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/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
51 #include <linux/memory.h>
52 #include <linux/compaction.h>
53 #include <trace/events/kmem.h>
54 #include <linux/ftrace_event.h>
56 #include <asm/tlbflush.h>
57 #include <asm/div64.h>
60 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
61 DEFINE_PER_CPU(int, numa_node);
62 EXPORT_PER_CPU_SYMBOL(numa_node);
66 * Array of node states.
68 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
69 [N_POSSIBLE] = NODE_MASK_ALL,
70 [N_ONLINE] = { { [0] = 1UL } },
72 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
74 [N_HIGH_MEMORY] = { { [0] = 1UL } },
76 [N_CPU] = { { [0] = 1UL } },
79 EXPORT_SYMBOL(node_states);
81 unsigned long totalram_pages __read_mostly;
82 unsigned long totalreserve_pages __read_mostly;
83 int percpu_pagelist_fraction;
84 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
86 #ifdef CONFIG_PM_SLEEP
88 * The following functions are used by the suspend/hibernate code to temporarily
89 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
90 * while devices are suspended. To avoid races with the suspend/hibernate code,
91 * they should always be called with pm_mutex held (gfp_allowed_mask also should
92 * only be modified with pm_mutex held, unless the suspend/hibernate code is
93 * guaranteed not to run in parallel with that modification).
95 void set_gfp_allowed_mask(gfp_t mask)
97 WARN_ON(!mutex_is_locked(&pm_mutex));
98 gfp_allowed_mask = mask;
101 gfp_t clear_gfp_allowed_mask(gfp_t mask)
103 gfp_t ret = gfp_allowed_mask;
105 WARN_ON(!mutex_is_locked(&pm_mutex));
106 gfp_allowed_mask &= ~mask;
109 #endif /* CONFIG_PM_SLEEP */
111 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
112 int pageblock_order __read_mostly;
115 static void __free_pages_ok(struct page *page, unsigned int order);
118 * results with 256, 32 in the lowmem_reserve sysctl:
119 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
120 * 1G machine -> (16M dma, 784M normal, 224M high)
121 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
122 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
123 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
125 * TBD: should special case ZONE_DMA32 machines here - in those we normally
126 * don't need any ZONE_NORMAL reservation
128 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
129 #ifdef CONFIG_ZONE_DMA
132 #ifdef CONFIG_ZONE_DMA32
135 #ifdef CONFIG_HIGHMEM
141 EXPORT_SYMBOL(totalram_pages);
143 static char * const zone_names[MAX_NR_ZONES] = {
144 #ifdef CONFIG_ZONE_DMA
147 #ifdef CONFIG_ZONE_DMA32
151 #ifdef CONFIG_HIGHMEM
157 int min_free_kbytes = 1024;
159 static unsigned long __meminitdata nr_kernel_pages;
160 static unsigned long __meminitdata nr_all_pages;
161 static unsigned long __meminitdata dma_reserve;
163 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
165 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
166 * ranges of memory (RAM) that may be registered with add_active_range().
167 * Ranges passed to add_active_range() will be merged if possible
168 * so the number of times add_active_range() can be called is
169 * related to the number of nodes and the number of holes
171 #ifdef CONFIG_MAX_ACTIVE_REGIONS
172 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
173 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
175 #if MAX_NUMNODES >= 32
176 /* If there can be many nodes, allow up to 50 holes per node */
177 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
179 /* By default, allow up to 256 distinct regions */
180 #define MAX_ACTIVE_REGIONS 256
184 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
185 static int __meminitdata nr_nodemap_entries;
186 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
187 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
188 static unsigned long __initdata required_kernelcore;
189 static unsigned long __initdata required_movablecore;
190 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
192 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
194 EXPORT_SYMBOL(movable_zone);
195 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
198 int nr_node_ids __read_mostly = MAX_NUMNODES;
199 int nr_online_nodes __read_mostly = 1;
200 EXPORT_SYMBOL(nr_node_ids);
201 EXPORT_SYMBOL(nr_online_nodes);
204 int page_group_by_mobility_disabled __read_mostly;
206 static void set_pageblock_migratetype(struct page *page, int migratetype)
209 if (unlikely(page_group_by_mobility_disabled))
210 migratetype = MIGRATE_UNMOVABLE;
212 set_pageblock_flags_group(page, (unsigned long)migratetype,
213 PB_migrate, PB_migrate_end);
216 bool oom_killer_disabled __read_mostly;
218 #ifdef CONFIG_DEBUG_VM
219 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
223 unsigned long pfn = page_to_pfn(page);
226 seq = zone_span_seqbegin(zone);
227 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
229 else if (pfn < zone->zone_start_pfn)
231 } while (zone_span_seqretry(zone, seq));
236 static int page_is_consistent(struct zone *zone, struct page *page)
238 if (!pfn_valid_within(page_to_pfn(page)))
240 if (zone != page_zone(page))
246 * Temporary debugging check for pages not lying within a given zone.
248 static int bad_range(struct zone *zone, struct page *page)
250 if (page_outside_zone_boundaries(zone, page))
252 if (!page_is_consistent(zone, page))
258 static inline int bad_range(struct zone *zone, struct page *page)
264 static void bad_page(struct page *page)
266 static unsigned long resume;
267 static unsigned long nr_shown;
268 static unsigned long nr_unshown;
270 /* Don't complain about poisoned pages */
271 if (PageHWPoison(page)) {
272 __ClearPageBuddy(page);
277 * Allow a burst of 60 reports, then keep quiet for that minute;
278 * or allow a steady drip of one report per second.
280 if (nr_shown == 60) {
281 if (time_before(jiffies, resume)) {
287 "BUG: Bad page state: %lu messages suppressed\n",
294 resume = jiffies + 60 * HZ;
296 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
297 current->comm, page_to_pfn(page));
302 /* Leave bad fields for debug, except PageBuddy could make trouble */
303 __ClearPageBuddy(page);
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 pages have their ->private pointing at
315 * the head page (even the head page has this).
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;
339 p->first_page = page;
343 static int destroy_compound_page(struct page *page, unsigned long order)
346 int nr_pages = 1 << order;
349 if (unlikely(compound_order(page) != order) ||
350 unlikely(!PageHead(page))) {
355 __ClearPageHead(page);
357 for (i = 1; i < nr_pages; i++) {
358 struct page *p = page + i;
360 if (unlikely(!PageTail(p) || (p->first_page != page))) {
370 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
375 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
376 * and __GFP_HIGHMEM from hard or soft interrupt context.
378 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
379 for (i = 0; i < (1 << order); i++)
380 clear_highpage(page + i);
383 static inline void set_page_order(struct page *page, int order)
385 set_page_private(page, order);
386 __SetPageBuddy(page);
389 static inline void rmv_page_order(struct page *page)
391 __ClearPageBuddy(page);
392 set_page_private(page, 0);
396 * Locate the struct page for both the matching buddy in our
397 * pair (buddy1) and the combined O(n+1) page they form (page).
399 * 1) Any buddy B1 will have an order O twin B2 which satisfies
400 * the following equation:
402 * For example, if the starting buddy (buddy2) is #8 its order
404 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
406 * 2) Any buddy B will have an order O+1 parent P which
407 * satisfies the following equation:
410 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
412 static inline struct page *
413 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
415 unsigned long buddy_idx = page_idx ^ (1 << order);
417 return page + (buddy_idx - page_idx);
420 static inline unsigned long
421 __find_combined_index(unsigned long page_idx, unsigned int order)
423 return (page_idx & ~(1 << order));
427 * This function checks whether a page is free && is the buddy
428 * we can do coalesce a page and its buddy if
429 * (a) the buddy is not in a hole &&
430 * (b) the buddy is in the buddy system &&
431 * (c) a page and its buddy have the same order &&
432 * (d) a page and its buddy are in the same zone.
434 * For recording whether a page is in the buddy system, we use PG_buddy.
435 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
437 * For recording page's order, we use page_private(page).
439 static inline int page_is_buddy(struct page *page, struct page *buddy,
442 if (!pfn_valid_within(page_to_pfn(buddy)))
445 if (page_zone_id(page) != page_zone_id(buddy))
448 if (PageBuddy(buddy) && page_order(buddy) == order) {
449 VM_BUG_ON(page_count(buddy) != 0);
456 * Freeing function for a buddy system allocator.
458 * The concept of a buddy system is to maintain direct-mapped table
459 * (containing bit values) for memory blocks of various "orders".
460 * The bottom level table contains the map for the smallest allocatable
461 * units of memory (here, pages), and each level above it describes
462 * pairs of units from the levels below, hence, "buddies".
463 * At a high level, all that happens here is marking the table entry
464 * at the bottom level available, and propagating the changes upward
465 * as necessary, plus some accounting needed to play nicely with other
466 * parts of the VM system.
467 * At each level, we keep a list of pages, which are heads of continuous
468 * free pages of length of (1 << order) and marked with PG_buddy. Page's
469 * order is recorded in page_private(page) field.
470 * So when we are allocating or freeing one, we can derive the state of the
471 * other. That is, if we allocate a small block, and both were
472 * free, the remainder of the region must be split into blocks.
473 * If a block is freed, and its buddy is also free, then this
474 * triggers coalescing into a block of larger size.
479 static inline void __free_one_page(struct page *page,
480 struct zone *zone, unsigned int order,
483 unsigned long page_idx;
484 unsigned long combined_idx;
487 if (unlikely(PageCompound(page)))
488 if (unlikely(destroy_compound_page(page, order)))
491 VM_BUG_ON(migratetype == -1);
493 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
495 VM_BUG_ON(page_idx & ((1 << order) - 1));
496 VM_BUG_ON(bad_range(zone, page));
498 while (order < MAX_ORDER-1) {
499 buddy = __page_find_buddy(page, page_idx, order);
500 if (!page_is_buddy(page, buddy, order))
503 /* Our buddy is free, merge with it and move up one order. */
504 list_del(&buddy->lru);
505 zone->free_area[order].nr_free--;
506 rmv_page_order(buddy);
507 combined_idx = __find_combined_index(page_idx, order);
508 page = page + (combined_idx - page_idx);
509 page_idx = combined_idx;
512 set_page_order(page, order);
515 * If this is not the largest possible page, check if the buddy
516 * of the next-highest order is free. If it is, it's possible
517 * that pages are being freed that will coalesce soon. In case,
518 * that is happening, add the free page to the tail of the list
519 * so it's less likely to be used soon and more likely to be merged
520 * as a higher order page
522 if ((order < MAX_ORDER-1) && pfn_valid_within(page_to_pfn(buddy))) {
523 struct page *higher_page, *higher_buddy;
524 combined_idx = __find_combined_index(page_idx, order);
525 higher_page = page + combined_idx - page_idx;
526 higher_buddy = __page_find_buddy(higher_page, combined_idx, order + 1);
527 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
528 list_add_tail(&page->lru,
529 &zone->free_area[order].free_list[migratetype]);
534 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
536 zone->free_area[order].nr_free++;
540 * free_page_mlock() -- clean up attempts to free and mlocked() page.
541 * Page should not be on lru, so no need to fix that up.
542 * free_pages_check() will verify...
544 static inline void free_page_mlock(struct page *page)
546 __dec_zone_page_state(page, NR_MLOCK);
547 __count_vm_event(UNEVICTABLE_MLOCKFREED);
550 static inline int free_pages_check(struct page *page)
552 if (unlikely(page_mapcount(page) |
553 (page->mapping != NULL) |
554 (atomic_read(&page->_count) != 0) |
555 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
559 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
560 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
565 * Frees a number of pages from the PCP lists
566 * Assumes all pages on list are in same zone, and of same order.
567 * count is the number of pages to free.
569 * If the zone was previously in an "all pages pinned" state then look to
570 * see if this freeing clears that state.
572 * And clear the zone's pages_scanned counter, to hold off the "all pages are
573 * pinned" detection logic.
575 static void free_pcppages_bulk(struct zone *zone, int count,
576 struct per_cpu_pages *pcp)
581 spin_lock(&zone->lock);
582 zone->all_unreclaimable = 0;
583 zone->pages_scanned = 0;
585 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
588 struct list_head *list;
591 * Remove pages from lists in a round-robin fashion. A
592 * batch_free count is maintained that is incremented when an
593 * empty list is encountered. This is so more pages are freed
594 * off fuller lists instead of spinning excessively around empty
599 if (++migratetype == MIGRATE_PCPTYPES)
601 list = &pcp->lists[migratetype];
602 } while (list_empty(list));
605 page = list_entry(list->prev, struct page, lru);
606 /* must delete as __free_one_page list manipulates */
607 list_del(&page->lru);
608 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
609 __free_one_page(page, zone, 0, page_private(page));
610 trace_mm_page_pcpu_drain(page, 0, page_private(page));
611 } while (--count && --batch_free && !list_empty(list));
613 spin_unlock(&zone->lock);
616 static void free_one_page(struct zone *zone, struct page *page, int order,
619 spin_lock(&zone->lock);
620 zone->all_unreclaimable = 0;
621 zone->pages_scanned = 0;
623 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
624 __free_one_page(page, zone, order, migratetype);
625 spin_unlock(&zone->lock);
628 static bool free_pages_prepare(struct page *page, unsigned int order)
633 trace_mm_page_free_direct(page, order);
634 kmemcheck_free_shadow(page, order);
636 for (i = 0; i < (1 << order); i++) {
637 struct page *pg = page + i;
641 bad += free_pages_check(pg);
646 if (!PageHighMem(page)) {
647 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
648 debug_check_no_obj_freed(page_address(page),
651 arch_free_page(page, order);
652 kernel_map_pages(page, 1 << order, 0);
657 static void __free_pages_ok(struct page *page, unsigned int order)
660 int wasMlocked = __TestClearPageMlocked(page);
662 if (!free_pages_prepare(page, order))
665 local_irq_save(flags);
666 if (unlikely(wasMlocked))
667 free_page_mlock(page);
668 __count_vm_events(PGFREE, 1 << order);
669 free_one_page(page_zone(page), page, order,
670 get_pageblock_migratetype(page));
671 local_irq_restore(flags);
675 * permit the bootmem allocator to evade page validation on high-order frees
677 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
680 __ClearPageReserved(page);
681 set_page_count(page, 0);
682 set_page_refcounted(page);
688 for (loop = 0; loop < BITS_PER_LONG; loop++) {
689 struct page *p = &page[loop];
691 if (loop + 1 < BITS_PER_LONG)
693 __ClearPageReserved(p);
694 set_page_count(p, 0);
697 set_page_refcounted(page);
698 __free_pages(page, order);
704 * The order of subdivision here is critical for the IO subsystem.
705 * Please do not alter this order without good reasons and regression
706 * testing. Specifically, as large blocks of memory are subdivided,
707 * the order in which smaller blocks are delivered depends on the order
708 * they're subdivided in this function. This is the primary factor
709 * influencing the order in which pages are delivered to the IO
710 * subsystem according to empirical testing, and this is also justified
711 * by considering the behavior of a buddy system containing a single
712 * large block of memory acted on by a series of small allocations.
713 * This behavior is a critical factor in sglist merging's success.
