2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/oom.h>
34 #include <linux/notifier.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/mempolicy.h>
43 #include <linux/stop_machine.h>
44 #include <linux/sort.h>
45 #include <linux/pfn.h>
46 #include <linux/backing-dev.h>
47 #include <linux/fault-inject.h>
48 #include <linux/page-isolation.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/debugobjects.h>
51 #include <linux/kmemleak.h>
52 #include <linux/memory.h>
53 #include <linux/compaction.h>
54 #include <trace/events/kmem.h>
55 #include <linux/ftrace_event.h>
57 #include <asm/tlbflush.h>
58 #include <asm/div64.h>
61 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
62 DEFINE_PER_CPU(int, numa_node);
63 EXPORT_PER_CPU_SYMBOL(numa_node);
66 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
68 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
69 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
70 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
71 * defined in <linux/topology.h>.
73 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
74 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
78 * Array of node states.
80 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
81 [N_POSSIBLE] = NODE_MASK_ALL,
82 [N_ONLINE] = { { [0] = 1UL } },
84 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
86 [N_HIGH_MEMORY] = { { [0] = 1UL } },
88 [N_CPU] = { { [0] = 1UL } },
91 EXPORT_SYMBOL(node_states);
93 unsigned long totalram_pages __read_mostly;
94 unsigned long totalreserve_pages __read_mostly;
95 int percpu_pagelist_fraction;
96 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
98 #ifdef CONFIG_PM_SLEEP
100 * The following functions are used by the suspend/hibernate code to temporarily
101 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
102 * while devices are suspended. To avoid races with the suspend/hibernate code,
103 * they should always be called with pm_mutex held (gfp_allowed_mask also should
104 * only be modified with pm_mutex held, unless the suspend/hibernate code is
105 * guaranteed not to run in parallel with that modification).
107 void set_gfp_allowed_mask(gfp_t mask)
109 WARN_ON(!mutex_is_locked(&pm_mutex));
110 gfp_allowed_mask = mask;
113 gfp_t clear_gfp_allowed_mask(gfp_t mask)
115 gfp_t ret = gfp_allowed_mask;
117 WARN_ON(!mutex_is_locked(&pm_mutex));
118 gfp_allowed_mask &= ~mask;
121 #endif /* CONFIG_PM_SLEEP */
123 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
124 int pageblock_order __read_mostly;
127 static void __free_pages_ok(struct page *page, unsigned int order);
130 * results with 256, 32 in the lowmem_reserve sysctl:
131 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
132 * 1G machine -> (16M dma, 784M normal, 224M high)
133 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
134 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
135 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
137 * TBD: should special case ZONE_DMA32 machines here - in those we normally
138 * don't need any ZONE_NORMAL reservation
140 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
141 #ifdef CONFIG_ZONE_DMA
144 #ifdef CONFIG_ZONE_DMA32
147 #ifdef CONFIG_HIGHMEM
153 EXPORT_SYMBOL(totalram_pages);
155 static char * const zone_names[MAX_NR_ZONES] = {
156 #ifdef CONFIG_ZONE_DMA
159 #ifdef CONFIG_ZONE_DMA32
163 #ifdef CONFIG_HIGHMEM
169 int min_free_kbytes = 1024;
171 static unsigned long __meminitdata nr_kernel_pages;
172 static unsigned long __meminitdata nr_all_pages;
173 static unsigned long __meminitdata dma_reserve;
175 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
177 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
178 * ranges of memory (RAM) that may be registered with add_active_range().
179 * Ranges passed to add_active_range() will be merged if possible
180 * so the number of times add_active_range() can be called is
181 * related to the number of nodes and the number of holes
183 #ifdef CONFIG_MAX_ACTIVE_REGIONS
184 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
185 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
187 #if MAX_NUMNODES >= 32
188 /* If there can be many nodes, allow up to 50 holes per node */
189 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
191 /* By default, allow up to 256 distinct regions */
192 #define MAX_ACTIVE_REGIONS 256
196 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
197 static int __meminitdata nr_nodemap_entries;
198 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
199 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
200 static unsigned long __initdata required_kernelcore;
201 static unsigned long __initdata required_movablecore;
202 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
204 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
206 EXPORT_SYMBOL(movable_zone);
207 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
210 int nr_node_ids __read_mostly = MAX_NUMNODES;
211 int nr_online_nodes __read_mostly = 1;
212 EXPORT_SYMBOL(nr_node_ids);
213 EXPORT_SYMBOL(nr_online_nodes);
216 int page_group_by_mobility_disabled __read_mostly;
218 static void set_pageblock_migratetype(struct page *page, int migratetype)
221 if (unlikely(page_group_by_mobility_disabled))
222 migratetype = MIGRATE_UNMOVABLE;
224 set_pageblock_flags_group(page, (unsigned long)migratetype,
225 PB_migrate, PB_migrate_end);
228 bool oom_killer_disabled __read_mostly;
230 #ifdef CONFIG_DEBUG_VM
231 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
235 unsigned long pfn = page_to_pfn(page);
238 seq = zone_span_seqbegin(zone);
239 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
241 else if (pfn < zone->zone_start_pfn)
243 } while (zone_span_seqretry(zone, seq));
248 static int page_is_consistent(struct zone *zone, struct page *page)
250 if (!pfn_valid_within(page_to_pfn(page)))
252 if (zone != page_zone(page))
258 * Temporary debugging check for pages not lying within a given zone.
260 static int bad_range(struct zone *zone, struct page *page)
262 if (page_outside_zone_boundaries(zone, page))
264 if (!page_is_consistent(zone, page))
270 static inline int bad_range(struct zone *zone, struct page *page)
276 static void bad_page(struct page *page)
278 static unsigned long resume;
279 static unsigned long nr_shown;
280 static unsigned long nr_unshown;
282 /* Don't complain about poisoned pages */
283 if (PageHWPoison(page)) {
284 __ClearPageBuddy(page);
289 * Allow a burst of 60 reports, then keep quiet for that minute;
290 * or allow a steady drip of one report per second.
292 if (nr_shown == 60) {
293 if (time_before(jiffies, resume)) {
299 "BUG: Bad page state: %lu messages suppressed\n",
306 resume = jiffies + 60 * HZ;
308 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
309 current->comm, page_to_pfn(page));
314 /* Leave bad fields for debug, except PageBuddy could make trouble */
315 __ClearPageBuddy(page);
316 add_taint(TAINT_BAD_PAGE);
320 * Higher-order pages are called "compound pages". They are structured thusly:
322 * The first PAGE_SIZE page is called the "head page".
324 * The remaining PAGE_SIZE pages are called "tail pages".
326 * All pages have PG_compound set. All pages have their ->private pointing at
327 * the head page (even the head page has this).
329 * The first tail page's ->lru.next holds the address of the compound page's
330 * put_page() function. Its ->lru.prev holds the order of allocation.
331 * This usage means that zero-order pages may not be compound.
334 static void free_compound_page(struct page *page)
336 __free_pages_ok(page, compound_order(page));
339 void prep_compound_page(struct page *page, unsigned long order)
342 int nr_pages = 1 << order;
344 set_compound_page_dtor(page, free_compound_page);
345 set_compound_order(page, order);
347 for (i = 1; i < nr_pages; i++) {
348 struct page *p = page + i;
351 p->first_page = page;
355 static int destroy_compound_page(struct page *page, unsigned long order)
358 int nr_pages = 1 << order;
361 if (unlikely(compound_order(page) != order) ||
362 unlikely(!PageHead(page))) {
367 __ClearPageHead(page);
369 for (i = 1; i < nr_pages; i++) {
370 struct page *p = page + i;
372 if (unlikely(!PageTail(p) || (p->first_page != page))) {
382 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
387 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
388 * and __GFP_HIGHMEM from hard or soft interrupt context.
390 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
391 for (i = 0; i < (1 << order); i++)
392 clear_highpage(page + i);
395 static inline void set_page_order(struct page *page, int order)
397 set_page_private(page, order);
398 __SetPageBuddy(page);
401 static inline void rmv_page_order(struct page *page)
403 __ClearPageBuddy(page);
404 set_page_private(page, 0);
408 * Locate the struct page for both the matching buddy in our
409 * pair (buddy1) and the combined O(n+1) page they form (page).
411 * 1) Any buddy B1 will have an order O twin B2 which satisfies
412 * the following equation:
414 * For example, if the starting buddy (buddy2) is #8 its order
416 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
418 * 2) Any buddy B will have an order O+1 parent P which
419 * satisfies the following equation:
422 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
424 static inline struct page *
425 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
427 unsigned long buddy_idx = page_idx ^ (1 << order);
429 return page + (buddy_idx - page_idx);
432 static inline unsigned long
433 __find_combined_index(unsigned long page_idx, unsigned int order)
435 return (page_idx & ~(1 << order));
439 * This function checks whether a page is free && is the buddy
440 * we can do coalesce a page and its buddy if
441 * (a) the buddy is not in a hole &&
442 * (b) the buddy is in the buddy system &&
443 * (c) a page and its buddy have the same order &&
444 * (d) a page and its buddy are in the same zone.
446 * For recording whether a page is in the buddy system, we use PG_buddy.
447 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
449 * For recording page's order, we use page_private(page).
451 static inline int page_is_buddy(struct page *page, struct page *buddy,
454 if (!pfn_valid_within(page_to_pfn(buddy)))
457 if (page_zone_id(page) != page_zone_id(buddy))
460 if (PageBuddy(buddy) && page_order(buddy) == order) {
461 VM_BUG_ON(page_count(buddy) != 0);
468 * Freeing function for a buddy system allocator.
470 * The concept of a buddy system is to maintain direct-mapped table
471 * (containing bit values) for memory blocks of various "orders".
472 * The bottom level table contains the map for the smallest allocatable
473 * units of memory (here, pages), and each level above it describes
474 * pairs of units from the levels below, hence, "buddies".
475 * At a high level, all that happens here is marking the table entry
476 * at the bottom level available, and propagating the changes upward
477 * as necessary, plus some accounting needed to play nicely with other
478 * parts of the VM system.
479 * At each level, we keep a list of pages, which are heads of continuous
480 * free pages of length of (1 << order) and marked with PG_buddy. Page's
481 * order is recorded in page_private(page) field.
482 * So when we are allocating or freeing one, we can derive the state of the
483 * other. That is, if we allocate a small block, and both were
484 * free, the remainder of the region must be split into blocks.
485 * If a block is freed, and its buddy is also free, then this
486 * triggers coalescing into a block of larger size.
491 static inline void __free_one_page(struct page *page,
492 struct zone *zone, unsigned int order,
495 unsigned long page_idx;
496 unsigned long combined_idx;
499 if (unlikely(PageCompound(page)))
500 if (unlikely(destroy_compound_page(page, order)))
503 VM_BUG_ON(migratetype == -1);
505 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
507 VM_BUG_ON(page_idx & ((1 << order) - 1));
508 VM_BUG_ON(bad_range(zone, page));
510 while (order < MAX_ORDER-1) {
511 buddy = __page_find_buddy(page, page_idx, order);
512 if (!page_is_buddy(page, buddy, order))
515 /* Our buddy is free, merge with it and move up one order. */
516 list_del(&buddy->lru);
517 zone->free_area[order].nr_free--;
518 rmv_page_order(buddy);
519 combined_idx = __find_combined_index(page_idx, order);
520 page = page + (combined_idx - page_idx);
521 page_idx = combined_idx;
524 set_page_order(page, order);
527 * If this is not the largest possible page, check if the buddy
528 * of the next-highest order is free. If it is, it's possible
529 * that pages are being freed that will coalesce soon. In case,
530 * that is happening, add the free page to the tail of the list
531 * so it's less likely to be used soon and more likely to be merged
532 * as a higher order page
534 if ((order < MAX_ORDER-1) && pfn_valid_within(page_to_pfn(buddy))) {
535 struct page *higher_page, *higher_buddy;
536 combined_idx = __find_combined_index(page_idx, order);
537 higher_page = page + combined_idx - page_idx;
538 higher_buddy = __page_find_buddy(higher_page, combined_idx, order + 1);
539 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
540 list_add_tail(&page->lru,
541 &zone->free_area[order].free_list[migratetype]);
546 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
548 zone->free_area[order].nr_free++;
552 * free_page_mlock() -- clean up attempts to free and mlocked() page.
553 * Page should not be on lru, so no need to fix that up.
554 * free_pages_check() will verify...
556 static inline void free_page_mlock(struct page *page)
558 __dec_zone_page_state(page, NR_MLOCK);
559 __count_vm_event(UNEVICTABLE_MLOCKFREED);
562 static inline int free_pages_check(struct page *page)
564 if (unlikely(page_mapcount(page) |
565 (page->mapping != NULL) |
566 (atomic_read(&page->_count) != 0) |
567 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
571 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
572 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
577 * Frees a number of pages from the PCP lists
578 * Assumes all pages on list are in same zone, and of same order.
579 * count is the number of pages to free.
581 * If the zone was previously in an "all pages pinned" state then look to
582 * see if this freeing clears that state.
584 * And clear the zone's pages_scanned counter, to hold off the "all pages are
585 * pinned" detection logic.
587 static void free_pcppages_bulk(struct zone *zone, int count,
588 struct per_cpu_pages *pcp)
593 spin_lock(&zone->lock);
594 zone->all_unreclaimable = 0;
595 zone->pages_scanned = 0;
597 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
600 struct list_head *list;
603 * Remove pages from lists in a round-robin fashion. A
604 * batch_free count is maintained that is incremented when an
605 * empty list is encountered. This is so more pages are freed
606 * off fuller lists instead of spinning excessively around empty
611 if (++migratetype == MIGRATE_PCPTYPES)
613 list = &pcp->lists[migratetype];
614 } while (list_empty(list));
617 page = list_entry(list->prev, struct page, lru);
618 /* must delete as __free_one_page list manipulates */
619 list_del(&page->lru);
620 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
621 __free_one_page(page, zone, 0, page_private(page));
622 trace_mm_page_pcpu_drain(page, 0, page_private(page));
623 } while (--count && --batch_free && !list_empty(list));
625 spin_unlock(&zone->lock);
628 static void free_one_page(struct zone *zone, struct page *page, int order,
631 spin_lock(&zone->lock);
632 zone->all_unreclaimable = 0;
633 zone->pages_scanned = 0;
635 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
636 __free_one_page(page, zone, order, migratetype);
637 spin_unlock(&zone->lock);
640 static bool free_pages_prepare(struct page *page, unsigned int order)
645 trace_mm_page_free_direct(page, order);
646 kmemcheck_free_shadow(page, order);
648 for (i = 0; i < (1 << order); i++) {
649 struct page *pg = page + i;
653 bad += free_pages_check(pg);
658 if (!PageHighMem(page)) {
659 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
660 debug_check_no_obj_freed(page_address(page),
663 arch_free_page(page, order);
664 kernel_map_pages(page, 1 << order, 0);
669 static void __free_pages_ok(struct page *page, unsigned int order)
672 int wasMlocked = __TestClearPageMlocked(page);
674 if (!free_pages_prepare(page, order))
677 local_irq_save(flags);
678 if (unlikely(wasMlocked))
679 free_page_mlock(page);
680 __count_vm_events(PGFREE, 1 << order);
681 free_one_page(page_zone(page), page, order,
682 get_pageblock_migratetype(page));
683 local_irq_restore(flags);
687 * permit the bootmem allocator to evade page validation on high-order frees
689 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
692 __ClearPageReserved(page);
693 set_page_count(page, 0);
694 set_page_refcounted(page);
700 for (loop = 0; loop < BITS_PER_LONG; loop++) {
701 struct page *p = &page[loop];
703 if (loop + 1 < BITS_PER_LONG)
705 __ClearPageReserved(p);
706 set_page_count(p, 0);
709 set_page_refcounted(page);
710 __free_pages(page, order);
716 * The order of subdivision here is critical for the IO subsystem.
717 * Please do not alter this order without good reasons and regression
718 * testing. Specifically, as large blocks of memory are subdivided,
719 * the order in which smaller blocks are delivered depends on the order
720 * they're subdivided in this function. This is the primary factor
721 * influencing the order in which pages are delivered to the IO
722 * subsystem according to empirical testing, and this is also justified
723 * by considering the behavior of a buddy system containing a single
724 * large block of memory acted on by a series of small allocations.
725 * This behavior is a critical factor in sglist merging's success.
