zonelists_init(void)
4338 __build_all_zonelists(NULL);
4339 mminit_verify_zonelist();
4340 cpuset_init_current_mems_allowed();
4344 * Called with zonelists_mutex held always
4345 * unless system_state == SYSTEM_BOOTING.
4347 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4348 * [we're only called with non-NULL zone through __meminit paths] and
4349 * (2) call of __init annotated helper build_all_zonelists_init
4350 * [protected by SYSTEM_BOOTING].
4352 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4354 set_zonelist_order();
4356 if (system_state == SYSTEM_BOOTING) {
4357 build_all_zonelists_init();
4359 #ifdef CONFIG_MEMORY_HOTPLUG
4361 setup_zone_pageset(zone);
4363 /* we have to stop all cpus to guarantee there is no user
4365 stop_machine(__build_all_zonelists, pgdat, NULL);
4366 /* cpuset refresh routine should be here */
4368 vm_total_pages = nr_free_pagecache_pages();
4370 * Disable grouping by mobility if the number of pages in the
4371 * system is too low to allow the mechanism to work. It would be
4372 * more accurate, but expensive to check per-zone. This check is
4373 * made on memory-hotadd so a system can start with mobility
4374 * disabled and enable it later
4376 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4377 page_group_by_mobility_disabled = 1;
4379 page_group_by_mobility_disabled = 0;
4381 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4382 "Total pages: %ld\n",
4384 zonelist_order_name[current_zonelist_order],
4385 page_group_by_mobility_disabled ? "off" : "on",
4388 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4393 * Helper functions to size the waitqueue hash table.
4394 * Essentially these want to choose hash table sizes sufficiently
4395 * large so that collisions trying to wait on pages are rare.
4396 * But in fact, the number of active page waitqueues on typical
4397 * systems is ridiculously low, less than 200. So this is even
4398 * conservative, even though it seems large.
4400 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4401 * waitqueues, i.e. the size of the waitq table given the number of pages.
4403 #define PAGES_PER_WAITQUEUE 256
4405 #ifndef CONFIG_MEMORY_HOTPLUG
4406 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4408 unsigned long size = 1;
4410 pages /= PAGES_PER_WAITQUEUE;
4412 while (size < pages)
4416 * Once we have dozens or even hundreds of threads sleeping
4417 * on IO we've got bigger problems than wait queue collision.
4418 * Limit the size of the wait table to a reasonable size.
4420 size = min(size, 4096UL);
4422 return max(size, 4UL);
4426 * A zone's size might be changed by hot-add, so it is not possible to determine
4427 * a suitable size for its wait_table. So we use the maximum size now.
4429 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4431 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4432 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4433 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4435 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4436 * or more by the traditional way. (See above). It equals:
4438 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4439 * ia64(16K page size) : = ( 8G + 4M)byte.
4440 * powerpc (64K page size) : = (32G +16M)byte.
4442 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4449 * This is an integer logarithm so that shifts can be used later
4450 * to extract the more random high bits from the multiplicative
4451 * hash function before the remainder is taken.
4453 static inline unsigned long wait_table_bits(unsigned long size)
4459 * Initially all pages are reserved - free ones are freed
4460 * up by free_all_bootmem() once the early boot process is
4461 * done. Non-atomic initialization, single-pass.
4463 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4464 unsigned long start_pfn, enum memmap_context context)
4466 pg_data_t *pgdat = NODE_DATA(nid);
4467 unsigned long end_pfn = start_pfn + size;
4470 unsigned long nr_initialised = 0;
4472 if (highest_memmap_pfn < end_pfn - 1)
4473 highest_memmap_pfn = end_pfn - 1;
4475 z = &pgdat->node_zones[zone];
4476 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4478 * There can be holes in boot-time mem_map[]s
4479 * handed to this function. They do not
4480 * exist on hotplugged memory.
4482 if (context == MEMMAP_EARLY) {
4483 if (!early_pfn_valid(pfn))
4485 if (!early_pfn_in_nid(pfn, nid))
4487 if (!update_defer_init(pgdat, pfn, end_pfn,
4493 * Mark the block movable so that blocks are reserved for
4494 * movable at startup. This will force kernel allocations
4495 * to reserve their blocks rather than leaking throughout
4496 * the address space during boot when many long-lived
4497 * kernel allocations are made.
4499 * bitmap is created for zone's valid pfn range. but memmap
4500 * can be created for invalid pages (for alignment)
4501 * check here not to call set_pageblock_migratetype() against
4504 if (!(pfn & (pageblock_nr_pages - 1))) {
4505 struct page *page = pfn_to_page(pfn);
4507 __init_single_page(page, pfn, zone, nid);
4508 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4510 __init_single_pfn(pfn, zone, nid);
4515 static void __meminit zone_init_free_lists(struct zone *zone)
4517 unsigned int order, t;
4518 for_each_migratetype_order(order, t) {
4519 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4520 zone->free_area[order].nr_free = 0;
4524 #ifndef __HAVE_ARCH_MEMMAP_INIT
4525 #define memmap_init(size, nid, zone, start_pfn) \
4526 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4529 static int zone_batchsize(struct zone *zone)
4535 * The per-cpu-pages pools are set to around 1000th of the
4536 * size of the zone. But no more than 1/2 of a meg.
4538 * OK, so we don't know how big the cache is. So guess.
4540 batch = zone->managed_pages / 1024;
4541 if (batch * PAGE_SIZE > 512 * 1024)
4542 batch = (512 * 1024) / PAGE_SIZE;
4543 batch /= 4; /* We effectively *= 4 below */
4548 * Clamp the batch to a 2^n - 1 value. Having a power
4549 * of 2 value was found to be more likely to have
4550 * suboptimal cache aliasing properties in some cases.
4552 * For example if 2 tasks are alternately allocating
4553 * batches of pages, one task can end up with a lot
4554 * of pages of one half of the possible page colors
4555 * and the other with pages of the other colors.
4557 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4562 /* The deferral and batching of frees should be suppressed under NOMMU
4565 * The problem is that NOMMU needs to be able to allocate large chunks
4566 * of contiguous memory as there's no hardware page translation to
4567 * assemble apparent contiguous memory from discontiguous pages.
4569 * Queueing large contiguous runs of pages for batching, however,
4570 * causes the pages to actually be freed in smaller chunks. As there
4571 * can be a significant delay between the individual batches being
4572 * recycled, this leads to the once large chunks of space being
4573 * fragmented and becoming unavailable for high-order allocations.
4580 * pcp->high and pcp->batch values are related and dependent on one another:
4581 * ->batch must never be higher then ->high.
4582 * The following function updates them in a safe manner without read side
4585 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4586 * those fields changing asynchronously (acording the the above rule).
4588 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4589 * outside of boot time (or some other assurance that no concurrent updaters
4592 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4593 unsigned long batch)
4595 /* start with a fail safe value for batch */
4599 /* Update high, then batch, in order */
4606 /* a companion to pageset_set_high() */
4607 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4609 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4612 static void pageset_init(struct per_cpu_pageset *p)
4614 struct per_cpu_pages *pcp;
4617 memset(p, 0, sizeof(*p));
4621 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4622 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4625 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4628 pageset_set_batch(p, batch);
4632 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4633 * to the value high for the pageset p.
4635 static void pageset_set_high(struct per_cpu_pageset *p,
4638 unsigned long batch = max(1UL, high / 4);
4639 if ((high / 4) > (PAGE_SHIFT * 8))
4640 batch = PAGE_SHIFT * 8;
4642 pageset_update(&p->pcp, high, batch);
4645 static void pageset_set_high_and_batch(struct zone *zone,
4646 struct per_cpu_pageset *pcp)
4648 if (percpu_pagelist_fraction)
4649 pageset_set_high(pcp,
4650 (zone->managed_pages /
4651 percpu_pagelist_fraction));
4653 pageset_set_batch(pcp, zone_batchsize(zone));
4656 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4658 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4661 pageset_set_high_and_batch(zone, pcp);
4664 static void __meminit setup_zone_pageset(struct zone *zone)
4667 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4668 for_each_possible_cpu(cpu)
4669 zone_pageset_init(zone, cpu);
4673 * Allocate per cpu pagesets and initialize them.
