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/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
62 #include <asm/tlbflush.h>
63 #include <asm/div64.h>
66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67 DEFINE_PER_CPU(int, numa_node);
68 EXPORT_PER_CPU_SYMBOL(numa_node);
71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
78 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
83 * Array of node states.
85 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86 [N_POSSIBLE] = NODE_MASK_ALL,
87 [N_ONLINE] = { { [0] = 1UL } },
89 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
91 [N_HIGH_MEMORY] = { { [0] = 1UL } },
93 [N_CPU] = { { [0] = 1UL } },
96 EXPORT_SYMBOL(node_states);
98 unsigned long totalram_pages __read_mostly;
99 unsigned long totalreserve_pages __read_mostly;
101 * When calculating the number of globally allowed dirty pages, there
102 * is a certain number of per-zone reserves that should not be
103 * considered dirtyable memory. This is the sum of those reserves
104 * over all existing zones that contribute dirtyable memory.
106 unsigned long dirty_balance_reserve __read_mostly;
108 int percpu_pagelist_fraction;
109 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
111 #ifdef CONFIG_PM_SLEEP
113 * The following functions are used by the suspend/hibernate code to temporarily
114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
115 * while devices are suspended. To avoid races with the suspend/hibernate code,
116 * they should always be called with pm_mutex held (gfp_allowed_mask also should
117 * only be modified with pm_mutex held, unless the suspend/hibernate code is
118 * guaranteed not to run in parallel with that modification).
121 static gfp_t saved_gfp_mask;
123 void pm_restore_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 if (saved_gfp_mask) {
127 gfp_allowed_mask = saved_gfp_mask;
132 void pm_restrict_gfp_mask(void)
134 WARN_ON(!mutex_is_locked(&pm_mutex));
135 WARN_ON(saved_gfp_mask);
136 saved_gfp_mask = gfp_allowed_mask;
137 gfp_allowed_mask &= ~GFP_IOFS;
140 bool pm_suspended_storage(void)
142 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
146 #endif /* CONFIG_PM_SLEEP */
148 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
149 int pageblock_order __read_mostly;
152 static void __free_pages_ok(struct page *page, unsigned int order);
155 * results with 256, 32 in the lowmem_reserve sysctl:
156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
157 * 1G machine -> (16M dma, 784M normal, 224M high)
158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
162 * TBD: should special case ZONE_DMA32 machines here - in those we normally
163 * don't need any ZONE_NORMAL reservation
165 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
166 #ifdef CONFIG_ZONE_DMA
169 #ifdef CONFIG_ZONE_DMA32
172 #ifdef CONFIG_HIGHMEM
178 EXPORT_SYMBOL(totalram_pages);
180 static char * const zone_names[MAX_NR_ZONES] = {
181 #ifdef CONFIG_ZONE_DMA
184 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 int min_free_kbytes = 1024;
196 static unsigned long __meminitdata nr_kernel_pages;
197 static unsigned long __meminitdata nr_all_pages;
198 static unsigned long __meminitdata dma_reserve;
200 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
201 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
202 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
203 static unsigned long __initdata required_kernelcore;
204 static unsigned long __initdata required_movablecore;
205 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
207 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
209 EXPORT_SYMBOL(movable_zone);
210 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
213 int nr_node_ids __read_mostly = MAX_NUMNODES;
214 int nr_online_nodes __read_mostly = 1;
215 EXPORT_SYMBOL(nr_node_ids);
216 EXPORT_SYMBOL(nr_online_nodes);
219 int page_group_by_mobility_disabled __read_mostly;
223 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
224 * Instead, use {un}set_pageblock_isolate.
226 void set_pageblock_migratetype(struct page *page, int migratetype)
229 if (unlikely(page_group_by_mobility_disabled))
230 migratetype = MIGRATE_UNMOVABLE;
232 set_pageblock_flags_group(page, (unsigned long)migratetype,
233 PB_migrate, PB_migrate_end);
236 bool oom_killer_disabled __read_mostly;
238 #ifdef CONFIG_DEBUG_VM
239 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
243 unsigned long pfn = page_to_pfn(page);
246 seq = zone_span_seqbegin(zone);
247 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
249 else if (pfn < zone->zone_start_pfn)
251 } while (zone_span_seqretry(zone, seq));
256 static int page_is_consistent(struct zone *zone, struct page *page)
258 if (!pfn_valid_within(page_to_pfn(page)))
260 if (zone != page_zone(page))
266 * Temporary debugging check for pages not lying within a given zone.
268 static int bad_range(struct zone *zone, struct page *page)
270 if (page_outside_zone_boundaries(zone, page))
272 if (!page_is_consistent(zone, page))
278 static inline int bad_range(struct zone *zone, struct page *page)
284 static void bad_page(struct page *page)
286 static unsigned long resume;
287 static unsigned long nr_shown;
288 static unsigned long nr_unshown;
290 /* Don't complain about poisoned pages */
291 if (PageHWPoison(page)) {
292 reset_page_mapcount(page); /* remove PageBuddy */
297 * Allow a burst of 60 reports, then keep quiet for that minute;
298 * or allow a steady drip of one report per second.
300 if (nr_shown == 60) {
301 if (time_before(jiffies, resume)) {
307 "BUG: Bad page state: %lu messages suppressed\n",
314 resume = jiffies + 60 * HZ;
316 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
317 current->comm, page_to_pfn(page));
323 /* Leave bad fields for debug, except PageBuddy could make trouble */
324 reset_page_mapcount(page); /* remove PageBuddy */
325 add_taint(TAINT_BAD_PAGE);
329 * Higher-order pages are called "compound pages". They are structured thusly:
331 * The first PAGE_SIZE page is called the "head page".
333 * The remaining PAGE_SIZE pages are called "tail pages".
335 * All pages have PG_compound set. All tail pages have their ->first_page
336 * pointing at the head page.
338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
343 static void free_compound_page(struct page *page)
345 __free_pages_ok(page, compound_order(page));
348 void prep_compound_page(struct page *page, unsigned long order)
351 int nr_pages = 1 << order;
353 set_compound_page_dtor(page, free_compound_page);
354 set_compound_order(page, order);
356 for (i = 1; i < nr_pages; i++) {
357 struct page *p = page + i;
359 set_page_count(p, 0);
360 p->first_page = page;
364 /* update __split_huge_page_refcount if you change this function */
365 static int destroy_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
371 if (unlikely(compound_order(page) != order) ||
372 unlikely(!PageHead(page))) {
377 __ClearPageHead(page);
379 for (i = 1; i < nr_pages; i++) {
380 struct page *p = page + i;
382 if (unlikely(!PageTail(p) || (p->first_page != page))) {
392 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
397 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
398 * and __GFP_HIGHMEM from hard or soft interrupt context.
400 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
401 for (i = 0; i < (1 << order); i++)
402 clear_highpage(page + i);
405 #ifdef CONFIG_DEBUG_PAGEALLOC
406 unsigned int _debug_guardpage_minorder;
408 static int __init debug_guardpage_minorder_setup(char *buf)
412 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
413 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
416 _debug_guardpage_minorder = res;
417 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
420 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
422 static inline void set_page_guard_flag(struct page *page)
424 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
427 static inline void clear_page_guard_flag(struct page *page)
429 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
432 static inline void set_page_guard_flag(struct page *page) { }
433 static inline void clear_page_guard_flag(struct page *page) { }
436 static inline void set_page_order(struct page *page, int order)
438 set_page_private(page, order);
439 __SetPageBuddy(page);
442 static inline void rmv_page_order(struct page *page)
444 __ClearPageBuddy(page);
445 set_page_private(page, 0);
449 * Locate the struct page for both the matching buddy in our
450 * pair (buddy1) and the combined O(n+1) page they form (page).
452 * 1) Any buddy B1 will have an order O twin B2 which satisfies
453 * the following equation:
455 * For example, if the starting buddy (buddy2) is #8 its order
457 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
459 * 2) Any buddy B will have an order O+1 parent P which
460 * satisfies the following equation:
463 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
465 static inline unsigned long
466 __find_buddy_index(unsigned long page_idx, unsigned int order)
468 return page_idx ^ (1 << order);
472 * This function checks whether a page is free && is the buddy
473 * we can do coalesce a page and its buddy if
474 * (a) the buddy is not in a hole &&
475 * (b) the buddy is in the buddy system &&
476 * (c) a page and its buddy have the same order &&
477 * (d) a page and its buddy are in the same zone.
479 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
480 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
482 * For recording page's order, we use page_private(page).
484 static inline int page_is_buddy(struct page *page, struct page *buddy,
487 if (!pfn_valid_within(page_to_pfn(buddy)))
490 if (page_zone_id(page) != page_zone_id(buddy))
493 if (page_is_guard(buddy) && page_order(buddy) == order) {
494 VM_BUG_ON(page_count(buddy) != 0);
498 if (PageBuddy(buddy) && page_order(buddy) == order) {
499 VM_BUG_ON(page_count(buddy) != 0);
506 * Freeing function for a buddy system allocator.
508 * The concept of a buddy system is to maintain direct-mapped table
509 * (containing bit values) for memory blocks of various "orders".
510 * The bottom level table contains the map for the smallest allocatable
511 * units of memory (here, pages), and each level above it describes
512 * pairs of units from the levels below, hence, "buddies".
513 * At a high level, all that happens here is marking the table entry
514 * at the bottom level available, and propagating the changes upward
515 * as necessary, plus some accounting needed to play nicely with other
516 * parts of the VM system.
517 * At each level, we keep a list of pages, which are heads of continuous
518 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
519 * order is recorded in page_private(page) field.
520 * So when we are allocating or freeing one, we can derive the state of the
521 * other. That is, if we allocate a small block, and both were
522 * free, the remainder of the region must be split into blocks.
523 * If a block is freed, and its buddy is also free, then this
524 * triggers coalescing into a block of larger size.
529 static inline void __free_one_page(struct page *page,
530 struct zone *zone, unsigned int order,
533 unsigned long page_idx;
534 unsigned long combined_idx;
535 unsigned long uninitialized_var(buddy_idx);
538 if (unlikely(PageCompound(page)))
539 if (unlikely(destroy_compound_page(page, order)))
542 VM_BUG_ON(migratetype == -1);
544 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
546 VM_BUG_ON(page_idx & ((1 << order) - 1));
547 VM_BUG_ON(bad_range(zone, page));
549 while (order < MAX_ORDER-1) {
550 buddy_idx = __find_buddy_index(page_idx, order);
551 buddy = page + (buddy_idx - page_idx);
552 if (!page_is_buddy(page, buddy, order))
555 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
556 * merge with it and move up one order.
558 if (page_is_guard(buddy)) {
559 clear_page_guard_flag(buddy);
560 set_page_private(page, 0);
561 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
563 list_del(&buddy->lru);
564 zone->free_area[order].nr_free--;
565 rmv_page_order(buddy);
567 combined_idx = buddy_idx & page_idx;
568 page = page + (combined_idx - page_idx);
569 page_idx = combined_idx;
572 set_page_order(page, order);
575 * If this is not the largest possible page, check if the buddy
576 * of the next-highest order is free. If it is, it's possible
577 * that pages are being freed that will coalesce soon. In case,
578 * that is happening, add the free page to the tail of the list
579 * so it's less likely to be used soon and more likely to be merged
580 * as a higher order page
582 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
583 struct page *higher_page, *higher_buddy;
584 combined_idx = buddy_idx & page_idx;
585 higher_page = page + (combined_idx - page_idx);
586 buddy_idx = __find_buddy_index(combined_idx, order + 1);
587 higher_buddy = page + (buddy_idx - combined_idx);
588 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
589 list_add_tail(&page->lru,
590 &zone->free_area[order].free_list[migratetype]);
595 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
597 zone->free_area[order].nr_free++;
601 * free_page_mlock() -- clean up attempts to free and mlocked() page.
602 * Page should not be on lru, so no need to fix that up.
603 * free_pages_check() will verify...
605 static inline void free_page_mlock(struct page *page)
607 __dec_zone_page_state(page, NR_MLOCK);
608 __count_vm_event(UNEVICTABLE_MLOCKFREED);
611 static inline int free_pages_check(struct page *page)
613 if (unlikely(page_mapcount(page) |
614 (page->mapping != NULL) |
615 (atomic_read(&page->_count) != 0) |
616 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
617 (mem_cgroup_bad_page_check(page)))) {
621 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
622 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
627 * Frees a number of pages from the PCP lists
628 * Assumes all pages on list are in same zone, and of same order.
629 * count is the number of pages to free.
631 * If the zone was previously in an "all pages pinned" state then look to
632 * see if this freeing clears that state.
634 * And clear the zone's pages_scanned counter, to hold off the "all pages are
635 * pinned" detection logic.
637 static void free_pcppages_bulk(struct zone *zone, int count,
638 struct per_cpu_pages *pcp)
644 spin_lock(&zone->lock);
645 zone->all_unreclaimable = 0;
646 zone->pages_scanned = 0;
650 struct list_head *list;
653 * Remove pages from lists in a round-robin fashion. A
654 * batch_free count is maintained that is incremented when an
655 * empty list is encountered. This is so more pages are freed
656 * off fuller lists instead of spinning excessively around empty
661 if (++migratetype == MIGRATE_PCPTYPES)
663 list = &pcp->lists[migratetype];
664 } while (list_empty(list));
666 /* This is the only non-empty list. Free them all. */
667 if (batch_free == MIGRATE_PCPTYPES)
668 batch_free = to_free;
671 page = list_entry(list->prev, struct page, lru);
672 /* must delete as __free_one_page list manipulates */
673 list_del(&page->lru);
674 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
675 __free_one_page(page, zone, 0, page_private(page));
676 trace_mm_page_pcpu_drain(page, 0, page_private(page));
677 } while (--to_free && --batch_free && !list_empty(list));
679 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
680 spin_unlock(&zone->lock);
683 static void free_one_page(struct zone *zone, struct page *page, int order,
686 spin_lock(&zone->lock);
687 zone->all_unreclaimable = 0;
688 zone->pages_scanned = 0;
690 __free_one_page(page, zone, order, migratetype);
691 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
692 spin_unlock(&zone->lock);
695 static bool free_pages_prepare(struct page *page, unsigned int order)
700 trace_mm_page_free(page, order);
701 kmemcheck_free_shadow(page, order);
704 page->mapping = NULL;
705 for (i = 0; i < (1 << order); i++)
706 bad += free_pages_check(page + i);
710 if (!PageHighMem(page)) {
711 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
712 debug_check_no_obj_freed(page_address(page),
715 arch_free_page(page, order);
716 kernel_map_pages(page, 1 << order, 0);
721 static void __free_pages_ok(struct page *page, unsigned int order)
724 int wasMlocked = __TestClearPageMlocked(page);
726 if (!free_pages_prepare(page, order))
729 local_irq_save(flags);
730 if (unlikely(wasMlocked))
731 free_page_mlock(page);
732 __count_vm_events(PGFREE, 1 << order);
733 free_one_page(page_zone(page), page, order,
734 get_pageblock_migratetype(page));
735 local_irq_restore(flags);
738 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
740 unsigned int nr_pages = 1 << order;
744 for (loop = 0; loop < nr_pages; loop++) {
745 struct page *p = &page[loop];
747 if (loop + 1 < nr_pages)
749 __ClearPageReserved(p);
750 set_page_count(p, 0);
753 set_page_refcounted(page);
754 __free_pages(page, order);
758 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
759 void __init init_cma_reserved_pageblock(struct page *page)
761 unsigned i = pageblock_nr_pages;
762 struct page *p = page;
765 __ClearPageReserved(p);
766 set_page_count(p, 0);
769 set_page_refcounted(page);
770 set_pageblock_migratetype(page, MIGRATE_CMA);
771 __free_pages(page, pageblock_order);
772 totalram_pages += pageblock_nr_pages;
777 * The order of subdivision here is critical for the IO subsystem.
778 * Please do not alter this order without good reasons and regression
779 * testing. Specifically, as large blocks of memory are subdivided,
780 * the order in which smaller blocks are delivered depends on the order
781 * they're subdivided in this function. This is the primary factor
782 * influencing the order in which pages are delivered to the IO
783 * subsystem according to empirical testing, and this is also justified
784 * by considering the behavior of a buddy system containing a single
785 * large block of memory acted on by a series of small allocations.
786 * This behavior is a critical factor in sglist merging's success.
790 static inline void expand(struct zone *zone, struct page *page,
791 int low, int high, struct free_area *area,
794 unsigned long size = 1 << high;
800 VM_BUG_ON(bad_range(zone, &page[size]));
802 #ifdef CONFIG_DEBUG_PAGEALLOC
803 if (high < debug_guardpage_minorder()) {
805 * Mark as guard pages (or page), that will allow to
806 * merge back to allocator when buddy will be freed.