729 static inline void expand(struct zone *zone, struct page *page,
730 int low, int high, struct free_area *area,
733 unsigned long size = 1 << high;
739 VM_BUG_ON(bad_range(zone, &page[size]));
740 list_add(&page[size].lru, &area->free_list[migratetype]);
742 set_page_order(&page[size], high);
747 * This page is about to be returned from the page allocator
749 static inline int check_new_page(struct page *page)
751 if (unlikely(page_mapcount(page) |
752 (page->mapping != NULL) |
753 (atomic_read(&page->_count) != 0) |
754 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
761 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
765 for (i = 0; i < (1 << order); i++) {
766 struct page *p = page + i;
767 if (unlikely(check_new_page(p)))
771 set_page_private(page, 0);
772 set_page_refcounted(page);
774 arch_alloc_page(page, order);
775 kernel_map_pages(page, 1 << order, 1);
777 if (gfp_flags & __GFP_ZERO)
778 prep_zero_page(page, order, gfp_flags);
780 if (order && (gfp_flags & __GFP_COMP))
781 prep_compound_page(page, order);
787 * Go through the free lists for the given migratetype and remove
788 * the smallest available page from the freelists
791 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
794 unsigned int current_order;
795 struct free_area * area;
798 /* Find a page of the appropriate size in the preferred list */
799 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
800 area = &(zone->free_area[current_order]);
801 if (list_empty(&area->free_list[migratetype]))
804 page = list_entry(area->free_list[migratetype].next,
806 list_del(&page->lru);
807 rmv_page_order(page);
809 expand(zone, page, order, current_order, area, migratetype);
818 * This array describes the order lists are fallen back to when
819 * the free lists for the desirable migrate type are depleted
821 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
822 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
823 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
824 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
825 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
829 * Move the free pages in a range to the free lists of the requested type.
830 * Note that start_page and end_pages are not aligned on a pageblock
831 * boundary. If alignment is required, use move_freepages_block()
833 static int move_freepages(struct zone *zone,
834 struct page *start_page, struct page *end_page,
841 #ifndef CONFIG_HOLES_IN_ZONE
843 * page_zone is not safe to call in this context when
844 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
845 * anyway as we check zone boundaries in move_freepages_block().
846 * Remove at a later date when no bug reports exist related to
847 * grouping pages by mobility
849 BUG_ON(page_zone(start_page) != page_zone(end_page));
852 for (page = start_page; page <= end_page;) {
853 /* Make sure we are not inadvertently changing nodes */
854 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
856 if (!pfn_valid_within(page_to_pfn(page))) {
861 if (!PageBuddy(page)) {
866 order = page_order(page);
867 list_del(&page->lru);
869 &zone->free_area[order].free_list[migratetype]);
871 pages_moved += 1 << order;
877 static int move_freepages_block(struct zone *zone, struct page *page,
880 unsigned long start_pfn, end_pfn;
881 struct page *start_page, *end_page;
883 start_pfn = page_to_pfn(page);
884 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
885 start_page = pfn_to_page(start_pfn);
886 end_page = start_page + pageblock_nr_pages - 1;
887 end_pfn = start_pfn + pageblock_nr_pages - 1;
889 /* Do not cross zone boundaries */
890 if (start_pfn < zone->zone_start_pfn)
892 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
895 return move_freepages(zone, start_page, end_page, migratetype);
898 static void change_pageblock_range(struct page *pageblock_page,
899 int start_order, int migratetype)
901 int nr_pageblocks = 1 << (start_order - pageblock_order);
903 while (nr_pageblocks--) {
904 set_pageblock_migratetype(pageblock_page, migratetype);
905 pageblock_page += pageblock_nr_pages;
909 /* Remove an element from the buddy allocator from the fallback list */
910 static inline struct page *
911 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
913 struct free_area * area;
918 /* Find the largest possible block of pages in the other list */
919 for (current_order = MAX_ORDER-1; current_order >= order;
921 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
922 migratetype = fallbacks[start_migratetype][i];
924 /* MIGRATE_RESERVE handled later if necessary */
925 if (migratetype == MIGRATE_RESERVE)
928 area = &(zone->free_area[current_order]);
929 if (list_empty(&area->free_list[migratetype]))
932 page = list_entry(area->free_list[migratetype].next,
937 * If breaking a large block of pages, move all free
938 * pages to the preferred allocation list. If falling
939 * back for a reclaimable kernel allocation, be more
940 * agressive about taking ownership of free pages
942 if (unlikely(current_order >= (pageblock_order >> 1)) ||
943 start_migratetype == MIGRATE_RECLAIMABLE ||
944 page_group_by_mobility_disabled) {
946 pages = move_freepages_block(zone, page,
949 /* Claim the whole block if over half of it is free */
950 if (pages >= (1 << (pageblock_order-1)) ||
951 page_group_by_mobility_disabled)
952 set_pageblock_migratetype(page,
955 migratetype = start_migratetype;
958 /* Remove the page from the freelists */
959 list_del(&page->lru);
960 rmv_page_order(page);
962 /* Take ownership for orders >= pageblock_order */
963 if (current_order >= pageblock_order)
964 change_pageblock_range(page, current_order,
967 expand(zone, page, order, current_order, area, migratetype);
969 trace_mm_page_alloc_extfrag(page, order, current_order,
970 start_migratetype, migratetype);
980 * Do the hard work of removing an element from the buddy allocator.
981 * Call me with the zone->lock already held.
983 static struct page *__rmqueue(struct zone *zone, unsigned int order,
989 page = __rmqueue_smallest(zone, order, migratetype);
991 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
992 page = __rmqueue_fallback(zone, order, migratetype);
995 * Use MIGRATE_RESERVE rather than fail an allocation. goto
996 * is used because __rmqueue_smallest is an inline function
997 * and we want just one call site
1000 migratetype = MIGRATE_RESERVE;
1005 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1010 * Obtain a specified number of elements from the buddy allocator, all under
1011 * a single hold of the lock, for efficiency. Add them to the supplied list.
1012 * Returns the number of new pages which were placed at *list.
1014 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1015 unsigned long count, struct list_head *list,
1016 int migratetype, int cold)
1020 spin_lock(&zone->lock);
1021 for (i = 0; i < count; ++i) {
1022 struct page *page = __rmqueue(zone, order, migratetype);
1023 if (unlikely(page == NULL))
1027 * Split buddy pages returned by expand() are received here
1028 * in physical page order. The page is added to the callers and
1029 * list and the list head then moves forward. From the callers
1030 * perspective, the linked list is ordered by page number in
1031 * some conditions. This is useful for IO devices that can
1032 * merge IO requests if the physical pages are ordered
1035 if (likely(cold == 0))
1036 list_add(&page->lru, list);
1038 list_add_tail(&page->lru, list);
1039 set_page_private(page, migratetype);
1042 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1043 spin_unlock(&zone->lock);
1049 * Called from the vmstat counter updater to drain pagesets of this
1050 * currently executing processor on remote nodes after they have
1053 * Note that this function must be called with the thread pinned to
1054 * a single processor.
1056 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1058 unsigned long flags;
1061 local_irq_save(flags);
1062 if (pcp->count >= pcp->batch)
1063 to_drain = pcp->batch;
1065 to_drain = pcp->count;
1066 free_pcppages_bulk(zone, to_drain, pcp);
1067 pcp->count -= to_drain;
1068 local_irq_restore(flags);
1073 * Drain pages of the indicated processor.
1075 * The processor must either be the current processor and the
1076 * thread pinned to the current processor or a processor that
1079 static void drain_pages(unsigned int cpu)
1081 unsigned long flags;
1084 for_each_populated_zone(zone) {
1085 struct per_cpu_pageset *pset;
1086 struct per_cpu_pages *pcp;
1088 local_irq_save(flags);
1089 pset = per_cpu_ptr(zone->pageset, cpu);
1092 free_pcppages_bulk(zone, pcp->count, pcp);
1094 local_irq_restore(flags);
1099 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1101 void drain_local_pages(void *arg)
1103 drain_pages(smp_processor_id());
1107 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1109 void drain_all_pages(void)
1111 on_each_cpu(drain_local_pages, NULL, 1);
1114 #ifdef CONFIG_HIBERNATION
1116 void mark_free_pages(struct zone *zone)
1118 unsigned long pfn, max_zone_pfn;
1119 unsigned long flags;
1121 struct list_head *curr;
1123 if (!zone->spanned_pages)
1126 spin_lock_irqsave(&zone->lock, flags);
1128 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1129 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1130 if (pfn_valid(pfn)) {
1131 struct page *page = pfn_to_page(pfn);
1133 if (!swsusp_page_is_forbidden(page))
1134 swsusp_unset_page_free(page);
1137 for_each_migratetype_order(order, t) {
1138 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1141 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1142 for (i = 0; i < (1UL << order); i++)
1143 swsusp_set_page_free(pfn_to_page(pfn + i));
1146 spin_unlock_irqrestore(&zone->lock, flags);
1148 #endif /* CONFIG_PM */
1151 * Free a 0-order page
1152 * cold == 1 ? free a cold page : free a hot page
1154 void free_hot_cold_page(struct page *page, int cold)
1156 struct zone *zone = page_zone(page);
1157 struct per_cpu_pages *pcp;
1158 unsigned long flags;
1160 int wasMlocked = __TestClearPageMlocked(page);
1162 if (!free_pages_prepare(page, 0))
1165 migratetype = get_pageblock_migratetype(page);
1166 set_page_private(page, migratetype);
1167 local_irq_save(flags);
1168 if (unlikely(wasMlocked))
1169 free_page_mlock(page);
1170 __count_vm_event(PGFREE);
1173 * We only track unmovable, reclaimable and movable on pcp lists.
1174 * Free ISOLATE pages back to the allocator because they are being
1175 * offlined but treat RESERVE as movable pages so we can get those
1176 * areas back if necessary. Otherwise, we may have to free
1177 * excessively into the page allocator
1179 if (migratetype >= MIGRATE_PCPTYPES) {
1180 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1181 free_one_page(zone, page, 0, migratetype);
1184 migratetype = MIGRATE_MOVABLE;
1187 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1189 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1191 list_add(&page->lru, &pcp->lists[migratetype]);
1193 if (pcp->count >= pcp->high) {
1194 free_pcppages_bulk(zone, pcp->batch, pcp);
1195 pcp->count -= pcp->batch;
1199 local_irq_restore(flags);
1203 * split_page takes a non-compound higher-order page, and splits it into
1204 * n (1<<order) sub-pages: page[0..n]
1205 * Each sub-page must be freed individually.
1207 * Note: this is probably too low level an operation for use in drivers.
1208 * Please consult with lkml before using this in your driver.
1210 void split_page(struct page *page, unsigned int order)
1214 VM_BUG_ON(PageCompound(page));
1215 VM_BUG_ON(!page_count(page));
1217 #ifdef CONFIG_KMEMCHECK
1219 * Split shadow pages too, because free(page[0]) would
1220 * otherwise free the whole shadow.
1222 if (kmemcheck_page_is_tracked(page))
1223 split_page(virt_to_page(page[0].shadow), order);
1226 for (i = 1; i < (1 << order); i++)
1227 set_page_refcounted(page + i);
1231 * Similar to split_page except the page is already free. As this is only
1232 * being used for migration, the migratetype of the block also changes.
1233 * As this is called with interrupts disabled, the caller is responsible
1234 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1237 * Note: this is probably too low level an operation for use in drivers.
1238 * Please consult with lkml before using this in your driver.
1240 int split_free_page(struct page *page)
1243 unsigned long watermark;
1246 BUG_ON(!PageBuddy(page));
1248 zone = page_zone(page);
1249 order = page_order(page);
1251 /* Obey watermarks as if the page was being allocated */
1252 watermark = low_wmark_pages(zone) + (1 << order);
1253 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1256 /* Remove page from free list */
1257 list_del(&page->lru);
1258 zone->free_area[order].nr_free--;
1259 rmv_page_order(page);
1260 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1262 /* Split into individual pages */
1263 set_page_refcounted(page);
1264 split_page(page, order);
1266 if (order >= pageblock_order - 1) {
1267 struct page *endpage = page + (1 << order) - 1;
1268 for (; page < endpage; page += pageblock_nr_pages)
1269 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1276 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1277 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1281 struct page *buffered_rmqueue(struct zone *preferred_zone,
1282 struct zone *zone, int order, gfp_t gfp_flags,
1285 unsigned long flags;
1287 int cold = !!(gfp_flags & __GFP_COLD);
1290 if (likely(order == 0)) {
1291 struct per_cpu_pages *pcp;
1292 struct list_head *list;
1294 local_irq_save(flags);
1295 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1296 list = &pcp->lists[migratetype];
1297 if (list_empty(list)) {
1298 pcp->count += rmqueue_bulk(zone, 0,
1301 if (unlikely(list_empty(list)))
1306 page = list_entry(list->prev, struct page, lru);
1308 page = list_entry(list->next, struct page, lru);
1310 list_del(&page->lru);
1313 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1315 * __GFP_NOFAIL is not to be used in new code.
1317 * All __GFP_NOFAIL callers should be fixed so that they
1318 * properly detect and handle allocation failures.
1320 * We most definitely don't want callers attempting to
1321 * allocate greater than order-1 page units with
1324 WARN_ON_ONCE(order > 1);
1326 spin_lock_irqsave(&zone->lock, flags);
1327 page = __rmqueue(zone, order, migratetype);
1328 spin_unlock(&zone->lock);
1331 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1334 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1335 zone_statistics(preferred_zone, zone);
1336 local_irq_restore(flags);
1338 VM_BUG_ON(bad_range(zone, page));
1339 if (prep_new_page(page, order, gfp_flags))
1344 local_irq_restore(flags);
1348 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1349 #define ALLOC_WMARK_MIN WMARK_MIN
1350 #define ALLOC_WMARK_LOW WMARK_LOW
1351 #define ALLOC_WMARK_HIGH WMARK_HIGH
1352 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1354 /* Mask to get the watermark bits */
1355 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1357 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1358 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1359 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1361 #ifdef CONFIG_FAIL_PAGE_ALLOC
1363 static struct fail_page_alloc_attr {
1364 struct fault_attr attr;
1366 u32 ignore_gfp_highmem;
1367 u32 ignore_gfp_wait;
1370 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1372 struct dentry *ignore_gfp_highmem_file;
1373 struct dentry *ignore_gfp_wait_file;
1374 struct dentry *min_order_file;
1376 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1378 } fail_page_alloc = {
1379 .attr = FAULT_ATTR_INITIALIZER,
1380 .ignore_gfp_wait = 1,
1381 .ignore_gfp_highmem = 1,
1385 static int __init setup_fail_page_alloc(char *str)
1387 return setup_fault_attr(&fail_page_alloc.attr, str);
1389 __setup("fail_page_alloc=", setup_fail_page_alloc);
1391 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1393 if (order < fail_page_alloc.min_order)
1395 if (gfp_mask & __GFP_NOFAIL)
1397 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1399 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1402 return should_fail(&fail_page_alloc.attr, 1 << order);
1405 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1407 static int __init fail_page_alloc_debugfs(void)
1409 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1413 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1417 dir = fail_page_alloc.attr.dentries.dir;
1419 fail_page_alloc.ignore_gfp_wait_file =
1420 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1421 &fail_page_alloc.ignore_gfp_wait);
1423 fail_page_alloc.ignore_gfp_highmem_file =
1424 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1425 &fail_page_alloc.ignore_gfp_highmem);
1426 fail_page_alloc.min_order_file =
1427 debugfs_create_u32("min-order", mode, dir,
1428 &fail_page_alloc.min_order);
1430 if (!fail_page_alloc.ignore_gfp_wait_file ||
1431 !fail_page_alloc.ignore_gfp_highmem_file ||
1432 !fail_page_alloc.min_order_file) {
1434 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1435 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1436 debugfs_remove(fail_page_alloc.min_order_file);
1437 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1443 late_initcall(fail_page_alloc_debugfs);
1445 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1447 #else /* CONFIG_FAIL_PAGE_ALLOC */
1449 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1454 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1457 * Return 1 if free pages are above 'mark'. This takes into account the order
1458 * of the allocation.
1460 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1461 int classzone_idx, int alloc_flags)
1463 /* free_pages my go negative - that's OK */
1465 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1468 if (alloc_flags & ALLOC_HIGH)
1470 if (alloc_flags & ALLOC_HARDER)
1473 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1475 for (o = 0; o < order; o++) {
1476 /* At the next order, this order's pages become unavailable */
1477 free_pages -= z->free_area[o].nr_free << o;
1479 /* Require fewer higher order pages to be free */
1482 if (free_pages <= min)
1490 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1491 * skip over zones that are not allowed by the cpuset, or that have
1492 * been recently (in last second) found to be nearly full. See further
1493 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1494 * that have to skip over a lot of full or unallowed zones.