717 static inline void expand(struct zone *zone, struct page *page,
718 int low, int high, struct free_area *area,
721 unsigned long size = 1 << high;
727 VM_BUG_ON(bad_range(zone, &page[size]));
728 list_add(&page[size].lru, &area->free_list[migratetype]);
730 set_page_order(&page[size], high);
735 * This page is about to be returned from the page allocator
737 static inline int check_new_page(struct page *page)
739 if (unlikely(page_mapcount(page) |
740 (page->mapping != NULL) |
741 (atomic_read(&page->_count) != 0) |
742 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
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_del(&page->lru);
857 &zone->free_area[order].free_list[migratetype]);
859 pages_moved += 1 << order;
865 static int move_freepages_block(struct zone *zone, struct page *page,
868 unsigned long start_pfn, end_pfn;
869 struct page *start_page, *end_page;
871 start_pfn = page_to_pfn(page);
872 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
873 start_page = pfn_to_page(start_pfn);
874 end_page = start_page + pageblock_nr_pages - 1;
875 end_pfn = start_pfn + pageblock_nr_pages - 1;
877 /* Do not cross zone boundaries */
878 if (start_pfn < zone->zone_start_pfn)
880 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
883 return move_freepages(zone, start_page, end_page, migratetype);
886 static void change_pageblock_range(struct page *pageblock_page,
887 int start_order, int migratetype)
889 int nr_pageblocks = 1 << (start_order - pageblock_order);
891 while (nr_pageblocks--) {
892 set_pageblock_migratetype(pageblock_page, migratetype);
893 pageblock_page += pageblock_nr_pages;
897 /* Remove an element from the buddy allocator from the fallback list */
898 static inline struct page *
899 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
901 struct free_area * area;
906 /* Find the largest possible block of pages in the other list */
907 for (current_order = MAX_ORDER-1; current_order >= order;
909 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
910 migratetype = fallbacks[start_migratetype][i];
912 /* MIGRATE_RESERVE handled later if necessary */
913 if (migratetype == MIGRATE_RESERVE)
916 area = &(zone->free_area[current_order]);
917 if (list_empty(&area->free_list[migratetype]))
920 page = list_entry(area->free_list[migratetype].next,
925 * If breaking a large block of pages, move all free
926 * pages to the preferred allocation list. If falling
927 * back for a reclaimable kernel allocation, be more
928 * agressive about taking ownership of free pages
930 if (unlikely(current_order >= (pageblock_order >> 1)) ||
931 start_migratetype == MIGRATE_RECLAIMABLE ||
932 page_group_by_mobility_disabled) {
934 pages = move_freepages_block(zone, page,
937 /* Claim the whole block if over half of it is free */
938 if (pages >= (1 << (pageblock_order-1)) ||
939 page_group_by_mobility_disabled)
940 set_pageblock_migratetype(page,
943 migratetype = start_migratetype;
946 /* Remove the page from the freelists */
947 list_del(&page->lru);
948 rmv_page_order(page);
950 /* Take ownership for orders >= pageblock_order */
951 if (current_order >= pageblock_order)
952 change_pageblock_range(page, current_order,
955 expand(zone, page, order, current_order, area, migratetype);
957 trace_mm_page_alloc_extfrag(page, order, current_order,
958 start_migratetype, migratetype);
968 * Do the hard work of removing an element from the buddy allocator.
969 * Call me with the zone->lock already held.
971 static struct page *__rmqueue(struct zone *zone, unsigned int order,
977 page = __rmqueue_smallest(zone, order, migratetype);
979 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
980 page = __rmqueue_fallback(zone, order, migratetype);
983 * Use MIGRATE_RESERVE rather than fail an allocation. goto
984 * is used because __rmqueue_smallest is an inline function
985 * and we want just one call site
988 migratetype = MIGRATE_RESERVE;
993 trace_mm_page_alloc_zone_locked(page, order, migratetype);
998 * Obtain a specified number of elements from the buddy allocator, all under
999 * a single hold of the lock, for efficiency. Add them to the supplied list.
1000 * Returns the number of new pages which were placed at *list.
1002 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1003 unsigned long count, struct list_head *list,
1004 int migratetype, int cold)
1008 spin_lock(&zone->lock);
1009 for (i = 0; i < count; ++i) {
1010 struct page *page = __rmqueue(zone, order, migratetype);
1011 if (unlikely(page == NULL))
1015 * Split buddy pages returned by expand() are received here
1016 * in physical page order. The page is added to the callers and
1017 * list and the list head then moves forward. From the callers
1018 * perspective, the linked list is ordered by page number in
1019 * some conditions. This is useful for IO devices that can
1020 * merge IO requests if the physical pages are ordered
1023 if (likely(cold == 0))
1024 list_add(&page->lru, list);
1026 list_add_tail(&page->lru, list);
1027 set_page_private(page, migratetype);
1030 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1031 spin_unlock(&zone->lock);
1037 * Called from the vmstat counter updater to drain pagesets of this
1038 * currently executing processor on remote nodes after they have
1041 * Note that this function must be called with the thread pinned to
1042 * a single processor.
1044 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1046 unsigned long flags;
1049 local_irq_save(flags);
1050 if (pcp->count >= pcp->batch)
1051 to_drain = pcp->batch;
1053 to_drain = pcp->count;
1054 free_pcppages_bulk(zone, to_drain, pcp);
1055 pcp->count -= to_drain;
1056 local_irq_restore(flags);
1061 * Drain pages of the indicated processor.
1063 * The processor must either be the current processor and the
1064 * thread pinned to the current processor or a processor that
1067 static void drain_pages(unsigned int cpu)
1069 unsigned long flags;
1072 for_each_populated_zone(zone) {
1073 struct per_cpu_pageset *pset;
1074 struct per_cpu_pages *pcp;
1076 local_irq_save(flags);
1077 pset = per_cpu_ptr(zone->pageset, cpu);
1080 free_pcppages_bulk(zone, pcp->count, pcp);
1082 local_irq_restore(flags);
1087 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1089 void drain_local_pages(void *arg)
1091 drain_pages(smp_processor_id());
1095 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1097 void drain_all_pages(void)
1099 on_each_cpu(drain_local_pages, NULL, 1);
1102 #ifdef CONFIG_HIBERNATION
1104 void mark_free_pages(struct zone *zone)
1106 unsigned long pfn, max_zone_pfn;
1107 unsigned long flags;
1109 struct list_head *curr;
1111 if (!zone->spanned_pages)
1114 spin_lock_irqsave(&zone->lock, flags);
1116 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1117 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1118 if (pfn_valid(pfn)) {
1119 struct page *page = pfn_to_page(pfn);
1121 if (!swsusp_page_is_forbidden(page))
1122 swsusp_unset_page_free(page);
1125 for_each_migratetype_order(order, t) {
1126 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1129 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1130 for (i = 0; i < (1UL << order); i++)
1131 swsusp_set_page_free(pfn_to_page(pfn + i));
1134 spin_unlock_irqrestore(&zone->lock, flags);
1136 #endif /* CONFIG_PM */
1139 * Free a 0-order page
1140 * cold == 1 ? free a cold page : free a hot page
1142 void free_hot_cold_page(struct page *page, int cold)
1144 struct zone *zone = page_zone(page);
1145 struct per_cpu_pages *pcp;
1146 unsigned long flags;
1148 int wasMlocked = __TestClearPageMlocked(page);
1150 if (!free_pages_prepare(page, 0))
1153 migratetype = get_pageblock_migratetype(page);
1154 set_page_private(page, migratetype);
1155 local_irq_save(flags);
1156 if (unlikely(wasMlocked))
1157 free_page_mlock(page);
1158 __count_vm_event(PGFREE);
1161 * We only track unmovable, reclaimable and movable on pcp lists.
1162 * Free ISOLATE pages back to the allocator because they are being
1163 * offlined but treat RESERVE as movable pages so we can get those
1164 * areas back if necessary. Otherwise, we may have to free
1165 * excessively into the page allocator
1167 if (migratetype >= MIGRATE_PCPTYPES) {
1168 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1169 free_one_page(zone, page, 0, migratetype);
1172 migratetype = MIGRATE_MOVABLE;
1175 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1177 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1179 list_add(&page->lru, &pcp->lists[migratetype]);
1181 if (pcp->count >= pcp->high) {
1182 free_pcppages_bulk(zone, pcp->batch, pcp);
1183 pcp->count -= pcp->batch;
1187 local_irq_restore(flags);
1191 * split_page takes a non-compound higher-order page, and splits it into
1192 * n (1<<order) sub-pages: page[0..n]
1193 * Each sub-page must be freed individually.
1195 * Note: this is probably too low level an operation for use in drivers.
1196 * Please consult with lkml before using this in your driver.
1198 void split_page(struct page *page, unsigned int order)
1202 VM_BUG_ON(PageCompound(page));
1203 VM_BUG_ON(!page_count(page));
1205 #ifdef CONFIG_KMEMCHECK
1207 * Split shadow pages too, because free(page[0]) would
1208 * otherwise free the whole shadow.
1210 if (kmemcheck_page_is_tracked(page))
1211 split_page(virt_to_page(page[0].shadow), order);
1214 for (i = 1; i < (1 << order); i++)
1215 set_page_refcounted(page + i);
1219 * Similar to split_page except the page is already free. As this is only
1220 * being used for migration, the migratetype of the block also changes.
1221 * As this is called with interrupts disabled, the caller is responsible
1222 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1225 * Note: this is probably too low level an operation for use in drivers.
1226 * Please consult with lkml before using this in your driver.
1228 int split_free_page(struct page *page)
1231 unsigned long watermark;
1234 BUG_ON(!PageBuddy(page));
1236 zone = page_zone(page);
1237 order = page_order(page);
1239 /* Obey watermarks as if the page was being allocated */
1240 watermark = low_wmark_pages(zone) + (1 << order);
1241 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1244 /* Remove page from free list */
1245 list_del(&page->lru);
1246 zone->free_area[order].nr_free--;
1247 rmv_page_order(page);
1248 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1250 /* Split into individual pages */
1251 set_page_refcounted(page);
1252 split_page(page, order);
1254 if (order >= pageblock_order - 1) {
1255 struct page *endpage = page + (1 << order) - 1;
1256 for (; page < endpage; page += pageblock_nr_pages)
1257 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1264 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1265 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1269 struct page *buffered_rmqueue(struct zone *preferred_zone,
1270 struct zone *zone, int order, gfp_t gfp_flags,
1273 unsigned long flags;
1275 int cold = !!(gfp_flags & __GFP_COLD);
1278 if (likely(order == 0)) {
1279 struct per_cpu_pages *pcp;
1280 struct list_head *list;
1282 local_irq_save(flags);
1283 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1284 list = &pcp->lists[migratetype];
1285 if (list_empty(list)) {
1286 pcp->count += rmqueue_bulk(zone, 0,
1289 if (unlikely(list_empty(list)))
1294 page = list_entry(list->prev, struct page, lru);
1296 page = list_entry(list->next, struct page, lru);
1298 list_del(&page->lru);
1301 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1303 * __GFP_NOFAIL is not to be used in new code.
1305 * All __GFP_NOFAIL callers should be fixed so that they
1306 * properly detect and handle allocation failures.
1308 * We most definitely don't want callers attempting to
1309 * allocate greater than order-1 page units with
1312 WARN_ON_ONCE(order > 1);
1314 spin_lock_irqsave(&zone->lock, flags);
1315 page = __rmqueue(zone, order, migratetype);
1316 spin_unlock(&zone->lock);
1319 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1322 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1323 zone_statistics(preferred_zone, zone);
1324 local_irq_restore(flags);
1326 VM_BUG_ON(bad_range(zone, page));
1327 if (prep_new_page(page, order, gfp_flags))
1332 local_irq_restore(flags);
1336 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1337 #define ALLOC_WMARK_MIN WMARK_MIN
1338 #define ALLOC_WMARK_LOW WMARK_LOW
1339 #define ALLOC_WMARK_HIGH WMARK_HIGH
1340 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1342 /* Mask to get the watermark bits */
1343 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1345 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1346 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1347 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1349 #ifdef CONFIG_FAIL_PAGE_ALLOC
1351 static struct fail_page_alloc_attr {
1352 struct fault_attr attr;
1354 u32 ignore_gfp_highmem;
1355 u32 ignore_gfp_wait;
1358 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1360 struct dentry *ignore_gfp_highmem_file;
1361 struct dentry *ignore_gfp_wait_file;
1362 struct dentry *min_order_file;
1364 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1366 } fail_page_alloc = {
1367 .attr = FAULT_ATTR_INITIALIZER,
1368 .ignore_gfp_wait = 1,
1369 .ignore_gfp_highmem = 1,
1373 static int __init setup_fail_page_alloc(char *str)
1375 return setup_fault_attr(&fail_page_alloc.attr, str);
1377 __setup("fail_page_alloc=", setup_fail_page_alloc);
1379 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1381 if (order < fail_page_alloc.min_order)
1383 if (gfp_mask & __GFP_NOFAIL)
1385 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1387 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1390 return should_fail(&fail_page_alloc.attr, 1 << order);
1393 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1395 static int __init fail_page_alloc_debugfs(void)
1397 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1401 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1405 dir = fail_page_alloc.attr.dentries.dir;
1407 fail_page_alloc.ignore_gfp_wait_file =
1408 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1409 &fail_page_alloc.ignore_gfp_wait);
1411 fail_page_alloc.ignore_gfp_highmem_file =
1412 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1413 &fail_page_alloc.ignore_gfp_highmem);
1414 fail_page_alloc.min_order_file =
1415 debugfs_create_u32("min-order", mode, dir,
1416 &fail_page_alloc.min_order);
1418 if (!fail_page_alloc.ignore_gfp_wait_file ||
1419 !fail_page_alloc.ignore_gfp_highmem_file ||
1420 !fail_page_alloc.min_order_file) {
1422 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1423 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1424 debugfs_remove(fail_page_alloc.min_order_file);
1425 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1431 late_initcall(fail_page_alloc_debugfs);
1433 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1435 #else /* CONFIG_FAIL_PAGE_ALLOC */
1437 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1442 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1445 * Return 1 if free pages are above 'mark'. This takes into account the order
1446 * of the allocation.
1448 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1449 int classzone_idx, int alloc_flags)
1451 /* free_pages my go negative - that's OK */
1453 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1456 if (alloc_flags & ALLOC_HIGH)
1458 if (alloc_flags & ALLOC_HARDER)
1461 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1463 for (o = 0; o < order; o++) {
1464 /* At the next order, this order's pages become unavailable */
1465 free_pages -= z->free_area[o].nr_free << o;
1467 /* Require fewer higher order pages to be free */
1470 if (free_pages <= min)
1478 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1479 * skip over zones that are not allowed by the cpuset, or that have
1480 * been recently (in last second) found to be nearly full. See further
1481 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1482 * that have to skip over a lot of full or unallowed zones.
1484 * If the zonelist cache is present in the passed in zonelist, then
1485 * returns a pointer to the allowed node mask (either the current
1486 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1488 * If the zonelist cache is not available for this zonelist, does
1489 * nothing and returns NULL.
1491 * If the fullzones BITMAP in the zonelist cache is stale (more than
1492 * a second since last zap'd) then we zap it out (clear its bits.)
1494 * We hold off even calling zlc_setup, until after we've checked the
1495 * first zone in the zonelist, on the theory that most allocations will
1496 * be satisfied from that first zone, so best to examine that zone as
1497 * quickly as we can.
1499 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1501 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1502 nodemask_t *allowednodes; /* zonelist_cache approximation */
1504 zlc = zonelist->zlcache_ptr;
1508 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1509 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1510 zlc->last_full_zap = jiffies;
1513 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1514 &cpuset_current_mems_allowed :
1515 &node_states[N_HIGH_MEMORY];
1516 return allowednodes;
1520 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1521 * if it is worth looking at further for free memory:
1522 * 1) Check that the zone isn't thought to be full (doesn't have its
1523 * bit set in the zonelist_cache fullzones BITMAP).
1524 * 2) Check that the zones node (obtained from the zonelist_cache
1525 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1526 * Return true (non-zero) if zone is worth looking at further, or
1527 * else return false (zero) if it is not.
1529 * This check -ignores- the distinction between various watermarks,
1530 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1531 * found to be full for any variation of these watermarks, it will
1532 * be considered full for up to one second by all requests, unless
1533 * we are so low on memory on all allowed nodes that we are forced
1534 * into the second scan of the zonelist.
1536 * In the second scan we ignore this zonelist cache and exactly
1537 * apply the watermarks to all zones, even it is slower to do so.
1538 * We are low on memory in the second scan, and should leave no stone
1539 * unturned looking for a free page.
1541 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1542 nodemask_t *allowednodes)
1544 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1545 int i; /* index of *z in zonelist zones */
1546 int n; /* node that zone *z is on */
1548 zlc = zonelist->zlcache_ptr;
1552 i = z - zonelist->_zonerefs;
1555 /* This zone is worth trying if it is allowed but not full */
1556 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1560 * Given 'z' scanning a zonelist, set the corresponding bit in
1561 * zlc->fullzones, so that subsequent attempts to allocate a page
1562 * from that zone don't waste time re-examining it.
