2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/memory.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/memcontrol.h>
59 #include <linux/prefetch.h>
61 #include <asm/tlbflush.h>
62 #include <asm/div64.h>
65 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
66 DEFINE_PER_CPU(int, numa_node);
67 EXPORT_PER_CPU_SYMBOL(numa_node);
70 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
72 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
73 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
74 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
75 * defined in <linux/topology.h>.
77 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
78 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
82 * Array of node states.
84 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
85 [N_POSSIBLE] = NODE_MASK_ALL,
86 [N_ONLINE] = { { [0] = 1UL } },
88 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
90 [N_HIGH_MEMORY] = { { [0] = 1UL } },
92 [N_CPU] = { { [0] = 1UL } },
95 EXPORT_SYMBOL(node_states);
97 unsigned long totalram_pages __read_mostly;
98 unsigned long totalreserve_pages __read_mostly;
99 int percpu_pagelist_fraction;
100 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
102 #ifdef CONFIG_PM_SLEEP
104 * The following functions are used by the suspend/hibernate code to temporarily
105 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
106 * while devices are suspended. To avoid races with the suspend/hibernate code,
107 * they should always be called with pm_mutex held (gfp_allowed_mask also should
108 * only be modified with pm_mutex held, unless the suspend/hibernate code is
109 * guaranteed not to run in parallel with that modification).
112 static gfp_t saved_gfp_mask;
114 void pm_restore_gfp_mask(void)
116 WARN_ON(!mutex_is_locked(&pm_mutex));
117 if (saved_gfp_mask) {
118 gfp_allowed_mask = saved_gfp_mask;
123 void pm_restrict_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 WARN_ON(saved_gfp_mask);
127 saved_gfp_mask = gfp_allowed_mask;
128 gfp_allowed_mask &= ~GFP_IOFS;
131 bool pm_suspended_storage(void)
133 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
137 #endif /* CONFIG_PM_SLEEP */
139 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
140 int pageblock_order __read_mostly;
143 static void __free_pages_ok(struct page *page, unsigned int order);
146 * results with 256, 32 in the lowmem_reserve sysctl:
147 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
148 * 1G machine -> (16M dma, 784M normal, 224M high)
149 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
150 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
151 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
153 * TBD: should special case ZONE_DMA32 machines here - in those we normally
154 * don't need any ZONE_NORMAL reservation
156 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
157 #ifdef CONFIG_ZONE_DMA
160 #ifdef CONFIG_ZONE_DMA32
163 #ifdef CONFIG_HIGHMEM
169 EXPORT_SYMBOL(totalram_pages);
171 static char * const zone_names[MAX_NR_ZONES] = {
172 #ifdef CONFIG_ZONE_DMA
175 #ifdef CONFIG_ZONE_DMA32
179 #ifdef CONFIG_HIGHMEM
185 int min_free_kbytes = 1024;
187 static unsigned long __meminitdata nr_kernel_pages;
188 static unsigned long __meminitdata nr_all_pages;
189 static unsigned long __meminitdata dma_reserve;
191 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
193 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
194 * ranges of memory (RAM) that may be registered with add_active_range().
195 * Ranges passed to add_active_range() will be merged if possible
196 * so the number of times add_active_range() can be called is
197 * related to the number of nodes and the number of holes
199 #ifdef CONFIG_MAX_ACTIVE_REGIONS
200 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
201 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
203 #if MAX_NUMNODES >= 32
204 /* If there can be many nodes, allow up to 50 holes per node */
205 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
207 /* By default, allow up to 256 distinct regions */
208 #define MAX_ACTIVE_REGIONS 256
212 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
213 static int __meminitdata nr_nodemap_entries;
214 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
215 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
216 static unsigned long __initdata required_kernelcore;
217 static unsigned long __initdata required_movablecore;
218 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
220 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
222 EXPORT_SYMBOL(movable_zone);
223 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
226 int nr_node_ids __read_mostly = MAX_NUMNODES;
227 int nr_online_nodes __read_mostly = 1;
228 EXPORT_SYMBOL(nr_node_ids);
229 EXPORT_SYMBOL(nr_online_nodes);
232 int page_group_by_mobility_disabled __read_mostly;
234 static void set_pageblock_migratetype(struct page *page, int migratetype)
237 if (unlikely(page_group_by_mobility_disabled))
238 migratetype = MIGRATE_UNMOVABLE;
240 set_pageblock_flags_group(page, (unsigned long)migratetype,
241 PB_migrate, PB_migrate_end);
244 bool oom_killer_disabled __read_mostly;
246 #ifdef CONFIG_DEBUG_VM
247 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
251 unsigned long pfn = page_to_pfn(page);
254 seq = zone_span_seqbegin(zone);
255 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
257 else if (pfn < zone->zone_start_pfn)
259 } while (zone_span_seqretry(zone, seq));
264 static int page_is_consistent(struct zone *zone, struct page *page)
266 if (!pfn_valid_within(page_to_pfn(page)))
268 if (zone != page_zone(page))
274 * Temporary debugging check for pages not lying within a given zone.
276 static int bad_range(struct zone *zone, struct page *page)
278 if (page_outside_zone_boundaries(zone, page))
280 if (!page_is_consistent(zone, page))
286 static inline int bad_range(struct zone *zone, struct page *page)
292 static void bad_page(struct page *page)
294 static unsigned long resume;
295 static unsigned long nr_shown;
296 static unsigned long nr_unshown;
298 /* Don't complain about poisoned pages */
299 if (PageHWPoison(page)) {
300 reset_page_mapcount(page); /* remove PageBuddy */
305 * Allow a burst of 60 reports, then keep quiet for that minute;
306 * or allow a steady drip of one report per second.
308 if (nr_shown == 60) {
309 if (time_before(jiffies, resume)) {
315 "BUG: Bad page state: %lu messages suppressed\n",
322 resume = jiffies + 60 * HZ;
324 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
325 current->comm, page_to_pfn(page));
331 /* Leave bad fields for debug, except PageBuddy could make trouble */
332 reset_page_mapcount(page); /* remove PageBuddy */
333 add_taint(TAINT_BAD_PAGE);
337 * Higher-order pages are called "compound pages". They are structured thusly:
339 * The first PAGE_SIZE page is called the "head page".
341 * The remaining PAGE_SIZE pages are called "tail pages".
343 * All pages have PG_compound set. All pages have their ->private pointing at
344 * the head page (even the head page has this).
346 * The first tail page's ->lru.next holds the address of the compound page's
347 * put_page() function. Its ->lru.prev holds the order of allocation.
348 * This usage means that zero-order pages may not be compound.
351 static void free_compound_page(struct page *page)
353 __free_pages_ok(page, compound_order(page));
356 void prep_compound_page(struct page *page, unsigned long order)
359 int nr_pages = 1 << order;
361 set_compound_page_dtor(page, free_compound_page);
362 set_compound_order(page, order);
364 for (i = 1; i < nr_pages; i++) {
365 struct page *p = page + i;
367 set_page_count(p, 0);
368 p->first_page = page;
372 /* update __split_huge_page_refcount if you change this function */
373 static int destroy_compound_page(struct page *page, unsigned long order)
376 int nr_pages = 1 << order;
379 if (unlikely(compound_order(page) != order) ||
380 unlikely(!PageHead(page))) {
385 __ClearPageHead(page);
387 for (i = 1; i < nr_pages; i++) {
388 struct page *p = page + i;
390 if (unlikely(!PageTail(p) || (p->first_page != page))) {
400 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
405 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
406 * and __GFP_HIGHMEM from hard or soft interrupt context.
408 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
409 for (i = 0; i < (1 << order); i++)
410 clear_highpage(page + i);
413 static inline void set_page_order(struct page *page, int order)
415 set_page_private(page, order);
416 __SetPageBuddy(page);
419 static inline void rmv_page_order(struct page *page)
421 __ClearPageBuddy(page);
422 set_page_private(page, 0);
426 * Locate the struct page for both the matching buddy in our
427 * pair (buddy1) and the combined O(n+1) page they form (page).
429 * 1) Any buddy B1 will have an order O twin B2 which satisfies
430 * the following equation:
432 * For example, if the starting buddy (buddy2) is #8 its order
434 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
436 * 2) Any buddy B will have an order O+1 parent P which
437 * satisfies the following equation:
440 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
442 static inline unsigned long
443 __find_buddy_index(unsigned long page_idx, unsigned int order)
445 return page_idx ^ (1 << order);
449 * This function checks whether a page is free && is the buddy
450 * we can do coalesce a page and its buddy if
451 * (a) the buddy is not in a hole &&
452 * (b) the buddy is in the buddy system &&
453 * (c) a page and its buddy have the same order &&
454 * (d) a page and its buddy are in the same zone.
456 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
457 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
459 * For recording page's order, we use page_private(page).
461 static inline int page_is_buddy(struct page *page, struct page *buddy,
464 if (!pfn_valid_within(page_to_pfn(buddy)))
467 if (page_zone_id(page) != page_zone_id(buddy))
470 if (PageBuddy(buddy) && page_order(buddy) == order) {
471 VM_BUG_ON(page_count(buddy) != 0);
478 * Freeing function for a buddy system allocator.
480 * The concept of a buddy system is to maintain direct-mapped table
481 * (containing bit values) for memory blocks of various "orders".
482 * The bottom level table contains the map for the smallest allocatable
483 * units of memory (here, pages), and each level above it describes
484 * pairs of units from the levels below, hence, "buddies".
485 * At a high level, all that happens here is marking the table entry
486 * at the bottom level available, and propagating the changes upward
487 * as necessary, plus some accounting needed to play nicely with other
488 * parts of the VM system.
489 * At each level, we keep a list of pages, which are heads of continuous
490 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
491 * order is recorded in page_private(page) field.
492 * So when we are allocating or freeing one, we can derive the state of the
493 * other. That is, if we allocate a small block, and both were
494 * free, the remainder of the region must be split into blocks.
495 * If a block is freed, and its buddy is also free, then this
496 * triggers coalescing into a block of larger size.
501 static inline void __free_one_page(struct page *page,
502 struct zone *zone, unsigned int order,
505 unsigned long page_idx;
506 unsigned long combined_idx;
507 unsigned long uninitialized_var(buddy_idx);
510 if (unlikely(PageCompound(page)))
511 if (unlikely(destroy_compound_page(page, order)))
514 VM_BUG_ON(migratetype == -1);
516 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
518 VM_BUG_ON(page_idx & ((1 << order) - 1));
519 VM_BUG_ON(bad_range(zone, page));
521 while (order < MAX_ORDER-1) {
522 buddy_idx = __find_buddy_index(page_idx, order);
523 buddy = page + (buddy_idx - page_idx);
524 if (!page_is_buddy(page, buddy, order))
527 /* Our buddy is free, merge with it and move up one order. */
528 list_del(&buddy->lru);
529 zone->free_area[order].nr_free--;
530 rmv_page_order(buddy);
531 combined_idx = buddy_idx & page_idx;
532 page = page + (combined_idx - page_idx);
533 page_idx = combined_idx;
536 set_page_order(page, order);
539 * If this is not the largest possible page, check if the buddy
540 * of the next-highest order is free. If it is, it's possible
541 * that pages are being freed that will coalesce soon. In case,
542 * that is happening, add the free page to the tail of the list
543 * so it's less likely to be used soon and more likely to be merged
544 * as a higher order page
546 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
547 struct page *higher_page, *higher_buddy;
548 combined_idx = buddy_idx & page_idx;
549 higher_page = page + (combined_idx - page_idx);
550 buddy_idx = __find_buddy_index(combined_idx, order + 1);
551 higher_buddy = higher_page + (buddy_idx - combined_idx);
552 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
553 list_add_tail(&page->lru,
554 &zone->free_area[order].free_list[migratetype]);
559 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
561 zone->free_area[order].nr_free++;
565 * free_page_mlock() -- clean up attempts to free and mlocked() page.
566 * Page should not be on lru, so no need to fix that up.
567 * free_pages_check() will verify...
569 static inline void free_page_mlock(struct page *page)
571 __dec_zone_page_state(page, NR_MLOCK);
572 __count_vm_event(UNEVICTABLE_MLOCKFREED);
575 static inline int free_pages_check(struct page *page)
577 if (unlikely(page_mapcount(page) |
578 (page->mapping != NULL) |
579 (atomic_read(&page->_count) != 0) |
580 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
581 (mem_cgroup_bad_page_check(page)))) {
585 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
586 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
591 * Frees a number of pages from the PCP lists
592 * Assumes all pages on list are in same zone, and of same order.
593 * count is the number of pages to free.
595 * If the zone was previously in an "all pages pinned" state then look to
596 * see if this freeing clears that state.
598 * And clear the zone's pages_scanned counter, to hold off the "all pages are
599 * pinned" detection logic.
601 static void free_pcppages_bulk(struct zone *zone, int count,
602 struct per_cpu_pages *pcp)
608 spin_lock(&zone->lock);
609 zone->all_unreclaimable = 0;
610 zone->pages_scanned = 0;
614 struct list_head *list;
617 * Remove pages from lists in a round-robin fashion. A
618 * batch_free count is maintained that is incremented when an
619 * empty list is encountered. This is so more pages are freed
620 * off fuller lists instead of spinning excessively around empty
625 if (++migratetype == MIGRATE_PCPTYPES)
627 list = &pcp->lists[migratetype];
628 } while (list_empty(list));
630 /* This is the only non-empty list. Free them all. */
631 if (batch_free == MIGRATE_PCPTYPES)
632 batch_free = to_free;
635 page = list_entry(list->prev, struct page, lru);
636 /* must delete as __free_one_page list manipulates */
637 list_del(&page->lru);
638 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
639 __free_one_page(page, zone, 0, page_private(page));
640 trace_mm_page_pcpu_drain(page, 0, page_private(page));
641 } while (--to_free && --batch_free && !list_empty(list));
643 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
644 spin_unlock(&zone->lock);
647 static void free_one_page(struct zone *zone, struct page *page, int order,
650 spin_lock(&zone->lock);
651 zone->all_unreclaimable = 0;
652 zone->pages_scanned = 0;
654 __free_one_page(page, zone, order, migratetype);
655 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
656 spin_unlock(&zone->lock);
659 static bool free_pages_prepare(struct page *page, unsigned int order)
664 trace_mm_page_free(page, order);
665 kmemcheck_free_shadow(page, order);
668 page->mapping = NULL;
669 for (i = 0; i < (1 << order); i++)
670 bad += free_pages_check(page + i);
674 if (!PageHighMem(page)) {
675 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
676 debug_check_no_obj_freed(page_address(page),
679 arch_free_page(page, order);
680 kernel_map_pages(page, 1 << order, 0);
685 static void __free_pages_ok(struct page *page, unsigned int order)
688 int wasMlocked = __TestClearPageMlocked(page);
690 if (!free_pages_prepare(page, order))
693 local_irq_save(flags);
694 if (unlikely(wasMlocked))
695 free_page_mlock(page);
696 __count_vm_events(PGFREE, 1 << order);
697 free_one_page(page_zone(page), page, order,
698 get_pageblock_migratetype(page));
699 local_irq_restore(flags);
703 * permit the bootmem allocator to evade page validation on high-order frees
705 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
708 __ClearPageReserved(page);
709 set_page_count(page, 0);
710 set_page_refcounted(page);
716 for (loop = 0; loop < BITS_PER_LONG; loop++) {
717 struct page *p = &page[loop];
719 if (loop + 1 < BITS_PER_LONG)
721 __ClearPageReserved(p);
722 set_page_count(p, 0);
725 set_page_refcounted(page);
726 __free_pages(page, order);
732 * The order of subdivision here is critical for the IO subsystem.
733 * Please do not alter this order without good reasons and regression
734 * testing. Specifically, as large blocks of memory are subdivided,
735 * the order in which smaller blocks are delivered depends on the order
736 * they're subdivided in this function. This is the primary factor
737 * influencing the order in which pages are delivered to the IO
738 * subsystem according to empirical testing, and this is also justified
739 * by considering the behavior of a buddy system containing a single
740 * large block of memory acted on by a series of small allocations.
741 * This behavior is a critical factor in sglist merging's success.
745 static inline void expand(struct zone *zone, struct page *page,
746 int low, int high, struct free_area *area,
749 unsigned long size = 1 << high;
755 VM_BUG_ON(bad_range(zone, &page[size]));
756 list_add(&page[size].lru, &area->free_list[migratetype]);
758 set_page_order(&page[size], high);
763 * This page is about to be returned from the page allocator
765 static inline int check_new_page(struct page *page)
767 if (unlikely(page_mapcount(page) |
768 (page->mapping != NULL) |
769 (atomic_read(&page->_count) != 0) |
770 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
771 (mem_cgroup_bad_page_check(page)))) {
778 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
782 for (i = 0; i < (1 << order); i++) {
783 struct page *p = page + i;
784 if (unlikely(check_new_page(p)))
788 set_page_private(page, 0);
789 set_page_refcounted(page);
791 arch_alloc_page(page, order);
792 kernel_map_pages(page, 1 << order, 1);
794 if (gfp_flags & __GFP_ZERO)
795 prep_zero_page(page, order, gfp_flags);
797 if (order && (gfp_flags & __GFP_COMP))
798 prep_compound_page(page, order);
804 * Go through the free lists for the given migratetype and remove
805 * the smallest available page from the freelists
808 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
811 unsigned int current_order;
812 struct free_area * area;
815 /* Find a page of the appropriate size in the preferred list */
816 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
817 area = &(zone->free_area[current_order]);
818 if (list_empty(&area->free_list[migratetype]))
821 page = list_entry(area->free_list[migratetype].next,
823 list_del(&page->lru);
824 rmv_page_order(page);
826 expand(zone, page, order, current_order, area, migratetype);
835 * This array describes the order lists are fallen back to when
836 * the free lists for the desirable migrate type are depleted
838 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
839 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
840 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
841 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
842 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
846 * Move the free pages in a range to the free lists of the requested type.