807 * Corresponding page table entries will not be touched,
808 * pages will stay not present in virtual address space
810 INIT_LIST_HEAD(&page[size].lru);
811 set_page_guard_flag(&page[size]);
812 set_page_private(&page[size], high);
813 /* Guard pages are not available for any usage */
814 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high));
818 list_add(&page[size].lru, &area->free_list[migratetype]);
820 set_page_order(&page[size], high);
825 * This page is about to be returned from the page allocator
827 static inline int check_new_page(struct page *page)
829 if (unlikely(page_mapcount(page) |
830 (page->mapping != NULL) |
831 (atomic_read(&page->_count) != 0) |
832 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
833 (mem_cgroup_bad_page_check(page)))) {
840 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
844 for (i = 0; i < (1 << order); i++) {
845 struct page *p = page + i;
846 if (unlikely(check_new_page(p)))
850 set_page_private(page, 0);
851 set_page_refcounted(page);
853 arch_alloc_page(page, order);
854 kernel_map_pages(page, 1 << order, 1);
856 if (gfp_flags & __GFP_ZERO)
857 prep_zero_page(page, order, gfp_flags);
859 if (order && (gfp_flags & __GFP_COMP))
860 prep_compound_page(page, order);
866 * Go through the free lists for the given migratetype and remove
867 * the smallest available page from the freelists
870 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
873 unsigned int current_order;
874 struct free_area * area;
877 /* Find a page of the appropriate size in the preferred list */
878 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
879 area = &(zone->free_area[current_order]);
880 if (list_empty(&area->free_list[migratetype]))
883 page = list_entry(area->free_list[migratetype].next,
885 list_del(&page->lru);
886 rmv_page_order(page);
888 expand(zone, page, order, current_order, area, migratetype);
897 * This array describes the order lists are fallen back to when
898 * the free lists for the desirable migrate type are depleted
900 static int fallbacks[MIGRATE_TYPES][4] = {
901 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
902 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
904 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
905 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
907 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
909 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
910 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
914 * Move the free pages in a range to the free lists of the requested type.
915 * Note that start_page and end_pages are not aligned on a pageblock
916 * boundary. If alignment is required, use move_freepages_block()
918 static int move_freepages(struct zone *zone,
919 struct page *start_page, struct page *end_page,
926 #ifndef CONFIG_HOLES_IN_ZONE
928 * page_zone is not safe to call in this context when
929 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
930 * anyway as we check zone boundaries in move_freepages_block().
931 * Remove at a later date when no bug reports exist related to
932 * grouping pages by mobility
934 BUG_ON(page_zone(start_page) != page_zone(end_page));
937 for (page = start_page; page <= end_page;) {
938 /* Make sure we are not inadvertently changing nodes */
939 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
941 if (!pfn_valid_within(page_to_pfn(page))) {
946 if (!PageBuddy(page)) {
951 order = page_order(page);
952 list_move(&page->lru,
953 &zone->free_area[order].free_list[migratetype]);
955 pages_moved += 1 << order;
961 int move_freepages_block(struct zone *zone, struct page *page,
964 unsigned long start_pfn, end_pfn;
965 struct page *start_page, *end_page;
967 start_pfn = page_to_pfn(page);
968 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
969 start_page = pfn_to_page(start_pfn);
970 end_page = start_page + pageblock_nr_pages - 1;
971 end_pfn = start_pfn + pageblock_nr_pages - 1;
973 /* Do not cross zone boundaries */
974 if (start_pfn < zone->zone_start_pfn)
976 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
979 return move_freepages(zone, start_page, end_page, migratetype);
982 static void change_pageblock_range(struct page *pageblock_page,
983 int start_order, int migratetype)
985 int nr_pageblocks = 1 << (start_order - pageblock_order);
987 while (nr_pageblocks--) {
988 set_pageblock_migratetype(pageblock_page, migratetype);
989 pageblock_page += pageblock_nr_pages;
993 /* Remove an element from the buddy allocator from the fallback list */
994 static inline struct page *
995 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
997 struct free_area * area;
1002 /* Find the largest possible block of pages in the other list */
1003 for (current_order = MAX_ORDER-1; current_order >= order;
1006 migratetype = fallbacks[start_migratetype][i];
1008 /* MIGRATE_RESERVE handled later if necessary */
1009 if (migratetype == MIGRATE_RESERVE)
1012 area = &(zone->free_area[current_order]);
1013 if (list_empty(&area->free_list[migratetype]))
1016 page = list_entry(area->free_list[migratetype].next,
1021 * If breaking a large block of pages, move all free
1022 * pages to the preferred allocation list. If falling
1023 * back for a reclaimable kernel allocation, be more
1024 * aggressive about taking ownership of free pages
1026 * On the other hand, never change migration
1027 * type of MIGRATE_CMA pageblocks nor move CMA
1028 * pages on different free lists. We don't
1029 * want unmovable pages to be allocated from
1030 * MIGRATE_CMA areas.
1032 if (!is_migrate_cma(migratetype) &&
1033 (unlikely(current_order >= pageblock_order / 2) ||
1034 start_migratetype == MIGRATE_RECLAIMABLE ||
1035 page_group_by_mobility_disabled)) {
1037 pages = move_freepages_block(zone, page,
1040 /* Claim the whole block if over half of it is free */
1041 if (pages >= (1 << (pageblock_order-1)) ||
1042 page_group_by_mobility_disabled)
1043 set_pageblock_migratetype(page,
1046 migratetype = start_migratetype;
1049 /* Remove the page from the freelists */
1050 list_del(&page->lru);
1051 rmv_page_order(page);
1053 /* Take ownership for orders >= pageblock_order */
1054 if (current_order >= pageblock_order &&
1055 !is_migrate_cma(migratetype))
1056 change_pageblock_range(page, current_order,
1059 expand(zone, page, order, current_order, area,
1060 is_migrate_cma(migratetype)
1061 ? migratetype : start_migratetype);
1063 trace_mm_page_alloc_extfrag(page, order, current_order,
1064 start_migratetype, migratetype);
1074 * Do the hard work of removing an element from the buddy allocator.
1075 * Call me with the zone->lock already held.
1077 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1083 page = __rmqueue_smallest(zone, order, migratetype);
1085 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1086 page = __rmqueue_fallback(zone, order, migratetype);
1089 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1090 * is used because __rmqueue_smallest is an inline function
1091 * and we want just one call site
1094 migratetype = MIGRATE_RESERVE;
1099 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1104 * Obtain a specified number of elements from the buddy allocator, all under
1105 * a single hold of the lock, for efficiency. Add them to the supplied list.
1106 * Returns the number of new pages which were placed at *list.
1108 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1109 unsigned long count, struct list_head *list,
1110 int migratetype, int cold)
1112 int mt = migratetype, i;
1114 spin_lock(&zone->lock);
1115 for (i = 0; i < count; ++i) {
1116 struct page *page = __rmqueue(zone, order, migratetype);
1117 if (unlikely(page == NULL))
1121 * Split buddy pages returned by expand() are received here
1122 * in physical page order. The page is added to the callers and
1123 * list and the list head then moves forward. From the callers
1124 * perspective, the linked list is ordered by page number in
1125 * some conditions. This is useful for IO devices that can
1126 * merge IO requests if the physical pages are ordered
1129 if (likely(cold == 0))
1130 list_add(&page->lru, list);
1132 list_add_tail(&page->lru, list);
1133 if (IS_ENABLED(CONFIG_CMA)) {
1134 mt = get_pageblock_migratetype(page);
1135 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1138 set_page_private(page, mt);
1141 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1142 spin_unlock(&zone->lock);
1148 * Called from the vmstat counter updater to drain pagesets of this
1149 * currently executing processor on remote nodes after they have
1152 * Note that this function must be called with the thread pinned to
1153 * a single processor.
1155 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1157 unsigned long flags;
1160 local_irq_save(flags);
1161 if (pcp->count >= pcp->batch)
1162 to_drain = pcp->batch;
1164 to_drain = pcp->count;
1166 free_pcppages_bulk(zone, to_drain, pcp);
1167 pcp->count -= to_drain;
1169 local_irq_restore(flags);
1174 * Drain pages of the indicated processor.
1176 * The processor must either be the current processor and the
1177 * thread pinned to the current processor or a processor that
1180 static void drain_pages(unsigned int cpu)
1182 unsigned long flags;
1185 for_each_populated_zone(zone) {
1186 struct per_cpu_pageset *pset;
1187 struct per_cpu_pages *pcp;
1189 local_irq_save(flags);
1190 pset = per_cpu_ptr(zone->pageset, cpu);
1194 free_pcppages_bulk(zone, pcp->count, pcp);
1197 local_irq_restore(flags);
1202 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1204 void drain_local_pages(void *arg)
1206 drain_pages(smp_processor_id());
1210 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1212 * Note that this code is protected against sending an IPI to an offline
1213 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1214 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1215 * nothing keeps CPUs from showing up after we populated the cpumask and
1216 * before the call to on_each_cpu_mask().
1218 void drain_all_pages(void)
1221 struct per_cpu_pageset *pcp;
1225 * Allocate in the BSS so we wont require allocation in
1226 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1228 static cpumask_t cpus_with_pcps;
1231 * We don't care about racing with CPU hotplug event
1232 * as offline notification will cause the notified
1233 * cpu to drain that CPU pcps and on_each_cpu_mask
1234 * disables preemption as part of its processing
1236 for_each_online_cpu(cpu) {
1237 bool has_pcps = false;
1238 for_each_populated_zone(zone) {
1239 pcp = per_cpu_ptr(zone->pageset, cpu);
1240 if (pcp->pcp.count) {
1246 cpumask_set_cpu(cpu, &cpus_with_pcps);
1248 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1250 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1253 #ifdef CONFIG_HIBERNATION
1255 void mark_free_pages(struct zone *zone)
1257 unsigned long pfn, max_zone_pfn;
1258 unsigned long flags;
1260 struct list_head *curr;
1262 if (!zone->spanned_pages)
1265 spin_lock_irqsave(&zone->lock, flags);
1267 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1268 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1269 if (pfn_valid(pfn)) {
1270 struct page *page = pfn_to_page(pfn);
1272 if (!swsusp_page_is_forbidden(page))
1273 swsusp_unset_page_free(page);
1276 for_each_migratetype_order(order, t) {
1277 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1280 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1281 for (i = 0; i < (1UL << order); i++)
1282 swsusp_set_page_free(pfn_to_page(pfn + i));
1285 spin_unlock_irqrestore(&zone->lock, flags);
1287 #endif /* CONFIG_PM */
1290 * Free a 0-order page
1291 * cold == 1 ? free a cold page : free a hot page
1293 void free_hot_cold_page(struct page *page, int cold)
1295 struct zone *zone = page_zone(page);
1296 struct per_cpu_pages *pcp;
1297 unsigned long flags;
1299 int wasMlocked = __TestClearPageMlocked(page);
1301 if (!free_pages_prepare(page, 0))
1304 migratetype = get_pageblock_migratetype(page);
1305 set_page_private(page, migratetype);
1306 local_irq_save(flags);
1307 if (unlikely(wasMlocked))
1308 free_page_mlock(page);
1309 __count_vm_event(PGFREE);
1312 * We only track unmovable, reclaimable and movable on pcp lists.
1313 * Free ISOLATE pages back to the allocator because they are being
1314 * offlined but treat RESERVE as movable pages so we can get those
1315 * areas back if necessary. Otherwise, we may have to free
1316 * excessively into the page allocator
1318 if (migratetype >= MIGRATE_PCPTYPES) {
1319 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1320 free_one_page(zone, page, 0, migratetype);
1323 migratetype = MIGRATE_MOVABLE;
1326 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1328 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1330 list_add(&page->lru, &pcp->lists[migratetype]);
1332 if (pcp->count >= pcp->high) {
1333 free_pcppages_bulk(zone, pcp->batch, pcp);
1334 pcp->count -= pcp->batch;
1338 local_irq_restore(flags);
1342 * Free a list of 0-order pages
1344 void free_hot_cold_page_list(struct list_head *list, int cold)
1346 struct page *page, *next;
1348 list_for_each_entry_safe(page, next, list, lru) {
1349 trace_mm_page_free_batched(page, cold);
1350 free_hot_cold_page(page, cold);
1355 * split_page takes a non-compound higher-order page, and splits it into
1356 * n (1<<order) sub-pages: page[0..n]
1357 * Each sub-page must be freed individually.
1359 * Note: this is probably too low level an operation for use in drivers.
1360 * Please consult with lkml before using this in your driver.
1362 void split_page(struct page *page, unsigned int order)
1366 VM_BUG_ON(PageCompound(page));
1367 VM_BUG_ON(!page_count(page));
1369 #ifdef CONFIG_KMEMCHECK
1371 * Split shadow pages too, because free(page[0]) would
1372 * otherwise free the whole shadow.
1374 if (kmemcheck_page_is_tracked(page))
1375 split_page(virt_to_page(page[0].shadow), order);
1378 for (i = 1; i < (1 << order); i++)
1379 set_page_refcounted(page + i);
1383 * Similar to split_page except the page is already free. As this is only
1384 * being used for migration, the migratetype of the block also changes.
1385 * As this is called with interrupts disabled, the caller is responsible
1386 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1389 * Note: this is probably too low level an operation for use in drivers.
1390 * Please consult with lkml before using this in your driver.
1392 int split_free_page(struct page *page)
1395 unsigned long watermark;
1398 BUG_ON(!PageBuddy(page));
1400 zone = page_zone(page);
1401 order = page_order(page);
1403 /* Obey watermarks as if the page was being allocated */
1404 watermark = low_wmark_pages(zone) + (1 << order);
1405 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1408 /* Remove page from free list */
1409 list_del(&page->lru);
1410 zone->free_area[order].nr_free--;
1411 rmv_page_order(page);
1412 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1414 /* Split into individual pages */
1415 set_page_refcounted(page);
1416 split_page(page, order);
1418 if (order >= pageblock_order - 1) {
1419 struct page *endpage = page + (1 << order) - 1;
1420 for (; page < endpage; page += pageblock_nr_pages) {
1421 int mt = get_pageblock_migratetype(page);
1422 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1423 set_pageblock_migratetype(page,
1432 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1433 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1437 struct page *buffered_rmqueue(struct zone *preferred_zone,
1438 struct zone *zone, int order, gfp_t gfp_flags,
1441 unsigned long flags;
1443 int cold = !!(gfp_flags & __GFP_COLD);
1446 if (likely(order == 0)) {
1447 struct per_cpu_pages *pcp;
1448 struct list_head *list;
1450 local_irq_save(flags);
1451 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1452 list = &pcp->lists[migratetype];
1453 if (list_empty(list)) {
1454 pcp->count += rmqueue_bulk(zone, 0,
1457 if (unlikely(list_empty(list)))
1462 page = list_entry(list->prev, struct page, lru);
1464 page = list_entry(list->next, struct page, lru);
1466 list_del(&page->lru);
1469 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1471 * __GFP_NOFAIL is not to be used in new code.
1473 * All __GFP_NOFAIL callers should be fixed so that they
1474 * properly detect and handle allocation failures.
1476 * We most definitely don't want callers attempting to
1477 * allocate greater than order-1 page units with
1480 WARN_ON_ONCE(order > 1);
1482 spin_lock_irqsave(&zone->lock, flags);
1483 page = __rmqueue(zone, order, migratetype);
1484 spin_unlock(&zone->lock);
1487 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1490 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1491 zone_statistics(preferred_zone, zone, gfp_flags);
1492 local_irq_restore(flags);
1494 VM_BUG_ON(bad_range(zone, page));
1495 if (prep_new_page(page, order, gfp_flags))
1500 local_irq_restore(flags);
1504 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1505 #define ALLOC_WMARK_MIN WMARK_MIN
1506 #define ALLOC_WMARK_LOW WMARK_LOW
1507 #define ALLOC_WMARK_HIGH WMARK_HIGH
1508 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1510 /* Mask to get the watermark bits */
1511 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1513 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1514 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1515 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1516 #define ALLOC_PFMEMALLOC 0x80 /* Caller has PF_MEMALLOC set */
1518 #ifdef CONFIG_FAIL_PAGE_ALLOC
1521 struct fault_attr attr;
1523 u32 ignore_gfp_highmem;
1524 u32 ignore_gfp_wait;
1526 } fail_page_alloc = {
1527 .attr = FAULT_ATTR_INITIALIZER,
1528 .ignore_gfp_wait = 1,
1529 .ignore_gfp_highmem = 1,
1533 static int __init setup_fail_page_alloc(char *str)
1535 return setup_fault_attr(&fail_page_alloc.attr, str);
1537 __setup("fail_page_alloc=", setup_fail_page_alloc);
1539 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1541 if (order < fail_page_alloc.min_order)
1543 if (gfp_mask & __GFP_NOFAIL)
1545 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1547 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1550 return should_fail(&fail_page_alloc.attr, 1 << order);
1553 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1555 static int __init fail_page_alloc_debugfs(void)
1557 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1560 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1561 &fail_page_alloc.attr);
1563 return PTR_ERR(dir);
1565 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1566 &fail_page_alloc.ignore_gfp_wait))
1568 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1569 &fail_page_alloc.ignore_gfp_highmem))
1571 if (!debugfs_create_u32("min-order", mode, dir,
1572 &fail_page_alloc.min_order))
1577 debugfs_remove_recursive(dir);
1582 late_initcall(fail_page_alloc_debugfs);
1584 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1586 #else /* CONFIG_FAIL_PAGE_ALLOC */
1588 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1593 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1596 * Return true if free pages are above 'mark'. This takes into account the order
1597 * of the allocation.
1599 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1600 int classzone_idx, int alloc_flags, long free_pages)
1602 /* free_pages my go negative - that's OK */
1604 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1607 free_pages -= (1 << order) - 1;
1608 if (alloc_flags & ALLOC_HIGH)
1610 if (alloc_flags & ALLOC_HARDER)
1613 if (free_pages <= min + lowmem_reserve)
1615 for (o = 0; o < order; o++) {
1616 /* At the next order, this order's pages become unavailable */
1617 free_pages -= z->free_area[o].nr_free << o;
1619 /* Require fewer higher order pages to be free */
1622 if (free_pages <= min)
1628 #ifdef CONFIG_MEMORY_ISOLATION
1629 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1631 if (unlikely(zone->nr_pageblock_isolate))
1632 return zone->nr_pageblock_isolate * pageblock_nr_pages;
1636 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1642 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1643 int classzone_idx, int alloc_flags)
1645 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1646 zone_page_state(z, NR_FREE_PAGES));
1649 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1650 int classzone_idx, int alloc_flags)
1652 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1654 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1655 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1658 * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1659 * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1660 * sleep although it could do so. But this is more desirable for memory
1661 * hotplug than sleeping which can cause a livelock in the direct
1664 free_pages -= nr_zone_isolate_freepages(z);
1665 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1671 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1672 * skip over zones that are not allowed by the cpuset, or that have
1673 * been recently (in last second) found to be nearly full. See further
1674 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1675 * that have to skip over a lot of full or unallowed zones.
