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".
450 * The remaining PAGE_SIZE pages are called "tail pages".
452 * All pages have PG_compound set. All tail pages have their ->first_page
453 * pointing at the head page.
455 * The first tail page's ->lru.next holds the address of the compound page's
456 * put_page() function. Its ->lru.prev 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 p->first_page = page;
477 /* Make sure p->first_page is always valid for PageTail() */
483 #ifdef CONFIG_DEBUG_PAGEALLOC
484 unsigned int _debug_guardpage_minorder;
485 bool _debug_pagealloc_enabled __read_mostly;
486 bool _debug_guardpage_enabled __read_mostly;
488 static int __init early_debug_pagealloc(char *buf)
493 if (strcmp(buf, "on") == 0)
494 _debug_pagealloc_enabled = true;
498 early_param("debug_pagealloc", early_debug_pagealloc);
500 static bool need_debug_guardpage(void)
502 /* If we don't use debug_pagealloc, we don't need guard page */
503 if (!debug_pagealloc_enabled())
509 static void init_debug_guardpage(void)
511 if (!debug_pagealloc_enabled())
514 _debug_guardpage_enabled = true;
517 struct page_ext_operations debug_guardpage_ops = {
518 .need = need_debug_guardpage,
519 .init = init_debug_guardpage,
522 static int __init debug_guardpage_minorder_setup(char *buf)
526 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
527 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
530 _debug_guardpage_minorder = res;
531 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
534 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
536 static inline void set_page_guard(struct zone *zone, struct page *page,
537 unsigned int order, int migratetype)
539 struct page_ext *page_ext;
541 if (!debug_guardpage_enabled())
544 page_ext = lookup_page_ext(page);
545 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
547 INIT_LIST_HEAD(&page->lru);
548 set_page_private(page, order);
549 /* Guard pages are not available for any usage */
550 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
553 static inline void clear_page_guard(struct zone *zone, struct page *page,
554 unsigned int order, int migratetype)
556 struct page_ext *page_ext;
558 if (!debug_guardpage_enabled())
561 page_ext = lookup_page_ext(page);
562 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
564 set_page_private(page, 0);
565 if (!is_migrate_isolate(migratetype))
566 __mod_zone_freepage_state(zone, (1 << order), migratetype);
569 struct page_ext_operations debug_guardpage_ops = { NULL, };
570 static inline void set_page_guard(struct zone *zone, struct page *page,
571 unsigned int order, int migratetype) {}
572 static inline void clear_page_guard(struct zone *zone, struct page *page,
573 unsigned int order, int migratetype) {}
576 static inline void set_page_order(struct page *page, unsigned int order)
578 set_page_private(page, order);
579 __SetPageBuddy(page);
582 static inline void rmv_page_order(struct page *page)
584 __ClearPageBuddy(page);
585 set_page_private(page, 0);
589 * This function checks whether a page is free && is the buddy
590 * we can do coalesce a page and its buddy if
591 * (a) the buddy is not in a hole &&
592 * (b) the buddy is in the buddy system &&
593 * (c) a page and its buddy have the same order &&
594 * (d) a page and its buddy are in the same zone.
596 * For recording whether a page is in the buddy system, we set ->_mapcount
597 * PAGE_BUDDY_MAPCOUNT_VALUE.
598 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
599 * serialized by zone->lock.
601 * For recording page's order, we use page_private(page).
603 static inline int page_is_buddy(struct page *page, struct page *buddy,
606 if (!pfn_valid_within(page_to_pfn(buddy)))
609 if (page_is_guard(buddy) && page_order(buddy) == order) {
610 if (page_zone_id(page) != page_zone_id(buddy))
613 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
618 if (PageBuddy(buddy) && page_order(buddy) == order) {
620 * zone check is done late to avoid uselessly
621 * calculating zone/node ids for pages that could
624 if (page_zone_id(page) != page_zone_id(buddy))
627 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
635 * Freeing function for a buddy system allocator.
637 * The concept of a buddy system is to maintain direct-mapped table
638 * (containing bit values) for memory blocks of various "orders".
639 * The bottom level table contains the map for the smallest allocatable
640 * units of memory (here, pages), and each level above it describes
641 * pairs of units from the levels below, hence, "buddies".
642 * At a high level, all that happens here is marking the table entry
643 * at the bottom level available, and propagating the changes upward
644 * as necessary, plus some accounting needed to play nicely with other
645 * parts of the VM system.
646 * At each level, we keep a list of pages, which are heads of continuous
647 * free pages of length of (1 << order) and marked with _mapcount
648 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
650 * So when we are allocating or freeing one, we can derive the state of the
651 * other. That is, if we allocate a small block, and both were
652 * free, the remainder of the region must be split into blocks.
653 * If a block is freed, and its buddy is also free, then this
654 * triggers coalescing into a block of larger size.
659 static inline void __free_one_page(struct page *page,
661 struct zone *zone, unsigned int order,
664 unsigned long page_idx;
665 unsigned long combined_idx;
666 unsigned long uninitialized_var(buddy_idx);
668 int max_order = MAX_ORDER;
670 VM_BUG_ON(!zone_is_initialized(zone));
671 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
673 VM_BUG_ON(migratetype == -1);
674 if (is_migrate_isolate(migratetype)) {
676 * We restrict max order of merging to prevent merge
677 * between freepages on isolate pageblock and normal
678 * pageblock. Without this, pageblock isolation
679 * could cause incorrect freepage accounting.
681 max_order = min(MAX_ORDER, pageblock_order + 1);
683 __mod_zone_freepage_state(zone, 1 << order, migratetype);
686 page_idx = pfn & ((1 << max_order) - 1);
688 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
689 VM_BUG_ON_PAGE(bad_range(zone, page), page);
691 while (order < max_order - 1) {
692 buddy_idx = __find_buddy_index(page_idx, order);
693 buddy = page + (buddy_idx - page_idx);
694 if (!page_is_buddy(page, buddy, order))
697 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
698 * merge with it and move up one order.
700 if (page_is_guard(buddy)) {
701 clear_page_guard(zone, buddy, order, migratetype);
703 list_del(&buddy->lru);
704 zone->free_area[order].nr_free--;
705 rmv_page_order(buddy);
707 combined_idx = buddy_idx & page_idx;
708 page = page + (combined_idx - page_idx);
709 page_idx = combined_idx;
712 set_page_order(page, order);
715 * If this is not the largest possible page, check if the buddy
716 * of the next-highest order is free. If it is, it's possible
717 * that pages are being freed that will coalesce soon. In case,
718 * that is happening, add the free page to the tail of the list
719 * so it's less likely to be used soon and more likely to be merged
720 * as a higher order page
722 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
723 struct page *higher_page, *higher_buddy;
724 combined_idx = buddy_idx & page_idx;
725 higher_page = page + (combined_idx - page_idx);
726 buddy_idx = __find_buddy_index(combined_idx, order + 1);
727 higher_buddy = higher_page + (buddy_idx - combined_idx);
728 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
729 list_add_tail(&page->lru,
730 &zone->free_area[order].free_list[migratetype]);
735 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
737 zone->free_area[order].nr_free++;
740 static inline int free_pages_check(struct page *page)
742 const char *bad_reason = NULL;
743 unsigned long bad_flags = 0;
745 if (unlikely(page_mapcount(page)))
746 bad_reason = "nonzero mapcount";
747 if (unlikely(page->mapping != NULL))
748 bad_reason = "non-NULL mapping";
749 if (unlikely(atomic_read(&page->_count) != 0))
750 bad_reason = "nonzero _count";
751 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
752 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
753 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
756 if (unlikely(page->mem_cgroup))
757 bad_reason = "page still charged to cgroup";
759 if (unlikely(bad_reason)) {
760 bad_page(page, bad_reason, bad_flags);
763 page_cpupid_reset_last(page);
764 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
765 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
770 * Frees a number of pages from the PCP lists
771 * Assumes all pages on list are in same zone, and of same order.
772 * count is the number of pages to free.
774 * If the zone was previously in an "all pages pinned" state then look to
775 * see if this freeing clears that state.
777 * And clear the zone's pages_scanned counter, to hold off the "all pages are
778 * pinned" detection logic.
780 static void free_pcppages_bulk(struct zone *zone, int count,
781 struct per_cpu_pages *pcp)
786 unsigned long nr_scanned;
788 spin_lock(&zone->lock);
789 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
791 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
795 struct list_head *list;
798 * Remove pages from lists in a round-robin fashion. A
799 * batch_free count is maintained that is incremented when an
800 * empty list is encountered. This is so more pages are freed
801 * off fuller lists instead of spinning excessively around empty
806 if (++migratetype == MIGRATE_PCPTYPES)
808 list = &pcp->lists[migratetype];
809 } while (list_empty(list));
811 /* This is the only non-empty list. Free them all. */
812 if (batch_free == MIGRATE_PCPTYPES)
813 batch_free = to_free;
816 int mt; /* migratetype of the to-be-freed page */
818 page = list_entry(list->prev, struct page, lru);
819 /* must delete as __free_one_page list manipulates */
820 list_del(&page->lru);
822 mt = get_pcppage_migratetype(page);
823 /* MIGRATE_ISOLATE page should not go to pcplists */
824 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
825 /* Pageblock could have been isolated meanwhile */
826 if (unlikely(has_isolate_pageblock(zone)))
827 mt = get_pageblock_migratetype(page);
829 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
830 trace_mm_page_pcpu_drain(page, 0, mt);
831 } while (--to_free && --batch_free && !list_empty(list));
833 spin_unlock(&zone->lock);
836 static void free_one_page(struct zone *zone,
837 struct page *page, unsigned long pfn,
841 unsigned long nr_scanned;
842 spin_lock(&zone->lock);
843 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
845 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
847 if (unlikely(has_isolate_pageblock(zone) ||
848 is_migrate_isolate(migratetype))) {
849 migratetype = get_pfnblock_migratetype(page, pfn);
851 __free_one_page(page, pfn, zone, order, migratetype);
852 spin_unlock(&zone->lock);
855 static int free_tail_pages_check(struct page *head_page, struct page *page)
857 if (!IS_ENABLED(CONFIG_DEBUG_VM))
859 if (unlikely(!PageTail(page))) {
860 bad_page(page, "PageTail not set", 0);
863 if (unlikely(page->first_page != head_page)) {
864 bad_page(page, "first_page not consistent", 0);
870 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
871 unsigned long zone, int nid)
873 set_page_links(page, zone, nid, pfn);
874 init_page_count(page);
875 page_mapcount_reset(page);
876 page_cpupid_reset_last(page);
878 INIT_LIST_HEAD(&page->lru);
879 #ifdef WANT_PAGE_VIRTUAL
880 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
881 if (!is_highmem_idx(zone))
882 set_page_address(page, __va(pfn << PAGE_SHIFT));
886 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
889 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
892 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
893 static void init_reserved_page(unsigned long pfn)
898 if (!early_page_uninitialised(pfn))
901 nid = early_pfn_to_nid(pfn);
902 pgdat = NODE_DATA(nid);
904 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
905 struct zone *zone = &pgdat->node_zones[zid];
907 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
910 __init_single_pfn(pfn, zid, nid);
913 static inline void init_reserved_page(unsigned long pfn)
916 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
919 * Initialised pages do not have PageReserved set. This function is
920 * called for each range allocated by the bootmem allocator and
921 * marks the pages PageReserved. The remaining valid pages are later
922 * sent to the buddy page allocator.
924 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
926 unsigned long start_pfn = PFN_DOWN(start);
927 unsigned long end_pfn = PFN_UP(end);
929 for (; start_pfn < end_pfn; start_pfn++) {
930 if (pfn_valid(start_pfn)) {
931 struct page *page = pfn_to_page(start_pfn);
933 init_reserved_page(start_pfn);
934 SetPageReserved(page);
939 static bool free_pages_prepare(struct page *page, unsigned int order)
941 bool compound = PageCompound(page);
944 VM_BUG_ON_PAGE(PageTail(page), page);
945 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
947 trace_mm_page_free(page, order);
948 kmemcheck_free_shadow(page, order);
949 kasan_free_pages(page, order);
952 page->mapping = NULL;
953 bad += free_pages_check(page);
954 for (i = 1; i < (1 << order); i++) {
956 bad += free_tail_pages_check(page, page + i);
957 bad += free_pages_check(page + i);
962 reset_page_owner(page, order);
964 if (!PageHighMem(page)) {
965 debug_check_no_locks_freed(page_address(page),
967 debug_check_no_obj_freed(page_address(page),
970 arch_free_page(page, order);
971 kernel_map_pages(page, 1 << order, 0);
976 static void __free_pages_ok(struct page *page, unsigned int order)
980 unsigned long pfn = page_to_pfn(page);
982 if (!free_pages_prepare(page, order))
985 migratetype = get_pfnblock_migratetype(page, pfn);
986 local_irq_save(flags);
987 __count_vm_events(PGFREE, 1 << order);
988 free_one_page(page_zone(page), page, pfn, order, migratetype);
989 local_irq_restore(flags);
992 static void __init __free_pages_boot_core(struct page *page,
993 unsigned long pfn, unsigned int order)
995 unsigned int nr_pages = 1 << order;
996 struct page *p = page;
1000 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1002 __ClearPageReserved(p);
1003 set_page_count(p, 0);
1005 __ClearPageReserved(p);
1006 set_page_count(p, 0);
1008 page_zone(page)->managed_pages += nr_pages;
1009 set_page_refcounted(page);
1010 __free_pages(page, order);
1013 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1014 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1016 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1018 int __meminit early_pfn_to_nid(unsigned long pfn)
1020 static DEFINE_SPINLOCK(early_pfn_lock);
1023 spin_lock(&early_pfn_lock);
1024 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1027 spin_unlock(&early_pfn_lock);
1033 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1034 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1035 struct mminit_pfnnid_cache *state)
1039 nid = __early_pfn_to_nid(pfn, state);
1040 if (nid >= 0 && nid != node)
1045 /* Only safe to use early in boot when initialisation is single-threaded */
1046 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1048 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1053 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1057 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1058 struct mminit_pfnnid_cache *state)
1065 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1068 if (early_page_uninitialised(pfn))
1070 return __free_pages_boot_core(page, pfn, order);
1073 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1074 static void __init deferred_free_range(struct page *page,
1075 unsigned long pfn, int nr_pages)
1082 /* Free a large naturally-aligned chunk if possible */
1083 if (nr_pages == MAX_ORDER_NR_PAGES &&
1084 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1085 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1086 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1090 for (i = 0; i < nr_pages; i++, page++, pfn++)
1091 __free_pages_boot_core(page, pfn, 0);
1094 /* Completion tracking for deferred_init_memmap() threads */
1095 static atomic_t pgdat_init_n_undone __initdata;
1096 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1098 static inline void __init pgdat_init_report_one_done(void)
1100 if (atomic_dec_and_test(&pgdat_init_n_undone))
1101 complete(&pgdat_init_all_done_comp);
1104 /* Initialise remaining memory on a node */
1105 static int __init deferred_init_memmap(void *data)
1107 pg_data_t *pgdat = data;
1108 int nid = pgdat->node_id;
1109 struct mminit_pfnnid_cache nid_init_state = { };
1110 unsigned long start = jiffies;
1111 unsigned long nr_pages = 0;
1112 unsigned long walk_start, walk_end;
1115 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1116 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1118 if (first_init_pfn == ULONG_MAX) {
1119 pgdat_init_report_one_done();
1123 /* Bind memory initialisation thread to a local node if possible */
1124 if (!cpumask_empty(cpumask))
1125 set_cpus_allowed_ptr(current, cpumask);
1127 /* Sanity check boundaries */
1128 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1129 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1130 pgdat->first_deferred_pfn = ULONG_MAX;
1132 /* Only the highest zone is deferred so find it */
1133 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1134 zone = pgdat->node_zones + zid;
1135 if (first_init_pfn < zone_end_pfn(zone))
1139 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1140 unsigned long pfn, end_pfn;
1141 struct page *page = NULL;
1142 struct page *free_base_page = NULL;
1143 unsigned long free_base_pfn = 0;
1146 end_pfn = min(walk_end, zone_end_pfn(zone));
1147 pfn = first_init_pfn;
1148 if (pfn < walk_start)
1150 if (pfn < zone->zone_start_pfn)
1151 pfn = zone->zone_start_pfn;
1153 for (; pfn < end_pfn; pfn++) {
1154 if (!pfn_valid_within(pfn))
1158 * Ensure pfn_valid is checked every
1159 * MAX_ORDER_NR_PAGES for memory holes
1161 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1162 if (!pfn_valid(pfn)) {
1168 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1173 /* Minimise pfn page lookups and scheduler checks */
1174 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1177 nr_pages += nr_to_free;
1178 deferred_free_range(free_base_page,
1179 free_base_pfn, nr_to_free);
1180 free_base_page = NULL;
1181 free_base_pfn = nr_to_free = 0;
1183 page = pfn_to_page(pfn);
1188 VM_BUG_ON(page_zone(page) != zone);
1192 __init_single_page(page, pfn, zid, nid);
1193 if (!free_base_page) {
1194 free_base_page = page;
1195 free_base_pfn = pfn;
1200 /* Where possible, batch up pages for a single free */
1203 /* Free the current block of pages to allocator */
1204 nr_pages += nr_to_free;
1205 deferred_free_range(free_base_page, free_base_pfn,
1207 free_base_page = NULL;
1208 free_base_pfn = nr_to_free = 0;
1211 first_init_pfn = max(end_pfn, first_init_pfn);
1214 /* Sanity check that the next zone really is unpopulated */
1215 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1217 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1218 jiffies_to_msecs(jiffies - start));
1220 pgdat_init_report_one_done();
1224 void __init page_alloc_init_late(void)
1228 /* There will be num_node_state(N_MEMORY) threads */
1229 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1230 for_each_node_state(nid, N_MEMORY) {
1231 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1234 /* Block until all are initialised */
1235 wait_for_completion(&pgdat_init_all_done_comp);
1237 /* Reinit limits that are based on free pages after the kernel is up */
1238 files_maxfiles_init();
1240 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1243 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1244 void __init init_cma_reserved_pageblock(struct page *page)
1246 unsigned i = pageblock_nr_pages;
1247 struct page *p = page;
1250 __ClearPageReserved(p);
1251 set_page_count(p, 0);
1254 set_pageblock_migratetype(page, MIGRATE_CMA);
1256 if (pageblock_order >= MAX_ORDER) {
1257 i = pageblock_nr_pages;
1260 set_page_refcounted(p);
1261 __free_pages(p, MAX_ORDER - 1);
1262 p += MAX_ORDER_NR_PAGES;
1263 } while (i -= MAX_ORDER_NR_PAGES);
1265 set_page_refcounted(page);
1266 __free_pages(page, pageblock_order);
1269 adjust_managed_page_count(page, pageblock_nr_pages);
1274 * The order of subdivision here is critical for the IO subsystem.