1564 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1566 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1567 int i; /* index of *z in zonelist zones */
1569 zlc = zonelist->zlcache_ptr;
1573 i = z - zonelist->_zonerefs;
1575 set_bit(i, zlc->fullzones);
1578 #else /* CONFIG_NUMA */
1580 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1585 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1586 nodemask_t *allowednodes)
1591 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1594 #endif /* CONFIG_NUMA */
1597 * get_page_from_freelist goes through the zonelist trying to allocate
1600 static struct page *
1601 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1602 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1603 struct zone *preferred_zone, int migratetype)
1606 struct page *page = NULL;
1609 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1610 int zlc_active = 0; /* set if using zonelist_cache */
1611 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1613 classzone_idx = zone_idx(preferred_zone);
1616 * Scan zonelist, looking for a zone with enough free.
1617 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1619 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1620 high_zoneidx, nodemask) {
1621 if (NUMA_BUILD && zlc_active &&
1622 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1624 if ((alloc_flags & ALLOC_CPUSET) &&
1625 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1628 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1629 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1633 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1634 if (zone_watermark_ok(zone, order, mark,
1635 classzone_idx, alloc_flags))
1638 if (zone_reclaim_mode == 0)
1639 goto this_zone_full;
1641 ret = zone_reclaim(zone, gfp_mask, order);
1643 case ZONE_RECLAIM_NOSCAN:
1646 case ZONE_RECLAIM_FULL:
1647 /* scanned but unreclaimable */
1648 goto this_zone_full;
1650 /* did we reclaim enough */
1651 if (!zone_watermark_ok(zone, order, mark,
1652 classzone_idx, alloc_flags))
1653 goto this_zone_full;
1658 page = buffered_rmqueue(preferred_zone, zone, order,
1659 gfp_mask, migratetype);
1664 zlc_mark_zone_full(zonelist, z);
1666 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1668 * we do zlc_setup after the first zone is tried but only
1669 * if there are multiple nodes make it worthwhile
1671 allowednodes = zlc_setup(zonelist, alloc_flags);
1677 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1678 /* Disable zlc cache for second zonelist scan */
1686 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1687 unsigned long pages_reclaimed)
1689 /* Do not loop if specifically requested */
1690 if (gfp_mask & __GFP_NORETRY)
1694 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1695 * means __GFP_NOFAIL, but that may not be true in other
1698 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1702 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1703 * specified, then we retry until we no longer reclaim any pages
1704 * (above), or we've reclaimed an order of pages at least as
1705 * large as the allocation's order. In both cases, if the
1706 * allocation still fails, we stop retrying.
1708 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1712 * Don't let big-order allocations loop unless the caller
1713 * explicitly requests that.
1715 if (gfp_mask & __GFP_NOFAIL)
1721 static inline struct page *
1722 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1723 struct zonelist *zonelist, enum zone_type high_zoneidx,
1724 nodemask_t *nodemask, struct zone *preferred_zone,
1729 /* Acquire the OOM killer lock for the zones in zonelist */
1730 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1731 schedule_timeout_uninterruptible(1);
1736 * Go through the zonelist yet one more time, keep very high watermark
1737 * here, this is only to catch a parallel oom killing, we must fail if
1738 * we're still under heavy pressure.
1740 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1741 order, zonelist, high_zoneidx,
1742 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1743 preferred_zone, migratetype);
1747 if (!(gfp_mask & __GFP_NOFAIL)) {
1748 /* The OOM killer will not help higher order allocs */
1749 if (order > PAGE_ALLOC_COSTLY_ORDER)
1752 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1753 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1754 * The caller should handle page allocation failure by itself if
1755 * it specifies __GFP_THISNODE.
1756 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1758 if (gfp_mask & __GFP_THISNODE)
1761 /* Exhausted what can be done so it's blamo time */
1762 out_of_memory(zonelist, gfp_mask, order, nodemask);
1765 clear_zonelist_oom(zonelist, gfp_mask);
1769 #ifdef CONFIG_COMPACTION
1770 /* Try memory compaction for high-order allocations before reclaim */
1771 static struct page *
1772 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1773 struct zonelist *zonelist, enum zone_type high_zoneidx,
1774 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1775 int migratetype, unsigned long *did_some_progress)
1779 if (!order || compaction_deferred(preferred_zone))
1782 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1784 if (*did_some_progress != COMPACT_SKIPPED) {
1786 /* Page migration frees to the PCP lists but we want merging */
1787 drain_pages(get_cpu());
1790 page = get_page_from_freelist(gfp_mask, nodemask,
1791 order, zonelist, high_zoneidx,
1792 alloc_flags, preferred_zone,
1795 preferred_zone->compact_considered = 0;
1796 preferred_zone->compact_defer_shift = 0;
1797 count_vm_event(COMPACTSUCCESS);
1802 * It's bad if compaction run occurs and fails.
1803 * The most likely reason is that pages exist,
1804 * but not enough to satisfy watermarks.
1806 count_vm_event(COMPACTFAIL);
1807 defer_compaction(preferred_zone);
1815 static inline struct page *
1816 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1817 struct zonelist *zonelist, enum zone_type high_zoneidx,
1818 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1819 int migratetype, unsigned long *did_some_progress)
1823 #endif /* CONFIG_COMPACTION */
1825 /* The really slow allocator path where we enter direct reclaim */
1826 static inline struct page *
1827 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1828 struct zonelist *zonelist, enum zone_type high_zoneidx,
1829 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1830 int migratetype, unsigned long *did_some_progress)
1832 struct page *page = NULL;
1833 struct reclaim_state reclaim_state;
1834 struct task_struct *p = current;
1838 /* We now go into synchronous reclaim */
1839 cpuset_memory_pressure_bump();
1840 p->flags |= PF_MEMALLOC;
1841 lockdep_set_current_reclaim_state(gfp_mask);
1842 reclaim_state.reclaimed_slab = 0;
1843 p->reclaim_state = &reclaim_state;
1845 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1847 p->reclaim_state = NULL;
1848 lockdep_clear_current_reclaim_state();
1849 p->flags &= ~PF_MEMALLOC;
1856 if (likely(*did_some_progress))
1857 page = get_page_from_freelist(gfp_mask, nodemask, order,
1858 zonelist, high_zoneidx,
1859 alloc_flags, preferred_zone,
1865 * This is called in the allocator slow-path if the allocation request is of
1866 * sufficient urgency to ignore watermarks and take other desperate measures
1868 static inline struct page *
1869 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1870 struct zonelist *zonelist, enum zone_type high_zoneidx,
1871 nodemask_t *nodemask, struct zone *preferred_zone,
1877 page = get_page_from_freelist(gfp_mask, nodemask, order,
1878 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1879 preferred_zone, migratetype);
1881 if (!page && gfp_mask & __GFP_NOFAIL)
1882 congestion_wait(BLK_RW_ASYNC, HZ/50);
1883 } while (!page && (gfp_mask & __GFP_NOFAIL));
1889 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1890 enum zone_type high_zoneidx)
1895 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1896 wakeup_kswapd(zone, order);
1900 gfp_to_alloc_flags(gfp_t gfp_mask)
1902 struct task_struct *p = current;
1903 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1904 const gfp_t wait = gfp_mask & __GFP_WAIT;
1906 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1907 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1910 * The caller may dip into page reserves a bit more if the caller
1911 * cannot run direct reclaim, or if the caller has realtime scheduling
1912 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1913 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1915 alloc_flags |= (gfp_mask & __GFP_HIGH);
1918 alloc_flags |= ALLOC_HARDER;
1920 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1921 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1923 alloc_flags &= ~ALLOC_CPUSET;
1924 } else if (unlikely(rt_task(p)) && !in_interrupt())
1925 alloc_flags |= ALLOC_HARDER;
1927 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1928 if (!in_interrupt() &&
1929 ((p->flags & PF_MEMALLOC) ||
1930 unlikely(test_thread_flag(TIF_MEMDIE))))
1931 alloc_flags |= ALLOC_NO_WATERMARKS;
1937 static inline struct page *
1938 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1939 struct zonelist *zonelist, enum zone_type high_zoneidx,
1940 nodemask_t *nodemask, struct zone *preferred_zone,
1943 const gfp_t wait = gfp_mask & __GFP_WAIT;
1944 struct page *page = NULL;
1946 unsigned long pages_reclaimed = 0;
1947 unsigned long did_some_progress;
1948 struct task_struct *p = current;
1951 * In the slowpath, we sanity check order to avoid ever trying to
1952 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1953 * be using allocators in order of preference for an area that is
1956 if (order >= MAX_ORDER) {
1957 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1962 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1963 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1964 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1965 * using a larger set of nodes after it has established that the
1966 * allowed per node queues are empty and that nodes are
1969 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1973 wake_all_kswapd(order, zonelist, high_zoneidx);
1976 * OK, we're below the kswapd watermark and have kicked background
1977 * reclaim. Now things get more complex, so set up alloc_flags according
1978 * to how we want to proceed.
1980 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1982 /* This is the last chance, in general, before the goto nopage. */
1983 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1984 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1985 preferred_zone, migratetype);
1990 /* Allocate without watermarks if the context allows */
1991 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1992 page = __alloc_pages_high_priority(gfp_mask, order,
1993 zonelist, high_zoneidx, nodemask,
1994 preferred_zone, migratetype);
1999 /* Atomic allocations - we can't balance anything */
2003 /* Avoid recursion of direct reclaim */
2004 if (p->flags & PF_MEMALLOC)
2007 /* Avoid allocations with no watermarks from looping endlessly */
2008 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2011 /* Try direct compaction */
2012 page = __alloc_pages_direct_compact(gfp_mask, order,
2013 zonelist, high_zoneidx,
2015 alloc_flags, preferred_zone,
2016 migratetype, &did_some_progress);
2020 /* Try direct reclaim and then allocating */
2021 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2022 zonelist, high_zoneidx,
2024 alloc_flags, preferred_zone,
2025 migratetype, &did_some_progress);
2030 * If we failed to make any progress reclaiming, then we are
2031 * running out of options and have to consider going OOM
2033 if (!did_some_progress) {
2034 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2035 if (oom_killer_disabled)
2037 page = __alloc_pages_may_oom(gfp_mask, order,
2038 zonelist, high_zoneidx,
2039 nodemask, preferred_zone,
2045 * The OOM killer does not trigger for high-order
2046 * ~__GFP_NOFAIL allocations so if no progress is being
2047 * made, there are no other options and retrying is
2050 if (order > PAGE_ALLOC_COSTLY_ORDER &&
2051 !(gfp_mask & __GFP_NOFAIL))
2058 /* Check if we should retry the allocation */
2059 pages_reclaimed += did_some_progress;
2060 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2061 /* Wait for some write requests to complete then retry */
2062 congestion_wait(BLK_RW_ASYNC, HZ/50);
2067 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
2068 printk(KERN_WARNING "%s: page allocation failure."
2069 " order:%d, mode:0x%x\n",
2070 p->comm, order, gfp_mask);
2076 if (kmemcheck_enabled)
2077 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2083 * This is the 'heart' of the zoned buddy allocator.
2086 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2087 struct zonelist *zonelist, nodemask_t *nodemask)
2089 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2090 struct zone *preferred_zone;
2092 int migratetype = allocflags_to_migratetype(gfp_mask);
2094 gfp_mask &= gfp_allowed_mask;
2096 lockdep_trace_alloc(gfp_mask);
2098 might_sleep_if(gfp_mask & __GFP_WAIT);
2100 if (should_fail_alloc_page(gfp_mask, order))
2104 * Check the zones suitable for the gfp_mask contain at least one
2105 * valid zone. It's possible to have an empty zonelist as a result
2106 * of GFP_THISNODE and a memoryless node
2108 if (unlikely(!zonelist->_zonerefs->zone))
2112 /* The preferred zone is used for statistics later */
2113 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
2114 if (!preferred_zone) {
2119 /* First allocation attempt */
2120 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2121 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2122 preferred_zone, migratetype);
2123 if (unlikely(!page))
2124 page = __alloc_pages_slowpath(gfp_mask, order,
2125 zonelist, high_zoneidx, nodemask,
2126 preferred_zone, migratetype);
2129 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2132 EXPORT_SYMBOL(__alloc_pages_nodemask);
2135 * Common helper functions.
2137 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2142 * __get_free_pages() returns a 32-bit address, which cannot represent
2145 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2147 page = alloc_pages(gfp_mask, order);
2150 return (unsigned long) page_address(page);
2152 EXPORT_SYMBOL(__get_free_pages);
2154 unsigned long get_zeroed_page(gfp_t gfp_mask)
2156 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2158 EXPORT_SYMBOL(get_zeroed_page);
2160 void __pagevec_free(struct pagevec *pvec)
2162 int i = pagevec_count(pvec);
2165 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2166 free_hot_cold_page(pvec->pages[i], pvec->cold);
2170 void __free_pages(struct page *page, unsigned int order)
2172 if (put_page_testzero(page)) {
2174 free_hot_cold_page(page, 0);
2176 __free_pages_ok(page, order);
2180 EXPORT_SYMBOL(__free_pages);
2182 void free_pages(unsigned long addr, unsigned int order)
2185 VM_BUG_ON(!virt_addr_valid((void *)addr));
2186 __free_pages(virt_to_page((void *)addr), order);
2190 EXPORT_SYMBOL(free_pages);
2193 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2194 * @size: the number of bytes to allocate
2195 * @gfp_mask: GFP flags for the allocation
2197 * This function is similar to alloc_pages(), except that it allocates the
2198 * minimum number of pages to satisfy the request. alloc_pages() can only
2199 * allocate memory in power-of-two pages.
2201 * This function is also limited by MAX_ORDER.
2203 * Memory allocated by this function must be released by free_pages_exact().
2205 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2207 unsigned int order = get_order(size);
2210 addr = __get_free_pages(gfp_mask, order);
2212 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2213 unsigned long used = addr + PAGE_ALIGN(size);
2215 split_page(virt_to_page((void *)addr), order);
2216 while (used < alloc_end) {
2222 return (void *)addr;
2224 EXPORT_SYMBOL(alloc_pages_exact);
2227 * free_pages_exact - release memory allocated via alloc_pages_exact()
2228 * @virt: the value returned by alloc_pages_exact.
2229 * @size: size of allocation, same value as passed to alloc_pages_exact().
2231 * Release the memory allocated by a previous call to alloc_pages_exact.
2233 void free_pages_exact(void *virt, size_t size)
2235 unsigned long addr = (unsigned long)virt;
2236 unsigned long end = addr + PAGE_ALIGN(size);
2238 while (addr < end) {
2243 EXPORT_SYMBOL(free_pages_exact);
2245 static unsigned int nr_free_zone_pages(int offset)
2250 /* Just pick one node, since fallback list is circular */
2251 unsigned int sum = 0;
2253 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2255 for_each_zone_zonelist(zone, z, zonelist, offset) {
2256 unsigned long size = zone->present_pages;
2257 unsigned long high = high_wmark_pages(zone);
2266 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2268 unsigned int nr_free_buffer_pages(void)
2270 return nr_free_zone_pages(gfp_zone(GFP_USER));
2272 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2275 * Amount of free RAM allocatable within all zones
2277 unsigned int nr_free_pagecache_pages(void)
2279 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2282 static inline void show_node(struct zone *zone)
2285 printk("Node %d ", zone_to_nid(zone));
2288 void si_meminfo(struct sysinfo *val)
2290 val->totalram = totalram_pages;
2292 val->freeram = global_page_state(NR_FREE_PAGES);
2293 val->bufferram = nr_blockdev_pages();
2294 val->totalhigh = totalhigh_pages;
2295 val->freehigh = nr_free_highpages();
2296 val->mem_unit = PAGE_SIZE;
2299 EXPORT_SYMBOL(si_meminfo);
2302 void si_meminfo_node(struct sysinfo *val, int nid)
2304 pg_data_t *pgdat = NODE_DATA(nid);
2306 val->totalram = pgdat->node_present_pages;
2307 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2308 #ifdef CONFIG_HIGHMEM
2309 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2310 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2316 val->mem_unit = PAGE_SIZE;
2320 #define K(x) ((x) << (PAGE_SHIFT-10))
2323 * Show free area list (used inside shift_scroll-lock stuff)
2324 * We also calculate the percentage fragmentation. We do this by counting the
2325 * memory on each free list with the exception of the first item on the list.