4674 * Before this call only boot pagesets were available.
4676 void __init setup_per_cpu_pageset(void)
4680 for_each_populated_zone(zone)
4681 setup_zone_pageset(zone);
4684 static noinline __init_refok
4685 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4691 * The per-page waitqueue mechanism uses hashed waitqueues
4694 zone->wait_table_hash_nr_entries =
4695 wait_table_hash_nr_entries(zone_size_pages);
4696 zone->wait_table_bits =
4697 wait_table_bits(zone->wait_table_hash_nr_entries);
4698 alloc_size = zone->wait_table_hash_nr_entries
4699 * sizeof(wait_queue_head_t);
4701 if (!slab_is_available()) {
4702 zone->wait_table = (wait_queue_head_t *)
4703 memblock_virt_alloc_node_nopanic(
4704 alloc_size, zone->zone_pgdat->node_id);
4707 * This case means that a zone whose size was 0 gets new memory
4708 * via memory hot-add.
4709 * But it may be the case that a new node was hot-added. In
4710 * this case vmalloc() will not be able to use this new node's
4711 * memory - this wait_table must be initialized to use this new
4712 * node itself as well.
4713 * To use this new node's memory, further consideration will be
4716 zone->wait_table = vmalloc(alloc_size);
4718 if (!zone->wait_table)
4721 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4722 init_waitqueue_head(zone->wait_table + i);
4727 static __meminit void zone_pcp_init(struct zone *zone)
4730 * per cpu subsystem is not up at this point. The following code
4731 * relies on the ability of the linker to provide the
4732 * offset of a (static) per cpu variable into the per cpu area.
4734 zone->pageset = &boot_pageset;
4736 if (populated_zone(zone))
4737 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4738 zone->name, zone->present_pages,
4739 zone_batchsize(zone));
4742 int __meminit init_currently_empty_zone(struct zone *zone,
4743 unsigned long zone_start_pfn,
4746 struct pglist_data *pgdat = zone->zone_pgdat;
4748 ret = zone_wait_table_init(zone, size);
4751 pgdat->nr_zones = zone_idx(zone) + 1;
4753 zone->zone_start_pfn = zone_start_pfn;
4755 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4756 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4758 (unsigned long)zone_idx(zone),
4759 zone_start_pfn, (zone_start_pfn + size));
4761 zone_init_free_lists(zone);
4766 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4767 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4770 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4772 int __meminit __early_pfn_to_nid(unsigned long pfn,
4773 struct mminit_pfnnid_cache *state)
4775 unsigned long start_pfn, end_pfn;
4778 if (state->last_start <= pfn && pfn < state->last_end)
4779 return state->last_nid;
4781 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4783 state->last_start = start_pfn;
4784 state->last_end = end_pfn;
4785 state->last_nid = nid;
4790 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4793 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4794 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4795 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4797 * If an architecture guarantees that all ranges registered contain no holes
4798 * and may be freed, this this function may be used instead of calling
4799 * memblock_free_early_nid() manually.
4801 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4803 unsigned long start_pfn, end_pfn;
4806 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4807 start_pfn = min(start_pfn, max_low_pfn);
4808 end_pfn = min(end_pfn, max_low_pfn);
4810 if (start_pfn < end_pfn)
4811 memblock_free_early_nid(PFN_PHYS(start_pfn),
4812 (end_pfn - start_pfn) << PAGE_SHIFT,
4818 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4819 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4821 * If an architecture guarantees that all ranges registered contain no holes and may
4822 * be freed, this function may be used instead of calling memory_present() manually.
4824 void __init sparse_memory_present_with_active_regions(int nid)
4826 unsigned long start_pfn, end_pfn;
4829 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4830 memory_present(this_nid, start_pfn, end_pfn);
4834 * get_pfn_range_for_nid - Return the start and end page frames for a node
4835 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4836 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4837 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4839 * It returns the start and end page frame of a node based on information
4840 * provided by memblock_set_node(). If called for a node
4841 * with no available memory, a warning is printed and the start and end
4844 void __meminit get_pfn_range_for_nid(unsigned int nid,
4845 unsigned long *start_pfn, unsigned long *end_pfn)
4847 unsigned long this_start_pfn, this_end_pfn;
4853 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4854 *start_pfn = min(*start_pfn, this_start_pfn);
4855 *end_pfn = max(*end_pfn, this_end_pfn);
4858 if (*start_pfn == -1UL)
4863 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4864 * assumption is made that zones within a node are ordered in monotonic
4865 * increasing memory addresses so that the "highest" populated zone is used
4867 static void __init find_usable_zone_for_movable(void)
4870 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4871 if (zone_index == ZONE_MOVABLE)
4874 if (arch_zone_highest_possible_pfn[zone_index] >
4875 arch_zone_lowest_possible_pfn[zone_index])
4879 VM_BUG_ON(zone_index == -1);
4880 movable_zone = zone_index;
4884 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4885 * because it is sized independent of architecture. Unlike the other zones,
4886 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4887 * in each node depending on the size of each node and how evenly kernelcore
4888 * is distributed. This helper function adjusts the zone ranges
4889 * provided by the architecture for a given node by using the end of the
4890 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4891 * zones within a node are in order of monotonic increases memory addresses
4893 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4894 unsigned long zone_type,
4895 unsigned long node_start_pfn,
4896 unsigned long node_end_pfn,
4897 unsigned long *zone_start_pfn,
4898 unsigned long *zone_end_pfn)
4900 /* Only adjust if ZONE_MOVABLE is on this node */
4901 if (zone_movable_pfn[nid]) {
4902 /* Size ZONE_MOVABLE */
4903 if (zone_type == ZONE_MOVABLE) {
4904 *zone_start_pfn = zone_movable_pfn[nid];
4905 *zone_end_pfn = min(node_end_pfn,
4906 arch_zone_highest_possible_pfn[movable_zone]);
4908 /* Adjust for ZONE_MOVABLE starting within this range */
4909 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4910 *zone_end_pfn > zone_movable_pfn[nid]) {
4911 *zone_end_pfn = zone_movable_pfn[nid];
4913 /* Check if this whole range is within ZONE_MOVABLE */
4914 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4915 *zone_start_pfn = *zone_end_pfn;
4920 * Return the number of pages a zone spans in a node, including holes
4921 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4923 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4924 unsigned long zone_type,
4925 unsigned long node_start_pfn,
4926 unsigned long node_end_pfn,
4927 unsigned long *ignored)
4929 unsigned long zone_start_pfn, zone_end_pfn;
4931 /* When hotadd a new node from cpu_up(), the node should be empty */
4932 if (!node_start_pfn && !node_end_pfn)
4935 /* Get the start and end of the zone */
4936 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4937 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4938 adjust_zone_range_for_zone_movable(nid, zone_type,
4939 node_start_pfn, node_end_pfn,
4940 &zone_start_pfn, &zone_end_pfn);
4942 /* Check that this node has pages within the zone's required range */
4943 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4946 /* Move the zone boundaries inside the node if necessary */
4947 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4948 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4950 /* Return the spanned pages */
4951 return zone_end_pfn - zone_start_pfn;
4955 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4956 * then all holes in the requested range will be accounted for.
4958 unsigned long __meminit __absent_pages_in_range(int nid,
4959 unsigned long range_start_pfn,
4960 unsigned long range_end_pfn)
4962 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4963 unsigned long start_pfn, end_pfn;
4966 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4967 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4968 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4969 nr_absent -= end_pfn - start_pfn;
4975 * absent_pages_in_range - Return number of page frames in holes within a range
4976 * @start_pfn: The start PFN to start searching for holes
4977 * @end_pfn: The end PFN to stop searching for holes
4979 * It returns the number of pages frames in memory holes within a range.