1275 * Please do not alter this order without good reasons and regression
1276 * testing. Specifically, as large blocks of memory are subdivided,
1277 * the order in which smaller blocks are delivered depends on the order
1278 * they're subdivided in this function. This is the primary factor
1279 * influencing the order in which pages are delivered to the IO
1280 * subsystem according to empirical testing, and this is also justified
1281 * by considering the behavior of a buddy system containing a single
1282 * large block of memory acted on by a series of small allocations.
1283 * This behavior is a critical factor in sglist merging's success.
1287 static inline void expand(struct zone *zone, struct page *page,
1288 int low, int high, struct free_area *area,
1291 unsigned long size = 1 << high;
1293 while (high > low) {
1297 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1299 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1300 debug_guardpage_enabled() &&
1301 high < debug_guardpage_minorder()) {
1303 * Mark as guard pages (or page), that will allow to
1304 * merge back to allocator when buddy will be freed.
1305 * Corresponding page table entries will not be touched,
1306 * pages will stay not present in virtual address space
1308 set_page_guard(zone, &page[size], high, migratetype);
1311 list_add(&page[size].lru, &area->free_list[migratetype]);
1313 set_page_order(&page[size], high);
1318 * This page is about to be returned from the page allocator
1320 static inline int check_new_page(struct page *page)
1322 const char *bad_reason = NULL;
1323 unsigned long bad_flags = 0;
1325 if (unlikely(page_mapcount(page)))
1326 bad_reason = "nonzero mapcount";
1327 if (unlikely(page->mapping != NULL))
1328 bad_reason = "non-NULL mapping";
1329 if (unlikely(atomic_read(&page->_count) != 0))
1330 bad_reason = "nonzero _count";
1331 if (unlikely(page->flags & __PG_HWPOISON)) {
1332 bad_reason = "HWPoisoned (hardware-corrupted)";
1333 bad_flags = __PG_HWPOISON;
1335 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1336 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1337 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1340 if (unlikely(page->mem_cgroup))
1341 bad_reason = "page still charged to cgroup";
1343 if (unlikely(bad_reason)) {
1344 bad_page(page, bad_reason, bad_flags);
1350 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1355 for (i = 0; i < (1 << order); i++) {
1356 struct page *p = page + i;
1357 if (unlikely(check_new_page(p)))
1361 set_page_private(page, 0);
1362 set_page_refcounted(page);
1364 arch_alloc_page(page, order);
1365 kernel_map_pages(page, 1 << order, 1);
1366 kasan_alloc_pages(page, order);
1368 if (gfp_flags & __GFP_ZERO)
1369 for (i = 0; i < (1 << order); i++)
1370 clear_highpage(page + i);
1372 if (order && (gfp_flags & __GFP_COMP))
1373 prep_compound_page(page, order);
1375 set_page_owner(page, order, gfp_flags);
1378 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1379 * allocate the page. The expectation is that the caller is taking
1380 * steps that will free more memory. The caller should avoid the page
1381 * being used for !PFMEMALLOC purposes.
1383 if (alloc_flags & ALLOC_NO_WATERMARKS)
1384 set_page_pfmemalloc(page);
1386 clear_page_pfmemalloc(page);
1392 * Go through the free lists for the given migratetype and remove
1393 * the smallest available page from the freelists
1396 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1399 unsigned int current_order;
1400 struct free_area *area;
1403 /* Find a page of the appropriate size in the preferred list */
1404 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1405 area = &(zone->free_area[current_order]);
1406 if (list_empty(&area->free_list[migratetype]))
1409 page = list_entry(area->free_list[migratetype].next,
1411 list_del(&page->lru);
1412 rmv_page_order(page);
1414 expand(zone, page, order, current_order, area, migratetype);
1415 set_pcppage_migratetype(page, migratetype);
1424 * This array describes the order lists are fallen back to when
1425 * the free lists for the desirable migrate type are depleted
1427 static int fallbacks[MIGRATE_TYPES][4] = {
1428 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1429 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1430 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1432 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1434 #ifdef CONFIG_MEMORY_ISOLATION
1435 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1440 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1443 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1446 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1447 unsigned int order) { return NULL; }
1451 * Move the free pages in a range to the free lists of the requested type.
1452 * Note that start_page and end_pages are not aligned on a pageblock
1453 * boundary. If alignment is required, use move_freepages_block()
1455 int move_freepages(struct zone *zone,
1456 struct page *start_page, struct page *end_page,
1460 unsigned long order;
1461 int pages_moved = 0;
1463 #ifndef CONFIG_HOLES_IN_ZONE
1465 * page_zone is not safe to call in this context when
1466 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1467 * anyway as we check zone boundaries in move_freepages_block().
1468 * Remove at a later date when no bug reports exist related to
1469 * grouping pages by mobility
1471 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1474 for (page = start_page; page <= end_page;) {
1475 /* Make sure we are not inadvertently changing nodes */
1476 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1478 if (!pfn_valid_within(page_to_pfn(page))) {
1483 if (!PageBuddy(page)) {
1488 order = page_order(page);
1489 list_move(&page->lru,
1490 &zone->free_area[order].free_list[migratetype]);
1492 pages_moved += 1 << order;
1498 int move_freepages_block(struct zone *zone, struct page *page,
1501 unsigned long start_pfn, end_pfn;
1502 struct page *start_page, *end_page;
1504 start_pfn = page_to_pfn(page);
1505 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1506 start_page = pfn_to_page(start_pfn);
1507 end_page = start_page + pageblock_nr_pages - 1;
1508 end_pfn = start_pfn + pageblock_nr_pages - 1;
1510 /* Do not cross zone boundaries */
1511 if (!zone_spans_pfn(zone, start_pfn))
1513 if (!zone_spans_pfn(zone, end_pfn))
1516 return move_freepages(zone, start_page, end_page, migratetype);
1519 static void change_pageblock_range(struct page *pageblock_page,
1520 int start_order, int migratetype)
1522 int nr_pageblocks = 1 << (start_order - pageblock_order);
1524 while (nr_pageblocks--) {
1525 set_pageblock_migratetype(pageblock_page, migratetype);
1526 pageblock_page += pageblock_nr_pages;
1531 * When we are falling back to another migratetype during allocation, try to
1532 * steal extra free pages from the same pageblocks to satisfy further
1533 * allocations, instead of polluting multiple pageblocks.
1535 * If we are stealing a relatively large buddy page, it is likely there will
1536 * be more free pages in the pageblock, so try to steal them all. For
1537 * reclaimable and unmovable allocations, we steal regardless of page size,
1538 * as fragmentation caused by those allocations polluting movable pageblocks
1539 * is worse than movable allocations stealing from unmovable and reclaimable
1542 static bool can_steal_fallback(unsigned int order, int start_mt)
1545 * Leaving this order check is intended, although there is
1546 * relaxed order check in next check. The reason is that
1547 * we can actually steal whole pageblock if this condition met,
1548 * but, below check doesn't guarantee it and that is just heuristic
1549 * so could be changed anytime.
1551 if (order >= pageblock_order)
1554 if (order >= pageblock_order / 2 ||
1555 start_mt == MIGRATE_RECLAIMABLE ||
1556 start_mt == MIGRATE_UNMOVABLE ||
1557 page_group_by_mobility_disabled)
1564 * This function implements actual steal behaviour. If order is large enough,
1565 * we can steal whole pageblock. If not, we first move freepages in this
1566 * pageblock and check whether half of pages are moved or not. If half of
1567 * pages are moved, we can change migratetype of pageblock and permanently
1568 * use it's pages as requested migratetype in the future.
1570 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1573 int current_order = page_order(page);
1576 /* Take ownership for orders >= pageblock_order */
1577 if (current_order >= pageblock_order) {
1578 change_pageblock_range(page, current_order, start_type);
1582 pages = move_freepages_block(zone, page, start_type);
1584 /* Claim the whole block if over half of it is free */
1585 if (pages >= (1 << (pageblock_order-1)) ||
1586 page_group_by_mobility_disabled)
1587 set_pageblock_migratetype(page, start_type);
1591 * Check whether there is a suitable fallback freepage with requested order.
1592 * If only_stealable is true, this function returns fallback_mt only if
1593 * we can steal other freepages all together. This would help to reduce
1594 * fragmentation due to mixed migratetype pages in one pageblock.
1596 int find_suitable_fallback(struct free_area *area, unsigned int order,
1597 int migratetype, bool only_stealable, bool *can_steal)
1602 if (area->nr_free == 0)
1607 fallback_mt = fallbacks[migratetype][i];
1608 if (fallback_mt == MIGRATE_TYPES)
1611 if (list_empty(&area->free_list[fallback_mt]))
1614 if (can_steal_fallback(order, migratetype))
1617 if (!only_stealable)
1628 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1629 * there are no empty page blocks that contain a page with a suitable order
1631 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1632 unsigned int alloc_order)
1635 unsigned long max_managed, flags;
1638 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1639 * Check is race-prone but harmless.
1641 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1642 if (zone->nr_reserved_highatomic >= max_managed)
1645 spin_lock_irqsave(&zone->lock, flags);
1647 /* Recheck the nr_reserved_highatomic limit under the lock */
1648 if (zone->nr_reserved_highatomic >= max_managed)
1652 mt = get_pageblock_migratetype(page);
1653 if (mt != MIGRATE_HIGHATOMIC &&
1654 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1655 zone->nr_reserved_highatomic += pageblock_nr_pages;
1656 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1657 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1661 spin_unlock_irqrestore(&zone->lock, flags);
1665 * Used when an allocation is about to fail under memory pressure. This
1666 * potentially hurts the reliability of high-order allocations when under
1667 * intense memory pressure but failed atomic allocations should be easier
1668 * to recover from than an OOM.
1670 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1672 struct zonelist *zonelist = ac->zonelist;
1673 unsigned long flags;
1679 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1681 /* Preserve at least one pageblock */
1682 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1685 spin_lock_irqsave(&zone->lock, flags);
1686 for (order = 0; order < MAX_ORDER; order++) {
1687 struct free_area *area = &(zone->free_area[order]);
1689 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1692 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1696 * It should never happen but changes to locking could
1697 * inadvertently allow a per-cpu drain to add pages
1698 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1699 * and watch for underflows.
1701 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1702 zone->nr_reserved_highatomic);
1705 * Convert to ac->migratetype and avoid the normal
1706 * pageblock stealing heuristics. Minimally, the caller
1707 * is doing the work and needs the pages. More
1708 * importantly, if the block was always converted to
1709 * MIGRATE_UNMOVABLE or another type then the number
1710 * of pageblocks that cannot be completely freed
1713 set_pageblock_migratetype(page, ac->migratetype);
1714 move_freepages_block(zone, page, ac->migratetype);
1715 spin_unlock_irqrestore(&zone->lock, flags);
1718 spin_unlock_irqrestore(&zone->lock, flags);
1722 /* Remove an element from the buddy allocator from the fallback list */
1723 static inline struct page *
1724 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1726 struct free_area *area;
1727 unsigned int current_order;
1732 /* Find the largest possible block of pages in the other list */
1733 for (current_order = MAX_ORDER-1;
1734 current_order >= order && current_order <= MAX_ORDER-1;
1736 area = &(zone->free_area[current_order]);
1737 fallback_mt = find_suitable_fallback(area, current_order,
1738 start_migratetype, false, &can_steal);
1739 if (fallback_mt == -1)
1742 page = list_entry(area->free_list[fallback_mt].next,
1745 steal_suitable_fallback(zone, page, start_migratetype);
1747 /* Remove the page from the freelists */
1749 list_del(&page->lru);
1750 rmv_page_order(page);
1752 expand(zone, page, order, current_order, area,
1755 * The pcppage_migratetype may differ from pageblock's
1756 * migratetype depending on the decisions in
1757 * find_suitable_fallback(). This is OK as long as it does not
1758 * differ for MIGRATE_CMA pageblocks. Those can be used as
1759 * fallback only via special __rmqueue_cma_fallback() function
1761 set_pcppage_migratetype(page, start_migratetype);
1763 trace_mm_page_alloc_extfrag(page, order, current_order,
1764 start_migratetype, fallback_mt);
1773 * Do the hard work of removing an element from the buddy allocator.
1774 * Call me with the zone->lock already held.
1776 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1777 int migratetype, gfp_t gfp_flags)
1781 page = __rmqueue_smallest(zone, order, migratetype);
1782 if (unlikely(!page)) {
1783 if (migratetype == MIGRATE_MOVABLE)
1784 page = __rmqueue_cma_fallback(zone, order);
1787 page = __rmqueue_fallback(zone, order, migratetype);
1790 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1795 * Obtain a specified number of elements from the buddy allocator, all under
1796 * a single hold of the lock, for efficiency. Add them to the supplied list.
1797 * Returns the number of new pages which were placed at *list.
1799 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1800 unsigned long count, struct list_head *list,
1801 int migratetype, bool cold)
1805 spin_lock(&zone->lock);
1806 for (i = 0; i < count; ++i) {
1807 struct page *page = __rmqueue(zone, order, migratetype, 0);
1808 if (unlikely(page == NULL))
1812 * Split buddy pages returned by expand() are received here
1813 * in physical page order. The page is added to the callers and
1814 * list and the list head then moves forward. From the callers
1815 * perspective, the linked list is ordered by page number in
1816 * some conditions. This is useful for IO devices that can
1817 * merge IO requests if the physical pages are ordered
1821 list_add(&page->lru, list);
1823 list_add_tail(&page->lru, list);
1825 if (is_migrate_cma(get_pcppage_migratetype(page)))
1826 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1829 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1830 spin_unlock(&zone->lock);
1836 * Called from the vmstat counter updater to drain pagesets of this
1837 * currently executing processor on remote nodes after they have
1840 * Note that this function must be called with the thread pinned to
1841 * a single processor.
1843 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1845 unsigned long flags;
1846 int to_drain, batch;
1848 local_irq_save(flags);
1849 batch = READ_ONCE(pcp->batch);
1850 to_drain = min(pcp->count, batch);
1852 free_pcppages_bulk(zone, to_drain, pcp);
1853 pcp->count -= to_drain;
1855 local_irq_restore(flags);
1860 * Drain pcplists of the indicated processor and zone.
1862 * The processor must either be the current processor and the
1863 * thread pinned to the current processor or a processor that
1866 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1868 unsigned long flags;
1869 struct per_cpu_pageset *pset;
1870 struct per_cpu_pages *pcp;
1872 local_irq_save(flags);
1873 pset = per_cpu_ptr(zone->pageset, cpu);
1877 free_pcppages_bulk(zone, pcp->count, pcp);
1880 local_irq_restore(flags);
1884 * Drain pcplists of all zones on the indicated processor.
1886 * The processor must either be the current processor and the
1887 * thread pinned to the current processor or a processor that
1890 static void drain_pages(unsigned int cpu)
1894 for_each_populated_zone(zone) {
1895 drain_pages_zone(cpu, zone);
1900 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1902 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1903 * the single zone's pages.
1905 void drain_local_pages(struct zone *zone)
1907 int cpu = smp_processor_id();
1910 drain_pages_zone(cpu, zone);
1916 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1918 * When zone parameter is non-NULL, spill just the single zone's pages.
1920 * Note that this code is protected against sending an IPI to an offline
1921 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1922 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1923 * nothing keeps CPUs from showing up after we populated the cpumask and
1924 * before the call to on_each_cpu_mask().
1926 void drain_all_pages(struct zone *zone)
1931 * Allocate in the BSS so we wont require allocation in
1932 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1934 static cpumask_t cpus_with_pcps;
1937 * We don't care about racing with CPU hotplug event
1938 * as offline notification will cause the notified
1939 * cpu to drain that CPU pcps and on_each_cpu_mask
1940 * disables preemption as part of its processing
1942 for_each_online_cpu(cpu) {
1943 struct per_cpu_pageset *pcp;
1945 bool has_pcps = false;
1948 pcp = per_cpu_ptr(zone->pageset, cpu);
1952 for_each_populated_zone(z) {
1953 pcp = per_cpu_ptr(z->pageset, cpu);
1954 if (pcp->pcp.count) {
1962 cpumask_set_cpu(cpu, &cpus_with_pcps);
1964 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1966 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1970 #ifdef CONFIG_HIBERNATION
1972 void mark_free_pages(struct zone *zone)
1974 unsigned long pfn, max_zone_pfn;
1975 unsigned long flags;
1976 unsigned int order, t;
1977 struct list_head *curr;
1979 if (zone_is_empty(zone))
1982 spin_lock_irqsave(&zone->lock, flags);
1984 max_zone_pfn = zone_end_pfn(zone);
1985 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1986 if (pfn_valid(pfn)) {
1987 struct page *page = pfn_to_page(pfn);
1989 if (!swsusp_page_is_forbidden(page))
1990 swsusp_unset_page_free(page);
1993 for_each_migratetype_order(order, t) {
1994 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1997 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1998 for (i = 0; i < (1UL << order); i++)
1999 swsusp_set_page_free(pfn_to_page(pfn + i));
2002 spin_unlock_irqrestore(&zone->lock, flags);
2004 #endif /* CONFIG_PM */
2007 * Free a 0-order page
2008 * cold == true ? free a cold page : free a hot page
2010 void free_hot_cold_page(struct page *page, bool cold)
2012 struct zone *zone = page_zone(page);
2013 struct per_cpu_pages *pcp;
2014 unsigned long flags;
2015 unsigned long pfn = page_to_pfn(page);
2018 if (!free_pages_prepare(page, 0))
2021 migratetype = get_pfnblock_migratetype(page, pfn);
2022 set_pcppage_migratetype(page, migratetype);
2023 local_irq_save(flags);
2024 __count_vm_event(PGFREE);
2027 * We only track unmovable, reclaimable and movable on pcp lists.
2028 * Free ISOLATE pages back to the allocator because they are being
2029 * offlined but treat RESERVE as movable pages so we can get those
2030 * areas back if necessary. Otherwise, we may have to free
2031 * excessively into the page allocator
2033 if (migratetype >= MIGRATE_PCPTYPES) {
2034 if (unlikely(is_migrate_isolate(migratetype))) {
2035 free_one_page(zone, page, pfn, 0, migratetype);
2038 migratetype = MIGRATE_MOVABLE;
2041 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2043 list_add(&page->lru, &pcp->lists[migratetype]);
2045 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2047 if (pcp->count >= pcp->high) {
2048 unsigned long batch = READ_ONCE(pcp->batch);
2049 free_pcppages_bulk(zone, batch, pcp);
2050 pcp->count -= batch;
2054 local_irq_restore(flags);
2058 * Free a list of 0-order pages
2060 void free_hot_cold_page_list(struct list_head *list, bool cold)
2062 struct page *page, *next;
2064 list_for_each_entry_safe(page, next, list, lru) {
2065 trace_mm_page_free_batched(page, cold);
2066 free_hot_cold_page(page, cold);
2071 * split_page takes a non-compound higher-order page, and splits it into
2072 * n (1<<order) sub-pages: page[0..n]
2073 * Each sub-page must be freed individually.
2075 * Note: this is probably too low level an operation for use in drivers.
2076 * Please consult with lkml before using this in your driver.
