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>
61 #include <asm/tlbflush.h>
62 #include <asm/div64.h>
65 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
66 DEFINE_PER_CPU(int, numa_node);
67 EXPORT_PER_CPU_SYMBOL(numa_node);
70 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
72 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
73 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
74 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
75 * defined in <linux/topology.h>.
77 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
78 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
82 * Array of node states.
84 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
85 [N_POSSIBLE] = NODE_MASK_ALL,
86 [N_ONLINE] = { { [0] = 1UL } },
88 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
90 [N_HIGH_MEMORY] = { { [0] = 1UL } },
92 [N_CPU] = { { [0] = 1UL } },
95 EXPORT_SYMBOL(node_states);
97 unsigned long totalram_pages __read_mostly;
98 unsigned long totalreserve_pages __read_mostly;
99 int percpu_pagelist_fraction;
100 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
102 #ifdef CONFIG_PM_SLEEP
104 * The following functions are used by the suspend/hibernate code to temporarily
105 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
106 * while devices are suspended. To avoid races with the suspend/hibernate code,
107 * they should always be called with pm_mutex held (gfp_allowed_mask also should
108 * only be modified with pm_mutex held, unless the suspend/hibernate code is
109 * guaranteed not to run in parallel with that modification).
112 static gfp_t saved_gfp_mask;
114 void pm_restore_gfp_mask(void)
116 WARN_ON(!mutex_is_locked(&pm_mutex));
117 if (saved_gfp_mask) {
118 gfp_allowed_mask = saved_gfp_mask;
123 void pm_restrict_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 WARN_ON(saved_gfp_mask);
127 saved_gfp_mask = gfp_allowed_mask;
128 gfp_allowed_mask &= ~GFP_IOFS;
130 #endif /* CONFIG_PM_SLEEP */
132 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
133 int pageblock_order __read_mostly;
136 static void __free_pages_ok(struct page *page, unsigned int order);
139 * results with 256, 32 in the lowmem_reserve sysctl:
140 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
141 * 1G machine -> (16M dma, 784M normal, 224M high)
142 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
143 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
144 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
146 * TBD: should special case ZONE_DMA32 machines here - in those we normally
147 * don't need any ZONE_NORMAL reservation
149 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
150 #ifdef CONFIG_ZONE_DMA
153 #ifdef CONFIG_ZONE_DMA32
156 #ifdef CONFIG_HIGHMEM
162 EXPORT_SYMBOL(totalram_pages);
164 static char * const zone_names[MAX_NR_ZONES] = {
165 #ifdef CONFIG_ZONE_DMA
168 #ifdef CONFIG_ZONE_DMA32
172 #ifdef CONFIG_HIGHMEM
178 int min_free_kbytes = 1024;
180 static unsigned long __meminitdata nr_kernel_pages;
181 static unsigned long __meminitdata nr_all_pages;
182 static unsigned long __meminitdata dma_reserve;
184 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
185 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
186 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
187 static unsigned long __initdata required_kernelcore;
188 static unsigned long __initdata required_movablecore;
189 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
191 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
193 EXPORT_SYMBOL(movable_zone);
194 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
197 int nr_node_ids __read_mostly = MAX_NUMNODES;
198 int nr_online_nodes __read_mostly = 1;
199 EXPORT_SYMBOL(nr_node_ids);
200 EXPORT_SYMBOL(nr_online_nodes);
203 int page_group_by_mobility_disabled __read_mostly;
205 static void set_pageblock_migratetype(struct page *page, int migratetype)
208 if (unlikely(page_group_by_mobility_disabled))
209 migratetype = MIGRATE_UNMOVABLE;
211 set_pageblock_flags_group(page, (unsigned long)migratetype,
212 PB_migrate, PB_migrate_end);
215 bool oom_killer_disabled __read_mostly;
217 #ifdef CONFIG_DEBUG_VM
218 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
222 unsigned long pfn = page_to_pfn(page);
225 seq = zone_span_seqbegin(zone);
226 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
228 else if (pfn < zone->zone_start_pfn)
230 } while (zone_span_seqretry(zone, seq));
235 static int page_is_consistent(struct zone *zone, struct page *page)
237 if (!pfn_valid_within(page_to_pfn(page)))
239 if (zone != page_zone(page))
245 * Temporary debugging check for pages not lying within a given zone.
247 static int bad_range(struct zone *zone, struct page *page)
249 if (page_outside_zone_boundaries(zone, page))
251 if (!page_is_consistent(zone, page))
257 static inline int bad_range(struct zone *zone, struct page *page)
263 static void bad_page(struct page *page)
265 static unsigned long resume;
266 static unsigned long nr_shown;
267 static unsigned long nr_unshown;
269 /* Don't complain about poisoned pages */
270 if (PageHWPoison(page)) {
271 reset_page_mapcount(page); /* remove PageBuddy */
276 * Allow a burst of 60 reports, then keep quiet for that minute;
277 * or allow a steady drip of one report per second.
279 if (nr_shown == 60) {
280 if (time_before(jiffies, resume)) {
286 "BUG: Bad page state: %lu messages suppressed\n",
293 resume = jiffies + 60 * HZ;
295 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
296 current->comm, page_to_pfn(page));
302 /* Leave bad fields for debug, except PageBuddy could make trouble */
303 reset_page_mapcount(page); /* remove PageBuddy */
304 add_taint(TAINT_BAD_PAGE);
308 * Higher-order pages are called "compound pages". They are structured thusly:
310 * The first PAGE_SIZE page is called the "head page".
312 * The remaining PAGE_SIZE pages are called "tail pages".
314 * All pages have PG_compound set. All tail pages have their ->first_page
315 * pointing at the head page.
317 * The first tail page's ->lru.next holds the address of the compound page's
318 * put_page() function. Its ->lru.prev holds the order of allocation.
319 * This usage means that zero-order pages may not be compound.
322 static void free_compound_page(struct page *page)
324 __free_pages_ok(page, compound_order(page));
327 void prep_compound_page(struct page *page, unsigned long order)
330 int nr_pages = 1 << order;
332 set_compound_page_dtor(page, free_compound_page);
333 set_compound_order(page, order);
335 for (i = 1; i < nr_pages; i++) {
336 struct page *p = page + i;
338 set_page_count(p, 0);
339 p->first_page = page;
343 /* update __split_huge_page_refcount if you change this function */
344 static int destroy_compound_page(struct page *page, unsigned long order)
347 int nr_pages = 1 << order;
350 if (unlikely(compound_order(page) != order) ||
351 unlikely(!PageHead(page))) {
356 __ClearPageHead(page);
358 for (i = 1; i < nr_pages; i++) {
359 struct page *p = page + i;
361 if (unlikely(!PageTail(p) || (p->first_page != page))) {
371 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
376 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
377 * and __GFP_HIGHMEM from hard or soft interrupt context.
379 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
380 for (i = 0; i < (1 << order); i++)
381 clear_highpage(page + i);
384 static inline void set_page_order(struct page *page, int order)
386 set_page_private(page, order);
387 __SetPageBuddy(page);
390 static inline void rmv_page_order(struct page *page)
392 __ClearPageBuddy(page);
393 set_page_private(page, 0);
397 * Locate the struct page for both the matching buddy in our
398 * pair (buddy1) and the combined O(n+1) page they form (page).
400 * 1) Any buddy B1 will have an order O twin B2 which satisfies
401 * the following equation:
403 * For example, if the starting buddy (buddy2) is #8 its order
405 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
407 * 2) Any buddy B will have an order O+1 parent P which
408 * satisfies the following equation:
411 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
413 static inline unsigned long
414 __find_buddy_index(unsigned long page_idx, unsigned int order)
416 return page_idx ^ (1 << order);
420 * This function checks whether a page is free && is the buddy
421 * we can do coalesce a page and its buddy if
422 * (a) the buddy is not in a hole &&
423 * (b) the buddy is in the buddy system &&
424 * (c) a page and its buddy have the same order &&
425 * (d) a page and its buddy are in the same zone.
427 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
428 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
430 * For recording page's order, we use page_private(page).
432 static inline int page_is_buddy(struct page *page, struct page *buddy,
435 if (!pfn_valid_within(page_to_pfn(buddy)))
438 if (page_zone_id(page) != page_zone_id(buddy))
441 if (PageBuddy(buddy) && page_order(buddy) == order) {
442 VM_BUG_ON(page_count(buddy) != 0);
449 * Freeing function for a buddy system allocator.
451 * The concept of a buddy system is to maintain direct-mapped table
452 * (containing bit values) for memory blocks of various "orders".
453 * The bottom level table contains the map for the smallest allocatable
454 * units of memory (here, pages), and each level above it describes
455 * pairs of units from the levels below, hence, "buddies".
456 * At a high level, all that happens here is marking the table entry
457 * at the bottom level available, and propagating the changes upward
458 * as necessary, plus some accounting needed to play nicely with other
459 * parts of the VM system.
460 * At each level, we keep a list of pages, which are heads of continuous
461 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
462 * order is recorded in page_private(page) field.
463 * So when we are allocating or freeing one, we can derive the state of the
464 * other. That is, if we allocate a small block, and both were
465 * free, the remainder of the region must be split into blocks.
466 * If a block is freed, and its buddy is also free, then this
467 * triggers coalescing into a block of larger size.
472 static inline void __free_one_page(struct page *page,
473 struct zone *zone, unsigned int order,
476 unsigned long page_idx;
477 unsigned long combined_idx;
478 unsigned long uninitialized_var(buddy_idx);
481 if (unlikely(PageCompound(page)))
482 if (unlikely(destroy_compound_page(page, order)))
485 VM_BUG_ON(migratetype == -1);
487 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
489 VM_BUG_ON(page_idx & ((1 << order) - 1));
490 VM_BUG_ON(bad_range(zone, page));
492 while (order < MAX_ORDER-1) {
493 buddy_idx = __find_buddy_index(page_idx, order);
494 buddy = page + (buddy_idx - page_idx);
495 if (!page_is_buddy(page, buddy, order))
498 /* Our buddy is free, merge with it and move up one order. */
499 list_del(&buddy->lru);
500 zone->free_area[order].nr_free--;
501 rmv_page_order(buddy);
502 combined_idx = buddy_idx & page_idx;
503 page = page + (combined_idx - page_idx);
504 page_idx = combined_idx;
507 set_page_order(page, order);
510 * If this is not the largest possible page, check if the buddy
511 * of the next-highest order is free. If it is, it's possible
512 * that pages are being freed that will coalesce soon. In case,
513 * that is happening, add the free page to the tail of the list
514 * so it's less likely to be used soon and more likely to be merged
515 * as a higher order page
517 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
518 struct page *higher_page, *higher_buddy;
519 combined_idx = buddy_idx & page_idx;
520 higher_page = page + (combined_idx - page_idx);
521 buddy_idx = __find_buddy_index(combined_idx, order + 1);
522 higher_buddy = page + (buddy_idx - combined_idx);
523 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
524 list_add_tail(&page->lru,
525 &zone->free_area[order].free_list[migratetype]);
530 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
532 zone->free_area[order].nr_free++;
536 * free_page_mlock() -- clean up attempts to free and mlocked() page.
537 * Page should not be on lru, so no need to fix that up.
538 * free_pages_check() will verify...
540 static inline void free_page_mlock(struct page *page)
542 __dec_zone_page_state(page, NR_MLOCK);
543 __count_vm_event(UNEVICTABLE_MLOCKFREED);
546 static inline int free_pages_check(struct page *page)
548 if (unlikely(page_mapcount(page) |
549 (page->mapping != NULL) |
550 (atomic_read(&page->_count) != 0) |
551 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
552 (mem_cgroup_bad_page_check(page)))) {
556 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
557 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
562 * Frees a number of pages from the PCP lists
563 * Assumes all pages on list are in same zone, and of same order.
564 * count is the number of pages to free.
566 * If the zone was previously in an "all pages pinned" state then look to
567 * see if this freeing clears that state.
569 * And clear the zone's pages_scanned counter, to hold off the "all pages are
570 * pinned" detection logic.
572 static void free_pcppages_bulk(struct zone *zone, int count,
573 struct per_cpu_pages *pcp)
579 spin_lock(&zone->lock);
580 zone->all_unreclaimable = 0;
581 zone->pages_scanned = 0;
585 struct list_head *list;
588 * Remove pages from lists in a round-robin fashion. A
589 * batch_free count is maintained that is incremented when an
590 * empty list is encountered. This is so more pages are freed
591 * off fuller lists instead of spinning excessively around empty
596 if (++migratetype == MIGRATE_PCPTYPES)
598 list = &pcp->lists[migratetype];
599 } while (list_empty(list));
601 /* This is the only non-empty list. Free them all. */
602 if (batch_free == MIGRATE_PCPTYPES)
603 batch_free = to_free;
606 page = list_entry(list->prev, struct page, lru);
607 /* must delete as __free_one_page list manipulates */
608 list_del(&page->lru);
609 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
610 __free_one_page(page, zone, 0, page_private(page));
611 trace_mm_page_pcpu_drain(page, 0, page_private(page));
612 } while (--to_free && --batch_free && !list_empty(list));
614 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
615 spin_unlock(&zone->lock);
618 static void free_one_page(struct zone *zone, struct page *page, int order,
621 spin_lock(&zone->lock);
622 zone->all_unreclaimable = 0;
623 zone->pages_scanned = 0;
625 __free_one_page(page, zone, order, migratetype);
626 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
627 spin_unlock(&zone->lock);
630 static bool free_pages_prepare(struct page *page, unsigned int order)
635 trace_mm_page_free(page, order);
636 kmemcheck_free_shadow(page, order);
639 page->mapping = NULL;
640 for (i = 0; i < (1 << order); i++)
641 bad += free_pages_check(page + i);
645 if (!PageHighMem(page)) {
646 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
647 debug_check_no_obj_freed(page_address(page),
650 arch_free_page(page, order);
651 kernel_map_pages(page, 1 << order, 0);
656 static void __free_pages_ok(struct page *page, unsigned int order)
659 int wasMlocked = __TestClearPageMlocked(page);
661 if (!free_pages_prepare(page, order))
664 local_irq_save(flags);
665 if (unlikely(wasMlocked))
666 free_page_mlock(page);
667 __count_vm_events(PGFREE, 1 << order);
668 free_one_page(page_zone(page), page, order,
669 get_pageblock_migratetype(page));
670 local_irq_restore(flags);
674 * permit the bootmem allocator to evade page validation on high-order frees
676 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
679 __ClearPageReserved(page);
680 set_page_count(page, 0);
681 set_page_refcounted(page);
687 for (loop = 0; loop < (1 << order); loop++) {
688 struct page *p = &page[loop];
690 if (loop + 1 < (1 << order))
692 __ClearPageReserved(p);
693 set_page_count(p, 0);
696 set_page_refcounted(page);
697 __free_pages(page, order);
703 * The order of subdivision here is critical for the IO subsystem.
704 * Please do not alter this order without good reasons and regression
705 * testing. Specifically, as large blocks of memory are subdivided,
706 * the order in which smaller blocks are delivered depends on the order
707 * they're subdivided in this function. This is the primary factor
708 * influencing the order in which pages are delivered to the IO
709 * subsystem according to empirical testing, and this is also justified
710 * by considering the behavior of a buddy system containing a single
711 * large block of memory acted on by a series of small allocations.
712 * This behavior is a critical factor in sglist merging's success.