4981 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4982 unsigned long end_pfn)
4984 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4987 /* Return the number of page frames in holes in a zone on a node */
4988 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4989 unsigned long zone_type,
4990 unsigned long node_start_pfn,
4991 unsigned long node_end_pfn,
4992 unsigned long *ignored)
4994 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4995 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4996 unsigned long zone_start_pfn, zone_end_pfn;
4998 /* When hotadd a new node from cpu_up(), the node should be empty */
4999 if (!node_start_pfn && !node_end_pfn)
5002 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5003 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5005 adjust_zone_range_for_zone_movable(nid, zone_type,
5006 node_start_pfn, node_end_pfn,
5007 &zone_start_pfn, &zone_end_pfn);
5008 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5011 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5012 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5013 unsigned long zone_type,
5014 unsigned long node_start_pfn,
5015 unsigned long node_end_pfn,
5016 unsigned long *zones_size)
5018 return zones_size[zone_type];
5021 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5022 unsigned long zone_type,
5023 unsigned long node_start_pfn,
5024 unsigned long node_end_pfn,
5025 unsigned long *zholes_size)
5030 return zholes_size[zone_type];
5033 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5035 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5036 unsigned long node_start_pfn,
5037 unsigned long node_end_pfn,
5038 unsigned long *zones_size,
5039 unsigned long *zholes_size)
5041 unsigned long realtotalpages = 0, totalpages = 0;
5044 for (i = 0; i < MAX_NR_ZONES; i++) {
5045 struct zone *zone = pgdat->node_zones + i;
5046 unsigned long size, real_size;
5048 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5052 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5053 node_start_pfn, node_end_pfn,
5055 zone->spanned_pages = size;
5056 zone->present_pages = real_size;
5059 realtotalpages += real_size;
5062 pgdat->node_spanned_pages = totalpages;
5063 pgdat->node_present_pages = realtotalpages;
5064 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5068 #ifndef CONFIG_SPARSEMEM
5070 * Calculate the size of the zone->blockflags rounded to an unsigned long
5071 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5072 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5073 * round what is now in bits to nearest long in bits, then return it in
5076 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5078 unsigned long usemapsize;
5080 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5081 usemapsize = roundup(zonesize, pageblock_nr_pages);
5082 usemapsize = usemapsize >> pageblock_order;
5083 usemapsize *= NR_PAGEBLOCK_BITS;
5084 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5086 return usemapsize / 8;
5089 static void __init setup_usemap(struct pglist_data *pgdat,
5091 unsigned long zone_start_pfn,
5092 unsigned long zonesize)
5094 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5095 zone->pageblock_flags = NULL;
5097 zone->pageblock_flags =
5098 memblock_virt_alloc_node_nopanic(usemapsize,
5102 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5103 unsigned long zone_start_pfn, unsigned long zonesize) {}
5104 #endif /* CONFIG_SPARSEMEM */
5106 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5108 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5109 void __paginginit set_pageblock_order(void)
5113 /* Check that pageblock_nr_pages has not already been setup */
5114 if (pageblock_order)
5117 if (HPAGE_SHIFT > PAGE_SHIFT)
5118 order = HUGETLB_PAGE_ORDER;
5120 order = MAX_ORDER - 1;
5123 * Assume the largest contiguous order of interest is a huge page.
5124 * This value may be variable depending on boot parameters on IA64 and
5127 pageblock_order = order;
5129 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5132 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5133 * is unused as pageblock_order is set at compile-time. See
5134 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5137 void __paginginit set_pageblock_order(void)
5141 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5143 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5144 unsigned long present_pages)
5146 unsigned long pages = spanned_pages;
5149 * Provide a more accurate estimation if there are holes within
5150 * the zone and SPARSEMEM is in use. If there are holes within the
5151 * zone, each populated memory region may cost us one or two extra
5152 * memmap pages due to alignment because memmap pages for each
5153 * populated regions may not naturally algined on page boundary.
5154 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5156 if (spanned_pages > present_pages + (present_pages >> 4) &&
5157 IS_ENABLED(CONFIG_SPARSEMEM))
5158 pages = present_pages;
5160 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5164 * Set up the zone data structures:
5165 * - mark all pages reserved
5166 * - mark all memory queues empty
5167 * - clear the memory bitmaps
5169 * NOTE: pgdat should get zeroed by caller.
5171 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5174 int nid = pgdat->node_id;
5175 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5178 pgdat_resize_init(pgdat);
5179 #ifdef CONFIG_NUMA_BALANCING
5180 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5181 pgdat->numabalancing_migrate_nr_pages = 0;
5182 pgdat->numabalancing_migrate_next_window = jiffies;
5184 init_waitqueue_head(&pgdat->kswapd_wait);
5185 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5186 pgdat_page_ext_init(pgdat);
5188 for (j = 0; j < MAX_NR_ZONES; j++) {
5189 struct zone *zone = pgdat->node_zones + j;
5190 unsigned long size, realsize, freesize, memmap_pages;
5192 size = zone->spanned_pages;
5193 realsize = freesize = zone->present_pages;
5196 * Adjust freesize so that it accounts for how much memory
5197 * is used by this zone for memmap. This affects the watermark
5198 * and per-cpu initialisations
5200 memmap_pages = calc_memmap_size(size, realsize);
5201 if (!is_highmem_idx(j)) {
5202 if (freesize >= memmap_pages) {
5203 freesize -= memmap_pages;
5206 " %s zone: %lu pages used for memmap\n",
5207 zone_names[j], memmap_pages);
5210 " %s zone: %lu pages exceeds freesize %lu\n",
5211 zone_names[j], memmap_pages, freesize);
5214 /* Account for reserved pages */
5215 if (j == 0 && freesize > dma_reserve) {
5216 freesize -= dma_reserve;
5217 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5218 zone_names[0], dma_reserve);
5221 if (!is_highmem_idx(j))
5222 nr_kernel_pages += freesize;
5223 /* Charge for highmem memmap if there are enough kernel pages */
5224 else if (nr_kernel_pages > memmap_pages * 2)
5225 nr_kernel_pages -= memmap_pages;
5226 nr_all_pages += freesize;
5229 * Set an approximate value for lowmem here, it will be adjusted
5230 * when the bootmem allocator frees pages into the buddy system.
5231 * And all highmem pages will be managed by the buddy system.
5233 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5236 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5238 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5240 zone->name = zone_names[j];
5241 spin_lock_init(&zone->lock);
5242 spin_lock_init(&zone->lru_lock);
5243 zone_seqlock_init(zone);
5244 zone->zone_pgdat = pgdat;
5245 zone_pcp_init(zone);
5247 /* For bootup, initialized properly in watermark setup */
5248 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5250 lruvec_init(&zone->lruvec);
5254 set_pageblock_order();
5255 setup_usemap(pgdat, zone, zone_start_pfn, size);
5256 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5258 memmap_init(size, nid, j, zone_start_pfn);
5259 zone_start_pfn += size;
5263 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5265 unsigned long __maybe_unused offset = 0;
5267 /* Skip empty nodes */
5268 if (!pgdat->node_spanned_pages)
5271 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5272 /* ia64 gets its own node_mem_map, before this, without bootmem */
5273 if (!pgdat->node_mem_map) {
5274 unsigned long size, start, end;
5278 * The zone's endpoints aren't required to be MAX_ORDER
5279 * aligned but the node_mem_map endpoints must be in order
5280 * for the buddy allocator to function correctly.