2078 void split_page(struct page *page, unsigned int order)
2083 VM_BUG_ON_PAGE(PageCompound(page), page);
2084 VM_BUG_ON_PAGE(!page_count(page), page);
2086 #ifdef CONFIG_KMEMCHECK
2088 * Split shadow pages too, because free(page[0]) would
2089 * otherwise free the whole shadow.
2091 if (kmemcheck_page_is_tracked(page))
2092 split_page(virt_to_page(page[0].shadow), order);
2095 gfp_mask = get_page_owner_gfp(page);
2096 set_page_owner(page, 0, gfp_mask);
2097 for (i = 1; i < (1 << order); i++) {
2098 set_page_refcounted(page + i);
2099 set_page_owner(page + i, 0, gfp_mask);
2102 EXPORT_SYMBOL_GPL(split_page);
2104 int __isolate_free_page(struct page *page, unsigned int order)
2106 unsigned long watermark;
2110 BUG_ON(!PageBuddy(page));
2112 zone = page_zone(page);
2113 mt = get_pageblock_migratetype(page);
2115 if (!is_migrate_isolate(mt)) {
2116 /* Obey watermarks as if the page was being allocated */
2117 watermark = low_wmark_pages(zone) + (1 << order);
2118 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2121 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2124 /* Remove page from free list */
2125 list_del(&page->lru);
2126 zone->free_area[order].nr_free--;
2127 rmv_page_order(page);
2129 set_page_owner(page, order, __GFP_MOVABLE);
2131 /* Set the pageblock if the isolated page is at least a pageblock */
2132 if (order >= pageblock_order - 1) {
2133 struct page *endpage = page + (1 << order) - 1;
2134 for (; page < endpage; page += pageblock_nr_pages) {
2135 int mt = get_pageblock_migratetype(page);
2136 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2137 set_pageblock_migratetype(page,
2143 return 1UL << order;
2147 * Similar to split_page except the page is already free. As this is only
2148 * being used for migration, the migratetype of the block also changes.
2149 * As this is called with interrupts disabled, the caller is responsible
2150 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2153 * Note: this is probably too low level an operation for use in drivers.
2154 * Please consult with lkml before using this in your driver.
2156 int split_free_page(struct page *page)
2161 order = page_order(page);
2163 nr_pages = __isolate_free_page(page, order);
2167 /* Split into individual pages */
2168 set_page_refcounted(page);
2169 split_page(page, order);
2174 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2177 struct page *buffered_rmqueue(struct zone *preferred_zone,
2178 struct zone *zone, unsigned int order,
2179 gfp_t gfp_flags, int alloc_flags, int migratetype)
2181 unsigned long flags;
2183 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2185 if (likely(order == 0)) {
2186 struct per_cpu_pages *pcp;
2187 struct list_head *list;
2189 local_irq_save(flags);
2190 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2191 list = &pcp->lists[migratetype];
2192 if (list_empty(list)) {
2193 pcp->count += rmqueue_bulk(zone, 0,
2196 if (unlikely(list_empty(list)))
2201 page = list_entry(list->prev, struct page, lru);
2203 page = list_entry(list->next, struct page, lru);
2205 list_del(&page->lru);
2208 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2210 * __GFP_NOFAIL is not to be used in new code.
2212 * All __GFP_NOFAIL callers should be fixed so that they
2213 * properly detect and handle allocation failures.
2215 * We most definitely don't want callers attempting to
2216 * allocate greater than order-1 page units with
2219 WARN_ON_ONCE(order > 1);
2221 spin_lock_irqsave(&zone->lock, flags);
2224 if (alloc_flags & ALLOC_HARDER) {
2225 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2227 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2230 page = __rmqueue(zone, order, migratetype, gfp_flags);
2231 spin_unlock(&zone->lock);
2234 __mod_zone_freepage_state(zone, -(1 << order),
2235 get_pcppage_migratetype(page));
2238 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2239 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2240 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2241 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2243 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2244 zone_statistics(preferred_zone, zone, gfp_flags);
2245 local_irq_restore(flags);
2247 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2251 local_irq_restore(flags);
2255 #ifdef CONFIG_FAIL_PAGE_ALLOC
2258 struct fault_attr attr;
2260 bool ignore_gfp_highmem;
2261 bool ignore_gfp_reclaim;
2263 } fail_page_alloc = {
2264 .attr = FAULT_ATTR_INITIALIZER,
2265 .ignore_gfp_reclaim = true,
2266 .ignore_gfp_highmem = true,
2270 static int __init setup_fail_page_alloc(char *str)
2272 return setup_fault_attr(&fail_page_alloc.attr, str);
2274 __setup("fail_page_alloc=", setup_fail_page_alloc);
2276 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2278 if (order < fail_page_alloc.min_order)
2280 if (gfp_mask & __GFP_NOFAIL)
2282 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2284 if (fail_page_alloc.ignore_gfp_reclaim &&
2285 (gfp_mask & __GFP_DIRECT_RECLAIM))
2288 return should_fail(&fail_page_alloc.attr, 1 << order);
2291 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2293 static int __init fail_page_alloc_debugfs(void)
2295 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2298 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2299 &fail_page_alloc.attr);
2301 return PTR_ERR(dir);
2303 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2304 &fail_page_alloc.ignore_gfp_reclaim))
2306 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2307 &fail_page_alloc.ignore_gfp_highmem))
2309 if (!debugfs_create_u32("min-order", mode, dir,
2310 &fail_page_alloc.min_order))
2315 debugfs_remove_recursive(dir);
2320 late_initcall(fail_page_alloc_debugfs);
2322 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2324 #else /* CONFIG_FAIL_PAGE_ALLOC */
2326 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2331 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2334 * Return true if free base pages are above 'mark'. For high-order checks it
2335 * will return true of the order-0 watermark is reached and there is at least
2336 * one free page of a suitable size. Checking now avoids taking the zone lock
2337 * to check in the allocation paths if no pages are free.
2339 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2340 unsigned long mark, int classzone_idx, int alloc_flags,
2345 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2347 /* free_pages may go negative - that's OK */
2348 free_pages -= (1 << order) - 1;
2350 if (alloc_flags & ALLOC_HIGH)
2354 * If the caller does not have rights to ALLOC_HARDER then subtract
2355 * the high-atomic reserves. This will over-estimate the size of the
2356 * atomic reserve but it avoids a search.
2358 if (likely(!alloc_harder))
2359 free_pages -= z->nr_reserved_highatomic;
2364 /* If allocation can't use CMA areas don't use free CMA pages */
2365 if (!(alloc_flags & ALLOC_CMA))
2366 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2370 * Check watermarks for an order-0 allocation request. If these
2371 * are not met, then a high-order request also cannot go ahead
2372 * even if a suitable page happened to be free.
2374 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2377 /* If this is an order-0 request then the watermark is fine */
2381 /* For a high-order request, check at least one suitable page is free */
2382 for (o = order; o < MAX_ORDER; o++) {
2383 struct free_area *area = &z->free_area[o];
2392 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2393 if (!list_empty(&area->free_list[mt]))
2398 if ((alloc_flags & ALLOC_CMA) &&
2399 !list_empty(&area->free_list[MIGRATE_CMA])) {
2407 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2408 int classzone_idx, int alloc_flags)
2410 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2411 zone_page_state(z, NR_FREE_PAGES));
2414 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2415 unsigned long mark, int classzone_idx)
2417 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2419 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2420 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2422 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2427 static bool zone_local(struct zone *local_zone, struct zone *zone)
2429 return local_zone->node == zone->node;
2432 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2434 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2437 #else /* CONFIG_NUMA */
2438 static bool zone_local(struct zone *local_zone, struct zone *zone)
2443 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2447 #endif /* CONFIG_NUMA */
2449 static void reset_alloc_batches(struct zone *preferred_zone)
2451 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2454 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2455 high_wmark_pages(zone) - low_wmark_pages(zone) -
2456 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2457 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2458 } while (zone++ != preferred_zone);
2462 * get_page_from_freelist goes through the zonelist trying to allocate
2465 static struct page *
2466 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2467 const struct alloc_context *ac)
2469 struct zonelist *zonelist = ac->zonelist;
2471 struct page *page = NULL;
2473 int nr_fair_skipped = 0;
2474 bool zonelist_rescan;
2477 zonelist_rescan = false;
2480 * Scan zonelist, looking for a zone with enough free.
2481 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2483 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2487 if (cpusets_enabled() &&
2488 (alloc_flags & ALLOC_CPUSET) &&
2489 !cpuset_zone_allowed(zone, gfp_mask))
2492 * Distribute pages in proportion to the individual
2493 * zone size to ensure fair page aging. The zone a
2494 * page was allocated in should have no effect on the
2495 * time the page has in memory before being reclaimed.
2497 if (alloc_flags & ALLOC_FAIR) {
2498 if (!zone_local(ac->preferred_zone, zone))
2500 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2506 * When allocating a page cache page for writing, we
2507 * want to get it from a zone that is within its dirty
2508 * limit, such that no single zone holds more than its
2509 * proportional share of globally allowed dirty pages.
2510 * The dirty limits take into account the zone's
2511 * lowmem reserves and high watermark so that kswapd
2512 * should be able to balance it without having to
2513 * write pages from its LRU list.
2515 * This may look like it could increase pressure on
2516 * lower zones by failing allocations in higher zones
2517 * before they are full. But the pages that do spill
2518 * over are limited as the lower zones are protected
2519 * by this very same mechanism. It should not become
2520 * a practical burden to them.
2522 * XXX: For now, allow allocations to potentially
2523 * exceed the per-zone dirty limit in the slowpath
2524 * (spread_dirty_pages unset) before going into reclaim,
2525 * which is important when on a NUMA setup the allowed
2526 * zones are together not big enough to reach the
2527 * global limit. The proper fix for these situations
2528 * will require awareness of zones in the
2529 * dirty-throttling and the flusher threads.
2531 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2534 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2535 if (!zone_watermark_ok(zone, order, mark,
2536 ac->classzone_idx, alloc_flags)) {
2539 /* Checked here to keep the fast path fast */
2540 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2541 if (alloc_flags & ALLOC_NO_WATERMARKS)
2544 if (zone_reclaim_mode == 0 ||
2545 !zone_allows_reclaim(ac->preferred_zone, zone))
2548 ret = zone_reclaim(zone, gfp_mask, order);
2550 case ZONE_RECLAIM_NOSCAN:
2553 case ZONE_RECLAIM_FULL:
2554 /* scanned but unreclaimable */
2557 /* did we reclaim enough */
2558 if (zone_watermark_ok(zone, order, mark,
2559 ac->classzone_idx, alloc_flags))
2567 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2568 gfp_mask, alloc_flags, ac->migratetype);
2570 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2574 * If this is a high-order atomic allocation then check
2575 * if the pageblock should be reserved for the future
2577 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2578 reserve_highatomic_pageblock(page, zone, order);
2585 * The first pass makes sure allocations are spread fairly within the
2586 * local node. However, the local node might have free pages left
2587 * after the fairness batches are exhausted, and remote zones haven't
2588 * even been considered yet. Try once more without fairness, and
2589 * include remote zones now, before entering the slowpath and waking
2590 * kswapd: prefer spilling to a remote zone over swapping locally.
2592 if (alloc_flags & ALLOC_FAIR) {
2593 alloc_flags &= ~ALLOC_FAIR;
2594 if (nr_fair_skipped) {
2595 zonelist_rescan = true;
2596 reset_alloc_batches(ac->preferred_zone);
2598 if (nr_online_nodes > 1)
2599 zonelist_rescan = true;
2602 if (zonelist_rescan)
2609 * Large machines with many possible nodes should not always dump per-node
2610 * meminfo in irq context.
2612 static inline bool should_suppress_show_mem(void)
2617 ret = in_interrupt();
2622 static DEFINE_RATELIMIT_STATE(nopage_rs,
2623 DEFAULT_RATELIMIT_INTERVAL,
2624 DEFAULT_RATELIMIT_BURST);
2626 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2628 unsigned int filter = SHOW_MEM_FILTER_NODES;
2630 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2631 debug_guardpage_minorder() > 0)
2635 * This documents exceptions given to allocations in certain
2636 * contexts that are allowed to allocate outside current's set
2639 if (!(gfp_mask & __GFP_NOMEMALLOC))
2640 if (test_thread_flag(TIF_MEMDIE) ||
2641 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2642 filter &= ~SHOW_MEM_FILTER_NODES;
2643 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2644 filter &= ~SHOW_MEM_FILTER_NODES;
2647 struct va_format vaf;
2650 va_start(args, fmt);
2655 pr_warn("%pV", &vaf);
2660 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2661 current->comm, order, gfp_mask);
2664 if (!should_suppress_show_mem())
2668 static inline struct page *
2669 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2670 const struct alloc_context *ac, unsigned long *did_some_progress)
2672 struct oom_control oc = {
2673 .zonelist = ac->zonelist,
2674 .nodemask = ac->nodemask,
2675 .gfp_mask = gfp_mask,
2680 *did_some_progress = 0;
2683 * Acquire the oom lock. If that fails, somebody else is
2684 * making progress for us.
2686 if (!mutex_trylock(&oom_lock)) {
2687 *did_some_progress = 1;
2688 schedule_timeout_uninterruptible(1);
2693 * Go through the zonelist yet one more time, keep very high watermark
2694 * here, this is only to catch a parallel oom killing, we must fail if
2695 * we're still under heavy pressure.
2697 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2698 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2702 if (!(gfp_mask & __GFP_NOFAIL)) {
2703 /* Coredumps can quickly deplete all memory reserves */
2704 if (current->flags & PF_DUMPCORE)
2706 /* The OOM killer will not help higher order allocs */
2707 if (order > PAGE_ALLOC_COSTLY_ORDER)
2709 /* The OOM killer does not needlessly kill tasks for lowmem */
2710 if (ac->high_zoneidx < ZONE_NORMAL)
2712 /* The OOM killer does not compensate for IO-less reclaim */
2713 if (!(gfp_mask & __GFP_FS)) {
2715 * XXX: Page reclaim didn't yield anything,
2716 * and the OOM killer can't be invoked, but
2717 * keep looping as per tradition.
2719 *did_some_progress = 1;
2722 if (pm_suspended_storage())
2724 /* The OOM killer may not free memory on a specific node */
2725 if (gfp_mask & __GFP_THISNODE)
2728 /* Exhausted what can be done so it's blamo time */
2729 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2730 *did_some_progress = 1;
2732 mutex_unlock(&oom_lock);
2736 #ifdef CONFIG_COMPACTION
2737 /* Try memory compaction for high-order allocations before reclaim */
2738 static struct page *
2739 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2740 int alloc_flags, const struct alloc_context *ac,
2741 enum migrate_mode mode, int *contended_compaction,
2742 bool *deferred_compaction)
2744 unsigned long compact_result;
2750 current->flags |= PF_MEMALLOC;
2751 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2752 mode, contended_compaction);
2753 current->flags &= ~PF_MEMALLOC;
2755 switch (compact_result) {
2756 case COMPACT_DEFERRED:
2757 *deferred_compaction = true;
2759 case COMPACT_SKIPPED:
2766 * At least in one zone compaction wasn't deferred or skipped, so let's
2767 * count a compaction stall
2769 count_vm_event(COMPACTSTALL);
2771 page = get_page_from_freelist(gfp_mask, order,
2772 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2775 struct zone *zone = page_zone(page);
2777 zone->compact_blockskip_flush = false;
2778 compaction_defer_reset(zone, order, true);
2779 count_vm_event(COMPACTSUCCESS);
2784 * It's bad if compaction run occurs and fails. The most likely reason
2785 * is that pages exist, but not enough to satisfy watermarks.
2787 count_vm_event(COMPACTFAIL);
2794 static inline struct page *
2795 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2796 int alloc_flags, const struct alloc_context *ac,
2797 enum migrate_mode mode, int *contended_compaction,
2798 bool *deferred_compaction)
2802 #endif /* CONFIG_COMPACTION */
2804 /* Perform direct synchronous page reclaim */
2806 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2807 const struct alloc_context *ac)
2809 struct reclaim_state reclaim_state;
2814 /* We now go into synchronous reclaim */
2815 cpuset_memory_pressure_bump();
2816 current->flags |= PF_MEMALLOC;
2817 lockdep_set_current_reclaim_state(gfp_mask);
2818 reclaim_state.reclaimed_slab = 0;
2819 current->reclaim_state = &reclaim_state;
2821 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2824 current->reclaim_state = NULL;
2825 lockdep_clear_current_reclaim_state();
2826 current->flags &= ~PF_MEMALLOC;
2833 /* The really slow allocator path where we enter direct reclaim */
2834 static inline struct page *
2835 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2836 int alloc_flags, const struct alloc_context *ac,
2837 unsigned long *did_some_progress)
2839 struct page *page = NULL;
2840 bool drained = false;
2842 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2843 if (unlikely(!(*did_some_progress)))
2847 page = get_page_from_freelist(gfp_mask, order,
2848 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2851 * If an allocation failed after direct reclaim, it could be because
2852 * pages are pinned on the per-cpu lists or in high alloc reserves.
2853 * Shrink them them and try again
2855 if (!page && !drained) {
2856 unreserve_highatomic_pageblock(ac);
2857 drain_all_pages(NULL);
2866 * This is called in the allocator slow-path if the allocation request is of
2867 * sufficient urgency to ignore watermarks and take other desperate measures
2869 static inline struct page *
2870 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2871 const struct alloc_context *ac)
2876 page = get_page_from_freelist(gfp_mask, order,
2877 ALLOC_NO_WATERMARKS, ac);
2879 if (!page && gfp_mask & __GFP_NOFAIL)
2880 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2882 } while (!page && (gfp_mask & __GFP_NOFAIL));
2887 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2892 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2893 ac->high_zoneidx, ac->nodemask)
2894 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2898 gfp_to_alloc_flags(gfp_t gfp_mask)
2900 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2902 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2903 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2906 * The caller may dip into page reserves a bit more if the caller
2907 * cannot run direct reclaim, or if the caller has realtime scheduling
2908 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2909 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2911 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2913 if (gfp_mask & __GFP_ATOMIC) {
2915 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2916 * if it can't schedule.
2918 if (!(gfp_mask & __GFP_NOMEMALLOC))
2919 alloc_flags |= ALLOC_HARDER;
2921 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2922 * comment for __cpuset_node_allowed().
2924 alloc_flags &= ~ALLOC_CPUSET;
2925 } else if (unlikely(rt_task(current)) && !in_interrupt())
2926 alloc_flags |= ALLOC_HARDER;
2928 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2929 if (gfp_mask & __GFP_MEMALLOC)
2930 alloc_flags |= ALLOC_NO_WATERMARKS;
2931 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2932 alloc_flags |= ALLOC_NO_WATERMARKS;
2933 else if (!in_interrupt() &&
2934 ((current->flags & PF_MEMALLOC) ||
2935 unlikely(test_thread_flag(TIF_MEMDIE))))
2936 alloc_flags |= ALLOC_NO_WATERMARKS;
2939 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2940 alloc_flags |= ALLOC_CMA;
2945 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2947 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2950 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2952 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2955 static inline struct page *
2956 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2957 struct alloc_context *ac)
2959 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2960 struct page *page = NULL;
2962 unsigned long pages_reclaimed = 0;
2963 unsigned long did_some_progress;
2964 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2965 bool deferred_compaction = false;
2966 int contended_compaction = COMPACT_CONTENDED_NONE;
2969 * In the slowpath, we sanity check order to avoid ever trying to
2970 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2971 * be using allocators in order of preference for an area that is
2974 if (order >= MAX_ORDER) {
2975 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2980 * We also sanity check to catch abuse of atomic reserves being used by
2981 * callers that are not in atomic context.