2327 void show_free_areas(void)
2332 for_each_populated_zone(zone) {
2334 printk("%s per-cpu:\n", zone->name);
2336 for_each_online_cpu(cpu) {
2337 struct per_cpu_pageset *pageset;
2339 pageset = per_cpu_ptr(zone->pageset, cpu);
2341 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2342 cpu, pageset->pcp.high,
2343 pageset->pcp.batch, pageset->pcp.count);
2347 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2348 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2350 " dirty:%lu writeback:%lu unstable:%lu\n"
2351 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2352 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2353 global_page_state(NR_ACTIVE_ANON),
2354 global_page_state(NR_INACTIVE_ANON),
2355 global_page_state(NR_ISOLATED_ANON),
2356 global_page_state(NR_ACTIVE_FILE),
2357 global_page_state(NR_INACTIVE_FILE),
2358 global_page_state(NR_ISOLATED_FILE),
2359 global_page_state(NR_UNEVICTABLE),
2360 global_page_state(NR_FILE_DIRTY),
2361 global_page_state(NR_WRITEBACK),
2362 global_page_state(NR_UNSTABLE_NFS),
2363 global_page_state(NR_FREE_PAGES),
2364 global_page_state(NR_SLAB_RECLAIMABLE),
2365 global_page_state(NR_SLAB_UNRECLAIMABLE),
2366 global_page_state(NR_FILE_MAPPED),
2367 global_page_state(NR_SHMEM),
2368 global_page_state(NR_PAGETABLE),
2369 global_page_state(NR_BOUNCE));
2371 for_each_populated_zone(zone) {
2380 " active_anon:%lukB"
2381 " inactive_anon:%lukB"
2382 " active_file:%lukB"
2383 " inactive_file:%lukB"
2384 " unevictable:%lukB"
2385 " isolated(anon):%lukB"
2386 " isolated(file):%lukB"
2393 " slab_reclaimable:%lukB"
2394 " slab_unreclaimable:%lukB"
2395 " kernel_stack:%lukB"
2399 " writeback_tmp:%lukB"
2400 " pages_scanned:%lu"
2401 " all_unreclaimable? %s"
2404 K(zone_page_state(zone, NR_FREE_PAGES)),
2405 K(min_wmark_pages(zone)),
2406 K(low_wmark_pages(zone)),
2407 K(high_wmark_pages(zone)),
2408 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2409 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2410 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2411 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2412 K(zone_page_state(zone, NR_UNEVICTABLE)),
2413 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2414 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2415 K(zone->present_pages),
2416 K(zone_page_state(zone, NR_MLOCK)),
2417 K(zone_page_state(zone, NR_FILE_DIRTY)),
2418 K(zone_page_state(zone, NR_WRITEBACK)),
2419 K(zone_page_state(zone, NR_FILE_MAPPED)),
2420 K(zone_page_state(zone, NR_SHMEM)),
2421 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2422 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2423 zone_page_state(zone, NR_KERNEL_STACK) *
2425 K(zone_page_state(zone, NR_PAGETABLE)),
2426 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2427 K(zone_page_state(zone, NR_BOUNCE)),
2428 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2429 zone->pages_scanned,
2430 (zone->all_unreclaimable ? "yes" : "no")
2432 printk("lowmem_reserve[]:");
2433 for (i = 0; i < MAX_NR_ZONES; i++)
2434 printk(" %lu", zone->lowmem_reserve[i]);
2438 for_each_populated_zone(zone) {
2439 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2442 printk("%s: ", zone->name);
2444 spin_lock_irqsave(&zone->lock, flags);
2445 for (order = 0; order < MAX_ORDER; order++) {
2446 nr[order] = zone->free_area[order].nr_free;
2447 total += nr[order] << order;
2449 spin_unlock_irqrestore(&zone->lock, flags);
2450 for (order = 0; order < MAX_ORDER; order++)
2451 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2452 printk("= %lukB\n", K(total));
2455 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2457 show_swap_cache_info();
2460 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2462 zoneref->zone = zone;
2463 zoneref->zone_idx = zone_idx(zone);
2467 * Builds allocation fallback zone lists.
2469 * Add all populated zones of a node to the zonelist.
2471 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2472 int nr_zones, enum zone_type zone_type)
2476 BUG_ON(zone_type >= MAX_NR_ZONES);
2481 zone = pgdat->node_zones + zone_type;
2482 if (populated_zone(zone)) {
2483 zoneref_set_zone(zone,
2484 &zonelist->_zonerefs[nr_zones++]);
2485 check_highest_zone(zone_type);
2488 } while (zone_type);
2495 * 0 = automatic detection of better ordering.
2496 * 1 = order by ([node] distance, -zonetype)
2497 * 2 = order by (-zonetype, [node] distance)
2499 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2500 * the same zonelist. So only NUMA can configure this param.
2502 #define ZONELIST_ORDER_DEFAULT 0
2503 #define ZONELIST_ORDER_NODE 1
2504 #define ZONELIST_ORDER_ZONE 2
2506 /* zonelist order in the kernel.
2507 * set_zonelist_order() will set this to NODE or ZONE.
2509 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2510 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2514 /* The value user specified ....changed by config */
2515 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2516 /* string for sysctl */
2517 #define NUMA_ZONELIST_ORDER_LEN 16
2518 char numa_zonelist_order[16] = "default";
2521 * interface for configure zonelist ordering.
2522 * command line option "numa_zonelist_order"
2523 * = "[dD]efault - default, automatic configuration.
2524 * = "[nN]ode - order by node locality, then by zone within node
2525 * = "[zZ]one - order by zone, then by locality within zone
2528 static int __parse_numa_zonelist_order(char *s)
2530 if (*s == 'd' || *s == 'D') {
2531 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2532 } else if (*s == 'n' || *s == 'N') {
2533 user_zonelist_order = ZONELIST_ORDER_NODE;
2534 } else if (*s == 'z' || *s == 'Z') {
2535 user_zonelist_order = ZONELIST_ORDER_ZONE;
2538 "Ignoring invalid numa_zonelist_order value: "
2545 static __init int setup_numa_zonelist_order(char *s)
2548 return __parse_numa_zonelist_order(s);
2551 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2554 * sysctl handler for numa_zonelist_order
2556 int numa_zonelist_order_handler(ctl_table *table, int write,
2557 void __user *buffer, size_t *length,
2560 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2562 static DEFINE_MUTEX(zl_order_mutex);
2564 mutex_lock(&zl_order_mutex);
2566 strcpy(saved_string, (char*)table->data);
2567 ret = proc_dostring(table, write, buffer, length, ppos);
2571 int oldval = user_zonelist_order;
2572 if (__parse_numa_zonelist_order((char*)table->data)) {
2574 * bogus value. restore saved string
2576 strncpy((char*)table->data, saved_string,
2577 NUMA_ZONELIST_ORDER_LEN);
2578 user_zonelist_order = oldval;
2579 } else if (oldval != user_zonelist_order) {
2580 mutex_lock(&zonelists_mutex);
2581 build_all_zonelists(NULL);
2582 mutex_unlock(&zonelists_mutex);
2586 mutex_unlock(&zl_order_mutex);
2591 #define MAX_NODE_LOAD (nr_online_nodes)
2592 static int node_load[MAX_NUMNODES];
2595 * find_next_best_node - find the next node that should appear in a given node's fallback list
2596 * @node: node whose fallback list we're appending
2597 * @used_node_mask: nodemask_t of already used nodes
2599 * We use a number of factors to determine which is the next node that should
2600 * appear on a given node's fallback list. The node should not have appeared
2601 * already in @node's fallback list, and it should be the next closest node
2602 * according to the distance array (which contains arbitrary distance values
2603 * from each node to each node in the system), and should also prefer nodes
2604 * with no CPUs, since presumably they'll have very little allocation pressure
2605 * on them otherwise.
2606 * It returns -1 if no node is found.
2608 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2611 int min_val = INT_MAX;
2613 const struct cpumask *tmp = cpumask_of_node(0);
2615 /* Use the local node if we haven't already */
2616 if (!node_isset(node, *used_node_mask)) {
2617 node_set(node, *used_node_mask);
2621 for_each_node_state(n, N_HIGH_MEMORY) {
2623 /* Don't want a node to appear more than once */
2624 if (node_isset(n, *used_node_mask))
2627 /* Use the distance array to find the distance */
2628 val = node_distance(node, n);
2630 /* Penalize nodes under us ("prefer the next node") */
2633 /* Give preference to headless and unused nodes */
2634 tmp = cpumask_of_node(n);
2635 if (!cpumask_empty(tmp))
2636 val += PENALTY_FOR_NODE_WITH_CPUS;
2638 /* Slight preference for less loaded node */
2639 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2640 val += node_load[n];
2642 if (val < min_val) {
2649 node_set(best_node, *used_node_mask);
2656 * Build zonelists ordered by node and zones within node.
2657 * This results in maximum locality--normal zone overflows into local
2658 * DMA zone, if any--but risks exhausting DMA zone.
2660 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2663 struct zonelist *zonelist;
2665 zonelist = &pgdat->node_zonelists[0];
2666 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2668 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2670 zonelist->_zonerefs[j].zone = NULL;
2671 zonelist->_zonerefs[j].zone_idx = 0;
2675 * Build gfp_thisnode zonelists
2677 static void build_thisnode_zonelists(pg_data_t *pgdat)
2680 struct zonelist *zonelist;
2682 zonelist = &pgdat->node_zonelists[1];
2683 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2684 zonelist->_zonerefs[j].zone = NULL;
2685 zonelist->_zonerefs[j].zone_idx = 0;
2689 * Build zonelists ordered by zone and nodes within zones.
2690 * This results in conserving DMA zone[s] until all Normal memory is
2691 * exhausted, but results in overflowing to remote node while memory
2692 * may still exist in local DMA zone.
2694 static int node_order[MAX_NUMNODES];
2696 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2699 int zone_type; /* needs to be signed */
2701 struct zonelist *zonelist;
2703 zonelist = &pgdat->node_zonelists[0];
2705 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2706 for (j = 0; j < nr_nodes; j++) {
2707 node = node_order[j];
2708 z = &NODE_DATA(node)->node_zones[zone_type];
2709 if (populated_zone(z)) {
2711 &zonelist->_zonerefs[pos++]);
2712 check_highest_zone(zone_type);
2716 zonelist->_zonerefs[pos].zone = NULL;
2717 zonelist->_zonerefs[pos].zone_idx = 0;
2720 static int default_zonelist_order(void)
2723 unsigned long low_kmem_size,total_size;
2727 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2728 * If they are really small and used heavily, the system can fall
2729 * into OOM very easily.
2730 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2732 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2735 for_each_online_node(nid) {
2736 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2737 z = &NODE_DATA(nid)->node_zones[zone_type];
2738 if (populated_zone(z)) {
2739 if (zone_type < ZONE_NORMAL)
2740 low_kmem_size += z->present_pages;
2741 total_size += z->present_pages;
2742 } else if (zone_type == ZONE_NORMAL) {
2744 * If any node has only lowmem, then node order
2745 * is preferred to allow kernel allocations
2746 * locally; otherwise, they can easily infringe
2747 * on other nodes when there is an abundance of
2748 * lowmem available to allocate from.
2750 return ZONELIST_ORDER_NODE;
2754 if (!low_kmem_size || /* there are no DMA area. */
2755 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2756 return ZONELIST_ORDER_NODE;
2758 * look into each node's config.
2759 * If there is a node whose DMA/DMA32 memory is very big area on
2760 * local memory, NODE_ORDER may be suitable.
2762 average_size = total_size /
2763 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2764 for_each_online_node(nid) {
2767 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2768 z = &NODE_DATA(nid)->node_zones[zone_type];
2769 if (populated_zone(z)) {
2770 if (zone_type < ZONE_NORMAL)
2771 low_kmem_size += z->present_pages;
2772 total_size += z->present_pages;
2775 if (low_kmem_size &&
2776 total_size > average_size && /* ignore small node */
2777 low_kmem_size > total_size * 70/100)
2778 return ZONELIST_ORDER_NODE;
2780 return ZONELIST_ORDER_ZONE;
2783 static void set_zonelist_order(void)
2785 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2786 current_zonelist_order = default_zonelist_order();
2788 current_zonelist_order = user_zonelist_order;
2791 static void build_zonelists(pg_data_t *pgdat)
2795 nodemask_t used_mask;
2796 int local_node, prev_node;
2797 struct zonelist *zonelist;
2798 int order = current_zonelist_order;
2800 /* initialize zonelists */
2801 for (i = 0; i < MAX_ZONELISTS; i++) {
2802 zonelist = pgdat->node_zonelists + i;
2803 zonelist->_zonerefs[0].zone = NULL;
2804 zonelist->_zonerefs[0].zone_idx = 0;
2807 /* NUMA-aware ordering of nodes */
2808 local_node = pgdat->node_id;
2809 load = nr_online_nodes;
2810 prev_node = local_node;
2811 nodes_clear(used_mask);
2813 memset(node_order, 0, sizeof(node_order));
2816 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2817 int distance = node_distance(local_node, node);
2820 * If another node is sufficiently far away then it is better
2821 * to reclaim pages in a zone before going off node.
2823 if (distance > RECLAIM_DISTANCE)
2824 zone_reclaim_mode = 1;
2827 * We don't want to pressure a particular node.
2828 * So adding penalty to the first node in same
2829 * distance group to make it round-robin.
2831 if (distance != node_distance(local_node, prev_node))
2832 node_load[node] = load;
2836 if (order == ZONELIST_ORDER_NODE)
2837 build_zonelists_in_node_order(pgdat, node);
2839 node_order[j++] = node; /* remember order */
2842 if (order == ZONELIST_ORDER_ZONE) {
2843 /* calculate node order -- i.e., DMA last! */
2844 build_zonelists_in_zone_order(pgdat, j);
2847 build_thisnode_zonelists(pgdat);
2850 /* Construct the zonelist performance cache - see further mmzone.h */
2851 static void build_zonelist_cache(pg_data_t *pgdat)
2853 struct zonelist *zonelist;
2854 struct zonelist_cache *zlc;
2857 zonelist = &pgdat->node_zonelists[0];
2858 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2859 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2860 for (z = zonelist->_zonerefs; z->zone; z++)
2861 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2865 #else /* CONFIG_NUMA */
2867 static void set_zonelist_order(void)
2869 current_zonelist_order = ZONELIST_ORDER_ZONE;
2872 static void build_zonelists(pg_data_t *pgdat)
2874 int node, local_node;
2876 struct zonelist *zonelist;
2878 local_node = pgdat->node_id;
2880 zonelist = &pgdat->node_zonelists[0];
2881 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2884 * Now we build the zonelist so that it contains the zones
2885 * of all the other nodes.
2886 * We don't want to pressure a particular node, so when
2887 * building the zones for node N, we make sure that the
2888 * zones coming right after the local ones are those from
2889 * node N+1 (modulo N)
2891 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2892 if (!node_online(node))
2894 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2897 for (node = 0; node < local_node; node++) {
2898 if (!node_online(node))
2900 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2904 zonelist->_zonerefs[j].zone = NULL;
2905 zonelist->_zonerefs[j].zone_idx = 0;
2908 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2909 static void build_zonelist_cache(pg_data_t *pgdat)
2911 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2914 #endif /* CONFIG_NUMA */
2917 * Boot pageset table. One per cpu which is going to be used for all
2918 * zones and all nodes. The parameters will be set in such a way
2919 * that an item put on a list will immediately be handed over to
2920 * the buddy list. This is safe since pageset manipulation is done
2921 * with interrupts disabled.
