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);
65 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
67 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
68 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
69 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
70 * defined in <linux/topology.h>.
72 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
73 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
77 * Array of node states.
79 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
80 [N_POSSIBLE] = NODE_MASK_ALL,
81 [N_ONLINE] = { { [0] = 1UL } },
83 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
85 [N_HIGH_MEMORY] = { { [0] = 1UL } },
87 [N_CPU] = { { [0] = 1UL } },
90 EXPORT_SYMBOL(node_states);
92 unsigned long totalram_pages __read_mostly;
93 unsigned long totalreserve_pages __read_mostly;
94 int percpu_pagelist_fraction;
95 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
97 #ifdef CONFIG_PM_SLEEP
99 * The following functions are used by the suspend/hibernate code to temporarily
100 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
101 * while devices are suspended. To avoid races with the suspend/hibernate code,
102 * they should always be called with pm_mutex held (gfp_allowed_mask also should
103 * only be modified with pm_mutex held, unless the suspend/hibernate code is
104 * guaranteed not to run in parallel with that modification).
107 static gfp_t saved_gfp_mask;
109 void pm_restore_gfp_mask(void)
111 WARN_ON(!mutex_is_locked(&pm_mutex));
112 if (saved_gfp_mask) {
113 gfp_allowed_mask = saved_gfp_mask;
118 void pm_restrict_gfp_mask(void)
120 WARN_ON(!mutex_is_locked(&pm_mutex));
121 WARN_ON(saved_gfp_mask);
122 saved_gfp_mask = gfp_allowed_mask;
123 gfp_allowed_mask &= ~GFP_IOFS;
125 #endif /* CONFIG_PM_SLEEP */
127 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
128 int pageblock_order __read_mostly;
131 static void __free_pages_ok(struct page *page, unsigned int order);
134 * results with 256, 32 in the lowmem_reserve sysctl:
135 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
136 * 1G machine -> (16M dma, 784M normal, 224M high)
137 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
138 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
139 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
141 * TBD: should special case ZONE_DMA32 machines here - in those we normally
142 * don't need any ZONE_NORMAL reservation
144 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
145 #ifdef CONFIG_ZONE_DMA
148 #ifdef CONFIG_ZONE_DMA32
151 #ifdef CONFIG_HIGHMEM
157 EXPORT_SYMBOL(totalram_pages);
159 static char * const zone_names[MAX_NR_ZONES] = {
160 #ifdef CONFIG_ZONE_DMA
163 #ifdef CONFIG_ZONE_DMA32
167 #ifdef CONFIG_HIGHMEM
173 int min_free_kbytes = 1024;
174 int min_free_order_shift = 1;
176 static unsigned long __meminitdata nr_kernel_pages;
177 static unsigned long __meminitdata nr_all_pages;
178 static unsigned long __meminitdata dma_reserve;
180 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
182 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
183 * ranges of memory (RAM) that may be registered with add_active_range().
184 * Ranges passed to add_active_range() will be merged if possible
185 * so the number of times add_active_range() can be called is
186 * related to the number of nodes and the number of holes
188 #ifdef CONFIG_MAX_ACTIVE_REGIONS
189 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
190 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
192 #if MAX_NUMNODES >= 32
193 /* If there can be many nodes, allow up to 50 holes per node */
194 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
196 /* By default, allow up to 256 distinct regions */
197 #define MAX_ACTIVE_REGIONS 256
201 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
202 static int __meminitdata nr_nodemap_entries;
203 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
204 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
205 static unsigned long __initdata required_kernelcore;
206 static unsigned long __initdata required_movablecore;
207 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
209 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
211 EXPORT_SYMBOL(movable_zone);
212 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
215 int nr_node_ids __read_mostly = MAX_NUMNODES;
216 int nr_online_nodes __read_mostly = 1;
217 EXPORT_SYMBOL(nr_node_ids);
218 EXPORT_SYMBOL(nr_online_nodes);
221 int page_group_by_mobility_disabled __read_mostly;
223 static void set_pageblock_migratetype(struct page *page, int migratetype)
226 if (unlikely(page_group_by_mobility_disabled))
227 migratetype = MIGRATE_UNMOVABLE;
229 set_pageblock_flags_group(page, (unsigned long)migratetype,
230 PB_migrate, PB_migrate_end);
233 bool oom_killer_disabled __read_mostly;
235 #ifdef CONFIG_DEBUG_VM
236 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
240 unsigned long pfn = page_to_pfn(page);
243 seq = zone_span_seqbegin(zone);
244 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
246 else if (pfn < zone->zone_start_pfn)
248 } while (zone_span_seqretry(zone, seq));
253 static int page_is_consistent(struct zone *zone, struct page *page)
255 if (!pfn_valid_within(page_to_pfn(page)))
257 if (zone != page_zone(page))
263 * Temporary debugging check for pages not lying within a given zone.
265 static int bad_range(struct zone *zone, struct page *page)
267 if (page_outside_zone_boundaries(zone, page))
269 if (!page_is_consistent(zone, page))
275 static inline int bad_range(struct zone *zone, struct page *page)
281 static void bad_page(struct page *page)
283 static unsigned long resume;
284 static unsigned long nr_shown;
285 static unsigned long nr_unshown;
287 /* Don't complain about poisoned pages */
288 if (PageHWPoison(page)) {
289 __ClearPageBuddy(page);
294 * Allow a burst of 60 reports, then keep quiet for that minute;
295 * or allow a steady drip of one report per second.
297 if (nr_shown == 60) {
298 if (time_before(jiffies, resume)) {
304 "BUG: Bad page state: %lu messages suppressed\n",
311 resume = jiffies + 60 * HZ;
313 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
314 current->comm, page_to_pfn(page));
319 /* Leave bad fields for debug, except PageBuddy could make trouble */
320 __ClearPageBuddy(page);
321 add_taint(TAINT_BAD_PAGE);
325 * Higher-order pages are called "compound pages". They are structured thusly:
327 * The first PAGE_SIZE page is called the "head page".
329 * The remaining PAGE_SIZE pages are called "tail pages".
331 * All pages have PG_compound set. All pages have their ->private pointing at
332 * the head page (even the head page has this).
334 * The first tail page's ->lru.next holds the address of the compound page's
335 * put_page() function. Its ->lru.prev holds the order of allocation.
336 * This usage means that zero-order pages may not be compound.
339 static void free_compound_page(struct page *page)
341 __free_pages_ok(page, compound_order(page));
344 void prep_compound_page(struct page *page, unsigned long order)
347 int nr_pages = 1 << order;
349 set_compound_page_dtor(page, free_compound_page);
350 set_compound_order(page, order);
352 for (i = 1; i < nr_pages; i++) {
353 struct page *p = page + i;
356 p->first_page = page;
360 static int destroy_compound_page(struct page *page, unsigned long order)
363 int nr_pages = 1 << order;
366 if (unlikely(compound_order(page) != order) ||
367 unlikely(!PageHead(page))) {
372 __ClearPageHead(page);
374 for (i = 1; i < nr_pages; i++) {
375 struct page *p = page + i;
377 if (unlikely(!PageTail(p) || (p->first_page != page))) {
387 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
392 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
393 * and __GFP_HIGHMEM from hard or soft interrupt context.
395 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
396 for (i = 0; i < (1 << order); i++)
397 clear_highpage(page + i);
400 static inline void set_page_order(struct page *page, int order)
402 set_page_private(page, order);
403 __SetPageBuddy(page);
406 static inline void rmv_page_order(struct page *page)
408 __ClearPageBuddy(page);
409 set_page_private(page, 0);
413 * Locate the struct page for both the matching buddy in our
414 * pair (buddy1) and the combined O(n+1) page they form (page).
416 * 1) Any buddy B1 will have an order O twin B2 which satisfies
417 * the following equation:
419 * For example, if the starting buddy (buddy2) is #8 its order
421 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
423 * 2) Any buddy B will have an order O+1 parent P which
424 * satisfies the following equation:
427 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
429 static inline struct page *
430 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
432 unsigned long buddy_idx = page_idx ^ (1 << order);
434 return page + (buddy_idx - page_idx);
437 static inline unsigned long
438 __find_combined_index(unsigned long page_idx, unsigned int order)
440 return (page_idx & ~(1 << order));
444 * This function checks whether a page is free && is the buddy
445 * we can do coalesce a page and its buddy if
446 * (a) the buddy is not in a hole &&
447 * (b) the buddy is in the buddy system &&
448 * (c) a page and its buddy have the same order &&
449 * (d) a page and its buddy are in the same zone.
451 * For recording whether a page is in the buddy system, we use PG_buddy.
452 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
454 * For recording page's order, we use page_private(page).
456 static inline int page_is_buddy(struct page *page, struct page *buddy,
459 if (!pfn_valid_within(page_to_pfn(buddy)))
462 if (page_zone_id(page) != page_zone_id(buddy))
465 if (PageBuddy(buddy) && page_order(buddy) == order) {
466 VM_BUG_ON(page_count(buddy) != 0);
473 * Freeing function for a buddy system allocator.
475 * The concept of a buddy system is to maintain direct-mapped table
476 * (containing bit values) for memory blocks of various "orders".
477 * The bottom level table contains the map for the smallest allocatable
478 * units of memory (here, pages), and each level above it describes
479 * pairs of units from the levels below, hence, "buddies".
480 * At a high level, all that happens here is marking the table entry
481 * at the bottom level available, and propagating the changes upward
482 * as necessary, plus some accounting needed to play nicely with other
483 * parts of the VM system.
484 * At each level, we keep a list of pages, which are heads of continuous
485 * free pages of length of (1 << order) and marked with PG_buddy. Page's
486 * order is recorded in page_private(page) field.
487 * So when we are allocating or freeing one, we can derive the state of the
488 * other. That is, if we allocate a small block, and both were
489 * free, the remainder of the region must be split into blocks.
490 * If a block is freed, and its buddy is also free, then this
491 * triggers coalescing into a block of larger size.
496 static inline void __free_one_page(struct page *page,
497 struct zone *zone, unsigned int order,
500 unsigned long page_idx;
501 unsigned long combined_idx;
504 if (unlikely(PageCompound(page)))
505 if (unlikely(destroy_compound_page(page, order)))
508 VM_BUG_ON(migratetype == -1);
510 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
512 VM_BUG_ON(page_idx & ((1 << order) - 1));
513 VM_BUG_ON(bad_range(zone, page));
515 while (order < MAX_ORDER-1) {
516 buddy = __page_find_buddy(page, page_idx, order);
517 if (!page_is_buddy(page, buddy, order))
520 /* Our buddy is free, merge with it and move up one order. */
521 list_del(&buddy->lru);
522 zone->free_area[order].nr_free--;
523 rmv_page_order(buddy);
524 combined_idx = __find_combined_index(page_idx, order);
525 page = page + (combined_idx - page_idx);
526 page_idx = combined_idx;
529 set_page_order(page, order);
532 * If this is not the largest possible page, check if the buddy
533 * of the next-highest order is free. If it is, it's possible
534 * that pages are being freed that will coalesce soon. In case,
535 * that is happening, add the free page to the tail of the list
536 * so it's less likely to be used soon and more likely to be merged
537 * as a higher order page
539 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
540 struct page *higher_page, *higher_buddy;
541 combined_idx = __find_combined_index(page_idx, order);
542 higher_page = page + combined_idx - page_idx;
543 higher_buddy = __page_find_buddy(higher_page, combined_idx, order + 1);
544 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
545 list_add_tail(&page->lru,
546 &zone->free_area[order].free_list[migratetype]);
551 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
553 zone->free_area[order].nr_free++;
557 * free_page_mlock() -- clean up attempts to free and mlocked() page.
558 * Page should not be on lru, so no need to fix that up.
559 * free_pages_check() will verify...
561 static inline void free_page_mlock(struct page *page)
563 __dec_zone_page_state(page, NR_MLOCK);
564 __count_vm_event(UNEVICTABLE_MLOCKFREED);
567 static inline int free_pages_check(struct page *page)
569 if (unlikely(page_mapcount(page) |
570 (page->mapping != NULL) |
571 (atomic_read(&page->_count) != 0) |
572 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
576 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
577 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
582 * Frees a number of pages from the PCP lists
583 * Assumes all pages on list are in same zone, and of same order.
584 * count is the number of pages to free.
586 * If the zone was previously in an "all pages pinned" state then look to
587 * see if this freeing clears that state.
589 * And clear the zone's pages_scanned counter, to hold off the "all pages are
590 * pinned" detection logic.
592 static void free_pcppages_bulk(struct zone *zone, int count,
593 struct per_cpu_pages *pcp)
599 spin_lock(&zone->lock);
600 zone->all_unreclaimable = 0;
601 zone->pages_scanned = 0;
605 struct list_head *list;
608 * Remove pages from lists in a round-robin fashion. A
609 * batch_free count is maintained that is incremented when an
610 * empty list is encountered. This is so more pages are freed
611 * off fuller lists instead of spinning excessively around empty
616 if (++migratetype == MIGRATE_PCPTYPES)
618 list = &pcp->lists[migratetype];
619 } while (list_empty(list));
622 page = list_entry(list->prev, struct page, lru);
623 /* must delete as __free_one_page list manipulates */
624 list_del(&page->lru);
625 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
626 __free_one_page(page, zone, 0, page_private(page));
627 trace_mm_page_pcpu_drain(page, 0, page_private(page));
628 } while (--to_free && --batch_free && !list_empty(list));
630 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
631 spin_unlock(&zone->lock);
634 static void free_one_page(struct zone *zone, struct page *page, int order,
637 spin_lock(&zone->lock);
638 zone->all_unreclaimable = 0;
639 zone->pages_scanned = 0;
641 __free_one_page(page, zone, order, migratetype);
642 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
643 spin_unlock(&zone->lock);
646 static bool free_pages_prepare(struct page *page, unsigned int order)
651 trace_mm_page_free_direct(page, order);
652 kmemcheck_free_shadow(page, order);
654 for (i = 0; i < (1 << order); i++) {
655 struct page *pg = page + i;
659 bad += free_pages_check(pg);
664 if (!PageHighMem(page)) {
665 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
666 debug_check_no_obj_freed(page_address(page),
669 arch_free_page(page, order);
670 kernel_map_pages(page, 1 << order, 0);
675 static void __free_pages_ok(struct page *page, unsigned int order)
678 int wasMlocked = __TestClearPageMlocked(page);
680 if (!free_pages_prepare(page, order))
683 local_irq_save(flags);
684 if (unlikely(wasMlocked))
685 free_page_mlock(page);
686 __count_vm_events(PGFREE, 1 << order);
687 free_one_page(page_zone(page), page, order,
688 get_pageblock_migratetype(page));
689 local_irq_restore(flags);
693 * permit the bootmem allocator to evade page validation on high-order frees
695 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
698 __ClearPageReserved(page);
699 set_page_count(page, 0);
700 set_page_refcounted(page);
706 for (loop = 0; loop < BITS_PER_LONG; loop++) {
707 struct page *p = &page[loop];
709 if (loop + 1 < BITS_PER_LONG)
711 __ClearPageReserved(p);
712 set_page_count(p, 0);
715 set_page_refcounted(page);
716 __free_pages(page, order);
722 * The order of subdivision here is critical for the IO subsystem.
723 * Please do not alter this order without good reasons and regression
724 * testing. Specifically, as large blocks of memory are subdivided,
725 * the order in which smaller blocks are delivered depends on the order
726 * they're subdivided in this function. This is the primary factor
727 * influencing the order in which pages are delivered to the IO
728 * subsystem according to empirical testing, and this is also justified
729 * by considering the behavior of a buddy system containing a single
730 * large block of memory acted on by a series of small allocations.
731 * This behavior is a critical factor in sglist merging's success.
