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>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
63 #include <asm/tlbflush.h>
64 #include <asm/div64.h>
67 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
68 DEFINE_PER_CPU(int, numa_node);
69 EXPORT_PER_CPU_SYMBOL(numa_node);
72 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
74 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
75 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
76 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
77 * defined in <linux/topology.h>.
79 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
80 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
84 * Array of node states.
86 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
87 [N_POSSIBLE] = NODE_MASK_ALL,
88 [N_ONLINE] = { { [0] = 1UL } },
90 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
92 [N_HIGH_MEMORY] = { { [0] = 1UL } },
94 [N_CPU] = { { [0] = 1UL } },
97 EXPORT_SYMBOL(node_states);
99 unsigned long totalram_pages __read_mostly;
100 unsigned long totalreserve_pages __read_mostly;
102 * When calculating the number of globally allowed dirty pages, there
103 * is a certain number of per-zone reserves that should not be
104 * considered dirtyable memory. This is the sum of those reserves
105 * over all existing zones that contribute dirtyable memory.
107 unsigned long dirty_balance_reserve __read_mostly;
109 int percpu_pagelist_fraction;
110 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
112 #ifdef CONFIG_PM_SLEEP
114 * The following functions are used by the suspend/hibernate code to temporarily
115 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
116 * while devices are suspended. To avoid races with the suspend/hibernate code,
117 * they should always be called with pm_mutex held (gfp_allowed_mask also should
118 * only be modified with pm_mutex held, unless the suspend/hibernate code is
119 * guaranteed not to run in parallel with that modification).
122 static gfp_t saved_gfp_mask;
124 void pm_restore_gfp_mask(void)
126 WARN_ON(!mutex_is_locked(&pm_mutex));
127 if (saved_gfp_mask) {
128 gfp_allowed_mask = saved_gfp_mask;
133 void pm_restrict_gfp_mask(void)
135 WARN_ON(!mutex_is_locked(&pm_mutex));
136 WARN_ON(saved_gfp_mask);
137 saved_gfp_mask = gfp_allowed_mask;
138 gfp_allowed_mask &= ~GFP_IOFS;
141 bool pm_suspended_storage(void)
143 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
147 #endif /* CONFIG_PM_SLEEP */
149 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
150 int pageblock_order __read_mostly;
153 static void __free_pages_ok(struct page *page, unsigned int order);
156 * results with 256, 32 in the lowmem_reserve sysctl:
157 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
158 * 1G machine -> (16M dma, 784M normal, 224M high)
159 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
160 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
161 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
163 * TBD: should special case ZONE_DMA32 machines here - in those we normally
164 * don't need any ZONE_NORMAL reservation
166 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
167 #ifdef CONFIG_ZONE_DMA
170 #ifdef CONFIG_ZONE_DMA32
173 #ifdef CONFIG_HIGHMEM
179 EXPORT_SYMBOL(totalram_pages);
181 static char * const zone_names[MAX_NR_ZONES] = {
182 #ifdef CONFIG_ZONE_DMA
185 #ifdef CONFIG_ZONE_DMA32
189 #ifdef CONFIG_HIGHMEM
195 int min_free_kbytes = 1024;
197 static unsigned long __meminitdata nr_kernel_pages;
198 static unsigned long __meminitdata nr_all_pages;
199 static unsigned long __meminitdata dma_reserve;
201 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
203 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
204 * ranges of memory (RAM) that may be registered with add_active_range().
205 * Ranges passed to add_active_range() will be merged if possible
206 * so the number of times add_active_range() can be called is
207 * related to the number of nodes and the number of holes
209 #ifdef CONFIG_MAX_ACTIVE_REGIONS
210 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
211 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
213 #if MAX_NUMNODES >= 32
214 /* If there can be many nodes, allow up to 50 holes per node */
215 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
217 /* By default, allow up to 256 distinct regions */
218 #define MAX_ACTIVE_REGIONS 256
222 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
223 static int __meminitdata nr_nodemap_entries;
224 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
225 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
226 static unsigned long __initdata required_kernelcore;
227 static unsigned long __initdata required_movablecore;
228 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
230 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
232 EXPORT_SYMBOL(movable_zone);
233 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
236 int nr_node_ids __read_mostly = MAX_NUMNODES;
237 int nr_online_nodes __read_mostly = 1;
238 EXPORT_SYMBOL(nr_node_ids);
239 EXPORT_SYMBOL(nr_online_nodes);
242 int page_group_by_mobility_disabled __read_mostly;
244 static void set_pageblock_migratetype(struct page *page, int migratetype)
247 if (unlikely(page_group_by_mobility_disabled))
248 migratetype = MIGRATE_UNMOVABLE;
250 set_pageblock_flags_group(page, (unsigned long)migratetype,
251 PB_migrate, PB_migrate_end);
254 bool oom_killer_disabled __read_mostly;
256 #ifdef CONFIG_DEBUG_VM
257 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
261 unsigned long pfn = page_to_pfn(page);
264 seq = zone_span_seqbegin(zone);
265 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
267 else if (pfn < zone->zone_start_pfn)
269 } while (zone_span_seqretry(zone, seq));
274 static int page_is_consistent(struct zone *zone, struct page *page)
276 if (!pfn_valid_within(page_to_pfn(page)))
278 if (zone != page_zone(page))
284 * Temporary debugging check for pages not lying within a given zone.
