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/page-debug-flags.h>
62 #include <asm/tlbflush.h>
63 #include <asm/div64.h>
66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67 DEFINE_PER_CPU(int, numa_node);
68 EXPORT_PER_CPU_SYMBOL(numa_node);
71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
78 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
83 * Array of node states.
85 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86 [N_POSSIBLE] = NODE_MASK_ALL,
87 [N_ONLINE] = { { [0] = 1UL } },
89 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
91 [N_HIGH_MEMORY] = { { [0] = 1UL } },
93 [N_CPU] = { { [0] = 1UL } },
96 EXPORT_SYMBOL(node_states);
98 unsigned long totalram_pages __read_mostly;
99 unsigned long totalreserve_pages __read_mostly;
100 int percpu_pagelist_fraction;
101 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
103 #ifdef CONFIG_PM_SLEEP
105 * The following functions are used by the suspend/hibernate code to temporarily
106 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
107 * while devices are suspended. To avoid races with the suspend/hibernate code,
108 * they should always be called with pm_mutex held (gfp_allowed_mask also should
109 * only be modified with pm_mutex held, unless the suspend/hibernate code is
110 * guaranteed not to run in parallel with that modification).
113 static gfp_t saved_gfp_mask;
115 void pm_restore_gfp_mask(void)
117 WARN_ON(!mutex_is_locked(&pm_mutex));
118 if (saved_gfp_mask) {
119 gfp_allowed_mask = saved_gfp_mask;
124 void pm_restrict_gfp_mask(void)
126 WARN_ON(!mutex_is_locked(&pm_mutex));
127 WARN_ON(saved_gfp_mask);
128 saved_gfp_mask = gfp_allowed_mask;
129 gfp_allowed_mask &= ~GFP_IOFS;
132 bool pm_suspended_storage(void)
134 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
138 #endif /* CONFIG_PM_SLEEP */
140 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
141 int pageblock_order __read_mostly;
144 static void __free_pages_ok(struct page *page, unsigned int order);
147 * results with 256, 32 in the lowmem_reserve sysctl:
148 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
149 * 1G machine -> (16M dma, 784M normal, 224M high)
150 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
151 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
152 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
154 * TBD: should special case ZONE_DMA32 machines here - in those we normally
155 * don't need any ZONE_NORMAL reservation
157 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
158 #ifdef CONFIG_ZONE_DMA
161 #ifdef CONFIG_ZONE_DMA32
164 #ifdef CONFIG_HIGHMEM
170 EXPORT_SYMBOL(totalram_pages);
172 static char * const zone_names[MAX_NR_ZONES] = {
173 #ifdef CONFIG_ZONE_DMA
176 #ifdef CONFIG_ZONE_DMA32
180 #ifdef CONFIG_HIGHMEM
186 int min_free_kbytes = 1024;
188 static unsigned long __meminitdata nr_kernel_pages;
189 static unsigned long __meminitdata nr_all_pages;
190 static unsigned long __meminitdata dma_reserve;
192 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
193 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
194 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
195 static unsigned long __initdata required_kernelcore;
196 static unsigned long __initdata required_movablecore;
197 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
199 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
201 EXPORT_SYMBOL(movable_zone);
202 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
205 int nr_node_ids __read_mostly = MAX_NUMNODES;
206 int nr_online_nodes __read_mostly = 1;
207 EXPORT_SYMBOL(nr_node_ids);
208 EXPORT_SYMBOL(nr_online_nodes);
211 int page_group_by_mobility_disabled __read_mostly;
213 static void set_pageblock_migratetype(struct page *page, int migratetype)
216 if (unlikely(page_group_by_mobility_disabled))
217 migratetype = MIGRATE_UNMOVABLE;
219 set_pageblock_flags_group(page, (unsigned long)migratetype,
220 PB_migrate, PB_migrate_end);
223 bool oom_killer_disabled __read_mostly;
225 #ifdef CONFIG_DEBUG_VM
226 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
230 unsigned long pfn = page_to_pfn(page);
233 seq = zone_span_seqbegin(zone);
234 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
236 else if (pfn < zone->zone_start_pfn)
238 } while (zone_span_seqretry(zone, seq));
243 static int page_is_consistent(struct zone *zone, struct page *page)
245 if (!pfn_valid_within(page_to_pfn(page)))
247 if (zone != page_zone(page))
253 * Temporary debugging check for pages not lying within a given zone.
255 static int bad_range(struct zone *zone, struct page *page)
257 if (page_outside_zone_boundaries(zone, page))
259 if (!page_is_consistent(zone, page))
265 static inline int bad_range(struct zone *zone, struct page *page)
271 static void bad_page(struct page *page)
273 static unsigned long resume;
274 static unsigned long nr_shown;
275 static unsigned long nr_unshown;
277 /* Don't complain about poisoned pages */
278 if (PageHWPoison(page)) {
279 reset_page_mapcount(page); /* remove PageBuddy */
284 * Allow a burst of 60 reports, then keep quiet for that minute;
285 * or allow a steady drip of one report per second.
287 if (nr_shown == 60) {
288 if (time_before(jiffies, resume)) {
294 "BUG: Bad page state: %lu messages suppressed\n",
301 resume = jiffies + 60 * HZ;
303 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
304 current->comm, page_to_pfn(page));
310 /* Leave bad fields for debug, except PageBuddy could make trouble */
311 reset_page_mapcount(page); /* remove PageBuddy */
312 add_taint(TAINT_BAD_PAGE);
316 * Higher-order pages are called "compound pages". They are structured thusly:
318 * The first PAGE_SIZE page is called the "head page".
320 * The remaining PAGE_SIZE pages are called "tail pages".
322 * All pages have PG_compound set. All tail pages have their ->first_page
323 * pointing at the head page.
325 * The first tail page's ->lru.next holds the address of the compound page's
326 * put_page() function. Its ->lru.prev holds the order of allocation.
327 * This usage means that zero-order pages may not be compound.
330 static void free_compound_page(struct page *page)
332 __free_pages_ok(page, compound_order(page));
335 void prep_compound_page(struct page *page, unsigned long order)
338 int nr_pages = 1 << order;
340 set_compound_page_dtor(page, free_compound_page);
341 set_compound_order(page, order);
343 for (i = 1; i < nr_pages; i++) {
344 struct page *p = page + i;
346 set_page_count(p, 0);
347 p->first_page = page;
351 /* update __split_huge_page_refcount if you change this function */
352 static int destroy_compound_page(struct page *page, unsigned long order)
355 int nr_pages = 1 << order;
358 if (unlikely(compound_order(page) != order) ||
359 unlikely(!PageHead(page))) {
364 __ClearPageHead(page);
366 for (i = 1; i < nr_pages; i++) {
367 struct page *p = page + i;
369 if (unlikely(!PageTail(p) || (p->first_page != page))) {
379 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
384 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
385 * and __GFP_HIGHMEM from hard or soft interrupt context.
387 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
388 for (i = 0; i < (1 << order); i++)
389 clear_highpage(page + i);
392 #ifdef CONFIG_DEBUG_PAGEALLOC
393 unsigned int _debug_guardpage_minorder;
395 static int __init debug_guardpage_minorder_setup(char *buf)
399 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
400 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
403 _debug_guardpage_minorder = res;
404 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
407 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
409 static inline void set_page_guard_flag(struct page *page)
411 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
414 static inline void clear_page_guard_flag(struct page *page)
416 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
419 static inline void set_page_guard_flag(struct page *page) { }
420 static inline void clear_page_guard_flag(struct page *page) { }
423 static inline void set_page_order(struct page *page, int order)
425 set_page_private(page, order);
426 __SetPageBuddy(page);
429 static inline void rmv_page_order(struct page *page)
431 __ClearPageBuddy(page);
432 set_page_private(page, 0);
436 * Locate the struct page for both the matching buddy in our
437 * pair (buddy1) and the combined O(n+1) page they form (page).
439 * 1) Any buddy B1 will have an order O twin B2 which satisfies
440 * the following equation:
442 * For example, if the starting buddy (buddy2) is #8 its order
444 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
446 * 2) Any buddy B will have an order O+1 parent P which
447 * satisfies the following equation:
450 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
452 static inline unsigned long
453 __find_buddy_index(unsigned long page_idx, unsigned int order)
455 return page_idx ^ (1 << order);
459 * This function checks whether a page is free && is the buddy
460 * we can do coalesce a page and its buddy if
461 * (a) the buddy is not in a hole &&
462 * (b) the buddy is in the buddy system &&
463 * (c) a page and its buddy have the same order &&
464 * (d) a page and its buddy are in the same zone.
466 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
467 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
469 * For recording page's order, we use page_private(page).
471 static inline int page_is_buddy(struct page *page, struct page *buddy,
474 if (!pfn_valid_within(page_to_pfn(buddy)))
477 if (page_zone_id(page) != page_zone_id(buddy))
480 if (page_is_guard(buddy) && page_order(buddy) == order) {
481 VM_BUG_ON(page_count(buddy) != 0);
485 if (PageBuddy(buddy) && page_order(buddy) == order) {
486 VM_BUG_ON(page_count(buddy) != 0);
493 * Freeing function for a buddy system allocator.
495 * The concept of a buddy system is to maintain direct-mapped table
496 * (containing bit values) for memory blocks of various "orders".
497 * The bottom level table contains the map for the smallest allocatable
498 * units of memory (here, pages), and each level above it describes
499 * pairs of units from the levels below, hence, "buddies".
500 * At a high level, all that happens here is marking the table entry
501 * at the bottom level available, and propagating the changes upward
502 * as necessary, plus some accounting needed to play nicely with other
503 * parts of the VM system.
504 * At each level, we keep a list of pages, which are heads of continuous
505 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
506 * order is recorded in page_private(page) field.
507 * So when we are allocating or freeing one, we can derive the state of the
508 * other. That is, if we allocate a small block, and both were
509 * free, the remainder of the region must be split into blocks.
510 * If a block is freed, and its buddy is also free, then this
511 * triggers coalescing into a block of larger size.
516 static inline void __free_one_page(struct page *page,
517 struct zone *zone, unsigned int order,
520 unsigned long page_idx;
521 unsigned long combined_idx;
522 unsigned long uninitialized_var(buddy_idx);
525 if (unlikely(PageCompound(page)))
526 if (unlikely(destroy_compound_page(page, order)))
529 VM_BUG_ON(migratetype == -1);
531 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
533 VM_BUG_ON(page_idx & ((1 << order) - 1));
534 VM_BUG_ON(bad_range(zone, page));
536 while (order < MAX_ORDER-1) {
537 buddy_idx = __find_buddy_index(page_idx, order);
538 buddy = page + (buddy_idx - page_idx);
539 if (!page_is_buddy(page, buddy, order))
542 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
543 * merge with it and move up one order.
545 if (page_is_guard(buddy)) {
546 clear_page_guard_flag(buddy);
547 set_page_private(page, 0);
548 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
550 list_del(&buddy->lru);
551 zone->free_area[order].nr_free--;
552 rmv_page_order(buddy);
554 combined_idx = buddy_idx & page_idx;
555 page = page + (combined_idx - page_idx);
556 page_idx = combined_idx;
559 set_page_order(page, order);
562 * If this is not the largest possible page, check if the buddy
563 * of the next-highest order is free. If it is, it's possible
564 * that pages are being freed that will coalesce soon. In case,
565 * that is happening, add the free page to the tail of the list
566 * so it's less likely to be used soon and more likely to be merged
567 * as a higher order page
569 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
570 struct page *higher_page, *higher_buddy;
571 combined_idx = buddy_idx & page_idx;
572 higher_page = page + (combined_idx - page_idx);
573 buddy_idx = __find_buddy_index(combined_idx, order + 1);
574 higher_buddy = page + (buddy_idx - combined_idx);
575 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
576 list_add_tail(&page->lru,
577 &zone->free_area[order].free_list[migratetype]);
582 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
584 zone->free_area[order].nr_free++;
588 * free_page_mlock() -- clean up attempts to free and mlocked() page.
589 * Page should not be on lru, so no need to fix that up.
590 * free_pages_check() will verify...
592 static inline void free_page_mlock(struct page *page)
594 __dec_zone_page_state(page, NR_MLOCK);
595 __count_vm_event(UNEVICTABLE_MLOCKFREED);
598 static inline int free_pages_check(struct page *page)
600 if (unlikely(page_mapcount(page) |
601 (page->mapping != NULL) |
602 (atomic_read(&page->_count) != 0) |
603 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
604 (mem_cgroup_bad_page_check(page)))) {
608 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
609 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
614 * Frees a number of pages from the PCP lists
615 * Assumes all pages on list are in same zone, and of same order.
616 * count is the number of pages to free.
618 * If the zone was previously in an "all pages pinned" state then look to
619 * see if this freeing clears that state.
621 * And clear the zone's pages_scanned counter, to hold off the "all pages are
622 * pinned" detection logic.
624 static void free_pcppages_bulk(struct zone *zone, int count,
625 struct per_cpu_pages *pcp)
631 spin_lock(&zone->lock);
632 zone->all_unreclaimable = 0;
633 zone->pages_scanned = 0;
637 struct list_head *list;
640 * Remove pages from lists in a round-robin fashion. A
641 * batch_free count is maintained that is incremented when an
642 * empty list is encountered. This is so more pages are freed
643 * off fuller lists instead of spinning excessively around empty
648 if (++migratetype == MIGRATE_PCPTYPES)
650 list = &pcp->lists[migratetype];
651 } while (list_empty(list));
653 /* This is the only non-empty list. Free them all. */
654 if (batch_free == MIGRATE_PCPTYPES)
655 batch_free = to_free;
658 page = list_entry(list->prev, struct page, lru);
659 /* must delete as __free_one_page list manipulates */
660 list_del(&page->lru);
661 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
662 __free_one_page(page, zone, 0, page_private(page));
663 trace_mm_page_pcpu_drain(page, 0, page_private(page));
664 } while (--to_free && --batch_free && !list_empty(list));
666 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
667 spin_unlock(&zone->lock);
670 static void free_one_page(struct zone *zone, struct page *page, int order,
673 spin_lock(&zone->lock);
674 zone->all_unreclaimable = 0;
675 zone->pages_scanned = 0;
677 __free_one_page(page, zone, order, migratetype);
678 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
679 spin_unlock(&zone->lock);
682 static bool free_pages_prepare(struct page *page, unsigned int order)
687 trace_mm_page_free(page, order);
688 kmemcheck_free_shadow(page, order);
691 page->mapping = NULL;
692 for (i = 0; i < (1 << order); i++)
693 bad += free_pages_check(page + i);
697 if (!PageHighMem(page)) {
698 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
699 debug_check_no_obj_freed(page_address(page),
702 arch_free_page(page, order);
703 kernel_map_pages(page, 1 << order, 0);
708 static void __free_pages_ok(struct page *page, unsigned int order)
711 int wasMlocked = __TestClearPageMlocked(page);
713 if (!free_pages_prepare(page, order))
716 local_irq_save(flags);
717 if (unlikely(wasMlocked))
718 free_page_mlock(page);
719 __count_vm_events(PGFREE, 1 << order);
720 free_one_page(page_zone(page), page, order,
721 get_pageblock_migratetype(page));
722 local_irq_restore(flags);
726 * permit the bootmem allocator to evade page validation on high-order frees
728 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
731 __ClearPageReserved(page);
732 set_page_count(page, 0);
733 set_page_refcounted(page);
739 for (loop = 0; loop < (1 << order); loop++) {
740 struct page *p = &page[loop];
742 if (loop + 1 < (1 << order))
744 __ClearPageReserved(p);
745 set_page_count(p, 0);
748 set_page_refcounted(page);
749 __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 void drain_all_pages(void)
1169 on_each_cpu(drain_local_pages, NULL, 1);
1172 #ifdef CONFIG_HIBERNATION
1174 void mark_free_pages(struct zone *zone)
1176 unsigned long pfn, max_zone_pfn;
1177 unsigned long flags;
1179 struct list_head *curr;
1181 if (!zone->spanned_pages)
1184 spin_lock_irqsave(&zone->lock, flags);
1186 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1187 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1188 if (pfn_valid(pfn)) {
1189 struct page *page = pfn_to_page(pfn);
1191 if (!swsusp_page_is_forbidden(page))
1192 swsusp_unset_page_free(page);
1195 for_each_migratetype_order(order, t) {
1196 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1199 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1200 for (i = 0; i < (1UL << order); i++)
1201 swsusp_set_page_free(pfn_to_page(pfn + i));
1204 spin_unlock_irqrestore(&zone->lock, flags);
1206 #endif /* CONFIG_PM */
1209 * Free a 0-order page
1210 * cold == 1 ? free a cold page : free a hot page
1212 void free_hot_cold_page(struct page *page, int cold)
1214 struct zone *zone = page_zone(page);
1215 struct per_cpu_pages *pcp;
1216 unsigned long flags;
1218 int wasMlocked = __TestClearPageMlocked(page);
1220 if (!free_pages_prepare(page, 0))
1223 migratetype = get_pageblock_migratetype(page);
1224 set_page_private(page, migratetype);
1225 local_irq_save(flags);
1226 if (unlikely(wasMlocked))
1227 free_page_mlock(page);
1228 __count_vm_event(PGFREE);
1231 * We only track unmovable, reclaimable and movable on pcp lists.