1496 * If the zonelist cache is present in the passed in zonelist, then
1497 * returns a pointer to the allowed node mask (either the current
1498 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1500 * If the zonelist cache is not available for this zonelist, does
1501 * nothing and returns NULL.
1503 * If the fullzones BITMAP in the zonelist cache is stale (more than
1504 * a second since last zap'd) then we zap it out (clear its bits.)
1506 * We hold off even calling zlc_setup, until after we've checked the
1507 * first zone in the zonelist, on the theory that most allocations will
1508 * be satisfied from that first zone, so best to examine that zone as
1509 * quickly as we can.
1511 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1513 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1514 nodemask_t *allowednodes; /* zonelist_cache approximation */
1516 zlc = zonelist->zlcache_ptr;
1520 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1521 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1522 zlc->last_full_zap = jiffies;
1525 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1526 &cpuset_current_mems_allowed :
1527 &node_states[N_HIGH_MEMORY];
1528 return allowednodes;
1532 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1533 * if it is worth looking at further for free memory:
1534 * 1) Check that the zone isn't thought to be full (doesn't have its
1535 * bit set in the zonelist_cache fullzones BITMAP).
1536 * 2) Check that the zones node (obtained from the zonelist_cache
1537 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1538 * Return true (non-zero) if zone is worth looking at further, or
1539 * else return false (zero) if it is not.
1541 * This check -ignores- the distinction between various watermarks,
1542 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1543 * found to be full for any variation of these watermarks, it will
1544 * be considered full for up to one second by all requests, unless
1545 * we are so low on memory on all allowed nodes that we are forced
1546 * into the second scan of the zonelist.
1548 * In the second scan we ignore this zonelist cache and exactly
1549 * apply the watermarks to all zones, even it is slower to do so.
1550 * We are low on memory in the second scan, and should leave no stone
1551 * unturned looking for a free page.
1553 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1554 nodemask_t *allowednodes)
1556 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1557 int i; /* index of *z in zonelist zones */
1558 int n; /* node that zone *z is on */
1560 zlc = zonelist->zlcache_ptr;
1564 i = z - zonelist->_zonerefs;
1567 /* This zone is worth trying if it is allowed but not full */
1568 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1572 * Given 'z' scanning a zonelist, set the corresponding bit in
1573 * zlc->fullzones, so that subsequent attempts to allocate a page
1574 * from that zone don't waste time re-examining it.
1576 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1578 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1579 int i; /* index of *z in zonelist zones */
1581 zlc = zonelist->zlcache_ptr;
1585 i = z - zonelist->_zonerefs;
1587 set_bit(i, zlc->fullzones);
1590 #else /* CONFIG_NUMA */
1592 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1597 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1598 nodemask_t *allowednodes)
1603 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1606 #endif /* CONFIG_NUMA */
1609 * get_page_from_freelist goes through the zonelist trying to allocate
1612 static struct page *
1613 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1614 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1615 struct zone *preferred_zone, int migratetype)
1618 struct page *page = NULL;
1621 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1622 int zlc_active = 0; /* set if using zonelist_cache */
1623 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1625 classzone_idx = zone_idx(preferred_zone);
1628 * Scan zonelist, looking for a zone with enough free.
1629 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1631 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1632 high_zoneidx, nodemask) {
1633 if (NUMA_BUILD && zlc_active &&
1634 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1636 if ((alloc_flags & ALLOC_CPUSET) &&
1637 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1640 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1641 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1645 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1646 if (zone_watermark_ok(zone, order, mark,
1647 classzone_idx, alloc_flags))
1650 if (zone_reclaim_mode == 0)
1651 goto this_zone_full;
1653 ret = zone_reclaim(zone, gfp_mask, order);
1655 case ZONE_RECLAIM_NOSCAN:
1658 case ZONE_RECLAIM_FULL:
1659 /* scanned but unreclaimable */
1660 goto this_zone_full;
1662 /* did we reclaim enough */
1663 if (!zone_watermark_ok(zone, order, mark,
1664 classzone_idx, alloc_flags))
1665 goto this_zone_full;
1670 page = buffered_rmqueue(preferred_zone, zone, order,
1671 gfp_mask, migratetype);
1676 zlc_mark_zone_full(zonelist, z);
1678 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1680 * we do zlc_setup after the first zone is tried but only
1681 * if there are multiple nodes make it worthwhile
1683 allowednodes = zlc_setup(zonelist, alloc_flags);
1689 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1690 /* Disable zlc cache for second zonelist scan */
1698 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1699 unsigned long pages_reclaimed)
1701 /* Do not loop if specifically requested */
1702 if (gfp_mask & __GFP_NORETRY)
1706 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1707 * means __GFP_NOFAIL, but that may not be true in other
1710 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1714 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1715 * specified, then we retry until we no longer reclaim any pages
1716 * (above), or we've reclaimed an order of pages at least as
1717 * large as the allocation's order. In both cases, if the
1718 * allocation still fails, we stop retrying.
1720 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1724 * Don't let big-order allocations loop unless the caller
1725 * explicitly requests that.
1727 if (gfp_mask & __GFP_NOFAIL)
1733 static inline struct page *
1734 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1735 struct zonelist *zonelist, enum zone_type high_zoneidx,
1736 nodemask_t *nodemask, struct zone *preferred_zone,
1741 /* Acquire the OOM killer lock for the zones in zonelist */
1742 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1743 schedule_timeout_uninterruptible(1);
1748 * Go through the zonelist yet one more time, keep very high watermark
1749 * here, this is only to catch a parallel oom killing, we must fail if
1750 * we're still under heavy pressure.
1752 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1753 order, zonelist, high_zoneidx,
1754 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1755 preferred_zone, migratetype);
1759 if (!(gfp_mask & __GFP_NOFAIL)) {
1760 /* The OOM killer will not help higher order allocs */
1761 if (order > PAGE_ALLOC_COSTLY_ORDER)
1764 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1765 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1766 * The caller should handle page allocation failure by itself if
1767 * it specifies __GFP_THISNODE.
1768 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1770 if (gfp_mask & __GFP_THISNODE)
1773 /* Exhausted what can be done so it's blamo time */
1774 out_of_memory(zonelist, gfp_mask, order, nodemask);
1777 clear_zonelist_oom(zonelist, gfp_mask);
1781 #ifdef CONFIG_COMPACTION
1782 /* Try memory compaction for high-order allocations before reclaim */
1783 static struct page *
1784 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1785 struct zonelist *zonelist, enum zone_type high_zoneidx,
1786 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1787 int migratetype, unsigned long *did_some_progress)
1791 if (!order || compaction_deferred(preferred_zone))
1794 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1796 if (*did_some_progress != COMPACT_SKIPPED) {
1798 /* Page migration frees to the PCP lists but we want merging */
1799 drain_pages(get_cpu());
1802 page = get_page_from_freelist(gfp_mask, nodemask,
1803 order, zonelist, high_zoneidx,
1804 alloc_flags, preferred_zone,
1807 preferred_zone->compact_considered = 0;
1808 preferred_zone->compact_defer_shift = 0;
1809 count_vm_event(COMPACTSUCCESS);
1814 * It's bad if compaction run occurs and fails.
1815 * The most likely reason is that pages exist,
1816 * but not enough to satisfy watermarks.
1818 count_vm_event(COMPACTFAIL);
1819 defer_compaction(preferred_zone);
1827 static inline struct page *
1828 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1829 struct zonelist *zonelist, enum zone_type high_zoneidx,
1830 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1831 int migratetype, unsigned long *did_some_progress)
1835 #endif /* CONFIG_COMPACTION */
1837 /* The really slow allocator path where we enter direct reclaim */
1838 static inline struct page *
1839 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1840 struct zonelist *zonelist, enum zone_type high_zoneidx,
1841 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1842 int migratetype, unsigned long *did_some_progress)
1844 struct page *page = NULL;
1845 struct reclaim_state reclaim_state;
1846 struct task_struct *p = current;
1850 /* We now go into synchronous reclaim */
1851 cpuset_memory_pressure_bump();
1852 p->flags |= PF_MEMALLOC;
1853 lockdep_set_current_reclaim_state(gfp_mask);
1854 reclaim_state.reclaimed_slab = 0;
1855 p->reclaim_state = &reclaim_state;
1857 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1859 p->reclaim_state = NULL;
1860 lockdep_clear_current_reclaim_state();
1861 p->flags &= ~PF_MEMALLOC;
1868 if (likely(*did_some_progress))
1869 page = get_page_from_freelist(gfp_mask, nodemask, order,
1870 zonelist, high_zoneidx,
1871 alloc_flags, preferred_zone,
1877 * This is called in the allocator slow-path if the allocation request is of
1878 * sufficient urgency to ignore watermarks and take other desperate measures
1880 static inline struct page *
1881 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1882 struct zonelist *zonelist, enum zone_type high_zoneidx,
1883 nodemask_t *nodemask, struct zone *preferred_zone,
1889 page = get_page_from_freelist(gfp_mask, nodemask, order,
1890 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1891 preferred_zone, migratetype);
1893 if (!page && gfp_mask & __GFP_NOFAIL)
1894 congestion_wait(BLK_RW_ASYNC, HZ/50);
1895 } while (!page && (gfp_mask & __GFP_NOFAIL));
1901 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1902 enum zone_type high_zoneidx)
1907 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1908 wakeup_kswapd(zone, order);
1912 gfp_to_alloc_flags(gfp_t gfp_mask)
1914 struct task_struct *p = current;
1915 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1916 const gfp_t wait = gfp_mask & __GFP_WAIT;
1918 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1919 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1922 * The caller may dip into page reserves a bit more if the caller
1923 * cannot run direct reclaim, or if the caller has realtime scheduling
1924 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1925 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1927 alloc_flags |= (gfp_mask & __GFP_HIGH);
1930 alloc_flags |= ALLOC_HARDER;
1932 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1933 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1935 alloc_flags &= ~ALLOC_CPUSET;
1936 } else if (unlikely(rt_task(p)) && !in_interrupt())
1937 alloc_flags |= ALLOC_HARDER;
1939 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1940 if (!in_interrupt() &&
1941 ((p->flags & PF_MEMALLOC) ||
1942 unlikely(test_thread_flag(TIF_MEMDIE))))
1943 alloc_flags |= ALLOC_NO_WATERMARKS;
1949 static inline struct page *
1950 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1951 struct zonelist *zonelist, enum zone_type high_zoneidx,
1952 nodemask_t *nodemask, struct zone *preferred_zone,
1955 const gfp_t wait = gfp_mask & __GFP_WAIT;
1956 struct page *page = NULL;
1958 unsigned long pages_reclaimed = 0;
1959 unsigned long did_some_progress;
1960 struct task_struct *p = current;
1963 * In the slowpath, we sanity check order to avoid ever trying to
1964 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1965 * be using allocators in order of preference for an area that is
1968 if (order >= MAX_ORDER) {
1969 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1974 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1975 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1976 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1977 * using a larger set of nodes after it has established that the
1978 * allowed per node queues are empty and that nodes are
1981 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1985 wake_all_kswapd(order, zonelist, high_zoneidx);
1988 * OK, we're below the kswapd watermark and have kicked background
1989 * reclaim. Now things get more complex, so set up alloc_flags according
1990 * to how we want to proceed.
1992 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1994 /* This is the last chance, in general, before the goto nopage. */
1995 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1996 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1997 preferred_zone, migratetype);
2002 /* Allocate without watermarks if the context allows */
2003 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2004 page = __alloc_pages_high_priority(gfp_mask, order,
2005 zonelist, high_zoneidx, nodemask,
2006 preferred_zone, migratetype);
2011 /* Atomic allocations - we can't balance anything */
2015 /* Avoid recursion of direct reclaim */
2016 if (p->flags & PF_MEMALLOC)
2019 /* Avoid allocations with no watermarks from looping endlessly */
2020 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2023 /* Try direct compaction */
2024 page = __alloc_pages_direct_compact(gfp_mask, order,
2025 zonelist, high_zoneidx,
2027 alloc_flags, preferred_zone,
2028 migratetype, &did_some_progress);
2032 /* Try direct reclaim and then allocating */
2033 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2034 zonelist, high_zoneidx,
2036 alloc_flags, preferred_zone,
2037 migratetype, &did_some_progress);
2042 * If we failed to make any progress reclaiming, then we are
2043 * running out of options and have to consider going OOM
2045 if (!did_some_progress) {
2046 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2047 if (oom_killer_disabled)
2049 page = __alloc_pages_may_oom(gfp_mask, order,
2050 zonelist, high_zoneidx,
2051 nodemask, preferred_zone,
2057 * The OOM killer does not trigger for high-order
2058 * ~__GFP_NOFAIL allocations so if no progress is being
2059 * made, there are no other options and retrying is
2062 if (order > PAGE_ALLOC_COSTLY_ORDER &&
2063 !(gfp_mask & __GFP_NOFAIL))
2070 /* Check if we should retry the allocation */
2071 pages_reclaimed += did_some_progress;
2072 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2073 /* Wait for some write requests to complete then retry */
2074 congestion_wait(BLK_RW_ASYNC, HZ/50);
2079 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
2080 printk(KERN_WARNING "%s: page allocation failure."
2081 " order:%d, mode:0x%x\n",
2082 p->comm, order, gfp_mask);
2088 if (kmemcheck_enabled)
2089 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2095 * This is the 'heart' of the zoned buddy allocator.
2098 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2099 struct zonelist *zonelist, nodemask_t *nodemask)
2101 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2102 struct zone *preferred_zone;
2104 int migratetype = allocflags_to_migratetype(gfp_mask);
2106 gfp_mask &= gfp_allowed_mask;
2108 lockdep_trace_alloc(gfp_mask);
2110 might_sleep_if(gfp_mask & __GFP_WAIT);
2112 if (should_fail_alloc_page(gfp_mask, order))
2116 * Check the zones suitable for the gfp_mask contain at least one
2117 * valid zone. It's possible to have an empty zonelist as a result
2118 * of GFP_THISNODE and a memoryless node
2120 if (unlikely(!zonelist->_zonerefs->zone))
2124 /* The preferred zone is used for statistics later */
2125 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
2126 if (!preferred_zone) {
2131 /* First allocation attempt */
2132 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2133 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2134 preferred_zone, migratetype);
2135 if (unlikely(!page))
2136 page = __alloc_pages_slowpath(gfp_mask, order,
2137 zonelist, high_zoneidx, nodemask,
2138 preferred_zone, migratetype);
2141 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2144 EXPORT_SYMBOL(__alloc_pages_nodemask);
2147 * Common helper functions.
2149 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2154 * __get_free_pages() returns a 32-bit address, which cannot represent
2157 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2159 page = alloc_pages(gfp_mask, order);
2162 return (unsigned long) page_address(page);
2164 EXPORT_SYMBOL(__get_free_pages);
2166 unsigned long get_zeroed_page(gfp_t gfp_mask)
2168 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2170 EXPORT_SYMBOL(get_zeroed_page);
2172 void __pagevec_free(struct pagevec *pvec)
2174 int i = pagevec_count(pvec);
2177 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2178 free_hot_cold_page(pvec->pages[i], pvec->cold);
2182 void __free_pages(struct page *page, unsigned int order)
2184 if (put_page_testzero(page)) {
2186 free_hot_cold_page(page, 0);
2188 __free_pages_ok(page, order);
2192 EXPORT_SYMBOL(__free_pages);
2194 void free_pages(unsigned long addr, unsigned int order)
2197 VM_BUG_ON(!virt_addr_valid((void *)addr));
2198 __free_pages(virt_to_page((void *)addr), order);
2202 EXPORT_SYMBOL(free_pages);
2205 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2206 * @size: the number of bytes to allocate
2207 * @gfp_mask: GFP flags for the allocation
2209 * This function is similar to alloc_pages(), except that it allocates the
2210 * minimum number of pages to satisfy the request. alloc_pages() can only
2211 * allocate memory in power-of-two pages.
2213 * This function is also limited by MAX_ORDER.
2215 * Memory allocated by this function must be released by free_pages_exact().
2217 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2219 unsigned int order = get_order(size);
2222 addr = __get_free_pages(gfp_mask, order);
2224 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2225 unsigned long used = addr + PAGE_ALIGN(size);
2227 split_page(virt_to_page((void *)addr), order);
2228 while (used < alloc_end) {
2234 return (void *)addr;
2236 EXPORT_SYMBOL(alloc_pages_exact);
2239 * free_pages_exact - release memory allocated via alloc_pages_exact()
2240 * @virt: the value returned by alloc_pages_exact.
