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/locallock.h>
64 #include <linux/page_owner.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;
128 #ifdef CONFIG_PM_SLEEP
130 * The following functions are used by the suspend/hibernate code to temporarily
131 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
132 * while devices are suspended. To avoid races with the suspend/hibernate code,
133 * they should always be called with pm_mutex held (gfp_allowed_mask also should
134 * only be modified with pm_mutex held, unless the suspend/hibernate code is
135 * guaranteed not to run in parallel with that modification).
138 static gfp_t saved_gfp_mask;
140 void pm_restore_gfp_mask(void)
142 WARN_ON(!mutex_is_locked(&pm_mutex));
143 if (saved_gfp_mask) {
144 gfp_allowed_mask = saved_gfp_mask;
149 void pm_restrict_gfp_mask(void)
151 WARN_ON(!mutex_is_locked(&pm_mutex));
152 WARN_ON(saved_gfp_mask);
153 saved_gfp_mask = gfp_allowed_mask;
154 gfp_allowed_mask &= ~GFP_IOFS;
157 bool pm_suspended_storage(void)
159 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
163 #endif /* CONFIG_PM_SLEEP */
165 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
166 int pageblock_order __read_mostly;
169 static void __free_pages_ok(struct page *page, unsigned int order);
172 * results with 256, 32 in the lowmem_reserve sysctl:
173 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
174 * 1G machine -> (16M dma, 784M normal, 224M high)
175 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
176 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
177 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
179 * TBD: should special case ZONE_DMA32 machines here - in those we normally
180 * don't need any ZONE_NORMAL reservation
182 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
183 #ifdef CONFIG_ZONE_DMA
186 #ifdef CONFIG_ZONE_DMA32
189 #ifdef CONFIG_HIGHMEM
195 EXPORT_SYMBOL(totalram_pages);
197 static char * const zone_names[MAX_NR_ZONES] = {
198 #ifdef CONFIG_ZONE_DMA
201 #ifdef CONFIG_ZONE_DMA32
205 #ifdef CONFIG_HIGHMEM
211 int min_free_kbytes = 1024;
212 int user_min_free_kbytes = -1;
214 static unsigned long __meminitdata nr_kernel_pages;
215 static unsigned long __meminitdata nr_all_pages;
216 static unsigned long __meminitdata dma_reserve;
218 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
219 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
220 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
221 static unsigned long __initdata required_kernelcore;
222 static unsigned long __initdata required_movablecore;
223 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
225 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
227 EXPORT_SYMBOL(movable_zone);
228 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
231 int nr_node_ids __read_mostly = MAX_NUMNODES;
232 int nr_online_nodes __read_mostly = 1;
233 EXPORT_SYMBOL(nr_node_ids);
234 EXPORT_SYMBOL(nr_online_nodes);
237 static DEFINE_LOCAL_IRQ_LOCK(pa_lock);
239 #ifdef CONFIG_PREEMPT_RT_BASE
240 # define cpu_lock_irqsave(cpu, flags) \
241 local_lock_irqsave_on(pa_lock, flags, cpu)
242 # define cpu_unlock_irqrestore(cpu, flags) \
243 local_unlock_irqrestore_on(pa_lock, flags, cpu)
245 # define cpu_lock_irqsave(cpu, flags) local_irq_save(flags)
246 # define cpu_unlock_irqrestore(cpu, flags) local_irq_restore(flags)
249 int page_group_by_mobility_disabled __read_mostly;
251 void set_pageblock_migratetype(struct page *page, int migratetype)
253 if (unlikely(page_group_by_mobility_disabled &&
254 migratetype < MIGRATE_PCPTYPES))
255 migratetype = MIGRATE_UNMOVABLE;
257 set_pageblock_flags_group(page, (unsigned long)migratetype,
258 PB_migrate, PB_migrate_end);
261 #ifdef CONFIG_DEBUG_VM
262 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
266 unsigned long pfn = page_to_pfn(page);
267 unsigned long sp, start_pfn;
270 seq = zone_span_seqbegin(zone);
271 start_pfn = zone->zone_start_pfn;
272 sp = zone->spanned_pages;
273 if (!zone_spans_pfn(zone, pfn))
275 } while (zone_span_seqretry(zone, seq));
278 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
279 pfn, zone_to_nid(zone), zone->name,
280 start_pfn, start_pfn + sp);
285 static int page_is_consistent(struct zone *zone, struct page *page)
287 if (!pfn_valid_within(page_to_pfn(page)))
289 if (zone != page_zone(page))
295 * Temporary debugging check for pages not lying within a given zone.
297 static int bad_range(struct zone *zone, struct page *page)
299 if (page_outside_zone_boundaries(zone, page))
301 if (!page_is_consistent(zone, page))
307 static inline int bad_range(struct zone *zone, struct page *page)
313 static void bad_page(struct page *page, const char *reason,
314 unsigned long bad_flags)
316 static unsigned long resume;
317 static unsigned long nr_shown;
318 static unsigned long nr_unshown;
320 /* Don't complain about poisoned pages */
321 if (PageHWPoison(page)) {
322 page_mapcount_reset(page); /* remove PageBuddy */
327 * Allow a burst of 60 reports, then keep quiet for that minute;
328 * or allow a steady drip of one report per second.
330 if (nr_shown == 60) {
331 if (time_before(jiffies, resume)) {
337 "BUG: Bad page state: %lu messages suppressed\n",
344 resume = jiffies + 60 * HZ;
346 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
347 current->comm, page_to_pfn(page));
348 dump_page_badflags(page, reason, bad_flags);
353 /* Leave bad fields for debug, except PageBuddy could make trouble */
354 page_mapcount_reset(page); /* remove PageBuddy */
355 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
359 * Higher-order pages are called "compound pages". They are structured thusly:
361 * The first PAGE_SIZE page is called the "head page".
363 * The remaining PAGE_SIZE pages are called "tail pages".
365 * All pages have PG_compound set. All tail pages have their ->first_page
366 * pointing at the head page.
368 * The first tail page's ->lru.next holds the address of the compound page's
369 * put_page() function. Its ->lru.prev holds the order of allocation.
370 * This usage means that zero-order pages may not be compound.
373 static void free_compound_page(struct page *page)
375 __free_pages_ok(page, compound_order(page));
378 void prep_compound_page(struct page *page, unsigned long order)
381 int nr_pages = 1 << order;
383 set_compound_page_dtor(page, free_compound_page);
384 set_compound_order(page, order);
386 for (i = 1; i < nr_pages; i++) {
387 struct page *p = page + i;
388 set_page_count(p, 0);
389 p->first_page = page;
390 /* Make sure p->first_page is always valid for PageTail() */
396 static inline void prep_zero_page(struct page *page, unsigned int order,
402 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
403 * and __GFP_HIGHMEM from hard or soft interrupt context.
405 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
406 for (i = 0; i < (1 << order); i++)
407 clear_highpage(page + i);
410 #ifdef CONFIG_DEBUG_PAGEALLOC
411 unsigned int _debug_guardpage_minorder;
412 bool _debug_pagealloc_enabled __read_mostly;
413 bool _debug_guardpage_enabled __read_mostly;
415 static int __init early_debug_pagealloc(char *buf)
420 if (strcmp(buf, "on") == 0)
421 _debug_pagealloc_enabled = true;
425 early_param("debug_pagealloc", early_debug_pagealloc);
427 static bool need_debug_guardpage(void)
429 /* If we don't use debug_pagealloc, we don't need guard page */
430 if (!debug_pagealloc_enabled())
436 static void init_debug_guardpage(void)
438 if (!debug_pagealloc_enabled())
441 _debug_guardpage_enabled = true;
444 struct page_ext_operations debug_guardpage_ops = {
445 .need = need_debug_guardpage,
446 .init = init_debug_guardpage,
449 static int __init debug_guardpage_minorder_setup(char *buf)
453 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
454 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
457 _debug_guardpage_minorder = res;
458 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
461 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
463 static inline void set_page_guard(struct zone *zone, struct page *page,
464 unsigned int order, int migratetype)
466 struct page_ext *page_ext;
468 if (!debug_guardpage_enabled())
471 page_ext = lookup_page_ext(page);
472 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
474 INIT_LIST_HEAD(&page->lru);
475 set_page_private(page, order);
476 /* Guard pages are not available for any usage */
477 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
480 static inline void clear_page_guard(struct zone *zone, struct page *page,
481 unsigned int order, int migratetype)
483 struct page_ext *page_ext;
485 if (!debug_guardpage_enabled())
488 page_ext = lookup_page_ext(page);
489 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
491 set_page_private(page, 0);
492 if (!is_migrate_isolate(migratetype))
493 __mod_zone_freepage_state(zone, (1 << order), migratetype);
496 struct page_ext_operations debug_guardpage_ops = { NULL, };
497 static inline void set_page_guard(struct zone *zone, struct page *page,
498 unsigned int order, int migratetype) {}
499 static inline void clear_page_guard(struct zone *zone, struct page *page,
500 unsigned int order, int migratetype) {}
503 static inline void set_page_order(struct page *page, unsigned int order)
505 set_page_private(page, order);
506 __SetPageBuddy(page);
509 static inline void rmv_page_order(struct page *page)
511 __ClearPageBuddy(page);
512 set_page_private(page, 0);
516 * This function checks whether a page is free && is the buddy
517 * we can do coalesce a page and its buddy if
518 * (a) the buddy is not in a hole &&
519 * (b) the buddy is in the buddy system &&
520 * (c) a page and its buddy have the same order &&
521 * (d) a page and its buddy are in the same zone.
523 * For recording whether a page is in the buddy system, we set ->_mapcount
524 * PAGE_BUDDY_MAPCOUNT_VALUE.
525 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
526 * serialized by zone->lock.
528 * For recording page's order, we use page_private(page).
530 static inline int page_is_buddy(struct page *page, struct page *buddy,
533 if (!pfn_valid_within(page_to_pfn(buddy)))
536 if (page_is_guard(buddy) && page_order(buddy) == order) {
537 if (page_zone_id(page) != page_zone_id(buddy))
540 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
545 if (PageBuddy(buddy) && page_order(buddy) == order) {
547 * zone check is done late to avoid uselessly
548 * calculating zone/node ids for pages that could
551 if (page_zone_id(page) != page_zone_id(buddy))
554 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
562 * Freeing function for a buddy system allocator.
564 * The concept of a buddy system is to maintain direct-mapped table
565 * (containing bit values) for memory blocks of various "orders".
566 * The bottom level table contains the map for the smallest allocatable
567 * units of memory (here, pages), and each level above it describes
568 * pairs of units from the levels below, hence, "buddies".
569 * At a high level, all that happens here is marking the table entry
570 * at the bottom level available, and propagating the changes upward
571 * as necessary, plus some accounting needed to play nicely with other
572 * parts of the VM system.
573 * At each level, we keep a list of pages, which are heads of continuous
574 * free pages of length of (1 << order) and marked with _mapcount
575 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
577 * So when we are allocating or freeing one, we can derive the state of the
578 * other. That is, if we allocate a small block, and both were
579 * free, the remainder of the region must be split into blocks.
580 * If a block is freed, and its buddy is also free, then this
581 * triggers coalescing into a block of larger size.
586 static inline void __free_one_page(struct page *page,
588 struct zone *zone, unsigned int order,
591 unsigned long page_idx;
592 unsigned long combined_idx;
593 unsigned long uninitialized_var(buddy_idx);
595 int max_order = MAX_ORDER;
597 VM_BUG_ON(!zone_is_initialized(zone));
598 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
600 VM_BUG_ON(migratetype == -1);
601 if (is_migrate_isolate(migratetype)) {
603 * We restrict max order of merging to prevent merge
604 * between freepages on isolate pageblock and normal
605 * pageblock. Without this, pageblock isolation
606 * could cause incorrect freepage accounting.
608 max_order = min(MAX_ORDER, pageblock_order + 1);
610 __mod_zone_freepage_state(zone, 1 << order, migratetype);
613 page_idx = pfn & ((1 << max_order) - 1);
615 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
616 VM_BUG_ON_PAGE(bad_range(zone, page), page);
618 while (order < max_order - 1) {
619 buddy_idx = __find_buddy_index(page_idx, order);
620 buddy = page + (buddy_idx - page_idx);
621 if (!page_is_buddy(page, buddy, order))
624 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
625 * merge with it and move up one order.
627 if (page_is_guard(buddy)) {
628 clear_page_guard(zone, buddy, order, migratetype);
630 list_del(&buddy->lru);
631 zone->free_area[order].nr_free--;
632 rmv_page_order(buddy);
634 combined_idx = buddy_idx & page_idx;
635 page = page + (combined_idx - page_idx);
636 page_idx = combined_idx;
639 set_page_order(page, order);
642 * If this is not the largest possible page, check if the buddy
643 * of the next-highest order is free. If it is, it's possible
644 * that pages are being freed that will coalesce soon. In case,
645 * that is happening, add the free page to the tail of the list
646 * so it's less likely to be used soon and more likely to be merged
647 * as a higher order page
649 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
650 struct page *higher_page, *higher_buddy;
651 combined_idx = buddy_idx & page_idx;
652 higher_page = page + (combined_idx - page_idx);
653 buddy_idx = __find_buddy_index(combined_idx, order + 1);
654 higher_buddy = higher_page + (buddy_idx - combined_idx);
655 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
656 list_add_tail(&page->lru,
657 &zone->free_area[order].free_list[migratetype]);
662 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
664 zone->free_area[order].nr_free++;
667 static inline int free_pages_check(struct page *page)
669 const char *bad_reason = NULL;
670 unsigned long bad_flags = 0;
672 if (unlikely(page_mapcount(page)))
673 bad_reason = "nonzero mapcount";
674 if (unlikely(page->mapping != NULL))
675 bad_reason = "non-NULL mapping";
676 if (unlikely(atomic_read(&page->_count) != 0))
677 bad_reason = "nonzero _count";
678 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
679 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
680 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
683 if (unlikely(page->mem_cgroup))
684 bad_reason = "page still charged to cgroup";
686 if (unlikely(bad_reason)) {
687 bad_page(page, bad_reason, bad_flags);
690 page_cpupid_reset_last(page);
691 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
692 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
697 * Frees a number of pages which have been collected from the pcp lists.
698 * Assumes all pages on list are in same zone, and of same order.
699 * count is the number of pages to free.
701 * If the zone was previously in an "all pages pinned" state then look to
702 * see if this freeing clears that state.
704 * And clear the zone's pages_scanned counter, to hold off the "all pages are
705 * pinned" detection logic.
707 static void free_pcppages_bulk(struct zone *zone, int count,
708 struct list_head *list)
711 unsigned long nr_scanned;
714 spin_lock_irqsave(&zone->lock, flags);
716 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
718 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
720 while (!list_empty(list)) {
721 struct page *page = list_first_entry(list, struct page, lru);
722 int mt; /* migratetype of the to-be-freed page */
724 /* must delete as __free_one_page list manipulates */
725 list_del(&page->lru);
727 mt = get_freepage_migratetype(page);
728 if (unlikely(has_isolate_pageblock(zone)))
729 mt = get_pageblock_migratetype(page);
731 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
732 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
733 trace_mm_page_pcpu_drain(page, 0, mt);
736 WARN_ON(to_free != 0);
737 spin_unlock_irqrestore(&zone->lock, flags);
741 * Moves a number of pages from the PCP lists to free list which
742 * is freed outside of the locked region.
744 * Assumes all pages on list are in same zone, and of same order.
745 * count is the number of pages to free.
747 static void isolate_pcp_pages(int to_free, struct per_cpu_pages *src,
748 struct list_head *dst)
755 struct list_head *list;
758 * Remove pages from lists in a round-robin fashion. A
759 * batch_free count is maintained that is incremented when an
760 * empty list is encountered. This is so more pages are freed
761 * off fuller lists instead of spinning excessively around empty
766 if (++migratetype == MIGRATE_PCPTYPES)
768 list = &src->lists[migratetype];
769 } while (list_empty(list));
771 /* This is the only non-empty list. Free them all. */
772 if (batch_free == MIGRATE_PCPTYPES)
773 batch_free = to_free;
776 page = list_last_entry(list, struct page, lru);
777 list_del(&page->lru);
778 list_add(&page->lru, dst);
779 } while (--to_free && --batch_free && !list_empty(list));
783 static void free_one_page(struct zone *zone,
784 struct page *page, unsigned long pfn,
788 unsigned long nr_scanned;
791 spin_lock_irqsave(&zone->lock, flags);
792 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
794 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
796 if (unlikely(has_isolate_pageblock(zone) ||
797 is_migrate_isolate(migratetype))) {
798 migratetype = get_pfnblock_migratetype(page, pfn);
800 __free_one_page(page, pfn, zone, order, migratetype);
801 spin_unlock_irqrestore(&zone->lock, flags);
804 static int free_tail_pages_check(struct page *head_page, struct page *page)
806 if (!IS_ENABLED(CONFIG_DEBUG_VM))
808 if (unlikely(!PageTail(page))) {
809 bad_page(page, "PageTail not set", 0);
812 if (unlikely(page->first_page != head_page)) {
813 bad_page(page, "first_page not consistent", 0);
819 static bool free_pages_prepare(struct page *page, unsigned int order)
821 bool compound = PageCompound(page);
824 VM_BUG_ON_PAGE(PageTail(page), page);
825 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
827 trace_mm_page_free(page, order);
828 kmemcheck_free_shadow(page, order);
829 kasan_free_pages(page, order);
832 page->mapping = NULL;
833 bad += free_pages_check(page);
834 for (i = 1; i < (1 << order); i++) {
836 bad += free_tail_pages_check(page, page + i);
837 bad += free_pages_check(page + i);
842 reset_page_owner(page, order);
844 if (!PageHighMem(page)) {
845 debug_check_no_locks_freed(page_address(page),
847 debug_check_no_obj_freed(page_address(page),
850 arch_free_page(page, order);
851 kernel_map_pages(page, 1 << order, 0);
856 static void __free_pages_ok(struct page *page, unsigned int order)
860 unsigned long pfn = page_to_pfn(page);
862 if (!free_pages_prepare(page, order))
865 migratetype = get_pfnblock_migratetype(page, pfn);
866 local_lock_irqsave(pa_lock, flags);
867 __count_vm_events(PGFREE, 1 << order);
868 set_freepage_migratetype(page, migratetype);
869 free_one_page(page_zone(page), page, pfn, order, migratetype);
870 local_unlock_irqrestore(pa_lock, flags);
873 void __init __free_pages_bootmem(struct page *page, unsigned int order)
875 unsigned int nr_pages = 1 << order;
876 struct page *p = page;
880 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
882 __ClearPageReserved(p);
883 set_page_count(p, 0);
885 __ClearPageReserved(p);
886 set_page_count(p, 0);
888 page_zone(page)->managed_pages += nr_pages;
889 set_page_refcounted(page);
890 __free_pages(page, order);
894 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
895 void __init init_cma_reserved_pageblock(struct page *page)
897 unsigned i = pageblock_nr_pages;
898 struct page *p = page;
901 __ClearPageReserved(p);
902 set_page_count(p, 0);
905 set_pageblock_migratetype(page, MIGRATE_CMA);
907 if (pageblock_order >= MAX_ORDER) {
908 i = pageblock_nr_pages;
911 set_page_refcounted(p);
912 __free_pages(p, MAX_ORDER - 1);
913 p += MAX_ORDER_NR_PAGES;
914 } while (i -= MAX_ORDER_NR_PAGES);
916 set_page_refcounted(page);
917 __free_pages(page, pageblock_order);
920 adjust_managed_page_count(page, pageblock_nr_pages);
925 * The order of subdivision here is critical for the IO subsystem.