1232 * Free ISOLATE pages back to the allocator because they are being
1233 * offlined but treat RESERVE as movable pages so we can get those
1234 * areas back if necessary. Otherwise, we may have to free
1235 * excessively into the page allocator
1237 if (migratetype >= MIGRATE_PCPTYPES) {
1238 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1239 free_one_page(zone, page, 0, migratetype);
1242 migratetype = MIGRATE_MOVABLE;
1245 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1247 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1249 list_add(&page->lru, &pcp->lists[migratetype]);
1251 if (pcp->count >= pcp->high) {
1252 free_pcppages_bulk(zone, pcp->batch, pcp);
1253 pcp->count -= pcp->batch;
1257 local_irq_restore(flags);
1261 * Free a list of 0-order pages
1263 void free_hot_cold_page_list(struct list_head *list, int cold)
1265 struct page *page, *next;
1267 list_for_each_entry_safe(page, next, list, lru) {
1268 trace_mm_page_free_batched(page, cold);
1269 free_hot_cold_page(page, cold);
1274 * split_page takes a non-compound higher-order page, and splits it into
1275 * n (1<<order) sub-pages: page[0..n]
1276 * Each sub-page must be freed individually.
1278 * Note: this is probably too low level an operation for use in drivers.
1279 * Please consult with lkml before using this in your driver.
1281 void split_page(struct page *page, unsigned int order)
1285 VM_BUG_ON(PageCompound(page));
1286 VM_BUG_ON(!page_count(page));
1288 #ifdef CONFIG_KMEMCHECK
1290 * Split shadow pages too, because free(page[0]) would
1291 * otherwise free the whole shadow.
1293 if (kmemcheck_page_is_tracked(page))
1294 split_page(virt_to_page(page[0].shadow), order);
1297 for (i = 1; i < (1 << order); i++)
1298 set_page_refcounted(page + i);
1302 * Similar to split_page except the page is already free. As this is only
1303 * being used for migration, the migratetype of the block also changes.
1304 * As this is called with interrupts disabled, the caller is responsible
1305 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1308 * Note: this is probably too low level an operation for use in drivers.
1309 * Please consult with lkml before using this in your driver.
1311 int split_free_page(struct page *page)
1314 unsigned long watermark;
1317 BUG_ON(!PageBuddy(page));
1319 zone = page_zone(page);
1320 order = page_order(page);
1322 /* Obey watermarks as if the page was being allocated */
1323 watermark = low_wmark_pages(zone) + (1 << order);
1324 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1327 /* Remove page from free list */
1328 list_del(&page->lru);
1329 zone->free_area[order].nr_free--;
1330 rmv_page_order(page);
1331 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1333 /* Split into individual pages */
1334 set_page_refcounted(page);
1335 split_page(page, order);
1337 if (order >= pageblock_order - 1) {
1338 struct page *endpage = page + (1 << order) - 1;
1339 for (; page < endpage; page += pageblock_nr_pages)
1340 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1347 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1348 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1352 struct page *buffered_rmqueue(struct zone *preferred_zone,
1353 struct zone *zone, int order, gfp_t gfp_flags,
1356 unsigned long flags;
1358 int cold = !!(gfp_flags & __GFP_COLD);
1361 if (likely(order == 0)) {
1362 struct per_cpu_pages *pcp;
1363 struct list_head *list;
1365 local_irq_save(flags);
1366 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1367 list = &pcp->lists[migratetype];
1368 if (list_empty(list)) {
1369 pcp->count += rmqueue_bulk(zone, 0,
1372 if (unlikely(list_empty(list)))
1377 page = list_entry(list->prev, struct page, lru);
1379 page = list_entry(list->next, struct page, lru);
1381 list_del(&page->lru);
1384 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1386 * __GFP_NOFAIL is not to be used in new code.
1388 * All __GFP_NOFAIL callers should be fixed so that they
1389 * properly detect and handle allocation failures.
1391 * We most definitely don't want callers attempting to
1392 * allocate greater than order-1 page units with
1395 WARN_ON_ONCE(order > 1);
1397 spin_lock_irqsave(&zone->lock, flags);
1398 page = __rmqueue(zone, order, migratetype);
1399 spin_unlock(&zone->lock);
1402 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1405 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1406 zone_statistics(preferred_zone, zone, gfp_flags);
1407 local_irq_restore(flags);
1409 VM_BUG_ON(bad_range(zone, page));
1410 if (prep_new_page(page, order, gfp_flags))
1415 local_irq_restore(flags);
1419 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1420 #define ALLOC_WMARK_MIN WMARK_MIN
1421 #define ALLOC_WMARK_LOW WMARK_LOW
1422 #define ALLOC_WMARK_HIGH WMARK_HIGH
1423 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1425 /* Mask to get the watermark bits */
1426 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1428 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1429 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1430 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1432 #ifdef CONFIG_FAIL_PAGE_ALLOC
1435 struct fault_attr attr;
1437 u32 ignore_gfp_highmem;
1438 u32 ignore_gfp_wait;
1440 } fail_page_alloc = {
1441 .attr = FAULT_ATTR_INITIALIZER,
1442 .ignore_gfp_wait = 1,
1443 .ignore_gfp_highmem = 1,
1447 static int __init setup_fail_page_alloc(char *str)
1449 return setup_fault_attr(&fail_page_alloc.attr, str);
1451 __setup("fail_page_alloc=", setup_fail_page_alloc);
1453 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1455 if (order < fail_page_alloc.min_order)
1457 if (gfp_mask & __GFP_NOFAIL)
1459 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1461 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1464 return should_fail(&fail_page_alloc.attr, 1 << order);
1467 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1469 static int __init fail_page_alloc_debugfs(void)
1471 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1474 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1475 &fail_page_alloc.attr);
1477 return PTR_ERR(dir);
1479 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1480 &fail_page_alloc.ignore_gfp_wait))
1482 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1483 &fail_page_alloc.ignore_gfp_highmem))
1485 if (!debugfs_create_u32("min-order", mode, dir,
1486 &fail_page_alloc.min_order))
1491 debugfs_remove_recursive(dir);
1496 late_initcall(fail_page_alloc_debugfs);
1498 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1500 #else /* CONFIG_FAIL_PAGE_ALLOC */
1502 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1507 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1510 * Return true if free pages are above 'mark'. This takes into account the order
1511 * of the allocation.
1513 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1514 int classzone_idx, int alloc_flags, long free_pages)
1516 /* free_pages my go negative - that's OK */
1520 free_pages -= (1 << order) + 1;
1521 if (alloc_flags & ALLOC_HIGH)
1523 if (alloc_flags & ALLOC_HARDER)
1526 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1528 for (o = 0; o < order; o++) {
1529 /* At the next order, this order's pages become unavailable */
1530 free_pages -= z->free_area[o].nr_free << o;
1532 /* Require fewer higher order pages to be free */
1535 if (free_pages <= min)
1541 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1542 int classzone_idx, int alloc_flags)
1544 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1545 zone_page_state(z, NR_FREE_PAGES));
1548 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1549 int classzone_idx, int alloc_flags)
1551 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1553 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1554 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1556 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1562 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1563 * skip over zones that are not allowed by the cpuset, or that have
1564 * been recently (in last second) found to be nearly full. See further
1565 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1566 * that have to skip over a lot of full or unallowed zones.
1568 * If the zonelist cache is present in the passed in zonelist, then
1569 * returns a pointer to the allowed node mask (either the current
1570 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1572 * If the zonelist cache is not available for this zonelist, does
1573 * nothing and returns NULL.
1575 * If the fullzones BITMAP in the zonelist cache is stale (more than
1576 * a second since last zap'd) then we zap it out (clear its bits.)
1578 * We hold off even calling zlc_setup, until after we've checked the
1579 * first zone in the zonelist, on the theory that most allocations will
1580 * be satisfied from that first zone, so best to examine that zone as
1581 * quickly as we can.
1583 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1585 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1586 nodemask_t *allowednodes; /* zonelist_cache approximation */
1588 zlc = zonelist->zlcache_ptr;
1592 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1593 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1594 zlc->last_full_zap = jiffies;
1597 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1598 &cpuset_current_mems_allowed :
1599 &node_states[N_HIGH_MEMORY];
1600 return allowednodes;
1604 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1605 * if it is worth looking at further for free memory:
1606 * 1) Check that the zone isn't thought to be full (doesn't have its
1607 * bit set in the zonelist_cache fullzones BITMAP).
1608 * 2) Check that the zones node (obtained from the zonelist_cache
1609 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1610 * Return true (non-zero) if zone is worth looking at further, or
1611 * else return false (zero) if it is not.
1613 * This check -ignores- the distinction between various watermarks,
1614 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1615 * found to be full for any variation of these watermarks, it will
1616 * be considered full for up to one second by all requests, unless
1617 * we are so low on memory on all allowed nodes that we are forced
1618 * into the second scan of the zonelist.
1620 * In the second scan we ignore this zonelist cache and exactly
1621 * apply the watermarks to all zones, even it is slower to do so.
1622 * We are low on memory in the second scan, and should leave no stone
1623 * unturned looking for a free page.
1625 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1626 nodemask_t *allowednodes)
1628 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1629 int i; /* index of *z in zonelist zones */
1630 int n; /* node that zone *z is on */
1632 zlc = zonelist->zlcache_ptr;
1636 i = z - zonelist->_zonerefs;
1639 /* This zone is worth trying if it is allowed but not full */
1640 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1644 * Given 'z' scanning a zonelist, set the corresponding bit in
1645 * zlc->fullzones, so that subsequent attempts to allocate a page
1646 * from that zone don't waste time re-examining it.
1648 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1650 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1651 int i; /* index of *z in zonelist zones */
1653 zlc = zonelist->zlcache_ptr;
1657 i = z - zonelist->_zonerefs;
1659 set_bit(i, zlc->fullzones);
1663 * clear all zones full, called after direct reclaim makes progress so that
1664 * a zone that was recently full is not skipped over for up to a second
1666 static void zlc_clear_zones_full(struct zonelist *zonelist)
1668 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1670 zlc = zonelist->zlcache_ptr;
1674 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1677 #else /* CONFIG_NUMA */
1679 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1684 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1685 nodemask_t *allowednodes)
1690 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1694 static void zlc_clear_zones_full(struct zonelist *zonelist)
1697 #endif /* CONFIG_NUMA */
1700 * get_page_from_freelist goes through the zonelist trying to allocate
1703 static struct page *
1704 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1705 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1706 struct zone *preferred_zone, int migratetype)
1709 struct page *page = NULL;
1712 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1713 int zlc_active = 0; /* set if using zonelist_cache */
1714 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1716 classzone_idx = zone_idx(preferred_zone);
1719 * Scan zonelist, looking for a zone with enough free.
1720 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1722 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1723 high_zoneidx, nodemask) {
1724 if (NUMA_BUILD && zlc_active &&
1725 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1727 if ((alloc_flags & ALLOC_CPUSET) &&
1728 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1731 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1732 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1736 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1737 if (zone_watermark_ok(zone, order, mark,
1738 classzone_idx, alloc_flags))
1741 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1743 * we do zlc_setup if there are multiple nodes
1744 * and before considering the first zone allowed
1747 allowednodes = zlc_setup(zonelist, alloc_flags);
1752 if (zone_reclaim_mode == 0)
1753 goto this_zone_full;
1756 * As we may have just activated ZLC, check if the first
1757 * eligible zone has failed zone_reclaim recently.
1759 if (NUMA_BUILD && zlc_active &&
1760 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1763 ret = zone_reclaim(zone, gfp_mask, order);
1765 case ZONE_RECLAIM_NOSCAN:
1768 case ZONE_RECLAIM_FULL:
1769 /* scanned but unreclaimable */
1772 /* did we reclaim enough */
1773 if (!zone_watermark_ok(zone, order, mark,
1774 classzone_idx, alloc_flags))
1775 goto this_zone_full;
1780 page = buffered_rmqueue(preferred_zone, zone, order,
1781 gfp_mask, migratetype);
1786 zlc_mark_zone_full(zonelist, z);
1789 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1790 /* Disable zlc cache for second zonelist scan */
1798 * Large machines with many possible nodes should not always dump per-node
1799 * meminfo in irq context.
1801 static inline bool should_suppress_show_mem(void)
1806 ret = in_interrupt();
1811 static DEFINE_RATELIMIT_STATE(nopage_rs,
1812 DEFAULT_RATELIMIT_INTERVAL,
1813 DEFAULT_RATELIMIT_BURST);
1815 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1817 unsigned int filter = SHOW_MEM_FILTER_NODES;
1819 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1820 debug_guardpage_minorder() > 0)
1824 * This documents exceptions given to allocations in certain
1825 * contexts that are allowed to allocate outside current's set
1828 if (!(gfp_mask & __GFP_NOMEMALLOC))
1829 if (test_thread_flag(TIF_MEMDIE) ||
1830 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1831 filter &= ~SHOW_MEM_FILTER_NODES;
1832 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1833 filter &= ~SHOW_MEM_FILTER_NODES;
1836 struct va_format vaf;
1839 va_start(args, fmt);
1844 pr_warn("%pV", &vaf);
1849 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1850 current->comm, order, gfp_mask);
1853 if (!should_suppress_show_mem())
1858 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1859 unsigned long did_some_progress,
1860 unsigned long pages_reclaimed)
1862 /* Do not loop if specifically requested */
1863 if (gfp_mask & __GFP_NORETRY)
1866 /* Always retry if specifically requested */
1867 if (gfp_mask & __GFP_NOFAIL)
1871 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
1872 * making forward progress without invoking OOM. Suspend also disables
1873 * storage devices so kswapd will not help. Bail if we are suspending.