716 static inline void expand(struct zone *zone, struct page *page,
717 int low, int high, struct free_area *area,
720 unsigned long size = 1 << high;
726 VM_BUG_ON(bad_range(zone, &page[size]));
727 list_add(&page[size].lru, &area->free_list[migratetype]);
729 set_page_order(&page[size], high);
734 * This page is about to be returned from the page allocator
736 static inline int check_new_page(struct page *page)
738 if (unlikely(page_mapcount(page) |
739 (page->mapping != NULL) |
740 (atomic_read(&page->_count) != 0) |
741 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
742 (mem_cgroup_bad_page_check(page)))) {
749 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
753 for (i = 0; i < (1 << order); i++) {
754 struct page *p = page + i;
755 if (unlikely(check_new_page(p)))
759 set_page_private(page, 0);
760 set_page_refcounted(page);
762 arch_alloc_page(page, order);
763 kernel_map_pages(page, 1 << order, 1);
765 if (gfp_flags & __GFP_ZERO)
766 prep_zero_page(page, order, gfp_flags);
768 if (order && (gfp_flags & __GFP_COMP))
769 prep_compound_page(page, order);
775 * Go through the free lists for the given migratetype and remove
776 * the smallest available page from the freelists
779 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
782 unsigned int current_order;
783 struct free_area * area;
786 /* Find a page of the appropriate size in the preferred list */
787 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
788 area = &(zone->free_area[current_order]);
789 if (list_empty(&area->free_list[migratetype]))
792 page = list_entry(area->free_list[migratetype].next,
794 list_del(&page->lru);
795 rmv_page_order(page);
797 expand(zone, page, order, current_order, area, migratetype);
806 * This array describes the order lists are fallen back to when
807 * the free lists for the desirable migrate type are depleted
809 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
810 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
811 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
812 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
813 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
817 * Move the free pages in a range to the free lists of the requested type.
818 * Note that start_page and end_pages are not aligned on a pageblock
819 * boundary. If alignment is required, use move_freepages_block()
821 static int move_freepages(struct zone *zone,
822 struct page *start_page, struct page *end_page,
829 #ifndef CONFIG_HOLES_IN_ZONE
831 * page_zone is not safe to call in this context when
832 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
833 * anyway as we check zone boundaries in move_freepages_block().
834 * Remove at a later date when no bug reports exist related to
835 * grouping pages by mobility
837 BUG_ON(page_zone(start_page) != page_zone(end_page));
840 for (page = start_page; page <= end_page;) {
841 /* Make sure we are not inadvertently changing nodes */
842 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
844 if (!pfn_valid_within(page_to_pfn(page))) {
849 if (!PageBuddy(page)) {
854 order = page_order(page);
855 list_move(&page->lru,
856 &zone->free_area[order].free_list[migratetype]);
858 pages_moved += 1 << order;
864 static int move_freepages_block(struct zone *zone, struct page *page,
867 unsigned long start_pfn, end_pfn;
868 struct page *start_page, *end_page;
870 start_pfn = page_to_pfn(page);
871 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
872 start_page = pfn_to_page(start_pfn);
873 end_page = start_page + pageblock_nr_pages - 1;
874 end_pfn = start_pfn + pageblock_nr_pages - 1;
876 /* Do not cross zone boundaries */
877 if (start_pfn < zone->zone_start_pfn)
879 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
882 return move_freepages(zone, start_page, end_page, migratetype);
885 static void change_pageblock_range(struct page *pageblock_page,
886 int start_order, int migratetype)
888 int nr_pageblocks = 1 << (start_order - pageblock_order);
890 while (nr_pageblocks--) {
891 set_pageblock_migratetype(pageblock_page, migratetype);
892 pageblock_page += pageblock_nr_pages;
896 /* Remove an element from the buddy allocator from the fallback list */
897 static inline struct page *
898 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
900 struct free_area * area;
905 /* Find the largest possible block of pages in the other list */
906 for (current_order = MAX_ORDER-1; current_order >= order;
908 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
909 migratetype = fallbacks[start_migratetype][i];
911 /* MIGRATE_RESERVE handled later if necessary */
912 if (migratetype == MIGRATE_RESERVE)
915 area = &(zone->free_area[current_order]);
916 if (list_empty(&area->free_list[migratetype]))
919 page = list_entry(area->free_list[migratetype].next,
924 * If breaking a large block of pages, move all free
925 * pages to the preferred allocation list. If falling
926 * back for a reclaimable kernel allocation, be more
927 * aggressive about taking ownership of free pages
929 if (unlikely(current_order >= (pageblock_order >> 1)) ||
930 start_migratetype == MIGRATE_RECLAIMABLE ||
931 page_group_by_mobility_disabled) {
933 pages = move_freepages_block(zone, page,
936 /* Claim the whole block if over half of it is free */
937 if (pages >= (1 << (pageblock_order-1)) ||
938 page_group_by_mobility_disabled)
939 set_pageblock_migratetype(page,
942 migratetype = start_migratetype;
945 /* Remove the page from the freelists */
946 list_del(&page->lru);
947 rmv_page_order(page);
949 /* Take ownership for orders >= pageblock_order */
950 if (current_order >= pageblock_order)
951 change_pageblock_range(page, current_order,
954 expand(zone, page, order, current_order, area, migratetype);
956 trace_mm_page_alloc_extfrag(page, order, current_order,
957 start_migratetype, migratetype);
967 * Do the hard work of removing an element from the buddy allocator.
968 * Call me with the zone->lock already held.
970 static struct page *__rmqueue(struct zone *zone, unsigned int order,
976 page = __rmqueue_smallest(zone, order, migratetype);
978 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
979 page = __rmqueue_fallback(zone, order, migratetype);
982 * Use MIGRATE_RESERVE rather than fail an allocation. goto
983 * is used because __rmqueue_smallest is an inline function
984 * and we want just one call site
987 migratetype = MIGRATE_RESERVE;
992 trace_mm_page_alloc_zone_locked(page, order, migratetype);
997 * Obtain a specified number of elements from the buddy allocator, all under
998 * a single hold of the lock, for efficiency. Add them to the supplied list.
999 * Returns the number of new pages which were placed at *list.
1001 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1002 unsigned long count, struct list_head *list,
1003 int migratetype, int cold)
1007 spin_lock(&zone->lock);
1008 for (i = 0; i < count; ++i) {
1009 struct page *page = __rmqueue(zone, order, migratetype);
1010 if (unlikely(page == NULL))
1014 * Split buddy pages returned by expand() are received here
1015 * in physical page order. The page is added to the callers and
1016 * list and the list head then moves forward. From the callers
1017 * perspective, the linked list is ordered by page number in
1018 * some conditions. This is useful for IO devices that can
1019 * merge IO requests if the physical pages are ordered
1022 if (likely(cold == 0))
1023 list_add(&page->lru, list);
1025 list_add_tail(&page->lru, list);
1026 set_page_private(page, migratetype);
1029 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1030 spin_unlock(&zone->lock);
1036 * Called from the vmstat counter updater to drain pagesets of this
1037 * currently executing processor on remote nodes after they have
1040 * Note that this function must be called with the thread pinned to
1041 * a single processor.
1043 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1045 unsigned long flags;
1048 local_irq_save(flags);
1049 if (pcp->count >= pcp->batch)
1050 to_drain = pcp->batch;
1052 to_drain = pcp->count;
1053 free_pcppages_bulk(zone, to_drain, pcp);
1054 pcp->count -= to_drain;
1055 local_irq_restore(flags);
1060 * Drain pages of the indicated processor.
1062 * The processor must either be the current processor and the
1063 * thread pinned to the current processor or a processor that
1066 static void drain_pages(unsigned int cpu)
1068 unsigned long flags;
1071 for_each_populated_zone(zone) {
1072 struct per_cpu_pageset *pset;
1073 struct per_cpu_pages *pcp;
1075 local_irq_save(flags);
1076 pset = per_cpu_ptr(zone->pageset, cpu);
1080 free_pcppages_bulk(zone, pcp->count, pcp);
1083 local_irq_restore(flags);
1088 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1090 void drain_local_pages(void *arg)
1092 drain_pages(smp_processor_id());
1096 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1098 void drain_all_pages(void)
1100 on_each_cpu(drain_local_pages, NULL, 1);
1103 #ifdef CONFIG_HIBERNATION
1105 void mark_free_pages(struct zone *zone)
1107 unsigned long pfn, max_zone_pfn;
1108 unsigned long flags;
1110 struct list_head *curr;
1112 if (!zone->spanned_pages)
1115 spin_lock_irqsave(&zone->lock, flags);
1117 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1118 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1119 if (pfn_valid(pfn)) {
1120 struct page *page = pfn_to_page(pfn);
1122 if (!swsusp_page_is_forbidden(page))
1123 swsusp_unset_page_free(page);
1126 for_each_migratetype_order(order, t) {
1127 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1130 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1131 for (i = 0; i < (1UL << order); i++)
1132 swsusp_set_page_free(pfn_to_page(pfn + i));
1135 spin_unlock_irqrestore(&zone->lock, flags);
1137 #endif /* CONFIG_PM */
1140 * Free a 0-order page
1141 * cold == 1 ? free a cold page : free a hot page
1143 void free_hot_cold_page(struct page *page, int cold)
1145 struct zone *zone = page_zone(page);
1146 struct per_cpu_pages *pcp;
1147 unsigned long flags;
1149 int wasMlocked = __TestClearPageMlocked(page);
1151 if (!free_pages_prepare(page, 0))
1154 migratetype = get_pageblock_migratetype(page);
1155 set_page_private(page, migratetype);
1156 local_irq_save(flags);
1157 if (unlikely(wasMlocked))
1158 free_page_mlock(page);
1159 __count_vm_event(PGFREE);
1162 * We only track unmovable, reclaimable and movable on pcp lists.
1163 * Free ISOLATE pages back to the allocator because they are being
1164 * offlined but treat RESERVE as movable pages so we can get those
1165 * areas back if necessary. Otherwise, we may have to free
1166 * excessively into the page allocator
1168 if (migratetype >= MIGRATE_PCPTYPES) {
1169 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1170 free_one_page(zone, page, 0, migratetype);
1173 migratetype = MIGRATE_MOVABLE;
1176 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1178 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1180 list_add(&page->lru, &pcp->lists[migratetype]);
1182 if (pcp->count >= pcp->high) {
1183 free_pcppages_bulk(zone, pcp->batch, pcp);
1184 pcp->count -= pcp->batch;
1188 local_irq_restore(flags);
1192 * Free a list of 0-order pages
1194 void free_hot_cold_page_list(struct list_head *list, int cold)
1196 struct page *page, *next;
1198 list_for_each_entry_safe(page, next, list, lru) {
1199 trace_mm_page_free_batched(page, cold);
1200 free_hot_cold_page(page, cold);
1205 * split_page takes a non-compound higher-order page, and splits it into
1206 * n (1<<order) sub-pages: page[0..n]
1207 * Each sub-page must be freed individually.
1209 * Note: this is probably too low level an operation for use in drivers.
1210 * Please consult with lkml before using this in your driver.
1212 void split_page(struct page *page, unsigned int order)
1216 VM_BUG_ON(PageCompound(page));
1217 VM_BUG_ON(!page_count(page));
1219 #ifdef CONFIG_KMEMCHECK
1221 * Split shadow pages too, because free(page[0]) would
1222 * otherwise free the whole shadow.
1224 if (kmemcheck_page_is_tracked(page))
1225 split_page(virt_to_page(page[0].shadow), order);
1228 for (i = 1; i < (1 << order); i++)
1229 set_page_refcounted(page + i);
1233 * Similar to split_page except the page is already free. As this is only
1234 * being used for migration, the migratetype of the block also changes.
1235 * As this is called with interrupts disabled, the caller is responsible
1236 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1239 * Note: this is probably too low level an operation for use in drivers.
1240 * Please consult with lkml before using this in your driver.
1242 int split_free_page(struct page *page)
1245 unsigned long watermark;
1248 BUG_ON(!PageBuddy(page));
1250 zone = page_zone(page);
1251 order = page_order(page);
1253 /* Obey watermarks as if the page was being allocated */
1254 watermark = low_wmark_pages(zone) + (1 << order);
1255 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1258 /* Remove page from free list */
1259 list_del(&page->lru);
1260 zone->free_area[order].nr_free--;
1261 rmv_page_order(page);
1262 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1264 /* Split into individual pages */
1265 set_page_refcounted(page);
1266 split_page(page, order);
1268 if (order >= pageblock_order - 1) {
1269 struct page *endpage = page + (1 << order) - 1;
1270 for (; page < endpage; page += pageblock_nr_pages)
1271 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1278 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1279 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1283 struct page *buffered_rmqueue(struct zone *preferred_zone,
1284 struct zone *zone, int order, gfp_t gfp_flags,
1287 unsigned long flags;
1289 int cold = !!(gfp_flags & __GFP_COLD);
1292 if (likely(order == 0)) {
1293 struct per_cpu_pages *pcp;
1294 struct list_head *list;
1296 local_irq_save(flags);
1297 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1298 list = &pcp->lists[migratetype];
1299 if (list_empty(list)) {
1300 pcp->count += rmqueue_bulk(zone, 0,
1303 if (unlikely(list_empty(list)))
1308 page = list_entry(list->prev, struct page, lru);
1310 page = list_entry(list->next, struct page, lru);
1312 list_del(&page->lru);
1315 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1317 * __GFP_NOFAIL is not to be used in new code.
1319 * All __GFP_NOFAIL callers should be fixed so that they
1320 * properly detect and handle allocation failures.
1322 * We most definitely don't want callers attempting to
1323 * allocate greater than order-1 page units with
1326 WARN_ON_ONCE(order > 1);
1328 spin_lock_irqsave(&zone->lock, flags);
1329 page = __rmqueue(zone, order, migratetype);
1330 spin_unlock(&zone->lock);
1333 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1336 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1337 zone_statistics(preferred_zone, zone, gfp_flags);
1338 local_irq_restore(flags);
1340 VM_BUG_ON(bad_range(zone, page));
1341 if (prep_new_page(page, order, gfp_flags))
1346 local_irq_restore(flags);
1350 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1351 #define ALLOC_WMARK_MIN WMARK_MIN
1352 #define ALLOC_WMARK_LOW WMARK_LOW
1353 #define ALLOC_WMARK_HIGH WMARK_HIGH
1354 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1356 /* Mask to get the watermark bits */
1357 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1359 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1360 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1361 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1363 #ifdef CONFIG_FAIL_PAGE_ALLOC
1366 struct fault_attr attr;
1368 u32 ignore_gfp_highmem;
1369 u32 ignore_gfp_wait;
1371 } fail_page_alloc = {
1372 .attr = FAULT_ATTR_INITIALIZER,
1373 .ignore_gfp_wait = 1,
1374 .ignore_gfp_highmem = 1,
1378 static int __init setup_fail_page_alloc(char *str)
1380 return setup_fault_attr(&fail_page_alloc.attr, str);
1382 __setup("fail_page_alloc=", setup_fail_page_alloc);
1384 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1386 if (order < fail_page_alloc.min_order)
1388 if (gfp_mask & __GFP_NOFAIL)
1390 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1392 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1395 return should_fail(&fail_page_alloc.attr, 1 << order);
1398 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1400 static int __init fail_page_alloc_debugfs(void)
1402 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1405 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1406 &fail_page_alloc.attr);
1408 return PTR_ERR(dir);
1410 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1411 &fail_page_alloc.ignore_gfp_wait))
1413 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1414 &fail_page_alloc.ignore_gfp_highmem))
1416 if (!debugfs_create_u32("min-order", mode, dir,
1417 &fail_page_alloc.min_order))
1422 debugfs_remove_recursive(dir);
1427 late_initcall(fail_page_alloc_debugfs);
1429 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1431 #else /* CONFIG_FAIL_PAGE_ALLOC */
1433 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1438 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1441 * Return true if free pages are above 'mark'. This takes into account the order
1442 * of the allocation.
1444 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1445 int classzone_idx, int alloc_flags, long free_pages)
1447 /* free_pages my go negative - that's OK */
1451 free_pages -= (1 << order) + 1;
1452 if (alloc_flags & ALLOC_HIGH)
1454 if (alloc_flags & ALLOC_HARDER)
1457 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1459 for (o = 0; o < order; o++) {
1460 /* At the next order, this order's pages become unavailable */
1461 free_pages -= z->free_area[o].nr_free << o;
1463 /* Require fewer higher order pages to be free */
1466 if (free_pages <= min)
1472 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1473 int classzone_idx, int alloc_flags)
1475 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1476 zone_page_state(z, NR_FREE_PAGES));
1479 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1480 int classzone_idx, int alloc_flags)
1482 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1484 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1485 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1487 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1493 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1494 * skip over zones that are not allowed by the cpuset, or that have
1495 * been recently (in last second) found to be nearly full. See further
1496 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1497 * that have to skip over a lot of full or unallowed zones.
