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/memory.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/memcontrol.h>
59 #include <linux/prefetch.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
63 #include <asm/tlbflush.h>
64 #include <asm/div64.h>
67 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
68 DEFINE_PER_CPU(int, numa_node);
69 EXPORT_PER_CPU_SYMBOL(numa_node);
72 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
74 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
75 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
76 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
77 * defined in <linux/topology.h>.
79 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
80 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
84 * Array of node states.
86 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
87 [N_POSSIBLE] = NODE_MASK_ALL,
88 [N_ONLINE] = { { [0] = 1UL } },
90 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
92 [N_HIGH_MEMORY] = { { [0] = 1UL } },
94 [N_CPU] = { { [0] = 1UL } },
97 EXPORT_SYMBOL(node_states);
99 unsigned long totalram_pages __read_mostly;
100 unsigned long totalreserve_pages __read_mostly;
102 * When calculating the number of globally allowed dirty pages, there
103 * is a certain number of per-zone reserves that should not be
104 * considered dirtyable memory. This is the sum of those reserves
105 * over all existing zones that contribute dirtyable memory.
107 unsigned long dirty_balance_reserve __read_mostly;
109 int percpu_pagelist_fraction;
110 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
112 #ifdef CONFIG_PM_SLEEP
114 * The following functions are used by the suspend/hibernate code to temporarily
115 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
116 * while devices are suspended. To avoid races with the suspend/hibernate code,
117 * they should always be called with pm_mutex held (gfp_allowed_mask also should
118 * only be modified with pm_mutex held, unless the suspend/hibernate code is
119 * guaranteed not to run in parallel with that modification).
122 static gfp_t saved_gfp_mask;
124 void pm_restore_gfp_mask(void)
126 WARN_ON(!mutex_is_locked(&pm_mutex));
127 if (saved_gfp_mask) {
128 gfp_allowed_mask = saved_gfp_mask;
133 void pm_restrict_gfp_mask(void)
135 WARN_ON(!mutex_is_locked(&pm_mutex));
136 WARN_ON(saved_gfp_mask);
137 saved_gfp_mask = gfp_allowed_mask;
138 gfp_allowed_mask &= ~GFP_IOFS;
141 bool pm_suspended_storage(void)
143 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
147 #endif /* CONFIG_PM_SLEEP */
149 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
150 int pageblock_order __read_mostly;
153 static void __free_pages_ok(struct page *page, unsigned int order);
156 * results with 256, 32 in the lowmem_reserve sysctl:
157 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
158 * 1G machine -> (16M dma, 784M normal, 224M high)
159 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
160 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
161 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
163 * TBD: should special case ZONE_DMA32 machines here - in those we normally
164 * don't need any ZONE_NORMAL reservation
166 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
167 #ifdef CONFIG_ZONE_DMA
170 #ifdef CONFIG_ZONE_DMA32
173 #ifdef CONFIG_HIGHMEM
179 EXPORT_SYMBOL(totalram_pages);
181 static char * const zone_names[MAX_NR_ZONES] = {
182 #ifdef CONFIG_ZONE_DMA
185 #ifdef CONFIG_ZONE_DMA32
189 #ifdef CONFIG_HIGHMEM
195 int min_free_kbytes = 1024;
197 static unsigned long __meminitdata nr_kernel_pages;
198 static unsigned long __meminitdata nr_all_pages;
199 static unsigned long __meminitdata dma_reserve;
201 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
202 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
203 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
204 static unsigned long __initdata required_kernelcore;
205 static unsigned long __initdata required_movablecore;
206 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
208 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
210 EXPORT_SYMBOL(movable_zone);
211 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
214 int nr_node_ids __read_mostly = MAX_NUMNODES;
215 int nr_online_nodes __read_mostly = 1;
216 EXPORT_SYMBOL(nr_node_ids);
217 EXPORT_SYMBOL(nr_online_nodes);
220 int page_group_by_mobility_disabled __read_mostly;
222 static void set_pageblock_migratetype(struct page *page, int migratetype)
225 if (unlikely(page_group_by_mobility_disabled))
226 migratetype = MIGRATE_UNMOVABLE;
228 set_pageblock_flags_group(page, (unsigned long)migratetype,
229 PB_migrate, PB_migrate_end);
232 bool oom_killer_disabled __read_mostly;
234 #ifdef CONFIG_DEBUG_VM
235 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
239 unsigned long pfn = page_to_pfn(page);
242 seq = zone_span_seqbegin(zone);
243 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
245 else if (pfn < zone->zone_start_pfn)
247 } while (zone_span_seqretry(zone, seq));
252 static int page_is_consistent(struct zone *zone, struct page *page)
254 if (!pfn_valid_within(page_to_pfn(page)))
256 if (zone != page_zone(page))
262 * Temporary debugging check for pages not lying within a given zone.
264 static int bad_range(struct zone *zone, struct page *page)
266 if (page_outside_zone_boundaries(zone, page))
268 if (!page_is_consistent(zone, page))
274 static inline int bad_range(struct zone *zone, struct page *page)
280 static void bad_page(struct page *page)
282 static unsigned long resume;
283 static unsigned long nr_shown;
284 static unsigned long nr_unshown;
286 /* Don't complain about poisoned pages */
287 if (PageHWPoison(page)) {
288 reset_page_mapcount(page); /* remove PageBuddy */
293 * Allow a burst of 60 reports, then keep quiet for that minute;
294 * or allow a steady drip of one report per second.
296 if (nr_shown == 60) {
297 if (time_before(jiffies, resume)) {
303 "BUG: Bad page state: %lu messages suppressed\n",
310 resume = jiffies + 60 * HZ;
312 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
313 current->comm, page_to_pfn(page));
319 /* Leave bad fields for debug, except PageBuddy could make trouble */
320 reset_page_mapcount(page); /* remove PageBuddy */
321 add_taint(TAINT_BAD_PAGE);
325 * Higher-order pages are called "compound pages". They are structured thusly:
327 * The first PAGE_SIZE page is called the "head page".
329 * The remaining PAGE_SIZE pages are called "tail pages".
331 * All pages have PG_compound set. All tail pages have their ->first_page
332 * pointing at the head page.
334 * The first tail page's ->lru.next holds the address of the compound page's
335 * put_page() function. Its ->lru.prev holds the order of allocation.
336 * This usage means that zero-order pages may not be compound.
339 static void free_compound_page(struct page *page)
341 __free_pages_ok(page, compound_order(page));
344 void prep_compound_page(struct page *page, unsigned long order)
347 int nr_pages = 1 << order;
349 set_compound_page_dtor(page, free_compound_page);
350 set_compound_order(page, order);
352 for (i = 1; i < nr_pages; i++) {
353 struct page *p = page + i;
355 set_page_count(p, 0);
356 p->first_page = page;
360 /* update __split_huge_page_refcount if you change this function */
361 static int destroy_compound_page(struct page *page, unsigned long order)
364 int nr_pages = 1 << order;
367 if (unlikely(compound_order(page) != order) ||
368 unlikely(!PageHead(page))) {
373 __ClearPageHead(page);
375 for (i = 1; i < nr_pages; i++) {
376 struct page *p = page + i;
378 if (unlikely(!PageTail(p) || (p->first_page != page))) {
388 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
393 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
394 * and __GFP_HIGHMEM from hard or soft interrupt context.
396 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
397 for (i = 0; i < (1 << order); i++)
398 clear_highpage(page + i);
401 #ifdef CONFIG_DEBUG_PAGEALLOC
402 unsigned int _debug_guardpage_minorder;
404 static int __init debug_guardpage_minorder_setup(char *buf)
408 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
409 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
412 _debug_guardpage_minorder = res;
413 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
416 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
418 static inline void set_page_guard_flag(struct page *page)
420 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
423 static inline void clear_page_guard_flag(struct page *page)
425 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
428 static inline void set_page_guard_flag(struct page *page) { }
429 static inline void clear_page_guard_flag(struct page *page) { }
432 static inline void set_page_order(struct page *page, int order)
434 set_page_private(page, order);
435 __SetPageBuddy(page);
438 static inline void rmv_page_order(struct page *page)
440 __ClearPageBuddy(page);
441 set_page_private(page, 0);
445 * Locate the struct page for both the matching buddy in our
446 * pair (buddy1) and the combined O(n+1) page they form (page).
448 * 1) Any buddy B1 will have an order O twin B2 which satisfies
449 * the following equation:
451 * For example, if the starting buddy (buddy2) is #8 its order
453 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
455 * 2) Any buddy B will have an order O+1 parent P which
456 * satisfies the following equation:
459 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
461 static inline unsigned long
462 __find_buddy_index(unsigned long page_idx, unsigned int order)
464 return page_idx ^ (1 << order);
468 * This function checks whether a page is free && is the buddy
469 * we can do coalesce a page and its buddy if
470 * (a) the buddy is not in a hole &&
471 * (b) the buddy is in the buddy system &&
472 * (c) a page and its buddy have the same order &&
473 * (d) a page and its buddy are in the same zone.
475 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
476 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
478 * For recording page's order, we use page_private(page).
480 static inline int page_is_buddy(struct page *page, struct page *buddy,
483 if (!pfn_valid_within(page_to_pfn(buddy)))
486 if (page_zone_id(page) != page_zone_id(buddy))
489 if (page_is_guard(buddy) && page_order(buddy) == order) {
490 VM_BUG_ON(page_count(buddy) != 0);
494 if (PageBuddy(buddy) && page_order(buddy) == order) {
495 VM_BUG_ON(page_count(buddy) != 0);
502 * Freeing function for a buddy system allocator.
504 * The concept of a buddy system is to maintain direct-mapped table
505 * (containing bit values) for memory blocks of various "orders".
506 * The bottom level table contains the map for the smallest allocatable
507 * units of memory (here, pages), and each level above it describes
508 * pairs of units from the levels below, hence, "buddies".
509 * At a high level, all that happens here is marking the table entry
510 * at the bottom level available, and propagating the changes upward
511 * as necessary, plus some accounting needed to play nicely with other
512 * parts of the VM system.
513 * At each level, we keep a list of pages, which are heads of continuous
514 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
515 * order is recorded in page_private(page) field.
516 * So when we are allocating or freeing one, we can derive the state of the
517 * other. That is, if we allocate a small block, and both were
518 * free, the remainder of the region must be split into blocks.
519 * If a block is freed, and its buddy is also free, then this
520 * triggers coalescing into a block of larger size.
525 static inline void __free_one_page(struct page *page,
526 struct zone *zone, unsigned int order,
529 unsigned long page_idx;
530 unsigned long combined_idx;
531 unsigned long uninitialized_var(buddy_idx);
534 if (unlikely(PageCompound(page)))
535 if (unlikely(destroy_compound_page(page, order)))
538 VM_BUG_ON(migratetype == -1);
540 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
542 VM_BUG_ON(page_idx & ((1 << order) - 1));
543 VM_BUG_ON(bad_range(zone, page));
545 while (order < MAX_ORDER-1) {
546 buddy_idx = __find_buddy_index(page_idx, order);
547 buddy = page + (buddy_idx - page_idx);
548 if (!page_is_buddy(page, buddy, order))
551 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
552 * merge with it and move up one order.
554 if (page_is_guard(buddy)) {
555 clear_page_guard_flag(buddy);
556 set_page_private(page, 0);
557 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
559 list_del(&buddy->lru);
560 zone->free_area[order].nr_free--;
561 rmv_page_order(buddy);
563 combined_idx = buddy_idx & page_idx;
564 page = page + (combined_idx - page_idx);
565 page_idx = combined_idx;
568 set_page_order(page, order);
571 * If this is not the largest possible page, check if the buddy
572 * of the next-highest order is free. If it is, it's possible
573 * that pages are being freed that will coalesce soon. In case,
574 * that is happening, add the free page to the tail of the list
575 * so it's less likely to be used soon and more likely to be merged
576 * as a higher order page
578 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
579 struct page *higher_page, *higher_buddy;
580 combined_idx = buddy_idx & page_idx;
581 higher_page = page + (combined_idx - page_idx);
582 buddy_idx = __find_buddy_index(combined_idx, order + 1);
583 higher_buddy = page + (buddy_idx - combined_idx);
584 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
585 list_add_tail(&page->lru,
586 &zone->free_area[order].free_list[migratetype]);
591 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
593 zone->free_area[order].nr_free++;
597 * free_page_mlock() -- clean up attempts to free and mlocked() page.
598 * Page should not be on lru, so no need to fix that up.
599 * free_pages_check() will verify...
601 static inline void free_page_mlock(struct page *page)
603 __dec_zone_page_state(page, NR_MLOCK);
604 __count_vm_event(UNEVICTABLE_MLOCKFREED);
607 static inline int free_pages_check(struct page *page)
609 if (unlikely(page_mapcount(page) |
610 (page->mapping != NULL) |
611 (atomic_read(&page->_count) != 0) |
612 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
613 (mem_cgroup_bad_page_check(page)))) {
617 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
618 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
623 * Frees a number of pages from the PCP lists
624 * Assumes all pages on list are in same zone, and of same order.
625 * count is the number of pages to free.
627 * If the zone was previously in an "all pages pinned" state then look to
628 * see if this freeing clears that state.
630 * And clear the zone's pages_scanned counter, to hold off the "all pages are
631 * pinned" detection logic.
633 static void free_pcppages_bulk(struct zone *zone, int count,
634 struct per_cpu_pages *pcp)
640 spin_lock(&zone->lock);
641 zone->all_unreclaimable = 0;
642 zone->pages_scanned = 0;
646 struct list_head *list;
649 * Remove pages from lists in a round-robin fashion. A
650 * batch_free count is maintained that is incremented when an
651 * empty list is encountered. This is so more pages are freed
652 * off fuller lists instead of spinning excessively around empty
657 if (++migratetype == MIGRATE_PCPTYPES)
659 list = &pcp->lists[migratetype];
660 } while (list_empty(list));
662 /* This is the only non-empty list. Free them all. */
663 if (batch_free == MIGRATE_PCPTYPES)
664 batch_free = to_free;
667 page = list_entry(list->prev, struct page, lru);
668 /* must delete as __free_one_page list manipulates */
669 list_del(&page->lru);
670 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
671 __free_one_page(page, zone, 0, page_private(page));
672 trace_mm_page_pcpu_drain(page, 0, page_private(page));
673 } while (--to_free && --batch_free && !list_empty(list));
675 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
676 spin_unlock(&zone->lock);
679 static void free_one_page(struct zone *zone, struct page *page, int order,
682 spin_lock(&zone->lock);
683 zone->all_unreclaimable = 0;
684 zone->pages_scanned = 0;
686 __free_one_page(page, zone, order, migratetype);
687 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
688 spin_unlock(&zone->lock);
691 static bool free_pages_prepare(struct page *page, unsigned int order)
696 trace_mm_page_free(page, order);
697 kmemcheck_free_shadow(page, order);
700 page->mapping = NULL;
701 for (i = 0; i < (1 << order); i++)
702 bad += free_pages_check(page + i);
706 if (!PageHighMem(page)) {
707 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
708 debug_check_no_obj_freed(page_address(page),
711 arch_free_page(page, order);
712 kernel_map_pages(page, 1 << order, 0);
717 static void __free_pages_ok(struct page *page, unsigned int order)
720 int wasMlocked = __TestClearPageMlocked(page);
722 if (!free_pages_prepare(page, order))
725 local_irq_save(flags);
726 if (unlikely(wasMlocked))
727 free_page_mlock(page);
728 __count_vm_events(PGFREE, 1 << order);
729 free_one_page(page_zone(page), page, order,
730 get_pageblock_migratetype(page));
731 local_irq_restore(flags);
734 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
736 unsigned int nr_pages = 1 << order;
740 for (loop = 0; loop < nr_pages; loop++) {
741 struct page *p = &page[loop];
743 if (loop + 1 < nr_pages)
745 __ClearPageReserved(p);
746 set_page_count(p, 0);
749 set_page_refcounted(page);
750 __free_pages(page, order);
755 * The order of subdivision here is critical for the IO subsystem.
756 * Please do not alter this order without good reasons and regression
757 * testing. Specifically, as large blocks of memory are subdivided,
758 * the order in which smaller blocks are delivered depends on the order
759 * they're subdivided in this function. This is the primary factor
760 * influencing the order in which pages are delivered to the IO
761 * subsystem according to empirical testing, and this is also justified
762 * by considering the behavior of a buddy system containing a single
763 * large block of memory acted on by a series of small allocations.
764 * This behavior is a critical factor in sglist merging's success.
768 static inline void expand(struct zone *zone, struct page *page,
769 int low, int high, struct free_area *area,
772 unsigned long size = 1 << high;
778 VM_BUG_ON(bad_range(zone, &page[size]));
780 #ifdef CONFIG_DEBUG_PAGEALLOC
781 if (high < debug_guardpage_minorder()) {
783 * Mark as guard pages (or page), that will allow to
784 * merge back to allocator when buddy will be freed.