735 static inline void expand(struct zone *zone, struct page *page,
736 int low, int high, struct free_area *area,
739 unsigned long size = 1 << high;
745 VM_BUG_ON(bad_range(zone, &page[size]));
746 list_add(&page[size].lru, &area->free_list[migratetype]);
748 set_page_order(&page[size], high);
753 * This page is about to be returned from the page allocator
755 static inline int check_new_page(struct page *page)
757 if (unlikely(page_mapcount(page) |
758 (page->mapping != NULL) |
759 (atomic_read(&page->_count) != 0) |
760 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
767 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
771 for (i = 0; i < (1 << order); i++) {
772 struct page *p = page + i;
773 if (unlikely(check_new_page(p)))
777 set_page_private(page, 0);
778 set_page_refcounted(page);
780 arch_alloc_page(page, order);
781 kernel_map_pages(page, 1 << order, 1);
783 if (gfp_flags & __GFP_ZERO)
784 prep_zero_page(page, order, gfp_flags);
786 if (order && (gfp_flags & __GFP_COMP))
787 prep_compound_page(page, order);
793 * Go through the free lists for the given migratetype and remove
794 * the smallest available page from the freelists
797 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
800 unsigned int current_order;
801 struct free_area * area;
804 /* Find a page of the appropriate size in the preferred list */
805 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
806 area = &(zone->free_area[current_order]);
807 if (list_empty(&area->free_list[migratetype]))
810 page = list_entry(area->free_list[migratetype].next,
812 list_del(&page->lru);
813 rmv_page_order(page);
815 expand(zone, page, order, current_order, area, migratetype);
824 * This array describes the order lists are fallen back to when
825 * the free lists for the desirable migrate type are depleted
827 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
828 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
829 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
830 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
831 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
835 * Move the free pages in a range to the free lists of the requested type.
836 * Note that start_page and end_pages are not aligned on a pageblock
837 * boundary. If alignment is required, use move_freepages_block()
839 static int move_freepages(struct zone *zone,
840 struct page *start_page, struct page *end_page,
847 #ifndef CONFIG_HOLES_IN_ZONE
849 * page_zone is not safe to call in this context when
850 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
851 * anyway as we check zone boundaries in move_freepages_block().
852 * Remove at a later date when no bug reports exist related to
853 * grouping pages by mobility
855 BUG_ON(page_zone(start_page) != page_zone(end_page));
858 for (page = start_page; page <= end_page;) {
859 /* Make sure we are not inadvertently changing nodes */
860 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
862 if (!pfn_valid_within(page_to_pfn(page))) {
867 if (!PageBuddy(page)) {
872 order = page_order(page);
873 list_del(&page->lru);
875 &zone->free_area[order].free_list[migratetype]);
877 pages_moved += 1 << order;
883 static int move_freepages_block(struct zone *zone, struct page *page,
886 unsigned long start_pfn, end_pfn;
887 struct page *start_page, *end_page;
889 start_pfn = page_to_pfn(page);
890 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
891 start_page = pfn_to_page(start_pfn);
892 end_page = start_page + pageblock_nr_pages - 1;
893 end_pfn = start_pfn + pageblock_nr_pages - 1;
895 /* Do not cross zone boundaries */
896 if (start_pfn < zone->zone_start_pfn)
898 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
901 return move_freepages(zone, start_page, end_page, migratetype);
904 static void change_pageblock_range(struct page *pageblock_page,
905 int start_order, int migratetype)
907 int nr_pageblocks = 1 << (start_order - pageblock_order);
909 while (nr_pageblocks--) {
910 set_pageblock_migratetype(pageblock_page, migratetype);
911 pageblock_page += pageblock_nr_pages;
915 /* Remove an element from the buddy allocator from the fallback list */
916 static inline struct page *
917 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
919 struct free_area * area;
924 /* Find the largest possible block of pages in the other list */
925 for (current_order = MAX_ORDER-1; current_order >= order;
927 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
928 migratetype = fallbacks[start_migratetype][i];
930 /* MIGRATE_RESERVE handled later if necessary */
931 if (migratetype == MIGRATE_RESERVE)
934 area = &(zone->free_area[current_order]);
935 if (list_empty(&area->free_list[migratetype]))
938 page = list_entry(area->free_list[migratetype].next,
943 * If breaking a large block of pages, move all free
944 * pages to the preferred allocation list. If falling
945 * back for a reclaimable kernel allocation, be more
946 * agressive about taking ownership of free pages
948 if (unlikely(current_order >= (pageblock_order >> 1)) ||
949 start_migratetype == MIGRATE_RECLAIMABLE ||
950 page_group_by_mobility_disabled) {
952 pages = move_freepages_block(zone, page,
955 /* Claim the whole block if over half of it is free */
956 if (pages >= (1 << (pageblock_order-1)) ||
957 page_group_by_mobility_disabled)
958 set_pageblock_migratetype(page,
961 migratetype = start_migratetype;
964 /* Remove the page from the freelists */
965 list_del(&page->lru);
966 rmv_page_order(page);
968 /* Take ownership for orders >= pageblock_order */
969 if (current_order >= pageblock_order)
970 change_pageblock_range(page, current_order,
973 expand(zone, page, order, current_order, area, migratetype);
975 trace_mm_page_alloc_extfrag(page, order, current_order,
976 start_migratetype, migratetype);
986 * Do the hard work of removing an element from the buddy allocator.
987 * Call me with the zone->lock already held.
989 static struct page *__rmqueue(struct zone *zone, unsigned int order,
995 page = __rmqueue_smallest(zone, order, migratetype);
997 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
998 page = __rmqueue_fallback(zone, order, migratetype);
1001 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1002 * is used because __rmqueue_smallest is an inline function
1003 * and we want just one call site
1006 migratetype = MIGRATE_RESERVE;
1011 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1016 * Obtain a specified number of elements from the buddy allocator, all under
1017 * a single hold of the lock, for efficiency. Add them to the supplied list.
1018 * Returns the number of new pages which were placed at *list.
1020 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1021 unsigned long count, struct list_head *list,
1022 int migratetype, int cold)
1026 spin_lock(&zone->lock);
1027 for (i = 0; i < count; ++i) {
1028 struct page *page = __rmqueue(zone, order, migratetype);
1029 if (unlikely(page == NULL))
1033 * Split buddy pages returned by expand() are received here
1034 * in physical page order. The page is added to the callers and
1035 * list and the list head then moves forward. From the callers
1036 * perspective, the linked list is ordered by page number in
1037 * some conditions. This is useful for IO devices that can
1038 * merge IO requests if the physical pages are ordered
1041 if (likely(cold == 0))
1042 list_add(&page->lru, list);
1044 list_add_tail(&page->lru, list);
1045 set_page_private(page, migratetype);
1048 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1049 spin_unlock(&zone->lock);
1055 * Called from the vmstat counter updater to drain pagesets of this
1056 * currently executing processor on remote nodes after they have
1059 * Note that this function must be called with the thread pinned to
1060 * a single processor.
1062 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1064 unsigned long flags;
1067 local_irq_save(flags);
1068 if (pcp->count >= pcp->batch)
1069 to_drain = pcp->batch;
1071 to_drain = pcp->count;
1072 free_pcppages_bulk(zone, to_drain, pcp);
1073 pcp->count -= to_drain;
1074 local_irq_restore(flags);
1079 * Drain pages of the indicated processor.
1081 * The processor must either be the current processor and the
1082 * thread pinned to the current processor or a processor that
1085 static void drain_pages(unsigned int cpu)
1087 unsigned long flags;
1090 for_each_populated_zone(zone) {
1091 struct per_cpu_pageset *pset;
1092 struct per_cpu_pages *pcp;
1094 local_irq_save(flags);
1095 pset = per_cpu_ptr(zone->pageset, cpu);
1098 free_pcppages_bulk(zone, pcp->count, pcp);
1100 local_irq_restore(flags);
1105 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1107 void drain_local_pages(void *arg)
1109 drain_pages(smp_processor_id());
1113 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1115 void drain_all_pages(void)
1117 on_each_cpu(drain_local_pages, NULL, 1);
1120 #ifdef CONFIG_HIBERNATION
1122 void mark_free_pages(struct zone *zone)
1124 unsigned long pfn, max_zone_pfn;
1125 unsigned long flags;
1127 struct list_head *curr;
1129 if (!zone->spanned_pages)
1132 spin_lock_irqsave(&zone->lock, flags);
1134 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1135 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1136 if (pfn_valid(pfn)) {
1137 struct page *page = pfn_to_page(pfn);
1139 if (!swsusp_page_is_forbidden(page))
1140 swsusp_unset_page_free(page);
1143 for_each_migratetype_order(order, t) {
1144 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1147 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1148 for (i = 0; i < (1UL << order); i++)
1149 swsusp_set_page_free(pfn_to_page(pfn + i));
1152 spin_unlock_irqrestore(&zone->lock, flags);
1154 #endif /* CONFIG_PM */
1157 * Free a 0-order page
1158 * cold == 1 ? free a cold page : free a hot page
1160 void free_hot_cold_page(struct page *page, int cold)
1162 struct zone *zone = page_zone(page);
1163 struct per_cpu_pages *pcp;
1164 unsigned long flags;
1166 int wasMlocked = __TestClearPageMlocked(page);
1168 if (!free_pages_prepare(page, 0))
1171 migratetype = get_pageblock_migratetype(page);
1172 set_page_private(page, migratetype);
1173 local_irq_save(flags);
1174 if (unlikely(wasMlocked))
1175 free_page_mlock(page);
1176 __count_vm_event(PGFREE);
1179 * We only track unmovable, reclaimable and movable on pcp lists.
1180 * Free ISOLATE pages back to the allocator because they are being
1181 * offlined but treat RESERVE as movable pages so we can get those
1182 * areas back if necessary. Otherwise, we may have to free
1183 * excessively into the page allocator
1185 if (migratetype >= MIGRATE_PCPTYPES) {
1186 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1187 free_one_page(zone, page, 0, migratetype);
1190 migratetype = MIGRATE_MOVABLE;
1193 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1195 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1197 list_add(&page->lru, &pcp->lists[migratetype]);
1199 if (pcp->count >= pcp->high) {
1200 free_pcppages_bulk(zone, pcp->batch, pcp);
1201 pcp->count -= pcp->batch;
1205 local_irq_restore(flags);
1209 * split_page takes a non-compound higher-order page, and splits it into
1210 * n (1<<order) sub-pages: page[0..n]
1211 * Each sub-page must be freed individually.
1213 * Note: this is probably too low level an operation for use in drivers.
1214 * Please consult with lkml before using this in your driver.
1216 void split_page(struct page *page, unsigned int order)
1220 VM_BUG_ON(PageCompound(page));
1221 VM_BUG_ON(!page_count(page));
1223 #ifdef CONFIG_KMEMCHECK
1225 * Split shadow pages too, because free(page[0]) would
1226 * otherwise free the whole shadow.
1228 if (kmemcheck_page_is_tracked(page))
1229 split_page(virt_to_page(page[0].shadow), order);
1232 for (i = 1; i < (1 << order); i++)
1233 set_page_refcounted(page + i);
1237 * Similar to split_page except the page is already free. As this is only
1238 * being used for migration, the migratetype of the block also changes.
1239 * As this is called with interrupts disabled, the caller is responsible
1240 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1243 * Note: this is probably too low level an operation for use in drivers.
1244 * Please consult with lkml before using this in your driver.
1246 int split_free_page(struct page *page)
1249 unsigned long watermark;
1252 BUG_ON(!PageBuddy(page));
1254 zone = page_zone(page);
1255 order = page_order(page);
1257 /* Obey watermarks as if the page was being allocated */
1258 watermark = low_wmark_pages(zone) + (1 << order);
1259 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1262 /* Remove page from free list */
1263 list_del(&page->lru);
1264 zone->free_area[order].nr_free--;
1265 rmv_page_order(page);
1266 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1268 /* Split into individual pages */
1269 set_page_refcounted(page);
1270 split_page(page, order);
1272 if (order >= pageblock_order - 1) {
1273 struct page *endpage = page + (1 << order) - 1;
1274 for (; page < endpage; page += pageblock_nr_pages)
1275 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1282 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1283 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1287 struct page *buffered_rmqueue(struct zone *preferred_zone,
1288 struct zone *zone, int order, gfp_t gfp_flags,
1291 unsigned long flags;
1293 int cold = !!(gfp_flags & __GFP_COLD);
1296 if (likely(order == 0)) {
1297 struct per_cpu_pages *pcp;
1298 struct list_head *list;
1300 local_irq_save(flags);
1301 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1302 list = &pcp->lists[migratetype];
1303 if (list_empty(list)) {
1304 pcp->count += rmqueue_bulk(zone, 0,
1307 if (unlikely(list_empty(list)))
1312 page = list_entry(list->prev, struct page, lru);
1314 page = list_entry(list->next, struct page, lru);
1316 list_del(&page->lru);
1319 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1321 * __GFP_NOFAIL is not to be used in new code.
1323 * All __GFP_NOFAIL callers should be fixed so that they
1324 * properly detect and handle allocation failures.
1326 * We most definitely don't want callers attempting to
1327 * allocate greater than order-1 page units with
1330 WARN_ON_ONCE(order > 1);
1332 spin_lock_irqsave(&zone->lock, flags);
1333 page = __rmqueue(zone, order, migratetype);
1334 spin_unlock(&zone->lock);
1337 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1340 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1341 zone_statistics(preferred_zone, zone);
1342 local_irq_restore(flags);
1344 VM_BUG_ON(bad_range(zone, page));
1345 if (prep_new_page(page, order, gfp_flags))
1350 local_irq_restore(flags);
1354 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1355 #define ALLOC_WMARK_MIN WMARK_MIN
1356 #define ALLOC_WMARK_LOW WMARK_LOW
1357 #define ALLOC_WMARK_HIGH WMARK_HIGH
1358 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1360 /* Mask to get the watermark bits */
1361 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1363 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1364 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1365 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1367 #ifdef CONFIG_FAIL_PAGE_ALLOC
1369 static struct fail_page_alloc_attr {
1370 struct fault_attr attr;
1372 u32 ignore_gfp_highmem;
1373 u32 ignore_gfp_wait;
1376 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1378 struct dentry *ignore_gfp_highmem_file;
1379 struct dentry *ignore_gfp_wait_file;
1380 struct dentry *min_order_file;
1382 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1384 } fail_page_alloc = {
1385 .attr = FAULT_ATTR_INITIALIZER,
1386 .ignore_gfp_wait = 1,
1387 .ignore_gfp_highmem = 1,
1391 static int __init setup_fail_page_alloc(char *str)
1393 return setup_fault_attr(&fail_page_alloc.attr, str);
1395 __setup("fail_page_alloc=", setup_fail_page_alloc);
1397 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1399 if (order < fail_page_alloc.min_order)
1401 if (gfp_mask & __GFP_NOFAIL)
1403 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1405 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1408 return should_fail(&fail_page_alloc.attr, 1 << order);
1411 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1413 static int __init fail_page_alloc_debugfs(void)
1415 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1419 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1423 dir = fail_page_alloc.attr.dentries.dir;
1425 fail_page_alloc.ignore_gfp_wait_file =
1426 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1427 &fail_page_alloc.ignore_gfp_wait);
1429 fail_page_alloc.ignore_gfp_highmem_file =
1430 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1431 &fail_page_alloc.ignore_gfp_highmem);
1432 fail_page_alloc.min_order_file =
1433 debugfs_create_u32("min-order", mode, dir,
1434 &fail_page_alloc.min_order);
1436 if (!fail_page_alloc.ignore_gfp_wait_file ||
1437 !fail_page_alloc.ignore_gfp_highmem_file ||
1438 !fail_page_alloc.min_order_file) {
1440 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1441 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1442 debugfs_remove(fail_page_alloc.min_order_file);
1443 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1449 late_initcall(fail_page_alloc_debugfs);
1451 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1453 #else /* CONFIG_FAIL_PAGE_ALLOC */
1455 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1460 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1463 * Return 1 if free pages are above 'mark'. This takes into account the order
1464 * of the allocation.
1466 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1467 int classzone_idx, int alloc_flags)
1469 /* free_pages my go negative - that's OK */
1471 long free_pages = zone_nr_free_pages(z) - (1 << order) + 1;
1474 if (alloc_flags & ALLOC_HIGH)
1476 if (alloc_flags & ALLOC_HARDER)
1479 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1481 for (o = 0; o < order; o++) {
1482 /* At the next order, this order's pages become unavailable */
1483 free_pages -= z->free_area[o].nr_free << o;
1485 /* Require fewer higher order pages to be free */
1486 min >>= min_free_order_shift;
1488 if (free_pages <= min)
1496 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1497 * skip over zones that are not allowed by the cpuset, or that have
1498 * been recently (in last second) found to be nearly full. See further
1499 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1500 * that have to skip over a lot of full or unallowed zones.
1502 * If the zonelist cache is present in the passed in zonelist, then
1503 * returns a pointer to the allowed node mask (either the current
1504 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1506 * If the zonelist cache is not available for this zonelist, does
1507 * nothing and returns NULL.
1509 * If the fullzones BITMAP in the zonelist cache is stale (more than
1510 * a second since last zap'd) then we zap it out (clear its bits.)
1512 * We hold off even calling zlc_setup, until after we've checked the
1513 * first zone in the zonelist, on the theory that most allocations will
1514 * be satisfied from that first zone, so best to examine that zone as
1515 * quickly as we can.