1499 * If the zonelist cache is present in the passed in zonelist, then
1500 * returns a pointer to the allowed node mask (either the current
1501 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1503 * If the zonelist cache is not available for this zonelist, does
1504 * nothing and returns NULL.
1506 * If the fullzones BITMAP in the zonelist cache is stale (more than
1507 * a second since last zap'd) then we zap it out (clear its bits.)
1509 * We hold off even calling zlc_setup, until after we've checked the
1510 * first zone in the zonelist, on the theory that most allocations will
1511 * be satisfied from that first zone, so best to examine that zone as
1512 * quickly as we can.
1514 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1516 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1517 nodemask_t *allowednodes; /* zonelist_cache approximation */
1519 zlc = zonelist->zlcache_ptr;
1523 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1524 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1525 zlc->last_full_zap = jiffies;
1528 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1529 &cpuset_current_mems_allowed :
1530 &node_states[N_HIGH_MEMORY];
1531 return allowednodes;
1535 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1536 * if it is worth looking at further for free memory:
1537 * 1) Check that the zone isn't thought to be full (doesn't have its
1538 * bit set in the zonelist_cache fullzones BITMAP).
1539 * 2) Check that the zones node (obtained from the zonelist_cache
1540 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1541 * Return true (non-zero) if zone is worth looking at further, or
1542 * else return false (zero) if it is not.
1544 * This check -ignores- the distinction between various watermarks,
1545 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1546 * found to be full for any variation of these watermarks, it will
1547 * be considered full for up to one second by all requests, unless
1548 * we are so low on memory on all allowed nodes that we are forced
1549 * into the second scan of the zonelist.
1551 * In the second scan we ignore this zonelist cache and exactly
1552 * apply the watermarks to all zones, even it is slower to do so.
1553 * We are low on memory in the second scan, and should leave no stone
1554 * unturned looking for a free page.
1556 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1557 nodemask_t *allowednodes)
1559 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1560 int i; /* index of *z in zonelist zones */
1561 int n; /* node that zone *z is on */
1563 zlc = zonelist->zlcache_ptr;
1567 i = z - zonelist->_zonerefs;
1570 /* This zone is worth trying if it is allowed but not full */
1571 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1575 * Given 'z' scanning a zonelist, set the corresponding bit in
1576 * zlc->fullzones, so that subsequent attempts to allocate a page
1577 * from that zone don't waste time re-examining it.
1579 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1581 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1582 int i; /* index of *z in zonelist zones */
1584 zlc = zonelist->zlcache_ptr;
1588 i = z - zonelist->_zonerefs;
1590 set_bit(i, zlc->fullzones);
1594 * clear all zones full, called after direct reclaim makes progress so that
1595 * a zone that was recently full is not skipped over for up to a second
1597 static void zlc_clear_zones_full(struct zonelist *zonelist)
1599 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1601 zlc = zonelist->zlcache_ptr;
1605 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1608 #else /* CONFIG_NUMA */
1610 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1615 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1616 nodemask_t *allowednodes)
1621 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1625 static void zlc_clear_zones_full(struct zonelist *zonelist)
1628 #endif /* CONFIG_NUMA */
1631 * get_page_from_freelist goes through the zonelist trying to allocate
1634 static struct page *
1635 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1636 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1637 struct zone *preferred_zone, int migratetype)
1640 struct page *page = NULL;
1643 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1644 int zlc_active = 0; /* set if using zonelist_cache */
1645 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1647 classzone_idx = zone_idx(preferred_zone);
1650 * Scan zonelist, looking for a zone with enough free.
1651 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1653 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1654 high_zoneidx, nodemask) {
1655 if (NUMA_BUILD && zlc_active &&
1656 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1658 if ((alloc_flags & ALLOC_CPUSET) &&
1659 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1662 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1663 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1667 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1668 if (zone_watermark_ok(zone, order, mark,
1669 classzone_idx, alloc_flags))
1672 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1674 * we do zlc_setup if there are multiple nodes
1675 * and before considering the first zone allowed
1678 allowednodes = zlc_setup(zonelist, alloc_flags);
1683 if (zone_reclaim_mode == 0)
1684 goto this_zone_full;
1687 * As we may have just activated ZLC, check if the first
1688 * eligible zone has failed zone_reclaim recently.
1690 if (NUMA_BUILD && zlc_active &&
1691 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1694 ret = zone_reclaim(zone, gfp_mask, order);
1696 case ZONE_RECLAIM_NOSCAN:
1699 case ZONE_RECLAIM_FULL:
1700 /* scanned but unreclaimable */
1703 /* did we reclaim enough */
1704 if (!zone_watermark_ok(zone, order, mark,
1705 classzone_idx, alloc_flags))
1706 goto this_zone_full;
1711 page = buffered_rmqueue(preferred_zone, zone, order,
1712 gfp_mask, migratetype);
1717 zlc_mark_zone_full(zonelist, z);
1720 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1721 /* Disable zlc cache for second zonelist scan */
1729 * Large machines with many possible nodes should not always dump per-node
1730 * meminfo in irq context.
1732 static inline bool should_suppress_show_mem(void)
1737 ret = in_interrupt();
1742 static DEFINE_RATELIMIT_STATE(nopage_rs,
1743 DEFAULT_RATELIMIT_INTERVAL,
1744 DEFAULT_RATELIMIT_BURST);
1746 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1748 unsigned int filter = SHOW_MEM_FILTER_NODES;
1750 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1754 * This documents exceptions given to allocations in certain
1755 * contexts that are allowed to allocate outside current's set
1758 if (!(gfp_mask & __GFP_NOMEMALLOC))
1759 if (test_thread_flag(TIF_MEMDIE) ||
1760 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1761 filter &= ~SHOW_MEM_FILTER_NODES;
1762 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1763 filter &= ~SHOW_MEM_FILTER_NODES;
1766 struct va_format vaf;
1769 va_start(args, fmt);
1774 pr_warn("%pV", &vaf);
1779 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1780 current->comm, order, gfp_mask);
1783 if (!should_suppress_show_mem())
1788 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1789 unsigned long pages_reclaimed)
1791 /* Do not loop if specifically requested */
1792 if (gfp_mask & __GFP_NORETRY)
1796 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1797 * means __GFP_NOFAIL, but that may not be true in other
1800 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1804 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1805 * specified, then we retry until we no longer reclaim any pages
1806 * (above), or we've reclaimed an order of pages at least as
1807 * large as the allocation's order. In both cases, if the
1808 * allocation still fails, we stop retrying.
1810 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1814 * Don't let big-order allocations loop unless the caller
1815 * explicitly requests that.
1817 if (gfp_mask & __GFP_NOFAIL)
1823 static inline struct page *
1824 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1825 struct zonelist *zonelist, enum zone_type high_zoneidx,
1826 nodemask_t *nodemask, struct zone *preferred_zone,
1831 /* Acquire the OOM killer lock for the zones in zonelist */
1832 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1833 schedule_timeout_uninterruptible(1);
1838 * Go through the zonelist yet one more time, keep very high watermark
1839 * here, this is only to catch a parallel oom killing, we must fail if
1840 * we're still under heavy pressure.
1842 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1843 order, zonelist, high_zoneidx,
1844 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1845 preferred_zone, migratetype);
1849 if (!(gfp_mask & __GFP_NOFAIL)) {
1850 /* The OOM killer will not help higher order allocs */
1851 if (order > PAGE_ALLOC_COSTLY_ORDER)
1853 /* The OOM killer does not needlessly kill tasks for lowmem */
1854 if (high_zoneidx < ZONE_NORMAL)
1857 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1858 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1859 * The caller should handle page allocation failure by itself if
1860 * it specifies __GFP_THISNODE.
1861 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1863 if (gfp_mask & __GFP_THISNODE)
1866 /* Exhausted what can be done so it's blamo time */
1867 out_of_memory(zonelist, gfp_mask, order, nodemask);
1870 clear_zonelist_oom(zonelist, gfp_mask);
1874 #ifdef CONFIG_COMPACTION
1875 /* Try memory compaction for high-order allocations before reclaim */
1876 static struct page *
1877 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1878 struct zonelist *zonelist, enum zone_type high_zoneidx,
1879 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1880 int migratetype, unsigned long *did_some_progress,
1881 bool sync_migration)
1885 if (!order || compaction_deferred(preferred_zone))
1888 current->flags |= PF_MEMALLOC;
1889 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1890 nodemask, sync_migration);
1891 current->flags &= ~PF_MEMALLOC;
1892 if (*did_some_progress != COMPACT_SKIPPED) {
1894 /* Page migration frees to the PCP lists but we want merging */
1895 drain_pages(get_cpu());
1898 page = get_page_from_freelist(gfp_mask, nodemask,
1899 order, zonelist, high_zoneidx,
1900 alloc_flags, preferred_zone,
1903 preferred_zone->compact_considered = 0;
1904 preferred_zone->compact_defer_shift = 0;
1905 count_vm_event(COMPACTSUCCESS);
1910 * It's bad if compaction run occurs and fails.
1911 * The most likely reason is that pages exist,
1912 * but not enough to satisfy watermarks.
1914 count_vm_event(COMPACTFAIL);
1915 defer_compaction(preferred_zone);
1923 static inline struct page *
1924 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1925 struct zonelist *zonelist, enum zone_type high_zoneidx,
1926 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1927 int migratetype, unsigned long *did_some_progress,
1928 bool sync_migration)
1932 #endif /* CONFIG_COMPACTION */
1934 /* The really slow allocator path where we enter direct reclaim */
1935 static inline struct page *
1936 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1937 struct zonelist *zonelist, enum zone_type high_zoneidx,
1938 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1939 int migratetype, unsigned long *did_some_progress)
1941 struct page *page = NULL;
1942 struct reclaim_state reclaim_state;
1943 bool drained = false;
1947 /* We now go into synchronous reclaim */
1948 cpuset_memory_pressure_bump();
1949 current->flags |= PF_MEMALLOC;
1950 lockdep_set_current_reclaim_state(gfp_mask);
1951 reclaim_state.reclaimed_slab = 0;
1952 current->reclaim_state = &reclaim_state;
1954 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1956 current->reclaim_state = NULL;
1957 lockdep_clear_current_reclaim_state();
1958 current->flags &= ~PF_MEMALLOC;
1962 if (unlikely(!(*did_some_progress)))
1965 /* After successful reclaim, reconsider all zones for allocation */
1967 zlc_clear_zones_full(zonelist);
1970 page = get_page_from_freelist(gfp_mask, nodemask, order,
1971 zonelist, high_zoneidx,
1972 alloc_flags, preferred_zone,
1976 * If an allocation failed after direct reclaim, it could be because
1977 * pages are pinned on the per-cpu lists. Drain them and try again
1979 if (!page && !drained) {
1989 * This is called in the allocator slow-path if the allocation request is of
1990 * sufficient urgency to ignore watermarks and take other desperate measures
1992 static inline struct page *
1993 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1994 struct zonelist *zonelist, enum zone_type high_zoneidx,
1995 nodemask_t *nodemask, struct zone *preferred_zone,
2001 page = get_page_from_freelist(gfp_mask, nodemask, order,
2002 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2003 preferred_zone, migratetype);
2005 if (!page && gfp_mask & __GFP_NOFAIL)
2006 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2007 } while (!page && (gfp_mask & __GFP_NOFAIL));
2013 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2014 enum zone_type high_zoneidx,
2015 enum zone_type classzone_idx)
2020 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2021 wakeup_kswapd(zone, order, classzone_idx);
2025 gfp_to_alloc_flags(gfp_t gfp_mask)
2027 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2028 const gfp_t wait = gfp_mask & __GFP_WAIT;
2030 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2031 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2034 * The caller may dip into page reserves a bit more if the caller
2035 * cannot run direct reclaim, or if the caller has realtime scheduling
2036 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2037 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2039 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2043 * Not worth trying to allocate harder for
2044 * __GFP_NOMEMALLOC even if it can't schedule.
2046 if (!(gfp_mask & __GFP_NOMEMALLOC))
2047 alloc_flags |= ALLOC_HARDER;
2049 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2050 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2052 alloc_flags &= ~ALLOC_CPUSET;
2053 } else if (unlikely(rt_task(current)) && !in_interrupt())
2054 alloc_flags |= ALLOC_HARDER;
2056 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2057 if (!in_interrupt() &&
2058 ((current->flags & PF_MEMALLOC) ||
2059 unlikely(test_thread_flag(TIF_MEMDIE))))
2060 alloc_flags |= ALLOC_NO_WATERMARKS;
2066 static inline struct page *
2067 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2068 struct zonelist *zonelist, enum zone_type high_zoneidx,
2069 nodemask_t *nodemask, struct zone *preferred_zone,
2072 const gfp_t wait = gfp_mask & __GFP_WAIT;
2073 struct page *page = NULL;
2075 unsigned long pages_reclaimed = 0;
2076 unsigned long did_some_progress;
2077 bool sync_migration = false;
2080 * In the slowpath, we sanity check order to avoid ever trying to
2081 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2082 * be using allocators in order of preference for an area that is
2085 if (order >= MAX_ORDER) {
2086 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2091 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2092 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2093 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2094 * using a larger set of nodes after it has established that the
2095 * allowed per node queues are empty and that nodes are
2098 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2102 if (!(gfp_mask & __GFP_NO_KSWAPD))
2103 wake_all_kswapd(order, zonelist, high_zoneidx,
2104 zone_idx(preferred_zone));
2107 * OK, we're below the kswapd watermark and have kicked background
2108 * reclaim. Now things get more complex, so set up alloc_flags according
2109 * to how we want to proceed.
2111 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2114 * Find the true preferred zone if the allocation is unconstrained by
2117 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2118 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2122 /* This is the last chance, in general, before the goto nopage. */
2123 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2124 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2125 preferred_zone, migratetype);
2129 /* Allocate without watermarks if the context allows */
2130 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2131 page = __alloc_pages_high_priority(gfp_mask, order,
2132 zonelist, high_zoneidx, nodemask,
2133 preferred_zone, migratetype);
2138 /* Atomic allocations - we can't balance anything */
2142 /* Avoid recursion of direct reclaim */
2143 if (current->flags & PF_MEMALLOC)
2146 /* Avoid allocations with no watermarks from looping endlessly */
2147 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2151 * Try direct compaction. The first pass is asynchronous. Subsequent
2152 * attempts after direct reclaim are synchronous
2154 page = __alloc_pages_direct_compact(gfp_mask, order,
2155 zonelist, high_zoneidx,
2157 alloc_flags, preferred_zone,
2158 migratetype, &did_some_progress,
2162 sync_migration = true;
2164 /* Try direct reclaim and then allocating */
2165 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2166 zonelist, high_zoneidx,
2168 alloc_flags, preferred_zone,
2169 migratetype, &did_some_progress);
2174 * If we failed to make any progress reclaiming, then we are
2175 * running out of options and have to consider going OOM
2177 if (!did_some_progress) {
2178 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2179 if (oom_killer_disabled)
2181 page = __alloc_pages_may_oom(gfp_mask, order,
2182 zonelist, high_zoneidx,
2183 nodemask, preferred_zone,
2188 if (!(gfp_mask & __GFP_NOFAIL)) {
2190 * The oom killer is not called for high-order
2191 * allocations that may fail, so if no progress
2192 * is being made, there are no other options and
2193 * retrying is unlikely to help.
2195 if (order > PAGE_ALLOC_COSTLY_ORDER)
2198 * The oom killer is not called for lowmem
2199 * allocations to prevent needlessly killing
2202 if (high_zoneidx < ZONE_NORMAL)
2210 /* Check if we should retry the allocation */
2211 pages_reclaimed += did_some_progress;
2212 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2213 /* Wait for some write requests to complete then retry */
2214 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2218 * High-order allocations do not necessarily loop after
2219 * direct reclaim and reclaim/compaction depends on compaction
2220 * being called after reclaim so call directly if necessary
2222 page = __alloc_pages_direct_compact(gfp_mask, order,
2223 zonelist, high_zoneidx,
2225 alloc_flags, preferred_zone,
2226 migratetype, &did_some_progress,
2233 warn_alloc_failed(gfp_mask, order, NULL);
2236 if (kmemcheck_enabled)
2237 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2243 * This is the 'heart' of the zoned buddy allocator.