785 * Corresponding page table entries will not be touched,
786 * pages will stay not present in virtual address space
788 INIT_LIST_HEAD(&page[size].lru);
789 set_page_guard_flag(&page[size]);
790 set_page_private(&page[size], high);
791 /* Guard pages are not available for any usage */
792 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high));
796 list_add(&page[size].lru, &area->free_list[migratetype]);
798 set_page_order(&page[size], high);
803 * This page is about to be returned from the page allocator
805 static inline int check_new_page(struct page *page)
807 if (unlikely(page_mapcount(page) |
808 (page->mapping != NULL) |
809 (atomic_read(&page->_count) != 0) |
810 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
811 (mem_cgroup_bad_page_check(page)))) {
818 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
822 for (i = 0; i < (1 << order); i++) {
823 struct page *p = page + i;
824 if (unlikely(check_new_page(p)))
828 set_page_private(page, 0);
829 set_page_refcounted(page);
831 arch_alloc_page(page, order);
832 kernel_map_pages(page, 1 << order, 1);
834 if (gfp_flags & __GFP_ZERO)
835 prep_zero_page(page, order, gfp_flags);
837 if (order && (gfp_flags & __GFP_COMP))
838 prep_compound_page(page, order);
844 * Go through the free lists for the given migratetype and remove
845 * the smallest available page from the freelists
848 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
851 unsigned int current_order;
852 struct free_area * area;
855 /* Find a page of the appropriate size in the preferred list */
856 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
857 area = &(zone->free_area[current_order]);
858 if (list_empty(&area->free_list[migratetype]))
861 page = list_entry(area->free_list[migratetype].next,
863 list_del(&page->lru);
864 rmv_page_order(page);
866 expand(zone, page, order, current_order, area, migratetype);
875 * This array describes the order lists are fallen back to when
876 * the free lists for the desirable migrate type are depleted
878 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
879 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
880 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
881 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
882 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
886 * Move the free pages in a range to the free lists of the requested type.
887 * Note that start_page and end_pages are not aligned on a pageblock
888 * boundary. If alignment is required, use move_freepages_block()
890 static int move_freepages(struct zone *zone,
891 struct page *start_page, struct page *end_page,
898 #ifndef CONFIG_HOLES_IN_ZONE
900 * page_zone is not safe to call in this context when
901 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
902 * anyway as we check zone boundaries in move_freepages_block().
903 * Remove at a later date when no bug reports exist related to
904 * grouping pages by mobility
906 BUG_ON(page_zone(start_page) != page_zone(end_page));
909 for (page = start_page; page <= end_page;) {
910 /* Make sure we are not inadvertently changing nodes */
911 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
913 if (!pfn_valid_within(page_to_pfn(page))) {
918 if (!PageBuddy(page)) {
923 order = page_order(page);
924 list_move(&page->lru,
925 &zone->free_area[order].free_list[migratetype]);
927 pages_moved += 1 << order;
933 static int move_freepages_block(struct zone *zone, struct page *page,
936 unsigned long start_pfn, end_pfn;
937 struct page *start_page, *end_page;
939 start_pfn = page_to_pfn(page);
940 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
941 start_page = pfn_to_page(start_pfn);
942 end_page = start_page + pageblock_nr_pages - 1;
943 end_pfn = start_pfn + pageblock_nr_pages - 1;
945 /* Do not cross zone boundaries */
946 if (start_pfn < zone->zone_start_pfn)
948 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
951 return move_freepages(zone, start_page, end_page, migratetype);
954 static void change_pageblock_range(struct page *pageblock_page,
955 int start_order, int migratetype)
957 int nr_pageblocks = 1 << (start_order - pageblock_order);
959 while (nr_pageblocks--) {
960 set_pageblock_migratetype(pageblock_page, migratetype);
961 pageblock_page += pageblock_nr_pages;
965 /* Remove an element from the buddy allocator from the fallback list */
966 static inline struct page *
967 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
969 struct free_area * area;
974 /* Find the largest possible block of pages in the other list */
975 for (current_order = MAX_ORDER-1; current_order >= order;
977 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
978 migratetype = fallbacks[start_migratetype][i];
980 /* MIGRATE_RESERVE handled later if necessary */
981 if (migratetype == MIGRATE_RESERVE)
984 area = &(zone->free_area[current_order]);
985 if (list_empty(&area->free_list[migratetype]))
988 page = list_entry(area->free_list[migratetype].next,
993 * If breaking a large block of pages, move all free
994 * pages to the preferred allocation list. If falling
995 * back for a reclaimable kernel allocation, be more
996 * aggressive about taking ownership of free pages
998 if (unlikely(current_order >= (pageblock_order >> 1)) ||
999 start_migratetype == MIGRATE_RECLAIMABLE ||
1000 page_group_by_mobility_disabled) {
1001 unsigned long pages;
1002 pages = move_freepages_block(zone, page,
1005 /* Claim the whole block if over half of it is free */
1006 if (pages >= (1 << (pageblock_order-1)) ||
1007 page_group_by_mobility_disabled)
1008 set_pageblock_migratetype(page,
1011 migratetype = start_migratetype;
1014 /* Remove the page from the freelists */
1015 list_del(&page->lru);
1016 rmv_page_order(page);
1018 /* Take ownership for orders >= pageblock_order */
1019 if (current_order >= pageblock_order)
1020 change_pageblock_range(page, current_order,
1023 expand(zone, page, order, current_order, area, migratetype);
1025 trace_mm_page_alloc_extfrag(page, order, current_order,
1026 start_migratetype, migratetype);
1036 * Do the hard work of removing an element from the buddy allocator.
1037 * Call me with the zone->lock already held.
1039 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1045 page = __rmqueue_smallest(zone, order, migratetype);
1047 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1048 page = __rmqueue_fallback(zone, order, migratetype);
1051 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1052 * is used because __rmqueue_smallest is an inline function
1053 * and we want just one call site
1056 migratetype = MIGRATE_RESERVE;
1061 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1066 * Obtain a specified number of elements from the buddy allocator, all under
1067 * a single hold of the lock, for efficiency. Add them to the supplied list.
1068 * Returns the number of new pages which were placed at *list.
1070 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1071 unsigned long count, struct list_head *list,
1072 int migratetype, int cold)
1076 spin_lock(&zone->lock);
1077 for (i = 0; i < count; ++i) {
1078 struct page *page = __rmqueue(zone, order, migratetype);
1079 if (unlikely(page == NULL))
1083 * Split buddy pages returned by expand() are received here
1084 * in physical page order. The page is added to the callers and
1085 * list and the list head then moves forward. From the callers
1086 * perspective, the linked list is ordered by page number in
1087 * some conditions. This is useful for IO devices that can
1088 * merge IO requests if the physical pages are ordered
1091 if (likely(cold == 0))
1092 list_add(&page->lru, list);
1094 list_add_tail(&page->lru, list);
1095 set_page_private(page, migratetype);
1098 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1099 spin_unlock(&zone->lock);
1105 * Called from the vmstat counter updater to drain pagesets of this
1106 * currently executing processor on remote nodes after they have
1109 * Note that this function must be called with the thread pinned to
1110 * a single processor.
1112 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1114 unsigned long flags;
1117 local_irq_save(flags);
1118 if (pcp->count >= pcp->batch)
1119 to_drain = pcp->batch;
1121 to_drain = pcp->count;
1122 free_pcppages_bulk(zone, to_drain, pcp);
1123 pcp->count -= to_drain;
1124 local_irq_restore(flags);
1129 * Drain pages of the indicated processor.
1131 * The processor must either be the current processor and the
1132 * thread pinned to the current processor or a processor that
1135 static void drain_pages(unsigned int cpu)
1137 unsigned long flags;
1140 for_each_populated_zone(zone) {
1141 struct per_cpu_pageset *pset;
1142 struct per_cpu_pages *pcp;
1144 local_irq_save(flags);
1145 pset = per_cpu_ptr(zone->pageset, cpu);
1149 free_pcppages_bulk(zone, pcp->count, pcp);
1152 local_irq_restore(flags);
1157 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1159 void drain_local_pages(void *arg)
1161 drain_pages(smp_processor_id());
1165 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1167 * Note that this code is protected against sending an IPI to an offline
1168 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1169 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1170 * nothing keeps CPUs from showing up after we populated the cpumask and
1171 * before the call to on_each_cpu_mask().
1173 void drain_all_pages(void)
1176 struct per_cpu_pageset *pcp;
1180 * Allocate in the BSS so we wont require allocation in
1181 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1183 static cpumask_t cpus_with_pcps;
1186 * We don't care about racing with CPU hotplug event
1187 * as offline notification will cause the notified
1188 * cpu to drain that CPU pcps and on_each_cpu_mask
1189 * disables preemption as part of its processing
1191 for_each_online_cpu(cpu) {
1192 bool has_pcps = false;
1193 for_each_populated_zone(zone) {
1194 pcp = per_cpu_ptr(zone->pageset, cpu);
1195 if (pcp->pcp.count) {
1201 cpumask_set_cpu(cpu, &cpus_with_pcps);
1203 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1205 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1208 #ifdef CONFIG_HIBERNATION
1210 void mark_free_pages(struct zone *zone)
1212 unsigned long pfn, max_zone_pfn;
1213 unsigned long flags;
1215 struct list_head *curr;
1217 if (!zone->spanned_pages)
1220 spin_lock_irqsave(&zone->lock, flags);
1222 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1223 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1224 if (pfn_valid(pfn)) {
1225 struct page *page = pfn_to_page(pfn);
1227 if (!swsusp_page_is_forbidden(page))
1228 swsusp_unset_page_free(page);
1231 for_each_migratetype_order(order, t) {
1232 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1235 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1236 for (i = 0; i < (1UL << order); i++)
1237 swsusp_set_page_free(pfn_to_page(pfn + i));
1240 spin_unlock_irqrestore(&zone->lock, flags);
1242 #endif /* CONFIG_PM */
1245 * Free a 0-order page
1246 * cold == 1 ? free a cold page : free a hot page
1248 void free_hot_cold_page(struct page *page, int cold)
1250 struct zone *zone = page_zone(page);
1251 struct per_cpu_pages *pcp;
1252 unsigned long flags;
1254 int wasMlocked = __TestClearPageMlocked(page);
1256 if (!free_pages_prepare(page, 0))
1259 migratetype = get_pageblock_migratetype(page);
1260 set_page_private(page, migratetype);
1261 local_irq_save(flags);
1262 if (unlikely(wasMlocked))
1263 free_page_mlock(page);
1264 __count_vm_event(PGFREE);
1267 * We only track unmovable, reclaimable and movable on pcp lists.
1268 * Free ISOLATE pages back to the allocator because they are being
1269 * offlined but treat RESERVE as movable pages so we can get those
1270 * areas back if necessary. Otherwise, we may have to free
1271 * excessively into the page allocator
1273 if (migratetype >= MIGRATE_PCPTYPES) {
1274 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1275 free_one_page(zone, page, 0, migratetype);
1278 migratetype = MIGRATE_MOVABLE;
1281 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1283 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1285 list_add(&page->lru, &pcp->lists[migratetype]);
1287 if (pcp->count >= pcp->high) {
1288 free_pcppages_bulk(zone, pcp->batch, pcp);
1289 pcp->count -= pcp->batch;
1293 local_irq_restore(flags);
1297 * Free a list of 0-order pages
1299 void free_hot_cold_page_list(struct list_head *list, int cold)
1301 struct page *page, *next;
1303 list_for_each_entry_safe(page, next, list, lru) {
1304 trace_mm_page_free_batched(page, cold);
1305 free_hot_cold_page(page, cold);
1310 * split_page takes a non-compound higher-order page, and splits it into
1311 * n (1<<order) sub-pages: page[0..n]
1312 * Each sub-page must be freed individually.
1314 * Note: this is probably too low level an operation for use in drivers.
1315 * Please consult with lkml before using this in your driver.
1317 void split_page(struct page *page, unsigned int order)
1321 VM_BUG_ON(PageCompound(page));
1322 VM_BUG_ON(!page_count(page));
1324 #ifdef CONFIG_KMEMCHECK
1326 * Split shadow pages too, because free(page[0]) would
1327 * otherwise free the whole shadow.
1329 if (kmemcheck_page_is_tracked(page))
1330 split_page(virt_to_page(page[0].shadow), order);
1333 for (i = 1; i < (1 << order); i++)
1334 set_page_refcounted(page + i);
1338 * Similar to split_page except the page is already free. As this is only
1339 * being used for migration, the migratetype of the block also changes.
1340 * As this is called with interrupts disabled, the caller is responsible
1341 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1344 * Note: this is probably too low level an operation for use in drivers.
1345 * Please consult with lkml before using this in your driver.
1347 int split_free_page(struct page *page)
1350 unsigned long watermark;
1353 BUG_ON(!PageBuddy(page));
1355 zone = page_zone(page);
1356 order = page_order(page);
1358 /* Obey watermarks as if the page was being allocated */
1359 watermark = low_wmark_pages(zone) + (1 << order);
1360 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1363 /* Remove page from free list */
1364 list_del(&page->lru);
1365 zone->free_area[order].nr_free--;
1366 rmv_page_order(page);
1367 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1369 /* Split into individual pages */
1370 set_page_refcounted(page);
1371 split_page(page, order);
1373 if (order >= pageblock_order - 1) {
1374 struct page *endpage = page + (1 << order) - 1;
1375 for (; page < endpage; page += pageblock_nr_pages)
1376 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1383 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1384 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1388 struct page *buffered_rmqueue(struct zone *preferred_zone,
1389 struct zone *zone, int order, gfp_t gfp_flags,
1392 unsigned long flags;
1394 int cold = !!(gfp_flags & __GFP_COLD);
1397 if (likely(order == 0)) {
1398 struct per_cpu_pages *pcp;
1399 struct list_head *list;
1401 local_irq_save(flags);
1402 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1403 list = &pcp->lists[migratetype];
1404 if (list_empty(list)) {
1405 pcp->count += rmqueue_bulk(zone, 0,
1408 if (unlikely(list_empty(list)))
1413 page = list_entry(list->prev, struct page, lru);
1415 page = list_entry(list->next, struct page, lru);
1417 list_del(&page->lru);
1420 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1422 * __GFP_NOFAIL is not to be used in new code.
1424 * All __GFP_NOFAIL callers should be fixed so that they
1425 * properly detect and handle allocation failures.
1427 * We most definitely don't want callers attempting to
1428 * allocate greater than order-1 page units with
1431 WARN_ON_ONCE(order > 1);
1433 spin_lock_irqsave(&zone->lock, flags);
1434 page = __rmqueue(zone, order, migratetype);
1435 spin_unlock(&zone->lock);
1438 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1441 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1442 zone_statistics(preferred_zone, zone, gfp_flags);
1443 local_irq_restore(flags);
1445 VM_BUG_ON(bad_range(zone, page));
1446 if (prep_new_page(page, order, gfp_flags))
1451 local_irq_restore(flags);
1455 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1456 #define ALLOC_WMARK_MIN WMARK_MIN
1457 #define ALLOC_WMARK_LOW WMARK_LOW
1458 #define ALLOC_WMARK_HIGH WMARK_HIGH
1459 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1461 /* Mask to get the watermark bits */
1462 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1464 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1465 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1466 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1468 #ifdef CONFIG_FAIL_PAGE_ALLOC
1471 struct fault_attr attr;
1473 u32 ignore_gfp_highmem;
1474 u32 ignore_gfp_wait;
1476 } fail_page_alloc = {
1477 .attr = FAULT_ATTR_INITIALIZER,
1478 .ignore_gfp_wait = 1,
1479 .ignore_gfp_highmem = 1,
1483 static int __init setup_fail_page_alloc(char *str)
1485 return setup_fault_attr(&fail_page_alloc.attr, str);
1487 __setup("fail_page_alloc=", setup_fail_page_alloc);
1489 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1491 if (order < fail_page_alloc.min_order)
1493 if (gfp_mask & __GFP_NOFAIL)
1495 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1497 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1500 return should_fail(&fail_page_alloc.attr, 1 << order);
1503 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1505 static int __init fail_page_alloc_debugfs(void)
1507 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1510 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1511 &fail_page_alloc.attr);
1513 return PTR_ERR(dir);
1515 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1516 &fail_page_alloc.ignore_gfp_wait))
1518 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1519 &fail_page_alloc.ignore_gfp_highmem))
1521 if (!debugfs_create_u32("min-order", mode, dir,
1522 &fail_page_alloc.min_order))
1527 debugfs_remove_recursive(dir);
1532 late_initcall(fail_page_alloc_debugfs);
1534 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1536 #else /* CONFIG_FAIL_PAGE_ALLOC */
1538 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1543 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1546 * Return true if free pages are above 'mark'. This takes into account the order
1547 * of the allocation.
1549 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1550 int classzone_idx, int alloc_flags, long free_pages)
1552 /* free_pages my go negative - that's OK */
1556 free_pages -= (1 << order) - 1;
1557 if (alloc_flags & ALLOC_HIGH)
1559 if (alloc_flags & ALLOC_HARDER)
1562 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1564 for (o = 0; o < order; o++) {
1565 /* At the next order, this order's pages become unavailable */
1566 free_pages -= z->free_area[o].nr_free << o;
1568 /* Require fewer higher order pages to be free */
1571 if (free_pages <= min)
1577 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1578 int classzone_idx, int alloc_flags)
1580 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1581 zone_page_state(z, NR_FREE_PAGES));
1584 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1585 int classzone_idx, int alloc_flags)
1587 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1589 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1590 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1592 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1598 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1599 * skip over zones that are not allowed by the cpuset, or that have
1600 * been recently (in last second) found to be nearly full. See further
1601 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1602 * that have to skip over a lot of full or unallowed zones.