2241 * @size: size of allocation, same value as passed to alloc_pages_exact().
2243 * Release the memory allocated by a previous call to alloc_pages_exact.
2245 void free_pages_exact(void *virt, size_t size)
2247 unsigned long addr = (unsigned long)virt;
2248 unsigned long end = addr + PAGE_ALIGN(size);
2250 while (addr < end) {
2255 EXPORT_SYMBOL(free_pages_exact);
2257 static unsigned int nr_free_zone_pages(int offset)
2262 /* Just pick one node, since fallback list is circular */
2263 unsigned int sum = 0;
2265 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2267 for_each_zone_zonelist(zone, z, zonelist, offset) {
2268 unsigned long size = zone->present_pages;
2269 unsigned long high = high_wmark_pages(zone);
2278 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2280 unsigned int nr_free_buffer_pages(void)
2282 return nr_free_zone_pages(gfp_zone(GFP_USER));
2284 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2287 * Amount of free RAM allocatable within all zones
2289 unsigned int nr_free_pagecache_pages(void)
2291 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2294 static inline void show_node(struct zone *zone)
2297 printk("Node %d ", zone_to_nid(zone));
2300 void si_meminfo(struct sysinfo *val)
2302 val->totalram = totalram_pages;
2304 val->freeram = global_page_state(NR_FREE_PAGES);
2305 val->bufferram = nr_blockdev_pages();
2306 val->totalhigh = totalhigh_pages;
2307 val->freehigh = nr_free_highpages();
2308 val->mem_unit = PAGE_SIZE;
2311 EXPORT_SYMBOL(si_meminfo);
2314 void si_meminfo_node(struct sysinfo *val, int nid)
2316 pg_data_t *pgdat = NODE_DATA(nid);
2318 val->totalram = pgdat->node_present_pages;
2319 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2320 #ifdef CONFIG_HIGHMEM
2321 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2322 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2328 val->mem_unit = PAGE_SIZE;
2332 #define K(x) ((x) << (PAGE_SHIFT-10))
2335 * Show free area list (used inside shift_scroll-lock stuff)
2336 * We also calculate the percentage fragmentation. We do this by counting the
2337 * memory on each free list with the exception of the first item on the list.
2339 void show_free_areas(void)
2344 for_each_populated_zone(zone) {
2346 printk("%s per-cpu:\n", zone->name);
2348 for_each_online_cpu(cpu) {
2349 struct per_cpu_pageset *pageset;
2351 pageset = per_cpu_ptr(zone->pageset, cpu);
2353 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2354 cpu, pageset->pcp.high,
2355 pageset->pcp.batch, pageset->pcp.count);
2359 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2360 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2362 " dirty:%lu writeback:%lu unstable:%lu\n"
2363 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2364 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2365 global_page_state(NR_ACTIVE_ANON),
2366 global_page_state(NR_INACTIVE_ANON),
2367 global_page_state(NR_ISOLATED_ANON),
2368 global_page_state(NR_ACTIVE_FILE),
2369 global_page_state(NR_INACTIVE_FILE),
2370 global_page_state(NR_ISOLATED_FILE),
2371 global_page_state(NR_UNEVICTABLE),
2372 global_page_state(NR_FILE_DIRTY),
2373 global_page_state(NR_WRITEBACK),
2374 global_page_state(NR_UNSTABLE_NFS),
2375 global_page_state(NR_FREE_PAGES),
2376 global_page_state(NR_SLAB_RECLAIMABLE),
2377 global_page_state(NR_SLAB_UNRECLAIMABLE),
2378 global_page_state(NR_FILE_MAPPED),
2379 global_page_state(NR_SHMEM),
2380 global_page_state(NR_PAGETABLE),
2381 global_page_state(NR_BOUNCE));
2383 for_each_populated_zone(zone) {
2392 " active_anon:%lukB"
2393 " inactive_anon:%lukB"
2394 " active_file:%lukB"
2395 " inactive_file:%lukB"
2396 " unevictable:%lukB"
2397 " isolated(anon):%lukB"
2398 " isolated(file):%lukB"
2405 " slab_reclaimable:%lukB"
2406 " slab_unreclaimable:%lukB"
2407 " kernel_stack:%lukB"
2411 " writeback_tmp:%lukB"
2412 " pages_scanned:%lu"
2413 " all_unreclaimable? %s"
2416 K(zone_page_state(zone, NR_FREE_PAGES)),
2417 K(min_wmark_pages(zone)),
2418 K(low_wmark_pages(zone)),
2419 K(high_wmark_pages(zone)),
2420 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2421 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2422 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2423 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2424 K(zone_page_state(zone, NR_UNEVICTABLE)),
2425 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2426 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2427 K(zone->present_pages),
2428 K(zone_page_state(zone, NR_MLOCK)),
2429 K(zone_page_state(zone, NR_FILE_DIRTY)),
2430 K(zone_page_state(zone, NR_WRITEBACK)),
2431 K(zone_page_state(zone, NR_FILE_MAPPED)),
2432 K(zone_page_state(zone, NR_SHMEM)),
2433 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2434 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2435 zone_page_state(zone, NR_KERNEL_STACK) *
2437 K(zone_page_state(zone, NR_PAGETABLE)),
2438 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2439 K(zone_page_state(zone, NR_BOUNCE)),
2440 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2441 zone->pages_scanned,
2442 (zone->all_unreclaimable ? "yes" : "no")
2444 printk("lowmem_reserve[]:");
2445 for (i = 0; i < MAX_NR_ZONES; i++)
2446 printk(" %lu", zone->lowmem_reserve[i]);
2450 for_each_populated_zone(zone) {
2451 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2454 printk("%s: ", zone->name);
2456 spin_lock_irqsave(&zone->lock, flags);
2457 for (order = 0; order < MAX_ORDER; order++) {
2458 nr[order] = zone->free_area[order].nr_free;
2459 total += nr[order] << order;
2461 spin_unlock_irqrestore(&zone->lock, flags);
2462 for (order = 0; order < MAX_ORDER; order++)
2463 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2464 printk("= %lukB\n", K(total));
2467 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2469 show_swap_cache_info();
2472 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2474 zoneref->zone = zone;
2475 zoneref->zone_idx = zone_idx(zone);
2479 * Builds allocation fallback zone lists.
2481 * Add all populated zones of a node to the zonelist.
2483 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2484 int nr_zones, enum zone_type zone_type)
2488 BUG_ON(zone_type >= MAX_NR_ZONES);
2493 zone = pgdat->node_zones + zone_type;
2494 if (populated_zone(zone)) {
2495 zoneref_set_zone(zone,
2496 &zonelist->_zonerefs[nr_zones++]);
2497 check_highest_zone(zone_type);
2500 } while (zone_type);
2507 * 0 = automatic detection of better ordering.
2508 * 1 = order by ([node] distance, -zonetype)
2509 * 2 = order by (-zonetype, [node] distance)
2511 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2512 * the same zonelist. So only NUMA can configure this param.
2514 #define ZONELIST_ORDER_DEFAULT 0
2515 #define ZONELIST_ORDER_NODE 1
2516 #define ZONELIST_ORDER_ZONE 2
2518 /* zonelist order in the kernel.
2519 * set_zonelist_order() will set this to NODE or ZONE.
2521 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2522 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2526 /* The value user specified ....changed by config */
2527 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2528 /* string for sysctl */
2529 #define NUMA_ZONELIST_ORDER_LEN 16
2530 char numa_zonelist_order[16] = "default";
2533 * interface for configure zonelist ordering.
2534 * command line option "numa_zonelist_order"
2535 * = "[dD]efault - default, automatic configuration.
2536 * = "[nN]ode - order by node locality, then by zone within node
2537 * = "[zZ]one - order by zone, then by locality within zone
2540 static int __parse_numa_zonelist_order(char *s)
2542 if (*s == 'd' || *s == 'D') {
2543 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2544 } else if (*s == 'n' || *s == 'N') {
2545 user_zonelist_order = ZONELIST_ORDER_NODE;
2546 } else if (*s == 'z' || *s == 'Z') {
2547 user_zonelist_order = ZONELIST_ORDER_ZONE;
2550 "Ignoring invalid numa_zonelist_order value: "
2557 static __init int setup_numa_zonelist_order(char *s)
2560 return __parse_numa_zonelist_order(s);
2563 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2566 * sysctl handler for numa_zonelist_order
2568 int numa_zonelist_order_handler(ctl_table *table, int write,
2569 void __user *buffer, size_t *length,
2572 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2574 static DEFINE_MUTEX(zl_order_mutex);
2576 mutex_lock(&zl_order_mutex);
2578 strcpy(saved_string, (char*)table->data);
2579 ret = proc_dostring(table, write, buffer, length, ppos);
2583 int oldval = user_zonelist_order;
2584 if (__parse_numa_zonelist_order((char*)table->data)) {
2586 * bogus value. restore saved string
2588 strncpy((char*)table->data, saved_string,
2589 NUMA_ZONELIST_ORDER_LEN);
2590 user_zonelist_order = oldval;
2591 } else if (oldval != user_zonelist_order) {
2592 mutex_lock(&zonelists_mutex);
2593 build_all_zonelists(NULL);
2594 mutex_unlock(&zonelists_mutex);
2598 mutex_unlock(&zl_order_mutex);
2603 #define MAX_NODE_LOAD (nr_online_nodes)
2604 static int node_load[MAX_NUMNODES];
2607 * find_next_best_node - find the next node that should appear in a given node's fallback list
2608 * @node: node whose fallback list we're appending
2609 * @used_node_mask: nodemask_t of already used nodes
2611 * We use a number of factors to determine which is the next node that should
2612 * appear on a given node's fallback list. The node should not have appeared
2613 * already in @node's fallback list, and it should be the next closest node
2614 * according to the distance array (which contains arbitrary distance values
2615 * from each node to each node in the system), and should also prefer nodes
2616 * with no CPUs, since presumably they'll have very little allocation pressure
2617 * on them otherwise.
2618 * It returns -1 if no node is found.
2620 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2623 int min_val = INT_MAX;
2625 const struct cpumask *tmp = cpumask_of_node(0);
2627 /* Use the local node if we haven't already */
2628 if (!node_isset(node, *used_node_mask)) {
2629 node_set(node, *used_node_mask);
2633 for_each_node_state(n, N_HIGH_MEMORY) {
2635 /* Don't want a node to appear more than once */
2636 if (node_isset(n, *used_node_mask))
2639 /* Use the distance array to find the distance */
2640 val = node_distance(node, n);
2642 /* Penalize nodes under us ("prefer the next node") */
2645 /* Give preference to headless and unused nodes */
2646 tmp = cpumask_of_node(n);
2647 if (!cpumask_empty(tmp))
2648 val += PENALTY_FOR_NODE_WITH_CPUS;
2650 /* Slight preference for less loaded node */
2651 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2652 val += node_load[n];
2654 if (val < min_val) {
2661 node_set(best_node, *used_node_mask);
2668 * Build zonelists ordered by node and zones within node.
2669 * This results in maximum locality--normal zone overflows into local
2670 * DMA zone, if any--but risks exhausting DMA zone.
2672 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2675 struct zonelist *zonelist;
2677 zonelist = &pgdat->node_zonelists[0];
2678 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2680 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2682 zonelist->_zonerefs[j].zone = NULL;
2683 zonelist->_zonerefs[j].zone_idx = 0;
2687 * Build gfp_thisnode zonelists
2689 static void build_thisnode_zonelists(pg_data_t *pgdat)
2692 struct zonelist *zonelist;
2694 zonelist = &pgdat->node_zonelists[1];
2695 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2696 zonelist->_zonerefs[j].zone = NULL;
2697 zonelist->_zonerefs[j].zone_idx = 0;
2701 * Build zonelists ordered by zone and nodes within zones.
2702 * This results in conserving DMA zone[s] until all Normal memory is
2703 * exhausted, but results in overflowing to remote node while memory
2704 * may still exist in local DMA zone.
2706 static int node_order[MAX_NUMNODES];
2708 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2711 int zone_type; /* needs to be signed */
2713 struct zonelist *zonelist;
2715 zonelist = &pgdat->node_zonelists[0];
2717 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2718 for (j = 0; j < nr_nodes; j++) {
2719 node = node_order[j];
2720 z = &NODE_DATA(node)->node_zones[zone_type];
2721 if (populated_zone(z)) {
2723 &zonelist->_zonerefs[pos++]);
2724 check_highest_zone(zone_type);
2728 zonelist->_zonerefs[pos].zone = NULL;
2729 zonelist->_zonerefs[pos].zone_idx = 0;
2732 static int default_zonelist_order(void)
2735 unsigned long low_kmem_size,total_size;
2739 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2740 * If they are really small and used heavily, the system can fall
2741 * into OOM very easily.
2742 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2744 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2747 for_each_online_node(nid) {
2748 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2749 z = &NODE_DATA(nid)->node_zones[zone_type];
2750 if (populated_zone(z)) {
2751 if (zone_type < ZONE_NORMAL)
2752 low_kmem_size += z->present_pages;
2753 total_size += z->present_pages;
2754 } else if (zone_type == ZONE_NORMAL) {
2756 * If any node has only lowmem, then node order
2757 * is preferred to allow kernel allocations
2758 * locally; otherwise, they can easily infringe
2759 * on other nodes when there is an abundance of
2760 * lowmem available to allocate from.
2762 return ZONELIST_ORDER_NODE;
2766 if (!low_kmem_size || /* there are no DMA area. */
2767 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2768 return ZONELIST_ORDER_NODE;
2770 * look into each node's config.
2771 * If there is a node whose DMA/DMA32 memory is very big area on
2772 * local memory, NODE_ORDER may be suitable.
2774 average_size = total_size /
2775 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2776 for_each_online_node(nid) {
2779 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2780 z = &NODE_DATA(nid)->node_zones[zone_type];
2781 if (populated_zone(z)) {
2782 if (zone_type < ZONE_NORMAL)
2783 low_kmem_size += z->present_pages;
2784 total_size += z->present_pages;
2787 if (low_kmem_size &&
2788 total_size > average_size && /* ignore small node */
2789 low_kmem_size > total_size * 70/100)
2790 return ZONELIST_ORDER_NODE;
2792 return ZONELIST_ORDER_ZONE;
2795 static void set_zonelist_order(void)
2797 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2798 current_zonelist_order = default_zonelist_order();
2800 current_zonelist_order = user_zonelist_order;
2803 static void build_zonelists(pg_data_t *pgdat)
2807 nodemask_t used_mask;
2808 int local_node, prev_node;
2809 struct zonelist *zonelist;
2810 int order = current_zonelist_order;
2812 /* initialize zonelists */
2813 for (i = 0; i < MAX_ZONELISTS; i++) {
2814 zonelist = pgdat->node_zonelists + i;
2815 zonelist->_zonerefs[0].zone = NULL;
2816 zonelist->_zonerefs[0].zone_idx = 0;
2819 /* NUMA-aware ordering of nodes */
2820 local_node = pgdat->node_id;
2821 load = nr_online_nodes;
2822 prev_node = local_node;
2823 nodes_clear(used_mask);
2825 memset(node_order, 0, sizeof(node_order));
2828 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2829 int distance = node_distance(local_node, node);
2832 * If another node is sufficiently far away then it is better
2833 * to reclaim pages in a zone before going off node.
2835 if (distance > RECLAIM_DISTANCE)
2836 zone_reclaim_mode = 1;
2839 * We don't want to pressure a particular node.
2840 * So adding penalty to the first node in same
2841 * distance group to make it round-robin.
2843 if (distance != node_distance(local_node, prev_node))
2844 node_load[node] = load;
2848 if (order == ZONELIST_ORDER_NODE)
2849 build_zonelists_in_node_order(pgdat, node);
2851 node_order[j++] = node; /* remember order */
2854 if (order == ZONELIST_ORDER_ZONE) {
2855 /* calculate node order -- i.e., DMA last! */
2856 build_zonelists_in_zone_order(pgdat, j);
2859 build_thisnode_zonelists(pgdat);
2862 /* Construct the zonelist performance cache - see further mmzone.h */
2863 static void build_zonelist_cache(pg_data_t *pgdat)
2865 struct zonelist *zonelist;
2866 struct zonelist_cache *zlc;
2869 zonelist = &pgdat->node_zonelists[0];
2870 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2871 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2872 for (z = zonelist->_zonerefs; z->zone; z++)
2873 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2876 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
2878 * Return node id of node used for "local" allocations.