926 * Please do not alter this order without good reasons and regression
927 * testing. Specifically, as large blocks of memory are subdivided,
928 * the order in which smaller blocks are delivered depends on the order
929 * they're subdivided in this function. This is the primary factor
930 * influencing the order in which pages are delivered to the IO
931 * subsystem according to empirical testing, and this is also justified
932 * by considering the behavior of a buddy system containing a single
933 * large block of memory acted on by a series of small allocations.
934 * This behavior is a critical factor in sglist merging's success.
938 static inline void expand(struct zone *zone, struct page *page,
939 int low, int high, struct free_area *area,
942 unsigned long size = 1 << high;
948 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
950 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
951 debug_guardpage_enabled() &&
952 high < debug_guardpage_minorder()) {
954 * Mark as guard pages (or page), that will allow to
955 * merge back to allocator when buddy will be freed.
956 * Corresponding page table entries will not be touched,
957 * pages will stay not present in virtual address space
959 set_page_guard(zone, &page[size], high, migratetype);
962 list_add(&page[size].lru, &area->free_list[migratetype]);
964 set_page_order(&page[size], high);
969 * This page is about to be returned from the page allocator
971 static inline int check_new_page(struct page *page)
973 const char *bad_reason = NULL;
974 unsigned long bad_flags = 0;
976 if (unlikely(page_mapcount(page)))
977 bad_reason = "nonzero mapcount";
978 if (unlikely(page->mapping != NULL))
979 bad_reason = "non-NULL mapping";
980 if (unlikely(atomic_read(&page->_count) != 0))
981 bad_reason = "nonzero _count";
982 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
983 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
984 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
987 if (unlikely(page->mem_cgroup))
988 bad_reason = "page still charged to cgroup";
990 if (unlikely(bad_reason)) {
991 bad_page(page, bad_reason, bad_flags);
997 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1002 for (i = 0; i < (1 << order); i++) {
1003 struct page *p = page + i;
1004 if (unlikely(check_new_page(p)))
1008 set_page_private(page, 0);
1009 set_page_refcounted(page);
1011 arch_alloc_page(page, order);
1012 kernel_map_pages(page, 1 << order, 1);
1013 kasan_alloc_pages(page, order);
1015 if (gfp_flags & __GFP_ZERO)
1016 prep_zero_page(page, order, gfp_flags);
1018 if (order && (gfp_flags & __GFP_COMP))
1019 prep_compound_page(page, order);
1021 set_page_owner(page, order, gfp_flags);
1024 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
1025 * allocate the page. The expectation is that the caller is taking
1026 * steps that will free more memory. The caller should avoid the page
1027 * being used for !PFMEMALLOC purposes.
1029 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1035 * Go through the free lists for the given migratetype and remove
1036 * the smallest available page from the freelists
1039 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1042 unsigned int current_order;
1043 struct free_area *area;
1046 /* Find a page of the appropriate size in the preferred list */
1047 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1048 area = &(zone->free_area[current_order]);
1049 if (list_empty(&area->free_list[migratetype]))
1052 page = list_entry(area->free_list[migratetype].next,
1054 list_del(&page->lru);
1055 rmv_page_order(page);
1057 expand(zone, page, order, current_order, area, migratetype);
1058 set_freepage_migratetype(page, migratetype);
1067 * This array describes the order lists are fallen back to when
1068 * the free lists for the desirable migrate type are depleted
1070 static int fallbacks[MIGRATE_TYPES][4] = {
1071 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1072 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1073 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1075 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1077 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1078 #ifdef CONFIG_MEMORY_ISOLATION
1079 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1084 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1087 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1090 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1091 unsigned int order) { return NULL; }
1095 * Move the free pages in a range to the free lists of the requested type.
1096 * Note that start_page and end_pages are not aligned on a pageblock
1097 * boundary. If alignment is required, use move_freepages_block()
1099 int move_freepages(struct zone *zone,
1100 struct page *start_page, struct page *end_page,
1104 unsigned long order;
1105 int pages_moved = 0;
1107 #ifndef CONFIG_HOLES_IN_ZONE
1109 * page_zone is not safe to call in this context when
1110 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1111 * anyway as we check zone boundaries in move_freepages_block().
1112 * Remove at a later date when no bug reports exist related to
1113 * grouping pages by mobility
1115 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1118 for (page = start_page; page <= end_page;) {
1119 /* Make sure we are not inadvertently changing nodes */
1120 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1122 if (!pfn_valid_within(page_to_pfn(page))) {
1127 if (!PageBuddy(page)) {
1132 order = page_order(page);
1133 list_move(&page->lru,
1134 &zone->free_area[order].free_list[migratetype]);
1135 set_freepage_migratetype(page, migratetype);
1137 pages_moved += 1 << order;
1143 int move_freepages_block(struct zone *zone, struct page *page,
1146 unsigned long start_pfn, end_pfn;
1147 struct page *start_page, *end_page;
1149 start_pfn = page_to_pfn(page);
1150 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1151 start_page = pfn_to_page(start_pfn);
1152 end_page = start_page + pageblock_nr_pages - 1;
1153 end_pfn = start_pfn + pageblock_nr_pages - 1;
1155 /* Do not cross zone boundaries */
1156 if (!zone_spans_pfn(zone, start_pfn))
1158 if (!zone_spans_pfn(zone, end_pfn))
1161 return move_freepages(zone, start_page, end_page, migratetype);
1164 static void change_pageblock_range(struct page *pageblock_page,
1165 int start_order, int migratetype)
1167 int nr_pageblocks = 1 << (start_order - pageblock_order);
1169 while (nr_pageblocks--) {
1170 set_pageblock_migratetype(pageblock_page, migratetype);
1171 pageblock_page += pageblock_nr_pages;
1176 * When we are falling back to another migratetype during allocation, try to
1177 * steal extra free pages from the same pageblocks to satisfy further
1178 * allocations, instead of polluting multiple pageblocks.
1180 * If we are stealing a relatively large buddy page, it is likely there will
1181 * be more free pages in the pageblock, so try to steal them all. For
1182 * reclaimable and unmovable allocations, we steal regardless of page size,
1183 * as fragmentation caused by those allocations polluting movable pageblocks
1184 * is worse than movable allocations stealing from unmovable and reclaimable
1187 static bool can_steal_fallback(unsigned int order, int start_mt)
1190 * Leaving this order check is intended, although there is
1191 * relaxed order check in next check. The reason is that
1192 * we can actually steal whole pageblock if this condition met,
1193 * but, below check doesn't guarantee it and that is just heuristic
1194 * so could be changed anytime.
1196 if (order >= pageblock_order)
1199 if (order >= pageblock_order / 2 ||
1200 start_mt == MIGRATE_RECLAIMABLE ||
1201 start_mt == MIGRATE_UNMOVABLE ||
1202 page_group_by_mobility_disabled)
1209 * This function implements actual steal behaviour. If order is large enough,
1210 * we can steal whole pageblock. If not, we first move freepages in this
1211 * pageblock and check whether half of pages are moved or not. If half of
1212 * pages are moved, we can change migratetype of pageblock and permanently
1213 * use it's pages as requested migratetype in the future.
1215 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1218 int current_order = page_order(page);
1221 /* Take ownership for orders >= pageblock_order */
1222 if (current_order >= pageblock_order) {
1223 change_pageblock_range(page, current_order, start_type);
1227 pages = move_freepages_block(zone, page, start_type);
1229 /* Claim the whole block if over half of it is free */
1230 if (pages >= (1 << (pageblock_order-1)) ||
1231 page_group_by_mobility_disabled)
1232 set_pageblock_migratetype(page, start_type);
1236 * Check whether there is a suitable fallback freepage with requested order.
1237 * If only_stealable is true, this function returns fallback_mt only if
1238 * we can steal other freepages all together. This would help to reduce
1239 * fragmentation due to mixed migratetype pages in one pageblock.
1241 int find_suitable_fallback(struct free_area *area, unsigned int order,
1242 int migratetype, bool only_stealable, bool *can_steal)
1247 if (area->nr_free == 0)
1252 fallback_mt = fallbacks[migratetype][i];
1253 if (fallback_mt == MIGRATE_RESERVE)
1256 if (list_empty(&area->free_list[fallback_mt]))
1259 if (can_steal_fallback(order, migratetype))
1262 if (!only_stealable)
1272 /* Remove an element from the buddy allocator from the fallback list */
1273 static inline struct page *
1274 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1276 struct free_area *area;
1277 unsigned int current_order;
1282 /* Find the largest possible block of pages in the other list */
1283 for (current_order = MAX_ORDER-1;
1284 current_order >= order && current_order <= MAX_ORDER-1;
1286 area = &(zone->free_area[current_order]);
1287 fallback_mt = find_suitable_fallback(area, current_order,
1288 start_migratetype, false, &can_steal);
1289 if (fallback_mt == -1)
1292 page = list_entry(area->free_list[fallback_mt].next,
1295 steal_suitable_fallback(zone, page, start_migratetype);
1297 /* Remove the page from the freelists */
1299 list_del(&page->lru);
1300 rmv_page_order(page);
1302 expand(zone, page, order, current_order, area,
1305 * The freepage_migratetype may differ from pageblock's
1306 * migratetype depending on the decisions in
1307 * try_to_steal_freepages(). This is OK as long as it
1308 * does not differ for MIGRATE_CMA pageblocks. For CMA
1309 * we need to make sure unallocated pages flushed from
1310 * pcp lists are returned to the correct freelist.
1312 set_freepage_migratetype(page, start_migratetype);
1314 trace_mm_page_alloc_extfrag(page, order, current_order,
1315 start_migratetype, fallback_mt);
1324 * Do the hard work of removing an element from the buddy allocator.
1325 * Call me with the zone->lock already held.
1327 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1333 page = __rmqueue_smallest(zone, order, migratetype);
1335 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1336 if (migratetype == MIGRATE_MOVABLE)
1337 page = __rmqueue_cma_fallback(zone, order);
1340 page = __rmqueue_fallback(zone, order, migratetype);
1343 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1344 * is used because __rmqueue_smallest is an inline function
1345 * and we want just one call site
1348 migratetype = MIGRATE_RESERVE;
1353 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1358 * Obtain a specified number of elements from the buddy allocator, all under
1359 * a single hold of the lock, for efficiency. Add them to the supplied list.
1360 * Returns the number of new pages which were placed at *list.
1362 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1363 unsigned long count, struct list_head *list,
1364 int migratetype, bool cold)
1368 spin_lock(&zone->lock);
1369 for (i = 0; i < count; ++i) {
1370 struct page *page = __rmqueue(zone, order, migratetype);
1371 if (unlikely(page == NULL))
1375 * Split buddy pages returned by expand() are received here
1376 * in physical page order. The page is added to the callers and
1377 * list and the list head then moves forward. From the callers
1378 * perspective, the linked list is ordered by page number in
1379 * some conditions. This is useful for IO devices that can
1380 * merge IO requests if the physical pages are ordered
1384 list_add(&page->lru, list);
1386 list_add_tail(&page->lru, list);
1388 if (is_migrate_cma(get_freepage_migratetype(page)))
1389 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1392 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1393 spin_unlock(&zone->lock);
1399 * Called from the vmstat counter updater to drain pagesets of this
1400 * currently executing processor on remote nodes after they have
1403 * Note that this function must be called with the thread pinned to
1404 * a single processor.
1406 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1408 unsigned long flags;
1410 int to_drain, batch;
1412 local_lock_irqsave(pa_lock, flags);
1413 batch = READ_ONCE(pcp->batch);
1414 to_drain = min(pcp->count, batch);
1416 isolate_pcp_pages(to_drain, pcp, &dst);
1417 pcp->count -= to_drain;
1419 local_unlock_irqrestore(pa_lock, flags);
1420 free_pcppages_bulk(zone, to_drain, &dst);
1425 * Drain pcplists of the indicated processor and zone.
1427 * The processor must either be the current processor and the
1428 * thread pinned to the current processor or a processor that
1431 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1433 unsigned long flags;
1434 struct per_cpu_pageset *pset;
1435 struct per_cpu_pages *pcp;
1439 cpu_lock_irqsave(cpu, flags);
1440 pset = per_cpu_ptr(zone->pageset, cpu);
1445 isolate_pcp_pages(count, pcp, &dst);
1448 cpu_unlock_irqrestore(cpu, flags);
1450 free_pcppages_bulk(zone, count, &dst);
1454 * Drain pcplists of all zones on the indicated processor.
1456 * The processor must either be the current processor and the
1457 * thread pinned to the current processor or a processor that
1460 static void drain_pages(unsigned int cpu)
1464 for_each_populated_zone(zone) {
1465 drain_pages_zone(cpu, zone);
1470 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1472 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1473 * the single zone's pages.
1475 void drain_local_pages(struct zone *zone)
1477 int cpu = smp_processor_id();
1480 drain_pages_zone(cpu, zone);
1486 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1488 * When zone parameter is non-NULL, spill just the single zone's pages.
1490 * Note that this code is protected against sending an IPI to an offline
1491 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1492 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1493 * nothing keeps CPUs from showing up after we populated the cpumask and
1494 * before the call to on_each_cpu_mask().
1496 void drain_all_pages(struct zone *zone)
1501 * Allocate in the BSS so we wont require allocation in
1502 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1504 static cpumask_t cpus_with_pcps;
1507 * We don't care about racing with CPU hotplug event
1508 * as offline notification will cause the notified
1509 * cpu to drain that CPU pcps and on_each_cpu_mask
1510 * disables preemption as part of its processing
1512 for_each_online_cpu(cpu) {
1513 struct per_cpu_pageset *pcp;
1515 bool has_pcps = false;
1518 pcp = per_cpu_ptr(zone->pageset, cpu);
1522 for_each_populated_zone(z) {
1523 pcp = per_cpu_ptr(z->pageset, cpu);
1524 if (pcp->pcp.count) {
1532 cpumask_set_cpu(cpu, &cpus_with_pcps);
1534 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1536 #ifndef CONFIG_PREEMPT_RT_BASE
1537 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1540 for_each_cpu(cpu, &cpus_with_pcps) {
1542 drain_pages_zone(cpu, zone);
1549 #ifdef CONFIG_HIBERNATION
1551 void mark_free_pages(struct zone *zone)
1553 unsigned long pfn, max_zone_pfn;
1554 unsigned long flags;
1555 unsigned int order, t;
1556 struct list_head *curr;
1558 if (zone_is_empty(zone))
1561 spin_lock_irqsave(&zone->lock, flags);
1563 max_zone_pfn = zone_end_pfn(zone);
1564 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1565 if (pfn_valid(pfn)) {
1566 struct page *page = pfn_to_page(pfn);
1568 if (!swsusp_page_is_forbidden(page))
1569 swsusp_unset_page_free(page);
1572 for_each_migratetype_order(order, t) {
1573 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1576 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1577 for (i = 0; i < (1UL << order); i++)
1578 swsusp_set_page_free(pfn_to_page(pfn + i));
1581 spin_unlock_irqrestore(&zone->lock, flags);
1583 #endif /* CONFIG_PM */
1586 * Free a 0-order page
1587 * cold == true ? free a cold page : free a hot page
1589 void free_hot_cold_page(struct page *page, bool cold)
1591 struct zone *zone = page_zone(page);
1592 struct per_cpu_pages *pcp;
1593 unsigned long flags;
1594 unsigned long pfn = page_to_pfn(page);
1597 if (!free_pages_prepare(page, 0))
1600 migratetype = get_pfnblock_migratetype(page, pfn);
1601 set_freepage_migratetype(page, migratetype);
1602 local_lock_irqsave(pa_lock, flags);
1603 __count_vm_event(PGFREE);
1606 * We only track unmovable, reclaimable and movable on pcp lists.
1607 * Free ISOLATE pages back to the allocator because they are being
1608 * offlined but treat RESERVE as movable pages so we can get those
1609 * areas back if necessary. Otherwise, we may have to free
1610 * excessively into the page allocator
1612 if (migratetype >= MIGRATE_PCPTYPES) {
1613 if (unlikely(is_migrate_isolate(migratetype))) {
1614 free_one_page(zone, page, pfn, 0, migratetype);
1617 migratetype = MIGRATE_MOVABLE;
1620 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1622 list_add(&page->lru, &pcp->lists[migratetype]);
1624 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1626 if (pcp->count >= pcp->high) {
1627 unsigned long batch = READ_ONCE(pcp->batch);
1630 isolate_pcp_pages(batch, pcp, &dst);
1631 pcp->count -= batch;
1632 local_unlock_irqrestore(pa_lock, flags);
1633 free_pcppages_bulk(zone, batch, &dst);
1638 local_unlock_irqrestore(pa_lock, flags);
1642 * Free a list of 0-order pages
1644 void free_hot_cold_page_list(struct list_head *list, bool cold)
1646 struct page *page, *next;
1648 list_for_each_entry_safe(page, next, list, lru) {
1649 trace_mm_page_free_batched(page, cold);
1650 free_hot_cold_page(page, cold);
1655 * split_page takes a non-compound higher-order page, and splits it into
1656 * n (1<<order) sub-pages: page[0..n]
1657 * Each sub-page must be freed individually.
1659 * Note: this is probably too low level an operation for use in drivers.
1660 * Please consult with lkml before using this in your driver.
1662 void split_page(struct page *page, unsigned int order)
1666 VM_BUG_ON_PAGE(PageCompound(page), page);
1667 VM_BUG_ON_PAGE(!page_count(page), page);
1669 #ifdef CONFIG_KMEMCHECK
1671 * Split shadow pages too, because free(page[0]) would
1672 * otherwise free the whole shadow.
1674 if (kmemcheck_page_is_tracked(page))
1675 split_page(virt_to_page(page[0].shadow), order);
1678 set_page_owner(page, 0, 0);
1679 for (i = 1; i < (1 << order); i++) {
1680 set_page_refcounted(page + i);
1681 set_page_owner(page + i, 0, 0);
1684 EXPORT_SYMBOL_GPL(split_page);
1686 int __isolate_free_page(struct page *page, unsigned int order)
1688 unsigned long watermark;
1692 BUG_ON(!PageBuddy(page));
1694 zone = page_zone(page);
1695 mt = get_pageblock_migratetype(page);
1697 if (!is_migrate_isolate(mt)) {
1698 /* Obey watermarks as if the page was being allocated */
1699 watermark = low_wmark_pages(zone) + (1 << order);
1700 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1703 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1706 /* Remove page from free list */
1707 list_del(&page->lru);
1708 zone->free_area[order].nr_free--;
1709 rmv_page_order(page);
1711 /* Set the pageblock if the isolated page is at least a pageblock */
1712 if (order >= pageblock_order - 1) {
1713 struct page *endpage = page + (1 << order) - 1;
1714 for (; page < endpage; page += pageblock_nr_pages) {
1715 int mt = get_pageblock_migratetype(page);
1716 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1717 set_pageblock_migratetype(page,
1722 set_page_owner(page, order, 0);
1723 return 1UL << order;
1727 * Similar to split_page except the page is already free. As this is only
1728 * being used for migration, the migratetype of the block also changes.
1729 * As this is called with interrupts disabled, the caller is responsible
1730 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1733 * Note: this is probably too low level an operation for use in drivers.
1734 * Please consult with lkml before using this in your driver.