1604 * If the zonelist cache is present in the passed in zonelist, then
1605 * returns a pointer to the allowed node mask (either the current
1606 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1608 * If the zonelist cache is not available for this zonelist, does
1609 * nothing and returns NULL.
1611 * If the fullzones BITMAP in the zonelist cache is stale (more than
1612 * a second since last zap'd) then we zap it out (clear its bits.)
1614 * We hold off even calling zlc_setup, until after we've checked the
1615 * first zone in the zonelist, on the theory that most allocations will
1616 * be satisfied from that first zone, so best to examine that zone as
1617 * quickly as we can.
1619 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1621 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1622 nodemask_t *allowednodes; /* zonelist_cache approximation */
1624 zlc = zonelist->zlcache_ptr;
1628 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1629 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1630 zlc->last_full_zap = jiffies;
1633 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1634 &cpuset_current_mems_allowed :
1635 &node_states[N_HIGH_MEMORY];
1636 return allowednodes;
1640 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1641 * if it is worth looking at further for free memory:
1642 * 1) Check that the zone isn't thought to be full (doesn't have its
1643 * bit set in the zonelist_cache fullzones BITMAP).
1644 * 2) Check that the zones node (obtained from the zonelist_cache
1645 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1646 * Return true (non-zero) if zone is worth looking at further, or
1647 * else return false (zero) if it is not.
1649 * This check -ignores- the distinction between various watermarks,
1650 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1651 * found to be full for any variation of these watermarks, it will
1652 * be considered full for up to one second by all requests, unless
1653 * we are so low on memory on all allowed nodes that we are forced
1654 * into the second scan of the zonelist.
1656 * In the second scan we ignore this zonelist cache and exactly
1657 * apply the watermarks to all zones, even it is slower to do so.
1658 * We are low on memory in the second scan, and should leave no stone
1659 * unturned looking for a free page.
1661 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1662 nodemask_t *allowednodes)
1664 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1665 int i; /* index of *z in zonelist zones */
1666 int n; /* node that zone *z is on */
1668 zlc = zonelist->zlcache_ptr;
1672 i = z - zonelist->_zonerefs;
1675 /* This zone is worth trying if it is allowed but not full */
1676 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1680 * Given 'z' scanning a zonelist, set the corresponding bit in
1681 * zlc->fullzones, so that subsequent attempts to allocate a page
1682 * from that zone don't waste time re-examining it.
1684 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1686 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1687 int i; /* index of *z in zonelist zones */
1689 zlc = zonelist->zlcache_ptr;
1693 i = z - zonelist->_zonerefs;
1695 set_bit(i, zlc->fullzones);
1699 * clear all zones full, called after direct reclaim makes progress so that
1700 * a zone that was recently full is not skipped over for up to a second
1702 static void zlc_clear_zones_full(struct zonelist *zonelist)
1704 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1706 zlc = zonelist->zlcache_ptr;
1710 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1713 #else /* CONFIG_NUMA */
1715 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1720 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1721 nodemask_t *allowednodes)
1726 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1730 static void zlc_clear_zones_full(struct zonelist *zonelist)
1733 #endif /* CONFIG_NUMA */
1736 * get_page_from_freelist goes through the zonelist trying to allocate
1739 static struct page *
1740 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1741 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1742 struct zone *preferred_zone, int migratetype)
1745 struct page *page = NULL;
1748 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1749 int zlc_active = 0; /* set if using zonelist_cache */
1750 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1752 classzone_idx = zone_idx(preferred_zone);
1755 * Scan zonelist, looking for a zone with enough free.
1756 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1758 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1759 high_zoneidx, nodemask) {
1760 if (NUMA_BUILD && zlc_active &&
1761 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1763 if ((alloc_flags & ALLOC_CPUSET) &&
1764 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1767 * When allocating a page cache page for writing, we
1768 * want to get it from a zone that is within its dirty
1769 * limit, such that no single zone holds more than its
1770 * proportional share of globally allowed dirty pages.
1771 * The dirty limits take into account the zone's
1772 * lowmem reserves and high watermark so that kswapd
1773 * should be able to balance it without having to
1774 * write pages from its LRU list.
1776 * This may look like it could increase pressure on
1777 * lower zones by failing allocations in higher zones
1778 * before they are full. But the pages that do spill
1779 * over are limited as the lower zones are protected
1780 * by this very same mechanism. It should not become
1781 * a practical burden to them.
1783 * XXX: For now, allow allocations to potentially
1784 * exceed the per-zone dirty limit in the slowpath
1785 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1786 * which is important when on a NUMA setup the allowed
1787 * zones are together not big enough to reach the
1788 * global limit. The proper fix for these situations
1789 * will require awareness of zones in the
1790 * dirty-throttling and the flusher threads.
1792 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1793 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1794 goto this_zone_full;
1796 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1797 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1801 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1802 if (zone_watermark_ok(zone, order, mark,
1803 classzone_idx, alloc_flags))
1806 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1808 * we do zlc_setup if there are multiple nodes
1809 * and before considering the first zone allowed
1812 allowednodes = zlc_setup(zonelist, alloc_flags);
1817 if (zone_reclaim_mode == 0)
1818 goto this_zone_full;
1821 * As we may have just activated ZLC, check if the first
1822 * eligible zone has failed zone_reclaim recently.
1824 if (NUMA_BUILD && zlc_active &&
1825 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1828 ret = zone_reclaim(zone, gfp_mask, order);
1830 case ZONE_RECLAIM_NOSCAN:
1833 case ZONE_RECLAIM_FULL:
1834 /* scanned but unreclaimable */
1837 /* did we reclaim enough */
1838 if (!zone_watermark_ok(zone, order, mark,
1839 classzone_idx, alloc_flags))
1840 goto this_zone_full;
1845 page = buffered_rmqueue(preferred_zone, zone, order,
1846 gfp_mask, migratetype);
1851 zlc_mark_zone_full(zonelist, z);
1854 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1855 /* Disable zlc cache for second zonelist scan */
1863 * Large machines with many possible nodes should not always dump per-node
1864 * meminfo in irq context.
1866 static inline bool should_suppress_show_mem(void)
1871 ret = in_interrupt();
1876 static DEFINE_RATELIMIT_STATE(nopage_rs,
1877 DEFAULT_RATELIMIT_INTERVAL,
1878 DEFAULT_RATELIMIT_BURST);
1880 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1882 unsigned int filter = SHOW_MEM_FILTER_NODES;
1884 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1885 debug_guardpage_minorder() > 0)
1889 * This documents exceptions given to allocations in certain
1890 * contexts that are allowed to allocate outside current's set
1893 if (!(gfp_mask & __GFP_NOMEMALLOC))
1894 if (test_thread_flag(TIF_MEMDIE) ||
1895 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1896 filter &= ~SHOW_MEM_FILTER_NODES;
1897 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1898 filter &= ~SHOW_MEM_FILTER_NODES;
1901 struct va_format vaf;
1904 va_start(args, fmt);
1909 pr_warn("%pV", &vaf);
1914 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1915 current->comm, order, gfp_mask);
1918 if (!should_suppress_show_mem())
1923 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1924 unsigned long did_some_progress,
1925 unsigned long pages_reclaimed)
1927 /* Do not loop if specifically requested */
1928 if (gfp_mask & __GFP_NORETRY)
1931 /* Always retry if specifically requested */
1932 if (gfp_mask & __GFP_NOFAIL)
1936 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
1937 * making forward progress without invoking OOM. Suspend also disables
1938 * storage devices so kswapd will not help. Bail if we are suspending.
1940 if (!did_some_progress && pm_suspended_storage())
1944 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1945 * means __GFP_NOFAIL, but that may not be true in other
1948 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1952 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1953 * specified, then we retry until we no longer reclaim any pages
1954 * (above), or we've reclaimed an order of pages at least as
1955 * large as the allocation's order. In both cases, if the
1956 * allocation still fails, we stop retrying.
1958 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1964 static inline struct page *
1965 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1966 struct zonelist *zonelist, enum zone_type high_zoneidx,
1967 nodemask_t *nodemask, struct zone *preferred_zone,
1972 /* Acquire the OOM killer lock for the zones in zonelist */
1973 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1974 schedule_timeout_uninterruptible(1);
1979 * Go through the zonelist yet one more time, keep very high watermark
1980 * here, this is only to catch a parallel oom killing, we must fail if
1981 * we're still under heavy pressure.
1983 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1984 order, zonelist, high_zoneidx,
1985 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1986 preferred_zone, migratetype);
1990 if (!(gfp_mask & __GFP_NOFAIL)) {
1991 /* The OOM killer will not help higher order allocs */
1992 if (order > PAGE_ALLOC_COSTLY_ORDER)
1994 /* The OOM killer does not needlessly kill tasks for lowmem */
1995 if (high_zoneidx < ZONE_NORMAL)
1998 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1999 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2000 * The caller should handle page allocation failure by itself if
2001 * it specifies __GFP_THISNODE.
2002 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2004 if (gfp_mask & __GFP_THISNODE)
2007 /* Exhausted what can be done so it's blamo time */
2008 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2011 clear_zonelist_oom(zonelist, gfp_mask);
2015 #ifdef CONFIG_COMPACTION
2016 /* Try memory compaction for high-order allocations before reclaim */
2017 static struct page *
2018 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2019 struct zonelist *zonelist, enum zone_type high_zoneidx,
2020 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2021 int migratetype, bool sync_migration,
2022 bool *deferred_compaction,
2023 unsigned long *did_some_progress)
2030 if (compaction_deferred(preferred_zone, order)) {
2031 *deferred_compaction = true;
2035 current->flags |= PF_MEMALLOC;
2036 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2037 nodemask, sync_migration);
2038 current->flags &= ~PF_MEMALLOC;
2039 if (*did_some_progress != COMPACT_SKIPPED) {
2041 /* Page migration frees to the PCP lists but we want merging */
2042 drain_pages(get_cpu());
2045 page = get_page_from_freelist(gfp_mask, nodemask,
2046 order, zonelist, high_zoneidx,
2047 alloc_flags, preferred_zone,
2050 preferred_zone->compact_considered = 0;
2051 preferred_zone->compact_defer_shift = 0;
2052 if (order >= preferred_zone->compact_order_failed)
2053 preferred_zone->compact_order_failed = order + 1;
2054 count_vm_event(COMPACTSUCCESS);
2059 * It's bad if compaction run occurs and fails.
2060 * The most likely reason is that pages exist,
2061 * but not enough to satisfy watermarks.
2063 count_vm_event(COMPACTFAIL);
2066 * As async compaction considers a subset of pageblocks, only
2067 * defer if the failure was a sync compaction failure.
2070 defer_compaction(preferred_zone, order);
2078 static inline struct page *
2079 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2080 struct zonelist *zonelist, enum zone_type high_zoneidx,
2081 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2082 int migratetype, bool sync_migration,
2083 bool *deferred_compaction,
2084 unsigned long *did_some_progress)
2088 #endif /* CONFIG_COMPACTION */
2090 /* The really slow allocator path where we enter direct reclaim */
2091 static inline struct page *
2092 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2093 struct zonelist *zonelist, enum zone_type high_zoneidx,
2094 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2095 int migratetype, unsigned long *did_some_progress)
2097 struct page *page = NULL;
2098 struct reclaim_state reclaim_state;
2099 bool drained = false;
2103 /* We now go into synchronous reclaim */
2104 cpuset_memory_pressure_bump();
2105 current->flags |= PF_MEMALLOC;
2106 lockdep_set_current_reclaim_state(gfp_mask);
2107 reclaim_state.reclaimed_slab = 0;
2108 current->reclaim_state = &reclaim_state;
2110 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2112 current->reclaim_state = NULL;
2113 lockdep_clear_current_reclaim_state();
2114 current->flags &= ~PF_MEMALLOC;
2118 if (unlikely(!(*did_some_progress)))
2121 /* After successful reclaim, reconsider all zones for allocation */
2123 zlc_clear_zones_full(zonelist);
2126 page = get_page_from_freelist(gfp_mask, nodemask, order,
2127 zonelist, high_zoneidx,
2128 alloc_flags, preferred_zone,
2132 * If an allocation failed after direct reclaim, it could be because
2133 * pages are pinned on the per-cpu lists. Drain them and try again
2135 if (!page && !drained) {
2145 * This is called in the allocator slow-path if the allocation request is of
2146 * sufficient urgency to ignore watermarks and take other desperate measures
2148 static inline struct page *
2149 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2150 struct zonelist *zonelist, enum zone_type high_zoneidx,
2151 nodemask_t *nodemask, struct zone *preferred_zone,
2157 page = get_page_from_freelist(gfp_mask, nodemask, order,
2158 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2159 preferred_zone, migratetype);
2161 if (!page && gfp_mask & __GFP_NOFAIL)
2162 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2163 } while (!page && (gfp_mask & __GFP_NOFAIL));
2169 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2170 enum zone_type high_zoneidx,
2171 enum zone_type classzone_idx)
2176 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2177 wakeup_kswapd(zone, order, classzone_idx);
2181 gfp_to_alloc_flags(gfp_t gfp_mask)
2183 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2184 const gfp_t wait = gfp_mask & __GFP_WAIT;
2186 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2187 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2190 * The caller may dip into page reserves a bit more if the caller
2191 * cannot run direct reclaim, or if the caller has realtime scheduling
2192 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2193 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2195 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2199 * Not worth trying to allocate harder for
2200 * __GFP_NOMEMALLOC even if it can't schedule.
2202 if (!(gfp_mask & __GFP_NOMEMALLOC))
2203 alloc_flags |= ALLOC_HARDER;
2205 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2206 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2208 alloc_flags &= ~ALLOC_CPUSET;
2209 } else if (unlikely(rt_task(current)) && !in_interrupt())
2210 alloc_flags |= ALLOC_HARDER;
2212 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2213 if (!in_interrupt() &&
2214 ((current->flags & PF_MEMALLOC) ||
2215 unlikely(test_thread_flag(TIF_MEMDIE))))
2216 alloc_flags |= ALLOC_NO_WATERMARKS;
2222 static inline struct page *
2223 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2224 struct zonelist *zonelist, enum zone_type high_zoneidx,
2225 nodemask_t *nodemask, struct zone *preferred_zone,
2228 const gfp_t wait = gfp_mask & __GFP_WAIT;
2229 struct page *page = NULL;
2231 unsigned long pages_reclaimed = 0;
2232 unsigned long did_some_progress;
2233 bool sync_migration = false;
2234 bool deferred_compaction = false;
2237 * In the slowpath, we sanity check order to avoid ever trying to
2238 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2239 * be using allocators in order of preference for an area that is
2242 if (order >= MAX_ORDER) {
2243 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2248 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2249 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2250 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2251 * using a larger set of nodes after it has established that the
2252 * allowed per node queues are empty and that nodes are
2255 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2259 if (!(gfp_mask & __GFP_NO_KSWAPD))
2260 wake_all_kswapd(order, zonelist, high_zoneidx,
2261 zone_idx(preferred_zone));
2264 * OK, we're below the kswapd watermark and have kicked background
2265 * reclaim. Now things get more complex, so set up alloc_flags according
2266 * to how we want to proceed.
2268 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2271 * Find the true preferred zone if the allocation is unconstrained by
2274 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2275 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2279 /* This is the last chance, in general, before the goto nopage. */
2280 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2281 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2282 preferred_zone, migratetype);
2286 /* Allocate without watermarks if the context allows */
2287 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2288 page = __alloc_pages_high_priority(gfp_mask, order,
2289 zonelist, high_zoneidx, nodemask,
2290 preferred_zone, migratetype);
2295 /* Atomic allocations - we can't balance anything */
2299 /* Avoid recursion of direct reclaim */
2300 if (current->flags & PF_MEMALLOC)
2303 /* Avoid allocations with no watermarks from looping endlessly */
2304 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2308 * Try direct compaction. The first pass is asynchronous. Subsequent
2309 * attempts after direct reclaim are synchronous
2311 page = __alloc_pages_direct_compact(gfp_mask, order,
2312 zonelist, high_zoneidx,
2314 alloc_flags, preferred_zone,
2315 migratetype, sync_migration,
2316 &deferred_compaction,
2317 &did_some_progress);
2320 sync_migration = true;
2323 * If compaction is deferred for high-order allocations, it is because
2324 * sync compaction recently failed. In this is the case and the caller
2325 * has requested the system not be heavily disrupted, fail the
2326 * allocation now instead of entering direct reclaim
2328 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2331 /* Try direct reclaim and then allocating */
2332 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2333 zonelist, high_zoneidx,
2335 alloc_flags, preferred_zone,
2336 migratetype, &did_some_progress);
2341 * If we failed to make any progress reclaiming, then we are
2342 * running out of options and have to consider going OOM
2344 if (!did_some_progress) {
2345 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2346 if (oom_killer_disabled)
2348 /* Coredumps can quickly deplete all memory reserves */
2349 if ((current->flags & PF_DUMPCORE) &&
2350 !(gfp_mask & __GFP_NOFAIL))
2352 page = __alloc_pages_may_oom(gfp_mask, order,
2353 zonelist, high_zoneidx,
2354 nodemask, preferred_zone,
2359 if (!(gfp_mask & __GFP_NOFAIL)) {
2361 * The oom killer is not called for high-order
2362 * allocations that may fail, so if no progress
2363 * is being made, there are no other options and
2364 * retrying is unlikely to help.