5282 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5283 offset = pgdat->node_start_pfn - start;
5284 end = pgdat_end_pfn(pgdat);
5285 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5286 size = (end - start) * sizeof(struct page);
5287 map = alloc_remap(pgdat->node_id, size);
5289 map = memblock_virt_alloc_node_nopanic(size,
5291 pgdat->node_mem_map = map + offset;
5293 #ifndef CONFIG_NEED_MULTIPLE_NODES
5295 * With no DISCONTIG, the global mem_map is just set as node 0's
5297 if (pgdat == NODE_DATA(0)) {
5298 mem_map = NODE_DATA(0)->node_mem_map;
5299 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5300 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5302 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5305 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5308 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5309 unsigned long node_start_pfn, unsigned long *zholes_size)
5311 pg_data_t *pgdat = NODE_DATA(nid);
5312 unsigned long start_pfn = 0;
5313 unsigned long end_pfn = 0;
5315 /* pg_data_t should be reset to zero when it's allocated */
5316 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5318 reset_deferred_meminit(pgdat);
5319 pgdat->node_id = nid;
5320 pgdat->node_start_pfn = node_start_pfn;
5321 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5322 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5323 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5324 (u64)start_pfn << PAGE_SHIFT,
5325 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5327 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5328 zones_size, zholes_size);
5330 alloc_node_mem_map(pgdat);
5331 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5332 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5333 nid, (unsigned long)pgdat,
5334 (unsigned long)pgdat->node_mem_map);
5337 free_area_init_core(pgdat);
5340 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5342 #if MAX_NUMNODES > 1
5344 * Figure out the number of possible node ids.
5346 void __init setup_nr_node_ids(void)
5348 unsigned int highest;
5350 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5351 nr_node_ids = highest + 1;
5356 * node_map_pfn_alignment - determine the maximum internode alignment
5358 * This function should be called after node map is populated and sorted.
5359 * It calculates the maximum power of two alignment which can distinguish
5362 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5363 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5364 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5365 * shifted, 1GiB is enough and this function will indicate so.
5367 * This is used to test whether pfn -> nid mapping of the chosen memory
5368 * model has fine enough granularity to avoid incorrect mapping for the
5369 * populated node map.
5371 * Returns the determined alignment in pfn's. 0 if there is no alignment
5372 * requirement (single node).
5374 unsigned long __init node_map_pfn_alignment(void)
5376 unsigned long accl_mask = 0, last_end = 0;
5377 unsigned long start, end, mask;
5381 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5382 if (!start || last_nid < 0 || last_nid == nid) {
5389 * Start with a mask granular enough to pin-point to the
5390 * start pfn and tick off bits one-by-one until it becomes
5391 * too coarse to separate the current node from the last.
5393 mask = ~((1 << __ffs(start)) - 1);
5394 while (mask && last_end <= (start & (mask << 1)))
5397 /* accumulate all internode masks */
5401 /* convert mask to number of pages */
5402 return ~accl_mask + 1;
5405 /* Find the lowest pfn for a node */
5406 static unsigned long __init find_min_pfn_for_node(int nid)
5408 unsigned long min_pfn = ULONG_MAX;
5409 unsigned long start_pfn;
5412 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5413 min_pfn = min(min_pfn, start_pfn);
5415 if (min_pfn == ULONG_MAX) {
5417 "Could not find start_pfn for node %d\n", nid);
5425 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5427 * It returns the minimum PFN based on information provided via
5428 * memblock_set_node().
5430 unsigned long __init find_min_pfn_with_active_regions(void)
5432 return find_min_pfn_for_node(MAX_NUMNODES);
5436 * early_calculate_totalpages()
5437 * Sum pages in active regions for movable zone.
5438 * Populate N_MEMORY for calculating usable_nodes.
5440 static unsigned long __init early_calculate_totalpages(void)
5442 unsigned long totalpages = 0;
5443 unsigned long start_pfn, end_pfn;
5446 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5447 unsigned long pages = end_pfn - start_pfn;
5449 totalpages += pages;
5451 node_set_state(nid, N_MEMORY);
5457 * Find the PFN the Movable zone begins in each node. Kernel memory
5458 * is spread evenly between nodes as long as the nodes have enough
5459 * memory. When they don't, some nodes will have more kernelcore than
5462 static void __init find_zone_movable_pfns_for_nodes(void)
5465 unsigned long usable_startpfn;
5466 unsigned long kernelcore_node, kernelcore_remaining;
5467 /* save the state before borrow the nodemask */
5468 nodemask_t saved_node_state = node_states[N_MEMORY];
5469 unsigned long totalpages = early_calculate_totalpages();
5470 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5471 struct memblock_region *r;
5473 /* Need to find movable_zone earlier when movable_node is specified. */
5474 find_usable_zone_for_movable();
5477 * If movable_node is specified, ignore kernelcore and movablecore
5480 if (movable_node_is_enabled()) {
5481 for_each_memblock(memory, r) {
5482 if (!memblock_is_hotpluggable(r))
5487 usable_startpfn = PFN_DOWN(r->base);
5488 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5489 min(usable_startpfn, zone_movable_pfn[nid]) :
5497 * If movablecore=nn[KMG] was specified, calculate what size of
5498 * kernelcore that corresponds so that memory usable for
5499 * any allocation type is evenly spread. If both kernelcore
5500 * and movablecore are specified, then the value of kernelcore
5501 * will be used for required_kernelcore if it's greater than
5502 * what movablecore would have allowed.
5504 if (required_movablecore) {
5505 unsigned long corepages;
5508 * Round-up so that ZONE_MOVABLE is at least as large as what
5509 * was requested by the user
5511 required_movablecore =
5512 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5513 required_movablecore = min(totalpages, required_movablecore);
5514 corepages = totalpages - required_movablecore;
5516 required_kernelcore = max(required_kernelcore, corepages);
5520 * If kernelcore was not specified or kernelcore size is larger
5521 * than totalpages, there is no ZONE_MOVABLE.
5523 if (!required_kernelcore || required_kernelcore >= totalpages)
5526 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5527 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5530 /* Spread kernelcore memory as evenly as possible throughout nodes */
5531 kernelcore_node = required_kernelcore / usable_nodes;
5532 for_each_node_state(nid, N_MEMORY) {
5533 unsigned long start_pfn, end_pfn;
5536 * Recalculate kernelcore_node if the division per node
5537 * now exceeds what is necessary to satisfy the requested
5538 * amount of memory for the kernel
5540 if (required_kernelcore < kernelcore_node)
5541 kernelcore_node = required_kernelcore / usable_nodes;
5544 * As the map is walked, we track how much memory is usable
5545 * by the kernel using kernelcore_remaining. When it is
5546 * 0, the rest of the node is usable by ZONE_MOVABLE
5548 kernelcore_remaining = kernelcore_node;
5550 /* Go through each range of PFNs within this node */
5551 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5552 unsigned long size_pages;
5554 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5555 if (start_pfn >= end_pfn)
5558 /* Account for what is only usable for kernelcore */
5559 if (start_pfn < usable_startpfn) {
5560 unsigned long kernel_pages;
5561 kernel_pages = min(end_pfn, usable_startpfn)
5564 kernelcore_remaining -= min(kernel_pages,
5565 kernelcore_remaining);
5566 required_kernelcore -= min(kernel_pages,
5567 required_kernelcore);
5569 /* Continue if range is now fully accounted */
5570 if (end_pfn <= usable_startpfn) {
5573 * Push zone_movable_pfn to the end so
5574 * that if we have to rebalance
5575 * kernelcore across nodes, we will
5576 * not double account here
5578 zone_movable_pfn[nid] = end_pfn;
5581 start_pfn = usable_startpfn;
5585 * The usable PFN range for ZONE_MOVABLE is from
5586 * start_pfn->end_pfn. Calculate size_pages as the
5587 * number of pages used as kernelcore
5589 size_pages = end_pfn - start_pfn;
5590 if (size_pages > kernelcore_remaining)
5591 size_pages = kernelcore_remaining;
5592 zone_movable_pfn[nid] = start_pfn + size_pages;
5595 * Some kernelcore has been met, update counts and
5596 * break if the kernelcore for this node has been
5599 required_kernelcore -= min(required_kernelcore,
5601 kernelcore_remaining -= size_pages;
5602 if (!kernelcore_remaining)
5608 * If there is still required_kernelcore, we do another pass with one
5609 * less node in the count. This will push zone_movable_pfn[nid] further
5610 * along on the nodes that still have memory until kernelcore is
5614 if (usable_nodes && required_kernelcore > usable_nodes)
5618 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5619 for (nid = 0; nid < MAX_NUMNODES; nid++)
5620 zone_movable_pfn[nid] =
5621 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5624 /* restore the node_state */
5625 node_states[N_MEMORY] = saved_node_state;
5628 /* Any regular or high memory on that node ? */
5629 static void check_for_memory(pg_data_t *pgdat, int nid)
5631 enum zone_type zone_type;
5633 if (N_MEMORY == N_NORMAL_MEMORY)
5636 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5637 struct zone *zone = &pgdat->node_zones[zone_type];
5638 if (populated_zone(zone)) {
5639 node_set_state(nid, N_HIGH_MEMORY);
5640 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5641 zone_type <= ZONE_NORMAL)
5642 node_set_state(nid, N_NORMAL_MEMORY);
5649 * free_area_init_nodes - Initialise all pg_data_t and zone data
5650 * @max_zone_pfn: an array of max PFNs for each zone
5652 * This will call free_area_init_node() for each active node in the system.