1736 int split_free_page(struct page *page)
1741 order = page_order(page);
1743 nr_pages = __isolate_free_page(page, order);
1747 /* Split into individual pages */
1748 set_page_refcounted(page);
1749 split_page(page, order);
1754 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1757 struct page *buffered_rmqueue(struct zone *preferred_zone,
1758 struct zone *zone, unsigned int order,
1759 gfp_t gfp_flags, int migratetype)
1761 unsigned long flags;
1763 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1765 if (likely(order == 0)) {
1766 struct per_cpu_pages *pcp;
1767 struct list_head *list;
1769 local_lock_irqsave(pa_lock, flags);
1770 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1771 list = &pcp->lists[migratetype];
1772 if (list_empty(list)) {
1773 pcp->count += rmqueue_bulk(zone, 0,
1776 if (unlikely(list_empty(list)))
1781 page = list_entry(list->prev, struct page, lru);
1783 page = list_entry(list->next, struct page, lru);
1785 list_del(&page->lru);
1788 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1790 * __GFP_NOFAIL is not to be used in new code.
1792 * All __GFP_NOFAIL callers should be fixed so that they
1793 * properly detect and handle allocation failures.
1795 * We most definitely don't want callers attempting to
1796 * allocate greater than order-1 page units with
1799 WARN_ON_ONCE(order > 1);
1801 local_spin_lock_irqsave(pa_lock, &zone->lock, flags);
1802 page = __rmqueue(zone, order, migratetype);
1804 spin_unlock(&zone->lock);
1807 __mod_zone_freepage_state(zone, -(1 << order),
1808 get_freepage_migratetype(page));
1809 spin_unlock(&zone->lock);
1812 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1813 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1814 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1815 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1817 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1818 zone_statistics(preferred_zone, zone, gfp_flags);
1819 local_unlock_irqrestore(pa_lock, flags);
1821 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1825 local_unlock_irqrestore(pa_lock, flags);
1829 #ifdef CONFIG_FAIL_PAGE_ALLOC
1832 struct fault_attr attr;
1834 u32 ignore_gfp_highmem;
1835 u32 ignore_gfp_wait;
1837 } fail_page_alloc = {
1838 .attr = FAULT_ATTR_INITIALIZER,
1839 .ignore_gfp_wait = 1,
1840 .ignore_gfp_highmem = 1,
1844 static int __init setup_fail_page_alloc(char *str)
1846 return setup_fault_attr(&fail_page_alloc.attr, str);
1848 __setup("fail_page_alloc=", setup_fail_page_alloc);
1850 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1852 if (order < fail_page_alloc.min_order)
1854 if (gfp_mask & __GFP_NOFAIL)
1856 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1858 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1861 return should_fail(&fail_page_alloc.attr, 1 << order);
1864 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1866 static int __init fail_page_alloc_debugfs(void)
1868 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1871 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1872 &fail_page_alloc.attr);
1874 return PTR_ERR(dir);
1876 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1877 &fail_page_alloc.ignore_gfp_wait))
1879 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1880 &fail_page_alloc.ignore_gfp_highmem))
1882 if (!debugfs_create_u32("min-order", mode, dir,
1883 &fail_page_alloc.min_order))
1888 debugfs_remove_recursive(dir);
1893 late_initcall(fail_page_alloc_debugfs);
1895 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1897 #else /* CONFIG_FAIL_PAGE_ALLOC */
1899 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1904 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1907 * Return true if free pages are above 'mark'. This takes into account the order
1908 * of the allocation.
1910 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1911 unsigned long mark, int classzone_idx, int alloc_flags,
1914 /* free_pages may go negative - that's OK */
1919 free_pages -= (1 << order) - 1;
1920 if (alloc_flags & ALLOC_HIGH)
1922 if (alloc_flags & ALLOC_HARDER)
1925 /* If allocation can't use CMA areas don't use free CMA pages */
1926 if (!(alloc_flags & ALLOC_CMA))
1927 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1930 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1932 for (o = 0; o < order; o++) {
1933 /* At the next order, this order's pages become unavailable */
1934 free_pages -= z->free_area[o].nr_free << o;
1936 /* Require fewer higher order pages to be free */
1939 if (free_pages <= min)
1945 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1946 int classzone_idx, int alloc_flags)
1948 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1949 zone_page_state(z, NR_FREE_PAGES));
1952 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1953 unsigned long mark, int classzone_idx, int alloc_flags)
1955 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1957 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1958 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1960 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1966 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1967 * skip over zones that are not allowed by the cpuset, or that have
1968 * been recently (in last second) found to be nearly full. See further
1969 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1970 * that have to skip over a lot of full or unallowed zones.
1972 * If the zonelist cache is present in the passed zonelist, then
1973 * returns a pointer to the allowed node mask (either the current
1974 * tasks mems_allowed, or node_states[N_MEMORY].)
1976 * If the zonelist cache is not available for this zonelist, does
1977 * nothing and returns NULL.
1979 * If the fullzones BITMAP in the zonelist cache is stale (more than
1980 * a second since last zap'd) then we zap it out (clear its bits.)
1982 * We hold off even calling zlc_setup, until after we've checked the
1983 * first zone in the zonelist, on the theory that most allocations will
1984 * be satisfied from that first zone, so best to examine that zone as
1985 * quickly as we can.
1987 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1989 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1990 nodemask_t *allowednodes; /* zonelist_cache approximation */
1992 zlc = zonelist->zlcache_ptr;
1996 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1997 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1998 zlc->last_full_zap = jiffies;
2001 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
2002 &cpuset_current_mems_allowed :
2003 &node_states[N_MEMORY];
2004 return allowednodes;
2008 * Given 'z' scanning a zonelist, run a couple of quick checks to see
2009 * if it is worth looking at further for free memory:
2010 * 1) Check that the zone isn't thought to be full (doesn't have its
2011 * bit set in the zonelist_cache fullzones BITMAP).
2012 * 2) Check that the zones node (obtained from the zonelist_cache
2013 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2014 * Return true (non-zero) if zone is worth looking at further, or
2015 * else return false (zero) if it is not.
2017 * This check -ignores- the distinction between various watermarks,
2018 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2019 * found to be full for any variation of these watermarks, it will
2020 * be considered full for up to one second by all requests, unless
2021 * we are so low on memory on all allowed nodes that we are forced
2022 * into the second scan of the zonelist.
2024 * In the second scan we ignore this zonelist cache and exactly
2025 * apply the watermarks to all zones, even it is slower to do so.
2026 * We are low on memory in the second scan, and should leave no stone
2027 * unturned looking for a free page.
2029 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2030 nodemask_t *allowednodes)
2032 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2033 int i; /* index of *z in zonelist zones */
2034 int n; /* node that zone *z is on */
2036 zlc = zonelist->zlcache_ptr;
2040 i = z - zonelist->_zonerefs;
2043 /* This zone is worth trying if it is allowed but not full */
2044 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
2048 * Given 'z' scanning a zonelist, set the corresponding bit in
2049 * zlc->fullzones, so that subsequent attempts to allocate a page
2050 * from that zone don't waste time re-examining it.
2052 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2054 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2055 int i; /* index of *z in zonelist zones */
2057 zlc = zonelist->zlcache_ptr;
2061 i = z - zonelist->_zonerefs;
2063 set_bit(i, zlc->fullzones);
2067 * clear all zones full, called after direct reclaim makes progress so that
2068 * a zone that was recently full is not skipped over for up to a second
2070 static void zlc_clear_zones_full(struct zonelist *zonelist)
2072 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2074 zlc = zonelist->zlcache_ptr;
2078 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2081 static bool zone_local(struct zone *local_zone, struct zone *zone)
2083 return local_zone->node == zone->node;
2086 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2088 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2092 #else /* CONFIG_NUMA */
2094 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2099 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2100 nodemask_t *allowednodes)
2105 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2109 static void zlc_clear_zones_full(struct zonelist *zonelist)
2113 static bool zone_local(struct zone *local_zone, struct zone *zone)
2118 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2123 #endif /* CONFIG_NUMA */
2125 static void reset_alloc_batches(struct zone *preferred_zone)
2127 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2130 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2131 high_wmark_pages(zone) - low_wmark_pages(zone) -
2132 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2133 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2134 } while (zone++ != preferred_zone);
2138 * get_page_from_freelist goes through the zonelist trying to allocate
2141 static struct page *
2142 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2143 const struct alloc_context *ac)
2145 struct zonelist *zonelist = ac->zonelist;
2147 struct page *page = NULL;
2149 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2150 int zlc_active = 0; /* set if using zonelist_cache */
2151 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2152 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2153 (gfp_mask & __GFP_WRITE);
2154 int nr_fair_skipped = 0;
2155 bool zonelist_rescan;
2158 zonelist_rescan = false;
2161 * Scan zonelist, looking for a zone with enough free.
2162 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2164 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2168 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2169 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2171 if (cpusets_enabled() &&
2172 (alloc_flags & ALLOC_CPUSET) &&
2173 !cpuset_zone_allowed(zone, gfp_mask))
2176 * Distribute pages in proportion to the individual
2177 * zone size to ensure fair page aging. The zone a
2178 * page was allocated in should have no effect on the
2179 * time the page has in memory before being reclaimed.
2181 if (alloc_flags & ALLOC_FAIR) {
2182 if (!zone_local(ac->preferred_zone, zone))
2184 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2190 * When allocating a page cache page for writing, we
2191 * want to get it from a zone that is within its dirty
2192 * limit, such that no single zone holds more than its
2193 * proportional share of globally allowed dirty pages.
2194 * The dirty limits take into account the zone's
2195 * lowmem reserves and high watermark so that kswapd
2196 * should be able to balance it without having to
2197 * write pages from its LRU list.
2199 * This may look like it could increase pressure on
2200 * lower zones by failing allocations in higher zones
2201 * before they are full. But the pages that do spill
2202 * over are limited as the lower zones are protected
2203 * by this very same mechanism. It should not become
2204 * a practical burden to them.
2206 * XXX: For now, allow allocations to potentially
2207 * exceed the per-zone dirty limit in the slowpath
2208 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2209 * which is important when on a NUMA setup the allowed
2210 * zones are together not big enough to reach the
2211 * global limit. The proper fix for these situations
2212 * will require awareness of zones in the
2213 * dirty-throttling and the flusher threads.
2215 if (consider_zone_dirty && !zone_dirty_ok(zone))
2218 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2219 if (!zone_watermark_ok(zone, order, mark,
2220 ac->classzone_idx, alloc_flags)) {
2223 /* Checked here to keep the fast path fast */
2224 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2225 if (alloc_flags & ALLOC_NO_WATERMARKS)
2228 if (IS_ENABLED(CONFIG_NUMA) &&
2229 !did_zlc_setup && nr_online_nodes > 1) {
2231 * we do zlc_setup if there are multiple nodes
2232 * and before considering the first zone allowed
2235 allowednodes = zlc_setup(zonelist, alloc_flags);
2240 if (zone_reclaim_mode == 0 ||
2241 !zone_allows_reclaim(ac->preferred_zone, zone))
2242 goto this_zone_full;
2245 * As we may have just activated ZLC, check if the first
2246 * eligible zone has failed zone_reclaim recently.
2248 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2249 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2252 ret = zone_reclaim(zone, gfp_mask, order);
2254 case ZONE_RECLAIM_NOSCAN:
2257 case ZONE_RECLAIM_FULL:
2258 /* scanned but unreclaimable */
2261 /* did we reclaim enough */
2262 if (zone_watermark_ok(zone, order, mark,
2263 ac->classzone_idx, alloc_flags))
2267 * Failed to reclaim enough to meet watermark.
2268 * Only mark the zone full if checking the min
2269 * watermark or if we failed to reclaim just
2270 * 1<<order pages or else the page allocator
2271 * fastpath will prematurely mark zones full
2272 * when the watermark is between the low and
2275 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2276 ret == ZONE_RECLAIM_SOME)
2277 goto this_zone_full;
2284 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2285 gfp_mask, ac->migratetype);
2287 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2292 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2293 zlc_mark_zone_full(zonelist, z);
2297 * The first pass makes sure allocations are spread fairly within the
2298 * local node. However, the local node might have free pages left
2299 * after the fairness batches are exhausted, and remote zones haven't
2300 * even been considered yet. Try once more without fairness, and
2301 * include remote zones now, before entering the slowpath and waking
2302 * kswapd: prefer spilling to a remote zone over swapping locally.
2304 if (alloc_flags & ALLOC_FAIR) {
2305 alloc_flags &= ~ALLOC_FAIR;
2306 if (nr_fair_skipped) {
2307 zonelist_rescan = true;
2308 reset_alloc_batches(ac->preferred_zone);
2310 if (nr_online_nodes > 1)
2311 zonelist_rescan = true;
2314 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2315 /* Disable zlc cache for second zonelist scan */
2317 zonelist_rescan = true;
2320 if (zonelist_rescan)
2327 * Large machines with many possible nodes should not always dump per-node
2328 * meminfo in irq context.
2330 static inline bool should_suppress_show_mem(void)
2335 ret = in_interrupt();
2340 static DEFINE_RATELIMIT_STATE(nopage_rs,
2341 DEFAULT_RATELIMIT_INTERVAL,
2342 DEFAULT_RATELIMIT_BURST);
2344 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2346 unsigned int filter = SHOW_MEM_FILTER_NODES;
2348 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2349 debug_guardpage_minorder() > 0)
2353 * This documents exceptions given to allocations in certain
2354 * contexts that are allowed to allocate outside current's set
2357 if (!(gfp_mask & __GFP_NOMEMALLOC))
2358 if (test_thread_flag(TIF_MEMDIE) ||
2359 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2360 filter &= ~SHOW_MEM_FILTER_NODES;
2361 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2362 filter &= ~SHOW_MEM_FILTER_NODES;
2365 struct va_format vaf;
2368 va_start(args, fmt);
2373 pr_warn("%pV", &vaf);
2378 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2379 current->comm, order, gfp_mask);
2382 if (!should_suppress_show_mem())
2387 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2388 unsigned long did_some_progress,
2389 unsigned long pages_reclaimed)
2391 /* Do not loop if specifically requested */
2392 if (gfp_mask & __GFP_NORETRY)
2395 /* Always retry if specifically requested */
2396 if (gfp_mask & __GFP_NOFAIL)
2400 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2401 * making forward progress without invoking OOM. Suspend also disables
2402 * storage devices so kswapd will not help. Bail if we are suspending.
2404 if (!did_some_progress && pm_suspended_storage())
2408 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2409 * means __GFP_NOFAIL, but that may not be true in other
2412 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2416 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2417 * specified, then we retry until we no longer reclaim any pages
2418 * (above), or we've reclaimed an order of pages at least as
2419 * large as the allocation's order. In both cases, if the
2420 * allocation still fails, we stop retrying.
2422 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2428 static inline struct page *
2429 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2430 const struct alloc_context *ac, unsigned long *did_some_progress)
2434 *did_some_progress = 0;
2437 * Acquire the per-zone oom lock for each zone. If that
2438 * fails, somebody else is making progress for us.
2440 if (!oom_zonelist_trylock(ac->zonelist, gfp_mask)) {
2441 *did_some_progress = 1;
2442 schedule_timeout_uninterruptible(1);
2447 * Go through the zonelist yet one more time, keep very high watermark
2448 * here, this is only to catch a parallel oom killing, we must fail if
2449 * we're still under heavy pressure.
2451 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2452 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2456 if (!(gfp_mask & __GFP_NOFAIL)) {
2457 /* Coredumps can quickly deplete all memory reserves */
2458 if (current->flags & PF_DUMPCORE)
2460 /* The OOM killer will not help higher order allocs */
2461 if (order > PAGE_ALLOC_COSTLY_ORDER)
2463 /* The OOM killer does not needlessly kill tasks for lowmem */
2464 if (ac->high_zoneidx < ZONE_NORMAL)
2466 /* The OOM killer does not compensate for light reclaim */
2467 if (!(gfp_mask & __GFP_FS)) {
2469 * XXX: Page reclaim didn't yield anything,
2470 * and the OOM killer can't be invoked, but
2471 * keep looping as per should_alloc_retry().
2473 *did_some_progress = 1;
2476 /* The OOM killer may not free memory on a specific node */
2477 if (gfp_mask & __GFP_THISNODE)
2480 /* Exhausted what can be done so it's blamo time */
2481 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false)
2482 || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2483 *did_some_progress = 1;
2485 oom_zonelist_unlock(ac->zonelist, gfp_mask);
2489 #ifdef CONFIG_COMPACTION
2490 /* Try memory compaction for high-order allocations before reclaim */
2491 static struct page *
2492 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2493 int alloc_flags, const struct alloc_context *ac,
2494 enum migrate_mode mode, int *contended_compaction,
2495 bool *deferred_compaction)
2497 unsigned long compact_result;
2503 current->flags |= PF_MEMALLOC;
2504 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2505 mode, contended_compaction);
2506 current->flags &= ~PF_MEMALLOC;
2508 switch (compact_result) {
2509 case COMPACT_DEFERRED:
2510 *deferred_compaction = true;
2512 case COMPACT_SKIPPED:
2519 * At least in one zone compaction wasn't deferred or skipped, so let's
2520 * count a compaction stall
2522 count_vm_event(COMPACTSTALL);
2524 page = get_page_from_freelist(gfp_mask, order,
2525 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2528 struct zone *zone = page_zone(page);
2530 zone->compact_blockskip_flush = false;
2531 compaction_defer_reset(zone, order, true);
2532 count_vm_event(COMPACTSUCCESS);
2537 * It's bad if compaction run occurs and fails. The most likely reason
2538 * is that pages exist, but not enough to satisfy watermarks.
2540 count_vm_event(COMPACTFAIL);
2547 static inline struct page *
2548 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2549 int alloc_flags, const struct alloc_context *ac,
2550 enum migrate_mode mode, int *contended_compaction,
2551 bool *deferred_compaction)
2555 #endif /* CONFIG_COMPACTION */
2557 /* Perform direct synchronous page reclaim */
2559 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2560 const struct alloc_context *ac)
2562 struct reclaim_state reclaim_state;
2567 /* We now go into synchronous reclaim */
2568 cpuset_memory_pressure_bump();
2569 current->flags |= PF_MEMALLOC;
2570 lockdep_set_current_reclaim_state(gfp_mask);
2571 reclaim_state.reclaimed_slab = 0;
2572 current->reclaim_state = &reclaim_state;
2574 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2577 current->reclaim_state = NULL;
2578 lockdep_clear_current_reclaim_state();
2579 current->flags &= ~PF_MEMALLOC;
2586 /* The really slow allocator path where we enter direct reclaim */
2587 static inline struct page *
2588 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2589 int alloc_flags, const struct alloc_context *ac,
2590 unsigned long *did_some_progress)
2592 struct page *page = NULL;
2593 bool drained = false;
2595 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2596 if (unlikely(!(*did_some_progress)))
2599 /* After successful reclaim, reconsider all zones for allocation */
2600 if (IS_ENABLED(CONFIG_NUMA))
2601 zlc_clear_zones_full(ac->zonelist);
2604 page = get_page_from_freelist(gfp_mask, order,
2605 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2608 * If an allocation failed after direct reclaim, it could be because
2609 * pages are pinned on the per-cpu lists. Drain them and try again
2611 if (!page && !drained) {
2612 drain_all_pages(NULL);
2621 * This is called in the allocator slow-path if the allocation request is of
2622 * sufficient urgency to ignore watermarks and take other desperate measures
2624 static inline struct page *
2625 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2626 const struct alloc_context *ac)
2631 page = get_page_from_freelist(gfp_mask, order,
2632 ALLOC_NO_WATERMARKS, ac);
2634 if (!page && gfp_mask & __GFP_NOFAIL)
2635 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2637 } while (!page && (gfp_mask & __GFP_NOFAIL));
2642 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2647 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2648 ac->high_zoneidx, ac->nodemask)
2649 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2653 gfp_to_alloc_flags(gfp_t gfp_mask)
2655 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2656 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2658 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2659 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2662 * The caller may dip into page reserves a bit more if the caller
2663 * cannot run direct reclaim, or if the caller has realtime scheduling
2664 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2665 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2667 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2671 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2672 * if it can't schedule.