2246 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2247 struct zonelist *zonelist, nodemask_t *nodemask)
2249 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2250 struct zone *preferred_zone;
2252 int migratetype = allocflags_to_migratetype(gfp_mask);
2254 gfp_mask &= gfp_allowed_mask;
2256 lockdep_trace_alloc(gfp_mask);
2258 might_sleep_if(gfp_mask & __GFP_WAIT);
2260 if (should_fail_alloc_page(gfp_mask, order))
2264 * Check the zones suitable for the gfp_mask contain at least one
2265 * valid zone. It's possible to have an empty zonelist as a result
2266 * of GFP_THISNODE and a memoryless node
2268 if (unlikely(!zonelist->_zonerefs->zone))
2272 /* The preferred zone is used for statistics later */
2273 first_zones_zonelist(zonelist, high_zoneidx,
2274 nodemask ? : &cpuset_current_mems_allowed,
2276 if (!preferred_zone) {
2281 /* First allocation attempt */
2282 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2283 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2284 preferred_zone, migratetype);
2285 if (unlikely(!page))
2286 page = __alloc_pages_slowpath(gfp_mask, order,
2287 zonelist, high_zoneidx, nodemask,
2288 preferred_zone, migratetype);
2291 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2294 EXPORT_SYMBOL(__alloc_pages_nodemask);
2297 * Common helper functions.
2299 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2304 * __get_free_pages() returns a 32-bit address, which cannot represent
2307 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2309 page = alloc_pages(gfp_mask, order);
2312 return (unsigned long) page_address(page);
2314 EXPORT_SYMBOL(__get_free_pages);
2316 unsigned long get_zeroed_page(gfp_t gfp_mask)
2318 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2320 EXPORT_SYMBOL(get_zeroed_page);
2322 void __free_pages(struct page *page, unsigned int order)
2324 if (put_page_testzero(page)) {
2326 free_hot_cold_page(page, 0);
2328 __free_pages_ok(page, order);
2332 EXPORT_SYMBOL(__free_pages);
2334 void free_pages(unsigned long addr, unsigned int order)
2337 VM_BUG_ON(!virt_addr_valid((void *)addr));
2338 __free_pages(virt_to_page((void *)addr), order);
2342 EXPORT_SYMBOL(free_pages);
2344 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2347 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2348 unsigned long used = addr + PAGE_ALIGN(size);
2350 split_page(virt_to_page((void *)addr), order);
2351 while (used < alloc_end) {
2356 return (void *)addr;
2360 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2361 * @size: the number of bytes to allocate
2362 * @gfp_mask: GFP flags for the allocation
2364 * This function is similar to alloc_pages(), except that it allocates the
2365 * minimum number of pages to satisfy the request. alloc_pages() can only
2366 * allocate memory in power-of-two pages.
2368 * This function is also limited by MAX_ORDER.
2370 * Memory allocated by this function must be released by free_pages_exact().
2372 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2374 unsigned int order = get_order(size);
2377 addr = __get_free_pages(gfp_mask, order);
2378 return make_alloc_exact(addr, order, size);
2380 EXPORT_SYMBOL(alloc_pages_exact);
2383 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2385 * @nid: the preferred node ID where memory should be allocated
2386 * @size: the number of bytes to allocate
2387 * @gfp_mask: GFP flags for the allocation
2389 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2391 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2394 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2396 unsigned order = get_order(size);
2397 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2400 return make_alloc_exact((unsigned long)page_address(p), order, size);
2402 EXPORT_SYMBOL(alloc_pages_exact_nid);
2405 * free_pages_exact - release memory allocated via alloc_pages_exact()
2406 * @virt: the value returned by alloc_pages_exact.
2407 * @size: size of allocation, same value as passed to alloc_pages_exact().
2409 * Release the memory allocated by a previous call to alloc_pages_exact.
2411 void free_pages_exact(void *virt, size_t size)
2413 unsigned long addr = (unsigned long)virt;
2414 unsigned long end = addr + PAGE_ALIGN(size);
2416 while (addr < end) {
2421 EXPORT_SYMBOL(free_pages_exact);
2423 static unsigned int nr_free_zone_pages(int offset)
2428 /* Just pick one node, since fallback list is circular */
2429 unsigned int sum = 0;
2431 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2433 for_each_zone_zonelist(zone, z, zonelist, offset) {
2434 unsigned long size = zone->present_pages;
2435 unsigned long high = high_wmark_pages(zone);
2444 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2446 unsigned int nr_free_buffer_pages(void)
2448 return nr_free_zone_pages(gfp_zone(GFP_USER));
2450 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2453 * Amount of free RAM allocatable within all zones
2455 unsigned int nr_free_pagecache_pages(void)
2457 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2460 static inline void show_node(struct zone *zone)
2463 printk("Node %d ", zone_to_nid(zone));
2466 void si_meminfo(struct sysinfo *val)
2468 val->totalram = totalram_pages;
2470 val->freeram = global_page_state(NR_FREE_PAGES);
2471 val->bufferram = nr_blockdev_pages();
2472 val->totalhigh = totalhigh_pages;
2473 val->freehigh = nr_free_highpages();
2474 val->mem_unit = PAGE_SIZE;
2477 EXPORT_SYMBOL(si_meminfo);
2480 void si_meminfo_node(struct sysinfo *val, int nid)
2482 pg_data_t *pgdat = NODE_DATA(nid);
2484 val->totalram = pgdat->node_present_pages;
2485 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2486 #ifdef CONFIG_HIGHMEM
2487 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2488 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2494 val->mem_unit = PAGE_SIZE;
2499 * Determine whether the node should be displayed or not, depending on whether
2500 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2502 bool skip_free_areas_node(unsigned int flags, int nid)
2506 if (!(flags & SHOW_MEM_FILTER_NODES))
2510 ret = !node_isset(nid, cpuset_current_mems_allowed);
2516 #define K(x) ((x) << (PAGE_SHIFT-10))
2519 * Show free area list (used inside shift_scroll-lock stuff)
2520 * We also calculate the percentage fragmentation. We do this by counting the
2521 * memory on each free list with the exception of the first item on the list.
2522 * Suppresses nodes that are not allowed by current's cpuset if
2523 * SHOW_MEM_FILTER_NODES is passed.
2525 void show_free_areas(unsigned int filter)
2530 for_each_populated_zone(zone) {
2531 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2534 printk("%s per-cpu:\n", zone->name);
2536 for_each_online_cpu(cpu) {
2537 struct per_cpu_pageset *pageset;
2539 pageset = per_cpu_ptr(zone->pageset, cpu);
2541 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2542 cpu, pageset->pcp.high,
2543 pageset->pcp.batch, pageset->pcp.count);
2547 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2548 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2550 " dirty:%lu writeback:%lu unstable:%lu\n"
2551 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2552 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2553 global_page_state(NR_ACTIVE_ANON),
2554 global_page_state(NR_INACTIVE_ANON),
2555 global_page_state(NR_ISOLATED_ANON),
2556 global_page_state(NR_ACTIVE_FILE),
2557 global_page_state(NR_INACTIVE_FILE),
2558 global_page_state(NR_ISOLATED_FILE),
2559 global_page_state(NR_UNEVICTABLE),
2560 global_page_state(NR_FILE_DIRTY),
2561 global_page_state(NR_WRITEBACK),
2562 global_page_state(NR_UNSTABLE_NFS),
2563 global_page_state(NR_FREE_PAGES),
2564 global_page_state(NR_SLAB_RECLAIMABLE),
2565 global_page_state(NR_SLAB_UNRECLAIMABLE),
2566 global_page_state(NR_FILE_MAPPED),
2567 global_page_state(NR_SHMEM),
2568 global_page_state(NR_PAGETABLE),
2569 global_page_state(NR_BOUNCE));
2571 for_each_populated_zone(zone) {
2574 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2582 " active_anon:%lukB"
2583 " inactive_anon:%lukB"
2584 " active_file:%lukB"
2585 " inactive_file:%lukB"
2586 " unevictable:%lukB"
2587 " isolated(anon):%lukB"
2588 " isolated(file):%lukB"
2595 " slab_reclaimable:%lukB"
2596 " slab_unreclaimable:%lukB"
2597 " kernel_stack:%lukB"
2601 " writeback_tmp:%lukB"
2602 " pages_scanned:%lu"
2603 " all_unreclaimable? %s"
2606 K(zone_page_state(zone, NR_FREE_PAGES)),
2607 K(min_wmark_pages(zone)),
2608 K(low_wmark_pages(zone)),
2609 K(high_wmark_pages(zone)),
2610 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2611 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2612 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2613 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2614 K(zone_page_state(zone, NR_UNEVICTABLE)),
2615 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2616 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2617 K(zone->present_pages),
2618 K(zone_page_state(zone, NR_MLOCK)),
2619 K(zone_page_state(zone, NR_FILE_DIRTY)),
2620 K(zone_page_state(zone, NR_WRITEBACK)),
2621 K(zone_page_state(zone, NR_FILE_MAPPED)),
2622 K(zone_page_state(zone, NR_SHMEM)),
2623 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2624 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2625 zone_page_state(zone, NR_KERNEL_STACK) *
2627 K(zone_page_state(zone, NR_PAGETABLE)),
2628 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2629 K(zone_page_state(zone, NR_BOUNCE)),
2630 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2631 zone->pages_scanned,
2632 (zone->all_unreclaimable ? "yes" : "no")
2634 printk("lowmem_reserve[]:");
2635 for (i = 0; i < MAX_NR_ZONES; i++)
2636 printk(" %lu", zone->lowmem_reserve[i]);
2640 for_each_populated_zone(zone) {
2641 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2643 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2646 printk("%s: ", zone->name);
2648 spin_lock_irqsave(&zone->lock, flags);
2649 for (order = 0; order < MAX_ORDER; order++) {
2650 nr[order] = zone->free_area[order].nr_free;
2651 total += nr[order] << order;
2653 spin_unlock_irqrestore(&zone->lock, flags);
2654 for (order = 0; order < MAX_ORDER; order++)
2655 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2656 printk("= %lukB\n", K(total));
2659 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2661 show_swap_cache_info();
2664 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2666 zoneref->zone = zone;
2667 zoneref->zone_idx = zone_idx(zone);
2671 * Builds allocation fallback zone lists.
2673 * Add all populated zones of a node to the zonelist.
2675 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2676 int nr_zones, enum zone_type zone_type)
2680 BUG_ON(zone_type >= MAX_NR_ZONES);
2685 zone = pgdat->node_zones + zone_type;
2686 if (populated_zone(zone)) {
2687 zoneref_set_zone(zone,
2688 &zonelist->_zonerefs[nr_zones++]);
2689 check_highest_zone(zone_type);
2692 } while (zone_type);
2699 * 0 = automatic detection of better ordering.
2700 * 1 = order by ([node] distance, -zonetype)
2701 * 2 = order by (-zonetype, [node] distance)
2703 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2704 * the same zonelist. So only NUMA can configure this param.
2706 #define ZONELIST_ORDER_DEFAULT 0
2707 #define ZONELIST_ORDER_NODE 1
2708 #define ZONELIST_ORDER_ZONE 2
2710 /* zonelist order in the kernel.
2711 * set_zonelist_order() will set this to NODE or ZONE.
2713 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2714 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2718 /* The value user specified ....changed by config */
2719 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2720 /* string for sysctl */
2721 #define NUMA_ZONELIST_ORDER_LEN 16
2722 char numa_zonelist_order[16] = "default";
2725 * interface for configure zonelist ordering.
2726 * command line option "numa_zonelist_order"
2727 * = "[dD]efault - default, automatic configuration.
2728 * = "[nN]ode - order by node locality, then by zone within node
2729 * = "[zZ]one - order by zone, then by locality within zone
2732 static int __parse_numa_zonelist_order(char *s)
2734 if (*s == 'd' || *s == 'D') {
2735 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2736 } else if (*s == 'n' || *s == 'N') {
2737 user_zonelist_order = ZONELIST_ORDER_NODE;
2738 } else if (*s == 'z' || *s == 'Z') {
2739 user_zonelist_order = ZONELIST_ORDER_ZONE;
2742 "Ignoring invalid numa_zonelist_order value: "
2749 static __init int setup_numa_zonelist_order(char *s)
2756 ret = __parse_numa_zonelist_order(s);
2758 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2762 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2765 * sysctl handler for numa_zonelist_order
2767 int numa_zonelist_order_handler(ctl_table *table, int write,
2768 void __user *buffer, size_t *length,
2771 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2773 static DEFINE_MUTEX(zl_order_mutex);
2775 mutex_lock(&zl_order_mutex);
2777 strcpy(saved_string, (char*)table->data);
2778 ret = proc_dostring(table, write, buffer, length, ppos);
2782 int oldval = user_zonelist_order;
2783 if (__parse_numa_zonelist_order((char*)table->data)) {
2785 * bogus value. restore saved string
2787 strncpy((char*)table->data, saved_string,
2788 NUMA_ZONELIST_ORDER_LEN);
2789 user_zonelist_order = oldval;
2790 } else if (oldval != user_zonelist_order) {
2791 mutex_lock(&zonelists_mutex);
2792 build_all_zonelists(NULL);
2793 mutex_unlock(&zonelists_mutex);
2797 mutex_unlock(&zl_order_mutex);
2802 #define MAX_NODE_LOAD (nr_online_nodes)
2803 static int node_load[MAX_NUMNODES];
2806 * find_next_best_node - find the next node that should appear in a given node's fallback list
2807 * @node: node whose fallback list we're appending
2808 * @used_node_mask: nodemask_t of already used nodes
2810 * We use a number of factors to determine which is the next node that should
2811 * appear on a given node's fallback list. The node should not have appeared
2812 * already in @node's fallback list, and it should be the next closest node
2813 * according to the distance array (which contains arbitrary distance values
2814 * from each node to each node in the system), and should also prefer nodes
2815 * with no CPUs, since presumably they'll have very little allocation pressure
2816 * on them otherwise.
2817 * It returns -1 if no node is found.
2819 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2822 int min_val = INT_MAX;
2824 const struct cpumask *tmp = cpumask_of_node(0);
2826 /* Use the local node if we haven't already */
2827 if (!node_isset(node, *used_node_mask)) {
2828 node_set(node, *used_node_mask);
2832 for_each_node_state(n, N_HIGH_MEMORY) {
2834 /* Don't want a node to appear more than once */
2835 if (node_isset(n, *used_node_mask))
2838 /* Use the distance array to find the distance */
2839 val = node_distance(node, n);
2841 /* Penalize nodes under us ("prefer the next node") */
2844 /* Give preference to headless and unused nodes */
2845 tmp = cpumask_of_node(n);
2846 if (!cpumask_empty(tmp))
2847 val += PENALTY_FOR_NODE_WITH_CPUS;
2849 /* Slight preference for less loaded node */
2850 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2851 val += node_load[n];
2853 if (val < min_val) {
2860 node_set(best_node, *used_node_mask);
2867 * Build zonelists ordered by node and zones within node.
2868 * This results in maximum locality--normal zone overflows into local
2869 * DMA zone, if any--but risks exhausting DMA zone.
2871 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2874 struct zonelist *zonelist;
2876 zonelist = &pgdat->node_zonelists[0];
2877 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2879 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2881 zonelist->_zonerefs[j].zone = NULL;
2882 zonelist->_zonerefs[j].zone_idx = 0;
2886 * Build gfp_thisnode zonelists
2888 static void build_thisnode_zonelists(pg_data_t *pgdat)
2891 struct zonelist *zonelist;
2893 zonelist = &pgdat->node_zonelists[1];
2894 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2895 zonelist->_zonerefs[j].zone = NULL;
2896 zonelist->_zonerefs[j].zone_idx = 0;
2900 * Build zonelists ordered by zone and nodes within zones.