1517 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1519 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1520 nodemask_t *allowednodes; /* zonelist_cache approximation */
1522 zlc = zonelist->zlcache_ptr;
1526 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1527 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1528 zlc->last_full_zap = jiffies;
1531 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1532 &cpuset_current_mems_allowed :
1533 &node_states[N_HIGH_MEMORY];
1534 return allowednodes;
1538 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1539 * if it is worth looking at further for free memory:
1540 * 1) Check that the zone isn't thought to be full (doesn't have its
1541 * bit set in the zonelist_cache fullzones BITMAP).
1542 * 2) Check that the zones node (obtained from the zonelist_cache
1543 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1544 * Return true (non-zero) if zone is worth looking at further, or
1545 * else return false (zero) if it is not.
1547 * This check -ignores- the distinction between various watermarks,
1548 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1549 * found to be full for any variation of these watermarks, it will
1550 * be considered full for up to one second by all requests, unless
1551 * we are so low on memory on all allowed nodes that we are forced
1552 * into the second scan of the zonelist.
1554 * In the second scan we ignore this zonelist cache and exactly
1555 * apply the watermarks to all zones, even it is slower to do so.
1556 * We are low on memory in the second scan, and should leave no stone
1557 * unturned looking for a free page.
1559 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1560 nodemask_t *allowednodes)
1562 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1563 int i; /* index of *z in zonelist zones */
1564 int n; /* node that zone *z is on */
1566 zlc = zonelist->zlcache_ptr;
1570 i = z - zonelist->_zonerefs;
1573 /* This zone is worth trying if it is allowed but not full */
1574 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1578 * Given 'z' scanning a zonelist, set the corresponding bit in
1579 * zlc->fullzones, so that subsequent attempts to allocate a page
1580 * from that zone don't waste time re-examining it.
1582 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1584 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1585 int i; /* index of *z in zonelist zones */
1587 zlc = zonelist->zlcache_ptr;
1591 i = z - zonelist->_zonerefs;
1593 set_bit(i, zlc->fullzones);
1596 #else /* CONFIG_NUMA */
1598 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1603 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1604 nodemask_t *allowednodes)
1609 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1612 #endif /* CONFIG_NUMA */
1615 * get_page_from_freelist goes through the zonelist trying to allocate
1618 static struct page *
1619 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1620 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1621 struct zone *preferred_zone, int migratetype)
1624 struct page *page = NULL;
1627 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1628 int zlc_active = 0; /* set if using zonelist_cache */
1629 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1631 classzone_idx = zone_idx(preferred_zone);
1634 * Scan zonelist, looking for a zone with enough free.
1635 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1637 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1638 high_zoneidx, nodemask) {
1639 if (NUMA_BUILD && zlc_active &&
1640 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1642 if ((alloc_flags & ALLOC_CPUSET) &&
1643 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1646 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1647 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1651 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1652 if (zone_watermark_ok(zone, order, mark,
1653 classzone_idx, alloc_flags))
1656 if (zone_reclaim_mode == 0)
1657 goto this_zone_full;
1659 ret = zone_reclaim(zone, gfp_mask, order);
1661 case ZONE_RECLAIM_NOSCAN:
1664 case ZONE_RECLAIM_FULL:
1665 /* scanned but unreclaimable */
1666 goto this_zone_full;
1668 /* did we reclaim enough */
1669 if (!zone_watermark_ok(zone, order, mark,
1670 classzone_idx, alloc_flags))
1671 goto this_zone_full;
1676 page = buffered_rmqueue(preferred_zone, zone, order,
1677 gfp_mask, migratetype);
1682 zlc_mark_zone_full(zonelist, z);
1684 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1686 * we do zlc_setup after the first zone is tried but only
1687 * if there are multiple nodes make it worthwhile
1689 allowednodes = zlc_setup(zonelist, alloc_flags);
1695 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1696 /* Disable zlc cache for second zonelist scan */
1704 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1705 unsigned long pages_reclaimed)
1707 /* Do not loop if specifically requested */
1708 if (gfp_mask & __GFP_NORETRY)
1712 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1713 * means __GFP_NOFAIL, but that may not be true in other
1716 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1720 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1721 * specified, then we retry until we no longer reclaim any pages
1722 * (above), or we've reclaimed an order of pages at least as
1723 * large as the allocation's order. In both cases, if the
1724 * allocation still fails, we stop retrying.
1726 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1730 * Don't let big-order allocations loop unless the caller
1731 * explicitly requests that.
1733 if (gfp_mask & __GFP_NOFAIL)
1739 static inline struct page *
1740 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1741 struct zonelist *zonelist, enum zone_type high_zoneidx,
1742 nodemask_t *nodemask, struct zone *preferred_zone,
1747 /* Acquire the OOM killer lock for the zones in zonelist */
1748 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1749 schedule_timeout_uninterruptible(1);
1754 * Go through the zonelist yet one more time, keep very high watermark
1755 * here, this is only to catch a parallel oom killing, we must fail if
1756 * we're still under heavy pressure.
1758 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1759 order, zonelist, high_zoneidx,
1760 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1761 preferred_zone, migratetype);
1765 if (!(gfp_mask & __GFP_NOFAIL)) {
1766 /* The OOM killer will not help higher order allocs */
1767 if (order > PAGE_ALLOC_COSTLY_ORDER)
1769 /* The OOM killer does not needlessly kill tasks for lowmem */
1770 if (high_zoneidx < ZONE_NORMAL)
1773 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1774 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1775 * The caller should handle page allocation failure by itself if
1776 * it specifies __GFP_THISNODE.
1777 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1779 if (gfp_mask & __GFP_THISNODE)
1782 /* Exhausted what can be done so it's blamo time */
1783 out_of_memory(zonelist, gfp_mask, order, nodemask);
1786 clear_zonelist_oom(zonelist, gfp_mask);
1790 #ifdef CONFIG_COMPACTION
1791 /* Try memory compaction for high-order allocations before reclaim */
1792 static struct page *
1793 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1794 struct zonelist *zonelist, enum zone_type high_zoneidx,
1795 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1796 int migratetype, unsigned long *did_some_progress)
1800 if (!order || compaction_deferred(preferred_zone))
1803 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1805 if (*did_some_progress != COMPACT_SKIPPED) {
1807 /* Page migration frees to the PCP lists but we want merging */
1808 drain_pages(get_cpu());
1811 page = get_page_from_freelist(gfp_mask, nodemask,
1812 order, zonelist, high_zoneidx,
1813 alloc_flags, preferred_zone,
1816 preferred_zone->compact_considered = 0;
1817 preferred_zone->compact_defer_shift = 0;
1818 count_vm_event(COMPACTSUCCESS);
1823 * It's bad if compaction run occurs and fails.
1824 * The most likely reason is that pages exist,
1825 * but not enough to satisfy watermarks.
1827 count_vm_event(COMPACTFAIL);
1828 defer_compaction(preferred_zone);
1836 static inline struct page *
1837 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1838 struct zonelist *zonelist, enum zone_type high_zoneidx,
1839 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1840 int migratetype, unsigned long *did_some_progress)
1844 #endif /* CONFIG_COMPACTION */
1846 /* The really slow allocator path where we enter direct reclaim */
1847 static inline struct page *
1848 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1849 struct zonelist *zonelist, enum zone_type high_zoneidx,
1850 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1851 int migratetype, unsigned long *did_some_progress)
1853 struct page *page = NULL;
1854 struct reclaim_state reclaim_state;
1855 struct task_struct *p = current;
1856 bool drained = false;
1860 /* We now go into synchronous reclaim */
1861 cpuset_memory_pressure_bump();
1862 p->flags |= PF_MEMALLOC;
1863 lockdep_set_current_reclaim_state(gfp_mask);
1864 reclaim_state.reclaimed_slab = 0;
1865 p->reclaim_state = &reclaim_state;
1867 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1869 p->reclaim_state = NULL;
1870 lockdep_clear_current_reclaim_state();
1871 p->flags &= ~PF_MEMALLOC;
1875 if (unlikely(!(*did_some_progress)))
1879 page = get_page_from_freelist(gfp_mask, nodemask, order,
1880 zonelist, high_zoneidx,
1881 alloc_flags, preferred_zone,
1885 * If an allocation failed after direct reclaim, it could be because
1886 * pages are pinned on the per-cpu lists. Drain them and try again
1888 if (!page && !drained) {
1898 * This is called in the allocator slow-path if the allocation request is of
1899 * sufficient urgency to ignore watermarks and take other desperate measures
1901 static inline struct page *
1902 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1903 struct zonelist *zonelist, enum zone_type high_zoneidx,
1904 nodemask_t *nodemask, struct zone *preferred_zone,
1910 page = get_page_from_freelist(gfp_mask, nodemask, order,
1911 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1912 preferred_zone, migratetype);
1914 if (!page && gfp_mask & __GFP_NOFAIL)
1915 congestion_wait(BLK_RW_ASYNC, HZ/50);
1916 } while (!page && (gfp_mask & __GFP_NOFAIL));
1922 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1923 enum zone_type high_zoneidx)
1928 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1929 wakeup_kswapd(zone, order);
1933 gfp_to_alloc_flags(gfp_t gfp_mask)
1935 struct task_struct *p = current;
1936 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1937 const gfp_t wait = gfp_mask & __GFP_WAIT;
1939 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1940 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1943 * The caller may dip into page reserves a bit more if the caller
1944 * cannot run direct reclaim, or if the caller has realtime scheduling
1945 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1946 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1948 alloc_flags |= (gfp_mask & __GFP_HIGH);
1951 alloc_flags |= ALLOC_HARDER;
1953 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1954 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1956 alloc_flags &= ~ALLOC_CPUSET;
1957 } else if (unlikely(rt_task(p)) && !in_interrupt())
1958 alloc_flags |= ALLOC_HARDER;
1960 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1961 if (!in_interrupt() &&
1962 ((p->flags & PF_MEMALLOC) ||
1963 unlikely(test_thread_flag(TIF_MEMDIE))))
1964 alloc_flags |= ALLOC_NO_WATERMARKS;
1970 static inline struct page *
1971 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1972 struct zonelist *zonelist, enum zone_type high_zoneidx,
1973 nodemask_t *nodemask, struct zone *preferred_zone,
1976 const gfp_t wait = gfp_mask & __GFP_WAIT;
1977 struct page *page = NULL;
1979 unsigned long pages_reclaimed = 0;
1980 unsigned long did_some_progress;
1981 struct task_struct *p = current;
1984 * In the slowpath, we sanity check order to avoid ever trying to
1985 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1986 * be using allocators in order of preference for an area that is
1989 if (order >= MAX_ORDER) {
1990 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1995 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1996 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1997 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1998 * using a larger set of nodes after it has established that the
1999 * allowed per node queues are empty and that nodes are
2002 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2006 if (!(gfp_mask & __GFP_NO_KSWAPD))
2007 wake_all_kswapd(order, zonelist, high_zoneidx);
2010 * OK, we're below the kswapd watermark and have kicked background
2011 * reclaim. Now things get more complex, so set up alloc_flags according
2012 * to how we want to proceed.
2014 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2016 /* This is the last chance, in general, before the goto nopage. */
2017 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2018 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2019 preferred_zone, migratetype);
2024 /* Allocate without watermarks if the context allows */
2025 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2026 page = __alloc_pages_high_priority(gfp_mask, order,
2027 zonelist, high_zoneidx, nodemask,
2028 preferred_zone, migratetype);
2033 /* Atomic allocations - we can't balance anything */
2037 /* Avoid recursion of direct reclaim */
2038 if (p->flags & PF_MEMALLOC)
2041 /* Avoid allocations with no watermarks from looping endlessly */
2042 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2045 /* Try direct compaction */
2046 page = __alloc_pages_direct_compact(gfp_mask, order,
2047 zonelist, high_zoneidx,
2049 alloc_flags, preferred_zone,
2050 migratetype, &did_some_progress);
2054 /* Try direct reclaim and then allocating */
2055 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2056 zonelist, high_zoneidx,
2058 alloc_flags, preferred_zone,
2059 migratetype, &did_some_progress);
2064 * If we failed to make any progress reclaiming, then we are
2065 * running out of options and have to consider going OOM
2067 if (!did_some_progress) {
2068 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2069 if (oom_killer_disabled)
2071 page = __alloc_pages_may_oom(gfp_mask, order,
2072 zonelist, high_zoneidx,
2073 nodemask, preferred_zone,
2078 if (!(gfp_mask & __GFP_NOFAIL)) {
2080 * The oom killer is not called for high-order
2081 * allocations that may fail, so if no progress
2082 * is being made, there are no other options and
2083 * retrying is unlikely to help.
2085 if (order > PAGE_ALLOC_COSTLY_ORDER)
2088 * The oom killer is not called for lowmem
2089 * allocations to prevent needlessly killing
2092 if (high_zoneidx < ZONE_NORMAL)
2100 /* Check if we should retry the allocation */
2101 pages_reclaimed += did_some_progress;
2102 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2103 /* Wait for some write requests to complete then retry */
2104 congestion_wait(BLK_RW_ASYNC, HZ/50);
2109 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
2110 printk(KERN_WARNING "%s: page allocation failure."
2111 " order:%d, mode:0x%x\n",
2112 p->comm, order, gfp_mask);
2118 if (kmemcheck_enabled)
2119 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2125 * This is the 'heart' of the zoned buddy allocator.
2128 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2129 struct zonelist *zonelist, nodemask_t *nodemask)
2131 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2132 struct zone *preferred_zone;
2134 int migratetype = allocflags_to_migratetype(gfp_mask);
2136 gfp_mask &= gfp_allowed_mask;
2138 lockdep_trace_alloc(gfp_mask);
2140 might_sleep_if(gfp_mask & __GFP_WAIT);
2142 if (should_fail_alloc_page(gfp_mask, order))
2146 * Check the zones suitable for the gfp_mask contain at least one
2147 * valid zone. It's possible to have an empty zonelist as a result
2148 * of GFP_THISNODE and a memoryless node
2150 if (unlikely(!zonelist->_zonerefs->zone))
2154 /* The preferred zone is used for statistics later */
2155 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
2156 if (!preferred_zone) {
2161 /* First allocation attempt */
2162 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2163 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2164 preferred_zone, migratetype);
2165 if (unlikely(!page))
2166 page = __alloc_pages_slowpath(gfp_mask, order,
2167 zonelist, high_zoneidx, nodemask,
2168 preferred_zone, migratetype);
2171 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2174 EXPORT_SYMBOL(__alloc_pages_nodemask);
2177 * Common helper functions.
2179 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2184 * __get_free_pages() returns a 32-bit address, which cannot represent
2187 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2189 page = alloc_pages(gfp_mask, order);
2192 return (unsigned long) page_address(page);
2194 EXPORT_SYMBOL(__get_free_pages);
2196 unsigned long get_zeroed_page(gfp_t gfp_mask)
2198 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2200 EXPORT_SYMBOL(get_zeroed_page);
2202 void __pagevec_free(struct pagevec *pvec)
2204 int i = pagevec_count(pvec);
2207 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2208 free_hot_cold_page(pvec->pages[i], pvec->cold);
2212 void __free_pages(struct page *page, unsigned int order)
2214 if (put_page_testzero(page)) {
2216 free_hot_cold_page(page, 0);
2218 __free_pages_ok(page, order);
2222 EXPORT_SYMBOL(__free_pages);
2224 void free_pages(unsigned long addr, unsigned int order)
2227 VM_BUG_ON(!virt_addr_valid((void *)addr));
2228 __free_pages(virt_to_page((void *)addr), order);
2232 EXPORT_SYMBOL(free_pages);
2235 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2236 * @size: the number of bytes to allocate
2237 * @gfp_mask: GFP flags for the allocation
2239 * This function is similar to alloc_pages(), except that it allocates the
2240 * minimum number of pages to satisfy the request. alloc_pages() can only
2241 * allocate memory in power-of-two pages.
2243 * This function is also limited by MAX_ORDER.
2245 * Memory allocated by this function must be released by free_pages_exact().
2247 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2249 unsigned int order = get_order(size);
2252 addr = __get_free_pages(gfp_mask, order);
2254 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2255 unsigned long used = addr + PAGE_ALIGN(size);
2257 split_page(virt_to_page((void *)addr), order);
2258 while (used < alloc_end) {
2264 return (void *)addr;
2266 EXPORT_SYMBOL(alloc_pages_exact);
2269 * free_pages_exact - release memory allocated via alloc_pages_exact()
2270 * @virt: the value returned by alloc_pages_exact.
