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_HAVE_MEMBLOCK_NODE_MAP
192 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
193 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
194 static unsigned long __initdata required_kernelcore;
195 static unsigned long __initdata required_movablecore;
196 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
198 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
200 EXPORT_SYMBOL(movable_zone);
201 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
204 int nr_node_ids __read_mostly = MAX_NUMNODES;
205 int nr_online_nodes __read_mostly = 1;
206 EXPORT_SYMBOL(nr_node_ids);
207 EXPORT_SYMBOL(nr_online_nodes);
210 int page_group_by_mobility_disabled __read_mostly;
212 static void set_pageblock_migratetype(struct page *page, int migratetype)
215 if (unlikely(page_group_by_mobility_disabled))
216 migratetype = MIGRATE_UNMOVABLE;
218 set_pageblock_flags_group(page, (unsigned long)migratetype,
219 PB_migrate, PB_migrate_end);
222 bool oom_killer_disabled __read_mostly;
224 #ifdef CONFIG_DEBUG_VM
225 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
229 unsigned long pfn = page_to_pfn(page);
232 seq = zone_span_seqbegin(zone);
233 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
235 else if (pfn < zone->zone_start_pfn)
237 } while (zone_span_seqretry(zone, seq));
242 static int page_is_consistent(struct zone *zone, struct page *page)
244 if (!pfn_valid_within(page_to_pfn(page)))
246 if (zone != page_zone(page))
252 * Temporary debugging check for pages not lying within a given zone.
254 static int bad_range(struct zone *zone, struct page *page)
256 if (page_outside_zone_boundaries(zone, page))
258 if (!page_is_consistent(zone, page))
264 static inline int bad_range(struct zone *zone, struct page *page)
270 static void bad_page(struct page *page)
272 static unsigned long resume;
273 static unsigned long nr_shown;
274 static unsigned long nr_unshown;
276 /* Don't complain about poisoned pages */
277 if (PageHWPoison(page)) {
278 reset_page_mapcount(page); /* remove PageBuddy */
283 * Allow a burst of 60 reports, then keep quiet for that minute;
284 * or allow a steady drip of one report per second.
286 if (nr_shown == 60) {
287 if (time_before(jiffies, resume)) {
293 "BUG: Bad page state: %lu messages suppressed\n",
300 resume = jiffies + 60 * HZ;
302 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
303 current->comm, page_to_pfn(page));
309 /* Leave bad fields for debug, except PageBuddy could make trouble */
310 reset_page_mapcount(page); /* remove PageBuddy */
311 add_taint(TAINT_BAD_PAGE);
315 * Higher-order pages are called "compound pages". They are structured thusly:
317 * The first PAGE_SIZE page is called the "head page".
319 * The remaining PAGE_SIZE pages are called "tail pages".
321 * All pages have PG_compound set. All tail pages have their ->first_page
322 * pointing at the head page.
324 * The first tail page's ->lru.next holds the address of the compound page's
325 * put_page() function. Its ->lru.prev holds the order of allocation.
326 * This usage means that zero-order pages may not be compound.
329 static void free_compound_page(struct page *page)
331 __free_pages_ok(page, compound_order(page));
334 void prep_compound_page(struct page *page, unsigned long order)
337 int nr_pages = 1 << order;
339 set_compound_page_dtor(page, free_compound_page);
340 set_compound_order(page, order);
342 for (i = 1; i < nr_pages; i++) {
343 struct page *p = page + i;
345 set_page_count(p, 0);
346 p->first_page = page;
350 /* update __split_huge_page_refcount if you change this function */
351 static int destroy_compound_page(struct page *page, unsigned long order)
354 int nr_pages = 1 << order;
357 if (unlikely(compound_order(page) != order) ||
358 unlikely(!PageHead(page))) {
363 __ClearPageHead(page);
365 for (i = 1; i < nr_pages; i++) {
366 struct page *p = page + i;
368 if (unlikely(!PageTail(p) || (p->first_page != page))) {
378 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
383 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
384 * and __GFP_HIGHMEM from hard or soft interrupt context.
386 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
387 for (i = 0; i < (1 << order); i++)
388 clear_highpage(page + i);
391 static inline void set_page_order(struct page *page, int order)
393 set_page_private(page, order);
394 __SetPageBuddy(page);
397 static inline void rmv_page_order(struct page *page)
399 __ClearPageBuddy(page);
400 set_page_private(page, 0);
404 * Locate the struct page for both the matching buddy in our
405 * pair (buddy1) and the combined O(n+1) page they form (page).
407 * 1) Any buddy B1 will have an order O twin B2 which satisfies
408 * the following equation:
410 * For example, if the starting buddy (buddy2) is #8 its order
412 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
414 * 2) Any buddy B will have an order O+1 parent P which
415 * satisfies the following equation:
418 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
420 static inline unsigned long
421 __find_buddy_index(unsigned long page_idx, unsigned int order)
423 return page_idx ^ (1 << order);
427 * This function checks whether a page is free && is the buddy
428 * we can do coalesce a page and its buddy if
429 * (a) the buddy is not in a hole &&
430 * (b) the buddy is in the buddy system &&
431 * (c) a page and its buddy have the same order &&
432 * (d) a page and its buddy are in the same zone.
434 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
435 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
437 * For recording page's order, we use page_private(page).
439 static inline int page_is_buddy(struct page *page, struct page *buddy,
442 if (!pfn_valid_within(page_to_pfn(buddy)))
445 if (page_zone_id(page) != page_zone_id(buddy))
448 if (PageBuddy(buddy) && page_order(buddy) == order) {
449 VM_BUG_ON(page_count(buddy) != 0);
456 * Freeing function for a buddy system allocator.
458 * The concept of a buddy system is to maintain direct-mapped table
459 * (containing bit values) for memory blocks of various "orders".
460 * The bottom level table contains the map for the smallest allocatable
461 * units of memory (here, pages), and each level above it describes
462 * pairs of units from the levels below, hence, "buddies".
463 * At a high level, all that happens here is marking the table entry
464 * at the bottom level available, and propagating the changes upward
465 * as necessary, plus some accounting needed to play nicely with other
466 * parts of the VM system.
467 * At each level, we keep a list of pages, which are heads of continuous
468 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
469 * order is recorded in page_private(page) field.
470 * So when we are allocating or freeing one, we can derive the state of the
471 * other. That is, if we allocate a small block, and both were
472 * free, the remainder of the region must be split into blocks.
473 * If a block is freed, and its buddy is also free, then this
474 * triggers coalescing into a block of larger size.
479 static inline void __free_one_page(struct page *page,
480 struct zone *zone, unsigned int order,
483 unsigned long page_idx;
484 unsigned long combined_idx;
485 unsigned long uninitialized_var(buddy_idx);
488 if (unlikely(PageCompound(page)))
489 if (unlikely(destroy_compound_page(page, order)))
492 VM_BUG_ON(migratetype == -1);
494 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
496 VM_BUG_ON(page_idx & ((1 << order) - 1));
497 VM_BUG_ON(bad_range(zone, page));
499 while (order < MAX_ORDER-1) {
500 buddy_idx = __find_buddy_index(page_idx, order);
501 buddy = page + (buddy_idx - page_idx);
502 if (!page_is_buddy(page, buddy, order))
505 /* Our buddy is free, merge with it and move up one order. */
506 list_del(&buddy->lru);
507 zone->free_area[order].nr_free--;
508 rmv_page_order(buddy);
509 combined_idx = buddy_idx & page_idx;
510 page = page + (combined_idx - page_idx);
511 page_idx = combined_idx;
514 set_page_order(page, order);
517 * If this is not the largest possible page, check if the buddy
518 * of the next-highest order is free. If it is, it's possible
519 * that pages are being freed that will coalesce soon. In case,
520 * that is happening, add the free page to the tail of the list
521 * so it's less likely to be used soon and more likely to be merged
522 * as a higher order page
524 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
525 struct page *higher_page, *higher_buddy;
526 combined_idx = buddy_idx & page_idx;
527 higher_page = page + (combined_idx - page_idx);
528 buddy_idx = __find_buddy_index(combined_idx, order + 1);
529 higher_buddy = page + (buddy_idx - combined_idx);
530 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
531 list_add_tail(&page->lru,
532 &zone->free_area[order].free_list[migratetype]);
537 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
539 zone->free_area[order].nr_free++;
543 * free_page_mlock() -- clean up attempts to free and mlocked() page.
544 * Page should not be on lru, so no need to fix that up.
545 * free_pages_check() will verify...
547 static inline void free_page_mlock(struct page *page)
549 __dec_zone_page_state(page, NR_MLOCK);
550 __count_vm_event(UNEVICTABLE_MLOCKFREED);
553 static inline int free_pages_check(struct page *page)
555 if (unlikely(page_mapcount(page) |
556 (page->mapping != NULL) |
557 (atomic_read(&page->_count) != 0) |
558 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
559 (mem_cgroup_bad_page_check(page)))) {
563 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
564 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
569 * Frees a number of pages from the PCP lists
570 * Assumes all pages on list are in same zone, and of same order.
571 * count is the number of pages to free.
573 * If the zone was previously in an "all pages pinned" state then look to
574 * see if this freeing clears that state.
576 * And clear the zone's pages_scanned counter, to hold off the "all pages are
577 * pinned" detection logic.
579 static void free_pcppages_bulk(struct zone *zone, int count,
580 struct per_cpu_pages *pcp)
586 spin_lock(&zone->lock);
587 zone->all_unreclaimable = 0;
588 zone->pages_scanned = 0;
592 struct list_head *list;
595 * Remove pages from lists in a round-robin fashion. A
596 * batch_free count is maintained that is incremented when an
597 * empty list is encountered. This is so more pages are freed
598 * off fuller lists instead of spinning excessively around empty
603 if (++migratetype == MIGRATE_PCPTYPES)
605 list = &pcp->lists[migratetype];
606 } while (list_empty(list));
608 /* This is the only non-empty list. Free them all. */
609 if (batch_free == MIGRATE_PCPTYPES)
610 batch_free = to_free;
613 page = list_entry(list->prev, struct page, lru);
614 /* must delete as __free_one_page list manipulates */
615 list_del(&page->lru);
616 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
617 __free_one_page(page, zone, 0, page_private(page));
618 trace_mm_page_pcpu_drain(page, 0, page_private(page));
619 } while (--to_free && --batch_free && !list_empty(list));
621 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
622 spin_unlock(&zone->lock);
625 static void free_one_page(struct zone *zone, struct page *page, int order,
628 spin_lock(&zone->lock);
629 zone->all_unreclaimable = 0;
630 zone->pages_scanned = 0;
632 __free_one_page(page, zone, order, migratetype);
633 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
634 spin_unlock(&zone->lock);
637 static bool free_pages_prepare(struct page *page, unsigned int order)
642 trace_mm_page_free(page, order);
643 kmemcheck_free_shadow(page, order);
646 page->mapping = NULL;
647 for (i = 0; i < (1 << order); i++)
648 bad += free_pages_check(page + i);
652 if (!PageHighMem(page)) {
653 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
654 debug_check_no_obj_freed(page_address(page),
657 arch_free_page(page, order);
658 kernel_map_pages(page, 1 << order, 0);
663 static void __free_pages_ok(struct page *page, unsigned int order)
666 int wasMlocked = __TestClearPageMlocked(page);
668 if (!free_pages_prepare(page, order))
671 local_irq_save(flags);
672 if (unlikely(wasMlocked))
673 free_page_mlock(page);
674 __count_vm_events(PGFREE, 1 << order);
675 free_one_page(page_zone(page), page, order,
676 get_pageblock_migratetype(page));
677 local_irq_restore(flags);
681 * permit the bootmem allocator to evade page validation on high-order frees
683 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
686 __ClearPageReserved(page);
687 set_page_count(page, 0);
688 set_page_refcounted(page);
694 for (loop = 0; loop < (1 << order); loop++) {
695 struct page *p = &page[loop];
697 if (loop + 1 < (1 << order))
699 __ClearPageReserved(p);
700 set_page_count(p, 0);
703 set_page_refcounted(page);
704 __free_pages(page, order);
710 * The order of subdivision here is critical for the IO subsystem.
711 * Please do not alter this order without good reasons and regression
712 * testing. Specifically, as large blocks of memory are subdivided,
713 * the order in which smaller blocks are delivered depends on the order
714 * they're subdivided in this function. This is the primary factor
715 * influencing the order in which pages are delivered to the IO
716 * subsystem according to empirical testing, and this is also justified
717 * by considering the behavior of a buddy system containing a single
718 * large block of memory acted on by a series of small allocations.
719 * This behavior is a critical factor in sglist merging's success.
723 static inline void expand(struct zone *zone, struct page *page,
724 int low, int high, struct free_area *area,
727 unsigned long size = 1 << high;
733 VM_BUG_ON(bad_range(zone, &page[size]));
734 list_add(&page[size].lru, &area->free_list[migratetype]);
736 set_page_order(&page[size], high);
741 * This page is about to be returned from the page allocator
743 static inline int check_new_page(struct page *page)
745 if (unlikely(page_mapcount(page) |
746 (page->mapping != NULL) |
747 (atomic_read(&page->_count) != 0) |
748 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
749 (mem_cgroup_bad_page_check(page)))) {
756 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
760 for (i = 0; i < (1 << order); i++) {
761 struct page *p = page + i;
762 if (unlikely(check_new_page(p)))
766 set_page_private(page, 0);
767 set_page_refcounted(page);
769 arch_alloc_page(page, order);
770 kernel_map_pages(page, 1 << order, 1);
772 if (gfp_flags & __GFP_ZERO)
773 prep_zero_page(page, order, gfp_flags);
775 if (order && (gfp_flags & __GFP_COMP))
776 prep_compound_page(page, order);
782 * Go through the free lists for the given migratetype and remove
783 * the smallest available page from the freelists
786 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
789 unsigned int current_order;
790 struct free_area * area;
793 /* Find a page of the appropriate size in the preferred list */
794 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
795 area = &(zone->free_area[current_order]);
796 if (list_empty(&area->free_list[migratetype]))
799 page = list_entry(area->free_list[migratetype].next,
801 list_del(&page->lru);
802 rmv_page_order(page);
804 expand(zone, page, order, current_order, area, migratetype);
813 * This array describes the order lists are fallen back to when
814 * the free lists for the desirable migrate type are depleted
816 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
817 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
818 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
819 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
820 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
824 * Move the free pages in a range to the free lists of the requested type.
825 * Note that start_page and end_pages are not aligned on a pageblock
826 * boundary. If alignment is required, use move_freepages_block()
828 static int move_freepages(struct zone *zone,
829 struct page *start_page, struct page *end_page,
836 #ifndef CONFIG_HOLES_IN_ZONE
838 * page_zone is not safe to call in this context when
839 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
840 * anyway as we check zone boundaries in move_freepages_block().