2901 * This results in conserving DMA zone[s] until all Normal memory is
2902 * exhausted, but results in overflowing to remote node while memory
2903 * may still exist in local DMA zone.
2905 static int node_order[MAX_NUMNODES];
2907 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2910 int zone_type; /* needs to be signed */
2912 struct zonelist *zonelist;
2914 zonelist = &pgdat->node_zonelists[0];
2916 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2917 for (j = 0; j < nr_nodes; j++) {
2918 node = node_order[j];
2919 z = &NODE_DATA(node)->node_zones[zone_type];
2920 if (populated_zone(z)) {
2922 &zonelist->_zonerefs[pos++]);
2923 check_highest_zone(zone_type);
2927 zonelist->_zonerefs[pos].zone = NULL;
2928 zonelist->_zonerefs[pos].zone_idx = 0;
2931 static int default_zonelist_order(void)
2934 unsigned long low_kmem_size,total_size;
2938 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2939 * If they are really small and used heavily, the system can fall
2940 * into OOM very easily.
2941 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2943 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2946 for_each_online_node(nid) {
2947 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2948 z = &NODE_DATA(nid)->node_zones[zone_type];
2949 if (populated_zone(z)) {
2950 if (zone_type < ZONE_NORMAL)
2951 low_kmem_size += z->present_pages;
2952 total_size += z->present_pages;
2953 } else if (zone_type == ZONE_NORMAL) {
2955 * If any node has only lowmem, then node order
2956 * is preferred to allow kernel allocations
2957 * locally; otherwise, they can easily infringe
2958 * on other nodes when there is an abundance of
2959 * lowmem available to allocate from.
2961 return ZONELIST_ORDER_NODE;
2965 if (!low_kmem_size || /* there are no DMA area. */
2966 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2967 return ZONELIST_ORDER_NODE;
2969 * look into each node's config.
2970 * If there is a node whose DMA/DMA32 memory is very big area on
2971 * local memory, NODE_ORDER may be suitable.
2973 average_size = total_size /
2974 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2975 for_each_online_node(nid) {
2978 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2979 z = &NODE_DATA(nid)->node_zones[zone_type];
2980 if (populated_zone(z)) {
2981 if (zone_type < ZONE_NORMAL)
2982 low_kmem_size += z->present_pages;
2983 total_size += z->present_pages;
2986 if (low_kmem_size &&
2987 total_size > average_size && /* ignore small node */
2988 low_kmem_size > total_size * 70/100)
2989 return ZONELIST_ORDER_NODE;
2991 return ZONELIST_ORDER_ZONE;
2994 static void set_zonelist_order(void)
2996 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2997 current_zonelist_order = default_zonelist_order();
2999 current_zonelist_order = user_zonelist_order;
3002 static void build_zonelists(pg_data_t *pgdat)
3006 nodemask_t used_mask;
3007 int local_node, prev_node;
3008 struct zonelist *zonelist;
3009 int order = current_zonelist_order;
3011 /* initialize zonelists */
3012 for (i = 0; i < MAX_ZONELISTS; i++) {
3013 zonelist = pgdat->node_zonelists + i;
3014 zonelist->_zonerefs[0].zone = NULL;
3015 zonelist->_zonerefs[0].zone_idx = 0;
3018 /* NUMA-aware ordering of nodes */
3019 local_node = pgdat->node_id;
3020 load = nr_online_nodes;
3021 prev_node = local_node;
3022 nodes_clear(used_mask);
3024 memset(node_order, 0, sizeof(node_order));
3027 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3028 int distance = node_distance(local_node, node);
3031 * If another node is sufficiently far away then it is better
3032 * to reclaim pages in a zone before going off node.
3034 if (distance > RECLAIM_DISTANCE)
3035 zone_reclaim_mode = 1;
3038 * We don't want to pressure a particular node.
3039 * So adding penalty to the first node in same
3040 * distance group to make it round-robin.
3042 if (distance != node_distance(local_node, prev_node))
3043 node_load[node] = load;
3047 if (order == ZONELIST_ORDER_NODE)
3048 build_zonelists_in_node_order(pgdat, node);
3050 node_order[j++] = node; /* remember order */
3053 if (order == ZONELIST_ORDER_ZONE) {
3054 /* calculate node order -- i.e., DMA last! */
3055 build_zonelists_in_zone_order(pgdat, j);
3058 build_thisnode_zonelists(pgdat);
3061 /* Construct the zonelist performance cache - see further mmzone.h */
3062 static void build_zonelist_cache(pg_data_t *pgdat)
3064 struct zonelist *zonelist;
3065 struct zonelist_cache *zlc;
3068 zonelist = &pgdat->node_zonelists[0];
3069 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3070 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3071 for (z = zonelist->_zonerefs; z->zone; z++)
3072 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3075 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3077 * Return node id of node used for "local" allocations.
3078 * I.e., first node id of first zone in arg node's generic zonelist.
3079 * Used for initializing percpu 'numa_mem', which is used primarily
3080 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3082 int local_memory_node(int node)
3086 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3087 gfp_zone(GFP_KERNEL),
3094 #else /* CONFIG_NUMA */
3096 static void set_zonelist_order(void)
3098 current_zonelist_order = ZONELIST_ORDER_ZONE;
3101 static void build_zonelists(pg_data_t *pgdat)
3103 int node, local_node;
3105 struct zonelist *zonelist;
3107 local_node = pgdat->node_id;
3109 zonelist = &pgdat->node_zonelists[0];
3110 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3113 * Now we build the zonelist so that it contains the zones
3114 * of all the other nodes.
3115 * We don't want to pressure a particular node, so when
3116 * building the zones for node N, we make sure that the
3117 * zones coming right after the local ones are those from
3118 * node N+1 (modulo N)
3120 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3121 if (!node_online(node))
3123 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3126 for (node = 0; node < local_node; node++) {
3127 if (!node_online(node))
3129 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3133 zonelist->_zonerefs[j].zone = NULL;
3134 zonelist->_zonerefs[j].zone_idx = 0;
3137 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3138 static void build_zonelist_cache(pg_data_t *pgdat)
3140 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3143 #endif /* CONFIG_NUMA */
3146 * Boot pageset table. One per cpu which is going to be used for all
3147 * zones and all nodes. The parameters will be set in such a way
3148 * that an item put on a list will immediately be handed over to
3149 * the buddy list. This is safe since pageset manipulation is done
3150 * with interrupts disabled.
3152 * The boot_pagesets must be kept even after bootup is complete for
3153 * unused processors and/or zones. They do play a role for bootstrapping
3154 * hotplugged processors.
3156 * zoneinfo_show() and maybe other functions do
3157 * not check if the processor is online before following the pageset pointer.
3158 * Other parts of the kernel may not check if the zone is available.
3160 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3161 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3162 static void setup_zone_pageset(struct zone *zone);
3165 * Global mutex to protect against size modification of zonelists
3166 * as well as to serialize pageset setup for the new populated zone.
3168 DEFINE_MUTEX(zonelists_mutex);
3170 /* return values int ....just for stop_machine() */
3171 static __init_refok int __build_all_zonelists(void *data)
3177 memset(node_load, 0, sizeof(node_load));
3179 for_each_online_node(nid) {
3180 pg_data_t *pgdat = NODE_DATA(nid);
3182 build_zonelists(pgdat);
3183 build_zonelist_cache(pgdat);
3187 * Initialize the boot_pagesets that are going to be used
3188 * for bootstrapping processors. The real pagesets for
3189 * each zone will be allocated later when the per cpu
3190 * allocator is available.
3192 * boot_pagesets are used also for bootstrapping offline
3193 * cpus if the system is already booted because the pagesets
3194 * are needed to initialize allocators on a specific cpu too.
3195 * F.e. the percpu allocator needs the page allocator which
3196 * needs the percpu allocator in order to allocate its pagesets
3197 * (a chicken-egg dilemma).
3199 for_each_possible_cpu(cpu) {
3200 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3202 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3204 * We now know the "local memory node" for each node--
3205 * i.e., the node of the first zone in the generic zonelist.
3206 * Set up numa_mem percpu variable for on-line cpus. During
3207 * boot, only the boot cpu should be on-line; we'll init the
3208 * secondary cpus' numa_mem as they come on-line. During
3209 * node/memory hotplug, we'll fixup all on-line cpus.
3211 if (cpu_online(cpu))
3212 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3220 * Called with zonelists_mutex held always
3221 * unless system_state == SYSTEM_BOOTING.
3223 void __ref build_all_zonelists(void *data)
3225 set_zonelist_order();
3227 if (system_state == SYSTEM_BOOTING) {
3228 __build_all_zonelists(NULL);
3229 mminit_verify_zonelist();
3230 cpuset_init_current_mems_allowed();
3232 /* we have to stop all cpus to guarantee there is no user
3234 #ifdef CONFIG_MEMORY_HOTPLUG
3236 setup_zone_pageset((struct zone *)data);
3238 stop_machine(__build_all_zonelists, NULL, NULL);
3239 /* cpuset refresh routine should be here */
3241 vm_total_pages = nr_free_pagecache_pages();
3243 * Disable grouping by mobility if the number of pages in the
3244 * system is too low to allow the mechanism to work. It would be
3245 * more accurate, but expensive to check per-zone. This check is
3246 * made on memory-hotadd so a system can start with mobility
3247 * disabled and enable it later
3249 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3250 page_group_by_mobility_disabled = 1;
3252 page_group_by_mobility_disabled = 0;
3254 printk("Built %i zonelists in %s order, mobility grouping %s. "
3255 "Total pages: %ld\n",
3257 zonelist_order_name[current_zonelist_order],
3258 page_group_by_mobility_disabled ? "off" : "on",
3261 printk("Policy zone: %s\n", zone_names[policy_zone]);
3266 * Helper functions to size the waitqueue hash table.
3267 * Essentially these want to choose hash table sizes sufficiently
3268 * large so that collisions trying to wait on pages are rare.
3269 * But in fact, the number of active page waitqueues on typical
3270 * systems is ridiculously low, less than 200. So this is even
3271 * conservative, even though it seems large.
3273 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3274 * waitqueues, i.e. the size of the waitq table given the number of pages.
3276 #define PAGES_PER_WAITQUEUE 256
3278 #ifndef CONFIG_MEMORY_HOTPLUG
3279 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3281 unsigned long size = 1;
3283 pages /= PAGES_PER_WAITQUEUE;
3285 while (size < pages)
3289 * Once we have dozens or even hundreds of threads sleeping
3290 * on IO we've got bigger problems than wait queue collision.
3291 * Limit the size of the wait table to a reasonable size.
3293 size = min(size, 4096UL);
3295 return max(size, 4UL);
3299 * A zone's size might be changed by hot-add, so it is not possible to determine
3300 * a suitable size for its wait_table. So we use the maximum size now.
3302 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3304 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3305 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3306 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3308 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3309 * or more by the traditional way. (See above). It equals:
3311 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3312 * ia64(16K page size) : = ( 8G + 4M)byte.
3313 * powerpc (64K page size) : = (32G +16M)byte.
3315 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3322 * This is an integer logarithm so that shifts can be used later
3323 * to extract the more random high bits from the multiplicative
3324 * hash function before the remainder is taken.
3326 static inline unsigned long wait_table_bits(unsigned long size)
3331 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3334 * Check if a pageblock contains reserved pages
3336 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3340 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3341 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3348 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3349 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3350 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3351 * higher will lead to a bigger reserve which will get freed as contiguous
3352 * blocks as reclaim kicks in
3354 static void setup_zone_migrate_reserve(struct zone *zone)
3356 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3358 unsigned long block_migratetype;
3362 * Get the start pfn, end pfn and the number of blocks to reserve
3363 * We have to be careful to be aligned to pageblock_nr_pages to
3364 * make sure that we always check pfn_valid for the first page in
3367 start_pfn = zone->zone_start_pfn;
3368 end_pfn = start_pfn + zone->spanned_pages;
3369 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3370 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3374 * Reserve blocks are generally in place to help high-order atomic
3375 * allocations that are short-lived. A min_free_kbytes value that
3376 * would result in more than 2 reserve blocks for atomic allocations
3377 * is assumed to be in place to help anti-fragmentation for the
3378 * future allocation of hugepages at runtime.
3380 reserve = min(2, reserve);
3382 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3383 if (!pfn_valid(pfn))
3385 page = pfn_to_page(pfn);
3387 /* Watch out for overlapping nodes */
3388 if (page_to_nid(page) != zone_to_nid(zone))
3391 block_migratetype = get_pageblock_migratetype(page);
3393 /* Only test what is necessary when the reserves are not met */
3396 * Blocks with reserved pages will never free, skip
3399 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3400 if (pageblock_is_reserved(pfn, block_end_pfn))
3403 /* If this block is reserved, account for it */
3404 if (block_migratetype == MIGRATE_RESERVE) {
3409 /* Suitable for reserving if this block is movable */
3410 if (block_migratetype == MIGRATE_MOVABLE) {
3411 set_pageblock_migratetype(page,
3413 move_freepages_block(zone, page,
3421 * If the reserve is met and this is a previous reserved block,
3424 if (block_migratetype == MIGRATE_RESERVE) {
3425 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3426 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3432 * Initially all pages are reserved - free ones are freed
3433 * up by free_all_bootmem() once the early boot process is
3434 * done. Non-atomic initialization, single-pass.
3436 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3437 unsigned long start_pfn, enum memmap_context context)
3440 unsigned long end_pfn = start_pfn + size;
3444 if (highest_memmap_pfn < end_pfn - 1)
3445 highest_memmap_pfn = end_pfn - 1;
3447 z = &NODE_DATA(nid)->node_zones[zone];
3448 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3450 * There can be holes in boot-time mem_map[]s
3451 * handed to this function. They do not
3452 * exist on hotplugged memory.
3454 if (context == MEMMAP_EARLY) {
3455 if (!early_pfn_valid(pfn))
3457 if (!early_pfn_in_nid(pfn, nid))
3460 page = pfn_to_page(pfn);
3461 set_page_links(page, zone, nid, pfn);
3462 mminit_verify_page_links(page, zone, nid, pfn);
3463 init_page_count(page);
3464 reset_page_mapcount(page);
3465 SetPageReserved(page);
3467 * Mark the block movable so that blocks are reserved for
3468 * movable at startup. This will force kernel allocations
3469 * to reserve their blocks rather than leaking throughout
3470 * the address space during boot when many long-lived
3471 * kernel allocations are made. Later some blocks near
3472 * the start are marked MIGRATE_RESERVE by
3473 * setup_zone_migrate_reserve()
3475 * bitmap is created for zone's valid pfn range. but memmap
3476 * can be created for invalid pages (for alignment)
3477 * check here not to call set_pageblock_migratetype() against
3480 if ((z->zone_start_pfn <= pfn)
3481 && (pfn < z->zone_start_pfn + z->spanned_pages)
3482 && !(pfn & (pageblock_nr_pages - 1)))
3483 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3485 INIT_LIST_HEAD(&page->lru);
3486 #ifdef WANT_PAGE_VIRTUAL
3487 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3488 if (!is_highmem_idx(zone))
3489 set_page_address(page, __va(pfn << PAGE_SHIFT));
3494 static void __meminit zone_init_free_lists(struct zone *zone)
3497 for_each_migratetype_order(order, t) {
3498 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3499 zone->free_area[order].nr_free = 0;
3503 #ifndef __HAVE_ARCH_MEMMAP_INIT
3504 #define memmap_init(size, nid, zone, start_pfn) \
3505 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3508 static int zone_batchsize(struct zone *zone)
3514 * The per-cpu-pages pools are set to around 1000th of the
3515 * size of the zone. But no more than 1/2 of a meg.
3517 * OK, so we don't know how big the cache is. So guess.
3519 batch = zone->present_pages / 1024;
3520 if (batch * PAGE_SIZE > 512 * 1024)
3521 batch = (512 * 1024) / PAGE_SIZE;
3522 batch /= 4; /* We effectively *= 4 below */
3527 * Clamp the batch to a 2^n - 1 value. Having a power
3528 * of 2 value was found to be more likely to have
3529 * suboptimal cache aliasing properties in some cases.