2674 if (!(gfp_mask & __GFP_NOMEMALLOC))
2675 alloc_flags |= ALLOC_HARDER;
2677 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2678 * comment for __cpuset_node_allowed().
2680 alloc_flags &= ~ALLOC_CPUSET;
2681 } else if (unlikely(rt_task(current)) && !in_interrupt())
2682 alloc_flags |= ALLOC_HARDER;
2684 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2685 if (gfp_mask & __GFP_MEMALLOC)
2686 alloc_flags |= ALLOC_NO_WATERMARKS;
2687 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2688 alloc_flags |= ALLOC_NO_WATERMARKS;
2689 else if (!in_interrupt() &&
2690 ((current->flags & PF_MEMALLOC) ||
2691 unlikely(test_thread_flag(TIF_MEMDIE))))
2692 alloc_flags |= ALLOC_NO_WATERMARKS;
2695 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2696 alloc_flags |= ALLOC_CMA;
2701 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2703 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2706 static inline struct page *
2707 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2708 struct alloc_context *ac)
2710 const gfp_t wait = gfp_mask & __GFP_WAIT;
2711 struct page *page = NULL;
2713 unsigned long pages_reclaimed = 0;
2714 unsigned long did_some_progress;
2715 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2716 bool deferred_compaction = false;
2717 int contended_compaction = COMPACT_CONTENDED_NONE;
2720 * In the slowpath, we sanity check order to avoid ever trying to
2721 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2722 * be using allocators in order of preference for an area that is
2725 if (order >= MAX_ORDER) {
2726 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2731 * If this allocation cannot block and it is for a specific node, then
2732 * fail early. There's no need to wakeup kswapd or retry for a
2733 * speculative node-specific allocation.
2735 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait)
2739 if (!(gfp_mask & __GFP_NO_KSWAPD))
2740 wake_all_kswapds(order, ac);
2743 * OK, we're below the kswapd watermark and have kicked background
2744 * reclaim. Now things get more complex, so set up alloc_flags according
2745 * to how we want to proceed.
2747 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2750 * Find the true preferred zone if the allocation is unconstrained by
2753 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
2754 struct zoneref *preferred_zoneref;
2755 preferred_zoneref = first_zones_zonelist(ac->zonelist,
2756 ac->high_zoneidx, NULL, &ac->preferred_zone);
2757 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
2760 /* This is the last chance, in general, before the goto nopage. */
2761 page = get_page_from_freelist(gfp_mask, order,
2762 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2766 /* Allocate without watermarks if the context allows */
2767 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2769 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2770 * the allocation is high priority and these type of
2771 * allocations are system rather than user orientated
2773 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
2775 page = __alloc_pages_high_priority(gfp_mask, order, ac);
2782 /* Atomic allocations - we can't balance anything */
2785 * All existing users of the deprecated __GFP_NOFAIL are
2786 * blockable, so warn of any new users that actually allow this
2787 * type of allocation to fail.
2789 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2793 /* Avoid recursion of direct reclaim */
2794 if (current->flags & PF_MEMALLOC)
2797 /* Avoid allocations with no watermarks from looping endlessly */
2798 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2802 * Try direct compaction. The first pass is asynchronous. Subsequent
2803 * attempts after direct reclaim are synchronous
2805 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
2807 &contended_compaction,
2808 &deferred_compaction);
2812 /* Checks for THP-specific high-order allocations */
2813 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2815 * If compaction is deferred for high-order allocations, it is
2816 * because sync compaction recently failed. If this is the case
2817 * and the caller requested a THP allocation, we do not want
2818 * to heavily disrupt the system, so we fail the allocation
2819 * instead of entering direct reclaim.
2821 if (deferred_compaction)
2825 * In all zones where compaction was attempted (and not
2826 * deferred or skipped), lock contention has been detected.
2827 * For THP allocation we do not want to disrupt the others
2828 * so we fallback to base pages instead.
2830 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2834 * If compaction was aborted due to need_resched(), we do not
2835 * want to further increase allocation latency, unless it is
2836 * khugepaged trying to collapse.
2838 if (contended_compaction == COMPACT_CONTENDED_SCHED
2839 && !(current->flags & PF_KTHREAD))
2844 * It can become very expensive to allocate transparent hugepages at
2845 * fault, so use asynchronous memory compaction for THP unless it is
2846 * khugepaged trying to collapse.
2848 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2849 (current->flags & PF_KTHREAD))
2850 migration_mode = MIGRATE_SYNC_LIGHT;
2852 /* Try direct reclaim and then allocating */
2853 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
2854 &did_some_progress);
2858 /* Check if we should retry the allocation */
2859 pages_reclaimed += did_some_progress;
2860 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2863 * If we fail to make progress by freeing individual
2864 * pages, but the allocation wants us to keep going,
2865 * start OOM killing tasks.
2867 if (!did_some_progress) {
2868 page = __alloc_pages_may_oom(gfp_mask, order, ac,
2869 &did_some_progress);
2872 if (!did_some_progress)
2875 /* Wait for some write requests to complete then retry */
2876 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
2880 * High-order allocations do not necessarily loop after
2881 * direct reclaim and reclaim/compaction depends on compaction
2882 * being called after reclaim so call directly if necessary
2884 page = __alloc_pages_direct_compact(gfp_mask, order,
2885 alloc_flags, ac, migration_mode,
2886 &contended_compaction,
2887 &deferred_compaction);
2893 warn_alloc_failed(gfp_mask, order, NULL);
2899 * This is the 'heart' of the zoned buddy allocator.
2902 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2903 struct zonelist *zonelist, nodemask_t *nodemask)
2905 struct zoneref *preferred_zoneref;
2906 struct page *page = NULL;
2907 unsigned int cpuset_mems_cookie;
2908 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2909 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
2910 struct alloc_context ac = {
2911 .high_zoneidx = gfp_zone(gfp_mask),
2912 .nodemask = nodemask,
2913 .migratetype = gfpflags_to_migratetype(gfp_mask),
2916 gfp_mask &= gfp_allowed_mask;
2918 lockdep_trace_alloc(gfp_mask);
2920 might_sleep_if(gfp_mask & __GFP_WAIT);
2922 if (should_fail_alloc_page(gfp_mask, order))
2926 * Check the zones suitable for the gfp_mask contain at least one
2927 * valid zone. It's possible to have an empty zonelist as a result
2928 * of __GFP_THISNODE and a memoryless node
2930 if (unlikely(!zonelist->_zonerefs->zone))
2933 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
2934 alloc_flags |= ALLOC_CMA;
2937 cpuset_mems_cookie = read_mems_allowed_begin();
2939 /* We set it here, as __alloc_pages_slowpath might have changed it */
2940 ac.zonelist = zonelist;
2941 /* The preferred zone is used for statistics later */
2942 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
2943 ac.nodemask ? : &cpuset_current_mems_allowed,
2944 &ac.preferred_zone);
2945 if (!ac.preferred_zone)
2947 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
2949 /* First allocation attempt */
2950 alloc_mask = gfp_mask|__GFP_HARDWALL;
2951 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
2952 if (unlikely(!page)) {
2954 * Runtime PM, block IO and its error handling path
2955 * can deadlock because I/O on the device might not
2958 alloc_mask = memalloc_noio_flags(gfp_mask);
2960 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
2963 if (kmemcheck_enabled && page)
2964 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2966 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
2970 * When updating a task's mems_allowed, it is possible to race with
2971 * parallel threads in such a way that an allocation can fail while
2972 * the mask is being updated. If a page allocation is about to fail,
2973 * check if the cpuset changed during allocation and if so, retry.
2975 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2980 EXPORT_SYMBOL(__alloc_pages_nodemask);
2983 * Common helper functions.
2985 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2990 * __get_free_pages() returns a 32-bit address, which cannot represent
2993 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2995 page = alloc_pages(gfp_mask, order);
2998 return (unsigned long) page_address(page);
3000 EXPORT_SYMBOL(__get_free_pages);
3002 unsigned long get_zeroed_page(gfp_t gfp_mask)
3004 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3006 EXPORT_SYMBOL(get_zeroed_page);
3008 void __free_pages(struct page *page, unsigned int order)
3010 if (put_page_testzero(page)) {
3012 free_hot_cold_page(page, false);
3014 __free_pages_ok(page, order);
3018 EXPORT_SYMBOL(__free_pages);
3020 void free_pages(unsigned long addr, unsigned int order)
3023 VM_BUG_ON(!virt_addr_valid((void *)addr));
3024 __free_pages(virt_to_page((void *)addr), order);
3028 EXPORT_SYMBOL(free_pages);
3031 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3032 * of the current memory cgroup.
3034 * It should be used when the caller would like to use kmalloc, but since the
3035 * allocation is large, it has to fall back to the page allocator.
3037 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3040 struct mem_cgroup *memcg = NULL;
3042 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3044 page = alloc_pages(gfp_mask, order);
3045 memcg_kmem_commit_charge(page, memcg, order);
3049 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3052 struct mem_cgroup *memcg = NULL;
3054 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3056 page = alloc_pages_node(nid, gfp_mask, order);
3057 memcg_kmem_commit_charge(page, memcg, order);
3062 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3065 void __free_kmem_pages(struct page *page, unsigned int order)
3067 memcg_kmem_uncharge_pages(page, order);
3068 __free_pages(page, order);
3071 void free_kmem_pages(unsigned long addr, unsigned int order)
3074 VM_BUG_ON(!virt_addr_valid((void *)addr));
3075 __free_kmem_pages(virt_to_page((void *)addr), order);
3079 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3082 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3083 unsigned long used = addr + PAGE_ALIGN(size);
3085 split_page(virt_to_page((void *)addr), order);
3086 while (used < alloc_end) {
3091 return (void *)addr;
3095 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3096 * @size: the number of bytes to allocate
3097 * @gfp_mask: GFP flags for the allocation
3099 * This function is similar to alloc_pages(), except that it allocates the
3100 * minimum number of pages to satisfy the request. alloc_pages() can only
3101 * allocate memory in power-of-two pages.
3103 * This function is also limited by MAX_ORDER.
3105 * Memory allocated by this function must be released by free_pages_exact().
3107 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3109 unsigned int order = get_order(size);
3112 addr = __get_free_pages(gfp_mask, order);
3113 return make_alloc_exact(addr, order, size);
3115 EXPORT_SYMBOL(alloc_pages_exact);
3118 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3120 * @nid: the preferred node ID where memory should be allocated
3121 * @size: the number of bytes to allocate
3122 * @gfp_mask: GFP flags for the allocation
3124 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3126 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3129 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3131 unsigned order = get_order(size);
3132 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3135 return make_alloc_exact((unsigned long)page_address(p), order, size);
3139 * free_pages_exact - release memory allocated via alloc_pages_exact()
3140 * @virt: the value returned by alloc_pages_exact.
3141 * @size: size of allocation, same value as passed to alloc_pages_exact().
3143 * Release the memory allocated by a previous call to alloc_pages_exact.
3145 void free_pages_exact(void *virt, size_t size)
3147 unsigned long addr = (unsigned long)virt;
3148 unsigned long end = addr + PAGE_ALIGN(size);
3150 while (addr < end) {
3155 EXPORT_SYMBOL(free_pages_exact);
3158 * nr_free_zone_pages - count number of pages beyond high watermark
3159 * @offset: The zone index of the highest zone
3161 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3162 * high watermark within all zones at or below a given zone index. For each
3163 * zone, the number of pages is calculated as:
3164 * managed_pages - high_pages
3166 static unsigned long nr_free_zone_pages(int offset)
3171 /* Just pick one node, since fallback list is circular */
3172 unsigned long sum = 0;
3174 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3176 for_each_zone_zonelist(zone, z, zonelist, offset) {
3177 unsigned long size = zone->managed_pages;
3178 unsigned long high = high_wmark_pages(zone);
3187 * nr_free_buffer_pages - count number of pages beyond high watermark
3189 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3190 * watermark within ZONE_DMA and ZONE_NORMAL.
3192 unsigned long nr_free_buffer_pages(void)
3194 return nr_free_zone_pages(gfp_zone(GFP_USER));
3196 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3199 * nr_free_pagecache_pages - count number of pages beyond high watermark
3201 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3202 * high watermark within all zones.
3204 unsigned long nr_free_pagecache_pages(void)
3206 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3209 static inline void show_node(struct zone *zone)
3211 if (IS_ENABLED(CONFIG_NUMA))
3212 printk("Node %d ", zone_to_nid(zone));
3215 void si_meminfo(struct sysinfo *val)
3217 val->totalram = totalram_pages;
3218 val->sharedram = global_page_state(NR_SHMEM);
3219 val->freeram = global_page_state(NR_FREE_PAGES);
3220 val->bufferram = nr_blockdev_pages();
3221 val->totalhigh = totalhigh_pages;
3222 val->freehigh = nr_free_highpages();
3223 val->mem_unit = PAGE_SIZE;
3226 EXPORT_SYMBOL(si_meminfo);
3229 void si_meminfo_node(struct sysinfo *val, int nid)
3231 int zone_type; /* needs to be signed */
3232 unsigned long managed_pages = 0;
3233 pg_data_t *pgdat = NODE_DATA(nid);
3235 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3236 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3237 val->totalram = managed_pages;
3238 val->sharedram = node_page_state(nid, NR_SHMEM);
3239 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3240 #ifdef CONFIG_HIGHMEM
3241 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3242 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3248 val->mem_unit = PAGE_SIZE;
3253 * Determine whether the node should be displayed or not, depending on whether
3254 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3256 bool skip_free_areas_node(unsigned int flags, int nid)
3259 unsigned int cpuset_mems_cookie;
3261 if (!(flags & SHOW_MEM_FILTER_NODES))
3265 cpuset_mems_cookie = read_mems_allowed_begin();
3266 ret = !node_isset(nid, cpuset_current_mems_allowed);
3267 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3272 #define K(x) ((x) << (PAGE_SHIFT-10))
3274 static void show_migration_types(unsigned char type)
3276 static const char types[MIGRATE_TYPES] = {
3277 [MIGRATE_UNMOVABLE] = 'U',
3278 [MIGRATE_RECLAIMABLE] = 'E',
3279 [MIGRATE_MOVABLE] = 'M',
3280 [MIGRATE_RESERVE] = 'R',
3282 [MIGRATE_CMA] = 'C',
3284 #ifdef CONFIG_MEMORY_ISOLATION
3285 [MIGRATE_ISOLATE] = 'I',
3288 char tmp[MIGRATE_TYPES + 1];
3292 for (i = 0; i < MIGRATE_TYPES; i++) {
3293 if (type & (1 << i))
3298 printk("(%s) ", tmp);
3302 * Show free area list (used inside shift_scroll-lock stuff)
3303 * We also calculate the percentage fragmentation. We do this by counting the
3304 * memory on each free list with the exception of the first item on the list.
3307 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3310 void show_free_areas(unsigned int filter)
3312 unsigned long free_pcp = 0;
3316 for_each_populated_zone(zone) {
3317 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3320 for_each_online_cpu(cpu)
3321 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3324 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3325 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3326 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3327 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3328 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3329 " free:%lu free_pcp:%lu free_cma:%lu\n",
3330 global_page_state(NR_ACTIVE_ANON),
3331 global_page_state(NR_INACTIVE_ANON),
3332 global_page_state(NR_ISOLATED_ANON),
3333 global_page_state(NR_ACTIVE_FILE),
3334 global_page_state(NR_INACTIVE_FILE),
3335 global_page_state(NR_ISOLATED_FILE),
3336 global_page_state(NR_UNEVICTABLE),
3337 global_page_state(NR_FILE_DIRTY),
3338 global_page_state(NR_WRITEBACK),
3339 global_page_state(NR_UNSTABLE_NFS),
3340 global_page_state(NR_SLAB_RECLAIMABLE),
3341 global_page_state(NR_SLAB_UNRECLAIMABLE),
3342 global_page_state(NR_FILE_MAPPED),
3343 global_page_state(NR_SHMEM),
3344 global_page_state(NR_PAGETABLE),
3345 global_page_state(NR_BOUNCE),
3346 global_page_state(NR_FREE_PAGES),
3348 global_page_state(NR_FREE_CMA_PAGES));
3350 for_each_populated_zone(zone) {
3353 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3357 for_each_online_cpu(cpu)
3358 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3366 " active_anon:%lukB"
3367 " inactive_anon:%lukB"
3368 " active_file:%lukB"
3369 " inactive_file:%lukB"
3370 " unevictable:%lukB"
3371 " isolated(anon):%lukB"
3372 " isolated(file):%lukB"
3380 " slab_reclaimable:%lukB"
3381 " slab_unreclaimable:%lukB"
3382 " kernel_stack:%lukB"
3389 " writeback_tmp:%lukB"
3390 " pages_scanned:%lu"
3391 " all_unreclaimable? %s"
3394 K(zone_page_state(zone, NR_FREE_PAGES)),
3395 K(min_wmark_pages(zone)),
3396 K(low_wmark_pages(zone)),
3397 K(high_wmark_pages(zone)),
3398 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3399 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3400 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3401 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3402 K(zone_page_state(zone, NR_UNEVICTABLE)),
3403 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3404 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3405 K(zone->present_pages),
3406 K(zone->managed_pages),
3407 K(zone_page_state(zone, NR_MLOCK)),
3408 K(zone_page_state(zone, NR_FILE_DIRTY)),
3409 K(zone_page_state(zone, NR_WRITEBACK)),
3410 K(zone_page_state(zone, NR_FILE_MAPPED)),
3411 K(zone_page_state(zone, NR_SHMEM)),
3412 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3413 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3414 zone_page_state(zone, NR_KERNEL_STACK) *
3416 K(zone_page_state(zone, NR_PAGETABLE)),
3417 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3418 K(zone_page_state(zone, NR_BOUNCE)),
3420 K(this_cpu_read(zone->pageset->pcp.count)),
3421 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3422 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3423 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3424 (!zone_reclaimable(zone) ? "yes" : "no")
3426 printk("lowmem_reserve[]:");
3427 for (i = 0; i < MAX_NR_ZONES; i++)
3428 printk(" %ld", zone->lowmem_reserve[i]);
3432 for_each_populated_zone(zone) {
3433 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3434 unsigned char types[MAX_ORDER];
3436 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3439 printk("%s: ", zone->name);
3441 spin_lock_irqsave(&zone->lock, flags);
3442 for (order = 0; order < MAX_ORDER; order++) {
3443 struct free_area *area = &zone->free_area[order];
3446 nr[order] = area->nr_free;
3447 total += nr[order] << order;
3450 for (type = 0; type < MIGRATE_TYPES; type++) {
3451 if (!list_empty(&area->free_list[type]))
3452 types[order] |= 1 << type;
3455 spin_unlock_irqrestore(&zone->lock, flags);
3456 for (order = 0; order < MAX_ORDER; order++) {
3457 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3459 show_migration_types(types[order]);
3461 printk("= %lukB\n", K(total));
3464 hugetlb_show_meminfo();
3466 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3468 show_swap_cache_info();
3471 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3473 zoneref->zone = zone;
3474 zoneref->zone_idx = zone_idx(zone);
3478 * Builds allocation fallback zone lists.
3480 * Add all populated zones of a node to the zonelist.
3482 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3486 enum zone_type zone_type = MAX_NR_ZONES;
3490 zone = pgdat->node_zones + zone_type;
3491 if (populated_zone(zone)) {
3492 zoneref_set_zone(zone,
3493 &zonelist->_zonerefs[nr_zones++]);
3494 check_highest_zone(zone_type);
3496 } while (zone_type);
3504 * 0 = automatic detection of better ordering.