841 * Remove at a later date when no bug reports exist related to
842 * grouping pages by mobility
844 BUG_ON(page_zone(start_page) != page_zone(end_page));
847 for (page = start_page; page <= end_page;) {
848 /* Make sure we are not inadvertently changing nodes */
849 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
851 if (!pfn_valid_within(page_to_pfn(page))) {
856 if (!PageBuddy(page)) {
861 order = page_order(page);
862 list_move(&page->lru,
863 &zone->free_area[order].free_list[migratetype]);
865 pages_moved += 1 << order;
871 static int move_freepages_block(struct zone *zone, struct page *page,
874 unsigned long start_pfn, end_pfn;
875 struct page *start_page, *end_page;
877 start_pfn = page_to_pfn(page);
878 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
879 start_page = pfn_to_page(start_pfn);
880 end_page = start_page + pageblock_nr_pages - 1;
881 end_pfn = start_pfn + pageblock_nr_pages - 1;
883 /* Do not cross zone boundaries */
884 if (start_pfn < zone->zone_start_pfn)
886 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
889 return move_freepages(zone, start_page, end_page, migratetype);
892 static void change_pageblock_range(struct page *pageblock_page,
893 int start_order, int migratetype)
895 int nr_pageblocks = 1 << (start_order - pageblock_order);
897 while (nr_pageblocks--) {
898 set_pageblock_migratetype(pageblock_page, migratetype);
899 pageblock_page += pageblock_nr_pages;
903 /* Remove an element from the buddy allocator from the fallback list */
904 static inline struct page *
905 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
907 struct free_area * area;
912 /* Find the largest possible block of pages in the other list */
913 for (current_order = MAX_ORDER-1; current_order >= order;
915 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
916 migratetype = fallbacks[start_migratetype][i];
918 /* MIGRATE_RESERVE handled later if necessary */
919 if (migratetype == MIGRATE_RESERVE)
922 area = &(zone->free_area[current_order]);
923 if (list_empty(&area->free_list[migratetype]))
926 page = list_entry(area->free_list[migratetype].next,
931 * If breaking a large block of pages, move all free
932 * pages to the preferred allocation list. If falling
933 * back for a reclaimable kernel allocation, be more
934 * aggressive about taking ownership of free pages
936 if (unlikely(current_order >= (pageblock_order >> 1)) ||
937 start_migratetype == MIGRATE_RECLAIMABLE ||
938 page_group_by_mobility_disabled) {
940 pages = move_freepages_block(zone, page,
943 /* Claim the whole block if over half of it is free */
944 if (pages >= (1 << (pageblock_order-1)) ||
945 page_group_by_mobility_disabled)
946 set_pageblock_migratetype(page,
949 migratetype = start_migratetype;
952 /* Remove the page from the freelists */
953 list_del(&page->lru);
954 rmv_page_order(page);
956 /* Take ownership for orders >= pageblock_order */
957 if (current_order >= pageblock_order)
958 change_pageblock_range(page, current_order,
961 expand(zone, page, order, current_order, area, migratetype);
963 trace_mm_page_alloc_extfrag(page, order, current_order,
964 start_migratetype, migratetype);
974 * Do the hard work of removing an element from the buddy allocator.
975 * Call me with the zone->lock already held.
977 static struct page *__rmqueue(struct zone *zone, unsigned int order,
983 page = __rmqueue_smallest(zone, order, migratetype);
985 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
986 page = __rmqueue_fallback(zone, order, migratetype);
989 * Use MIGRATE_RESERVE rather than fail an allocation. goto
990 * is used because __rmqueue_smallest is an inline function
991 * and we want just one call site
994 migratetype = MIGRATE_RESERVE;
999 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1004 * Obtain a specified number of elements from the buddy allocator, all under
1005 * a single hold of the lock, for efficiency. Add them to the supplied list.
1006 * Returns the number of new pages which were placed at *list.
1008 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1009 unsigned long count, struct list_head *list,
1010 int migratetype, int cold)
1014 spin_lock(&zone->lock);
1015 for (i = 0; i < count; ++i) {
1016 struct page *page = __rmqueue(zone, order, migratetype);
1017 if (unlikely(page == NULL))
1021 * Split buddy pages returned by expand() are received here
1022 * in physical page order. The page is added to the callers and
1023 * list and the list head then moves forward. From the callers
1024 * perspective, the linked list is ordered by page number in
1025 * some conditions. This is useful for IO devices that can
1026 * merge IO requests if the physical pages are ordered
1029 if (likely(cold == 0))
1030 list_add(&page->lru, list);
1032 list_add_tail(&page->lru, list);
1033 set_page_private(page, migratetype);
1036 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1037 spin_unlock(&zone->lock);
1043 * Called from the vmstat counter updater to drain pagesets of this
1044 * currently executing processor on remote nodes after they have
1047 * Note that this function must be called with the thread pinned to
1048 * a single processor.
1050 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1052 unsigned long flags;
1055 local_irq_save(flags);
1056 if (pcp->count >= pcp->batch)
1057 to_drain = pcp->batch;
1059 to_drain = pcp->count;
1060 free_pcppages_bulk(zone, to_drain, pcp);
1061 pcp->count -= to_drain;
1062 local_irq_restore(flags);
1067 * Drain pages of the indicated processor.
1069 * The processor must either be the current processor and the
1070 * thread pinned to the current processor or a processor that
1073 static void drain_pages(unsigned int cpu)
1075 unsigned long flags;
1078 for_each_populated_zone(zone) {
1079 struct per_cpu_pageset *pset;
1080 struct per_cpu_pages *pcp;
1082 local_irq_save(flags);
1083 pset = per_cpu_ptr(zone->pageset, cpu);
1087 free_pcppages_bulk(zone, pcp->count, pcp);
1090 local_irq_restore(flags);
1095 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1097 void drain_local_pages(void *arg)
1099 drain_pages(smp_processor_id());
1103 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1105 void drain_all_pages(void)
1107 on_each_cpu(drain_local_pages, NULL, 1);
1110 #ifdef CONFIG_HIBERNATION
1112 void mark_free_pages(struct zone *zone)
1114 unsigned long pfn, max_zone_pfn;
1115 unsigned long flags;
1117 struct list_head *curr;
1119 if (!zone->spanned_pages)
1122 spin_lock_irqsave(&zone->lock, flags);
1124 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1125 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1126 if (pfn_valid(pfn)) {
1127 struct page *page = pfn_to_page(pfn);
1129 if (!swsusp_page_is_forbidden(page))
1130 swsusp_unset_page_free(page);
1133 for_each_migratetype_order(order, t) {
1134 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1137 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1138 for (i = 0; i < (1UL << order); i++)
1139 swsusp_set_page_free(pfn_to_page(pfn + i));
1142 spin_unlock_irqrestore(&zone->lock, flags);
1144 #endif /* CONFIG_PM */
1147 * Free a 0-order page
1148 * cold == 1 ? free a cold page : free a hot page
1150 void free_hot_cold_page(struct page *page, int cold)
1152 struct zone *zone = page_zone(page);
1153 struct per_cpu_pages *pcp;
1154 unsigned long flags;
1156 int wasMlocked = __TestClearPageMlocked(page);
1158 if (!free_pages_prepare(page, 0))
1161 migratetype = get_pageblock_migratetype(page);
1162 set_page_private(page, migratetype);
1163 local_irq_save(flags);
1164 if (unlikely(wasMlocked))
1165 free_page_mlock(page);
1166 __count_vm_event(PGFREE);
1169 * We only track unmovable, reclaimable and movable on pcp lists.
1170 * Free ISOLATE pages back to the allocator because they are being
1171 * offlined but treat RESERVE as movable pages so we can get those
1172 * areas back if necessary. Otherwise, we may have to free
1173 * excessively into the page allocator
1175 if (migratetype >= MIGRATE_PCPTYPES) {
1176 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1177 free_one_page(zone, page, 0, migratetype);
1180 migratetype = MIGRATE_MOVABLE;
1183 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1185 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1187 list_add(&page->lru, &pcp->lists[migratetype]);
1189 if (pcp->count >= pcp->high) {
1190 free_pcppages_bulk(zone, pcp->batch, pcp);
1191 pcp->count -= pcp->batch;
1195 local_irq_restore(flags);
1199 * Free a list of 0-order pages
1201 void free_hot_cold_page_list(struct list_head *list, int cold)
1203 struct page *page, *next;
1205 list_for_each_entry_safe(page, next, list, lru) {
1206 trace_mm_page_free_batched(page, cold);
1207 free_hot_cold_page(page, cold);
1212 * split_page takes a non-compound higher-order page, and splits it into
1213 * n (1<<order) sub-pages: page[0..n]
1214 * Each sub-page must be freed individually.
1216 * Note: this is probably too low level an operation for use in drivers.
1217 * Please consult with lkml before using this in your driver.
1219 void split_page(struct page *page, unsigned int order)
1223 VM_BUG_ON(PageCompound(page));
1224 VM_BUG_ON(!page_count(page));
1226 #ifdef CONFIG_KMEMCHECK
1228 * Split shadow pages too, because free(page[0]) would
1229 * otherwise free the whole shadow.
1231 if (kmemcheck_page_is_tracked(page))
1232 split_page(virt_to_page(page[0].shadow), order);
1235 for (i = 1; i < (1 << order); i++)
1236 set_page_refcounted(page + i);
1240 * Similar to split_page except the page is already free. As this is only
1241 * being used for migration, the migratetype of the block also changes.
1242 * As this is called with interrupts disabled, the caller is responsible
1243 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1246 * Note: this is probably too low level an operation for use in drivers.
1247 * Please consult with lkml before using this in your driver.
1249 int split_free_page(struct page *page)
1252 unsigned long watermark;
1255 BUG_ON(!PageBuddy(page));
1257 zone = page_zone(page);
1258 order = page_order(page);
1260 /* Obey watermarks as if the page was being allocated */
1261 watermark = low_wmark_pages(zone) + (1 << order);
1262 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1265 /* Remove page from free list */
1266 list_del(&page->lru);
1267 zone->free_area[order].nr_free--;
1268 rmv_page_order(page);
1269 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1271 /* Split into individual pages */
1272 set_page_refcounted(page);
1273 split_page(page, order);
1275 if (order >= pageblock_order - 1) {
1276 struct page *endpage = page + (1 << order) - 1;
1277 for (; page < endpage; page += pageblock_nr_pages)
1278 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1285 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1286 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1290 struct page *buffered_rmqueue(struct zone *preferred_zone,
1291 struct zone *zone, int order, gfp_t gfp_flags,
1294 unsigned long flags;
1296 int cold = !!(gfp_flags & __GFP_COLD);
1299 if (likely(order == 0)) {
1300 struct per_cpu_pages *pcp;
1301 struct list_head *list;
1303 local_irq_save(flags);
1304 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1305 list = &pcp->lists[migratetype];
1306 if (list_empty(list)) {
1307 pcp->count += rmqueue_bulk(zone, 0,
1310 if (unlikely(list_empty(list)))
1315 page = list_entry(list->prev, struct page, lru);
1317 page = list_entry(list->next, struct page, lru);
1319 list_del(&page->lru);
1322 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1324 * __GFP_NOFAIL is not to be used in new code.
1326 * All __GFP_NOFAIL callers should be fixed so that they
1327 * properly detect and handle allocation failures.
1329 * We most definitely don't want callers attempting to
1330 * allocate greater than order-1 page units with
1333 WARN_ON_ONCE(order > 1);
1335 spin_lock_irqsave(&zone->lock, flags);
1336 page = __rmqueue(zone, order, migratetype);
1337 spin_unlock(&zone->lock);
1340 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1343 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1344 zone_statistics(preferred_zone, zone, gfp_flags);
1345 local_irq_restore(flags);
1347 VM_BUG_ON(bad_range(zone, page));
1348 if (prep_new_page(page, order, gfp_flags))
1353 local_irq_restore(flags);
1357 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1358 #define ALLOC_WMARK_MIN WMARK_MIN
1359 #define ALLOC_WMARK_LOW WMARK_LOW
1360 #define ALLOC_WMARK_HIGH WMARK_HIGH
1361 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1363 /* Mask to get the watermark bits */
1364 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1366 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1367 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1368 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1370 #ifdef CONFIG_FAIL_PAGE_ALLOC
1373 struct fault_attr attr;
1375 u32 ignore_gfp_highmem;
1376 u32 ignore_gfp_wait;
1378 } fail_page_alloc = {
1379 .attr = FAULT_ATTR_INITIALIZER,
1380 .ignore_gfp_wait = 1,
1381 .ignore_gfp_highmem = 1,
1385 static int __init setup_fail_page_alloc(char *str)
1387 return setup_fault_attr(&fail_page_alloc.attr, str);
1389 __setup("fail_page_alloc=", setup_fail_page_alloc);
1391 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1393 if (order < fail_page_alloc.min_order)
1395 if (gfp_mask & __GFP_NOFAIL)
1397 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1399 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1402 return should_fail(&fail_page_alloc.attr, 1 << order);
1405 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1407 static int __init fail_page_alloc_debugfs(void)
1409 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1412 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1413 &fail_page_alloc.attr);
1415 return PTR_ERR(dir);
1417 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1418 &fail_page_alloc.ignore_gfp_wait))
1420 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1421 &fail_page_alloc.ignore_gfp_highmem))
1423 if (!debugfs_create_u32("min-order", mode, dir,
1424 &fail_page_alloc.min_order))
1429 debugfs_remove_recursive(dir);
1434 late_initcall(fail_page_alloc_debugfs);
1436 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1438 #else /* CONFIG_FAIL_PAGE_ALLOC */
1440 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1445 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1448 * Return true if free pages are above 'mark'. This takes into account the order
1449 * of the allocation.
1451 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1452 int classzone_idx, int alloc_flags, long free_pages)
1454 /* free_pages my go negative - that's OK */
1458 free_pages -= (1 << order) + 1;
1459 if (alloc_flags & ALLOC_HIGH)
1461 if (alloc_flags & ALLOC_HARDER)
1464 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1466 for (o = 0; o < order; o++) {
1467 /* At the next order, this order's pages become unavailable */
1468 free_pages -= z->free_area[o].nr_free << o;
1470 /* Require fewer higher order pages to be free */
1473 if (free_pages <= min)
1479 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1480 int classzone_idx, int alloc_flags)
1482 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1483 zone_page_state(z, NR_FREE_PAGES));
1486 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1487 int classzone_idx, int alloc_flags)
1489 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1491 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1492 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1494 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1500 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1501 * skip over zones that are not allowed by the cpuset, or that have
1502 * been recently (in last second) found to be nearly full. See further
1503 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1504 * that have to skip over a lot of full or unallowed zones.
1506 * If the zonelist cache is present in the passed in zonelist, then
1507 * returns a pointer to the allowed node mask (either the current
1508 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1510 * If the zonelist cache is not available for this zonelist, does
1511 * nothing and returns NULL.
1513 * If the fullzones BITMAP in the zonelist cache is stale (more than
1514 * a second since last zap'd) then we zap it out (clear its bits.)
1516 * We hold off even calling zlc_setup, until after we've checked the
1517 * first zone in the zonelist, on the theory that most allocations will
1518 * be satisfied from that first zone, so best to examine that zone as
1519 * quickly as we can.
1521 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1523 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1524 nodemask_t *allowednodes; /* zonelist_cache approximation */
1526 zlc = zonelist->zlcache_ptr;
1530 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1531 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1532 zlc->last_full_zap = jiffies;
1535 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1536 &cpuset_current_mems_allowed :
1537 &node_states[N_HIGH_MEMORY];
1538 return allowednodes;
1542 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1543 * if it is worth looking at further for free memory:
1544 * 1) Check that the zone isn't thought to be full (doesn't have its
1545 * bit set in the zonelist_cache fullzones BITMAP).
1546 * 2) Check that the zones node (obtained from the zonelist_cache
1547 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1548 * Return true (non-zero) if zone is worth looking at further, or
1549 * else return false (zero) if it is not.
1551 * This check -ignores- the distinction between various watermarks,
1552 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1553 * found to be full for any variation of these watermarks, it will
1554 * be considered full for up to one second by all requests, unless
1555 * we are so low on memory on all allowed nodes that we are forced
1556 * into the second scan of the zonelist.
1558 * In the second scan we ignore this zonelist cache and exactly
1559 * apply the watermarks to all zones, even it is slower to do so.
1560 * We are low on memory in the second scan, and should leave no stone
1561 * unturned looking for a free page.