2366 if (order > PAGE_ALLOC_COSTLY_ORDER)
2369 * The oom killer is not called for lowmem
2370 * allocations to prevent needlessly killing
2373 if (high_zoneidx < ZONE_NORMAL)
2381 /* Check if we should retry the allocation */
2382 pages_reclaimed += did_some_progress;
2383 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2385 /* Wait for some write requests to complete then retry */
2386 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2390 * High-order allocations do not necessarily loop after
2391 * direct reclaim and reclaim/compaction depends on compaction
2392 * being called after reclaim so call directly if necessary
2394 page = __alloc_pages_direct_compact(gfp_mask, order,
2395 zonelist, high_zoneidx,
2397 alloc_flags, preferred_zone,
2398 migratetype, sync_migration,
2399 &deferred_compaction,
2400 &did_some_progress);
2406 warn_alloc_failed(gfp_mask, order, NULL);
2409 if (kmemcheck_enabled)
2410 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2416 * This is the 'heart' of the zoned buddy allocator.
2419 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2420 struct zonelist *zonelist, nodemask_t *nodemask)
2422 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2423 struct zone *preferred_zone;
2424 struct page *page = NULL;
2425 int migratetype = allocflags_to_migratetype(gfp_mask);
2426 unsigned int cpuset_mems_cookie;
2428 gfp_mask &= gfp_allowed_mask;
2430 lockdep_trace_alloc(gfp_mask);
2432 might_sleep_if(gfp_mask & __GFP_WAIT);
2434 if (should_fail_alloc_page(gfp_mask, order))
2438 * Check the zones suitable for the gfp_mask contain at least one
2439 * valid zone. It's possible to have an empty zonelist as a result
2440 * of GFP_THISNODE and a memoryless node
2442 if (unlikely(!zonelist->_zonerefs->zone))
2446 cpuset_mems_cookie = get_mems_allowed();
2448 /* The preferred zone is used for statistics later */
2449 first_zones_zonelist(zonelist, high_zoneidx,
2450 nodemask ? : &cpuset_current_mems_allowed,
2452 if (!preferred_zone)
2455 /* First allocation attempt */
2456 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2457 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2458 preferred_zone, migratetype);
2459 if (unlikely(!page))
2460 page = __alloc_pages_slowpath(gfp_mask, order,
2461 zonelist, high_zoneidx, nodemask,
2462 preferred_zone, migratetype);
2464 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2468 * When updating a task's mems_allowed, it is possible to race with
2469 * parallel threads in such a way that an allocation can fail while
2470 * the mask is being updated. If a page allocation is about to fail,
2471 * check if the cpuset changed during allocation and if so, retry.
2473 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2478 EXPORT_SYMBOL(__alloc_pages_nodemask);
2481 * Common helper functions.
2483 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2488 * __get_free_pages() returns a 32-bit address, which cannot represent
2491 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2493 page = alloc_pages(gfp_mask, order);
2496 return (unsigned long) page_address(page);
2498 EXPORT_SYMBOL(__get_free_pages);
2500 unsigned long get_zeroed_page(gfp_t gfp_mask)
2502 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2504 EXPORT_SYMBOL(get_zeroed_page);
2506 void __free_pages(struct page *page, unsigned int order)
2508 if (put_page_testzero(page)) {
2510 free_hot_cold_page(page, 0);
2512 __free_pages_ok(page, order);
2516 EXPORT_SYMBOL(__free_pages);
2518 void free_pages(unsigned long addr, unsigned int order)
2521 VM_BUG_ON(!virt_addr_valid((void *)addr));
2522 __free_pages(virt_to_page((void *)addr), order);
2526 EXPORT_SYMBOL(free_pages);
2528 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2531 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2532 unsigned long used = addr + PAGE_ALIGN(size);
2534 split_page(virt_to_page((void *)addr), order);
2535 while (used < alloc_end) {
2540 return (void *)addr;
2544 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2545 * @size: the number of bytes to allocate
2546 * @gfp_mask: GFP flags for the allocation
2548 * This function is similar to alloc_pages(), except that it allocates the
2549 * minimum number of pages to satisfy the request. alloc_pages() can only
2550 * allocate memory in power-of-two pages.
2552 * This function is also limited by MAX_ORDER.
2554 * Memory allocated by this function must be released by free_pages_exact().
2556 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2558 unsigned int order = get_order(size);
2561 addr = __get_free_pages(gfp_mask, order);
2562 return make_alloc_exact(addr, order, size);
2564 EXPORT_SYMBOL(alloc_pages_exact);
2567 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2569 * @nid: the preferred node ID where memory should be allocated
2570 * @size: the number of bytes to allocate
2571 * @gfp_mask: GFP flags for the allocation
2573 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2575 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2578 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2580 unsigned order = get_order(size);
2581 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2584 return make_alloc_exact((unsigned long)page_address(p), order, size);
2586 EXPORT_SYMBOL(alloc_pages_exact_nid);
2589 * free_pages_exact - release memory allocated via alloc_pages_exact()
2590 * @virt: the value returned by alloc_pages_exact.
2591 * @size: size of allocation, same value as passed to alloc_pages_exact().
2593 * Release the memory allocated by a previous call to alloc_pages_exact.
2595 void free_pages_exact(void *virt, size_t size)
2597 unsigned long addr = (unsigned long)virt;
2598 unsigned long end = addr + PAGE_ALIGN(size);
2600 while (addr < end) {
2605 EXPORT_SYMBOL(free_pages_exact);
2607 static unsigned int nr_free_zone_pages(int offset)
2612 /* Just pick one node, since fallback list is circular */
2613 unsigned int sum = 0;
2615 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2617 for_each_zone_zonelist(zone, z, zonelist, offset) {
2618 unsigned long size = zone->present_pages;
2619 unsigned long high = high_wmark_pages(zone);
2628 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2630 unsigned int nr_free_buffer_pages(void)
2632 return nr_free_zone_pages(gfp_zone(GFP_USER));
2634 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2637 * Amount of free RAM allocatable within all zones
2639 unsigned int nr_free_pagecache_pages(void)
2641 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2644 static inline void show_node(struct zone *zone)
2647 printk("Node %d ", zone_to_nid(zone));
2650 void si_meminfo(struct sysinfo *val)
2652 val->totalram = totalram_pages;
2654 val->freeram = global_page_state(NR_FREE_PAGES);
2655 val->bufferram = nr_blockdev_pages();
2656 val->totalhigh = totalhigh_pages;
2657 val->freehigh = nr_free_highpages();
2658 val->mem_unit = PAGE_SIZE;
2661 EXPORT_SYMBOL(si_meminfo);
2664 void si_meminfo_node(struct sysinfo *val, int nid)
2666 pg_data_t *pgdat = NODE_DATA(nid);
2668 val->totalram = pgdat->node_present_pages;
2669 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2670 #ifdef CONFIG_HIGHMEM
2671 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2672 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2678 val->mem_unit = PAGE_SIZE;
2683 * Determine whether the node should be displayed or not, depending on whether
2684 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2686 bool skip_free_areas_node(unsigned int flags, int nid)
2689 unsigned int cpuset_mems_cookie;
2691 if (!(flags & SHOW_MEM_FILTER_NODES))
2695 cpuset_mems_cookie = get_mems_allowed();
2696 ret = !node_isset(nid, cpuset_current_mems_allowed);
2697 } while (!put_mems_allowed(cpuset_mems_cookie));
2702 #define K(x) ((x) << (PAGE_SHIFT-10))
2705 * Show free area list (used inside shift_scroll-lock stuff)
2706 * We also calculate the percentage fragmentation. We do this by counting the
2707 * memory on each free list with the exception of the first item on the list.
2708 * Suppresses nodes that are not allowed by current's cpuset if
2709 * SHOW_MEM_FILTER_NODES is passed.
2711 void show_free_areas(unsigned int filter)
2716 for_each_populated_zone(zone) {
2717 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2720 printk("%s per-cpu:\n", zone->name);
2722 for_each_online_cpu(cpu) {
2723 struct per_cpu_pageset *pageset;
2725 pageset = per_cpu_ptr(zone->pageset, cpu);
2727 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2728 cpu, pageset->pcp.high,
2729 pageset->pcp.batch, pageset->pcp.count);
2733 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2734 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2736 " dirty:%lu writeback:%lu unstable:%lu\n"
2737 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2738 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2739 global_page_state(NR_ACTIVE_ANON),
2740 global_page_state(NR_INACTIVE_ANON),
2741 global_page_state(NR_ISOLATED_ANON),
2742 global_page_state(NR_ACTIVE_FILE),
2743 global_page_state(NR_INACTIVE_FILE),
2744 global_page_state(NR_ISOLATED_FILE),
2745 global_page_state(NR_UNEVICTABLE),
2746 global_page_state(NR_FILE_DIRTY),
2747 global_page_state(NR_WRITEBACK),
2748 global_page_state(NR_UNSTABLE_NFS),
2749 global_page_state(NR_FREE_PAGES),
2750 global_page_state(NR_SLAB_RECLAIMABLE),
2751 global_page_state(NR_SLAB_UNRECLAIMABLE),
2752 global_page_state(NR_FILE_MAPPED),
2753 global_page_state(NR_SHMEM),
2754 global_page_state(NR_PAGETABLE),
2755 global_page_state(NR_BOUNCE));
2757 for_each_populated_zone(zone) {
2760 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2768 " active_anon:%lukB"
2769 " inactive_anon:%lukB"
2770 " active_file:%lukB"
2771 " inactive_file:%lukB"
2772 " unevictable:%lukB"
2773 " isolated(anon):%lukB"
2774 " isolated(file):%lukB"
2781 " slab_reclaimable:%lukB"
2782 " slab_unreclaimable:%lukB"
2783 " kernel_stack:%lukB"
2787 " writeback_tmp:%lukB"
2788 " pages_scanned:%lu"
2789 " all_unreclaimable? %s"
2792 K(zone_page_state(zone, NR_FREE_PAGES)),
2793 K(min_wmark_pages(zone)),
2794 K(low_wmark_pages(zone)),
2795 K(high_wmark_pages(zone)),
2796 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2797 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2798 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2799 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2800 K(zone_page_state(zone, NR_UNEVICTABLE)),
2801 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2802 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2803 K(zone->present_pages),
2804 K(zone_page_state(zone, NR_MLOCK)),
2805 K(zone_page_state(zone, NR_FILE_DIRTY)),
2806 K(zone_page_state(zone, NR_WRITEBACK)),
2807 K(zone_page_state(zone, NR_FILE_MAPPED)),
2808 K(zone_page_state(zone, NR_SHMEM)),
2809 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2810 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2811 zone_page_state(zone, NR_KERNEL_STACK) *
2813 K(zone_page_state(zone, NR_PAGETABLE)),
2814 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2815 K(zone_page_state(zone, NR_BOUNCE)),
2816 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2817 zone->pages_scanned,
2818 (zone->all_unreclaimable ? "yes" : "no")
2820 printk("lowmem_reserve[]:");
2821 for (i = 0; i < MAX_NR_ZONES; i++)
2822 printk(" %lu", zone->lowmem_reserve[i]);
2826 for_each_populated_zone(zone) {
2827 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2829 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2832 printk("%s: ", zone->name);
2834 spin_lock_irqsave(&zone->lock, flags);
2835 for (order = 0; order < MAX_ORDER; order++) {
2836 nr[order] = zone->free_area[order].nr_free;
2837 total += nr[order] << order;
2839 spin_unlock_irqrestore(&zone->lock, flags);
2840 for (order = 0; order < MAX_ORDER; order++)
2841 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2842 printk("= %lukB\n", K(total));
2845 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2847 show_swap_cache_info();
2850 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2852 zoneref->zone = zone;
2853 zoneref->zone_idx = zone_idx(zone);
2857 * Builds allocation fallback zone lists.
2859 * Add all populated zones of a node to the zonelist.
2861 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2862 int nr_zones, enum zone_type zone_type)
2866 BUG_ON(zone_type >= MAX_NR_ZONES);
2871 zone = pgdat->node_zones + zone_type;
2872 if (populated_zone(zone)) {
2873 zoneref_set_zone(zone,
2874 &zonelist->_zonerefs[nr_zones++]);
2875 check_highest_zone(zone_type);
2878 } while (zone_type);
2885 * 0 = automatic detection of better ordering.
2886 * 1 = order by ([node] distance, -zonetype)
2887 * 2 = order by (-zonetype, [node] distance)
2889 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2890 * the same zonelist. So only NUMA can configure this param.
2892 #define ZONELIST_ORDER_DEFAULT 0
2893 #define ZONELIST_ORDER_NODE 1
2894 #define ZONELIST_ORDER_ZONE 2
2896 /* zonelist order in the kernel.
2897 * set_zonelist_order() will set this to NODE or ZONE.
2899 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2900 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2904 /* The value user specified ....changed by config */
2905 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2906 /* string for sysctl */
2907 #define NUMA_ZONELIST_ORDER_LEN 16
2908 char numa_zonelist_order[16] = "default";
2911 * interface for configure zonelist ordering.
2912 * command line option "numa_zonelist_order"
2913 * = "[dD]efault - default, automatic configuration.
2914 * = "[nN]ode - order by node locality, then by zone within node
2915 * = "[zZ]one - order by zone, then by locality within zone
2918 static int __parse_numa_zonelist_order(char *s)
2920 if (*s == 'd' || *s == 'D') {
2921 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2922 } else if (*s == 'n' || *s == 'N') {
2923 user_zonelist_order = ZONELIST_ORDER_NODE;
2924 } else if (*s == 'z' || *s == 'Z') {
2925 user_zonelist_order = ZONELIST_ORDER_ZONE;
2928 "Ignoring invalid numa_zonelist_order value: "
2935 static __init int setup_numa_zonelist_order(char *s)
2942 ret = __parse_numa_zonelist_order(s);
2944 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2948 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2951 * sysctl handler for numa_zonelist_order
2953 int numa_zonelist_order_handler(ctl_table *table, int write,
2954 void __user *buffer, size_t *length,
2957 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2959 static DEFINE_MUTEX(zl_order_mutex);
2961 mutex_lock(&zl_order_mutex);
2963 strcpy(saved_string, (char*)table->data);
2964 ret = proc_dostring(table, write, buffer, length, ppos);
2968 int oldval = user_zonelist_order;
2969 if (__parse_numa_zonelist_order((char*)table->data)) {
2971 * bogus value. restore saved string
2973 strncpy((char*)table->data, saved_string,
2974 NUMA_ZONELIST_ORDER_LEN);
2975 user_zonelist_order = oldval;
2976 } else if (oldval != user_zonelist_order) {
2977 mutex_lock(&zonelists_mutex);
2978 build_all_zonelists(NULL);
2979 mutex_unlock(&zonelists_mutex);
2983 mutex_unlock(&zl_order_mutex);
2988 #define MAX_NODE_LOAD (nr_online_nodes)
2989 static int node_load[MAX_NUMNODES];
2992 * find_next_best_node - find the next node that should appear in a given node's fallback list
2993 * @node: node whose fallback list we're appending
2994 * @used_node_mask: nodemask_t of already used nodes
2996 * We use a number of factors to determine which is the next node that should
2997 * appear on a given node's fallback list. The node should not have appeared
2998 * already in @node's fallback list, and it should be the next closest node
2999 * according to the distance array (which contains arbitrary distance values
3000 * from each node to each node in the system), and should also prefer nodes
3001 * with no CPUs, since presumably they'll have very little allocation pressure
3002 * on them otherwise.
3003 * It returns -1 if no node is found.
3005 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3008 int min_val = INT_MAX;
3010 const struct cpumask *tmp = cpumask_of_node(0);
3012 /* Use the local node if we haven't already */
3013 if (!node_isset(node, *used_node_mask)) {
3014 node_set(node, *used_node_mask);
3018 for_each_node_state(n, N_HIGH_MEMORY) {
3020 /* Don't want a node to appear more than once */
3021 if (node_isset(n, *used_node_mask))
3024 /* Use the distance array to find the distance */
3025 val = node_distance(node, n);
3027 /* Penalize nodes under us ("prefer the next node") */
3030 /* Give preference to headless and unused nodes */
3031 tmp = cpumask_of_node(n);
3032 if (!cpumask_empty(tmp))
3033 val += PENALTY_FOR_NODE_WITH_CPUS;
3035 /* Slight preference for less loaded node */
3036 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3037 val += node_load[n];
3039 if (val < min_val) {
3046 node_set(best_node, *used_node_mask);
3053 * Build zonelists ordered by node and zones within node.
3054 * This results in maximum locality--normal zone overflows into local
3055 * DMA zone, if any--but risks exhausting DMA zone.
3057 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3060 struct zonelist *zonelist;
3062 zonelist = &pgdat->node_zonelists[0];
3063 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3065 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3067 zonelist->_zonerefs[j].zone = NULL;
3068 zonelist->_zonerefs[j].zone_idx = 0;
3072 * Build gfp_thisnode zonelists
3074 static void build_thisnode_zonelists(pg_data_t *pgdat)
3077 struct zonelist *zonelist;
3079 zonelist = &pgdat->node_zonelists[1];
3080 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3081 zonelist->_zonerefs[j].zone = NULL;
3082 zonelist->_zonerefs[j].zone_idx = 0;
3086 * Build zonelists ordered by zone and nodes within zones.