1875 if (!did_some_progress && pm_suspended_storage())
1879 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1880 * means __GFP_NOFAIL, but that may not be true in other
1883 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1887 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1888 * specified, then we retry until we no longer reclaim any pages
1889 * (above), or we've reclaimed an order of pages at least as
1890 * large as the allocation's order. In both cases, if the
1891 * allocation still fails, we stop retrying.
1893 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1899 static inline struct page *
1900 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1901 struct zonelist *zonelist, enum zone_type high_zoneidx,
1902 nodemask_t *nodemask, struct zone *preferred_zone,
1907 /* Acquire the OOM killer lock for the zones in zonelist */
1908 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1909 schedule_timeout_uninterruptible(1);
1914 * Go through the zonelist yet one more time, keep very high watermark
1915 * here, this is only to catch a parallel oom killing, we must fail if
1916 * we're still under heavy pressure.
1918 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1919 order, zonelist, high_zoneidx,
1920 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1921 preferred_zone, migratetype);
1925 if (!(gfp_mask & __GFP_NOFAIL)) {
1926 /* The OOM killer will not help higher order allocs */
1927 if (order > PAGE_ALLOC_COSTLY_ORDER)
1929 /* The OOM killer does not needlessly kill tasks for lowmem */
1930 if (high_zoneidx < ZONE_NORMAL)
1933 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1934 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1935 * The caller should handle page allocation failure by itself if
1936 * it specifies __GFP_THISNODE.
1937 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1939 if (gfp_mask & __GFP_THISNODE)
1942 /* Exhausted what can be done so it's blamo time */
1943 out_of_memory(zonelist, gfp_mask, order, nodemask);
1946 clear_zonelist_oom(zonelist, gfp_mask);
1950 #ifdef CONFIG_COMPACTION
1951 /* Try memory compaction for high-order allocations before reclaim */
1952 static struct page *
1953 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1954 struct zonelist *zonelist, enum zone_type high_zoneidx,
1955 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1956 int migratetype, unsigned long *did_some_progress,
1957 bool sync_migration)
1961 if (!order || compaction_deferred(preferred_zone))
1964 current->flags |= PF_MEMALLOC;
1965 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1966 nodemask, sync_migration);
1967 current->flags &= ~PF_MEMALLOC;
1968 if (*did_some_progress != COMPACT_SKIPPED) {
1970 /* Page migration frees to the PCP lists but we want merging */
1971 drain_pages(get_cpu());
1974 page = get_page_from_freelist(gfp_mask, nodemask,
1975 order, zonelist, high_zoneidx,
1976 alloc_flags, preferred_zone,
1979 preferred_zone->compact_considered = 0;
1980 preferred_zone->compact_defer_shift = 0;
1981 count_vm_event(COMPACTSUCCESS);
1986 * It's bad if compaction run occurs and fails.
1987 * The most likely reason is that pages exist,
1988 * but not enough to satisfy watermarks.
1990 count_vm_event(COMPACTFAIL);
1991 defer_compaction(preferred_zone);
1999 static inline struct page *
2000 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2001 struct zonelist *zonelist, enum zone_type high_zoneidx,
2002 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2003 int migratetype, unsigned long *did_some_progress,
2004 bool sync_migration)
2008 #endif /* CONFIG_COMPACTION */
2010 /* The really slow allocator path where we enter direct reclaim */
2011 static inline struct page *
2012 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2013 struct zonelist *zonelist, enum zone_type high_zoneidx,
2014 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2015 int migratetype, unsigned long *did_some_progress)
2017 struct page *page = NULL;
2018 struct reclaim_state reclaim_state;
2019 bool drained = false;
2023 /* We now go into synchronous reclaim */
2024 cpuset_memory_pressure_bump();
2025 current->flags |= PF_MEMALLOC;
2026 lockdep_set_current_reclaim_state(gfp_mask);
2027 reclaim_state.reclaimed_slab = 0;
2028 current->reclaim_state = &reclaim_state;
2030 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2032 current->reclaim_state = NULL;
2033 lockdep_clear_current_reclaim_state();
2034 current->flags &= ~PF_MEMALLOC;
2038 if (unlikely(!(*did_some_progress)))
2041 /* After successful reclaim, reconsider all zones for allocation */
2043 zlc_clear_zones_full(zonelist);
2046 page = get_page_from_freelist(gfp_mask, nodemask, order,
2047 zonelist, high_zoneidx,
2048 alloc_flags, preferred_zone,
2052 * If an allocation failed after direct reclaim, it could be because
2053 * pages are pinned on the per-cpu lists. Drain them and try again
2055 if (!page && !drained) {
2065 * This is called in the allocator slow-path if the allocation request is of
2066 * sufficient urgency to ignore watermarks and take other desperate measures
2068 static inline struct page *
2069 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2070 struct zonelist *zonelist, enum zone_type high_zoneidx,
2071 nodemask_t *nodemask, struct zone *preferred_zone,
2077 page = get_page_from_freelist(gfp_mask, nodemask, order,
2078 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2079 preferred_zone, migratetype);
2081 if (!page && gfp_mask & __GFP_NOFAIL)
2082 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2083 } while (!page && (gfp_mask & __GFP_NOFAIL));
2089 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2090 enum zone_type high_zoneidx,
2091 enum zone_type classzone_idx)
2096 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2097 wakeup_kswapd(zone, order, classzone_idx);
2101 gfp_to_alloc_flags(gfp_t gfp_mask)
2103 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2104 const gfp_t wait = gfp_mask & __GFP_WAIT;
2106 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2107 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2110 * The caller may dip into page reserves a bit more if the caller
2111 * cannot run direct reclaim, or if the caller has realtime scheduling
2112 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2113 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2115 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2119 * Not worth trying to allocate harder for
2120 * __GFP_NOMEMALLOC even if it can't schedule.
2122 if (!(gfp_mask & __GFP_NOMEMALLOC))
2123 alloc_flags |= ALLOC_HARDER;
2125 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2126 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2128 alloc_flags &= ~ALLOC_CPUSET;
2129 } else if (unlikely(rt_task(current)) && !in_interrupt())
2130 alloc_flags |= ALLOC_HARDER;
2132 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2133 if (!in_interrupt() &&
2134 ((current->flags & PF_MEMALLOC) ||
2135 unlikely(test_thread_flag(TIF_MEMDIE))))
2136 alloc_flags |= ALLOC_NO_WATERMARKS;
2142 static inline struct page *
2143 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2144 struct zonelist *zonelist, enum zone_type high_zoneidx,
2145 nodemask_t *nodemask, struct zone *preferred_zone,
2148 const gfp_t wait = gfp_mask & __GFP_WAIT;
2149 struct page *page = NULL;
2151 unsigned long pages_reclaimed = 0;
2152 unsigned long did_some_progress;
2153 bool sync_migration = false;
2156 * In the slowpath, we sanity check order to avoid ever trying to
2157 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2158 * be using allocators in order of preference for an area that is
2161 if (order >= MAX_ORDER) {
2162 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2167 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2168 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2169 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2170 * using a larger set of nodes after it has established that the
2171 * allowed per node queues are empty and that nodes are
2174 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2178 if (!(gfp_mask & __GFP_NO_KSWAPD))
2179 wake_all_kswapd(order, zonelist, high_zoneidx,
2180 zone_idx(preferred_zone));
2183 * OK, we're below the kswapd watermark and have kicked background
2184 * reclaim. Now things get more complex, so set up alloc_flags according
2185 * to how we want to proceed.
2187 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2190 * Find the true preferred zone if the allocation is unconstrained by
2193 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2194 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2198 /* This is the last chance, in general, before the goto nopage. */
2199 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2200 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2201 preferred_zone, migratetype);
2205 /* Allocate without watermarks if the context allows */
2206 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2207 page = __alloc_pages_high_priority(gfp_mask, order,
2208 zonelist, high_zoneidx, nodemask,
2209 preferred_zone, migratetype);
2214 /* Atomic allocations - we can't balance anything */
2218 /* Avoid recursion of direct reclaim */
2219 if (current->flags & PF_MEMALLOC)
2222 /* Avoid allocations with no watermarks from looping endlessly */
2223 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2227 * Try direct compaction. The first pass is asynchronous. Subsequent
2228 * attempts after direct reclaim are synchronous
2230 page = __alloc_pages_direct_compact(gfp_mask, order,
2231 zonelist, high_zoneidx,
2233 alloc_flags, preferred_zone,
2234 migratetype, &did_some_progress,
2238 sync_migration = true;
2240 /* Try direct reclaim and then allocating */
2241 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2242 zonelist, high_zoneidx,
2244 alloc_flags, preferred_zone,
2245 migratetype, &did_some_progress);
2250 * If we failed to make any progress reclaiming, then we are
2251 * running out of options and have to consider going OOM
2253 if (!did_some_progress) {
2254 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2255 if (oom_killer_disabled)
2257 page = __alloc_pages_may_oom(gfp_mask, order,
2258 zonelist, high_zoneidx,
2259 nodemask, preferred_zone,
2264 if (!(gfp_mask & __GFP_NOFAIL)) {
2266 * The oom killer is not called for high-order
2267 * allocations that may fail, so if no progress
2268 * is being made, there are no other options and
2269 * retrying is unlikely to help.
2271 if (order > PAGE_ALLOC_COSTLY_ORDER)
2274 * The oom killer is not called for lowmem
2275 * allocations to prevent needlessly killing
2278 if (high_zoneidx < ZONE_NORMAL)
2286 /* Check if we should retry the allocation */
2287 pages_reclaimed += did_some_progress;
2288 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2290 /* Wait for some write requests to complete then retry */
2291 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2295 * High-order allocations do not necessarily loop after
2296 * direct reclaim and reclaim/compaction depends on compaction
2297 * being called after reclaim so call directly if necessary
2299 page = __alloc_pages_direct_compact(gfp_mask, order,
2300 zonelist, high_zoneidx,
2302 alloc_flags, preferred_zone,
2303 migratetype, &did_some_progress,
2310 warn_alloc_failed(gfp_mask, order, NULL);
2313 if (kmemcheck_enabled)
2314 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2320 * This is the 'heart' of the zoned buddy allocator.
2323 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2324 struct zonelist *zonelist, nodemask_t *nodemask)
2326 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2327 struct zone *preferred_zone;
2329 int migratetype = allocflags_to_migratetype(gfp_mask);
2331 gfp_mask &= gfp_allowed_mask;
2333 lockdep_trace_alloc(gfp_mask);
2335 might_sleep_if(gfp_mask & __GFP_WAIT);
2337 if (should_fail_alloc_page(gfp_mask, order))
2341 * Check the zones suitable for the gfp_mask contain at least one
2342 * valid zone. It's possible to have an empty zonelist as a result
2343 * of GFP_THISNODE and a memoryless node
2345 if (unlikely(!zonelist->_zonerefs->zone))
2349 /* The preferred zone is used for statistics later */
2350 first_zones_zonelist(zonelist, high_zoneidx,
2351 nodemask ? : &cpuset_current_mems_allowed,
2353 if (!preferred_zone) {
2358 /* First allocation attempt */
2359 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2360 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2361 preferred_zone, migratetype);
2362 if (unlikely(!page))
2363 page = __alloc_pages_slowpath(gfp_mask, order,
2364 zonelist, high_zoneidx, nodemask,
2365 preferred_zone, migratetype);
2368 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2371 EXPORT_SYMBOL(__alloc_pages_nodemask);
2374 * Common helper functions.
2376 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2381 * __get_free_pages() returns a 32-bit address, which cannot represent
2384 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2386 page = alloc_pages(gfp_mask, order);
2389 return (unsigned long) page_address(page);
2391 EXPORT_SYMBOL(__get_free_pages);
2393 unsigned long get_zeroed_page(gfp_t gfp_mask)
2395 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2397 EXPORT_SYMBOL(get_zeroed_page);
2399 void __free_pages(struct page *page, unsigned int order)
2401 if (put_page_testzero(page)) {
2403 free_hot_cold_page(page, 0);
2405 __free_pages_ok(page, order);
2409 EXPORT_SYMBOL(__free_pages);
2411 void free_pages(unsigned long addr, unsigned int order)
2414 VM_BUG_ON(!virt_addr_valid((void *)addr));
2415 __free_pages(virt_to_page((void *)addr), order);
2419 EXPORT_SYMBOL(free_pages);
2421 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2424 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2425 unsigned long used = addr + PAGE_ALIGN(size);
2427 split_page(virt_to_page((void *)addr), order);
2428 while (used < alloc_end) {
2433 return (void *)addr;
2437 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2438 * @size: the number of bytes to allocate
2439 * @gfp_mask: GFP flags for the allocation
2441 * This function is similar to alloc_pages(), except that it allocates the
2442 * minimum number of pages to satisfy the request. alloc_pages() can only
2443 * allocate memory in power-of-two pages.
2445 * This function is also limited by MAX_ORDER.
2447 * Memory allocated by this function must be released by free_pages_exact().
2449 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2451 unsigned int order = get_order(size);
2454 addr = __get_free_pages(gfp_mask, order);
2455 return make_alloc_exact(addr, order, size);
2457 EXPORT_SYMBOL(alloc_pages_exact);
2460 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2462 * @nid: the preferred node ID where memory should be allocated
2463 * @size: the number of bytes to allocate
2464 * @gfp_mask: GFP flags for the allocation
2466 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2468 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2471 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2473 unsigned order = get_order(size);
2474 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2477 return make_alloc_exact((unsigned long)page_address(p), order, size);
2479 EXPORT_SYMBOL(alloc_pages_exact_nid);
2482 * free_pages_exact - release memory allocated via alloc_pages_exact()
2483 * @virt: the value returned by alloc_pages_exact.
2484 * @size: size of allocation, same value as passed to alloc_pages_exact().
2486 * Release the memory allocated by a previous call to alloc_pages_exact.