3531 * For example if 2 tasks are alternately allocating
3532 * batches of pages, one task can end up with a lot
3533 * of pages of one half of the possible page colors
3534 * and the other with pages of the other colors.
3536 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3541 /* The deferral and batching of frees should be suppressed under NOMMU
3544 * The problem is that NOMMU needs to be able to allocate large chunks
3545 * of contiguous memory as there's no hardware page translation to
3546 * assemble apparent contiguous memory from discontiguous pages.
3548 * Queueing large contiguous runs of pages for batching, however,
3549 * causes the pages to actually be freed in smaller chunks. As there
3550 * can be a significant delay between the individual batches being
3551 * recycled, this leads to the once large chunks of space being
3552 * fragmented and becoming unavailable for high-order allocations.
3558 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3560 struct per_cpu_pages *pcp;
3563 memset(p, 0, sizeof(*p));
3567 pcp->high = 6 * batch;
3568 pcp->batch = max(1UL, 1 * batch);
3569 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3570 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3574 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3575 * to the value high for the pageset p.
3578 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3581 struct per_cpu_pages *pcp;
3585 pcp->batch = max(1UL, high/4);
3586 if ((high/4) > (PAGE_SHIFT * 8))
3587 pcp->batch = PAGE_SHIFT * 8;
3590 static void setup_zone_pageset(struct zone *zone)
3594 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3596 for_each_possible_cpu(cpu) {
3597 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3599 setup_pageset(pcp, zone_batchsize(zone));
3601 if (percpu_pagelist_fraction)
3602 setup_pagelist_highmark(pcp,
3603 (zone->present_pages /
3604 percpu_pagelist_fraction));
3609 * Allocate per cpu pagesets and initialize them.
3610 * Before this call only boot pagesets were available.
3612 void __init setup_per_cpu_pageset(void)
3616 for_each_populated_zone(zone)
3617 setup_zone_pageset(zone);
3620 static noinline __init_refok
3621 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3624 struct pglist_data *pgdat = zone->zone_pgdat;
3628 * The per-page waitqueue mechanism uses hashed waitqueues
3631 zone->wait_table_hash_nr_entries =
3632 wait_table_hash_nr_entries(zone_size_pages);
3633 zone->wait_table_bits =
3634 wait_table_bits(zone->wait_table_hash_nr_entries);
3635 alloc_size = zone->wait_table_hash_nr_entries
3636 * sizeof(wait_queue_head_t);
3638 if (!slab_is_available()) {
3639 zone->wait_table = (wait_queue_head_t *)
3640 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3643 * This case means that a zone whose size was 0 gets new memory
3644 * via memory hot-add.
3645 * But it may be the case that a new node was hot-added. In
3646 * this case vmalloc() will not be able to use this new node's
3647 * memory - this wait_table must be initialized to use this new
3648 * node itself as well.
3649 * To use this new node's memory, further consideration will be
3652 zone->wait_table = vmalloc(alloc_size);
3654 if (!zone->wait_table)
3657 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3658 init_waitqueue_head(zone->wait_table + i);
3663 static int __zone_pcp_update(void *data)
3665 struct zone *zone = data;
3667 unsigned long batch = zone_batchsize(zone), flags;
3669 for_each_possible_cpu(cpu) {
3670 struct per_cpu_pageset *pset;
3671 struct per_cpu_pages *pcp;
3673 pset = per_cpu_ptr(zone->pageset, cpu);
3676 local_irq_save(flags);
3677 free_pcppages_bulk(zone, pcp->count, pcp);
3678 setup_pageset(pset, batch);
3679 local_irq_restore(flags);
3684 void zone_pcp_update(struct zone *zone)
3686 stop_machine(__zone_pcp_update, zone, NULL);
3689 static __meminit void zone_pcp_init(struct zone *zone)
3692 * per cpu subsystem is not up at this point. The following code
3693 * relies on the ability of the linker to provide the
3694 * offset of a (static) per cpu variable into the per cpu area.
3696 zone->pageset = &boot_pageset;
3698 if (zone->present_pages)
3699 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3700 zone->name, zone->present_pages,
3701 zone_batchsize(zone));
3704 __meminit int init_currently_empty_zone(struct zone *zone,
3705 unsigned long zone_start_pfn,
3707 enum memmap_context context)
3709 struct pglist_data *pgdat = zone->zone_pgdat;
3711 ret = zone_wait_table_init(zone, size);
3714 pgdat->nr_zones = zone_idx(zone) + 1;
3716 zone->zone_start_pfn = zone_start_pfn;
3718 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3719 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3721 (unsigned long)zone_idx(zone),
3722 zone_start_pfn, (zone_start_pfn + size));
3724 zone_init_free_lists(zone);
3729 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
3730 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3732 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3733 * Architectures may implement their own version but if add_active_range()
3734 * was used and there are no special requirements, this is a convenient
3737 int __meminit __early_pfn_to_nid(unsigned long pfn)
3739 unsigned long start_pfn, end_pfn;
3742 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
3743 if (start_pfn <= pfn && pfn < end_pfn)
3745 /* This is a memory hole */
3748 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3750 int __meminit early_pfn_to_nid(unsigned long pfn)
3754 nid = __early_pfn_to_nid(pfn);
3757 /* just returns 0 */
3761 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3762 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3766 nid = __early_pfn_to_nid(pfn);
3767 if (nid >= 0 && nid != node)
3774 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3775 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3776 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3778 * If an architecture guarantees that all ranges registered with
3779 * add_active_ranges() contain no holes and may be freed, this
3780 * this function may be used instead of calling free_bootmem() manually.
3782 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
3784 unsigned long start_pfn, end_pfn;
3787 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
3788 start_pfn = min(start_pfn, max_low_pfn);
3789 end_pfn = min(end_pfn, max_low_pfn);
3791 if (start_pfn < end_pfn)
3792 free_bootmem_node(NODE_DATA(this_nid),
3793 PFN_PHYS(start_pfn),
3794 (end_pfn - start_pfn) << PAGE_SHIFT);
3798 int __init add_from_early_node_map(struct range *range, int az,
3799 int nr_range, int nid)
3801 unsigned long start_pfn, end_pfn;
3804 /* need to go over early_node_map to find out good range for node */
3805 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL)
3806 nr_range = add_range(range, az, nr_range, start_pfn, end_pfn);
3811 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3812 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3814 * If an architecture guarantees that all ranges registered with
3815 * add_active_ranges() contain no holes and may be freed, this
3816 * function may be used instead of calling memory_present() manually.
3818 void __init sparse_memory_present_with_active_regions(int nid)
3820 unsigned long start_pfn, end_pfn;
3823 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
3824 memory_present(this_nid, start_pfn, end_pfn);
3828 * get_pfn_range_for_nid - Return the start and end page frames for a node
3829 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3830 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3831 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3833 * It returns the start and end page frame of a node based on information
3834 * provided by an arch calling add_active_range(). If called for a node
3835 * with no available memory, a warning is printed and the start and end
3838 void __meminit get_pfn_range_for_nid(unsigned int nid,
3839 unsigned long *start_pfn, unsigned long *end_pfn)
3841 unsigned long this_start_pfn, this_end_pfn;
3847 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
3848 *start_pfn = min(*start_pfn, this_start_pfn);
3849 *end_pfn = max(*end_pfn, this_end_pfn);
3852 if (*start_pfn == -1UL)
3857 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3858 * assumption is made that zones within a node are ordered in monotonic
3859 * increasing memory addresses so that the "highest" populated zone is used
3861 static void __init find_usable_zone_for_movable(void)
3864 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3865 if (zone_index == ZONE_MOVABLE)
3868 if (arch_zone_highest_possible_pfn[zone_index] >
3869 arch_zone_lowest_possible_pfn[zone_index])
3873 VM_BUG_ON(zone_index == -1);
3874 movable_zone = zone_index;
3878 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3879 * because it is sized independent of architecture. Unlike the other zones,
3880 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3881 * in each node depending on the size of each node and how evenly kernelcore
3882 * is distributed. This helper function adjusts the zone ranges
3883 * provided by the architecture for a given node by using the end of the
3884 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3885 * zones within a node are in order of monotonic increases memory addresses
3887 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3888 unsigned long zone_type,
3889 unsigned long node_start_pfn,
3890 unsigned long node_end_pfn,
3891 unsigned long *zone_start_pfn,
3892 unsigned long *zone_end_pfn)
3894 /* Only adjust if ZONE_MOVABLE is on this node */
3895 if (zone_movable_pfn[nid]) {
3896 /* Size ZONE_MOVABLE */
3897 if (zone_type == ZONE_MOVABLE) {
3898 *zone_start_pfn = zone_movable_pfn[nid];
3899 *zone_end_pfn = min(node_end_pfn,
3900 arch_zone_highest_possible_pfn[movable_zone]);
3902 /* Adjust for ZONE_MOVABLE starting within this range */
3903 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3904 *zone_end_pfn > zone_movable_pfn[nid]) {
3905 *zone_end_pfn = zone_movable_pfn[nid];
3907 /* Check if this whole range is within ZONE_MOVABLE */
3908 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3909 *zone_start_pfn = *zone_end_pfn;
3914 * Return the number of pages a zone spans in a node, including holes
3915 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3917 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3918 unsigned long zone_type,
3919 unsigned long *ignored)
3921 unsigned long node_start_pfn, node_end_pfn;
3922 unsigned long zone_start_pfn, zone_end_pfn;
3924 /* Get the start and end of the node and zone */
3925 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3926 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3927 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3928 adjust_zone_range_for_zone_movable(nid, zone_type,
3929 node_start_pfn, node_end_pfn,
3930 &zone_start_pfn, &zone_end_pfn);
3932 /* Check that this node has pages within the zone's required range */
3933 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3936 /* Move the zone boundaries inside the node if necessary */
3937 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3938 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3940 /* Return the spanned pages */
3941 return zone_end_pfn - zone_start_pfn;
3945 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3946 * then all holes in the requested range will be accounted for.
3948 unsigned long __meminit __absent_pages_in_range(int nid,
3949 unsigned long range_start_pfn,
3950 unsigned long range_end_pfn)
3952 unsigned long nr_absent = range_end_pfn - range_start_pfn;
3953 unsigned long start_pfn, end_pfn;
3956 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
3957 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
3958 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
3959 nr_absent -= end_pfn - start_pfn;
3965 * absent_pages_in_range - Return number of page frames in holes within a range
3966 * @start_pfn: The start PFN to start searching for holes
3967 * @end_pfn: The end PFN to stop searching for holes
3969 * It returns the number of pages frames in memory holes within a range.
3971 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3972 unsigned long end_pfn)
3974 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3977 /* Return the number of page frames in holes in a zone on a node */
3978 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3979 unsigned long zone_type,
3980 unsigned long *ignored)
3982 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
3983 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
3984 unsigned long node_start_pfn, node_end_pfn;
3985 unsigned long zone_start_pfn, zone_end_pfn;
3987 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3988 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
3989 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
3991 adjust_zone_range_for_zone_movable(nid, zone_type,
3992 node_start_pfn, node_end_pfn,
3993 &zone_start_pfn, &zone_end_pfn);
3994 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3997 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
3998 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3999 unsigned long zone_type,
4000 unsigned long *zones_size)
4002 return zones_size[zone_type];
4005 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4006 unsigned long zone_type,
4007 unsigned long *zholes_size)
4012 return zholes_size[zone_type];
4015 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4017 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4018 unsigned long *zones_size, unsigned long *zholes_size)
4020 unsigned long realtotalpages, totalpages = 0;
4023 for (i = 0; i < MAX_NR_ZONES; i++)
4024 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4026 pgdat->node_spanned_pages = totalpages;
4028 realtotalpages = totalpages;
4029 for (i = 0; i < MAX_NR_ZONES; i++)
4031 zone_absent_pages_in_node(pgdat->node_id, i,
4033 pgdat->node_present_pages = realtotalpages;
4034 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4038 #ifndef CONFIG_SPARSEMEM
4040 * Calculate the size of the zone->blockflags rounded to an unsigned long
4041 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4042 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4043 * round what is now in bits to nearest long in bits, then return it in
4046 static unsigned long __init usemap_size(unsigned long zonesize)
4048 unsigned long usemapsize;
4050 usemapsize = roundup(zonesize, pageblock_nr_pages);
4051 usemapsize = usemapsize >> pageblock_order;
4052 usemapsize *= NR_PAGEBLOCK_BITS;
4053 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4055 return usemapsize / 8;
4058 static void __init setup_usemap(struct pglist_data *pgdat,
4059 struct zone *zone, unsigned long zonesize)
4061 unsigned long usemapsize = usemap_size(zonesize);
4062 zone->pageblock_flags = NULL;
4064 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4068 static inline void setup_usemap(struct pglist_data *pgdat,
4069 struct zone *zone, unsigned long zonesize) {}
4070 #endif /* CONFIG_SPARSEMEM */
4072 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4074 /* Return a sensible default order for the pageblock size. */
4075 static inline int pageblock_default_order(void)
4077 if (HPAGE_SHIFT > PAGE_SHIFT)
4078 return HUGETLB_PAGE_ORDER;
4083 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4084 static inline void __init set_pageblock_order(unsigned int order)
4086 /* Check that pageblock_nr_pages has not already been setup */
4087 if (pageblock_order)
4091 * Assume the largest contiguous order of interest is a huge page.
4092 * This value may be variable depending on boot parameters on IA64
4094 pageblock_order = order;
4096 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4099 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4100 * and pageblock_default_order() are unused as pageblock_order is set
4101 * at compile-time. See include/linux/pageblock-flags.h for the values of
4102 * pageblock_order based on the kernel config
4104 static inline int pageblock_default_order(unsigned int order)
4108 #define set_pageblock_order(x) do {} while (0)
4110 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4113 * Set up the zone data structures:
4114 * - mark all pages reserved
4115 * - mark all memory queues empty
4116 * - clear the memory bitmaps
4118 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4119 unsigned long *zones_size, unsigned long *zholes_size)
4122 int nid = pgdat->node_id;
4123 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4126 pgdat_resize_init(pgdat);
4127 pgdat->nr_zones = 0;
4128 init_waitqueue_head(&pgdat->kswapd_wait);
4129 pgdat->kswapd_max_order = 0;
4130 pgdat_page_cgroup_init(pgdat);
4132 for (j = 0; j < MAX_NR_ZONES; j++) {
4133 struct zone *zone = pgdat->node_zones + j;
4134 unsigned long size, realsize, memmap_pages;
4137 size = zone_spanned_pages_in_node(nid, j, zones_size);
4138 realsize = size - zone_absent_pages_in_node(nid, j,
4142 * Adjust realsize so that it accounts for how much memory
4143 * is used by this zone for memmap. This affects the watermark
4144 * and per-cpu initialisations
4147 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4148 if (realsize >= memmap_pages) {
4149 realsize -= memmap_pages;
4152 " %s zone: %lu pages used for memmap\n",
4153 zone_names[j], memmap_pages);
4156 " %s zone: %lu pages exceeds realsize %lu\n",
4157 zone_names[j], memmap_pages, realsize);
4159 /* Account for reserved pages */
4160 if (j == 0 && realsize > dma_reserve) {
4161 realsize -= dma_reserve;
4162 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4163 zone_names[0], dma_reserve);
4166 if (!is_highmem_idx(j))
4167 nr_kernel_pages += realsize;
4168 nr_all_pages += realsize;
4170 zone->spanned_pages = size;
4171 zone->present_pages = realsize;
4174 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4176 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4178 zone->name = zone_names[j];
4179 spin_lock_init(&zone->lock);
4180 spin_lock_init(&zone->lru_lock);
4181 zone_seqlock_init(zone);
4182 zone->zone_pgdat = pgdat;
4184 zone_pcp_init(zone);
4186 INIT_LIST_HEAD(&zone->lru[l].list);
4187 zone->reclaim_stat.recent_rotated[0] = 0;
4188 zone->reclaim_stat.recent_rotated[1] = 0;
4189 zone->reclaim_stat.recent_scanned[0] = 0;
4190 zone->reclaim_stat.recent_scanned[1] = 0;
4191 zap_zone_vm_stats(zone);
4196 set_pageblock_order(pageblock_default_order());
4197 setup_usemap(pgdat, zone, size);
4198 ret = init_currently_empty_zone(zone, zone_start_pfn,
4199 size, MEMMAP_EARLY);
4201 memmap_init(size, nid, j, zone_start_pfn);
4202 zone_start_pfn += size;
4206 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4208 /* Skip empty nodes */
4209 if (!pgdat->node_spanned_pages)
4212 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4213 /* ia64 gets its own node_mem_map, before this, without bootmem */
4214 if (!pgdat->node_mem_map) {
4215 unsigned long size, start, end;
4219 * The zone's endpoints aren't required to be MAX_ORDER
4220 * aligned but the node_mem_map endpoints must be in order
4221 * for the buddy allocator to function correctly.