286 static int bad_range(struct zone *zone, struct page *page)
288 if (page_outside_zone_boundaries(zone, page))
290 if (!page_is_consistent(zone, page))
296 static inline int bad_range(struct zone *zone, struct page *page)
302 static void bad_page(struct page *page)
304 static unsigned long resume;
305 static unsigned long nr_shown;
306 static unsigned long nr_unshown;
308 /* Don't complain about poisoned pages */
309 if (PageHWPoison(page)) {
310 reset_page_mapcount(page); /* remove PageBuddy */
315 * Allow a burst of 60 reports, then keep quiet for that minute;
316 * or allow a steady drip of one report per second.
318 if (nr_shown == 60) {
319 if (time_before(jiffies, resume)) {
325 "BUG: Bad page state: %lu messages suppressed\n",
332 resume = jiffies + 60 * HZ;
334 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
335 current->comm, page_to_pfn(page));
341 /* Leave bad fields for debug, except PageBuddy could make trouble */
342 reset_page_mapcount(page); /* remove PageBuddy */
343 add_taint(TAINT_BAD_PAGE);
347 * Higher-order pages are called "compound pages". They are structured thusly:
349 * The first PAGE_SIZE page is called the "head page".
351 * The remaining PAGE_SIZE pages are called "tail pages".
353 * All pages have PG_compound set. All pages have their ->private pointing at
354 * the head page (even the head page has this).
356 * The first tail page's ->lru.next holds the address of the compound page's
357 * put_page() function. Its ->lru.prev holds the order of allocation.
358 * This usage means that zero-order pages may not be compound.
361 static void free_compound_page(struct page *page)
363 __free_pages_ok(page, compound_order(page));
366 void prep_compound_page(struct page *page, unsigned long order)
369 int nr_pages = 1 << order;
371 set_compound_page_dtor(page, free_compound_page);
372 set_compound_order(page, order);
374 for (i = 1; i < nr_pages; i++) {
375 struct page *p = page + i;
377 set_page_count(p, 0);
378 p->first_page = page;
382 /* update __split_huge_page_refcount if you change this function */
383 static int destroy_compound_page(struct page *page, unsigned long order)
386 int nr_pages = 1 << order;
389 if (unlikely(compound_order(page) != order) ||
390 unlikely(!PageHead(page))) {
395 __ClearPageHead(page);
397 for (i = 1; i < nr_pages; i++) {
398 struct page *p = page + i;
400 if (unlikely(!PageTail(p) || (p->first_page != page))) {
410 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
415 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
416 * and __GFP_HIGHMEM from hard or soft interrupt context.
418 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
419 for (i = 0; i < (1 << order); i++)
420 clear_highpage(page + i);
423 #ifdef CONFIG_DEBUG_PAGEALLOC
424 unsigned int _debug_guardpage_minorder;
426 static int __init debug_guardpage_minorder_setup(char *buf)
430 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
431 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
434 _debug_guardpage_minorder = res;
435 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
438 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
440 static inline void set_page_guard_flag(struct page *page)
442 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
445 static inline void clear_page_guard_flag(struct page *page)
447 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
450 static inline void set_page_guard_flag(struct page *page) { }
451 static inline void clear_page_guard_flag(struct page *page) { }
454 static inline void set_page_order(struct page *page, int order)
456 set_page_private(page, order);
457 __SetPageBuddy(page);
460 static inline void rmv_page_order(struct page *page)
462 __ClearPageBuddy(page);
463 set_page_private(page, 0);
467 * Locate the struct page for both the matching buddy in our
468 * pair (buddy1) and the combined O(n+1) page they form (page).