1677 * If the zonelist cache is present in the passed in zonelist, then
1678 * returns a pointer to the allowed node mask (either the current
1679 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1681 * If the zonelist cache is not available for this zonelist, does
1682 * nothing and returns NULL.
1684 * If the fullzones BITMAP in the zonelist cache is stale (more than
1685 * a second since last zap'd) then we zap it out (clear its bits.)
1687 * We hold off even calling zlc_setup, until after we've checked the
1688 * first zone in the zonelist, on the theory that most allocations will
1689 * be satisfied from that first zone, so best to examine that zone as
1690 * quickly as we can.
1692 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1694 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1695 nodemask_t *allowednodes; /* zonelist_cache approximation */
1697 zlc = zonelist->zlcache_ptr;
1701 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1702 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1703 zlc->last_full_zap = jiffies;
1706 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1707 &cpuset_current_mems_allowed :
1708 &node_states[N_HIGH_MEMORY];
1709 return allowednodes;
1713 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1714 * if it is worth looking at further for free memory:
1715 * 1) Check that the zone isn't thought to be full (doesn't have its
1716 * bit set in the zonelist_cache fullzones BITMAP).
1717 * 2) Check that the zones node (obtained from the zonelist_cache
1718 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1719 * Return true (non-zero) if zone is worth looking at further, or
1720 * else return false (zero) if it is not.
1722 * This check -ignores- the distinction between various watermarks,
1723 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1724 * found to be full for any variation of these watermarks, it will
1725 * be considered full for up to one second by all requests, unless
1726 * we are so low on memory on all allowed nodes that we are forced
1727 * into the second scan of the zonelist.
1729 * In the second scan we ignore this zonelist cache and exactly
1730 * apply the watermarks to all zones, even it is slower to do so.
1731 * We are low on memory in the second scan, and should leave no stone
1732 * unturned looking for a free page.
1734 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1735 nodemask_t *allowednodes)
1737 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1738 int i; /* index of *z in zonelist zones */
1739 int n; /* node that zone *z is on */
1741 zlc = zonelist->zlcache_ptr;
1745 i = z - zonelist->_zonerefs;
1748 /* This zone is worth trying if it is allowed but not full */
1749 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1753 * Given 'z' scanning a zonelist, set the corresponding bit in
1754 * zlc->fullzones, so that subsequent attempts to allocate a page
1755 * from that zone don't waste time re-examining it.
1757 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1759 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1760 int i; /* index of *z in zonelist zones */
1762 zlc = zonelist->zlcache_ptr;
1766 i = z - zonelist->_zonerefs;
1768 set_bit(i, zlc->fullzones);
1772 * clear all zones full, called after direct reclaim makes progress so that
1773 * a zone that was recently full is not skipped over for up to a second
1775 static void zlc_clear_zones_full(struct zonelist *zonelist)
1777 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1779 zlc = zonelist->zlcache_ptr;
1783 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1786 #else /* CONFIG_NUMA */
1788 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1793 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1794 nodemask_t *allowednodes)
1799 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1803 static void zlc_clear_zones_full(struct zonelist *zonelist)
1806 #endif /* CONFIG_NUMA */
1809 * get_page_from_freelist goes through the zonelist trying to allocate
1812 static struct page *
1813 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1814 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1815 struct zone *preferred_zone, int migratetype)
1818 struct page *page = NULL;
1821 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1822 int zlc_active = 0; /* set if using zonelist_cache */
1823 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1825 classzone_idx = zone_idx(preferred_zone);
1828 * Scan zonelist, looking for a zone with enough free.
1829 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1831 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1832 high_zoneidx, nodemask) {
1833 if (NUMA_BUILD && zlc_active &&
1834 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1836 if ((alloc_flags & ALLOC_CPUSET) &&
1837 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1840 * When allocating a page cache page for writing, we
1841 * want to get it from a zone that is within its dirty
1842 * limit, such that no single zone holds more than its
1843 * proportional share of globally allowed dirty pages.
1844 * The dirty limits take into account the zone's
1845 * lowmem reserves and high watermark so that kswapd
1846 * should be able to balance it without having to
1847 * write pages from its LRU list.
1849 * This may look like it could increase pressure on
1850 * lower zones by failing allocations in higher zones
1851 * before they are full. But the pages that do spill
1852 * over are limited as the lower zones are protected
1853 * by this very same mechanism. It should not become
1854 * a practical burden to them.
1856 * XXX: For now, allow allocations to potentially
1857 * exceed the per-zone dirty limit in the slowpath
1858 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1859 * which is important when on a NUMA setup the allowed
1860 * zones are together not big enough to reach the
1861 * global limit. The proper fix for these situations
1862 * will require awareness of zones in the
1863 * dirty-throttling and the flusher threads.
1865 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1866 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1867 goto this_zone_full;
1869 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1870 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1874 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1875 if (zone_watermark_ok(zone, order, mark,
1876 classzone_idx, alloc_flags))
1879 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1881 * we do zlc_setup if there are multiple nodes
1882 * and before considering the first zone allowed
1885 allowednodes = zlc_setup(zonelist, alloc_flags);
1890 if (zone_reclaim_mode == 0)
1891 goto this_zone_full;
1894 * As we may have just activated ZLC, check if the first
1895 * eligible zone has failed zone_reclaim recently.
1897 if (NUMA_BUILD && zlc_active &&
1898 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1901 ret = zone_reclaim(zone, gfp_mask, order);
1903 case ZONE_RECLAIM_NOSCAN:
1906 case ZONE_RECLAIM_FULL:
1907 /* scanned but unreclaimable */
1910 /* did we reclaim enough */
1911 if (!zone_watermark_ok(zone, order, mark,
1912 classzone_idx, alloc_flags))
1913 goto this_zone_full;
1918 page = buffered_rmqueue(preferred_zone, zone, order,
1919 gfp_mask, migratetype);
1924 zlc_mark_zone_full(zonelist, z);
1927 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1928 /* Disable zlc cache for second zonelist scan */
1936 * Large machines with many possible nodes should not always dump per-node
1937 * meminfo in irq context.
1939 static inline bool should_suppress_show_mem(void)
1944 ret = in_interrupt();
1949 static DEFINE_RATELIMIT_STATE(nopage_rs,
1950 DEFAULT_RATELIMIT_INTERVAL,
1951 DEFAULT_RATELIMIT_BURST);
1953 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1955 unsigned int filter = SHOW_MEM_FILTER_NODES;
1957 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1958 debug_guardpage_minorder() > 0)
1962 * This documents exceptions given to allocations in certain
1963 * contexts that are allowed to allocate outside current's set
1966 if (!(gfp_mask & __GFP_NOMEMALLOC))
1967 if (test_thread_flag(TIF_MEMDIE) ||
1968 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1969 filter &= ~SHOW_MEM_FILTER_NODES;
1970 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1971 filter &= ~SHOW_MEM_FILTER_NODES;
1974 struct va_format vaf;
1977 va_start(args, fmt);
1982 pr_warn("%pV", &vaf);
1987 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1988 current->comm, order, gfp_mask);
1991 if (!should_suppress_show_mem())
1996 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1997 unsigned long did_some_progress,
1998 unsigned long pages_reclaimed)
2000 /* Do not loop if specifically requested */
2001 if (gfp_mask & __GFP_NORETRY)
2004 /* Always retry if specifically requested */
2005 if (gfp_mask & __GFP_NOFAIL)
2009 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2010 * making forward progress without invoking OOM. Suspend also disables
2011 * storage devices so kswapd will not help. Bail if we are suspending.
2013 if (!did_some_progress && pm_suspended_storage())
2017 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2018 * means __GFP_NOFAIL, but that may not be true in other
2021 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2025 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2026 * specified, then we retry until we no longer reclaim any pages
2027 * (above), or we've reclaimed an order of pages at least as
2028 * large as the allocation's order. In both cases, if the
2029 * allocation still fails, we stop retrying.
2031 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2037 static inline struct page *
2038 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2039 struct zonelist *zonelist, enum zone_type high_zoneidx,
2040 nodemask_t *nodemask, struct zone *preferred_zone,
2045 /* Acquire the OOM killer lock for the zones in zonelist */
2046 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2047 schedule_timeout_uninterruptible(1);
2052 * Go through the zonelist yet one more time, keep very high watermark
2053 * here, this is only to catch a parallel oom killing, we must fail if
2054 * we're still under heavy pressure.
2056 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2057 order, zonelist, high_zoneidx,
2058 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2059 preferred_zone, migratetype);
2063 if (!(gfp_mask & __GFP_NOFAIL)) {
2064 /* The OOM killer will not help higher order allocs */
2065 if (order > PAGE_ALLOC_COSTLY_ORDER)
2067 /* The OOM killer does not needlessly kill tasks for lowmem */
2068 if (high_zoneidx < ZONE_NORMAL)
2071 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2072 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2073 * The caller should handle page allocation failure by itself if
2074 * it specifies __GFP_THISNODE.
2075 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2077 if (gfp_mask & __GFP_THISNODE)
2080 /* Exhausted what can be done so it's blamo time */
2081 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2084 clear_zonelist_oom(zonelist, gfp_mask);
2088 #ifdef CONFIG_COMPACTION
2089 /* Try memory compaction for high-order allocations before reclaim */
2090 static struct page *
2091 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2092 struct zonelist *zonelist, enum zone_type high_zoneidx,
2093 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2094 int migratetype, bool sync_migration,
2095 bool *deferred_compaction,
2096 unsigned long *did_some_progress)
2103 if (compaction_deferred(preferred_zone, order)) {
2104 *deferred_compaction = true;
2108 current->flags |= PF_MEMALLOC;
2109 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2110 nodemask, sync_migration);
2111 current->flags &= ~PF_MEMALLOC;
2112 if (*did_some_progress != COMPACT_SKIPPED) {
2114 /* Page migration frees to the PCP lists but we want merging */
2115 drain_pages(get_cpu());
2118 page = get_page_from_freelist(gfp_mask, nodemask,
2119 order, zonelist, high_zoneidx,
2120 alloc_flags, preferred_zone,
2123 preferred_zone->compact_considered = 0;
2124 preferred_zone->compact_defer_shift = 0;
2125 if (order >= preferred_zone->compact_order_failed)
2126 preferred_zone->compact_order_failed = order + 1;
2127 count_vm_event(COMPACTSUCCESS);
2132 * It's bad if compaction run occurs and fails.
2133 * The most likely reason is that pages exist,
2134 * but not enough to satisfy watermarks.
2136 count_vm_event(COMPACTFAIL);
2139 * As async compaction considers a subset of pageblocks, only
2140 * defer if the failure was a sync compaction failure.
2143 defer_compaction(preferred_zone, order);
2151 static inline struct page *
2152 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2153 struct zonelist *zonelist, enum zone_type high_zoneidx,
2154 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2155 int migratetype, bool sync_migration,
2156 bool *deferred_compaction,
2157 unsigned long *did_some_progress)
2161 #endif /* CONFIG_COMPACTION */
2163 /* Perform direct synchronous page reclaim */
2165 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2166 nodemask_t *nodemask)
2168 struct reclaim_state reclaim_state;
2173 /* We now go into synchronous reclaim */
2174 cpuset_memory_pressure_bump();
2175 current->flags |= PF_MEMALLOC;
2176 lockdep_set_current_reclaim_state(gfp_mask);
2177 reclaim_state.reclaimed_slab = 0;
2178 current->reclaim_state = &reclaim_state;
2180 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2182 current->reclaim_state = NULL;
2183 lockdep_clear_current_reclaim_state();
2184 current->flags &= ~PF_MEMALLOC;
2191 /* The really slow allocator path where we enter direct reclaim */
2192 static inline struct page *
2193 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2194 struct zonelist *zonelist, enum zone_type high_zoneidx,
2195 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2196 int migratetype, unsigned long *did_some_progress)
2198 struct page *page = NULL;
2199 bool drained = false;
2201 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2203 if (unlikely(!(*did_some_progress)))
2206 /* After successful reclaim, reconsider all zones for allocation */
2208 zlc_clear_zones_full(zonelist);
2211 page = get_page_from_freelist(gfp_mask, nodemask, order,
2212 zonelist, high_zoneidx,
2213 alloc_flags, preferred_zone,
2217 * If an allocation failed after direct reclaim, it could be because
2218 * pages are pinned on the per-cpu lists. Drain them and try again
2220 if (!page && !drained) {
2230 * This is called in the allocator slow-path if the allocation request is of
2231 * sufficient urgency to ignore watermarks and take other desperate measures
2233 static inline struct page *
2234 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2235 struct zonelist *zonelist, enum zone_type high_zoneidx,
2236 nodemask_t *nodemask, struct zone *preferred_zone,
2242 page = get_page_from_freelist(gfp_mask, nodemask, order,
2243 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2244 preferred_zone, migratetype);
2246 if (!page && gfp_mask & __GFP_NOFAIL)
2247 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2248 } while (!page && (gfp_mask & __GFP_NOFAIL));
2254 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2255 enum zone_type high_zoneidx,
2256 enum zone_type classzone_idx)
2261 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2262 wakeup_kswapd(zone, order, classzone_idx);
2266 gfp_to_alloc_flags(gfp_t gfp_mask)
2268 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2269 const gfp_t wait = gfp_mask & __GFP_WAIT;
2271 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2272 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2275 * The caller may dip into page reserves a bit more if the caller
2276 * cannot run direct reclaim, or if the caller has realtime scheduling
2277 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2278 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2280 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2284 * Not worth trying to allocate harder for
2285 * __GFP_NOMEMALLOC even if it can't schedule.
2287 if (!(gfp_mask & __GFP_NOMEMALLOC))
2288 alloc_flags |= ALLOC_HARDER;
2290 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2291 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2293 alloc_flags &= ~ALLOC_CPUSET;
2294 } else if (unlikely(rt_task(current)) && !in_interrupt())
2295 alloc_flags |= ALLOC_HARDER;
2297 if ((current->flags & PF_MEMALLOC) ||
2298 unlikely(test_thread_flag(TIF_MEMDIE))) {
2299 alloc_flags |= ALLOC_PFMEMALLOC;
2301 if (likely(!(gfp_mask & __GFP_NOMEMALLOC)) && !in_interrupt())
2302 alloc_flags |= ALLOC_NO_WATERMARKS;
2308 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2310 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_PFMEMALLOC);
2313 static inline struct page *
2314 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2315 struct zonelist *zonelist, enum zone_type high_zoneidx,
2316 nodemask_t *nodemask, struct zone *preferred_zone,
2319 const gfp_t wait = gfp_mask & __GFP_WAIT;
2320 struct page *page = NULL;
2322 unsigned long pages_reclaimed = 0;
2323 unsigned long did_some_progress;
2324 bool sync_migration = false;
2325 bool deferred_compaction = false;
2328 * In the slowpath, we sanity check order to avoid ever trying to
2329 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2330 * be using allocators in order of preference for an area that is
2333 if (order >= MAX_ORDER) {
2334 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2339 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2340 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2341 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2342 * using a larger set of nodes after it has established that the
2343 * allowed per node queues are empty and that nodes are
2346 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2350 if (!(gfp_mask & __GFP_NO_KSWAPD))
2351 wake_all_kswapd(order, zonelist, high_zoneidx,
2352 zone_idx(preferred_zone));
2355 * OK, we're below the kswapd watermark and have kicked background
2356 * reclaim. Now things get more complex, so set up alloc_flags according
2357 * to how we want to proceed.
2359 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2362 * Find the true preferred zone if the allocation is unconstrained by
2365 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2366 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2370 /* This is the last chance, in general, before the goto nopage. */
2371 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2372 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2373 preferred_zone, migratetype);
2377 /* Allocate without watermarks if the context allows */
2378 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2379 page = __alloc_pages_high_priority(gfp_mask, order,
2380 zonelist, high_zoneidx, nodemask,
2381 preferred_zone, migratetype);
2386 /* Atomic allocations - we can't balance anything */
2390 /* Avoid recursion of direct reclaim */
2391 if (current->flags & PF_MEMALLOC)
2394 /* Avoid allocations with no watermarks from looping endlessly */
2395 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2399 * Try direct compaction. The first pass is asynchronous. Subsequent
2400 * attempts after direct reclaim are synchronous
2402 page = __alloc_pages_direct_compact(gfp_mask, order,
2403 zonelist, high_zoneidx,
2405 alloc_flags, preferred_zone,
2406 migratetype, sync_migration,
2407 &deferred_compaction,
2408 &did_some_progress);
2411 sync_migration = true;
2414 * If compaction is deferred for high-order allocations, it is because
2415 * sync compaction recently failed. In this is the case and the caller
2416 * has requested the system not be heavily disrupted, fail the
2417 * allocation now instead of entering direct reclaim
2419 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2422 /* Try direct reclaim and then allocating */
2423 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2424 zonelist, high_zoneidx,
2426 alloc_flags, preferred_zone,
2427 migratetype, &did_some_progress);
2432 * If we failed to make any progress reclaiming, then we are
2433 * running out of options and have to consider going OOM
2435 if (!did_some_progress) {
2436 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2437 if (oom_killer_disabled)
2439 /* Coredumps can quickly deplete all memory reserves */
2440 if ((current->flags & PF_DUMPCORE) &&
2441 !(gfp_mask & __GFP_NOFAIL))
2443 page = __alloc_pages_may_oom(gfp_mask, order,
2444 zonelist, high_zoneidx,
2445 nodemask, preferred_zone,
2450 if (!(gfp_mask & __GFP_NOFAIL)) {
2452 * The oom killer is not called for high-order
2453 * allocations that may fail, so if no progress
2454 * is being made, there are no other options and
2455 * retrying is unlikely to help.