5653 * Using the page ranges provided by memblock_set_node(), the size of each
5654 * zone in each node and their holes is calculated. If the maximum PFN
5655 * between two adjacent zones match, it is assumed that the zone is empty.
5656 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5657 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5658 * starts where the previous one ended. For example, ZONE_DMA32 starts
5659 * at arch_max_dma_pfn.
5661 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5663 unsigned long start_pfn, end_pfn;
5666 /* Record where the zone boundaries are */
5667 memset(arch_zone_lowest_possible_pfn, 0,
5668 sizeof(arch_zone_lowest_possible_pfn));
5669 memset(arch_zone_highest_possible_pfn, 0,
5670 sizeof(arch_zone_highest_possible_pfn));
5671 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5672 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5673 for (i = 1; i < MAX_NR_ZONES; i++) {
5674 if (i == ZONE_MOVABLE)
5676 arch_zone_lowest_possible_pfn[i] =
5677 arch_zone_highest_possible_pfn[i-1];
5678 arch_zone_highest_possible_pfn[i] =
5679 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5681 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5682 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5684 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5685 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5686 find_zone_movable_pfns_for_nodes();
5688 /* Print out the zone ranges */
5689 pr_info("Zone ranges:\n");
5690 for (i = 0; i < MAX_NR_ZONES; i++) {
5691 if (i == ZONE_MOVABLE)
5693 pr_info(" %-8s ", zone_names[i]);
5694 if (arch_zone_lowest_possible_pfn[i] ==
5695 arch_zone_highest_possible_pfn[i])
5698 pr_cont("[mem %#018Lx-%#018Lx]\n",
5699 (u64)arch_zone_lowest_possible_pfn[i]
5701 ((u64)arch_zone_highest_possible_pfn[i]
5702 << PAGE_SHIFT) - 1);
5705 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5706 pr_info("Movable zone start for each node\n");
5707 for (i = 0; i < MAX_NUMNODES; i++) {
5708 if (zone_movable_pfn[i])
5709 pr_info(" Node %d: %#018Lx\n", i,
5710 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5713 /* Print out the early node map */
5714 pr_info("Early memory node ranges\n");
5715 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5716 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5717 (u64)start_pfn << PAGE_SHIFT,
5718 ((u64)end_pfn << PAGE_SHIFT) - 1);
5720 /* Initialise every node */
5721 mminit_verify_pageflags_layout();
5722 setup_nr_node_ids();
5723 for_each_online_node(nid) {
5724 pg_data_t *pgdat = NODE_DATA(nid);
5725 free_area_init_node(nid, NULL,
5726 find_min_pfn_for_node(nid), NULL);
5728 /* Any memory on that node */
5729 if (pgdat->node_present_pages)
5730 node_set_state(nid, N_MEMORY);
5731 check_for_memory(pgdat, nid);
5735 static int __init cmdline_parse_core(char *p, unsigned long *core)
5737 unsigned long long coremem;
5741 coremem = memparse(p, &p);
5742 *core = coremem >> PAGE_SHIFT;
5744 /* Paranoid check that UL is enough for the coremem value */
5745 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5751 * kernelcore=size sets the amount of memory for use for allocations that
5752 * cannot be reclaimed or migrated.
5754 static int __init cmdline_parse_kernelcore(char *p)
5756 return cmdline_parse_core(p, &required_kernelcore);
5760 * movablecore=size sets the amount of memory for use for allocations that
5761 * can be reclaimed or migrated.
5763 static int __init cmdline_parse_movablecore(char *p)
5765 return cmdline_parse_core(p, &required_movablecore);
5768 early_param("kernelcore", cmdline_parse_kernelcore);
5769 early_param("movablecore", cmdline_parse_movablecore);
5771 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5773 void adjust_managed_page_count(struct page *page, long count)
5775 spin_lock(&managed_page_count_lock);
5776 page_zone(page)->managed_pages += count;
5777 totalram_pages += count;
5778 #ifdef CONFIG_HIGHMEM
5779 if (PageHighMem(page))
5780 totalhigh_pages += count;
5782 spin_unlock(&managed_page_count_lock);
5784 EXPORT_SYMBOL(adjust_managed_page_count);
5786 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5789 unsigned long pages = 0;
5791 start = (void *)PAGE_ALIGN((unsigned long)start);
5792 end = (void *)((unsigned long)end & PAGE_MASK);
5793 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5794 if ((unsigned int)poison <= 0xFF)
5795 memset(pos, poison, PAGE_SIZE);
5796 free_reserved_page(virt_to_page(pos));
5800 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5801 s, pages << (PAGE_SHIFT - 10), start, end);
5805 EXPORT_SYMBOL(free_reserved_area);
5807 #ifdef CONFIG_HIGHMEM
5808 void free_highmem_page(struct page *page)
5810 __free_reserved_page(page);
5812 page_zone(page)->managed_pages++;
5818 void __init mem_init_print_info(const char *str)
5820 unsigned long physpages, codesize, datasize, rosize, bss_size;
5821 unsigned long init_code_size, init_data_size;
5823 physpages = get_num_physpages();
5824 codesize = _etext - _stext;
5825 datasize = _edata - _sdata;
5826 rosize = __end_rodata - __start_rodata;
5827 bss_size = __bss_stop - __bss_start;
5828 init_data_size = __init_end - __init_begin;
5829 init_code_size = _einittext - _sinittext;
5832 * Detect special cases and adjust section sizes accordingly:
5833 * 1) .init.* may be embedded into .data sections
5834 * 2) .init.text.* may be out of [__init_begin, __init_end],
5835 * please refer to arch/tile/kernel/vmlinux.lds.S.
5836 * 3) .rodata.* may be embedded into .text or .data sections.
5838 #define adj_init_size(start, end, size, pos, adj) \
5840 if (start <= pos && pos < end && size > adj) \
5844 adj_init_size(__init_begin, __init_end, init_data_size,
5845 _sinittext, init_code_size);
5846 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5847 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5848 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5849 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5851 #undef adj_init_size
5853 pr_info("Memory: %luK/%luK available "
5854 "(%luK kernel code, %luK rwdata, %luK rodata, "
5855 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5856 #ifdef CONFIG_HIGHMEM
5860 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5861 codesize >> 10, datasize >> 10, rosize >> 10,
5862 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5863 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5864 totalcma_pages << (PAGE_SHIFT-10),
5865 #ifdef CONFIG_HIGHMEM
5866 totalhigh_pages << (PAGE_SHIFT-10),
5868 str ? ", " : "", str ? str : "");
5872 * set_dma_reserve - set the specified number of pages reserved in the first zone
5873 * @new_dma_reserve: The number of pages to mark reserved
5875 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5876 * In the DMA zone, a significant percentage may be consumed by kernel image
5877 * and other unfreeable allocations which can skew the watermarks badly. This
5878 * function may optionally be used to account for unfreeable pages in the
5879 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5880 * smaller per-cpu batchsize.