2488 void free_pages_exact(void *virt, size_t size)
2490 unsigned long addr = (unsigned long)virt;
2491 unsigned long end = addr + PAGE_ALIGN(size);
2493 while (addr < end) {
2498 EXPORT_SYMBOL(free_pages_exact);
2500 static unsigned int nr_free_zone_pages(int offset)
2505 /* Just pick one node, since fallback list is circular */
2506 unsigned int sum = 0;
2508 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2510 for_each_zone_zonelist(zone, z, zonelist, offset) {
2511 unsigned long size = zone->present_pages;
2512 unsigned long high = high_wmark_pages(zone);
2521 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2523 unsigned int nr_free_buffer_pages(void)
2525 return nr_free_zone_pages(gfp_zone(GFP_USER));
2527 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2530 * Amount of free RAM allocatable within all zones
2532 unsigned int nr_free_pagecache_pages(void)
2534 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2537 static inline void show_node(struct zone *zone)
2540 printk("Node %d ", zone_to_nid(zone));
2543 void si_meminfo(struct sysinfo *val)
2545 val->totalram = totalram_pages;
2547 val->freeram = global_page_state(NR_FREE_PAGES);
2548 val->bufferram = nr_blockdev_pages();
2549 val->totalhigh = totalhigh_pages;
2550 val->freehigh = nr_free_highpages();
2551 val->mem_unit = PAGE_SIZE;
2554 EXPORT_SYMBOL(si_meminfo);
2557 void si_meminfo_node(struct sysinfo *val, int nid)
2559 pg_data_t *pgdat = NODE_DATA(nid);
2561 val->totalram = pgdat->node_present_pages;
2562 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2563 #ifdef CONFIG_HIGHMEM
2564 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2565 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2571 val->mem_unit = PAGE_SIZE;
2576 * Determine whether the node should be displayed or not, depending on whether
2577 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2579 bool skip_free_areas_node(unsigned int flags, int nid)
2583 if (!(flags & SHOW_MEM_FILTER_NODES))
2587 ret = !node_isset(nid, cpuset_current_mems_allowed);
2593 #define K(x) ((x) << (PAGE_SHIFT-10))
2596 * Show free area list (used inside shift_scroll-lock stuff)
2597 * We also calculate the percentage fragmentation. We do this by counting the
2598 * memory on each free list with the exception of the first item on the list.
2599 * Suppresses nodes that are not allowed by current's cpuset if
2600 * SHOW_MEM_FILTER_NODES is passed.
2602 void show_free_areas(unsigned int filter)
2607 for_each_populated_zone(zone) {
2608 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2611 printk("%s per-cpu:\n", zone->name);
2613 for_each_online_cpu(cpu) {
2614 struct per_cpu_pageset *pageset;
2616 pageset = per_cpu_ptr(zone->pageset, cpu);
2618 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2619 cpu, pageset->pcp.high,
2620 pageset->pcp.batch, pageset->pcp.count);
2624 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2625 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2627 " dirty:%lu writeback:%lu unstable:%lu\n"
2628 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2629 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2630 global_page_state(NR_ACTIVE_ANON),
2631 global_page_state(NR_INACTIVE_ANON),
2632 global_page_state(NR_ISOLATED_ANON),
2633 global_page_state(NR_ACTIVE_FILE),
2634 global_page_state(NR_INACTIVE_FILE),
2635 global_page_state(NR_ISOLATED_FILE),
2636 global_page_state(NR_UNEVICTABLE),
2637 global_page_state(NR_FILE_DIRTY),
2638 global_page_state(NR_WRITEBACK),
2639 global_page_state(NR_UNSTABLE_NFS),
2640 global_page_state(NR_FREE_PAGES),
2641 global_page_state(NR_SLAB_RECLAIMABLE),
2642 global_page_state(NR_SLAB_UNRECLAIMABLE),
2643 global_page_state(NR_FILE_MAPPED),
2644 global_page_state(NR_SHMEM),
2645 global_page_state(NR_PAGETABLE),
2646 global_page_state(NR_BOUNCE));
2648 for_each_populated_zone(zone) {
2651 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2659 " active_anon:%lukB"
2660 " inactive_anon:%lukB"
2661 " active_file:%lukB"
2662 " inactive_file:%lukB"
2663 " unevictable:%lukB"
2664 " isolated(anon):%lukB"
2665 " isolated(file):%lukB"
2672 " slab_reclaimable:%lukB"
2673 " slab_unreclaimable:%lukB"
2674 " kernel_stack:%lukB"
2678 " writeback_tmp:%lukB"
2679 " pages_scanned:%lu"
2680 " all_unreclaimable? %s"
2683 K(zone_page_state(zone, NR_FREE_PAGES)),
2684 K(min_wmark_pages(zone)),
2685 K(low_wmark_pages(zone)),
2686 K(high_wmark_pages(zone)),
2687 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2688 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2689 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2690 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2691 K(zone_page_state(zone, NR_UNEVICTABLE)),
2692 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2693 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2694 K(zone->present_pages),
2695 K(zone_page_state(zone, NR_MLOCK)),
2696 K(zone_page_state(zone, NR_FILE_DIRTY)),
2697 K(zone_page_state(zone, NR_WRITEBACK)),
2698 K(zone_page_state(zone, NR_FILE_MAPPED)),
2699 K(zone_page_state(zone, NR_SHMEM)),
2700 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2701 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2702 zone_page_state(zone, NR_KERNEL_STACK) *
2704 K(zone_page_state(zone, NR_PAGETABLE)),
2705 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2706 K(zone_page_state(zone, NR_BOUNCE)),
2707 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2708 zone->pages_scanned,
2709 (zone->all_unreclaimable ? "yes" : "no")
2711 printk("lowmem_reserve[]:");
2712 for (i = 0; i < MAX_NR_ZONES; i++)
2713 printk(" %lu", zone->lowmem_reserve[i]);
2717 for_each_populated_zone(zone) {
2718 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2720 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2723 printk("%s: ", zone->name);
2725 spin_lock_irqsave(&zone->lock, flags);
2726 for (order = 0; order < MAX_ORDER; order++) {
2727 nr[order] = zone->free_area[order].nr_free;
2728 total += nr[order] << order;
2730 spin_unlock_irqrestore(&zone->lock, flags);
2731 for (order = 0; order < MAX_ORDER; order++)
2732 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2733 printk("= %lukB\n", K(total));
2736 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2738 show_swap_cache_info();
2741 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2743 zoneref->zone = zone;
2744 zoneref->zone_idx = zone_idx(zone);
2748 * Builds allocation fallback zone lists.
2750 * Add all populated zones of a node to the zonelist.
2752 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2753 int nr_zones, enum zone_type zone_type)
2757 BUG_ON(zone_type >= MAX_NR_ZONES);
2762 zone = pgdat->node_zones + zone_type;
2763 if (populated_zone(zone)) {
2764 zoneref_set_zone(zone,
2765 &zonelist->_zonerefs[nr_zones++]);
2766 check_highest_zone(zone_type);
2769 } while (zone_type);
2776 * 0 = automatic detection of better ordering.
2777 * 1 = order by ([node] distance, -zonetype)
2778 * 2 = order by (-zonetype, [node] distance)
2780 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2781 * the same zonelist. So only NUMA can configure this param.
2783 #define ZONELIST_ORDER_DEFAULT 0
2784 #define ZONELIST_ORDER_NODE 1
2785 #define ZONELIST_ORDER_ZONE 2
2787 /* zonelist order in the kernel.
2788 * set_zonelist_order() will set this to NODE or ZONE.
2790 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2791 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2795 /* The value user specified ....changed by config */
2796 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2797 /* string for sysctl */
2798 #define NUMA_ZONELIST_ORDER_LEN 16
2799 char numa_zonelist_order[16] = "default";
2802 * interface for configure zonelist ordering.
2803 * command line option "numa_zonelist_order"
2804 * = "[dD]efault - default, automatic configuration.
2805 * = "[nN]ode - order by node locality, then by zone within node
2806 * = "[zZ]one - order by zone, then by locality within zone
2809 static int __parse_numa_zonelist_order(char *s)
2811 if (*s == 'd' || *s == 'D') {
2812 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2813 } else if (*s == 'n' || *s == 'N') {
2814 user_zonelist_order = ZONELIST_ORDER_NODE;
2815 } else if (*s == 'z' || *s == 'Z') {
2816 user_zonelist_order = ZONELIST_ORDER_ZONE;
2819 "Ignoring invalid numa_zonelist_order value: "
2826 static __init int setup_numa_zonelist_order(char *s)
2833 ret = __parse_numa_zonelist_order(s);
2835 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2839 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2842 * sysctl handler for numa_zonelist_order
2844 int numa_zonelist_order_handler(ctl_table *table, int write,
2845 void __user *buffer, size_t *length,
2848 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2850 static DEFINE_MUTEX(zl_order_mutex);
2852 mutex_lock(&zl_order_mutex);
2854 strcpy(saved_string, (char*)table->data);
2855 ret = proc_dostring(table, write, buffer, length, ppos);
2859 int oldval = user_zonelist_order;
2860 if (__parse_numa_zonelist_order((char*)table->data)) {
2862 * bogus value. restore saved string
2864 strncpy((char*)table->data, saved_string,
2865 NUMA_ZONELIST_ORDER_LEN);
2866 user_zonelist_order = oldval;
2867 } else if (oldval != user_zonelist_order) {
2868 mutex_lock(&zonelists_mutex);
2869 build_all_zonelists(NULL);
2870 mutex_unlock(&zonelists_mutex);
2874 mutex_unlock(&zl_order_mutex);
2879 #define MAX_NODE_LOAD (nr_online_nodes)
2880 static int node_load[MAX_NUMNODES];
2883 * find_next_best_node - find the next node that should appear in a given node's fallback list
2884 * @node: node whose fallback list we're appending
2885 * @used_node_mask: nodemask_t of already used nodes
2887 * We use a number of factors to determine which is the next node that should
2888 * appear on a given node's fallback list. The node should not have appeared
2889 * already in @node's fallback list, and it should be the next closest node
2890 * according to the distance array (which contains arbitrary distance values
2891 * from each node to each node in the system), and should also prefer nodes
2892 * with no CPUs, since presumably they'll have very little allocation pressure
2893 * on them otherwise.
2894 * It returns -1 if no node is found.
2896 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2899 int min_val = INT_MAX;
2901 const struct cpumask *tmp = cpumask_of_node(0);
2903 /* Use the local node if we haven't already */
2904 if (!node_isset(node, *used_node_mask)) {
2905 node_set(node, *used_node_mask);
2909 for_each_node_state(n, N_HIGH_MEMORY) {
2911 /* Don't want a node to appear more than once */
2912 if (node_isset(n, *used_node_mask))
2915 /* Use the distance array to find the distance */
2916 val = node_distance(node, n);
2918 /* Penalize nodes under us ("prefer the next node") */
2921 /* Give preference to headless and unused nodes */
2922 tmp = cpumask_of_node(n);
2923 if (!cpumask_empty(tmp))
2924 val += PENALTY_FOR_NODE_WITH_CPUS;
2926 /* Slight preference for less loaded node */
2927 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2928 val += node_load[n];
2930 if (val < min_val) {
2937 node_set(best_node, *used_node_mask);
2944 * Build zonelists ordered by node and zones within node.
2945 * This results in maximum locality--normal zone overflows into local
2946 * DMA zone, if any--but risks exhausting DMA zone.
2948 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2951 struct zonelist *zonelist;
2953 zonelist = &pgdat->node_zonelists[0];
2954 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2956 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2958 zonelist->_zonerefs[j].zone = NULL;
2959 zonelist->_zonerefs[j].zone_idx = 0;
2963 * Build gfp_thisnode zonelists
2965 static void build_thisnode_zonelists(pg_data_t *pgdat)
2968 struct zonelist *zonelist;
2970 zonelist = &pgdat->node_zonelists[1];
2971 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2972 zonelist->_zonerefs[j].zone = NULL;
2973 zonelist->_zonerefs[j].zone_idx = 0;
2977 * Build zonelists ordered by zone and nodes within zones.
2978 * This results in conserving DMA zone[s] until all Normal memory is
2979 * exhausted, but results in overflowing to remote node while memory
2980 * may still exist in local DMA zone.
2982 static int node_order[MAX_NUMNODES];
2984 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2987 int zone_type; /* needs to be signed */
2989 struct zonelist *zonelist;
2991 zonelist = &pgdat->node_zonelists[0];
2993 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2994 for (j = 0; j < nr_nodes; j++) {
2995 node = node_order[j];
2996 z = &NODE_DATA(node)->node_zones[zone_type];
2997 if (populated_zone(z)) {
2999 &zonelist->_zonerefs[pos++]);
3000 check_highest_zone(zone_type);
3004 zonelist->_zonerefs[pos].zone = NULL;
3005 zonelist->_zonerefs[pos].zone_idx = 0;
3008 static int default_zonelist_order(void)
3011 unsigned long low_kmem_size,total_size;
3015 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3016 * If they are really small and used heavily, the system can fall
3017 * into OOM very easily.
3018 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3020 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3023 for_each_online_node(nid) {
3024 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3025 z = &NODE_DATA(nid)->node_zones[zone_type];
3026 if (populated_zone(z)) {
3027 if (zone_type < ZONE_NORMAL)
3028 low_kmem_size += z->present_pages;
3029 total_size += z->present_pages;
3030 } else if (zone_type == ZONE_NORMAL) {
3032 * If any node has only lowmem, then node order
3033 * is preferred to allow kernel allocations
3034 * locally; otherwise, they can easily infringe
3035 * on other nodes when there is an abundance of
3036 * lowmem available to allocate from.
3038 return ZONELIST_ORDER_NODE;
3042 if (!low_kmem_size || /* there are no DMA area. */
3043 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3044 return ZONELIST_ORDER_NODE;
3046 * look into each node's config.
3047 * If there is a node whose DMA/DMA32 memory is very big area on
3048 * local memory, NODE_ORDER may be suitable.
3050 average_size = total_size /
3051 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3052 for_each_online_node(nid) {
3055 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3056 z = &NODE_DATA(nid)->node_zones[zone_type];
3057 if (populated_zone(z)) {
3058 if (zone_type < ZONE_NORMAL)
3059 low_kmem_size += z->present_pages;
3060 total_size += z->present_pages;
3063 if (low_kmem_size &&
3064 total_size > average_size && /* ignore small node */
3065 low_kmem_size > total_size * 70/100)
3066 return ZONELIST_ORDER_NODE;
3068 return ZONELIST_ORDER_ZONE;
3071 static void set_zonelist_order(void)
3073 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3074 current_zonelist_order = default_zonelist_order();
3076 current_zonelist_order = user_zonelist_order;
3079 static void build_zonelists(pg_data_t *pgdat)
3083 nodemask_t used_mask;
3084 int local_node, prev_node;
3085 struct zonelist *zonelist;
3086 int order = current_zonelist_order;
3088 /* initialize zonelists */
3089 for (i = 0; i < MAX_ZONELISTS; i++) {
3090 zonelist = pgdat->node_zonelists + i;
3091 zonelist->_zonerefs[0].zone = NULL;
3092 zonelist->_zonerefs[0].zone_idx = 0;
3095 /* NUMA-aware ordering of nodes */
3096 local_node = pgdat->node_id;
3097 load = nr_online_nodes;
3098 prev_node = local_node;
3099 nodes_clear(used_mask);
3101 memset(node_order, 0, sizeof(node_order));
3104 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3105 int distance = node_distance(local_node, node);
3108 * If another node is sufficiently far away then it is better
3109 * to reclaim pages in a zone before going off node.