2457 if (order > PAGE_ALLOC_COSTLY_ORDER)
2460 * The oom killer is not called for lowmem
2461 * allocations to prevent needlessly killing
2464 if (high_zoneidx < ZONE_NORMAL)
2472 /* Check if we should retry the allocation */
2473 pages_reclaimed += did_some_progress;
2474 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2476 /* Wait for some write requests to complete then retry */
2477 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2481 * High-order allocations do not necessarily loop after
2482 * direct reclaim and reclaim/compaction depends on compaction
2483 * being called after reclaim so call directly if necessary
2485 page = __alloc_pages_direct_compact(gfp_mask, order,
2486 zonelist, high_zoneidx,
2488 alloc_flags, preferred_zone,
2489 migratetype, sync_migration,
2490 &deferred_compaction,
2491 &did_some_progress);
2497 warn_alloc_failed(gfp_mask, order, NULL);
2501 * page->pfmemalloc is set when the caller had PFMEMALLOC set or is
2502 * been OOM killed. The expectation is that the caller is taking
2503 * steps that will free more memory. The caller should avoid the
2504 * page being used for !PFMEMALLOC purposes.
2506 page->pfmemalloc = !!(alloc_flags & ALLOC_PFMEMALLOC);
2508 if (kmemcheck_enabled)
2509 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2515 * This is the 'heart' of the zoned buddy allocator.
2518 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2519 struct zonelist *zonelist, nodemask_t *nodemask)
2521 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2522 struct zone *preferred_zone;
2523 struct page *page = NULL;
2524 int migratetype = allocflags_to_migratetype(gfp_mask);
2525 unsigned int cpuset_mems_cookie;
2527 gfp_mask &= gfp_allowed_mask;
2529 lockdep_trace_alloc(gfp_mask);
2531 might_sleep_if(gfp_mask & __GFP_WAIT);
2533 if (should_fail_alloc_page(gfp_mask, order))
2537 * Check the zones suitable for the gfp_mask contain at least one
2538 * valid zone. It's possible to have an empty zonelist as a result
2539 * of GFP_THISNODE and a memoryless node
2541 if (unlikely(!zonelist->_zonerefs->zone))
2545 cpuset_mems_cookie = get_mems_allowed();
2547 /* The preferred zone is used for statistics later */
2548 first_zones_zonelist(zonelist, high_zoneidx,
2549 nodemask ? : &cpuset_current_mems_allowed,
2551 if (!preferred_zone)
2554 /* First allocation attempt */
2555 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2556 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2557 preferred_zone, migratetype);
2558 if (unlikely(!page))
2559 page = __alloc_pages_slowpath(gfp_mask, order,
2560 zonelist, high_zoneidx, nodemask,
2561 preferred_zone, migratetype);
2563 page->pfmemalloc = false;
2565 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2569 * When updating a task's mems_allowed, it is possible to race with
2570 * parallel threads in such a way that an allocation can fail while
2571 * the mask is being updated. If a page allocation is about to fail,
2572 * check if the cpuset changed during allocation and if so, retry.
2574 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2579 EXPORT_SYMBOL(__alloc_pages_nodemask);
2582 * Common helper functions.
2584 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2589 * __get_free_pages() returns a 32-bit address, which cannot represent
2592 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2594 page = alloc_pages(gfp_mask, order);
2597 return (unsigned long) page_address(page);
2599 EXPORT_SYMBOL(__get_free_pages);
2601 unsigned long get_zeroed_page(gfp_t gfp_mask)
2603 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2605 EXPORT_SYMBOL(get_zeroed_page);
2607 void __free_pages(struct page *page, unsigned int order)
2609 if (put_page_testzero(page)) {
2611 free_hot_cold_page(page, 0);
2613 __free_pages_ok(page, order);
2617 EXPORT_SYMBOL(__free_pages);
2619 void free_pages(unsigned long addr, unsigned int order)
2622 VM_BUG_ON(!virt_addr_valid((void *)addr));
2623 __free_pages(virt_to_page((void *)addr), order);
2627 EXPORT_SYMBOL(free_pages);
2629 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2632 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2633 unsigned long used = addr + PAGE_ALIGN(size);
2635 split_page(virt_to_page((void *)addr), order);
2636 while (used < alloc_end) {
2641 return (void *)addr;
2645 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2646 * @size: the number of bytes to allocate
2647 * @gfp_mask: GFP flags for the allocation
2649 * This function is similar to alloc_pages(), except that it allocates the
2650 * minimum number of pages to satisfy the request. alloc_pages() can only
2651 * allocate memory in power-of-two pages.
2653 * This function is also limited by MAX_ORDER.
2655 * Memory allocated by this function must be released by free_pages_exact().
2657 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2659 unsigned int order = get_order(size);
2662 addr = __get_free_pages(gfp_mask, order);
2663 return make_alloc_exact(addr, order, size);
2665 EXPORT_SYMBOL(alloc_pages_exact);
2668 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2670 * @nid: the preferred node ID where memory should be allocated
2671 * @size: the number of bytes to allocate
2672 * @gfp_mask: GFP flags for the allocation
2674 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2676 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2679 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2681 unsigned order = get_order(size);
2682 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2685 return make_alloc_exact((unsigned long)page_address(p), order, size);
2687 EXPORT_SYMBOL(alloc_pages_exact_nid);
2690 * free_pages_exact - release memory allocated via alloc_pages_exact()
2691 * @virt: the value returned by alloc_pages_exact.
2692 * @size: size of allocation, same value as passed to alloc_pages_exact().
2694 * Release the memory allocated by a previous call to alloc_pages_exact.
2696 void free_pages_exact(void *virt, size_t size)
2698 unsigned long addr = (unsigned long)virt;
2699 unsigned long end = addr + PAGE_ALIGN(size);
2701 while (addr < end) {
2706 EXPORT_SYMBOL(free_pages_exact);
2708 static unsigned int nr_free_zone_pages(int offset)
2713 /* Just pick one node, since fallback list is circular */
2714 unsigned int sum = 0;
2716 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2718 for_each_zone_zonelist(zone, z, zonelist, offset) {
2719 unsigned long size = zone->present_pages;
2720 unsigned long high = high_wmark_pages(zone);
2729 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2731 unsigned int nr_free_buffer_pages(void)
2733 return nr_free_zone_pages(gfp_zone(GFP_USER));
2735 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2738 * Amount of free RAM allocatable within all zones
2740 unsigned int nr_free_pagecache_pages(void)
2742 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2745 static inline void show_node(struct zone *zone)
2748 printk("Node %d ", zone_to_nid(zone));
2751 void si_meminfo(struct sysinfo *val)
2753 val->totalram = totalram_pages;
2755 val->freeram = global_page_state(NR_FREE_PAGES);
2756 val->bufferram = nr_blockdev_pages();
2757 val->totalhigh = totalhigh_pages;
2758 val->freehigh = nr_free_highpages();
2759 val->mem_unit = PAGE_SIZE;
2762 EXPORT_SYMBOL(si_meminfo);
2765 void si_meminfo_node(struct sysinfo *val, int nid)
2767 pg_data_t *pgdat = NODE_DATA(nid);
2769 val->totalram = pgdat->node_present_pages;
2770 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2771 #ifdef CONFIG_HIGHMEM
2772 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2773 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2779 val->mem_unit = PAGE_SIZE;
2784 * Determine whether the node should be displayed or not, depending on whether
2785 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2787 bool skip_free_areas_node(unsigned int flags, int nid)
2790 unsigned int cpuset_mems_cookie;
2792 if (!(flags & SHOW_MEM_FILTER_NODES))
2796 cpuset_mems_cookie = get_mems_allowed();
2797 ret = !node_isset(nid, cpuset_current_mems_allowed);
2798 } while (!put_mems_allowed(cpuset_mems_cookie));
2803 #define K(x) ((x) << (PAGE_SHIFT-10))
2806 * Show free area list (used inside shift_scroll-lock stuff)
2807 * We also calculate the percentage fragmentation. We do this by counting the
2808 * memory on each free list with the exception of the first item on the list.
2809 * Suppresses nodes that are not allowed by current's cpuset if
2810 * SHOW_MEM_FILTER_NODES is passed.
2812 void show_free_areas(unsigned int filter)
2817 for_each_populated_zone(zone) {
2818 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2821 printk("%s per-cpu:\n", zone->name);
2823 for_each_online_cpu(cpu) {
2824 struct per_cpu_pageset *pageset;
2826 pageset = per_cpu_ptr(zone->pageset, cpu);
2828 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2829 cpu, pageset->pcp.high,
2830 pageset->pcp.batch, pageset->pcp.count);
2834 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2835 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2837 " dirty:%lu writeback:%lu unstable:%lu\n"
2838 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2839 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2840 global_page_state(NR_ACTIVE_ANON),
2841 global_page_state(NR_INACTIVE_ANON),
2842 global_page_state(NR_ISOLATED_ANON),
2843 global_page_state(NR_ACTIVE_FILE),
2844 global_page_state(NR_INACTIVE_FILE),
2845 global_page_state(NR_ISOLATED_FILE),
2846 global_page_state(NR_UNEVICTABLE),
2847 global_page_state(NR_FILE_DIRTY),
2848 global_page_state(NR_WRITEBACK),
2849 global_page_state(NR_UNSTABLE_NFS),
2850 global_page_state(NR_FREE_PAGES),
2851 global_page_state(NR_SLAB_RECLAIMABLE),
2852 global_page_state(NR_SLAB_UNRECLAIMABLE),
2853 global_page_state(NR_FILE_MAPPED),
2854 global_page_state(NR_SHMEM),
2855 global_page_state(NR_PAGETABLE),
2856 global_page_state(NR_BOUNCE));
2858 for_each_populated_zone(zone) {
2861 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2869 " active_anon:%lukB"
2870 " inactive_anon:%lukB"
2871 " active_file:%lukB"
2872 " inactive_file:%lukB"
2873 " unevictable:%lukB"
2874 " isolated(anon):%lukB"
2875 " isolated(file):%lukB"
2882 " slab_reclaimable:%lukB"
2883 " slab_unreclaimable:%lukB"
2884 " kernel_stack:%lukB"
2888 " writeback_tmp:%lukB"
2889 " pages_scanned:%lu"
2890 " all_unreclaimable? %s"
2893 K(zone_page_state(zone, NR_FREE_PAGES)),
2894 K(min_wmark_pages(zone)),
2895 K(low_wmark_pages(zone)),
2896 K(high_wmark_pages(zone)),
2897 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2898 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2899 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2900 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2901 K(zone_page_state(zone, NR_UNEVICTABLE)),
2902 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2903 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2904 K(zone->present_pages),
2905 K(zone_page_state(zone, NR_MLOCK)),
2906 K(zone_page_state(zone, NR_FILE_DIRTY)),
2907 K(zone_page_state(zone, NR_WRITEBACK)),
2908 K(zone_page_state(zone, NR_FILE_MAPPED)),
2909 K(zone_page_state(zone, NR_SHMEM)),
2910 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2911 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2912 zone_page_state(zone, NR_KERNEL_STACK) *
2914 K(zone_page_state(zone, NR_PAGETABLE)),
2915 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2916 K(zone_page_state(zone, NR_BOUNCE)),
2917 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2918 zone->pages_scanned,
2919 (zone->all_unreclaimable ? "yes" : "no")
2921 printk("lowmem_reserve[]:");
2922 for (i = 0; i < MAX_NR_ZONES; i++)
2923 printk(" %lu", zone->lowmem_reserve[i]);
2927 for_each_populated_zone(zone) {
2928 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2930 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2933 printk("%s: ", zone->name);
2935 spin_lock_irqsave(&zone->lock, flags);
2936 for (order = 0; order < MAX_ORDER; order++) {
2937 nr[order] = zone->free_area[order].nr_free;
2938 total += nr[order] << order;
2940 spin_unlock_irqrestore(&zone->lock, flags);
2941 for (order = 0; order < MAX_ORDER; order++)
2942 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2943 printk("= %lukB\n", K(total));
2946 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2948 show_swap_cache_info();
2951 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2953 zoneref->zone = zone;
2954 zoneref->zone_idx = zone_idx(zone);
2958 * Builds allocation fallback zone lists.
2960 * Add all populated zones of a node to the zonelist.
2962 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2963 int nr_zones, enum zone_type zone_type)
2967 BUG_ON(zone_type >= MAX_NR_ZONES);
2972 zone = pgdat->node_zones + zone_type;
2973 if (populated_zone(zone)) {
2974 zoneref_set_zone(zone,
2975 &zonelist->_zonerefs[nr_zones++]);
2976 check_highest_zone(zone_type);
2979 } while (zone_type);
2986 * 0 = automatic detection of better ordering.
2987 * 1 = order by ([node] distance, -zonetype)
2988 * 2 = order by (-zonetype, [node] distance)
2990 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2991 * the same zonelist. So only NUMA can configure this param.
2993 #define ZONELIST_ORDER_DEFAULT 0
2994 #define ZONELIST_ORDER_NODE 1
2995 #define ZONELIST_ORDER_ZONE 2
2997 /* zonelist order in the kernel.
2998 * set_zonelist_order() will set this to NODE or ZONE.
3000 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3001 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3005 /* The value user specified ....changed by config */
3006 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3007 /* string for sysctl */
3008 #define NUMA_ZONELIST_ORDER_LEN 16
3009 char numa_zonelist_order[16] = "default";
3012 * interface for configure zonelist ordering.
3013 * command line option "numa_zonelist_order"
3014 * = "[dD]efault - default, automatic configuration.
3015 * = "[nN]ode - order by node locality, then by zone within node
3016 * = "[zZ]one - order by zone, then by locality within zone
3019 static int __parse_numa_zonelist_order(char *s)
3021 if (*s == 'd' || *s == 'D') {
3022 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3023 } else if (*s == 'n' || *s == 'N') {
3024 user_zonelist_order = ZONELIST_ORDER_NODE;
3025 } else if (*s == 'z' || *s == 'Z') {
3026 user_zonelist_order = ZONELIST_ORDER_ZONE;
3029 "Ignoring invalid numa_zonelist_order value: "
3036 static __init int setup_numa_zonelist_order(char *s)
3043 ret = __parse_numa_zonelist_order(s);
3045 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3049 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3052 * sysctl handler for numa_zonelist_order
3054 int numa_zonelist_order_handler(ctl_table *table, int write,
3055 void __user *buffer, size_t *length,
3058 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3060 static DEFINE_MUTEX(zl_order_mutex);
3062 mutex_lock(&zl_order_mutex);
3064 strcpy(saved_string, (char*)table->data);
3065 ret = proc_dostring(table, write, buffer, length, ppos);
3069 int oldval = user_zonelist_order;
3070 if (__parse_numa_zonelist_order((char*)table->data)) {
3072 * bogus value. restore saved string
3074 strncpy((char*)table->data, saved_string,
3075 NUMA_ZONELIST_ORDER_LEN);
3076 user_zonelist_order = oldval;
3077 } else if (oldval != user_zonelist_order) {
3078 mutex_lock(&zonelists_mutex);
3079 build_all_zonelists(NULL, NULL);
3080 mutex_unlock(&zonelists_mutex);
3084 mutex_unlock(&zl_order_mutex);
3089 #define MAX_NODE_LOAD (nr_online_nodes)
3090 static int node_load[MAX_NUMNODES];
3093 * find_next_best_node - find the next node that should appear in a given node's fallback list
3094 * @node: node whose fallback list we're appending
3095 * @used_node_mask: nodemask_t of already used nodes
3097 * We use a number of factors to determine which is the next node that should
3098 * appear on a given node's fallback list. The node should not have appeared
3099 * already in @node's fallback list, and it should be the next closest node
3100 * according to the distance array (which contains arbitrary distance values
3101 * from each node to each node in the system), and should also prefer nodes
3102 * with no CPUs, since presumably they'll have very little allocation pressure
3103 * on them otherwise.
3104 * It returns -1 if no node is found.
3106 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3109 int min_val = INT_MAX;
3111 const struct cpumask *tmp = cpumask_of_node(0);
3113 /* Use the local node if we haven't already */
3114 if (!node_isset(node, *used_node_mask)) {
3115 node_set(node, *used_node_mask);
3119 for_each_node_state(n, N_HIGH_MEMORY) {
3121 /* Don't want a node to appear more than once */
3122 if (node_isset(n, *used_node_mask))
3125 /* Use the distance array to find the distance */
3126 val = node_distance(node, n);
3128 /* Penalize nodes under us ("prefer the next node") */
3131 /* Give preference to headless and unused nodes */
3132 tmp = cpumask_of_node(n);
3133 if (!cpumask_empty(tmp))
3134 val += PENALTY_FOR_NODE_WITH_CPUS;
3136 /* Slight preference for less loaded node */
3137 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3138 val += node_load[n];
3140 if (val < min_val) {
3147 node_set(best_node, *used_node_mask);
3154 * Build zonelists ordered by node and zones within node.
3155 * This results in maximum locality--normal zone overflows into local
3156 * DMA zone, if any--but risks exhausting DMA zone.
3158 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3161 struct zonelist *zonelist;
3163 zonelist = &pgdat->node_zonelists[0];
3164 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3166 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3168 zonelist->_zonerefs[j].zone = NULL;
3169 zonelist->_zonerefs[j].zone_idx = 0;
3173 * Build gfp_thisnode zonelists
3175 static void build_thisnode_zonelists(pg_data_t *pgdat)
3178 struct zonelist *zonelist;
3180 zonelist = &pgdat->node_zonelists[1];
3181 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3182 zonelist->_zonerefs[j].zone = NULL;
3183 zonelist->_zonerefs[j].zone_idx = 0;
3187 * Build zonelists ordered by zone and nodes within zones.