847 * Note that start_page and end_pages are not aligned on a pageblock
848 * boundary. If alignment is required, use move_freepages_block()
850 static int move_freepages(struct zone *zone,
851 struct page *start_page, struct page *end_page,
858 #ifndef CONFIG_HOLES_IN_ZONE
860 * page_zone is not safe to call in this context when
861 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
862 * anyway as we check zone boundaries in move_freepages_block().
863 * Remove at a later date when no bug reports exist related to
864 * grouping pages by mobility
866 BUG_ON(page_zone(start_page) != page_zone(end_page));
869 for (page = start_page; page <= end_page;) {
870 /* Make sure we are not inadvertently changing nodes */
871 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
873 if (!pfn_valid_within(page_to_pfn(page))) {
878 if (!PageBuddy(page)) {
883 order = page_order(page);
884 list_move(&page->lru,
885 &zone->free_area[order].free_list[migratetype]);
887 pages_moved += 1 << order;
893 static int move_freepages_block(struct zone *zone, struct page *page,
896 unsigned long start_pfn, end_pfn;
897 struct page *start_page, *end_page;
899 start_pfn = page_to_pfn(page);
900 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
901 start_page = pfn_to_page(start_pfn);
902 end_page = start_page + pageblock_nr_pages - 1;
903 end_pfn = start_pfn + pageblock_nr_pages - 1;
905 /* Do not cross zone boundaries */
906 if (start_pfn < zone->zone_start_pfn)
908 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
911 return move_freepages(zone, start_page, end_page, migratetype);
914 static void change_pageblock_range(struct page *pageblock_page,
915 int start_order, int migratetype)
917 int nr_pageblocks = 1 << (start_order - pageblock_order);
919 while (nr_pageblocks--) {
920 set_pageblock_migratetype(pageblock_page, migratetype);
921 pageblock_page += pageblock_nr_pages;
925 /* Remove an element from the buddy allocator from the fallback list */
926 static inline struct page *
927 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
929 struct free_area * area;
934 /* Find the largest possible block of pages in the other list */
935 for (current_order = MAX_ORDER-1; current_order >= order;
937 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
938 migratetype = fallbacks[start_migratetype][i];
940 /* MIGRATE_RESERVE handled later if necessary */
941 if (migratetype == MIGRATE_RESERVE)
944 area = &(zone->free_area[current_order]);
945 if (list_empty(&area->free_list[migratetype]))
948 page = list_entry(area->free_list[migratetype].next,
953 * If breaking a large block of pages, move all free
954 * pages to the preferred allocation list. If falling
955 * back for a reclaimable kernel allocation, be more
956 * aggressive about taking ownership of free pages
958 if (unlikely(current_order >= (pageblock_order >> 1)) ||
959 start_migratetype == MIGRATE_RECLAIMABLE ||
960 page_group_by_mobility_disabled) {
962 pages = move_freepages_block(zone, page,
965 /* Claim the whole block if over half of it is free */
966 if (pages >= (1 << (pageblock_order-1)) ||
967 page_group_by_mobility_disabled)
968 set_pageblock_migratetype(page,
971 migratetype = start_migratetype;
974 /* Remove the page from the freelists */
975 list_del(&page->lru);
976 rmv_page_order(page);
978 /* Take ownership for orders >= pageblock_order */
979 if (current_order >= pageblock_order)
980 change_pageblock_range(page, current_order,
983 expand(zone, page, order, current_order, area, migratetype);
985 trace_mm_page_alloc_extfrag(page, order, current_order,
986 start_migratetype, migratetype);
996 * Do the hard work of removing an element from the buddy allocator.
997 * Call me with the zone->lock already held.
999 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1005 page = __rmqueue_smallest(zone, order, migratetype);
1007 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1008 page = __rmqueue_fallback(zone, order, migratetype);
1011 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1012 * is used because __rmqueue_smallest is an inline function
1013 * and we want just one call site
1016 migratetype = MIGRATE_RESERVE;
1021 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1026 * Obtain a specified number of elements from the buddy allocator, all under
1027 * a single hold of the lock, for efficiency. Add them to the supplied list.
1028 * Returns the number of new pages which were placed at *list.
1030 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1031 unsigned long count, struct list_head *list,
1032 int migratetype, int cold)
1036 spin_lock(&zone->lock);
1037 for (i = 0; i < count; ++i) {
1038 struct page *page = __rmqueue(zone, order, migratetype);
1039 if (unlikely(page == NULL))
1043 * Split buddy pages returned by expand() are received here
1044 * in physical page order. The page is added to the callers and
1045 * list and the list head then moves forward. From the callers
1046 * perspective, the linked list is ordered by page number in
1047 * some conditions. This is useful for IO devices that can
1048 * merge IO requests if the physical pages are ordered
1051 if (likely(cold == 0))
1052 list_add(&page->lru, list);
1054 list_add_tail(&page->lru, list);
1055 set_page_private(page, migratetype);
1058 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1059 spin_unlock(&zone->lock);
1065 * Called from the vmstat counter updater to drain pagesets of this
1066 * currently executing processor on remote nodes after they have
1069 * Note that this function must be called with the thread pinned to
1070 * a single processor.
1072 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1074 unsigned long flags;
1077 local_irq_save(flags);
1078 if (pcp->count >= pcp->batch)
1079 to_drain = pcp->batch;
1081 to_drain = pcp->count;
1082 free_pcppages_bulk(zone, to_drain, pcp);
1083 pcp->count -= to_drain;
1084 local_irq_restore(flags);
1089 * Drain pages of the indicated processor.
1091 * The processor must either be the current processor and the
1092 * thread pinned to the current processor or a processor that
1095 static void drain_pages(unsigned int cpu)
1097 unsigned long flags;
1100 for_each_populated_zone(zone) {
1101 struct per_cpu_pageset *pset;
1102 struct per_cpu_pages *pcp;
1104 local_irq_save(flags);
1105 pset = per_cpu_ptr(zone->pageset, cpu);
1109 free_pcppages_bulk(zone, pcp->count, pcp);
1112 local_irq_restore(flags);
1117 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1119 void drain_local_pages(void *arg)
1121 drain_pages(smp_processor_id());
1125 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1127 void drain_all_pages(void)
1129 on_each_cpu(drain_local_pages, NULL, 1);
1132 #ifdef CONFIG_HIBERNATION
1134 void mark_free_pages(struct zone *zone)
1136 unsigned long pfn, max_zone_pfn;
1137 unsigned long flags;
1139 struct list_head *curr;
1141 if (!zone->spanned_pages)
1144 spin_lock_irqsave(&zone->lock, flags);
1146 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1147 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1148 if (pfn_valid(pfn)) {
1149 struct page *page = pfn_to_page(pfn);
1151 if (!swsusp_page_is_forbidden(page))
1152 swsusp_unset_page_free(page);
1155 for_each_migratetype_order(order, t) {
1156 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1159 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1160 for (i = 0; i < (1UL << order); i++)
1161 swsusp_set_page_free(pfn_to_page(pfn + i));
1164 spin_unlock_irqrestore(&zone->lock, flags);
1166 #endif /* CONFIG_PM */
1169 * Free a 0-order page
1170 * cold == 1 ? free a cold page : free a hot page
1172 void free_hot_cold_page(struct page *page, int cold)
1174 struct zone *zone = page_zone(page);
1175 struct per_cpu_pages *pcp;
1176 unsigned long flags;
1178 int wasMlocked = __TestClearPageMlocked(page);
1180 if (!free_pages_prepare(page, 0))
1183 migratetype = get_pageblock_migratetype(page);
1184 set_page_private(page, migratetype);
1185 local_irq_save(flags);
1186 if (unlikely(wasMlocked))
1187 free_page_mlock(page);
1188 __count_vm_event(PGFREE);
1191 * We only track unmovable, reclaimable and movable on pcp lists.
1192 * Free ISOLATE pages back to the allocator because they are being
1193 * offlined but treat RESERVE as movable pages so we can get those
1194 * areas back if necessary. Otherwise, we may have to free
1195 * excessively into the page allocator
1197 if (migratetype >= MIGRATE_PCPTYPES) {
1198 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1199 free_one_page(zone, page, 0, migratetype);
1202 migratetype = MIGRATE_MOVABLE;
1205 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1207 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1209 list_add(&page->lru, &pcp->lists[migratetype]);
1211 if (pcp->count >= pcp->high) {
1212 free_pcppages_bulk(zone, pcp->batch, pcp);
1213 pcp->count -= pcp->batch;
1217 local_irq_restore(flags);
1221 * Free a list of 0-order pages
1223 void free_hot_cold_page_list(struct list_head *list, int cold)
1225 struct page *page, *next;
1227 list_for_each_entry_safe(page, next, list, lru) {
1228 trace_mm_page_free_batched(page, cold);
1229 free_hot_cold_page(page, cold);
1234 * split_page takes a non-compound higher-order page, and splits it into
1235 * n (1<<order) sub-pages: page[0..n]
1236 * Each sub-page must be freed individually.
1238 * Note: this is probably too low level an operation for use in drivers.
1239 * Please consult with lkml before using this in your driver.
1241 void split_page(struct page *page, unsigned int order)
1245 VM_BUG_ON(PageCompound(page));
1246 VM_BUG_ON(!page_count(page));
1248 #ifdef CONFIG_KMEMCHECK
1250 * Split shadow pages too, because free(page[0]) would
1251 * otherwise free the whole shadow.
1253 if (kmemcheck_page_is_tracked(page))
1254 split_page(virt_to_page(page[0].shadow), order);
1257 for (i = 1; i < (1 << order); i++)
1258 set_page_refcounted(page + i);
1262 * Similar to split_page except the page is already free. As this is only
1263 * being used for migration, the migratetype of the block also changes.
1264 * As this is called with interrupts disabled, the caller is responsible
1265 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1268 * Note: this is probably too low level an operation for use in drivers.
1269 * Please consult with lkml before using this in your driver.
1271 int split_free_page(struct page *page)
1274 unsigned long watermark;
1277 BUG_ON(!PageBuddy(page));
1279 zone = page_zone(page);
1280 order = page_order(page);
1282 /* Obey watermarks as if the page was being allocated */
1283 watermark = low_wmark_pages(zone) + (1 << order);
1284 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1287 /* Remove page from free list */
1288 list_del(&page->lru);
1289 zone->free_area[order].nr_free--;
1290 rmv_page_order(page);
1291 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1293 /* Split into individual pages */
1294 set_page_refcounted(page);
1295 split_page(page, order);
1297 if (order >= pageblock_order - 1) {
1298 struct page *endpage = page + (1 << order) - 1;
1299 for (; page < endpage; page += pageblock_nr_pages)
1300 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1307 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1308 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1312 struct page *buffered_rmqueue(struct zone *preferred_zone,
1313 struct zone *zone, int order, gfp_t gfp_flags,
1316 unsigned long flags;
1318 int cold = !!(gfp_flags & __GFP_COLD);
1321 if (likely(order == 0)) {
1322 struct per_cpu_pages *pcp;
1323 struct list_head *list;
1325 local_irq_save(flags);
1326 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1327 list = &pcp->lists[migratetype];
1328 if (list_empty(list)) {
1329 pcp->count += rmqueue_bulk(zone, 0,
1332 if (unlikely(list_empty(list)))
1337 page = list_entry(list->prev, struct page, lru);
1339 page = list_entry(list->next, struct page, lru);
1341 list_del(&page->lru);
1344 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1346 * __GFP_NOFAIL is not to be used in new code.
1348 * All __GFP_NOFAIL callers should be fixed so that they
1349 * properly detect and handle allocation failures.
1351 * We most definitely don't want callers attempting to
1352 * allocate greater than order-1 page units with
1355 WARN_ON_ONCE(order > 1);
1357 spin_lock_irqsave(&zone->lock, flags);
1358 page = __rmqueue(zone, order, migratetype);
1359 spin_unlock(&zone->lock);
1362 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1365 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1366 zone_statistics(preferred_zone, zone, gfp_flags);
1367 local_irq_restore(flags);
1369 VM_BUG_ON(bad_range(zone, page));
1370 if (prep_new_page(page, order, gfp_flags))
1375 local_irq_restore(flags);
1379 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1380 #define ALLOC_WMARK_MIN WMARK_MIN
1381 #define ALLOC_WMARK_LOW WMARK_LOW
1382 #define ALLOC_WMARK_HIGH WMARK_HIGH
1383 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1385 /* Mask to get the watermark bits */
1386 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1388 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1389 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1390 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1392 #ifdef CONFIG_FAIL_PAGE_ALLOC
1395 struct fault_attr attr;
1397 u32 ignore_gfp_highmem;
1398 u32 ignore_gfp_wait;
1400 } fail_page_alloc = {
1401 .attr = FAULT_ATTR_INITIALIZER,
1402 .ignore_gfp_wait = 1,
1403 .ignore_gfp_highmem = 1,
1407 static int __init setup_fail_page_alloc(char *str)
1409 return setup_fault_attr(&fail_page_alloc.attr, str);
1411 __setup("fail_page_alloc=", setup_fail_page_alloc);
1413 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1415 if (order < fail_page_alloc.min_order)
1417 if (gfp_mask & __GFP_NOFAIL)
1419 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1421 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1424 return should_fail(&fail_page_alloc.attr, 1 << order);
1427 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1429 static int __init fail_page_alloc_debugfs(void)
1431 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1434 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1435 &fail_page_alloc.attr);
1437 return PTR_ERR(dir);
1439 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1440 &fail_page_alloc.ignore_gfp_wait))
1442 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1443 &fail_page_alloc.ignore_gfp_highmem))
1445 if (!debugfs_create_u32("min-order", mode, dir,
1446 &fail_page_alloc.min_order))
1451 debugfs_remove_recursive(dir);
1456 late_initcall(fail_page_alloc_debugfs);
1458 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1460 #else /* CONFIG_FAIL_PAGE_ALLOC */
1462 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1467 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1470 * Return true if free pages are above 'mark'. This takes into account the order
1471 * of the allocation.
1473 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1474 int classzone_idx, int alloc_flags, long free_pages)
1476 /* free_pages my go negative - that's OK */
1480 free_pages -= (1 << order) + 1;
1481 if (alloc_flags & ALLOC_HIGH)
1483 if (alloc_flags & ALLOC_HARDER)
1486 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1488 for (o = 0; o < order; o++) {
1489 /* At the next order, this order's pages become unavailable */
1490 free_pages -= z->free_area[o].nr_free << o;
1492 /* Require fewer higher order pages to be free */
1495 if (free_pages <= min)
1501 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1502 int classzone_idx, int alloc_flags)
1504 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1505 zone_page_state(z, NR_FREE_PAGES));
1508 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1509 int classzone_idx, int alloc_flags)
1511 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1513 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1514 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1516 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1522 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1523 * skip over zones that are not allowed by the cpuset, or that have
1524 * been recently (in last second) found to be nearly full. See further
1525 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1526 * that have to skip over a lot of full or unallowed zones.
1528 * If the zonelist cache is present in the passed in zonelist, then
1529 * returns a pointer to the allowed node mask (either the current
1530 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1532 * If the zonelist cache is not available for this zonelist, does
1533 * nothing and returns NULL.
1535 * If the fullzones BITMAP in the zonelist cache is stale (more than
1536 * a second since last zap'd) then we zap it out (clear its bits.)
1538 * We hold off even calling zlc_setup, until after we've checked the
1539 * first zone in the zonelist, on the theory that most allocations will
1540 * be satisfied from that first zone, so best to examine that zone as
1541 * quickly as we can.
1543 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1545 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1546 nodemask_t *allowednodes; /* zonelist_cache approximation */
1548 zlc = zonelist->zlcache_ptr;
1552 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1553 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1554 zlc->last_full_zap = jiffies;
1557 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1558 &cpuset_current_mems_allowed :
1559 &node_states[N_HIGH_MEMORY];
1560 return allowednodes;
1564 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1565 * if it is worth looking at further for free memory:
1566 * 1) Check that the zone isn't thought to be full (doesn't have its
1567 * bit set in the zonelist_cache fullzones BITMAP).
1568 * 2) Check that the zones node (obtained from the zonelist_cache
1569 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1570 * Return true (non-zero) if zone is worth looking at further, or
1571 * else return false (zero) if it is not.
1573 * This check -ignores- the distinction between various watermarks,
1574 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1575 * found to be full for any variation of these watermarks, it will
1576 * be considered full for up to one second by all requests, unless
1577 * we are so low on memory on all allowed nodes that we are forced
1578 * into the second scan of the zonelist.
1580 * In the second scan we ignore this zonelist cache and exactly
1581 * apply the watermarks to all zones, even it is slower to do so.
1582 * We are low on memory in the second scan, and should leave no stone
1583 * unturned looking for a free page.