2923 * The boot_pagesets must be kept even after bootup is complete for
2924 * unused processors and/or zones. They do play a role for bootstrapping
2925 * hotplugged processors.
2927 * zoneinfo_show() and maybe other functions do
2928 * not check if the processor is online before following the pageset pointer.
2929 * Other parts of the kernel may not check if the zone is available.
2931 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
2932 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
2933 static void setup_zone_pageset(struct zone *zone);
2936 * Global mutex to protect against size modification of zonelists
2937 * as well as to serialize pageset setup for the new populated zone.
2939 DEFINE_MUTEX(zonelists_mutex);
2941 /* return values int ....just for stop_machine() */
2942 static __init_refok int __build_all_zonelists(void *data)
2948 memset(node_load, 0, sizeof(node_load));
2950 for_each_online_node(nid) {
2951 pg_data_t *pgdat = NODE_DATA(nid);
2953 build_zonelists(pgdat);
2954 build_zonelist_cache(pgdat);
2957 #ifdef CONFIG_MEMORY_HOTPLUG
2958 /* Setup real pagesets for the new zone */
2960 struct zone *zone = data;
2961 setup_zone_pageset(zone);
2966 * Initialize the boot_pagesets that are going to be used
2967 * for bootstrapping processors. The real pagesets for
2968 * each zone will be allocated later when the per cpu
2969 * allocator is available.
2971 * boot_pagesets are used also for bootstrapping offline
2972 * cpus if the system is already booted because the pagesets
2973 * are needed to initialize allocators on a specific cpu too.
2974 * F.e. the percpu allocator needs the page allocator which
2975 * needs the percpu allocator in order to allocate its pagesets
2976 * (a chicken-egg dilemma).
2978 for_each_possible_cpu(cpu)
2979 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
2985 * Called with zonelists_mutex held always
2986 * unless system_state == SYSTEM_BOOTING.
2988 void build_all_zonelists(void *data)
2990 set_zonelist_order();
2992 if (system_state == SYSTEM_BOOTING) {
2993 __build_all_zonelists(NULL);
2994 mminit_verify_zonelist();
2995 cpuset_init_current_mems_allowed();
2997 /* we have to stop all cpus to guarantee there is no user
2999 stop_machine(__build_all_zonelists, data, NULL);
3000 /* cpuset refresh routine should be here */
3002 vm_total_pages = nr_free_pagecache_pages();
3004 * Disable grouping by mobility if the number of pages in the
3005 * system is too low to allow the mechanism to work. It would be
3006 * more accurate, but expensive to check per-zone. This check is
3007 * made on memory-hotadd so a system can start with mobility
3008 * disabled and enable it later
3010 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3011 page_group_by_mobility_disabled = 1;
3013 page_group_by_mobility_disabled = 0;
3015 printk("Built %i zonelists in %s order, mobility grouping %s. "
3016 "Total pages: %ld\n",
3018 zonelist_order_name[current_zonelist_order],
3019 page_group_by_mobility_disabled ? "off" : "on",
3022 printk("Policy zone: %s\n", zone_names[policy_zone]);
3027 * Helper functions to size the waitqueue hash table.
3028 * Essentially these want to choose hash table sizes sufficiently
3029 * large so that collisions trying to wait on pages are rare.
3030 * But in fact, the number of active page waitqueues on typical
3031 * systems is ridiculously low, less than 200. So this is even
3032 * conservative, even though it seems large.
3034 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3035 * waitqueues, i.e. the size of the waitq table given the number of pages.
3037 #define PAGES_PER_WAITQUEUE 256
3039 #ifndef CONFIG_MEMORY_HOTPLUG
3040 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3042 unsigned long size = 1;
3044 pages /= PAGES_PER_WAITQUEUE;
3046 while (size < pages)
3050 * Once we have dozens or even hundreds of threads sleeping
3051 * on IO we've got bigger problems than wait queue collision.
3052 * Limit the size of the wait table to a reasonable size.
3054 size = min(size, 4096UL);
3056 return max(size, 4UL);
3060 * A zone's size might be changed by hot-add, so it is not possible to determine
3061 * a suitable size for its wait_table. So we use the maximum size now.
3063 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3065 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3066 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3067 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3069 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3070 * or more by the traditional way. (See above). It equals:
3072 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3073 * ia64(16K page size) : = ( 8G + 4M)byte.
3074 * powerpc (64K page size) : = (32G +16M)byte.
3076 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3083 * This is an integer logarithm so that shifts can be used later
3084 * to extract the more random high bits from the multiplicative
3085 * hash function before the remainder is taken.
3087 static inline unsigned long wait_table_bits(unsigned long size)
3092 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3095 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3096 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3097 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3098 * higher will lead to a bigger reserve which will get freed as contiguous
3099 * blocks as reclaim kicks in
3101 static void setup_zone_migrate_reserve(struct zone *zone)
3103 unsigned long start_pfn, pfn, end_pfn;
3105 unsigned long block_migratetype;
3108 /* Get the start pfn, end pfn and the number of blocks to reserve */
3109 start_pfn = zone->zone_start_pfn;
3110 end_pfn = start_pfn + zone->spanned_pages;
3111 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3115 * Reserve blocks are generally in place to help high-order atomic
3116 * allocations that are short-lived. A min_free_kbytes value that
3117 * would result in more than 2 reserve blocks for atomic allocations
3118 * is assumed to be in place to help anti-fragmentation for the
3119 * future allocation of hugepages at runtime.
3121 reserve = min(2, reserve);
3123 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3124 if (!pfn_valid(pfn))
3126 page = pfn_to_page(pfn);
3128 /* Watch out for overlapping nodes */
3129 if (page_to_nid(page) != zone_to_nid(zone))
3132 /* Blocks with reserved pages will never free, skip them. */
3133 if (PageReserved(page))
3136 block_migratetype = get_pageblock_migratetype(page);
3138 /* If this block is reserved, account for it */
3139 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3144 /* Suitable for reserving if this block is movable */
3145 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3146 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3147 move_freepages_block(zone, page, MIGRATE_RESERVE);
3153 * If the reserve is met and this is a previous reserved block,
3156 if (block_migratetype == MIGRATE_RESERVE) {
3157 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3158 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3164 * Initially all pages are reserved - free ones are freed
3165 * up by free_all_bootmem() once the early boot process is
3166 * done. Non-atomic initialization, single-pass.
3168 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3169 unsigned long start_pfn, enum memmap_context context)
3172 unsigned long end_pfn = start_pfn + size;
3176 if (highest_memmap_pfn < end_pfn - 1)
3177 highest_memmap_pfn = end_pfn - 1;
3179 z = &NODE_DATA(nid)->node_zones[zone];
3180 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3182 * There can be holes in boot-time mem_map[]s
3183 * handed to this function. They do not
3184 * exist on hotplugged memory.
3186 if (context == MEMMAP_EARLY) {
3187 if (!early_pfn_valid(pfn))
3189 if (!early_pfn_in_nid(pfn, nid))
3192 page = pfn_to_page(pfn);
3193 set_page_links(page, zone, nid, pfn);
3194 mminit_verify_page_links(page, zone, nid, pfn);
3195 init_page_count(page);
3196 reset_page_mapcount(page);
3197 SetPageReserved(page);
3199 * Mark the block movable so that blocks are reserved for
3200 * movable at startup. This will force kernel allocations
3201 * to reserve their blocks rather than leaking throughout
3202 * the address space during boot when many long-lived
3203 * kernel allocations are made. Later some blocks near
3204 * the start are marked MIGRATE_RESERVE by
3205 * setup_zone_migrate_reserve()
3207 * bitmap is created for zone's valid pfn range. but memmap
3208 * can be created for invalid pages (for alignment)
3209 * check here not to call set_pageblock_migratetype() against
3212 if ((z->zone_start_pfn <= pfn)
3213 && (pfn < z->zone_start_pfn + z->spanned_pages)
3214 && !(pfn & (pageblock_nr_pages - 1)))
3215 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3217 INIT_LIST_HEAD(&page->lru);
3218 #ifdef WANT_PAGE_VIRTUAL
3219 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3220 if (!is_highmem_idx(zone))
3221 set_page_address(page, __va(pfn << PAGE_SHIFT));
3226 static void __meminit zone_init_free_lists(struct zone *zone)
3229 for_each_migratetype_order(order, t) {
3230 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3231 zone->free_area[order].nr_free = 0;
3235 #ifndef __HAVE_ARCH_MEMMAP_INIT
3236 #define memmap_init(size, nid, zone, start_pfn) \
3237 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3240 static int zone_batchsize(struct zone *zone)
3246 * The per-cpu-pages pools are set to around 1000th of the
3247 * size of the zone. But no more than 1/2 of a meg.
3249 * OK, so we don't know how big the cache is. So guess.
3251 batch = zone->present_pages / 1024;
3252 if (batch * PAGE_SIZE > 512 * 1024)
3253 batch = (512 * 1024) / PAGE_SIZE;
3254 batch /= 4; /* We effectively *= 4 below */
3259 * Clamp the batch to a 2^n - 1 value. Having a power
3260 * of 2 value was found to be more likely to have
3261 * suboptimal cache aliasing properties in some cases.
3263 * For example if 2 tasks are alternately allocating
3264 * batches of pages, one task can end up with a lot
3265 * of pages of one half of the possible page colors
3266 * and the other with pages of the other colors.
3268 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3273 /* The deferral and batching of frees should be suppressed under NOMMU
3276 * The problem is that NOMMU needs to be able to allocate large chunks
3277 * of contiguous memory as there's no hardware page translation to
3278 * assemble apparent contiguous memory from discontiguous pages.
3280 * Queueing large contiguous runs of pages for batching, however,
3281 * causes the pages to actually be freed in smaller chunks. As there
3282 * can be a significant delay between the individual batches being
3283 * recycled, this leads to the once large chunks of space being
3284 * fragmented and becoming unavailable for high-order allocations.
3290 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3292 struct per_cpu_pages *pcp;
3295 memset(p, 0, sizeof(*p));
3299 pcp->high = 6 * batch;
3300 pcp->batch = max(1UL, 1 * batch);
3301 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3302 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3306 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3307 * to the value high for the pageset p.
3310 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3313 struct per_cpu_pages *pcp;
3317 pcp->batch = max(1UL, high/4);
3318 if ((high/4) > (PAGE_SHIFT * 8))
3319 pcp->batch = PAGE_SHIFT * 8;
3322 static __meminit void setup_zone_pageset(struct zone *zone)
3326 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3328 for_each_possible_cpu(cpu) {
3329 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3331 setup_pageset(pcp, zone_batchsize(zone));
3333 if (percpu_pagelist_fraction)
3334 setup_pagelist_highmark(pcp,
3335 (zone->present_pages /
3336 percpu_pagelist_fraction));
3341 * Allocate per cpu pagesets and initialize them.
3342 * Before this call only boot pagesets were available.
3344 void __init setup_per_cpu_pageset(void)
3348 for_each_populated_zone(zone)
3349 setup_zone_pageset(zone);
3352 static noinline __init_refok
3353 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3356 struct pglist_data *pgdat = zone->zone_pgdat;
3360 * The per-page waitqueue mechanism uses hashed waitqueues
3363 zone->wait_table_hash_nr_entries =
3364 wait_table_hash_nr_entries(zone_size_pages);
3365 zone->wait_table_bits =
3366 wait_table_bits(zone->wait_table_hash_nr_entries);
3367 alloc_size = zone->wait_table_hash_nr_entries
3368 * sizeof(wait_queue_head_t);
3370 if (!slab_is_available()) {
3371 zone->wait_table = (wait_queue_head_t *)
3372 alloc_bootmem_node(pgdat, alloc_size);
3375 * This case means that a zone whose size was 0 gets new memory
3376 * via memory hot-add.
3377 * But it may be the case that a new node was hot-added. In
3378 * this case vmalloc() will not be able to use this new node's
3379 * memory - this wait_table must be initialized to use this new
3380 * node itself as well.
3381 * To use this new node's memory, further consideration will be
3384 zone->wait_table = vmalloc(alloc_size);
3386 if (!zone->wait_table)
3389 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3390 init_waitqueue_head(zone->wait_table + i);
3395 static int __zone_pcp_update(void *data)
3397 struct zone *zone = data;
3399 unsigned long batch = zone_batchsize(zone), flags;
3401 for_each_possible_cpu(cpu) {
3402 struct per_cpu_pageset *pset;
3403 struct per_cpu_pages *pcp;
3405 pset = per_cpu_ptr(zone->pageset, cpu);
3408 local_irq_save(flags);
3409 free_pcppages_bulk(zone, pcp->count, pcp);
3410 setup_pageset(pset, batch);
3411 local_irq_restore(flags);
3416 void zone_pcp_update(struct zone *zone)
3418 stop_machine(__zone_pcp_update, zone, NULL);
3421 static __meminit void zone_pcp_init(struct zone *zone)
3424 * per cpu subsystem is not up at this point. The following code
3425 * relies on the ability of the linker to provide the
3426 * offset of a (static) per cpu variable into the per cpu area.
3428 zone->pageset = &boot_pageset;
3430 if (zone->present_pages)
3431 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3432 zone->name, zone->present_pages,
3433 zone_batchsize(zone));
3436 __meminit int init_currently_empty_zone(struct zone *zone,
3437 unsigned long zone_start_pfn,
3439 enum memmap_context context)
3441 struct pglist_data *pgdat = zone->zone_pgdat;
3443 ret = zone_wait_table_init(zone, size);
3446 pgdat->nr_zones = zone_idx(zone) + 1;
3448 zone->zone_start_pfn = zone_start_pfn;
3450 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3451 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3453 (unsigned long)zone_idx(zone),
3454 zone_start_pfn, (zone_start_pfn + size));
3456 zone_init_free_lists(zone);
3461 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3463 * Basic iterator support. Return the first range of PFNs for a node
3464 * Note: nid == MAX_NUMNODES returns first region regardless of node
3466 static int __meminit first_active_region_index_in_nid(int nid)
3470 for (i = 0; i < nr_nodemap_entries; i++)
3471 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3478 * Basic iterator support. Return the next active range of PFNs for a node
3479 * Note: nid == MAX_NUMNODES returns next region regardless of node
3481 static int __meminit next_active_region_index_in_nid(int index, int nid)
3483 for (index = index + 1; index < nr_nodemap_entries; index++)
3484 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3490 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3492 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3493 * Architectures may implement their own version but if add_active_range()
3494 * was used and there are no special requirements, this is a convenient
3497 int __meminit __early_pfn_to_nid(unsigned long pfn)
3501 for (i = 0; i < nr_nodemap_entries; i++) {
3502 unsigned long start_pfn = early_node_map[i].start_pfn;
3503 unsigned long end_pfn = early_node_map[i].end_pfn;
3505 if (start_pfn <= pfn && pfn < end_pfn)
3506 return early_node_map[i].nid;
3508 /* This is a memory hole */
3511 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3513 int __meminit early_pfn_to_nid(unsigned long pfn)
3517 nid = __early_pfn_to_nid(pfn);
3520 /* just returns 0 */
3524 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3525 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3529 nid = __early_pfn_to_nid(pfn);
3530 if (nid >= 0 && nid != node)
3536 /* Basic iterator support to walk early_node_map[] */
3537 #define for_each_active_range_index_in_nid(i, nid) \
3538 for (i = first_active_region_index_in_nid(nid); i != -1; \
3539 i = next_active_region_index_in_nid(i, nid))
3542 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3543 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3544 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3546 * If an architecture guarantees that all ranges registered with
3547 * add_active_ranges() contain no holes and may be freed, this
3548 * this function may be used instead of calling free_bootmem() manually.