2271 * @size: size of allocation, same value as passed to alloc_pages_exact().
2273 * Release the memory allocated by a previous call to alloc_pages_exact.
2275 void free_pages_exact(void *virt, size_t size)
2277 unsigned long addr = (unsigned long)virt;
2278 unsigned long end = addr + PAGE_ALIGN(size);
2280 while (addr < end) {
2285 EXPORT_SYMBOL(free_pages_exact);
2287 static unsigned int nr_free_zone_pages(int offset)
2292 /* Just pick one node, since fallback list is circular */
2293 unsigned int sum = 0;
2295 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2297 for_each_zone_zonelist(zone, z, zonelist, offset) {
2298 unsigned long size = zone->present_pages;
2299 unsigned long high = high_wmark_pages(zone);
2308 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2310 unsigned int nr_free_buffer_pages(void)
2312 return nr_free_zone_pages(gfp_zone(GFP_USER));
2314 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2317 * Amount of free RAM allocatable within all zones
2319 unsigned int nr_free_pagecache_pages(void)
2321 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2324 static inline void show_node(struct zone *zone)
2327 printk("Node %d ", zone_to_nid(zone));
2330 void si_meminfo(struct sysinfo *val)
2332 val->totalram = totalram_pages;
2334 val->freeram = global_page_state(NR_FREE_PAGES);
2335 val->bufferram = nr_blockdev_pages();
2336 val->totalhigh = totalhigh_pages;
2337 val->freehigh = nr_free_highpages();
2338 val->mem_unit = PAGE_SIZE;
2341 EXPORT_SYMBOL(si_meminfo);
2344 void si_meminfo_node(struct sysinfo *val, int nid)
2346 pg_data_t *pgdat = NODE_DATA(nid);
2348 val->totalram = pgdat->node_present_pages;
2349 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2350 #ifdef CONFIG_HIGHMEM
2351 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2352 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2358 val->mem_unit = PAGE_SIZE;
2362 #define K(x) ((x) << (PAGE_SHIFT-10))
2365 * Show free area list (used inside shift_scroll-lock stuff)
2366 * We also calculate the percentage fragmentation. We do this by counting the
2367 * memory on each free list with the exception of the first item on the list.
2369 void show_free_areas(void)
2374 for_each_populated_zone(zone) {
2376 printk("%s per-cpu:\n", zone->name);
2378 for_each_online_cpu(cpu) {
2379 struct per_cpu_pageset *pageset;
2381 pageset = per_cpu_ptr(zone->pageset, cpu);
2383 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2384 cpu, pageset->pcp.high,
2385 pageset->pcp.batch, pageset->pcp.count);
2389 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2390 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2392 " dirty:%lu writeback:%lu unstable:%lu\n"
2393 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2394 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2395 global_page_state(NR_ACTIVE_ANON),
2396 global_page_state(NR_INACTIVE_ANON),
2397 global_page_state(NR_ISOLATED_ANON),
2398 global_page_state(NR_ACTIVE_FILE),
2399 global_page_state(NR_INACTIVE_FILE),
2400 global_page_state(NR_ISOLATED_FILE),
2401 global_page_state(NR_UNEVICTABLE),
2402 global_page_state(NR_FILE_DIRTY),
2403 global_page_state(NR_WRITEBACK),
2404 global_page_state(NR_UNSTABLE_NFS),
2405 global_page_state(NR_FREE_PAGES),
2406 global_page_state(NR_SLAB_RECLAIMABLE),
2407 global_page_state(NR_SLAB_UNRECLAIMABLE),
2408 global_page_state(NR_FILE_MAPPED),
2409 global_page_state(NR_SHMEM),
2410 global_page_state(NR_PAGETABLE),
2411 global_page_state(NR_BOUNCE));
2413 for_each_populated_zone(zone) {
2422 " active_anon:%lukB"
2423 " inactive_anon:%lukB"
2424 " active_file:%lukB"
2425 " inactive_file:%lukB"
2426 " unevictable:%lukB"
2427 " isolated(anon):%lukB"
2428 " isolated(file):%lukB"
2435 " slab_reclaimable:%lukB"
2436 " slab_unreclaimable:%lukB"
2437 " kernel_stack:%lukB"
2441 " writeback_tmp:%lukB"
2442 " pages_scanned:%lu"
2443 " all_unreclaimable? %s"
2446 K(zone_nr_free_pages(zone)),
2447 K(min_wmark_pages(zone)),
2448 K(low_wmark_pages(zone)),
2449 K(high_wmark_pages(zone)),
2450 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2451 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2452 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2453 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2454 K(zone_page_state(zone, NR_UNEVICTABLE)),
2455 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2456 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2457 K(zone->present_pages),
2458 K(zone_page_state(zone, NR_MLOCK)),
2459 K(zone_page_state(zone, NR_FILE_DIRTY)),
2460 K(zone_page_state(zone, NR_WRITEBACK)),
2461 K(zone_page_state(zone, NR_FILE_MAPPED)),
2462 K(zone_page_state(zone, NR_SHMEM)),
2463 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2464 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2465 zone_page_state(zone, NR_KERNEL_STACK) *
2467 K(zone_page_state(zone, NR_PAGETABLE)),
2468 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2469 K(zone_page_state(zone, NR_BOUNCE)),
2470 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2471 zone->pages_scanned,
2472 (zone->all_unreclaimable ? "yes" : "no")
2474 printk("lowmem_reserve[]:");
2475 for (i = 0; i < MAX_NR_ZONES; i++)
2476 printk(" %lu", zone->lowmem_reserve[i]);
2480 for_each_populated_zone(zone) {
2481 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2484 printk("%s: ", zone->name);
2486 spin_lock_irqsave(&zone->lock, flags);
2487 for (order = 0; order < MAX_ORDER; order++) {
2488 nr[order] = zone->free_area[order].nr_free;
2489 total += nr[order] << order;
2491 spin_unlock_irqrestore(&zone->lock, flags);
2492 for (order = 0; order < MAX_ORDER; order++)
2493 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2494 printk("= %lukB\n", K(total));
2497 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2499 show_swap_cache_info();
2502 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2504 zoneref->zone = zone;
2505 zoneref->zone_idx = zone_idx(zone);
2509 * Builds allocation fallback zone lists.
2511 * Add all populated zones of a node to the zonelist.
2513 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2514 int nr_zones, enum zone_type zone_type)
2518 BUG_ON(zone_type >= MAX_NR_ZONES);
2523 zone = pgdat->node_zones + zone_type;
2524 if (populated_zone(zone)) {
2525 zoneref_set_zone(zone,
2526 &zonelist->_zonerefs[nr_zones++]);
2527 check_highest_zone(zone_type);
2530 } while (zone_type);
2537 * 0 = automatic detection of better ordering.
2538 * 1 = order by ([node] distance, -zonetype)
2539 * 2 = order by (-zonetype, [node] distance)
2541 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2542 * the same zonelist. So only NUMA can configure this param.
2544 #define ZONELIST_ORDER_DEFAULT 0
2545 #define ZONELIST_ORDER_NODE 1
2546 #define ZONELIST_ORDER_ZONE 2
2548 /* zonelist order in the kernel.
2549 * set_zonelist_order() will set this to NODE or ZONE.
2551 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2552 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2556 /* The value user specified ....changed by config */
2557 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2558 /* string for sysctl */
2559 #define NUMA_ZONELIST_ORDER_LEN 16
2560 char numa_zonelist_order[16] = "default";
2563 * interface for configure zonelist ordering.
2564 * command line option "numa_zonelist_order"
2565 * = "[dD]efault - default, automatic configuration.
2566 * = "[nN]ode - order by node locality, then by zone within node
2567 * = "[zZ]one - order by zone, then by locality within zone
2570 static int __parse_numa_zonelist_order(char *s)
2572 if (*s == 'd' || *s == 'D') {
2573 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2574 } else if (*s == 'n' || *s == 'N') {
2575 user_zonelist_order = ZONELIST_ORDER_NODE;
2576 } else if (*s == 'z' || *s == 'Z') {
2577 user_zonelist_order = ZONELIST_ORDER_ZONE;
2580 "Ignoring invalid numa_zonelist_order value: "
2587 static __init int setup_numa_zonelist_order(char *s)
2590 return __parse_numa_zonelist_order(s);
2593 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2596 * sysctl handler for numa_zonelist_order
2598 int numa_zonelist_order_handler(ctl_table *table, int write,
2599 void __user *buffer, size_t *length,
2602 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2604 static DEFINE_MUTEX(zl_order_mutex);
2606 mutex_lock(&zl_order_mutex);
2608 strcpy(saved_string, (char*)table->data);
2609 ret = proc_dostring(table, write, buffer, length, ppos);
2613 int oldval = user_zonelist_order;
2614 if (__parse_numa_zonelist_order((char*)table->data)) {
2616 * bogus value. restore saved string
2618 strncpy((char*)table->data, saved_string,
2619 NUMA_ZONELIST_ORDER_LEN);
2620 user_zonelist_order = oldval;
2621 } else if (oldval != user_zonelist_order) {
2622 mutex_lock(&zonelists_mutex);
2623 build_all_zonelists(NULL);
2624 mutex_unlock(&zonelists_mutex);
2628 mutex_unlock(&zl_order_mutex);
2633 #define MAX_NODE_LOAD (nr_online_nodes)
2634 static int node_load[MAX_NUMNODES];
2637 * find_next_best_node - find the next node that should appear in a given node's fallback list
2638 * @node: node whose fallback list we're appending
2639 * @used_node_mask: nodemask_t of already used nodes
2641 * We use a number of factors to determine which is the next node that should
2642 * appear on a given node's fallback list. The node should not have appeared
2643 * already in @node's fallback list, and it should be the next closest node
2644 * according to the distance array (which contains arbitrary distance values
2645 * from each node to each node in the system), and should also prefer nodes
2646 * with no CPUs, since presumably they'll have very little allocation pressure
2647 * on them otherwise.
2648 * It returns -1 if no node is found.
2650 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2653 int min_val = INT_MAX;
2655 const struct cpumask *tmp = cpumask_of_node(0);
2657 /* Use the local node if we haven't already */
2658 if (!node_isset(node, *used_node_mask)) {
2659 node_set(node, *used_node_mask);
2663 for_each_node_state(n, N_HIGH_MEMORY) {
2665 /* Don't want a node to appear more than once */
2666 if (node_isset(n, *used_node_mask))
2669 /* Use the distance array to find the distance */
2670 val = node_distance(node, n);
2672 /* Penalize nodes under us ("prefer the next node") */
2675 /* Give preference to headless and unused nodes */
2676 tmp = cpumask_of_node(n);
2677 if (!cpumask_empty(tmp))
2678 val += PENALTY_FOR_NODE_WITH_CPUS;
2680 /* Slight preference for less loaded node */
2681 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2682 val += node_load[n];
2684 if (val < min_val) {
2691 node_set(best_node, *used_node_mask);
2698 * Build zonelists ordered by node and zones within node.
2699 * This results in maximum locality--normal zone overflows into local
2700 * DMA zone, if any--but risks exhausting DMA zone.
2702 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2705 struct zonelist *zonelist;
2707 zonelist = &pgdat->node_zonelists[0];
2708 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2710 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2712 zonelist->_zonerefs[j].zone = NULL;
2713 zonelist->_zonerefs[j].zone_idx = 0;
2717 * Build gfp_thisnode zonelists
2719 static void build_thisnode_zonelists(pg_data_t *pgdat)
2722 struct zonelist *zonelist;
2724 zonelist = &pgdat->node_zonelists[1];
2725 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2726 zonelist->_zonerefs[j].zone = NULL;
2727 zonelist->_zonerefs[j].zone_idx = 0;
2731 * Build zonelists ordered by zone and nodes within zones.
2732 * This results in conserving DMA zone[s] until all Normal memory is
2733 * exhausted, but results in overflowing to remote node while memory
2734 * may still exist in local DMA zone.
2736 static int node_order[MAX_NUMNODES];
2738 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2741 int zone_type; /* needs to be signed */
2743 struct zonelist *zonelist;
2745 zonelist = &pgdat->node_zonelists[0];
2747 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2748 for (j = 0; j < nr_nodes; j++) {
2749 node = node_order[j];
2750 z = &NODE_DATA(node)->node_zones[zone_type];
2751 if (populated_zone(z)) {
2753 &zonelist->_zonerefs[pos++]);
2754 check_highest_zone(zone_type);
2758 zonelist->_zonerefs[pos].zone = NULL;
2759 zonelist->_zonerefs[pos].zone_idx = 0;
2762 static int default_zonelist_order(void)
2765 unsigned long low_kmem_size,total_size;
2769 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2770 * If they are really small and used heavily, the system can fall
2771 * into OOM very easily.
2772 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2774 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2777 for_each_online_node(nid) {
2778 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2779 z = &NODE_DATA(nid)->node_zones[zone_type];
2780 if (populated_zone(z)) {
2781 if (zone_type < ZONE_NORMAL)
2782 low_kmem_size += z->present_pages;
2783 total_size += z->present_pages;
2784 } else if (zone_type == ZONE_NORMAL) {
2786 * If any node has only lowmem, then node order
2787 * is preferred to allow kernel allocations
2788 * locally; otherwise, they can easily infringe
2789 * on other nodes when there is an abundance of
2790 * lowmem available to allocate from.
2792 return ZONELIST_ORDER_NODE;
2796 if (!low_kmem_size || /* there are no DMA area. */
2797 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2798 return ZONELIST_ORDER_NODE;
2800 * look into each node's config.
2801 * If there is a node whose DMA/DMA32 memory is very big area on
2802 * local memory, NODE_ORDER may be suitable.
2804 average_size = total_size /
2805 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2806 for_each_online_node(nid) {
2809 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2810 z = &NODE_DATA(nid)->node_zones[zone_type];
2811 if (populated_zone(z)) {
2812 if (zone_type < ZONE_NORMAL)
2813 low_kmem_size += z->present_pages;
2814 total_size += z->present_pages;
2817 if (low_kmem_size &&
2818 total_size > average_size && /* ignore small node */
2819 low_kmem_size > total_size * 70/100)
2820 return ZONELIST_ORDER_NODE;
2822 return ZONELIST_ORDER_ZONE;
2825 static void set_zonelist_order(void)
2827 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2828 current_zonelist_order = default_zonelist_order();
2830 current_zonelist_order = user_zonelist_order;
2833 static void build_zonelists(pg_data_t *pgdat)
2837 nodemask_t used_mask;
2838 int local_node, prev_node;
2839 struct zonelist *zonelist;
2840 int order = current_zonelist_order;
2842 /* initialize zonelists */
2843 for (i = 0; i < MAX_ZONELISTS; i++) {
2844 zonelist = pgdat->node_zonelists + i;
2845 zonelist->_zonerefs[0].zone = NULL;
2846 zonelist->_zonerefs[0].zone_idx = 0;
2849 /* NUMA-aware ordering of nodes */
2850 local_node = pgdat->node_id;
2851 load = nr_online_nodes;
2852 prev_node = local_node;
2853 nodes_clear(used_mask);
2855 memset(node_order, 0, sizeof(node_order));
2858 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2859 int distance = node_distance(local_node, node);
2862 * If another node is sufficiently far away then it is better
2863 * to reclaim pages in a zone before going off node.
2865 if (distance > RECLAIM_DISTANCE)
2866 zone_reclaim_mode = 1;
2869 * We don't want to pressure a particular node.
2870 * So adding penalty to the first node in same
2871 * distance group to make it round-robin.
2873 if (distance != node_distance(local_node, prev_node))
2874 node_load[node] = load;
2878 if (order == ZONELIST_ORDER_NODE)
2879 build_zonelists_in_node_order(pgdat, node);
2881 node_order[j++] = node; /* remember order */
2884 if (order == ZONELIST_ORDER_ZONE) {
2885 /* calculate node order -- i.e., DMA last! */
2886 build_zonelists_in_zone_order(pgdat, j);
2889 build_thisnode_zonelists(pgdat);
2892 /* Construct the zonelist performance cache - see further mmzone.h */
2893 static void build_zonelist_cache(pg_data_t *pgdat)
2895 struct zonelist *zonelist;
2896 struct zonelist_cache *zlc;
2899 zonelist = &pgdat->node_zonelists[0];
2900 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2901 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2902 for (z = zonelist->_zonerefs; z->zone; z++)
2903 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2906 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
2908 * Return node id of node used for "local" allocations.
2909 * I.e., first node id of first zone in arg node's generic zonelist.