4223 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4224 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4225 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4226 size = (end - start) * sizeof(struct page);
4227 map = alloc_remap(pgdat->node_id, size);
4229 map = alloc_bootmem_node_nopanic(pgdat, size);
4230 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4232 #ifndef CONFIG_NEED_MULTIPLE_NODES
4234 * With no DISCONTIG, the global mem_map is just set as node 0's
4236 if (pgdat == NODE_DATA(0)) {
4237 mem_map = NODE_DATA(0)->node_mem_map;
4238 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4239 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4240 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4241 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4244 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4247 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4248 unsigned long node_start_pfn, unsigned long *zholes_size)
4250 pg_data_t *pgdat = NODE_DATA(nid);
4252 pgdat->node_id = nid;
4253 pgdat->node_start_pfn = node_start_pfn;
4254 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4256 alloc_node_mem_map(pgdat);
4257 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4258 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4259 nid, (unsigned long)pgdat,
4260 (unsigned long)pgdat->node_mem_map);
4263 free_area_init_core(pgdat, zones_size, zholes_size);
4266 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4268 #if MAX_NUMNODES > 1
4270 * Figure out the number of possible node ids.
4272 static void __init setup_nr_node_ids(void)
4275 unsigned int highest = 0;
4277 for_each_node_mask(node, node_possible_map)
4279 nr_node_ids = highest + 1;
4282 static inline void setup_nr_node_ids(void)
4288 * node_map_pfn_alignment - determine the maximum internode alignment
4290 * This function should be called after node map is populated and sorted.
4291 * It calculates the maximum power of two alignment which can distinguish
4294 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4295 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4296 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4297 * shifted, 1GiB is enough and this function will indicate so.
4299 * This is used to test whether pfn -> nid mapping of the chosen memory
4300 * model has fine enough granularity to avoid incorrect mapping for the
4301 * populated node map.
4303 * Returns the determined alignment in pfn's. 0 if there is no alignment
4304 * requirement (single node).
4306 unsigned long __init node_map_pfn_alignment(void)
4308 unsigned long accl_mask = 0, last_end = 0;
4309 unsigned long start, end, mask;
4313 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4314 if (!start || last_nid < 0 || last_nid == nid) {
4321 * Start with a mask granular enough to pin-point to the
4322 * start pfn and tick off bits one-by-one until it becomes
4323 * too coarse to separate the current node from the last.
4325 mask = ~((1 << __ffs(start)) - 1);
4326 while (mask && last_end <= (start & (mask << 1)))
4329 /* accumulate all internode masks */
4333 /* convert mask to number of pages */
4334 return ~accl_mask + 1;
4337 /* Find the lowest pfn for a node */
4338 static unsigned long __init find_min_pfn_for_node(int nid)
4340 unsigned long min_pfn = ULONG_MAX;
4341 unsigned long start_pfn;
4344 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4345 min_pfn = min(min_pfn, start_pfn);
4347 if (min_pfn == ULONG_MAX) {
4349 "Could not find start_pfn for node %d\n", nid);
4357 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4359 * It returns the minimum PFN based on information provided via
4360 * add_active_range().
4362 unsigned long __init find_min_pfn_with_active_regions(void)
4364 return find_min_pfn_for_node(MAX_NUMNODES);
4368 * early_calculate_totalpages()
4369 * Sum pages in active regions for movable zone.
4370 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4372 static unsigned long __init early_calculate_totalpages(void)
4374 unsigned long totalpages = 0;
4375 unsigned long start_pfn, end_pfn;
4378 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4379 unsigned long pages = end_pfn - start_pfn;
4381 totalpages += pages;
4383 node_set_state(nid, N_HIGH_MEMORY);
4389 * Find the PFN the Movable zone begins in each node. Kernel memory
4390 * is spread evenly between nodes as long as the nodes have enough
4391 * memory. When they don't, some nodes will have more kernelcore than
4394 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4397 unsigned long usable_startpfn;
4398 unsigned long kernelcore_node, kernelcore_remaining;
4399 /* save the state before borrow the nodemask */
4400 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4401 unsigned long totalpages = early_calculate_totalpages();
4402 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4405 * If movablecore was specified, calculate what size of
4406 * kernelcore that corresponds so that memory usable for
4407 * any allocation type is evenly spread. If both kernelcore
4408 * and movablecore are specified, then the value of kernelcore
4409 * will be used for required_kernelcore if it's greater than
4410 * what movablecore would have allowed.
4412 if (required_movablecore) {
4413 unsigned long corepages;
4416 * Round-up so that ZONE_MOVABLE is at least as large as what
4417 * was requested by the user
4419 required_movablecore =
4420 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4421 corepages = totalpages - required_movablecore;
4423 required_kernelcore = max(required_kernelcore, corepages);
4426 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4427 if (!required_kernelcore)
4430 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4431 find_usable_zone_for_movable();
4432 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4435 /* Spread kernelcore memory as evenly as possible throughout nodes */
4436 kernelcore_node = required_kernelcore / usable_nodes;
4437 for_each_node_state(nid, N_HIGH_MEMORY) {
4438 unsigned long start_pfn, end_pfn;
4441 * Recalculate kernelcore_node if the division per node
4442 * now exceeds what is necessary to satisfy the requested
4443 * amount of memory for the kernel
4445 if (required_kernelcore < kernelcore_node)
4446 kernelcore_node = required_kernelcore / usable_nodes;
4449 * As the map is walked, we track how much memory is usable
4450 * by the kernel using kernelcore_remaining. When it is
4451 * 0, the rest of the node is usable by ZONE_MOVABLE
4453 kernelcore_remaining = kernelcore_node;
4455 /* Go through each range of PFNs within this node */
4456 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4457 unsigned long size_pages;
4459 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4460 if (start_pfn >= end_pfn)
4463 /* Account for what is only usable for kernelcore */
4464 if (start_pfn < usable_startpfn) {
4465 unsigned long kernel_pages;
4466 kernel_pages = min(end_pfn, usable_startpfn)
4469 kernelcore_remaining -= min(kernel_pages,
4470 kernelcore_remaining);
4471 required_kernelcore -= min(kernel_pages,
4472 required_kernelcore);
4474 /* Continue if range is now fully accounted */
4475 if (end_pfn <= usable_startpfn) {
4478 * Push zone_movable_pfn to the end so
4479 * that if we have to rebalance
4480 * kernelcore across nodes, we will
4481 * not double account here
4483 zone_movable_pfn[nid] = end_pfn;
4486 start_pfn = usable_startpfn;
4490 * The usable PFN range for ZONE_MOVABLE is from
4491 * start_pfn->end_pfn. Calculate size_pages as the
4492 * number of pages used as kernelcore
4494 size_pages = end_pfn - start_pfn;
4495 if (size_pages > kernelcore_remaining)
4496 size_pages = kernelcore_remaining;
4497 zone_movable_pfn[nid] = start_pfn + size_pages;
4500 * Some kernelcore has been met, update counts and
4501 * break if the kernelcore for this node has been
4504 required_kernelcore -= min(required_kernelcore,
4506 kernelcore_remaining -= size_pages;
4507 if (!kernelcore_remaining)
4513 * If there is still required_kernelcore, we do another pass with one
4514 * less node in the count. This will push zone_movable_pfn[nid] further
4515 * along on the nodes that still have memory until kernelcore is
4519 if (usable_nodes && required_kernelcore > usable_nodes)
4522 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4523 for (nid = 0; nid < MAX_NUMNODES; nid++)
4524 zone_movable_pfn[nid] =
4525 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4528 /* restore the node_state */
4529 node_states[N_HIGH_MEMORY] = saved_node_state;
4532 /* Any regular memory on that node ? */
4533 static void check_for_regular_memory(pg_data_t *pgdat)
4535 #ifdef CONFIG_HIGHMEM
4536 enum zone_type zone_type;
4538 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4539 struct zone *zone = &pgdat->node_zones[zone_type];
4540 if (zone->present_pages)
4541 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4547 * free_area_init_nodes - Initialise all pg_data_t and zone data
4548 * @max_zone_pfn: an array of max PFNs for each zone
4550 * This will call free_area_init_node() for each active node in the system.
4551 * Using the page ranges provided by add_active_range(), the size of each
4552 * zone in each node and their holes is calculated. If the maximum PFN
4553 * between two adjacent zones match, it is assumed that the zone is empty.
4554 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4555 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4556 * starts where the previous one ended. For example, ZONE_DMA32 starts
4557 * at arch_max_dma_pfn.
4559 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4561 unsigned long start_pfn, end_pfn;
4564 /* Record where the zone boundaries are */
4565 memset(arch_zone_lowest_possible_pfn, 0,
4566 sizeof(arch_zone_lowest_possible_pfn));
4567 memset(arch_zone_highest_possible_pfn, 0,
4568 sizeof(arch_zone_highest_possible_pfn));
4569 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4570 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4571 for (i = 1; i < MAX_NR_ZONES; i++) {
4572 if (i == ZONE_MOVABLE)
4574 arch_zone_lowest_possible_pfn[i] =
4575 arch_zone_highest_possible_pfn[i-1];
4576 arch_zone_highest_possible_pfn[i] =
4577 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4579 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4580 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4582 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4583 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4584 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4586 /* Print out the zone ranges */
4587 printk("Zone PFN ranges:\n");
4588 for (i = 0; i < MAX_NR_ZONES; i++) {
4589 if (i == ZONE_MOVABLE)
4591 printk(" %-8s ", zone_names[i]);
4592 if (arch_zone_lowest_possible_pfn[i] ==
4593 arch_zone_highest_possible_pfn[i])
4596 printk("%0#10lx -> %0#10lx\n",
4597 arch_zone_lowest_possible_pfn[i],
4598 arch_zone_highest_possible_pfn[i]);
4601 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4602 printk("Movable zone start PFN for each node\n");
4603 for (i = 0; i < MAX_NUMNODES; i++) {
4604 if (zone_movable_pfn[i])
4605 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4608 /* Print out the early_node_map[] */
4609 printk("Early memory PFN ranges\n");
4610 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4611 printk(" %3d: %0#10lx -> %0#10lx\n", nid, start_pfn, end_pfn);
4613 /* Initialise every node */
4614 mminit_verify_pageflags_layout();
4615 setup_nr_node_ids();
4616 for_each_online_node(nid) {
4617 pg_data_t *pgdat = NODE_DATA(nid);
4618 free_area_init_node(nid, NULL,
4619 find_min_pfn_for_node(nid), NULL);
4621 /* Any memory on that node */
4622 if (pgdat->node_present_pages)
4623 node_set_state(nid, N_HIGH_MEMORY);
4624 check_for_regular_memory(pgdat);
4628 static int __init cmdline_parse_core(char *p, unsigned long *core)
4630 unsigned long long coremem;
4634 coremem = memparse(p, &p);
4635 *core = coremem >> PAGE_SHIFT;
4637 /* Paranoid check that UL is enough for the coremem value */
4638 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4644 * kernelcore=size sets the amount of memory for use for allocations that
4645 * cannot be reclaimed or migrated.
4647 static int __init cmdline_parse_kernelcore(char *p)
4649 return cmdline_parse_core(p, &required_kernelcore);
4653 * movablecore=size sets the amount of memory for use for allocations that
4654 * can be reclaimed or migrated.
4656 static int __init cmdline_parse_movablecore(char *p)
4658 return cmdline_parse_core(p, &required_movablecore);
4661 early_param("kernelcore", cmdline_parse_kernelcore);
4662 early_param("movablecore", cmdline_parse_movablecore);
4664 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4667 * set_dma_reserve - set the specified number of pages reserved in the first zone
4668 * @new_dma_reserve: The number of pages to mark reserved
4670 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4671 * In the DMA zone, a significant percentage may be consumed by kernel image
4672 * and other unfreeable allocations which can skew the watermarks badly. This
4673 * function may optionally be used to account for unfreeable pages in the
4674 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4675 * smaller per-cpu batchsize.
4677 void __init set_dma_reserve(unsigned long new_dma_reserve)
4679 dma_reserve = new_dma_reserve;
4682 void __init free_area_init(unsigned long *zones_size)
4684 free_area_init_node(0, zones_size,
4685 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4688 static int page_alloc_cpu_notify(struct notifier_block *self,
4689 unsigned long action, void *hcpu)
4691 int cpu = (unsigned long)hcpu;
4693 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4697 * Spill the event counters of the dead processor
4698 * into the current processors event counters.
4699 * This artificially elevates the count of the current
4702 vm_events_fold_cpu(cpu);
4705 * Zero the differential counters of the dead processor
4706 * so that the vm statistics are consistent.
4708 * This is only okay since the processor is dead and cannot
4709 * race with what we are doing.
4711 refresh_cpu_vm_stats(cpu);
4716 void __init page_alloc_init(void)
4718 hotcpu_notifier(page_alloc_cpu_notify, 0);
4722 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4723 * or min_free_kbytes changes.
4725 static void calculate_totalreserve_pages(void)
4727 struct pglist_data *pgdat;
4728 unsigned long reserve_pages = 0;
4729 enum zone_type i, j;
4731 for_each_online_pgdat(pgdat) {
4732 for (i = 0; i < MAX_NR_ZONES; i++) {
4733 struct zone *zone = pgdat->node_zones + i;
4734 unsigned long max = 0;
4736 /* Find valid and maximum lowmem_reserve in the zone */
4737 for (j = i; j < MAX_NR_ZONES; j++) {
4738 if (zone->lowmem_reserve[j] > max)
4739 max = zone->lowmem_reserve[j];
4742 /* we treat the high watermark as reserved pages. */
4743 max += high_wmark_pages(zone);
4745 if (max > zone->present_pages)
4746 max = zone->present_pages;
4747 reserve_pages += max;
4750 totalreserve_pages = reserve_pages;
4754 * setup_per_zone_lowmem_reserve - called whenever
4755 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4756 * has a correct pages reserved value, so an adequate number of
4757 * pages are left in the zone after a successful __alloc_pages().
4759 static void setup_per_zone_lowmem_reserve(void)
4761 struct pglist_data *pgdat;
4762 enum zone_type j, idx;
4764 for_each_online_pgdat(pgdat) {
4765 for (j = 0; j < MAX_NR_ZONES; j++) {
4766 struct zone *zone = pgdat->node_zones + j;
4767 unsigned long present_pages = zone->present_pages;
4769 zone->lowmem_reserve[j] = 0;
4773 struct zone *lower_zone;
4777 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4778 sysctl_lowmem_reserve_ratio[idx] = 1;
4780 lower_zone = pgdat->node_zones + idx;
4781 lower_zone->lowmem_reserve[j] = present_pages /
4782 sysctl_lowmem_reserve_ratio[idx];
4783 present_pages += lower_zone->present_pages;
4788 /* update totalreserve_pages */
4789 calculate_totalreserve_pages();
4793 * setup_per_zone_wmarks - called when min_free_kbytes changes
4794 * or when memory is hot-{added|removed}
4796 * Ensures that the watermark[min,low,high] values for each zone are set
4797 * correctly with respect to min_free_kbytes.