3111 if (distance > RECLAIM_DISTANCE)
3112 zone_reclaim_mode = 1;
3115 * We don't want to pressure a particular node.
3116 * So adding penalty to the first node in same
3117 * distance group to make it round-robin.
3119 if (distance != node_distance(local_node, prev_node))
3120 node_load[node] = load;
3124 if (order == ZONELIST_ORDER_NODE)
3125 build_zonelists_in_node_order(pgdat, node);
3127 node_order[j++] = node; /* remember order */
3130 if (order == ZONELIST_ORDER_ZONE) {
3131 /* calculate node order -- i.e., DMA last! */
3132 build_zonelists_in_zone_order(pgdat, j);
3135 build_thisnode_zonelists(pgdat);
3138 /* Construct the zonelist performance cache - see further mmzone.h */
3139 static void build_zonelist_cache(pg_data_t *pgdat)
3141 struct zonelist *zonelist;
3142 struct zonelist_cache *zlc;
3145 zonelist = &pgdat->node_zonelists[0];
3146 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3147 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3148 for (z = zonelist->_zonerefs; z->zone; z++)
3149 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3152 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3154 * Return node id of node used for "local" allocations.
3155 * I.e., first node id of first zone in arg node's generic zonelist.
3156 * Used for initializing percpu 'numa_mem', which is used primarily
3157 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3159 int local_memory_node(int node)
3163 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3164 gfp_zone(GFP_KERNEL),
3171 #else /* CONFIG_NUMA */
3173 static void set_zonelist_order(void)
3175 current_zonelist_order = ZONELIST_ORDER_ZONE;
3178 static void build_zonelists(pg_data_t *pgdat)
3180 int node, local_node;
3182 struct zonelist *zonelist;
3184 local_node = pgdat->node_id;
3186 zonelist = &pgdat->node_zonelists[0];
3187 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3190 * Now we build the zonelist so that it contains the zones
3191 * of all the other nodes.
3192 * We don't want to pressure a particular node, so when
3193 * building the zones for node N, we make sure that the
3194 * zones coming right after the local ones are those from
3195 * node N+1 (modulo N)
3197 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3198 if (!node_online(node))
3200 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3203 for (node = 0; node < local_node; node++) {
3204 if (!node_online(node))
3206 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3210 zonelist->_zonerefs[j].zone = NULL;
3211 zonelist->_zonerefs[j].zone_idx = 0;
3214 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3215 static void build_zonelist_cache(pg_data_t *pgdat)
3217 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3220 #endif /* CONFIG_NUMA */
3223 * Boot pageset table. One per cpu which is going to be used for all
3224 * zones and all nodes. The parameters will be set in such a way
3225 * that an item put on a list will immediately be handed over to
3226 * the buddy list. This is safe since pageset manipulation is done
3227 * with interrupts disabled.
3229 * The boot_pagesets must be kept even after bootup is complete for
3230 * unused processors and/or zones. They do play a role for bootstrapping
3231 * hotplugged processors.
3233 * zoneinfo_show() and maybe other functions do
3234 * not check if the processor is online before following the pageset pointer.
3235 * Other parts of the kernel may not check if the zone is available.
3237 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3238 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3239 static void setup_zone_pageset(struct zone *zone);
3242 * Global mutex to protect against size modification of zonelists
3243 * as well as to serialize pageset setup for the new populated zone.
3245 DEFINE_MUTEX(zonelists_mutex);
3247 /* return values int ....just for stop_machine() */
3248 static __init_refok int __build_all_zonelists(void *data)
3254 memset(node_load, 0, sizeof(node_load));
3256 for_each_online_node(nid) {
3257 pg_data_t *pgdat = NODE_DATA(nid);
3259 build_zonelists(pgdat);
3260 build_zonelist_cache(pgdat);
3264 * Initialize the boot_pagesets that are going to be used
3265 * for bootstrapping processors. The real pagesets for
3266 * each zone will be allocated later when the per cpu
3267 * allocator is available.
3269 * boot_pagesets are used also for bootstrapping offline
3270 * cpus if the system is already booted because the pagesets
3271 * are needed to initialize allocators on a specific cpu too.
3272 * F.e. the percpu allocator needs the page allocator which
3273 * needs the percpu allocator in order to allocate its pagesets
3274 * (a chicken-egg dilemma).
3276 for_each_possible_cpu(cpu) {
3277 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3279 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3281 * We now know the "local memory node" for each node--
3282 * i.e., the node of the first zone in the generic zonelist.
3283 * Set up numa_mem percpu variable for on-line cpus. During
3284 * boot, only the boot cpu should be on-line; we'll init the
3285 * secondary cpus' numa_mem as they come on-line. During
3286 * node/memory hotplug, we'll fixup all on-line cpus.
3288 if (cpu_online(cpu))
3289 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3297 * Called with zonelists_mutex held always
3298 * unless system_state == SYSTEM_BOOTING.
3300 void __ref build_all_zonelists(void *data)
3302 set_zonelist_order();
3304 if (system_state == SYSTEM_BOOTING) {
3305 __build_all_zonelists(NULL);
3306 mminit_verify_zonelist();
3307 cpuset_init_current_mems_allowed();
3309 /* we have to stop all cpus to guarantee there is no user
3311 #ifdef CONFIG_MEMORY_HOTPLUG
3313 setup_zone_pageset((struct zone *)data);
3315 stop_machine(__build_all_zonelists, NULL, NULL);
3316 /* cpuset refresh routine should be here */
3318 vm_total_pages = nr_free_pagecache_pages();
3320 * Disable grouping by mobility if the number of pages in the
3321 * system is too low to allow the mechanism to work. It would be
3322 * more accurate, but expensive to check per-zone. This check is
3323 * made on memory-hotadd so a system can start with mobility
3324 * disabled and enable it later
3326 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3327 page_group_by_mobility_disabled = 1;
3329 page_group_by_mobility_disabled = 0;
3331 printk("Built %i zonelists in %s order, mobility grouping %s. "
3332 "Total pages: %ld\n",
3334 zonelist_order_name[current_zonelist_order],
3335 page_group_by_mobility_disabled ? "off" : "on",
3338 printk("Policy zone: %s\n", zone_names[policy_zone]);
3343 * Helper functions to size the waitqueue hash table.
3344 * Essentially these want to choose hash table sizes sufficiently
3345 * large so that collisions trying to wait on pages are rare.
3346 * But in fact, the number of active page waitqueues on typical
3347 * systems is ridiculously low, less than 200. So this is even
3348 * conservative, even though it seems large.
3350 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3351 * waitqueues, i.e. the size of the waitq table given the number of pages.
3353 #define PAGES_PER_WAITQUEUE 256
3355 #ifndef CONFIG_MEMORY_HOTPLUG
3356 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3358 unsigned long size = 1;
3360 pages /= PAGES_PER_WAITQUEUE;
3362 while (size < pages)
3366 * Once we have dozens or even hundreds of threads sleeping
3367 * on IO we've got bigger problems than wait queue collision.
3368 * Limit the size of the wait table to a reasonable size.
3370 size = min(size, 4096UL);
3372 return max(size, 4UL);
3376 * A zone's size might be changed by hot-add, so it is not possible to determine
3377 * a suitable size for its wait_table. So we use the maximum size now.
3379 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3381 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3382 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3383 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3385 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3386 * or more by the traditional way. (See above). It equals:
3388 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3389 * ia64(16K page size) : = ( 8G + 4M)byte.
3390 * powerpc (64K page size) : = (32G +16M)byte.
3392 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3399 * This is an integer logarithm so that shifts can be used later
3400 * to extract the more random high bits from the multiplicative
3401 * hash function before the remainder is taken.
3403 static inline unsigned long wait_table_bits(unsigned long size)
3408 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3411 * Check if a pageblock contains reserved pages
3413 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3417 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3418 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3425 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3426 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3427 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3428 * higher will lead to a bigger reserve which will get freed as contiguous
3429 * blocks as reclaim kicks in
3431 static void setup_zone_migrate_reserve(struct zone *zone)
3433 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3435 unsigned long block_migratetype;
3439 * Get the start pfn, end pfn and the number of blocks to reserve
3440 * We have to be careful to be aligned to pageblock_nr_pages to
3441 * make sure that we always check pfn_valid for the first page in
3444 start_pfn = zone->zone_start_pfn;
3445 end_pfn = start_pfn + zone->spanned_pages;
3446 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3447 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3451 * Reserve blocks are generally in place to help high-order atomic
3452 * allocations that are short-lived. A min_free_kbytes value that
3453 * would result in more than 2 reserve blocks for atomic allocations
3454 * is assumed to be in place to help anti-fragmentation for the
3455 * future allocation of hugepages at runtime.
3457 reserve = min(2, reserve);
3459 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3460 if (!pfn_valid(pfn))
3462 page = pfn_to_page(pfn);
3464 /* Watch out for overlapping nodes */
3465 if (page_to_nid(page) != zone_to_nid(zone))
3468 block_migratetype = get_pageblock_migratetype(page);
3470 /* Only test what is necessary when the reserves are not met */
3473 * Blocks with reserved pages will never free, skip
3476 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3477 if (pageblock_is_reserved(pfn, block_end_pfn))
3480 /* If this block is reserved, account for it */
3481 if (block_migratetype == MIGRATE_RESERVE) {
3486 /* Suitable for reserving if this block is movable */
3487 if (block_migratetype == MIGRATE_MOVABLE) {
3488 set_pageblock_migratetype(page,
3490 move_freepages_block(zone, page,
3498 * If the reserve is met and this is a previous reserved block,
3501 if (block_migratetype == MIGRATE_RESERVE) {
3502 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3503 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3509 * Initially all pages are reserved - free ones are freed
3510 * up by free_all_bootmem() once the early boot process is
3511 * done. Non-atomic initialization, single-pass.
3513 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3514 unsigned long start_pfn, enum memmap_context context)
3517 unsigned long end_pfn = start_pfn + size;
3521 if (highest_memmap_pfn < end_pfn - 1)
3522 highest_memmap_pfn = end_pfn - 1;
3524 z = &NODE_DATA(nid)->node_zones[zone];
3525 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3527 * There can be holes in boot-time mem_map[]s
3528 * handed to this function. They do not
3529 * exist on hotplugged memory.
3531 if (context == MEMMAP_EARLY) {
3532 if (!early_pfn_valid(pfn))
3534 if (!early_pfn_in_nid(pfn, nid))
3537 page = pfn_to_page(pfn);
3538 set_page_links(page, zone, nid, pfn);
3539 mminit_verify_page_links(page, zone, nid, pfn);
3540 init_page_count(page);
3541 reset_page_mapcount(page);
3542 SetPageReserved(page);
3544 * Mark the block movable so that blocks are reserved for
3545 * movable at startup. This will force kernel allocations
3546 * to reserve their blocks rather than leaking throughout
3547 * the address space during boot when many long-lived
3548 * kernel allocations are made. Later some blocks near
3549 * the start are marked MIGRATE_RESERVE by
3550 * setup_zone_migrate_reserve()
3552 * bitmap is created for zone's valid pfn range. but memmap
3553 * can be created for invalid pages (for alignment)
3554 * check here not to call set_pageblock_migratetype() against
3557 if ((z->zone_start_pfn <= pfn)
3558 && (pfn < z->zone_start_pfn + z->spanned_pages)
3559 && !(pfn & (pageblock_nr_pages - 1)))
3560 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3562 INIT_LIST_HEAD(&page->lru);
3563 #ifdef WANT_PAGE_VIRTUAL
3564 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3565 if (!is_highmem_idx(zone))
3566 set_page_address(page, __va(pfn << PAGE_SHIFT));
3571 static void __meminit zone_init_free_lists(struct zone *zone)
3574 for_each_migratetype_order(order, t) {
3575 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3576 zone->free_area[order].nr_free = 0;
3580 #ifndef __HAVE_ARCH_MEMMAP_INIT
3581 #define memmap_init(size, nid, zone, start_pfn) \
3582 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3585 static int zone_batchsize(struct zone *zone)
3591 * The per-cpu-pages pools are set to around 1000th of the
3592 * size of the zone. But no more than 1/2 of a meg.
3594 * OK, so we don't know how big the cache is. So guess.
3596 batch = zone->present_pages / 1024;
3597 if (batch * PAGE_SIZE > 512 * 1024)
3598 batch = (512 * 1024) / PAGE_SIZE;
3599 batch /= 4; /* We effectively *= 4 below */
3604 * Clamp the batch to a 2^n - 1 value. Having a power
3605 * of 2 value was found to be more likely to have
3606 * suboptimal cache aliasing properties in some cases.
3608 * For example if 2 tasks are alternately allocating
3609 * batches of pages, one task can end up with a lot
3610 * of pages of one half of the possible page colors
3611 * and the other with pages of the other colors.
3613 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3618 /* The deferral and batching of frees should be suppressed under NOMMU
3621 * The problem is that NOMMU needs to be able to allocate large chunks
3622 * of contiguous memory as there's no hardware page translation to
3623 * assemble apparent contiguous memory from discontiguous pages.
3625 * Queueing large contiguous runs of pages for batching, however,
3626 * causes the pages to actually be freed in smaller chunks. As there
3627 * can be a significant delay between the individual batches being
3628 * recycled, this leads to the once large chunks of space being
3629 * fragmented and becoming unavailable for high-order allocations.
3635 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3637 struct per_cpu_pages *pcp;
3640 memset(p, 0, sizeof(*p));
3644 pcp->high = 6 * batch;
3645 pcp->batch = max(1UL, 1 * batch);
3646 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3647 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3651 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3652 * to the value high for the pageset p.
3655 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3658 struct per_cpu_pages *pcp;
3662 pcp->batch = max(1UL, high/4);
3663 if ((high/4) > (PAGE_SHIFT * 8))
3664 pcp->batch = PAGE_SHIFT * 8;
3667 static void setup_zone_pageset(struct zone *zone)
3671 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3673 for_each_possible_cpu(cpu) {
3674 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3676 setup_pageset(pcp, zone_batchsize(zone));
3678 if (percpu_pagelist_fraction)
3679 setup_pagelist_highmark(pcp,
3680 (zone->present_pages /
3681 percpu_pagelist_fraction));
3686 * Allocate per cpu pagesets and initialize them.
3687 * Before this call only boot pagesets were available.
3689 void __init setup_per_cpu_pageset(void)
3693 for_each_populated_zone(zone)
3694 setup_zone_pageset(zone);
3697 static noinline __init_refok
3698 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3701 struct pglist_data *pgdat = zone->zone_pgdat;
3705 * The per-page waitqueue mechanism uses hashed waitqueues
3708 zone->wait_table_hash_nr_entries =
3709 wait_table_hash_nr_entries(zone_size_pages);
3710 zone->wait_table_bits =
3711 wait_table_bits(zone->wait_table_hash_nr_entries);
3712 alloc_size = zone->wait_table_hash_nr_entries
3713 * sizeof(wait_queue_head_t);
3715 if (!slab_is_available()) {
3716 zone->wait_table = (wait_queue_head_t *)
3717 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3720 * This case means that a zone whose size was 0 gets new memory
3721 * via memory hot-add.