5882 void __init set_dma_reserve(unsigned long new_dma_reserve)
5884 dma_reserve = new_dma_reserve;
5887 void __init free_area_init(unsigned long *zones_size)
5889 free_area_init_node(0, zones_size,
5890 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5893 static int page_alloc_cpu_notify(struct notifier_block *self,
5894 unsigned long action, void *hcpu)
5896 int cpu = (unsigned long)hcpu;
5898 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5899 lru_add_drain_cpu(cpu);
5903 * Spill the event counters of the dead processor
5904 * into the current processors event counters.
5905 * This artificially elevates the count of the current
5908 vm_events_fold_cpu(cpu);
5911 * Zero the differential counters of the dead processor
5912 * so that the vm statistics are consistent.
5914 * This is only okay since the processor is dead and cannot
5915 * race with what we are doing.
5917 cpu_vm_stats_fold(cpu);
5922 void __init page_alloc_init(void)
5924 hotcpu_notifier(page_alloc_cpu_notify, 0);
5928 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5929 * or min_free_kbytes changes.
5931 static void calculate_totalreserve_pages(void)
5933 struct pglist_data *pgdat;
5934 unsigned long reserve_pages = 0;
5935 enum zone_type i, j;
5937 for_each_online_pgdat(pgdat) {
5938 for (i = 0; i < MAX_NR_ZONES; i++) {
5939 struct zone *zone = pgdat->node_zones + i;
5942 /* Find valid and maximum lowmem_reserve in the zone */
5943 for (j = i; j < MAX_NR_ZONES; j++) {
5944 if (zone->lowmem_reserve[j] > max)
5945 max = zone->lowmem_reserve[j];
5948 /* we treat the high watermark as reserved pages. */
5949 max += high_wmark_pages(zone);
5951 if (max > zone->managed_pages)
5952 max = zone->managed_pages;
5953 reserve_pages += max;
5955 * Lowmem reserves are not available to
5956 * GFP_HIGHUSER page cache allocations and
5957 * kswapd tries to balance zones to their high
5958 * watermark. As a result, neither should be
5959 * regarded as dirtyable memory, to prevent a
5960 * situation where reclaim has to clean pages
5961 * in order to balance the zones.
5963 zone->dirty_balance_reserve = max;
5966 dirty_balance_reserve = reserve_pages;
5967 totalreserve_pages = reserve_pages;
5971 * setup_per_zone_lowmem_reserve - called whenever
5972 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5973 * has a correct pages reserved value, so an adequate number of
5974 * pages are left in the zone after a successful __alloc_pages().
5976 static void setup_per_zone_lowmem_reserve(void)
5978 struct pglist_data *pgdat;
5979 enum zone_type j, idx;
5981 for_each_online_pgdat(pgdat) {
5982 for (j = 0; j < MAX_NR_ZONES; j++) {
5983 struct zone *zone = pgdat->node_zones + j;
5984 unsigned long managed_pages = zone->managed_pages;
5986 zone->lowmem_reserve[j] = 0;
5990 struct zone *lower_zone;
5994 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5995 sysctl_lowmem_reserve_ratio[idx] = 1;
5997 lower_zone = pgdat->node_zones + idx;
5998 lower_zone->lowmem_reserve[j] = managed_pages /
5999 sysctl_lowmem_reserve_ratio[idx];
6000 managed_pages += lower_zone->managed_pages;
6005 /* update totalreserve_pages */
6006 calculate_totalreserve_pages();
6009 static void __setup_per_zone_wmarks(void)
6011 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6012 unsigned long lowmem_pages = 0;
6014 unsigned long flags;
6016 /* Calculate total number of !ZONE_HIGHMEM pages */
6017 for_each_zone(zone) {
6018 if (!is_highmem(zone))
6019 lowmem_pages += zone->managed_pages;
6022 for_each_zone(zone) {
6025 spin_lock_irqsave(&zone->lock, flags);
6026 tmp = (u64)pages_min * zone->managed_pages;
6027 do_div(tmp, lowmem_pages);
6028 if (is_highmem(zone)) {
6030 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6031 * need highmem pages, so cap pages_min to a small
6034 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6035 * deltas control asynch page reclaim, and so should
6036 * not be capped for highmem.
6038 unsigned long min_pages;
6040 min_pages = zone->managed_pages / 1024;
6041 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6042 zone->watermark[WMARK_MIN] = min_pages;
6045 * If it's a lowmem zone, reserve a number of pages
6046 * proportionate to the zone's size.
6048 zone->watermark[WMARK_MIN] = tmp;
6051 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6052 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6054 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6055 high_wmark_pages(zone) - low_wmark_pages(zone) -
6056 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6058 spin_unlock_irqrestore(&zone->lock, flags);
6061 /* update totalreserve_pages */
6062 calculate_totalreserve_pages();
6066 * setup_per_zone_wmarks - called when min_free_kbytes changes
6067 * or when memory is hot-{added|removed}
6069 * Ensures that the watermark[min,low,high] values for each zone are set
6070 * correctly with respect to min_free_kbytes.
6072 void setup_per_zone_wmarks(void)
6074 mutex_lock(&zonelists_mutex);
6075 __setup_per_zone_wmarks();
6076 mutex_unlock(&zonelists_mutex);
6080 * The inactive anon list should be small enough that the VM never has to
6081 * do too much work, but large enough that each inactive page has a chance
6082 * to be referenced again before it is swapped out.
6084 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6085 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6086 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6087 * the anonymous pages are kept on the inactive list.
6090 * memory ratio inactive anon
6091 * -------------------------------------
6100 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6102 unsigned int gb, ratio;
6104 /* Zone size in gigabytes */
6105 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6107 ratio = int_sqrt(10 * gb);
6111 zone->inactive_ratio = ratio;
6114 static void __meminit setup_per_zone_inactive_ratio(void)
6119 calculate_zone_inactive_ratio(zone);
6123 * Initialise min_free_kbytes.
6125 * For small machines we want it small (128k min). For large machines
6126 * we want it large (64MB max). But it is not linear, because network
6127 * bandwidth does not increase linearly with machine size. We use
6129 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6130 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6146 int __meminit init_per_zone_wmark_min(void)
6148 unsigned long lowmem_kbytes;
6149 int new_min_free_kbytes;
6151 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6152 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6154 if (new_min_free_kbytes > user_min_free_kbytes) {
6155 min_free_kbytes = new_min_free_kbytes;
6156 if (min_free_kbytes < 128)
6157 min_free_kbytes = 128;
6158 if (min_free_kbytes > 65536)
6159 min_free_kbytes = 65536;
6161 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6162 new_min_free_kbytes, user_min_free_kbytes);
6164 setup_per_zone_wmarks();
6165 refresh_zone_stat_thresholds();
6166 setup_per_zone_lowmem_reserve();
6167 setup_per_zone_inactive_ratio();
6170 module_init(init_per_zone_wmark_min)
6173 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6174 * that we can call two helper functions whenever min_free_kbytes
6177 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6178 void __user *buffer, size_t *length, loff_t *ppos)
6182 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6187 user_min_free_kbytes = min_free_kbytes;
6188 setup_per_zone_wmarks();
6194 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6195 void __user *buffer, size_t *length, loff_t *ppos)
6200 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6205 zone->min_unmapped_pages = (zone->managed_pages *
6206 sysctl_min_unmapped_ratio) / 100;
6210 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6211 void __user *buffer, size_t *length, loff_t *ppos)
6216 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6221 zone->min_slab_pages = (zone->managed_pages *
6222 sysctl_min_slab_ratio) / 100;
6228 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6229 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6230 * whenever sysctl_lowmem_reserve_ratio changes.
6232 * The reserve ratio obviously has absolutely no relation with the
6233 * minimum watermarks. The lowmem reserve ratio can only make sense
6234 * if in function of the boot time zone sizes.
6236 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6237 void __user *buffer, size_t *length, loff_t *ppos)
6239 proc_dointvec_minmax(table, write, buffer, length, ppos);
6240 setup_per_zone_lowmem_reserve();
6245 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6246 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6247 * pagelist can have before it gets flushed back to buddy allocator.