3188 * This results in conserving DMA zone[s] until all Normal memory is
3189 * exhausted, but results in overflowing to remote node while memory
3190 * may still exist in local DMA zone.
3192 static int node_order[MAX_NUMNODES];
3194 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3197 int zone_type; /* needs to be signed */
3199 struct zonelist *zonelist;
3201 zonelist = &pgdat->node_zonelists[0];
3203 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3204 for (j = 0; j < nr_nodes; j++) {
3205 node = node_order[j];
3206 z = &NODE_DATA(node)->node_zones[zone_type];
3207 if (populated_zone(z)) {
3209 &zonelist->_zonerefs[pos++]);
3210 check_highest_zone(zone_type);
3214 zonelist->_zonerefs[pos].zone = NULL;
3215 zonelist->_zonerefs[pos].zone_idx = 0;
3218 static int default_zonelist_order(void)
3221 unsigned long low_kmem_size,total_size;
3225 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3226 * If they are really small and used heavily, the system can fall
3227 * into OOM very easily.
3228 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3230 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3233 for_each_online_node(nid) {
3234 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3235 z = &NODE_DATA(nid)->node_zones[zone_type];
3236 if (populated_zone(z)) {
3237 if (zone_type < ZONE_NORMAL)
3238 low_kmem_size += z->present_pages;
3239 total_size += z->present_pages;
3240 } else if (zone_type == ZONE_NORMAL) {
3242 * If any node has only lowmem, then node order
3243 * is preferred to allow kernel allocations
3244 * locally; otherwise, they can easily infringe
3245 * on other nodes when there is an abundance of
3246 * lowmem available to allocate from.
3248 return ZONELIST_ORDER_NODE;
3252 if (!low_kmem_size || /* there are no DMA area. */
3253 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3254 return ZONELIST_ORDER_NODE;
3256 * look into each node's config.
3257 * If there is a node whose DMA/DMA32 memory is very big area on
3258 * local memory, NODE_ORDER may be suitable.
3260 average_size = total_size /
3261 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3262 for_each_online_node(nid) {
3265 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3266 z = &NODE_DATA(nid)->node_zones[zone_type];
3267 if (populated_zone(z)) {
3268 if (zone_type < ZONE_NORMAL)
3269 low_kmem_size += z->present_pages;
3270 total_size += z->present_pages;
3273 if (low_kmem_size &&
3274 total_size > average_size && /* ignore small node */
3275 low_kmem_size > total_size * 70/100)
3276 return ZONELIST_ORDER_NODE;
3278 return ZONELIST_ORDER_ZONE;
3281 static void set_zonelist_order(void)
3283 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3284 current_zonelist_order = default_zonelist_order();
3286 current_zonelist_order = user_zonelist_order;
3289 static void build_zonelists(pg_data_t *pgdat)
3293 nodemask_t used_mask;
3294 int local_node, prev_node;
3295 struct zonelist *zonelist;
3296 int order = current_zonelist_order;
3298 /* initialize zonelists */
3299 for (i = 0; i < MAX_ZONELISTS; i++) {
3300 zonelist = pgdat->node_zonelists + i;
3301 zonelist->_zonerefs[0].zone = NULL;
3302 zonelist->_zonerefs[0].zone_idx = 0;
3305 /* NUMA-aware ordering of nodes */
3306 local_node = pgdat->node_id;
3307 load = nr_online_nodes;
3308 prev_node = local_node;
3309 nodes_clear(used_mask);
3311 memset(node_order, 0, sizeof(node_order));
3314 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3315 int distance = node_distance(local_node, node);
3318 * If another node is sufficiently far away then it is better
3319 * to reclaim pages in a zone before going off node.
3321 if (distance > RECLAIM_DISTANCE)
3322 zone_reclaim_mode = 1;
3325 * We don't want to pressure a particular node.
3326 * So adding penalty to the first node in same
3327 * distance group to make it round-robin.
3329 if (distance != node_distance(local_node, prev_node))
3330 node_load[node] = load;
3334 if (order == ZONELIST_ORDER_NODE)
3335 build_zonelists_in_node_order(pgdat, node);
3337 node_order[j++] = node; /* remember order */
3340 if (order == ZONELIST_ORDER_ZONE) {
3341 /* calculate node order -- i.e., DMA last! */
3342 build_zonelists_in_zone_order(pgdat, j);
3345 build_thisnode_zonelists(pgdat);
3348 /* Construct the zonelist performance cache - see further mmzone.h */
3349 static void build_zonelist_cache(pg_data_t *pgdat)
3351 struct zonelist *zonelist;
3352 struct zonelist_cache *zlc;
3355 zonelist = &pgdat->node_zonelists[0];
3356 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3357 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3358 for (z = zonelist->_zonerefs; z->zone; z++)
3359 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3362 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3364 * Return node id of node used for "local" allocations.
3365 * I.e., first node id of first zone in arg node's generic zonelist.
3366 * Used for initializing percpu 'numa_mem', which is used primarily
3367 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3369 int local_memory_node(int node)
3373 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3374 gfp_zone(GFP_KERNEL),
3381 #else /* CONFIG_NUMA */
3383 static void set_zonelist_order(void)
3385 current_zonelist_order = ZONELIST_ORDER_ZONE;
3388 static void build_zonelists(pg_data_t *pgdat)
3390 int node, local_node;
3392 struct zonelist *zonelist;
3394 local_node = pgdat->node_id;
3396 zonelist = &pgdat->node_zonelists[0];
3397 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3400 * Now we build the zonelist so that it contains the zones
3401 * of all the other nodes.
3402 * We don't want to pressure a particular node, so when
3403 * building the zones for node N, we make sure that the
3404 * zones coming right after the local ones are those from
3405 * node N+1 (modulo N)
3407 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3408 if (!node_online(node))
3410 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3413 for (node = 0; node < local_node; node++) {
3414 if (!node_online(node))
3416 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3420 zonelist->_zonerefs[j].zone = NULL;
3421 zonelist->_zonerefs[j].zone_idx = 0;
3424 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3425 static void build_zonelist_cache(pg_data_t *pgdat)
3427 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3430 #endif /* CONFIG_NUMA */
3433 * Boot pageset table. One per cpu which is going to be used for all
3434 * zones and all nodes. The parameters will be set in such a way
3435 * that an item put on a list will immediately be handed over to
3436 * the buddy list. This is safe since pageset manipulation is done
3437 * with interrupts disabled.
3439 * The boot_pagesets must be kept even after bootup is complete for
3440 * unused processors and/or zones. They do play a role for bootstrapping
3441 * hotplugged processors.
3443 * zoneinfo_show() and maybe other functions do
3444 * not check if the processor is online before following the pageset pointer.
3445 * Other parts of the kernel may not check if the zone is available.
3447 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3448 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3449 static void setup_zone_pageset(struct zone *zone);
3452 * Global mutex to protect against size modification of zonelists
3453 * as well as to serialize pageset setup for the new populated zone.
3455 DEFINE_MUTEX(zonelists_mutex);
3457 /* return values int ....just for stop_machine() */
3458 static int __build_all_zonelists(void *data)
3462 pg_data_t *self = data;
3465 memset(node_load, 0, sizeof(node_load));
3468 if (self && !node_online(self->node_id)) {
3469 build_zonelists(self);
3470 build_zonelist_cache(self);
3473 for_each_online_node(nid) {
3474 pg_data_t *pgdat = NODE_DATA(nid);
3476 build_zonelists(pgdat);
3477 build_zonelist_cache(pgdat);
3481 * Initialize the boot_pagesets that are going to be used
3482 * for bootstrapping processors. The real pagesets for
3483 * each zone will be allocated later when the per cpu
3484 * allocator is available.
3486 * boot_pagesets are used also for bootstrapping offline
3487 * cpus if the system is already booted because the pagesets
3488 * are needed to initialize allocators on a specific cpu too.
3489 * F.e. the percpu allocator needs the page allocator which
3490 * needs the percpu allocator in order to allocate its pagesets
3491 * (a chicken-egg dilemma).
3493 for_each_possible_cpu(cpu) {
3494 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3496 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3498 * We now know the "local memory node" for each node--
3499 * i.e., the node of the first zone in the generic zonelist.
3500 * Set up numa_mem percpu variable for on-line cpus. During
3501 * boot, only the boot cpu should be on-line; we'll init the
3502 * secondary cpus' numa_mem as they come on-line. During
3503 * node/memory hotplug, we'll fixup all on-line cpus.
3505 if (cpu_online(cpu))
3506 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3514 * Called with zonelists_mutex held always
3515 * unless system_state == SYSTEM_BOOTING.
3517 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3519 set_zonelist_order();
3521 if (system_state == SYSTEM_BOOTING) {
3522 __build_all_zonelists(NULL);
3523 mminit_verify_zonelist();
3524 cpuset_init_current_mems_allowed();
3526 /* we have to stop all cpus to guarantee there is no user
3528 #ifdef CONFIG_MEMORY_HOTPLUG
3530 setup_zone_pageset(zone);
3532 stop_machine(__build_all_zonelists, pgdat, NULL);
3533 /* cpuset refresh routine should be here */
3535 vm_total_pages = nr_free_pagecache_pages();
3537 * Disable grouping by mobility if the number of pages in the
3538 * system is too low to allow the mechanism to work. It would be
3539 * more accurate, but expensive to check per-zone. This check is
3540 * made on memory-hotadd so a system can start with mobility
3541 * disabled and enable it later
3543 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3544 page_group_by_mobility_disabled = 1;
3546 page_group_by_mobility_disabled = 0;
3548 printk("Built %i zonelists in %s order, mobility grouping %s. "
3549 "Total pages: %ld\n",
3551 zonelist_order_name[current_zonelist_order],
3552 page_group_by_mobility_disabled ? "off" : "on",
3555 printk("Policy zone: %s\n", zone_names[policy_zone]);
3560 * Helper functions to size the waitqueue hash table.
3561 * Essentially these want to choose hash table sizes sufficiently
3562 * large so that collisions trying to wait on pages are rare.
3563 * But in fact, the number of active page waitqueues on typical
3564 * systems is ridiculously low, less than 200. So this is even
3565 * conservative, even though it seems large.
3567 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3568 * waitqueues, i.e. the size of the waitq table given the number of pages.
3570 #define PAGES_PER_WAITQUEUE 256
3572 #ifndef CONFIG_MEMORY_HOTPLUG
3573 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3575 unsigned long size = 1;
3577 pages /= PAGES_PER_WAITQUEUE;
3579 while (size < pages)
3583 * Once we have dozens or even hundreds of threads sleeping
3584 * on IO we've got bigger problems than wait queue collision.
3585 * Limit the size of the wait table to a reasonable size.
3587 size = min(size, 4096UL);
3589 return max(size, 4UL);
3593 * A zone's size might be changed by hot-add, so it is not possible to determine
3594 * a suitable size for its wait_table. So we use the maximum size now.
3596 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3598 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3599 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3600 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3602 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3603 * or more by the traditional way. (See above). It equals:
3605 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3606 * ia64(16K page size) : = ( 8G + 4M)byte.
3607 * powerpc (64K page size) : = (32G +16M)byte.
3609 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3616 * This is an integer logarithm so that shifts can be used later
3617 * to extract the more random high bits from the multiplicative
3618 * hash function before the remainder is taken.
3620 static inline unsigned long wait_table_bits(unsigned long size)
3625 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3628 * Check if a pageblock contains reserved pages
3630 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3634 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3635 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3642 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3643 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3644 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3645 * higher will lead to a bigger reserve which will get freed as contiguous
3646 * blocks as reclaim kicks in
3648 static void setup_zone_migrate_reserve(struct zone *zone)
3650 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3652 unsigned long block_migratetype;
3656 * Get the start pfn, end pfn and the number of blocks to reserve
3657 * We have to be careful to be aligned to pageblock_nr_pages to
3658 * make sure that we always check pfn_valid for the first page in
3661 start_pfn = zone->zone_start_pfn;
3662 end_pfn = start_pfn + zone->spanned_pages;
3663 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3664 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3668 * Reserve blocks are generally in place to help high-order atomic
3669 * allocations that are short-lived. A min_free_kbytes value that
3670 * would result in more than 2 reserve blocks for atomic allocations
3671 * is assumed to be in place to help anti-fragmentation for the
3672 * future allocation of hugepages at runtime.
3674 reserve = min(2, reserve);
3676 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3677 if (!pfn_valid(pfn))
3679 page = pfn_to_page(pfn);
3681 /* Watch out for overlapping nodes */
3682 if (page_to_nid(page) != zone_to_nid(zone))
3685 block_migratetype = get_pageblock_migratetype(page);
3687 /* Only test what is necessary when the reserves are not met */
3690 * Blocks with reserved pages will never free, skip
3693 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3694 if (pageblock_is_reserved(pfn, block_end_pfn))
3697 /* If this block is reserved, account for it */
3698 if (block_migratetype == MIGRATE_RESERVE) {
3703 /* Suitable for reserving if this block is movable */
3704 if (block_migratetype == MIGRATE_MOVABLE) {
3705 set_pageblock_migratetype(page,
3707 move_freepages_block(zone, page,
3715 * If the reserve is met and this is a previous reserved block,
3718 if (block_migratetype == MIGRATE_RESERVE) {
3719 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3720 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3726 * Initially all pages are reserved - free ones are freed
3727 * up by free_all_bootmem() once the early boot process is
3728 * done. Non-atomic initialization, single-pass.
3730 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3731 unsigned long start_pfn, enum memmap_context context)
3734 unsigned long end_pfn = start_pfn + size;
3738 if (highest_memmap_pfn < end_pfn - 1)
3739 highest_memmap_pfn = end_pfn - 1;
3741 z = &NODE_DATA(nid)->node_zones[zone];
3742 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3744 * There can be holes in boot-time mem_map[]s
3745 * handed to this function. They do not
3746 * exist on hotplugged memory.
3748 if (context == MEMMAP_EARLY) {
3749 if (!early_pfn_valid(pfn))
3751 if (!early_pfn_in_nid(pfn, nid))
3754 page = pfn_to_page(pfn);
3755 set_page_links(page, zone, nid, pfn);
3756 mminit_verify_page_links(page, zone, nid, pfn);
3757 init_page_count(page);
3758 reset_page_mapcount(page);
3759 SetPageReserved(page);
3761 * Mark the block movable so that blocks are reserved for
3762 * movable at startup. This will force kernel allocations
3763 * to reserve their blocks rather than leaking throughout
3764 * the address space during boot when many long-lived
3765 * kernel allocations are made. Later some blocks near
3766 * the start are marked MIGRATE_RESERVE by
3767 * setup_zone_migrate_reserve()
3769 * bitmap is created for zone's valid pfn range. but memmap
3770 * can be created for invalid pages (for alignment)
3771 * check here not to call set_pageblock_migratetype() against
3774 if ((z->zone_start_pfn <= pfn)
3775 && (pfn < z->zone_start_pfn + z->spanned_pages)
3776 && !(pfn & (pageblock_nr_pages - 1)))
3777 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3779 INIT_LIST_HEAD(&page->lru);
3780 #ifdef WANT_PAGE_VIRTUAL
3781 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3782 if (!is_highmem_idx(zone))
3783 set_page_address(page, __va(pfn << PAGE_SHIFT));
3788 static void __meminit zone_init_free_lists(struct zone *zone)
3791 for_each_migratetype_order(order, t) {
3792 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3793 zone->free_area[order].nr_free = 0;
3797 #ifndef __HAVE_ARCH_MEMMAP_INIT
3798 #define memmap_init(size, nid, zone, start_pfn) \
3799 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3802 static int __meminit zone_batchsize(struct zone *zone)
3808 * The per-cpu-pages pools are set to around 1000th of the
3809 * size of the zone. But no more than 1/2 of a meg.
3811 * OK, so we don't know how big the cache is. So guess.
3813 batch = zone->present_pages / 1024;
3814 if (batch * PAGE_SIZE > 512 * 1024)
3815 batch = (512 * 1024) / PAGE_SIZE;
3816 batch /= 4; /* We effectively *= 4 below */
3821 * Clamp the batch to a 2^n - 1 value. Having a power
3822 * of 2 value was found to be more likely to have
3823 * suboptimal cache aliasing properties in some cases.
3825 * For example if 2 tasks are alternately allocating
3826 * batches of pages, one task can end up with a lot
3827 * of pages of one half of the possible page colors
3828 * and the other with pages of the other colors.
3830 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3835 /* The deferral and batching of frees should be suppressed under NOMMU
3838 * The problem is that NOMMU needs to be able to allocate large chunks
3839 * of contiguous memory as there's no hardware page translation to
3840 * assemble apparent contiguous memory from discontiguous pages.
3842 * Queueing large contiguous runs of pages for batching, however,
3843 * causes the pages to actually be freed in smaller chunks. As there
3844 * can be a significant delay between the individual batches being
3845 * recycled, this leads to the once large chunks of space being
3846 * fragmented and becoming unavailable for high-order allocations.
3852 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3854 struct per_cpu_pages *pcp;
3857 memset(p, 0, sizeof(*p));
3861 pcp->high = 6 * batch;
3862 pcp->batch = max(1UL, 1 * batch);
3863 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3864 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3868 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3869 * to the value high for the pageset p.