2879 * I.e., first node id of first zone in arg node's generic zonelist.
2880 * Used for initializing percpu 'numa_mem', which is used primarily
2881 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
2883 int local_memory_node(int node)
2887 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
2888 gfp_zone(GFP_KERNEL),
2895 #else /* CONFIG_NUMA */
2897 static void set_zonelist_order(void)
2899 current_zonelist_order = ZONELIST_ORDER_ZONE;
2902 static void build_zonelists(pg_data_t *pgdat)
2904 int node, local_node;
2906 struct zonelist *zonelist;
2908 local_node = pgdat->node_id;
2910 zonelist = &pgdat->node_zonelists[0];
2911 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2914 * Now we build the zonelist so that it contains the zones
2915 * of all the other nodes.
2916 * We don't want to pressure a particular node, so when
2917 * building the zones for node N, we make sure that the
2918 * zones coming right after the local ones are those from
2919 * node N+1 (modulo N)
2921 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2922 if (!node_online(node))
2924 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2927 for (node = 0; node < local_node; node++) {
2928 if (!node_online(node))
2930 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2934 zonelist->_zonerefs[j].zone = NULL;
2935 zonelist->_zonerefs[j].zone_idx = 0;
2938 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2939 static void build_zonelist_cache(pg_data_t *pgdat)
2941 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2944 #endif /* CONFIG_NUMA */
2947 * Boot pageset table. One per cpu which is going to be used for all
2948 * zones and all nodes. The parameters will be set in such a way
2949 * that an item put on a list will immediately be handed over to
2950 * the buddy list. This is safe since pageset manipulation is done
2951 * with interrupts disabled.
2953 * The boot_pagesets must be kept even after bootup is complete for
2954 * unused processors and/or zones. They do play a role for bootstrapping
2955 * hotplugged processors.
2957 * zoneinfo_show() and maybe other functions do
2958 * not check if the processor is online before following the pageset pointer.
2959 * Other parts of the kernel may not check if the zone is available.
2961 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
2962 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
2963 static void setup_zone_pageset(struct zone *zone);
2966 * Global mutex to protect against size modification of zonelists
2967 * as well as to serialize pageset setup for the new populated zone.
2969 DEFINE_MUTEX(zonelists_mutex);
2971 /* return values int ....just for stop_machine() */
2972 static __init_refok int __build_all_zonelists(void *data)
2978 memset(node_load, 0, sizeof(node_load));
2980 for_each_online_node(nid) {
2981 pg_data_t *pgdat = NODE_DATA(nid);
2983 build_zonelists(pgdat);
2984 build_zonelist_cache(pgdat);
2987 #ifdef CONFIG_MEMORY_HOTPLUG
2988 /* Setup real pagesets for the new zone */
2990 struct zone *zone = data;
2991 setup_zone_pageset(zone);
2996 * Initialize the boot_pagesets that are going to be used
2997 * for bootstrapping processors. The real pagesets for
2998 * each zone will be allocated later when the per cpu
2999 * allocator is available.
3001 * boot_pagesets are used also for bootstrapping offline
3002 * cpus if the system is already booted because the pagesets
3003 * are needed to initialize allocators on a specific cpu too.
3004 * F.e. the percpu allocator needs the page allocator which
3005 * needs the percpu allocator in order to allocate its pagesets
3006 * (a chicken-egg dilemma).
3008 for_each_possible_cpu(cpu) {
3009 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3011 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3013 * We now know the "local memory node" for each node--
3014 * i.e., the node of the first zone in the generic zonelist.
3015 * Set up numa_mem percpu variable for on-line cpus. During
3016 * boot, only the boot cpu should be on-line; we'll init the
3017 * secondary cpus' numa_mem as they come on-line. During
3018 * node/memory hotplug, we'll fixup all on-line cpus.
3020 if (cpu_online(cpu))
3021 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3029 * Called with zonelists_mutex held always
3030 * unless system_state == SYSTEM_BOOTING.
3032 void build_all_zonelists(void *data)
3034 set_zonelist_order();
3036 if (system_state == SYSTEM_BOOTING) {
3037 __build_all_zonelists(NULL);
3038 mminit_verify_zonelist();
3039 cpuset_init_current_mems_allowed();
3041 /* we have to stop all cpus to guarantee there is no user
3043 stop_machine(__build_all_zonelists, data, NULL);
3044 /* cpuset refresh routine should be here */
3046 vm_total_pages = nr_free_pagecache_pages();
3048 * Disable grouping by mobility if the number of pages in the
3049 * system is too low to allow the mechanism to work. It would be
3050 * more accurate, but expensive to check per-zone. This check is
3051 * made on memory-hotadd so a system can start with mobility
3052 * disabled and enable it later
3054 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3055 page_group_by_mobility_disabled = 1;
3057 page_group_by_mobility_disabled = 0;
3059 printk("Built %i zonelists in %s order, mobility grouping %s. "
3060 "Total pages: %ld\n",
3062 zonelist_order_name[current_zonelist_order],
3063 page_group_by_mobility_disabled ? "off" : "on",
3066 printk("Policy zone: %s\n", zone_names[policy_zone]);
3071 * Helper functions to size the waitqueue hash table.
3072 * Essentially these want to choose hash table sizes sufficiently
3073 * large so that collisions trying to wait on pages are rare.
3074 * But in fact, the number of active page waitqueues on typical
3075 * systems is ridiculously low, less than 200. So this is even
3076 * conservative, even though it seems large.
3078 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3079 * waitqueues, i.e. the size of the waitq table given the number of pages.
3081 #define PAGES_PER_WAITQUEUE 256
3083 #ifndef CONFIG_MEMORY_HOTPLUG
3084 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3086 unsigned long size = 1;
3088 pages /= PAGES_PER_WAITQUEUE;
3090 while (size < pages)
3094 * Once we have dozens or even hundreds of threads sleeping
3095 * on IO we've got bigger problems than wait queue collision.
3096 * Limit the size of the wait table to a reasonable size.
3098 size = min(size, 4096UL);
3100 return max(size, 4UL);
3104 * A zone's size might be changed by hot-add, so it is not possible to determine
3105 * a suitable size for its wait_table. So we use the maximum size now.
3107 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3109 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3110 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3111 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3113 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3114 * or more by the traditional way. (See above). It equals:
3116 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3117 * ia64(16K page size) : = ( 8G + 4M)byte.
3118 * powerpc (64K page size) : = (32G +16M)byte.
3120 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3127 * This is an integer logarithm so that shifts can be used later
3128 * to extract the more random high bits from the multiplicative
3129 * hash function before the remainder is taken.
3131 static inline unsigned long wait_table_bits(unsigned long size)
3136 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3139 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3140 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3141 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3142 * higher will lead to a bigger reserve which will get freed as contiguous
3143 * blocks as reclaim kicks in
3145 static void setup_zone_migrate_reserve(struct zone *zone)
3147 unsigned long start_pfn, pfn, end_pfn;
3149 unsigned long block_migratetype;
3152 /* Get the start pfn, end pfn and the number of blocks to reserve */
3153 start_pfn = zone->zone_start_pfn;
3154 end_pfn = start_pfn + zone->spanned_pages;
3155 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3159 * Reserve blocks are generally in place to help high-order atomic
3160 * allocations that are short-lived. A min_free_kbytes value that
3161 * would result in more than 2 reserve blocks for atomic allocations
3162 * is assumed to be in place to help anti-fragmentation for the
3163 * future allocation of hugepages at runtime.
3165 reserve = min(2, reserve);
3167 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3168 if (!pfn_valid(pfn))
3170 page = pfn_to_page(pfn);
3172 /* Watch out for overlapping nodes */
3173 if (page_to_nid(page) != zone_to_nid(zone))
3176 /* Blocks with reserved pages will never free, skip them. */
3177 if (PageReserved(page))
3180 block_migratetype = get_pageblock_migratetype(page);
3182 /* If this block is reserved, account for it */
3183 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3188 /* Suitable for reserving if this block is movable */
3189 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3190 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3191 move_freepages_block(zone, page, MIGRATE_RESERVE);
3197 * If the reserve is met and this is a previous reserved block,
3200 if (block_migratetype == MIGRATE_RESERVE) {
3201 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3202 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3208 * Initially all pages are reserved - free ones are freed
3209 * up by free_all_bootmem() once the early boot process is
3210 * done. Non-atomic initialization, single-pass.
3212 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3213 unsigned long start_pfn, enum memmap_context context)
3216 unsigned long end_pfn = start_pfn + size;
3220 if (highest_memmap_pfn < end_pfn - 1)
3221 highest_memmap_pfn = end_pfn - 1;
3223 z = &NODE_DATA(nid)->node_zones[zone];
3224 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3226 * There can be holes in boot-time mem_map[]s
3227 * handed to this function. They do not
3228 * exist on hotplugged memory.
3230 if (context == MEMMAP_EARLY) {
3231 if (!early_pfn_valid(pfn))
3233 if (!early_pfn_in_nid(pfn, nid))
3236 page = pfn_to_page(pfn);
3237 set_page_links(page, zone, nid, pfn);
3238 mminit_verify_page_links(page, zone, nid, pfn);
3239 init_page_count(page);
3240 reset_page_mapcount(page);
3241 SetPageReserved(page);
3243 * Mark the block movable so that blocks are reserved for
3244 * movable at startup. This will force kernel allocations
3245 * to reserve their blocks rather than leaking throughout
3246 * the address space during boot when many long-lived
3247 * kernel allocations are made. Later some blocks near
3248 * the start are marked MIGRATE_RESERVE by
3249 * setup_zone_migrate_reserve()
3251 * bitmap is created for zone's valid pfn range. but memmap
3252 * can be created for invalid pages (for alignment)
3253 * check here not to call set_pageblock_migratetype() against
3256 if ((z->zone_start_pfn <= pfn)
3257 && (pfn < z->zone_start_pfn + z->spanned_pages)
3258 && !(pfn & (pageblock_nr_pages - 1)))
3259 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3261 INIT_LIST_HEAD(&page->lru);
3262 #ifdef WANT_PAGE_VIRTUAL
3263 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3264 if (!is_highmem_idx(zone))
3265 set_page_address(page, __va(pfn << PAGE_SHIFT));
3270 static void __meminit zone_init_free_lists(struct zone *zone)
3273 for_each_migratetype_order(order, t) {
3274 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3275 zone->free_area[order].nr_free = 0;
3279 #ifndef __HAVE_ARCH_MEMMAP_INIT
3280 #define memmap_init(size, nid, zone, start_pfn) \
3281 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3284 static int zone_batchsize(struct zone *zone)
3290 * The per-cpu-pages pools are set to around 1000th of the
3291 * size of the zone. But no more than 1/2 of a meg.
3293 * OK, so we don't know how big the cache is. So guess.
3295 batch = zone->present_pages / 1024;
3296 if (batch * PAGE_SIZE > 512 * 1024)
3297 batch = (512 * 1024) / PAGE_SIZE;
3298 batch /= 4; /* We effectively *= 4 below */
3303 * Clamp the batch to a 2^n - 1 value. Having a power
3304 * of 2 value was found to be more likely to have
3305 * suboptimal cache aliasing properties in some cases.
3307 * For example if 2 tasks are alternately allocating
3308 * batches of pages, one task can end up with a lot
3309 * of pages of one half of the possible page colors
3310 * and the other with pages of the other colors.
3312 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3317 /* The deferral and batching of frees should be suppressed under NOMMU
3320 * The problem is that NOMMU needs to be able to allocate large chunks
3321 * of contiguous memory as there's no hardware page translation to
3322 * assemble apparent contiguous memory from discontiguous pages.
3324 * Queueing large contiguous runs of pages for batching, however,
3325 * causes the pages to actually be freed in smaller chunks. As there
3326 * can be a significant delay between the individual batches being
3327 * recycled, this leads to the once large chunks of space being
3328 * fragmented and becoming unavailable for high-order allocations.
3334 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3336 struct per_cpu_pages *pcp;
3339 memset(p, 0, sizeof(*p));
3343 pcp->high = 6 * batch;
3344 pcp->batch = max(1UL, 1 * batch);
3345 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3346 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3350 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3351 * to the value high for the pageset p.
3354 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3357 struct per_cpu_pages *pcp;
3361 pcp->batch = max(1UL, high/4);
3362 if ((high/4) > (PAGE_SHIFT * 8))
3363 pcp->batch = PAGE_SHIFT * 8;
3366 static __meminit void setup_zone_pageset(struct zone *zone)
3370 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3372 for_each_possible_cpu(cpu) {
3373 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3375 setup_pageset(pcp, zone_batchsize(zone));
3377 if (percpu_pagelist_fraction)
3378 setup_pagelist_highmark(pcp,
3379 (zone->present_pages /
3380 percpu_pagelist_fraction));
3385 * Allocate per cpu pagesets and initialize them.
3386 * Before this call only boot pagesets were available.
3388 void __init setup_per_cpu_pageset(void)
3392 for_each_populated_zone(zone)
3393 setup_zone_pageset(zone);
3396 static noinline __init_refok
3397 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3400 struct pglist_data *pgdat = zone->zone_pgdat;
3404 * The per-page waitqueue mechanism uses hashed waitqueues
3407 zone->wait_table_hash_nr_entries =
3408 wait_table_hash_nr_entries(zone_size_pages);
3409 zone->wait_table_bits =
3410 wait_table_bits(zone->wait_table_hash_nr_entries);
3411 alloc_size = zone->wait_table_hash_nr_entries
3412 * sizeof(wait_queue_head_t);
3414 if (!slab_is_available()) {
3415 zone->wait_table = (wait_queue_head_t *)
3416 alloc_bootmem_node(pgdat, alloc_size);
3419 * This case means that a zone whose size was 0 gets new memory
3420 * via memory hot-add.
3421 * But it may be the case that a new node was hot-added. In
3422 * this case vmalloc() will not be able to use this new node's
3423 * memory - this wait_table must be initialized to use this new
3424 * node itself as well.
3425 * To use this new node's memory, further consideration will be
3428 zone->wait_table = vmalloc(alloc_size);
3430 if (!zone->wait_table)
3433 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3434 init_waitqueue_head(zone->wait_table + i);
3439 static int __zone_pcp_update(void *data)
3441 struct zone *zone = data;
3443 unsigned long batch = zone_batchsize(zone), flags;
3445 for_each_possible_cpu(cpu) {
3446 struct per_cpu_pageset *pset;
3447 struct per_cpu_pages *pcp;
3449 pset = per_cpu_ptr(zone->pageset, cpu);
3452 local_irq_save(flags);
3453 free_pcppages_bulk(zone, pcp->count, pcp);
3454 setup_pageset(pset, batch);
3455 local_irq_restore(flags);
3460 void zone_pcp_update(struct zone *zone)
3462 stop_machine(__zone_pcp_update, zone, NULL);
3465 static __meminit void zone_pcp_init(struct zone *zone)
3468 * per cpu subsystem is not up at this point. The following code
3469 * relies on the ability of the linker to provide the
3470 * offset of a (static) per cpu variable into the per cpu area.
3472 zone->pageset = &boot_pageset;
3474 if (zone->present_pages)
3475 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3476 zone->name, zone->present_pages,
3477 zone_batchsize(zone));
3480 __meminit int init_currently_empty_zone(struct zone *zone,
3481 unsigned long zone_start_pfn,
3483 enum memmap_context context)
3485 struct pglist_data *pgdat = zone->zone_pgdat;
3487 ret = zone_wait_table_init(zone, size);
3490 pgdat->nr_zones = zone_idx(zone) + 1;
3492 zone->zone_start_pfn = zone_start_pfn;
3494 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3495 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3497 (unsigned long)zone_idx(zone),
3498 zone_start_pfn, (zone_start_pfn + size));
3500 zone_init_free_lists(zone);
3505 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3507 * Basic iterator support. Return the first range of PFNs for a node
3508 * Note: nid == MAX_NUMNODES returns first region regardless of node
3510 static int __meminit first_active_region_index_in_nid(int nid)
3514 for (i = 0; i < nr_nodemap_entries; i++)
3515 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3522 * Basic iterator support. Return the next active range of PFNs for a node
3523 * Note: nid == MAX_NUMNODES returns next region regardless of node
3525 static int __meminit next_active_region_index_in_nid(int index, int nid)
3527 for (index = index + 1; index < nr_nodemap_entries; index++)
3528 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3534 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3536 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3537 * Architectures may implement their own version but if add_active_range()
3538 * was used and there are no special requirements, this is a convenient
3541 int __meminit __early_pfn_to_nid(unsigned long pfn)
3545 for (i = 0; i < nr_nodemap_entries; i++) {
3546 unsigned long start_pfn = early_node_map[i].start_pfn;
3547 unsigned long end_pfn = early_node_map[i].end_pfn;
3549 if (start_pfn <= pfn && pfn < end_pfn)
3550 return early_node_map[i].nid;
3552 /* This is a memory hole */
3555 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3557 int __meminit early_pfn_to_nid(unsigned long pfn)
3561 nid = __early_pfn_to_nid(pfn);
3564 /* just returns 0 */
3568 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3569 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3573 nid = __early_pfn_to_nid(pfn);
3574 if (nid >= 0 && nid != node)
3580 /* Basic iterator support to walk early_node_map[] */
3581 #define for_each_active_range_index_in_nid(i, nid) \
3582 for (i = first_active_region_index_in_nid(nid); i != -1; \
3583 i = next_active_region_index_in_nid(i, nid))
3586 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3587 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3588 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3590 * If an architecture guarantees that all ranges registered with
3591 * add_active_ranges() contain no holes and may be freed, this
3592 * this function may be used instead of calling free_bootmem() manually.