1585 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1586 nodemask_t *allowednodes)
1588 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1589 int i; /* index of *z in zonelist zones */
1590 int n; /* node that zone *z is on */
1592 zlc = zonelist->zlcache_ptr;
1596 i = z - zonelist->_zonerefs;
1599 /* This zone is worth trying if it is allowed but not full */
1600 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1604 * Given 'z' scanning a zonelist, set the corresponding bit in
1605 * zlc->fullzones, so that subsequent attempts to allocate a page
1606 * from that zone don't waste time re-examining it.
1608 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1610 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1611 int i; /* index of *z in zonelist zones */
1613 zlc = zonelist->zlcache_ptr;
1617 i = z - zonelist->_zonerefs;
1619 set_bit(i, zlc->fullzones);
1623 * clear all zones full, called after direct reclaim makes progress so that
1624 * a zone that was recently full is not skipped over for up to a second
1626 static void zlc_clear_zones_full(struct zonelist *zonelist)
1628 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1630 zlc = zonelist->zlcache_ptr;
1634 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1637 #else /* CONFIG_NUMA */
1639 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1644 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1645 nodemask_t *allowednodes)
1650 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1654 static void zlc_clear_zones_full(struct zonelist *zonelist)
1657 #endif /* CONFIG_NUMA */
1660 * get_page_from_freelist goes through the zonelist trying to allocate
1663 static struct page *
1664 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1665 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1666 struct zone *preferred_zone, int migratetype)
1669 struct page *page = NULL;
1672 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1673 int zlc_active = 0; /* set if using zonelist_cache */
1674 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1676 classzone_idx = zone_idx(preferred_zone);
1679 * Scan zonelist, looking for a zone with enough free.
1680 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1682 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1683 high_zoneidx, nodemask) {
1684 if (NUMA_BUILD && zlc_active &&
1685 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1687 if ((alloc_flags & ALLOC_CPUSET) &&
1688 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1691 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1692 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1696 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1697 if (zone_watermark_ok(zone, order, mark,
1698 classzone_idx, alloc_flags))
1701 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1703 * we do zlc_setup if there are multiple nodes
1704 * and before considering the first zone allowed
1707 allowednodes = zlc_setup(zonelist, alloc_flags);
1712 if (zone_reclaim_mode == 0)
1713 goto this_zone_full;
1716 * As we may have just activated ZLC, check if the first
1717 * eligible zone has failed zone_reclaim recently.
1719 if (NUMA_BUILD && zlc_active &&
1720 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1723 ret = zone_reclaim(zone, gfp_mask, order);
1725 case ZONE_RECLAIM_NOSCAN:
1728 case ZONE_RECLAIM_FULL:
1729 /* scanned but unreclaimable */
1732 /* did we reclaim enough */
1733 if (!zone_watermark_ok(zone, order, mark,
1734 classzone_idx, alloc_flags))
1735 goto this_zone_full;
1740 page = buffered_rmqueue(preferred_zone, zone, order,
1741 gfp_mask, migratetype);
1746 zlc_mark_zone_full(zonelist, z);
1749 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1750 /* Disable zlc cache for second zonelist scan */
1758 * Large machines with many possible nodes should not always dump per-node
1759 * meminfo in irq context.
1761 static inline bool should_suppress_show_mem(void)
1766 ret = in_interrupt();
1771 static DEFINE_RATELIMIT_STATE(nopage_rs,
1772 DEFAULT_RATELIMIT_INTERVAL,
1773 DEFAULT_RATELIMIT_BURST);
1775 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1777 unsigned int filter = SHOW_MEM_FILTER_NODES;
1779 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1783 * This documents exceptions given to allocations in certain
1784 * contexts that are allowed to allocate outside current's set
1787 if (!(gfp_mask & __GFP_NOMEMALLOC))
1788 if (test_thread_flag(TIF_MEMDIE) ||
1789 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1790 filter &= ~SHOW_MEM_FILTER_NODES;
1791 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1792 filter &= ~SHOW_MEM_FILTER_NODES;
1795 struct va_format vaf;
1798 va_start(args, fmt);
1803 pr_warn("%pV", &vaf);
1808 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1809 current->comm, order, gfp_mask);
1812 if (!should_suppress_show_mem())
1817 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1818 unsigned long did_some_progress,
1819 unsigned long pages_reclaimed)
1821 /* Do not loop if specifically requested */
1822 if (gfp_mask & __GFP_NORETRY)
1825 /* Always retry if specifically requested */
1826 if (gfp_mask & __GFP_NOFAIL)
1830 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
1831 * making forward progress without invoking OOM. Suspend also disables
1832 * storage devices so kswapd will not help. Bail if we are suspending.
1834 if (!did_some_progress && pm_suspended_storage())
1838 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1839 * means __GFP_NOFAIL, but that may not be true in other
1842 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1846 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1847 * specified, then we retry until we no longer reclaim any pages
1848 * (above), or we've reclaimed an order of pages at least as
1849 * large as the allocation's order. In both cases, if the
1850 * allocation still fails, we stop retrying.
1852 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1858 static inline struct page *
1859 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1860 struct zonelist *zonelist, enum zone_type high_zoneidx,
1861 nodemask_t *nodemask, struct zone *preferred_zone,
1866 /* Acquire the OOM killer lock for the zones in zonelist */
1867 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1868 schedule_timeout_uninterruptible(1);
1873 * Go through the zonelist yet one more time, keep very high watermark
1874 * here, this is only to catch a parallel oom killing, we must fail if
1875 * we're still under heavy pressure.
1877 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1878 order, zonelist, high_zoneidx,
1879 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1880 preferred_zone, migratetype);
1884 if (!(gfp_mask & __GFP_NOFAIL)) {
1885 /* The OOM killer will not help higher order allocs */
1886 if (order > PAGE_ALLOC_COSTLY_ORDER)
1888 /* The OOM killer does not needlessly kill tasks for lowmem */
1889 if (high_zoneidx < ZONE_NORMAL)
1892 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1893 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1894 * The caller should handle page allocation failure by itself if
1895 * it specifies __GFP_THISNODE.
1896 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1898 if (gfp_mask & __GFP_THISNODE)
1901 /* Exhausted what can be done so it's blamo time */
1902 out_of_memory(zonelist, gfp_mask, order, nodemask);
1905 clear_zonelist_oom(zonelist, gfp_mask);
1909 #ifdef CONFIG_COMPACTION
1910 /* Try memory compaction for high-order allocations before reclaim */
1911 static struct page *
1912 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1913 struct zonelist *zonelist, enum zone_type high_zoneidx,
1914 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1915 int migratetype, bool sync_migration,
1916 bool *deferred_compaction,
1917 unsigned long *did_some_progress)
1924 if (compaction_deferred(preferred_zone)) {
1925 *deferred_compaction = true;
1929 current->flags |= PF_MEMALLOC;
1930 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1931 nodemask, sync_migration);
1932 current->flags &= ~PF_MEMALLOC;
1933 if (*did_some_progress != COMPACT_SKIPPED) {
1935 /* Page migration frees to the PCP lists but we want merging */
1936 drain_pages(get_cpu());
1939 page = get_page_from_freelist(gfp_mask, nodemask,
1940 order, zonelist, high_zoneidx,
1941 alloc_flags, preferred_zone,
1944 preferred_zone->compact_considered = 0;
1945 preferred_zone->compact_defer_shift = 0;
1946 count_vm_event(COMPACTSUCCESS);
1951 * It's bad if compaction run occurs and fails.
1952 * The most likely reason is that pages exist,
1953 * but not enough to satisfy watermarks.
1955 count_vm_event(COMPACTFAIL);
1958 * As async compaction considers a subset of pageblocks, only
1959 * defer if the failure was a sync compaction failure.
1962 defer_compaction(preferred_zone);
1970 static inline struct page *
1971 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1972 struct zonelist *zonelist, enum zone_type high_zoneidx,
1973 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1974 int migratetype, bool sync_migration,
1975 bool *deferred_compaction,
1976 unsigned long *did_some_progress)
1980 #endif /* CONFIG_COMPACTION */
1982 /* The really slow allocator path where we enter direct reclaim */
1983 static inline struct page *
1984 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1985 struct zonelist *zonelist, enum zone_type high_zoneidx,
1986 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1987 int migratetype, unsigned long *did_some_progress)
1989 struct page *page = NULL;
1990 struct reclaim_state reclaim_state;
1991 bool drained = false;
1995 /* We now go into synchronous reclaim */
1996 cpuset_memory_pressure_bump();
1997 current->flags |= PF_MEMALLOC;
1998 lockdep_set_current_reclaim_state(gfp_mask);
1999 reclaim_state.reclaimed_slab = 0;
2000 current->reclaim_state = &reclaim_state;
2002 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2004 current->reclaim_state = NULL;
2005 lockdep_clear_current_reclaim_state();
2006 current->flags &= ~PF_MEMALLOC;
2010 if (unlikely(!(*did_some_progress)))
2013 /* After successful reclaim, reconsider all zones for allocation */
2015 zlc_clear_zones_full(zonelist);
2018 page = get_page_from_freelist(gfp_mask, nodemask, order,
2019 zonelist, high_zoneidx,
2020 alloc_flags, preferred_zone,
2024 * If an allocation failed after direct reclaim, it could be because
2025 * pages are pinned on the per-cpu lists. Drain them and try again
2027 if (!page && !drained) {
2037 * This is called in the allocator slow-path if the allocation request is of
2038 * sufficient urgency to ignore watermarks and take other desperate measures
2040 static inline struct page *
2041 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2042 struct zonelist *zonelist, enum zone_type high_zoneidx,
2043 nodemask_t *nodemask, struct zone *preferred_zone,
2049 page = get_page_from_freelist(gfp_mask, nodemask, order,
2050 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2051 preferred_zone, migratetype);
2053 if (!page && gfp_mask & __GFP_NOFAIL)
2054 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2055 } while (!page && (gfp_mask & __GFP_NOFAIL));
2061 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2062 enum zone_type high_zoneidx,
2063 enum zone_type classzone_idx)
2068 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2069 wakeup_kswapd(zone, order, classzone_idx);
2073 gfp_to_alloc_flags(gfp_t gfp_mask)
2075 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2076 const gfp_t wait = gfp_mask & __GFP_WAIT;
2078 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2079 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2082 * The caller may dip into page reserves a bit more if the caller
2083 * cannot run direct reclaim, or if the caller has realtime scheduling
2084 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2085 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2087 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2091 * Not worth trying to allocate harder for
2092 * __GFP_NOMEMALLOC even if it can't schedule.
2094 if (!(gfp_mask & __GFP_NOMEMALLOC))
2095 alloc_flags |= ALLOC_HARDER;
2097 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2098 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2100 alloc_flags &= ~ALLOC_CPUSET;
2101 } else if (unlikely(rt_task(current)) && !in_interrupt())
2102 alloc_flags |= ALLOC_HARDER;
2104 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2105 if (!in_interrupt() &&
2106 ((current->flags & PF_MEMALLOC) ||
2107 unlikely(test_thread_flag(TIF_MEMDIE))))
2108 alloc_flags |= ALLOC_NO_WATERMARKS;
2114 static inline struct page *
2115 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2116 struct zonelist *zonelist, enum zone_type high_zoneidx,
2117 nodemask_t *nodemask, struct zone *preferred_zone,
2120 const gfp_t wait = gfp_mask & __GFP_WAIT;
2121 struct page *page = NULL;
2123 unsigned long pages_reclaimed = 0;
2124 unsigned long did_some_progress;
2125 bool sync_migration = false;
2126 bool deferred_compaction = false;
2129 * In the slowpath, we sanity check order to avoid ever trying to
2130 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2131 * be using allocators in order of preference for an area that is
2134 if (order >= MAX_ORDER) {
2135 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2140 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2141 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2142 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2143 * using a larger set of nodes after it has established that the
2144 * allowed per node queues are empty and that nodes are
2147 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2151 if (!(gfp_mask & __GFP_NO_KSWAPD))
2152 wake_all_kswapd(order, zonelist, high_zoneidx,
2153 zone_idx(preferred_zone));
2156 * OK, we're below the kswapd watermark and have kicked background
2157 * reclaim. Now things get more complex, so set up alloc_flags according
2158 * to how we want to proceed.
2160 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2163 * Find the true preferred zone if the allocation is unconstrained by
2166 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2167 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2171 /* This is the last chance, in general, before the goto nopage. */
2172 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2173 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2174 preferred_zone, migratetype);
2178 /* Allocate without watermarks if the context allows */
2179 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2180 page = __alloc_pages_high_priority(gfp_mask, order,
2181 zonelist, high_zoneidx, nodemask,
2182 preferred_zone, migratetype);
2187 /* Atomic allocations - we can't balance anything */
2191 /* Avoid recursion of direct reclaim */
2192 if (current->flags & PF_MEMALLOC)
2195 /* Avoid allocations with no watermarks from looping endlessly */
2196 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2200 * Try direct compaction. The first pass is asynchronous. Subsequent
2201 * attempts after direct reclaim are synchronous
2203 page = __alloc_pages_direct_compact(gfp_mask, order,
2204 zonelist, high_zoneidx,
2206 alloc_flags, preferred_zone,
2207 migratetype, sync_migration,
2208 &deferred_compaction,
2209 &did_some_progress);
2212 sync_migration = true;
2215 * If compaction is deferred for high-order allocations, it is because
2216 * sync compaction recently failed. In this is the case and the caller
2217 * has requested the system not be heavily disrupted, fail the
2218 * allocation now instead of entering direct reclaim
2220 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2223 /* Try direct reclaim and then allocating */
2224 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2225 zonelist, high_zoneidx,
2227 alloc_flags, preferred_zone,
2228 migratetype, &did_some_progress);
2233 * If we failed to make any progress reclaiming, then we are
2234 * running out of options and have to consider going OOM
2236 if (!did_some_progress) {
2237 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2238 if (oom_killer_disabled)
2240 page = __alloc_pages_may_oom(gfp_mask, order,
2241 zonelist, high_zoneidx,
2242 nodemask, preferred_zone,
2247 if (!(gfp_mask & __GFP_NOFAIL)) {
2249 * The oom killer is not called for high-order
2250 * allocations that may fail, so if no progress
2251 * is being made, there are no other options and
2252 * retrying is unlikely to help.
2254 if (order > PAGE_ALLOC_COSTLY_ORDER)
2257 * The oom killer is not called for lowmem
2258 * allocations to prevent needlessly killing
2261 if (high_zoneidx < ZONE_NORMAL)
2269 /* Check if we should retry the allocation */
2270 pages_reclaimed += did_some_progress;
2271 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2273 /* Wait for some write requests to complete then retry */
2274 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2278 * High-order allocations do not necessarily loop after
2279 * direct reclaim and reclaim/compaction depends on compaction
2280 * being called after reclaim so call directly if necessary
2282 page = __alloc_pages_direct_compact(gfp_mask, order,
2283 zonelist, high_zoneidx,
2285 alloc_flags, preferred_zone,
2286 migratetype, sync_migration,
2287 &deferred_compaction,
2288 &did_some_progress);
2294 warn_alloc_failed(gfp_mask, order, NULL);
2297 if (kmemcheck_enabled)
2298 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2304 * This is the 'heart' of the zoned buddy allocator.
2307 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2308 struct zonelist *zonelist, nodemask_t *nodemask)
2310 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2311 struct zone *preferred_zone;
2312 struct page *page = NULL;
2313 int migratetype = allocflags_to_migratetype(gfp_mask);
2314 unsigned int cpuset_mems_cookie;
2316 gfp_mask &= gfp_allowed_mask;
2318 lockdep_trace_alloc(gfp_mask);
2320 might_sleep_if(gfp_mask & __GFP_WAIT);
2322 if (should_fail_alloc_page(gfp_mask, order))
2326 * Check the zones suitable for the gfp_mask contain at least one
2327 * valid zone. It's possible to have an empty zonelist as a result
2328 * of GFP_THISNODE and a memoryless node
2330 if (unlikely(!zonelist->_zonerefs->zone))
2334 cpuset_mems_cookie = get_mems_allowed();
2336 /* The preferred zone is used for statistics later */
2337 first_zones_zonelist(zonelist, high_zoneidx,
2338 nodemask ? : &cpuset_current_mems_allowed,
2340 if (!preferred_zone)
2343 /* First allocation attempt */
2344 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2345 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2346 preferred_zone, migratetype);
2347 if (unlikely(!page))
2348 page = __alloc_pages_slowpath(gfp_mask, order,
2349 zonelist, high_zoneidx, nodemask,
2350 preferred_zone, migratetype);
2352 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2356 * When updating a task's mems_allowed, it is possible to race with
2357 * parallel threads in such a way that an allocation can fail while
2358 * the mask is being updated. If a page allocation is about to fail,
2359 * check if the cpuset changed during allocation and if so, retry.
2361 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2366 EXPORT_SYMBOL(__alloc_pages_nodemask);
2369 * Common helper functions.