2910 * Used for initializing percpu 'numa_mem', which is used primarily
2911 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
2913 int local_memory_node(int node)
2917 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
2918 gfp_zone(GFP_KERNEL),
2925 #else /* CONFIG_NUMA */
2927 static void set_zonelist_order(void)
2929 current_zonelist_order = ZONELIST_ORDER_ZONE;
2932 static void build_zonelists(pg_data_t *pgdat)
2934 int node, local_node;
2936 struct zonelist *zonelist;
2938 local_node = pgdat->node_id;
2940 zonelist = &pgdat->node_zonelists[0];
2941 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2944 * Now we build the zonelist so that it contains the zones
2945 * of all the other nodes.
2946 * We don't want to pressure a particular node, so when
2947 * building the zones for node N, we make sure that the
2948 * zones coming right after the local ones are those from
2949 * node N+1 (modulo N)
2951 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2952 if (!node_online(node))
2954 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2957 for (node = 0; node < local_node; node++) {
2958 if (!node_online(node))
2960 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2964 zonelist->_zonerefs[j].zone = NULL;
2965 zonelist->_zonerefs[j].zone_idx = 0;
2968 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2969 static void build_zonelist_cache(pg_data_t *pgdat)
2971 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2974 #endif /* CONFIG_NUMA */
2977 * Boot pageset table. One per cpu which is going to be used for all
2978 * zones and all nodes. The parameters will be set in such a way
2979 * that an item put on a list will immediately be handed over to
2980 * the buddy list. This is safe since pageset manipulation is done
2981 * with interrupts disabled.
2983 * The boot_pagesets must be kept even after bootup is complete for
2984 * unused processors and/or zones. They do play a role for bootstrapping
2985 * hotplugged processors.
2987 * zoneinfo_show() and maybe other functions do
2988 * not check if the processor is online before following the pageset pointer.
2989 * Other parts of the kernel may not check if the zone is available.
2991 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
2992 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
2993 static void setup_zone_pageset(struct zone *zone);
2996 * Global mutex to protect against size modification of zonelists
2997 * as well as to serialize pageset setup for the new populated zone.
2999 DEFINE_MUTEX(zonelists_mutex);
3001 /* return values int ....just for stop_machine() */
3002 static __init_refok int __build_all_zonelists(void *data)
3008 memset(node_load, 0, sizeof(node_load));
3010 for_each_online_node(nid) {
3011 pg_data_t *pgdat = NODE_DATA(nid);
3013 build_zonelists(pgdat);
3014 build_zonelist_cache(pgdat);
3017 #ifdef CONFIG_MEMORY_HOTPLUG
3018 /* Setup real pagesets for the new zone */
3020 struct zone *zone = data;
3021 setup_zone_pageset(zone);
3026 * Initialize the boot_pagesets that are going to be used
3027 * for bootstrapping processors. The real pagesets for
3028 * each zone will be allocated later when the per cpu
3029 * allocator is available.
3031 * boot_pagesets are used also for bootstrapping offline
3032 * cpus if the system is already booted because the pagesets
3033 * are needed to initialize allocators on a specific cpu too.
3034 * F.e. the percpu allocator needs the page allocator which
3035 * needs the percpu allocator in order to allocate its pagesets
3036 * (a chicken-egg dilemma).
3038 for_each_possible_cpu(cpu) {
3039 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3041 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3043 * We now know the "local memory node" for each node--
3044 * i.e., the node of the first zone in the generic zonelist.
3045 * Set up numa_mem percpu variable for on-line cpus. During
3046 * boot, only the boot cpu should be on-line; we'll init the
3047 * secondary cpus' numa_mem as they come on-line. During
3048 * node/memory hotplug, we'll fixup all on-line cpus.
3050 if (cpu_online(cpu))
3051 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3059 * Called with zonelists_mutex held always
3060 * unless system_state == SYSTEM_BOOTING.
3062 void build_all_zonelists(void *data)
3064 set_zonelist_order();
3066 if (system_state == SYSTEM_BOOTING) {
3067 __build_all_zonelists(NULL);
3068 mminit_verify_zonelist();
3069 cpuset_init_current_mems_allowed();
3071 /* we have to stop all cpus to guarantee there is no user
3073 stop_machine(__build_all_zonelists, data, NULL);
3074 /* cpuset refresh routine should be here */
3076 vm_total_pages = nr_free_pagecache_pages();
3078 * Disable grouping by mobility if the number of pages in the
3079 * system is too low to allow the mechanism to work. It would be
3080 * more accurate, but expensive to check per-zone. This check is
3081 * made on memory-hotadd so a system can start with mobility
3082 * disabled and enable it later
3084 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3085 page_group_by_mobility_disabled = 1;
3087 page_group_by_mobility_disabled = 0;
3089 printk("Built %i zonelists in %s order, mobility grouping %s. "
3090 "Total pages: %ld\n",
3092 zonelist_order_name[current_zonelist_order],
3093 page_group_by_mobility_disabled ? "off" : "on",
3096 printk("Policy zone: %s\n", zone_names[policy_zone]);
3101 * Helper functions to size the waitqueue hash table.
3102 * Essentially these want to choose hash table sizes sufficiently
3103 * large so that collisions trying to wait on pages are rare.
3104 * But in fact, the number of active page waitqueues on typical
3105 * systems is ridiculously low, less than 200. So this is even
3106 * conservative, even though it seems large.
3108 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3109 * waitqueues, i.e. the size of the waitq table given the number of pages.
3111 #define PAGES_PER_WAITQUEUE 256
3113 #ifndef CONFIG_MEMORY_HOTPLUG
3114 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3116 unsigned long size = 1;
3118 pages /= PAGES_PER_WAITQUEUE;
3120 while (size < pages)
3124 * Once we have dozens or even hundreds of threads sleeping
3125 * on IO we've got bigger problems than wait queue collision.
3126 * Limit the size of the wait table to a reasonable size.
3128 size = min(size, 4096UL);
3130 return max(size, 4UL);
3134 * A zone's size might be changed by hot-add, so it is not possible to determine
3135 * a suitable size for its wait_table. So we use the maximum size now.
3137 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3139 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3140 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3141 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3143 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3144 * or more by the traditional way. (See above). It equals:
3146 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3147 * ia64(16K page size) : = ( 8G + 4M)byte.
3148 * powerpc (64K page size) : = (32G +16M)byte.
3150 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3157 * This is an integer logarithm so that shifts can be used later
3158 * to extract the more random high bits from the multiplicative
3159 * hash function before the remainder is taken.
3161 static inline unsigned long wait_table_bits(unsigned long size)
3166 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3169 * Check if a pageblock contains reserved pages
3171 static int pageblock_is_reserved(unsigned long start_pfn)
3173 unsigned long end_pfn = start_pfn + pageblock_nr_pages;
3176 for (pfn = start_pfn; pfn < end_pfn; pfn++)
3177 if (PageReserved(pfn_to_page(pfn)))
3183 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3184 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3185 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3186 * higher will lead to a bigger reserve which will get freed as contiguous
3187 * blocks as reclaim kicks in
3189 static void setup_zone_migrate_reserve(struct zone *zone)
3191 unsigned long start_pfn, pfn, end_pfn;
3193 unsigned long block_migratetype;
3196 /* Get the start pfn, end pfn and the number of blocks to reserve */
3197 start_pfn = zone->zone_start_pfn;
3198 end_pfn = start_pfn + zone->spanned_pages;
3199 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3203 * Reserve blocks are generally in place to help high-order atomic
3204 * allocations that are short-lived. A min_free_kbytes value that
3205 * would result in more than 2 reserve blocks for atomic allocations
3206 * is assumed to be in place to help anti-fragmentation for the
3207 * future allocation of hugepages at runtime.
3209 reserve = min(2, reserve);
3211 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3212 if (!pfn_valid(pfn))
3214 page = pfn_to_page(pfn);
3216 /* Watch out for overlapping nodes */
3217 if (page_to_nid(page) != zone_to_nid(zone))
3220 /* Blocks with reserved pages will never free, skip them. */
3221 if (pageblock_is_reserved(pfn))
3224 block_migratetype = get_pageblock_migratetype(page);
3226 /* If this block is reserved, account for it */
3227 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3232 /* Suitable for reserving if this block is movable */
3233 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3234 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3235 move_freepages_block(zone, page, MIGRATE_RESERVE);
3241 * If the reserve is met and this is a previous reserved block,
3244 if (block_migratetype == MIGRATE_RESERVE) {
3245 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3246 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3252 * Initially all pages are reserved - free ones are freed
3253 * up by free_all_bootmem() once the early boot process is
3254 * done. Non-atomic initialization, single-pass.
3256 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3257 unsigned long start_pfn, enum memmap_context context)
3260 unsigned long end_pfn = start_pfn + size;
3264 if (highest_memmap_pfn < end_pfn - 1)
3265 highest_memmap_pfn = end_pfn - 1;
3267 z = &NODE_DATA(nid)->node_zones[zone];
3268 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3270 * There can be holes in boot-time mem_map[]s
3271 * handed to this function. They do not
3272 * exist on hotplugged memory.
3274 if (context == MEMMAP_EARLY) {
3275 if (!early_pfn_valid(pfn))
3277 if (!early_pfn_in_nid(pfn, nid))
3280 page = pfn_to_page(pfn);
3281 set_page_links(page, zone, nid, pfn);
3282 mminit_verify_page_links(page, zone, nid, pfn);
3283 init_page_count(page);
3284 reset_page_mapcount(page);
3285 SetPageReserved(page);
3287 * Mark the block movable so that blocks are reserved for
3288 * movable at startup. This will force kernel allocations
3289 * to reserve their blocks rather than leaking throughout
3290 * the address space during boot when many long-lived
3291 * kernel allocations are made. Later some blocks near
3292 * the start are marked MIGRATE_RESERVE by
3293 * setup_zone_migrate_reserve()
3295 * bitmap is created for zone's valid pfn range. but memmap
3296 * can be created for invalid pages (for alignment)
3297 * check here not to call set_pageblock_migratetype() against
3300 if ((z->zone_start_pfn <= pfn)
3301 && (pfn < z->zone_start_pfn + z->spanned_pages)
3302 && !(pfn & (pageblock_nr_pages - 1)))
3303 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3305 INIT_LIST_HEAD(&page->lru);
3306 #ifdef WANT_PAGE_VIRTUAL
3307 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3308 if (!is_highmem_idx(zone))
3309 set_page_address(page, __va(pfn << PAGE_SHIFT));
3314 static void __meminit zone_init_free_lists(struct zone *zone)
3317 for_each_migratetype_order(order, t) {
3318 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3319 zone->free_area[order].nr_free = 0;
3323 #ifndef __HAVE_ARCH_MEMMAP_INIT
3324 #define memmap_init(size, nid, zone, start_pfn) \
3325 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3328 static int zone_batchsize(struct zone *zone)
3334 * The per-cpu-pages pools are set to around 1000th of the
3335 * size of the zone. But no more than 1/2 of a meg.
3337 * OK, so we don't know how big the cache is. So guess.
3339 batch = zone->present_pages / 1024;
3340 if (batch * PAGE_SIZE > 512 * 1024)
3341 batch = (512 * 1024) / PAGE_SIZE;
3342 batch /= 4; /* We effectively *= 4 below */
3347 * Clamp the batch to a 2^n - 1 value. Having a power
3348 * of 2 value was found to be more likely to have
3349 * suboptimal cache aliasing properties in some cases.
3351 * For example if 2 tasks are alternately allocating
3352 * batches of pages, one task can end up with a lot
3353 * of pages of one half of the possible page colors
3354 * and the other with pages of the other colors.
3356 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3361 /* The deferral and batching of frees should be suppressed under NOMMU
3364 * The problem is that NOMMU needs to be able to allocate large chunks
3365 * of contiguous memory as there's no hardware page translation to
3366 * assemble apparent contiguous memory from discontiguous pages.
3368 * Queueing large contiguous runs of pages for batching, however,
3369 * causes the pages to actually be freed in smaller chunks. As there
3370 * can be a significant delay between the individual batches being
3371 * recycled, this leads to the once large chunks of space being
3372 * fragmented and becoming unavailable for high-order allocations.
3378 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3380 struct per_cpu_pages *pcp;
3383 memset(p, 0, sizeof(*p));
3387 pcp->high = 6 * batch;
3388 pcp->batch = max(1UL, 1 * batch);
3389 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3390 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3394 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3395 * to the value high for the pageset p.
3398 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3401 struct per_cpu_pages *pcp;
3405 pcp->batch = max(1UL, high/4);
3406 if ((high/4) > (PAGE_SHIFT * 8))
3407 pcp->batch = PAGE_SHIFT * 8;
3410 static __meminit void setup_zone_pageset(struct zone *zone)
3414 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3416 for_each_possible_cpu(cpu) {
3417 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3419 setup_pageset(pcp, zone_batchsize(zone));
3421 if (percpu_pagelist_fraction)
3422 setup_pagelist_highmark(pcp,
3423 (zone->present_pages /
3424 percpu_pagelist_fraction));
3429 * Allocate per cpu pagesets and initialize them.
3430 * Before this call only boot pagesets were available.
3432 void __init setup_per_cpu_pageset(void)
3436 for_each_populated_zone(zone)
3437 setup_zone_pageset(zone);
3440 static noinline __init_refok
3441 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3444 struct pglist_data *pgdat = zone->zone_pgdat;
3448 * The per-page waitqueue mechanism uses hashed waitqueues
3451 zone->wait_table_hash_nr_entries =
3452 wait_table_hash_nr_entries(zone_size_pages);
3453 zone->wait_table_bits =
3454 wait_table_bits(zone->wait_table_hash_nr_entries);
3455 alloc_size = zone->wait_table_hash_nr_entries
3456 * sizeof(wait_queue_head_t);
3458 if (!slab_is_available()) {
3459 zone->wait_table = (wait_queue_head_t *)
3460 alloc_bootmem_node(pgdat, alloc_size);
3463 * This case means that a zone whose size was 0 gets new memory
3464 * via memory hot-add.
3465 * But it may be the case that a new node was hot-added. In
3466 * this case vmalloc() will not be able to use this new node's
3467 * memory - this wait_table must be initialized to use this new
3468 * node itself as well.
3469 * To use this new node's memory, further consideration will be
3472 zone->wait_table = vmalloc(alloc_size);
3474 if (!zone->wait_table)
3477 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3478 init_waitqueue_head(zone->wait_table + i);
3483 static int __zone_pcp_update(void *data)
3485 struct zone *zone = data;
3487 unsigned long batch = zone_batchsize(zone), flags;
3489 for_each_possible_cpu(cpu) {
3490 struct per_cpu_pageset *pset;
3491 struct per_cpu_pages *pcp;
3493 pset = per_cpu_ptr(zone->pageset, cpu);
3496 local_irq_save(flags);
3497 free_pcppages_bulk(zone, pcp->count, pcp);
3498 setup_pageset(pset, batch);
3499 local_irq_restore(flags);
3504 void zone_pcp_update(struct zone *zone)
3506 stop_machine(__zone_pcp_update, zone, NULL);
3509 static __meminit void zone_pcp_init(struct zone *zone)
3512 * per cpu subsystem is not up at this point. The following code
3513 * relies on the ability of the linker to provide the
3514 * offset of a (static) per cpu variable into the per cpu area.
3516 zone->pageset = &boot_pageset;
3518 if (zone->present_pages)
3519 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3520 zone->name, zone->present_pages,
3521 zone_batchsize(zone));
3524 __meminit int init_currently_empty_zone(struct zone *zone,
3525 unsigned long zone_start_pfn,
3527 enum memmap_context context)
3529 struct pglist_data *pgdat = zone->zone_pgdat;
3531 ret = zone_wait_table_init(zone, size);
3534 pgdat->nr_zones = zone_idx(zone) + 1;
3536 zone->zone_start_pfn = zone_start_pfn;
3538 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3539 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3541 (unsigned long)zone_idx(zone),
3542 zone_start_pfn, (zone_start_pfn + size));
3544 zone_init_free_lists(zone);
3549 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3551 * Basic iterator support. Return the first range of PFNs for a node
3552 * Note: nid == MAX_NUMNODES returns first region regardless of node
3554 static int __meminit first_active_region_index_in_nid(int nid)
3558 for (i = 0; i < nr_nodemap_entries; i++)
3559 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3566 * Basic iterator support. Return the next active range of PFNs for a node
3567 * Note: nid == MAX_NUMNODES returns next region regardless of node
3569 static int __meminit next_active_region_index_in_nid(int index, int nid)
3571 for (index = index + 1; index < nr_nodemap_entries; index++)
3572 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3578 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3580 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3581 * Architectures may implement their own version but if add_active_range()
3582 * was used and there are no special requirements, this is a convenient
3585 int __meminit __early_pfn_to_nid(unsigned long pfn)
3589 for (i = 0; i < nr_nodemap_entries; i++) {
3590 unsigned long start_pfn = early_node_map[i].start_pfn;
3591 unsigned long end_pfn = early_node_map[i].end_pfn;
3593 if (start_pfn <= pfn && pfn < end_pfn)
3594 return early_node_map[i].nid;
3596 /* This is a memory hole */
3599 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3601 int __meminit early_pfn_to_nid(unsigned long pfn)
3605 nid = __early_pfn_to_nid(pfn);
3608 /* just returns 0 */
3612 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3613 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3617 nid = __early_pfn_to_nid(pfn);
3618 if (nid >= 0 && nid != node)
3624 /* Basic iterator support to walk early_node_map[] */
3625 #define for_each_active_range_index_in_nid(i, nid) \
3626 for (i = first_active_region_index_in_nid(nid); i != -1; \
3627 i = next_active_region_index_in_nid(i, nid))
3630 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3631 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3632 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3634 * If an architecture guarantees that all ranges registered with
3635 * add_active_ranges() contain no holes and may be freed, this
3636 * this function may be used instead of calling free_bootmem() manually.