1563 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1564 nodemask_t *allowednodes)
1566 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1567 int i; /* index of *z in zonelist zones */
1568 int n; /* node that zone *z is on */
1570 zlc = zonelist->zlcache_ptr;
1574 i = z - zonelist->_zonerefs;
1577 /* This zone is worth trying if it is allowed but not full */
1578 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1582 * Given 'z' scanning a zonelist, set the corresponding bit in
1583 * zlc->fullzones, so that subsequent attempts to allocate a page
1584 * from that zone don't waste time re-examining it.
1586 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1588 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1589 int i; /* index of *z in zonelist zones */
1591 zlc = zonelist->zlcache_ptr;
1595 i = z - zonelist->_zonerefs;
1597 set_bit(i, zlc->fullzones);
1601 * clear all zones full, called after direct reclaim makes progress so that
1602 * a zone that was recently full is not skipped over for up to a second
1604 static void zlc_clear_zones_full(struct zonelist *zonelist)
1606 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1608 zlc = zonelist->zlcache_ptr;
1612 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1615 #else /* CONFIG_NUMA */
1617 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1622 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1623 nodemask_t *allowednodes)
1628 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1632 static void zlc_clear_zones_full(struct zonelist *zonelist)
1635 #endif /* CONFIG_NUMA */
1638 * get_page_from_freelist goes through the zonelist trying to allocate
1641 static struct page *
1642 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1643 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1644 struct zone *preferred_zone, int migratetype)
1647 struct page *page = NULL;
1650 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1651 int zlc_active = 0; /* set if using zonelist_cache */
1652 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1654 classzone_idx = zone_idx(preferred_zone);
1657 * Scan zonelist, looking for a zone with enough free.
1658 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1660 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1661 high_zoneidx, nodemask) {
1662 if (NUMA_BUILD && zlc_active &&
1663 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1665 if ((alloc_flags & ALLOC_CPUSET) &&
1666 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1669 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1670 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1674 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1675 if (zone_watermark_ok(zone, order, mark,
1676 classzone_idx, alloc_flags))
1679 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1681 * we do zlc_setup if there are multiple nodes
1682 * and before considering the first zone allowed
1685 allowednodes = zlc_setup(zonelist, alloc_flags);
1690 if (zone_reclaim_mode == 0)
1691 goto this_zone_full;
1694 * As we may have just activated ZLC, check if the first
1695 * eligible zone has failed zone_reclaim recently.
1697 if (NUMA_BUILD && zlc_active &&
1698 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1701 ret = zone_reclaim(zone, gfp_mask, order);
1703 case ZONE_RECLAIM_NOSCAN:
1706 case ZONE_RECLAIM_FULL:
1707 /* scanned but unreclaimable */
1710 /* did we reclaim enough */
1711 if (!zone_watermark_ok(zone, order, mark,
1712 classzone_idx, alloc_flags))
1713 goto this_zone_full;
1718 page = buffered_rmqueue(preferred_zone, zone, order,
1719 gfp_mask, migratetype);
1724 zlc_mark_zone_full(zonelist, z);
1727 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1728 /* Disable zlc cache for second zonelist scan */
1736 * Large machines with many possible nodes should not always dump per-node
1737 * meminfo in irq context.
1739 static inline bool should_suppress_show_mem(void)
1744 ret = in_interrupt();
1749 static DEFINE_RATELIMIT_STATE(nopage_rs,
1750 DEFAULT_RATELIMIT_INTERVAL,
1751 DEFAULT_RATELIMIT_BURST);
1753 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1755 unsigned int filter = SHOW_MEM_FILTER_NODES;
1757 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1761 * This documents exceptions given to allocations in certain
1762 * contexts that are allowed to allocate outside current's set
1765 if (!(gfp_mask & __GFP_NOMEMALLOC))
1766 if (test_thread_flag(TIF_MEMDIE) ||
1767 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1768 filter &= ~SHOW_MEM_FILTER_NODES;
1769 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1770 filter &= ~SHOW_MEM_FILTER_NODES;
1773 struct va_format vaf;
1776 va_start(args, fmt);
1781 pr_warn("%pV", &vaf);
1786 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1787 current->comm, order, gfp_mask);
1790 if (!should_suppress_show_mem())
1795 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1796 unsigned long did_some_progress,
1797 unsigned long pages_reclaimed)
1799 /* Do not loop if specifically requested */
1800 if (gfp_mask & __GFP_NORETRY)
1803 /* Always retry if specifically requested */
1804 if (gfp_mask & __GFP_NOFAIL)
1808 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
1809 * making forward progress without invoking OOM. Suspend also disables
1810 * storage devices so kswapd will not help. Bail if we are suspending.
1812 if (!did_some_progress && pm_suspended_storage())
1816 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1817 * means __GFP_NOFAIL, but that may not be true in other
1820 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1824 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1825 * specified, then we retry until we no longer reclaim any pages
1826 * (above), or we've reclaimed an order of pages at least as
1827 * large as the allocation's order. In both cases, if the
1828 * allocation still fails, we stop retrying.
1830 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1836 static inline struct page *
1837 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1838 struct zonelist *zonelist, enum zone_type high_zoneidx,
1839 nodemask_t *nodemask, struct zone *preferred_zone,
1844 /* Acquire the OOM killer lock for the zones in zonelist */
1845 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1846 schedule_timeout_uninterruptible(1);
1851 * Go through the zonelist yet one more time, keep very high watermark
1852 * here, this is only to catch a parallel oom killing, we must fail if
1853 * we're still under heavy pressure.
1855 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1856 order, zonelist, high_zoneidx,
1857 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1858 preferred_zone, migratetype);
1862 if (!(gfp_mask & __GFP_NOFAIL)) {
1863 /* The OOM killer will not help higher order allocs */
1864 if (order > PAGE_ALLOC_COSTLY_ORDER)
1866 /* The OOM killer does not needlessly kill tasks for lowmem */
1867 if (high_zoneidx < ZONE_NORMAL)
1870 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1871 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1872 * The caller should handle page allocation failure by itself if
1873 * it specifies __GFP_THISNODE.
1874 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1876 if (gfp_mask & __GFP_THISNODE)
1879 /* Exhausted what can be done so it's blamo time */
1880 out_of_memory(zonelist, gfp_mask, order, nodemask);
1883 clear_zonelist_oom(zonelist, gfp_mask);
1887 #ifdef CONFIG_COMPACTION
1888 /* Try memory compaction for high-order allocations before reclaim */
1889 static struct page *
1890 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1891 struct zonelist *zonelist, enum zone_type high_zoneidx,
1892 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1893 int migratetype, unsigned long *did_some_progress,
1894 bool sync_migration)
1898 if (!order || compaction_deferred(preferred_zone))
1901 current->flags |= PF_MEMALLOC;
1902 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1903 nodemask, sync_migration);
1904 current->flags &= ~PF_MEMALLOC;
1905 if (*did_some_progress != COMPACT_SKIPPED) {
1907 /* Page migration frees to the PCP lists but we want merging */
1908 drain_pages(get_cpu());
1911 page = get_page_from_freelist(gfp_mask, nodemask,
1912 order, zonelist, high_zoneidx,
1913 alloc_flags, preferred_zone,
1916 preferred_zone->compact_considered = 0;
1917 preferred_zone->compact_defer_shift = 0;
1918 count_vm_event(COMPACTSUCCESS);
1923 * It's bad if compaction run occurs and fails.
1924 * The most likely reason is that pages exist,
1925 * but not enough to satisfy watermarks.
1927 count_vm_event(COMPACTFAIL);
1928 defer_compaction(preferred_zone);
1936 static inline struct page *
1937 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1938 struct zonelist *zonelist, enum zone_type high_zoneidx,
1939 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1940 int migratetype, unsigned long *did_some_progress,
1941 bool sync_migration)
1945 #endif /* CONFIG_COMPACTION */
1947 /* The really slow allocator path where we enter direct reclaim */
1948 static inline struct page *
1949 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1950 struct zonelist *zonelist, enum zone_type high_zoneidx,
1951 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1952 int migratetype, unsigned long *did_some_progress)
1954 struct page *page = NULL;
1955 struct reclaim_state reclaim_state;
1956 bool drained = false;
1960 /* We now go into synchronous reclaim */
1961 cpuset_memory_pressure_bump();
1962 current->flags |= PF_MEMALLOC;
1963 lockdep_set_current_reclaim_state(gfp_mask);
1964 reclaim_state.reclaimed_slab = 0;
1965 current->reclaim_state = &reclaim_state;
1967 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1969 current->reclaim_state = NULL;
1970 lockdep_clear_current_reclaim_state();
1971 current->flags &= ~PF_MEMALLOC;
1975 if (unlikely(!(*did_some_progress)))
1978 /* After successful reclaim, reconsider all zones for allocation */
1980 zlc_clear_zones_full(zonelist);
1983 page = get_page_from_freelist(gfp_mask, nodemask, order,
1984 zonelist, high_zoneidx,
1985 alloc_flags, preferred_zone,
1989 * If an allocation failed after direct reclaim, it could be because
1990 * pages are pinned on the per-cpu lists. Drain them and try again
1992 if (!page && !drained) {
2002 * This is called in the allocator slow-path if the allocation request is of
2003 * sufficient urgency to ignore watermarks and take other desperate measures
2005 static inline struct page *
2006 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2007 struct zonelist *zonelist, enum zone_type high_zoneidx,
2008 nodemask_t *nodemask, struct zone *preferred_zone,
2014 page = get_page_from_freelist(gfp_mask, nodemask, order,
2015 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2016 preferred_zone, migratetype);
2018 if (!page && gfp_mask & __GFP_NOFAIL)
2019 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2020 } while (!page && (gfp_mask & __GFP_NOFAIL));
2026 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2027 enum zone_type high_zoneidx,
2028 enum zone_type classzone_idx)
2033 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2034 wakeup_kswapd(zone, order, classzone_idx);
2038 gfp_to_alloc_flags(gfp_t gfp_mask)
2040 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2041 const gfp_t wait = gfp_mask & __GFP_WAIT;
2043 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2044 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2047 * The caller may dip into page reserves a bit more if the caller
2048 * cannot run direct reclaim, or if the caller has realtime scheduling
2049 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2050 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2052 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2056 * Not worth trying to allocate harder for
2057 * __GFP_NOMEMALLOC even if it can't schedule.
2059 if (!(gfp_mask & __GFP_NOMEMALLOC))
2060 alloc_flags |= ALLOC_HARDER;
2062 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2063 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2065 alloc_flags &= ~ALLOC_CPUSET;
2066 } else if (unlikely(rt_task(current)) && !in_interrupt())
2067 alloc_flags |= ALLOC_HARDER;
2069 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2070 if (!in_interrupt() &&
2071 ((current->flags & PF_MEMALLOC) ||
2072 unlikely(test_thread_flag(TIF_MEMDIE))))
2073 alloc_flags |= ALLOC_NO_WATERMARKS;
2079 static inline struct page *
2080 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2081 struct zonelist *zonelist, enum zone_type high_zoneidx,
2082 nodemask_t *nodemask, struct zone *preferred_zone,
2085 const gfp_t wait = gfp_mask & __GFP_WAIT;
2086 struct page *page = NULL;
2088 unsigned long pages_reclaimed = 0;
2089 unsigned long did_some_progress;
2090 bool sync_migration = false;
2093 * In the slowpath, we sanity check order to avoid ever trying to
2094 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2095 * be using allocators in order of preference for an area that is
2098 if (order >= MAX_ORDER) {
2099 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2104 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2105 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2106 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2107 * using a larger set of nodes after it has established that the
2108 * allowed per node queues are empty and that nodes are
2111 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2115 if (!(gfp_mask & __GFP_NO_KSWAPD))
2116 wake_all_kswapd(order, zonelist, high_zoneidx,
2117 zone_idx(preferred_zone));
2120 * OK, we're below the kswapd watermark and have kicked background
2121 * reclaim. Now things get more complex, so set up alloc_flags according
2122 * to how we want to proceed.
2124 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2127 * Find the true preferred zone if the allocation is unconstrained by
2130 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2131 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2135 /* This is the last chance, in general, before the goto nopage. */
2136 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2137 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2138 preferred_zone, migratetype);
2142 /* Allocate without watermarks if the context allows */
2143 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2144 page = __alloc_pages_high_priority(gfp_mask, order,
2145 zonelist, high_zoneidx, nodemask,
2146 preferred_zone, migratetype);
2151 /* Atomic allocations - we can't balance anything */
2155 /* Avoid recursion of direct reclaim */
2156 if (current->flags & PF_MEMALLOC)
2159 /* Avoid allocations with no watermarks from looping endlessly */
2160 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2164 * Try direct compaction. The first pass is asynchronous. Subsequent
2165 * attempts after direct reclaim are synchronous
2167 page = __alloc_pages_direct_compact(gfp_mask, order,
2168 zonelist, high_zoneidx,
2170 alloc_flags, preferred_zone,
2171 migratetype, &did_some_progress,
2175 sync_migration = true;
2177 /* Try direct reclaim and then allocating */
2178 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2179 zonelist, high_zoneidx,
2181 alloc_flags, preferred_zone,
2182 migratetype, &did_some_progress);
2187 * If we failed to make any progress reclaiming, then we are
2188 * running out of options and have to consider going OOM
2190 if (!did_some_progress) {
2191 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2192 if (oom_killer_disabled)
2194 page = __alloc_pages_may_oom(gfp_mask, order,
2195 zonelist, high_zoneidx,
2196 nodemask, preferred_zone,
2201 if (!(gfp_mask & __GFP_NOFAIL)) {
2203 * The oom killer is not called for high-order
2204 * allocations that may fail, so if no progress
2205 * is being made, there are no other options and
2206 * retrying is unlikely to help.
2208 if (order > PAGE_ALLOC_COSTLY_ORDER)
2211 * The oom killer is not called for lowmem
2212 * allocations to prevent needlessly killing
2215 if (high_zoneidx < ZONE_NORMAL)
2223 /* Check if we should retry the allocation */
2224 pages_reclaimed += did_some_progress;
2225 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2227 /* Wait for some write requests to complete then retry */
2228 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2232 * High-order allocations do not necessarily loop after
2233 * direct reclaim and reclaim/compaction depends on compaction
2234 * being called after reclaim so call directly if necessary
2236 page = __alloc_pages_direct_compact(gfp_mask, order,
2237 zonelist, high_zoneidx,
2239 alloc_flags, preferred_zone,
2240 migratetype, &did_some_progress,
2247 warn_alloc_failed(gfp_mask, order, NULL);
2250 if (kmemcheck_enabled)
2251 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2257 * This is the 'heart' of the zoned buddy allocator.
2260 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2261 struct zonelist *zonelist, nodemask_t *nodemask)
2263 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2264 struct zone *preferred_zone;
2266 int migratetype = allocflags_to_migratetype(gfp_mask);
2268 gfp_mask &= gfp_allowed_mask;
2270 lockdep_trace_alloc(gfp_mask);
2272 might_sleep_if(gfp_mask & __GFP_WAIT);
2274 if (should_fail_alloc_page(gfp_mask, order))
2278 * Check the zones suitable for the gfp_mask contain at least one
2279 * valid zone. It's possible to have an empty zonelist as a result
2280 * of GFP_THISNODE and a memoryless node
2282 if (unlikely(!zonelist->_zonerefs->zone))
2286 /* The preferred zone is used for statistics later */
2287 first_zones_zonelist(zonelist, high_zoneidx,
2288 nodemask ? : &cpuset_current_mems_allowed,
2290 if (!preferred_zone) {
2295 /* First allocation attempt */
2296 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2297 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2298 preferred_zone, migratetype);
2299 if (unlikely(!page))
2300 page = __alloc_pages_slowpath(gfp_mask, order,
2301 zonelist, high_zoneidx, nodemask,
2302 preferred_zone, migratetype);
2305 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2308 EXPORT_SYMBOL(__alloc_pages_nodemask);
2311 * Common helper functions.