2371 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2376 * __get_free_pages() returns a 32-bit address, which cannot represent
2379 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2381 page = alloc_pages(gfp_mask, order);
2384 return (unsigned long) page_address(page);
2386 EXPORT_SYMBOL(__get_free_pages);
2388 unsigned long get_zeroed_page(gfp_t gfp_mask)
2390 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2392 EXPORT_SYMBOL(get_zeroed_page);
2394 void __free_pages(struct page *page, unsigned int order)
2396 if (put_page_testzero(page)) {
2398 free_hot_cold_page(page, 0);
2400 __free_pages_ok(page, order);
2404 EXPORT_SYMBOL(__free_pages);
2406 void free_pages(unsigned long addr, unsigned int order)
2409 VM_BUG_ON(!virt_addr_valid((void *)addr));
2410 __free_pages(virt_to_page((void *)addr), order);
2414 EXPORT_SYMBOL(free_pages);
2416 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2419 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2420 unsigned long used = addr + PAGE_ALIGN(size);
2422 split_page(virt_to_page((void *)addr), order);
2423 while (used < alloc_end) {
2428 return (void *)addr;
2432 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2433 * @size: the number of bytes to allocate
2434 * @gfp_mask: GFP flags for the allocation
2436 * This function is similar to alloc_pages(), except that it allocates the
2437 * minimum number of pages to satisfy the request. alloc_pages() can only
2438 * allocate memory in power-of-two pages.
2440 * This function is also limited by MAX_ORDER.
2442 * Memory allocated by this function must be released by free_pages_exact().
2444 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2446 unsigned int order = get_order(size);
2449 addr = __get_free_pages(gfp_mask, order);
2450 return make_alloc_exact(addr, order, size);
2452 EXPORT_SYMBOL(alloc_pages_exact);
2455 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2457 * @nid: the preferred node ID where memory should be allocated
2458 * @size: the number of bytes to allocate
2459 * @gfp_mask: GFP flags for the allocation
2461 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2463 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2466 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2468 unsigned order = get_order(size);
2469 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2472 return make_alloc_exact((unsigned long)page_address(p), order, size);
2474 EXPORT_SYMBOL(alloc_pages_exact_nid);
2477 * free_pages_exact - release memory allocated via alloc_pages_exact()
2478 * @virt: the value returned by alloc_pages_exact.
2479 * @size: size of allocation, same value as passed to alloc_pages_exact().
2481 * Release the memory allocated by a previous call to alloc_pages_exact.
2483 void free_pages_exact(void *virt, size_t size)
2485 unsigned long addr = (unsigned long)virt;
2486 unsigned long end = addr + PAGE_ALIGN(size);
2488 while (addr < end) {
2493 EXPORT_SYMBOL(free_pages_exact);
2495 static unsigned int nr_free_zone_pages(int offset)
2500 /* Just pick one node, since fallback list is circular */
2501 unsigned int sum = 0;
2503 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2505 for_each_zone_zonelist(zone, z, zonelist, offset) {
2506 unsigned long size = zone->present_pages;
2507 unsigned long high = high_wmark_pages(zone);
2516 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2518 unsigned int nr_free_buffer_pages(void)
2520 return nr_free_zone_pages(gfp_zone(GFP_USER));
2522 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2525 * Amount of free RAM allocatable within all zones
2527 unsigned int nr_free_pagecache_pages(void)
2529 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2532 static inline void show_node(struct zone *zone)
2535 printk("Node %d ", zone_to_nid(zone));
2538 void si_meminfo(struct sysinfo *val)
2540 val->totalram = totalram_pages;
2542 val->freeram = global_page_state(NR_FREE_PAGES);
2543 val->bufferram = nr_blockdev_pages();
2544 val->totalhigh = totalhigh_pages;
2545 val->freehigh = nr_free_highpages();
2546 val->mem_unit = PAGE_SIZE;
2549 EXPORT_SYMBOL(si_meminfo);
2552 void si_meminfo_node(struct sysinfo *val, int nid)
2554 pg_data_t *pgdat = NODE_DATA(nid);
2556 val->totalram = pgdat->node_present_pages;
2557 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2558 #ifdef CONFIG_HIGHMEM
2559 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2560 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2566 val->mem_unit = PAGE_SIZE;
2571 * Determine whether the node should be displayed or not, depending on whether
2572 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2574 bool skip_free_areas_node(unsigned int flags, int nid)
2577 unsigned int cpuset_mems_cookie;
2579 if (!(flags & SHOW_MEM_FILTER_NODES))
2583 cpuset_mems_cookie = get_mems_allowed();
2584 ret = !node_isset(nid, cpuset_current_mems_allowed);
2585 } while (!put_mems_allowed(cpuset_mems_cookie));
2590 #define K(x) ((x) << (PAGE_SHIFT-10))
2593 * Show free area list (used inside shift_scroll-lock stuff)
2594 * We also calculate the percentage fragmentation. We do this by counting the
2595 * memory on each free list with the exception of the first item on the list.
2596 * Suppresses nodes that are not allowed by current's cpuset if
2597 * SHOW_MEM_FILTER_NODES is passed.
2599 void show_free_areas(unsigned int filter)
2604 for_each_populated_zone(zone) {
2605 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2608 printk("%s per-cpu:\n", zone->name);
2610 for_each_online_cpu(cpu) {
2611 struct per_cpu_pageset *pageset;
2613 pageset = per_cpu_ptr(zone->pageset, cpu);
2615 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2616 cpu, pageset->pcp.high,
2617 pageset->pcp.batch, pageset->pcp.count);
2621 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2622 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2624 " dirty:%lu writeback:%lu unstable:%lu\n"
2625 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2626 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2627 global_page_state(NR_ACTIVE_ANON),
2628 global_page_state(NR_INACTIVE_ANON),
2629 global_page_state(NR_ISOLATED_ANON),
2630 global_page_state(NR_ACTIVE_FILE),
2631 global_page_state(NR_INACTIVE_FILE),
2632 global_page_state(NR_ISOLATED_FILE),
2633 global_page_state(NR_UNEVICTABLE),
2634 global_page_state(NR_FILE_DIRTY),
2635 global_page_state(NR_WRITEBACK),
2636 global_page_state(NR_UNSTABLE_NFS),
2637 global_page_state(NR_FREE_PAGES),
2638 global_page_state(NR_SLAB_RECLAIMABLE),
2639 global_page_state(NR_SLAB_UNRECLAIMABLE),
2640 global_page_state(NR_FILE_MAPPED),
2641 global_page_state(NR_SHMEM),
2642 global_page_state(NR_PAGETABLE),
2643 global_page_state(NR_BOUNCE));
2645 for_each_populated_zone(zone) {
2648 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2656 " active_anon:%lukB"
2657 " inactive_anon:%lukB"
2658 " active_file:%lukB"
2659 " inactive_file:%lukB"
2660 " unevictable:%lukB"
2661 " isolated(anon):%lukB"
2662 " isolated(file):%lukB"
2669 " slab_reclaimable:%lukB"
2670 " slab_unreclaimable:%lukB"
2671 " kernel_stack:%lukB"
2675 " writeback_tmp:%lukB"
2676 " pages_scanned:%lu"
2677 " all_unreclaimable? %s"
2680 K(zone_page_state(zone, NR_FREE_PAGES)),
2681 K(min_wmark_pages(zone)),
2682 K(low_wmark_pages(zone)),
2683 K(high_wmark_pages(zone)),
2684 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2685 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2686 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2687 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2688 K(zone_page_state(zone, NR_UNEVICTABLE)),
2689 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2690 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2691 K(zone->present_pages),
2692 K(zone_page_state(zone, NR_MLOCK)),
2693 K(zone_page_state(zone, NR_FILE_DIRTY)),
2694 K(zone_page_state(zone, NR_WRITEBACK)),
2695 K(zone_page_state(zone, NR_FILE_MAPPED)),
2696 K(zone_page_state(zone, NR_SHMEM)),
2697 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2698 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2699 zone_page_state(zone, NR_KERNEL_STACK) *
2701 K(zone_page_state(zone, NR_PAGETABLE)),
2702 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2703 K(zone_page_state(zone, NR_BOUNCE)),
2704 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2705 zone->pages_scanned,
2706 (zone->all_unreclaimable ? "yes" : "no")
2708 printk("lowmem_reserve[]:");
2709 for (i = 0; i < MAX_NR_ZONES; i++)
2710 printk(" %lu", zone->lowmem_reserve[i]);
2714 for_each_populated_zone(zone) {
2715 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2717 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2720 printk("%s: ", zone->name);
2722 spin_lock_irqsave(&zone->lock, flags);
2723 for (order = 0; order < MAX_ORDER; order++) {
2724 nr[order] = zone->free_area[order].nr_free;
2725 total += nr[order] << order;
2727 spin_unlock_irqrestore(&zone->lock, flags);
2728 for (order = 0; order < MAX_ORDER; order++)
2729 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2730 printk("= %lukB\n", K(total));
2733 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2735 show_swap_cache_info();
2738 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2740 zoneref->zone = zone;
2741 zoneref->zone_idx = zone_idx(zone);
2745 * Builds allocation fallback zone lists.
2747 * Add all populated zones of a node to the zonelist.
2749 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2750 int nr_zones, enum zone_type zone_type)
2754 BUG_ON(zone_type >= MAX_NR_ZONES);
2759 zone = pgdat->node_zones + zone_type;
2760 if (populated_zone(zone)) {
2761 zoneref_set_zone(zone,
2762 &zonelist->_zonerefs[nr_zones++]);
2763 check_highest_zone(zone_type);
2766 } while (zone_type);
2773 * 0 = automatic detection of better ordering.
2774 * 1 = order by ([node] distance, -zonetype)
2775 * 2 = order by (-zonetype, [node] distance)
2777 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2778 * the same zonelist. So only NUMA can configure this param.
2780 #define ZONELIST_ORDER_DEFAULT 0
2781 #define ZONELIST_ORDER_NODE 1
2782 #define ZONELIST_ORDER_ZONE 2
2784 /* zonelist order in the kernel.
2785 * set_zonelist_order() will set this to NODE or ZONE.
2787 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2788 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2792 /* The value user specified ....changed by config */
2793 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2794 /* string for sysctl */
2795 #define NUMA_ZONELIST_ORDER_LEN 16
2796 char numa_zonelist_order[16] = "default";
2799 * interface for configure zonelist ordering.
2800 * command line option "numa_zonelist_order"
2801 * = "[dD]efault - default, automatic configuration.
2802 * = "[nN]ode - order by node locality, then by zone within node
2803 * = "[zZ]one - order by zone, then by locality within zone
2806 static int __parse_numa_zonelist_order(char *s)
2808 if (*s == 'd' || *s == 'D') {
2809 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2810 } else if (*s == 'n' || *s == 'N') {
2811 user_zonelist_order = ZONELIST_ORDER_NODE;
2812 } else if (*s == 'z' || *s == 'Z') {
2813 user_zonelist_order = ZONELIST_ORDER_ZONE;
2816 "Ignoring invalid numa_zonelist_order value: "
2823 static __init int setup_numa_zonelist_order(char *s)
2830 ret = __parse_numa_zonelist_order(s);
2832 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2836 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2839 * sysctl handler for numa_zonelist_order
2841 int numa_zonelist_order_handler(ctl_table *table, int write,
2842 void __user *buffer, size_t *length,
2845 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2847 static DEFINE_MUTEX(zl_order_mutex);
2849 mutex_lock(&zl_order_mutex);
2851 strcpy(saved_string, (char*)table->data);
2852 ret = proc_dostring(table, write, buffer, length, ppos);
2856 int oldval = user_zonelist_order;
2857 if (__parse_numa_zonelist_order((char*)table->data)) {
2859 * bogus value. restore saved string
2861 strncpy((char*)table->data, saved_string,
2862 NUMA_ZONELIST_ORDER_LEN);
2863 user_zonelist_order = oldval;
2864 } else if (oldval != user_zonelist_order) {
2865 mutex_lock(&zonelists_mutex);
2866 build_all_zonelists(NULL);
2867 mutex_unlock(&zonelists_mutex);
2871 mutex_unlock(&zl_order_mutex);
2876 #define MAX_NODE_LOAD (nr_online_nodes)
2877 static int node_load[MAX_NUMNODES];
2880 * find_next_best_node - find the next node that should appear in a given node's fallback list
2881 * @node: node whose fallback list we're appending
2882 * @used_node_mask: nodemask_t of already used nodes
2884 * We use a number of factors to determine which is the next node that should
2885 * appear on a given node's fallback list. The node should not have appeared
2886 * already in @node's fallback list, and it should be the next closest node
2887 * according to the distance array (which contains arbitrary distance values
2888 * from each node to each node in the system), and should also prefer nodes
2889 * with no CPUs, since presumably they'll have very little allocation pressure
2890 * on them otherwise.
2891 * It returns -1 if no node is found.
2893 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2896 int min_val = INT_MAX;
2898 const struct cpumask *tmp = cpumask_of_node(0);
2900 /* Use the local node if we haven't already */
2901 if (!node_isset(node, *used_node_mask)) {
2902 node_set(node, *used_node_mask);
2906 for_each_node_state(n, N_HIGH_MEMORY) {
2908 /* Don't want a node to appear more than once */
2909 if (node_isset(n, *used_node_mask))
2912 /* Use the distance array to find the distance */
2913 val = node_distance(node, n);
2915 /* Penalize nodes under us ("prefer the next node") */
2918 /* Give preference to headless and unused nodes */
2919 tmp = cpumask_of_node(n);
2920 if (!cpumask_empty(tmp))
2921 val += PENALTY_FOR_NODE_WITH_CPUS;
2923 /* Slight preference for less loaded node */
2924 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2925 val += node_load[n];
2927 if (val < min_val) {
2934 node_set(best_node, *used_node_mask);
2941 * Build zonelists ordered by node and zones within node.
2942 * This results in maximum locality--normal zone overflows into local
2943 * DMA zone, if any--but risks exhausting DMA zone.
2945 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2948 struct zonelist *zonelist;
2950 zonelist = &pgdat->node_zonelists[0];
2951 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2953 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2955 zonelist->_zonerefs[j].zone = NULL;
2956 zonelist->_zonerefs[j].zone_idx = 0;
2960 * Build gfp_thisnode zonelists
2962 static void build_thisnode_zonelists(pg_data_t *pgdat)
2965 struct zonelist *zonelist;
2967 zonelist = &pgdat->node_zonelists[1];
2968 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2969 zonelist->_zonerefs[j].zone = NULL;
2970 zonelist->_zonerefs[j].zone_idx = 0;
2974 * Build zonelists ordered by zone and nodes within zones.
2975 * This results in conserving DMA zone[s] until all Normal memory is
2976 * exhausted, but results in overflowing to remote node while memory
2977 * may still exist in local DMA zone.
2979 static int node_order[MAX_NUMNODES];
2981 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2984 int zone_type; /* needs to be signed */
2986 struct zonelist *zonelist;
2988 zonelist = &pgdat->node_zonelists[0];
2990 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2991 for (j = 0; j < nr_nodes; j++) {
2992 node = node_order[j];
2993 z = &NODE_DATA(node)->node_zones[zone_type];
2994 if (populated_zone(z)) {
2996 &zonelist->_zonerefs[pos++]);
2997 check_highest_zone(zone_type);
3001 zonelist->_zonerefs[pos].zone = NULL;
3002 zonelist->_zonerefs[pos].zone_idx = 0;
3005 static int default_zonelist_order(void)
3008 unsigned long low_kmem_size,total_size;
3012 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3013 * If they are really small and used heavily, the system can fall
3014 * into OOM very easily.
3015 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3017 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3020 for_each_online_node(nid) {
3021 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3022 z = &NODE_DATA(nid)->node_zones[zone_type];
3023 if (populated_zone(z)) {
3024 if (zone_type < ZONE_NORMAL)
3025 low_kmem_size += z->present_pages;
3026 total_size += z->present_pages;
3027 } else if (zone_type == ZONE_NORMAL) {
3029 * If any node has only lowmem, then node order
3030 * is preferred to allow kernel allocations
3031 * locally; otherwise, they can easily infringe
3032 * on other nodes when there is an abundance of
3033 * lowmem available to allocate from.
3035 return ZONELIST_ORDER_NODE;
3039 if (!low_kmem_size || /* there are no DMA area. */
3040 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3041 return ZONELIST_ORDER_NODE;
3043 * look into each node's config.
3044 * If there is a node whose DMA/DMA32 memory is very big area on
3045 * local memory, NODE_ORDER may be suitable.
3047 average_size = total_size /
3048 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3049 for_each_online_node(nid) {
3052 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3053 z = &NODE_DATA(nid)->node_zones[zone_type];
3054 if (populated_zone(z)) {
3055 if (zone_type < ZONE_NORMAL)
3056 low_kmem_size += z->present_pages;
3057 total_size += z->present_pages;
3060 if (low_kmem_size &&
3061 total_size > average_size && /* ignore small node */
3062 low_kmem_size > total_size * 70/100)
3063 return ZONELIST_ORDER_NODE;
3065 return ZONELIST_ORDER_ZONE;
3068 static void set_zonelist_order(void)
3070 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3071 current_zonelist_order = default_zonelist_order();
3073 current_zonelist_order = user_zonelist_order;
3076 static void build_zonelists(pg_data_t *pgdat)
3080 nodemask_t used_mask;
3081 int local_node, prev_node;
3082 struct zonelist *zonelist;
3083 int order = current_zonelist_order;
3085 /* initialize zonelists */
3086 for (i = 0; i < MAX_ZONELISTS; i++) {
3087 zonelist = pgdat->node_zonelists + i;
3088 zonelist->_zonerefs[0].zone = NULL;
3089 zonelist->_zonerefs[0].zone_idx = 0;
3092 /* NUMA-aware ordering of nodes */
3093 local_node = pgdat->node_id;
3094 load = nr_online_nodes;
3095 prev_node = local_node;
3096 nodes_clear(used_mask);
3098 memset(node_order, 0, sizeof(node_order));
3101 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3102 int distance = node_distance(local_node, node);
3105 * If another node is sufficiently far away then it is better
3106 * to reclaim pages in a zone before going off node.