3505 * 1 = order by ([node] distance, -zonetype)
3506 * 2 = order by (-zonetype, [node] distance)
3508 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3509 * the same zonelist. So only NUMA can configure this param.
3511 #define ZONELIST_ORDER_DEFAULT 0
3512 #define ZONELIST_ORDER_NODE 1
3513 #define ZONELIST_ORDER_ZONE 2
3515 /* zonelist order in the kernel.
3516 * set_zonelist_order() will set this to NODE or ZONE.
3518 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3519 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3523 /* The value user specified ....changed by config */
3524 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3525 /* string for sysctl */
3526 #define NUMA_ZONELIST_ORDER_LEN 16
3527 char numa_zonelist_order[16] = "default";
3530 * interface for configure zonelist ordering.
3531 * command line option "numa_zonelist_order"
3532 * = "[dD]efault - default, automatic configuration.
3533 * = "[nN]ode - order by node locality, then by zone within node
3534 * = "[zZ]one - order by zone, then by locality within zone
3537 static int __parse_numa_zonelist_order(char *s)
3539 if (*s == 'd' || *s == 'D') {
3540 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3541 } else if (*s == 'n' || *s == 'N') {
3542 user_zonelist_order = ZONELIST_ORDER_NODE;
3543 } else if (*s == 'z' || *s == 'Z') {
3544 user_zonelist_order = ZONELIST_ORDER_ZONE;
3547 "Ignoring invalid numa_zonelist_order value: "
3554 static __init int setup_numa_zonelist_order(char *s)
3561 ret = __parse_numa_zonelist_order(s);
3563 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3567 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3570 * sysctl handler for numa_zonelist_order
3572 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3573 void __user *buffer, size_t *length,
3576 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3578 static DEFINE_MUTEX(zl_order_mutex);
3580 mutex_lock(&zl_order_mutex);
3582 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3586 strcpy(saved_string, (char *)table->data);
3588 ret = proc_dostring(table, write, buffer, length, ppos);
3592 int oldval = user_zonelist_order;
3594 ret = __parse_numa_zonelist_order((char *)table->data);
3597 * bogus value. restore saved string
3599 strncpy((char *)table->data, saved_string,
3600 NUMA_ZONELIST_ORDER_LEN);
3601 user_zonelist_order = oldval;
3602 } else if (oldval != user_zonelist_order) {
3603 mutex_lock(&zonelists_mutex);
3604 build_all_zonelists(NULL, NULL);
3605 mutex_unlock(&zonelists_mutex);
3609 mutex_unlock(&zl_order_mutex);
3614 #define MAX_NODE_LOAD (nr_online_nodes)
3615 static int node_load[MAX_NUMNODES];
3618 * find_next_best_node - find the next node that should appear in a given node's fallback list
3619 * @node: node whose fallback list we're appending
3620 * @used_node_mask: nodemask_t of already used nodes
3622 * We use a number of factors to determine which is the next node that should
3623 * appear on a given node's fallback list. The node should not have appeared
3624 * already in @node's fallback list, and it should be the next closest node
3625 * according to the distance array (which contains arbitrary distance values
3626 * from each node to each node in the system), and should also prefer nodes
3627 * with no CPUs, since presumably they'll have very little allocation pressure
3628 * on them otherwise.
3629 * It returns -1 if no node is found.
3631 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3634 int min_val = INT_MAX;
3635 int best_node = NUMA_NO_NODE;
3636 const struct cpumask *tmp = cpumask_of_node(0);
3638 /* Use the local node if we haven't already */
3639 if (!node_isset(node, *used_node_mask)) {
3640 node_set(node, *used_node_mask);
3644 for_each_node_state(n, N_MEMORY) {
3646 /* Don't want a node to appear more than once */
3647 if (node_isset(n, *used_node_mask))
3650 /* Use the distance array to find the distance */
3651 val = node_distance(node, n);
3653 /* Penalize nodes under us ("prefer the next node") */
3656 /* Give preference to headless and unused nodes */
3657 tmp = cpumask_of_node(n);
3658 if (!cpumask_empty(tmp))
3659 val += PENALTY_FOR_NODE_WITH_CPUS;
3661 /* Slight preference for less loaded node */
3662 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3663 val += node_load[n];
3665 if (val < min_val) {
3672 node_set(best_node, *used_node_mask);
3679 * Build zonelists ordered by node and zones within node.
3680 * This results in maximum locality--normal zone overflows into local
3681 * DMA zone, if any--but risks exhausting DMA zone.
3683 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3686 struct zonelist *zonelist;
3688 zonelist = &pgdat->node_zonelists[0];
3689 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3691 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3692 zonelist->_zonerefs[j].zone = NULL;
3693 zonelist->_zonerefs[j].zone_idx = 0;
3697 * Build gfp_thisnode zonelists
3699 static void build_thisnode_zonelists(pg_data_t *pgdat)
3702 struct zonelist *zonelist;
3704 zonelist = &pgdat->node_zonelists[1];
3705 j = build_zonelists_node(pgdat, zonelist, 0);
3706 zonelist->_zonerefs[j].zone = NULL;
3707 zonelist->_zonerefs[j].zone_idx = 0;
3711 * Build zonelists ordered by zone and nodes within zones.
3712 * This results in conserving DMA zone[s] until all Normal memory is
3713 * exhausted, but results in overflowing to remote node while memory
3714 * may still exist in local DMA zone.
3716 static int node_order[MAX_NUMNODES];
3718 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3721 int zone_type; /* needs to be signed */
3723 struct zonelist *zonelist;
3725 zonelist = &pgdat->node_zonelists[0];
3727 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3728 for (j = 0; j < nr_nodes; j++) {
3729 node = node_order[j];
3730 z = &NODE_DATA(node)->node_zones[zone_type];
3731 if (populated_zone(z)) {
3733 &zonelist->_zonerefs[pos++]);
3734 check_highest_zone(zone_type);
3738 zonelist->_zonerefs[pos].zone = NULL;
3739 zonelist->_zonerefs[pos].zone_idx = 0;
3742 #if defined(CONFIG_64BIT)
3744 * Devices that require DMA32/DMA are relatively rare and do not justify a
3745 * penalty to every machine in case the specialised case applies. Default
3746 * to Node-ordering on 64-bit NUMA machines
3748 static int default_zonelist_order(void)
3750 return ZONELIST_ORDER_NODE;
3754 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3755 * by the kernel. If processes running on node 0 deplete the low memory zone
3756 * then reclaim will occur more frequency increasing stalls and potentially
3757 * be easier to OOM if a large percentage of the zone is under writeback or
3758 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3759 * Hence, default to zone ordering on 32-bit.
3761 static int default_zonelist_order(void)
3763 return ZONELIST_ORDER_ZONE;
3765 #endif /* CONFIG_64BIT */
3767 static void set_zonelist_order(void)
3769 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3770 current_zonelist_order = default_zonelist_order();
3772 current_zonelist_order = user_zonelist_order;
3775 static void build_zonelists(pg_data_t *pgdat)
3779 nodemask_t used_mask;
3780 int local_node, prev_node;
3781 struct zonelist *zonelist;
3782 int order = current_zonelist_order;
3784 /* initialize zonelists */
3785 for (i = 0; i < MAX_ZONELISTS; i++) {
3786 zonelist = pgdat->node_zonelists + i;
3787 zonelist->_zonerefs[0].zone = NULL;
3788 zonelist->_zonerefs[0].zone_idx = 0;
3791 /* NUMA-aware ordering of nodes */
3792 local_node = pgdat->node_id;
3793 load = nr_online_nodes;
3794 prev_node = local_node;
3795 nodes_clear(used_mask);
3797 memset(node_order, 0, sizeof(node_order));
3800 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3802 * We don't want to pressure a particular node.
3803 * So adding penalty to the first node in same
3804 * distance group to make it round-robin.
3806 if (node_distance(local_node, node) !=
3807 node_distance(local_node, prev_node))
3808 node_load[node] = load;
3812 if (order == ZONELIST_ORDER_NODE)
3813 build_zonelists_in_node_order(pgdat, node);
3815 node_order[j++] = node; /* remember order */
3818 if (order == ZONELIST_ORDER_ZONE) {
3819 /* calculate node order -- i.e., DMA last! */
3820 build_zonelists_in_zone_order(pgdat, j);
3823 build_thisnode_zonelists(pgdat);
3826 /* Construct the zonelist performance cache - see further mmzone.h */
3827 static void build_zonelist_cache(pg_data_t *pgdat)
3829 struct zonelist *zonelist;
3830 struct zonelist_cache *zlc;
3833 zonelist = &pgdat->node_zonelists[0];
3834 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3835 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3836 for (z = zonelist->_zonerefs; z->zone; z++)
3837 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3840 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3842 * Return node id of node used for "local" allocations.
3843 * I.e., first node id of first zone in arg node's generic zonelist.
3844 * Used for initializing percpu 'numa_mem', which is used primarily
3845 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3847 int local_memory_node(int node)
3851 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3852 gfp_zone(GFP_KERNEL),
3859 #else /* CONFIG_NUMA */
3861 static void set_zonelist_order(void)
3863 current_zonelist_order = ZONELIST_ORDER_ZONE;
3866 static void build_zonelists(pg_data_t *pgdat)
3868 int node, local_node;
3870 struct zonelist *zonelist;
3872 local_node = pgdat->node_id;
3874 zonelist = &pgdat->node_zonelists[0];
3875 j = build_zonelists_node(pgdat, zonelist, 0);
3878 * Now we build the zonelist so that it contains the zones
3879 * of all the other nodes.
3880 * We don't want to pressure a particular node, so when
3881 * building the zones for node N, we make sure that the
3882 * zones coming right after the local ones are those from
3883 * node N+1 (modulo N)
3885 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3886 if (!node_online(node))
3888 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3890 for (node = 0; node < local_node; node++) {
3891 if (!node_online(node))
3893 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3896 zonelist->_zonerefs[j].zone = NULL;
3897 zonelist->_zonerefs[j].zone_idx = 0;
3900 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3901 static void build_zonelist_cache(pg_data_t *pgdat)
3903 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3906 #endif /* CONFIG_NUMA */
3909 * Boot pageset table. One per cpu which is going to be used for all
3910 * zones and all nodes. The parameters will be set in such a way
3911 * that an item put on a list will immediately be handed over to
3912 * the buddy list. This is safe since pageset manipulation is done
3913 * with interrupts disabled.
3915 * The boot_pagesets must be kept even after bootup is complete for
3916 * unused processors and/or zones. They do play a role for bootstrapping
3917 * hotplugged processors.
3919 * zoneinfo_show() and maybe other functions do
3920 * not check if the processor is online before following the pageset pointer.
3921 * Other parts of the kernel may not check if the zone is available.
3923 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3924 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3925 static void setup_zone_pageset(struct zone *zone);
3928 * Global mutex to protect against size modification of zonelists
3929 * as well as to serialize pageset setup for the new populated zone.
3931 DEFINE_MUTEX(zonelists_mutex);
3933 /* return values int ....just for stop_machine() */
3934 static int __build_all_zonelists(void *data)
3938 pg_data_t *self = data;
3941 memset(node_load, 0, sizeof(node_load));
3944 if (self && !node_online(self->node_id)) {
3945 build_zonelists(self);
3946 build_zonelist_cache(self);
3949 for_each_online_node(nid) {
3950 pg_data_t *pgdat = NODE_DATA(nid);
3952 build_zonelists(pgdat);
3953 build_zonelist_cache(pgdat);
3957 * Initialize the boot_pagesets that are going to be used
3958 * for bootstrapping processors. The real pagesets for
3959 * each zone will be allocated later when the per cpu
3960 * allocator is available.
3962 * boot_pagesets are used also for bootstrapping offline
3963 * cpus if the system is already booted because the pagesets
3964 * are needed to initialize allocators on a specific cpu too.
3965 * F.e. the percpu allocator needs the page allocator which
3966 * needs the percpu allocator in order to allocate its pagesets
3967 * (a chicken-egg dilemma).
3969 for_each_possible_cpu(cpu) {
3970 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3972 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3974 * We now know the "local memory node" for each node--
3975 * i.e., the node of the first zone in the generic zonelist.
3976 * Set up numa_mem percpu variable for on-line cpus. During
3977 * boot, only the boot cpu should be on-line; we'll init the
3978 * secondary cpus' numa_mem as they come on-line. During
3979 * node/memory hotplug, we'll fixup all on-line cpus.
3981 if (cpu_online(cpu))
3982 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3989 static noinline void __init
3990 build_all_zonelists_init(void)
3992 __build_all_zonelists(NULL);
3993 mminit_verify_zonelist();
3994 cpuset_init_current_mems_allowed();
3998 * Called with zonelists_mutex held always
3999 * unless system_state == SYSTEM_BOOTING.
4001 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4002 * [we're only called with non-NULL zone through __meminit paths] and
4003 * (2) call of __init annotated helper build_all_zonelists_init
4004 * [protected by SYSTEM_BOOTING].
4006 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4008 set_zonelist_order();
4010 if (system_state == SYSTEM_BOOTING) {
4011 build_all_zonelists_init();
4013 #ifdef CONFIG_MEMORY_HOTPLUG
4015 setup_zone_pageset(zone);
4017 /* we have to stop all cpus to guarantee there is no user
4019 stop_machine(__build_all_zonelists, pgdat, NULL);
4020 /* cpuset refresh routine should be here */
4022 vm_total_pages = nr_free_pagecache_pages();
4024 * Disable grouping by mobility if the number of pages in the
4025 * system is too low to allow the mechanism to work. It would be
4026 * more accurate, but expensive to check per-zone. This check is
4027 * made on memory-hotadd so a system can start with mobility
4028 * disabled and enable it later
4030 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4031 page_group_by_mobility_disabled = 1;
4033 page_group_by_mobility_disabled = 0;
4035 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4036 "Total pages: %ld\n",
4038 zonelist_order_name[current_zonelist_order],
4039 page_group_by_mobility_disabled ? "off" : "on",
4042 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4047 * Helper functions to size the waitqueue hash table.
4048 * Essentially these want to choose hash table sizes sufficiently
4049 * large so that collisions trying to wait on pages are rare.
4050 * But in fact, the number of active page waitqueues on typical
4051 * systems is ridiculously low, less than 200. So this is even
4052 * conservative, even though it seems large.
4054 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4055 * waitqueues, i.e. the size of the waitq table given the number of pages.
4057 #define PAGES_PER_WAITQUEUE 256
4059 #ifndef CONFIG_MEMORY_HOTPLUG
4060 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4062 unsigned long size = 1;
4064 pages /= PAGES_PER_WAITQUEUE;
4066 while (size < pages)
4070 * Once we have dozens or even hundreds of threads sleeping
4071 * on IO we've got bigger problems than wait queue collision.
4072 * Limit the size of the wait table to a reasonable size.
4074 size = min(size, 4096UL);
4076 return max(size, 4UL);
4080 * A zone's size might be changed by hot-add, so it is not possible to determine
4081 * a suitable size for its wait_table. So we use the maximum size now.
4083 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4085 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4086 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4087 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4089 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4090 * or more by the traditional way. (See above). It equals:
4092 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4093 * ia64(16K page size) : = ( 8G + 4M)byte.
4094 * powerpc (64K page size) : = (32G +16M)byte.
4096 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4103 * This is an integer logarithm so that shifts can be used later
4104 * to extract the more random high bits from the multiplicative
4105 * hash function before the remainder is taken.
4107 static inline unsigned long wait_table_bits(unsigned long size)
4113 * Check if a pageblock contains reserved pages
4115 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4119 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4120 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4127 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4128 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4129 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4130 * higher will lead to a bigger reserve which will get freed as contiguous
4131 * blocks as reclaim kicks in
4133 static void setup_zone_migrate_reserve(struct zone *zone)
4135 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4137 unsigned long block_migratetype;
4142 * Get the start pfn, end pfn and the number of blocks to reserve
4143 * We have to be careful to be aligned to pageblock_nr_pages to
4144 * make sure that we always check pfn_valid for the first page in
4147 start_pfn = zone->zone_start_pfn;
4148 end_pfn = zone_end_pfn(zone);
4149 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4150 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4154 * Reserve blocks are generally in place to help high-order atomic
4155 * allocations that are short-lived. A min_free_kbytes value that
4156 * would result in more than 2 reserve blocks for atomic allocations
4157 * is assumed to be in place to help anti-fragmentation for the
4158 * future allocation of hugepages at runtime.
4160 reserve = min(2, reserve);
4161 old_reserve = zone->nr_migrate_reserve_block;
4163 /* When memory hot-add, we almost always need to do nothing */
4164 if (reserve == old_reserve)
4166 zone->nr_migrate_reserve_block = reserve;
4168 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4169 if (!pfn_valid(pfn))
4171 page = pfn_to_page(pfn);
4173 /* Watch out for overlapping nodes */
4174 if (page_to_nid(page) != zone_to_nid(zone))
4177 block_migratetype = get_pageblock_migratetype(page);
4179 /* Only test what is necessary when the reserves are not met */
4182 * Blocks with reserved pages will never free, skip
4185 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4186 if (pageblock_is_reserved(pfn, block_end_pfn))
4189 /* If this block is reserved, account for it */
4190 if (block_migratetype == MIGRATE_RESERVE) {
4195 /* Suitable for reserving if this block is movable */
4196 if (block_migratetype == MIGRATE_MOVABLE) {
4197 set_pageblock_migratetype(page,
4199 move_freepages_block(zone, page,
4204 } else if (!old_reserve) {
4206 * At boot time we don't need to scan the whole zone
4207 * for turning off MIGRATE_RESERVE.
4213 * If the reserve is met and this is a previous reserved block,
4216 if (block_migratetype == MIGRATE_RESERVE) {
4217 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4218 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4224 * Initially all pages are reserved - free ones are freed
4225 * up by free_all_bootmem() once the early boot process is
4226 * done. Non-atomic initialization, single-pass.
4228 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4229 unsigned long start_pfn, enum memmap_context context)
4232 unsigned long end_pfn = start_pfn + size;
4236 if (highest_memmap_pfn < end_pfn - 1)
4237 highest_memmap_pfn = end_pfn - 1;
4239 z = &NODE_DATA(nid)->node_zones[zone];
4240 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4242 * There can be holes in boot-time mem_map[]s
4243 * handed to this function. They do not
4244 * exist on hotplugged memory.
4246 if (context == MEMMAP_EARLY) {
4247 if (!early_pfn_valid(pfn))
4249 if (!early_pfn_in_nid(pfn, nid))
4252 page = pfn_to_page(pfn);
4253 set_page_links(page, zone, nid, pfn);
4254 mminit_verify_page_links(page, zone, nid, pfn);
4255 init_page_count(page);
4256 page_mapcount_reset(page);
4257 page_cpupid_reset_last(page);
4258 SetPageReserved(page);
4260 * Mark the block movable so that blocks are reserved for
4261 * movable at startup. This will force kernel allocations
4262 * to reserve their blocks rather than leaking throughout
4263 * the address space during boot when many long-lived
4264 * kernel allocations are made. Later some blocks near
4265 * the start are marked MIGRATE_RESERVE by
4266 * setup_zone_migrate_reserve()
4268 * bitmap is created for zone's valid pfn range. but memmap
4269 * can be created for invalid pages (for alignment)
4270 * check here not to call set_pageblock_migratetype() against
4273 if ((z->zone_start_pfn <= pfn)
4274 && (pfn < zone_end_pfn(z))
4275 && !(pfn & (pageblock_nr_pages - 1)))
4276 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4278 INIT_LIST_HEAD(&page->lru);
4279 #ifdef WANT_PAGE_VIRTUAL
4280 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4281 if (!is_highmem_idx(zone))
4282 set_page_address(page, __va(pfn << PAGE_SHIFT));
4287 static void __meminit zone_init_free_lists(struct zone *zone)
4289 unsigned int order, t;
4290 for_each_migratetype_order(order, t) {
4291 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4292 zone->free_area[order].nr_free = 0;
4296 #ifndef __HAVE_ARCH_MEMMAP_INIT
4297 #define memmap_init(size, nid, zone, start_pfn) \
4298 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4301 static int zone_batchsize(struct zone *zone)
4307 * The per-cpu-pages pools are set to around 1000th of the
4308 * size of the zone. But no more than 1/2 of a meg.