2313 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2318 * __get_free_pages() returns a 32-bit address, which cannot represent
2321 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2323 page = alloc_pages(gfp_mask, order);
2326 return (unsigned long) page_address(page);
2328 EXPORT_SYMBOL(__get_free_pages);
2330 unsigned long get_zeroed_page(gfp_t gfp_mask)
2332 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2334 EXPORT_SYMBOL(get_zeroed_page);
2336 void __free_pages(struct page *page, unsigned int order)
2338 if (put_page_testzero(page)) {
2340 free_hot_cold_page(page, 0);
2342 __free_pages_ok(page, order);
2346 EXPORT_SYMBOL(__free_pages);
2348 void free_pages(unsigned long addr, unsigned int order)
2351 VM_BUG_ON(!virt_addr_valid((void *)addr));
2352 __free_pages(virt_to_page((void *)addr), order);
2356 EXPORT_SYMBOL(free_pages);
2358 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2361 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2362 unsigned long used = addr + PAGE_ALIGN(size);
2364 split_page(virt_to_page((void *)addr), order);
2365 while (used < alloc_end) {
2370 return (void *)addr;
2374 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2375 * @size: the number of bytes to allocate
2376 * @gfp_mask: GFP flags for the allocation
2378 * This function is similar to alloc_pages(), except that it allocates the
2379 * minimum number of pages to satisfy the request. alloc_pages() can only
2380 * allocate memory in power-of-two pages.
2382 * This function is also limited by MAX_ORDER.
2384 * Memory allocated by this function must be released by free_pages_exact().
2386 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2388 unsigned int order = get_order(size);
2391 addr = __get_free_pages(gfp_mask, order);
2392 return make_alloc_exact(addr, order, size);
2394 EXPORT_SYMBOL(alloc_pages_exact);
2397 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2399 * @nid: the preferred node ID where memory should be allocated
2400 * @size: the number of bytes to allocate
2401 * @gfp_mask: GFP flags for the allocation
2403 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2405 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2408 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2410 unsigned order = get_order(size);
2411 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2414 return make_alloc_exact((unsigned long)page_address(p), order, size);
2416 EXPORT_SYMBOL(alloc_pages_exact_nid);
2419 * free_pages_exact - release memory allocated via alloc_pages_exact()
2420 * @virt: the value returned by alloc_pages_exact.
2421 * @size: size of allocation, same value as passed to alloc_pages_exact().
2423 * Release the memory allocated by a previous call to alloc_pages_exact.
2425 void free_pages_exact(void *virt, size_t size)
2427 unsigned long addr = (unsigned long)virt;
2428 unsigned long end = addr + PAGE_ALIGN(size);
2430 while (addr < end) {
2435 EXPORT_SYMBOL(free_pages_exact);
2437 static unsigned int nr_free_zone_pages(int offset)
2442 /* Just pick one node, since fallback list is circular */
2443 unsigned int sum = 0;
2445 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2447 for_each_zone_zonelist(zone, z, zonelist, offset) {
2448 unsigned long size = zone->present_pages;
2449 unsigned long high = high_wmark_pages(zone);
2458 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2460 unsigned int nr_free_buffer_pages(void)
2462 return nr_free_zone_pages(gfp_zone(GFP_USER));
2464 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2467 * Amount of free RAM allocatable within all zones
2469 unsigned int nr_free_pagecache_pages(void)
2471 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2474 static inline void show_node(struct zone *zone)
2477 printk("Node %d ", zone_to_nid(zone));
2480 void si_meminfo(struct sysinfo *val)
2482 val->totalram = totalram_pages;
2484 val->freeram = global_page_state(NR_FREE_PAGES);
2485 val->bufferram = nr_blockdev_pages();
2486 val->totalhigh = totalhigh_pages;
2487 val->freehigh = nr_free_highpages();
2488 val->mem_unit = PAGE_SIZE;
2491 EXPORT_SYMBOL(si_meminfo);
2494 void si_meminfo_node(struct sysinfo *val, int nid)
2496 pg_data_t *pgdat = NODE_DATA(nid);
2498 val->totalram = pgdat->node_present_pages;
2499 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2500 #ifdef CONFIG_HIGHMEM
2501 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2502 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2508 val->mem_unit = PAGE_SIZE;
2513 * Determine whether the node should be displayed or not, depending on whether
2514 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2516 bool skip_free_areas_node(unsigned int flags, int nid)
2520 if (!(flags & SHOW_MEM_FILTER_NODES))
2524 ret = !node_isset(nid, cpuset_current_mems_allowed);
2530 #define K(x) ((x) << (PAGE_SHIFT-10))
2533 * Show free area list (used inside shift_scroll-lock stuff)
2534 * We also calculate the percentage fragmentation. We do this by counting the
2535 * memory on each free list with the exception of the first item on the list.
2536 * Suppresses nodes that are not allowed by current's cpuset if
2537 * SHOW_MEM_FILTER_NODES is passed.
2539 void show_free_areas(unsigned int filter)
2544 for_each_populated_zone(zone) {
2545 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2548 printk("%s per-cpu:\n", zone->name);
2550 for_each_online_cpu(cpu) {
2551 struct per_cpu_pageset *pageset;
2553 pageset = per_cpu_ptr(zone->pageset, cpu);
2555 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2556 cpu, pageset->pcp.high,
2557 pageset->pcp.batch, pageset->pcp.count);
2561 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2562 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2564 " dirty:%lu writeback:%lu unstable:%lu\n"
2565 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2566 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2567 global_page_state(NR_ACTIVE_ANON),
2568 global_page_state(NR_INACTIVE_ANON),
2569 global_page_state(NR_ISOLATED_ANON),
2570 global_page_state(NR_ACTIVE_FILE),
2571 global_page_state(NR_INACTIVE_FILE),
2572 global_page_state(NR_ISOLATED_FILE),
2573 global_page_state(NR_UNEVICTABLE),
2574 global_page_state(NR_FILE_DIRTY),
2575 global_page_state(NR_WRITEBACK),
2576 global_page_state(NR_UNSTABLE_NFS),
2577 global_page_state(NR_FREE_PAGES),
2578 global_page_state(NR_SLAB_RECLAIMABLE),
2579 global_page_state(NR_SLAB_UNRECLAIMABLE),
2580 global_page_state(NR_FILE_MAPPED),
2581 global_page_state(NR_SHMEM),
2582 global_page_state(NR_PAGETABLE),
2583 global_page_state(NR_BOUNCE));
2585 for_each_populated_zone(zone) {
2588 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2596 " active_anon:%lukB"
2597 " inactive_anon:%lukB"
2598 " active_file:%lukB"
2599 " inactive_file:%lukB"
2600 " unevictable:%lukB"
2601 " isolated(anon):%lukB"
2602 " isolated(file):%lukB"
2609 " slab_reclaimable:%lukB"
2610 " slab_unreclaimable:%lukB"
2611 " kernel_stack:%lukB"
2615 " writeback_tmp:%lukB"
2616 " pages_scanned:%lu"
2617 " all_unreclaimable? %s"
2620 K(zone_page_state(zone, NR_FREE_PAGES)),
2621 K(min_wmark_pages(zone)),
2622 K(low_wmark_pages(zone)),
2623 K(high_wmark_pages(zone)),
2624 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2625 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2626 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2627 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2628 K(zone_page_state(zone, NR_UNEVICTABLE)),
2629 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2630 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2631 K(zone->present_pages),
2632 K(zone_page_state(zone, NR_MLOCK)),
2633 K(zone_page_state(zone, NR_FILE_DIRTY)),
2634 K(zone_page_state(zone, NR_WRITEBACK)),
2635 K(zone_page_state(zone, NR_FILE_MAPPED)),
2636 K(zone_page_state(zone, NR_SHMEM)),
2637 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2638 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2639 zone_page_state(zone, NR_KERNEL_STACK) *
2641 K(zone_page_state(zone, NR_PAGETABLE)),
2642 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2643 K(zone_page_state(zone, NR_BOUNCE)),
2644 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2645 zone->pages_scanned,
2646 (zone->all_unreclaimable ? "yes" : "no")
2648 printk("lowmem_reserve[]:");
2649 for (i = 0; i < MAX_NR_ZONES; i++)
2650 printk(" %lu", zone->lowmem_reserve[i]);
2654 for_each_populated_zone(zone) {
2655 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2657 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2660 printk("%s: ", zone->name);
2662 spin_lock_irqsave(&zone->lock, flags);
2663 for (order = 0; order < MAX_ORDER; order++) {
2664 nr[order] = zone->free_area[order].nr_free;
2665 total += nr[order] << order;
2667 spin_unlock_irqrestore(&zone->lock, flags);
2668 for (order = 0; order < MAX_ORDER; order++)
2669 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2670 printk("= %lukB\n", K(total));
2673 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2675 show_swap_cache_info();
2678 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2680 zoneref->zone = zone;
2681 zoneref->zone_idx = zone_idx(zone);
2685 * Builds allocation fallback zone lists.
2687 * Add all populated zones of a node to the zonelist.
2689 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2690 int nr_zones, enum zone_type zone_type)
2694 BUG_ON(zone_type >= MAX_NR_ZONES);
2699 zone = pgdat->node_zones + zone_type;
2700 if (populated_zone(zone)) {
2701 zoneref_set_zone(zone,
2702 &zonelist->_zonerefs[nr_zones++]);
2703 check_highest_zone(zone_type);
2706 } while (zone_type);
2713 * 0 = automatic detection of better ordering.
2714 * 1 = order by ([node] distance, -zonetype)
2715 * 2 = order by (-zonetype, [node] distance)
2717 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2718 * the same zonelist. So only NUMA can configure this param.
2720 #define ZONELIST_ORDER_DEFAULT 0
2721 #define ZONELIST_ORDER_NODE 1
2722 #define ZONELIST_ORDER_ZONE 2
2724 /* zonelist order in the kernel.
2725 * set_zonelist_order() will set this to NODE or ZONE.
2727 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2728 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2732 /* The value user specified ....changed by config */
2733 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2734 /* string for sysctl */
2735 #define NUMA_ZONELIST_ORDER_LEN 16
2736 char numa_zonelist_order[16] = "default";
2739 * interface for configure zonelist ordering.
2740 * command line option "numa_zonelist_order"
2741 * = "[dD]efault - default, automatic configuration.
2742 * = "[nN]ode - order by node locality, then by zone within node
2743 * = "[zZ]one - order by zone, then by locality within zone
2746 static int __parse_numa_zonelist_order(char *s)
2748 if (*s == 'd' || *s == 'D') {
2749 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2750 } else if (*s == 'n' || *s == 'N') {
2751 user_zonelist_order = ZONELIST_ORDER_NODE;
2752 } else if (*s == 'z' || *s == 'Z') {
2753 user_zonelist_order = ZONELIST_ORDER_ZONE;
2756 "Ignoring invalid numa_zonelist_order value: "
2763 static __init int setup_numa_zonelist_order(char *s)
2770 ret = __parse_numa_zonelist_order(s);
2772 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2776 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2779 * sysctl handler for numa_zonelist_order
2781 int numa_zonelist_order_handler(ctl_table *table, int write,
2782 void __user *buffer, size_t *length,
2785 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2787 static DEFINE_MUTEX(zl_order_mutex);
2789 mutex_lock(&zl_order_mutex);
2791 strcpy(saved_string, (char*)table->data);
2792 ret = proc_dostring(table, write, buffer, length, ppos);
2796 int oldval = user_zonelist_order;
2797 if (__parse_numa_zonelist_order((char*)table->data)) {
2799 * bogus value. restore saved string
2801 strncpy((char*)table->data, saved_string,
2802 NUMA_ZONELIST_ORDER_LEN);
2803 user_zonelist_order = oldval;
2804 } else if (oldval != user_zonelist_order) {
2805 mutex_lock(&zonelists_mutex);
2806 build_all_zonelists(NULL);
2807 mutex_unlock(&zonelists_mutex);
2811 mutex_unlock(&zl_order_mutex);
2816 #define MAX_NODE_LOAD (nr_online_nodes)
2817 static int node_load[MAX_NUMNODES];
2820 * find_next_best_node - find the next node that should appear in a given node's fallback list
2821 * @node: node whose fallback list we're appending
2822 * @used_node_mask: nodemask_t of already used nodes
2824 * We use a number of factors to determine which is the next node that should
2825 * appear on a given node's fallback list. The node should not have appeared
2826 * already in @node's fallback list, and it should be the next closest node
2827 * according to the distance array (which contains arbitrary distance values
2828 * from each node to each node in the system), and should also prefer nodes
2829 * with no CPUs, since presumably they'll have very little allocation pressure
2830 * on them otherwise.
2831 * It returns -1 if no node is found.
2833 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2836 int min_val = INT_MAX;
2838 const struct cpumask *tmp = cpumask_of_node(0);
2840 /* Use the local node if we haven't already */
2841 if (!node_isset(node, *used_node_mask)) {
2842 node_set(node, *used_node_mask);
2846 for_each_node_state(n, N_HIGH_MEMORY) {
2848 /* Don't want a node to appear more than once */
2849 if (node_isset(n, *used_node_mask))
2852 /* Use the distance array to find the distance */
2853 val = node_distance(node, n);
2855 /* Penalize nodes under us ("prefer the next node") */
2858 /* Give preference to headless and unused nodes */
2859 tmp = cpumask_of_node(n);
2860 if (!cpumask_empty(tmp))
2861 val += PENALTY_FOR_NODE_WITH_CPUS;
2863 /* Slight preference for less loaded node */
2864 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2865 val += node_load[n];
2867 if (val < min_val) {
2874 node_set(best_node, *used_node_mask);
2881 * Build zonelists ordered by node and zones within node.
2882 * This results in maximum locality--normal zone overflows into local
2883 * DMA zone, if any--but risks exhausting DMA zone.
2885 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2888 struct zonelist *zonelist;
2890 zonelist = &pgdat->node_zonelists[0];
2891 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2893 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2895 zonelist->_zonerefs[j].zone = NULL;
2896 zonelist->_zonerefs[j].zone_idx = 0;
2900 * Build gfp_thisnode zonelists
2902 static void build_thisnode_zonelists(pg_data_t *pgdat)
2905 struct zonelist *zonelist;
2907 zonelist = &pgdat->node_zonelists[1];
2908 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2909 zonelist->_zonerefs[j].zone = NULL;
2910 zonelist->_zonerefs[j].zone_idx = 0;
2914 * Build zonelists ordered by zone and nodes within zones.
2915 * This results in conserving DMA zone[s] until all Normal memory is
2916 * exhausted, but results in overflowing to remote node while memory
2917 * may still exist in local DMA zone.
2919 static int node_order[MAX_NUMNODES];
2921 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2924 int zone_type; /* needs to be signed */
2926 struct zonelist *zonelist;
2928 zonelist = &pgdat->node_zonelists[0];
2930 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2931 for (j = 0; j < nr_nodes; j++) {
2932 node = node_order[j];
2933 z = &NODE_DATA(node)->node_zones[zone_type];
2934 if (populated_zone(z)) {
2936 &zonelist->_zonerefs[pos++]);
2937 check_highest_zone(zone_type);
2941 zonelist->_zonerefs[pos].zone = NULL;
2942 zonelist->_zonerefs[pos].zone_idx = 0;
2945 static int default_zonelist_order(void)
2948 unsigned long low_kmem_size,total_size;
2952 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2953 * If they are really small and used heavily, the system can fall
2954 * into OOM very easily.