3550 void __init free_bootmem_with_active_regions(int nid,
3551 unsigned long max_low_pfn)
3555 for_each_active_range_index_in_nid(i, nid) {
3556 unsigned long size_pages = 0;
3557 unsigned long end_pfn = early_node_map[i].end_pfn;
3559 if (early_node_map[i].start_pfn >= max_low_pfn)
3562 if (end_pfn > max_low_pfn)
3563 end_pfn = max_low_pfn;
3565 size_pages = end_pfn - early_node_map[i].start_pfn;
3566 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3567 PFN_PHYS(early_node_map[i].start_pfn),
3568 size_pages << PAGE_SHIFT);
3572 int __init add_from_early_node_map(struct range *range, int az,
3573 int nr_range, int nid)
3578 /* need to go over early_node_map to find out good range for node */
3579 for_each_active_range_index_in_nid(i, nid) {
3580 start = early_node_map[i].start_pfn;
3581 end = early_node_map[i].end_pfn;
3582 nr_range = add_range(range, az, nr_range, start, end);
3587 #ifdef CONFIG_NO_BOOTMEM
3588 void * __init __alloc_memory_core_early(int nid, u64 size, u64 align,
3589 u64 goal, u64 limit)
3594 /* need to go over early_node_map to find out good range for node */
3595 for_each_active_range_index_in_nid(i, nid) {
3597 u64 ei_start, ei_last;
3599 ei_last = early_node_map[i].end_pfn;
3600 ei_last <<= PAGE_SHIFT;
3601 ei_start = early_node_map[i].start_pfn;
3602 ei_start <<= PAGE_SHIFT;
3603 addr = find_early_area(ei_start, ei_last,
3604 goal, limit, size, align);
3610 printk(KERN_DEBUG "alloc (nid=%d %llx - %llx) (%llx - %llx) %llx %llx => %llx\n",
3612 ei_start, ei_last, goal, limit, size,
3616 ptr = phys_to_virt(addr);
3617 memset(ptr, 0, size);
3618 reserve_early_without_check(addr, addr + size, "BOOTMEM");
3627 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3632 for_each_active_range_index_in_nid(i, nid) {
3633 ret = work_fn(early_node_map[i].start_pfn,
3634 early_node_map[i].end_pfn, data);
3640 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3641 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3643 * If an architecture guarantees that all ranges registered with
3644 * add_active_ranges() contain no holes and may be freed, this
3645 * function may be used instead of calling memory_present() manually.
3647 void __init sparse_memory_present_with_active_regions(int nid)
3651 for_each_active_range_index_in_nid(i, nid)
3652 memory_present(early_node_map[i].nid,
3653 early_node_map[i].start_pfn,
3654 early_node_map[i].end_pfn);
3658 * get_pfn_range_for_nid - Return the start and end page frames for a node
3659 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3660 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3661 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3663 * It returns the start and end page frame of a node based on information
3664 * provided by an arch calling add_active_range(). If called for a node
3665 * with no available memory, a warning is printed and the start and end
3668 void __meminit get_pfn_range_for_nid(unsigned int nid,
3669 unsigned long *start_pfn, unsigned long *end_pfn)
3675 for_each_active_range_index_in_nid(i, nid) {
3676 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3677 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3680 if (*start_pfn == -1UL)
3685 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3686 * assumption is made that zones within a node are ordered in monotonic
3687 * increasing memory addresses so that the "highest" populated zone is used
3689 static void __init find_usable_zone_for_movable(void)
3692 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3693 if (zone_index == ZONE_MOVABLE)
3696 if (arch_zone_highest_possible_pfn[zone_index] >
3697 arch_zone_lowest_possible_pfn[zone_index])
3701 VM_BUG_ON(zone_index == -1);
3702 movable_zone = zone_index;
3706 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3707 * because it is sized independant of architecture. Unlike the other zones,
3708 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3709 * in each node depending on the size of each node and how evenly kernelcore
3710 * is distributed. This helper function adjusts the zone ranges
3711 * provided by the architecture for a given node by using the end of the
3712 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3713 * zones within a node are in order of monotonic increases memory addresses
3715 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3716 unsigned long zone_type,
3717 unsigned long node_start_pfn,
3718 unsigned long node_end_pfn,
3719 unsigned long *zone_start_pfn,
3720 unsigned long *zone_end_pfn)
3722 /* Only adjust if ZONE_MOVABLE is on this node */
3723 if (zone_movable_pfn[nid]) {
3724 /* Size ZONE_MOVABLE */
3725 if (zone_type == ZONE_MOVABLE) {
3726 *zone_start_pfn = zone_movable_pfn[nid];
3727 *zone_end_pfn = min(node_end_pfn,
3728 arch_zone_highest_possible_pfn[movable_zone]);
3730 /* Adjust for ZONE_MOVABLE starting within this range */
3731 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3732 *zone_end_pfn > zone_movable_pfn[nid]) {
3733 *zone_end_pfn = zone_movable_pfn[nid];
3735 /* Check if this whole range is within ZONE_MOVABLE */
3736 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3737 *zone_start_pfn = *zone_end_pfn;
3742 * Return the number of pages a zone spans in a node, including holes
3743 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3745 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3746 unsigned long zone_type,
3747 unsigned long *ignored)
3749 unsigned long node_start_pfn, node_end_pfn;
3750 unsigned long zone_start_pfn, zone_end_pfn;
3752 /* Get the start and end of the node and zone */
3753 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3754 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3755 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3756 adjust_zone_range_for_zone_movable(nid, zone_type,
3757 node_start_pfn, node_end_pfn,
3758 &zone_start_pfn, &zone_end_pfn);
3760 /* Check that this node has pages within the zone's required range */
3761 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3764 /* Move the zone boundaries inside the node if necessary */
3765 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3766 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3768 /* Return the spanned pages */
3769 return zone_end_pfn - zone_start_pfn;
3773 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3774 * then all holes in the requested range will be accounted for.
3776 unsigned long __meminit __absent_pages_in_range(int nid,
3777 unsigned long range_start_pfn,
3778 unsigned long range_end_pfn)
3781 unsigned long prev_end_pfn = 0, hole_pages = 0;
3782 unsigned long start_pfn;
3784 /* Find the end_pfn of the first active range of pfns in the node */
3785 i = first_active_region_index_in_nid(nid);
3789 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3791 /* Account for ranges before physical memory on this node */
3792 if (early_node_map[i].start_pfn > range_start_pfn)
3793 hole_pages = prev_end_pfn - range_start_pfn;
3795 /* Find all holes for the zone within the node */
3796 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3798 /* No need to continue if prev_end_pfn is outside the zone */
3799 if (prev_end_pfn >= range_end_pfn)
3802 /* Make sure the end of the zone is not within the hole */
3803 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3804 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3806 /* Update the hole size cound and move on */
3807 if (start_pfn > range_start_pfn) {
3808 BUG_ON(prev_end_pfn > start_pfn);
3809 hole_pages += start_pfn - prev_end_pfn;
3811 prev_end_pfn = early_node_map[i].end_pfn;
3814 /* Account for ranges past physical memory on this node */
3815 if (range_end_pfn > prev_end_pfn)
3816 hole_pages += range_end_pfn -
3817 max(range_start_pfn, prev_end_pfn);
3823 * absent_pages_in_range - Return number of page frames in holes within a range
3824 * @start_pfn: The start PFN to start searching for holes
3825 * @end_pfn: The end PFN to stop searching for holes
3827 * It returns the number of pages frames in memory holes within a range.
3829 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3830 unsigned long end_pfn)
3832 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3835 /* Return the number of page frames in holes in a zone on a node */
3836 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3837 unsigned long zone_type,
3838 unsigned long *ignored)
3840 unsigned long node_start_pfn, node_end_pfn;
3841 unsigned long zone_start_pfn, zone_end_pfn;
3843 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3844 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3846 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3849 adjust_zone_range_for_zone_movable(nid, zone_type,
3850 node_start_pfn, node_end_pfn,
3851 &zone_start_pfn, &zone_end_pfn);
3852 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3856 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3857 unsigned long zone_type,
3858 unsigned long *zones_size)
3860 return zones_size[zone_type];
3863 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3864 unsigned long zone_type,
3865 unsigned long *zholes_size)
3870 return zholes_size[zone_type];
3875 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3876 unsigned long *zones_size, unsigned long *zholes_size)
3878 unsigned long realtotalpages, totalpages = 0;
3881 for (i = 0; i < MAX_NR_ZONES; i++)
3882 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3884 pgdat->node_spanned_pages = totalpages;
3886 realtotalpages = totalpages;
3887 for (i = 0; i < MAX_NR_ZONES; i++)
3889 zone_absent_pages_in_node(pgdat->node_id, i,
3891 pgdat->node_present_pages = realtotalpages;
3892 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3896 #ifndef CONFIG_SPARSEMEM
3898 * Calculate the size of the zone->blockflags rounded to an unsigned long
3899 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3900 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3901 * round what is now in bits to nearest long in bits, then return it in
3904 static unsigned long __init usemap_size(unsigned long zonesize)
3906 unsigned long usemapsize;
3908 usemapsize = roundup(zonesize, pageblock_nr_pages);
3909 usemapsize = usemapsize >> pageblock_order;
3910 usemapsize *= NR_PAGEBLOCK_BITS;
3911 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3913 return usemapsize / 8;
3916 static void __init setup_usemap(struct pglist_data *pgdat,
3917 struct zone *zone, unsigned long zonesize)
3919 unsigned long usemapsize = usemap_size(zonesize);
3920 zone->pageblock_flags = NULL;
3922 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3925 static void inline setup_usemap(struct pglist_data *pgdat,
3926 struct zone *zone, unsigned long zonesize) {}
3927 #endif /* CONFIG_SPARSEMEM */
3929 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3931 /* Return a sensible default order for the pageblock size. */
3932 static inline int pageblock_default_order(void)
3934 if (HPAGE_SHIFT > PAGE_SHIFT)
3935 return HUGETLB_PAGE_ORDER;
3940 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3941 static inline void __init set_pageblock_order(unsigned int order)
3943 /* Check that pageblock_nr_pages has not already been setup */
3944 if (pageblock_order)
3948 * Assume the largest contiguous order of interest is a huge page.
3949 * This value may be variable depending on boot parameters on IA64
3951 pageblock_order = order;
3953 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3956 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3957 * and pageblock_default_order() are unused as pageblock_order is set
3958 * at compile-time. See include/linux/pageblock-flags.h for the values of
3959 * pageblock_order based on the kernel config
3961 static inline int pageblock_default_order(unsigned int order)
3965 #define set_pageblock_order(x) do {} while (0)
3967 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3970 * Set up the zone data structures:
3971 * - mark all pages reserved
3972 * - mark all memory queues empty
3973 * - clear the memory bitmaps
3975 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3976 unsigned long *zones_size, unsigned long *zholes_size)
3979 int nid = pgdat->node_id;
3980 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3983 pgdat_resize_init(pgdat);
3984 pgdat->nr_zones = 0;
3985 init_waitqueue_head(&pgdat->kswapd_wait);
3986 pgdat->kswapd_max_order = 0;
3987 pgdat_page_cgroup_init(pgdat);
3989 for (j = 0; j < MAX_NR_ZONES; j++) {
3990 struct zone *zone = pgdat->node_zones + j;
3991 unsigned long size, realsize, memmap_pages;
3994 size = zone_spanned_pages_in_node(nid, j, zones_size);
3995 realsize = size - zone_absent_pages_in_node(nid, j,
3999 * Adjust realsize so that it accounts for how much memory
4000 * is used by this zone for memmap. This affects the watermark
4001 * and per-cpu initialisations
4004 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4005 if (realsize >= memmap_pages) {
4006 realsize -= memmap_pages;
4009 " %s zone: %lu pages used for memmap\n",
4010 zone_names[j], memmap_pages);
4013 " %s zone: %lu pages exceeds realsize %lu\n",
4014 zone_names[j], memmap_pages, realsize);
4016 /* Account for reserved pages */
4017 if (j == 0 && realsize > dma_reserve) {
4018 realsize -= dma_reserve;
4019 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4020 zone_names[0], dma_reserve);
4023 if (!is_highmem_idx(j))
4024 nr_kernel_pages += realsize;
4025 nr_all_pages += realsize;
4027 zone->spanned_pages = size;
4028 zone->present_pages = realsize;
4031 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4033 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4035 zone->name = zone_names[j];
4036 spin_lock_init(&zone->lock);
4037 spin_lock_init(&zone->lru_lock);
4038 zone_seqlock_init(zone);
4039 zone->zone_pgdat = pgdat;
4041 zone->prev_priority = DEF_PRIORITY;
4043 zone_pcp_init(zone);
4045 INIT_LIST_HEAD(&zone->lru[l].list);
4046 zone->reclaim_stat.nr_saved_scan[l] = 0;
4048 zone->reclaim_stat.recent_rotated[0] = 0;
4049 zone->reclaim_stat.recent_rotated[1] = 0;
4050 zone->reclaim_stat.recent_scanned[0] = 0;
4051 zone->reclaim_stat.recent_scanned[1] = 0;
4052 zap_zone_vm_stats(zone);
4057 set_pageblock_order(pageblock_default_order());
4058 setup_usemap(pgdat, zone, size);
4059 ret = init_currently_empty_zone(zone, zone_start_pfn,
4060 size, MEMMAP_EARLY);
4062 memmap_init(size, nid, j, zone_start_pfn);
4063 zone_start_pfn += size;
4067 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4069 /* Skip empty nodes */
4070 if (!pgdat->node_spanned_pages)
4073 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4074 /* ia64 gets its own node_mem_map, before this, without bootmem */
4075 if (!pgdat->node_mem_map) {
4076 unsigned long size, start, end;
4080 * The zone's endpoints aren't required to be MAX_ORDER
4081 * aligned but the node_mem_map endpoints must be in order
4082 * for the buddy allocator to function correctly.
4084 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4085 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4086 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4087 size = (end - start) * sizeof(struct page);
4088 map = alloc_remap(pgdat->node_id, size);
4090 map = alloc_bootmem_node(pgdat, size);
4091 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4093 #ifndef CONFIG_NEED_MULTIPLE_NODES
4095 * With no DISCONTIG, the global mem_map is just set as node 0's
4097 if (pgdat == NODE_DATA(0)) {
4098 mem_map = NODE_DATA(0)->node_mem_map;
4099 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4100 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4101 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4102 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4105 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4108 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4109 unsigned long node_start_pfn, unsigned long *zholes_size)
4111 pg_data_t *pgdat = NODE_DATA(nid);
4113 pgdat->node_id = nid;
4114 pgdat->node_start_pfn = node_start_pfn;
4115 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4117 alloc_node_mem_map(pgdat);
4118 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4119 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4120 nid, (unsigned long)pgdat,
4121 (unsigned long)pgdat->node_mem_map);
4124 free_area_init_core(pgdat, zones_size, zholes_size);
4127 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4129 #if MAX_NUMNODES > 1
4131 * Figure out the number of possible node ids.
4133 static void __init setup_nr_node_ids(void)
4136 unsigned int highest = 0;
4138 for_each_node_mask(node, node_possible_map)
4140 nr_node_ids = highest + 1;
4143 static inline void setup_nr_node_ids(void)
4149 * add_active_range - Register a range of PFNs backed by physical memory
4150 * @nid: The node ID the range resides on
4151 * @start_pfn: The start PFN of the available physical memory
4152 * @end_pfn: The end PFN of the available physical memory
4154 * These ranges are stored in an early_node_map[] and later used by
4155 * free_area_init_nodes() to calculate zone sizes and holes. If the
4156 * range spans a memory hole, it is up to the architecture to ensure
4157 * the memory is not freed by the bootmem allocator. If possible
4158 * the range being registered will be merged with existing ranges.