6249 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6250 void __user *buffer, size_t *length, loff_t *ppos)
6253 int old_percpu_pagelist_fraction;
6256 mutex_lock(&pcp_batch_high_lock);
6257 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6259 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6260 if (!write || ret < 0)
6263 /* Sanity checking to avoid pcp imbalance */
6264 if (percpu_pagelist_fraction &&
6265 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6266 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6272 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6275 for_each_populated_zone(zone) {
6278 for_each_possible_cpu(cpu)
6279 pageset_set_high_and_batch(zone,
6280 per_cpu_ptr(zone->pageset, cpu));
6283 mutex_unlock(&pcp_batch_high_lock);
6288 int hashdist = HASHDIST_DEFAULT;
6290 static int __init set_hashdist(char *str)
6294 hashdist = simple_strtoul(str, &str, 0);
6297 __setup("hashdist=", set_hashdist);
6301 * allocate a large system hash table from bootmem
6302 * - it is assumed that the hash table must contain an exact power-of-2
6303 * quantity of entries
6304 * - limit is the number of hash buckets, not the total allocation size
6306 void *__init alloc_large_system_hash(const char *tablename,
6307 unsigned long bucketsize,
6308 unsigned long numentries,
6311 unsigned int *_hash_shift,
6312 unsigned int *_hash_mask,
6313 unsigned long low_limit,
6314 unsigned long high_limit)
6316 unsigned long long max = high_limit;
6317 unsigned long log2qty, size;
6320 /* allow the kernel cmdline to have a say */
6322 /* round applicable memory size up to nearest megabyte */
6323 numentries = nr_kernel_pages;
6325 /* It isn't necessary when PAGE_SIZE >= 1MB */
6326 if (PAGE_SHIFT < 20)
6327 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6329 /* limit to 1 bucket per 2^scale bytes of low memory */
6330 if (scale > PAGE_SHIFT)
6331 numentries >>= (scale - PAGE_SHIFT);
6333 numentries <<= (PAGE_SHIFT - scale);
6335 /* Make sure we've got at least a 0-order allocation.. */
6336 if (unlikely(flags & HASH_SMALL)) {
6337 /* Makes no sense without HASH_EARLY */
6338 WARN_ON(!(flags & HASH_EARLY));
6339 if (!(numentries >> *_hash_shift)) {
6340 numentries = 1UL << *_hash_shift;
6341 BUG_ON(!numentries);
6343 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6344 numentries = PAGE_SIZE / bucketsize;
6346 numentries = roundup_pow_of_two(numentries);
6348 /* limit allocation size to 1/16 total memory by default */
6350 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6351 do_div(max, bucketsize);
6353 max = min(max, 0x80000000ULL);
6355 if (numentries < low_limit)
6356 numentries = low_limit;
6357 if (numentries > max)
6360 log2qty = ilog2(numentries);
6363 size = bucketsize << log2qty;
6364 if (flags & HASH_EARLY)
6365 table = memblock_virt_alloc_nopanic(size, 0);
6367 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6370 * If bucketsize is not a power-of-two, we may free
6371 * some pages at the end of hash table which
6372 * alloc_pages_exact() automatically does
6374 if (get_order(size) < MAX_ORDER) {
6375 table = alloc_pages_exact(size, GFP_ATOMIC);
6376 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6379 } while (!table && size > PAGE_SIZE && --log2qty);
6382 panic("Failed to allocate %s hash table\n", tablename);
6384 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6387 ilog2(size) - PAGE_SHIFT,
6391 *_hash_shift = log2qty;
6393 *_hash_mask = (1 << log2qty) - 1;
6398 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6399 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6402 #ifdef CONFIG_SPARSEMEM
6403 return __pfn_to_section(pfn)->pageblock_flags;
6405 return zone->pageblock_flags;
6406 #endif /* CONFIG_SPARSEMEM */
6409 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6411 #ifdef CONFIG_SPARSEMEM
6412 pfn &= (PAGES_PER_SECTION-1);
6413 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6415 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6416 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6417 #endif /* CONFIG_SPARSEMEM */
6421 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6422 * @page: The page within the block of interest
6423 * @pfn: The target page frame number
6424 * @end_bitidx: The last bit of interest to retrieve
6425 * @mask: mask of bits that the caller is interested in
6427 * Return: pageblock_bits flags
6429 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6430 unsigned long end_bitidx,
6434 unsigned long *bitmap;
6435 unsigned long bitidx, word_bitidx;
6438 zone = page_zone(page);
6439 bitmap = get_pageblock_bitmap(zone, pfn);
6440 bitidx = pfn_to_bitidx(zone, pfn);
6441 word_bitidx = bitidx / BITS_PER_LONG;
6442 bitidx &= (BITS_PER_LONG-1);
6444 word = bitmap[word_bitidx];
6445 bitidx += end_bitidx;
6446 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6450 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6451 * @page: The page within the block of interest
6452 * @flags: The flags to set
6453 * @pfn: The target page frame number
6454 * @end_bitidx: The last bit of interest
6455 * @mask: mask of bits that the caller is interested in
6457 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6459 unsigned long end_bitidx,
6463 unsigned long *bitmap;
6464 unsigned long bitidx, word_bitidx;
6465 unsigned long old_word, word;
6467 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6469 zone = page_zone(page);
6470 bitmap = get_pageblock_bitmap(zone, pfn);
6471 bitidx = pfn_to_bitidx(zone, pfn);
6472 word_bitidx = bitidx / BITS_PER_LONG;
6473 bitidx &= (BITS_PER_LONG-1);
6475 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6477 bitidx += end_bitidx;
6478 mask <<= (BITS_PER_LONG - bitidx - 1);
6479 flags <<= (BITS_PER_LONG - bitidx - 1);
6481 word = READ_ONCE(bitmap[word_bitidx]);
6483 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6484 if (word == old_word)
6491 * This function checks whether pageblock includes unmovable pages or not.
6492 * If @count is not zero, it is okay to include less @count unmovable pages
6494 * PageLRU check without isolation or lru_lock could race so that
6495 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6496 * expect this function should be exact.
6498 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6499 bool skip_hwpoisoned_pages)
6501 unsigned long pfn, iter, found;
6505 * For avoiding noise data, lru_add_drain_all() should be called
6506 * If ZONE_MOVABLE, the zone never contains unmovable pages
6508 if (zone_idx(zone) == ZONE_MOVABLE)
6510 mt = get_pageblock_migratetype(page);
6511 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6514 pfn = page_to_pfn(page);
6515 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6516 unsigned long check = pfn + iter;
6518 if (!pfn_valid_within(check))
6521 page = pfn_to_page(check);
6524 * Hugepages are not in LRU lists, but they're movable.
6525 * We need not scan over tail pages bacause we don't
6526 * handle each tail page individually in migration.
6528 if (PageHuge(page)) {
6529 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6534 * We can't use page_count without pin a page
6535 * because another CPU can free compound page.
6536 * This check already skips compound tails of THP
6537 * because their page->_count is zero at all time.
6539 if (!atomic_read(&page->_count)) {
6540 if (PageBuddy(page))
6541 iter += (1 << page_order(page)) - 1;
6546 * The HWPoisoned page may be not in buddy system, and
6547 * page_count() is not 0.
6549 if (skip_hwpoisoned_pages && PageHWPoison(page))
6555 * If there are RECLAIMABLE pages, we need to check
6556 * it. But now, memory offline itself doesn't call
6557 * shrink_node_slabs() and it still to be fixed.
6560 * If the page is not RAM, page_count()should be 0.
6561 * we don't need more check. This is an _used_ not-movable page.
6563 * The problematic thing here is PG_reserved pages. PG_reserved
6564 * is set to both of a memory hole page and a _used_ kernel
6573 bool is_pageblock_removable_nolock(struct page *page)
6579 * We have to be careful here because we are iterating over memory
6580 * sections which are not zone aware so we might end up outside of
6581 * the zone but still within the section.
6582 * We have to take care about the node as well. If the node is offline
6583 * its NODE_DATA will be NULL - see page_zone.