3638 void __init free_bootmem_with_active_regions(int nid,
3639 unsigned long max_low_pfn)
3643 for_each_active_range_index_in_nid(i, nid) {
3644 unsigned long size_pages = 0;
3645 unsigned long end_pfn = early_node_map[i].end_pfn;
3647 if (early_node_map[i].start_pfn >= max_low_pfn)
3650 if (end_pfn > max_low_pfn)
3651 end_pfn = max_low_pfn;
3653 size_pages = end_pfn - early_node_map[i].start_pfn;
3654 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3655 PFN_PHYS(early_node_map[i].start_pfn),
3656 size_pages << PAGE_SHIFT);
3660 int __init add_from_early_node_map(struct range *range, int az,
3661 int nr_range, int nid)
3666 /* need to go over early_node_map to find out good range for node */
3667 for_each_active_range_index_in_nid(i, nid) {
3668 start = early_node_map[i].start_pfn;
3669 end = early_node_map[i].end_pfn;
3670 nr_range = add_range(range, az, nr_range, start, end);
3675 #ifdef CONFIG_NO_BOOTMEM
3676 void * __init __alloc_memory_core_early(int nid, u64 size, u64 align,
3677 u64 goal, u64 limit)
3682 if (limit > get_max_mapped())
3683 limit = get_max_mapped();
3685 /* need to go over early_node_map to find out good range for node */
3686 for_each_active_range_index_in_nid(i, nid) {
3688 u64 ei_start, ei_last;
3690 ei_last = early_node_map[i].end_pfn;
3691 ei_last <<= PAGE_SHIFT;
3692 ei_start = early_node_map[i].start_pfn;
3693 ei_start <<= PAGE_SHIFT;
3694 addr = find_early_area(ei_start, ei_last,
3695 goal, limit, size, align);
3701 printk(KERN_DEBUG "alloc (nid=%d %llx - %llx) (%llx - %llx) %llx %llx => %llx\n",
3703 ei_start, ei_last, goal, limit, size,
3707 ptr = phys_to_virt(addr);
3708 memset(ptr, 0, size);
3709 reserve_early_without_check(addr, addr + size, "BOOTMEM");
3711 * The min_count is set to 0 so that bootmem allocated blocks
3712 * are never reported as leaks.
3714 kmemleak_alloc(ptr, size, 0, 0);
3723 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3728 for_each_active_range_index_in_nid(i, nid) {
3729 ret = work_fn(early_node_map[i].start_pfn,
3730 early_node_map[i].end_pfn, data);
3736 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3737 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3739 * If an architecture guarantees that all ranges registered with
3740 * add_active_ranges() contain no holes and may be freed, this
3741 * function may be used instead of calling memory_present() manually.
3743 void __init sparse_memory_present_with_active_regions(int nid)
3747 for_each_active_range_index_in_nid(i, nid)
3748 memory_present(early_node_map[i].nid,
3749 early_node_map[i].start_pfn,
3750 early_node_map[i].end_pfn);
3754 * get_pfn_range_for_nid - Return the start and end page frames for a node
3755 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3756 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3757 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3759 * It returns the start and end page frame of a node based on information
3760 * provided by an arch calling add_active_range(). If called for a node
3761 * with no available memory, a warning is printed and the start and end
3764 void __meminit get_pfn_range_for_nid(unsigned int nid,
3765 unsigned long *start_pfn, unsigned long *end_pfn)
3771 for_each_active_range_index_in_nid(i, nid) {
3772 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3773 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3776 if (*start_pfn == -1UL)
3781 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3782 * assumption is made that zones within a node are ordered in monotonic
3783 * increasing memory addresses so that the "highest" populated zone is used
3785 static void __init find_usable_zone_for_movable(void)
3788 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3789 if (zone_index == ZONE_MOVABLE)
3792 if (arch_zone_highest_possible_pfn[zone_index] >
3793 arch_zone_lowest_possible_pfn[zone_index])
3797 VM_BUG_ON(zone_index == -1);
3798 movable_zone = zone_index;
3802 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3803 * because it is sized independant of architecture. Unlike the other zones,
3804 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3805 * in each node depending on the size of each node and how evenly kernelcore
3806 * is distributed. This helper function adjusts the zone ranges
3807 * provided by the architecture for a given node by using the end of the
3808 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3809 * zones within a node are in order of monotonic increases memory addresses
3811 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3812 unsigned long zone_type,
3813 unsigned long node_start_pfn,
3814 unsigned long node_end_pfn,
3815 unsigned long *zone_start_pfn,
3816 unsigned long *zone_end_pfn)
3818 /* Only adjust if ZONE_MOVABLE is on this node */
3819 if (zone_movable_pfn[nid]) {
3820 /* Size ZONE_MOVABLE */
3821 if (zone_type == ZONE_MOVABLE) {
3822 *zone_start_pfn = zone_movable_pfn[nid];
3823 *zone_end_pfn = min(node_end_pfn,
3824 arch_zone_highest_possible_pfn[movable_zone]);
3826 /* Adjust for ZONE_MOVABLE starting within this range */
3827 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3828 *zone_end_pfn > zone_movable_pfn[nid]) {
3829 *zone_end_pfn = zone_movable_pfn[nid];
3831 /* Check if this whole range is within ZONE_MOVABLE */
3832 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3833 *zone_start_pfn = *zone_end_pfn;
3838 * Return the number of pages a zone spans in a node, including holes
3839 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3841 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3842 unsigned long zone_type,
3843 unsigned long *ignored)
3845 unsigned long node_start_pfn, node_end_pfn;
3846 unsigned long zone_start_pfn, zone_end_pfn;
3848 /* Get the start and end of the node and zone */
3849 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3850 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3851 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3852 adjust_zone_range_for_zone_movable(nid, zone_type,
3853 node_start_pfn, node_end_pfn,
3854 &zone_start_pfn, &zone_end_pfn);
3856 /* Check that this node has pages within the zone's required range */
3857 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3860 /* Move the zone boundaries inside the node if necessary */
3861 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3862 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3864 /* Return the spanned pages */
3865 return zone_end_pfn - zone_start_pfn;
3869 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3870 * then all holes in the requested range will be accounted for.
3872 unsigned long __meminit __absent_pages_in_range(int nid,
3873 unsigned long range_start_pfn,
3874 unsigned long range_end_pfn)
3877 unsigned long prev_end_pfn = 0, hole_pages = 0;
3878 unsigned long start_pfn;
3880 /* Find the end_pfn of the first active range of pfns in the node */
3881 i = first_active_region_index_in_nid(nid);
3885 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3887 /* Account for ranges before physical memory on this node */
3888 if (early_node_map[i].start_pfn > range_start_pfn)
3889 hole_pages = prev_end_pfn - range_start_pfn;
3891 /* Find all holes for the zone within the node */
3892 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3894 /* No need to continue if prev_end_pfn is outside the zone */
3895 if (prev_end_pfn >= range_end_pfn)
3898 /* Make sure the end of the zone is not within the hole */
3899 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3900 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3902 /* Update the hole size cound and move on */
3903 if (start_pfn > range_start_pfn) {
3904 BUG_ON(prev_end_pfn > start_pfn);
3905 hole_pages += start_pfn - prev_end_pfn;
3907 prev_end_pfn = early_node_map[i].end_pfn;
3910 /* Account for ranges past physical memory on this node */
3911 if (range_end_pfn > prev_end_pfn)
3912 hole_pages += range_end_pfn -
3913 max(range_start_pfn, prev_end_pfn);
3919 * absent_pages_in_range - Return number of page frames in holes within a range
3920 * @start_pfn: The start PFN to start searching for holes
3921 * @end_pfn: The end PFN to stop searching for holes
3923 * It returns the number of pages frames in memory holes within a range.
3925 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3926 unsigned long end_pfn)
3928 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3931 /* Return the number of page frames in holes in a zone on a node */
3932 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3933 unsigned long zone_type,
3934 unsigned long *ignored)
3936 unsigned long node_start_pfn, node_end_pfn;
3937 unsigned long zone_start_pfn, zone_end_pfn;
3939 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3940 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3942 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3945 adjust_zone_range_for_zone_movable(nid, zone_type,
3946 node_start_pfn, node_end_pfn,
3947 &zone_start_pfn, &zone_end_pfn);
3948 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3952 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3953 unsigned long zone_type,
3954 unsigned long *zones_size)
3956 return zones_size[zone_type];
3959 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3960 unsigned long zone_type,
3961 unsigned long *zholes_size)
3966 return zholes_size[zone_type];
3971 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3972 unsigned long *zones_size, unsigned long *zholes_size)
3974 unsigned long realtotalpages, totalpages = 0;
3977 for (i = 0; i < MAX_NR_ZONES; i++)
3978 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3980 pgdat->node_spanned_pages = totalpages;
3982 realtotalpages = totalpages;
3983 for (i = 0; i < MAX_NR_ZONES; i++)
3985 zone_absent_pages_in_node(pgdat->node_id, i,
3987 pgdat->node_present_pages = realtotalpages;
3988 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3992 #ifndef CONFIG_SPARSEMEM
3994 * Calculate the size of the zone->blockflags rounded to an unsigned long
3995 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3996 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3997 * round what is now in bits to nearest long in bits, then return it in
4000 static unsigned long __init usemap_size(unsigned long zonesize)
4002 unsigned long usemapsize;
4004 usemapsize = roundup(zonesize, pageblock_nr_pages);
4005 usemapsize = usemapsize >> pageblock_order;
4006 usemapsize *= NR_PAGEBLOCK_BITS;
4007 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4009 return usemapsize / 8;
4012 static void __init setup_usemap(struct pglist_data *pgdat,
4013 struct zone *zone, unsigned long zonesize)
4015 unsigned long usemapsize = usemap_size(zonesize);
4016 zone->pageblock_flags = NULL;
4018 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
4021 static void inline setup_usemap(struct pglist_data *pgdat,
4022 struct zone *zone, unsigned long zonesize) {}
4023 #endif /* CONFIG_SPARSEMEM */
4025 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4027 /* Return a sensible default order for the pageblock size. */
4028 static inline int pageblock_default_order(void)
4030 if (HPAGE_SHIFT > PAGE_SHIFT)
4031 return HUGETLB_PAGE_ORDER;
4036 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4037 static inline void __init set_pageblock_order(unsigned int order)
4039 /* Check that pageblock_nr_pages has not already been setup */
4040 if (pageblock_order)
4044 * Assume the largest contiguous order of interest is a huge page.
4045 * This value may be variable depending on boot parameters on IA64
4047 pageblock_order = order;
4049 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4052 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4053 * and pageblock_default_order() are unused as pageblock_order is set
4054 * at compile-time. See include/linux/pageblock-flags.h for the values of
4055 * pageblock_order based on the kernel config
4057 static inline int pageblock_default_order(unsigned int order)
4061 #define set_pageblock_order(x) do {} while (0)
4063 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4066 * Set up the zone data structures:
4067 * - mark all pages reserved
4068 * - mark all memory queues empty
4069 * - clear the memory bitmaps
4071 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4072 unsigned long *zones_size, unsigned long *zholes_size)
4075 int nid = pgdat->node_id;
4076 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4079 pgdat_resize_init(pgdat);
4080 pgdat->nr_zones = 0;
4081 init_waitqueue_head(&pgdat->kswapd_wait);
4082 pgdat->kswapd_max_order = 0;
4083 pgdat_page_cgroup_init(pgdat);
4085 for (j = 0; j < MAX_NR_ZONES; j++) {
4086 struct zone *zone = pgdat->node_zones + j;
4087 unsigned long size, realsize, memmap_pages;
4090 size = zone_spanned_pages_in_node(nid, j, zones_size);
4091 realsize = size - zone_absent_pages_in_node(nid, j,
4095 * Adjust realsize so that it accounts for how much memory
4096 * is used by this zone for memmap. This affects the watermark
4097 * and per-cpu initialisations
4100 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4101 if (realsize >= memmap_pages) {
4102 realsize -= memmap_pages;
4105 " %s zone: %lu pages used for memmap\n",
4106 zone_names[j], memmap_pages);
4109 " %s zone: %lu pages exceeds realsize %lu\n",
4110 zone_names[j], memmap_pages, realsize);
4112 /* Account for reserved pages */
4113 if (j == 0 && realsize > dma_reserve) {
4114 realsize -= dma_reserve;
4115 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4116 zone_names[0], dma_reserve);
4119 if (!is_highmem_idx(j))
4120 nr_kernel_pages += realsize;
4121 nr_all_pages += realsize;
4123 zone->spanned_pages = size;
4124 zone->present_pages = realsize;
4127 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4129 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4131 zone->name = zone_names[j];
4132 spin_lock_init(&zone->lock);
4133 spin_lock_init(&zone->lru_lock);
4134 zone_seqlock_init(zone);
4135 zone->zone_pgdat = pgdat;
4137 zone_pcp_init(zone);
4139 INIT_LIST_HEAD(&zone->lru[l].list);
4140 zone->reclaim_stat.nr_saved_scan[l] = 0;
4142 zone->reclaim_stat.recent_rotated[0] = 0;
4143 zone->reclaim_stat.recent_rotated[1] = 0;
4144 zone->reclaim_stat.recent_scanned[0] = 0;
4145 zone->reclaim_stat.recent_scanned[1] = 0;
4146 zap_zone_vm_stats(zone);
4151 set_pageblock_order(pageblock_default_order());
4152 setup_usemap(pgdat, zone, size);
4153 ret = init_currently_empty_zone(zone, zone_start_pfn,
4154 size, MEMMAP_EARLY);
4156 memmap_init(size, nid, j, zone_start_pfn);
4157 zone_start_pfn += size;
4161 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4163 /* Skip empty nodes */
4164 if (!pgdat->node_spanned_pages)
4167 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4168 /* ia64 gets its own node_mem_map, before this, without bootmem */
4169 if (!pgdat->node_mem_map) {
4170 unsigned long size, start, end;
4174 * The zone's endpoints aren't required to be MAX_ORDER
4175 * aligned but the node_mem_map endpoints must be in order
4176 * for the buddy allocator to function correctly.
4178 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4179 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4180 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4181 size = (end - start) * sizeof(struct page);
4182 map = alloc_remap(pgdat->node_id, size);
4184 map = alloc_bootmem_node(pgdat, size);
4185 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4187 #ifndef CONFIG_NEED_MULTIPLE_NODES
4189 * With no DISCONTIG, the global mem_map is just set as node 0's
4191 if (pgdat == NODE_DATA(0)) {
4192 mem_map = NODE_DATA(0)->node_mem_map;
4193 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4194 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4195 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4196 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4199 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4202 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4203 unsigned long node_start_pfn, unsigned long *zholes_size)
4205 pg_data_t *pgdat = NODE_DATA(nid);
4207 pgdat->node_id = nid;
4208 pgdat->node_start_pfn = node_start_pfn;
4209 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4211 alloc_node_mem_map(pgdat);
4212 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4213 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4214 nid, (unsigned long)pgdat,
4215 (unsigned long)pgdat->node_mem_map);
4218 free_area_init_core(pgdat, zones_size, zholes_size);
4221 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4223 #if MAX_NUMNODES > 1
4225 * Figure out the number of possible node ids.