3087 * This results in conserving DMA zone[s] until all Normal memory is
3088 * exhausted, but results in overflowing to remote node while memory
3089 * may still exist in local DMA zone.
3091 static int node_order[MAX_NUMNODES];
3093 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3096 int zone_type; /* needs to be signed */
3098 struct zonelist *zonelist;
3100 zonelist = &pgdat->node_zonelists[0];
3102 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3103 for (j = 0; j < nr_nodes; j++) {
3104 node = node_order[j];
3105 z = &NODE_DATA(node)->node_zones[zone_type];
3106 if (populated_zone(z)) {
3108 &zonelist->_zonerefs[pos++]);
3109 check_highest_zone(zone_type);
3113 zonelist->_zonerefs[pos].zone = NULL;
3114 zonelist->_zonerefs[pos].zone_idx = 0;
3117 static int default_zonelist_order(void)
3120 unsigned long low_kmem_size,total_size;
3124 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3125 * If they are really small and used heavily, the system can fall
3126 * into OOM very easily.
3127 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3129 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3132 for_each_online_node(nid) {
3133 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3134 z = &NODE_DATA(nid)->node_zones[zone_type];
3135 if (populated_zone(z)) {
3136 if (zone_type < ZONE_NORMAL)
3137 low_kmem_size += z->present_pages;
3138 total_size += z->present_pages;
3139 } else if (zone_type == ZONE_NORMAL) {
3141 * If any node has only lowmem, then node order
3142 * is preferred to allow kernel allocations
3143 * locally; otherwise, they can easily infringe
3144 * on other nodes when there is an abundance of
3145 * lowmem available to allocate from.
3147 return ZONELIST_ORDER_NODE;
3151 if (!low_kmem_size || /* there are no DMA area. */
3152 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3153 return ZONELIST_ORDER_NODE;
3155 * look into each node's config.
3156 * If there is a node whose DMA/DMA32 memory is very big area on
3157 * local memory, NODE_ORDER may be suitable.
3159 average_size = total_size /
3160 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3161 for_each_online_node(nid) {
3164 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3165 z = &NODE_DATA(nid)->node_zones[zone_type];
3166 if (populated_zone(z)) {
3167 if (zone_type < ZONE_NORMAL)
3168 low_kmem_size += z->present_pages;
3169 total_size += z->present_pages;
3172 if (low_kmem_size &&
3173 total_size > average_size && /* ignore small node */
3174 low_kmem_size > total_size * 70/100)
3175 return ZONELIST_ORDER_NODE;
3177 return ZONELIST_ORDER_ZONE;
3180 static void set_zonelist_order(void)
3182 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3183 current_zonelist_order = default_zonelist_order();
3185 current_zonelist_order = user_zonelist_order;
3188 static void build_zonelists(pg_data_t *pgdat)
3192 nodemask_t used_mask;
3193 int local_node, prev_node;
3194 struct zonelist *zonelist;
3195 int order = current_zonelist_order;
3197 /* initialize zonelists */
3198 for (i = 0; i < MAX_ZONELISTS; i++) {
3199 zonelist = pgdat->node_zonelists + i;
3200 zonelist->_zonerefs[0].zone = NULL;
3201 zonelist->_zonerefs[0].zone_idx = 0;
3204 /* NUMA-aware ordering of nodes */
3205 local_node = pgdat->node_id;
3206 load = nr_online_nodes;
3207 prev_node = local_node;
3208 nodes_clear(used_mask);
3210 memset(node_order, 0, sizeof(node_order));
3213 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3214 int distance = node_distance(local_node, node);
3217 * If another node is sufficiently far away then it is better
3218 * to reclaim pages in a zone before going off node.
3220 if (distance > RECLAIM_DISTANCE)
3221 zone_reclaim_mode = 1;
3224 * We don't want to pressure a particular node.
3225 * So adding penalty to the first node in same
3226 * distance group to make it round-robin.
3228 if (distance != node_distance(local_node, prev_node))
3229 node_load[node] = load;
3233 if (order == ZONELIST_ORDER_NODE)
3234 build_zonelists_in_node_order(pgdat, node);
3236 node_order[j++] = node; /* remember order */
3239 if (order == ZONELIST_ORDER_ZONE) {
3240 /* calculate node order -- i.e., DMA last! */
3241 build_zonelists_in_zone_order(pgdat, j);
3244 build_thisnode_zonelists(pgdat);
3247 /* Construct the zonelist performance cache - see further mmzone.h */
3248 static void build_zonelist_cache(pg_data_t *pgdat)
3250 struct zonelist *zonelist;
3251 struct zonelist_cache *zlc;
3254 zonelist = &pgdat->node_zonelists[0];
3255 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3256 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3257 for (z = zonelist->_zonerefs; z->zone; z++)
3258 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3261 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3263 * Return node id of node used for "local" allocations.
3264 * I.e., first node id of first zone in arg node's generic zonelist.
3265 * Used for initializing percpu 'numa_mem', which is used primarily
3266 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3268 int local_memory_node(int node)
3272 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3273 gfp_zone(GFP_KERNEL),
3280 #else /* CONFIG_NUMA */
3282 static void set_zonelist_order(void)
3284 current_zonelist_order = ZONELIST_ORDER_ZONE;
3287 static void build_zonelists(pg_data_t *pgdat)
3289 int node, local_node;
3291 struct zonelist *zonelist;
3293 local_node = pgdat->node_id;
3295 zonelist = &pgdat->node_zonelists[0];
3296 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3299 * Now we build the zonelist so that it contains the zones
3300 * of all the other nodes.
3301 * We don't want to pressure a particular node, so when
3302 * building the zones for node N, we make sure that the
3303 * zones coming right after the local ones are those from
3304 * node N+1 (modulo N)
3306 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3307 if (!node_online(node))
3309 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3312 for (node = 0; node < local_node; node++) {
3313 if (!node_online(node))
3315 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3319 zonelist->_zonerefs[j].zone = NULL;
3320 zonelist->_zonerefs[j].zone_idx = 0;
3323 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3324 static void build_zonelist_cache(pg_data_t *pgdat)
3326 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3329 #endif /* CONFIG_NUMA */
3332 * Boot pageset table. One per cpu which is going to be used for all
3333 * zones and all nodes. The parameters will be set in such a way
3334 * that an item put on a list will immediately be handed over to
3335 * the buddy list. This is safe since pageset manipulation is done
3336 * with interrupts disabled.
3338 * The boot_pagesets must be kept even after bootup is complete for
3339 * unused processors and/or zones. They do play a role for bootstrapping
3340 * hotplugged processors.
3342 * zoneinfo_show() and maybe other functions do
3343 * not check if the processor is online before following the pageset pointer.
3344 * Other parts of the kernel may not check if the zone is available.
3346 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3347 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3348 static void setup_zone_pageset(struct zone *zone);
3351 * Global mutex to protect against size modification of zonelists
3352 * as well as to serialize pageset setup for the new populated zone.
3354 DEFINE_MUTEX(zonelists_mutex);
3356 /* return values int ....just for stop_machine() */
3357 static __init_refok int __build_all_zonelists(void *data)
3363 memset(node_load, 0, sizeof(node_load));
3365 for_each_online_node(nid) {
3366 pg_data_t *pgdat = NODE_DATA(nid);
3368 build_zonelists(pgdat);
3369 build_zonelist_cache(pgdat);
3373 * Initialize the boot_pagesets that are going to be used
3374 * for bootstrapping processors. The real pagesets for
3375 * each zone will be allocated later when the per cpu
3376 * allocator is available.
3378 * boot_pagesets are used also for bootstrapping offline
3379 * cpus if the system is already booted because the pagesets
3380 * are needed to initialize allocators on a specific cpu too.
3381 * F.e. the percpu allocator needs the page allocator which
3382 * needs the percpu allocator in order to allocate its pagesets
3383 * (a chicken-egg dilemma).
3385 for_each_possible_cpu(cpu) {
3386 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3388 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3390 * We now know the "local memory node" for each node--
3391 * i.e., the node of the first zone in the generic zonelist.
3392 * Set up numa_mem percpu variable for on-line cpus. During
3393 * boot, only the boot cpu should be on-line; we'll init the
3394 * secondary cpus' numa_mem as they come on-line. During
3395 * node/memory hotplug, we'll fixup all on-line cpus.
3397 if (cpu_online(cpu))
3398 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3406 * Called with zonelists_mutex held always
3407 * unless system_state == SYSTEM_BOOTING.
3409 void __ref build_all_zonelists(void *data)
3411 set_zonelist_order();
3413 if (system_state == SYSTEM_BOOTING) {
3414 __build_all_zonelists(NULL);
3415 mminit_verify_zonelist();
3416 cpuset_init_current_mems_allowed();
3418 /* we have to stop all cpus to guarantee there is no user
3420 #ifdef CONFIG_MEMORY_HOTPLUG
3422 setup_zone_pageset((struct zone *)data);
3424 stop_machine(__build_all_zonelists, NULL, NULL);
3425 /* cpuset refresh routine should be here */
3427 vm_total_pages = nr_free_pagecache_pages();
3429 * Disable grouping by mobility if the number of pages in the
3430 * system is too low to allow the mechanism to work. It would be
3431 * more accurate, but expensive to check per-zone. This check is
3432 * made on memory-hotadd so a system can start with mobility
3433 * disabled and enable it later
3435 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3436 page_group_by_mobility_disabled = 1;
3438 page_group_by_mobility_disabled = 0;
3440 printk("Built %i zonelists in %s order, mobility grouping %s. "
3441 "Total pages: %ld\n",
3443 zonelist_order_name[current_zonelist_order],
3444 page_group_by_mobility_disabled ? "off" : "on",
3447 printk("Policy zone: %s\n", zone_names[policy_zone]);
3452 * Helper functions to size the waitqueue hash table.
3453 * Essentially these want to choose hash table sizes sufficiently
3454 * large so that collisions trying to wait on pages are rare.
3455 * But in fact, the number of active page waitqueues on typical
3456 * systems is ridiculously low, less than 200. So this is even
3457 * conservative, even though it seems large.
3459 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3460 * waitqueues, i.e. the size of the waitq table given the number of pages.
3462 #define PAGES_PER_WAITQUEUE 256
3464 #ifndef CONFIG_MEMORY_HOTPLUG
3465 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3467 unsigned long size = 1;
3469 pages /= PAGES_PER_WAITQUEUE;
3471 while (size < pages)
3475 * Once we have dozens or even hundreds of threads sleeping
3476 * on IO we've got bigger problems than wait queue collision.
3477 * Limit the size of the wait table to a reasonable size.
3479 size = min(size, 4096UL);
3481 return max(size, 4UL);
3485 * A zone's size might be changed by hot-add, so it is not possible to determine
3486 * a suitable size for its wait_table. So we use the maximum size now.
3488 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3490 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3491 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3492 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3494 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3495 * or more by the traditional way. (See above). It equals:
3497 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3498 * ia64(16K page size) : = ( 8G + 4M)byte.
3499 * powerpc (64K page size) : = (32G +16M)byte.
3501 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3508 * This is an integer logarithm so that shifts can be used later
3509 * to extract the more random high bits from the multiplicative
3510 * hash function before the remainder is taken.
3512 static inline unsigned long wait_table_bits(unsigned long size)
3517 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3520 * Check if a pageblock contains reserved pages
3522 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3526 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3527 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3534 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3535 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3536 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3537 * higher will lead to a bigger reserve which will get freed as contiguous
3538 * blocks as reclaim kicks in
3540 static void setup_zone_migrate_reserve(struct zone *zone)
3542 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3544 unsigned long block_migratetype;
3548 * Get the start pfn, end pfn and the number of blocks to reserve
3549 * We have to be careful to be aligned to pageblock_nr_pages to
3550 * make sure that we always check pfn_valid for the first page in
3553 start_pfn = zone->zone_start_pfn;
3554 end_pfn = start_pfn + zone->spanned_pages;
3555 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3556 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3560 * Reserve blocks are generally in place to help high-order atomic
3561 * allocations that are short-lived. A min_free_kbytes value that
3562 * would result in more than 2 reserve blocks for atomic allocations
3563 * is assumed to be in place to help anti-fragmentation for the
3564 * future allocation of hugepages at runtime.
3566 reserve = min(2, reserve);
3568 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3569 if (!pfn_valid(pfn))
3571 page = pfn_to_page(pfn);
3573 /* Watch out for overlapping nodes */
3574 if (page_to_nid(page) != zone_to_nid(zone))
3577 block_migratetype = get_pageblock_migratetype(page);
3579 /* Only test what is necessary when the reserves are not met */
3582 * Blocks with reserved pages will never free, skip
3585 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3586 if (pageblock_is_reserved(pfn, block_end_pfn))
3589 /* If this block is reserved, account for it */
3590 if (block_migratetype == MIGRATE_RESERVE) {
3595 /* Suitable for reserving if this block is movable */
3596 if (block_migratetype == MIGRATE_MOVABLE) {
3597 set_pageblock_migratetype(page,
3599 move_freepages_block(zone, page,
3607 * If the reserve is met and this is a previous reserved block,
3610 if (block_migratetype == MIGRATE_RESERVE) {
3611 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3612 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3618 * Initially all pages are reserved - free ones are freed
3619 * up by free_all_bootmem() once the early boot process is
3620 * done. Non-atomic initialization, single-pass.
3622 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3623 unsigned long start_pfn, enum memmap_context context)
3626 unsigned long end_pfn = start_pfn + size;
3630 if (highest_memmap_pfn < end_pfn - 1)
3631 highest_memmap_pfn = end_pfn - 1;
3633 z = &NODE_DATA(nid)->node_zones[zone];
3634 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3636 * There can be holes in boot-time mem_map[]s
3637 * handed to this function. They do not
3638 * exist on hotplugged memory.
3640 if (context == MEMMAP_EARLY) {
3641 if (!early_pfn_valid(pfn))
3643 if (!early_pfn_in_nid(pfn, nid))
3646 page = pfn_to_page(pfn);
3647 set_page_links(page, zone, nid, pfn);
3648 mminit_verify_page_links(page, zone, nid, pfn);
3649 init_page_count(page);
3650 reset_page_mapcount(page);
3651 SetPageReserved(page);
3653 * Mark the block movable so that blocks are reserved for
3654 * movable at startup. This will force kernel allocations
3655 * to reserve their blocks rather than leaking throughout
3656 * the address space during boot when many long-lived
3657 * kernel allocations are made. Later some blocks near
3658 * the start are marked MIGRATE_RESERVE by
3659 * setup_zone_migrate_reserve()
3661 * bitmap is created for zone's valid pfn range. but memmap
3662 * can be created for invalid pages (for alignment)
3663 * check here not to call set_pageblock_migratetype() against
3666 if ((z->zone_start_pfn <= pfn)
3667 && (pfn < z->zone_start_pfn + z->spanned_pages)
3668 && !(pfn & (pageblock_nr_pages - 1)))
3669 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3671 INIT_LIST_HEAD(&page->lru);
3672 #ifdef WANT_PAGE_VIRTUAL
3673 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3674 if (!is_highmem_idx(zone))
3675 set_page_address(page, __va(pfn << PAGE_SHIFT));
3680 static void __meminit zone_init_free_lists(struct zone *zone)
3683 for_each_migratetype_order(order, t) {
3684 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3685 zone->free_area[order].nr_free = 0;
3689 #ifndef __HAVE_ARCH_MEMMAP_INIT
3690 #define memmap_init(size, nid, zone, start_pfn) \
3691 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3694 static int zone_batchsize(struct zone *zone)
3700 * The per-cpu-pages pools are set to around 1000th of the
3701 * size of the zone. But no more than 1/2 of a meg.
3703 * OK, so we don't know how big the cache is. So guess.
3705 batch = zone->present_pages / 1024;
3706 if (batch * PAGE_SIZE > 512 * 1024)
3707 batch = (512 * 1024) / PAGE_SIZE;
3708 batch /= 4; /* We effectively *= 4 below */
3713 * Clamp the batch to a 2^n - 1 value. Having a power
3714 * of 2 value was found to be more likely to have
3715 * suboptimal cache aliasing properties in some cases.
3717 * For example if 2 tasks are alternately allocating
3718 * batches of pages, one task can end up with a lot
3719 * of pages of one half of the possible page colors
3720 * and the other with pages of the other colors.
3722 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3727 /* The deferral and batching of frees should be suppressed under NOMMU
3730 * The problem is that NOMMU needs to be able to allocate large chunks
3731 * of contiguous memory as there's no hardware page translation to
3732 * assemble apparent contiguous memory from discontiguous pages.
3734 * Queueing large contiguous runs of pages for batching, however,
3735 * causes the pages to actually be freed in smaller chunks. As there
3736 * can be a significant delay between the individual batches being
3737 * recycled, this leads to the once large chunks of space being
3738 * fragmented and becoming unavailable for high-order allocations.
3744 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3746 struct per_cpu_pages *pcp;
3749 memset(p, 0, sizeof(*p));
3753 pcp->high = 6 * batch;
3754 pcp->batch = max(1UL, 1 * batch);
3755 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3756 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3760 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3761 * to the value high for the pageset p.