2983 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
2984 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
2985 gfp_mask &= ~__GFP_ATOMIC;
2988 * If this allocation cannot block and it is for a specific node, then
2989 * fail early. There's no need to wakeup kswapd or retry for a
2990 * speculative node-specific allocation.
2992 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
2996 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
2997 wake_all_kswapds(order, ac);
3000 * OK, we're below the kswapd watermark and have kicked background
3001 * reclaim. Now things get more complex, so set up alloc_flags according
3002 * to how we want to proceed.
3004 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3007 * Find the true preferred zone if the allocation is unconstrained by
3010 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3011 struct zoneref *preferred_zoneref;
3012 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3013 ac->high_zoneidx, NULL, &ac->preferred_zone);
3014 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3017 /* This is the last chance, in general, before the goto nopage. */
3018 page = get_page_from_freelist(gfp_mask, order,
3019 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3023 /* Allocate without watermarks if the context allows */
3024 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3026 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3027 * the allocation is high priority and these type of
3028 * allocations are system rather than user orientated
3030 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3032 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3039 /* Caller is not willing to reclaim, we can't balance anything */
3040 if (!can_direct_reclaim) {
3042 * All existing users of the deprecated __GFP_NOFAIL are
3043 * blockable, so warn of any new users that actually allow this
3044 * type of allocation to fail.
3046 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3050 /* Avoid recursion of direct reclaim */
3051 if (current->flags & PF_MEMALLOC)
3054 /* Avoid allocations with no watermarks from looping endlessly */
3055 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3059 * Try direct compaction. The first pass is asynchronous. Subsequent
3060 * attempts after direct reclaim are synchronous
3062 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3064 &contended_compaction,
3065 &deferred_compaction);
3069 /* Checks for THP-specific high-order allocations */
3070 if (is_thp_gfp_mask(gfp_mask)) {
3072 * If compaction is deferred for high-order allocations, it is
3073 * because sync compaction recently failed. If this is the case
3074 * and the caller requested a THP allocation, we do not want
3075 * to heavily disrupt the system, so we fail the allocation
3076 * instead of entering direct reclaim.
3078 if (deferred_compaction)
3082 * In all zones where compaction was attempted (and not
3083 * deferred or skipped), lock contention has been detected.
3084 * For THP allocation we do not want to disrupt the others
3085 * so we fallback to base pages instead.
3087 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3091 * If compaction was aborted due to need_resched(), we do not
3092 * want to further increase allocation latency, unless it is
3093 * khugepaged trying to collapse.
3095 if (contended_compaction == COMPACT_CONTENDED_SCHED
3096 && !(current->flags & PF_KTHREAD))
3101 * It can become very expensive to allocate transparent hugepages at
3102 * fault, so use asynchronous memory compaction for THP unless it is
3103 * khugepaged trying to collapse.
3105 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3106 migration_mode = MIGRATE_SYNC_LIGHT;
3108 /* Try direct reclaim and then allocating */
3109 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3110 &did_some_progress);
3114 /* Do not loop if specifically requested */
3115 if (gfp_mask & __GFP_NORETRY)
3118 /* Keep reclaiming pages as long as there is reasonable progress */
3119 pages_reclaimed += did_some_progress;
3120 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3121 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3122 /* Wait for some write requests to complete then retry */
3123 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3127 /* Reclaim has failed us, start killing things */
3128 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3132 /* Retry as long as the OOM killer is making progress */
3133 if (did_some_progress)
3138 * High-order allocations do not necessarily loop after
3139 * direct reclaim and reclaim/compaction depends on compaction
3140 * being called after reclaim so call directly if necessary
3142 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3144 &contended_compaction,
3145 &deferred_compaction);
3149 warn_alloc_failed(gfp_mask, order, NULL);
3155 * This is the 'heart' of the zoned buddy allocator.
3158 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3159 struct zonelist *zonelist, nodemask_t *nodemask)
3161 struct zoneref *preferred_zoneref;
3162 struct page *page = NULL;
3163 unsigned int cpuset_mems_cookie;
3164 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3165 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3166 struct alloc_context ac = {
3167 .high_zoneidx = gfp_zone(gfp_mask),
3168 .nodemask = nodemask,
3169 .migratetype = gfpflags_to_migratetype(gfp_mask),
3172 gfp_mask &= gfp_allowed_mask;
3174 lockdep_trace_alloc(gfp_mask);
3176 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3178 if (should_fail_alloc_page(gfp_mask, order))
3182 * Check the zones suitable for the gfp_mask contain at least one
3183 * valid zone. It's possible to have an empty zonelist as a result
3184 * of __GFP_THISNODE and a memoryless node
3186 if (unlikely(!zonelist->_zonerefs->zone))
3189 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3190 alloc_flags |= ALLOC_CMA;
3193 cpuset_mems_cookie = read_mems_allowed_begin();
3195 /* We set it here, as __alloc_pages_slowpath might have changed it */
3196 ac.zonelist = zonelist;
3198 /* Dirty zone balancing only done in the fast path */
3199 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3201 /* The preferred zone is used for statistics later */
3202 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3203 ac.nodemask ? : &cpuset_current_mems_allowed,
3204 &ac.preferred_zone);
3205 if (!ac.preferred_zone)
3207 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3209 /* First allocation attempt */
3210 alloc_mask = gfp_mask|__GFP_HARDWALL;
3211 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3212 if (unlikely(!page)) {
3214 * Runtime PM, block IO and its error handling path
3215 * can deadlock because I/O on the device might not
3218 alloc_mask = memalloc_noio_flags(gfp_mask);
3219 ac.spread_dirty_pages = false;
3221 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3224 if (kmemcheck_enabled && page)
3225 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3227 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3231 * When updating a task's mems_allowed, it is possible to race with
3232 * parallel threads in such a way that an allocation can fail while
3233 * the mask is being updated. If a page allocation is about to fail,
3234 * check if the cpuset changed during allocation and if so, retry.
3236 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3241 EXPORT_SYMBOL(__alloc_pages_nodemask);
3244 * Common helper functions.
3246 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3251 * __get_free_pages() returns a 32-bit address, which cannot represent
3254 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3256 page = alloc_pages(gfp_mask, order);
3259 return (unsigned long) page_address(page);
3261 EXPORT_SYMBOL(__get_free_pages);
3263 unsigned long get_zeroed_page(gfp_t gfp_mask)
3265 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3267 EXPORT_SYMBOL(get_zeroed_page);
3269 void __free_pages(struct page *page, unsigned int order)
3271 if (put_page_testzero(page)) {
3273 free_hot_cold_page(page, false);
3275 __free_pages_ok(page, order);
3279 EXPORT_SYMBOL(__free_pages);
3281 void free_pages(unsigned long addr, unsigned int order)
3284 VM_BUG_ON(!virt_addr_valid((void *)addr));
3285 __free_pages(virt_to_page((void *)addr), order);
3289 EXPORT_SYMBOL(free_pages);
3293 * An arbitrary-length arbitrary-offset area of memory which resides
3294 * within a 0 or higher order page. Multiple fragments within that page
3295 * are individually refcounted, in the page's reference counter.
3297 * The page_frag functions below provide a simple allocation framework for
3298 * page fragments. This is used by the network stack and network device
3299 * drivers to provide a backing region of memory for use as either an
3300 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3302 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3305 struct page *page = NULL;
3306 gfp_t gfp = gfp_mask;
3308 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3309 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3311 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3312 PAGE_FRAG_CACHE_MAX_ORDER);
3313 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3315 if (unlikely(!page))
3316 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3318 nc->va = page ? page_address(page) : NULL;
3323 void *__alloc_page_frag(struct page_frag_cache *nc,
3324 unsigned int fragsz, gfp_t gfp_mask)
3326 unsigned int size = PAGE_SIZE;
3330 if (unlikely(!nc->va)) {
3332 page = __page_frag_refill(nc, gfp_mask);
3336 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3337 /* if size can vary use size else just use PAGE_SIZE */
3340 /* Even if we own the page, we do not use atomic_set().
3341 * This would break get_page_unless_zero() users.
3343 atomic_add(size - 1, &page->_count);
3345 /* reset page count bias and offset to start of new frag */
3346 nc->pfmemalloc = page_is_pfmemalloc(page);
3347 nc->pagecnt_bias = size;
3351 offset = nc->offset - fragsz;
3352 if (unlikely(offset < 0)) {
3353 page = virt_to_page(nc->va);
3355 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3358 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3359 /* if size can vary use size else just use PAGE_SIZE */
3362 /* OK, page count is 0, we can safely set it */
3363 atomic_set(&page->_count, size);
3365 /* reset page count bias and offset to start of new frag */
3366 nc->pagecnt_bias = size;
3367 offset = size - fragsz;
3371 nc->offset = offset;
3373 return nc->va + offset;
3375 EXPORT_SYMBOL(__alloc_page_frag);
3378 * Frees a page fragment allocated out of either a compound or order 0 page.
3380 void __free_page_frag(void *addr)
3382 struct page *page = virt_to_head_page(addr);
3384 if (unlikely(put_page_testzero(page)))
3385 __free_pages_ok(page, compound_order(page));
3387 EXPORT_SYMBOL(__free_page_frag);
3390 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3391 * of the current memory cgroup.
3393 * It should be used when the caller would like to use kmalloc, but since the
3394 * allocation is large, it has to fall back to the page allocator.
3396 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3400 page = alloc_pages(gfp_mask, order);
3401 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3402 __free_pages(page, order);
3408 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3412 page = alloc_pages_node(nid, gfp_mask, order);
3413 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3414 __free_pages(page, order);
3421 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3424 void __free_kmem_pages(struct page *page, unsigned int order)
3426 memcg_kmem_uncharge(page, order);
3427 __free_pages(page, order);
3430 void free_kmem_pages(unsigned long addr, unsigned int order)
3433 VM_BUG_ON(!virt_addr_valid((void *)addr));
3434 __free_kmem_pages(virt_to_page((void *)addr), order);
3438 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3441 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3442 unsigned long used = addr + PAGE_ALIGN(size);
3444 split_page(virt_to_page((void *)addr), order);
3445 while (used < alloc_end) {
3450 return (void *)addr;
3454 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3455 * @size: the number of bytes to allocate
3456 * @gfp_mask: GFP flags for the allocation
3458 * This function is similar to alloc_pages(), except that it allocates the
3459 * minimum number of pages to satisfy the request. alloc_pages() can only
3460 * allocate memory in power-of-two pages.
3462 * This function is also limited by MAX_ORDER.
3464 * Memory allocated by this function must be released by free_pages_exact().
3466 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3468 unsigned int order = get_order(size);
3471 addr = __get_free_pages(gfp_mask, order);
3472 return make_alloc_exact(addr, order, size);
3474 EXPORT_SYMBOL(alloc_pages_exact);
3477 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3479 * @nid: the preferred node ID where memory should be allocated
3480 * @size: the number of bytes to allocate
3481 * @gfp_mask: GFP flags for the allocation
3483 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3486 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3488 unsigned order = get_order(size);
3489 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3492 return make_alloc_exact((unsigned long)page_address(p), order, size);
3496 * free_pages_exact - release memory allocated via alloc_pages_exact()
3497 * @virt: the value returned by alloc_pages_exact.
3498 * @size: size of allocation, same value as passed to alloc_pages_exact().
3500 * Release the memory allocated by a previous call to alloc_pages_exact.
3502 void free_pages_exact(void *virt, size_t size)
3504 unsigned long addr = (unsigned long)virt;
3505 unsigned long end = addr + PAGE_ALIGN(size);
3507 while (addr < end) {
3512 EXPORT_SYMBOL(free_pages_exact);
3515 * nr_free_zone_pages - count number of pages beyond high watermark
3516 * @offset: The zone index of the highest zone
3518 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3519 * high watermark within all zones at or below a given zone index. For each
3520 * zone, the number of pages is calculated as:
3521 * managed_pages - high_pages
3523 static unsigned long nr_free_zone_pages(int offset)
3528 /* Just pick one node, since fallback list is circular */
3529 unsigned long sum = 0;
3531 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3533 for_each_zone_zonelist(zone, z, zonelist, offset) {
3534 unsigned long size = zone->managed_pages;
3535 unsigned long high = high_wmark_pages(zone);
3544 * nr_free_buffer_pages - count number of pages beyond high watermark
3546 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3547 * watermark within ZONE_DMA and ZONE_NORMAL.
3549 unsigned long nr_free_buffer_pages(void)
3551 return nr_free_zone_pages(gfp_zone(GFP_USER));
3553 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3556 * nr_free_pagecache_pages - count number of pages beyond high watermark
3558 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3559 * high watermark within all zones.
3561 unsigned long nr_free_pagecache_pages(void)
3563 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3566 static inline void show_node(struct zone *zone)
3568 if (IS_ENABLED(CONFIG_NUMA))
3569 printk("Node %d ", zone_to_nid(zone));
3572 void si_meminfo(struct sysinfo *val)
3574 val->totalram = totalram_pages;
3575 val->sharedram = global_page_state(NR_SHMEM);
3576 val->freeram = global_page_state(NR_FREE_PAGES);
3577 val->bufferram = nr_blockdev_pages();
3578 val->totalhigh = totalhigh_pages;
3579 val->freehigh = nr_free_highpages();
3580 val->mem_unit = PAGE_SIZE;
3583 EXPORT_SYMBOL(si_meminfo);
3586 void si_meminfo_node(struct sysinfo *val, int nid)
3588 int zone_type; /* needs to be signed */
3589 unsigned long managed_pages = 0;
3590 pg_data_t *pgdat = NODE_DATA(nid);
3592 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3593 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3594 val->totalram = managed_pages;
3595 val->sharedram = node_page_state(nid, NR_SHMEM);
3596 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3597 #ifdef CONFIG_HIGHMEM
3598 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3599 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3605 val->mem_unit = PAGE_SIZE;
3610 * Determine whether the node should be displayed or not, depending on whether
3611 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3613 bool skip_free_areas_node(unsigned int flags, int nid)
3616 unsigned int cpuset_mems_cookie;
3618 if (!(flags & SHOW_MEM_FILTER_NODES))
3622 cpuset_mems_cookie = read_mems_allowed_begin();
3623 ret = !node_isset(nid, cpuset_current_mems_allowed);
3624 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3629 #define K(x) ((x) << (PAGE_SHIFT-10))
3631 static void show_migration_types(unsigned char type)
3633 static const char types[MIGRATE_TYPES] = {
3634 [MIGRATE_UNMOVABLE] = 'U',
3635 [MIGRATE_RECLAIMABLE] = 'E',
3636 [MIGRATE_MOVABLE] = 'M',
3638 [MIGRATE_CMA] = 'C',
3640 #ifdef CONFIG_MEMORY_ISOLATION
3641 [MIGRATE_ISOLATE] = 'I',
3644 char tmp[MIGRATE_TYPES + 1];
3648 for (i = 0; i < MIGRATE_TYPES; i++) {
3649 if (type & (1 << i))
3654 printk("(%s) ", tmp);
3658 * Show free area list (used inside shift_scroll-lock stuff)
3659 * We also calculate the percentage fragmentation. We do this by counting the
3660 * memory on each free list with the exception of the first item on the list.
3663 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3666 void show_free_areas(unsigned int filter)
3668 unsigned long free_pcp = 0;
3672 for_each_populated_zone(zone) {
3673 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3676 for_each_online_cpu(cpu)
3677 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3680 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3681 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3682 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3683 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3684 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3685 " free:%lu free_pcp:%lu free_cma:%lu\n",
3686 global_page_state(NR_ACTIVE_ANON),
3687 global_page_state(NR_INACTIVE_ANON),
3688 global_page_state(NR_ISOLATED_ANON),
3689 global_page_state(NR_ACTIVE_FILE),
3690 global_page_state(NR_INACTIVE_FILE),
3691 global_page_state(NR_ISOLATED_FILE),
3692 global_page_state(NR_UNEVICTABLE),
3693 global_page_state(NR_FILE_DIRTY),
3694 global_page_state(NR_WRITEBACK),
3695 global_page_state(NR_UNSTABLE_NFS),
3696 global_page_state(NR_SLAB_RECLAIMABLE),
3697 global_page_state(NR_SLAB_UNRECLAIMABLE),
3698 global_page_state(NR_FILE_MAPPED),
3699 global_page_state(NR_SHMEM),
3700 global_page_state(NR_PAGETABLE),
3701 global_page_state(NR_BOUNCE),
3702 global_page_state(NR_FREE_PAGES),
3704 global_page_state(NR_FREE_CMA_PAGES));
3706 for_each_populated_zone(zone) {
3709 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3713 for_each_online_cpu(cpu)
3714 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3722 " active_anon:%lukB"
3723 " inactive_anon:%lukB"
3724 " active_file:%lukB"
3725 " inactive_file:%lukB"
3726 " unevictable:%lukB"
3727 " isolated(anon):%lukB"
3728 " isolated(file):%lukB"
3736 " slab_reclaimable:%lukB"
3737 " slab_unreclaimable:%lukB"
3738 " kernel_stack:%lukB"
3745 " writeback_tmp:%lukB"
3746 " pages_scanned:%lu"
3747 " all_unreclaimable? %s"
3750 K(zone_page_state(zone, NR_FREE_PAGES)),
3751 K(min_wmark_pages(zone)),
3752 K(low_wmark_pages(zone)),
3753 K(high_wmark_pages(zone)),
3754 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3755 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3756 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3757 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3758 K(zone_page_state(zone, NR_UNEVICTABLE)),
3759 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3760 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3761 K(zone->present_pages),
3762 K(zone->managed_pages),
3763 K(zone_page_state(zone, NR_MLOCK)),
3764 K(zone_page_state(zone, NR_FILE_DIRTY)),
3765 K(zone_page_state(zone, NR_WRITEBACK)),
3766 K(zone_page_state(zone, NR_FILE_MAPPED)),
3767 K(zone_page_state(zone, NR_SHMEM)),
3768 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3769 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3770 zone_page_state(zone, NR_KERNEL_STACK) *
3772 K(zone_page_state(zone, NR_PAGETABLE)),
3773 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3774 K(zone_page_state(zone, NR_BOUNCE)),
3776 K(this_cpu_read(zone->pageset->pcp.count)),
3777 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3778 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3779 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3780 (!zone_reclaimable(zone) ? "yes" : "no")
3782 printk("lowmem_reserve[]:");
3783 for (i = 0; i < MAX_NR_ZONES; i++)
3784 printk(" %ld", zone->lowmem_reserve[i]);
3788 for_each_populated_zone(zone) {
3789 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3790 unsigned char types[MAX_ORDER];
3792 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3795 printk("%s: ", zone->name);
3797 spin_lock_irqsave(&zone->lock, flags);
3798 for (order = 0; order < MAX_ORDER; order++) {
3799 struct free_area *area = &zone->free_area[order];
3802 nr[order] = area->nr_free;
3803 total += nr[order] << order;
3806 for (type = 0; type < MIGRATE_TYPES; type++) {
3807 if (!list_empty(&area->free_list[type]))
3808 types[order] |= 1 << type;
3811 spin_unlock_irqrestore(&zone->lock, flags);
3812 for (order = 0; order < MAX_ORDER; order++) {
3813 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3815 show_migration_types(types[order]);
3817 printk("= %lukB\n", K(total));
3820 hugetlb_show_meminfo();
3822 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3824 show_swap_cache_info();
3827 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3829 zoneref->zone = zone;
3830 zoneref->zone_idx = zone_idx(zone);
3834 * Builds allocation fallback zone lists.