2955 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2957 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2960 for_each_online_node(nid) {
2961 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2962 z = &NODE_DATA(nid)->node_zones[zone_type];
2963 if (populated_zone(z)) {
2964 if (zone_type < ZONE_NORMAL)
2965 low_kmem_size += z->present_pages;
2966 total_size += z->present_pages;
2967 } else if (zone_type == ZONE_NORMAL) {
2969 * If any node has only lowmem, then node order
2970 * is preferred to allow kernel allocations
2971 * locally; otherwise, they can easily infringe
2972 * on other nodes when there is an abundance of
2973 * lowmem available to allocate from.
2975 return ZONELIST_ORDER_NODE;
2979 if (!low_kmem_size || /* there are no DMA area. */
2980 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2981 return ZONELIST_ORDER_NODE;
2983 * look into each node's config.
2984 * If there is a node whose DMA/DMA32 memory is very big area on
2985 * local memory, NODE_ORDER may be suitable.
2987 average_size = total_size /
2988 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2989 for_each_online_node(nid) {
2992 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2993 z = &NODE_DATA(nid)->node_zones[zone_type];
2994 if (populated_zone(z)) {
2995 if (zone_type < ZONE_NORMAL)
2996 low_kmem_size += z->present_pages;
2997 total_size += z->present_pages;
3000 if (low_kmem_size &&
3001 total_size > average_size && /* ignore small node */
3002 low_kmem_size > total_size * 70/100)
3003 return ZONELIST_ORDER_NODE;
3005 return ZONELIST_ORDER_ZONE;
3008 static void set_zonelist_order(void)
3010 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3011 current_zonelist_order = default_zonelist_order();
3013 current_zonelist_order = user_zonelist_order;
3016 static void build_zonelists(pg_data_t *pgdat)
3020 nodemask_t used_mask;
3021 int local_node, prev_node;
3022 struct zonelist *zonelist;
3023 int order = current_zonelist_order;
3025 /* initialize zonelists */
3026 for (i = 0; i < MAX_ZONELISTS; i++) {
3027 zonelist = pgdat->node_zonelists + i;
3028 zonelist->_zonerefs[0].zone = NULL;
3029 zonelist->_zonerefs[0].zone_idx = 0;
3032 /* NUMA-aware ordering of nodes */
3033 local_node = pgdat->node_id;
3034 load = nr_online_nodes;
3035 prev_node = local_node;
3036 nodes_clear(used_mask);
3038 memset(node_order, 0, sizeof(node_order));
3041 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3042 int distance = node_distance(local_node, node);
3045 * If another node is sufficiently far away then it is better
3046 * to reclaim pages in a zone before going off node.
3048 if (distance > RECLAIM_DISTANCE)
3049 zone_reclaim_mode = 1;
3052 * We don't want to pressure a particular node.
3053 * So adding penalty to the first node in same
3054 * distance group to make it round-robin.
3056 if (distance != node_distance(local_node, prev_node))
3057 node_load[node] = load;
3061 if (order == ZONELIST_ORDER_NODE)
3062 build_zonelists_in_node_order(pgdat, node);
3064 node_order[j++] = node; /* remember order */
3067 if (order == ZONELIST_ORDER_ZONE) {
3068 /* calculate node order -- i.e., DMA last! */
3069 build_zonelists_in_zone_order(pgdat, j);
3072 build_thisnode_zonelists(pgdat);
3075 /* Construct the zonelist performance cache - see further mmzone.h */
3076 static void build_zonelist_cache(pg_data_t *pgdat)
3078 struct zonelist *zonelist;
3079 struct zonelist_cache *zlc;
3082 zonelist = &pgdat->node_zonelists[0];
3083 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3084 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3085 for (z = zonelist->_zonerefs; z->zone; z++)
3086 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3089 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3091 * Return node id of node used for "local" allocations.
3092 * I.e., first node id of first zone in arg node's generic zonelist.
3093 * Used for initializing percpu 'numa_mem', which is used primarily
3094 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3096 int local_memory_node(int node)
3100 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3101 gfp_zone(GFP_KERNEL),
3108 #else /* CONFIG_NUMA */
3110 static void set_zonelist_order(void)
3112 current_zonelist_order = ZONELIST_ORDER_ZONE;
3115 static void build_zonelists(pg_data_t *pgdat)
3117 int node, local_node;
3119 struct zonelist *zonelist;
3121 local_node = pgdat->node_id;
3123 zonelist = &pgdat->node_zonelists[0];
3124 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3127 * Now we build the zonelist so that it contains the zones
3128 * of all the other nodes.
3129 * We don't want to pressure a particular node, so when
3130 * building the zones for node N, we make sure that the
3131 * zones coming right after the local ones are those from
3132 * node N+1 (modulo N)
3134 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3135 if (!node_online(node))
3137 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3140 for (node = 0; node < local_node; node++) {
3141 if (!node_online(node))
3143 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3147 zonelist->_zonerefs[j].zone = NULL;
3148 zonelist->_zonerefs[j].zone_idx = 0;
3151 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3152 static void build_zonelist_cache(pg_data_t *pgdat)
3154 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3157 #endif /* CONFIG_NUMA */
3160 * Boot pageset table. One per cpu which is going to be used for all
3161 * zones and all nodes. The parameters will be set in such a way
3162 * that an item put on a list will immediately be handed over to
3163 * the buddy list. This is safe since pageset manipulation is done
3164 * with interrupts disabled.
3166 * The boot_pagesets must be kept even after bootup is complete for
3167 * unused processors and/or zones. They do play a role for bootstrapping
3168 * hotplugged processors.
3170 * zoneinfo_show() and maybe other functions do
3171 * not check if the processor is online before following the pageset pointer.
3172 * Other parts of the kernel may not check if the zone is available.
3174 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3175 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3176 static void setup_zone_pageset(struct zone *zone);
3179 * Global mutex to protect against size modification of zonelists
3180 * as well as to serialize pageset setup for the new populated zone.
3182 DEFINE_MUTEX(zonelists_mutex);
3184 /* return values int ....just for stop_machine() */
3185 static __init_refok int __build_all_zonelists(void *data)
3191 memset(node_load, 0, sizeof(node_load));
3193 for_each_online_node(nid) {
3194 pg_data_t *pgdat = NODE_DATA(nid);
3196 build_zonelists(pgdat);
3197 build_zonelist_cache(pgdat);
3201 * Initialize the boot_pagesets that are going to be used
3202 * for bootstrapping processors. The real pagesets for
3203 * each zone will be allocated later when the per cpu
3204 * allocator is available.
3206 * boot_pagesets are used also for bootstrapping offline
3207 * cpus if the system is already booted because the pagesets
3208 * are needed to initialize allocators on a specific cpu too.
3209 * F.e. the percpu allocator needs the page allocator which
3210 * needs the percpu allocator in order to allocate its pagesets
3211 * (a chicken-egg dilemma).
3213 for_each_possible_cpu(cpu) {
3214 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3216 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3218 * We now know the "local memory node" for each node--
3219 * i.e., the node of the first zone in the generic zonelist.
3220 * Set up numa_mem percpu variable for on-line cpus. During
3221 * boot, only the boot cpu should be on-line; we'll init the
3222 * secondary cpus' numa_mem as they come on-line. During
3223 * node/memory hotplug, we'll fixup all on-line cpus.
3225 if (cpu_online(cpu))
3226 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3234 * Called with zonelists_mutex held always
3235 * unless system_state == SYSTEM_BOOTING.
3237 void __ref build_all_zonelists(void *data)
3239 set_zonelist_order();
3241 if (system_state == SYSTEM_BOOTING) {
3242 __build_all_zonelists(NULL);
3243 mminit_verify_zonelist();
3244 cpuset_init_current_mems_allowed();
3246 /* we have to stop all cpus to guarantee there is no user
3248 #ifdef CONFIG_MEMORY_HOTPLUG
3250 setup_zone_pageset((struct zone *)data);
3252 stop_machine(__build_all_zonelists, NULL, NULL);
3253 /* cpuset refresh routine should be here */
3255 vm_total_pages = nr_free_pagecache_pages();
3257 * Disable grouping by mobility if the number of pages in the
3258 * system is too low to allow the mechanism to work. It would be
3259 * more accurate, but expensive to check per-zone. This check is
3260 * made on memory-hotadd so a system can start with mobility
3261 * disabled and enable it later
3263 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3264 page_group_by_mobility_disabled = 1;
3266 page_group_by_mobility_disabled = 0;
3268 printk("Built %i zonelists in %s order, mobility grouping %s. "
3269 "Total pages: %ld\n",
3271 zonelist_order_name[current_zonelist_order],
3272 page_group_by_mobility_disabled ? "off" : "on",
3275 printk("Policy zone: %s\n", zone_names[policy_zone]);
3280 * Helper functions to size the waitqueue hash table.
3281 * Essentially these want to choose hash table sizes sufficiently
3282 * large so that collisions trying to wait on pages are rare.
3283 * But in fact, the number of active page waitqueues on typical
3284 * systems is ridiculously low, less than 200. So this is even
3285 * conservative, even though it seems large.
3287 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3288 * waitqueues, i.e. the size of the waitq table given the number of pages.
3290 #define PAGES_PER_WAITQUEUE 256
3292 #ifndef CONFIG_MEMORY_HOTPLUG
3293 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3295 unsigned long size = 1;
3297 pages /= PAGES_PER_WAITQUEUE;
3299 while (size < pages)
3303 * Once we have dozens or even hundreds of threads sleeping
3304 * on IO we've got bigger problems than wait queue collision.
3305 * Limit the size of the wait table to a reasonable size.
3307 size = min(size, 4096UL);
3309 return max(size, 4UL);
3313 * A zone's size might be changed by hot-add, so it is not possible to determine
3314 * a suitable size for its wait_table. So we use the maximum size now.
3316 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3318 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3319 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3320 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3322 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3323 * or more by the traditional way. (See above). It equals:
3325 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3326 * ia64(16K page size) : = ( 8G + 4M)byte.
3327 * powerpc (64K page size) : = (32G +16M)byte.
3329 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3336 * This is an integer logarithm so that shifts can be used later
3337 * to extract the more random high bits from the multiplicative
3338 * hash function before the remainder is taken.
3340 static inline unsigned long wait_table_bits(unsigned long size)
3345 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3348 * Check if a pageblock contains reserved pages
3350 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3354 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3355 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3362 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3363 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3364 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3365 * higher will lead to a bigger reserve which will get freed as contiguous
3366 * blocks as reclaim kicks in
3368 static void setup_zone_migrate_reserve(struct zone *zone)
3370 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3372 unsigned long block_migratetype;
3376 * Get the start pfn, end pfn and the number of blocks to reserve
3377 * We have to be careful to be aligned to pageblock_nr_pages to
3378 * make sure that we always check pfn_valid for the first page in
3381 start_pfn = zone->zone_start_pfn;
3382 end_pfn = start_pfn + zone->spanned_pages;
3383 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3384 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3388 * Reserve blocks are generally in place to help high-order atomic
3389 * allocations that are short-lived. A min_free_kbytes value that
3390 * would result in more than 2 reserve blocks for atomic allocations
3391 * is assumed to be in place to help anti-fragmentation for the
3392 * future allocation of hugepages at runtime.
3394 reserve = min(2, reserve);
3396 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3397 if (!pfn_valid(pfn))
3399 page = pfn_to_page(pfn);
3401 /* Watch out for overlapping nodes */
3402 if (page_to_nid(page) != zone_to_nid(zone))
3405 block_migratetype = get_pageblock_migratetype(page);
3407 /* Only test what is necessary when the reserves are not met */
3410 * Blocks with reserved pages will never free, skip
3413 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3414 if (pageblock_is_reserved(pfn, block_end_pfn))
3417 /* If this block is reserved, account for it */
3418 if (block_migratetype == MIGRATE_RESERVE) {
3423 /* Suitable for reserving if this block is movable */
3424 if (block_migratetype == MIGRATE_MOVABLE) {
3425 set_pageblock_migratetype(page,
3427 move_freepages_block(zone, page,
3435 * If the reserve is met and this is a previous reserved block,
3438 if (block_migratetype == MIGRATE_RESERVE) {
3439 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3440 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3446 * Initially all pages are reserved - free ones are freed
3447 * up by free_all_bootmem() once the early boot process is
3448 * done. Non-atomic initialization, single-pass.
3450 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3451 unsigned long start_pfn, enum memmap_context context)
3454 unsigned long end_pfn = start_pfn + size;
3458 if (highest_memmap_pfn < end_pfn - 1)
3459 highest_memmap_pfn = end_pfn - 1;
3461 z = &NODE_DATA(nid)->node_zones[zone];
3462 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3464 * There can be holes in boot-time mem_map[]s
3465 * handed to this function. They do not
3466 * exist on hotplugged memory.
3468 if (context == MEMMAP_EARLY) {
3469 if (!early_pfn_valid(pfn))
3471 if (!early_pfn_in_nid(pfn, nid))
3474 page = pfn_to_page(pfn);
3475 set_page_links(page, zone, nid, pfn);
3476 mminit_verify_page_links(page, zone, nid, pfn);
3477 init_page_count(page);
3478 reset_page_mapcount(page);
3479 SetPageReserved(page);
3481 * Mark the block movable so that blocks are reserved for
3482 * movable at startup. This will force kernel allocations
3483 * to reserve their blocks rather than leaking throughout
3484 * the address space during boot when many long-lived
3485 * kernel allocations are made. Later some blocks near
3486 * the start are marked MIGRATE_RESERVE by
3487 * setup_zone_migrate_reserve()
3489 * bitmap is created for zone's valid pfn range. but memmap
3490 * can be created for invalid pages (for alignment)
3491 * check here not to call set_pageblock_migratetype() against
3494 if ((z->zone_start_pfn <= pfn)
3495 && (pfn < z->zone_start_pfn + z->spanned_pages)
3496 && !(pfn & (pageblock_nr_pages - 1)))
3497 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3499 INIT_LIST_HEAD(&page->lru);
3500 #ifdef WANT_PAGE_VIRTUAL
3501 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3502 if (!is_highmem_idx(zone))
3503 set_page_address(page, __va(pfn << PAGE_SHIFT));
3508 static void __meminit zone_init_free_lists(struct zone *zone)
3511 for_each_migratetype_order(order, t) {
3512 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3513 zone->free_area[order].nr_free = 0;
3517 #ifndef __HAVE_ARCH_MEMMAP_INIT
3518 #define memmap_init(size, nid, zone, start_pfn) \
3519 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3522 static int zone_batchsize(struct zone *zone)
3528 * The per-cpu-pages pools are set to around 1000th of the
3529 * size of the zone. But no more than 1/2 of a meg.
3531 * OK, so we don't know how big the cache is. So guess.
3533 batch = zone->present_pages / 1024;
3534 if (batch * PAGE_SIZE > 512 * 1024)
3535 batch = (512 * 1024) / PAGE_SIZE;
3536 batch /= 4; /* We effectively *= 4 below */
3541 * Clamp the batch to a 2^n - 1 value. Having a power
3542 * of 2 value was found to be more likely to have
3543 * suboptimal cache aliasing properties in some cases.
3545 * For example if 2 tasks are alternately allocating
3546 * batches of pages, one task can end up with a lot
3547 * of pages of one half of the possible page colors
3548 * and the other with pages of the other colors.
3550 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3555 /* The deferral and batching of frees should be suppressed under NOMMU
3558 * The problem is that NOMMU needs to be able to allocate large chunks
3559 * of contiguous memory as there's no hardware page translation to
3560 * assemble apparent contiguous memory from discontiguous pages.
3562 * Queueing large contiguous runs of pages for batching, however,
3563 * causes the pages to actually be freed in smaller chunks. As there
3564 * can be a significant delay between the individual batches being
3565 * recycled, this leads to the once large chunks of space being
3566 * fragmented and becoming unavailable for high-order allocations.