3872 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3875 struct per_cpu_pages *pcp;
3879 pcp->batch = max(1UL, high/4);
3880 if ((high/4) > (PAGE_SHIFT * 8))
3881 pcp->batch = PAGE_SHIFT * 8;
3884 static void __meminit setup_zone_pageset(struct zone *zone)
3888 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3890 for_each_possible_cpu(cpu) {
3891 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3893 setup_pageset(pcp, zone_batchsize(zone));
3895 if (percpu_pagelist_fraction)
3896 setup_pagelist_highmark(pcp,
3897 (zone->present_pages /
3898 percpu_pagelist_fraction));
3903 * Allocate per cpu pagesets and initialize them.
3904 * Before this call only boot pagesets were available.
3906 void __init setup_per_cpu_pageset(void)
3910 for_each_populated_zone(zone)
3911 setup_zone_pageset(zone);
3914 static noinline __init_refok
3915 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3918 struct pglist_data *pgdat = zone->zone_pgdat;
3922 * The per-page waitqueue mechanism uses hashed waitqueues
3925 zone->wait_table_hash_nr_entries =
3926 wait_table_hash_nr_entries(zone_size_pages);
3927 zone->wait_table_bits =
3928 wait_table_bits(zone->wait_table_hash_nr_entries);
3929 alloc_size = zone->wait_table_hash_nr_entries
3930 * sizeof(wait_queue_head_t);
3932 if (!slab_is_available()) {
3933 zone->wait_table = (wait_queue_head_t *)
3934 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3937 * This case means that a zone whose size was 0 gets new memory
3938 * via memory hot-add.
3939 * But it may be the case that a new node was hot-added. In
3940 * this case vmalloc() will not be able to use this new node's
3941 * memory - this wait_table must be initialized to use this new
3942 * node itself as well.
3943 * To use this new node's memory, further consideration will be
3946 zone->wait_table = vmalloc(alloc_size);
3948 if (!zone->wait_table)
3951 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3952 init_waitqueue_head(zone->wait_table + i);
3957 static __meminit void zone_pcp_init(struct zone *zone)
3960 * per cpu subsystem is not up at this point. The following code
3961 * relies on the ability of the linker to provide the
3962 * offset of a (static) per cpu variable into the per cpu area.
3964 zone->pageset = &boot_pageset;
3966 if (zone->present_pages)
3967 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3968 zone->name, zone->present_pages,
3969 zone_batchsize(zone));
3972 int __meminit init_currently_empty_zone(struct zone *zone,
3973 unsigned long zone_start_pfn,
3975 enum memmap_context context)
3977 struct pglist_data *pgdat = zone->zone_pgdat;
3979 ret = zone_wait_table_init(zone, size);
3982 pgdat->nr_zones = zone_idx(zone) + 1;
3984 zone->zone_start_pfn = zone_start_pfn;
3986 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3987 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3989 (unsigned long)zone_idx(zone),
3990 zone_start_pfn, (zone_start_pfn + size));
3992 zone_init_free_lists(zone);
3997 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
3998 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4000 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4001 * Architectures may implement their own version but if add_active_range()
4002 * was used and there are no special requirements, this is a convenient
4005 int __meminit __early_pfn_to_nid(unsigned long pfn)
4007 unsigned long start_pfn, end_pfn;
4010 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4011 if (start_pfn <= pfn && pfn < end_pfn)
4013 /* This is a memory hole */
4016 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4018 int __meminit early_pfn_to_nid(unsigned long pfn)
4022 nid = __early_pfn_to_nid(pfn);
4025 /* just returns 0 */
4029 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4030 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4034 nid = __early_pfn_to_nid(pfn);
4035 if (nid >= 0 && nid != node)
4042 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4043 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4044 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4046 * If an architecture guarantees that all ranges registered with
4047 * add_active_ranges() contain no holes and may be freed, this
4048 * this function may be used instead of calling free_bootmem() manually.
4050 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4052 unsigned long start_pfn, end_pfn;
4055 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4056 start_pfn = min(start_pfn, max_low_pfn);
4057 end_pfn = min(end_pfn, max_low_pfn);
4059 if (start_pfn < end_pfn)
4060 free_bootmem_node(NODE_DATA(this_nid),
4061 PFN_PHYS(start_pfn),
4062 (end_pfn - start_pfn) << PAGE_SHIFT);
4067 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4068 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4070 * If an architecture guarantees that all ranges registered with
4071 * add_active_ranges() contain no holes and may be freed, this
4072 * function may be used instead of calling memory_present() manually.
4074 void __init sparse_memory_present_with_active_regions(int nid)
4076 unsigned long start_pfn, end_pfn;
4079 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4080 memory_present(this_nid, start_pfn, end_pfn);
4084 * get_pfn_range_for_nid - Return the start and end page frames for a node
4085 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4086 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4087 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4089 * It returns the start and end page frame of a node based on information
4090 * provided by an arch calling add_active_range(). If called for a node
4091 * with no available memory, a warning is printed and the start and end
4094 void __meminit get_pfn_range_for_nid(unsigned int nid,
4095 unsigned long *start_pfn, unsigned long *end_pfn)
4097 unsigned long this_start_pfn, this_end_pfn;
4103 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4104 *start_pfn = min(*start_pfn, this_start_pfn);
4105 *end_pfn = max(*end_pfn, this_end_pfn);
4108 if (*start_pfn == -1UL)
4113 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4114 * assumption is made that zones within a node are ordered in monotonic
4115 * increasing memory addresses so that the "highest" populated zone is used
4117 static void __init find_usable_zone_for_movable(void)
4120 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4121 if (zone_index == ZONE_MOVABLE)
4124 if (arch_zone_highest_possible_pfn[zone_index] >
4125 arch_zone_lowest_possible_pfn[zone_index])
4129 VM_BUG_ON(zone_index == -1);
4130 movable_zone = zone_index;
4134 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4135 * because it is sized independent of architecture. Unlike the other zones,
4136 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4137 * in each node depending on the size of each node and how evenly kernelcore
4138 * is distributed. This helper function adjusts the zone ranges
4139 * provided by the architecture for a given node by using the end of the
4140 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4141 * zones within a node are in order of monotonic increases memory addresses
4143 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4144 unsigned long zone_type,
4145 unsigned long node_start_pfn,
4146 unsigned long node_end_pfn,
4147 unsigned long *zone_start_pfn,
4148 unsigned long *zone_end_pfn)
4150 /* Only adjust if ZONE_MOVABLE is on this node */
4151 if (zone_movable_pfn[nid]) {
4152 /* Size ZONE_MOVABLE */
4153 if (zone_type == ZONE_MOVABLE) {
4154 *zone_start_pfn = zone_movable_pfn[nid];
4155 *zone_end_pfn = min(node_end_pfn,
4156 arch_zone_highest_possible_pfn[movable_zone]);
4158 /* Adjust for ZONE_MOVABLE starting within this range */
4159 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4160 *zone_end_pfn > zone_movable_pfn[nid]) {
4161 *zone_end_pfn = zone_movable_pfn[nid];
4163 /* Check if this whole range is within ZONE_MOVABLE */
4164 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4165 *zone_start_pfn = *zone_end_pfn;
4170 * Return the number of pages a zone spans in a node, including holes
4171 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4173 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4174 unsigned long zone_type,
4175 unsigned long *ignored)
4177 unsigned long node_start_pfn, node_end_pfn;
4178 unsigned long zone_start_pfn, zone_end_pfn;
4180 /* Get the start and end of the node and zone */
4181 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4182 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4183 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4184 adjust_zone_range_for_zone_movable(nid, zone_type,
4185 node_start_pfn, node_end_pfn,
4186 &zone_start_pfn, &zone_end_pfn);
4188 /* Check that this node has pages within the zone's required range */
4189 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4192 /* Move the zone boundaries inside the node if necessary */
4193 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4194 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4196 /* Return the spanned pages */
4197 return zone_end_pfn - zone_start_pfn;
4201 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4202 * then all holes in the requested range will be accounted for.
4204 unsigned long __meminit __absent_pages_in_range(int nid,
4205 unsigned long range_start_pfn,
4206 unsigned long range_end_pfn)
4208 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4209 unsigned long start_pfn, end_pfn;
4212 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4213 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4214 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4215 nr_absent -= end_pfn - start_pfn;
4221 * absent_pages_in_range - Return number of page frames in holes within a range
4222 * @start_pfn: The start PFN to start searching for holes
4223 * @end_pfn: The end PFN to stop searching for holes
4225 * It returns the number of pages frames in memory holes within a range.
4227 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4228 unsigned long end_pfn)
4230 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4233 /* Return the number of page frames in holes in a zone on a node */
4234 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4235 unsigned long zone_type,
4236 unsigned long *ignored)
4238 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4239 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4240 unsigned long node_start_pfn, node_end_pfn;
4241 unsigned long zone_start_pfn, zone_end_pfn;
4243 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4244 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4245 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4247 adjust_zone_range_for_zone_movable(nid, zone_type,
4248 node_start_pfn, node_end_pfn,
4249 &zone_start_pfn, &zone_end_pfn);
4250 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4253 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4254 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4255 unsigned long zone_type,
4256 unsigned long *zones_size)
4258 return zones_size[zone_type];
4261 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4262 unsigned long zone_type,
4263 unsigned long *zholes_size)
4268 return zholes_size[zone_type];
4271 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4273 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4274 unsigned long *zones_size, unsigned long *zholes_size)
4276 unsigned long realtotalpages, totalpages = 0;
4279 for (i = 0; i < MAX_NR_ZONES; i++)
4280 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4282 pgdat->node_spanned_pages = totalpages;
4284 realtotalpages = totalpages;
4285 for (i = 0; i < MAX_NR_ZONES; i++)
4287 zone_absent_pages_in_node(pgdat->node_id, i,
4289 pgdat->node_present_pages = realtotalpages;
4290 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4294 #ifndef CONFIG_SPARSEMEM
4296 * Calculate the size of the zone->blockflags rounded to an unsigned long
4297 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4298 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4299 * round what is now in bits to nearest long in bits, then return it in
4302 static unsigned long __init usemap_size(unsigned long zonesize)
4304 unsigned long usemapsize;
4306 usemapsize = roundup(zonesize, pageblock_nr_pages);
4307 usemapsize = usemapsize >> pageblock_order;
4308 usemapsize *= NR_PAGEBLOCK_BITS;
4309 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4311 return usemapsize / 8;
4314 static void __init setup_usemap(struct pglist_data *pgdat,
4315 struct zone *zone, unsigned long zonesize)
4317 unsigned long usemapsize = usemap_size(zonesize);
4318 zone->pageblock_flags = NULL;
4320 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4324 static inline void setup_usemap(struct pglist_data *pgdat,
4325 struct zone *zone, unsigned long zonesize) {}
4326 #endif /* CONFIG_SPARSEMEM */
4328 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4330 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4331 void __init set_pageblock_order(void)
4335 /* Check that pageblock_nr_pages has not already been setup */
4336 if (pageblock_order)
4339 if (HPAGE_SHIFT > PAGE_SHIFT)
4340 order = HUGETLB_PAGE_ORDER;
4342 order = MAX_ORDER - 1;
4345 * Assume the largest contiguous order of interest is a huge page.
4346 * This value may be variable depending on boot parameters on IA64 and
4349 pageblock_order = order;
4351 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4354 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4355 * is unused as pageblock_order is set at compile-time. See
4356 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4359 void __init set_pageblock_order(void)
4363 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4366 * Set up the zone data structures:
4367 * - mark all pages reserved
4368 * - mark all memory queues empty
4369 * - clear the memory bitmaps
4371 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4372 unsigned long *zones_size, unsigned long *zholes_size)
4375 int nid = pgdat->node_id;
4376 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4379 pgdat_resize_init(pgdat);
4380 pgdat->nr_zones = 0;
4381 init_waitqueue_head(&pgdat->kswapd_wait);
4382 pgdat->kswapd_max_order = 0;
4383 pgdat_page_cgroup_init(pgdat);
4385 for (j = 0; j < MAX_NR_ZONES; j++) {
4386 struct zone *zone = pgdat->node_zones + j;
4387 unsigned long size, realsize, memmap_pages;
4389 size = zone_spanned_pages_in_node(nid, j, zones_size);
4390 realsize = size - zone_absent_pages_in_node(nid, j,
4394 * Adjust realsize so that it accounts for how much memory
4395 * is used by this zone for memmap. This affects the watermark
4396 * and per-cpu initialisations
4399 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4400 if (realsize >= memmap_pages) {
4401 realsize -= memmap_pages;
4404 " %s zone: %lu pages used for memmap\n",
4405 zone_names[j], memmap_pages);
4408 " %s zone: %lu pages exceeds realsize %lu\n",
4409 zone_names[j], memmap_pages, realsize);
4411 /* Account for reserved pages */
4412 if (j == 0 && realsize > dma_reserve) {
4413 realsize -= dma_reserve;
4414 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4415 zone_names[0], dma_reserve);
4418 if (!is_highmem_idx(j))
4419 nr_kernel_pages += realsize;
4420 nr_all_pages += realsize;
4422 zone->spanned_pages = size;
4423 zone->present_pages = realsize;
4424 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
4425 zone->compact_cached_free_pfn = zone->zone_start_pfn +
4426 zone->spanned_pages;
4427 zone->compact_cached_free_pfn &= ~(pageblock_nr_pages-1);
4431 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4433 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4435 zone->name = zone_names[j];
4436 spin_lock_init(&zone->lock);
4437 spin_lock_init(&zone->lru_lock);
4438 zone_seqlock_init(zone);
4439 zone->zone_pgdat = pgdat;
4441 zone_pcp_init(zone);
4442 lruvec_init(&zone->lruvec, zone);
4443 zap_zone_vm_stats(zone);
4445 #ifdef CONFIG_MEMORY_ISOLATION
4446 zone->nr_pageblock_isolate = 0;
4451 set_pageblock_order();
4452 setup_usemap(pgdat, zone, size);
4453 ret = init_currently_empty_zone(zone, zone_start_pfn,
4454 size, MEMMAP_EARLY);
4456 memmap_init(size, nid, j, zone_start_pfn);
4457 zone_start_pfn += size;
4461 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4463 /* Skip empty nodes */
4464 if (!pgdat->node_spanned_pages)
4467 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4468 /* ia64 gets its own node_mem_map, before this, without bootmem */
4469 if (!pgdat->node_mem_map) {
4470 unsigned long size, start, end;
4474 * The zone's endpoints aren't required to be MAX_ORDER
4475 * aligned but the node_mem_map endpoints must be in order
4476 * for the buddy allocator to function correctly.
4478 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4479 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4480 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4481 size = (end - start) * sizeof(struct page);
4482 map = alloc_remap(pgdat->node_id, size);
4484 map = alloc_bootmem_node_nopanic(pgdat, size);
4485 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4487 #ifndef CONFIG_NEED_MULTIPLE_NODES
4489 * With no DISCONTIG, the global mem_map is just set as node 0's
4491 if (pgdat == NODE_DATA(0)) {
4492 mem_map = NODE_DATA(0)->node_mem_map;
4493 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4494 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4495 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4496 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4499 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4502 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4503 unsigned long node_start_pfn, unsigned long *zholes_size)
4505 pg_data_t *pgdat = NODE_DATA(nid);
4507 pgdat->node_id = nid;
4508 pgdat->node_start_pfn = node_start_pfn;
4509 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4511 alloc_node_mem_map(pgdat);
4512 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4513 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4514 nid, (unsigned long)pgdat,
4515 (unsigned long)pgdat->node_mem_map);
4518 free_area_init_core(pgdat, zones_size, zholes_size);
4521 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4523 #if MAX_NUMNODES > 1
4525 * Figure out the number of possible node ids.
4527 static void __init setup_nr_node_ids(void)
4530 unsigned int highest = 0;
4532 for_each_node_mask(node, node_possible_map)
4534 nr_node_ids = highest + 1;
4537 static inline void setup_nr_node_ids(void)
4543 * node_map_pfn_alignment - determine the maximum internode alignment
4545 * This function should be called after node map is populated and sorted.
4546 * It calculates the maximum power of two alignment which can distinguish
4549 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4550 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4551 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4552 * shifted, 1GiB is enough and this function will indicate so.
4554 * This is used to test whether pfn -> nid mapping of the chosen memory
4555 * model has fine enough granularity to avoid incorrect mapping for the
4556 * populated node map.
4558 * Returns the determined alignment in pfn's. 0 if there is no alignment
4559 * requirement (single node).
4561 unsigned long __init node_map_pfn_alignment(void)
4563 unsigned long accl_mask = 0, last_end = 0;
4564 unsigned long start, end, mask;
4568 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4569 if (!start || last_nid < 0 || last_nid == nid) {
4576 * Start with a mask granular enough to pin-point to the
4577 * start pfn and tick off bits one-by-one until it becomes
4578 * too coarse to separate the current node from the last.
4580 mask = ~((1 << __ffs(start)) - 1);
4581 while (mask && last_end <= (start & (mask << 1)))
4584 /* accumulate all internode masks */
4588 /* convert mask to number of pages */
4589 return ~accl_mask + 1;
4592 /* Find the lowest pfn for a node */
4593 static unsigned long __init find_min_pfn_for_node(int nid)
4595 unsigned long min_pfn = ULONG_MAX;
4596 unsigned long start_pfn;
4599 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4600 min_pfn = min(min_pfn, start_pfn);
4602 if (min_pfn == ULONG_MAX) {
4604 "Could not find start_pfn for node %d\n", nid);
4612 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4614 * It returns the minimum PFN based on information provided via
4615 * add_active_range().
4617 unsigned long __init find_min_pfn_with_active_regions(void)
4619 return find_min_pfn_for_node(MAX_NUMNODES);
4623 * early_calculate_totalpages()
4624 * Sum pages in active regions for movable zone.