3764 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3767 struct per_cpu_pages *pcp;
3771 pcp->batch = max(1UL, high/4);
3772 if ((high/4) > (PAGE_SHIFT * 8))
3773 pcp->batch = PAGE_SHIFT * 8;
3776 static void setup_zone_pageset(struct zone *zone)
3780 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3782 for_each_possible_cpu(cpu) {
3783 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3785 setup_pageset(pcp, zone_batchsize(zone));
3787 if (percpu_pagelist_fraction)
3788 setup_pagelist_highmark(pcp,
3789 (zone->present_pages /
3790 percpu_pagelist_fraction));
3795 * Allocate per cpu pagesets and initialize them.
3796 * Before this call only boot pagesets were available.
3798 void __init setup_per_cpu_pageset(void)
3802 for_each_populated_zone(zone)
3803 setup_zone_pageset(zone);
3806 static noinline __init_refok
3807 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3810 struct pglist_data *pgdat = zone->zone_pgdat;
3814 * The per-page waitqueue mechanism uses hashed waitqueues
3817 zone->wait_table_hash_nr_entries =
3818 wait_table_hash_nr_entries(zone_size_pages);
3819 zone->wait_table_bits =
3820 wait_table_bits(zone->wait_table_hash_nr_entries);
3821 alloc_size = zone->wait_table_hash_nr_entries
3822 * sizeof(wait_queue_head_t);
3824 if (!slab_is_available()) {
3825 zone->wait_table = (wait_queue_head_t *)
3826 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3829 * This case means that a zone whose size was 0 gets new memory
3830 * via memory hot-add.
3831 * But it may be the case that a new node was hot-added. In
3832 * this case vmalloc() will not be able to use this new node's
3833 * memory - this wait_table must be initialized to use this new
3834 * node itself as well.
3835 * To use this new node's memory, further consideration will be
3838 zone->wait_table = vmalloc(alloc_size);
3840 if (!zone->wait_table)
3843 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3844 init_waitqueue_head(zone->wait_table + i);
3849 static int __zone_pcp_update(void *data)
3851 struct zone *zone = data;
3853 unsigned long batch = zone_batchsize(zone), flags;
3855 for_each_possible_cpu(cpu) {
3856 struct per_cpu_pageset *pset;
3857 struct per_cpu_pages *pcp;
3859 pset = per_cpu_ptr(zone->pageset, cpu);
3862 local_irq_save(flags);
3863 free_pcppages_bulk(zone, pcp->count, pcp);
3864 setup_pageset(pset, batch);
3865 local_irq_restore(flags);
3870 void zone_pcp_update(struct zone *zone)
3872 stop_machine(__zone_pcp_update, zone, NULL);
3875 static __meminit void zone_pcp_init(struct zone *zone)
3878 * per cpu subsystem is not up at this point. The following code
3879 * relies on the ability of the linker to provide the
3880 * offset of a (static) per cpu variable into the per cpu area.
3882 zone->pageset = &boot_pageset;
3884 if (zone->present_pages)
3885 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3886 zone->name, zone->present_pages,
3887 zone_batchsize(zone));
3890 __meminit int init_currently_empty_zone(struct zone *zone,
3891 unsigned long zone_start_pfn,
3893 enum memmap_context context)
3895 struct pglist_data *pgdat = zone->zone_pgdat;
3897 ret = zone_wait_table_init(zone, size);
3900 pgdat->nr_zones = zone_idx(zone) + 1;
3902 zone->zone_start_pfn = zone_start_pfn;
3904 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3905 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3907 (unsigned long)zone_idx(zone),
3908 zone_start_pfn, (zone_start_pfn + size));
3910 zone_init_free_lists(zone);
3915 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
3916 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3918 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3919 * Architectures may implement their own version but if add_active_range()
3920 * was used and there are no special requirements, this is a convenient
3923 int __meminit __early_pfn_to_nid(unsigned long pfn)
3925 unsigned long start_pfn, end_pfn;
3928 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
3929 if (start_pfn <= pfn && pfn < end_pfn)
3931 /* This is a memory hole */
3934 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3936 int __meminit early_pfn_to_nid(unsigned long pfn)
3940 nid = __early_pfn_to_nid(pfn);
3943 /* just returns 0 */
3947 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3948 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3952 nid = __early_pfn_to_nid(pfn);
3953 if (nid >= 0 && nid != node)
3960 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3961 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3962 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3964 * If an architecture guarantees that all ranges registered with
3965 * add_active_ranges() contain no holes and may be freed, this
3966 * this function may be used instead of calling free_bootmem() manually.
3968 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
3970 unsigned long start_pfn, end_pfn;
3973 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
3974 start_pfn = min(start_pfn, max_low_pfn);
3975 end_pfn = min(end_pfn, max_low_pfn);
3977 if (start_pfn < end_pfn)
3978 free_bootmem_node(NODE_DATA(this_nid),
3979 PFN_PHYS(start_pfn),
3980 (end_pfn - start_pfn) << PAGE_SHIFT);
3985 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3986 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3988 * If an architecture guarantees that all ranges registered with
3989 * add_active_ranges() contain no holes and may be freed, this
3990 * function may be used instead of calling memory_present() manually.
3992 void __init sparse_memory_present_with_active_regions(int nid)
3994 unsigned long start_pfn, end_pfn;
3997 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
3998 memory_present(this_nid, start_pfn, end_pfn);
4002 * get_pfn_range_for_nid - Return the start and end page frames for a node
4003 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4004 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4005 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4007 * It returns the start and end page frame of a node based on information
4008 * provided by an arch calling add_active_range(). If called for a node
4009 * with no available memory, a warning is printed and the start and end
4012 void __meminit get_pfn_range_for_nid(unsigned int nid,
4013 unsigned long *start_pfn, unsigned long *end_pfn)
4015 unsigned long this_start_pfn, this_end_pfn;
4021 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4022 *start_pfn = min(*start_pfn, this_start_pfn);
4023 *end_pfn = max(*end_pfn, this_end_pfn);
4026 if (*start_pfn == -1UL)
4031 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4032 * assumption is made that zones within a node are ordered in monotonic
4033 * increasing memory addresses so that the "highest" populated zone is used
4035 static void __init find_usable_zone_for_movable(void)
4038 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4039 if (zone_index == ZONE_MOVABLE)
4042 if (arch_zone_highest_possible_pfn[zone_index] >
4043 arch_zone_lowest_possible_pfn[zone_index])
4047 VM_BUG_ON(zone_index == -1);
4048 movable_zone = zone_index;
4052 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4053 * because it is sized independent of architecture. Unlike the other zones,
4054 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4055 * in each node depending on the size of each node and how evenly kernelcore
4056 * is distributed. This helper function adjusts the zone ranges
4057 * provided by the architecture for a given node by using the end of the
4058 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4059 * zones within a node are in order of monotonic increases memory addresses
4061 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4062 unsigned long zone_type,
4063 unsigned long node_start_pfn,
4064 unsigned long node_end_pfn,
4065 unsigned long *zone_start_pfn,
4066 unsigned long *zone_end_pfn)
4068 /* Only adjust if ZONE_MOVABLE is on this node */
4069 if (zone_movable_pfn[nid]) {
4070 /* Size ZONE_MOVABLE */
4071 if (zone_type == ZONE_MOVABLE) {
4072 *zone_start_pfn = zone_movable_pfn[nid];
4073 *zone_end_pfn = min(node_end_pfn,
4074 arch_zone_highest_possible_pfn[movable_zone]);
4076 /* Adjust for ZONE_MOVABLE starting within this range */
4077 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4078 *zone_end_pfn > zone_movable_pfn[nid]) {
4079 *zone_end_pfn = zone_movable_pfn[nid];
4081 /* Check if this whole range is within ZONE_MOVABLE */
4082 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4083 *zone_start_pfn = *zone_end_pfn;
4088 * Return the number of pages a zone spans in a node, including holes
4089 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4091 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4092 unsigned long zone_type,
4093 unsigned long *ignored)
4095 unsigned long node_start_pfn, node_end_pfn;
4096 unsigned long zone_start_pfn, zone_end_pfn;
4098 /* Get the start and end of the node and zone */
4099 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4100 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4101 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4102 adjust_zone_range_for_zone_movable(nid, zone_type,
4103 node_start_pfn, node_end_pfn,
4104 &zone_start_pfn, &zone_end_pfn);
4106 /* Check that this node has pages within the zone's required range */
4107 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4110 /* Move the zone boundaries inside the node if necessary */
4111 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4112 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4114 /* Return the spanned pages */
4115 return zone_end_pfn - zone_start_pfn;
4119 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4120 * then all holes in the requested range will be accounted for.
4122 unsigned long __meminit __absent_pages_in_range(int nid,
4123 unsigned long range_start_pfn,
4124 unsigned long range_end_pfn)
4126 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4127 unsigned long start_pfn, end_pfn;
4130 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4131 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4132 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4133 nr_absent -= end_pfn - start_pfn;
4139 * absent_pages_in_range - Return number of page frames in holes within a range
4140 * @start_pfn: The start PFN to start searching for holes
4141 * @end_pfn: The end PFN to stop searching for holes
4143 * It returns the number of pages frames in memory holes within a range.
4145 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4146 unsigned long end_pfn)
4148 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4151 /* Return the number of page frames in holes in a zone on a node */
4152 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4153 unsigned long zone_type,
4154 unsigned long *ignored)
4156 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4157 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4158 unsigned long node_start_pfn, node_end_pfn;
4159 unsigned long zone_start_pfn, zone_end_pfn;
4161 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4162 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4163 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4165 adjust_zone_range_for_zone_movable(nid, zone_type,
4166 node_start_pfn, node_end_pfn,
4167 &zone_start_pfn, &zone_end_pfn);
4168 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4171 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4172 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4173 unsigned long zone_type,
4174 unsigned long *zones_size)
4176 return zones_size[zone_type];
4179 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4180 unsigned long zone_type,
4181 unsigned long *zholes_size)
4186 return zholes_size[zone_type];
4189 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4191 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4192 unsigned long *zones_size, unsigned long *zholes_size)
4194 unsigned long realtotalpages, totalpages = 0;
4197 for (i = 0; i < MAX_NR_ZONES; i++)
4198 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4200 pgdat->node_spanned_pages = totalpages;
4202 realtotalpages = totalpages;
4203 for (i = 0; i < MAX_NR_ZONES; i++)
4205 zone_absent_pages_in_node(pgdat->node_id, i,
4207 pgdat->node_present_pages = realtotalpages;
4208 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4212 #ifndef CONFIG_SPARSEMEM
4214 * Calculate the size of the zone->blockflags rounded to an unsigned long
4215 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4216 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4217 * round what is now in bits to nearest long in bits, then return it in
4220 static unsigned long __init usemap_size(unsigned long zonesize)
4222 unsigned long usemapsize;
4224 usemapsize = roundup(zonesize, pageblock_nr_pages);
4225 usemapsize = usemapsize >> pageblock_order;
4226 usemapsize *= NR_PAGEBLOCK_BITS;
4227 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4229 return usemapsize / 8;
4232 static void __init setup_usemap(struct pglist_data *pgdat,
4233 struct zone *zone, unsigned long zonesize)
4235 unsigned long usemapsize = usemap_size(zonesize);
4236 zone->pageblock_flags = NULL;
4238 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4242 static inline void setup_usemap(struct pglist_data *pgdat,
4243 struct zone *zone, unsigned long zonesize) {}
4244 #endif /* CONFIG_SPARSEMEM */
4246 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4248 /* Return a sensible default order for the pageblock size. */
4249 static inline int pageblock_default_order(void)
4251 if (HPAGE_SHIFT > PAGE_SHIFT)
4252 return HUGETLB_PAGE_ORDER;
4257 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4258 static inline void __init set_pageblock_order(unsigned int order)
4260 /* Check that pageblock_nr_pages has not already been setup */
4261 if (pageblock_order)
4265 * Assume the largest contiguous order of interest is a huge page.
4266 * This value may be variable depending on boot parameters on IA64
4268 pageblock_order = order;
4270 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4273 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4274 * and pageblock_default_order() are unused as pageblock_order is set
4275 * at compile-time. See include/linux/pageblock-flags.h for the values of
4276 * pageblock_order based on the kernel config
4278 static inline int pageblock_default_order(unsigned int order)
4282 #define set_pageblock_order(x) do {} while (0)
4284 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4287 * Set up the zone data structures:
4288 * - mark all pages reserved
4289 * - mark all memory queues empty
4290 * - clear the memory bitmaps
4292 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4293 unsigned long *zones_size, unsigned long *zholes_size)
4296 int nid = pgdat->node_id;
4297 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4300 pgdat_resize_init(pgdat);
4301 pgdat->nr_zones = 0;
4302 init_waitqueue_head(&pgdat->kswapd_wait);
4303 pgdat->kswapd_max_order = 0;
4304 pgdat_page_cgroup_init(pgdat);
4306 for (j = 0; j < MAX_NR_ZONES; j++) {
4307 struct zone *zone = pgdat->node_zones + j;
4308 unsigned long size, realsize, memmap_pages;
4311 size = zone_spanned_pages_in_node(nid, j, zones_size);
4312 realsize = size - zone_absent_pages_in_node(nid, j,
4316 * Adjust realsize so that it accounts for how much memory
4317 * is used by this zone for memmap. This affects the watermark
4318 * and per-cpu initialisations
4321 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4322 if (realsize >= memmap_pages) {
4323 realsize -= memmap_pages;
4326 " %s zone: %lu pages used for memmap\n",
4327 zone_names[j], memmap_pages);
4330 " %s zone: %lu pages exceeds realsize %lu\n",
4331 zone_names[j], memmap_pages, realsize);
4333 /* Account for reserved pages */
4334 if (j == 0 && realsize > dma_reserve) {
4335 realsize -= dma_reserve;
4336 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4337 zone_names[0], dma_reserve);
4340 if (!is_highmem_idx(j))
4341 nr_kernel_pages += realsize;
4342 nr_all_pages += realsize;
4344 zone->spanned_pages = size;
4345 zone->present_pages = realsize;
4348 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4350 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4352 zone->name = zone_names[j];
4353 spin_lock_init(&zone->lock);
4354 spin_lock_init(&zone->lru_lock);
4355 zone_seqlock_init(zone);
4356 zone->zone_pgdat = pgdat;
4358 zone_pcp_init(zone);
4360 INIT_LIST_HEAD(&zone->lruvec.lists[lru]);
4361 zone->reclaim_stat.recent_rotated[0] = 0;
4362 zone->reclaim_stat.recent_rotated[1] = 0;
4363 zone->reclaim_stat.recent_scanned[0] = 0;
4364 zone->reclaim_stat.recent_scanned[1] = 0;
4365 zap_zone_vm_stats(zone);
4370 set_pageblock_order(pageblock_default_order());
4371 setup_usemap(pgdat, zone, size);
4372 ret = init_currently_empty_zone(zone, zone_start_pfn,
4373 size, MEMMAP_EARLY);
4375 memmap_init(size, nid, j, zone_start_pfn);
4376 zone_start_pfn += size;
4380 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4382 /* Skip empty nodes */
4383 if (!pgdat->node_spanned_pages)
4386 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4387 /* ia64 gets its own node_mem_map, before this, without bootmem */
4388 if (!pgdat->node_mem_map) {
4389 unsigned long size, start, end;
4393 * The zone's endpoints aren't required to be MAX_ORDER
4394 * aligned but the node_mem_map endpoints must be in order
4395 * for the buddy allocator to function correctly.
4397 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4398 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4399 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4400 size = (end - start) * sizeof(struct page);
4401 map = alloc_remap(pgdat->node_id, size);
4403 map = alloc_bootmem_node_nopanic(pgdat, size);
4404 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4406 #ifndef CONFIG_NEED_MULTIPLE_NODES
4408 * With no DISCONTIG, the global mem_map is just set as node 0's
4410 if (pgdat == NODE_DATA(0)) {
4411 mem_map = NODE_DATA(0)->node_mem_map;
4412 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4413 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4414 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4415 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4418 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4421 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4422 unsigned long node_start_pfn, unsigned long *zholes_size)
4424 pg_data_t *pgdat = NODE_DATA(nid);
4426 pgdat->node_id = nid;
4427 pgdat->node_start_pfn = node_start_pfn;
4428 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4430 alloc_node_mem_map(pgdat);
4431 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4432 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4433 nid, (unsigned long)pgdat,
4434 (unsigned long)pgdat->node_mem_map);
4437 free_area_init_core(pgdat, zones_size, zholes_size);
4440 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4442 #if MAX_NUMNODES > 1
4444 * Figure out the number of possible node ids.
4446 static void __init setup_nr_node_ids(void)
4449 unsigned int highest = 0;
4451 for_each_node_mask(node, node_possible_map)
4453 nr_node_ids = highest + 1;
4456 static inline void setup_nr_node_ids(void)
4462 * node_map_pfn_alignment - determine the maximum internode alignment
4464 * This function should be called after node map is populated and sorted.
4465 * It calculates the maximum power of two alignment which can distinguish
4468 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4469 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4470 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4471 * shifted, 1GiB is enough and this function will indicate so.
4473 * This is used to test whether pfn -> nid mapping of the chosen memory
4474 * model has fine enough granularity to avoid incorrect mapping for the
4475 * populated node map.
4477 * Returns the determined alignment in pfn's. 0 if there is no alignment
4478 * requirement (single node).