3108 if (distance > RECLAIM_DISTANCE)
3109 zone_reclaim_mode = 1;
3112 * We don't want to pressure a particular node.
3113 * So adding penalty to the first node in same
3114 * distance group to make it round-robin.
3116 if (distance != node_distance(local_node, prev_node))
3117 node_load[node] = load;
3121 if (order == ZONELIST_ORDER_NODE)
3122 build_zonelists_in_node_order(pgdat, node);
3124 node_order[j++] = node; /* remember order */
3127 if (order == ZONELIST_ORDER_ZONE) {
3128 /* calculate node order -- i.e., DMA last! */
3129 build_zonelists_in_zone_order(pgdat, j);
3132 build_thisnode_zonelists(pgdat);
3135 /* Construct the zonelist performance cache - see further mmzone.h */
3136 static void build_zonelist_cache(pg_data_t *pgdat)
3138 struct zonelist *zonelist;
3139 struct zonelist_cache *zlc;
3142 zonelist = &pgdat->node_zonelists[0];
3143 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3144 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3145 for (z = zonelist->_zonerefs; z->zone; z++)
3146 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3149 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3151 * Return node id of node used for "local" allocations.
3152 * I.e., first node id of first zone in arg node's generic zonelist.
3153 * Used for initializing percpu 'numa_mem', which is used primarily
3154 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3156 int local_memory_node(int node)
3160 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3161 gfp_zone(GFP_KERNEL),
3168 #else /* CONFIG_NUMA */
3170 static void set_zonelist_order(void)
3172 current_zonelist_order = ZONELIST_ORDER_ZONE;
3175 static void build_zonelists(pg_data_t *pgdat)
3177 int node, local_node;
3179 struct zonelist *zonelist;
3181 local_node = pgdat->node_id;
3183 zonelist = &pgdat->node_zonelists[0];
3184 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3187 * Now we build the zonelist so that it contains the zones
3188 * of all the other nodes.
3189 * We don't want to pressure a particular node, so when
3190 * building the zones for node N, we make sure that the
3191 * zones coming right after the local ones are those from
3192 * node N+1 (modulo N)
3194 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3195 if (!node_online(node))
3197 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3200 for (node = 0; node < local_node; node++) {
3201 if (!node_online(node))
3203 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3207 zonelist->_zonerefs[j].zone = NULL;
3208 zonelist->_zonerefs[j].zone_idx = 0;
3211 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3212 static void build_zonelist_cache(pg_data_t *pgdat)
3214 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3217 #endif /* CONFIG_NUMA */
3220 * Boot pageset table. One per cpu which is going to be used for all
3221 * zones and all nodes. The parameters will be set in such a way
3222 * that an item put on a list will immediately be handed over to
3223 * the buddy list. This is safe since pageset manipulation is done
3224 * with interrupts disabled.
3226 * The boot_pagesets must be kept even after bootup is complete for
3227 * unused processors and/or zones. They do play a role for bootstrapping
3228 * hotplugged processors.
3230 * zoneinfo_show() and maybe other functions do
3231 * not check if the processor is online before following the pageset pointer.
3232 * Other parts of the kernel may not check if the zone is available.
3234 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3235 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3236 static void setup_zone_pageset(struct zone *zone);
3239 * Global mutex to protect against size modification of zonelists
3240 * as well as to serialize pageset setup for the new populated zone.
3242 DEFINE_MUTEX(zonelists_mutex);
3244 /* return values int ....just for stop_machine() */
3245 static __init_refok int __build_all_zonelists(void *data)
3251 memset(node_load, 0, sizeof(node_load));
3253 for_each_online_node(nid) {
3254 pg_data_t *pgdat = NODE_DATA(nid);
3256 build_zonelists(pgdat);
3257 build_zonelist_cache(pgdat);
3261 * Initialize the boot_pagesets that are going to be used
3262 * for bootstrapping processors. The real pagesets for
3263 * each zone will be allocated later when the per cpu
3264 * allocator is available.
3266 * boot_pagesets are used also for bootstrapping offline
3267 * cpus if the system is already booted because the pagesets
3268 * are needed to initialize allocators on a specific cpu too.
3269 * F.e. the percpu allocator needs the page allocator which
3270 * needs the percpu allocator in order to allocate its pagesets
3271 * (a chicken-egg dilemma).
3273 for_each_possible_cpu(cpu) {
3274 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3276 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3278 * We now know the "local memory node" for each node--
3279 * i.e., the node of the first zone in the generic zonelist.
3280 * Set up numa_mem percpu variable for on-line cpus. During
3281 * boot, only the boot cpu should be on-line; we'll init the
3282 * secondary cpus' numa_mem as they come on-line. During
3283 * node/memory hotplug, we'll fixup all on-line cpus.
3285 if (cpu_online(cpu))
3286 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3294 * Called with zonelists_mutex held always
3295 * unless system_state == SYSTEM_BOOTING.
3297 void __ref build_all_zonelists(void *data)
3299 set_zonelist_order();
3301 if (system_state == SYSTEM_BOOTING) {
3302 __build_all_zonelists(NULL);
3303 mminit_verify_zonelist();
3304 cpuset_init_current_mems_allowed();
3306 /* we have to stop all cpus to guarantee there is no user
3308 #ifdef CONFIG_MEMORY_HOTPLUG
3310 setup_zone_pageset((struct zone *)data);
3312 stop_machine(__build_all_zonelists, NULL, NULL);
3313 /* cpuset refresh routine should be here */
3315 vm_total_pages = nr_free_pagecache_pages();
3317 * Disable grouping by mobility if the number of pages in the
3318 * system is too low to allow the mechanism to work. It would be
3319 * more accurate, but expensive to check per-zone. This check is
3320 * made on memory-hotadd so a system can start with mobility
3321 * disabled and enable it later
3323 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3324 page_group_by_mobility_disabled = 1;
3326 page_group_by_mobility_disabled = 0;
3328 printk("Built %i zonelists in %s order, mobility grouping %s. "
3329 "Total pages: %ld\n",
3331 zonelist_order_name[current_zonelist_order],
3332 page_group_by_mobility_disabled ? "off" : "on",
3335 printk("Policy zone: %s\n", zone_names[policy_zone]);
3340 * Helper functions to size the waitqueue hash table.
3341 * Essentially these want to choose hash table sizes sufficiently
3342 * large so that collisions trying to wait on pages are rare.
3343 * But in fact, the number of active page waitqueues on typical
3344 * systems is ridiculously low, less than 200. So this is even
3345 * conservative, even though it seems large.
3347 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3348 * waitqueues, i.e. the size of the waitq table given the number of pages.
3350 #define PAGES_PER_WAITQUEUE 256
3352 #ifndef CONFIG_MEMORY_HOTPLUG
3353 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3355 unsigned long size = 1;
3357 pages /= PAGES_PER_WAITQUEUE;
3359 while (size < pages)
3363 * Once we have dozens or even hundreds of threads sleeping
3364 * on IO we've got bigger problems than wait queue collision.
3365 * Limit the size of the wait table to a reasonable size.
3367 size = min(size, 4096UL);
3369 return max(size, 4UL);
3373 * A zone's size might be changed by hot-add, so it is not possible to determine
3374 * a suitable size for its wait_table. So we use the maximum size now.
3376 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3378 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3379 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3380 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3382 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3383 * or more by the traditional way. (See above). It equals:
3385 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3386 * ia64(16K page size) : = ( 8G + 4M)byte.
3387 * powerpc (64K page size) : = (32G +16M)byte.
3389 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3396 * This is an integer logarithm so that shifts can be used later
3397 * to extract the more random high bits from the multiplicative
3398 * hash function before the remainder is taken.
3400 static inline unsigned long wait_table_bits(unsigned long size)
3405 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3408 * Check if a pageblock contains reserved pages
3410 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3414 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3415 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3422 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3423 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3424 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3425 * higher will lead to a bigger reserve which will get freed as contiguous
3426 * blocks as reclaim kicks in
3428 static void setup_zone_migrate_reserve(struct zone *zone)
3430 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3432 unsigned long block_migratetype;
3436 * Get the start pfn, end pfn and the number of blocks to reserve
3437 * We have to be careful to be aligned to pageblock_nr_pages to
3438 * make sure that we always check pfn_valid for the first page in
3441 start_pfn = zone->zone_start_pfn;
3442 end_pfn = start_pfn + zone->spanned_pages;
3443 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3444 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3448 * Reserve blocks are generally in place to help high-order atomic
3449 * allocations that are short-lived. A min_free_kbytes value that
3450 * would result in more than 2 reserve blocks for atomic allocations
3451 * is assumed to be in place to help anti-fragmentation for the
3452 * future allocation of hugepages at runtime.
3454 reserve = min(2, reserve);
3456 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3457 if (!pfn_valid(pfn))
3459 page = pfn_to_page(pfn);
3461 /* Watch out for overlapping nodes */
3462 if (page_to_nid(page) != zone_to_nid(zone))
3465 block_migratetype = get_pageblock_migratetype(page);
3467 /* Only test what is necessary when the reserves are not met */
3470 * Blocks with reserved pages will never free, skip
3473 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3474 if (pageblock_is_reserved(pfn, block_end_pfn))
3477 /* If this block is reserved, account for it */
3478 if (block_migratetype == MIGRATE_RESERVE) {
3483 /* Suitable for reserving if this block is movable */
3484 if (block_migratetype == MIGRATE_MOVABLE) {
3485 set_pageblock_migratetype(page,
3487 move_freepages_block(zone, page,
3495 * If the reserve is met and this is a previous reserved block,
3498 if (block_migratetype == MIGRATE_RESERVE) {
3499 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3500 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3506 * Initially all pages are reserved - free ones are freed
3507 * up by free_all_bootmem() once the early boot process is
3508 * done. Non-atomic initialization, single-pass.
3510 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3511 unsigned long start_pfn, enum memmap_context context)
3514 unsigned long end_pfn = start_pfn + size;
3518 if (highest_memmap_pfn < end_pfn - 1)
3519 highest_memmap_pfn = end_pfn - 1;
3521 z = &NODE_DATA(nid)->node_zones[zone];
3522 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3524 * There can be holes in boot-time mem_map[]s
3525 * handed to this function. They do not
3526 * exist on hotplugged memory.
3528 if (context == MEMMAP_EARLY) {
3529 if (!early_pfn_valid(pfn))
3531 if (!early_pfn_in_nid(pfn, nid))
3534 page = pfn_to_page(pfn);
3535 set_page_links(page, zone, nid, pfn);
3536 mminit_verify_page_links(page, zone, nid, pfn);
3537 init_page_count(page);
3538 reset_page_mapcount(page);
3539 SetPageReserved(page);
3541 * Mark the block movable so that blocks are reserved for
3542 * movable at startup. This will force kernel allocations
3543 * to reserve their blocks rather than leaking throughout
3544 * the address space during boot when many long-lived
3545 * kernel allocations are made. Later some blocks near
3546 * the start are marked MIGRATE_RESERVE by
3547 * setup_zone_migrate_reserve()
3549 * bitmap is created for zone's valid pfn range. but memmap
3550 * can be created for invalid pages (for alignment)
3551 * check here not to call set_pageblock_migratetype() against
3554 if ((z->zone_start_pfn <= pfn)
3555 && (pfn < z->zone_start_pfn + z->spanned_pages)
3556 && !(pfn & (pageblock_nr_pages - 1)))
3557 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3559 INIT_LIST_HEAD(&page->lru);
3560 #ifdef WANT_PAGE_VIRTUAL
3561 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3562 if (!is_highmem_idx(zone))
3563 set_page_address(page, __va(pfn << PAGE_SHIFT));
3568 static void __meminit zone_init_free_lists(struct zone *zone)
3571 for_each_migratetype_order(order, t) {
3572 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3573 zone->free_area[order].nr_free = 0;
3577 #ifndef __HAVE_ARCH_MEMMAP_INIT
3578 #define memmap_init(size, nid, zone, start_pfn) \
3579 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3582 static int zone_batchsize(struct zone *zone)
3588 * The per-cpu-pages pools are set to around 1000th of the
3589 * size of the zone. But no more than 1/2 of a meg.
3591 * OK, so we don't know how big the cache is. So guess.
3593 batch = zone->present_pages / 1024;
3594 if (batch * PAGE_SIZE > 512 * 1024)
3595 batch = (512 * 1024) / PAGE_SIZE;
3596 batch /= 4; /* We effectively *= 4 below */
3601 * Clamp the batch to a 2^n - 1 value. Having a power
3602 * of 2 value was found to be more likely to have
3603 * suboptimal cache aliasing properties in some cases.
3605 * For example if 2 tasks are alternately allocating
3606 * batches of pages, one task can end up with a lot
3607 * of pages of one half of the possible page colors
3608 * and the other with pages of the other colors.
3610 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3615 /* The deferral and batching of frees should be suppressed under NOMMU
3618 * The problem is that NOMMU needs to be able to allocate large chunks
3619 * of contiguous memory as there's no hardware page translation to
3620 * assemble apparent contiguous memory from discontiguous pages.
3622 * Queueing large contiguous runs of pages for batching, however,
3623 * causes the pages to actually be freed in smaller chunks. As there
3624 * can be a significant delay between the individual batches being
3625 * recycled, this leads to the once large chunks of space being
3626 * fragmented and becoming unavailable for high-order allocations.
3632 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3634 struct per_cpu_pages *pcp;
3637 memset(p, 0, sizeof(*p));
3641 pcp->high = 6 * batch;
3642 pcp->batch = max(1UL, 1 * batch);
3643 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3644 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3648 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3649 * to the value high for the pageset p.
3652 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3655 struct per_cpu_pages *pcp;
3659 pcp->batch = max(1UL, high/4);
3660 if ((high/4) > (PAGE_SHIFT * 8))
3661 pcp->batch = PAGE_SHIFT * 8;
3664 static void setup_zone_pageset(struct zone *zone)
3668 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3670 for_each_possible_cpu(cpu) {
3671 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3673 setup_pageset(pcp, zone_batchsize(zone));
3675 if (percpu_pagelist_fraction)
3676 setup_pagelist_highmark(pcp,
3677 (zone->present_pages /
3678 percpu_pagelist_fraction));
3683 * Allocate per cpu pagesets and initialize them.
3684 * Before this call only boot pagesets were available.
3686 void __init setup_per_cpu_pageset(void)
3690 for_each_populated_zone(zone)
3691 setup_zone_pageset(zone);
3694 static noinline __init_refok
3695 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3698 struct pglist_data *pgdat = zone->zone_pgdat;
3702 * The per-page waitqueue mechanism uses hashed waitqueues
3705 zone->wait_table_hash_nr_entries =
3706 wait_table_hash_nr_entries(zone_size_pages);
3707 zone->wait_table_bits =
3708 wait_table_bits(zone->wait_table_hash_nr_entries);
3709 alloc_size = zone->wait_table_hash_nr_entries
3710 * sizeof(wait_queue_head_t);
3712 if (!slab_is_available()) {
3713 zone->wait_table = (wait_queue_head_t *)
3714 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3717 * This case means that a zone whose size was 0 gets new memory
3718 * via memory hot-add.
3719 * But it may be the case that a new node was hot-added. In
3720 * this case vmalloc() will not be able to use this new node's
3721 * memory - this wait_table must be initialized to use this new
3722 * node itself as well.
3723 * To use this new node's memory, further consideration will be
3726 zone->wait_table = vmalloc(alloc_size);
3728 if (!zone->wait_table)
3731 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3732 init_waitqueue_head(zone->wait_table + i);
3737 static int __zone_pcp_update(void *data)
3739 struct zone *zone = data;
3741 unsigned long batch = zone_batchsize(zone), flags;
3743 for_each_possible_cpu(cpu) {
3744 struct per_cpu_pageset *pset;
3745 struct per_cpu_pages *pcp;
3747 pset = per_cpu_ptr(zone->pageset, cpu);
3750 local_irq_save(flags);
3751 free_pcppages_bulk(zone, pcp->count, pcp);
3752 setup_pageset(pset, batch);
3753 local_irq_restore(flags);
3758 void zone_pcp_update(struct zone *zone)
3760 stop_machine(__zone_pcp_update, zone, NULL);
3763 static __meminit void zone_pcp_init(struct zone *zone)
3766 * per cpu subsystem is not up at this point. The following code
3767 * relies on the ability of the linker to provide the
3768 * offset of a (static) per cpu variable into the per cpu area.