3722 * But it may be the case that a new node was hot-added. In
3723 * this case vmalloc() will not be able to use this new node's
3724 * memory - this wait_table must be initialized to use this new
3725 * node itself as well.
3726 * To use this new node's memory, further consideration will be
3729 zone->wait_table = vmalloc(alloc_size);
3731 if (!zone->wait_table)
3734 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3735 init_waitqueue_head(zone->wait_table + i);
3740 static int __zone_pcp_update(void *data)
3742 struct zone *zone = data;
3744 unsigned long batch = zone_batchsize(zone), flags;
3746 for_each_possible_cpu(cpu) {
3747 struct per_cpu_pageset *pset;
3748 struct per_cpu_pages *pcp;
3750 pset = per_cpu_ptr(zone->pageset, cpu);
3753 local_irq_save(flags);
3754 free_pcppages_bulk(zone, pcp->count, pcp);
3755 setup_pageset(pset, batch);
3756 local_irq_restore(flags);
3761 void zone_pcp_update(struct zone *zone)
3763 stop_machine(__zone_pcp_update, zone, NULL);
3766 static __meminit void zone_pcp_init(struct zone *zone)
3769 * per cpu subsystem is not up at this point. The following code
3770 * relies on the ability of the linker to provide the
3771 * offset of a (static) per cpu variable into the per cpu area.
3773 zone->pageset = &boot_pageset;
3775 if (zone->present_pages)
3776 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3777 zone->name, zone->present_pages,
3778 zone_batchsize(zone));
3781 __meminit int init_currently_empty_zone(struct zone *zone,
3782 unsigned long zone_start_pfn,
3784 enum memmap_context context)
3786 struct pglist_data *pgdat = zone->zone_pgdat;
3788 ret = zone_wait_table_init(zone, size);
3791 pgdat->nr_zones = zone_idx(zone) + 1;
3793 zone->zone_start_pfn = zone_start_pfn;
3795 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3796 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3798 (unsigned long)zone_idx(zone),
3799 zone_start_pfn, (zone_start_pfn + size));
3801 zone_init_free_lists(zone);
3806 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
3807 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3809 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3810 * Architectures may implement their own version but if add_active_range()
3811 * was used and there are no special requirements, this is a convenient
3814 int __meminit __early_pfn_to_nid(unsigned long pfn)
3816 unsigned long start_pfn, end_pfn;
3819 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
3820 if (start_pfn <= pfn && pfn < end_pfn)
3822 /* This is a memory hole */
3825 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3827 int __meminit early_pfn_to_nid(unsigned long pfn)
3831 nid = __early_pfn_to_nid(pfn);
3834 /* just returns 0 */
3838 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3839 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3843 nid = __early_pfn_to_nid(pfn);
3844 if (nid >= 0 && nid != node)
3851 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3852 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3853 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3855 * If an architecture guarantees that all ranges registered with
3856 * add_active_ranges() contain no holes and may be freed, this
3857 * this function may be used instead of calling free_bootmem() manually.
3859 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
3861 unsigned long start_pfn, end_pfn;
3864 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
3865 start_pfn = min(start_pfn, max_low_pfn);
3866 end_pfn = min(end_pfn, max_low_pfn);
3868 if (start_pfn < end_pfn)
3869 free_bootmem_node(NODE_DATA(this_nid),
3870 PFN_PHYS(start_pfn),
3871 (end_pfn - start_pfn) << PAGE_SHIFT);
3875 int __init add_from_early_node_map(struct range *range, int az,
3876 int nr_range, int nid)
3878 unsigned long start_pfn, end_pfn;
3881 /* need to go over early_node_map to find out good range for node */
3882 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL)
3883 nr_range = add_range(range, az, nr_range, start_pfn, end_pfn);
3888 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3889 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3891 * If an architecture guarantees that all ranges registered with
3892 * add_active_ranges() contain no holes and may be freed, this
3893 * function may be used instead of calling memory_present() manually.
3895 void __init sparse_memory_present_with_active_regions(int nid)
3897 unsigned long start_pfn, end_pfn;
3900 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
3901 memory_present(this_nid, start_pfn, end_pfn);
3905 * get_pfn_range_for_nid - Return the start and end page frames for a node
3906 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3907 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3908 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3910 * It returns the start and end page frame of a node based on information
3911 * provided by an arch calling add_active_range(). If called for a node
3912 * with no available memory, a warning is printed and the start and end
3915 void __meminit get_pfn_range_for_nid(unsigned int nid,
3916 unsigned long *start_pfn, unsigned long *end_pfn)
3918 unsigned long this_start_pfn, this_end_pfn;
3924 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
3925 *start_pfn = min(*start_pfn, this_start_pfn);
3926 *end_pfn = max(*end_pfn, this_end_pfn);
3929 if (*start_pfn == -1UL)
3934 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3935 * assumption is made that zones within a node are ordered in monotonic
3936 * increasing memory addresses so that the "highest" populated zone is used
3938 static void __init find_usable_zone_for_movable(void)
3941 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3942 if (zone_index == ZONE_MOVABLE)
3945 if (arch_zone_highest_possible_pfn[zone_index] >
3946 arch_zone_lowest_possible_pfn[zone_index])
3950 VM_BUG_ON(zone_index == -1);
3951 movable_zone = zone_index;
3955 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3956 * because it is sized independent of architecture. Unlike the other zones,
3957 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3958 * in each node depending on the size of each node and how evenly kernelcore
3959 * is distributed. This helper function adjusts the zone ranges
3960 * provided by the architecture for a given node by using the end of the
3961 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3962 * zones within a node are in order of monotonic increases memory addresses
3964 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3965 unsigned long zone_type,
3966 unsigned long node_start_pfn,
3967 unsigned long node_end_pfn,
3968 unsigned long *zone_start_pfn,
3969 unsigned long *zone_end_pfn)
3971 /* Only adjust if ZONE_MOVABLE is on this node */
3972 if (zone_movable_pfn[nid]) {
3973 /* Size ZONE_MOVABLE */
3974 if (zone_type == ZONE_MOVABLE) {
3975 *zone_start_pfn = zone_movable_pfn[nid];
3976 *zone_end_pfn = min(node_end_pfn,
3977 arch_zone_highest_possible_pfn[movable_zone]);
3979 /* Adjust for ZONE_MOVABLE starting within this range */
3980 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3981 *zone_end_pfn > zone_movable_pfn[nid]) {
3982 *zone_end_pfn = zone_movable_pfn[nid];
3984 /* Check if this whole range is within ZONE_MOVABLE */
3985 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3986 *zone_start_pfn = *zone_end_pfn;
3991 * Return the number of pages a zone spans in a node, including holes
3992 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3994 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3995 unsigned long zone_type,
3996 unsigned long *ignored)
3998 unsigned long node_start_pfn, node_end_pfn;
3999 unsigned long zone_start_pfn, zone_end_pfn;
4001 /* Get the start and end of the node and zone */
4002 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4003 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4004 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4005 adjust_zone_range_for_zone_movable(nid, zone_type,
4006 node_start_pfn, node_end_pfn,
4007 &zone_start_pfn, &zone_end_pfn);
4009 /* Check that this node has pages within the zone's required range */
4010 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4013 /* Move the zone boundaries inside the node if necessary */
4014 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4015 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4017 /* Return the spanned pages */
4018 return zone_end_pfn - zone_start_pfn;
4022 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4023 * then all holes in the requested range will be accounted for.
4025 unsigned long __meminit __absent_pages_in_range(int nid,
4026 unsigned long range_start_pfn,
4027 unsigned long range_end_pfn)
4029 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4030 unsigned long start_pfn, end_pfn;
4033 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4034 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4035 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4036 nr_absent -= end_pfn - start_pfn;
4042 * absent_pages_in_range - Return number of page frames in holes within a range
4043 * @start_pfn: The start PFN to start searching for holes
4044 * @end_pfn: The end PFN to stop searching for holes
4046 * It returns the number of pages frames in memory holes within a range.
4048 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4049 unsigned long end_pfn)
4051 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4054 /* Return the number of page frames in holes in a zone on a node */
4055 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4056 unsigned long zone_type,
4057 unsigned long *ignored)
4059 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4060 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4061 unsigned long node_start_pfn, node_end_pfn;
4062 unsigned long zone_start_pfn, zone_end_pfn;
4064 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4065 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4066 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4068 adjust_zone_range_for_zone_movable(nid, zone_type,
4069 node_start_pfn, node_end_pfn,
4070 &zone_start_pfn, &zone_end_pfn);
4071 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4074 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4075 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4076 unsigned long zone_type,
4077 unsigned long *zones_size)
4079 return zones_size[zone_type];
4082 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4083 unsigned long zone_type,
4084 unsigned long *zholes_size)
4089 return zholes_size[zone_type];
4092 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4094 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4095 unsigned long *zones_size, unsigned long *zholes_size)
4097 unsigned long realtotalpages, totalpages = 0;
4100 for (i = 0; i < MAX_NR_ZONES; i++)
4101 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4103 pgdat->node_spanned_pages = totalpages;
4105 realtotalpages = totalpages;
4106 for (i = 0; i < MAX_NR_ZONES; i++)
4108 zone_absent_pages_in_node(pgdat->node_id, i,
4110 pgdat->node_present_pages = realtotalpages;
4111 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4115 #ifndef CONFIG_SPARSEMEM
4117 * Calculate the size of the zone->blockflags rounded to an unsigned long
4118 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4119 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4120 * round what is now in bits to nearest long in bits, then return it in
4123 static unsigned long __init usemap_size(unsigned long zonesize)
4125 unsigned long usemapsize;
4127 usemapsize = roundup(zonesize, pageblock_nr_pages);
4128 usemapsize = usemapsize >> pageblock_order;
4129 usemapsize *= NR_PAGEBLOCK_BITS;
4130 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4132 return usemapsize / 8;
4135 static void __init setup_usemap(struct pglist_data *pgdat,
4136 struct zone *zone, unsigned long zonesize)
4138 unsigned long usemapsize = usemap_size(zonesize);
4139 zone->pageblock_flags = NULL;
4141 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4145 static inline void setup_usemap(struct pglist_data *pgdat,
4146 struct zone *zone, unsigned long zonesize) {}
4147 #endif /* CONFIG_SPARSEMEM */
4149 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4151 /* Return a sensible default order for the pageblock size. */
4152 static inline int pageblock_default_order(void)
4154 if (HPAGE_SHIFT > PAGE_SHIFT)
4155 return HUGETLB_PAGE_ORDER;
4160 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4161 static inline void __init set_pageblock_order(unsigned int order)
4163 /* Check that pageblock_nr_pages has not already been setup */
4164 if (pageblock_order)
4168 * Assume the largest contiguous order of interest is a huge page.
4169 * This value may be variable depending on boot parameters on IA64
4171 pageblock_order = order;
4173 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4176 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4177 * and pageblock_default_order() are unused as pageblock_order is set
4178 * at compile-time. See include/linux/pageblock-flags.h for the values of
4179 * pageblock_order based on the kernel config
4181 static inline int pageblock_default_order(unsigned int order)
4185 #define set_pageblock_order(x) do {} while (0)
4187 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4190 * Set up the zone data structures:
4191 * - mark all pages reserved
4192 * - mark all memory queues empty
4193 * - clear the memory bitmaps
4195 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4196 unsigned long *zones_size, unsigned long *zholes_size)
4199 int nid = pgdat->node_id;
4200 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4203 pgdat_resize_init(pgdat);
4204 pgdat->nr_zones = 0;
4205 init_waitqueue_head(&pgdat->kswapd_wait);
4206 pgdat->kswapd_max_order = 0;
4207 pgdat_page_cgroup_init(pgdat);
4209 for (j = 0; j < MAX_NR_ZONES; j++) {
4210 struct zone *zone = pgdat->node_zones + j;
4211 unsigned long size, realsize, memmap_pages;
4214 size = zone_spanned_pages_in_node(nid, j, zones_size);
4215 realsize = size - zone_absent_pages_in_node(nid, j,
4219 * Adjust realsize so that it accounts for how much memory
4220 * is used by this zone for memmap. This affects the watermark
4221 * and per-cpu initialisations
4224 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4225 if (realsize >= memmap_pages) {
4226 realsize -= memmap_pages;
4229 " %s zone: %lu pages used for memmap\n",
4230 zone_names[j], memmap_pages);
4233 " %s zone: %lu pages exceeds realsize %lu\n",
4234 zone_names[j], memmap_pages, realsize);
4236 /* Account for reserved pages */
4237 if (j == 0 && realsize > dma_reserve) {
4238 realsize -= dma_reserve;
4239 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4240 zone_names[0], dma_reserve);
4243 if (!is_highmem_idx(j))
4244 nr_kernel_pages += realsize;
4245 nr_all_pages += realsize;
4247 zone->spanned_pages = size;
4248 zone->present_pages = realsize;
4251 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4253 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4255 zone->name = zone_names[j];
4256 spin_lock_init(&zone->lock);
4257 spin_lock_init(&zone->lru_lock);
4258 zone_seqlock_init(zone);
4259 zone->zone_pgdat = pgdat;
4261 zone_pcp_init(zone);
4263 INIT_LIST_HEAD(&zone->lru[l].list);
4264 zone->reclaim_stat.recent_rotated[0] = 0;
4265 zone->reclaim_stat.recent_rotated[1] = 0;
4266 zone->reclaim_stat.recent_scanned[0] = 0;
4267 zone->reclaim_stat.recent_scanned[1] = 0;
4268 zap_zone_vm_stats(zone);
4273 set_pageblock_order(pageblock_default_order());
4274 setup_usemap(pgdat, zone, size);
4275 ret = init_currently_empty_zone(zone, zone_start_pfn,
4276 size, MEMMAP_EARLY);
4278 memmap_init(size, nid, j, zone_start_pfn);
4279 zone_start_pfn += size;
4283 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4285 /* Skip empty nodes */
4286 if (!pgdat->node_spanned_pages)
4289 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4290 /* ia64 gets its own node_mem_map, before this, without bootmem */
4291 if (!pgdat->node_mem_map) {
4292 unsigned long size, start, end;
4296 * The zone's endpoints aren't required to be MAX_ORDER
4297 * aligned but the node_mem_map endpoints must be in order
4298 * for the buddy allocator to function correctly.
4300 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4301 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4302 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4303 size = (end - start) * sizeof(struct page);
4304 map = alloc_remap(pgdat->node_id, size);
4306 map = alloc_bootmem_node_nopanic(pgdat, size);
4307 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4309 #ifndef CONFIG_NEED_MULTIPLE_NODES
4311 * With no DISCONTIG, the global mem_map is just set as node 0's
4313 if (pgdat == NODE_DATA(0)) {
4314 mem_map = NODE_DATA(0)->node_mem_map;
4315 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4316 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4317 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4318 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4321 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4324 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4325 unsigned long node_start_pfn, unsigned long *zholes_size)
4327 pg_data_t *pgdat = NODE_DATA(nid);
4329 pgdat->node_id = nid;
4330 pgdat->node_start_pfn = node_start_pfn;
4331 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4333 alloc_node_mem_map(pgdat);
4334 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4335 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4336 nid, (unsigned long)pgdat,
4337 (unsigned long)pgdat->node_mem_map);
4340 free_area_init_core(pgdat, zones_size, zholes_size);
4343 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4345 #if MAX_NUMNODES > 1
4347 * Figure out the number of possible node ids.