3572 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3574 struct per_cpu_pages *pcp;
3577 memset(p, 0, sizeof(*p));
3581 pcp->high = 6 * batch;
3582 pcp->batch = max(1UL, 1 * batch);
3583 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3584 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3588 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3589 * to the value high for the pageset p.
3592 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3595 struct per_cpu_pages *pcp;
3599 pcp->batch = max(1UL, high/4);
3600 if ((high/4) > (PAGE_SHIFT * 8))
3601 pcp->batch = PAGE_SHIFT * 8;
3604 static void setup_zone_pageset(struct zone *zone)
3608 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3610 for_each_possible_cpu(cpu) {
3611 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3613 setup_pageset(pcp, zone_batchsize(zone));
3615 if (percpu_pagelist_fraction)
3616 setup_pagelist_highmark(pcp,
3617 (zone->present_pages /
3618 percpu_pagelist_fraction));
3623 * Allocate per cpu pagesets and initialize them.
3624 * Before this call only boot pagesets were available.
3626 void __init setup_per_cpu_pageset(void)
3630 for_each_populated_zone(zone)
3631 setup_zone_pageset(zone);
3634 static noinline __init_refok
3635 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3638 struct pglist_data *pgdat = zone->zone_pgdat;
3642 * The per-page waitqueue mechanism uses hashed waitqueues
3645 zone->wait_table_hash_nr_entries =
3646 wait_table_hash_nr_entries(zone_size_pages);
3647 zone->wait_table_bits =
3648 wait_table_bits(zone->wait_table_hash_nr_entries);
3649 alloc_size = zone->wait_table_hash_nr_entries
3650 * sizeof(wait_queue_head_t);
3652 if (!slab_is_available()) {
3653 zone->wait_table = (wait_queue_head_t *)
3654 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3657 * This case means that a zone whose size was 0 gets new memory
3658 * via memory hot-add.
3659 * But it may be the case that a new node was hot-added. In
3660 * this case vmalloc() will not be able to use this new node's
3661 * memory - this wait_table must be initialized to use this new
3662 * node itself as well.
3663 * To use this new node's memory, further consideration will be
3666 zone->wait_table = vmalloc(alloc_size);
3668 if (!zone->wait_table)
3671 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3672 init_waitqueue_head(zone->wait_table + i);
3677 static int __zone_pcp_update(void *data)
3679 struct zone *zone = data;
3681 unsigned long batch = zone_batchsize(zone), flags;
3683 for_each_possible_cpu(cpu) {
3684 struct per_cpu_pageset *pset;
3685 struct per_cpu_pages *pcp;
3687 pset = per_cpu_ptr(zone->pageset, cpu);
3690 local_irq_save(flags);
3691 free_pcppages_bulk(zone, pcp->count, pcp);
3692 setup_pageset(pset, batch);
3693 local_irq_restore(flags);
3698 void zone_pcp_update(struct zone *zone)
3700 stop_machine(__zone_pcp_update, zone, NULL);
3703 static __meminit void zone_pcp_init(struct zone *zone)
3706 * per cpu subsystem is not up at this point. The following code
3707 * relies on the ability of the linker to provide the
3708 * offset of a (static) per cpu variable into the per cpu area.
3710 zone->pageset = &boot_pageset;
3712 if (zone->present_pages)
3713 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3714 zone->name, zone->present_pages,
3715 zone_batchsize(zone));
3718 __meminit int init_currently_empty_zone(struct zone *zone,
3719 unsigned long zone_start_pfn,
3721 enum memmap_context context)
3723 struct pglist_data *pgdat = zone->zone_pgdat;
3725 ret = zone_wait_table_init(zone, size);
3728 pgdat->nr_zones = zone_idx(zone) + 1;
3730 zone->zone_start_pfn = zone_start_pfn;
3732 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3733 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3735 (unsigned long)zone_idx(zone),
3736 zone_start_pfn, (zone_start_pfn + size));
3738 zone_init_free_lists(zone);
3743 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
3744 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3746 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3747 * Architectures may implement their own version but if add_active_range()
3748 * was used and there are no special requirements, this is a convenient
3751 int __meminit __early_pfn_to_nid(unsigned long pfn)
3753 unsigned long start_pfn, end_pfn;
3756 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
3757 if (start_pfn <= pfn && pfn < end_pfn)
3759 /* This is a memory hole */
3762 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3764 int __meminit early_pfn_to_nid(unsigned long pfn)
3768 nid = __early_pfn_to_nid(pfn);
3771 /* just returns 0 */
3775 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3776 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3780 nid = __early_pfn_to_nid(pfn);
3781 if (nid >= 0 && nid != node)
3788 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3789 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3790 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3792 * If an architecture guarantees that all ranges registered with
3793 * add_active_ranges() contain no holes and may be freed, this
3794 * this function may be used instead of calling free_bootmem() manually.
3796 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
3798 unsigned long start_pfn, end_pfn;
3801 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
3802 start_pfn = min(start_pfn, max_low_pfn);
3803 end_pfn = min(end_pfn, max_low_pfn);
3805 if (start_pfn < end_pfn)
3806 free_bootmem_node(NODE_DATA(this_nid),
3807 PFN_PHYS(start_pfn),
3808 (end_pfn - start_pfn) << PAGE_SHIFT);
3812 int __init add_from_early_node_map(struct range *range, int az,
3813 int nr_range, int nid)
3815 unsigned long start_pfn, end_pfn;
3818 /* need to go over early_node_map to find out good range for node */
3819 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL)
3820 nr_range = add_range(range, az, nr_range, start_pfn, end_pfn);
3825 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3826 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3828 * If an architecture guarantees that all ranges registered with
3829 * add_active_ranges() contain no holes and may be freed, this
3830 * function may be used instead of calling memory_present() manually.
3832 void __init sparse_memory_present_with_active_regions(int nid)
3834 unsigned long start_pfn, end_pfn;
3837 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
3838 memory_present(this_nid, start_pfn, end_pfn);
3842 * get_pfn_range_for_nid - Return the start and end page frames for a node
3843 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3844 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3845 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3847 * It returns the start and end page frame of a node based on information
3848 * provided by an arch calling add_active_range(). If called for a node
3849 * with no available memory, a warning is printed and the start and end
3852 void __meminit get_pfn_range_for_nid(unsigned int nid,
3853 unsigned long *start_pfn, unsigned long *end_pfn)
3855 unsigned long this_start_pfn, this_end_pfn;
3861 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
3862 *start_pfn = min(*start_pfn, this_start_pfn);
3863 *end_pfn = max(*end_pfn, this_end_pfn);
3866 if (*start_pfn == -1UL)
3871 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3872 * assumption is made that zones within a node are ordered in monotonic
3873 * increasing memory addresses so that the "highest" populated zone is used
3875 static void __init find_usable_zone_for_movable(void)
3878 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3879 if (zone_index == ZONE_MOVABLE)
3882 if (arch_zone_highest_possible_pfn[zone_index] >
3883 arch_zone_lowest_possible_pfn[zone_index])
3887 VM_BUG_ON(zone_index == -1);
3888 movable_zone = zone_index;
3892 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3893 * because it is sized independent of architecture. Unlike the other zones,
3894 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3895 * in each node depending on the size of each node and how evenly kernelcore
3896 * is distributed. This helper function adjusts the zone ranges
3897 * provided by the architecture for a given node by using the end of the
3898 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3899 * zones within a node are in order of monotonic increases memory addresses
3901 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3902 unsigned long zone_type,
3903 unsigned long node_start_pfn,
3904 unsigned long node_end_pfn,
3905 unsigned long *zone_start_pfn,
3906 unsigned long *zone_end_pfn)
3908 /* Only adjust if ZONE_MOVABLE is on this node */
3909 if (zone_movable_pfn[nid]) {
3910 /* Size ZONE_MOVABLE */
3911 if (zone_type == ZONE_MOVABLE) {
3912 *zone_start_pfn = zone_movable_pfn[nid];
3913 *zone_end_pfn = min(node_end_pfn,
3914 arch_zone_highest_possible_pfn[movable_zone]);
3916 /* Adjust for ZONE_MOVABLE starting within this range */
3917 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3918 *zone_end_pfn > zone_movable_pfn[nid]) {
3919 *zone_end_pfn = zone_movable_pfn[nid];
3921 /* Check if this whole range is within ZONE_MOVABLE */
3922 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3923 *zone_start_pfn = *zone_end_pfn;
3928 * Return the number of pages a zone spans in a node, including holes
3929 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3931 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3932 unsigned long zone_type,
3933 unsigned long *ignored)
3935 unsigned long node_start_pfn, node_end_pfn;
3936 unsigned long zone_start_pfn, zone_end_pfn;
3938 /* Get the start and end of the node and zone */
3939 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3940 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3941 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3942 adjust_zone_range_for_zone_movable(nid, zone_type,
3943 node_start_pfn, node_end_pfn,
3944 &zone_start_pfn, &zone_end_pfn);
3946 /* Check that this node has pages within the zone's required range */
3947 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3950 /* Move the zone boundaries inside the node if necessary */
3951 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3952 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3954 /* Return the spanned pages */
3955 return zone_end_pfn - zone_start_pfn;
3959 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3960 * then all holes in the requested range will be accounted for.
3962 unsigned long __meminit __absent_pages_in_range(int nid,
3963 unsigned long range_start_pfn,
3964 unsigned long range_end_pfn)
3966 unsigned long nr_absent = range_end_pfn - range_start_pfn;
3967 unsigned long start_pfn, end_pfn;
3970 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
3971 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
3972 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
3973 nr_absent -= end_pfn - start_pfn;
3979 * absent_pages_in_range - Return number of page frames in holes within a range
3980 * @start_pfn: The start PFN to start searching for holes
3981 * @end_pfn: The end PFN to stop searching for holes
3983 * It returns the number of pages frames in memory holes within a range.
3985 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3986 unsigned long end_pfn)
3988 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3991 /* Return the number of page frames in holes in a zone on a node */
3992 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3993 unsigned long zone_type,
3994 unsigned long *ignored)
3996 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
3997 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
3998 unsigned long node_start_pfn, node_end_pfn;
3999 unsigned long zone_start_pfn, zone_end_pfn;
4001 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4002 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4003 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4005 adjust_zone_range_for_zone_movable(nid, zone_type,
4006 node_start_pfn, node_end_pfn,
4007 &zone_start_pfn, &zone_end_pfn);
4008 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4011 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4012 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4013 unsigned long zone_type,
4014 unsigned long *zones_size)
4016 return zones_size[zone_type];
4019 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4020 unsigned long zone_type,
4021 unsigned long *zholes_size)
4026 return zholes_size[zone_type];
4029 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4031 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4032 unsigned long *zones_size, unsigned long *zholes_size)
4034 unsigned long realtotalpages, totalpages = 0;
4037 for (i = 0; i < MAX_NR_ZONES; i++)
4038 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4040 pgdat->node_spanned_pages = totalpages;
4042 realtotalpages = totalpages;
4043 for (i = 0; i < MAX_NR_ZONES; i++)
4045 zone_absent_pages_in_node(pgdat->node_id, i,
4047 pgdat->node_present_pages = realtotalpages;
4048 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4052 #ifndef CONFIG_SPARSEMEM
4054 * Calculate the size of the zone->blockflags rounded to an unsigned long
4055 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4056 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4057 * round what is now in bits to nearest long in bits, then return it in
4060 static unsigned long __init usemap_size(unsigned long zonesize)
4062 unsigned long usemapsize;
4064 usemapsize = roundup(zonesize, pageblock_nr_pages);
4065 usemapsize = usemapsize >> pageblock_order;
4066 usemapsize *= NR_PAGEBLOCK_BITS;
4067 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4069 return usemapsize / 8;
4072 static void __init setup_usemap(struct pglist_data *pgdat,
4073 struct zone *zone, unsigned long zonesize)
4075 unsigned long usemapsize = usemap_size(zonesize);
4076 zone->pageblock_flags = NULL;
4078 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4082 static inline void setup_usemap(struct pglist_data *pgdat,
4083 struct zone *zone, unsigned long zonesize) {}
4084 #endif /* CONFIG_SPARSEMEM */
4086 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4088 /* Return a sensible default order for the pageblock size. */
4089 static inline int pageblock_default_order(void)
4091 if (HPAGE_SHIFT > PAGE_SHIFT)
4092 return HUGETLB_PAGE_ORDER;
4097 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4098 static inline void __init set_pageblock_order(unsigned int order)
4100 /* Check that pageblock_nr_pages has not already been setup */
4101 if (pageblock_order)
4105 * Assume the largest contiguous order of interest is a huge page.
4106 * This value may be variable depending on boot parameters on IA64
4108 pageblock_order = order;
4110 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4113 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4114 * and pageblock_default_order() are unused as pageblock_order is set
4115 * at compile-time. See include/linux/pageblock-flags.h for the values of
4116 * pageblock_order based on the kernel config
4118 static inline int pageblock_default_order(unsigned int order)
4122 #define set_pageblock_order(x) do {} while (0)
4124 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4127 * Set up the zone data structures:
4128 * - mark all pages reserved
4129 * - mark all memory queues empty
4130 * - clear the memory bitmaps
4132 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4133 unsigned long *zones_size, unsigned long *zholes_size)
4136 int nid = pgdat->node_id;
4137 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4140 pgdat_resize_init(pgdat);
4141 pgdat->nr_zones = 0;
4142 init_waitqueue_head(&pgdat->kswapd_wait);
4143 pgdat->kswapd_max_order = 0;
4144 pgdat_page_cgroup_init(pgdat);
4146 for (j = 0; j < MAX_NR_ZONES; j++) {
4147 struct zone *zone = pgdat->node_zones + j;
4148 unsigned long size, realsize, memmap_pages;
4151 size = zone_spanned_pages_in_node(nid, j, zones_size);
4152 realsize = size - zone_absent_pages_in_node(nid, j,
4156 * Adjust realsize so that it accounts for how much memory
4157 * is used by this zone for memmap. This affects the watermark
4158 * and per-cpu initialisations
4161 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4162 if (realsize >= memmap_pages) {
4163 realsize -= memmap_pages;
4166 " %s zone: %lu pages used for memmap\n",
4167 zone_names[j], memmap_pages);
4170 " %s zone: %lu pages exceeds realsize %lu\n",
4171 zone_names[j], memmap_pages, realsize);
4173 /* Account for reserved pages */
4174 if (j == 0 && realsize > dma_reserve) {
4175 realsize -= dma_reserve;
4176 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4177 zone_names[0], dma_reserve);
4180 if (!is_highmem_idx(j))
4181 nr_kernel_pages += realsize;
4182 nr_all_pages += realsize;
4184 zone->spanned_pages = size;
4185 zone->present_pages = realsize;
4188 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4190 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4192 zone->name = zone_names[j];
4193 spin_lock_init(&zone->lock);
4194 spin_lock_init(&zone->lru_lock);
4195 zone_seqlock_init(zone);
4196 zone->zone_pgdat = pgdat;
4198 zone_pcp_init(zone);
4200 INIT_LIST_HEAD(&zone->lru[l].list);
4201 zone->reclaim_stat.recent_rotated[0] = 0;
4202 zone->reclaim_stat.recent_rotated[1] = 0;
4203 zone->reclaim_stat.recent_scanned[0] = 0;
4204 zone->reclaim_stat.recent_scanned[1] = 0;
4205 zap_zone_vm_stats(zone);
4210 set_pageblock_order(pageblock_default_order());
4211 setup_usemap(pgdat, zone, size);
4212 ret = init_currently_empty_zone(zone, zone_start_pfn,
4213 size, MEMMAP_EARLY);
4215 memmap_init(size, nid, j, zone_start_pfn);
4216 zone_start_pfn += size;
4220 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4222 /* Skip empty nodes */
4223 if (!pgdat->node_spanned_pages)
4226 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4227 /* ia64 gets its own node_mem_map, before this, without bootmem */
4228 if (!pgdat->node_mem_map) {
4229 unsigned long size, start, end;
4233 * The zone's endpoints aren't required to be MAX_ORDER
4234 * aligned but the node_mem_map endpoints must be in order
4235 * for the buddy allocator to function correctly.