3836 * Add all populated zones of a node to the zonelist.
3838 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3842 enum zone_type zone_type = MAX_NR_ZONES;
3846 zone = pgdat->node_zones + zone_type;
3847 if (populated_zone(zone)) {
3848 zoneref_set_zone(zone,
3849 &zonelist->_zonerefs[nr_zones++]);
3850 check_highest_zone(zone_type);
3852 } while (zone_type);
3860 * 0 = automatic detection of better ordering.
3861 * 1 = order by ([node] distance, -zonetype)
3862 * 2 = order by (-zonetype, [node] distance)
3864 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3865 * the same zonelist. So only NUMA can configure this param.
3867 #define ZONELIST_ORDER_DEFAULT 0
3868 #define ZONELIST_ORDER_NODE 1
3869 #define ZONELIST_ORDER_ZONE 2
3871 /* zonelist order in the kernel.
3872 * set_zonelist_order() will set this to NODE or ZONE.
3874 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3875 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3879 /* The value user specified ....changed by config */
3880 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3881 /* string for sysctl */
3882 #define NUMA_ZONELIST_ORDER_LEN 16
3883 char numa_zonelist_order[16] = "default";
3886 * interface for configure zonelist ordering.
3887 * command line option "numa_zonelist_order"
3888 * = "[dD]efault - default, automatic configuration.
3889 * = "[nN]ode - order by node locality, then by zone within node
3890 * = "[zZ]one - order by zone, then by locality within zone
3893 static int __parse_numa_zonelist_order(char *s)
3895 if (*s == 'd' || *s == 'D') {
3896 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3897 } else if (*s == 'n' || *s == 'N') {
3898 user_zonelist_order = ZONELIST_ORDER_NODE;
3899 } else if (*s == 'z' || *s == 'Z') {
3900 user_zonelist_order = ZONELIST_ORDER_ZONE;
3903 "Ignoring invalid numa_zonelist_order value: "
3910 static __init int setup_numa_zonelist_order(char *s)
3917 ret = __parse_numa_zonelist_order(s);
3919 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3923 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3926 * sysctl handler for numa_zonelist_order
3928 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3929 void __user *buffer, size_t *length,
3932 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3934 static DEFINE_MUTEX(zl_order_mutex);
3936 mutex_lock(&zl_order_mutex);
3938 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3942 strcpy(saved_string, (char *)table->data);
3944 ret = proc_dostring(table, write, buffer, length, ppos);
3948 int oldval = user_zonelist_order;
3950 ret = __parse_numa_zonelist_order((char *)table->data);
3953 * bogus value. restore saved string
3955 strncpy((char *)table->data, saved_string,
3956 NUMA_ZONELIST_ORDER_LEN);
3957 user_zonelist_order = oldval;
3958 } else if (oldval != user_zonelist_order) {
3959 mutex_lock(&zonelists_mutex);
3960 build_all_zonelists(NULL, NULL);
3961 mutex_unlock(&zonelists_mutex);
3965 mutex_unlock(&zl_order_mutex);
3970 #define MAX_NODE_LOAD (nr_online_nodes)
3971 static int node_load[MAX_NUMNODES];
3974 * find_next_best_node - find the next node that should appear in a given node's fallback list
3975 * @node: node whose fallback list we're appending
3976 * @used_node_mask: nodemask_t of already used nodes
3978 * We use a number of factors to determine which is the next node that should
3979 * appear on a given node's fallback list. The node should not have appeared
3980 * already in @node's fallback list, and it should be the next closest node
3981 * according to the distance array (which contains arbitrary distance values
3982 * from each node to each node in the system), and should also prefer nodes
3983 * with no CPUs, since presumably they'll have very little allocation pressure
3984 * on them otherwise.
3985 * It returns -1 if no node is found.
3987 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3990 int min_val = INT_MAX;
3991 int best_node = NUMA_NO_NODE;
3992 const struct cpumask *tmp = cpumask_of_node(0);
3994 /* Use the local node if we haven't already */
3995 if (!node_isset(node, *used_node_mask)) {
3996 node_set(node, *used_node_mask);
4000 for_each_node_state(n, N_MEMORY) {
4002 /* Don't want a node to appear more than once */
4003 if (node_isset(n, *used_node_mask))
4006 /* Use the distance array to find the distance */
4007 val = node_distance(node, n);
4009 /* Penalize nodes under us ("prefer the next node") */
4012 /* Give preference to headless and unused nodes */
4013 tmp = cpumask_of_node(n);
4014 if (!cpumask_empty(tmp))
4015 val += PENALTY_FOR_NODE_WITH_CPUS;
4017 /* Slight preference for less loaded node */
4018 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4019 val += node_load[n];
4021 if (val < min_val) {
4028 node_set(best_node, *used_node_mask);
4035 * Build zonelists ordered by node and zones within node.
4036 * This results in maximum locality--normal zone overflows into local
4037 * DMA zone, if any--but risks exhausting DMA zone.
4039 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4042 struct zonelist *zonelist;
4044 zonelist = &pgdat->node_zonelists[0];
4045 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4047 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4048 zonelist->_zonerefs[j].zone = NULL;
4049 zonelist->_zonerefs[j].zone_idx = 0;
4053 * Build gfp_thisnode zonelists
4055 static void build_thisnode_zonelists(pg_data_t *pgdat)
4058 struct zonelist *zonelist;
4060 zonelist = &pgdat->node_zonelists[1];
4061 j = build_zonelists_node(pgdat, zonelist, 0);
4062 zonelist->_zonerefs[j].zone = NULL;
4063 zonelist->_zonerefs[j].zone_idx = 0;
4067 * Build zonelists ordered by zone and nodes within zones.
4068 * This results in conserving DMA zone[s] until all Normal memory is
4069 * exhausted, but results in overflowing to remote node while memory
4070 * may still exist in local DMA zone.
4072 static int node_order[MAX_NUMNODES];
4074 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4077 int zone_type; /* needs to be signed */
4079 struct zonelist *zonelist;
4081 zonelist = &pgdat->node_zonelists[0];
4083 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4084 for (j = 0; j < nr_nodes; j++) {
4085 node = node_order[j];
4086 z = &NODE_DATA(node)->node_zones[zone_type];
4087 if (populated_zone(z)) {
4089 &zonelist->_zonerefs[pos++]);
4090 check_highest_zone(zone_type);
4094 zonelist->_zonerefs[pos].zone = NULL;
4095 zonelist->_zonerefs[pos].zone_idx = 0;
4098 #if defined(CONFIG_64BIT)
4100 * Devices that require DMA32/DMA are relatively rare and do not justify a
4101 * penalty to every machine in case the specialised case applies. Default
4102 * to Node-ordering on 64-bit NUMA machines
4104 static int default_zonelist_order(void)
4106 return ZONELIST_ORDER_NODE;
4110 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4111 * by the kernel. If processes running on node 0 deplete the low memory zone
4112 * then reclaim will occur more frequency increasing stalls and potentially
4113 * be easier to OOM if a large percentage of the zone is under writeback or
4114 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4115 * Hence, default to zone ordering on 32-bit.
4117 static int default_zonelist_order(void)
4119 return ZONELIST_ORDER_ZONE;
4121 #endif /* CONFIG_64BIT */
4123 static void set_zonelist_order(void)
4125 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4126 current_zonelist_order = default_zonelist_order();
4128 current_zonelist_order = user_zonelist_order;
4131 static void build_zonelists(pg_data_t *pgdat)
4135 nodemask_t used_mask;
4136 int local_node, prev_node;
4137 struct zonelist *zonelist;
4138 int order = current_zonelist_order;
4140 /* initialize zonelists */
4141 for (i = 0; i < MAX_ZONELISTS; i++) {
4142 zonelist = pgdat->node_zonelists + i;
4143 zonelist->_zonerefs[0].zone = NULL;
4144 zonelist->_zonerefs[0].zone_idx = 0;
4147 /* NUMA-aware ordering of nodes */
4148 local_node = pgdat->node_id;
4149 load = nr_online_nodes;
4150 prev_node = local_node;
4151 nodes_clear(used_mask);
4153 memset(node_order, 0, sizeof(node_order));
4156 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4158 * We don't want to pressure a particular node.
4159 * So adding penalty to the first node in same
4160 * distance group to make it round-robin.
4162 if (node_distance(local_node, node) !=
4163 node_distance(local_node, prev_node))
4164 node_load[node] = load;
4168 if (order == ZONELIST_ORDER_NODE)
4169 build_zonelists_in_node_order(pgdat, node);
4171 node_order[j++] = node; /* remember order */
4174 if (order == ZONELIST_ORDER_ZONE) {
4175 /* calculate node order -- i.e., DMA last! */
4176 build_zonelists_in_zone_order(pgdat, j);
4179 build_thisnode_zonelists(pgdat);
4182 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4184 * Return node id of node used for "local" allocations.
4185 * I.e., first node id of first zone in arg node's generic zonelist.
4186 * Used for initializing percpu 'numa_mem', which is used primarily
4187 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4189 int local_memory_node(int node)
4193 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4194 gfp_zone(GFP_KERNEL),
4201 #else /* CONFIG_NUMA */
4203 static void set_zonelist_order(void)
4205 current_zonelist_order = ZONELIST_ORDER_ZONE;
4208 static void build_zonelists(pg_data_t *pgdat)
4210 int node, local_node;
4212 struct zonelist *zonelist;
4214 local_node = pgdat->node_id;
4216 zonelist = &pgdat->node_zonelists[0];
4217 j = build_zonelists_node(pgdat, zonelist, 0);
4220 * Now we build the zonelist so that it contains the zones
4221 * of all the other nodes.
4222 * We don't want to pressure a particular node, so when
4223 * building the zones for node N, we make sure that the
4224 * zones coming right after the local ones are those from
4225 * node N+1 (modulo N)
4227 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4228 if (!node_online(node))
4230 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4232 for (node = 0; node < local_node; node++) {
4233 if (!node_online(node))
4235 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4238 zonelist->_zonerefs[j].zone = NULL;
4239 zonelist->_zonerefs[j].zone_idx = 0;
4242 #endif /* CONFIG_NUMA */
4245 * Boot pageset table. One per cpu which is going to be used for all
4246 * zones and all nodes. The parameters will be set in such a way
4247 * that an item put on a list will immediately be handed over to
4248 * the buddy list. This is safe since pageset manipulation is done
4249 * with interrupts disabled.
4251 * The boot_pagesets must be kept even after bootup is complete for
4252 * unused processors and/or zones. They do play a role for bootstrapping
4253 * hotplugged processors.
4255 * zoneinfo_show() and maybe other functions do
4256 * not check if the processor is online before following the pageset pointer.
4257 * Other parts of the kernel may not check if the zone is available.
4259 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4260 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4261 static void setup_zone_pageset(struct zone *zone);
4264 * Global mutex to protect against size modification of zonelists
4265 * as well as to serialize pageset setup for the new populated zone.
4267 DEFINE_MUTEX(zonelists_mutex);
4269 /* return values int ....just for stop_machine() */
4270 static int __build_all_zonelists(void *data)
4274 pg_data_t *self = data;
4277 memset(node_load, 0, sizeof(node_load));
4280 if (self && !node_online(self->node_id)) {
4281 build_zonelists(self);
4284 for_each_online_node(nid) {
4285 pg_data_t *pgdat = NODE_DATA(nid);
4287 build_zonelists(pgdat);
4291 * Initialize the boot_pagesets that are going to be used
4292 * for bootstrapping processors. The real pagesets for
4293 * each zone will be allocated later when the per cpu
4294 * allocator is available.
4296 * boot_pagesets are used also for bootstrapping offline
4297 * cpus if the system is already booted because the pagesets
4298 * are needed to initialize allocators on a specific cpu too.
4299 * F.e. the percpu allocator needs the page allocator which
4300 * needs the percpu allocator in order to allocate its pagesets
4301 * (a chicken-egg dilemma).
4303 for_each_possible_cpu(cpu) {
4304 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4306 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4308 * We now know the "local memory node" for each node--
4309 * i.e., the node of the first zone in the generic zonelist.
4310 * Set up numa_mem percpu variable for on-line cpus. During
4311 * boot, only the boot cpu should be on-line; we'll init the
4312 * secondary cpus' numa_mem as they come on-line. During
4313 * node/memory hotplug, we'll fixup all on-line cpus.
4315 if (cpu_online(cpu))
4316 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4323 static noinline void __init
4324 build_all_zonelists_init(void)
4326 __build_all_zonelists(NULL);
4327 mminit_verify_zonelist();
4328 cpuset_init_current_mems_allowed();
4332 * Called with zonelists_mutex held always
4333 * unless system_state == SYSTEM_BOOTING.
4335 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4336 * [we're only called with non-NULL zone through __meminit paths] and
4337 * (2) call of __init annotated helper build_all_zonelists_init
4338 * [protected by SYSTEM_BOOTING].
4340 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4342 set_zonelist_order();
4344 if (system_state == SYSTEM_BOOTING) {
4345 build_all_zonelists_init();
4347 #ifdef CONFIG_MEMORY_HOTPLUG
4349 setup_zone_pageset(zone);
4351 /* we have to stop all cpus to guarantee there is no user
4353 stop_machine(__build_all_zonelists, pgdat, NULL);
4354 /* cpuset refresh routine should be here */
4356 vm_total_pages = nr_free_pagecache_pages();
4358 * Disable grouping by mobility if the number of pages in the
4359 * system is too low to allow the mechanism to work. It would be
4360 * more accurate, but expensive to check per-zone. This check is
4361 * made on memory-hotadd so a system can start with mobility
4362 * disabled and enable it later
4364 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4365 page_group_by_mobility_disabled = 1;
4367 page_group_by_mobility_disabled = 0;
4369 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4370 "Total pages: %ld\n",
4372 zonelist_order_name[current_zonelist_order],
4373 page_group_by_mobility_disabled ? "off" : "on",
4376 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4381 * Helper functions to size the waitqueue hash table.
4382 * Essentially these want to choose hash table sizes sufficiently
4383 * large so that collisions trying to wait on pages are rare.
4384 * But in fact, the number of active page waitqueues on typical
4385 * systems is ridiculously low, less than 200. So this is even
4386 * conservative, even though it seems large.
4388 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4389 * waitqueues, i.e. the size of the waitq table given the number of pages.
4391 #define PAGES_PER_WAITQUEUE 256
4393 #ifndef CONFIG_MEMORY_HOTPLUG
4394 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4396 unsigned long size = 1;
4398 pages /= PAGES_PER_WAITQUEUE;
4400 while (size < pages)
4404 * Once we have dozens or even hundreds of threads sleeping
4405 * on IO we've got bigger problems than wait queue collision.
4406 * Limit the size of the wait table to a reasonable size.
4408 size = min(size, 4096UL);
4410 return max(size, 4UL);
4414 * A zone's size might be changed by hot-add, so it is not possible to determine
4415 * a suitable size for its wait_table. So we use the maximum size now.
4417 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4419 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4420 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4421 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4423 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4424 * or more by the traditional way. (See above). It equals:
4426 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4427 * ia64(16K page size) : = ( 8G + 4M)byte.
4428 * powerpc (64K page size) : = (32G +16M)byte.
4430 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4437 * This is an integer logarithm so that shifts can be used later
4438 * to extract the more random high bits from the multiplicative
4439 * hash function before the remainder is taken.
4441 static inline unsigned long wait_table_bits(unsigned long size)
4447 * Initially all pages are reserved - free ones are freed
4448 * up by free_all_bootmem() once the early boot process is
4449 * done. Non-atomic initialization, single-pass.
4451 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4452 unsigned long start_pfn, enum memmap_context context)
4454 pg_data_t *pgdat = NODE_DATA(nid);
4455 unsigned long end_pfn = start_pfn + size;
4458 unsigned long nr_initialised = 0;
4460 if (highest_memmap_pfn < end_pfn - 1)
4461 highest_memmap_pfn = end_pfn - 1;
4463 z = &pgdat->node_zones[zone];
4464 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4466 * There can be holes in boot-time mem_map[]s
4467 * handed to this function. They do not
4468 * exist on hotplugged memory.
4470 if (context == MEMMAP_EARLY) {
4471 if (!early_pfn_valid(pfn))
4473 if (!early_pfn_in_nid(pfn, nid))
4475 if (!update_defer_init(pgdat, pfn, end_pfn,
4481 * Mark the block movable so that blocks are reserved for
4482 * movable at startup. This will force kernel allocations
4483 * to reserve their blocks rather than leaking throughout
4484 * the address space during boot when many long-lived
4485 * kernel allocations are made.
4487 * bitmap is created for zone's valid pfn range. but memmap
4488 * can be created for invalid pages (for alignment)
4489 * check here not to call set_pageblock_migratetype() against
4492 if (!(pfn & (pageblock_nr_pages - 1))) {
4493 struct page *page = pfn_to_page(pfn);
4495 __init_single_page(page, pfn, zone, nid);
4496 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4498 __init_single_pfn(pfn, zone, nid);
4503 static void __meminit zone_init_free_lists(struct zone *zone)
4505 unsigned int order, t;
4506 for_each_migratetype_order(order, t) {
4507 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4508 zone->free_area[order].nr_free = 0;
4512 #ifndef __HAVE_ARCH_MEMMAP_INIT
4513 #define memmap_init(size, nid, zone, start_pfn) \
4514 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4517 static int zone_batchsize(struct zone *zone)
4523 * The per-cpu-pages pools are set to around 1000th of the
4524 * size of the zone. But no more than 1/2 of a meg.
4526 * OK, so we don't know how big the cache is. So guess.
4528 batch = zone->managed_pages / 1024;
4529 if (batch * PAGE_SIZE > 512 * 1024)
4530 batch = (512 * 1024) / PAGE_SIZE;
4531 batch /= 4; /* We effectively *= 4 below */
4536 * Clamp the batch to a 2^n - 1 value. Having a power
4537 * of 2 value was found to be more likely to have
4538 * suboptimal cache aliasing properties in some cases.
4540 * For example if 2 tasks are alternately allocating
4541 * batches of pages, one task can end up with a lot
4542 * of pages of one half of the possible page colors
4543 * and the other with pages of the other colors.