4349 static void __init setup_nr_node_ids(void)
4352 unsigned int highest = 0;
4354 for_each_node_mask(node, node_possible_map)
4356 nr_node_ids = highest + 1;
4359 static inline void setup_nr_node_ids(void)
4365 * node_map_pfn_alignment - determine the maximum internode alignment
4367 * This function should be called after node map is populated and sorted.
4368 * It calculates the maximum power of two alignment which can distinguish
4371 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4372 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4373 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4374 * shifted, 1GiB is enough and this function will indicate so.
4376 * This is used to test whether pfn -> nid mapping of the chosen memory
4377 * model has fine enough granularity to avoid incorrect mapping for the
4378 * populated node map.
4380 * Returns the determined alignment in pfn's. 0 if there is no alignment
4381 * requirement (single node).
4383 unsigned long __init node_map_pfn_alignment(void)
4385 unsigned long accl_mask = 0, last_end = 0;
4386 unsigned long start, end, mask;
4390 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4391 if (!start || last_nid < 0 || last_nid == nid) {
4398 * Start with a mask granular enough to pin-point to the
4399 * start pfn and tick off bits one-by-one until it becomes
4400 * too coarse to separate the current node from the last.
4402 mask = ~((1 << __ffs(start)) - 1);
4403 while (mask && last_end <= (start & (mask << 1)))
4406 /* accumulate all internode masks */
4410 /* convert mask to number of pages */
4411 return ~accl_mask + 1;
4414 /* Find the lowest pfn for a node */
4415 static unsigned long __init find_min_pfn_for_node(int nid)
4417 unsigned long min_pfn = ULONG_MAX;
4418 unsigned long start_pfn;
4421 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4422 min_pfn = min(min_pfn, start_pfn);
4424 if (min_pfn == ULONG_MAX) {
4426 "Could not find start_pfn for node %d\n", nid);
4434 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4436 * It returns the minimum PFN based on information provided via
4437 * add_active_range().
4439 unsigned long __init find_min_pfn_with_active_regions(void)
4441 return find_min_pfn_for_node(MAX_NUMNODES);
4445 * early_calculate_totalpages()
4446 * Sum pages in active regions for movable zone.
4447 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4449 static unsigned long __init early_calculate_totalpages(void)
4451 unsigned long totalpages = 0;
4452 unsigned long start_pfn, end_pfn;
4455 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4456 unsigned long pages = end_pfn - start_pfn;
4458 totalpages += pages;
4460 node_set_state(nid, N_HIGH_MEMORY);
4466 * Find the PFN the Movable zone begins in each node. Kernel memory
4467 * is spread evenly between nodes as long as the nodes have enough
4468 * memory. When they don't, some nodes will have more kernelcore than
4471 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4474 unsigned long usable_startpfn;
4475 unsigned long kernelcore_node, kernelcore_remaining;
4476 /* save the state before borrow the nodemask */
4477 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4478 unsigned long totalpages = early_calculate_totalpages();
4479 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4482 * If movablecore was specified, calculate what size of
4483 * kernelcore that corresponds so that memory usable for
4484 * any allocation type is evenly spread. If both kernelcore
4485 * and movablecore are specified, then the value of kernelcore
4486 * will be used for required_kernelcore if it's greater than
4487 * what movablecore would have allowed.
4489 if (required_movablecore) {
4490 unsigned long corepages;
4493 * Round-up so that ZONE_MOVABLE is at least as large as what
4494 * was requested by the user
4496 required_movablecore =
4497 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4498 corepages = totalpages - required_movablecore;
4500 required_kernelcore = max(required_kernelcore, corepages);
4503 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4504 if (!required_kernelcore)
4507 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4508 find_usable_zone_for_movable();
4509 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4512 /* Spread kernelcore memory as evenly as possible throughout nodes */
4513 kernelcore_node = required_kernelcore / usable_nodes;
4514 for_each_node_state(nid, N_HIGH_MEMORY) {
4515 unsigned long start_pfn, end_pfn;
4518 * Recalculate kernelcore_node if the division per node
4519 * now exceeds what is necessary to satisfy the requested
4520 * amount of memory for the kernel
4522 if (required_kernelcore < kernelcore_node)
4523 kernelcore_node = required_kernelcore / usable_nodes;
4526 * As the map is walked, we track how much memory is usable
4527 * by the kernel using kernelcore_remaining. When it is
4528 * 0, the rest of the node is usable by ZONE_MOVABLE
4530 kernelcore_remaining = kernelcore_node;
4532 /* Go through each range of PFNs within this node */
4533 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4534 unsigned long size_pages;
4536 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4537 if (start_pfn >= end_pfn)
4540 /* Account for what is only usable for kernelcore */
4541 if (start_pfn < usable_startpfn) {
4542 unsigned long kernel_pages;
4543 kernel_pages = min(end_pfn, usable_startpfn)
4546 kernelcore_remaining -= min(kernel_pages,
4547 kernelcore_remaining);
4548 required_kernelcore -= min(kernel_pages,
4549 required_kernelcore);
4551 /* Continue if range is now fully accounted */
4552 if (end_pfn <= usable_startpfn) {
4555 * Push zone_movable_pfn to the end so
4556 * that if we have to rebalance
4557 * kernelcore across nodes, we will
4558 * not double account here
4560 zone_movable_pfn[nid] = end_pfn;
4563 start_pfn = usable_startpfn;
4567 * The usable PFN range for ZONE_MOVABLE is from
4568 * start_pfn->end_pfn. Calculate size_pages as the
4569 * number of pages used as kernelcore
4571 size_pages = end_pfn - start_pfn;
4572 if (size_pages > kernelcore_remaining)
4573 size_pages = kernelcore_remaining;
4574 zone_movable_pfn[nid] = start_pfn + size_pages;
4577 * Some kernelcore has been met, update counts and
4578 * break if the kernelcore for this node has been
4581 required_kernelcore -= min(required_kernelcore,
4583 kernelcore_remaining -= size_pages;
4584 if (!kernelcore_remaining)
4590 * If there is still required_kernelcore, we do another pass with one
4591 * less node in the count. This will push zone_movable_pfn[nid] further
4592 * along on the nodes that still have memory until kernelcore is
4596 if (usable_nodes && required_kernelcore > usable_nodes)
4599 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4600 for (nid = 0; nid < MAX_NUMNODES; nid++)
4601 zone_movable_pfn[nid] =
4602 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4605 /* restore the node_state */
4606 node_states[N_HIGH_MEMORY] = saved_node_state;
4609 /* Any regular memory on that node ? */
4610 static void check_for_regular_memory(pg_data_t *pgdat)
4612 #ifdef CONFIG_HIGHMEM
4613 enum zone_type zone_type;
4615 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4616 struct zone *zone = &pgdat->node_zones[zone_type];
4617 if (zone->present_pages)
4618 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4624 * free_area_init_nodes - Initialise all pg_data_t and zone data
4625 * @max_zone_pfn: an array of max PFNs for each zone
4627 * This will call free_area_init_node() for each active node in the system.
4628 * Using the page ranges provided by add_active_range(), the size of each
4629 * zone in each node and their holes is calculated. If the maximum PFN
4630 * between two adjacent zones match, it is assumed that the zone is empty.
4631 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4632 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4633 * starts where the previous one ended. For example, ZONE_DMA32 starts
4634 * at arch_max_dma_pfn.
4636 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4638 unsigned long start_pfn, end_pfn;
4641 /* Record where the zone boundaries are */
4642 memset(arch_zone_lowest_possible_pfn, 0,
4643 sizeof(arch_zone_lowest_possible_pfn));
4644 memset(arch_zone_highest_possible_pfn, 0,
4645 sizeof(arch_zone_highest_possible_pfn));
4646 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4647 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4648 for (i = 1; i < MAX_NR_ZONES; i++) {
4649 if (i == ZONE_MOVABLE)
4651 arch_zone_lowest_possible_pfn[i] =
4652 arch_zone_highest_possible_pfn[i-1];
4653 arch_zone_highest_possible_pfn[i] =
4654 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4656 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4657 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4659 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4660 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4661 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4663 /* Print out the zone ranges */
4664 printk("Zone PFN ranges:\n");
4665 for (i = 0; i < MAX_NR_ZONES; i++) {
4666 if (i == ZONE_MOVABLE)
4668 printk(" %-8s ", zone_names[i]);
4669 if (arch_zone_lowest_possible_pfn[i] ==
4670 arch_zone_highest_possible_pfn[i])
4673 printk("%0#10lx -> %0#10lx\n",
4674 arch_zone_lowest_possible_pfn[i],
4675 arch_zone_highest_possible_pfn[i]);
4678 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4679 printk("Movable zone start PFN for each node\n");
4680 for (i = 0; i < MAX_NUMNODES; i++) {
4681 if (zone_movable_pfn[i])
4682 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4685 /* Print out the early_node_map[] */
4686 printk("Early memory PFN ranges\n");
4687 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4688 printk(" %3d: %0#10lx -> %0#10lx\n", nid, start_pfn, end_pfn);
4690 /* Initialise every node */
4691 mminit_verify_pageflags_layout();
4692 setup_nr_node_ids();
4693 for_each_online_node(nid) {
4694 pg_data_t *pgdat = NODE_DATA(nid);
4695 free_area_init_node(nid, NULL,
4696 find_min_pfn_for_node(nid), NULL);
4698 /* Any memory on that node */
4699 if (pgdat->node_present_pages)
4700 node_set_state(nid, N_HIGH_MEMORY);
4701 check_for_regular_memory(pgdat);
4705 static int __init cmdline_parse_core(char *p, unsigned long *core)
4707 unsigned long long coremem;
4711 coremem = memparse(p, &p);
4712 *core = coremem >> PAGE_SHIFT;
4714 /* Paranoid check that UL is enough for the coremem value */
4715 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4721 * kernelcore=size sets the amount of memory for use for allocations that
4722 * cannot be reclaimed or migrated.
4724 static int __init cmdline_parse_kernelcore(char *p)
4726 return cmdline_parse_core(p, &required_kernelcore);
4730 * movablecore=size sets the amount of memory for use for allocations that
4731 * can be reclaimed or migrated.
4733 static int __init cmdline_parse_movablecore(char *p)
4735 return cmdline_parse_core(p, &required_movablecore);
4738 early_param("kernelcore", cmdline_parse_kernelcore);
4739 early_param("movablecore", cmdline_parse_movablecore);
4741 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4744 * set_dma_reserve - set the specified number of pages reserved in the first zone
4745 * @new_dma_reserve: The number of pages to mark reserved
4747 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4748 * In the DMA zone, a significant percentage may be consumed by kernel image
4749 * and other unfreeable allocations which can skew the watermarks badly. This
4750 * function may optionally be used to account for unfreeable pages in the
4751 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4752 * smaller per-cpu batchsize.
4754 void __init set_dma_reserve(unsigned long new_dma_reserve)
4756 dma_reserve = new_dma_reserve;
4759 void __init free_area_init(unsigned long *zones_size)
4761 free_area_init_node(0, zones_size,
4762 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4765 static int page_alloc_cpu_notify(struct notifier_block *self,
4766 unsigned long action, void *hcpu)
4768 int cpu = (unsigned long)hcpu;
4770 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4774 * Spill the event counters of the dead processor
4775 * into the current processors event counters.
4776 * This artificially elevates the count of the current
4779 vm_events_fold_cpu(cpu);
4782 * Zero the differential counters of the dead processor
4783 * so that the vm statistics are consistent.
4785 * This is only okay since the processor is dead and cannot
4786 * race with what we are doing.
4788 refresh_cpu_vm_stats(cpu);
4793 void __init page_alloc_init(void)
4795 hotcpu_notifier(page_alloc_cpu_notify, 0);
4799 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4800 * or min_free_kbytes changes.
4802 static void calculate_totalreserve_pages(void)
4804 struct pglist_data *pgdat;
4805 unsigned long reserve_pages = 0;
4806 enum zone_type i, j;
4808 for_each_online_pgdat(pgdat) {
4809 for (i = 0; i < MAX_NR_ZONES; i++) {
4810 struct zone *zone = pgdat->node_zones + i;
4811 unsigned long max = 0;
4813 /* Find valid and maximum lowmem_reserve in the zone */
4814 for (j = i; j < MAX_NR_ZONES; j++) {
4815 if (zone->lowmem_reserve[j] > max)
4816 max = zone->lowmem_reserve[j];
4819 /* we treat the high watermark as reserved pages. */
4820 max += high_wmark_pages(zone);
4822 if (max > zone->present_pages)
4823 max = zone->present_pages;
4824 reserve_pages += max;
4827 totalreserve_pages = reserve_pages;
4831 * setup_per_zone_lowmem_reserve - called whenever
4832 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4833 * has a correct pages reserved value, so an adequate number of
4834 * pages are left in the zone after a successful __alloc_pages().
4836 static void setup_per_zone_lowmem_reserve(void)
4838 struct pglist_data *pgdat;
4839 enum zone_type j, idx;
4841 for_each_online_pgdat(pgdat) {
4842 for (j = 0; j < MAX_NR_ZONES; j++) {
4843 struct zone *zone = pgdat->node_zones + j;
4844 unsigned long present_pages = zone->present_pages;
4846 zone->lowmem_reserve[j] = 0;
4850 struct zone *lower_zone;
4854 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4855 sysctl_lowmem_reserve_ratio[idx] = 1;
4857 lower_zone = pgdat->node_zones + idx;
4858 lower_zone->lowmem_reserve[j] = present_pages /
4859 sysctl_lowmem_reserve_ratio[idx];
4860 present_pages += lower_zone->present_pages;
4865 /* update totalreserve_pages */
4866 calculate_totalreserve_pages();
4870 * setup_per_zone_wmarks - called when min_free_kbytes changes
4871 * or when memory is hot-{added|removed}
4873 * Ensures that the watermark[min,low,high] values for each zone are set
4874 * correctly with respect to min_free_kbytes.
4876 void setup_per_zone_wmarks(void)
4878 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4879 unsigned long lowmem_pages = 0;
4881 unsigned long flags;
4883 /* Calculate total number of !ZONE_HIGHMEM pages */
4884 for_each_zone(zone) {
4885 if (!is_highmem(zone))
4886 lowmem_pages += zone->present_pages;
4889 for_each_zone(zone) {
4892 spin_lock_irqsave(&zone->lock, flags);
4893 tmp = (u64)pages_min * zone->present_pages;
4894 do_div(tmp, lowmem_pages);
4895 if (is_highmem(zone)) {
4897 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4898 * need highmem pages, so cap pages_min to a small
4901 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4902 * deltas controls asynch page reclaim, and so should
4903 * not be capped for highmem.
4907 min_pages = zone->present_pages / 1024;
4908 if (min_pages < SWAP_CLUSTER_MAX)
4909 min_pages = SWAP_CLUSTER_MAX;
4910 if (min_pages > 128)
4912 zone->watermark[WMARK_MIN] = min_pages;
4915 * If it's a lowmem zone, reserve a number of pages
4916 * proportionate to the zone's size.