4480 unsigned long __init node_map_pfn_alignment(void)
4482 unsigned long accl_mask = 0, last_end = 0;
4483 unsigned long start, end, mask;
4487 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4488 if (!start || last_nid < 0 || last_nid == nid) {
4495 * Start with a mask granular enough to pin-point to the
4496 * start pfn and tick off bits one-by-one until it becomes
4497 * too coarse to separate the current node from the last.
4499 mask = ~((1 << __ffs(start)) - 1);
4500 while (mask && last_end <= (start & (mask << 1)))
4503 /* accumulate all internode masks */
4507 /* convert mask to number of pages */
4508 return ~accl_mask + 1;
4511 /* Find the lowest pfn for a node */
4512 static unsigned long __init find_min_pfn_for_node(int nid)
4514 unsigned long min_pfn = ULONG_MAX;
4515 unsigned long start_pfn;
4518 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4519 min_pfn = min(min_pfn, start_pfn);
4521 if (min_pfn == ULONG_MAX) {
4523 "Could not find start_pfn for node %d\n", nid);
4531 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4533 * It returns the minimum PFN based on information provided via
4534 * add_active_range().
4536 unsigned long __init find_min_pfn_with_active_regions(void)
4538 return find_min_pfn_for_node(MAX_NUMNODES);
4542 * early_calculate_totalpages()
4543 * Sum pages in active regions for movable zone.
4544 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4546 static unsigned long __init early_calculate_totalpages(void)
4548 unsigned long totalpages = 0;
4549 unsigned long start_pfn, end_pfn;
4552 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4553 unsigned long pages = end_pfn - start_pfn;
4555 totalpages += pages;
4557 node_set_state(nid, N_HIGH_MEMORY);
4563 * Find the PFN the Movable zone begins in each node. Kernel memory
4564 * is spread evenly between nodes as long as the nodes have enough
4565 * memory. When they don't, some nodes will have more kernelcore than
4568 static void __init find_zone_movable_pfns_for_nodes(void)
4571 unsigned long usable_startpfn;
4572 unsigned long kernelcore_node, kernelcore_remaining;
4573 /* save the state before borrow the nodemask */
4574 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4575 unsigned long totalpages = early_calculate_totalpages();
4576 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4579 * If movablecore was specified, calculate what size of
4580 * kernelcore that corresponds so that memory usable for
4581 * any allocation type is evenly spread. If both kernelcore
4582 * and movablecore are specified, then the value of kernelcore
4583 * will be used for required_kernelcore if it's greater than
4584 * what movablecore would have allowed.
4586 if (required_movablecore) {
4587 unsigned long corepages;
4590 * Round-up so that ZONE_MOVABLE is at least as large as what
4591 * was requested by the user
4593 required_movablecore =
4594 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4595 corepages = totalpages - required_movablecore;
4597 required_kernelcore = max(required_kernelcore, corepages);
4600 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4601 if (!required_kernelcore)
4604 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4605 find_usable_zone_for_movable();
4606 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4609 /* Spread kernelcore memory as evenly as possible throughout nodes */
4610 kernelcore_node = required_kernelcore / usable_nodes;
4611 for_each_node_state(nid, N_HIGH_MEMORY) {
4612 unsigned long start_pfn, end_pfn;
4615 * Recalculate kernelcore_node if the division per node
4616 * now exceeds what is necessary to satisfy the requested
4617 * amount of memory for the kernel
4619 if (required_kernelcore < kernelcore_node)
4620 kernelcore_node = required_kernelcore / usable_nodes;
4623 * As the map is walked, we track how much memory is usable
4624 * by the kernel using kernelcore_remaining. When it is
4625 * 0, the rest of the node is usable by ZONE_MOVABLE
4627 kernelcore_remaining = kernelcore_node;
4629 /* Go through each range of PFNs within this node */
4630 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4631 unsigned long size_pages;
4633 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4634 if (start_pfn >= end_pfn)
4637 /* Account for what is only usable for kernelcore */
4638 if (start_pfn < usable_startpfn) {
4639 unsigned long kernel_pages;
4640 kernel_pages = min(end_pfn, usable_startpfn)
4643 kernelcore_remaining -= min(kernel_pages,
4644 kernelcore_remaining);
4645 required_kernelcore -= min(kernel_pages,
4646 required_kernelcore);
4648 /* Continue if range is now fully accounted */
4649 if (end_pfn <= usable_startpfn) {
4652 * Push zone_movable_pfn to the end so
4653 * that if we have to rebalance
4654 * kernelcore across nodes, we will
4655 * not double account here
4657 zone_movable_pfn[nid] = end_pfn;
4660 start_pfn = usable_startpfn;
4664 * The usable PFN range for ZONE_MOVABLE is from
4665 * start_pfn->end_pfn. Calculate size_pages as the
4666 * number of pages used as kernelcore
4668 size_pages = end_pfn - start_pfn;
4669 if (size_pages > kernelcore_remaining)
4670 size_pages = kernelcore_remaining;
4671 zone_movable_pfn[nid] = start_pfn + size_pages;
4674 * Some kernelcore has been met, update counts and
4675 * break if the kernelcore for this node has been
4678 required_kernelcore -= min(required_kernelcore,
4680 kernelcore_remaining -= size_pages;
4681 if (!kernelcore_remaining)
4687 * If there is still required_kernelcore, we do another pass with one
4688 * less node in the count. This will push zone_movable_pfn[nid] further
4689 * along on the nodes that still have memory until kernelcore is
4693 if (usable_nodes && required_kernelcore > usable_nodes)
4696 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4697 for (nid = 0; nid < MAX_NUMNODES; nid++)
4698 zone_movable_pfn[nid] =
4699 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4702 /* restore the node_state */
4703 node_states[N_HIGH_MEMORY] = saved_node_state;
4706 /* Any regular memory on that node ? */
4707 static void check_for_regular_memory(pg_data_t *pgdat)
4709 #ifdef CONFIG_HIGHMEM
4710 enum zone_type zone_type;
4712 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4713 struct zone *zone = &pgdat->node_zones[zone_type];
4714 if (zone->present_pages) {
4715 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4723 * free_area_init_nodes - Initialise all pg_data_t and zone data
4724 * @max_zone_pfn: an array of max PFNs for each zone
4726 * This will call free_area_init_node() for each active node in the system.
4727 * Using the page ranges provided by add_active_range(), the size of each
4728 * zone in each node and their holes is calculated. If the maximum PFN
4729 * between two adjacent zones match, it is assumed that the zone is empty.
4730 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4731 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4732 * starts where the previous one ended. For example, ZONE_DMA32 starts
4733 * at arch_max_dma_pfn.
4735 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4737 unsigned long start_pfn, end_pfn;
4740 /* Record where the zone boundaries are */
4741 memset(arch_zone_lowest_possible_pfn, 0,
4742 sizeof(arch_zone_lowest_possible_pfn));
4743 memset(arch_zone_highest_possible_pfn, 0,
4744 sizeof(arch_zone_highest_possible_pfn));
4745 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4746 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4747 for (i = 1; i < MAX_NR_ZONES; i++) {
4748 if (i == ZONE_MOVABLE)
4750 arch_zone_lowest_possible_pfn[i] =
4751 arch_zone_highest_possible_pfn[i-1];
4752 arch_zone_highest_possible_pfn[i] =
4753 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4755 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4756 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4758 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4759 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4760 find_zone_movable_pfns_for_nodes();
4762 /* Print out the zone ranges */
4763 printk("Zone PFN ranges:\n");
4764 for (i = 0; i < MAX_NR_ZONES; i++) {
4765 if (i == ZONE_MOVABLE)
4767 printk(" %-8s ", zone_names[i]);
4768 if (arch_zone_lowest_possible_pfn[i] ==
4769 arch_zone_highest_possible_pfn[i])
4772 printk("%0#10lx -> %0#10lx\n",
4773 arch_zone_lowest_possible_pfn[i],
4774 arch_zone_highest_possible_pfn[i]);
4777 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4778 printk("Movable zone start PFN for each node\n");
4779 for (i = 0; i < MAX_NUMNODES; i++) {
4780 if (zone_movable_pfn[i])
4781 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4784 /* Print out the early_node_map[] */
4785 printk("Early memory PFN ranges\n");
4786 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4787 printk(" %3d: %0#10lx -> %0#10lx\n", nid, start_pfn, end_pfn);
4789 /* Initialise every node */
4790 mminit_verify_pageflags_layout();
4791 setup_nr_node_ids();
4792 for_each_online_node(nid) {
4793 pg_data_t *pgdat = NODE_DATA(nid);
4794 free_area_init_node(nid, NULL,
4795 find_min_pfn_for_node(nid), NULL);
4797 /* Any memory on that node */
4798 if (pgdat->node_present_pages)
4799 node_set_state(nid, N_HIGH_MEMORY);
4800 check_for_regular_memory(pgdat);
4804 static int __init cmdline_parse_core(char *p, unsigned long *core)
4806 unsigned long long coremem;
4810 coremem = memparse(p, &p);
4811 *core = coremem >> PAGE_SHIFT;
4813 /* Paranoid check that UL is enough for the coremem value */
4814 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4820 * kernelcore=size sets the amount of memory for use for allocations that
4821 * cannot be reclaimed or migrated.
4823 static int __init cmdline_parse_kernelcore(char *p)
4825 return cmdline_parse_core(p, &required_kernelcore);
4829 * movablecore=size sets the amount of memory for use for allocations that
4830 * can be reclaimed or migrated.
4832 static int __init cmdline_parse_movablecore(char *p)
4834 return cmdline_parse_core(p, &required_movablecore);
4837 early_param("kernelcore", cmdline_parse_kernelcore);
4838 early_param("movablecore", cmdline_parse_movablecore);
4840 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4843 * set_dma_reserve - set the specified number of pages reserved in the first zone
4844 * @new_dma_reserve: The number of pages to mark reserved
4846 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4847 * In the DMA zone, a significant percentage may be consumed by kernel image
4848 * and other unfreeable allocations which can skew the watermarks badly. This
4849 * function may optionally be used to account for unfreeable pages in the
4850 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4851 * smaller per-cpu batchsize.
4853 void __init set_dma_reserve(unsigned long new_dma_reserve)
4855 dma_reserve = new_dma_reserve;
4858 void __init free_area_init(unsigned long *zones_size)
4860 free_area_init_node(0, zones_size,
4861 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4864 static int page_alloc_cpu_notify(struct notifier_block *self,
4865 unsigned long action, void *hcpu)
4867 int cpu = (unsigned long)hcpu;
4869 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4870 lru_add_drain_cpu(cpu);
4874 * Spill the event counters of the dead processor
4875 * into the current processors event counters.
4876 * This artificially elevates the count of the current
4879 vm_events_fold_cpu(cpu);
4882 * Zero the differential counters of the dead processor
4883 * so that the vm statistics are consistent.
4885 * This is only okay since the processor is dead and cannot
4886 * race with what we are doing.
4888 refresh_cpu_vm_stats(cpu);
4893 void __init page_alloc_init(void)
4895 hotcpu_notifier(page_alloc_cpu_notify, 0);
4899 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4900 * or min_free_kbytes changes.
4902 static void calculate_totalreserve_pages(void)
4904 struct pglist_data *pgdat;
4905 unsigned long reserve_pages = 0;
4906 enum zone_type i, j;
4908 for_each_online_pgdat(pgdat) {
4909 for (i = 0; i < MAX_NR_ZONES; i++) {
4910 struct zone *zone = pgdat->node_zones + i;
4911 unsigned long max = 0;
4913 /* Find valid and maximum lowmem_reserve in the zone */
4914 for (j = i; j < MAX_NR_ZONES; j++) {
4915 if (zone->lowmem_reserve[j] > max)
4916 max = zone->lowmem_reserve[j];
4919 /* we treat the high watermark as reserved pages. */
4920 max += high_wmark_pages(zone);
4922 if (max > zone->present_pages)
4923 max = zone->present_pages;
4924 reserve_pages += max;
4926 * Lowmem reserves are not available to
4927 * GFP_HIGHUSER page cache allocations and
4928 * kswapd tries to balance zones to their high
4929 * watermark. As a result, neither should be
4930 * regarded as dirtyable memory, to prevent a
4931 * situation where reclaim has to clean pages
4932 * in order to balance the zones.
4934 zone->dirty_balance_reserve = max;
4937 dirty_balance_reserve = reserve_pages;
4938 totalreserve_pages = reserve_pages;
4942 * setup_per_zone_lowmem_reserve - called whenever
4943 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4944 * has a correct pages reserved value, so an adequate number of
4945 * pages are left in the zone after a successful __alloc_pages().
4947 static void setup_per_zone_lowmem_reserve(void)
4949 struct pglist_data *pgdat;
4950 enum zone_type j, idx;
4952 for_each_online_pgdat(pgdat) {
4953 for (j = 0; j < MAX_NR_ZONES; j++) {
4954 struct zone *zone = pgdat->node_zones + j;
4955 unsigned long present_pages = zone->present_pages;
4957 zone->lowmem_reserve[j] = 0;
4961 struct zone *lower_zone;
4965 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4966 sysctl_lowmem_reserve_ratio[idx] = 1;
4968 lower_zone = pgdat->node_zones + idx;
4969 lower_zone->lowmem_reserve[j] = present_pages /
4970 sysctl_lowmem_reserve_ratio[idx];
4971 present_pages += lower_zone->present_pages;
4976 /* update totalreserve_pages */
4977 calculate_totalreserve_pages();
4981 * setup_per_zone_wmarks - called when min_free_kbytes changes
4982 * or when memory is hot-{added|removed}
4984 * Ensures that the watermark[min,low,high] values for each zone are set
4985 * correctly with respect to min_free_kbytes.
4987 void setup_per_zone_wmarks(void)
4989 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4990 unsigned long lowmem_pages = 0;
4992 unsigned long flags;
4994 /* Calculate total number of !ZONE_HIGHMEM pages */
4995 for_each_zone(zone) {
4996 if (!is_highmem(zone))
4997 lowmem_pages += zone->present_pages;
5000 for_each_zone(zone) {
5003 spin_lock_irqsave(&zone->lock, flags);
5004 tmp = (u64)pages_min * zone->present_pages;
5005 do_div(tmp, lowmem_pages);
5006 if (is_highmem(zone)) {
5008 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5009 * need highmem pages, so cap pages_min to a small
5012 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5013 * deltas controls asynch page reclaim, and so should
5014 * not be capped for highmem.
5018 min_pages = zone->present_pages / 1024;
5019 if (min_pages < SWAP_CLUSTER_MAX)
5020 min_pages = SWAP_CLUSTER_MAX;
5021 if (min_pages > 128)
5023 zone->watermark[WMARK_MIN] = min_pages;
5026 * If it's a lowmem zone, reserve a number of pages
5027 * proportionate to the zone's size.
5029 zone->watermark[WMARK_MIN] = tmp;
5032 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5033 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5034 setup_zone_migrate_reserve(zone);
5035 spin_unlock_irqrestore(&zone->lock, flags);
5038 /* update totalreserve_pages */
5039 calculate_totalreserve_pages();
5043 * The inactive anon list should be small enough that the VM never has to
5044 * do too much work, but large enough that each inactive page has a chance
5045 * to be referenced again before it is swapped out.
5047 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5048 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5049 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5050 * the anonymous pages are kept on the inactive list.
5053 * memory ratio inactive anon
5054 * -------------------------------------
5063 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5065 unsigned int gb, ratio;
5067 /* Zone size in gigabytes */
5068 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5070 ratio = int_sqrt(10 * gb);
5074 zone->inactive_ratio = ratio;
5077 static void __meminit setup_per_zone_inactive_ratio(void)
5082 calculate_zone_inactive_ratio(zone);
5086 * Initialise min_free_kbytes.
5088 * For small machines we want it small (128k min). For large machines
5089 * we want it large (64MB max). But it is not linear, because network
5090 * bandwidth does not increase linearly with machine size. We use
5092 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5093 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5109 int __meminit init_per_zone_wmark_min(void)
5111 unsigned long lowmem_kbytes;
5113 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5115 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5116 if (min_free_kbytes < 128)
5117 min_free_kbytes = 128;
5118 if (min_free_kbytes > 65536)
5119 min_free_kbytes = 65536;
5120 setup_per_zone_wmarks();
5121 refresh_zone_stat_thresholds();
5122 setup_per_zone_lowmem_reserve();
5123 setup_per_zone_inactive_ratio();
5126 module_init(init_per_zone_wmark_min)
5129 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5130 * that we can call two helper functions whenever min_free_kbytes
5133 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5134 void __user *buffer, size_t *length, loff_t *ppos)
5136 proc_dointvec(table, write, buffer, length, ppos);
5138 setup_per_zone_wmarks();
5143 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5144 void __user *buffer, size_t *length, loff_t *ppos)
5149 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5154 zone->min_unmapped_pages = (zone->present_pages *
5155 sysctl_min_unmapped_ratio) / 100;
5159 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5160 void __user *buffer, size_t *length, loff_t *ppos)
5165 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5170 zone->min_slab_pages = (zone->present_pages *
5171 sysctl_min_slab_ratio) / 100;
5177 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5178 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5179 * whenever sysctl_lowmem_reserve_ratio changes.
5181 * The reserve ratio obviously has absolutely no relation with the
5182 * minimum watermarks. The lowmem reserve ratio can only make sense
5183 * if in function of the boot time zone sizes.