4545 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4550 /* The deferral and batching of frees should be suppressed under NOMMU
4553 * The problem is that NOMMU needs to be able to allocate large chunks
4554 * of contiguous memory as there's no hardware page translation to
4555 * assemble apparent contiguous memory from discontiguous pages.
4557 * Queueing large contiguous runs of pages for batching, however,
4558 * causes the pages to actually be freed in smaller chunks. As there
4559 * can be a significant delay between the individual batches being
4560 * recycled, this leads to the once large chunks of space being
4561 * fragmented and becoming unavailable for high-order allocations.
4568 * pcp->high and pcp->batch values are related and dependent on one another:
4569 * ->batch must never be higher then ->high.
4570 * The following function updates them in a safe manner without read side
4573 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4574 * those fields changing asynchronously (acording the the above rule).
4576 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4577 * outside of boot time (or some other assurance that no concurrent updaters
4580 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4581 unsigned long batch)
4583 /* start with a fail safe value for batch */
4587 /* Update high, then batch, in order */
4594 /* a companion to pageset_set_high() */
4595 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4597 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4600 static void pageset_init(struct per_cpu_pageset *p)
4602 struct per_cpu_pages *pcp;
4605 memset(p, 0, sizeof(*p));
4609 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4610 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4613 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4616 pageset_set_batch(p, batch);
4620 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4621 * to the value high for the pageset p.
4623 static void pageset_set_high(struct per_cpu_pageset *p,
4626 unsigned long batch = max(1UL, high / 4);
4627 if ((high / 4) > (PAGE_SHIFT * 8))
4628 batch = PAGE_SHIFT * 8;
4630 pageset_update(&p->pcp, high, batch);
4633 static void pageset_set_high_and_batch(struct zone *zone,
4634 struct per_cpu_pageset *pcp)
4636 if (percpu_pagelist_fraction)
4637 pageset_set_high(pcp,
4638 (zone->managed_pages /
4639 percpu_pagelist_fraction));
4641 pageset_set_batch(pcp, zone_batchsize(zone));
4644 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4646 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4649 pageset_set_high_and_batch(zone, pcp);
4652 static void __meminit setup_zone_pageset(struct zone *zone)
4655 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4656 for_each_possible_cpu(cpu)
4657 zone_pageset_init(zone, cpu);
4661 * Allocate per cpu pagesets and initialize them.
4662 * Before this call only boot pagesets were available.
4664 void __init setup_per_cpu_pageset(void)
4668 for_each_populated_zone(zone)
4669 setup_zone_pageset(zone);
4672 static noinline __init_refok
4673 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4679 * The per-page waitqueue mechanism uses hashed waitqueues
4682 zone->wait_table_hash_nr_entries =
4683 wait_table_hash_nr_entries(zone_size_pages);
4684 zone->wait_table_bits =
4685 wait_table_bits(zone->wait_table_hash_nr_entries);
4686 alloc_size = zone->wait_table_hash_nr_entries
4687 * sizeof(wait_queue_head_t);
4689 if (!slab_is_available()) {
4690 zone->wait_table = (wait_queue_head_t *)
4691 memblock_virt_alloc_node_nopanic(
4692 alloc_size, zone->zone_pgdat->node_id);
4695 * This case means that a zone whose size was 0 gets new memory
4696 * via memory hot-add.
4697 * But it may be the case that a new node was hot-added. In
4698 * this case vmalloc() will not be able to use this new node's
4699 * memory - this wait_table must be initialized to use this new
4700 * node itself as well.
4701 * To use this new node's memory, further consideration will be
4704 zone->wait_table = vmalloc(alloc_size);
4706 if (!zone->wait_table)
4709 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4710 init_waitqueue_head(zone->wait_table + i);
4715 static __meminit void zone_pcp_init(struct zone *zone)
4718 * per cpu subsystem is not up at this point. The following code
4719 * relies on the ability of the linker to provide the
4720 * offset of a (static) per cpu variable into the per cpu area.
4722 zone->pageset = &boot_pageset;
4724 if (populated_zone(zone))
4725 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4726 zone->name, zone->present_pages,
4727 zone_batchsize(zone));
4730 int __meminit init_currently_empty_zone(struct zone *zone,
4731 unsigned long zone_start_pfn,
4734 struct pglist_data *pgdat = zone->zone_pgdat;
4736 ret = zone_wait_table_init(zone, size);
4739 pgdat->nr_zones = zone_idx(zone) + 1;
4741 zone->zone_start_pfn = zone_start_pfn;
4743 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4744 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4746 (unsigned long)zone_idx(zone),
4747 zone_start_pfn, (zone_start_pfn + size));
4749 zone_init_free_lists(zone);
4754 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4755 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4758 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4760 int __meminit __early_pfn_to_nid(unsigned long pfn,
4761 struct mminit_pfnnid_cache *state)
4763 unsigned long start_pfn, end_pfn;
4766 if (state->last_start <= pfn && pfn < state->last_end)
4767 return state->last_nid;
4769 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4771 state->last_start = start_pfn;
4772 state->last_end = end_pfn;
4773 state->last_nid = nid;
4778 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4781 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4782 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4783 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4785 * If an architecture guarantees that all ranges registered contain no holes
4786 * and may be freed, this this function may be used instead of calling
4787 * memblock_free_early_nid() manually.
4789 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4791 unsigned long start_pfn, end_pfn;
4794 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4795 start_pfn = min(start_pfn, max_low_pfn);
4796 end_pfn = min(end_pfn, max_low_pfn);
4798 if (start_pfn < end_pfn)
4799 memblock_free_early_nid(PFN_PHYS(start_pfn),
4800 (end_pfn - start_pfn) << PAGE_SHIFT,
4806 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4807 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4809 * If an architecture guarantees that all ranges registered contain no holes and may
4810 * be freed, this function may be used instead of calling memory_present() manually.
4812 void __init sparse_memory_present_with_active_regions(int nid)
4814 unsigned long start_pfn, end_pfn;
4817 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4818 memory_present(this_nid, start_pfn, end_pfn);
4822 * get_pfn_range_for_nid - Return the start and end page frames for a node
4823 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4824 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4825 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4827 * It returns the start and end page frame of a node based on information
4828 * provided by memblock_set_node(). If called for a node
4829 * with no available memory, a warning is printed and the start and end
4832 void __meminit get_pfn_range_for_nid(unsigned int nid,
4833 unsigned long *start_pfn, unsigned long *end_pfn)
4835 unsigned long this_start_pfn, this_end_pfn;
4841 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4842 *start_pfn = min(*start_pfn, this_start_pfn);
4843 *end_pfn = max(*end_pfn, this_end_pfn);
4846 if (*start_pfn == -1UL)
4851 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4852 * assumption is made that zones within a node are ordered in monotonic
4853 * increasing memory addresses so that the "highest" populated zone is used
4855 static void __init find_usable_zone_for_movable(void)
4858 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4859 if (zone_index == ZONE_MOVABLE)
4862 if (arch_zone_highest_possible_pfn[zone_index] >
4863 arch_zone_lowest_possible_pfn[zone_index])
4867 VM_BUG_ON(zone_index == -1);
4868 movable_zone = zone_index;
4872 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4873 * because it is sized independent of architecture. Unlike the other zones,
4874 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4875 * in each node depending on the size of each node and how evenly kernelcore
4876 * is distributed. This helper function adjusts the zone ranges
4877 * provided by the architecture for a given node by using the end of the
4878 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4879 * zones within a node are in order of monotonic increases memory addresses
4881 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4882 unsigned long zone_type,
4883 unsigned long node_start_pfn,
4884 unsigned long node_end_pfn,
4885 unsigned long *zone_start_pfn,
4886 unsigned long *zone_end_pfn)
4888 /* Only adjust if ZONE_MOVABLE is on this node */
4889 if (zone_movable_pfn[nid]) {
4890 /* Size ZONE_MOVABLE */
4891 if (zone_type == ZONE_MOVABLE) {
4892 *zone_start_pfn = zone_movable_pfn[nid];
4893 *zone_end_pfn = min(node_end_pfn,
4894 arch_zone_highest_possible_pfn[movable_zone]);
4896 /* Adjust for ZONE_MOVABLE starting within this range */
4897 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4898 *zone_end_pfn > zone_movable_pfn[nid]) {
4899 *zone_end_pfn = zone_movable_pfn[nid];
4901 /* Check if this whole range is within ZONE_MOVABLE */
4902 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4903 *zone_start_pfn = *zone_end_pfn;
4908 * Return the number of pages a zone spans in a node, including holes
4909 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4911 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4912 unsigned long zone_type,
4913 unsigned long node_start_pfn,
4914 unsigned long node_end_pfn,
4915 unsigned long *ignored)
4917 unsigned long zone_start_pfn, zone_end_pfn;
4919 /* When hotadd a new node from cpu_up(), the node should be empty */
4920 if (!node_start_pfn && !node_end_pfn)
4923 /* Get the start and end of the zone */
4924 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4925 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4926 adjust_zone_range_for_zone_movable(nid, zone_type,
4927 node_start_pfn, node_end_pfn,
4928 &zone_start_pfn, &zone_end_pfn);
4930 /* Check that this node has pages within the zone's required range */
4931 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4934 /* Move the zone boundaries inside the node if necessary */
4935 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4936 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4938 /* Return the spanned pages */
4939 return zone_end_pfn - zone_start_pfn;
4943 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4944 * then all holes in the requested range will be accounted for.
4946 unsigned long __meminit __absent_pages_in_range(int nid,
4947 unsigned long range_start_pfn,
4948 unsigned long range_end_pfn)
4950 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4951 unsigned long start_pfn, end_pfn;
4954 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4955 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4956 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4957 nr_absent -= end_pfn - start_pfn;
4963 * absent_pages_in_range - Return number of page frames in holes within a range
4964 * @start_pfn: The start PFN to start searching for holes
4965 * @end_pfn: The end PFN to stop searching for holes
4967 * It returns the number of pages frames in memory holes within a range.
4969 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4970 unsigned long end_pfn)
4972 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4975 /* Return the number of page frames in holes in a zone on a node */
4976 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4977 unsigned long zone_type,
4978 unsigned long node_start_pfn,
4979 unsigned long node_end_pfn,
4980 unsigned long *ignored)
4982 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4983 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4984 unsigned long zone_start_pfn, zone_end_pfn;
4986 /* When hotadd a new node from cpu_up(), the node should be empty */
4987 if (!node_start_pfn && !node_end_pfn)
4990 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4991 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4993 adjust_zone_range_for_zone_movable(nid, zone_type,
4994 node_start_pfn, node_end_pfn,
4995 &zone_start_pfn, &zone_end_pfn);
4996 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4999 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5000 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5001 unsigned long zone_type,
5002 unsigned long node_start_pfn,
5003 unsigned long node_end_pfn,
5004 unsigned long *zones_size)
5006 return zones_size[zone_type];
5009 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5010 unsigned long zone_type,
5011 unsigned long node_start_pfn,
5012 unsigned long node_end_pfn,
5013 unsigned long *zholes_size)
5018 return zholes_size[zone_type];
5021 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5023 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5024 unsigned long node_start_pfn,
5025 unsigned long node_end_pfn,
5026 unsigned long *zones_size,
5027 unsigned long *zholes_size)
5029 unsigned long realtotalpages = 0, totalpages = 0;
5032 for (i = 0; i < MAX_NR_ZONES; i++) {
5033 struct zone *zone = pgdat->node_zones + i;
5034 unsigned long size, real_size;
5036 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5040 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5041 node_start_pfn, node_end_pfn,
5043 zone->spanned_pages = size;
5044 zone->present_pages = real_size;
5047 realtotalpages += real_size;
5050 pgdat->node_spanned_pages = totalpages;
5051 pgdat->node_present_pages = realtotalpages;
5052 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5056 #ifndef CONFIG_SPARSEMEM
5058 * Calculate the size of the zone->blockflags rounded to an unsigned long
5059 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5060 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5061 * round what is now in bits to nearest long in bits, then return it in
5064 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5066 unsigned long usemapsize;
5068 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5069 usemapsize = roundup(zonesize, pageblock_nr_pages);
5070 usemapsize = usemapsize >> pageblock_order;
5071 usemapsize *= NR_PAGEBLOCK_BITS;
5072 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5074 return usemapsize / 8;
5077 static void __init setup_usemap(struct pglist_data *pgdat,
5079 unsigned long zone_start_pfn,
5080 unsigned long zonesize)
5082 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5083 zone->pageblock_flags = NULL;
5085 zone->pageblock_flags =
5086 memblock_virt_alloc_node_nopanic(usemapsize,
5090 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5091 unsigned long zone_start_pfn, unsigned long zonesize) {}
5092 #endif /* CONFIG_SPARSEMEM */
5094 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5096 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5097 void __paginginit set_pageblock_order(void)
5101 /* Check that pageblock_nr_pages has not already been setup */
5102 if (pageblock_order)
5105 if (HPAGE_SHIFT > PAGE_SHIFT)
5106 order = HUGETLB_PAGE_ORDER;
5108 order = MAX_ORDER - 1;
5111 * Assume the largest contiguous order of interest is a huge page.
5112 * This value may be variable depending on boot parameters on IA64 and
5115 pageblock_order = order;
5117 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5120 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5121 * is unused as pageblock_order is set at compile-time. See
5122 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5125 void __paginginit set_pageblock_order(void)
5129 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5131 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5132 unsigned long present_pages)
5134 unsigned long pages = spanned_pages;
5137 * Provide a more accurate estimation if there are holes within
5138 * the zone and SPARSEMEM is in use. If there are holes within the
5139 * zone, each populated memory region may cost us one or two extra
5140 * memmap pages due to alignment because memmap pages for each
5141 * populated regions may not naturally algined on page boundary.
5142 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5144 if (spanned_pages > present_pages + (present_pages >> 4) &&
5145 IS_ENABLED(CONFIG_SPARSEMEM))
5146 pages = present_pages;
5148 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5152 * Set up the zone data structures:
5153 * - mark all pages reserved
5154 * - mark all memory queues empty
5155 * - clear the memory bitmaps
5157 * NOTE: pgdat should get zeroed by caller.
5159 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5162 int nid = pgdat->node_id;
5163 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5166 pgdat_resize_init(pgdat);
5167 #ifdef CONFIG_NUMA_BALANCING
5168 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5169 pgdat->numabalancing_migrate_nr_pages = 0;
5170 pgdat->numabalancing_migrate_next_window = jiffies;
5172 init_waitqueue_head(&pgdat->kswapd_wait);
5173 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5174 pgdat_page_ext_init(pgdat);
5176 for (j = 0; j < MAX_NR_ZONES; j++) {
5177 struct zone *zone = pgdat->node_zones + j;
5178 unsigned long size, realsize, freesize, memmap_pages;
5180 size = zone->spanned_pages;
5181 realsize = freesize = zone->present_pages;
5184 * Adjust freesize so that it accounts for how much memory
5185 * is used by this zone for memmap. This affects the watermark
5186 * and per-cpu initialisations
5188 memmap_pages = calc_memmap_size(size, realsize);
5189 if (!is_highmem_idx(j)) {
5190 if (freesize >= memmap_pages) {
5191 freesize -= memmap_pages;
5194 " %s zone: %lu pages used for memmap\n",
5195 zone_names[j], memmap_pages);
5198 " %s zone: %lu pages exceeds freesize %lu\n",
5199 zone_names[j], memmap_pages, freesize);
5202 /* Account for reserved pages */
5203 if (j == 0 && freesize > dma_reserve) {
5204 freesize -= dma_reserve;
5205 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5206 zone_names[0], dma_reserve);
5209 if (!is_highmem_idx(j))
5210 nr_kernel_pages += freesize;
5211 /* Charge for highmem memmap if there are enough kernel pages */
5212 else if (nr_kernel_pages > memmap_pages * 2)
5213 nr_kernel_pages -= memmap_pages;
5214 nr_all_pages += freesize;
5217 * Set an approximate value for lowmem here, it will be adjusted
5218 * when the bootmem allocator frees pages into the buddy system.
5219 * And all highmem pages will be managed by the buddy system.
5221 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5224 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5226 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5228 zone->name = zone_names[j];
5229 spin_lock_init(&zone->lock);
5230 spin_lock_init(&zone->lru_lock);
5231 zone_seqlock_init(zone);
5232 zone->zone_pgdat = pgdat;
5233 zone_pcp_init(zone);
5235 /* For bootup, initialized properly in watermark setup */
5236 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5238 lruvec_init(&zone->lruvec);
5242 set_pageblock_order();
5243 setup_usemap(pgdat, zone, zone_start_pfn, size);
5244 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5246 memmap_init(size, nid, j, zone_start_pfn);
5247 zone_start_pfn += size;
5251 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5253 unsigned long __maybe_unused offset = 0;
5255 /* Skip empty nodes */
5256 if (!pgdat->node_spanned_pages)
5259 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5260 /* ia64 gets its own node_mem_map, before this, without bootmem */
5261 if (!pgdat->node_mem_map) {
5262 unsigned long size, start, end;
5266 * The zone's endpoints aren't required to be MAX_ORDER
5267 * aligned but the node_mem_map endpoints must be in order
5268 * for the buddy allocator to function correctly.
5270 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5271 offset = pgdat->node_start_pfn - start;
5272 end = pgdat_end_pfn(pgdat);
5273 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5274 size = (end - start) * sizeof(struct page);
5275 map = alloc_remap(pgdat->node_id, size);
5277 map = memblock_virt_alloc_node_nopanic(size,
5279 pgdat->node_mem_map = map + offset;
5281 #ifndef CONFIG_NEED_MULTIPLE_NODES
5283 * With no DISCONTIG, the global mem_map is just set as node 0's
5285 if (pgdat == NODE_DATA(0)) {
5286 mem_map = NODE_DATA(0)->node_mem_map;
5287 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5288 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5290 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5293 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5296 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5297 unsigned long node_start_pfn, unsigned long *zholes_size)
5299 pg_data_t *pgdat = NODE_DATA(nid);
5300 unsigned long start_pfn = 0;
5301 unsigned long end_pfn = 0;
5303 /* pg_data_t should be reset to zero when it's allocated */
5304 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5306 reset_deferred_meminit(pgdat);
5307 pgdat->node_id = nid;
5308 pgdat->node_start_pfn = node_start_pfn;
5309 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5310 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5311 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5312 (u64)start_pfn << PAGE_SHIFT,
5313 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5315 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5316 zones_size, zholes_size);
5318 alloc_node_mem_map(pgdat);
5319 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5320 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5321 nid, (unsigned long)pgdat,
5322 (unsigned long)pgdat->node_mem_map);
5325 free_area_init_core(pgdat);
5328 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5330 #if MAX_NUMNODES > 1
5332 * Figure out the number of possible node ids.
5334 void __init setup_nr_node_ids(void)
5336 unsigned int highest;
5338 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5339 nr_node_ids = highest + 1;
5344 * node_map_pfn_alignment - determine the maximum internode alignment
5346 * This function should be called after node map is populated and sorted.
5347 * It calculates the maximum power of two alignment which can distinguish
5350 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5351 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5352 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5353 * shifted, 1GiB is enough and this function will indicate so.