3594 void __init free_bootmem_with_active_regions(int nid,
3595 unsigned long max_low_pfn)
3599 for_each_active_range_index_in_nid(i, nid) {
3600 unsigned long size_pages = 0;
3601 unsigned long end_pfn = early_node_map[i].end_pfn;
3603 if (early_node_map[i].start_pfn >= max_low_pfn)
3606 if (end_pfn > max_low_pfn)
3607 end_pfn = max_low_pfn;
3609 size_pages = end_pfn - early_node_map[i].start_pfn;
3610 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3611 PFN_PHYS(early_node_map[i].start_pfn),
3612 size_pages << PAGE_SHIFT);
3616 #ifdef CONFIG_HAVE_MEMBLOCK
3617 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3618 u64 goal, u64 limit)
3622 /* Need to go over early_node_map to find out good range for node */
3623 for_each_active_range_index_in_nid(i, nid) {
3625 u64 ei_start, ei_last;
3626 u64 final_start, final_end;
3628 ei_last = early_node_map[i].end_pfn;
3629 ei_last <<= PAGE_SHIFT;
3630 ei_start = early_node_map[i].start_pfn;
3631 ei_start <<= PAGE_SHIFT;
3633 final_start = max(ei_start, goal);
3634 final_end = min(ei_last, limit);
3636 if (final_start >= final_end)
3639 addr = memblock_find_in_range(final_start, final_end, size, align);
3641 if (addr == MEMBLOCK_ERROR)
3647 return MEMBLOCK_ERROR;
3651 int __init add_from_early_node_map(struct range *range, int az,
3652 int nr_range, int nid)
3657 /* need to go over early_node_map to find out good range for node */
3658 for_each_active_range_index_in_nid(i, nid) {
3659 start = early_node_map[i].start_pfn;
3660 end = early_node_map[i].end_pfn;
3661 nr_range = add_range(range, az, nr_range, start, end);
3666 #ifdef CONFIG_NO_BOOTMEM
3667 void * __init __alloc_memory_core_early(int nid, u64 size, u64 align,
3668 u64 goal, u64 limit)
3673 if (limit > memblock.current_limit)
3674 limit = memblock.current_limit;
3676 addr = find_memory_core_early(nid, size, align, goal, limit);
3678 if (addr == MEMBLOCK_ERROR)
3681 ptr = phys_to_virt(addr);
3682 memset(ptr, 0, size);
3683 memblock_x86_reserve_range(addr, addr + size, "BOOTMEM");
3685 * The min_count is set to 0 so that bootmem allocated blocks
3686 * are never reported as leaks.
3688 kmemleak_alloc(ptr, size, 0, 0);
3694 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3699 for_each_active_range_index_in_nid(i, nid) {
3700 ret = work_fn(early_node_map[i].start_pfn,
3701 early_node_map[i].end_pfn, data);
3707 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3708 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3710 * If an architecture guarantees that all ranges registered with
3711 * add_active_ranges() contain no holes and may be freed, this
3712 * function may be used instead of calling memory_present() manually.
3714 void __init sparse_memory_present_with_active_regions(int nid)
3718 for_each_active_range_index_in_nid(i, nid)
3719 memory_present(early_node_map[i].nid,
3720 early_node_map[i].start_pfn,
3721 early_node_map[i].end_pfn);
3725 * get_pfn_range_for_nid - Return the start and end page frames for a node
3726 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3727 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3728 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3730 * It returns the start and end page frame of a node based on information
3731 * provided by an arch calling add_active_range(). If called for a node
3732 * with no available memory, a warning is printed and the start and end
3735 void __meminit get_pfn_range_for_nid(unsigned int nid,
3736 unsigned long *start_pfn, unsigned long *end_pfn)
3742 for_each_active_range_index_in_nid(i, nid) {
3743 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3744 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3747 if (*start_pfn == -1UL)
3752 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3753 * assumption is made that zones within a node are ordered in monotonic
3754 * increasing memory addresses so that the "highest" populated zone is used
3756 static void __init find_usable_zone_for_movable(void)
3759 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3760 if (zone_index == ZONE_MOVABLE)
3763 if (arch_zone_highest_possible_pfn[zone_index] >
3764 arch_zone_lowest_possible_pfn[zone_index])
3768 VM_BUG_ON(zone_index == -1);
3769 movable_zone = zone_index;
3773 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3774 * because it is sized independant of architecture. Unlike the other zones,
3775 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3776 * in each node depending on the size of each node and how evenly kernelcore
3777 * is distributed. This helper function adjusts the zone ranges
3778 * provided by the architecture for a given node by using the end of the
3779 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3780 * zones within a node are in order of monotonic increases memory addresses
3782 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3783 unsigned long zone_type,
3784 unsigned long node_start_pfn,
3785 unsigned long node_end_pfn,
3786 unsigned long *zone_start_pfn,
3787 unsigned long *zone_end_pfn)
3789 /* Only adjust if ZONE_MOVABLE is on this node */
3790 if (zone_movable_pfn[nid]) {
3791 /* Size ZONE_MOVABLE */
3792 if (zone_type == ZONE_MOVABLE) {
3793 *zone_start_pfn = zone_movable_pfn[nid];
3794 *zone_end_pfn = min(node_end_pfn,
3795 arch_zone_highest_possible_pfn[movable_zone]);
3797 /* Adjust for ZONE_MOVABLE starting within this range */
3798 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3799 *zone_end_pfn > zone_movable_pfn[nid]) {
3800 *zone_end_pfn = zone_movable_pfn[nid];
3802 /* Check if this whole range is within ZONE_MOVABLE */
3803 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3804 *zone_start_pfn = *zone_end_pfn;
3809 * Return the number of pages a zone spans in a node, including holes
3810 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3812 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3813 unsigned long zone_type,
3814 unsigned long *ignored)
3816 unsigned long node_start_pfn, node_end_pfn;
3817 unsigned long zone_start_pfn, zone_end_pfn;
3819 /* Get the start and end of the node and zone */
3820 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3821 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3822 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3823 adjust_zone_range_for_zone_movable(nid, zone_type,
3824 node_start_pfn, node_end_pfn,
3825 &zone_start_pfn, &zone_end_pfn);
3827 /* Check that this node has pages within the zone's required range */
3828 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3831 /* Move the zone boundaries inside the node if necessary */
3832 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3833 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3835 /* Return the spanned pages */
3836 return zone_end_pfn - zone_start_pfn;
3840 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3841 * then all holes in the requested range will be accounted for.
3843 unsigned long __meminit __absent_pages_in_range(int nid,
3844 unsigned long range_start_pfn,
3845 unsigned long range_end_pfn)
3848 unsigned long prev_end_pfn = 0, hole_pages = 0;
3849 unsigned long start_pfn;
3851 /* Find the end_pfn of the first active range of pfns in the node */
3852 i = first_active_region_index_in_nid(nid);
3856 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3858 /* Account for ranges before physical memory on this node */
3859 if (early_node_map[i].start_pfn > range_start_pfn)
3860 hole_pages = prev_end_pfn - range_start_pfn;
3862 /* Find all holes for the zone within the node */
3863 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3865 /* No need to continue if prev_end_pfn is outside the zone */
3866 if (prev_end_pfn >= range_end_pfn)
3869 /* Make sure the end of the zone is not within the hole */
3870 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3871 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3873 /* Update the hole size cound and move on */
3874 if (start_pfn > range_start_pfn) {
3875 BUG_ON(prev_end_pfn > start_pfn);
3876 hole_pages += start_pfn - prev_end_pfn;
3878 prev_end_pfn = early_node_map[i].end_pfn;
3881 /* Account for ranges past physical memory on this node */
3882 if (range_end_pfn > prev_end_pfn)
3883 hole_pages += range_end_pfn -
3884 max(range_start_pfn, prev_end_pfn);
3890 * absent_pages_in_range - Return number of page frames in holes within a range
3891 * @start_pfn: The start PFN to start searching for holes
3892 * @end_pfn: The end PFN to stop searching for holes
3894 * It returns the number of pages frames in memory holes within a range.
3896 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3897 unsigned long end_pfn)
3899 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3902 /* Return the number of page frames in holes in a zone on a node */
3903 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3904 unsigned long zone_type,
3905 unsigned long *ignored)
3907 unsigned long node_start_pfn, node_end_pfn;
3908 unsigned long zone_start_pfn, zone_end_pfn;
3910 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3911 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3913 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3916 adjust_zone_range_for_zone_movable(nid, zone_type,
3917 node_start_pfn, node_end_pfn,
3918 &zone_start_pfn, &zone_end_pfn);
3919 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3923 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3924 unsigned long zone_type,
3925 unsigned long *zones_size)
3927 return zones_size[zone_type];
3930 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3931 unsigned long zone_type,
3932 unsigned long *zholes_size)
3937 return zholes_size[zone_type];
3942 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3943 unsigned long *zones_size, unsigned long *zholes_size)
3945 unsigned long realtotalpages, totalpages = 0;
3948 for (i = 0; i < MAX_NR_ZONES; i++)
3949 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3951 pgdat->node_spanned_pages = totalpages;
3953 realtotalpages = totalpages;
3954 for (i = 0; i < MAX_NR_ZONES; i++)
3956 zone_absent_pages_in_node(pgdat->node_id, i,
3958 pgdat->node_present_pages = realtotalpages;
3959 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3963 #ifndef CONFIG_SPARSEMEM
3965 * Calculate the size of the zone->blockflags rounded to an unsigned long
3966 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3967 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3968 * round what is now in bits to nearest long in bits, then return it in
3971 static unsigned long __init usemap_size(unsigned long zonesize)
3973 unsigned long usemapsize;
3975 usemapsize = roundup(zonesize, pageblock_nr_pages);
3976 usemapsize = usemapsize >> pageblock_order;
3977 usemapsize *= NR_PAGEBLOCK_BITS;
3978 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3980 return usemapsize / 8;
3983 static void __init setup_usemap(struct pglist_data *pgdat,
3984 struct zone *zone, unsigned long zonesize)
3986 unsigned long usemapsize = usemap_size(zonesize);
3987 zone->pageblock_flags = NULL;
3989 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3992 static void inline setup_usemap(struct pglist_data *pgdat,
3993 struct zone *zone, unsigned long zonesize) {}
3994 #endif /* CONFIG_SPARSEMEM */
3996 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3998 /* Return a sensible default order for the pageblock size. */
3999 static inline int pageblock_default_order(void)
4001 if (HPAGE_SHIFT > PAGE_SHIFT)
4002 return HUGETLB_PAGE_ORDER;
4007 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4008 static inline void __init set_pageblock_order(unsigned int order)
4010 /* Check that pageblock_nr_pages has not already been setup */
4011 if (pageblock_order)
4015 * Assume the largest contiguous order of interest is a huge page.
4016 * This value may be variable depending on boot parameters on IA64
4018 pageblock_order = order;
4020 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4023 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4024 * and pageblock_default_order() are unused as pageblock_order is set
4025 * at compile-time. See include/linux/pageblock-flags.h for the values of
4026 * pageblock_order based on the kernel config
4028 static inline int pageblock_default_order(unsigned int order)
4032 #define set_pageblock_order(x) do {} while (0)
4034 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4037 * Set up the zone data structures:
4038 * - mark all pages reserved
4039 * - mark all memory queues empty
4040 * - clear the memory bitmaps
4042 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4043 unsigned long *zones_size, unsigned long *zholes_size)
4046 int nid = pgdat->node_id;
4047 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4050 pgdat_resize_init(pgdat);
4051 pgdat->nr_zones = 0;
4052 init_waitqueue_head(&pgdat->kswapd_wait);
4053 pgdat->kswapd_max_order = 0;
4054 pgdat_page_cgroup_init(pgdat);
4056 for (j = 0; j < MAX_NR_ZONES; j++) {
4057 struct zone *zone = pgdat->node_zones + j;
4058 unsigned long size, realsize, memmap_pages;
4061 size = zone_spanned_pages_in_node(nid, j, zones_size);
4062 realsize = size - zone_absent_pages_in_node(nid, j,
4066 * Adjust realsize so that it accounts for how much memory
4067 * is used by this zone for memmap. This affects the watermark
4068 * and per-cpu initialisations
4071 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4072 if (realsize >= memmap_pages) {
4073 realsize -= memmap_pages;
4076 " %s zone: %lu pages used for memmap\n",
4077 zone_names[j], memmap_pages);
4080 " %s zone: %lu pages exceeds realsize %lu\n",
4081 zone_names[j], memmap_pages, realsize);
4083 /* Account for reserved pages */
4084 if (j == 0 && realsize > dma_reserve) {
4085 realsize -= dma_reserve;
4086 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4087 zone_names[0], dma_reserve);
4090 if (!is_highmem_idx(j))
4091 nr_kernel_pages += realsize;
4092 nr_all_pages += realsize;
4094 zone->spanned_pages = size;
4095 zone->present_pages = realsize;
4098 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4100 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4102 zone->name = zone_names[j];
4103 spin_lock_init(&zone->lock);
4104 spin_lock_init(&zone->lru_lock);
4105 zone_seqlock_init(zone);
4106 zone->zone_pgdat = pgdat;
4108 zone->prev_priority = DEF_PRIORITY;
4110 zone_pcp_init(zone);
4112 INIT_LIST_HEAD(&zone->lru[l].list);
4113 zone->reclaim_stat.nr_saved_scan[l] = 0;
4115 zone->reclaim_stat.recent_rotated[0] = 0;
4116 zone->reclaim_stat.recent_rotated[1] = 0;
4117 zone->reclaim_stat.recent_scanned[0] = 0;
4118 zone->reclaim_stat.recent_scanned[1] = 0;
4119 zap_zone_vm_stats(zone);
4124 set_pageblock_order(pageblock_default_order());
4125 setup_usemap(pgdat, zone, size);
4126 ret = init_currently_empty_zone(zone, zone_start_pfn,
4127 size, MEMMAP_EARLY);
4129 memmap_init(size, nid, j, zone_start_pfn);
4130 zone_start_pfn += size;
4134 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4136 /* Skip empty nodes */
4137 if (!pgdat->node_spanned_pages)
4140 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4141 /* ia64 gets its own node_mem_map, before this, without bootmem */
4142 if (!pgdat->node_mem_map) {
4143 unsigned long size, start, end;
4147 * The zone's endpoints aren't required to be MAX_ORDER
4148 * aligned but the node_mem_map endpoints must be in order
4149 * for the buddy allocator to function correctly.
4151 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4152 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4153 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4154 size = (end - start) * sizeof(struct page);
4155 map = alloc_remap(pgdat->node_id, size);
4157 map = alloc_bootmem_node(pgdat, size);
4158 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4160 #ifndef CONFIG_NEED_MULTIPLE_NODES
4162 * With no DISCONTIG, the global mem_map is just set as node 0's
4164 if (pgdat == NODE_DATA(0)) {
4165 mem_map = NODE_DATA(0)->node_mem_map;
4166 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4167 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4168 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4169 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4172 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4175 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4176 unsigned long node_start_pfn, unsigned long *zholes_size)
4178 pg_data_t *pgdat = NODE_DATA(nid);
4180 pgdat->node_id = nid;
4181 pgdat->node_start_pfn = node_start_pfn;
4182 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4184 alloc_node_mem_map(pgdat);
4185 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4186 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4187 nid, (unsigned long)pgdat,
4188 (unsigned long)pgdat->node_mem_map);
4191 free_area_init_core(pgdat, zones_size, zholes_size);
4194 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4196 #if MAX_NUMNODES > 1
4198 * Figure out the number of possible node ids.