4310 * OK, so we don't know how big the cache is. So guess.
4312 batch = zone->managed_pages / 1024;
4313 if (batch * PAGE_SIZE > 512 * 1024)
4314 batch = (512 * 1024) / PAGE_SIZE;
4315 batch /= 4; /* We effectively *= 4 below */
4320 * Clamp the batch to a 2^n - 1 value. Having a power
4321 * of 2 value was found to be more likely to have
4322 * suboptimal cache aliasing properties in some cases.
4324 * For example if 2 tasks are alternately allocating
4325 * batches of pages, one task can end up with a lot
4326 * of pages of one half of the possible page colors
4327 * and the other with pages of the other colors.
4329 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4334 /* The deferral and batching of frees should be suppressed under NOMMU
4337 * The problem is that NOMMU needs to be able to allocate large chunks
4338 * of contiguous memory as there's no hardware page translation to
4339 * assemble apparent contiguous memory from discontiguous pages.
4341 * Queueing large contiguous runs of pages for batching, however,
4342 * causes the pages to actually be freed in smaller chunks. As there
4343 * can be a significant delay between the individual batches being
4344 * recycled, this leads to the once large chunks of space being
4345 * fragmented and becoming unavailable for high-order allocations.
4352 * pcp->high and pcp->batch values are related and dependent on one another:
4353 * ->batch must never be higher then ->high.
4354 * The following function updates them in a safe manner without read side
4357 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4358 * those fields changing asynchronously (acording the the above rule).
4360 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4361 * outside of boot time (or some other assurance that no concurrent updaters
4364 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4365 unsigned long batch)
4367 /* start with a fail safe value for batch */
4371 /* Update high, then batch, in order */
4378 /* a companion to pageset_set_high() */
4379 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4381 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4384 static void pageset_init(struct per_cpu_pageset *p)
4386 struct per_cpu_pages *pcp;
4389 memset(p, 0, sizeof(*p));
4393 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4394 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4397 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4400 pageset_set_batch(p, batch);
4404 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4405 * to the value high for the pageset p.
4407 static void pageset_set_high(struct per_cpu_pageset *p,
4410 unsigned long batch = max(1UL, high / 4);
4411 if ((high / 4) > (PAGE_SHIFT * 8))
4412 batch = PAGE_SHIFT * 8;
4414 pageset_update(&p->pcp, high, batch);
4417 static void pageset_set_high_and_batch(struct zone *zone,
4418 struct per_cpu_pageset *pcp)
4420 if (percpu_pagelist_fraction)
4421 pageset_set_high(pcp,
4422 (zone->managed_pages /
4423 percpu_pagelist_fraction));
4425 pageset_set_batch(pcp, zone_batchsize(zone));
4428 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4430 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4433 pageset_set_high_and_batch(zone, pcp);
4436 static void __meminit setup_zone_pageset(struct zone *zone)
4439 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4440 for_each_possible_cpu(cpu)
4441 zone_pageset_init(zone, cpu);
4445 * Allocate per cpu pagesets and initialize them.
4446 * Before this call only boot pagesets were available.
4448 void __init setup_per_cpu_pageset(void)
4452 for_each_populated_zone(zone)
4453 setup_zone_pageset(zone);
4456 static noinline __init_refok
4457 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4463 * The per-page waitqueue mechanism uses hashed waitqueues
4466 zone->wait_table_hash_nr_entries =
4467 wait_table_hash_nr_entries(zone_size_pages);
4468 zone->wait_table_bits =
4469 wait_table_bits(zone->wait_table_hash_nr_entries);
4470 alloc_size = zone->wait_table_hash_nr_entries
4471 * sizeof(wait_queue_head_t);
4473 if (!slab_is_available()) {
4474 zone->wait_table = (wait_queue_head_t *)
4475 memblock_virt_alloc_node_nopanic(
4476 alloc_size, zone->zone_pgdat->node_id);
4479 * This case means that a zone whose size was 0 gets new memory
4480 * via memory hot-add.
4481 * But it may be the case that a new node was hot-added. In
4482 * this case vmalloc() will not be able to use this new node's
4483 * memory - this wait_table must be initialized to use this new
4484 * node itself as well.
4485 * To use this new node's memory, further consideration will be
4488 zone->wait_table = vmalloc(alloc_size);
4490 if (!zone->wait_table)
4493 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4494 init_waitqueue_head(zone->wait_table + i);
4499 static __meminit void zone_pcp_init(struct zone *zone)
4502 * per cpu subsystem is not up at this point. The following code
4503 * relies on the ability of the linker to provide the
4504 * offset of a (static) per cpu variable into the per cpu area.
4506 zone->pageset = &boot_pageset;
4508 if (populated_zone(zone))
4509 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4510 zone->name, zone->present_pages,
4511 zone_batchsize(zone));
4514 int __meminit init_currently_empty_zone(struct zone *zone,
4515 unsigned long zone_start_pfn,
4517 enum memmap_context context)
4519 struct pglist_data *pgdat = zone->zone_pgdat;
4521 ret = zone_wait_table_init(zone, size);
4524 pgdat->nr_zones = zone_idx(zone) + 1;
4526 zone->zone_start_pfn = zone_start_pfn;
4528 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4529 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4531 (unsigned long)zone_idx(zone),
4532 zone_start_pfn, (zone_start_pfn + size));
4534 zone_init_free_lists(zone);
4539 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4540 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4542 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4544 int __meminit __early_pfn_to_nid(unsigned long pfn)
4546 unsigned long start_pfn, end_pfn;
4549 * NOTE: The following SMP-unsafe globals are only used early in boot
4550 * when the kernel is running single-threaded.
4552 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4553 static int __meminitdata last_nid;
4555 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4558 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4560 last_start_pfn = start_pfn;
4561 last_end_pfn = end_pfn;
4567 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4569 int __meminit early_pfn_to_nid(unsigned long pfn)
4573 nid = __early_pfn_to_nid(pfn);
4576 /* just returns 0 */
4580 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4581 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4585 nid = __early_pfn_to_nid(pfn);
4586 if (nid >= 0 && nid != node)
4593 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4594 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4595 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4597 * If an architecture guarantees that all ranges registered contain no holes
4598 * and may be freed, this this function may be used instead of calling
4599 * memblock_free_early_nid() manually.
4601 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4603 unsigned long start_pfn, end_pfn;
4606 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4607 start_pfn = min(start_pfn, max_low_pfn);
4608 end_pfn = min(end_pfn, max_low_pfn);
4610 if (start_pfn < end_pfn)
4611 memblock_free_early_nid(PFN_PHYS(start_pfn),
4612 (end_pfn - start_pfn) << PAGE_SHIFT,
4618 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4619 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4621 * If an architecture guarantees that all ranges registered contain no holes and may
4622 * be freed, this function may be used instead of calling memory_present() manually.
4624 void __init sparse_memory_present_with_active_regions(int nid)
4626 unsigned long start_pfn, end_pfn;
4629 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4630 memory_present(this_nid, start_pfn, end_pfn);
4634 * get_pfn_range_for_nid - Return the start and end page frames for a node
4635 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4636 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4637 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4639 * It returns the start and end page frame of a node based on information
4640 * provided by memblock_set_node(). If called for a node
4641 * with no available memory, a warning is printed and the start and end
4644 void __meminit get_pfn_range_for_nid(unsigned int nid,
4645 unsigned long *start_pfn, unsigned long *end_pfn)
4647 unsigned long this_start_pfn, this_end_pfn;
4653 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4654 *start_pfn = min(*start_pfn, this_start_pfn);
4655 *end_pfn = max(*end_pfn, this_end_pfn);
4658 if (*start_pfn == -1UL)
4663 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4664 * assumption is made that zones within a node are ordered in monotonic
4665 * increasing memory addresses so that the "highest" populated zone is used
4667 static void __init find_usable_zone_for_movable(void)
4670 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4671 if (zone_index == ZONE_MOVABLE)
4674 if (arch_zone_highest_possible_pfn[zone_index] >
4675 arch_zone_lowest_possible_pfn[zone_index])
4679 VM_BUG_ON(zone_index == -1);
4680 movable_zone = zone_index;
4684 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4685 * because it is sized independent of architecture. Unlike the other zones,
4686 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4687 * in each node depending on the size of each node and how evenly kernelcore
4688 * is distributed. This helper function adjusts the zone ranges
4689 * provided by the architecture for a given node by using the end of the
4690 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4691 * zones within a node are in order of monotonic increases memory addresses
4693 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4694 unsigned long zone_type,
4695 unsigned long node_start_pfn,
4696 unsigned long node_end_pfn,
4697 unsigned long *zone_start_pfn,
4698 unsigned long *zone_end_pfn)
4700 /* Only adjust if ZONE_MOVABLE is on this node */
4701 if (zone_movable_pfn[nid]) {
4702 /* Size ZONE_MOVABLE */
4703 if (zone_type == ZONE_MOVABLE) {
4704 *zone_start_pfn = zone_movable_pfn[nid];
4705 *zone_end_pfn = min(node_end_pfn,
4706 arch_zone_highest_possible_pfn[movable_zone]);
4708 /* Adjust for ZONE_MOVABLE starting within this range */
4709 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4710 *zone_end_pfn > zone_movable_pfn[nid]) {
4711 *zone_end_pfn = zone_movable_pfn[nid];
4713 /* Check if this whole range is within ZONE_MOVABLE */
4714 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4715 *zone_start_pfn = *zone_end_pfn;
4720 * Return the number of pages a zone spans in a node, including holes
4721 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4723 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4724 unsigned long zone_type,
4725 unsigned long node_start_pfn,
4726 unsigned long node_end_pfn,
4727 unsigned long *ignored)
4729 unsigned long zone_start_pfn, zone_end_pfn;
4731 /* Get the start and end of the zone */
4732 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4733 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4734 adjust_zone_range_for_zone_movable(nid, zone_type,
4735 node_start_pfn, node_end_pfn,
4736 &zone_start_pfn, &zone_end_pfn);
4738 /* Check that this node has pages within the zone's required range */
4739 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4742 /* Move the zone boundaries inside the node if necessary */
4743 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4744 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4746 /* Return the spanned pages */
4747 return zone_end_pfn - zone_start_pfn;
4751 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4752 * then all holes in the requested range will be accounted for.
4754 unsigned long __meminit __absent_pages_in_range(int nid,
4755 unsigned long range_start_pfn,
4756 unsigned long range_end_pfn)
4758 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4759 unsigned long start_pfn, end_pfn;
4762 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4763 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4764 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4765 nr_absent -= end_pfn - start_pfn;
4771 * absent_pages_in_range - Return number of page frames in holes within a range
4772 * @start_pfn: The start PFN to start searching for holes
4773 * @end_pfn: The end PFN to stop searching for holes
4775 * It returns the number of pages frames in memory holes within a range.
4777 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4778 unsigned long end_pfn)
4780 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4783 /* Return the number of page frames in holes in a zone on a node */
4784 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4785 unsigned long zone_type,
4786 unsigned long node_start_pfn,
4787 unsigned long node_end_pfn,
4788 unsigned long *ignored)
4790 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4791 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4792 unsigned long zone_start_pfn, zone_end_pfn;
4794 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4795 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4797 adjust_zone_range_for_zone_movable(nid, zone_type,
4798 node_start_pfn, node_end_pfn,
4799 &zone_start_pfn, &zone_end_pfn);
4800 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4803 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4804 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4805 unsigned long zone_type,
4806 unsigned long node_start_pfn,
4807 unsigned long node_end_pfn,
4808 unsigned long *zones_size)
4810 return zones_size[zone_type];
4813 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4814 unsigned long zone_type,
4815 unsigned long node_start_pfn,
4816 unsigned long node_end_pfn,
4817 unsigned long *zholes_size)
4822 return zholes_size[zone_type];
4825 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4827 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4828 unsigned long node_start_pfn,
4829 unsigned long node_end_pfn,
4830 unsigned long *zones_size,
4831 unsigned long *zholes_size)
4833 unsigned long realtotalpages, totalpages = 0;
4836 for (i = 0; i < MAX_NR_ZONES; i++)
4837 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4841 pgdat->node_spanned_pages = totalpages;
4843 realtotalpages = totalpages;
4844 for (i = 0; i < MAX_NR_ZONES; i++)
4846 zone_absent_pages_in_node(pgdat->node_id, i,
4847 node_start_pfn, node_end_pfn,
4849 pgdat->node_present_pages = realtotalpages;
4850 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4854 #ifndef CONFIG_SPARSEMEM
4856 * Calculate the size of the zone->blockflags rounded to an unsigned long
4857 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4858 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4859 * round what is now in bits to nearest long in bits, then return it in
4862 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4864 unsigned long usemapsize;
4866 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4867 usemapsize = roundup(zonesize, pageblock_nr_pages);
4868 usemapsize = usemapsize >> pageblock_order;
4869 usemapsize *= NR_PAGEBLOCK_BITS;
4870 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4872 return usemapsize / 8;
4875 static void __init setup_usemap(struct pglist_data *pgdat,
4877 unsigned long zone_start_pfn,
4878 unsigned long zonesize)
4880 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4881 zone->pageblock_flags = NULL;
4883 zone->pageblock_flags =
4884 memblock_virt_alloc_node_nopanic(usemapsize,
4888 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4889 unsigned long zone_start_pfn, unsigned long zonesize) {}
4890 #endif /* CONFIG_SPARSEMEM */
4892 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4894 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4895 void __paginginit set_pageblock_order(void)
4899 /* Check that pageblock_nr_pages has not already been setup */
4900 if (pageblock_order)
4903 if (HPAGE_SHIFT > PAGE_SHIFT)
4904 order = HUGETLB_PAGE_ORDER;
4906 order = MAX_ORDER - 1;
4909 * Assume the largest contiguous order of interest is a huge page.
4910 * This value may be variable depending on boot parameters on IA64 and
4913 pageblock_order = order;
4915 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4918 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4919 * is unused as pageblock_order is set at compile-time. See
4920 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4923 void __paginginit set_pageblock_order(void)
4927 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4929 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4930 unsigned long present_pages)
4932 unsigned long pages = spanned_pages;
4935 * Provide a more accurate estimation if there are holes within
4936 * the zone and SPARSEMEM is in use. If there are holes within the
4937 * zone, each populated memory region may cost us one or two extra
4938 * memmap pages due to alignment because memmap pages for each
4939 * populated regions may not naturally algined on page boundary.
4940 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4942 if (spanned_pages > present_pages + (present_pages >> 4) &&
4943 IS_ENABLED(CONFIG_SPARSEMEM))
4944 pages = present_pages;
4946 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4950 * Set up the zone data structures:
4951 * - mark all pages reserved
4952 * - mark all memory queues empty
4953 * - clear the memory bitmaps
4955 * NOTE: pgdat should get zeroed by caller.
4957 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4958 unsigned long node_start_pfn, unsigned long node_end_pfn,
4959 unsigned long *zones_size, unsigned long *zholes_size)
4962 int nid = pgdat->node_id;
4963 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4966 pgdat_resize_init(pgdat);
4967 #ifdef CONFIG_NUMA_BALANCING
4968 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4969 pgdat->numabalancing_migrate_nr_pages = 0;
4970 pgdat->numabalancing_migrate_next_window = jiffies;
4972 init_waitqueue_head(&pgdat->kswapd_wait);
4973 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4974 pgdat_page_ext_init(pgdat);
4976 for (j = 0; j < MAX_NR_ZONES; j++) {
4977 struct zone *zone = pgdat->node_zones + j;
4978 unsigned long size, realsize, freesize, memmap_pages;
4980 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4981 node_end_pfn, zones_size);
4982 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4988 * Adjust freesize so that it accounts for how much memory
4989 * is used by this zone for memmap. This affects the watermark
4990 * and per-cpu initialisations
4992 memmap_pages = calc_memmap_size(size, realsize);
4993 if (!is_highmem_idx(j)) {
4994 if (freesize >= memmap_pages) {
4995 freesize -= memmap_pages;
4998 " %s zone: %lu pages used for memmap\n",
4999 zone_names[j], memmap_pages);
5002 " %s zone: %lu pages exceeds freesize %lu\n",
5003 zone_names[j], memmap_pages, freesize);
5006 /* Account for reserved pages */
5007 if (j == 0 && freesize > dma_reserve) {
5008 freesize -= dma_reserve;
5009 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5010 zone_names[0], dma_reserve);
5013 if (!is_highmem_idx(j))
5014 nr_kernel_pages += freesize;
5015 /* Charge for highmem memmap if there are enough kernel pages */
5016 else if (nr_kernel_pages > memmap_pages * 2)
5017 nr_kernel_pages -= memmap_pages;
5018 nr_all_pages += freesize;
5020 zone->spanned_pages = size;
5021 zone->present_pages = realsize;
5023 * Set an approximate value for lowmem here, it will be adjusted
5024 * when the bootmem allocator frees pages into the buddy system.
5025 * And all highmem pages will be managed by the buddy system.
5027 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5030 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5032 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5034 zone->name = zone_names[j];
5035 spin_lock_init(&zone->lock);
5036 spin_lock_init(&zone->lru_lock);
5037 zone_seqlock_init(zone);
5038 zone->zone_pgdat = pgdat;
5039 zone_pcp_init(zone);
5041 /* For bootup, initialized properly in watermark setup */
5042 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5044 lruvec_init(&zone->lruvec);
5048 set_pageblock_order();
5049 setup_usemap(pgdat, zone, zone_start_pfn, size);
5050 ret = init_currently_empty_zone(zone, zone_start_pfn,
5051 size, MEMMAP_EARLY);
5053 memmap_init(size, nid, j, zone_start_pfn);
5054 zone_start_pfn += size;
5058 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5060 /* Skip empty nodes */
5061 if (!pgdat->node_spanned_pages)
5064 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5065 /* ia64 gets its own node_mem_map, before this, without bootmem */
5066 if (!pgdat->node_mem_map) {
5067 unsigned long size, start, end;
5071 * The zone's endpoints aren't required to be MAX_ORDER
5072 * aligned but the node_mem_map endpoints must be in order
5073 * for the buddy allocator to function correctly.