5355 * This is used to test whether pfn -> nid mapping of the chosen memory
5356 * model has fine enough granularity to avoid incorrect mapping for the
5357 * populated node map.
5359 * Returns the determined alignment in pfn's. 0 if there is no alignment
5360 * requirement (single node).
5362 unsigned long __init node_map_pfn_alignment(void)
5364 unsigned long accl_mask = 0, last_end = 0;
5365 unsigned long start, end, mask;
5369 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5370 if (!start || last_nid < 0 || last_nid == nid) {
5377 * Start with a mask granular enough to pin-point to the
5378 * start pfn and tick off bits one-by-one until it becomes
5379 * too coarse to separate the current node from the last.
5381 mask = ~((1 << __ffs(start)) - 1);
5382 while (mask && last_end <= (start & (mask << 1)))
5385 /* accumulate all internode masks */
5389 /* convert mask to number of pages */
5390 return ~accl_mask + 1;
5393 /* Find the lowest pfn for a node */
5394 static unsigned long __init find_min_pfn_for_node(int nid)
5396 unsigned long min_pfn = ULONG_MAX;
5397 unsigned long start_pfn;
5400 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5401 min_pfn = min(min_pfn, start_pfn);
5403 if (min_pfn == ULONG_MAX) {
5405 "Could not find start_pfn for node %d\n", nid);
5413 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5415 * It returns the minimum PFN based on information provided via
5416 * memblock_set_node().
5418 unsigned long __init find_min_pfn_with_active_regions(void)
5420 return find_min_pfn_for_node(MAX_NUMNODES);
5424 * early_calculate_totalpages()
5425 * Sum pages in active regions for movable zone.
5426 * Populate N_MEMORY for calculating usable_nodes.
5428 static unsigned long __init early_calculate_totalpages(void)
5430 unsigned long totalpages = 0;
5431 unsigned long start_pfn, end_pfn;
5434 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5435 unsigned long pages = end_pfn - start_pfn;
5437 totalpages += pages;
5439 node_set_state(nid, N_MEMORY);
5445 * Find the PFN the Movable zone begins in each node. Kernel memory
5446 * is spread evenly between nodes as long as the nodes have enough
5447 * memory. When they don't, some nodes will have more kernelcore than
5450 static void __init find_zone_movable_pfns_for_nodes(void)
5453 unsigned long usable_startpfn;
5454 unsigned long kernelcore_node, kernelcore_remaining;
5455 /* save the state before borrow the nodemask */
5456 nodemask_t saved_node_state = node_states[N_MEMORY];
5457 unsigned long totalpages = early_calculate_totalpages();
5458 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5459 struct memblock_region *r;
5461 /* Need to find movable_zone earlier when movable_node is specified. */
5462 find_usable_zone_for_movable();
5465 * If movable_node is specified, ignore kernelcore and movablecore
5468 if (movable_node_is_enabled()) {
5469 for_each_memblock(memory, r) {
5470 if (!memblock_is_hotpluggable(r))
5475 usable_startpfn = PFN_DOWN(r->base);
5476 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5477 min(usable_startpfn, zone_movable_pfn[nid]) :
5485 * If movablecore=nn[KMG] was specified, calculate what size of
5486 * kernelcore that corresponds so that memory usable for
5487 * any allocation type is evenly spread. If both kernelcore
5488 * and movablecore are specified, then the value of kernelcore
5489 * will be used for required_kernelcore if it's greater than
5490 * what movablecore would have allowed.
5492 if (required_movablecore) {
5493 unsigned long corepages;
5496 * Round-up so that ZONE_MOVABLE is at least as large as what
5497 * was requested by the user
5499 required_movablecore =
5500 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5501 required_movablecore = min(totalpages, required_movablecore);
5502 corepages = totalpages - required_movablecore;
5504 required_kernelcore = max(required_kernelcore, corepages);
5508 * If kernelcore was not specified or kernelcore size is larger
5509 * than totalpages, there is no ZONE_MOVABLE.
5511 if (!required_kernelcore || required_kernelcore >= totalpages)
5514 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5515 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5518 /* Spread kernelcore memory as evenly as possible throughout nodes */
5519 kernelcore_node = required_kernelcore / usable_nodes;
5520 for_each_node_state(nid, N_MEMORY) {
5521 unsigned long start_pfn, end_pfn;
5524 * Recalculate kernelcore_node if the division per node
5525 * now exceeds what is necessary to satisfy the requested
5526 * amount of memory for the kernel
5528 if (required_kernelcore < kernelcore_node)
5529 kernelcore_node = required_kernelcore / usable_nodes;
5532 * As the map is walked, we track how much memory is usable
5533 * by the kernel using kernelcore_remaining. When it is
5534 * 0, the rest of the node is usable by ZONE_MOVABLE
5536 kernelcore_remaining = kernelcore_node;
5538 /* Go through each range of PFNs within this node */
5539 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5540 unsigned long size_pages;
5542 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5543 if (start_pfn >= end_pfn)
5546 /* Account for what is only usable for kernelcore */
5547 if (start_pfn < usable_startpfn) {
5548 unsigned long kernel_pages;
5549 kernel_pages = min(end_pfn, usable_startpfn)
5552 kernelcore_remaining -= min(kernel_pages,
5553 kernelcore_remaining);
5554 required_kernelcore -= min(kernel_pages,
5555 required_kernelcore);
5557 /* Continue if range is now fully accounted */
5558 if (end_pfn <= usable_startpfn) {
5561 * Push zone_movable_pfn to the end so
5562 * that if we have to rebalance
5563 * kernelcore across nodes, we will
5564 * not double account here
5566 zone_movable_pfn[nid] = end_pfn;
5569 start_pfn = usable_startpfn;
5573 * The usable PFN range for ZONE_MOVABLE is from
5574 * start_pfn->end_pfn. Calculate size_pages as the
5575 * number of pages used as kernelcore
5577 size_pages = end_pfn - start_pfn;
5578 if (size_pages > kernelcore_remaining)
5579 size_pages = kernelcore_remaining;
5580 zone_movable_pfn[nid] = start_pfn + size_pages;
5583 * Some kernelcore has been met, update counts and
5584 * break if the kernelcore for this node has been
5587 required_kernelcore -= min(required_kernelcore,
5589 kernelcore_remaining -= size_pages;
5590 if (!kernelcore_remaining)
5596 * If there is still required_kernelcore, we do another pass with one
5597 * less node in the count. This will push zone_movable_pfn[nid] further
5598 * along on the nodes that still have memory until kernelcore is
5602 if (usable_nodes && required_kernelcore > usable_nodes)
5606 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5607 for (nid = 0; nid < MAX_NUMNODES; nid++)
5608 zone_movable_pfn[nid] =
5609 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5612 /* restore the node_state */
5613 node_states[N_MEMORY] = saved_node_state;
5616 /* Any regular or high memory on that node ? */
5617 static void check_for_memory(pg_data_t *pgdat, int nid)
5619 enum zone_type zone_type;
5621 if (N_MEMORY == N_NORMAL_MEMORY)
5624 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5625 struct zone *zone = &pgdat->node_zones[zone_type];
5626 if (populated_zone(zone)) {
5627 node_set_state(nid, N_HIGH_MEMORY);
5628 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5629 zone_type <= ZONE_NORMAL)
5630 node_set_state(nid, N_NORMAL_MEMORY);
5637 * free_area_init_nodes - Initialise all pg_data_t and zone data
5638 * @max_zone_pfn: an array of max PFNs for each zone
5640 * This will call free_area_init_node() for each active node in the system.
5641 * Using the page ranges provided by memblock_set_node(), the size of each
5642 * zone in each node and their holes is calculated. If the maximum PFN
5643 * between two adjacent zones match, it is assumed that the zone is empty.
5644 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5645 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5646 * starts where the previous one ended. For example, ZONE_DMA32 starts
5647 * at arch_max_dma_pfn.
5649 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5651 unsigned long start_pfn, end_pfn;
5654 /* Record where the zone boundaries are */
5655 memset(arch_zone_lowest_possible_pfn, 0,
5656 sizeof(arch_zone_lowest_possible_pfn));
5657 memset(arch_zone_highest_possible_pfn, 0,
5658 sizeof(arch_zone_highest_possible_pfn));
5659 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5660 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5661 for (i = 1; i < MAX_NR_ZONES; i++) {
5662 if (i == ZONE_MOVABLE)
5664 arch_zone_lowest_possible_pfn[i] =
5665 arch_zone_highest_possible_pfn[i-1];
5666 arch_zone_highest_possible_pfn[i] =
5667 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5669 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5670 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5672 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5673 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5674 find_zone_movable_pfns_for_nodes();
5676 /* Print out the zone ranges */
5677 pr_info("Zone ranges:\n");
5678 for (i = 0; i < MAX_NR_ZONES; i++) {
5679 if (i == ZONE_MOVABLE)
5681 pr_info(" %-8s ", zone_names[i]);
5682 if (arch_zone_lowest_possible_pfn[i] ==
5683 arch_zone_highest_possible_pfn[i])
5686 pr_cont("[mem %#018Lx-%#018Lx]\n",
5687 (u64)arch_zone_lowest_possible_pfn[i]
5689 ((u64)arch_zone_highest_possible_pfn[i]
5690 << PAGE_SHIFT) - 1);
5693 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5694 pr_info("Movable zone start for each node\n");
5695 for (i = 0; i < MAX_NUMNODES; i++) {
5696 if (zone_movable_pfn[i])
5697 pr_info(" Node %d: %#018Lx\n", i,
5698 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5701 /* Print out the early node map */
5702 pr_info("Early memory node ranges\n");
5703 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5704 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5705 (u64)start_pfn << PAGE_SHIFT,
5706 ((u64)end_pfn << PAGE_SHIFT) - 1);
5708 /* Initialise every node */
5709 mminit_verify_pageflags_layout();
5710 setup_nr_node_ids();
5711 for_each_online_node(nid) {
5712 pg_data_t *pgdat = NODE_DATA(nid);
5713 free_area_init_node(nid, NULL,
5714 find_min_pfn_for_node(nid), NULL);
5716 /* Any memory on that node */
5717 if (pgdat->node_present_pages)
5718 node_set_state(nid, N_MEMORY);
5719 check_for_memory(pgdat, nid);
5723 static int __init cmdline_parse_core(char *p, unsigned long *core)
5725 unsigned long long coremem;
5729 coremem = memparse(p, &p);
5730 *core = coremem >> PAGE_SHIFT;
5732 /* Paranoid check that UL is enough for the coremem value */
5733 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5739 * kernelcore=size sets the amount of memory for use for allocations that
5740 * cannot be reclaimed or migrated.
5742 static int __init cmdline_parse_kernelcore(char *p)
5744 return cmdline_parse_core(p, &required_kernelcore);
5748 * movablecore=size sets the amount of memory for use for allocations that
5749 * can be reclaimed or migrated.
5751 static int __init cmdline_parse_movablecore(char *p)
5753 return cmdline_parse_core(p, &required_movablecore);
5756 early_param("kernelcore", cmdline_parse_kernelcore);
5757 early_param("movablecore", cmdline_parse_movablecore);
5759 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5761 void adjust_managed_page_count(struct page *page, long count)
5763 spin_lock(&managed_page_count_lock);
5764 page_zone(page)->managed_pages += count;
5765 totalram_pages += count;
5766 #ifdef CONFIG_HIGHMEM
5767 if (PageHighMem(page))
5768 totalhigh_pages += count;
5770 spin_unlock(&managed_page_count_lock);
5772 EXPORT_SYMBOL(adjust_managed_page_count);
5774 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5777 unsigned long pages = 0;
5779 start = (void *)PAGE_ALIGN((unsigned long)start);
5780 end = (void *)((unsigned long)end & PAGE_MASK);
5781 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5782 if ((unsigned int)poison <= 0xFF)
5783 memset(pos, poison, PAGE_SIZE);
5784 free_reserved_page(virt_to_page(pos));
5788 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5789 s, pages << (PAGE_SHIFT - 10), start, end);
5793 EXPORT_SYMBOL(free_reserved_area);
5795 #ifdef CONFIG_HIGHMEM
5796 void free_highmem_page(struct page *page)
5798 __free_reserved_page(page);
5800 page_zone(page)->managed_pages++;
5806 void __init mem_init_print_info(const char *str)
5808 unsigned long physpages, codesize, datasize, rosize, bss_size;
5809 unsigned long init_code_size, init_data_size;
5811 physpages = get_num_physpages();
5812 codesize = _etext - _stext;
5813 datasize = _edata - _sdata;
5814 rosize = __end_rodata - __start_rodata;
5815 bss_size = __bss_stop - __bss_start;
5816 init_data_size = __init_end - __init_begin;
5817 init_code_size = _einittext - _sinittext;
5820 * Detect special cases and adjust section sizes accordingly:
5821 * 1) .init.* may be embedded into .data sections
5822 * 2) .init.text.* may be out of [__init_begin, __init_end],
5823 * please refer to arch/tile/kernel/vmlinux.lds.S.
5824 * 3) .rodata.* may be embedded into .text or .data sections.
5826 #define adj_init_size(start, end, size, pos, adj) \
5828 if (start <= pos && pos < end && size > adj) \
5832 adj_init_size(__init_begin, __init_end, init_data_size,
5833 _sinittext, init_code_size);
5834 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5835 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5836 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5837 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5839 #undef adj_init_size
5841 pr_info("Memory: %luK/%luK available "
5842 "(%luK kernel code, %luK rwdata, %luK rodata, "
5843 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5844 #ifdef CONFIG_HIGHMEM
5848 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5849 codesize >> 10, datasize >> 10, rosize >> 10,
5850 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5851 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5852 totalcma_pages << (PAGE_SHIFT-10),
5853 #ifdef CONFIG_HIGHMEM
5854 totalhigh_pages << (PAGE_SHIFT-10),
5856 str ? ", " : "", str ? str : "");
5860 * set_dma_reserve - set the specified number of pages reserved in the first zone
5861 * @new_dma_reserve: The number of pages to mark reserved
5863 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5864 * In the DMA zone, a significant percentage may be consumed by kernel image
5865 * and other unfreeable allocations which can skew the watermarks badly. This
5866 * function may optionally be used to account for unfreeable pages in the
5867 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5868 * smaller per-cpu batchsize.
5870 void __init set_dma_reserve(unsigned long new_dma_reserve)
5872 dma_reserve = new_dma_reserve;
5875 void __init free_area_init(unsigned long *zones_size)
5877 free_area_init_node(0, zones_size,
5878 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5881 static int page_alloc_cpu_notify(struct notifier_block *self,
5882 unsigned long action, void *hcpu)
5884 int cpu = (unsigned long)hcpu;
5886 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5887 lru_add_drain_cpu(cpu);
5891 * Spill the event counters of the dead processor
5892 * into the current processors event counters.
5893 * This artificially elevates the count of the current
5896 vm_events_fold_cpu(cpu);
5899 * Zero the differential counters of the dead processor
5900 * so that the vm statistics are consistent.
5902 * This is only okay since the processor is dead and cannot
5903 * race with what we are doing.
5905 cpu_vm_stats_fold(cpu);
5910 void __init page_alloc_init(void)
5912 hotcpu_notifier(page_alloc_cpu_notify, 0);
5916 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5917 * or min_free_kbytes changes.
5919 static void calculate_totalreserve_pages(void)
5921 struct pglist_data *pgdat;
5922 unsigned long reserve_pages = 0;
5923 enum zone_type i, j;
5925 for_each_online_pgdat(pgdat) {
5926 for (i = 0; i < MAX_NR_ZONES; i++) {
5927 struct zone *zone = pgdat->node_zones + i;
5930 /* Find valid and maximum lowmem_reserve in the zone */
5931 for (j = i; j < MAX_NR_ZONES; j++) {
5932 if (zone->lowmem_reserve[j] > max)
5933 max = zone->lowmem_reserve[j];
5936 /* we treat the high watermark as reserved pages. */
5937 max += high_wmark_pages(zone);
5939 if (max > zone->managed_pages)
5940 max = zone->managed_pages;
5941 reserve_pages += max;
5943 * Lowmem reserves are not available to
5944 * GFP_HIGHUSER page cache allocations and
5945 * kswapd tries to balance zones to their high
5946 * watermark. As a result, neither should be
5947 * regarded as dirtyable memory, to prevent a
5948 * situation where reclaim has to clean pages
5949 * in order to balance the zones.
5951 zone->dirty_balance_reserve = max;
5954 dirty_balance_reserve = reserve_pages;
5955 totalreserve_pages = reserve_pages;
5959 * setup_per_zone_lowmem_reserve - called whenever
5960 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5961 * has a correct pages reserved value, so an adequate number of
5962 * pages are left in the zone after a successful __alloc_pages().
5964 static void setup_per_zone_lowmem_reserve(void)
5966 struct pglist_data *pgdat;
5967 enum zone_type j, idx;
5969 for_each_online_pgdat(pgdat) {
5970 for (j = 0; j < MAX_NR_ZONES; j++) {
5971 struct zone *zone = pgdat->node_zones + j;
5972 unsigned long managed_pages = zone->managed_pages;
5974 zone->lowmem_reserve[j] = 0;
5978 struct zone *lower_zone;
5982 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5983 sysctl_lowmem_reserve_ratio[idx] = 1;
5985 lower_zone = pgdat->node_zones + idx;
5986 lower_zone->lowmem_reserve[j] = managed_pages /
5987 sysctl_lowmem_reserve_ratio[idx];
5988 managed_pages += lower_zone->managed_pages;
5993 /* update totalreserve_pages */
5994 calculate_totalreserve_pages();
5997 static void __setup_per_zone_wmarks(void)
5999 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6000 unsigned long lowmem_pages = 0;
6002 unsigned long flags;
6004 /* Calculate total number of !ZONE_HIGHMEM pages */
6005 for_each_zone(zone) {
6006 if (!is_highmem(zone))
6007 lowmem_pages += zone->managed_pages;
6010 for_each_zone(zone) {
6013 spin_lock_irqsave(&zone->lock, flags);
6014 tmp = (u64)pages_min * zone->managed_pages;
6015 do_div(tmp, lowmem_pages);
6016 if (is_highmem(zone)) {
6018 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6019 * need highmem pages, so cap pages_min to a small
6022 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6023 * deltas control asynch page reclaim, and so should
6024 * not be capped for highmem.
6026 unsigned long min_pages;
6028 min_pages = zone->managed_pages / 1024;
6029 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6030 zone->watermark[WMARK_MIN] = min_pages;
6033 * If it's a lowmem zone, reserve a number of pages
6034 * proportionate to the zone's size.
6036 zone->watermark[WMARK_MIN] = tmp;
6039 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6040 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6042 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6043 high_wmark_pages(zone) - low_wmark_pages(zone) -
6044 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6046 spin_unlock_irqrestore(&zone->lock, flags);
6049 /* update totalreserve_pages */
6050 calculate_totalreserve_pages();
6054 * setup_per_zone_wmarks - called when min_free_kbytes changes
6055 * or when memory is hot-{added|removed}
6057 * Ensures that the watermark[min,low,high] values for each zone are set
6058 * correctly with respect to min_free_kbytes.