4918 zone->watermark[WMARK_MIN] = tmp;
4921 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4922 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4923 setup_zone_migrate_reserve(zone);
4924 spin_unlock_irqrestore(&zone->lock, flags);
4927 /* update totalreserve_pages */
4928 calculate_totalreserve_pages();
4932 * The inactive anon list should be small enough that the VM never has to
4933 * do too much work, but large enough that each inactive page has a chance
4934 * to be referenced again before it is swapped out.
4936 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4937 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4938 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4939 * the anonymous pages are kept on the inactive list.
4942 * memory ratio inactive anon
4943 * -------------------------------------
4952 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
4954 unsigned int gb, ratio;
4956 /* Zone size in gigabytes */
4957 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4959 ratio = int_sqrt(10 * gb);
4963 zone->inactive_ratio = ratio;
4966 static void __meminit setup_per_zone_inactive_ratio(void)
4971 calculate_zone_inactive_ratio(zone);
4975 * Initialise min_free_kbytes.
4977 * For small machines we want it small (128k min). For large machines
4978 * we want it large (64MB max). But it is not linear, because network
4979 * bandwidth does not increase linearly with machine size. We use
4981 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4982 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4998 int __meminit init_per_zone_wmark_min(void)
5000 unsigned long lowmem_kbytes;
5002 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5004 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5005 if (min_free_kbytes < 128)
5006 min_free_kbytes = 128;
5007 if (min_free_kbytes > 65536)
5008 min_free_kbytes = 65536;
5009 setup_per_zone_wmarks();
5010 refresh_zone_stat_thresholds();
5011 setup_per_zone_lowmem_reserve();
5012 setup_per_zone_inactive_ratio();
5015 module_init(init_per_zone_wmark_min)
5018 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5019 * that we can call two helper functions whenever min_free_kbytes
5022 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5023 void __user *buffer, size_t *length, loff_t *ppos)
5025 proc_dointvec(table, write, buffer, length, ppos);
5027 setup_per_zone_wmarks();
5032 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5033 void __user *buffer, size_t *length, loff_t *ppos)
5038 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5043 zone->min_unmapped_pages = (zone->present_pages *
5044 sysctl_min_unmapped_ratio) / 100;
5048 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5049 void __user *buffer, size_t *length, loff_t *ppos)
5054 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5059 zone->min_slab_pages = (zone->present_pages *
5060 sysctl_min_slab_ratio) / 100;
5066 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5067 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5068 * whenever sysctl_lowmem_reserve_ratio changes.
5070 * The reserve ratio obviously has absolutely no relation with the
5071 * minimum watermarks. The lowmem reserve ratio can only make sense
5072 * if in function of the boot time zone sizes.
5074 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5075 void __user *buffer, size_t *length, loff_t *ppos)
5077 proc_dointvec_minmax(table, write, buffer, length, ppos);
5078 setup_per_zone_lowmem_reserve();
5083 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5084 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5085 * can have before it gets flushed back to buddy allocator.
5088 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5089 void __user *buffer, size_t *length, loff_t *ppos)
5095 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5096 if (!write || (ret == -EINVAL))
5098 for_each_populated_zone(zone) {
5099 for_each_possible_cpu(cpu) {
5101 high = zone->present_pages / percpu_pagelist_fraction;
5102 setup_pagelist_highmark(
5103 per_cpu_ptr(zone->pageset, cpu), high);
5109 int hashdist = HASHDIST_DEFAULT;
5112 static int __init set_hashdist(char *str)
5116 hashdist = simple_strtoul(str, &str, 0);
5119 __setup("hashdist=", set_hashdist);
5123 * allocate a large system hash table from bootmem
5124 * - it is assumed that the hash table must contain an exact power-of-2
5125 * quantity of entries
5126 * - limit is the number of hash buckets, not the total allocation size
5128 void *__init alloc_large_system_hash(const char *tablename,
5129 unsigned long bucketsize,
5130 unsigned long numentries,
5133 unsigned int *_hash_shift,
5134 unsigned int *_hash_mask,
5135 unsigned long limit)
5137 unsigned long long max = limit;
5138 unsigned long log2qty, size;
5141 /* allow the kernel cmdline to have a say */
5143 /* round applicable memory size up to nearest megabyte */
5144 numentries = nr_kernel_pages;
5145 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5146 numentries >>= 20 - PAGE_SHIFT;
5147 numentries <<= 20 - PAGE_SHIFT;
5149 /* limit to 1 bucket per 2^scale bytes of low memory */
5150 if (scale > PAGE_SHIFT)
5151 numentries >>= (scale - PAGE_SHIFT);
5153 numentries <<= (PAGE_SHIFT - scale);
5155 /* Make sure we've got at least a 0-order allocation.. */
5156 if (unlikely(flags & HASH_SMALL)) {
5157 /* Makes no sense without HASH_EARLY */
5158 WARN_ON(!(flags & HASH_EARLY));
5159 if (!(numentries >> *_hash_shift)) {
5160 numentries = 1UL << *_hash_shift;
5161 BUG_ON(!numentries);
5163 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5164 numentries = PAGE_SIZE / bucketsize;
5166 numentries = roundup_pow_of_two(numentries);
5168 /* limit allocation size to 1/16 total memory by default */
5170 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5171 do_div(max, bucketsize);
5174 if (numentries > max)
5177 log2qty = ilog2(numentries);
5180 size = bucketsize << log2qty;
5181 if (flags & HASH_EARLY)
5182 table = alloc_bootmem_nopanic(size);
5184 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5187 * If bucketsize is not a power-of-two, we may free
5188 * some pages at the end of hash table which
5189 * alloc_pages_exact() automatically does
5191 if (get_order(size) < MAX_ORDER) {
5192 table = alloc_pages_exact(size, GFP_ATOMIC);
5193 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5196 } while (!table && size > PAGE_SIZE && --log2qty);
5199 panic("Failed to allocate %s hash table\n", tablename);
5201 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5204 ilog2(size) - PAGE_SHIFT,
5208 *_hash_shift = log2qty;
5210 *_hash_mask = (1 << log2qty) - 1;
5215 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5216 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5219 #ifdef CONFIG_SPARSEMEM
5220 return __pfn_to_section(pfn)->pageblock_flags;
5222 return zone->pageblock_flags;
5223 #endif /* CONFIG_SPARSEMEM */
5226 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5228 #ifdef CONFIG_SPARSEMEM
5229 pfn &= (PAGES_PER_SECTION-1);
5230 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5232 pfn = pfn - zone->zone_start_pfn;
5233 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5234 #endif /* CONFIG_SPARSEMEM */
5238 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5239 * @page: The page within the block of interest
5240 * @start_bitidx: The first bit of interest to retrieve
5241 * @end_bitidx: The last bit of interest
5242 * returns pageblock_bits flags
5244 unsigned long get_pageblock_flags_group(struct page *page,
5245 int start_bitidx, int end_bitidx)
5248 unsigned long *bitmap;
5249 unsigned long pfn, bitidx;
5250 unsigned long flags = 0;
5251 unsigned long value = 1;
5253 zone = page_zone(page);
5254 pfn = page_to_pfn(page);
5255 bitmap = get_pageblock_bitmap(zone, pfn);
5256 bitidx = pfn_to_bitidx(zone, pfn);
5258 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5259 if (test_bit(bitidx + start_bitidx, bitmap))
5266 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5267 * @page: The page within the block of interest
5268 * @start_bitidx: The first bit of interest
5269 * @end_bitidx: The last bit of interest
5270 * @flags: The flags to set
5272 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5273 int start_bitidx, int end_bitidx)
5276 unsigned long *bitmap;
5277 unsigned long pfn, bitidx;
5278 unsigned long value = 1;
5280 zone = page_zone(page);
5281 pfn = page_to_pfn(page);
5282 bitmap = get_pageblock_bitmap(zone, pfn);
5283 bitidx = pfn_to_bitidx(zone, pfn);
5284 VM_BUG_ON(pfn < zone->zone_start_pfn);
5285 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5287 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5289 __set_bit(bitidx + start_bitidx, bitmap);
5291 __clear_bit(bitidx + start_bitidx, bitmap);
5295 * This is designed as sub function...plz see page_isolation.c also.
5296 * set/clear page block's type to be ISOLATE.
5297 * page allocater never alloc memory from ISOLATE block.
5301 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5303 unsigned long pfn, iter, found;
5305 * For avoiding noise data, lru_add_drain_all() should be called
5306 * If ZONE_MOVABLE, the zone never contains immobile pages
5308 if (zone_idx(zone) == ZONE_MOVABLE)
5311 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5314 pfn = page_to_pfn(page);
5315 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5316 unsigned long check = pfn + iter;
5318 if (!pfn_valid_within(check))
5321 page = pfn_to_page(check);
5322 if (!page_count(page)) {
5323 if (PageBuddy(page))
5324 iter += (1 << page_order(page)) - 1;
5330 * If there are RECLAIMABLE pages, we need to check it.
5331 * But now, memory offline itself doesn't call shrink_slab()
5332 * and it still to be fixed.
5335 * If the page is not RAM, page_count()should be 0.
5336 * we don't need more check. This is an _used_ not-movable page.
5338 * The problematic thing here is PG_reserved pages. PG_reserved
5339 * is set to both of a memory hole page and a _used_ kernel
5348 bool is_pageblock_removable_nolock(struct page *page)
5350 struct zone *zone = page_zone(page);
5351 return __count_immobile_pages(zone, page, 0);
5354 int set_migratetype_isolate(struct page *page)
5357 unsigned long flags, pfn;
5358 struct memory_isolate_notify arg;
5362 zone = page_zone(page);
5364 spin_lock_irqsave(&zone->lock, flags);
5366 pfn = page_to_pfn(page);
5367 arg.start_pfn = pfn;
5368 arg.nr_pages = pageblock_nr_pages;
5369 arg.pages_found = 0;
5372 * It may be possible to isolate a pageblock even if the
5373 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5374 * notifier chain is used by balloon drivers to return the
5375 * number of pages in a range that are held by the balloon
5376 * driver to shrink memory. If all the pages are accounted for
5377 * by balloons, are free, or on the LRU, isolation can continue.
5378 * Later, for example, when memory hotplug notifier runs, these
5379 * pages reported as "can be isolated" should be isolated(freed)
5380 * by the balloon driver through the memory notifier chain.
5382 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5383 notifier_ret = notifier_to_errno(notifier_ret);
5387 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5388 * We just check MOVABLE pages.
5390 if (__count_immobile_pages(zone, page, arg.pages_found))
5394 * immobile means "not-on-lru" paes. If immobile is larger than
5395 * removable-by-driver pages reported by notifier, we'll fail.
5400 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5401 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5404 spin_unlock_irqrestore(&zone->lock, flags);
5410 void unset_migratetype_isolate(struct page *page)
5413 unsigned long flags;
5414 zone = page_zone(page);
5415 spin_lock_irqsave(&zone->lock, flags);
5416 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5418 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5419 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5421 spin_unlock_irqrestore(&zone->lock, flags);
5424 #ifdef CONFIG_MEMORY_HOTREMOVE
5426 * All pages in the range must be isolated before calling this.
5429 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5435 unsigned long flags;
5436 /* find the first valid pfn */
5437 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5442 zone = page_zone(pfn_to_page(pfn));
5443 spin_lock_irqsave(&zone->lock, flags);
5445 while (pfn < end_pfn) {
5446 if (!pfn_valid(pfn)) {
5450 page = pfn_to_page(pfn);
5451 BUG_ON(page_count(page));
5452 BUG_ON(!PageBuddy(page));
5453 order = page_order(page);
5454 #ifdef CONFIG_DEBUG_VM
5455 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5456 pfn, 1 << order, end_pfn);
5458 list_del(&page->lru);
5459 rmv_page_order(page);
5460 zone->free_area[order].nr_free--;
5461 __mod_zone_page_state(zone, NR_FREE_PAGES,
5463 for (i = 0; i < (1 << order); i++)
5464 SetPageReserved((page+i));
5465 pfn += (1 << order);
5467 spin_unlock_irqrestore(&zone->lock, flags);
5471 #ifdef CONFIG_MEMORY_FAILURE
5472 bool is_free_buddy_page(struct page *page)
5474 struct zone *zone = page_zone(page);
5475 unsigned long pfn = page_to_pfn(page);
5476 unsigned long flags;
5479 spin_lock_irqsave(&zone->lock, flags);
5480 for (order = 0; order < MAX_ORDER; order++) {
5481 struct page *page_head = page - (pfn & ((1 << order) - 1));
5483 if (PageBuddy(page_head) && page_order(page_head) >= order)
5486 spin_unlock_irqrestore(&zone->lock, flags);
5488 return order < MAX_ORDER;
5492 static struct trace_print_flags pageflag_names[] = {
5493 {1UL << PG_locked, "locked" },
5494 {1UL << PG_error, "error" },
5495 {1UL << PG_referenced, "referenced" },
5496 {1UL << PG_uptodate, "uptodate" },
5497 {1UL << PG_dirty, "dirty" },
5498 {1UL << PG_lru, "lru" },
5499 {1UL << PG_active, "active" },
5500 {1UL << PG_slab, "slab" },
5501 {1UL << PG_owner_priv_1, "owner_priv_1" },
5502 {1UL << PG_arch_1, "arch_1" },
5503 {1UL << PG_reserved, "reserved" },
5504 {1UL << PG_private, "private" },
5505 {1UL << PG_private_2, "private_2" },
5506 {1UL << PG_writeback, "writeback" },
5507 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5508 {1UL << PG_head, "head" },
5509 {1UL << PG_tail, "tail" },
5511 {1UL << PG_compound, "compound" },
5513 {1UL << PG_swapcache, "swapcache" },
5514 {1UL << PG_mappedtodisk, "mappedtodisk" },
5515 {1UL << PG_reclaim, "reclaim" },
5516 {1UL << PG_swapbacked, "swapbacked" },
5517 {1UL << PG_unevictable, "unevictable" },
5519 {1UL << PG_mlocked, "mlocked" },
5521 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5522 {1UL << PG_uncached, "uncached" },
5524 #ifdef CONFIG_MEMORY_FAILURE
5525 {1UL << PG_hwpoison, "hwpoison" },
5530 static void dump_page_flags(unsigned long flags)
5532 const char *delim = "";
5536 printk(KERN_ALERT "page flags: %#lx(", flags);
5538 /* remove zone id */
5539 flags &= (1UL << NR_PAGEFLAGS) - 1;
5541 for (i = 0; pageflag_names[i].name && flags; i++) {
5543 mask = pageflag_names[i].mask;
5544 if ((flags & mask) != mask)
5548 printk("%s%s", delim, pageflag_names[i].name);
5552 /* check for left over flags */
5554 printk("%s%#lx", delim, flags);
5559 void dump_page(struct page *page)
5562 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5563 page, atomic_read(&page->_count), page_mapcount(page),
5564 page->mapping, page->index);
5565 dump_page_flags(page->flags);
5566 mem_cgroup_print_bad_page(page);