4160 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4161 unsigned long end_pfn)
4165 mminit_dprintk(MMINIT_TRACE, "memory_register",
4166 "Entering add_active_range(%d, %#lx, %#lx) "
4167 "%d entries of %d used\n",
4168 nid, start_pfn, end_pfn,
4169 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4171 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4173 /* Merge with existing active regions if possible */
4174 for (i = 0; i < nr_nodemap_entries; i++) {
4175 if (early_node_map[i].nid != nid)
4178 /* Skip if an existing region covers this new one */
4179 if (start_pfn >= early_node_map[i].start_pfn &&
4180 end_pfn <= early_node_map[i].end_pfn)
4183 /* Merge forward if suitable */
4184 if (start_pfn <= early_node_map[i].end_pfn &&
4185 end_pfn > early_node_map[i].end_pfn) {
4186 early_node_map[i].end_pfn = end_pfn;
4190 /* Merge backward if suitable */
4191 if (start_pfn < early_node_map[i].start_pfn &&
4192 end_pfn >= early_node_map[i].start_pfn) {
4193 early_node_map[i].start_pfn = start_pfn;
4198 /* Check that early_node_map is large enough */
4199 if (i >= MAX_ACTIVE_REGIONS) {
4200 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4201 MAX_ACTIVE_REGIONS);
4205 early_node_map[i].nid = nid;
4206 early_node_map[i].start_pfn = start_pfn;
4207 early_node_map[i].end_pfn = end_pfn;
4208 nr_nodemap_entries = i + 1;
4212 * remove_active_range - Shrink an existing registered range of PFNs
4213 * @nid: The node id the range is on that should be shrunk
4214 * @start_pfn: The new PFN of the range
4215 * @end_pfn: The new PFN of the range
4217 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4218 * The map is kept near the end physical page range that has already been
4219 * registered. This function allows an arch to shrink an existing registered
4222 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4223 unsigned long end_pfn)
4228 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4229 nid, start_pfn, end_pfn);
4231 /* Find the old active region end and shrink */
4232 for_each_active_range_index_in_nid(i, nid) {
4233 if (early_node_map[i].start_pfn >= start_pfn &&
4234 early_node_map[i].end_pfn <= end_pfn) {
4236 early_node_map[i].start_pfn = 0;
4237 early_node_map[i].end_pfn = 0;
4241 if (early_node_map[i].start_pfn < start_pfn &&
4242 early_node_map[i].end_pfn > start_pfn) {
4243 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4244 early_node_map[i].end_pfn = start_pfn;
4245 if (temp_end_pfn > end_pfn)
4246 add_active_range(nid, end_pfn, temp_end_pfn);
4249 if (early_node_map[i].start_pfn >= start_pfn &&
4250 early_node_map[i].end_pfn > end_pfn &&
4251 early_node_map[i].start_pfn < end_pfn) {
4252 early_node_map[i].start_pfn = end_pfn;
4260 /* remove the blank ones */
4261 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4262 if (early_node_map[i].nid != nid)
4264 if (early_node_map[i].end_pfn)
4266 /* we found it, get rid of it */
4267 for (j = i; j < nr_nodemap_entries - 1; j++)
4268 memcpy(&early_node_map[j], &early_node_map[j+1],
4269 sizeof(early_node_map[j]));
4270 j = nr_nodemap_entries - 1;
4271 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4272 nr_nodemap_entries--;
4277 * remove_all_active_ranges - Remove all currently registered regions
4279 * During discovery, it may be found that a table like SRAT is invalid
4280 * and an alternative discovery method must be used. This function removes
4281 * all currently registered regions.
4283 void __init remove_all_active_ranges(void)
4285 memset(early_node_map, 0, sizeof(early_node_map));
4286 nr_nodemap_entries = 0;
4289 /* Compare two active node_active_regions */
4290 static int __init cmp_node_active_region(const void *a, const void *b)
4292 struct node_active_region *arange = (struct node_active_region *)a;
4293 struct node_active_region *brange = (struct node_active_region *)b;
4295 /* Done this way to avoid overflows */
4296 if (arange->start_pfn > brange->start_pfn)
4298 if (arange->start_pfn < brange->start_pfn)
4304 /* sort the node_map by start_pfn */
4305 void __init sort_node_map(void)
4307 sort(early_node_map, (size_t)nr_nodemap_entries,
4308 sizeof(struct node_active_region),
4309 cmp_node_active_region, NULL);
4312 /* Find the lowest pfn for a node */
4313 static unsigned long __init find_min_pfn_for_node(int nid)
4316 unsigned long min_pfn = ULONG_MAX;
4318 /* Assuming a sorted map, the first range found has the starting pfn */
4319 for_each_active_range_index_in_nid(i, nid)
4320 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4322 if (min_pfn == ULONG_MAX) {
4324 "Could not find start_pfn for node %d\n", nid);
4332 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4334 * It returns the minimum PFN based on information provided via
4335 * add_active_range().
4337 unsigned long __init find_min_pfn_with_active_regions(void)
4339 return find_min_pfn_for_node(MAX_NUMNODES);
4343 * early_calculate_totalpages()
4344 * Sum pages in active regions for movable zone.
4345 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4347 static unsigned long __init early_calculate_totalpages(void)
4350 unsigned long totalpages = 0;
4352 for (i = 0; i < nr_nodemap_entries; i++) {
4353 unsigned long pages = early_node_map[i].end_pfn -
4354 early_node_map[i].start_pfn;
4355 totalpages += pages;
4357 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4363 * Find the PFN the Movable zone begins in each node. Kernel memory
4364 * is spread evenly between nodes as long as the nodes have enough
4365 * memory. When they don't, some nodes will have more kernelcore than
4368 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4371 unsigned long usable_startpfn;
4372 unsigned long kernelcore_node, kernelcore_remaining;
4373 /* save the state before borrow the nodemask */
4374 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4375 unsigned long totalpages = early_calculate_totalpages();
4376 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4379 * If movablecore was specified, calculate what size of
4380 * kernelcore that corresponds so that memory usable for
4381 * any allocation type is evenly spread. If both kernelcore
4382 * and movablecore are specified, then the value of kernelcore
4383 * will be used for required_kernelcore if it's greater than
4384 * what movablecore would have allowed.
4386 if (required_movablecore) {
4387 unsigned long corepages;
4390 * Round-up so that ZONE_MOVABLE is at least as large as what
4391 * was requested by the user
4393 required_movablecore =
4394 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4395 corepages = totalpages - required_movablecore;
4397 required_kernelcore = max(required_kernelcore, corepages);
4400 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4401 if (!required_kernelcore)
4404 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4405 find_usable_zone_for_movable();
4406 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4409 /* Spread kernelcore memory as evenly as possible throughout nodes */
4410 kernelcore_node = required_kernelcore / usable_nodes;
4411 for_each_node_state(nid, N_HIGH_MEMORY) {
4413 * Recalculate kernelcore_node if the division per node
4414 * now exceeds what is necessary to satisfy the requested
4415 * amount of memory for the kernel
4417 if (required_kernelcore < kernelcore_node)
4418 kernelcore_node = required_kernelcore / usable_nodes;
4421 * As the map is walked, we track how much memory is usable
4422 * by the kernel using kernelcore_remaining. When it is
4423 * 0, the rest of the node is usable by ZONE_MOVABLE
4425 kernelcore_remaining = kernelcore_node;
4427 /* Go through each range of PFNs within this node */
4428 for_each_active_range_index_in_nid(i, nid) {
4429 unsigned long start_pfn, end_pfn;
4430 unsigned long size_pages;
4432 start_pfn = max(early_node_map[i].start_pfn,
4433 zone_movable_pfn[nid]);
4434 end_pfn = early_node_map[i].end_pfn;
4435 if (start_pfn >= end_pfn)
4438 /* Account for what is only usable for kernelcore */
4439 if (start_pfn < usable_startpfn) {
4440 unsigned long kernel_pages;
4441 kernel_pages = min(end_pfn, usable_startpfn)
4444 kernelcore_remaining -= min(kernel_pages,
4445 kernelcore_remaining);
4446 required_kernelcore -= min(kernel_pages,
4447 required_kernelcore);
4449 /* Continue if range is now fully accounted */
4450 if (end_pfn <= usable_startpfn) {
4453 * Push zone_movable_pfn to the end so
4454 * that if we have to rebalance
4455 * kernelcore across nodes, we will
4456 * not double account here
4458 zone_movable_pfn[nid] = end_pfn;
4461 start_pfn = usable_startpfn;
4465 * The usable PFN range for ZONE_MOVABLE is from
4466 * start_pfn->end_pfn. Calculate size_pages as the
4467 * number of pages used as kernelcore
4469 size_pages = end_pfn - start_pfn;
4470 if (size_pages > kernelcore_remaining)
4471 size_pages = kernelcore_remaining;
4472 zone_movable_pfn[nid] = start_pfn + size_pages;
4475 * Some kernelcore has been met, update counts and
4476 * break if the kernelcore for this node has been
4479 required_kernelcore -= min(required_kernelcore,
4481 kernelcore_remaining -= size_pages;
4482 if (!kernelcore_remaining)
4488 * If there is still required_kernelcore, we do another pass with one
4489 * less node in the count. This will push zone_movable_pfn[nid] further
4490 * along on the nodes that still have memory until kernelcore is
4494 if (usable_nodes && required_kernelcore > usable_nodes)
4497 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4498 for (nid = 0; nid < MAX_NUMNODES; nid++)
4499 zone_movable_pfn[nid] =
4500 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4503 /* restore the node_state */
4504 node_states[N_HIGH_MEMORY] = saved_node_state;
4507 /* Any regular memory on that node ? */
4508 static void check_for_regular_memory(pg_data_t *pgdat)
4510 #ifdef CONFIG_HIGHMEM
4511 enum zone_type zone_type;
4513 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4514 struct zone *zone = &pgdat->node_zones[zone_type];
4515 if (zone->present_pages)
4516 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4522 * free_area_init_nodes - Initialise all pg_data_t and zone data
4523 * @max_zone_pfn: an array of max PFNs for each zone
4525 * This will call free_area_init_node() for each active node in the system.
4526 * Using the page ranges provided by add_active_range(), the size of each
4527 * zone in each node and their holes is calculated. If the maximum PFN
4528 * between two adjacent zones match, it is assumed that the zone is empty.
4529 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4530 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4531 * starts where the previous one ended. For example, ZONE_DMA32 starts
4532 * at arch_max_dma_pfn.
4534 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4539 /* Sort early_node_map as initialisation assumes it is sorted */
4542 /* Record where the zone boundaries are */
4543 memset(arch_zone_lowest_possible_pfn, 0,
4544 sizeof(arch_zone_lowest_possible_pfn));
4545 memset(arch_zone_highest_possible_pfn, 0,
4546 sizeof(arch_zone_highest_possible_pfn));
4547 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4548 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4549 for (i = 1; i < MAX_NR_ZONES; i++) {
4550 if (i == ZONE_MOVABLE)
4552 arch_zone_lowest_possible_pfn[i] =
4553 arch_zone_highest_possible_pfn[i-1];
4554 arch_zone_highest_possible_pfn[i] =
4555 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4557 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4558 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4560 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4561 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4562 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4564 /* Print out the zone ranges */
4565 printk("Zone PFN ranges:\n");
4566 for (i = 0; i < MAX_NR_ZONES; i++) {
4567 if (i == ZONE_MOVABLE)
4569 printk(" %-8s ", zone_names[i]);
4570 if (arch_zone_lowest_possible_pfn[i] ==
4571 arch_zone_highest_possible_pfn[i])
4574 printk("%0#10lx -> %0#10lx\n",
4575 arch_zone_lowest_possible_pfn[i],
4576 arch_zone_highest_possible_pfn[i]);
4579 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4580 printk("Movable zone start PFN for each node\n");
4581 for (i = 0; i < MAX_NUMNODES; i++) {
4582 if (zone_movable_pfn[i])
4583 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4586 /* Print out the early_node_map[] */
4587 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4588 for (i = 0; i < nr_nodemap_entries; i++)
4589 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4590 early_node_map[i].start_pfn,
4591 early_node_map[i].end_pfn);
4593 /* Initialise every node */
4594 mminit_verify_pageflags_layout();
4595 setup_nr_node_ids();
4596 for_each_online_node(nid) {
4597 pg_data_t *pgdat = NODE_DATA(nid);
4598 free_area_init_node(nid, NULL,
4599 find_min_pfn_for_node(nid), NULL);
4601 /* Any memory on that node */
4602 if (pgdat->node_present_pages)
4603 node_set_state(nid, N_HIGH_MEMORY);
4604 check_for_regular_memory(pgdat);
4608 static int __init cmdline_parse_core(char *p, unsigned long *core)
4610 unsigned long long coremem;
4614 coremem = memparse(p, &p);
4615 *core = coremem >> PAGE_SHIFT;
4617 /* Paranoid check that UL is enough for the coremem value */
4618 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4624 * kernelcore=size sets the amount of memory for use for allocations that
4625 * cannot be reclaimed or migrated.
4627 static int __init cmdline_parse_kernelcore(char *p)
4629 return cmdline_parse_core(p, &required_kernelcore);
4633 * movablecore=size sets the amount of memory for use for allocations that
4634 * can be reclaimed or migrated.
4636 static int __init cmdline_parse_movablecore(char *p)
4638 return cmdline_parse_core(p, &required_movablecore);
4641 early_param("kernelcore", cmdline_parse_kernelcore);
4642 early_param("movablecore", cmdline_parse_movablecore);
4644 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4647 * set_dma_reserve - set the specified number of pages reserved in the first zone
4648 * @new_dma_reserve: The number of pages to mark reserved
4650 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4651 * In the DMA zone, a significant percentage may be consumed by kernel image
4652 * and other unfreeable allocations which can skew the watermarks badly. This
4653 * function may optionally be used to account for unfreeable pages in the
4654 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4655 * smaller per-cpu batchsize.
4657 void __init set_dma_reserve(unsigned long new_dma_reserve)
4659 dma_reserve = new_dma_reserve;
4662 #ifndef CONFIG_NEED_MULTIPLE_NODES
4663 struct pglist_data __refdata contig_page_data = {
4664 #ifndef CONFIG_NO_BOOTMEM
4665 .bdata = &bootmem_node_data[0]
4668 EXPORT_SYMBOL(contig_page_data);
4671 void __init free_area_init(unsigned long *zones_size)
4673 free_area_init_node(0, zones_size,
4674 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4677 static int page_alloc_cpu_notify(struct notifier_block *self,
4678 unsigned long action, void *hcpu)
4680 int cpu = (unsigned long)hcpu;
4682 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4686 * Spill the event counters of the dead processor
4687 * into the current processors event counters.
4688 * This artificially elevates the count of the current
4691 vm_events_fold_cpu(cpu);
4694 * Zero the differential counters of the dead processor
4695 * so that the vm statistics are consistent.
4697 * This is only okay since the processor is dead and cannot
4698 * race with what we are doing.
4700 refresh_cpu_vm_stats(cpu);
4705 void __init page_alloc_init(void)
4707 hotcpu_notifier(page_alloc_cpu_notify, 0);
4711 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4712 * or min_free_kbytes changes.
4714 static void calculate_totalreserve_pages(void)
4716 struct pglist_data *pgdat;
4717 unsigned long reserve_pages = 0;
4718 enum zone_type i, j;
4720 for_each_online_pgdat(pgdat) {
4721 for (i = 0; i < MAX_NR_ZONES; i++) {
4722 struct zone *zone = pgdat->node_zones + i;
4723 unsigned long max = 0;
4725 /* Find valid and maximum lowmem_reserve in the zone */
4726 for (j = i; j < MAX_NR_ZONES; j++) {
4727 if (zone->lowmem_reserve[j] > max)
4728 max = zone->lowmem_reserve[j];
4731 /* we treat the high watermark as reserved pages. */
4732 max += high_wmark_pages(zone);
4734 if (max > zone->present_pages)
4735 max = zone->present_pages;
4736 reserve_pages += max;
4739 totalreserve_pages = reserve_pages;
4743 * setup_per_zone_lowmem_reserve - called whenever
4744 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4745 * has a correct pages reserved value, so an adequate number of
4746 * pages are left in the zone after a successful __alloc_pages().