5075 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5076 end = pgdat_end_pfn(pgdat);
5077 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5078 size = (end - start) * sizeof(struct page);
5079 map = alloc_remap(pgdat->node_id, size);
5081 map = memblock_virt_alloc_node_nopanic(size,
5083 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5085 #ifndef CONFIG_NEED_MULTIPLE_NODES
5087 * With no DISCONTIG, the global mem_map is just set as node 0's
5089 if (pgdat == NODE_DATA(0)) {
5090 mem_map = NODE_DATA(0)->node_mem_map;
5091 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5092 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5093 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5094 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5097 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5100 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5101 unsigned long node_start_pfn, unsigned long *zholes_size)
5103 pg_data_t *pgdat = NODE_DATA(nid);
5104 unsigned long start_pfn = 0;
5105 unsigned long end_pfn = 0;
5107 /* pg_data_t should be reset to zero when it's allocated */
5108 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5110 pgdat->node_id = nid;
5111 pgdat->node_start_pfn = node_start_pfn;
5112 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5113 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5114 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5115 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
5117 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5118 zones_size, zholes_size);
5120 alloc_node_mem_map(pgdat);
5121 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5122 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5123 nid, (unsigned long)pgdat,
5124 (unsigned long)pgdat->node_mem_map);
5127 free_area_init_core(pgdat, start_pfn, end_pfn,
5128 zones_size, zholes_size);
5131 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5133 #if MAX_NUMNODES > 1
5135 * Figure out the number of possible node ids.
5137 void __init setup_nr_node_ids(void)
5140 unsigned int highest = 0;
5142 for_each_node_mask(node, node_possible_map)
5144 nr_node_ids = highest + 1;
5149 * node_map_pfn_alignment - determine the maximum internode alignment
5151 * This function should be called after node map is populated and sorted.
5152 * It calculates the maximum power of two alignment which can distinguish
5155 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5156 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5157 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5158 * shifted, 1GiB is enough and this function will indicate so.
5160 * This is used to test whether pfn -> nid mapping of the chosen memory
5161 * model has fine enough granularity to avoid incorrect mapping for the
5162 * populated node map.
5164 * Returns the determined alignment in pfn's. 0 if there is no alignment
5165 * requirement (single node).
5167 unsigned long __init node_map_pfn_alignment(void)
5169 unsigned long accl_mask = 0, last_end = 0;
5170 unsigned long start, end, mask;
5174 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5175 if (!start || last_nid < 0 || last_nid == nid) {
5182 * Start with a mask granular enough to pin-point to the
5183 * start pfn and tick off bits one-by-one until it becomes
5184 * too coarse to separate the current node from the last.
5186 mask = ~((1 << __ffs(start)) - 1);
5187 while (mask && last_end <= (start & (mask << 1)))
5190 /* accumulate all internode masks */
5194 /* convert mask to number of pages */
5195 return ~accl_mask + 1;
5198 /* Find the lowest pfn for a node */
5199 static unsigned long __init find_min_pfn_for_node(int nid)
5201 unsigned long min_pfn = ULONG_MAX;
5202 unsigned long start_pfn;
5205 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5206 min_pfn = min(min_pfn, start_pfn);
5208 if (min_pfn == ULONG_MAX) {
5210 "Could not find start_pfn for node %d\n", nid);
5218 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5220 * It returns the minimum PFN based on information provided via
5221 * memblock_set_node().
5223 unsigned long __init find_min_pfn_with_active_regions(void)
5225 return find_min_pfn_for_node(MAX_NUMNODES);
5229 * early_calculate_totalpages()
5230 * Sum pages in active regions for movable zone.
5231 * Populate N_MEMORY for calculating usable_nodes.
5233 static unsigned long __init early_calculate_totalpages(void)
5235 unsigned long totalpages = 0;
5236 unsigned long start_pfn, end_pfn;
5239 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5240 unsigned long pages = end_pfn - start_pfn;
5242 totalpages += pages;
5244 node_set_state(nid, N_MEMORY);
5250 * Find the PFN the Movable zone begins in each node. Kernel memory
5251 * is spread evenly between nodes as long as the nodes have enough
5252 * memory. When they don't, some nodes will have more kernelcore than
5255 static void __init find_zone_movable_pfns_for_nodes(void)
5258 unsigned long usable_startpfn;
5259 unsigned long kernelcore_node, kernelcore_remaining;
5260 /* save the state before borrow the nodemask */
5261 nodemask_t saved_node_state = node_states[N_MEMORY];
5262 unsigned long totalpages = early_calculate_totalpages();
5263 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5264 struct memblock_region *r;
5266 /* Need to find movable_zone earlier when movable_node is specified. */
5267 find_usable_zone_for_movable();
5270 * If movable_node is specified, ignore kernelcore and movablecore
5273 if (movable_node_is_enabled()) {
5274 for_each_memblock(memory, r) {
5275 if (!memblock_is_hotpluggable(r))
5280 usable_startpfn = PFN_DOWN(r->base);
5281 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5282 min(usable_startpfn, zone_movable_pfn[nid]) :
5290 * If movablecore=nn[KMG] was specified, calculate what size of
5291 * kernelcore that corresponds so that memory usable for
5292 * any allocation type is evenly spread. If both kernelcore
5293 * and movablecore are specified, then the value of kernelcore
5294 * will be used for required_kernelcore if it's greater than
5295 * what movablecore would have allowed.
5297 if (required_movablecore) {
5298 unsigned long corepages;
5301 * Round-up so that ZONE_MOVABLE is at least as large as what
5302 * was requested by the user
5304 required_movablecore =
5305 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5306 corepages = totalpages - required_movablecore;
5308 required_kernelcore = max(required_kernelcore, corepages);
5311 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5312 if (!required_kernelcore)
5315 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5316 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5319 /* Spread kernelcore memory as evenly as possible throughout nodes */
5320 kernelcore_node = required_kernelcore / usable_nodes;
5321 for_each_node_state(nid, N_MEMORY) {
5322 unsigned long start_pfn, end_pfn;
5325 * Recalculate kernelcore_node if the division per node
5326 * now exceeds what is necessary to satisfy the requested
5327 * amount of memory for the kernel
5329 if (required_kernelcore < kernelcore_node)
5330 kernelcore_node = required_kernelcore / usable_nodes;
5333 * As the map is walked, we track how much memory is usable
5334 * by the kernel using kernelcore_remaining. When it is
5335 * 0, the rest of the node is usable by ZONE_MOVABLE
5337 kernelcore_remaining = kernelcore_node;
5339 /* Go through each range of PFNs within this node */
5340 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5341 unsigned long size_pages;
5343 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5344 if (start_pfn >= end_pfn)
5347 /* Account for what is only usable for kernelcore */
5348 if (start_pfn < usable_startpfn) {
5349 unsigned long kernel_pages;
5350 kernel_pages = min(end_pfn, usable_startpfn)
5353 kernelcore_remaining -= min(kernel_pages,
5354 kernelcore_remaining);
5355 required_kernelcore -= min(kernel_pages,
5356 required_kernelcore);
5358 /* Continue if range is now fully accounted */
5359 if (end_pfn <= usable_startpfn) {
5362 * Push zone_movable_pfn to the end so
5363 * that if we have to rebalance
5364 * kernelcore across nodes, we will
5365 * not double account here
5367 zone_movable_pfn[nid] = end_pfn;
5370 start_pfn = usable_startpfn;
5374 * The usable PFN range for ZONE_MOVABLE is from
5375 * start_pfn->end_pfn. Calculate size_pages as the
5376 * number of pages used as kernelcore
5378 size_pages = end_pfn - start_pfn;
5379 if (size_pages > kernelcore_remaining)
5380 size_pages = kernelcore_remaining;
5381 zone_movable_pfn[nid] = start_pfn + size_pages;
5384 * Some kernelcore has been met, update counts and
5385 * break if the kernelcore for this node has been
5388 required_kernelcore -= min(required_kernelcore,
5390 kernelcore_remaining -= size_pages;
5391 if (!kernelcore_remaining)
5397 * If there is still required_kernelcore, we do another pass with one
5398 * less node in the count. This will push zone_movable_pfn[nid] further
5399 * along on the nodes that still have memory until kernelcore is
5403 if (usable_nodes && required_kernelcore > usable_nodes)
5407 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5408 for (nid = 0; nid < MAX_NUMNODES; nid++)
5409 zone_movable_pfn[nid] =
5410 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5413 /* restore the node_state */
5414 node_states[N_MEMORY] = saved_node_state;
5417 /* Any regular or high memory on that node ? */
5418 static void check_for_memory(pg_data_t *pgdat, int nid)
5420 enum zone_type zone_type;
5422 if (N_MEMORY == N_NORMAL_MEMORY)
5425 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5426 struct zone *zone = &pgdat->node_zones[zone_type];
5427 if (populated_zone(zone)) {
5428 node_set_state(nid, N_HIGH_MEMORY);
5429 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5430 zone_type <= ZONE_NORMAL)
5431 node_set_state(nid, N_NORMAL_MEMORY);
5438 * free_area_init_nodes - Initialise all pg_data_t and zone data
5439 * @max_zone_pfn: an array of max PFNs for each zone
5441 * This will call free_area_init_node() for each active node in the system.
5442 * Using the page ranges provided by memblock_set_node(), the size of each
5443 * zone in each node and their holes is calculated. If the maximum PFN
5444 * between two adjacent zones match, it is assumed that the zone is empty.
5445 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5446 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5447 * starts where the previous one ended. For example, ZONE_DMA32 starts
5448 * at arch_max_dma_pfn.
5450 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5452 unsigned long start_pfn, end_pfn;
5455 /* Record where the zone boundaries are */
5456 memset(arch_zone_lowest_possible_pfn, 0,
5457 sizeof(arch_zone_lowest_possible_pfn));
5458 memset(arch_zone_highest_possible_pfn, 0,
5459 sizeof(arch_zone_highest_possible_pfn));
5460 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5461 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5462 for (i = 1; i < MAX_NR_ZONES; i++) {
5463 if (i == ZONE_MOVABLE)
5465 arch_zone_lowest_possible_pfn[i] =
5466 arch_zone_highest_possible_pfn[i-1];
5467 arch_zone_highest_possible_pfn[i] =
5468 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5470 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5471 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5473 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5474 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5475 find_zone_movable_pfns_for_nodes();
5477 /* Print out the zone ranges */
5478 pr_info("Zone ranges:\n");
5479 for (i = 0; i < MAX_NR_ZONES; i++) {
5480 if (i == ZONE_MOVABLE)
5482 pr_info(" %-8s ", zone_names[i]);
5483 if (arch_zone_lowest_possible_pfn[i] ==
5484 arch_zone_highest_possible_pfn[i])
5487 pr_cont("[mem %#018Lx-%#018Lx]\n",
5488 (u64)arch_zone_lowest_possible_pfn[i]
5490 ((u64)arch_zone_highest_possible_pfn[i]
5491 << PAGE_SHIFT) - 1);
5494 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5495 pr_info("Movable zone start for each node\n");
5496 for (i = 0; i < MAX_NUMNODES; i++) {
5497 if (zone_movable_pfn[i])
5498 pr_info(" Node %d: %#018Lx\n", i,
5499 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5502 /* Print out the early node map */
5503 pr_info("Early memory node ranges\n");
5504 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5505 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5506 (u64)start_pfn << PAGE_SHIFT,
5507 ((u64)end_pfn << PAGE_SHIFT) - 1);
5509 /* Initialise every node */
5510 mminit_verify_pageflags_layout();
5511 setup_nr_node_ids();
5512 for_each_online_node(nid) {
5513 pg_data_t *pgdat = NODE_DATA(nid);
5514 free_area_init_node(nid, NULL,
5515 find_min_pfn_for_node(nid), NULL);
5517 /* Any memory on that node */
5518 if (pgdat->node_present_pages)
5519 node_set_state(nid, N_MEMORY);
5520 check_for_memory(pgdat, nid);
5524 static int __init cmdline_parse_core(char *p, unsigned long *core)
5526 unsigned long long coremem;
5530 coremem = memparse(p, &p);
5531 *core = coremem >> PAGE_SHIFT;
5533 /* Paranoid check that UL is enough for the coremem value */
5534 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5540 * kernelcore=size sets the amount of memory for use for allocations that
5541 * cannot be reclaimed or migrated.
5543 static int __init cmdline_parse_kernelcore(char *p)
5545 return cmdline_parse_core(p, &required_kernelcore);
5549 * movablecore=size sets the amount of memory for use for allocations that
5550 * can be reclaimed or migrated.
5552 static int __init cmdline_parse_movablecore(char *p)
5554 return cmdline_parse_core(p, &required_movablecore);
5557 early_param("kernelcore", cmdline_parse_kernelcore);
5558 early_param("movablecore", cmdline_parse_movablecore);
5560 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5562 void adjust_managed_page_count(struct page *page, long count)
5564 spin_lock(&managed_page_count_lock);
5565 page_zone(page)->managed_pages += count;
5566 totalram_pages += count;
5567 #ifdef CONFIG_HIGHMEM
5568 if (PageHighMem(page))
5569 totalhigh_pages += count;
5571 spin_unlock(&managed_page_count_lock);
5573 EXPORT_SYMBOL(adjust_managed_page_count);
5575 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5578 unsigned long pages = 0;
5580 start = (void *)PAGE_ALIGN((unsigned long)start);
5581 end = (void *)((unsigned long)end & PAGE_MASK);
5582 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5583 if ((unsigned int)poison <= 0xFF)
5584 memset(pos, poison, PAGE_SIZE);
5585 free_reserved_page(virt_to_page(pos));
5589 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5590 s, pages << (PAGE_SHIFT - 10), start, end);
5594 EXPORT_SYMBOL(free_reserved_area);
5596 #ifdef CONFIG_HIGHMEM
5597 void free_highmem_page(struct page *page)
5599 __free_reserved_page(page);
5601 page_zone(page)->managed_pages++;
5607 void __init mem_init_print_info(const char *str)
5609 unsigned long physpages, codesize, datasize, rosize, bss_size;
5610 unsigned long init_code_size, init_data_size;
5612 physpages = get_num_physpages();
5613 codesize = _etext - _stext;
5614 datasize = _edata - _sdata;
5615 rosize = __end_rodata - __start_rodata;
5616 bss_size = __bss_stop - __bss_start;
5617 init_data_size = __init_end - __init_begin;
5618 init_code_size = _einittext - _sinittext;
5621 * Detect special cases and adjust section sizes accordingly:
5622 * 1) .init.* may be embedded into .data sections
5623 * 2) .init.text.* may be out of [__init_begin, __init_end],
5624 * please refer to arch/tile/kernel/vmlinux.lds.S.
5625 * 3) .rodata.* may be embedded into .text or .data sections.
5627 #define adj_init_size(start, end, size, pos, adj) \
5629 if (start <= pos && pos < end && size > adj) \
5633 adj_init_size(__init_begin, __init_end, init_data_size,
5634 _sinittext, init_code_size);
5635 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5636 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5637 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5638 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5640 #undef adj_init_size
5642 pr_info("Memory: %luK/%luK available "
5643 "(%luK kernel code, %luK rwdata, %luK rodata, "
5644 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5645 #ifdef CONFIG_HIGHMEM
5649 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5650 codesize >> 10, datasize >> 10, rosize >> 10,
5651 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5652 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5653 totalcma_pages << (PAGE_SHIFT-10),
5654 #ifdef CONFIG_HIGHMEM
5655 totalhigh_pages << (PAGE_SHIFT-10),
5657 str ? ", " : "", str ? str : "");
5661 * set_dma_reserve - set the specified number of pages reserved in the first zone
5662 * @new_dma_reserve: The number of pages to mark reserved
5664 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5665 * In the DMA zone, a significant percentage may be consumed by kernel image
5666 * and other unfreeable allocations which can skew the watermarks badly. This
5667 * function may optionally be used to account for unfreeable pages in the
5668 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5669 * smaller per-cpu batchsize.
5671 void __init set_dma_reserve(unsigned long new_dma_reserve)
5673 dma_reserve = new_dma_reserve;
5676 void __init free_area_init(unsigned long *zones_size)
5678 free_area_init_node(0, zones_size,
5679 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5682 static int page_alloc_cpu_notify(struct notifier_block *self,
5683 unsigned long action, void *hcpu)
5685 int cpu = (unsigned long)hcpu;
5687 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5688 lru_add_drain_cpu(cpu);
5692 * Spill the event counters of the dead processor
5693 * into the current processors event counters.
5694 * This artificially elevates the count of the current
5697 vm_events_fold_cpu(cpu);
5700 * Zero the differential counters of the dead processor
5701 * so that the vm statistics are consistent.
5703 * This is only okay since the processor is dead and cannot
5704 * race with what we are doing.
5706 cpu_vm_stats_fold(cpu);
5711 void __init page_alloc_init(void)
5713 hotcpu_notifier(page_alloc_cpu_notify, 0);
5714 local_irq_lock_init(pa_lock);
5718 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5719 * or min_free_kbytes changes.
5721 static void calculate_totalreserve_pages(void)
5723 struct pglist_data *pgdat;
5724 unsigned long reserve_pages = 0;
5725 enum zone_type i, j;
5727 for_each_online_pgdat(pgdat) {
5728 for (i = 0; i < MAX_NR_ZONES; i++) {
5729 struct zone *zone = pgdat->node_zones + i;
5732 /* Find valid and maximum lowmem_reserve in the zone */
5733 for (j = i; j < MAX_NR_ZONES; j++) {
5734 if (zone->lowmem_reserve[j] > max)
5735 max = zone->lowmem_reserve[j];
5738 /* we treat the high watermark as reserved pages. */
5739 max += high_wmark_pages(zone);
5741 if (max > zone->managed_pages)
5742 max = zone->managed_pages;
5743 reserve_pages += max;
5745 * Lowmem reserves are not available to
5746 * GFP_HIGHUSER page cache allocations and
5747 * kswapd tries to balance zones to their high
5748 * watermark. As a result, neither should be
5749 * regarded as dirtyable memory, to prevent a
5750 * situation where reclaim has to clean pages
5751 * in order to balance the zones.
5753 zone->dirty_balance_reserve = max;
5756 dirty_balance_reserve = reserve_pages;
5757 totalreserve_pages = reserve_pages;
5761 * setup_per_zone_lowmem_reserve - called whenever
5762 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5763 * has a correct pages reserved value, so an adequate number of
5764 * pages are left in the zone after a successful __alloc_pages().
5766 static void setup_per_zone_lowmem_reserve(void)
5768 struct pglist_data *pgdat;
5769 enum zone_type j, idx;
5771 for_each_online_pgdat(pgdat) {
5772 for (j = 0; j < MAX_NR_ZONES; j++) {
5773 struct zone *zone = pgdat->node_zones + j;
5774 unsigned long managed_pages = zone->managed_pages;
5776 zone->lowmem_reserve[j] = 0;
5780 struct zone *lower_zone;
5784 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5785 sysctl_lowmem_reserve_ratio[idx] = 1;
5787 lower_zone = pgdat->node_zones + idx;
5788 lower_zone->lowmem_reserve[j] = managed_pages /
5789 sysctl_lowmem_reserve_ratio[idx];
5790 managed_pages += lower_zone->managed_pages;
5795 /* update totalreserve_pages */
5796 calculate_totalreserve_pages();
5799 static void __setup_per_zone_wmarks(void)
5801 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5802 unsigned long lowmem_pages = 0;
5804 unsigned long flags;
5806 /* Calculate total number of !ZONE_HIGHMEM pages */
5807 for_each_zone(zone) {
5808 if (!is_highmem(zone))
5809 lowmem_pages += zone->managed_pages;
5812 for_each_zone(zone) {
5815 spin_lock_irqsave(&zone->lock, flags);
5816 tmp = (u64)pages_min * zone->managed_pages;
5817 do_div(tmp, lowmem_pages);
5818 if (is_highmem(zone)) {
5820 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5821 * need highmem pages, so cap pages_min to a small
5824 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5825 * deltas control asynch page reclaim, and so should
5826 * not be capped for highmem.
5828 unsigned long min_pages;
5830 min_pages = zone->managed_pages / 1024;
5831 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5832 zone->watermark[WMARK_MIN] = min_pages;
5835 * If it's a lowmem zone, reserve a number of pages
5836 * proportionate to the zone's size.