4799 void setup_per_zone_wmarks(void)
4801 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4802 unsigned long lowmem_pages = 0;
4804 unsigned long flags;
4806 /* Calculate total number of !ZONE_HIGHMEM pages */
4807 for_each_zone(zone) {
4808 if (!is_highmem(zone))
4809 lowmem_pages += zone->present_pages;
4812 for_each_zone(zone) {
4815 spin_lock_irqsave(&zone->lock, flags);
4816 tmp = (u64)pages_min * zone->present_pages;
4817 do_div(tmp, lowmem_pages);
4818 if (is_highmem(zone)) {
4820 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4821 * need highmem pages, so cap pages_min to a small
4824 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4825 * deltas controls asynch page reclaim, and so should
4826 * not be capped for highmem.
4830 min_pages = zone->present_pages / 1024;
4831 if (min_pages < SWAP_CLUSTER_MAX)
4832 min_pages = SWAP_CLUSTER_MAX;
4833 if (min_pages > 128)
4835 zone->watermark[WMARK_MIN] = min_pages;
4838 * If it's a lowmem zone, reserve a number of pages
4839 * proportionate to the zone's size.
4841 zone->watermark[WMARK_MIN] = tmp;
4844 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4845 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4846 setup_zone_migrate_reserve(zone);
4847 spin_unlock_irqrestore(&zone->lock, flags);
4850 /* update totalreserve_pages */
4851 calculate_totalreserve_pages();
4855 * The inactive anon list should be small enough that the VM never has to
4856 * do too much work, but large enough that each inactive page has a chance
4857 * to be referenced again before it is swapped out.
4859 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4860 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4861 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4862 * the anonymous pages are kept on the inactive list.
4865 * memory ratio inactive anon
4866 * -------------------------------------
4875 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
4877 unsigned int gb, ratio;
4879 /* Zone size in gigabytes */
4880 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4882 ratio = int_sqrt(10 * gb);
4886 zone->inactive_ratio = ratio;
4889 static void __meminit setup_per_zone_inactive_ratio(void)
4894 calculate_zone_inactive_ratio(zone);
4898 * Initialise min_free_kbytes.
4900 * For small machines we want it small (128k min). For large machines
4901 * we want it large (64MB max). But it is not linear, because network
4902 * bandwidth does not increase linearly with machine size. We use
4904 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4905 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4921 int __meminit init_per_zone_wmark_min(void)
4923 unsigned long lowmem_kbytes;
4925 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4927 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4928 if (min_free_kbytes < 128)
4929 min_free_kbytes = 128;
4930 if (min_free_kbytes > 65536)
4931 min_free_kbytes = 65536;
4932 setup_per_zone_wmarks();
4933 refresh_zone_stat_thresholds();
4934 setup_per_zone_lowmem_reserve();
4935 setup_per_zone_inactive_ratio();
4938 module_init(init_per_zone_wmark_min)
4941 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4942 * that we can call two helper functions whenever min_free_kbytes
4945 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4946 void __user *buffer, size_t *length, loff_t *ppos)
4948 proc_dointvec(table, write, buffer, length, ppos);
4950 setup_per_zone_wmarks();
4955 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4956 void __user *buffer, size_t *length, loff_t *ppos)
4961 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4966 zone->min_unmapped_pages = (zone->present_pages *
4967 sysctl_min_unmapped_ratio) / 100;
4971 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4972 void __user *buffer, size_t *length, loff_t *ppos)
4977 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4982 zone->min_slab_pages = (zone->present_pages *
4983 sysctl_min_slab_ratio) / 100;
4989 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4990 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4991 * whenever sysctl_lowmem_reserve_ratio changes.
4993 * The reserve ratio obviously has absolutely no relation with the
4994 * minimum watermarks. The lowmem reserve ratio can only make sense
4995 * if in function of the boot time zone sizes.
4997 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4998 void __user *buffer, size_t *length, loff_t *ppos)
5000 proc_dointvec_minmax(table, write, buffer, length, ppos);
5001 setup_per_zone_lowmem_reserve();
5006 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5007 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5008 * can have before it gets flushed back to buddy allocator.
5011 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5012 void __user *buffer, size_t *length, loff_t *ppos)
5018 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5019 if (!write || (ret == -EINVAL))
5021 for_each_populated_zone(zone) {
5022 for_each_possible_cpu(cpu) {
5024 high = zone->present_pages / percpu_pagelist_fraction;
5025 setup_pagelist_highmark(
5026 per_cpu_ptr(zone->pageset, cpu), high);
5032 int hashdist = HASHDIST_DEFAULT;
5035 static int __init set_hashdist(char *str)
5039 hashdist = simple_strtoul(str, &str, 0);
5042 __setup("hashdist=", set_hashdist);
5046 * allocate a large system hash table from bootmem
5047 * - it is assumed that the hash table must contain an exact power-of-2
5048 * quantity of entries
5049 * - limit is the number of hash buckets, not the total allocation size
5051 void *__init alloc_large_system_hash(const char *tablename,
5052 unsigned long bucketsize,
5053 unsigned long numentries,
5056 unsigned int *_hash_shift,
5057 unsigned int *_hash_mask,
5058 unsigned long limit)
5060 unsigned long long max = limit;
5061 unsigned long log2qty, size;
5064 /* allow the kernel cmdline to have a say */
5066 /* round applicable memory size up to nearest megabyte */
5067 numentries = nr_kernel_pages;
5068 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5069 numentries >>= 20 - PAGE_SHIFT;
5070 numentries <<= 20 - PAGE_SHIFT;
5072 /* limit to 1 bucket per 2^scale bytes of low memory */
5073 if (scale > PAGE_SHIFT)
5074 numentries >>= (scale - PAGE_SHIFT);
5076 numentries <<= (PAGE_SHIFT - scale);
5078 /* Make sure we've got at least a 0-order allocation.. */
5079 if (unlikely(flags & HASH_SMALL)) {
5080 /* Makes no sense without HASH_EARLY */
5081 WARN_ON(!(flags & HASH_EARLY));
5082 if (!(numentries >> *_hash_shift)) {
5083 numentries = 1UL << *_hash_shift;
5084 BUG_ON(!numentries);
5086 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5087 numentries = PAGE_SIZE / bucketsize;
5089 numentries = roundup_pow_of_two(numentries);
5091 /* limit allocation size to 1/16 total memory by default */
5093 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5094 do_div(max, bucketsize);
5097 if (numentries > max)
5100 log2qty = ilog2(numentries);
5103 size = bucketsize << log2qty;
5104 if (flags & HASH_EARLY)
5105 table = alloc_bootmem_nopanic(size);
5107 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5110 * If bucketsize is not a power-of-two, we may free
5111 * some pages at the end of hash table which
5112 * alloc_pages_exact() automatically does
5114 if (get_order(size) < MAX_ORDER) {
5115 table = alloc_pages_exact(size, GFP_ATOMIC);
5116 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5119 } while (!table && size > PAGE_SIZE && --log2qty);
5122 panic("Failed to allocate %s hash table\n", tablename);
5124 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5127 ilog2(size) - PAGE_SHIFT,
5131 *_hash_shift = log2qty;
5133 *_hash_mask = (1 << log2qty) - 1;
5138 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5139 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5142 #ifdef CONFIG_SPARSEMEM
5143 return __pfn_to_section(pfn)->pageblock_flags;
5145 return zone->pageblock_flags;
5146 #endif /* CONFIG_SPARSEMEM */
5149 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5151 #ifdef CONFIG_SPARSEMEM
5152 pfn &= (PAGES_PER_SECTION-1);
5153 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5155 pfn = pfn - zone->zone_start_pfn;
5156 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5157 #endif /* CONFIG_SPARSEMEM */
5161 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5162 * @page: The page within the block of interest
5163 * @start_bitidx: The first bit of interest to retrieve
5164 * @end_bitidx: The last bit of interest
5165 * returns pageblock_bits flags
5167 unsigned long get_pageblock_flags_group(struct page *page,
5168 int start_bitidx, int end_bitidx)
5171 unsigned long *bitmap;
5172 unsigned long pfn, bitidx;
5173 unsigned long flags = 0;
5174 unsigned long value = 1;
5176 zone = page_zone(page);
5177 pfn = page_to_pfn(page);
5178 bitmap = get_pageblock_bitmap(zone, pfn);
5179 bitidx = pfn_to_bitidx(zone, pfn);
5181 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5182 if (test_bit(bitidx + start_bitidx, bitmap))
5189 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5190 * @page: The page within the block of interest
5191 * @start_bitidx: The first bit of interest
5192 * @end_bitidx: The last bit of interest
5193 * @flags: The flags to set
5195 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5196 int start_bitidx, int end_bitidx)
5199 unsigned long *bitmap;
5200 unsigned long pfn, bitidx;
5201 unsigned long value = 1;
5203 zone = page_zone(page);
5204 pfn = page_to_pfn(page);
5205 bitmap = get_pageblock_bitmap(zone, pfn);
5206 bitidx = pfn_to_bitidx(zone, pfn);
5207 VM_BUG_ON(pfn < zone->zone_start_pfn);
5208 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5210 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5212 __set_bit(bitidx + start_bitidx, bitmap);
5214 __clear_bit(bitidx + start_bitidx, bitmap);
5218 * This is designed as sub function...plz see page_isolation.c also.
5219 * set/clear page block's type to be ISOLATE.
5220 * page allocater never alloc memory from ISOLATE block.
5224 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5226 unsigned long pfn, iter, found;
5228 * For avoiding noise data, lru_add_drain_all() should be called
5229 * If ZONE_MOVABLE, the zone never contains immobile pages
5231 if (zone_idx(zone) == ZONE_MOVABLE)
5234 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5237 pfn = page_to_pfn(page);
5238 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5239 unsigned long check = pfn + iter;
5241 if (!pfn_valid_within(check))
5244 page = pfn_to_page(check);
5245 if (!page_count(page)) {
5246 if (PageBuddy(page))
5247 iter += (1 << page_order(page)) - 1;
5253 * If there are RECLAIMABLE pages, we need to check it.
5254 * But now, memory offline itself doesn't call shrink_slab()
5255 * and it still to be fixed.
5258 * If the page is not RAM, page_count()should be 0.
5259 * we don't need more check. This is an _used_ not-movable page.
5261 * The problematic thing here is PG_reserved pages. PG_reserved
5262 * is set to both of a memory hole page and a _used_ kernel
5271 bool is_pageblock_removable_nolock(struct page *page)
5273 struct zone *zone = page_zone(page);
5274 return __count_immobile_pages(zone, page, 0);
5277 int set_migratetype_isolate(struct page *page)
5280 unsigned long flags, pfn;
5281 struct memory_isolate_notify arg;
5285 zone = page_zone(page);
5287 spin_lock_irqsave(&zone->lock, flags);
5289 pfn = page_to_pfn(page);
5290 arg.start_pfn = pfn;
5291 arg.nr_pages = pageblock_nr_pages;
5292 arg.pages_found = 0;
5295 * It may be possible to isolate a pageblock even if the
5296 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5297 * notifier chain is used by balloon drivers to return the
5298 * number of pages in a range that are held by the balloon
5299 * driver to shrink memory. If all the pages are accounted for
5300 * by balloons, are free, or on the LRU, isolation can continue.
5301 * Later, for example, when memory hotplug notifier runs, these
5302 * pages reported as "can be isolated" should be isolated(freed)
5303 * by the balloon driver through the memory notifier chain.
5305 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5306 notifier_ret = notifier_to_errno(notifier_ret);
5310 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5311 * We just check MOVABLE pages.
5313 if (__count_immobile_pages(zone, page, arg.pages_found))
5317 * immobile means "not-on-lru" paes. If immobile is larger than
5318 * removable-by-driver pages reported by notifier, we'll fail.
5323 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5324 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5327 spin_unlock_irqrestore(&zone->lock, flags);
5333 void unset_migratetype_isolate(struct page *page)
5336 unsigned long flags;
5337 zone = page_zone(page);
5338 spin_lock_irqsave(&zone->lock, flags);
5339 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5341 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5342 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5344 spin_unlock_irqrestore(&zone->lock, flags);
5347 #ifdef CONFIG_MEMORY_HOTREMOVE
5349 * All pages in the range must be isolated before calling this.
5352 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5358 unsigned long flags;
5359 /* find the first valid pfn */
5360 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5365 zone = page_zone(pfn_to_page(pfn));
5366 spin_lock_irqsave(&zone->lock, flags);
5368 while (pfn < end_pfn) {
5369 if (!pfn_valid(pfn)) {
5373 page = pfn_to_page(pfn);
5374 BUG_ON(page_count(page));
5375 BUG_ON(!PageBuddy(page));
5376 order = page_order(page);
5377 #ifdef CONFIG_DEBUG_VM
5378 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5379 pfn, 1 << order, end_pfn);
5381 list_del(&page->lru);
5382 rmv_page_order(page);
5383 zone->free_area[order].nr_free--;
5384 __mod_zone_page_state(zone, NR_FREE_PAGES,
5386 for (i = 0; i < (1 << order); i++)
5387 SetPageReserved((page+i));
5388 pfn += (1 << order);
5390 spin_unlock_irqrestore(&zone->lock, flags);
5394 #ifdef CONFIG_MEMORY_FAILURE
5395 bool is_free_buddy_page(struct page *page)
5397 struct zone *zone = page_zone(page);
5398 unsigned long pfn = page_to_pfn(page);
5399 unsigned long flags;
5402 spin_lock_irqsave(&zone->lock, flags);
5403 for (order = 0; order < MAX_ORDER; order++) {
5404 struct page *page_head = page - (pfn & ((1 << order) - 1));
5406 if (PageBuddy(page_head) && page_order(page_head) >= order)
5409 spin_unlock_irqrestore(&zone->lock, flags);
5411 return order < MAX_ORDER;
5415 static struct trace_print_flags pageflag_names[] = {
5416 {1UL << PG_locked, "locked" },
5417 {1UL << PG_error, "error" },
5418 {1UL << PG_referenced, "referenced" },
5419 {1UL << PG_uptodate, "uptodate" },
5420 {1UL << PG_dirty, "dirty" },
5421 {1UL << PG_lru, "lru" },
5422 {1UL << PG_active, "active" },
5423 {1UL << PG_slab, "slab" },
5424 {1UL << PG_owner_priv_1, "owner_priv_1" },
5425 {1UL << PG_arch_1, "arch_1" },
5426 {1UL << PG_reserved, "reserved" },
5427 {1UL << PG_private, "private" },
5428 {1UL << PG_private_2, "private_2" },
5429 {1UL << PG_writeback, "writeback" },
5430 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5431 {1UL << PG_head, "head" },
5432 {1UL << PG_tail, "tail" },
5434 {1UL << PG_compound, "compound" },
5436 {1UL << PG_swapcache, "swapcache" },
5437 {1UL << PG_mappedtodisk, "mappedtodisk" },
5438 {1UL << PG_reclaim, "reclaim" },
5439 {1UL << PG_swapbacked, "swapbacked" },
5440 {1UL << PG_unevictable, "unevictable" },
5442 {1UL << PG_mlocked, "mlocked" },
5444 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5445 {1UL << PG_uncached, "uncached" },
5447 #ifdef CONFIG_MEMORY_FAILURE
5448 {1UL << PG_hwpoison, "hwpoison" },
5453 static void dump_page_flags(unsigned long flags)
5455 const char *delim = "";
5459 printk(KERN_ALERT "page flags: %#lx(", flags);
5461 /* remove zone id */
5462 flags &= (1UL << NR_PAGEFLAGS) - 1;
5464 for (i = 0; pageflag_names[i].name && flags; i++) {
5466 mask = pageflag_names[i].mask;
5467 if ((flags & mask) != mask)
5471 printk("%s%s", delim, pageflag_names[i].name);
5475 /* check for left over flags */
5477 printk("%s%#lx", delim, flags);
5482 void dump_page(struct page *page)
5485 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5486 page, atomic_read(&page->_count), page_mapcount(page),
5487 page->mapping, page->index);
5488 dump_page_flags(page->flags);
5489 mem_cgroup_print_bad_page(page);