4748 static void setup_per_zone_lowmem_reserve(void)
4750 struct pglist_data *pgdat;
4751 enum zone_type j, idx;
4753 for_each_online_pgdat(pgdat) {
4754 for (j = 0; j < MAX_NR_ZONES; j++) {
4755 struct zone *zone = pgdat->node_zones + j;
4756 unsigned long present_pages = zone->present_pages;
4758 zone->lowmem_reserve[j] = 0;
4762 struct zone *lower_zone;
4766 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4767 sysctl_lowmem_reserve_ratio[idx] = 1;
4769 lower_zone = pgdat->node_zones + idx;
4770 lower_zone->lowmem_reserve[j] = present_pages /
4771 sysctl_lowmem_reserve_ratio[idx];
4772 present_pages += lower_zone->present_pages;
4777 /* update totalreserve_pages */
4778 calculate_totalreserve_pages();
4782 * setup_per_zone_wmarks - called when min_free_kbytes changes
4783 * or when memory is hot-{added|removed}
4785 * Ensures that the watermark[min,low,high] values for each zone are set
4786 * correctly with respect to min_free_kbytes.
4788 void setup_per_zone_wmarks(void)
4790 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4791 unsigned long lowmem_pages = 0;
4793 unsigned long flags;
4795 /* Calculate total number of !ZONE_HIGHMEM pages */
4796 for_each_zone(zone) {
4797 if (!is_highmem(zone))
4798 lowmem_pages += zone->present_pages;
4801 for_each_zone(zone) {
4804 spin_lock_irqsave(&zone->lock, flags);
4805 tmp = (u64)pages_min * zone->present_pages;
4806 do_div(tmp, lowmem_pages);
4807 if (is_highmem(zone)) {
4809 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4810 * need highmem pages, so cap pages_min to a small
4813 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4814 * deltas controls asynch page reclaim, and so should
4815 * not be capped for highmem.
4819 min_pages = zone->present_pages / 1024;
4820 if (min_pages < SWAP_CLUSTER_MAX)
4821 min_pages = SWAP_CLUSTER_MAX;
4822 if (min_pages > 128)
4824 zone->watermark[WMARK_MIN] = min_pages;
4827 * If it's a lowmem zone, reserve a number of pages
4828 * proportionate to the zone's size.
4830 zone->watermark[WMARK_MIN] = tmp;
4833 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4834 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4835 setup_zone_migrate_reserve(zone);
4836 spin_unlock_irqrestore(&zone->lock, flags);
4839 /* update totalreserve_pages */
4840 calculate_totalreserve_pages();
4844 * The inactive anon list should be small enough that the VM never has to
4845 * do too much work, but large enough that each inactive page has a chance
4846 * to be referenced again before it is swapped out.
4848 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4849 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4850 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4851 * the anonymous pages are kept on the inactive list.
4854 * memory ratio inactive anon
4855 * -------------------------------------
4864 void calculate_zone_inactive_ratio(struct zone *zone)
4866 unsigned int gb, ratio;
4868 /* Zone size in gigabytes */
4869 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4871 ratio = int_sqrt(10 * gb);
4875 zone->inactive_ratio = ratio;
4878 static void __init setup_per_zone_inactive_ratio(void)
4883 calculate_zone_inactive_ratio(zone);
4887 * Initialise min_free_kbytes.
4889 * For small machines we want it small (128k min). For large machines
4890 * we want it large (64MB max). But it is not linear, because network
4891 * bandwidth does not increase linearly with machine size. We use
4893 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4894 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4910 static int __init init_per_zone_wmark_min(void)
4912 unsigned long lowmem_kbytes;
4914 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4916 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4917 if (min_free_kbytes < 128)
4918 min_free_kbytes = 128;
4919 if (min_free_kbytes > 65536)
4920 min_free_kbytes = 65536;
4921 setup_per_zone_wmarks();
4922 setup_per_zone_lowmem_reserve();
4923 setup_per_zone_inactive_ratio();
4926 module_init(init_per_zone_wmark_min)
4929 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4930 * that we can call two helper functions whenever min_free_kbytes
4933 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4934 void __user *buffer, size_t *length, loff_t *ppos)
4936 proc_dointvec(table, write, buffer, length, ppos);
4938 setup_per_zone_wmarks();
4943 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4944 void __user *buffer, size_t *length, loff_t *ppos)
4949 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4954 zone->min_unmapped_pages = (zone->present_pages *
4955 sysctl_min_unmapped_ratio) / 100;
4959 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4960 void __user *buffer, size_t *length, loff_t *ppos)
4965 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4970 zone->min_slab_pages = (zone->present_pages *
4971 sysctl_min_slab_ratio) / 100;
4977 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4978 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4979 * whenever sysctl_lowmem_reserve_ratio changes.
4981 * The reserve ratio obviously has absolutely no relation with the
4982 * minimum watermarks. The lowmem reserve ratio can only make sense
4983 * if in function of the boot time zone sizes.
4985 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4986 void __user *buffer, size_t *length, loff_t *ppos)
4988 proc_dointvec_minmax(table, write, buffer, length, ppos);
4989 setup_per_zone_lowmem_reserve();
4994 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4995 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4996 * can have before it gets flushed back to buddy allocator.
4999 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5000 void __user *buffer, size_t *length, loff_t *ppos)
5006 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5007 if (!write || (ret == -EINVAL))
5009 for_each_populated_zone(zone) {
5010 for_each_possible_cpu(cpu) {
5012 high = zone->present_pages / percpu_pagelist_fraction;
5013 setup_pagelist_highmark(
5014 per_cpu_ptr(zone->pageset, cpu), high);
5020 int hashdist = HASHDIST_DEFAULT;
5023 static int __init set_hashdist(char *str)
5027 hashdist = simple_strtoul(str, &str, 0);
5030 __setup("hashdist=", set_hashdist);
5034 * allocate a large system hash table from bootmem
5035 * - it is assumed that the hash table must contain an exact power-of-2
5036 * quantity of entries
5037 * - limit is the number of hash buckets, not the total allocation size
5039 void *__init alloc_large_system_hash(const char *tablename,
5040 unsigned long bucketsize,
5041 unsigned long numentries,
5044 unsigned int *_hash_shift,
5045 unsigned int *_hash_mask,
5046 unsigned long limit)
5048 unsigned long long max = limit;
5049 unsigned long log2qty, size;
5052 /* allow the kernel cmdline to have a say */
5054 /* round applicable memory size up to nearest megabyte */
5055 numentries = nr_kernel_pages;
5056 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5057 numentries >>= 20 - PAGE_SHIFT;
5058 numentries <<= 20 - PAGE_SHIFT;
5060 /* limit to 1 bucket per 2^scale bytes of low memory */
5061 if (scale > PAGE_SHIFT)
5062 numentries >>= (scale - PAGE_SHIFT);
5064 numentries <<= (PAGE_SHIFT - scale);
5066 /* Make sure we've got at least a 0-order allocation.. */
5067 if (unlikely(flags & HASH_SMALL)) {
5068 /* Makes no sense without HASH_EARLY */
5069 WARN_ON(!(flags & HASH_EARLY));
5070 if (!(numentries >> *_hash_shift)) {
5071 numentries = 1UL << *_hash_shift;
5072 BUG_ON(!numentries);
5074 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5075 numentries = PAGE_SIZE / bucketsize;
5077 numentries = roundup_pow_of_two(numentries);
5079 /* limit allocation size to 1/16 total memory by default */
5081 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5082 do_div(max, bucketsize);
5085 if (numentries > max)
5088 log2qty = ilog2(numentries);
5091 size = bucketsize << log2qty;
5092 if (flags & HASH_EARLY)
5093 table = alloc_bootmem_nopanic(size);
5095 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5098 * If bucketsize is not a power-of-two, we may free
5099 * some pages at the end of hash table which
5100 * alloc_pages_exact() automatically does
5102 if (get_order(size) < MAX_ORDER) {
5103 table = alloc_pages_exact(size, GFP_ATOMIC);
5104 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5107 } while (!table && size > PAGE_SIZE && --log2qty);
5110 panic("Failed to allocate %s hash table\n", tablename);
5112 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
5115 ilog2(size) - PAGE_SHIFT,
5119 *_hash_shift = log2qty;
5121 *_hash_mask = (1 << log2qty) - 1;
5126 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5127 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5130 #ifdef CONFIG_SPARSEMEM
5131 return __pfn_to_section(pfn)->pageblock_flags;
5133 return zone->pageblock_flags;
5134 #endif /* CONFIG_SPARSEMEM */
5137 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5139 #ifdef CONFIG_SPARSEMEM
5140 pfn &= (PAGES_PER_SECTION-1);
5141 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5143 pfn = pfn - zone->zone_start_pfn;
5144 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5145 #endif /* CONFIG_SPARSEMEM */
5149 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5150 * @page: The page within the block of interest
5151 * @start_bitidx: The first bit of interest to retrieve
5152 * @end_bitidx: The last bit of interest
5153 * returns pageblock_bits flags
5155 unsigned long get_pageblock_flags_group(struct page *page,
5156 int start_bitidx, int end_bitidx)
5159 unsigned long *bitmap;
5160 unsigned long pfn, bitidx;
5161 unsigned long flags = 0;
5162 unsigned long value = 1;
5164 zone = page_zone(page);
5165 pfn = page_to_pfn(page);
5166 bitmap = get_pageblock_bitmap(zone, pfn);
5167 bitidx = pfn_to_bitidx(zone, pfn);
5169 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5170 if (test_bit(bitidx + start_bitidx, bitmap))
5177 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5178 * @page: The page within the block of interest
5179 * @start_bitidx: The first bit of interest
5180 * @end_bitidx: The last bit of interest
5181 * @flags: The flags to set
5183 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5184 int start_bitidx, int end_bitidx)
5187 unsigned long *bitmap;
5188 unsigned long pfn, bitidx;
5189 unsigned long value = 1;
5191 zone = page_zone(page);
5192 pfn = page_to_pfn(page);
5193 bitmap = get_pageblock_bitmap(zone, pfn);
5194 bitidx = pfn_to_bitidx(zone, pfn);
5195 VM_BUG_ON(pfn < zone->zone_start_pfn);
5196 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5198 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5200 __set_bit(bitidx + start_bitidx, bitmap);
5202 __clear_bit(bitidx + start_bitidx, bitmap);
5206 * This is designed as sub function...plz see page_isolation.c also.
5207 * set/clear page block's type to be ISOLATE.
5208 * page allocater never alloc memory from ISOLATE block.
5211 int set_migratetype_isolate(struct page *page)
5214 struct page *curr_page;
5215 unsigned long flags, pfn, iter;
5216 unsigned long immobile = 0;
5217 struct memory_isolate_notify arg;
5222 zone = page_zone(page);
5223 zone_idx = zone_idx(zone);
5225 spin_lock_irqsave(&zone->lock, flags);
5226 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE ||
5227 zone_idx == ZONE_MOVABLE) {
5232 pfn = page_to_pfn(page);
5233 arg.start_pfn = pfn;
5234 arg.nr_pages = pageblock_nr_pages;
5235 arg.pages_found = 0;
5238 * It may be possible to isolate a pageblock even if the
5239 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5240 * notifier chain is used by balloon drivers to return the
5241 * number of pages in a range that are held by the balloon
5242 * driver to shrink memory. If all the pages are accounted for
5243 * by balloons, are free, or on the LRU, isolation can continue.
5244 * Later, for example, when memory hotplug notifier runs, these
5245 * pages reported as "can be isolated" should be isolated(freed)
5246 * by the balloon driver through the memory notifier chain.
5248 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5249 notifier_ret = notifier_to_errno(notifier_ret);
5250 if (notifier_ret || !arg.pages_found)
5253 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) {
5254 if (!pfn_valid_within(pfn))
5257 curr_page = pfn_to_page(iter);
5258 if (!page_count(curr_page) || PageLRU(curr_page))
5264 if (arg.pages_found == immobile)
5269 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5270 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5273 spin_unlock_irqrestore(&zone->lock, flags);
5279 void unset_migratetype_isolate(struct page *page)
5282 unsigned long flags;
5283 zone = page_zone(page);
5284 spin_lock_irqsave(&zone->lock, flags);
5285 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5287 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5288 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5290 spin_unlock_irqrestore(&zone->lock, flags);
5293 #ifdef CONFIG_MEMORY_HOTREMOVE
5295 * All pages in the range must be isolated before calling this.
5298 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5304 unsigned long flags;
5305 /* find the first valid pfn */
5306 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5311 zone = page_zone(pfn_to_page(pfn));
5312 spin_lock_irqsave(&zone->lock, flags);
5314 while (pfn < end_pfn) {
5315 if (!pfn_valid(pfn)) {
5319 page = pfn_to_page(pfn);
5320 BUG_ON(page_count(page));
5321 BUG_ON(!PageBuddy(page));
5322 order = page_order(page);
5323 #ifdef CONFIG_DEBUG_VM
5324 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5325 pfn, 1 << order, end_pfn);
5327 list_del(&page->lru);
5328 rmv_page_order(page);
5329 zone->free_area[order].nr_free--;
5330 __mod_zone_page_state(zone, NR_FREE_PAGES,
5332 for (i = 0; i < (1 << order); i++)
5333 SetPageReserved((page+i));
5334 pfn += (1 << order);
5336 spin_unlock_irqrestore(&zone->lock, flags);
5340 #ifdef CONFIG_MEMORY_FAILURE
5341 bool is_free_buddy_page(struct page *page)
5343 struct zone *zone = page_zone(page);
5344 unsigned long pfn = page_to_pfn(page);
5345 unsigned long flags;
5348 spin_lock_irqsave(&zone->lock, flags);
5349 for (order = 0; order < MAX_ORDER; order++) {
5350 struct page *page_head = page - (pfn & ((1 << order) - 1));
5352 if (PageBuddy(page_head) && page_order(page_head) >= order)
5355 spin_unlock_irqrestore(&zone->lock, flags);
5357 return order < MAX_ORDER;
5361 static struct trace_print_flags pageflag_names[] = {
5362 {1UL << PG_locked, "locked" },
5363 {1UL << PG_error, "error" },
5364 {1UL << PG_referenced, "referenced" },
5365 {1UL << PG_uptodate, "uptodate" },
5366 {1UL << PG_dirty, "dirty" },
5367 {1UL << PG_lru, "lru" },
5368 {1UL << PG_active, "active" },
5369 {1UL << PG_slab, "slab" },
5370 {1UL << PG_owner_priv_1, "owner_priv_1" },
5371 {1UL << PG_arch_1, "arch_1" },
5372 {1UL << PG_reserved, "reserved" },
5373 {1UL << PG_private, "private" },
5374 {1UL << PG_private_2, "private_2" },
5375 {1UL << PG_writeback, "writeback" },
5376 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5377 {1UL << PG_head, "head" },
5378 {1UL << PG_tail, "tail" },
5380 {1UL << PG_compound, "compound" },
5382 {1UL << PG_swapcache, "swapcache" },
5383 {1UL << PG_mappedtodisk, "mappedtodisk" },
5384 {1UL << PG_reclaim, "reclaim" },
5385 {1UL << PG_buddy, "buddy" },
5386 {1UL << PG_swapbacked, "swapbacked" },
5387 {1UL << PG_unevictable, "unevictable" },
5389 {1UL << PG_mlocked, "mlocked" },
5391 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5392 {1UL << PG_uncached, "uncached" },
5394 #ifdef CONFIG_MEMORY_FAILURE
5395 {1UL << PG_hwpoison, "hwpoison" },
5400 static void dump_page_flags(unsigned long flags)
5402 const char *delim = "";
5406 printk(KERN_ALERT "page flags: %#lx(", flags);
5408 /* remove zone id */
5409 flags &= (1UL << NR_PAGEFLAGS) - 1;
5411 for (i = 0; pageflag_names[i].name && flags; i++) {
5413 mask = pageflag_names[i].mask;
5414 if ((flags & mask) != mask)
5418 printk("%s%s", delim, pageflag_names[i].name);
5422 /* check for left over flags */
5424 printk("%s%#lx", delim, flags);
5429 void dump_page(struct page *page)
5432 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5433 page, page_count(page), page_mapcount(page),
5434 page->mapping, page->index);
5435 dump_page_flags(page->flags);