4625 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4627 static unsigned long __init early_calculate_totalpages(void)
4629 unsigned long totalpages = 0;
4630 unsigned long start_pfn, end_pfn;
4633 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4634 unsigned long pages = end_pfn - start_pfn;
4636 totalpages += pages;
4638 node_set_state(nid, N_HIGH_MEMORY);
4644 * Find the PFN the Movable zone begins in each node. Kernel memory
4645 * is spread evenly between nodes as long as the nodes have enough
4646 * memory. When they don't, some nodes will have more kernelcore than
4649 static void __init find_zone_movable_pfns_for_nodes(void)
4652 unsigned long usable_startpfn;
4653 unsigned long kernelcore_node, kernelcore_remaining;
4654 /* save the state before borrow the nodemask */
4655 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4656 unsigned long totalpages = early_calculate_totalpages();
4657 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4660 * If movablecore was specified, calculate what size of
4661 * kernelcore that corresponds so that memory usable for
4662 * any allocation type is evenly spread. If both kernelcore
4663 * and movablecore are specified, then the value of kernelcore
4664 * will be used for required_kernelcore if it's greater than
4665 * what movablecore would have allowed.
4667 if (required_movablecore) {
4668 unsigned long corepages;
4671 * Round-up so that ZONE_MOVABLE is at least as large as what
4672 * was requested by the user
4674 required_movablecore =
4675 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4676 corepages = totalpages - required_movablecore;
4678 required_kernelcore = max(required_kernelcore, corepages);
4681 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4682 if (!required_kernelcore)
4685 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4686 find_usable_zone_for_movable();
4687 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4690 /* Spread kernelcore memory as evenly as possible throughout nodes */
4691 kernelcore_node = required_kernelcore / usable_nodes;
4692 for_each_node_state(nid, N_HIGH_MEMORY) {
4693 unsigned long start_pfn, end_pfn;
4696 * Recalculate kernelcore_node if the division per node
4697 * now exceeds what is necessary to satisfy the requested
4698 * amount of memory for the kernel
4700 if (required_kernelcore < kernelcore_node)
4701 kernelcore_node = required_kernelcore / usable_nodes;
4704 * As the map is walked, we track how much memory is usable
4705 * by the kernel using kernelcore_remaining. When it is
4706 * 0, the rest of the node is usable by ZONE_MOVABLE
4708 kernelcore_remaining = kernelcore_node;
4710 /* Go through each range of PFNs within this node */
4711 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4712 unsigned long size_pages;
4714 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4715 if (start_pfn >= end_pfn)
4718 /* Account for what is only usable for kernelcore */
4719 if (start_pfn < usable_startpfn) {
4720 unsigned long kernel_pages;
4721 kernel_pages = min(end_pfn, usable_startpfn)
4724 kernelcore_remaining -= min(kernel_pages,
4725 kernelcore_remaining);
4726 required_kernelcore -= min(kernel_pages,
4727 required_kernelcore);
4729 /* Continue if range is now fully accounted */
4730 if (end_pfn <= usable_startpfn) {
4733 * Push zone_movable_pfn to the end so
4734 * that if we have to rebalance
4735 * kernelcore across nodes, we will
4736 * not double account here
4738 zone_movable_pfn[nid] = end_pfn;
4741 start_pfn = usable_startpfn;
4745 * The usable PFN range for ZONE_MOVABLE is from
4746 * start_pfn->end_pfn. Calculate size_pages as the
4747 * number of pages used as kernelcore
4749 size_pages = end_pfn - start_pfn;
4750 if (size_pages > kernelcore_remaining)
4751 size_pages = kernelcore_remaining;
4752 zone_movable_pfn[nid] = start_pfn + size_pages;
4755 * Some kernelcore has been met, update counts and
4756 * break if the kernelcore for this node has been
4759 required_kernelcore -= min(required_kernelcore,
4761 kernelcore_remaining -= size_pages;
4762 if (!kernelcore_remaining)
4768 * If there is still required_kernelcore, we do another pass with one
4769 * less node in the count. This will push zone_movable_pfn[nid] further
4770 * along on the nodes that still have memory until kernelcore is
4774 if (usable_nodes && required_kernelcore > usable_nodes)
4777 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4778 for (nid = 0; nid < MAX_NUMNODES; nid++)
4779 zone_movable_pfn[nid] =
4780 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4783 /* restore the node_state */
4784 node_states[N_HIGH_MEMORY] = saved_node_state;
4787 /* Any regular memory on that node ? */
4788 static void __init check_for_regular_memory(pg_data_t *pgdat)
4790 #ifdef CONFIG_HIGHMEM
4791 enum zone_type zone_type;
4793 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4794 struct zone *zone = &pgdat->node_zones[zone_type];
4795 if (zone->present_pages) {
4796 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4804 * free_area_init_nodes - Initialise all pg_data_t and zone data
4805 * @max_zone_pfn: an array of max PFNs for each zone
4807 * This will call free_area_init_node() for each active node in the system.
4808 * Using the page ranges provided by add_active_range(), the size of each
4809 * zone in each node and their holes is calculated. If the maximum PFN
4810 * between two adjacent zones match, it is assumed that the zone is empty.
4811 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4812 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4813 * starts where the previous one ended. For example, ZONE_DMA32 starts
4814 * at arch_max_dma_pfn.
4816 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4818 unsigned long start_pfn, end_pfn;
4821 /* Record where the zone boundaries are */
4822 memset(arch_zone_lowest_possible_pfn, 0,
4823 sizeof(arch_zone_lowest_possible_pfn));
4824 memset(arch_zone_highest_possible_pfn, 0,
4825 sizeof(arch_zone_highest_possible_pfn));
4826 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4827 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4828 for (i = 1; i < MAX_NR_ZONES; i++) {
4829 if (i == ZONE_MOVABLE)
4831 arch_zone_lowest_possible_pfn[i] =
4832 arch_zone_highest_possible_pfn[i-1];
4833 arch_zone_highest_possible_pfn[i] =
4834 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4836 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4837 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4839 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4840 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4841 find_zone_movable_pfns_for_nodes();
4843 /* Print out the zone ranges */
4844 printk("Zone ranges:\n");
4845 for (i = 0; i < MAX_NR_ZONES; i++) {
4846 if (i == ZONE_MOVABLE)
4848 printk(KERN_CONT " %-8s ", zone_names[i]);
4849 if (arch_zone_lowest_possible_pfn[i] ==
4850 arch_zone_highest_possible_pfn[i])
4851 printk(KERN_CONT "empty\n");
4853 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4854 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4855 (arch_zone_highest_possible_pfn[i]
4856 << PAGE_SHIFT) - 1);
4859 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4860 printk("Movable zone start for each node\n");
4861 for (i = 0; i < MAX_NUMNODES; i++) {
4862 if (zone_movable_pfn[i])
4863 printk(" Node %d: %#010lx\n", i,
4864 zone_movable_pfn[i] << PAGE_SHIFT);
4867 /* Print out the early_node_map[] */
4868 printk("Early memory node ranges\n");
4869 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4870 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
4871 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
4873 /* Initialise every node */
4874 mminit_verify_pageflags_layout();
4875 setup_nr_node_ids();
4876 for_each_online_node(nid) {
4877 pg_data_t *pgdat = NODE_DATA(nid);
4878 free_area_init_node(nid, NULL,
4879 find_min_pfn_for_node(nid), NULL);
4881 /* Any memory on that node */
4882 if (pgdat->node_present_pages)
4883 node_set_state(nid, N_HIGH_MEMORY);
4884 check_for_regular_memory(pgdat);
4888 static int __init cmdline_parse_core(char *p, unsigned long *core)
4890 unsigned long long coremem;
4894 coremem = memparse(p, &p);
4895 *core = coremem >> PAGE_SHIFT;
4897 /* Paranoid check that UL is enough for the coremem value */
4898 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4904 * kernelcore=size sets the amount of memory for use for allocations that
4905 * cannot be reclaimed or migrated.
4907 static int __init cmdline_parse_kernelcore(char *p)
4909 return cmdline_parse_core(p, &required_kernelcore);
4913 * movablecore=size sets the amount of memory for use for allocations that
4914 * can be reclaimed or migrated.
4916 static int __init cmdline_parse_movablecore(char *p)
4918 return cmdline_parse_core(p, &required_movablecore);
4921 early_param("kernelcore", cmdline_parse_kernelcore);
4922 early_param("movablecore", cmdline_parse_movablecore);
4924 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4927 * set_dma_reserve - set the specified number of pages reserved in the first zone
4928 * @new_dma_reserve: The number of pages to mark reserved
4930 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4931 * In the DMA zone, a significant percentage may be consumed by kernel image
4932 * and other unfreeable allocations which can skew the watermarks badly. This
4933 * function may optionally be used to account for unfreeable pages in the
4934 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4935 * smaller per-cpu batchsize.
4937 void __init set_dma_reserve(unsigned long new_dma_reserve)
4939 dma_reserve = new_dma_reserve;
4942 void __init free_area_init(unsigned long *zones_size)
4944 free_area_init_node(0, zones_size,
4945 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4948 static int page_alloc_cpu_notify(struct notifier_block *self,
4949 unsigned long action, void *hcpu)
4951 int cpu = (unsigned long)hcpu;
4953 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4954 lru_add_drain_cpu(cpu);
4958 * Spill the event counters of the dead processor
4959 * into the current processors event counters.
4960 * This artificially elevates the count of the current
4963 vm_events_fold_cpu(cpu);
4966 * Zero the differential counters of the dead processor
4967 * so that the vm statistics are consistent.
4969 * This is only okay since the processor is dead and cannot
4970 * race with what we are doing.
4972 refresh_cpu_vm_stats(cpu);
4977 void __init page_alloc_init(void)
4979 hotcpu_notifier(page_alloc_cpu_notify, 0);
4983 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4984 * or min_free_kbytes changes.
4986 static void calculate_totalreserve_pages(void)
4988 struct pglist_data *pgdat;
4989 unsigned long reserve_pages = 0;
4990 enum zone_type i, j;
4992 for_each_online_pgdat(pgdat) {
4993 for (i = 0; i < MAX_NR_ZONES; i++) {
4994 struct zone *zone = pgdat->node_zones + i;
4995 unsigned long max = 0;
4997 /* Find valid and maximum lowmem_reserve in the zone */
4998 for (j = i; j < MAX_NR_ZONES; j++) {
4999 if (zone->lowmem_reserve[j] > max)
5000 max = zone->lowmem_reserve[j];
5003 /* we treat the high watermark as reserved pages. */
5004 max += high_wmark_pages(zone);
5006 if (max > zone->present_pages)
5007 max = zone->present_pages;
5008 reserve_pages += max;
5010 * Lowmem reserves are not available to
5011 * GFP_HIGHUSER page cache allocations and
5012 * kswapd tries to balance zones to their high
5013 * watermark. As a result, neither should be
5014 * regarded as dirtyable memory, to prevent a
5015 * situation where reclaim has to clean pages
5016 * in order to balance the zones.
5018 zone->dirty_balance_reserve = max;
5021 dirty_balance_reserve = reserve_pages;
5022 totalreserve_pages = reserve_pages;
5026 * setup_per_zone_lowmem_reserve - called whenever
5027 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5028 * has a correct pages reserved value, so an adequate number of
5029 * pages are left in the zone after a successful __alloc_pages().
5031 static void setup_per_zone_lowmem_reserve(void)
5033 struct pglist_data *pgdat;
5034 enum zone_type j, idx;
5036 for_each_online_pgdat(pgdat) {
5037 for (j = 0; j < MAX_NR_ZONES; j++) {
5038 struct zone *zone = pgdat->node_zones + j;
5039 unsigned long present_pages = zone->present_pages;
5041 zone->lowmem_reserve[j] = 0;
5045 struct zone *lower_zone;
5049 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5050 sysctl_lowmem_reserve_ratio[idx] = 1;
5052 lower_zone = pgdat->node_zones + idx;
5053 lower_zone->lowmem_reserve[j] = present_pages /
5054 sysctl_lowmem_reserve_ratio[idx];
5055 present_pages += lower_zone->present_pages;
5060 /* update totalreserve_pages */
5061 calculate_totalreserve_pages();
5064 static void __setup_per_zone_wmarks(void)
5066 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5067 unsigned long lowmem_pages = 0;
5069 unsigned long flags;
5071 /* Calculate total number of !ZONE_HIGHMEM pages */
5072 for_each_zone(zone) {
5073 if (!is_highmem(zone))
5074 lowmem_pages += zone->present_pages;
5077 for_each_zone(zone) {
5080 spin_lock_irqsave(&zone->lock, flags);
5081 tmp = (u64)pages_min * zone->present_pages;
5082 do_div(tmp, lowmem_pages);
5083 if (is_highmem(zone)) {
5085 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5086 * need highmem pages, so cap pages_min to a small
5089 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5090 * deltas controls asynch page reclaim, and so should
5091 * not be capped for highmem.
5095 min_pages = zone->present_pages / 1024;
5096 if (min_pages < SWAP_CLUSTER_MAX)
5097 min_pages = SWAP_CLUSTER_MAX;
5098 if (min_pages > 128)
5100 zone->watermark[WMARK_MIN] = min_pages;
5103 * If it's a lowmem zone, reserve a number of pages
5104 * proportionate to the zone's size.
5106 zone->watermark[WMARK_MIN] = tmp;
5109 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5110 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5112 zone->watermark[WMARK_MIN] += cma_wmark_pages(zone);
5113 zone->watermark[WMARK_LOW] += cma_wmark_pages(zone);
5114 zone->watermark[WMARK_HIGH] += cma_wmark_pages(zone);
5116 setup_zone_migrate_reserve(zone);
5117 spin_unlock_irqrestore(&zone->lock, flags);
5120 /* update totalreserve_pages */
5121 calculate_totalreserve_pages();
5125 * setup_per_zone_wmarks - called when min_free_kbytes changes
5126 * or when memory is hot-{added|removed}
5128 * Ensures that the watermark[min,low,high] values for each zone are set
5129 * correctly with respect to min_free_kbytes.
5131 void setup_per_zone_wmarks(void)
5133 mutex_lock(&zonelists_mutex);
5134 __setup_per_zone_wmarks();
5135 mutex_unlock(&zonelists_mutex);
5139 * The inactive anon list should be small enough that the VM never has to
5140 * do too much work, but large enough that each inactive page has a chance
5141 * to be referenced again before it is swapped out.
5143 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5144 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5145 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5146 * the anonymous pages are kept on the inactive list.
5149 * memory ratio inactive anon
5150 * -------------------------------------
5159 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5161 unsigned int gb, ratio;
5163 /* Zone size in gigabytes */
5164 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5166 ratio = int_sqrt(10 * gb);
5170 zone->inactive_ratio = ratio;
5173 static void __meminit setup_per_zone_inactive_ratio(void)
5178 calculate_zone_inactive_ratio(zone);
5182 * Initialise min_free_kbytes.
5184 * For small machines we want it small (128k min). For large machines
5185 * we want it large (64MB max). But it is not linear, because network
5186 * bandwidth does not increase linearly with machine size. We use
5188 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5189 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5205 int __meminit init_per_zone_wmark_min(void)
5207 unsigned long lowmem_kbytes;
5209 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5211 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5212 if (min_free_kbytes < 128)
5213 min_free_kbytes = 128;
5214 if (min_free_kbytes > 65536)
5215 min_free_kbytes = 65536;
5216 setup_per_zone_wmarks();
5217 refresh_zone_stat_thresholds();
5218 setup_per_zone_lowmem_reserve();
5219 setup_per_zone_inactive_ratio();
5222 module_init(init_per_zone_wmark_min)
5225 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5226 * that we can call two helper functions whenever min_free_kbytes
5229 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5230 void __user *buffer, size_t *length, loff_t *ppos)
5232 proc_dointvec(table, write, buffer, length, ppos);
5234 setup_per_zone_wmarks();
5239 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5240 void __user *buffer, size_t *length, loff_t *ppos)
5245 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5250 zone->min_unmapped_pages = (zone->present_pages *
5251 sysctl_min_unmapped_ratio) / 100;
5255 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5256 void __user *buffer, size_t *length, loff_t *ppos)
5261 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5266 zone->min_slab_pages = (zone->present_pages *
5267 sysctl_min_slab_ratio) / 100;
5273 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5274 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5275 * whenever sysctl_lowmem_reserve_ratio changes.
5277 * The reserve ratio obviously has absolutely no relation with the
5278 * minimum watermarks. The lowmem reserve ratio can only make sense
5279 * if in function of the boot time zone sizes.
5281 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5282 void __user *buffer, size_t *length, loff_t *ppos)
5284 proc_dointvec_minmax(table, write, buffer, length, ppos);
5285 setup_per_zone_lowmem_reserve();
5290 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5291 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5292 * can have before it gets flushed back to buddy allocator.