5185 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5186 void __user *buffer, size_t *length, loff_t *ppos)
5188 proc_dointvec_minmax(table, write, buffer, length, ppos);
5189 setup_per_zone_lowmem_reserve();
5194 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5195 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5196 * can have before it gets flushed back to buddy allocator.
5199 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5200 void __user *buffer, size_t *length, loff_t *ppos)
5206 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5207 if (!write || (ret < 0))
5209 for_each_populated_zone(zone) {
5210 for_each_possible_cpu(cpu) {
5212 high = zone->present_pages / percpu_pagelist_fraction;
5213 setup_pagelist_highmark(
5214 per_cpu_ptr(zone->pageset, cpu), high);
5220 int hashdist = HASHDIST_DEFAULT;
5223 static int __init set_hashdist(char *str)
5227 hashdist = simple_strtoul(str, &str, 0);
5230 __setup("hashdist=", set_hashdist);
5234 * allocate a large system hash table from bootmem
5235 * - it is assumed that the hash table must contain an exact power-of-2
5236 * quantity of entries
5237 * - limit is the number of hash buckets, not the total allocation size
5239 void *__init alloc_large_system_hash(const char *tablename,
5240 unsigned long bucketsize,
5241 unsigned long numentries,
5244 unsigned int *_hash_shift,
5245 unsigned int *_hash_mask,
5246 unsigned long limit)
5248 unsigned long long max = limit;
5249 unsigned long log2qty, size;
5252 /* allow the kernel cmdline to have a say */
5254 /* round applicable memory size up to nearest megabyte */
5255 numentries = nr_kernel_pages;
5256 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5257 numentries >>= 20 - PAGE_SHIFT;
5258 numentries <<= 20 - PAGE_SHIFT;
5260 /* limit to 1 bucket per 2^scale bytes of low memory */
5261 if (scale > PAGE_SHIFT)
5262 numentries >>= (scale - PAGE_SHIFT);
5264 numentries <<= (PAGE_SHIFT - scale);
5266 /* Make sure we've got at least a 0-order allocation.. */
5267 if (unlikely(flags & HASH_SMALL)) {
5268 /* Makes no sense without HASH_EARLY */
5269 WARN_ON(!(flags & HASH_EARLY));
5270 if (!(numentries >> *_hash_shift)) {
5271 numentries = 1UL << *_hash_shift;
5272 BUG_ON(!numentries);
5274 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5275 numentries = PAGE_SIZE / bucketsize;
5277 numentries = roundup_pow_of_two(numentries);
5279 /* limit allocation size to 1/16 total memory by default */
5281 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5282 do_div(max, bucketsize);
5284 max = min(max, 0x80000000ULL);
5286 if (numentries > max)
5289 log2qty = ilog2(numentries);
5292 size = bucketsize << log2qty;
5293 if (flags & HASH_EARLY)
5294 table = alloc_bootmem_nopanic(size);
5296 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5299 * If bucketsize is not a power-of-two, we may free
5300 * some pages at the end of hash table which
5301 * alloc_pages_exact() automatically does
5303 if (get_order(size) < MAX_ORDER) {
5304 table = alloc_pages_exact(size, GFP_ATOMIC);
5305 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5308 } while (!table && size > PAGE_SIZE && --log2qty);
5311 panic("Failed to allocate %s hash table\n", tablename);
5313 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5316 ilog2(size) - PAGE_SHIFT,
5320 *_hash_shift = log2qty;
5322 *_hash_mask = (1 << log2qty) - 1;
5327 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5328 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5331 #ifdef CONFIG_SPARSEMEM
5332 return __pfn_to_section(pfn)->pageblock_flags;
5334 return zone->pageblock_flags;
5335 #endif /* CONFIG_SPARSEMEM */
5338 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5340 #ifdef CONFIG_SPARSEMEM
5341 pfn &= (PAGES_PER_SECTION-1);
5342 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5344 pfn = pfn - zone->zone_start_pfn;
5345 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5346 #endif /* CONFIG_SPARSEMEM */
5350 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5351 * @page: The page within the block of interest
5352 * @start_bitidx: The first bit of interest to retrieve
5353 * @end_bitidx: The last bit of interest
5354 * returns pageblock_bits flags
5356 unsigned long get_pageblock_flags_group(struct page *page,
5357 int start_bitidx, int end_bitidx)
5360 unsigned long *bitmap;
5361 unsigned long pfn, bitidx;
5362 unsigned long flags = 0;
5363 unsigned long value = 1;
5365 zone = page_zone(page);
5366 pfn = page_to_pfn(page);
5367 bitmap = get_pageblock_bitmap(zone, pfn);
5368 bitidx = pfn_to_bitidx(zone, pfn);
5370 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5371 if (test_bit(bitidx + start_bitidx, bitmap))
5378 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5379 * @page: The page within the block of interest
5380 * @start_bitidx: The first bit of interest
5381 * @end_bitidx: The last bit of interest
5382 * @flags: The flags to set
5384 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5385 int start_bitidx, int end_bitidx)
5388 unsigned long *bitmap;
5389 unsigned long pfn, bitidx;
5390 unsigned long value = 1;
5392 zone = page_zone(page);
5393 pfn = page_to_pfn(page);
5394 bitmap = get_pageblock_bitmap(zone, pfn);
5395 bitidx = pfn_to_bitidx(zone, pfn);
5396 VM_BUG_ON(pfn < zone->zone_start_pfn);
5397 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5399 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5401 __set_bit(bitidx + start_bitidx, bitmap);
5403 __clear_bit(bitidx + start_bitidx, bitmap);
5407 * This is designed as sub function...plz see page_isolation.c also.
5408 * set/clear page block's type to be ISOLATE.
5409 * page allocater never alloc memory from ISOLATE block.
5413 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5415 unsigned long pfn, iter, found;
5417 * For avoiding noise data, lru_add_drain_all() should be called
5418 * If ZONE_MOVABLE, the zone never contains immobile pages
5420 if (zone_idx(zone) == ZONE_MOVABLE)
5423 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5426 pfn = page_to_pfn(page);
5427 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5428 unsigned long check = pfn + iter;
5430 if (!pfn_valid_within(check))
5433 page = pfn_to_page(check);
5434 if (!page_count(page)) {
5435 if (PageBuddy(page))
5436 iter += (1 << page_order(page)) - 1;
5442 * If there are RECLAIMABLE pages, we need to check it.
5443 * But now, memory offline itself doesn't call shrink_slab()
5444 * and it still to be fixed.
5447 * If the page is not RAM, page_count()should be 0.
5448 * we don't need more check. This is an _used_ not-movable page.
5450 * The problematic thing here is PG_reserved pages. PG_reserved
5451 * is set to both of a memory hole page and a _used_ kernel
5460 bool is_pageblock_removable_nolock(struct page *page)
5466 * We have to be careful here because we are iterating over memory
5467 * sections which are not zone aware so we might end up outside of
5468 * the zone but still within the section.
5469 * We have to take care about the node as well. If the node is offline
5470 * its NODE_DATA will be NULL - see page_zone.
5472 if (!node_online(page_to_nid(page)))
5475 zone = page_zone(page);
5476 pfn = page_to_pfn(page);
5477 if (zone->zone_start_pfn > pfn ||
5478 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5481 return __count_immobile_pages(zone, page, 0);
5484 int set_migratetype_isolate(struct page *page)
5487 unsigned long flags, pfn;
5488 struct memory_isolate_notify arg;
5492 zone = page_zone(page);
5494 spin_lock_irqsave(&zone->lock, flags);
5496 pfn = page_to_pfn(page);
5497 arg.start_pfn = pfn;
5498 arg.nr_pages = pageblock_nr_pages;
5499 arg.pages_found = 0;
5502 * It may be possible to isolate a pageblock even if the
5503 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5504 * notifier chain is used by balloon drivers to return the
5505 * number of pages in a range that are held by the balloon
5506 * driver to shrink memory. If all the pages are accounted for
5507 * by balloons, are free, or on the LRU, isolation can continue.
5508 * Later, for example, when memory hotplug notifier runs, these
5509 * pages reported as "can be isolated" should be isolated(freed)
5510 * by the balloon driver through the memory notifier chain.
5512 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5513 notifier_ret = notifier_to_errno(notifier_ret);
5517 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5518 * We just check MOVABLE pages.
5520 if (__count_immobile_pages(zone, page, arg.pages_found))
5524 * immobile means "not-on-lru" paes. If immobile is larger than
5525 * removable-by-driver pages reported by notifier, we'll fail.
5530 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5531 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5534 spin_unlock_irqrestore(&zone->lock, flags);
5540 void unset_migratetype_isolate(struct page *page)
5543 unsigned long flags;
5544 zone = page_zone(page);
5545 spin_lock_irqsave(&zone->lock, flags);
5546 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5548 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5549 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5551 spin_unlock_irqrestore(&zone->lock, flags);
5556 static unsigned long pfn_max_align_down(unsigned long pfn)
5558 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5559 pageblock_nr_pages) - 1);
5562 static unsigned long pfn_max_align_up(unsigned long pfn)
5564 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5565 pageblock_nr_pages));
5568 static struct page *
5569 __alloc_contig_migrate_alloc(struct page *page, unsigned long private,
5572 return alloc_page(GFP_HIGHUSER_MOVABLE);
5575 /* [start, end) must belong to a single zone. */
5576 static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5578 /* This function is based on compact_zone() from compaction.c. */
5580 unsigned long pfn = start;
5581 unsigned int tries = 0;
5584 struct compact_control cc = {
5585 .nr_migratepages = 0,
5587 .zone = page_zone(pfn_to_page(start)),
5590 INIT_LIST_HEAD(&cc.migratepages);
5592 migrate_prep_local();
5594 while (pfn < end || !list_empty(&cc.migratepages)) {
5595 if (fatal_signal_pending(current)) {
5600 if (list_empty(&cc.migratepages)) {
5601 cc.nr_migratepages = 0;
5602 pfn = isolate_migratepages_range(cc.zone, &cc,
5609 } else if (++tries == 5) {
5610 ret = ret < 0 ? ret : -EBUSY;
5614 ret = migrate_pages(&cc.migratepages,
5615 __alloc_contig_migrate_alloc,
5619 putback_lru_pages(&cc.migratepages);
5620 return ret > 0 ? 0 : ret;
5624 * alloc_contig_range() -- tries to allocate given range of pages
5625 * @start: start PFN to allocate
5626 * @end: one-past-the-last PFN to allocate
5628 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5629 * aligned, however it's the caller's responsibility to guarantee that
5630 * we are the only thread that changes migrate type of pageblocks the
5633 * The PFN range must belong to a single zone.
5635 * Returns zero on success or negative error code. On success all
5636 * pages which PFN is in [start, end) are allocated for the caller and
5637 * need to be freed with free_contig_range().
5639 int alloc_contig_range(unsigned long start, unsigned long end)
5641 struct zone *zone = page_zone(pfn_to_page(start));
5642 unsigned long outer_start, outer_end;
5646 * What we do here is we mark all pageblocks in range as
5647 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5648 * have different sizes, and due to the way page allocator
5649 * work, we align the range to biggest of the two pages so
5650 * that page allocator won't try to merge buddies from
5651 * different pageblocks and change MIGRATE_ISOLATE to some
5652 * other migration type.
5654 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5655 * migrate the pages from an unaligned range (ie. pages that
5656 * we are interested in). This will put all the pages in
5657 * range back to page allocator as MIGRATE_ISOLATE.
5659 * When this is done, we take the pages in range from page
5660 * allocator removing them from the buddy system. This way
5661 * page allocator will never consider using them.
5663 * This lets us mark the pageblocks back as
5664 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5665 * aligned range but not in the unaligned, original range are
5666 * put back to page allocator so that buddy can use them.
5669 ret = start_isolate_page_range(pfn_max_align_down(start),
5670 pfn_max_align_up(end));
5674 ret = __alloc_contig_migrate_range(start, end);
5679 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5680 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5681 * more, all pages in [start, end) are free in page allocator.
5682 * What we are going to do is to allocate all pages from
5683 * [start, end) (that is remove them from page allocator).
5685 * The only problem is that pages at the beginning and at the
5686 * end of interesting range may be not aligned with pages that
5687 * page allocator holds, ie. they can be part of higher order
5688 * pages. Because of this, we reserve the bigger range and
5689 * once this is done free the pages we are not interested in.
5691 * We don't have to hold zone->lock here because the pages are
5692 * isolated thus they won't get removed from buddy.
5695 lru_add_drain_all();
5699 outer_start = start;
5700 while (!PageBuddy(pfn_to_page(outer_start))) {
5701 if (++order >= MAX_ORDER) {
5705 outer_start &= ~0UL << order;
5708 /* Make sure the range is really isolated. */
5709 if (test_pages_isolated(outer_start, end)) {
5710 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5716 outer_end = isolate_freepages_range(outer_start, end);
5722 /* Free head and tail (if any) */
5723 if (start != outer_start)
5724 free_contig_range(outer_start, start - outer_start);
5725 if (end != outer_end)
5726 free_contig_range(end, outer_end - end);
5729 undo_isolate_page_range(pfn_max_align_down(start),
5730 pfn_max_align_up(end));
5734 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5736 for (; nr_pages--; ++pfn)
5737 __free_page(pfn_to_page(pfn));
5741 #ifdef CONFIG_MEMORY_HOTREMOVE
5743 * All pages in the range must be isolated before calling this.
5746 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5752 unsigned long flags;
5753 /* find the first valid pfn */
5754 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5759 zone = page_zone(pfn_to_page(pfn));
5760 spin_lock_irqsave(&zone->lock, flags);
5762 while (pfn < end_pfn) {
5763 if (!pfn_valid(pfn)) {
5767 page = pfn_to_page(pfn);
5768 BUG_ON(page_count(page));
5769 BUG_ON(!PageBuddy(page));
5770 order = page_order(page);
5771 #ifdef CONFIG_DEBUG_VM
5772 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5773 pfn, 1 << order, end_pfn);
5775 list_del(&page->lru);
5776 rmv_page_order(page);
5777 zone->free_area[order].nr_free--;
5778 __mod_zone_page_state(zone, NR_FREE_PAGES,
5780 for (i = 0; i < (1 << order); i++)
5781 SetPageReserved((page+i));
5782 pfn += (1 << order);
5784 spin_unlock_irqrestore(&zone->lock, flags);
5788 #ifdef CONFIG_MEMORY_FAILURE
5789 bool is_free_buddy_page(struct page *page)
5791 struct zone *zone = page_zone(page);
5792 unsigned long pfn = page_to_pfn(page);
5793 unsigned long flags;
5796 spin_lock_irqsave(&zone->lock, flags);
5797 for (order = 0; order < MAX_ORDER; order++) {
5798 struct page *page_head = page - (pfn & ((1 << order) - 1));
5800 if (PageBuddy(page_head) && page_order(page_head) >= order)
5803 spin_unlock_irqrestore(&zone->lock, flags);
5805 return order < MAX_ORDER;
5809 static struct trace_print_flags pageflag_names[] = {
5810 {1UL << PG_locked, "locked" },
5811 {1UL << PG_error, "error" },
5812 {1UL << PG_referenced, "referenced" },
5813 {1UL << PG_uptodate, "uptodate" },
5814 {1UL << PG_dirty, "dirty" },
5815 {1UL << PG_lru, "lru" },
5816 {1UL << PG_active, "active" },
5817 {1UL << PG_slab, "slab" },
5818 {1UL << PG_owner_priv_1, "owner_priv_1" },
5819 {1UL << PG_arch_1, "arch_1" },
5820 {1UL << PG_reserved, "reserved" },
5821 {1UL << PG_private, "private" },
5822 {1UL << PG_private_2, "private_2" },
5823 {1UL << PG_writeback, "writeback" },
5824 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5825 {1UL << PG_head, "head" },
5826 {1UL << PG_tail, "tail" },
5828 {1UL << PG_compound, "compound" },
5830 {1UL << PG_swapcache, "swapcache" },
5831 {1UL << PG_mappedtodisk, "mappedtodisk" },
5832 {1UL << PG_reclaim, "reclaim" },
5833 {1UL << PG_swapbacked, "swapbacked" },
5834 {1UL << PG_unevictable, "unevictable" },
5836 {1UL << PG_mlocked, "mlocked" },
5838 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5839 {1UL << PG_uncached, "uncached" },
5841 #ifdef CONFIG_MEMORY_FAILURE
5842 {1UL << PG_hwpoison, "hwpoison" },
5847 static void dump_page_flags(unsigned long flags)
5849 const char *delim = "";
5853 printk(KERN_ALERT "page flags: %#lx(", flags);
5855 /* remove zone id */
5856 flags &= (1UL << NR_PAGEFLAGS) - 1;
5858 for (i = 0; pageflag_names[i].name && flags; i++) {
5860 mask = pageflag_names[i].mask;
5861 if ((flags & mask) != mask)
5865 printk("%s%s", delim, pageflag_names[i].name);
5869 /* check for left over flags */
5871 printk("%s%#lx", delim, flags);
5876 void dump_page(struct page *page)
5879 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5880 page, atomic_read(&page->_count), page_mapcount(page),
5881 page->mapping, page->index);
5882 dump_page_flags(page->flags);
5883 mem_cgroup_print_bad_page(page);