4227 static void __init setup_nr_node_ids(void)
4230 unsigned int highest = 0;
4232 for_each_node_mask(node, node_possible_map)
4234 nr_node_ids = highest + 1;
4237 static inline void setup_nr_node_ids(void)
4243 * add_active_range - Register a range of PFNs backed by physical memory
4244 * @nid: The node ID the range resides on
4245 * @start_pfn: The start PFN of the available physical memory
4246 * @end_pfn: The end PFN of the available physical memory
4248 * These ranges are stored in an early_node_map[] and later used by
4249 * free_area_init_nodes() to calculate zone sizes and holes. If the
4250 * range spans a memory hole, it is up to the architecture to ensure
4251 * the memory is not freed by the bootmem allocator. If possible
4252 * the range being registered will be merged with existing ranges.
4254 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4255 unsigned long end_pfn)
4259 mminit_dprintk(MMINIT_TRACE, "memory_register",
4260 "Entering add_active_range(%d, %#lx, %#lx) "
4261 "%d entries of %d used\n",
4262 nid, start_pfn, end_pfn,
4263 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4265 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4267 /* Merge with existing active regions if possible */
4268 for (i = 0; i < nr_nodemap_entries; i++) {
4269 if (early_node_map[i].nid != nid)
4272 /* Skip if an existing region covers this new one */
4273 if (start_pfn >= early_node_map[i].start_pfn &&
4274 end_pfn <= early_node_map[i].end_pfn)
4277 /* Merge forward if suitable */
4278 if (start_pfn <= early_node_map[i].end_pfn &&
4279 end_pfn > early_node_map[i].end_pfn) {
4280 early_node_map[i].end_pfn = end_pfn;
4284 /* Merge backward if suitable */
4285 if (start_pfn < early_node_map[i].start_pfn &&
4286 end_pfn >= early_node_map[i].start_pfn) {
4287 early_node_map[i].start_pfn = start_pfn;
4292 /* Check that early_node_map is large enough */
4293 if (i >= MAX_ACTIVE_REGIONS) {
4294 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4295 MAX_ACTIVE_REGIONS);
4299 early_node_map[i].nid = nid;
4300 early_node_map[i].start_pfn = start_pfn;
4301 early_node_map[i].end_pfn = end_pfn;
4302 nr_nodemap_entries = i + 1;
4306 * remove_active_range - Shrink an existing registered range of PFNs
4307 * @nid: The node id the range is on that should be shrunk
4308 * @start_pfn: The new PFN of the range
4309 * @end_pfn: The new PFN of the range
4311 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4312 * The map is kept near the end physical page range that has already been
4313 * registered. This function allows an arch to shrink an existing registered
4316 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4317 unsigned long end_pfn)
4322 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4323 nid, start_pfn, end_pfn);
4325 /* Find the old active region end and shrink */
4326 for_each_active_range_index_in_nid(i, nid) {
4327 if (early_node_map[i].start_pfn >= start_pfn &&
4328 early_node_map[i].end_pfn <= end_pfn) {
4330 early_node_map[i].start_pfn = 0;
4331 early_node_map[i].end_pfn = 0;
4335 if (early_node_map[i].start_pfn < start_pfn &&
4336 early_node_map[i].end_pfn > start_pfn) {
4337 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4338 early_node_map[i].end_pfn = start_pfn;
4339 if (temp_end_pfn > end_pfn)
4340 add_active_range(nid, end_pfn, temp_end_pfn);
4343 if (early_node_map[i].start_pfn >= start_pfn &&
4344 early_node_map[i].end_pfn > end_pfn &&
4345 early_node_map[i].start_pfn < end_pfn) {
4346 early_node_map[i].start_pfn = end_pfn;
4354 /* remove the blank ones */
4355 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4356 if (early_node_map[i].nid != nid)
4358 if (early_node_map[i].end_pfn)
4360 /* we found it, get rid of it */
4361 for (j = i; j < nr_nodemap_entries - 1; j++)
4362 memcpy(&early_node_map[j], &early_node_map[j+1],
4363 sizeof(early_node_map[j]));
4364 j = nr_nodemap_entries - 1;
4365 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4366 nr_nodemap_entries--;
4371 * remove_all_active_ranges - Remove all currently registered regions
4373 * During discovery, it may be found that a table like SRAT is invalid
4374 * and an alternative discovery method must be used. This function removes
4375 * all currently registered regions.
4377 void __init remove_all_active_ranges(void)
4379 memset(early_node_map, 0, sizeof(early_node_map));
4380 nr_nodemap_entries = 0;
4383 /* Compare two active node_active_regions */
4384 static int __init cmp_node_active_region(const void *a, const void *b)
4386 struct node_active_region *arange = (struct node_active_region *)a;
4387 struct node_active_region *brange = (struct node_active_region *)b;
4389 /* Done this way to avoid overflows */
4390 if (arange->start_pfn > brange->start_pfn)
4392 if (arange->start_pfn < brange->start_pfn)
4398 /* sort the node_map by start_pfn */
4399 void __init sort_node_map(void)
4401 sort(early_node_map, (size_t)nr_nodemap_entries,
4402 sizeof(struct node_active_region),
4403 cmp_node_active_region, NULL);
4406 /* Find the lowest pfn for a node */
4407 static unsigned long __init find_min_pfn_for_node(int nid)
4410 unsigned long min_pfn = ULONG_MAX;
4412 /* Assuming a sorted map, the first range found has the starting pfn */
4413 for_each_active_range_index_in_nid(i, nid)
4414 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4416 if (min_pfn == ULONG_MAX) {
4418 "Could not find start_pfn for node %d\n", nid);
4426 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4428 * It returns the minimum PFN based on information provided via
4429 * add_active_range().
4431 unsigned long __init find_min_pfn_with_active_regions(void)
4433 return find_min_pfn_for_node(MAX_NUMNODES);
4437 * early_calculate_totalpages()
4438 * Sum pages in active regions for movable zone.
4439 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4441 static unsigned long __init early_calculate_totalpages(void)
4444 unsigned long totalpages = 0;
4446 for (i = 0; i < nr_nodemap_entries; i++) {
4447 unsigned long pages = early_node_map[i].end_pfn -
4448 early_node_map[i].start_pfn;
4449 totalpages += pages;
4451 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4457 * Find the PFN the Movable zone begins in each node. Kernel memory
4458 * is spread evenly between nodes as long as the nodes have enough
4459 * memory. When they don't, some nodes will have more kernelcore than
4462 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4465 unsigned long usable_startpfn;
4466 unsigned long kernelcore_node, kernelcore_remaining;
4467 /* save the state before borrow the nodemask */
4468 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4469 unsigned long totalpages = early_calculate_totalpages();
4470 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4473 * If movablecore was specified, calculate what size of
4474 * kernelcore that corresponds so that memory usable for
4475 * any allocation type is evenly spread. If both kernelcore
4476 * and movablecore are specified, then the value of kernelcore
4477 * will be used for required_kernelcore if it's greater than
4478 * what movablecore would have allowed.
4480 if (required_movablecore) {
4481 unsigned long corepages;
4484 * Round-up so that ZONE_MOVABLE is at least as large as what
4485 * was requested by the user
4487 required_movablecore =
4488 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4489 corepages = totalpages - required_movablecore;
4491 required_kernelcore = max(required_kernelcore, corepages);
4494 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4495 if (!required_kernelcore)
4498 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4499 find_usable_zone_for_movable();
4500 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4503 /* Spread kernelcore memory as evenly as possible throughout nodes */
4504 kernelcore_node = required_kernelcore / usable_nodes;
4505 for_each_node_state(nid, N_HIGH_MEMORY) {
4507 * Recalculate kernelcore_node if the division per node
4508 * now exceeds what is necessary to satisfy the requested
4509 * amount of memory for the kernel
4511 if (required_kernelcore < kernelcore_node)
4512 kernelcore_node = required_kernelcore / usable_nodes;
4515 * As the map is walked, we track how much memory is usable
4516 * by the kernel using kernelcore_remaining. When it is
4517 * 0, the rest of the node is usable by ZONE_MOVABLE
4519 kernelcore_remaining = kernelcore_node;
4521 /* Go through each range of PFNs within this node */
4522 for_each_active_range_index_in_nid(i, nid) {
4523 unsigned long start_pfn, end_pfn;
4524 unsigned long size_pages;
4526 start_pfn = max(early_node_map[i].start_pfn,
4527 zone_movable_pfn[nid]);
4528 end_pfn = early_node_map[i].end_pfn;
4529 if (start_pfn >= end_pfn)
4532 /* Account for what is only usable for kernelcore */
4533 if (start_pfn < usable_startpfn) {
4534 unsigned long kernel_pages;
4535 kernel_pages = min(end_pfn, usable_startpfn)
4538 kernelcore_remaining -= min(kernel_pages,
4539 kernelcore_remaining);
4540 required_kernelcore -= min(kernel_pages,
4541 required_kernelcore);
4543 /* Continue if range is now fully accounted */
4544 if (end_pfn <= usable_startpfn) {
4547 * Push zone_movable_pfn to the end so
4548 * that if we have to rebalance
4549 * kernelcore across nodes, we will
4550 * not double account here
4552 zone_movable_pfn[nid] = end_pfn;
4555 start_pfn = usable_startpfn;
4559 * The usable PFN range for ZONE_MOVABLE is from
4560 * start_pfn->end_pfn. Calculate size_pages as the
4561 * number of pages used as kernelcore
4563 size_pages = end_pfn - start_pfn;
4564 if (size_pages > kernelcore_remaining)
4565 size_pages = kernelcore_remaining;
4566 zone_movable_pfn[nid] = start_pfn + size_pages;
4569 * Some kernelcore has been met, update counts and
4570 * break if the kernelcore for this node has been
4573 required_kernelcore -= min(required_kernelcore,
4575 kernelcore_remaining -= size_pages;
4576 if (!kernelcore_remaining)
4582 * If there is still required_kernelcore, we do another pass with one
4583 * less node in the count. This will push zone_movable_pfn[nid] further
4584 * along on the nodes that still have memory until kernelcore is
4588 if (usable_nodes && required_kernelcore > usable_nodes)
4591 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4592 for (nid = 0; nid < MAX_NUMNODES; nid++)
4593 zone_movable_pfn[nid] =
4594 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4597 /* restore the node_state */
4598 node_states[N_HIGH_MEMORY] = saved_node_state;
4601 /* Any regular memory on that node ? */
4602 static void check_for_regular_memory(pg_data_t *pgdat)
4604 #ifdef CONFIG_HIGHMEM
4605 enum zone_type zone_type;
4607 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4608 struct zone *zone = &pgdat->node_zones[zone_type];
4609 if (zone->present_pages)
4610 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4616 * free_area_init_nodes - Initialise all pg_data_t and zone data
4617 * @max_zone_pfn: an array of max PFNs for each zone
4619 * This will call free_area_init_node() for each active node in the system.
4620 * Using the page ranges provided by add_active_range(), the size of each
4621 * zone in each node and their holes is calculated. If the maximum PFN
4622 * between two adjacent zones match, it is assumed that the zone is empty.
4623 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4624 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4625 * starts where the previous one ended. For example, ZONE_DMA32 starts
4626 * at arch_max_dma_pfn.
4628 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4633 /* Sort early_node_map as initialisation assumes it is sorted */
4636 /* Record where the zone boundaries are */
4637 memset(arch_zone_lowest_possible_pfn, 0,
4638 sizeof(arch_zone_lowest_possible_pfn));
4639 memset(arch_zone_highest_possible_pfn, 0,
4640 sizeof(arch_zone_highest_possible_pfn));
4641 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4642 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4643 for (i = 1; i < MAX_NR_ZONES; i++) {
4644 if (i == ZONE_MOVABLE)
4646 arch_zone_lowest_possible_pfn[i] =
4647 arch_zone_highest_possible_pfn[i-1];
4648 arch_zone_highest_possible_pfn[i] =
4649 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4651 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4652 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4654 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4655 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4656 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4658 /* Print out the zone ranges */
4659 printk("Zone PFN ranges:\n");
4660 for (i = 0; i < MAX_NR_ZONES; i++) {
4661 if (i == ZONE_MOVABLE)
4663 printk(" %-8s ", zone_names[i]);
4664 if (arch_zone_lowest_possible_pfn[i] ==
4665 arch_zone_highest_possible_pfn[i])
4668 printk("%0#10lx -> %0#10lx\n",
4669 arch_zone_lowest_possible_pfn[i],
4670 arch_zone_highest_possible_pfn[i]);
4673 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4674 printk("Movable zone start PFN for each node\n");
4675 for (i = 0; i < MAX_NUMNODES; i++) {
4676 if (zone_movable_pfn[i])
4677 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4680 /* Print out the early_node_map[] */
4681 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4682 for (i = 0; i < nr_nodemap_entries; i++)
4683 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4684 early_node_map[i].start_pfn,
4685 early_node_map[i].end_pfn);
4687 /* Initialise every node */
4688 mminit_verify_pageflags_layout();
4689 setup_nr_node_ids();
4690 for_each_online_node(nid) {
4691 pg_data_t *pgdat = NODE_DATA(nid);
4692 free_area_init_node(nid, NULL,
4693 find_min_pfn_for_node(nid), NULL);
4695 /* Any memory on that node */
4696 if (pgdat->node_present_pages)
4697 node_set_state(nid, N_HIGH_MEMORY);
4698 check_for_regular_memory(pgdat);
4702 static int __init cmdline_parse_core(char *p, unsigned long *core)
4704 unsigned long long coremem;
4708 coremem = memparse(p, &p);
4709 *core = coremem >> PAGE_SHIFT;
4711 /* Paranoid check that UL is enough for the coremem value */
4712 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4718 * kernelcore=size sets the amount of memory for use for allocations that
4719 * cannot be reclaimed or migrated.
4721 static int __init cmdline_parse_kernelcore(char *p)
4723 return cmdline_parse_core(p, &required_kernelcore);
4727 * movablecore=size sets the amount of memory for use for allocations that
4728 * can be reclaimed or migrated.
4730 static int __init cmdline_parse_movablecore(char *p)
4732 return cmdline_parse_core(p, &required_movablecore);
4735 early_param("kernelcore", cmdline_parse_kernelcore);
4736 early_param("movablecore", cmdline_parse_movablecore);
4738 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4741 * set_dma_reserve - set the specified number of pages reserved in the first zone
4742 * @new_dma_reserve: The number of pages to mark reserved
4744 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4745 * In the DMA zone, a significant percentage may be consumed by kernel image
4746 * and other unfreeable allocations which can skew the watermarks badly. This
4747 * function may optionally be used to account for unfreeable pages in the
4748 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4749 * smaller per-cpu batchsize.
4751 void __init set_dma_reserve(unsigned long new_dma_reserve)
4753 dma_reserve = new_dma_reserve;
4756 #ifndef CONFIG_NEED_MULTIPLE_NODES
4757 struct pglist_data __refdata contig_page_data = {
4758 #ifndef CONFIG_NO_BOOTMEM
4759 .bdata = &bootmem_node_data[0]
4762 EXPORT_SYMBOL(contig_page_data);
4765 void __init free_area_init(unsigned long *zones_size)
4767 free_area_init_node(0, zones_size,
4768 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4771 static int page_alloc_cpu_notify(struct notifier_block *self,
4772 unsigned long action, void *hcpu)
4774 int cpu = (unsigned long)hcpu;
4776 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4780 * Spill the event counters of the dead processor
4781 * into the current processors event counters.
4782 * This artificially elevates the count of the current
4785 vm_events_fold_cpu(cpu);
4788 * Zero the differential counters of the dead processor
4789 * so that the vm statistics are consistent.
4791 * This is only okay since the processor is dead and cannot
4792 * race with what we are doing.
4794 refresh_cpu_vm_stats(cpu);
4799 void __init page_alloc_init(void)
4801 hotcpu_notifier(page_alloc_cpu_notify, 0);
4805 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4806 * or min_free_kbytes changes.
4808 static void calculate_totalreserve_pages(void)
4810 struct pglist_data *pgdat;
4811 unsigned long reserve_pages = 0;
4812 enum zone_type i, j;
4814 for_each_online_pgdat(pgdat) {
4815 for (i = 0; i < MAX_NR_ZONES; i++) {
4816 struct zone *zone = pgdat->node_zones + i;
4817 unsigned long max = 0;
4819 /* Find valid and maximum lowmem_reserve in the zone */
4820 for (j = i; j < MAX_NR_ZONES; j++) {
4821 if (zone->lowmem_reserve[j] > max)
4822 max = zone->lowmem_reserve[j];
4825 /* we treat the high watermark as reserved pages. */
4826 max += high_wmark_pages(zone);
4828 if (max > zone->present_pages)
4829 max = zone->present_pages;
4830 reserve_pages += max;
4833 totalreserve_pages = reserve_pages;
4837 * setup_per_zone_lowmem_reserve - called whenever
4838 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4839 * has a correct pages reserved value, so an adequate number of
4840 * pages are left in the zone after a successful __alloc_pages().