3770 zone->pageset = &boot_pageset;
3772 if (zone->present_pages)
3773 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3774 zone->name, zone->present_pages,
3775 zone_batchsize(zone));
3778 __meminit int init_currently_empty_zone(struct zone *zone,
3779 unsigned long zone_start_pfn,
3781 enum memmap_context context)
3783 struct pglist_data *pgdat = zone->zone_pgdat;
3785 ret = zone_wait_table_init(zone, size);
3788 pgdat->nr_zones = zone_idx(zone) + 1;
3790 zone->zone_start_pfn = zone_start_pfn;
3792 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3793 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3795 (unsigned long)zone_idx(zone),
3796 zone_start_pfn, (zone_start_pfn + size));
3798 zone_init_free_lists(zone);
3803 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3805 * Basic iterator support. Return the first range of PFNs for a node
3806 * Note: nid == MAX_NUMNODES returns first region regardless of node
3808 static int __meminit first_active_region_index_in_nid(int nid)
3812 for (i = 0; i < nr_nodemap_entries; i++)
3813 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3820 * Basic iterator support. Return the next active range of PFNs for a node
3821 * Note: nid == MAX_NUMNODES returns next region regardless of node
3823 static int __meminit next_active_region_index_in_nid(int index, int nid)
3825 for (index = index + 1; index < nr_nodemap_entries; index++)
3826 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3832 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3834 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3835 * Architectures may implement their own version but if add_active_range()
3836 * was used and there are no special requirements, this is a convenient
3839 int __meminit __early_pfn_to_nid(unsigned long pfn)
3843 for (i = 0; i < nr_nodemap_entries; i++) {
3844 unsigned long start_pfn = early_node_map[i].start_pfn;
3845 unsigned long end_pfn = early_node_map[i].end_pfn;
3847 if (start_pfn <= pfn && pfn < end_pfn)
3848 return early_node_map[i].nid;
3850 /* This is a memory hole */
3853 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3855 int __meminit early_pfn_to_nid(unsigned long pfn)
3859 nid = __early_pfn_to_nid(pfn);
3862 /* just returns 0 */
3866 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3867 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3871 nid = __early_pfn_to_nid(pfn);
3872 if (nid >= 0 && nid != node)
3878 /* Basic iterator support to walk early_node_map[] */
3879 #define for_each_active_range_index_in_nid(i, nid) \
3880 for (i = first_active_region_index_in_nid(nid); i != -1; \
3881 i = next_active_region_index_in_nid(i, nid))
3884 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3885 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3886 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3888 * If an architecture guarantees that all ranges registered with
3889 * add_active_ranges() contain no holes and may be freed, this
3890 * this function may be used instead of calling free_bootmem() manually.
3892 void __init free_bootmem_with_active_regions(int nid,
3893 unsigned long max_low_pfn)
3897 for_each_active_range_index_in_nid(i, nid) {
3898 unsigned long size_pages = 0;
3899 unsigned long end_pfn = early_node_map[i].end_pfn;
3901 if (early_node_map[i].start_pfn >= max_low_pfn)
3904 if (end_pfn > max_low_pfn)
3905 end_pfn = max_low_pfn;
3907 size_pages = end_pfn - early_node_map[i].start_pfn;
3908 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3909 PFN_PHYS(early_node_map[i].start_pfn),
3910 size_pages << PAGE_SHIFT);
3914 #ifdef CONFIG_HAVE_MEMBLOCK
3916 * Basic iterator support. Return the last range of PFNs for a node
3917 * Note: nid == MAX_NUMNODES returns last region regardless of node
3919 static int __meminit last_active_region_index_in_nid(int nid)
3923 for (i = nr_nodemap_entries - 1; i >= 0; i--)
3924 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3931 * Basic iterator support. Return the previous active range of PFNs for a node
3932 * Note: nid == MAX_NUMNODES returns next region regardless of node
3934 static int __meminit previous_active_region_index_in_nid(int index, int nid)
3936 for (index = index - 1; index >= 0; index--)
3937 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3943 #define for_each_active_range_index_in_nid_reverse(i, nid) \
3944 for (i = last_active_region_index_in_nid(nid); i != -1; \
3945 i = previous_active_region_index_in_nid(i, nid))
3947 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3948 u64 goal, u64 limit)
3952 /* Need to go over early_node_map to find out good range for node */
3953 for_each_active_range_index_in_nid_reverse(i, nid) {
3955 u64 ei_start, ei_last;
3956 u64 final_start, final_end;
3958 ei_last = early_node_map[i].end_pfn;
3959 ei_last <<= PAGE_SHIFT;
3960 ei_start = early_node_map[i].start_pfn;
3961 ei_start <<= PAGE_SHIFT;
3963 final_start = max(ei_start, goal);
3964 final_end = min(ei_last, limit);
3966 if (final_start >= final_end)
3969 addr = memblock_find_in_range(final_start, final_end, size, align);
3971 if (addr == MEMBLOCK_ERROR)
3977 return MEMBLOCK_ERROR;
3981 int __init add_from_early_node_map(struct range *range, int az,
3982 int nr_range, int nid)
3987 /* need to go over early_node_map to find out good range for node */
3988 for_each_active_range_index_in_nid(i, nid) {
3989 start = early_node_map[i].start_pfn;
3990 end = early_node_map[i].end_pfn;
3991 nr_range = add_range(range, az, nr_range, start, end);
3996 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
4001 for_each_active_range_index_in_nid(i, nid) {
4002 ret = work_fn(early_node_map[i].start_pfn,
4003 early_node_map[i].end_pfn, data);
4009 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4010 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4012 * If an architecture guarantees that all ranges registered with
4013 * add_active_ranges() contain no holes and may be freed, this
4014 * function may be used instead of calling memory_present() manually.
4016 void __init sparse_memory_present_with_active_regions(int nid)
4020 for_each_active_range_index_in_nid(i, nid)
4021 memory_present(early_node_map[i].nid,
4022 early_node_map[i].start_pfn,
4023 early_node_map[i].end_pfn);
4027 * get_pfn_range_for_nid - Return the start and end page frames for a node
4028 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4029 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4030 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4032 * It returns the start and end page frame of a node based on information
4033 * provided by an arch calling add_active_range(). If called for a node
4034 * with no available memory, a warning is printed and the start and end
4037 void __meminit get_pfn_range_for_nid(unsigned int nid,
4038 unsigned long *start_pfn, unsigned long *end_pfn)
4044 for_each_active_range_index_in_nid(i, nid) {
4045 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
4046 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
4049 if (*start_pfn == -1UL)
4054 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4055 * assumption is made that zones within a node are ordered in monotonic
4056 * increasing memory addresses so that the "highest" populated zone is used
4058 static void __init find_usable_zone_for_movable(void)
4061 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4062 if (zone_index == ZONE_MOVABLE)
4065 if (arch_zone_highest_possible_pfn[zone_index] >
4066 arch_zone_lowest_possible_pfn[zone_index])
4070 VM_BUG_ON(zone_index == -1);
4071 movable_zone = zone_index;
4075 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4076 * because it is sized independent of architecture. Unlike the other zones,
4077 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4078 * in each node depending on the size of each node and how evenly kernelcore
4079 * is distributed. This helper function adjusts the zone ranges
4080 * provided by the architecture for a given node by using the end of the
4081 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4082 * zones within a node are in order of monotonic increases memory addresses
4084 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4085 unsigned long zone_type,
4086 unsigned long node_start_pfn,
4087 unsigned long node_end_pfn,
4088 unsigned long *zone_start_pfn,
4089 unsigned long *zone_end_pfn)
4091 /* Only adjust if ZONE_MOVABLE is on this node */
4092 if (zone_movable_pfn[nid]) {
4093 /* Size ZONE_MOVABLE */
4094 if (zone_type == ZONE_MOVABLE) {
4095 *zone_start_pfn = zone_movable_pfn[nid];
4096 *zone_end_pfn = min(node_end_pfn,
4097 arch_zone_highest_possible_pfn[movable_zone]);
4099 /* Adjust for ZONE_MOVABLE starting within this range */
4100 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4101 *zone_end_pfn > zone_movable_pfn[nid]) {
4102 *zone_end_pfn = zone_movable_pfn[nid];
4104 /* Check if this whole range is within ZONE_MOVABLE */
4105 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4106 *zone_start_pfn = *zone_end_pfn;
4111 * Return the number of pages a zone spans in a node, including holes
4112 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4114 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4115 unsigned long zone_type,
4116 unsigned long *ignored)
4118 unsigned long node_start_pfn, node_end_pfn;
4119 unsigned long zone_start_pfn, zone_end_pfn;
4121 /* Get the start and end of the node and zone */
4122 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4123 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4124 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4125 adjust_zone_range_for_zone_movable(nid, zone_type,
4126 node_start_pfn, node_end_pfn,
4127 &zone_start_pfn, &zone_end_pfn);
4129 /* Check that this node has pages within the zone's required range */
4130 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4133 /* Move the zone boundaries inside the node if necessary */
4134 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4135 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4137 /* Return the spanned pages */
4138 return zone_end_pfn - zone_start_pfn;
4142 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4143 * then all holes in the requested range will be accounted for.
4145 unsigned long __meminit __absent_pages_in_range(int nid,
4146 unsigned long range_start_pfn,
4147 unsigned long range_end_pfn)
4150 unsigned long prev_end_pfn = 0, hole_pages = 0;
4151 unsigned long start_pfn;
4153 /* Find the end_pfn of the first active range of pfns in the node */
4154 i = first_active_region_index_in_nid(nid);
4158 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4160 /* Account for ranges before physical memory on this node */
4161 if (early_node_map[i].start_pfn > range_start_pfn)
4162 hole_pages = prev_end_pfn - range_start_pfn;
4164 /* Find all holes for the zone within the node */
4165 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4167 /* No need to continue if prev_end_pfn is outside the zone */
4168 if (prev_end_pfn >= range_end_pfn)
4171 /* Make sure the end of the zone is not within the hole */
4172 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4173 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4175 /* Update the hole size cound and move on */
4176 if (start_pfn > range_start_pfn) {
4177 BUG_ON(prev_end_pfn > start_pfn);
4178 hole_pages += start_pfn - prev_end_pfn;
4180 prev_end_pfn = early_node_map[i].end_pfn;
4183 /* Account for ranges past physical memory on this node */
4184 if (range_end_pfn > prev_end_pfn)
4185 hole_pages += range_end_pfn -
4186 max(range_start_pfn, prev_end_pfn);
4192 * absent_pages_in_range - Return number of page frames in holes within a range
4193 * @start_pfn: The start PFN to start searching for holes
4194 * @end_pfn: The end PFN to stop searching for holes
4196 * It returns the number of pages frames in memory holes within a range.
4198 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4199 unsigned long end_pfn)
4201 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4204 /* Return the number of page frames in holes in a zone on a node */
4205 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4206 unsigned long zone_type,
4207 unsigned long *ignored)
4209 unsigned long node_start_pfn, node_end_pfn;
4210 unsigned long zone_start_pfn, zone_end_pfn;
4212 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4213 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4215 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4218 adjust_zone_range_for_zone_movable(nid, zone_type,
4219 node_start_pfn, node_end_pfn,
4220 &zone_start_pfn, &zone_end_pfn);
4221 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4225 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4226 unsigned long zone_type,
4227 unsigned long *zones_size)
4229 return zones_size[zone_type];
4232 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4233 unsigned long zone_type,
4234 unsigned long *zholes_size)
4239 return zholes_size[zone_type];
4244 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4245 unsigned long *zones_size, unsigned long *zholes_size)
4247 unsigned long realtotalpages, totalpages = 0;
4250 for (i = 0; i < MAX_NR_ZONES; i++)
4251 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4253 pgdat->node_spanned_pages = totalpages;
4255 realtotalpages = totalpages;
4256 for (i = 0; i < MAX_NR_ZONES; i++)
4258 zone_absent_pages_in_node(pgdat->node_id, i,
4260 pgdat->node_present_pages = realtotalpages;
4261 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4265 #ifndef CONFIG_SPARSEMEM
4267 * Calculate the size of the zone->blockflags rounded to an unsigned long
4268 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4269 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4270 * round what is now in bits to nearest long in bits, then return it in
4273 static unsigned long __init usemap_size(unsigned long zonesize)
4275 unsigned long usemapsize;
4277 usemapsize = roundup(zonesize, pageblock_nr_pages);
4278 usemapsize = usemapsize >> pageblock_order;
4279 usemapsize *= NR_PAGEBLOCK_BITS;
4280 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4282 return usemapsize / 8;
4285 static void __init setup_usemap(struct pglist_data *pgdat,
4286 struct zone *zone, unsigned long zonesize)
4288 unsigned long usemapsize = usemap_size(zonesize);
4289 zone->pageblock_flags = NULL;
4291 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4295 static inline void setup_usemap(struct pglist_data *pgdat,
4296 struct zone *zone, unsigned long zonesize) {}
4297 #endif /* CONFIG_SPARSEMEM */
4299 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4301 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4302 void __init set_pageblock_order(void)
4306 /* Check that pageblock_nr_pages has not already been setup */
4307 if (pageblock_order)
4310 if (HPAGE_SHIFT > PAGE_SHIFT)
4311 order = HUGETLB_PAGE_ORDER;
4313 order = MAX_ORDER - 1;
4316 * Assume the largest contiguous order of interest is a huge page.
4317 * This value may be variable depending on boot parameters on IA64 and
4320 pageblock_order = order;
4322 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4325 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4326 * is unused as pageblock_order is set at compile-time. See
4327 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4330 void __init set_pageblock_order(void)
4334 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4337 * Set up the zone data structures:
4338 * - mark all pages reserved
4339 * - mark all memory queues empty
4340 * - clear the memory bitmaps
4342 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4343 unsigned long *zones_size, unsigned long *zholes_size)
4346 int nid = pgdat->node_id;
4347 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4350 pgdat_resize_init(pgdat);
4351 pgdat->nr_zones = 0;
4352 init_waitqueue_head(&pgdat->kswapd_wait);
4353 pgdat->kswapd_max_order = 0;
4354 pgdat_page_cgroup_init(pgdat);
4356 for (j = 0; j < MAX_NR_ZONES; j++) {
4357 struct zone *zone = pgdat->node_zones + j;
4358 unsigned long size, realsize, memmap_pages;
4361 size = zone_spanned_pages_in_node(nid, j, zones_size);
4362 realsize = size - zone_absent_pages_in_node(nid, j,
4366 * Adjust realsize so that it accounts for how much memory
4367 * is used by this zone for memmap. This affects the watermark
4368 * and per-cpu initialisations
4371 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4372 if (realsize >= memmap_pages) {
4373 realsize -= memmap_pages;
4376 " %s zone: %lu pages used for memmap\n",
4377 zone_names[j], memmap_pages);
4380 " %s zone: %lu pages exceeds realsize %lu\n",
4381 zone_names[j], memmap_pages, realsize);
4383 /* Account for reserved pages */
4384 if (j == 0 && realsize > dma_reserve) {
4385 realsize -= dma_reserve;
4386 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4387 zone_names[0], dma_reserve);
4390 if (!is_highmem_idx(j))
4391 nr_kernel_pages += realsize;
4392 nr_all_pages += realsize;
4394 zone->spanned_pages = size;
4395 zone->present_pages = realsize;
4398 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4400 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4402 zone->name = zone_names[j];
4403 spin_lock_init(&zone->lock);
4404 spin_lock_init(&zone->lru_lock);
4405 zone_seqlock_init(zone);
4406 zone->zone_pgdat = pgdat;
4408 zone_pcp_init(zone);
4410 INIT_LIST_HEAD(&zone->lru[l].list);
4411 zone->reclaim_stat.recent_rotated[0] = 0;
4412 zone->reclaim_stat.recent_rotated[1] = 0;
4413 zone->reclaim_stat.recent_scanned[0] = 0;
4414 zone->reclaim_stat.recent_scanned[1] = 0;
4415 zap_zone_vm_stats(zone);
4420 set_pageblock_order();
4421 setup_usemap(pgdat, zone, size);
4422 ret = init_currently_empty_zone(zone, zone_start_pfn,
4423 size, MEMMAP_EARLY);
4425 memmap_init(size, nid, j, zone_start_pfn);
4426 zone_start_pfn += size;
4430 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4432 /* Skip empty nodes */
4433 if (!pgdat->node_spanned_pages)
4436 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4437 /* ia64 gets its own node_mem_map, before this, without bootmem */
4438 if (!pgdat->node_mem_map) {
4439 unsigned long size, start, end;
4443 * The zone's endpoints aren't required to be MAX_ORDER
4444 * aligned but the node_mem_map endpoints must be in order
4445 * for the buddy allocator to function correctly.
4447 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4448 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4449 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4450 size = (end - start) * sizeof(struct page);
4451 map = alloc_remap(pgdat->node_id, size);
4453 map = alloc_bootmem_node_nopanic(pgdat, size);
4454 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4456 #ifndef CONFIG_NEED_MULTIPLE_NODES
4458 * With no DISCONTIG, the global mem_map is just set as node 0's
4460 if (pgdat == NODE_DATA(0)) {
4461 mem_map = NODE_DATA(0)->node_mem_map;
4462 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4463 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4464 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4465 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4468 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4471 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4472 unsigned long node_start_pfn, unsigned long *zholes_size)
4474 pg_data_t *pgdat = NODE_DATA(nid);
4476 pgdat->node_id = nid;
4477 pgdat->node_start_pfn = node_start_pfn;
4478 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4480 alloc_node_mem_map(pgdat);
4481 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4482 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4483 nid, (unsigned long)pgdat,
4484 (unsigned long)pgdat->node_mem_map);
4487 free_area_init_core(pgdat, zones_size, zholes_size);
4490 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4492 #if MAX_NUMNODES > 1
4494 * Figure out the number of possible node ids.
4496 static void __init setup_nr_node_ids(void)
4499 unsigned int highest = 0;
4501 for_each_node_mask(node, node_possible_map)
4503 nr_node_ids = highest + 1;
4506 static inline void setup_nr_node_ids(void)
4512 * add_active_range - Register a range of PFNs backed by physical memory
4513 * @nid: The node ID the range resides on
4514 * @start_pfn: The start PFN of the available physical memory
4515 * @end_pfn: The end PFN of the available physical memory
4517 * These ranges are stored in an early_node_map[] and later used by
4518 * free_area_init_nodes() to calculate zone sizes and holes. If the
4519 * range spans a memory hole, it is up to the architecture to ensure
4520 * the memory is not freed by the bootmem allocator. If possible
4521 * the range being registered will be merged with existing ranges.