470 * 1) Any buddy B1 will have an order O twin B2 which satisfies
471 * the following equation:
473 * For example, if the starting buddy (buddy2) is #8 its order
475 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
477 * 2) Any buddy B will have an order O+1 parent P which
478 * satisfies the following equation:
481 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
483 static inline unsigned long
484 __find_buddy_index(unsigned long page_idx, unsigned int order)
486 return page_idx ^ (1 << order);
490 * This function checks whether a page is free && is the buddy
491 * we can do coalesce a page and its buddy if
492 * (a) the buddy is not in a hole &&
493 * (b) the buddy is in the buddy system &&
494 * (c) a page and its buddy have the same order &&
495 * (d) a page and its buddy are in the same zone.
497 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
498 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
500 * For recording page's order, we use page_private(page).
502 static inline int page_is_buddy(struct page *page, struct page *buddy,
505 if (!pfn_valid_within(page_to_pfn(buddy)))
508 if (page_zone_id(page) != page_zone_id(buddy))
511 if (page_is_guard(buddy) && page_order(buddy) == order) {
512 VM_BUG_ON(page_count(buddy) != 0);
516 if (PageBuddy(buddy) && page_order(buddy) == order) {
517 VM_BUG_ON(page_count(buddy) != 0);
524 * Freeing function for a buddy system allocator.
526 * The concept of a buddy system is to maintain direct-mapped table
527 * (containing bit values) for memory blocks of various "orders".
528 * The bottom level table contains the map for the smallest allocatable
529 * units of memory (here, pages), and each level above it describes
530 * pairs of units from the levels below, hence, "buddies".
531 * At a high level, all that happens here is marking the table entry
532 * at the bottom level available, and propagating the changes upward
533 * as necessary, plus some accounting needed to play nicely with other
534 * parts of the VM system.
535 * At each level, we keep a list of pages, which are heads of continuous
536 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
537 * order is recorded in page_private(page) field.
538 * So when we are allocating or freeing one, we can derive the state of the
539 * other. That is, if we allocate a small block, and both were
540 * free, the remainder of the region must be split into blocks.
541 * If a block is freed, and its buddy is also free, then this
542 * triggers coalescing into a block of larger size.
547 static inline void __free_one_page(struct page *page,
548 struct zone *zone, unsigned int order,
551 unsigned long page_idx;
552 unsigned long combined_idx;
553 unsigned long uninitialized_var(buddy_idx);
556 if (unlikely(PageCompound(page)))
557 if (unlikely(destroy_compound_page(page, order)))
560 VM_BUG_ON(migratetype == -1);
562 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
564 VM_BUG_ON(page_idx & ((1 << order) - 1));
565 VM_BUG_ON(bad_range(zone, page));
567 while (order < MAX_ORDER-1) {
568 buddy_idx = __find_buddy_index(page_idx, order);
569 buddy = page + (buddy_idx - page_idx);
570 if (!page_is_buddy(page, buddy, order))
573 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
574 * merge with it and move up one order.
576 if (page_is_guard(buddy)) {
577 clear_page_guard_flag(buddy);
578 set_page_private(page, 0);
579 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
581 list_del(&buddy->lru);
582 zone->free_area[order].nr_free--;
583 rmv_page_order(buddy);
585 combined_idx = buddy_idx & page_idx;
586 page = page + (combined_idx - page_idx);
587 page_idx = combined_idx;
590 set_page_order(page, order);
593 * If this is not the largest possible page, check if the buddy
594 * of the next-highest order is free. If it is, it's possible
595 * that pages are being freed that will coalesce soon. In case,
596 * that is happening, add the free page to the tail of the list
597 * so it's less likely to be used soon and more likely to be merged
598 * as a higher order page
600 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
601 struct page *higher_page, *higher_buddy;
602 combined_idx = buddy_idx & page_idx;
603 higher_page = page + (combined_idx - page_idx);
604 buddy_idx = __find_buddy_index(combined_idx, order + 1);
605 higher_buddy = higher_page + (buddy_idx - combined_idx);
606 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
607 list_add_tail(&page->lru,
608 &zone->free_area[order].free_list[migratetype]);
613 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
615 zone->free_area[order].nr_free++;
619 * free_page_mlock() -- clean up attempts to free and mlocked() page.
620 * Page should not be on lru, so no need to fix that up.