4200 static void __init setup_nr_node_ids(void)
4203 unsigned int highest = 0;
4205 for_each_node_mask(node, node_possible_map)
4207 nr_node_ids = highest + 1;
4210 static inline void setup_nr_node_ids(void)
4216 * add_active_range - Register a range of PFNs backed by physical memory
4217 * @nid: The node ID the range resides on
4218 * @start_pfn: The start PFN of the available physical memory
4219 * @end_pfn: The end PFN of the available physical memory
4221 * These ranges are stored in an early_node_map[] and later used by
4222 * free_area_init_nodes() to calculate zone sizes and holes. If the
4223 * range spans a memory hole, it is up to the architecture to ensure
4224 * the memory is not freed by the bootmem allocator. If possible
4225 * the range being registered will be merged with existing ranges.
4227 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4228 unsigned long end_pfn)
4232 mminit_dprintk(MMINIT_TRACE, "memory_register",
4233 "Entering add_active_range(%d, %#lx, %#lx) "
4234 "%d entries of %d used\n",
4235 nid, start_pfn, end_pfn,
4236 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4238 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4240 /* Merge with existing active regions if possible */
4241 for (i = 0; i < nr_nodemap_entries; i++) {
4242 if (early_node_map[i].nid != nid)
4245 /* Skip if an existing region covers this new one */
4246 if (start_pfn >= early_node_map[i].start_pfn &&
4247 end_pfn <= early_node_map[i].end_pfn)
4250 /* Merge forward if suitable */
4251 if (start_pfn <= early_node_map[i].end_pfn &&
4252 end_pfn > early_node_map[i].end_pfn) {
4253 early_node_map[i].end_pfn = end_pfn;
4257 /* Merge backward if suitable */
4258 if (start_pfn < early_node_map[i].start_pfn &&
4259 end_pfn >= early_node_map[i].start_pfn) {
4260 early_node_map[i].start_pfn = start_pfn;
4265 /* Check that early_node_map is large enough */
4266 if (i >= MAX_ACTIVE_REGIONS) {
4267 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4268 MAX_ACTIVE_REGIONS);
4272 early_node_map[i].nid = nid;
4273 early_node_map[i].start_pfn = start_pfn;
4274 early_node_map[i].end_pfn = end_pfn;
4275 nr_nodemap_entries = i + 1;
4279 * remove_active_range - Shrink an existing registered range of PFNs
4280 * @nid: The node id the range is on that should be shrunk
4281 * @start_pfn: The new PFN of the range
4282 * @end_pfn: The new PFN of the range
4284 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4285 * The map is kept near the end physical page range that has already been
4286 * registered. This function allows an arch to shrink an existing registered
4289 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4290 unsigned long end_pfn)
4295 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4296 nid, start_pfn, end_pfn);
4298 /* Find the old active region end and shrink */
4299 for_each_active_range_index_in_nid(i, nid) {
4300 if (early_node_map[i].start_pfn >= start_pfn &&
4301 early_node_map[i].end_pfn <= end_pfn) {
4303 early_node_map[i].start_pfn = 0;
4304 early_node_map[i].end_pfn = 0;
4308 if (early_node_map[i].start_pfn < start_pfn &&
4309 early_node_map[i].end_pfn > start_pfn) {
4310 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4311 early_node_map[i].end_pfn = start_pfn;
4312 if (temp_end_pfn > end_pfn)
4313 add_active_range(nid, end_pfn, temp_end_pfn);
4316 if (early_node_map[i].start_pfn >= start_pfn &&
4317 early_node_map[i].end_pfn > end_pfn &&
4318 early_node_map[i].start_pfn < end_pfn) {
4319 early_node_map[i].start_pfn = end_pfn;
4327 /* remove the blank ones */
4328 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4329 if (early_node_map[i].nid != nid)
4331 if (early_node_map[i].end_pfn)
4333 /* we found it, get rid of it */
4334 for (j = i; j < nr_nodemap_entries - 1; j++)
4335 memcpy(&early_node_map[j], &early_node_map[j+1],
4336 sizeof(early_node_map[j]));
4337 j = nr_nodemap_entries - 1;
4338 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4339 nr_nodemap_entries--;
4344 * remove_all_active_ranges - Remove all currently registered regions
4346 * During discovery, it may be found that a table like SRAT is invalid
4347 * and an alternative discovery method must be used. This function removes
4348 * all currently registered regions.
4350 void __init remove_all_active_ranges(void)
4352 memset(early_node_map, 0, sizeof(early_node_map));
4353 nr_nodemap_entries = 0;
4356 /* Compare two active node_active_regions */
4357 static int __init cmp_node_active_region(const void *a, const void *b)
4359 struct node_active_region *arange = (struct node_active_region *)a;
4360 struct node_active_region *brange = (struct node_active_region *)b;
4362 /* Done this way to avoid overflows */
4363 if (arange->start_pfn > brange->start_pfn)
4365 if (arange->start_pfn < brange->start_pfn)
4371 /* sort the node_map by start_pfn */
4372 void __init sort_node_map(void)
4374 sort(early_node_map, (size_t)nr_nodemap_entries,
4375 sizeof(struct node_active_region),
4376 cmp_node_active_region, NULL);
4379 /* Find the lowest pfn for a node */
4380 static unsigned long __init find_min_pfn_for_node(int nid)
4383 unsigned long min_pfn = ULONG_MAX;
4385 /* Assuming a sorted map, the first range found has the starting pfn */
4386 for_each_active_range_index_in_nid(i, nid)
4387 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4389 if (min_pfn == ULONG_MAX) {
4391 "Could not find start_pfn for node %d\n", nid);
4399 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4401 * It returns the minimum PFN based on information provided via
4402 * add_active_range().
4404 unsigned long __init find_min_pfn_with_active_regions(void)
4406 return find_min_pfn_for_node(MAX_NUMNODES);
4410 * early_calculate_totalpages()
4411 * Sum pages in active regions for movable zone.
4412 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4414 static unsigned long __init early_calculate_totalpages(void)
4417 unsigned long totalpages = 0;
4419 for (i = 0; i < nr_nodemap_entries; i++) {
4420 unsigned long pages = early_node_map[i].end_pfn -
4421 early_node_map[i].start_pfn;
4422 totalpages += pages;
4424 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4430 * Find the PFN the Movable zone begins in each node. Kernel memory
4431 * is spread evenly between nodes as long as the nodes have enough
4432 * memory. When they don't, some nodes will have more kernelcore than
4435 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4438 unsigned long usable_startpfn;
4439 unsigned long kernelcore_node, kernelcore_remaining;
4440 /* save the state before borrow the nodemask */
4441 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4442 unsigned long totalpages = early_calculate_totalpages();
4443 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4446 * If movablecore was specified, calculate what size of
4447 * kernelcore that corresponds so that memory usable for
4448 * any allocation type is evenly spread. If both kernelcore
4449 * and movablecore are specified, then the value of kernelcore
4450 * will be used for required_kernelcore if it's greater than
4451 * what movablecore would have allowed.
4453 if (required_movablecore) {
4454 unsigned long corepages;
4457 * Round-up so that ZONE_MOVABLE is at least as large as what
4458 * was requested by the user
4460 required_movablecore =
4461 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4462 corepages = totalpages - required_movablecore;
4464 required_kernelcore = max(required_kernelcore, corepages);
4467 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4468 if (!required_kernelcore)
4471 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4472 find_usable_zone_for_movable();
4473 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4476 /* Spread kernelcore memory as evenly as possible throughout nodes */
4477 kernelcore_node = required_kernelcore / usable_nodes;
4478 for_each_node_state(nid, N_HIGH_MEMORY) {
4480 * Recalculate kernelcore_node if the division per node
4481 * now exceeds what is necessary to satisfy the requested
4482 * amount of memory for the kernel
4484 if (required_kernelcore < kernelcore_node)
4485 kernelcore_node = required_kernelcore / usable_nodes;
4488 * As the map is walked, we track how much memory is usable
4489 * by the kernel using kernelcore_remaining. When it is
4490 * 0, the rest of the node is usable by ZONE_MOVABLE
4492 kernelcore_remaining = kernelcore_node;
4494 /* Go through each range of PFNs within this node */
4495 for_each_active_range_index_in_nid(i, nid) {
4496 unsigned long start_pfn, end_pfn;
4497 unsigned long size_pages;
4499 start_pfn = max(early_node_map[i].start_pfn,
4500 zone_movable_pfn[nid]);
4501 end_pfn = early_node_map[i].end_pfn;
4502 if (start_pfn >= end_pfn)
4505 /* Account for what is only usable for kernelcore */
4506 if (start_pfn < usable_startpfn) {
4507 unsigned long kernel_pages;
4508 kernel_pages = min(end_pfn, usable_startpfn)
4511 kernelcore_remaining -= min(kernel_pages,
4512 kernelcore_remaining);
4513 required_kernelcore -= min(kernel_pages,
4514 required_kernelcore);
4516 /* Continue if range is now fully accounted */
4517 if (end_pfn <= usable_startpfn) {
4520 * Push zone_movable_pfn to the end so
4521 * that if we have to rebalance
4522 * kernelcore across nodes, we will
4523 * not double account here
4525 zone_movable_pfn[nid] = end_pfn;
4528 start_pfn = usable_startpfn;
4532 * The usable PFN range for ZONE_MOVABLE is from
4533 * start_pfn->end_pfn. Calculate size_pages as the
4534 * number of pages used as kernelcore
4536 size_pages = end_pfn - start_pfn;
4537 if (size_pages > kernelcore_remaining)
4538 size_pages = kernelcore_remaining;
4539 zone_movable_pfn[nid] = start_pfn + size_pages;
4542 * Some kernelcore has been met, update counts and
4543 * break if the kernelcore for this node has been
4546 required_kernelcore -= min(required_kernelcore,
4548 kernelcore_remaining -= size_pages;
4549 if (!kernelcore_remaining)
4555 * If there is still required_kernelcore, we do another pass with one
4556 * less node in the count. This will push zone_movable_pfn[nid] further
4557 * along on the nodes that still have memory until kernelcore is
4561 if (usable_nodes && required_kernelcore > usable_nodes)
4564 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4565 for (nid = 0; nid < MAX_NUMNODES; nid++)
4566 zone_movable_pfn[nid] =
4567 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4570 /* restore the node_state */
4571 node_states[N_HIGH_MEMORY] = saved_node_state;
4574 /* Any regular memory on that node ? */
4575 static void check_for_regular_memory(pg_data_t *pgdat)
4577 #ifdef CONFIG_HIGHMEM
4578 enum zone_type zone_type;
4580 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4581 struct zone *zone = &pgdat->node_zones[zone_type];
4582 if (zone->present_pages)
4583 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4589 * free_area_init_nodes - Initialise all pg_data_t and zone data
4590 * @max_zone_pfn: an array of max PFNs for each zone
4592 * This will call free_area_init_node() for each active node in the system.
4593 * Using the page ranges provided by add_active_range(), the size of each
4594 * zone in each node and their holes is calculated. If the maximum PFN
4595 * between two adjacent zones match, it is assumed that the zone is empty.
4596 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4597 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4598 * starts where the previous one ended. For example, ZONE_DMA32 starts
4599 * at arch_max_dma_pfn.
4601 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4606 /* Sort early_node_map as initialisation assumes it is sorted */
4609 /* Record where the zone boundaries are */
4610 memset(arch_zone_lowest_possible_pfn, 0,
4611 sizeof(arch_zone_lowest_possible_pfn));
4612 memset(arch_zone_highest_possible_pfn, 0,
4613 sizeof(arch_zone_highest_possible_pfn));
4614 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4615 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4616 for (i = 1; i < MAX_NR_ZONES; i++) {
4617 if (i == ZONE_MOVABLE)
4619 arch_zone_lowest_possible_pfn[i] =
4620 arch_zone_highest_possible_pfn[i-1];
4621 arch_zone_highest_possible_pfn[i] =
4622 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4624 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4625 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4627 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4628 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4629 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4631 /* Print out the zone ranges */
4632 printk("Zone PFN ranges:\n");
4633 for (i = 0; i < MAX_NR_ZONES; i++) {
4634 if (i == ZONE_MOVABLE)
4636 printk(" %-8s ", zone_names[i]);
4637 if (arch_zone_lowest_possible_pfn[i] ==
4638 arch_zone_highest_possible_pfn[i])
4641 printk("%0#10lx -> %0#10lx\n",
4642 arch_zone_lowest_possible_pfn[i],
4643 arch_zone_highest_possible_pfn[i]);
4646 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4647 printk("Movable zone start PFN for each node\n");
4648 for (i = 0; i < MAX_NUMNODES; i++) {
4649 if (zone_movable_pfn[i])
4650 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4653 /* Print out the early_node_map[] */
4654 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4655 for (i = 0; i < nr_nodemap_entries; i++)
4656 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4657 early_node_map[i].start_pfn,
4658 early_node_map[i].end_pfn);
4660 /* Initialise every node */
4661 mminit_verify_pageflags_layout();
4662 setup_nr_node_ids();
4663 for_each_online_node(nid) {
4664 pg_data_t *pgdat = NODE_DATA(nid);
4665 free_area_init_node(nid, NULL,
4666 find_min_pfn_for_node(nid), NULL);
4668 /* Any memory on that node */
4669 if (pgdat->node_present_pages)
4670 node_set_state(nid, N_HIGH_MEMORY);
4671 check_for_regular_memory(pgdat);
4675 static int __init cmdline_parse_core(char *p, unsigned long *core)
4677 unsigned long long coremem;
4681 coremem = memparse(p, &p);
4682 *core = coremem >> PAGE_SHIFT;
4684 /* Paranoid check that UL is enough for the coremem value */
4685 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4691 * kernelcore=size sets the amount of memory for use for allocations that
4692 * cannot be reclaimed or migrated.
4694 static int __init cmdline_parse_kernelcore(char *p)
4696 return cmdline_parse_core(p, &required_kernelcore);
4700 * movablecore=size sets the amount of memory for use for allocations that
4701 * can be reclaimed or migrated.
4703 static int __init cmdline_parse_movablecore(char *p)
4705 return cmdline_parse_core(p, &required_movablecore);
4708 early_param("kernelcore", cmdline_parse_kernelcore);
4709 early_param("movablecore", cmdline_parse_movablecore);
4711 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4714 * set_dma_reserve - set the specified number of pages reserved in the first zone
4715 * @new_dma_reserve: The number of pages to mark reserved
4717 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4718 * In the DMA zone, a significant percentage may be consumed by kernel image
4719 * and other unfreeable allocations which can skew the watermarks badly. This
4720 * function may optionally be used to account for unfreeable pages in the
4721 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4722 * smaller per-cpu batchsize.
4724 void __init set_dma_reserve(unsigned long new_dma_reserve)
4726 dma_reserve = new_dma_reserve;
4729 #ifndef CONFIG_NEED_MULTIPLE_NODES
4730 struct pglist_data __refdata contig_page_data = {
4731 #ifndef CONFIG_NO_BOOTMEM
4732 .bdata = &bootmem_node_data[0]
4735 EXPORT_SYMBOL(contig_page_data);
4738 void __init free_area_init(unsigned long *zones_size)
4740 free_area_init_node(0, zones_size,
4741 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4744 static int page_alloc_cpu_notify(struct notifier_block *self,
4745 unsigned long action, void *hcpu)
4747 int cpu = (unsigned long)hcpu;
4749 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4753 * Spill the event counters of the dead processor
4754 * into the current processors event counters.
4755 * This artificially elevates the count of the current
4758 vm_events_fold_cpu(cpu);
4761 * Zero the differential counters of the dead processor
4762 * so that the vm statistics are consistent.
4764 * This is only okay since the processor is dead and cannot
4765 * race with what we are doing.
4767 refresh_cpu_vm_stats(cpu);
4772 void __init page_alloc_init(void)
4774 hotcpu_notifier(page_alloc_cpu_notify, 0);
4778 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4779 * or min_free_kbytes changes.
4781 static void calculate_totalreserve_pages(void)
4783 struct pglist_data *pgdat;
4784 unsigned long reserve_pages = 0;
4785 enum zone_type i, j;
4787 for_each_online_pgdat(pgdat) {
4788 for (i = 0; i < MAX_NR_ZONES; i++) {
4789 struct zone *zone = pgdat->node_zones + i;
4790 unsigned long max = 0;
4792 /* Find valid and maximum lowmem_reserve in the zone */
4793 for (j = i; j < MAX_NR_ZONES; j++) {
4794 if (zone->lowmem_reserve[j] > max)
4795 max = zone->lowmem_reserve[j];
4798 /* we treat the high watermark as reserved pages. */
4799 max += high_wmark_pages(zone);
4801 if (max > zone->present_pages)
4802 max = zone->present_pages;
4803 reserve_pages += max;
4806 totalreserve_pages = reserve_pages;
4810 * setup_per_zone_lowmem_reserve - called whenever
4811 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4812 * has a correct pages reserved value, so an adequate number of
4813 * pages are left in the zone after a successful __alloc_pages().