4842 static void setup_per_zone_lowmem_reserve(void)
4844 struct pglist_data *pgdat;
4845 enum zone_type j, idx;
4847 for_each_online_pgdat(pgdat) {
4848 for (j = 0; j < MAX_NR_ZONES; j++) {
4849 struct zone *zone = pgdat->node_zones + j;
4850 unsigned long present_pages = zone->present_pages;
4852 zone->lowmem_reserve[j] = 0;
4856 struct zone *lower_zone;
4860 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4861 sysctl_lowmem_reserve_ratio[idx] = 1;
4863 lower_zone = pgdat->node_zones + idx;
4864 lower_zone->lowmem_reserve[j] = present_pages /
4865 sysctl_lowmem_reserve_ratio[idx];
4866 present_pages += lower_zone->present_pages;
4871 /* update totalreserve_pages */
4872 calculate_totalreserve_pages();
4876 * setup_per_zone_wmarks - called when min_free_kbytes changes
4877 * or when memory is hot-{added|removed}
4879 * Ensures that the watermark[min,low,high] values for each zone are set
4880 * correctly with respect to min_free_kbytes.
4882 void setup_per_zone_wmarks(void)
4884 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4885 unsigned long lowmem_pages = 0;
4887 unsigned long flags;
4889 /* Calculate total number of !ZONE_HIGHMEM pages */
4890 for_each_zone(zone) {
4891 if (!is_highmem(zone))
4892 lowmem_pages += zone->present_pages;
4895 for_each_zone(zone) {
4898 spin_lock_irqsave(&zone->lock, flags);
4899 tmp = (u64)pages_min * zone->present_pages;
4900 do_div(tmp, lowmem_pages);
4901 if (is_highmem(zone)) {
4903 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4904 * need highmem pages, so cap pages_min to a small
4907 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4908 * deltas controls asynch page reclaim, and so should
4909 * not be capped for highmem.
4913 min_pages = zone->present_pages / 1024;
4914 if (min_pages < SWAP_CLUSTER_MAX)
4915 min_pages = SWAP_CLUSTER_MAX;
4916 if (min_pages > 128)
4918 zone->watermark[WMARK_MIN] = min_pages;
4921 * If it's a lowmem zone, reserve a number of pages
4922 * proportionate to the zone's size.
4924 zone->watermark[WMARK_MIN] = tmp;
4927 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4928 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4929 setup_zone_migrate_reserve(zone);
4930 spin_unlock_irqrestore(&zone->lock, flags);
4933 /* update totalreserve_pages */
4934 calculate_totalreserve_pages();
4938 * The inactive anon list should be small enough that the VM never has to
4939 * do too much work, but large enough that each inactive page has a chance
4940 * to be referenced again before it is swapped out.
4942 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4943 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4944 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4945 * the anonymous pages are kept on the inactive list.
4948 * memory ratio inactive anon
4949 * -------------------------------------
4958 void calculate_zone_inactive_ratio(struct zone *zone)
4960 unsigned int gb, ratio;
4962 /* Zone size in gigabytes */
4963 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4965 ratio = int_sqrt(10 * gb);
4969 zone->inactive_ratio = ratio;
4972 static void __init setup_per_zone_inactive_ratio(void)
4977 calculate_zone_inactive_ratio(zone);
4981 * Initialise min_free_kbytes.
4983 * For small machines we want it small (128k min). For large machines
4984 * we want it large (64MB max). But it is not linear, because network
4985 * bandwidth does not increase linearly with machine size. We use
4987 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4988 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5004 static int __init init_per_zone_wmark_min(void)
5006 unsigned long lowmem_kbytes;
5008 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5010 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5011 if (min_free_kbytes < 128)
5012 min_free_kbytes = 128;
5013 if (min_free_kbytes > 65536)
5014 min_free_kbytes = 65536;
5015 setup_per_zone_wmarks();
5016 setup_per_zone_lowmem_reserve();
5017 setup_per_zone_inactive_ratio();
5020 module_init(init_per_zone_wmark_min)
5023 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5024 * that we can call two helper functions whenever min_free_kbytes
5027 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5028 void __user *buffer, size_t *length, loff_t *ppos)
5030 proc_dointvec(table, write, buffer, length, ppos);
5032 setup_per_zone_wmarks();
5037 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5038 void __user *buffer, size_t *length, loff_t *ppos)
5043 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5048 zone->min_unmapped_pages = (zone->present_pages *
5049 sysctl_min_unmapped_ratio) / 100;
5053 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5054 void __user *buffer, size_t *length, loff_t *ppos)
5059 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5064 zone->min_slab_pages = (zone->present_pages *
5065 sysctl_min_slab_ratio) / 100;
5071 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5072 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5073 * whenever sysctl_lowmem_reserve_ratio changes.
5075 * The reserve ratio obviously has absolutely no relation with the
5076 * minimum watermarks. The lowmem reserve ratio can only make sense
5077 * if in function of the boot time zone sizes.
5079 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5080 void __user *buffer, size_t *length, loff_t *ppos)
5082 proc_dointvec_minmax(table, write, buffer, length, ppos);
5083 setup_per_zone_lowmem_reserve();
5088 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5089 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5090 * can have before it gets flushed back to buddy allocator.
5093 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5094 void __user *buffer, size_t *length, loff_t *ppos)
5100 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5101 if (!write || (ret == -EINVAL))
5103 for_each_populated_zone(zone) {
5104 for_each_possible_cpu(cpu) {
5106 high = zone->present_pages / percpu_pagelist_fraction;
5107 setup_pagelist_highmark(
5108 per_cpu_ptr(zone->pageset, cpu), high);
5114 int hashdist = HASHDIST_DEFAULT;
5117 static int __init set_hashdist(char *str)
5121 hashdist = simple_strtoul(str, &str, 0);
5124 __setup("hashdist=", set_hashdist);
5128 * allocate a large system hash table from bootmem
5129 * - it is assumed that the hash table must contain an exact power-of-2
5130 * quantity of entries
5131 * - limit is the number of hash buckets, not the total allocation size
5133 void *__init alloc_large_system_hash(const char *tablename,
5134 unsigned long bucketsize,
5135 unsigned long numentries,
5138 unsigned int *_hash_shift,
5139 unsigned int *_hash_mask,
5140 unsigned long limit)
5142 unsigned long long max = limit;
5143 unsigned long log2qty, size;
5146 /* allow the kernel cmdline to have a say */
5148 /* round applicable memory size up to nearest megabyte */
5149 numentries = nr_kernel_pages;
5150 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5151 numentries >>= 20 - PAGE_SHIFT;
5152 numentries <<= 20 - PAGE_SHIFT;
5154 /* limit to 1 bucket per 2^scale bytes of low memory */
5155 if (scale > PAGE_SHIFT)
5156 numentries >>= (scale - PAGE_SHIFT);
5158 numentries <<= (PAGE_SHIFT - scale);
5160 /* Make sure we've got at least a 0-order allocation.. */
5161 if (unlikely(flags & HASH_SMALL)) {
5162 /* Makes no sense without HASH_EARLY */
5163 WARN_ON(!(flags & HASH_EARLY));
5164 if (!(numentries >> *_hash_shift)) {
5165 numentries = 1UL << *_hash_shift;
5166 BUG_ON(!numentries);
5168 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5169 numentries = PAGE_SIZE / bucketsize;
5171 numentries = roundup_pow_of_two(numentries);
5173 /* limit allocation size to 1/16 total memory by default */
5175 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5176 do_div(max, bucketsize);
5179 if (numentries > max)
5182 log2qty = ilog2(numentries);
5185 size = bucketsize << log2qty;
5186 if (flags & HASH_EARLY)
5187 table = alloc_bootmem_nopanic(size);
5189 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5192 * If bucketsize is not a power-of-two, we may free
5193 * some pages at the end of hash table which
5194 * alloc_pages_exact() automatically does
5196 if (get_order(size) < MAX_ORDER) {
5197 table = alloc_pages_exact(size, GFP_ATOMIC);
5198 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5201 } while (!table && size > PAGE_SIZE && --log2qty);
5204 panic("Failed to allocate %s hash table\n", tablename);
5206 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5209 ilog2(size) - PAGE_SHIFT,
5213 *_hash_shift = log2qty;
5215 *_hash_mask = (1 << log2qty) - 1;
5220 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5221 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5224 #ifdef CONFIG_SPARSEMEM
5225 return __pfn_to_section(pfn)->pageblock_flags;
5227 return zone->pageblock_flags;
5228 #endif /* CONFIG_SPARSEMEM */
5231 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5233 #ifdef CONFIG_SPARSEMEM
5234 pfn &= (PAGES_PER_SECTION-1);
5235 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5237 pfn = pfn - zone->zone_start_pfn;
5238 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5239 #endif /* CONFIG_SPARSEMEM */
5243 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5244 * @page: The page within the block of interest
5245 * @start_bitidx: The first bit of interest to retrieve
5246 * @end_bitidx: The last bit of interest
5247 * returns pageblock_bits flags
5249 unsigned long get_pageblock_flags_group(struct page *page,
5250 int start_bitidx, int end_bitidx)
5253 unsigned long *bitmap;
5254 unsigned long pfn, bitidx;
5255 unsigned long flags = 0;
5256 unsigned long value = 1;
5258 zone = page_zone(page);
5259 pfn = page_to_pfn(page);
5260 bitmap = get_pageblock_bitmap(zone, pfn);
5261 bitidx = pfn_to_bitidx(zone, pfn);
5263 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5264 if (test_bit(bitidx + start_bitidx, bitmap))
5271 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5272 * @page: The page within the block of interest
5273 * @start_bitidx: The first bit of interest
5274 * @end_bitidx: The last bit of interest
5275 * @flags: The flags to set
5277 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5278 int start_bitidx, int end_bitidx)
5281 unsigned long *bitmap;
5282 unsigned long pfn, bitidx;
5283 unsigned long value = 1;
5285 zone = page_zone(page);
5286 pfn = page_to_pfn(page);
5287 bitmap = get_pageblock_bitmap(zone, pfn);
5288 bitidx = pfn_to_bitidx(zone, pfn);
5289 VM_BUG_ON(pfn < zone->zone_start_pfn);
5290 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5292 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5294 __set_bit(bitidx + start_bitidx, bitmap);
5296 __clear_bit(bitidx + start_bitidx, bitmap);
5300 * This is designed as sub function...plz see page_isolation.c also.
5301 * set/clear page block's type to be ISOLATE.
5302 * page allocater never alloc memory from ISOLATE block.
5305 int set_migratetype_isolate(struct page *page)
5308 struct page *curr_page;
5309 unsigned long flags, pfn, iter;
5310 unsigned long immobile = 0;
5311 struct memory_isolate_notify arg;
5316 zone = page_zone(page);
5317 zone_idx = zone_idx(zone);
5319 spin_lock_irqsave(&zone->lock, flags);
5320 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE ||
5321 zone_idx == ZONE_MOVABLE) {
5326 pfn = page_to_pfn(page);
5327 arg.start_pfn = pfn;
5328 arg.nr_pages = pageblock_nr_pages;
5329 arg.pages_found = 0;
5332 * It may be possible to isolate a pageblock even if the
5333 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5334 * notifier chain is used by balloon drivers to return the
5335 * number of pages in a range that are held by the balloon
5336 * driver to shrink memory. If all the pages are accounted for
5337 * by balloons, are free, or on the LRU, isolation can continue.
5338 * Later, for example, when memory hotplug notifier runs, these
5339 * pages reported as "can be isolated" should be isolated(freed)
5340 * by the balloon driver through the memory notifier chain.
5342 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5343 notifier_ret = notifier_to_errno(notifier_ret);
5344 if (notifier_ret || !arg.pages_found)
5347 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) {
5348 if (!pfn_valid_within(pfn))
5351 curr_page = pfn_to_page(iter);
5352 if (!page_count(curr_page) || PageLRU(curr_page))
5358 if (arg.pages_found == immobile)
5363 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5364 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5367 spin_unlock_irqrestore(&zone->lock, flags);
5373 void unset_migratetype_isolate(struct page *page)
5376 unsigned long flags;
5377 zone = page_zone(page);
5378 spin_lock_irqsave(&zone->lock, flags);
5379 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5381 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5382 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5384 spin_unlock_irqrestore(&zone->lock, flags);
5387 #ifdef CONFIG_MEMORY_HOTREMOVE
5389 * All pages in the range must be isolated before calling this.
5392 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5398 unsigned long flags;
5399 /* find the first valid pfn */
5400 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5405 zone = page_zone(pfn_to_page(pfn));
5406 spin_lock_irqsave(&zone->lock, flags);
5408 while (pfn < end_pfn) {
5409 if (!pfn_valid(pfn)) {
5413 page = pfn_to_page(pfn);
5414 BUG_ON(page_count(page));
5415 BUG_ON(!PageBuddy(page));
5416 order = page_order(page);
5417 #ifdef CONFIG_DEBUG_VM
5418 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5419 pfn, 1 << order, end_pfn);
5421 list_del(&page->lru);
5422 rmv_page_order(page);
5423 zone->free_area[order].nr_free--;
5424 __mod_zone_page_state(zone, NR_FREE_PAGES,
5426 for (i = 0; i < (1 << order); i++)
5427 SetPageReserved((page+i));
5428 pfn += (1 << order);
5430 spin_unlock_irqrestore(&zone->lock, flags);
5434 #ifdef CONFIG_MEMORY_FAILURE
5435 bool is_free_buddy_page(struct page *page)
5437 struct zone *zone = page_zone(page);
5438 unsigned long pfn = page_to_pfn(page);
5439 unsigned long flags;
5442 spin_lock_irqsave(&zone->lock, flags);
5443 for (order = 0; order < MAX_ORDER; order++) {
5444 struct page *page_head = page - (pfn & ((1 << order) - 1));
5446 if (PageBuddy(page_head) && page_order(page_head) >= order)
5449 spin_unlock_irqrestore(&zone->lock, flags);
5451 return order < MAX_ORDER;
5455 static struct trace_print_flags pageflag_names[] = {
5456 {1UL << PG_locked, "locked" },
5457 {1UL << PG_error, "error" },
5458 {1UL << PG_referenced, "referenced" },
5459 {1UL << PG_uptodate, "uptodate" },
5460 {1UL << PG_dirty, "dirty" },
5461 {1UL << PG_lru, "lru" },
5462 {1UL << PG_active, "active" },
5463 {1UL << PG_slab, "slab" },
5464 {1UL << PG_owner_priv_1, "owner_priv_1" },
5465 {1UL << PG_arch_1, "arch_1" },
5466 {1UL << PG_reserved, "reserved" },
5467 {1UL << PG_private, "private" },
5468 {1UL << PG_private_2, "private_2" },
5469 {1UL << PG_writeback, "writeback" },
5470 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5471 {1UL << PG_head, "head" },
5472 {1UL << PG_tail, "tail" },
5474 {1UL << PG_compound, "compound" },
5476 {1UL << PG_swapcache, "swapcache" },
5477 {1UL << PG_mappedtodisk, "mappedtodisk" },
5478 {1UL << PG_reclaim, "reclaim" },
5479 {1UL << PG_buddy, "buddy" },
5480 {1UL << PG_swapbacked, "swapbacked" },
5481 {1UL << PG_unevictable, "unevictable" },
5483 {1UL << PG_mlocked, "mlocked" },
5485 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5486 {1UL << PG_uncached, "uncached" },
5488 #ifdef CONFIG_MEMORY_FAILURE
5489 {1UL << PG_hwpoison, "hwpoison" },
5494 static void dump_page_flags(unsigned long flags)
5496 const char *delim = "";
5500 printk(KERN_ALERT "page flags: %#lx(", flags);
5502 /* remove zone id */
5503 flags &= (1UL << NR_PAGEFLAGS) - 1;
5505 for (i = 0; pageflag_names[i].name && flags; i++) {
5507 mask = pageflag_names[i].mask;
5508 if ((flags & mask) != mask)
5512 printk("%s%s", delim, pageflag_names[i].name);
5516 /* check for left over flags */
5518 printk("%s%#lx", delim, flags);
5523 void dump_page(struct page *page)
5526 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5527 page, page_count(page), page_mapcount(page),
5528 page->mapping, page->index);
5529 dump_page_flags(page->flags);