4523 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4524 unsigned long end_pfn)
4528 mminit_dprintk(MMINIT_TRACE, "memory_register",
4529 "Entering add_active_range(%d, %#lx, %#lx) "
4530 "%d entries of %d used\n",
4531 nid, start_pfn, end_pfn,
4532 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4534 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4536 /* Merge with existing active regions if possible */
4537 for (i = 0; i < nr_nodemap_entries; i++) {
4538 if (early_node_map[i].nid != nid)
4541 /* Skip if an existing region covers this new one */
4542 if (start_pfn >= early_node_map[i].start_pfn &&
4543 end_pfn <= early_node_map[i].end_pfn)
4546 /* Merge forward if suitable */
4547 if (start_pfn <= early_node_map[i].end_pfn &&
4548 end_pfn > early_node_map[i].end_pfn) {
4549 early_node_map[i].end_pfn = end_pfn;
4553 /* Merge backward if suitable */
4554 if (start_pfn < early_node_map[i].start_pfn &&
4555 end_pfn >= early_node_map[i].start_pfn) {
4556 early_node_map[i].start_pfn = start_pfn;
4561 /* Check that early_node_map is large enough */
4562 if (i >= MAX_ACTIVE_REGIONS) {
4563 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4564 MAX_ACTIVE_REGIONS);
4568 early_node_map[i].nid = nid;
4569 early_node_map[i].start_pfn = start_pfn;
4570 early_node_map[i].end_pfn = end_pfn;
4571 nr_nodemap_entries = i + 1;
4575 * remove_active_range - Shrink an existing registered range of PFNs
4576 * @nid: The node id the range is on that should be shrunk
4577 * @start_pfn: The new PFN of the range
4578 * @end_pfn: The new PFN of the range
4580 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4581 * The map is kept near the end physical page range that has already been
4582 * registered. This function allows an arch to shrink an existing registered
4585 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4586 unsigned long end_pfn)
4591 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4592 nid, start_pfn, end_pfn);
4594 /* Find the old active region end and shrink */
4595 for_each_active_range_index_in_nid(i, nid) {
4596 if (early_node_map[i].start_pfn >= start_pfn &&
4597 early_node_map[i].end_pfn <= end_pfn) {
4599 early_node_map[i].start_pfn = 0;
4600 early_node_map[i].end_pfn = 0;
4604 if (early_node_map[i].start_pfn < start_pfn &&
4605 early_node_map[i].end_pfn > start_pfn) {
4606 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4607 early_node_map[i].end_pfn = start_pfn;
4608 if (temp_end_pfn > end_pfn)
4609 add_active_range(nid, end_pfn, temp_end_pfn);
4612 if (early_node_map[i].start_pfn >= start_pfn &&
4613 early_node_map[i].end_pfn > end_pfn &&
4614 early_node_map[i].start_pfn < end_pfn) {
4615 early_node_map[i].start_pfn = end_pfn;
4623 /* remove the blank ones */
4624 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4625 if (early_node_map[i].nid != nid)
4627 if (early_node_map[i].end_pfn)
4629 /* we found it, get rid of it */
4630 for (j = i; j < nr_nodemap_entries - 1; j++)
4631 memcpy(&early_node_map[j], &early_node_map[j+1],
4632 sizeof(early_node_map[j]));
4633 j = nr_nodemap_entries - 1;
4634 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4635 nr_nodemap_entries--;
4640 * remove_all_active_ranges - Remove all currently registered regions
4642 * During discovery, it may be found that a table like SRAT is invalid
4643 * and an alternative discovery method must be used. This function removes
4644 * all currently registered regions.
4646 void __init remove_all_active_ranges(void)
4648 memset(early_node_map, 0, sizeof(early_node_map));
4649 nr_nodemap_entries = 0;
4652 /* Compare two active node_active_regions */
4653 static int __init cmp_node_active_region(const void *a, const void *b)
4655 struct node_active_region *arange = (struct node_active_region *)a;
4656 struct node_active_region *brange = (struct node_active_region *)b;
4658 /* Done this way to avoid overflows */
4659 if (arange->start_pfn > brange->start_pfn)
4661 if (arange->start_pfn < brange->start_pfn)
4667 /* sort the node_map by start_pfn */
4668 void __init sort_node_map(void)
4670 sort(early_node_map, (size_t)nr_nodemap_entries,
4671 sizeof(struct node_active_region),
4672 cmp_node_active_region, NULL);
4676 * node_map_pfn_alignment - determine the maximum internode alignment
4678 * This function should be called after node map is populated and sorted.
4679 * It calculates the maximum power of two alignment which can distinguish
4682 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4683 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4684 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4685 * shifted, 1GiB is enough and this function will indicate so.
4687 * This is used to test whether pfn -> nid mapping of the chosen memory
4688 * model has fine enough granularity to avoid incorrect mapping for the
4689 * populated node map.
4691 * Returns the determined alignment in pfn's. 0 if there is no alignment
4692 * requirement (single node).
4694 unsigned long __init node_map_pfn_alignment(void)
4696 unsigned long accl_mask = 0, last_end = 0;
4700 for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4701 int nid = early_node_map[i].nid;
4702 unsigned long start = early_node_map[i].start_pfn;
4703 unsigned long end = early_node_map[i].end_pfn;
4706 if (!start || last_nid < 0 || last_nid == nid) {
4713 * Start with a mask granular enough to pin-point to the
4714 * start pfn and tick off bits one-by-one until it becomes
4715 * too coarse to separate the current node from the last.
4717 mask = ~((1 << __ffs(start)) - 1);
4718 while (mask && last_end <= (start & (mask << 1)))
4721 /* accumulate all internode masks */
4725 /* convert mask to number of pages */
4726 return ~accl_mask + 1;
4729 /* Find the lowest pfn for a node */
4730 static unsigned long __init find_min_pfn_for_node(int nid)
4733 unsigned long min_pfn = ULONG_MAX;
4735 /* Assuming a sorted map, the first range found has the starting pfn */
4736 for_each_active_range_index_in_nid(i, nid)
4737 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4739 if (min_pfn == ULONG_MAX) {
4741 "Could not find start_pfn for node %d\n", nid);
4749 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4751 * It returns the minimum PFN based on information provided via
4752 * add_active_range().
4754 unsigned long __init find_min_pfn_with_active_regions(void)
4756 return find_min_pfn_for_node(MAX_NUMNODES);
4760 * early_calculate_totalpages()
4761 * Sum pages in active regions for movable zone.
4762 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4764 static unsigned long __init early_calculate_totalpages(void)
4767 unsigned long totalpages = 0;
4769 for (i = 0; i < nr_nodemap_entries; i++) {
4770 unsigned long pages = early_node_map[i].end_pfn -
4771 early_node_map[i].start_pfn;
4772 totalpages += pages;
4774 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4780 * Find the PFN the Movable zone begins in each node. Kernel memory
4781 * is spread evenly between nodes as long as the nodes have enough
4782 * memory. When they don't, some nodes will have more kernelcore than
4785 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4788 unsigned long usable_startpfn;
4789 unsigned long kernelcore_node, kernelcore_remaining;
4790 /* save the state before borrow the nodemask */
4791 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4792 unsigned long totalpages = early_calculate_totalpages();
4793 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4796 * If movablecore was specified, calculate what size of
4797 * kernelcore that corresponds so that memory usable for
4798 * any allocation type is evenly spread. If both kernelcore
4799 * and movablecore are specified, then the value of kernelcore
4800 * will be used for required_kernelcore if it's greater than
4801 * what movablecore would have allowed.
4803 if (required_movablecore) {
4804 unsigned long corepages;
4807 * Round-up so that ZONE_MOVABLE is at least as large as what
4808 * was requested by the user
4810 required_movablecore =
4811 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4812 corepages = totalpages - required_movablecore;
4814 required_kernelcore = max(required_kernelcore, corepages);
4817 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4818 if (!required_kernelcore)
4821 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4822 find_usable_zone_for_movable();
4823 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4826 /* Spread kernelcore memory as evenly as possible throughout nodes */
4827 kernelcore_node = required_kernelcore / usable_nodes;
4828 for_each_node_state(nid, N_HIGH_MEMORY) {
4830 * Recalculate kernelcore_node if the division per node
4831 * now exceeds what is necessary to satisfy the requested
4832 * amount of memory for the kernel
4834 if (required_kernelcore < kernelcore_node)
4835 kernelcore_node = required_kernelcore / usable_nodes;
4838 * As the map is walked, we track how much memory is usable
4839 * by the kernel using kernelcore_remaining. When it is
4840 * 0, the rest of the node is usable by ZONE_MOVABLE
4842 kernelcore_remaining = kernelcore_node;
4844 /* Go through each range of PFNs within this node */
4845 for_each_active_range_index_in_nid(i, nid) {
4846 unsigned long start_pfn, end_pfn;
4847 unsigned long size_pages;
4849 start_pfn = max(early_node_map[i].start_pfn,
4850 zone_movable_pfn[nid]);
4851 end_pfn = early_node_map[i].end_pfn;
4852 if (start_pfn >= end_pfn)
4855 /* Account for what is only usable for kernelcore */
4856 if (start_pfn < usable_startpfn) {
4857 unsigned long kernel_pages;
4858 kernel_pages = min(end_pfn, usable_startpfn)
4861 kernelcore_remaining -= min(kernel_pages,
4862 kernelcore_remaining);
4863 required_kernelcore -= min(kernel_pages,
4864 required_kernelcore);
4866 /* Continue if range is now fully accounted */
4867 if (end_pfn <= usable_startpfn) {
4870 * Push zone_movable_pfn to the end so
4871 * that if we have to rebalance
4872 * kernelcore across nodes, we will
4873 * not double account here
4875 zone_movable_pfn[nid] = end_pfn;
4878 start_pfn = usable_startpfn;
4882 * The usable PFN range for ZONE_MOVABLE is from
4883 * start_pfn->end_pfn. Calculate size_pages as the
4884 * number of pages used as kernelcore
4886 size_pages = end_pfn - start_pfn;
4887 if (size_pages > kernelcore_remaining)
4888 size_pages = kernelcore_remaining;
4889 zone_movable_pfn[nid] = start_pfn + size_pages;
4892 * Some kernelcore has been met, update counts and
4893 * break if the kernelcore for this node has been
4896 required_kernelcore -= min(required_kernelcore,
4898 kernelcore_remaining -= size_pages;
4899 if (!kernelcore_remaining)
4905 * If there is still required_kernelcore, we do another pass with one
4906 * less node in the count. This will push zone_movable_pfn[nid] further
4907 * along on the nodes that still have memory until kernelcore is
4911 if (usable_nodes && required_kernelcore > usable_nodes)
4914 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4915 for (nid = 0; nid < MAX_NUMNODES; nid++)
4916 zone_movable_pfn[nid] =
4917 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4920 /* restore the node_state */
4921 node_states[N_HIGH_MEMORY] = saved_node_state;
4924 /* Any regular memory on that node ? */
4925 static void check_for_regular_memory(pg_data_t *pgdat)
4927 #ifdef CONFIG_HIGHMEM
4928 enum zone_type zone_type;
4930 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4931 struct zone *zone = &pgdat->node_zones[zone_type];
4932 if (zone->present_pages)
4933 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4939 * free_area_init_nodes - Initialise all pg_data_t and zone data
4940 * @max_zone_pfn: an array of max PFNs for each zone
4942 * This will call free_area_init_node() for each active node in the system.
4943 * Using the page ranges provided by add_active_range(), the size of each
4944 * zone in each node and their holes is calculated. If the maximum PFN
4945 * between two adjacent zones match, it is assumed that the zone is empty.
4946 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4947 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4948 * starts where the previous one ended. For example, ZONE_DMA32 starts
4949 * at arch_max_dma_pfn.
4951 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4956 /* Sort early_node_map as initialisation assumes it is sorted */
4959 /* Record where the zone boundaries are */
4960 memset(arch_zone_lowest_possible_pfn, 0,
4961 sizeof(arch_zone_lowest_possible_pfn));
4962 memset(arch_zone_highest_possible_pfn, 0,
4963 sizeof(arch_zone_highest_possible_pfn));
4964 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4965 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4966 for (i = 1; i < MAX_NR_ZONES; i++) {
4967 if (i == ZONE_MOVABLE)
4969 arch_zone_lowest_possible_pfn[i] =
4970 arch_zone_highest_possible_pfn[i-1];
4971 arch_zone_highest_possible_pfn[i] =
4972 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4974 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4975 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4977 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4978 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4979 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4981 /* Print out the zone ranges */
4982 printk("Zone PFN ranges:\n");
4983 for (i = 0; i < MAX_NR_ZONES; i++) {
4984 if (i == ZONE_MOVABLE)
4986 printk(" %-8s ", zone_names[i]);
4987 if (arch_zone_lowest_possible_pfn[i] ==
4988 arch_zone_highest_possible_pfn[i])
4991 printk("%0#10lx -> %0#10lx\n",
4992 arch_zone_lowest_possible_pfn[i],
4993 arch_zone_highest_possible_pfn[i]);
4996 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4997 printk("Movable zone start PFN for each node\n");
4998 for (i = 0; i < MAX_NUMNODES; i++) {
4999 if (zone_movable_pfn[i])
5000 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
5003 /* Print out the early_node_map[] */
5004 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
5005 for (i = 0; i < nr_nodemap_entries; i++)
5006 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
5007 early_node_map[i].start_pfn,
5008 early_node_map[i].end_pfn);
5010 /* Initialise every node */
5011 mminit_verify_pageflags_layout();
5012 setup_nr_node_ids();
5013 for_each_online_node(nid) {
5014 pg_data_t *pgdat = NODE_DATA(nid);
5015 free_area_init_node(nid, NULL,
5016 find_min_pfn_for_node(nid), NULL);
5018 /* Any memory on that node */
5019 if (pgdat->node_present_pages)
5020 node_set_state(nid, N_HIGH_MEMORY);
5021 check_for_regular_memory(pgdat);
5025 static int __init cmdline_parse_core(char *p, unsigned long *core)
5027 unsigned long long coremem;
5031 coremem = memparse(p, &p);
5032 *core = coremem >> PAGE_SHIFT;
5034 /* Paranoid check that UL is enough for the coremem value */
5035 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5041 * kernelcore=size sets the amount of memory for use for allocations that
5042 * cannot be reclaimed or migrated.
5044 static int __init cmdline_parse_kernelcore(char *p)
5046 return cmdline_parse_core(p, &required_kernelcore);
5050 * movablecore=size sets the amount of memory for use for allocations that
5051 * can be reclaimed or migrated.
5053 static int __init cmdline_parse_movablecore(char *p)
5055 return cmdline_parse_core(p, &required_movablecore);
5058 early_param("kernelcore", cmdline_parse_kernelcore);
5059 early_param("movablecore", cmdline_parse_movablecore);
5061 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
5064 * set_dma_reserve - set the specified number of pages reserved in the first zone
5065 * @new_dma_reserve: The number of pages to mark reserved
5067 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5068 * In the DMA zone, a significant percentage may be consumed by kernel image
5069 * and other unfreeable allocations which can skew the watermarks badly. This
5070 * function may optionally be used to account for unfreeable pages in the
5071 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5072 * smaller per-cpu batchsize.
5074 void __init set_dma_reserve(unsigned long new_dma_reserve)
5076 dma_reserve = new_dma_reserve;
5079 void __init free_area_init(unsigned long *zones_size)
5081 free_area_init_node(0, zones_size,
5082 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5085 static int page_alloc_cpu_notify(struct notifier_block *self,
5086 unsigned long action, void *hcpu)
5088 int cpu = (unsigned long)hcpu;
5090 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5094 * Spill the event counters of the dead processor
5095 * into the current processors event counters.
5096 * This artificially elevates the count of the current
5099 vm_events_fold_cpu(cpu);
5102 * Zero the differential counters of the dead processor
5103 * so that the vm statistics are consistent.
5105 * This is only okay since the processor is dead and cannot
5106 * race with what we are doing.
5108 refresh_cpu_vm_stats(cpu);
5113 void __init page_alloc_init(void)
5115 hotcpu_notifier(page_alloc_cpu_notify, 0);
5119 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5120 * or min_free_kbytes changes.
5122 static void calculate_totalreserve_pages(void)
5124 struct pglist_data *pgdat;
5125 unsigned long reserve_pages = 0;
5126 enum zone_type i, j;
5128 for_each_online_pgdat(pgdat) {
5129 for (i = 0; i < MAX_NR_ZONES; i++) {
5130 struct zone *zone = pgdat->node_zones + i;
5131 unsigned long max = 0;
5133 /* Find valid and maximum lowmem_reserve in the zone */
5134 for (j = i; j < MAX_NR_ZONES; j++) {
5135 if (zone->lowmem_reserve[j] > max)
5136 max = zone->lowmem_reserve[j];
5139 /* we treat the high watermark as reserved pages. */
5140 max += high_wmark_pages(zone);
5142 if (max > zone->present_pages)
5143 max = zone->present_pages;
5144 reserve_pages += max;
5147 totalreserve_pages = reserve_pages;
5151 * setup_per_zone_lowmem_reserve - called whenever
5152 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5153 * has a correct pages reserved value, so an adequate number of
5154 * pages are left in the zone after a successful __alloc_pages().