4815 static void setup_per_zone_lowmem_reserve(void)
4817 struct pglist_data *pgdat;
4818 enum zone_type j, idx;
4820 for_each_online_pgdat(pgdat) {
4821 for (j = 0; j < MAX_NR_ZONES; j++) {
4822 struct zone *zone = pgdat->node_zones + j;
4823 unsigned long present_pages = zone->present_pages;
4825 zone->lowmem_reserve[j] = 0;
4829 struct zone *lower_zone;
4833 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4834 sysctl_lowmem_reserve_ratio[idx] = 1;
4836 lower_zone = pgdat->node_zones + idx;
4837 lower_zone->lowmem_reserve[j] = present_pages /
4838 sysctl_lowmem_reserve_ratio[idx];
4839 present_pages += lower_zone->present_pages;
4844 /* update totalreserve_pages */
4845 calculate_totalreserve_pages();
4849 * setup_per_zone_wmarks - called when min_free_kbytes changes
4850 * or when memory is hot-{added|removed}
4852 * Ensures that the watermark[min,low,high] values for each zone are set
4853 * correctly with respect to min_free_kbytes.
4855 void setup_per_zone_wmarks(void)
4857 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4858 unsigned long lowmem_pages = 0;
4860 unsigned long flags;
4862 /* Calculate total number of !ZONE_HIGHMEM pages */
4863 for_each_zone(zone) {
4864 if (!is_highmem(zone))
4865 lowmem_pages += zone->present_pages;
4868 for_each_zone(zone) {
4871 spin_lock_irqsave(&zone->lock, flags);
4872 tmp = (u64)pages_min * zone->present_pages;
4873 do_div(tmp, lowmem_pages);
4874 if (is_highmem(zone)) {
4876 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4877 * need highmem pages, so cap pages_min to a small
4880 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4881 * deltas controls asynch page reclaim, and so should
4882 * not be capped for highmem.
4886 min_pages = zone->present_pages / 1024;
4887 if (min_pages < SWAP_CLUSTER_MAX)
4888 min_pages = SWAP_CLUSTER_MAX;
4889 if (min_pages > 128)
4891 zone->watermark[WMARK_MIN] = min_pages;
4894 * If it's a lowmem zone, reserve a number of pages
4895 * proportionate to the zone's size.
4897 zone->watermark[WMARK_MIN] = tmp;
4900 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4901 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4902 setup_zone_migrate_reserve(zone);
4903 spin_unlock_irqrestore(&zone->lock, flags);
4906 /* update totalreserve_pages */
4907 calculate_totalreserve_pages();
4911 * The inactive anon list should be small enough that the VM never has to
4912 * do too much work, but large enough that each inactive page has a chance
4913 * to be referenced again before it is swapped out.
4915 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4916 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4917 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4918 * the anonymous pages are kept on the inactive list.
4921 * memory ratio inactive anon
4922 * -------------------------------------
4931 void calculate_zone_inactive_ratio(struct zone *zone)
4933 unsigned int gb, ratio;
4935 /* Zone size in gigabytes */
4936 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4938 ratio = int_sqrt(10 * gb);
4942 zone->inactive_ratio = ratio;
4945 static void __init setup_per_zone_inactive_ratio(void)
4950 calculate_zone_inactive_ratio(zone);
4954 * Initialise min_free_kbytes.
4956 * For small machines we want it small (128k min). For large machines
4957 * we want it large (64MB max). But it is not linear, because network
4958 * bandwidth does not increase linearly with machine size. We use
4960 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4961 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4977 static int __init init_per_zone_wmark_min(void)
4979 unsigned long lowmem_kbytes;
4981 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4983 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4984 if (min_free_kbytes < 128)
4985 min_free_kbytes = 128;
4986 if (min_free_kbytes > 65536)
4987 min_free_kbytes = 65536;
4988 setup_per_zone_wmarks();
4989 setup_per_zone_lowmem_reserve();
4990 setup_per_zone_inactive_ratio();
4993 module_init(init_per_zone_wmark_min)
4996 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4997 * that we can call two helper functions whenever min_free_kbytes
5000 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5001 void __user *buffer, size_t *length, loff_t *ppos)
5003 proc_dointvec(table, write, buffer, length, ppos);
5005 setup_per_zone_wmarks();
5010 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5011 void __user *buffer, size_t *length, loff_t *ppos)
5016 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5021 zone->min_unmapped_pages = (zone->present_pages *
5022 sysctl_min_unmapped_ratio) / 100;
5026 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5027 void __user *buffer, size_t *length, loff_t *ppos)
5032 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5037 zone->min_slab_pages = (zone->present_pages *
5038 sysctl_min_slab_ratio) / 100;
5044 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5045 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5046 * whenever sysctl_lowmem_reserve_ratio changes.
5048 * The reserve ratio obviously has absolutely no relation with the
5049 * minimum watermarks. The lowmem reserve ratio can only make sense
5050 * if in function of the boot time zone sizes.
5052 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5053 void __user *buffer, size_t *length, loff_t *ppos)
5055 proc_dointvec_minmax(table, write, buffer, length, ppos);
5056 setup_per_zone_lowmem_reserve();
5061 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5062 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5063 * can have before it gets flushed back to buddy allocator.
5066 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5067 void __user *buffer, size_t *length, loff_t *ppos)
5073 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5074 if (!write || (ret == -EINVAL))
5076 for_each_populated_zone(zone) {
5077 for_each_possible_cpu(cpu) {
5079 high = zone->present_pages / percpu_pagelist_fraction;
5080 setup_pagelist_highmark(
5081 per_cpu_ptr(zone->pageset, cpu), high);
5087 int hashdist = HASHDIST_DEFAULT;
5090 static int __init set_hashdist(char *str)
5094 hashdist = simple_strtoul(str, &str, 0);
5097 __setup("hashdist=", set_hashdist);
5101 * allocate a large system hash table from bootmem
5102 * - it is assumed that the hash table must contain an exact power-of-2
5103 * quantity of entries
5104 * - limit is the number of hash buckets, not the total allocation size
5106 void *__init alloc_large_system_hash(const char *tablename,
5107 unsigned long bucketsize,
5108 unsigned long numentries,
5111 unsigned int *_hash_shift,
5112 unsigned int *_hash_mask,
5113 unsigned long limit)
5115 unsigned long long max = limit;
5116 unsigned long log2qty, size;
5119 /* allow the kernel cmdline to have a say */
5121 /* round applicable memory size up to nearest megabyte */
5122 numentries = nr_kernel_pages;
5123 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5124 numentries >>= 20 - PAGE_SHIFT;
5125 numentries <<= 20 - PAGE_SHIFT;
5127 /* limit to 1 bucket per 2^scale bytes of low memory */
5128 if (scale > PAGE_SHIFT)
5129 numentries >>= (scale - PAGE_SHIFT);
5131 numentries <<= (PAGE_SHIFT - scale);
5133 /* Make sure we've got at least a 0-order allocation.. */
5134 if (unlikely(flags & HASH_SMALL)) {
5135 /* Makes no sense without HASH_EARLY */
5136 WARN_ON(!(flags & HASH_EARLY));
5137 if (!(numentries >> *_hash_shift)) {
5138 numentries = 1UL << *_hash_shift;
5139 BUG_ON(!numentries);
5141 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5142 numentries = PAGE_SIZE / bucketsize;
5144 numentries = roundup_pow_of_two(numentries);
5146 /* limit allocation size to 1/16 total memory by default */
5148 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5149 do_div(max, bucketsize);
5152 if (numentries > max)
5155 log2qty = ilog2(numentries);
5158 size = bucketsize << log2qty;
5159 if (flags & HASH_EARLY)
5160 table = alloc_bootmem_nopanic(size);
5162 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5165 * If bucketsize is not a power-of-two, we may free
5166 * some pages at the end of hash table which
5167 * alloc_pages_exact() automatically does
5169 if (get_order(size) < MAX_ORDER) {
5170 table = alloc_pages_exact(size, GFP_ATOMIC);
5171 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5174 } while (!table && size > PAGE_SIZE && --log2qty);
5177 panic("Failed to allocate %s hash table\n", tablename);
5179 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
5182 ilog2(size) - PAGE_SHIFT,
5186 *_hash_shift = log2qty;
5188 *_hash_mask = (1 << log2qty) - 1;
5193 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5194 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5197 #ifdef CONFIG_SPARSEMEM
5198 return __pfn_to_section(pfn)->pageblock_flags;
5200 return zone->pageblock_flags;
5201 #endif /* CONFIG_SPARSEMEM */
5204 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5206 #ifdef CONFIG_SPARSEMEM
5207 pfn &= (PAGES_PER_SECTION-1);
5208 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5210 pfn = pfn - zone->zone_start_pfn;
5211 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5212 #endif /* CONFIG_SPARSEMEM */
5216 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5217 * @page: The page within the block of interest
5218 * @start_bitidx: The first bit of interest to retrieve
5219 * @end_bitidx: The last bit of interest
5220 * returns pageblock_bits flags
5222 unsigned long get_pageblock_flags_group(struct page *page,
5223 int start_bitidx, int end_bitidx)
5226 unsigned long *bitmap;
5227 unsigned long pfn, bitidx;
5228 unsigned long flags = 0;
5229 unsigned long value = 1;
5231 zone = page_zone(page);
5232 pfn = page_to_pfn(page);
5233 bitmap = get_pageblock_bitmap(zone, pfn);
5234 bitidx = pfn_to_bitidx(zone, pfn);
5236 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5237 if (test_bit(bitidx + start_bitidx, bitmap))
5244 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5245 * @page: The page within the block of interest
5246 * @start_bitidx: The first bit of interest
5247 * @end_bitidx: The last bit of interest
5248 * @flags: The flags to set
5250 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5251 int start_bitidx, int end_bitidx)
5254 unsigned long *bitmap;
5255 unsigned long pfn, bitidx;
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);
5262 VM_BUG_ON(pfn < zone->zone_start_pfn);
5263 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5265 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5267 __set_bit(bitidx + start_bitidx, bitmap);
5269 __clear_bit(bitidx + start_bitidx, bitmap);
5273 * This is designed as sub function...plz see page_isolation.c also.
5274 * set/clear page block's type to be ISOLATE.
5275 * page allocater never alloc memory from ISOLATE block.
5278 int set_migratetype_isolate(struct page *page)
5281 struct page *curr_page;
5282 unsigned long flags, pfn, iter;
5283 unsigned long immobile = 0;
5284 struct memory_isolate_notify arg;
5289 zone = page_zone(page);
5290 zone_idx = zone_idx(zone);
5292 spin_lock_irqsave(&zone->lock, flags);
5293 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE ||
5294 zone_idx == ZONE_MOVABLE) {
5299 pfn = page_to_pfn(page);
5300 arg.start_pfn = pfn;
5301 arg.nr_pages = pageblock_nr_pages;
5302 arg.pages_found = 0;
5305 * It may be possible to isolate a pageblock even if the
5306 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5307 * notifier chain is used by balloon drivers to return the
5308 * number of pages in a range that are held by the balloon
5309 * driver to shrink memory. If all the pages are accounted for
5310 * by balloons, are free, or on the LRU, isolation can continue.
5311 * Later, for example, when memory hotplug notifier runs, these
5312 * pages reported as "can be isolated" should be isolated(freed)
5313 * by the balloon driver through the memory notifier chain.
5315 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5316 notifier_ret = notifier_to_errno(notifier_ret);
5317 if (notifier_ret || !arg.pages_found)
5320 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) {
5321 if (!pfn_valid_within(pfn))
5324 curr_page = pfn_to_page(iter);
5325 if (!page_count(curr_page) || PageLRU(curr_page))
5331 if (arg.pages_found == immobile)
5336 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5337 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5340 spin_unlock_irqrestore(&zone->lock, flags);
5346 void unset_migratetype_isolate(struct page *page)
5349 unsigned long flags;
5350 zone = page_zone(page);
5351 spin_lock_irqsave(&zone->lock, flags);
5352 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5354 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5355 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5357 spin_unlock_irqrestore(&zone->lock, flags);
5360 #ifdef CONFIG_MEMORY_HOTREMOVE
5362 * All pages in the range must be isolated before calling this.
5365 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5371 unsigned long flags;
5372 /* find the first valid pfn */
5373 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5378 zone = page_zone(pfn_to_page(pfn));
5379 spin_lock_irqsave(&zone->lock, flags);
5381 while (pfn < end_pfn) {
5382 if (!pfn_valid(pfn)) {
5386 page = pfn_to_page(pfn);
5387 BUG_ON(page_count(page));
5388 BUG_ON(!PageBuddy(page));
5389 order = page_order(page);
5390 #ifdef CONFIG_DEBUG_VM
5391 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5392 pfn, 1 << order, end_pfn);
5394 list_del(&page->lru);
5395 rmv_page_order(page);
5396 zone->free_area[order].nr_free--;
5397 __mod_zone_page_state(zone, NR_FREE_PAGES,
5399 for (i = 0; i < (1 << order); i++)
5400 SetPageReserved((page+i));
5401 pfn += (1 << order);
5403 spin_unlock_irqrestore(&zone->lock, flags);
5407 #ifdef CONFIG_MEMORY_FAILURE
5408 bool is_free_buddy_page(struct page *page)
5410 struct zone *zone = page_zone(page);
5411 unsigned long pfn = page_to_pfn(page);
5412 unsigned long flags;
5415 spin_lock_irqsave(&zone->lock, flags);
5416 for (order = 0; order < MAX_ORDER; order++) {
5417 struct page *page_head = page - (pfn & ((1 << order) - 1));
5419 if (PageBuddy(page_head) && page_order(page_head) >= order)
5422 spin_unlock_irqrestore(&zone->lock, flags);
5424 return order < MAX_ORDER;
5428 static struct trace_print_flags pageflag_names[] = {
5429 {1UL << PG_locked, "locked" },
5430 {1UL << PG_error, "error" },
5431 {1UL << PG_referenced, "referenced" },
5432 {1UL << PG_uptodate, "uptodate" },
5433 {1UL << PG_dirty, "dirty" },
5434 {1UL << PG_lru, "lru" },
5435 {1UL << PG_active, "active" },
5436 {1UL << PG_slab, "slab" },
5437 {1UL << PG_owner_priv_1, "owner_priv_1" },
5438 {1UL << PG_arch_1, "arch_1" },
5439 {1UL << PG_reserved, "reserved" },
5440 {1UL << PG_private, "private" },
5441 {1UL << PG_private_2, "private_2" },
5442 {1UL << PG_writeback, "writeback" },
5443 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5444 {1UL << PG_head, "head" },
5445 {1UL << PG_tail, "tail" },
5447 {1UL << PG_compound, "compound" },
5449 {1UL << PG_swapcache, "swapcache" },
5450 {1UL << PG_mappedtodisk, "mappedtodisk" },
5451 {1UL << PG_reclaim, "reclaim" },
5452 {1UL << PG_buddy, "buddy" },
5453 {1UL << PG_swapbacked, "swapbacked" },
5454 {1UL << PG_unevictable, "unevictable" },
5456 {1UL << PG_mlocked, "mlocked" },
5458 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5459 {1UL << PG_uncached, "uncached" },
5461 #ifdef CONFIG_MEMORY_FAILURE
5462 {1UL << PG_hwpoison, "hwpoison" },
5467 static void dump_page_flags(unsigned long flags)
5469 const char *delim = "";
5473 printk(KERN_ALERT "page flags: %#lx(", flags);
5475 /* remove zone id */
5476 flags &= (1UL << NR_PAGEFLAGS) - 1;
5478 for (i = 0; pageflag_names[i].name && flags; i++) {
5480 mask = pageflag_names[i].mask;
5481 if ((flags & mask) != mask)
5485 printk("%s%s", delim, pageflag_names[i].name);
5489 /* check for left over flags */
5491 printk("%s%#lx", delim, flags);
5496 void dump_page(struct page *page)
5499 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5500 page, page_count(page), page_mapcount(page),
5501 page->mapping, page->index);
5502 dump_page_flags(page->flags);