4237 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4238 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4239 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4240 size = (end - start) * sizeof(struct page);
4241 map = alloc_remap(pgdat->node_id, size);
4243 map = alloc_bootmem_node_nopanic(pgdat, size);
4244 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4246 #ifndef CONFIG_NEED_MULTIPLE_NODES
4248 * With no DISCONTIG, the global mem_map is just set as node 0's
4250 if (pgdat == NODE_DATA(0)) {
4251 mem_map = NODE_DATA(0)->node_mem_map;
4252 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4253 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4254 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4255 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4258 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4261 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4262 unsigned long node_start_pfn, unsigned long *zholes_size)
4264 pg_data_t *pgdat = NODE_DATA(nid);
4266 pgdat->node_id = nid;
4267 pgdat->node_start_pfn = node_start_pfn;
4268 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4270 alloc_node_mem_map(pgdat);
4271 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4272 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4273 nid, (unsigned long)pgdat,
4274 (unsigned long)pgdat->node_mem_map);
4277 free_area_init_core(pgdat, zones_size, zholes_size);
4280 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4282 #if MAX_NUMNODES > 1
4284 * Figure out the number of possible node ids.
4286 static void __init setup_nr_node_ids(void)
4289 unsigned int highest = 0;
4291 for_each_node_mask(node, node_possible_map)
4293 nr_node_ids = highest + 1;
4296 static inline void setup_nr_node_ids(void)
4302 * node_map_pfn_alignment - determine the maximum internode alignment
4304 * This function should be called after node map is populated and sorted.
4305 * It calculates the maximum power of two alignment which can distinguish
4308 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4309 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4310 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4311 * shifted, 1GiB is enough and this function will indicate so.
4313 * This is used to test whether pfn -> nid mapping of the chosen memory
4314 * model has fine enough granularity to avoid incorrect mapping for the
4315 * populated node map.
4317 * Returns the determined alignment in pfn's. 0 if there is no alignment
4318 * requirement (single node).
4320 unsigned long __init node_map_pfn_alignment(void)
4322 unsigned long accl_mask = 0, last_end = 0;
4323 unsigned long start, end, mask;
4327 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4328 if (!start || last_nid < 0 || last_nid == nid) {
4335 * Start with a mask granular enough to pin-point to the
4336 * start pfn and tick off bits one-by-one until it becomes
4337 * too coarse to separate the current node from the last.
4339 mask = ~((1 << __ffs(start)) - 1);
4340 while (mask && last_end <= (start & (mask << 1)))
4343 /* accumulate all internode masks */
4347 /* convert mask to number of pages */
4348 return ~accl_mask + 1;
4351 /* Find the lowest pfn for a node */
4352 static unsigned long __init find_min_pfn_for_node(int nid)
4354 unsigned long min_pfn = ULONG_MAX;
4355 unsigned long start_pfn;
4358 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4359 min_pfn = min(min_pfn, start_pfn);
4361 if (min_pfn == ULONG_MAX) {
4363 "Could not find start_pfn for node %d\n", nid);
4371 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4373 * It returns the minimum PFN based on information provided via
4374 * add_active_range().
4376 unsigned long __init find_min_pfn_with_active_regions(void)
4378 return find_min_pfn_for_node(MAX_NUMNODES);
4382 * early_calculate_totalpages()
4383 * Sum pages in active regions for movable zone.
4384 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4386 static unsigned long __init early_calculate_totalpages(void)
4388 unsigned long totalpages = 0;
4389 unsigned long start_pfn, end_pfn;
4392 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4393 unsigned long pages = end_pfn - start_pfn;
4395 totalpages += pages;
4397 node_set_state(nid, N_HIGH_MEMORY);
4403 * Find the PFN the Movable zone begins in each node. Kernel memory
4404 * is spread evenly between nodes as long as the nodes have enough
4405 * memory. When they don't, some nodes will have more kernelcore than
4408 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4411 unsigned long usable_startpfn;
4412 unsigned long kernelcore_node, kernelcore_remaining;
4413 /* save the state before borrow the nodemask */
4414 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4415 unsigned long totalpages = early_calculate_totalpages();
4416 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4419 * If movablecore was specified, calculate what size of
4420 * kernelcore that corresponds so that memory usable for
4421 * any allocation type is evenly spread. If both kernelcore
4422 * and movablecore are specified, then the value of kernelcore
4423 * will be used for required_kernelcore if it's greater than
4424 * what movablecore would have allowed.
4426 if (required_movablecore) {
4427 unsigned long corepages;
4430 * Round-up so that ZONE_MOVABLE is at least as large as what
4431 * was requested by the user
4433 required_movablecore =
4434 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4435 corepages = totalpages - required_movablecore;
4437 required_kernelcore = max(required_kernelcore, corepages);
4440 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4441 if (!required_kernelcore)
4444 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4445 find_usable_zone_for_movable();
4446 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4449 /* Spread kernelcore memory as evenly as possible throughout nodes */
4450 kernelcore_node = required_kernelcore / usable_nodes;
4451 for_each_node_state(nid, N_HIGH_MEMORY) {
4452 unsigned long start_pfn, end_pfn;
4455 * Recalculate kernelcore_node if the division per node
4456 * now exceeds what is necessary to satisfy the requested
4457 * amount of memory for the kernel
4459 if (required_kernelcore < kernelcore_node)
4460 kernelcore_node = required_kernelcore / usable_nodes;
4463 * As the map is walked, we track how much memory is usable
4464 * by the kernel using kernelcore_remaining. When it is
4465 * 0, the rest of the node is usable by ZONE_MOVABLE
4467 kernelcore_remaining = kernelcore_node;
4469 /* Go through each range of PFNs within this node */
4470 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4471 unsigned long size_pages;
4473 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4474 if (start_pfn >= end_pfn)
4477 /* Account for what is only usable for kernelcore */
4478 if (start_pfn < usable_startpfn) {
4479 unsigned long kernel_pages;
4480 kernel_pages = min(end_pfn, usable_startpfn)
4483 kernelcore_remaining -= min(kernel_pages,
4484 kernelcore_remaining);
4485 required_kernelcore -= min(kernel_pages,
4486 required_kernelcore);
4488 /* Continue if range is now fully accounted */
4489 if (end_pfn <= usable_startpfn) {
4492 * Push zone_movable_pfn to the end so
4493 * that if we have to rebalance
4494 * kernelcore across nodes, we will
4495 * not double account here
4497 zone_movable_pfn[nid] = end_pfn;
4500 start_pfn = usable_startpfn;
4504 * The usable PFN range for ZONE_MOVABLE is from
4505 * start_pfn->end_pfn. Calculate size_pages as the
4506 * number of pages used as kernelcore
4508 size_pages = end_pfn - start_pfn;
4509 if (size_pages > kernelcore_remaining)
4510 size_pages = kernelcore_remaining;
4511 zone_movable_pfn[nid] = start_pfn + size_pages;
4514 * Some kernelcore has been met, update counts and
4515 * break if the kernelcore for this node has been
4518 required_kernelcore -= min(required_kernelcore,
4520 kernelcore_remaining -= size_pages;
4521 if (!kernelcore_remaining)
4527 * If there is still required_kernelcore, we do another pass with one
4528 * less node in the count. This will push zone_movable_pfn[nid] further
4529 * along on the nodes that still have memory until kernelcore is
4533 if (usable_nodes && required_kernelcore > usable_nodes)
4536 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4537 for (nid = 0; nid < MAX_NUMNODES; nid++)
4538 zone_movable_pfn[nid] =
4539 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4542 /* restore the node_state */
4543 node_states[N_HIGH_MEMORY] = saved_node_state;
4546 /* Any regular memory on that node ? */
4547 static void check_for_regular_memory(pg_data_t *pgdat)
4549 #ifdef CONFIG_HIGHMEM
4550 enum zone_type zone_type;
4552 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4553 struct zone *zone = &pgdat->node_zones[zone_type];
4554 if (zone->present_pages)
4555 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4561 * free_area_init_nodes - Initialise all pg_data_t and zone data
4562 * @max_zone_pfn: an array of max PFNs for each zone
4564 * This will call free_area_init_node() for each active node in the system.
4565 * Using the page ranges provided by add_active_range(), the size of each
4566 * zone in each node and their holes is calculated. If the maximum PFN
4567 * between two adjacent zones match, it is assumed that the zone is empty.
4568 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4569 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4570 * starts where the previous one ended. For example, ZONE_DMA32 starts
4571 * at arch_max_dma_pfn.
4573 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4575 unsigned long start_pfn, end_pfn;
4578 /* Record where the zone boundaries are */
4579 memset(arch_zone_lowest_possible_pfn, 0,
4580 sizeof(arch_zone_lowest_possible_pfn));
4581 memset(arch_zone_highest_possible_pfn, 0,
4582 sizeof(arch_zone_highest_possible_pfn));
4583 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4584 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4585 for (i = 1; i < MAX_NR_ZONES; i++) {
4586 if (i == ZONE_MOVABLE)
4588 arch_zone_lowest_possible_pfn[i] =
4589 arch_zone_highest_possible_pfn[i-1];
4590 arch_zone_highest_possible_pfn[i] =
4591 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4593 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4594 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4596 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4597 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4598 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4600 /* Print out the zone ranges */
4601 printk("Zone PFN ranges:\n");
4602 for (i = 0; i < MAX_NR_ZONES; i++) {
4603 if (i == ZONE_MOVABLE)
4605 printk(" %-8s ", zone_names[i]);
4606 if (arch_zone_lowest_possible_pfn[i] ==
4607 arch_zone_highest_possible_pfn[i])
4610 printk("%0#10lx -> %0#10lx\n",
4611 arch_zone_lowest_possible_pfn[i],
4612 arch_zone_highest_possible_pfn[i]);
4615 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4616 printk("Movable zone start PFN for each node\n");
4617 for (i = 0; i < MAX_NUMNODES; i++) {
4618 if (zone_movable_pfn[i])
4619 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4622 /* Print out the early_node_map[] */
4623 printk("Early memory PFN ranges\n");
4624 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4625 printk(" %3d: %0#10lx -> %0#10lx\n", nid, start_pfn, end_pfn);
4627 /* Initialise every node */
4628 mminit_verify_pageflags_layout();
4629 setup_nr_node_ids();
4630 for_each_online_node(nid) {
4631 pg_data_t *pgdat = NODE_DATA(nid);
4632 free_area_init_node(nid, NULL,
4633 find_min_pfn_for_node(nid), NULL);
4635 /* Any memory on that node */
4636 if (pgdat->node_present_pages)
4637 node_set_state(nid, N_HIGH_MEMORY);
4638 check_for_regular_memory(pgdat);
4642 static int __init cmdline_parse_core(char *p, unsigned long *core)
4644 unsigned long long coremem;
4648 coremem = memparse(p, &p);
4649 *core = coremem >> PAGE_SHIFT;
4651 /* Paranoid check that UL is enough for the coremem value */
4652 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4658 * kernelcore=size sets the amount of memory for use for allocations that
4659 * cannot be reclaimed or migrated.
4661 static int __init cmdline_parse_kernelcore(char *p)
4663 return cmdline_parse_core(p, &required_kernelcore);
4667 * movablecore=size sets the amount of memory for use for allocations that
4668 * can be reclaimed or migrated.
4670 static int __init cmdline_parse_movablecore(char *p)
4672 return cmdline_parse_core(p, &required_movablecore);
4675 early_param("kernelcore", cmdline_parse_kernelcore);
4676 early_param("movablecore", cmdline_parse_movablecore);
4678 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4681 * set_dma_reserve - set the specified number of pages reserved in the first zone
4682 * @new_dma_reserve: The number of pages to mark reserved
4684 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4685 * In the DMA zone, a significant percentage may be consumed by kernel image
4686 * and other unfreeable allocations which can skew the watermarks badly. This
4687 * function may optionally be used to account for unfreeable pages in the
4688 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4689 * smaller per-cpu batchsize.
4691 void __init set_dma_reserve(unsigned long new_dma_reserve)
4693 dma_reserve = new_dma_reserve;
4696 void __init free_area_init(unsigned long *zones_size)
4698 free_area_init_node(0, zones_size,
4699 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4702 static int page_alloc_cpu_notify(struct notifier_block *self,
4703 unsigned long action, void *hcpu)
4705 int cpu = (unsigned long)hcpu;
4707 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4711 * Spill the event counters of the dead processor
4712 * into the current processors event counters.
4713 * This artificially elevates the count of the current
4716 vm_events_fold_cpu(cpu);
4719 * Zero the differential counters of the dead processor
4720 * so that the vm statistics are consistent.
4722 * This is only okay since the processor is dead and cannot
4723 * race with what we are doing.
4725 refresh_cpu_vm_stats(cpu);
4730 void __init page_alloc_init(void)
4732 hotcpu_notifier(page_alloc_cpu_notify, 0);
4736 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4737 * or min_free_kbytes changes.
4739 static void calculate_totalreserve_pages(void)
4741 struct pglist_data *pgdat;
4742 unsigned long reserve_pages = 0;
4743 enum zone_type i, j;
4745 for_each_online_pgdat(pgdat) {
4746 for (i = 0; i < MAX_NR_ZONES; i++) {
4747 struct zone *zone = pgdat->node_zones + i;
4748 unsigned long max = 0;
4750 /* Find valid and maximum lowmem_reserve in the zone */
4751 for (j = i; j < MAX_NR_ZONES; j++) {
4752 if (zone->lowmem_reserve[j] > max)
4753 max = zone->lowmem_reserve[j];
4756 /* we treat the high watermark as reserved pages. */
4757 max += high_wmark_pages(zone);
4759 if (max > zone->present_pages)
4760 max = zone->present_pages;
4761 reserve_pages += max;
4764 totalreserve_pages = reserve_pages;
4768 * setup_per_zone_lowmem_reserve - called whenever
4769 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4770 * has a correct pages reserved value, so an adequate number of
4771 * pages are left in the zone after a successful __alloc_pages().
4773 static void setup_per_zone_lowmem_reserve(void)
4775 struct pglist_data *pgdat;
4776 enum zone_type j, idx;
4778 for_each_online_pgdat(pgdat) {
4779 for (j = 0; j < MAX_NR_ZONES; j++) {
4780 struct zone *zone = pgdat->node_zones + j;
4781 unsigned long present_pages = zone->present_pages;
4783 zone->lowmem_reserve[j] = 0;
4787 struct zone *lower_zone;
4791 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4792 sysctl_lowmem_reserve_ratio[idx] = 1;
4794 lower_zone = pgdat->node_zones + idx;
4795 lower_zone->lowmem_reserve[j] = present_pages /
4796 sysctl_lowmem_reserve_ratio[idx];
4797 present_pages += lower_zone->present_pages;
4802 /* update totalreserve_pages */
4803 calculate_totalreserve_pages();
4807 * setup_per_zone_wmarks - called when min_free_kbytes changes
4808 * or when memory is hot-{added|removed}
4810 * Ensures that the watermark[min,low,high] values for each zone are set
4811 * correctly with respect to min_free_kbytes.