5156 static void setup_per_zone_lowmem_reserve(void)
5158 struct pglist_data *pgdat;
5159 enum zone_type j, idx;
5161 for_each_online_pgdat(pgdat) {
5162 for (j = 0; j < MAX_NR_ZONES; j++) {
5163 struct zone *zone = pgdat->node_zones + j;
5164 unsigned long present_pages = zone->present_pages;
5166 zone->lowmem_reserve[j] = 0;
5170 struct zone *lower_zone;
5174 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5175 sysctl_lowmem_reserve_ratio[idx] = 1;
5177 lower_zone = pgdat->node_zones + idx;
5178 lower_zone->lowmem_reserve[j] = present_pages /
5179 sysctl_lowmem_reserve_ratio[idx];
5180 present_pages += lower_zone->present_pages;
5185 /* update totalreserve_pages */
5186 calculate_totalreserve_pages();
5190 * setup_per_zone_wmarks - called when min_free_kbytes changes
5191 * or when memory is hot-{added|removed}
5193 * Ensures that the watermark[min,low,high] values for each zone are set
5194 * correctly with respect to min_free_kbytes.
5196 void setup_per_zone_wmarks(void)
5198 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5199 unsigned long lowmem_pages = 0;
5201 unsigned long flags;
5203 /* Calculate total number of !ZONE_HIGHMEM pages */
5204 for_each_zone(zone) {
5205 if (!is_highmem(zone))
5206 lowmem_pages += zone->present_pages;
5209 for_each_zone(zone) {
5212 spin_lock_irqsave(&zone->lock, flags);
5213 tmp = (u64)pages_min * zone->present_pages;
5214 do_div(tmp, lowmem_pages);
5215 if (is_highmem(zone)) {
5217 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5218 * need highmem pages, so cap pages_min to a small
5221 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5222 * deltas controls asynch page reclaim, and so should
5223 * not be capped for highmem.
5227 min_pages = zone->present_pages / 1024;
5228 if (min_pages < SWAP_CLUSTER_MAX)
5229 min_pages = SWAP_CLUSTER_MAX;
5230 if (min_pages > 128)
5232 zone->watermark[WMARK_MIN] = min_pages;
5235 * If it's a lowmem zone, reserve a number of pages
5236 * proportionate to the zone's size.
5238 zone->watermark[WMARK_MIN] = tmp;
5241 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5242 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5243 setup_zone_migrate_reserve(zone);
5244 spin_unlock_irqrestore(&zone->lock, flags);
5247 /* update totalreserve_pages */
5248 calculate_totalreserve_pages();
5252 * The inactive anon list should be small enough that the VM never has to
5253 * do too much work, but large enough that each inactive page has a chance
5254 * to be referenced again before it is swapped out.
5256 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5257 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5258 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5259 * the anonymous pages are kept on the inactive list.
5262 * memory ratio inactive anon
5263 * -------------------------------------
5272 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5274 unsigned int gb, ratio;
5276 /* Zone size in gigabytes */
5277 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5279 ratio = int_sqrt(10 * gb);
5283 zone->inactive_ratio = ratio;
5286 static void __meminit setup_per_zone_inactive_ratio(void)
5291 calculate_zone_inactive_ratio(zone);
5295 * Initialise min_free_kbytes.
5297 * For small machines we want it small (128k min). For large machines
5298 * we want it large (64MB max). But it is not linear, because network
5299 * bandwidth does not increase linearly with machine size. We use
5301 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5302 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5318 int __meminit init_per_zone_wmark_min(void)
5320 unsigned long lowmem_kbytes;
5322 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5324 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5325 if (min_free_kbytes < 128)
5326 min_free_kbytes = 128;
5327 if (min_free_kbytes > 65536)
5328 min_free_kbytes = 65536;
5329 setup_per_zone_wmarks();
5330 refresh_zone_stat_thresholds();
5331 setup_per_zone_lowmem_reserve();
5332 setup_per_zone_inactive_ratio();
5335 module_init(init_per_zone_wmark_min)
5338 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5339 * that we can call two helper functions whenever min_free_kbytes
5342 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5343 void __user *buffer, size_t *length, loff_t *ppos)
5345 proc_dointvec(table, write, buffer, length, ppos);
5347 setup_per_zone_wmarks();
5352 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5353 void __user *buffer, size_t *length, loff_t *ppos)
5358 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5363 zone->min_unmapped_pages = (zone->present_pages *
5364 sysctl_min_unmapped_ratio) / 100;
5368 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5369 void __user *buffer, size_t *length, loff_t *ppos)
5374 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5379 zone->min_slab_pages = (zone->present_pages *
5380 sysctl_min_slab_ratio) / 100;
5386 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5387 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5388 * whenever sysctl_lowmem_reserve_ratio changes.
5390 * The reserve ratio obviously has absolutely no relation with the
5391 * minimum watermarks. The lowmem reserve ratio can only make sense
5392 * if in function of the boot time zone sizes.
5394 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5395 void __user *buffer, size_t *length, loff_t *ppos)
5397 proc_dointvec_minmax(table, write, buffer, length, ppos);
5398 setup_per_zone_lowmem_reserve();
5403 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5404 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5405 * can have before it gets flushed back to buddy allocator.
5408 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5409 void __user *buffer, size_t *length, loff_t *ppos)
5415 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5416 if (!write || (ret == -EINVAL))
5418 for_each_populated_zone(zone) {
5419 for_each_possible_cpu(cpu) {
5421 high = zone->present_pages / percpu_pagelist_fraction;
5422 setup_pagelist_highmark(
5423 per_cpu_ptr(zone->pageset, cpu), high);
5429 int hashdist = HASHDIST_DEFAULT;
5432 static int __init set_hashdist(char *str)
5436 hashdist = simple_strtoul(str, &str, 0);
5439 __setup("hashdist=", set_hashdist);
5443 * allocate a large system hash table from bootmem
5444 * - it is assumed that the hash table must contain an exact power-of-2
5445 * quantity of entries
5446 * - limit is the number of hash buckets, not the total allocation size
5448 void *__init alloc_large_system_hash(const char *tablename,
5449 unsigned long bucketsize,
5450 unsigned long numentries,
5453 unsigned int *_hash_shift,
5454 unsigned int *_hash_mask,
5455 unsigned long limit)
5457 unsigned long long max = limit;
5458 unsigned long log2qty, size;
5461 /* allow the kernel cmdline to have a say */
5463 /* round applicable memory size up to nearest megabyte */
5464 numentries = nr_kernel_pages;
5465 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5466 numentries >>= 20 - PAGE_SHIFT;
5467 numentries <<= 20 - PAGE_SHIFT;
5469 /* limit to 1 bucket per 2^scale bytes of low memory */
5470 if (scale > PAGE_SHIFT)
5471 numentries >>= (scale - PAGE_SHIFT);
5473 numentries <<= (PAGE_SHIFT - scale);
5475 /* Make sure we've got at least a 0-order allocation.. */
5476 if (unlikely(flags & HASH_SMALL)) {
5477 /* Makes no sense without HASH_EARLY */
5478 WARN_ON(!(flags & HASH_EARLY));
5479 if (!(numentries >> *_hash_shift)) {
5480 numentries = 1UL << *_hash_shift;
5481 BUG_ON(!numentries);
5483 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5484 numentries = PAGE_SIZE / bucketsize;
5486 numentries = roundup_pow_of_two(numentries);
5488 /* limit allocation size to 1/16 total memory by default */
5490 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5491 do_div(max, bucketsize);
5494 if (numentries > max)
5497 log2qty = ilog2(numentries);
5500 size = bucketsize << log2qty;
5501 if (flags & HASH_EARLY)
5502 table = alloc_bootmem_nopanic(size);
5504 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5507 * If bucketsize is not a power-of-two, we may free
5508 * some pages at the end of hash table which
5509 * alloc_pages_exact() automatically does
5511 if (get_order(size) < MAX_ORDER) {
5512 table = alloc_pages_exact(size, GFP_ATOMIC);
5513 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5516 } while (!table && size > PAGE_SIZE && --log2qty);
5519 panic("Failed to allocate %s hash table\n", tablename);
5521 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5524 ilog2(size) - PAGE_SHIFT,
5528 *_hash_shift = log2qty;
5530 *_hash_mask = (1 << log2qty) - 1;
5535 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5536 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5539 #ifdef CONFIG_SPARSEMEM
5540 return __pfn_to_section(pfn)->pageblock_flags;
5542 return zone->pageblock_flags;
5543 #endif /* CONFIG_SPARSEMEM */
5546 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5548 #ifdef CONFIG_SPARSEMEM
5549 pfn &= (PAGES_PER_SECTION-1);
5550 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5552 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5553 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5554 #endif /* CONFIG_SPARSEMEM */
5558 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5559 * @page: The page within the block of interest
5560 * @start_bitidx: The first bit of interest to retrieve
5561 * @end_bitidx: The last bit of interest
5562 * returns pageblock_bits flags
5564 unsigned long get_pageblock_flags_group(struct page *page,
5565 int start_bitidx, int end_bitidx)
5568 unsigned long *bitmap;
5569 unsigned long pfn, bitidx;
5570 unsigned long flags = 0;
5571 unsigned long value = 1;
5573 zone = page_zone(page);
5574 pfn = page_to_pfn(page);
5575 bitmap = get_pageblock_bitmap(zone, pfn);
5576 bitidx = pfn_to_bitidx(zone, pfn);
5578 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5579 if (test_bit(bitidx + start_bitidx, bitmap))
5586 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5587 * @page: The page within the block of interest
5588 * @start_bitidx: The first bit of interest
5589 * @end_bitidx: The last bit of interest
5590 * @flags: The flags to set
5592 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5593 int start_bitidx, int end_bitidx)
5596 unsigned long *bitmap;
5597 unsigned long pfn, bitidx;
5598 unsigned long value = 1;
5600 zone = page_zone(page);
5601 pfn = page_to_pfn(page);
5602 bitmap = get_pageblock_bitmap(zone, pfn);
5603 bitidx = pfn_to_bitidx(zone, pfn);
5604 VM_BUG_ON(pfn < zone->zone_start_pfn);
5605 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5607 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5609 __set_bit(bitidx + start_bitidx, bitmap);
5611 __clear_bit(bitidx + start_bitidx, bitmap);
5615 * This is designed as sub function...plz see page_isolation.c also.
5616 * set/clear page block's type to be ISOLATE.
5617 * page allocater never alloc memory from ISOLATE block.
5621 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5623 unsigned long pfn, iter, found;
5625 * For avoiding noise data, lru_add_drain_all() should be called
5626 * If ZONE_MOVABLE, the zone never contains immobile pages
5628 if (zone_idx(zone) == ZONE_MOVABLE)
5631 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5634 pfn = page_to_pfn(page);
5635 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5636 unsigned long check = pfn + iter;
5638 if (!pfn_valid_within(check))
5641 page = pfn_to_page(check);
5642 if (!page_count(page)) {
5643 if (PageBuddy(page))
5644 iter += (1 << page_order(page)) - 1;
5650 * If there are RECLAIMABLE pages, we need to check it.
5651 * But now, memory offline itself doesn't call shrink_slab()
5652 * and it still to be fixed.
5655 * If the page is not RAM, page_count()should be 0.
5656 * we don't need more check. This is an _used_ not-movable page.
5658 * The problematic thing here is PG_reserved pages. PG_reserved
5659 * is set to both of a memory hole page and a _used_ kernel
5668 bool is_pageblock_removable_nolock(struct page *page)
5670 struct zone *zone = page_zone(page);
5671 unsigned long pfn = page_to_pfn(page);
5674 * We have to be careful here because we are iterating over memory
5675 * sections which are not zone aware so we might end up outside of
5676 * the zone but still within the section.
5678 if (!zone || zone->zone_start_pfn > pfn ||
5679 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5682 return __count_immobile_pages(zone, page, 0);
5685 int set_migratetype_isolate(struct page *page)
5688 unsigned long flags, pfn;
5689 struct memory_isolate_notify arg;
5693 zone = page_zone(page);
5695 spin_lock_irqsave(&zone->lock, flags);
5697 pfn = page_to_pfn(page);
5698 arg.start_pfn = pfn;
5699 arg.nr_pages = pageblock_nr_pages;
5700 arg.pages_found = 0;
5703 * It may be possible to isolate a pageblock even if the
5704 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5705 * notifier chain is used by balloon drivers to return the
5706 * number of pages in a range that are held by the balloon
5707 * driver to shrink memory. If all the pages are accounted for
5708 * by balloons, are free, or on the LRU, isolation can continue.
5709 * Later, for example, when memory hotplug notifier runs, these
5710 * pages reported as "can be isolated" should be isolated(freed)
5711 * by the balloon driver through the memory notifier chain.
5713 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5714 notifier_ret = notifier_to_errno(notifier_ret);
5718 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5719 * We just check MOVABLE pages.
5721 if (__count_immobile_pages(zone, page, arg.pages_found))
5725 * immobile means "not-on-lru" paes. If immobile is larger than
5726 * removable-by-driver pages reported by notifier, we'll fail.
5731 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5732 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5735 spin_unlock_irqrestore(&zone->lock, flags);
5741 void unset_migratetype_isolate(struct page *page)
5744 unsigned long flags;
5745 zone = page_zone(page);
5746 spin_lock_irqsave(&zone->lock, flags);
5747 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5749 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5750 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5752 spin_unlock_irqrestore(&zone->lock, flags);
5755 #ifdef CONFIG_MEMORY_HOTREMOVE
5757 * All pages in the range must be isolated before calling this.
5760 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5766 unsigned long flags;
5767 /* find the first valid pfn */
5768 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5773 zone = page_zone(pfn_to_page(pfn));
5774 spin_lock_irqsave(&zone->lock, flags);
5776 while (pfn < end_pfn) {
5777 if (!pfn_valid(pfn)) {
5781 page = pfn_to_page(pfn);
5782 BUG_ON(page_count(page));
5783 BUG_ON(!PageBuddy(page));
5784 order = page_order(page);
5785 #ifdef CONFIG_DEBUG_VM
5786 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5787 pfn, 1 << order, end_pfn);
5789 list_del(&page->lru);
5790 rmv_page_order(page);
5791 zone->free_area[order].nr_free--;
5792 __mod_zone_page_state(zone, NR_FREE_PAGES,
5794 for (i = 0; i < (1 << order); i++)
5795 SetPageReserved((page+i));
5796 pfn += (1 << order);
5798 spin_unlock_irqrestore(&zone->lock, flags);
5802 #ifdef CONFIG_MEMORY_FAILURE
5803 bool is_free_buddy_page(struct page *page)
5805 struct zone *zone = page_zone(page);
5806 unsigned long pfn = page_to_pfn(page);
5807 unsigned long flags;
5810 spin_lock_irqsave(&zone->lock, flags);
5811 for (order = 0; order < MAX_ORDER; order++) {
5812 struct page *page_head = page - (pfn & ((1 << order) - 1));
5814 if (PageBuddy(page_head) && page_order(page_head) >= order)
5817 spin_unlock_irqrestore(&zone->lock, flags);
5819 return order < MAX_ORDER;
5823 static struct trace_print_flags pageflag_names[] = {
5824 {1UL << PG_locked, "locked" },
5825 {1UL << PG_error, "error" },
5826 {1UL << PG_referenced, "referenced" },
5827 {1UL << PG_uptodate, "uptodate" },
5828 {1UL << PG_dirty, "dirty" },
5829 {1UL << PG_lru, "lru" },
5830 {1UL << PG_active, "active" },
5831 {1UL << PG_slab, "slab" },
5832 {1UL << PG_owner_priv_1, "owner_priv_1" },
5833 {1UL << PG_arch_1, "arch_1" },
5834 {1UL << PG_reserved, "reserved" },
5835 {1UL << PG_private, "private" },
5836 {1UL << PG_private_2, "private_2" },
5837 {1UL << PG_writeback, "writeback" },
5838 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5839 {1UL << PG_head, "head" },
5840 {1UL << PG_tail, "tail" },
5842 {1UL << PG_compound, "compound" },
5844 {1UL << PG_swapcache, "swapcache" },
5845 {1UL << PG_mappedtodisk, "mappedtodisk" },
5846 {1UL << PG_reclaim, "reclaim" },
5847 {1UL << PG_swapbacked, "swapbacked" },
5848 {1UL << PG_unevictable, "unevictable" },
5850 {1UL << PG_mlocked, "mlocked" },
5852 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5853 {1UL << PG_uncached, "uncached" },
5855 #ifdef CONFIG_MEMORY_FAILURE
5856 {1UL << PG_hwpoison, "hwpoison" },
5861 static void dump_page_flags(unsigned long flags)
5863 const char *delim = "";
5867 printk(KERN_ALERT "page flags: %#lx(", flags);
5869 /* remove zone id */
5870 flags &= (1UL << NR_PAGEFLAGS) - 1;
5872 for (i = 0; pageflag_names[i].name && flags; i++) {
5874 mask = pageflag_names[i].mask;
5875 if ((flags & mask) != mask)
5879 printk("%s%s", delim, pageflag_names[i].name);
5883 /* check for left over flags */
5885 printk("%s%#lx", delim, flags);
5890 void dump_page(struct page *page)
5893 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5894 page, atomic_read(&page->_count), page_mapcount(page),
5895 page->mapping, page->index);
5896 dump_page_flags(page->flags);
5897 mem_cgroup_print_bad_page(page);