5295 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5296 void __user *buffer, size_t *length, loff_t *ppos)
5302 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5303 if (!write || (ret < 0))
5305 for_each_populated_zone(zone) {
5306 for_each_possible_cpu(cpu) {
5308 high = zone->present_pages / percpu_pagelist_fraction;
5309 setup_pagelist_highmark(
5310 per_cpu_ptr(zone->pageset, cpu), high);
5316 int hashdist = HASHDIST_DEFAULT;
5319 static int __init set_hashdist(char *str)
5323 hashdist = simple_strtoul(str, &str, 0);
5326 __setup("hashdist=", set_hashdist);
5330 * allocate a large system hash table from bootmem
5331 * - it is assumed that the hash table must contain an exact power-of-2
5332 * quantity of entries
5333 * - limit is the number of hash buckets, not the total allocation size
5335 void *__init alloc_large_system_hash(const char *tablename,
5336 unsigned long bucketsize,
5337 unsigned long numentries,
5340 unsigned int *_hash_shift,
5341 unsigned int *_hash_mask,
5342 unsigned long low_limit,
5343 unsigned long high_limit)
5345 unsigned long long max = high_limit;
5346 unsigned long log2qty, size;
5349 /* allow the kernel cmdline to have a say */
5351 /* round applicable memory size up to nearest megabyte */
5352 numentries = nr_kernel_pages;
5353 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5354 numentries >>= 20 - PAGE_SHIFT;
5355 numentries <<= 20 - PAGE_SHIFT;
5357 /* limit to 1 bucket per 2^scale bytes of low memory */
5358 if (scale > PAGE_SHIFT)
5359 numentries >>= (scale - PAGE_SHIFT);
5361 numentries <<= (PAGE_SHIFT - scale);
5363 /* Make sure we've got at least a 0-order allocation.. */
5364 if (unlikely(flags & HASH_SMALL)) {
5365 /* Makes no sense without HASH_EARLY */
5366 WARN_ON(!(flags & HASH_EARLY));
5367 if (!(numentries >> *_hash_shift)) {
5368 numentries = 1UL << *_hash_shift;
5369 BUG_ON(!numentries);
5371 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5372 numentries = PAGE_SIZE / bucketsize;
5374 numentries = roundup_pow_of_two(numentries);
5376 /* limit allocation size to 1/16 total memory by default */
5378 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5379 do_div(max, bucketsize);
5381 max = min(max, 0x80000000ULL);
5383 if (numentries < low_limit)
5384 numentries = low_limit;
5385 if (numentries > max)
5388 log2qty = ilog2(numentries);
5391 size = bucketsize << log2qty;
5392 if (flags & HASH_EARLY)
5393 table = alloc_bootmem_nopanic(size);
5395 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5398 * If bucketsize is not a power-of-two, we may free
5399 * some pages at the end of hash table which
5400 * alloc_pages_exact() automatically does
5402 if (get_order(size) < MAX_ORDER) {
5403 table = alloc_pages_exact(size, GFP_ATOMIC);
5404 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5407 } while (!table && size > PAGE_SIZE && --log2qty);
5410 panic("Failed to allocate %s hash table\n", tablename);
5412 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5415 ilog2(size) - PAGE_SHIFT,
5419 *_hash_shift = log2qty;
5421 *_hash_mask = (1 << log2qty) - 1;
5426 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5427 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5430 #ifdef CONFIG_SPARSEMEM
5431 return __pfn_to_section(pfn)->pageblock_flags;
5433 return zone->pageblock_flags;
5434 #endif /* CONFIG_SPARSEMEM */
5437 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5439 #ifdef CONFIG_SPARSEMEM
5440 pfn &= (PAGES_PER_SECTION-1);
5441 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5443 pfn = pfn - zone->zone_start_pfn;
5444 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5445 #endif /* CONFIG_SPARSEMEM */
5449 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5450 * @page: The page within the block of interest
5451 * @start_bitidx: The first bit of interest to retrieve
5452 * @end_bitidx: The last bit of interest
5453 * returns pageblock_bits flags
5455 unsigned long get_pageblock_flags_group(struct page *page,
5456 int start_bitidx, int end_bitidx)
5459 unsigned long *bitmap;
5460 unsigned long pfn, bitidx;
5461 unsigned long flags = 0;
5462 unsigned long value = 1;
5464 zone = page_zone(page);
5465 pfn = page_to_pfn(page);
5466 bitmap = get_pageblock_bitmap(zone, pfn);
5467 bitidx = pfn_to_bitidx(zone, pfn);
5469 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5470 if (test_bit(bitidx + start_bitidx, bitmap))
5477 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5478 * @page: The page within the block of interest
5479 * @start_bitidx: The first bit of interest
5480 * @end_bitidx: The last bit of interest
5481 * @flags: The flags to set
5483 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5484 int start_bitidx, int end_bitidx)
5487 unsigned long *bitmap;
5488 unsigned long pfn, bitidx;
5489 unsigned long value = 1;
5491 zone = page_zone(page);
5492 pfn = page_to_pfn(page);
5493 bitmap = get_pageblock_bitmap(zone, pfn);
5494 bitidx = pfn_to_bitidx(zone, pfn);
5495 VM_BUG_ON(pfn < zone->zone_start_pfn);
5496 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5498 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5500 __set_bit(bitidx + start_bitidx, bitmap);
5502 __clear_bit(bitidx + start_bitidx, bitmap);
5506 * This function checks whether pageblock includes unmovable pages or not.
5507 * If @count is not zero, it is okay to include less @count unmovable pages
5509 * PageLRU check wihtout isolation or lru_lock could race so that
5510 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5511 * expect this function should be exact.
5513 bool has_unmovable_pages(struct zone *zone, struct page *page, int count)
5515 unsigned long pfn, iter, found;
5519 * For avoiding noise data, lru_add_drain_all() should be called
5520 * If ZONE_MOVABLE, the zone never contains unmovable pages
5522 if (zone_idx(zone) == ZONE_MOVABLE)
5524 mt = get_pageblock_migratetype(page);
5525 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5528 pfn = page_to_pfn(page);
5529 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5530 unsigned long check = pfn + iter;
5532 if (!pfn_valid_within(check))
5535 page = pfn_to_page(check);
5537 * We can't use page_count without pin a page
5538 * because another CPU can free compound page.
5539 * This check already skips compound tails of THP
5540 * because their page->_count is zero at all time.
5542 if (!atomic_read(&page->_count)) {
5543 if (PageBuddy(page))
5544 iter += (1 << page_order(page)) - 1;
5551 * If there are RECLAIMABLE pages, we need to check it.
5552 * But now, memory offline itself doesn't call shrink_slab()
5553 * and it still to be fixed.
5556 * If the page is not RAM, page_count()should be 0.
5557 * we don't need more check. This is an _used_ not-movable page.
5559 * The problematic thing here is PG_reserved pages. PG_reserved
5560 * is set to both of a memory hole page and a _used_ kernel
5569 bool is_pageblock_removable_nolock(struct page *page)
5575 * We have to be careful here because we are iterating over memory
5576 * sections which are not zone aware so we might end up outside of
5577 * the zone but still within the section.
5578 * We have to take care about the node as well. If the node is offline
5579 * its NODE_DATA will be NULL - see page_zone.
5581 if (!node_online(page_to_nid(page)))
5584 zone = page_zone(page);
5585 pfn = page_to_pfn(page);
5586 if (zone->zone_start_pfn > pfn ||
5587 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5590 return !has_unmovable_pages(zone, page, 0);
5595 static unsigned long pfn_max_align_down(unsigned long pfn)
5597 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5598 pageblock_nr_pages) - 1);
5601 static unsigned long pfn_max_align_up(unsigned long pfn)
5603 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5604 pageblock_nr_pages));
5607 static struct page *
5608 __alloc_contig_migrate_alloc(struct page *page, unsigned long private,
5611 gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
5613 if (PageHighMem(page))
5614 gfp_mask |= __GFP_HIGHMEM;
5616 return alloc_page(gfp_mask);
5619 /* [start, end) must belong to a single zone. */
5620 static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5622 /* This function is based on compact_zone() from compaction.c. */
5624 unsigned long pfn = start;
5625 unsigned int tries = 0;
5628 struct compact_control cc = {
5629 .nr_migratepages = 0,
5631 .zone = page_zone(pfn_to_page(start)),
5634 INIT_LIST_HEAD(&cc.migratepages);
5636 migrate_prep_local();
5638 while (pfn < end || !list_empty(&cc.migratepages)) {
5639 if (fatal_signal_pending(current)) {
5644 if (list_empty(&cc.migratepages)) {
5645 cc.nr_migratepages = 0;
5646 pfn = isolate_migratepages_range(cc.zone, &cc,
5653 } else if (++tries == 5) {
5654 ret = ret < 0 ? ret : -EBUSY;
5658 ret = migrate_pages(&cc.migratepages,
5659 __alloc_contig_migrate_alloc,
5660 0, false, MIGRATE_SYNC);
5663 putback_lru_pages(&cc.migratepages);
5664 return ret > 0 ? 0 : ret;
5668 * Update zone's cma pages counter used for watermark level calculation.
5670 static inline void __update_cma_watermarks(struct zone *zone, int count)
5672 unsigned long flags;
5673 spin_lock_irqsave(&zone->lock, flags);
5674 zone->min_cma_pages += count;
5675 spin_unlock_irqrestore(&zone->lock, flags);
5676 setup_per_zone_wmarks();
5680 * Trigger memory pressure bump to reclaim some pages in order to be able to
5681 * allocate 'count' pages in single page units. Does similar work as
5682 *__alloc_pages_slowpath() function.
5684 static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
5686 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
5687 struct zonelist *zonelist = node_zonelist(0, gfp_mask);
5688 int did_some_progress = 0;
5692 * Increase level of watermarks to force kswapd do his job
5693 * to stabilise at new watermark level.
5695 __update_cma_watermarks(zone, count);
5697 /* Obey watermarks as if the page was being allocated */
5698 while (!zone_watermark_ok(zone, 0, low_wmark_pages(zone), 0, 0)) {
5699 wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));
5701 did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
5703 if (!did_some_progress) {
5704 /* Exhausted what can be done so it's blamo time */
5705 out_of_memory(zonelist, gfp_mask, order, NULL, false);
5709 /* Restore original watermark levels. */
5710 __update_cma_watermarks(zone, -count);
5716 * alloc_contig_range() -- tries to allocate given range of pages
5717 * @start: start PFN to allocate
5718 * @end: one-past-the-last PFN to allocate
5719 * @migratetype: migratetype of the underlaying pageblocks (either
5720 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5721 * in range must have the same migratetype and it must
5722 * be either of the two.
5724 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5725 * aligned, however it's the caller's responsibility to guarantee that
5726 * we are the only thread that changes migrate type of pageblocks the
5729 * The PFN range must belong to a single zone.
5731 * Returns zero on success or negative error code. On success all
5732 * pages which PFN is in [start, end) are allocated for the caller and
5733 * need to be freed with free_contig_range().
5735 int alloc_contig_range(unsigned long start, unsigned long end,
5736 unsigned migratetype)
5738 struct zone *zone = page_zone(pfn_to_page(start));
5739 unsigned long outer_start, outer_end;
5743 * What we do here is we mark all pageblocks in range as
5744 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5745 * have different sizes, and due to the way page allocator
5746 * work, we align the range to biggest of the two pages so
5747 * that page allocator won't try to merge buddies from
5748 * different pageblocks and change MIGRATE_ISOLATE to some
5749 * other migration type.
5751 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5752 * migrate the pages from an unaligned range (ie. pages that
5753 * we are interested in). This will put all the pages in
5754 * range back to page allocator as MIGRATE_ISOLATE.
5756 * When this is done, we take the pages in range from page
5757 * allocator removing them from the buddy system. This way
5758 * page allocator will never consider using them.
5760 * This lets us mark the pageblocks back as
5761 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5762 * aligned range but not in the unaligned, original range are
5763 * put back to page allocator so that buddy can use them.
5766 ret = start_isolate_page_range(pfn_max_align_down(start),
5767 pfn_max_align_up(end), migratetype);
5771 ret = __alloc_contig_migrate_range(start, end);
5776 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5777 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5778 * more, all pages in [start, end) are free in page allocator.
5779 * What we are going to do is to allocate all pages from
5780 * [start, end) (that is remove them from page allocator).
5782 * The only problem is that pages at the beginning and at the
5783 * end of interesting range may be not aligned with pages that
5784 * page allocator holds, ie. they can be part of higher order
5785 * pages. Because of this, we reserve the bigger range and
5786 * once this is done free the pages we are not interested in.
5788 * We don't have to hold zone->lock here because the pages are
5789 * isolated thus they won't get removed from buddy.
5792 lru_add_drain_all();
5796 outer_start = start;
5797 while (!PageBuddy(pfn_to_page(outer_start))) {
5798 if (++order >= MAX_ORDER) {
5802 outer_start &= ~0UL << order;
5805 /* Make sure the range is really isolated. */
5806 if (test_pages_isolated(outer_start, end)) {
5807 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5814 * Reclaim enough pages to make sure that contiguous allocation
5815 * will not starve the system.
5817 __reclaim_pages(zone, GFP_HIGHUSER_MOVABLE, end-start);
5819 /* Grab isolated pages from freelists. */
5820 outer_end = isolate_freepages_range(outer_start, end);
5826 /* Free head and tail (if any) */
5827 if (start != outer_start)
5828 free_contig_range(outer_start, start - outer_start);
5829 if (end != outer_end)
5830 free_contig_range(end, outer_end - end);
5833 undo_isolate_page_range(pfn_max_align_down(start),
5834 pfn_max_align_up(end), migratetype);
5838 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5840 for (; nr_pages--; ++pfn)
5841 __free_page(pfn_to_page(pfn));
5845 #ifdef CONFIG_MEMORY_HOTPLUG
5846 static int __meminit __zone_pcp_update(void *data)
5848 struct zone *zone = data;
5850 unsigned long batch = zone_batchsize(zone), flags;
5852 for_each_possible_cpu(cpu) {
5853 struct per_cpu_pageset *pset;
5854 struct per_cpu_pages *pcp;
5856 pset = per_cpu_ptr(zone->pageset, cpu);
5859 local_irq_save(flags);
5861 free_pcppages_bulk(zone, pcp->count, pcp);
5862 setup_pageset(pset, batch);
5863 local_irq_restore(flags);
5868 void __meminit zone_pcp_update(struct zone *zone)
5870 stop_machine(__zone_pcp_update, zone, NULL);
5874 #ifdef CONFIG_MEMORY_HOTREMOVE
5875 void zone_pcp_reset(struct zone *zone)
5877 unsigned long flags;
5879 /* avoid races with drain_pages() */
5880 local_irq_save(flags);
5881 if (zone->pageset != &boot_pageset) {
5882 free_percpu(zone->pageset);
5883 zone->pageset = &boot_pageset;
5885 local_irq_restore(flags);
5889 * All pages in the range must be isolated before calling this.
5892 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5898 unsigned long flags;
5899 /* find the first valid pfn */
5900 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5905 zone = page_zone(pfn_to_page(pfn));
5906 spin_lock_irqsave(&zone->lock, flags);
5908 while (pfn < end_pfn) {
5909 if (!pfn_valid(pfn)) {
5913 page = pfn_to_page(pfn);
5914 BUG_ON(page_count(page));
5915 BUG_ON(!PageBuddy(page));
5916 order = page_order(page);
5917 #ifdef CONFIG_DEBUG_VM
5918 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5919 pfn, 1 << order, end_pfn);
5921 list_del(&page->lru);
5922 rmv_page_order(page);
5923 zone->free_area[order].nr_free--;
5924 __mod_zone_page_state(zone, NR_FREE_PAGES,
5926 for (i = 0; i < (1 << order); i++)
5927 SetPageReserved((page+i));
5928 pfn += (1 << order);
5930 spin_unlock_irqrestore(&zone->lock, flags);
5934 #ifdef CONFIG_MEMORY_FAILURE
5935 bool is_free_buddy_page(struct page *page)
5937 struct zone *zone = page_zone(page);
5938 unsigned long pfn = page_to_pfn(page);
5939 unsigned long flags;
5942 spin_lock_irqsave(&zone->lock, flags);
5943 for (order = 0; order < MAX_ORDER; order++) {
5944 struct page *page_head = page - (pfn & ((1 << order) - 1));
5946 if (PageBuddy(page_head) && page_order(page_head) >= order)
5949 spin_unlock_irqrestore(&zone->lock, flags);
5951 return order < MAX_ORDER;
5955 static const struct trace_print_flags pageflag_names[] = {
5956 {1UL << PG_locked, "locked" },
5957 {1UL << PG_error, "error" },
5958 {1UL << PG_referenced, "referenced" },
5959 {1UL << PG_uptodate, "uptodate" },
5960 {1UL << PG_dirty, "dirty" },
5961 {1UL << PG_lru, "lru" },
5962 {1UL << PG_active, "active" },
5963 {1UL << PG_slab, "slab" },
5964 {1UL << PG_owner_priv_1, "owner_priv_1" },
5965 {1UL << PG_arch_1, "arch_1" },
5966 {1UL << PG_reserved, "reserved" },
5967 {1UL << PG_private, "private" },
5968 {1UL << PG_private_2, "private_2" },
5969 {1UL << PG_writeback, "writeback" },
5970 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5971 {1UL << PG_head, "head" },
5972 {1UL << PG_tail, "tail" },
5974 {1UL << PG_compound, "compound" },
5976 {1UL << PG_swapcache, "swapcache" },
5977 {1UL << PG_mappedtodisk, "mappedtodisk" },
5978 {1UL << PG_reclaim, "reclaim" },
5979 {1UL << PG_swapbacked, "swapbacked" },
5980 {1UL << PG_unevictable, "unevictable" },
5982 {1UL << PG_mlocked, "mlocked" },
5984 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5985 {1UL << PG_uncached, "uncached" },
5987 #ifdef CONFIG_MEMORY_FAILURE
5988 {1UL << PG_hwpoison, "hwpoison" },
5990 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5991 {1UL << PG_compound_lock, "compound_lock" },
5995 static void dump_page_flags(unsigned long flags)
5997 const char *delim = "";
6001 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6003 printk(KERN_ALERT "page flags: %#lx(", flags);
6005 /* remove zone id */
6006 flags &= (1UL << NR_PAGEFLAGS) - 1;
6008 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6010 mask = pageflag_names[i].mask;
6011 if ((flags & mask) != mask)
6015 printk("%s%s", delim, pageflag_names[i].name);
6019 /* check for left over flags */
6021 printk("%s%#lx", delim, flags);
6026 void dump_page(struct page *page)
6029 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6030 page, atomic_read(&page->_count), page_mapcount(page),
6031 page->mapping, page->index);
6032 dump_page_flags(page->flags);
6033 mem_cgroup_print_bad_page(page);