6585 if (!node_online(page_to_nid(page)))
6588 zone = page_zone(page);
6589 pfn = page_to_pfn(page);
6590 if (!zone_spans_pfn(zone, pfn))
6593 return !has_unmovable_pages(zone, page, 0, true);
6598 static unsigned long pfn_max_align_down(unsigned long pfn)
6600 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6601 pageblock_nr_pages) - 1);
6604 static unsigned long pfn_max_align_up(unsigned long pfn)
6606 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6607 pageblock_nr_pages));
6610 /* [start, end) must belong to a single zone. */
6611 static int __alloc_contig_migrate_range(struct compact_control *cc,
6612 unsigned long start, unsigned long end)
6614 /* This function is based on compact_zone() from compaction.c. */
6615 unsigned long nr_reclaimed;
6616 unsigned long pfn = start;
6617 unsigned int tries = 0;
6622 while (pfn < end || !list_empty(&cc->migratepages)) {
6623 if (fatal_signal_pending(current)) {
6628 if (list_empty(&cc->migratepages)) {
6629 cc->nr_migratepages = 0;
6630 pfn = isolate_migratepages_range(cc, pfn, end);
6636 } else if (++tries == 5) {
6637 ret = ret < 0 ? ret : -EBUSY;
6641 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6643 cc->nr_migratepages -= nr_reclaimed;
6645 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6646 NULL, 0, cc->mode, MR_CMA);
6649 putback_movable_pages(&cc->migratepages);
6656 * alloc_contig_range() -- tries to allocate given range of pages
6657 * @start: start PFN to allocate
6658 * @end: one-past-the-last PFN to allocate
6659 * @migratetype: migratetype of the underlaying pageblocks (either
6660 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6661 * in range must have the same migratetype and it must
6662 * be either of the two.
6664 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6665 * aligned, however it's the caller's responsibility to guarantee that
6666 * we are the only thread that changes migrate type of pageblocks the
6669 * The PFN range must belong to a single zone.
6671 * Returns zero on success or negative error code. On success all
6672 * pages which PFN is in [start, end) are allocated for the caller and
6673 * need to be freed with free_contig_range().
6675 int alloc_contig_range(unsigned long start, unsigned long end,
6676 unsigned migratetype)
6678 unsigned long outer_start, outer_end;
6681 struct compact_control cc = {
6682 .nr_migratepages = 0,
6684 .zone = page_zone(pfn_to_page(start)),
6685 .mode = MIGRATE_SYNC,
6686 .ignore_skip_hint = true,
6688 INIT_LIST_HEAD(&cc.migratepages);
6691 * What we do here is we mark all pageblocks in range as
6692 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6693 * have different sizes, and due to the way page allocator
6694 * work, we align the range to biggest of the two pages so
6695 * that page allocator won't try to merge buddies from
6696 * different pageblocks and change MIGRATE_ISOLATE to some
6697 * other migration type.
6699 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6700 * migrate the pages from an unaligned range (ie. pages that
6701 * we are interested in). This will put all the pages in
6702 * range back to page allocator as MIGRATE_ISOLATE.
6704 * When this is done, we take the pages in range from page
6705 * allocator removing them from the buddy system. This way
6706 * page allocator will never consider using them.
6708 * This lets us mark the pageblocks back as
6709 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6710 * aligned range but not in the unaligned, original range are
6711 * put back to page allocator so that buddy can use them.
6714 ret = start_isolate_page_range(pfn_max_align_down(start),
6715 pfn_max_align_up(end), migratetype,
6720 ret = __alloc_contig_migrate_range(&cc, start, end);
6725 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6726 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6727 * more, all pages in [start, end) are free in page allocator.
6728 * What we are going to do is to allocate all pages from
6729 * [start, end) (that is remove them from page allocator).
6731 * The only problem is that pages at the beginning and at the
6732 * end of interesting range may be not aligned with pages that
6733 * page allocator holds, ie. they can be part of higher order
6734 * pages. Because of this, we reserve the bigger range and
6735 * once this is done free the pages we are not interested in.
6737 * We don't have to hold zone->lock here because the pages are
6738 * isolated thus they won't get removed from buddy.
6741 lru_add_drain_all();
6742 drain_all_pages(cc.zone);
6745 outer_start = start;
6746 while (!PageBuddy(pfn_to_page(outer_start))) {
6747 if (++order >= MAX_ORDER) {
6751 outer_start &= ~0UL << order;
6754 /* Make sure the range is really isolated. */
6755 if (test_pages_isolated(outer_start, end, false)) {
6756 pr_info("%s: [%lx, %lx) PFNs busy\n",
6757 __func__, outer_start, end);
6762 /* Grab isolated pages from freelists. */
6763 outer_end = isolate_freepages_range(&cc, outer_start, end);
6769 /* Free head and tail (if any) */
6770 if (start != outer_start)
6771 free_contig_range(outer_start, start - outer_start);
6772 if (end != outer_end)
6773 free_contig_range(end, outer_end - end);
6776 undo_isolate_page_range(pfn_max_align_down(start),
6777 pfn_max_align_up(end), migratetype);
6781 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6783 unsigned int count = 0;
6785 for (; nr_pages--; pfn++) {
6786 struct page *page = pfn_to_page(pfn);
6788 count += page_count(page) != 1;
6791 WARN(count != 0, "%d pages are still in use!\n", count);
6795 #ifdef CONFIG_MEMORY_HOTPLUG
6797 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6798 * page high values need to be recalulated.
6800 void __meminit zone_pcp_update(struct zone *zone)
6803 mutex_lock(&pcp_batch_high_lock);
6804 for_each_possible_cpu(cpu)
6805 pageset_set_high_and_batch(zone,
6806 per_cpu_ptr(zone->pageset, cpu));
6807 mutex_unlock(&pcp_batch_high_lock);
6811 void zone_pcp_reset(struct zone *zone)
6813 unsigned long flags;
6815 struct per_cpu_pageset *pset;
6817 /* avoid races with drain_pages() */
6818 local_irq_save(flags);
6819 if (zone->pageset != &boot_pageset) {
6820 for_each_online_cpu(cpu) {
6821 pset = per_cpu_ptr(zone->pageset, cpu);
6822 drain_zonestat(zone, pset);
6824 free_percpu(zone->pageset);
6825 zone->pageset = &boot_pageset;
6827 local_irq_restore(flags);
6830 #ifdef CONFIG_MEMORY_HOTREMOVE
6832 * All pages in the range must be isolated before calling this.
6835 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6839 unsigned int order, i;
6841 unsigned long flags;
6842 /* find the first valid pfn */
6843 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6848 zone = page_zone(pfn_to_page(pfn));
6849 spin_lock_irqsave(&zone->lock, flags);
6851 while (pfn < end_pfn) {
6852 if (!pfn_valid(pfn)) {
6856 page = pfn_to_page(pfn);
6858 * The HWPoisoned page may be not in buddy system, and
6859 * page_count() is not 0.
6861 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6863 SetPageReserved(page);
6867 BUG_ON(page_count(page));
6868 BUG_ON(!PageBuddy(page));
6869 order = page_order(page);
6870 #ifdef CONFIG_DEBUG_VM
6871 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6872 pfn, 1 << order, end_pfn);
6874 list_del(&page->lru);
6875 rmv_page_order(page);
6876 zone->free_area[order].nr_free--;
6877 for (i = 0; i < (1 << order); i++)
6878 SetPageReserved((page+i));
6879 pfn += (1 << order);
6881 spin_unlock_irqrestore(&zone->lock, flags);
6885 #ifdef CONFIG_MEMORY_FAILURE
6886 bool is_free_buddy_page(struct page *page)
6888 struct zone *zone = page_zone(page);
6889 unsigned long pfn = page_to_pfn(page);
6890 unsigned long flags;
6893 spin_lock_irqsave(&zone->lock, flags);
6894 for (order = 0; order < MAX_ORDER; order++) {
6895 struct page *page_head = page - (pfn & ((1 << order) - 1));
6897 if (PageBuddy(page_head) && page_order(page_head) >= order)
6900 spin_unlock_irqrestore(&zone->lock, flags);
6902 return order < MAX_ORDER;