5838 zone->watermark[WMARK_MIN] = tmp;
5841 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5842 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5844 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5845 high_wmark_pages(zone) - low_wmark_pages(zone) -
5846 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5848 setup_zone_migrate_reserve(zone);
5849 spin_unlock_irqrestore(&zone->lock, flags);
5852 /* update totalreserve_pages */
5853 calculate_totalreserve_pages();
5857 * setup_per_zone_wmarks - called when min_free_kbytes changes
5858 * or when memory is hot-{added|removed}
5860 * Ensures that the watermark[min,low,high] values for each zone are set
5861 * correctly with respect to min_free_kbytes.
5863 void setup_per_zone_wmarks(void)
5865 mutex_lock(&zonelists_mutex);
5866 __setup_per_zone_wmarks();
5867 mutex_unlock(&zonelists_mutex);
5871 * The inactive anon list should be small enough that the VM never has to
5872 * do too much work, but large enough that each inactive page has a chance
5873 * to be referenced again before it is swapped out.
5875 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5876 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5877 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5878 * the anonymous pages are kept on the inactive list.
5881 * memory ratio inactive anon
5882 * -------------------------------------
5891 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5893 unsigned int gb, ratio;
5895 /* Zone size in gigabytes */
5896 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5898 ratio = int_sqrt(10 * gb);
5902 zone->inactive_ratio = ratio;
5905 static void __meminit setup_per_zone_inactive_ratio(void)
5910 calculate_zone_inactive_ratio(zone);
5914 * Initialise min_free_kbytes.
5916 * For small machines we want it small (128k min). For large machines
5917 * we want it large (64MB max). But it is not linear, because network
5918 * bandwidth does not increase linearly with machine size. We use
5920 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5921 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5937 int __meminit init_per_zone_wmark_min(void)
5939 unsigned long lowmem_kbytes;
5940 int new_min_free_kbytes;
5942 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5943 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5945 if (new_min_free_kbytes > user_min_free_kbytes) {
5946 min_free_kbytes = new_min_free_kbytes;
5947 if (min_free_kbytes < 128)
5948 min_free_kbytes = 128;
5949 if (min_free_kbytes > 65536)
5950 min_free_kbytes = 65536;
5952 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5953 new_min_free_kbytes, user_min_free_kbytes);
5955 setup_per_zone_wmarks();
5956 refresh_zone_stat_thresholds();
5957 setup_per_zone_lowmem_reserve();
5958 setup_per_zone_inactive_ratio();
5961 module_init(init_per_zone_wmark_min)
5964 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5965 * that we can call two helper functions whenever min_free_kbytes
5968 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5969 void __user *buffer, size_t *length, loff_t *ppos)
5973 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5978 user_min_free_kbytes = min_free_kbytes;
5979 setup_per_zone_wmarks();
5985 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5986 void __user *buffer, size_t *length, loff_t *ppos)
5991 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5996 zone->min_unmapped_pages = (zone->managed_pages *
5997 sysctl_min_unmapped_ratio) / 100;
6001 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6002 void __user *buffer, size_t *length, loff_t *ppos)
6007 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6012 zone->min_slab_pages = (zone->managed_pages *
6013 sysctl_min_slab_ratio) / 100;
6019 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6020 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6021 * whenever sysctl_lowmem_reserve_ratio changes.
6023 * The reserve ratio obviously has absolutely no relation with the
6024 * minimum watermarks. The lowmem reserve ratio can only make sense
6025 * if in function of the boot time zone sizes.
6027 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6028 void __user *buffer, size_t *length, loff_t *ppos)
6030 proc_dointvec_minmax(table, write, buffer, length, ppos);
6031 setup_per_zone_lowmem_reserve();
6036 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6037 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6038 * pagelist can have before it gets flushed back to buddy allocator.
6040 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6041 void __user *buffer, size_t *length, loff_t *ppos)
6044 int old_percpu_pagelist_fraction;
6047 mutex_lock(&pcp_batch_high_lock);
6048 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6050 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6051 if (!write || ret < 0)
6054 /* Sanity checking to avoid pcp imbalance */
6055 if (percpu_pagelist_fraction &&
6056 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6057 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6063 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6066 for_each_populated_zone(zone) {
6069 for_each_possible_cpu(cpu)
6070 pageset_set_high_and_batch(zone,
6071 per_cpu_ptr(zone->pageset, cpu));
6074 mutex_unlock(&pcp_batch_high_lock);
6078 int hashdist = HASHDIST_DEFAULT;
6081 static int __init set_hashdist(char *str)
6085 hashdist = simple_strtoul(str, &str, 0);
6088 __setup("hashdist=", set_hashdist);
6092 * allocate a large system hash table from bootmem
6093 * - it is assumed that the hash table must contain an exact power-of-2
6094 * quantity of entries
6095 * - limit is the number of hash buckets, not the total allocation size
6097 void *__init alloc_large_system_hash(const char *tablename,
6098 unsigned long bucketsize,
6099 unsigned long numentries,
6102 unsigned int *_hash_shift,
6103 unsigned int *_hash_mask,
6104 unsigned long low_limit,
6105 unsigned long high_limit)
6107 unsigned long long max = high_limit;
6108 unsigned long log2qty, size;
6111 /* allow the kernel cmdline to have a say */
6113 /* round applicable memory size up to nearest megabyte */
6114 numentries = nr_kernel_pages;
6116 /* It isn't necessary when PAGE_SIZE >= 1MB */
6117 if (PAGE_SHIFT < 20)
6118 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6120 /* limit to 1 bucket per 2^scale bytes of low memory */
6121 if (scale > PAGE_SHIFT)
6122 numentries >>= (scale - PAGE_SHIFT);
6124 numentries <<= (PAGE_SHIFT - scale);
6126 /* Make sure we've got at least a 0-order allocation.. */
6127 if (unlikely(flags & HASH_SMALL)) {
6128 /* Makes no sense without HASH_EARLY */
6129 WARN_ON(!(flags & HASH_EARLY));
6130 if (!(numentries >> *_hash_shift)) {
6131 numentries = 1UL << *_hash_shift;
6132 BUG_ON(!numentries);
6134 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6135 numentries = PAGE_SIZE / bucketsize;
6137 numentries = roundup_pow_of_two(numentries);
6139 /* limit allocation size to 1/16 total memory by default */
6141 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6142 do_div(max, bucketsize);
6144 max = min(max, 0x80000000ULL);
6146 if (numentries < low_limit)
6147 numentries = low_limit;
6148 if (numentries > max)
6151 log2qty = ilog2(numentries);
6154 size = bucketsize << log2qty;
6155 if (flags & HASH_EARLY)
6156 table = memblock_virt_alloc_nopanic(size, 0);
6158 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6161 * If bucketsize is not a power-of-two, we may free
6162 * some pages at the end of hash table which
6163 * alloc_pages_exact() automatically does
6165 if (get_order(size) < MAX_ORDER) {
6166 table = alloc_pages_exact(size, GFP_ATOMIC);
6167 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6170 } while (!table && size > PAGE_SIZE && --log2qty);
6173 panic("Failed to allocate %s hash table\n", tablename);
6175 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6178 ilog2(size) - PAGE_SHIFT,
6182 *_hash_shift = log2qty;
6184 *_hash_mask = (1 << log2qty) - 1;
6189 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6190 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6193 #ifdef CONFIG_SPARSEMEM
6194 return __pfn_to_section(pfn)->pageblock_flags;
6196 return zone->pageblock_flags;
6197 #endif /* CONFIG_SPARSEMEM */
6200 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6202 #ifdef CONFIG_SPARSEMEM
6203 pfn &= (PAGES_PER_SECTION-1);
6204 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6206 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6207 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6208 #endif /* CONFIG_SPARSEMEM */
6212 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6213 * @page: The page within the block of interest
6214 * @pfn: The target page frame number
6215 * @end_bitidx: The last bit of interest to retrieve
6216 * @mask: mask of bits that the caller is interested in
6218 * Return: pageblock_bits flags
6220 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6221 unsigned long end_bitidx,
6225 unsigned long *bitmap;
6226 unsigned long bitidx, word_bitidx;
6229 zone = page_zone(page);
6230 bitmap = get_pageblock_bitmap(zone, pfn);
6231 bitidx = pfn_to_bitidx(zone, pfn);
6232 word_bitidx = bitidx / BITS_PER_LONG;
6233 bitidx &= (BITS_PER_LONG-1);
6235 word = bitmap[word_bitidx];
6236 bitidx += end_bitidx;
6237 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6241 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6242 * @page: The page within the block of interest
6243 * @flags: The flags to set
6244 * @pfn: The target page frame number
6245 * @end_bitidx: The last bit of interest
6246 * @mask: mask of bits that the caller is interested in
6248 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6250 unsigned long end_bitidx,
6254 unsigned long *bitmap;
6255 unsigned long bitidx, word_bitidx;
6256 unsigned long old_word, word;
6258 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6260 zone = page_zone(page);
6261 bitmap = get_pageblock_bitmap(zone, pfn);
6262 bitidx = pfn_to_bitidx(zone, pfn);
6263 word_bitidx = bitidx / BITS_PER_LONG;
6264 bitidx &= (BITS_PER_LONG-1);
6266 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6268 bitidx += end_bitidx;
6269 mask <<= (BITS_PER_LONG - bitidx - 1);
6270 flags <<= (BITS_PER_LONG - bitidx - 1);
6272 word = READ_ONCE(bitmap[word_bitidx]);
6274 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6275 if (word == old_word)
6282 * This function checks whether pageblock includes unmovable pages or not.
6283 * If @count is not zero, it is okay to include less @count unmovable pages
6285 * PageLRU check without isolation or lru_lock could race so that
6286 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6287 * expect this function should be exact.
6289 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6290 bool skip_hwpoisoned_pages)
6292 unsigned long pfn, iter, found;
6296 * For avoiding noise data, lru_add_drain_all() should be called
6297 * If ZONE_MOVABLE, the zone never contains unmovable pages
6299 if (zone_idx(zone) == ZONE_MOVABLE)
6301 mt = get_pageblock_migratetype(page);
6302 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6305 pfn = page_to_pfn(page);
6306 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6307 unsigned long check = pfn + iter;
6309 if (!pfn_valid_within(check))
6312 page = pfn_to_page(check);
6315 * Hugepages are not in LRU lists, but they're movable.
6316 * We need not scan over tail pages bacause we don't
6317 * handle each tail page individually in migration.
6319 if (PageHuge(page)) {
6320 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6325 * We can't use page_count without pin a page
6326 * because another CPU can free compound page.
6327 * This check already skips compound tails of THP
6328 * because their page->_count is zero at all time.
6330 if (!atomic_read(&page->_count)) {
6331 if (PageBuddy(page))
6332 iter += (1 << page_order(page)) - 1;
6337 * The HWPoisoned page may be not in buddy system, and
6338 * page_count() is not 0.
6340 if (skip_hwpoisoned_pages && PageHWPoison(page))
6346 * If there are RECLAIMABLE pages, we need to check
6347 * it. But now, memory offline itself doesn't call
6348 * shrink_node_slabs() and it still to be fixed.
6351 * If the page is not RAM, page_count()should be 0.
6352 * we don't need more check. This is an _used_ not-movable page.
6354 * The problematic thing here is PG_reserved pages. PG_reserved
6355 * is set to both of a memory hole page and a _used_ kernel
6364 bool is_pageblock_removable_nolock(struct page *page)
6370 * We have to be careful here because we are iterating over memory
6371 * sections which are not zone aware so we might end up outside of
6372 * the zone but still within the section.
6373 * We have to take care about the node as well. If the node is offline
6374 * its NODE_DATA will be NULL - see page_zone.
6376 if (!node_online(page_to_nid(page)))
6379 zone = page_zone(page);
6380 pfn = page_to_pfn(page);
6381 if (!zone_spans_pfn(zone, pfn))
6384 return !has_unmovable_pages(zone, page, 0, true);
6389 static unsigned long pfn_max_align_down(unsigned long pfn)
6391 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6392 pageblock_nr_pages) - 1);
6395 static unsigned long pfn_max_align_up(unsigned long pfn)
6397 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6398 pageblock_nr_pages));
6401 /* [start, end) must belong to a single zone. */
6402 static int __alloc_contig_migrate_range(struct compact_control *cc,
6403 unsigned long start, unsigned long end)
6405 /* This function is based on compact_zone() from compaction.c. */
6406 unsigned long nr_reclaimed;
6407 unsigned long pfn = start;
6408 unsigned int tries = 0;
6413 while (pfn < end || !list_empty(&cc->migratepages)) {
6414 if (fatal_signal_pending(current)) {
6419 if (list_empty(&cc->migratepages)) {
6420 cc->nr_migratepages = 0;
6421 pfn = isolate_migratepages_range(cc, pfn, end);
6427 } else if (++tries == 5) {
6428 ret = ret < 0 ? ret : -EBUSY;
6432 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6434 cc->nr_migratepages -= nr_reclaimed;
6436 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6437 NULL, 0, cc->mode, MR_CMA);
6440 putback_movable_pages(&cc->migratepages);
6447 * alloc_contig_range() -- tries to allocate given range of pages
6448 * @start: start PFN to allocate
6449 * @end: one-past-the-last PFN to allocate
6450 * @migratetype: migratetype of the underlaying pageblocks (either
6451 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6452 * in range must have the same migratetype and it must
6453 * be either of the two.
6455 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6456 * aligned, however it's the caller's responsibility to guarantee that
6457 * we are the only thread that changes migrate type of pageblocks the
6460 * The PFN range must belong to a single zone.
6462 * Returns zero on success or negative error code. On success all
6463 * pages which PFN is in [start, end) are allocated for the caller and
6464 * need to be freed with free_contig_range().
6466 int alloc_contig_range(unsigned long start, unsigned long end,
6467 unsigned migratetype)
6469 unsigned long outer_start, outer_end;
6472 struct compact_control cc = {
6473 .nr_migratepages = 0,
6475 .zone = page_zone(pfn_to_page(start)),
6476 .mode = MIGRATE_SYNC,
6477 .ignore_skip_hint = true,
6479 INIT_LIST_HEAD(&cc.migratepages);
6482 * What we do here is we mark all pageblocks in range as
6483 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6484 * have different sizes, and due to the way page allocator
6485 * work, we align the range to biggest of the two pages so
6486 * that page allocator won't try to merge buddies from
6487 * different pageblocks and change MIGRATE_ISOLATE to some
6488 * other migration type.
6490 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6491 * migrate the pages from an unaligned range (ie. pages that
6492 * we are interested in). This will put all the pages in
6493 * range back to page allocator as MIGRATE_ISOLATE.
6495 * When this is done, we take the pages in range from page
6496 * allocator removing them from the buddy system. This way
6497 * page allocator will never consider using them.
6499 * This lets us mark the pageblocks back as
6500 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6501 * aligned range but not in the unaligned, original range are
6502 * put back to page allocator so that buddy can use them.
6505 ret = start_isolate_page_range(pfn_max_align_down(start),
6506 pfn_max_align_up(end), migratetype,
6511 ret = __alloc_contig_migrate_range(&cc, start, end);
6516 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6517 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6518 * more, all pages in [start, end) are free in page allocator.
6519 * What we are going to do is to allocate all pages from
6520 * [start, end) (that is remove them from page allocator).
6522 * The only problem is that pages at the beginning and at the
6523 * end of interesting range may be not aligned with pages that
6524 * page allocator holds, ie. they can be part of higher order
6525 * pages. Because of this, we reserve the bigger range and
6526 * once this is done free the pages we are not interested in.
6528 * We don't have to hold zone->lock here because the pages are
6529 * isolated thus they won't get removed from buddy.
6532 lru_add_drain_all();
6533 drain_all_pages(cc.zone);
6536 outer_start = start;
6537 while (!PageBuddy(pfn_to_page(outer_start))) {
6538 if (++order >= MAX_ORDER) {
6542 outer_start &= ~0UL << order;
6545 /* Make sure the range is really isolated. */
6546 if (test_pages_isolated(outer_start, end, false)) {
6547 pr_info("%s: [%lx, %lx) PFNs busy\n",
6548 __func__, outer_start, end);
6553 /* Grab isolated pages from freelists. */
6554 outer_end = isolate_freepages_range(&cc, outer_start, end);
6560 /* Free head and tail (if any) */
6561 if (start != outer_start)
6562 free_contig_range(outer_start, start - outer_start);
6563 if (end != outer_end)
6564 free_contig_range(end, outer_end - end);
6567 undo_isolate_page_range(pfn_max_align_down(start),
6568 pfn_max_align_up(end), migratetype);
6572 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6574 unsigned int count = 0;
6576 for (; nr_pages--; pfn++) {
6577 struct page *page = pfn_to_page(pfn);
6579 count += page_count(page) != 1;
6582 WARN(count != 0, "%d pages are still in use!\n", count);
6586 #ifdef CONFIG_MEMORY_HOTPLUG
6588 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6589 * page high values need to be recalulated.
6591 void __meminit zone_pcp_update(struct zone *zone)
6594 mutex_lock(&pcp_batch_high_lock);
6595 for_each_possible_cpu(cpu)
6596 pageset_set_high_and_batch(zone,
6597 per_cpu_ptr(zone->pageset, cpu));
6598 mutex_unlock(&pcp_batch_high_lock);
6602 void zone_pcp_reset(struct zone *zone)
6604 unsigned long flags;
6606 struct per_cpu_pageset *pset;
6608 /* avoid races with drain_pages() */
6609 local_lock_irqsave(pa_lock, flags);
6610 if (zone->pageset != &boot_pageset) {
6611 for_each_online_cpu(cpu) {
6612 pset = per_cpu_ptr(zone->pageset, cpu);
6613 drain_zonestat(zone, pset);
6615 free_percpu(zone->pageset);
6616 zone->pageset = &boot_pageset;
6618 local_unlock_irqrestore(pa_lock, flags);
6621 #ifdef CONFIG_MEMORY_HOTREMOVE
6623 * All pages in the range must be isolated before calling this.
6626 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6630 unsigned int order, i;
6632 unsigned long flags;
6633 /* find the first valid pfn */
6634 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6639 zone = page_zone(pfn_to_page(pfn));
6640 spin_lock_irqsave(&zone->lock, flags);
6642 while (pfn < end_pfn) {
6643 if (!pfn_valid(pfn)) {
6647 page = pfn_to_page(pfn);
6649 * The HWPoisoned page may be not in buddy system, and
6650 * page_count() is not 0.
6652 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6654 SetPageReserved(page);
6658 BUG_ON(page_count(page));
6659 BUG_ON(!PageBuddy(page));
6660 order = page_order(page);
6661 #ifdef CONFIG_DEBUG_VM
6662 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6663 pfn, 1 << order, end_pfn);
6665 list_del(&page->lru);
6666 rmv_page_order(page);
6667 zone->free_area[order].nr_free--;
6668 for (i = 0; i < (1 << order); i++)
6669 SetPageReserved((page+i));
6670 pfn += (1 << order);
6672 spin_unlock_irqrestore(&zone->lock, flags);
6676 #ifdef CONFIG_MEMORY_FAILURE
6677 bool is_free_buddy_page(struct page *page)
6679 struct zone *zone = page_zone(page);
6680 unsigned long pfn = page_to_pfn(page);
6681 unsigned long flags;
6684 spin_lock_irqsave(&zone->lock, flags);
6685 for (order = 0; order < MAX_ORDER; order++) {
6686 struct page *page_head = page - (pfn & ((1 << order) - 1));
6688 if (PageBuddy(page_head) && page_order(page_head) >= order)
6691 spin_unlock_irqrestore(&zone->lock, flags);
6693 return order < MAX_ORDER;