6060 void setup_per_zone_wmarks(void)
6062 mutex_lock(&zonelists_mutex);
6063 __setup_per_zone_wmarks();
6064 mutex_unlock(&zonelists_mutex);
6068 * The inactive anon list should be small enough that the VM never has to
6069 * do too much work, but large enough that each inactive page has a chance
6070 * to be referenced again before it is swapped out.
6072 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6073 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6074 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6075 * the anonymous pages are kept on the inactive list.
6078 * memory ratio inactive anon
6079 * -------------------------------------
6088 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6090 unsigned int gb, ratio;
6092 /* Zone size in gigabytes */
6093 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6095 ratio = int_sqrt(10 * gb);
6099 zone->inactive_ratio = ratio;
6102 static void __meminit setup_per_zone_inactive_ratio(void)
6107 calculate_zone_inactive_ratio(zone);
6111 * Initialise min_free_kbytes.
6113 * For small machines we want it small (128k min). For large machines
6114 * we want it large (64MB max). But it is not linear, because network
6115 * bandwidth does not increase linearly with machine size. We use
6117 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6118 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6134 int __meminit init_per_zone_wmark_min(void)
6136 unsigned long lowmem_kbytes;
6137 int new_min_free_kbytes;
6139 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6140 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6142 if (new_min_free_kbytes > user_min_free_kbytes) {
6143 min_free_kbytes = new_min_free_kbytes;
6144 if (min_free_kbytes < 128)
6145 min_free_kbytes = 128;
6146 if (min_free_kbytes > 65536)
6147 min_free_kbytes = 65536;
6149 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6150 new_min_free_kbytes, user_min_free_kbytes);
6152 setup_per_zone_wmarks();
6153 refresh_zone_stat_thresholds();
6154 setup_per_zone_lowmem_reserve();
6155 setup_per_zone_inactive_ratio();
6158 module_init(init_per_zone_wmark_min)
6161 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6162 * that we can call two helper functions whenever min_free_kbytes
6165 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6166 void __user *buffer, size_t *length, loff_t *ppos)
6170 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6175 user_min_free_kbytes = min_free_kbytes;
6176 setup_per_zone_wmarks();
6182 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6183 void __user *buffer, size_t *length, loff_t *ppos)
6188 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6193 zone->min_unmapped_pages = (zone->managed_pages *
6194 sysctl_min_unmapped_ratio) / 100;
6198 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6199 void __user *buffer, size_t *length, loff_t *ppos)
6204 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6209 zone->min_slab_pages = (zone->managed_pages *
6210 sysctl_min_slab_ratio) / 100;
6216 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6217 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6218 * whenever sysctl_lowmem_reserve_ratio changes.
6220 * The reserve ratio obviously has absolutely no relation with the
6221 * minimum watermarks. The lowmem reserve ratio can only make sense
6222 * if in function of the boot time zone sizes.
6224 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6225 void __user *buffer, size_t *length, loff_t *ppos)
6227 proc_dointvec_minmax(table, write, buffer, length, ppos);
6228 setup_per_zone_lowmem_reserve();
6233 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6234 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6235 * pagelist can have before it gets flushed back to buddy allocator.
6237 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6238 void __user *buffer, size_t *length, loff_t *ppos)
6241 int old_percpu_pagelist_fraction;
6244 mutex_lock(&pcp_batch_high_lock);
6245 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6247 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6248 if (!write || ret < 0)
6251 /* Sanity checking to avoid pcp imbalance */
6252 if (percpu_pagelist_fraction &&
6253 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6254 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6260 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6263 for_each_populated_zone(zone) {
6266 for_each_possible_cpu(cpu)
6267 pageset_set_high_and_batch(zone,
6268 per_cpu_ptr(zone->pageset, cpu));
6271 mutex_unlock(&pcp_batch_high_lock);
6276 int hashdist = HASHDIST_DEFAULT;
6278 static int __init set_hashdist(char *str)
6282 hashdist = simple_strtoul(str, &str, 0);
6285 __setup("hashdist=", set_hashdist);
6289 * allocate a large system hash table from bootmem
6290 * - it is assumed that the hash table must contain an exact power-of-2
6291 * quantity of entries
6292 * - limit is the number of hash buckets, not the total allocation size
6294 void *__init alloc_large_system_hash(const char *tablename,
6295 unsigned long bucketsize,
6296 unsigned long numentries,
6299 unsigned int *_hash_shift,
6300 unsigned int *_hash_mask,
6301 unsigned long low_limit,
6302 unsigned long high_limit)
6304 unsigned long long max = high_limit;
6305 unsigned long log2qty, size;
6308 /* allow the kernel cmdline to have a say */
6310 /* round applicable memory size up to nearest megabyte */
6311 numentries = nr_kernel_pages;
6313 /* It isn't necessary when PAGE_SIZE >= 1MB */
6314 if (PAGE_SHIFT < 20)
6315 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6317 /* limit to 1 bucket per 2^scale bytes of low memory */
6318 if (scale > PAGE_SHIFT)
6319 numentries >>= (scale - PAGE_SHIFT);
6321 numentries <<= (PAGE_SHIFT - scale);
6323 /* Make sure we've got at least a 0-order allocation.. */
6324 if (unlikely(flags & HASH_SMALL)) {
6325 /* Makes no sense without HASH_EARLY */
6326 WARN_ON(!(flags & HASH_EARLY));
6327 if (!(numentries >> *_hash_shift)) {
6328 numentries = 1UL << *_hash_shift;
6329 BUG_ON(!numentries);
6331 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6332 numentries = PAGE_SIZE / bucketsize;
6334 numentries = roundup_pow_of_two(numentries);
6336 /* limit allocation size to 1/16 total memory by default */
6338 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6339 do_div(max, bucketsize);
6341 max = min(max, 0x80000000ULL);
6343 if (numentries < low_limit)
6344 numentries = low_limit;
6345 if (numentries > max)
6348 log2qty = ilog2(numentries);
6351 size = bucketsize << log2qty;
6352 if (flags & HASH_EARLY)
6353 table = memblock_virt_alloc_nopanic(size, 0);
6355 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6358 * If bucketsize is not a power-of-two, we may free
6359 * some pages at the end of hash table which
6360 * alloc_pages_exact() automatically does
6362 if (get_order(size) < MAX_ORDER) {
6363 table = alloc_pages_exact(size, GFP_ATOMIC);
6364 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6367 } while (!table && size > PAGE_SIZE && --log2qty);
6370 panic("Failed to allocate %s hash table\n", tablename);
6372 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6375 ilog2(size) - PAGE_SHIFT,
6379 *_hash_shift = log2qty;
6381 *_hash_mask = (1 << log2qty) - 1;
6386 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6387 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6390 #ifdef CONFIG_SPARSEMEM
6391 return __pfn_to_section(pfn)->pageblock_flags;
6393 return zone->pageblock_flags;
6394 #endif /* CONFIG_SPARSEMEM */
6397 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6399 #ifdef CONFIG_SPARSEMEM
6400 pfn &= (PAGES_PER_SECTION-1);
6401 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6403 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6404 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6405 #endif /* CONFIG_SPARSEMEM */
6409 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6410 * @page: The page within the block of interest
6411 * @pfn: The target page frame number
6412 * @end_bitidx: The last bit of interest to retrieve
6413 * @mask: mask of bits that the caller is interested in
6415 * Return: pageblock_bits flags
6417 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6418 unsigned long end_bitidx,
6422 unsigned long *bitmap;
6423 unsigned long bitidx, word_bitidx;
6426 zone = page_zone(page);
6427 bitmap = get_pageblock_bitmap(zone, pfn);
6428 bitidx = pfn_to_bitidx(zone, pfn);
6429 word_bitidx = bitidx / BITS_PER_LONG;
6430 bitidx &= (BITS_PER_LONG-1);
6432 word = bitmap[word_bitidx];
6433 bitidx += end_bitidx;
6434 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6438 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6439 * @page: The page within the block of interest
6440 * @flags: The flags to set
6441 * @pfn: The target page frame number
6442 * @end_bitidx: The last bit of interest
6443 * @mask: mask of bits that the caller is interested in
6445 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6447 unsigned long end_bitidx,
6451 unsigned long *bitmap;
6452 unsigned long bitidx, word_bitidx;
6453 unsigned long old_word, word;
6455 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6457 zone = page_zone(page);
6458 bitmap = get_pageblock_bitmap(zone, pfn);
6459 bitidx = pfn_to_bitidx(zone, pfn);
6460 word_bitidx = bitidx / BITS_PER_LONG;
6461 bitidx &= (BITS_PER_LONG-1);
6463 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6465 bitidx += end_bitidx;
6466 mask <<= (BITS_PER_LONG - bitidx - 1);
6467 flags <<= (BITS_PER_LONG - bitidx - 1);
6469 word = READ_ONCE(bitmap[word_bitidx]);
6471 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6472 if (word == old_word)
6479 * This function checks whether pageblock includes unmovable pages or not.
6480 * If @count is not zero, it is okay to include less @count unmovable pages
6482 * PageLRU check without isolation or lru_lock could race so that
6483 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6484 * expect this function should be exact.
6486 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6487 bool skip_hwpoisoned_pages)
6489 unsigned long pfn, iter, found;
6493 * For avoiding noise data, lru_add_drain_all() should be called
6494 * If ZONE_MOVABLE, the zone never contains unmovable pages
6496 if (zone_idx(zone) == ZONE_MOVABLE)
6498 mt = get_pageblock_migratetype(page);
6499 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6502 pfn = page_to_pfn(page);
6503 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6504 unsigned long check = pfn + iter;
6506 if (!pfn_valid_within(check))
6509 page = pfn_to_page(check);
6512 * Hugepages are not in LRU lists, but they're movable.
6513 * We need not scan over tail pages bacause we don't
6514 * handle each tail page individually in migration.
6516 if (PageHuge(page)) {
6517 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6522 * We can't use page_count without pin a page
6523 * because another CPU can free compound page.
6524 * This check already skips compound tails of THP
6525 * because their page->_count is zero at all time.
6527 if (!atomic_read(&page->_count)) {
6528 if (PageBuddy(page))
6529 iter += (1 << page_order(page)) - 1;
6534 * The HWPoisoned page may be not in buddy system, and
6535 * page_count() is not 0.
6537 if (skip_hwpoisoned_pages && PageHWPoison(page))
6543 * If there are RECLAIMABLE pages, we need to check
6544 * it. But now, memory offline itself doesn't call
6545 * shrink_node_slabs() and it still to be fixed.
6548 * If the page is not RAM, page_count()should be 0.
6549 * we don't need more check. This is an _used_ not-movable page.
6551 * The problematic thing here is PG_reserved pages. PG_reserved
6552 * is set to both of a memory hole page and a _used_ kernel
6561 bool is_pageblock_removable_nolock(struct page *page)
6567 * We have to be careful here because we are iterating over memory
6568 * sections which are not zone aware so we might end up outside of
6569 * the zone but still within the section.
6570 * We have to take care about the node as well. If the node is offline
6571 * its NODE_DATA will be NULL - see page_zone.
6573 if (!node_online(page_to_nid(page)))
6576 zone = page_zone(page);
6577 pfn = page_to_pfn(page);
6578 if (!zone_spans_pfn(zone, pfn))
6581 return !has_unmovable_pages(zone, page, 0, true);
6586 static unsigned long pfn_max_align_down(unsigned long pfn)
6588 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6589 pageblock_nr_pages) - 1);
6592 static unsigned long pfn_max_align_up(unsigned long pfn)
6594 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6595 pageblock_nr_pages));
6598 /* [start, end) must belong to a single zone. */
6599 static int __alloc_contig_migrate_range(struct compact_control *cc,
6600 unsigned long start, unsigned long end)
6602 /* This function is based on compact_zone() from compaction.c. */
6603 unsigned long nr_reclaimed;
6604 unsigned long pfn = start;
6605 unsigned int tries = 0;
6610 while (pfn < end || !list_empty(&cc->migratepages)) {
6611 if (fatal_signal_pending(current)) {
6616 if (list_empty(&cc->migratepages)) {
6617 cc->nr_migratepages = 0;
6618 pfn = isolate_migratepages_range(cc, pfn, end);
6624 } else if (++tries == 5) {
6625 ret = ret < 0 ? ret : -EBUSY;
6629 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6631 cc->nr_migratepages -= nr_reclaimed;
6633 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6634 NULL, 0, cc->mode, MR_CMA);
6637 putback_movable_pages(&cc->migratepages);
6644 * alloc_contig_range() -- tries to allocate given range of pages
6645 * @start: start PFN to allocate
6646 * @end: one-past-the-last PFN to allocate
6647 * @migratetype: migratetype of the underlaying pageblocks (either
6648 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6649 * in range must have the same migratetype and it must
6650 * be either of the two.
6652 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6653 * aligned, however it's the caller's responsibility to guarantee that
6654 * we are the only thread that changes migrate type of pageblocks the
6657 * The PFN range must belong to a single zone.
6659 * Returns zero on success or negative error code. On success all
6660 * pages which PFN is in [start, end) are allocated for the caller and
6661 * need to be freed with free_contig_range().
6663 int alloc_contig_range(unsigned long start, unsigned long end,
6664 unsigned migratetype)
6666 unsigned long outer_start, outer_end;
6669 struct compact_control cc = {
6670 .nr_migratepages = 0,
6672 .zone = page_zone(pfn_to_page(start)),
6673 .mode = MIGRATE_SYNC,
6674 .ignore_skip_hint = true,
6676 INIT_LIST_HEAD(&cc.migratepages);
6679 * What we do here is we mark all pageblocks in range as
6680 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6681 * have different sizes, and due to the way page allocator
6682 * work, we align the range to biggest of the two pages so
6683 * that page allocator won't try to merge buddies from
6684 * different pageblocks and change MIGRATE_ISOLATE to some
6685 * other migration type.
6687 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6688 * migrate the pages from an unaligned range (ie. pages that
6689 * we are interested in). This will put all the pages in
6690 * range back to page allocator as MIGRATE_ISOLATE.
6692 * When this is done, we take the pages in range from page
6693 * allocator removing them from the buddy system. This way
6694 * page allocator will never consider using them.
6696 * This lets us mark the pageblocks back as
6697 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6698 * aligned range but not in the unaligned, original range are
6699 * put back to page allocator so that buddy can use them.
6702 ret = start_isolate_page_range(pfn_max_align_down(start),
6703 pfn_max_align_up(end), migratetype,
6708 ret = __alloc_contig_migrate_range(&cc, start, end);
6713 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6714 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6715 * more, all pages in [start, end) are free in page allocator.
6716 * What we are going to do is to allocate all pages from
6717 * [start, end) (that is remove them from page allocator).
6719 * The only problem is that pages at the beginning and at the
6720 * end of interesting range may be not aligned with pages that
6721 * page allocator holds, ie. they can be part of higher order
6722 * pages. Because of this, we reserve the bigger range and
6723 * once this is done free the pages we are not interested in.
6725 * We don't have to hold zone->lock here because the pages are
6726 * isolated thus they won't get removed from buddy.
6729 lru_add_drain_all();
6730 drain_all_pages(cc.zone);
6733 outer_start = start;
6734 while (!PageBuddy(pfn_to_page(outer_start))) {
6735 if (++order >= MAX_ORDER) {
6739 outer_start &= ~0UL << order;
6742 /* Make sure the range is really isolated. */
6743 if (test_pages_isolated(outer_start, end, false)) {
6744 pr_info("%s: [%lx, %lx) PFNs busy\n",
6745 __func__, outer_start, end);
6750 /* Grab isolated pages from freelists. */
6751 outer_end = isolate_freepages_range(&cc, outer_start, end);
6757 /* Free head and tail (if any) */
6758 if (start != outer_start)
6759 free_contig_range(outer_start, start - outer_start);
6760 if (end != outer_end)
6761 free_contig_range(end, outer_end - end);
6764 undo_isolate_page_range(pfn_max_align_down(start),
6765 pfn_max_align_up(end), migratetype);
6769 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6771 unsigned int count = 0;
6773 for (; nr_pages--; pfn++) {
6774 struct page *page = pfn_to_page(pfn);
6776 count += page_count(page) != 1;
6779 WARN(count != 0, "%d pages are still in use!\n", count);
6783 #ifdef CONFIG_MEMORY_HOTPLUG
6785 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6786 * page high values need to be recalulated.
6788 void __meminit zone_pcp_update(struct zone *zone)
6791 mutex_lock(&pcp_batch_high_lock);
6792 for_each_possible_cpu(cpu)
6793 pageset_set_high_and_batch(zone,
6794 per_cpu_ptr(zone->pageset, cpu));
6795 mutex_unlock(&pcp_batch_high_lock);
6799 void zone_pcp_reset(struct zone *zone)
6801 unsigned long flags;
6803 struct per_cpu_pageset *pset;
6805 /* avoid races with drain_pages() */
6806 local_irq_save(flags);
6807 if (zone->pageset != &boot_pageset) {
6808 for_each_online_cpu(cpu) {
6809 pset = per_cpu_ptr(zone->pageset, cpu);
6810 drain_zonestat(zone, pset);
6812 free_percpu(zone->pageset);
6813 zone->pageset = &boot_pageset;
6815 local_irq_restore(flags);
6818 #ifdef CONFIG_MEMORY_HOTREMOVE
6820 * All pages in the range must be isolated before calling this.
6823 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6827 unsigned int order, i;
6829 unsigned long flags;
6830 /* find the first valid pfn */
6831 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6836 zone = page_zone(pfn_to_page(pfn));
6837 spin_lock_irqsave(&zone->lock, flags);
6839 while (pfn < end_pfn) {
6840 if (!pfn_valid(pfn)) {
6844 page = pfn_to_page(pfn);
6846 * The HWPoisoned page may be not in buddy system, and
6847 * page_count() is not 0.
6849 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6851 SetPageReserved(page);
6855 BUG_ON(page_count(page));
6856 BUG_ON(!PageBuddy(page));
6857 order = page_order(page);
6858 #ifdef CONFIG_DEBUG_VM
6859 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6860 pfn, 1 << order, end_pfn);
6862 list_del(&page->lru);
6863 rmv_page_order(page);
6864 zone->free_area[order].nr_free--;
6865 for (i = 0; i < (1 << order); i++)
6866 SetPageReserved((page+i));
6867 pfn += (1 << order);
6869 spin_unlock_irqrestore(&zone->lock, flags);
6873 #ifdef CONFIG_MEMORY_FAILURE
6874 bool is_free_buddy_page(struct page *page)
6876 struct zone *zone = page_zone(page);
6877 unsigned long pfn = page_to_pfn(page);
6878 unsigned long flags;
6881 spin_lock_irqsave(&zone->lock, flags);
6882 for (order = 0; order < MAX_ORDER; order++) {
6883 struct page *page_head = page - (pfn & ((1 << order) - 1));
6885 if (PageBuddy(page_head) && page_order(page_head) >= order)
6888 spin_unlock_irqrestore(&zone->lock, flags);
6890 return order < MAX_ORDER;