4813 void setup_per_zone_wmarks(void)
4815 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4816 unsigned long lowmem_pages = 0;
4818 unsigned long flags;
4820 /* Calculate total number of !ZONE_HIGHMEM pages */
4821 for_each_zone(zone) {
4822 if (!is_highmem(zone))
4823 lowmem_pages += zone->present_pages;
4826 for_each_zone(zone) {
4829 spin_lock_irqsave(&zone->lock, flags);
4830 tmp = (u64)pages_min * zone->present_pages;
4831 do_div(tmp, lowmem_pages);
4832 if (is_highmem(zone)) {
4834 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4835 * need highmem pages, so cap pages_min to a small
4838 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4839 * deltas controls asynch page reclaim, and so should
4840 * not be capped for highmem.
4844 min_pages = zone->present_pages / 1024;
4845 if (min_pages < SWAP_CLUSTER_MAX)
4846 min_pages = SWAP_CLUSTER_MAX;
4847 if (min_pages > 128)
4849 zone->watermark[WMARK_MIN] = min_pages;
4852 * If it's a lowmem zone, reserve a number of pages
4853 * proportionate to the zone's size.
4855 zone->watermark[WMARK_MIN] = tmp;
4858 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4859 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4860 setup_zone_migrate_reserve(zone);
4861 spin_unlock_irqrestore(&zone->lock, flags);
4864 /* update totalreserve_pages */
4865 calculate_totalreserve_pages();
4869 * The inactive anon list should be small enough that the VM never has to
4870 * do too much work, but large enough that each inactive page has a chance
4871 * to be referenced again before it is swapped out.
4873 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4874 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4875 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4876 * the anonymous pages are kept on the inactive list.
4879 * memory ratio inactive anon
4880 * -------------------------------------
4889 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
4891 unsigned int gb, ratio;
4893 /* Zone size in gigabytes */
4894 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4896 ratio = int_sqrt(10 * gb);
4900 zone->inactive_ratio = ratio;
4903 static void __meminit setup_per_zone_inactive_ratio(void)
4908 calculate_zone_inactive_ratio(zone);
4912 * Initialise min_free_kbytes.
4914 * For small machines we want it small (128k min). For large machines
4915 * we want it large (64MB max). But it is not linear, because network
4916 * bandwidth does not increase linearly with machine size. We use
4918 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4919 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4935 int __meminit init_per_zone_wmark_min(void)
4937 unsigned long lowmem_kbytes;
4939 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4941 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4942 if (min_free_kbytes < 128)
4943 min_free_kbytes = 128;
4944 if (min_free_kbytes > 65536)
4945 min_free_kbytes = 65536;
4946 setup_per_zone_wmarks();
4947 refresh_zone_stat_thresholds();
4948 setup_per_zone_lowmem_reserve();
4949 setup_per_zone_inactive_ratio();
4952 module_init(init_per_zone_wmark_min)
4955 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4956 * that we can call two helper functions whenever min_free_kbytes
4959 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4960 void __user *buffer, size_t *length, loff_t *ppos)
4962 proc_dointvec(table, write, buffer, length, ppos);
4964 setup_per_zone_wmarks();
4969 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4970 void __user *buffer, size_t *length, loff_t *ppos)
4975 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4980 zone->min_unmapped_pages = (zone->present_pages *
4981 sysctl_min_unmapped_ratio) / 100;
4985 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4986 void __user *buffer, size_t *length, loff_t *ppos)
4991 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4996 zone->min_slab_pages = (zone->present_pages *
4997 sysctl_min_slab_ratio) / 100;
5003 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5004 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5005 * whenever sysctl_lowmem_reserve_ratio changes.
5007 * The reserve ratio obviously has absolutely no relation with the
5008 * minimum watermarks. The lowmem reserve ratio can only make sense
5009 * if in function of the boot time zone sizes.
5011 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5012 void __user *buffer, size_t *length, loff_t *ppos)
5014 proc_dointvec_minmax(table, write, buffer, length, ppos);
5015 setup_per_zone_lowmem_reserve();
5020 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5021 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5022 * can have before it gets flushed back to buddy allocator.
5025 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5026 void __user *buffer, size_t *length, loff_t *ppos)
5032 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5033 if (!write || (ret == -EINVAL))
5035 for_each_populated_zone(zone) {
5036 for_each_possible_cpu(cpu) {
5038 high = zone->present_pages / percpu_pagelist_fraction;
5039 setup_pagelist_highmark(
5040 per_cpu_ptr(zone->pageset, cpu), high);
5046 int hashdist = HASHDIST_DEFAULT;
5049 static int __init set_hashdist(char *str)
5053 hashdist = simple_strtoul(str, &str, 0);
5056 __setup("hashdist=", set_hashdist);
5060 * allocate a large system hash table from bootmem
5061 * - it is assumed that the hash table must contain an exact power-of-2
5062 * quantity of entries
5063 * - limit is the number of hash buckets, not the total allocation size
5065 void *__init alloc_large_system_hash(const char *tablename,
5066 unsigned long bucketsize,
5067 unsigned long numentries,
5070 unsigned int *_hash_shift,
5071 unsigned int *_hash_mask,
5072 unsigned long limit)
5074 unsigned long long max = limit;
5075 unsigned long log2qty, size;
5078 /* allow the kernel cmdline to have a say */
5080 /* round applicable memory size up to nearest megabyte */
5081 numentries = nr_kernel_pages;
5082 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5083 numentries >>= 20 - PAGE_SHIFT;
5084 numentries <<= 20 - PAGE_SHIFT;
5086 /* limit to 1 bucket per 2^scale bytes of low memory */
5087 if (scale > PAGE_SHIFT)
5088 numentries >>= (scale - PAGE_SHIFT);
5090 numentries <<= (PAGE_SHIFT - scale);
5092 /* Make sure we've got at least a 0-order allocation.. */
5093 if (unlikely(flags & HASH_SMALL)) {
5094 /* Makes no sense without HASH_EARLY */
5095 WARN_ON(!(flags & HASH_EARLY));
5096 if (!(numentries >> *_hash_shift)) {
5097 numentries = 1UL << *_hash_shift;
5098 BUG_ON(!numentries);
5100 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5101 numentries = PAGE_SIZE / bucketsize;
5103 numentries = roundup_pow_of_two(numentries);
5105 /* limit allocation size to 1/16 total memory by default */
5107 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5108 do_div(max, bucketsize);
5111 if (numentries > max)
5114 log2qty = ilog2(numentries);
5117 size = bucketsize << log2qty;
5118 if (flags & HASH_EARLY)
5119 table = alloc_bootmem_nopanic(size);
5121 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5124 * If bucketsize is not a power-of-two, we may free
5125 * some pages at the end of hash table which
5126 * alloc_pages_exact() automatically does
5128 if (get_order(size) < MAX_ORDER) {
5129 table = alloc_pages_exact(size, GFP_ATOMIC);
5130 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5133 } while (!table && size > PAGE_SIZE && --log2qty);
5136 panic("Failed to allocate %s hash table\n", tablename);
5138 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5141 ilog2(size) - PAGE_SHIFT,
5145 *_hash_shift = log2qty;
5147 *_hash_mask = (1 << log2qty) - 1;
5152 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5153 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5156 #ifdef CONFIG_SPARSEMEM
5157 return __pfn_to_section(pfn)->pageblock_flags;
5159 return zone->pageblock_flags;
5160 #endif /* CONFIG_SPARSEMEM */
5163 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5165 #ifdef CONFIG_SPARSEMEM
5166 pfn &= (PAGES_PER_SECTION-1);
5167 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5169 pfn = pfn - zone->zone_start_pfn;
5170 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5171 #endif /* CONFIG_SPARSEMEM */
5175 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5176 * @page: The page within the block of interest
5177 * @start_bitidx: The first bit of interest to retrieve
5178 * @end_bitidx: The last bit of interest
5179 * returns pageblock_bits flags
5181 unsigned long get_pageblock_flags_group(struct page *page,
5182 int start_bitidx, int end_bitidx)
5185 unsigned long *bitmap;
5186 unsigned long pfn, bitidx;
5187 unsigned long flags = 0;
5188 unsigned long value = 1;
5190 zone = page_zone(page);
5191 pfn = page_to_pfn(page);
5192 bitmap = get_pageblock_bitmap(zone, pfn);
5193 bitidx = pfn_to_bitidx(zone, pfn);
5195 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5196 if (test_bit(bitidx + start_bitidx, bitmap))
5203 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5204 * @page: The page within the block of interest
5205 * @start_bitidx: The first bit of interest
5206 * @end_bitidx: The last bit of interest
5207 * @flags: The flags to set
5209 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5210 int start_bitidx, int end_bitidx)
5213 unsigned long *bitmap;
5214 unsigned long pfn, bitidx;
5215 unsigned long value = 1;
5217 zone = page_zone(page);
5218 pfn = page_to_pfn(page);
5219 bitmap = get_pageblock_bitmap(zone, pfn);
5220 bitidx = pfn_to_bitidx(zone, pfn);
5221 VM_BUG_ON(pfn < zone->zone_start_pfn);
5222 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5224 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5226 __set_bit(bitidx + start_bitidx, bitmap);
5228 __clear_bit(bitidx + start_bitidx, bitmap);
5232 * This is designed as sub function...plz see page_isolation.c also.
5233 * set/clear page block's type to be ISOLATE.
5234 * page allocater never alloc memory from ISOLATE block.
5238 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5240 unsigned long pfn, iter, found;
5242 * For avoiding noise data, lru_add_drain_all() should be called
5243 * If ZONE_MOVABLE, the zone never contains immobile pages
5245 if (zone_idx(zone) == ZONE_MOVABLE)
5248 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5251 pfn = page_to_pfn(page);
5252 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5253 unsigned long check = pfn + iter;
5255 if (!pfn_valid_within(check))
5258 page = pfn_to_page(check);
5259 if (!page_count(page)) {
5260 if (PageBuddy(page))
5261 iter += (1 << page_order(page)) - 1;
5267 * If there are RECLAIMABLE pages, we need to check it.
5268 * But now, memory offline itself doesn't call shrink_slab()
5269 * and it still to be fixed.
5272 * If the page is not RAM, page_count()should be 0.
5273 * we don't need more check. This is an _used_ not-movable page.
5275 * The problematic thing here is PG_reserved pages. PG_reserved
5276 * is set to both of a memory hole page and a _used_ kernel
5285 bool is_pageblock_removable_nolock(struct page *page)
5287 struct zone *zone = page_zone(page);
5288 return __count_immobile_pages(zone, page, 0);
5291 int set_migratetype_isolate(struct page *page)
5294 unsigned long flags, pfn;
5295 struct memory_isolate_notify arg;
5299 zone = page_zone(page);
5301 spin_lock_irqsave(&zone->lock, flags);
5303 pfn = page_to_pfn(page);
5304 arg.start_pfn = pfn;
5305 arg.nr_pages = pageblock_nr_pages;
5306 arg.pages_found = 0;
5309 * It may be possible to isolate a pageblock even if the
5310 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5311 * notifier chain is used by balloon drivers to return the
5312 * number of pages in a range that are held by the balloon
5313 * driver to shrink memory. If all the pages are accounted for
5314 * by balloons, are free, or on the LRU, isolation can continue.
5315 * Later, for example, when memory hotplug notifier runs, these
5316 * pages reported as "can be isolated" should be isolated(freed)
5317 * by the balloon driver through the memory notifier chain.
5319 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5320 notifier_ret = notifier_to_errno(notifier_ret);
5324 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5325 * We just check MOVABLE pages.
5327 if (__count_immobile_pages(zone, page, arg.pages_found))
5331 * immobile means "not-on-lru" paes. If immobile is larger than
5332 * removable-by-driver pages reported by notifier, we'll fail.
5337 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5338 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5341 spin_unlock_irqrestore(&zone->lock, flags);
5347 void unset_migratetype_isolate(struct page *page)
5350 unsigned long flags;
5351 zone = page_zone(page);
5352 spin_lock_irqsave(&zone->lock, flags);
5353 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5355 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5356 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5358 spin_unlock_irqrestore(&zone->lock, flags);
5361 #ifdef CONFIG_MEMORY_HOTREMOVE
5363 * All pages in the range must be isolated before calling this.
5366 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5372 unsigned long flags;
5373 /* find the first valid pfn */
5374 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5379 zone = page_zone(pfn_to_page(pfn));
5380 spin_lock_irqsave(&zone->lock, flags);
5382 while (pfn < end_pfn) {
5383 if (!pfn_valid(pfn)) {
5387 page = pfn_to_page(pfn);
5388 BUG_ON(page_count(page));
5389 BUG_ON(!PageBuddy(page));
5390 order = page_order(page);
5391 #ifdef CONFIG_DEBUG_VM
5392 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5393 pfn, 1 << order, end_pfn);
5395 list_del(&page->lru);
5396 rmv_page_order(page);
5397 zone->free_area[order].nr_free--;
5398 __mod_zone_page_state(zone, NR_FREE_PAGES,
5400 for (i = 0; i < (1 << order); i++)
5401 SetPageReserved((page+i));
5402 pfn += (1 << order);
5404 spin_unlock_irqrestore(&zone->lock, flags);
5408 #ifdef CONFIG_MEMORY_FAILURE
5409 bool is_free_buddy_page(struct page *page)
5411 struct zone *zone = page_zone(page);
5412 unsigned long pfn = page_to_pfn(page);
5413 unsigned long flags;
5416 spin_lock_irqsave(&zone->lock, flags);
5417 for (order = 0; order < MAX_ORDER; order++) {
5418 struct page *page_head = page - (pfn & ((1 << order) - 1));
5420 if (PageBuddy(page_head) && page_order(page_head) >= order)
5423 spin_unlock_irqrestore(&zone->lock, flags);
5425 return order < MAX_ORDER;
5429 static struct trace_print_flags pageflag_names[] = {
5430 {1UL << PG_locked, "locked" },
5431 {1UL << PG_error, "error" },
5432 {1UL << PG_referenced, "referenced" },
5433 {1UL << PG_uptodate, "uptodate" },
5434 {1UL << PG_dirty, "dirty" },
5435 {1UL << PG_lru, "lru" },
5436 {1UL << PG_active, "active" },
5437 {1UL << PG_slab, "slab" },
5438 {1UL << PG_owner_priv_1, "owner_priv_1" },
5439 {1UL << PG_arch_1, "arch_1" },
5440 {1UL << PG_reserved, "reserved" },
5441 {1UL << PG_private, "private" },
5442 {1UL << PG_private_2, "private_2" },
5443 {1UL << PG_writeback, "writeback" },
5444 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5445 {1UL << PG_head, "head" },
5446 {1UL << PG_tail, "tail" },
5448 {1UL << PG_compound, "compound" },
5450 {1UL << PG_swapcache, "swapcache" },
5451 {1UL << PG_mappedtodisk, "mappedtodisk" },
5452 {1UL << PG_reclaim, "reclaim" },
5453 {1UL << PG_swapbacked, "swapbacked" },
5454 {1UL << PG_unevictable, "unevictable" },
5456 {1UL << PG_mlocked, "mlocked" },
5458 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5459 {1UL << PG_uncached, "uncached" },
5461 #ifdef CONFIG_MEMORY_FAILURE
5462 {1UL << PG_hwpoison, "hwpoison" },
5467 static void dump_page_flags(unsigned long flags)
5469 const char *delim = "";
5473 printk(KERN_ALERT "page flags: %#lx(", flags);
5475 /* remove zone id */
5476 flags &= (1UL << NR_PAGEFLAGS) - 1;
5478 for (i = 0; pageflag_names[i].name && flags; i++) {
5480 mask = pageflag_names[i].mask;
5481 if ((flags & mask) != mask)
5485 printk("%s%s", delim, pageflag_names[i].name);
5489 /* check for left over flags */
5491 printk("%s%#lx", delim, flags);
5496 void dump_page(struct page *page)
5499 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5500 page, atomic_read(&page->_count), page_mapcount(page),
5501 page->mapping, page->index);
5502 dump_page_flags(page->flags);
5503 mem_cgroup_print_bad_page(page);