621 * free_pages_check() will verify...
623 static inline void free_page_mlock(struct page *page)
625 __dec_zone_page_state(page, NR_MLOCK);
626 __count_vm_event(UNEVICTABLE_MLOCKFREED);
629 static inline int free_pages_check(struct page *page)
631 if (unlikely(page_mapcount(page) |
632 (page->mapping != NULL) |
633 (atomic_read(&page->_count) != 0) |
634 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
635 (mem_cgroup_bad_page_check(page)))) {
639 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
640 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
645 * Frees a number of pages from the PCP lists
646 * Assumes all pages on list are in same zone, and of same order.
647 * count is the number of pages to free.
649 * If the zone was previously in an "all pages pinned" state then look to
650 * see if this freeing clears that state.
652 * And clear the zone's pages_scanned counter, to hold off the "all pages are
653 * pinned" detection logic.
655 static void free_pcppages_bulk(struct zone *zone, int count,
656 struct per_cpu_pages *pcp)
662 spin_lock(&zone->lock);
663 zone->all_unreclaimable = 0;
664 zone->pages_scanned = 0;
668 struct list_head *list;
671 * Remove pages from lists in a round-robin fashion. A
672 * batch_free count is maintained that is incremented when an
673 * empty list is encountered. This is so more pages are freed
674 * off fuller lists instead of spinning excessively around empty
679 if (++migratetype == MIGRATE_PCPTYPES)
681 list = &pcp->lists[migratetype];
682 } while (list_empty(list));
684 /* This is the only non-empty list. Free them all. */
685 if (batch_free == MIGRATE_PCPTYPES)
686 batch_free = to_free;
689 page = list_entry(list->prev, struct page, lru);
690 /* must delete as __free_one_page list manipulates */
691 list_del(&page->lru);
692 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
693 __free_one_page(page, zone, 0, page_private(page));
694 trace_mm_page_pcpu_drain(page, 0, page_private(page));
695 } while (--to_free && --batch_free && !list_empty(list));
697 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
698 spin_unlock(&zone->lock);
701 static void free_one_page(struct zone *zone, struct page *page, int order,
704 spin_lock(&zone->lock);
705 zone->all_unreclaimable = 0;
706 zone->pages_scanned = 0;
708 __free_one_page(page, zone, order, migratetype);
709 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
710 spin_unlock(&zone->lock);
713 static bool free_pages_prepare(struct page *page, unsigned int order)
718 trace_mm_page_free(page, order);
719 kmemcheck_free_shadow(page, order);
722 page->mapping = NULL;
723 for (i = 0; i < (1 << order); i++)
724 bad += free_pages_check(page + i);
728 if (!PageHighMem(page)) {
729 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
730 debug_check_no_obj_freed(page_address(page),
733 arch_free_page(page, order);
734 kernel_map_pages(page, 1 << order, 0);
739 static void __free_pages_ok(struct page *page, unsigned int order)
742 int wasMlocked = __TestClearPageMlocked(page);
744 if (!free_pages_prepare(page, order))
747 local_irq_save(flags);
748 if (unlikely(wasMlocked))
749 free_page_mlock(page);
750 __count_vm_events(PGFREE, 1 << order);
751 free_one_page(page_zone(page), page, order,
752 get_pageblock_migratetype(page));
753 local_irq_restore(flags);
757 * permit the bootmem allocator to evade page validation on high-order frees
759 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
762 __ClearPageReserved(page);
763 set_page_count(page, 0);
764 set_page_refcounted(page);
770 for (loop = 0; loop < BITS_PER_LONG; loop++) {
771 struct page *p = &page[loop];
773 if (loop + 1 < BITS_PER_LONG)
775 __ClearPageReserved(p);
776 set_page_count(p, 0);
779 set_page_refcounted(page);
780 __free_pages(page, order);
786 * The order of subdivision here is critical for the IO subsystem.
787 * Please do not alter this order without good reasons and regression
788 * testing. Specifically, as large blocks of memory are subdivided,
789 * the order in which smaller blocks are delivered depends on the order
790 * they're subdivided in this function. This is the primary factor
791 * influencing the order in which pages are delivered to the IO
792 * subsystem according to empirical testing, and this is also justified
793 * by considering the behavior of a buddy system containing a single
794 * large block of memory acted on by a series of small allocations.
795 * This behavior is a critical factor in sglist merging's success.
799 static inline void expand(struct zone *zone, struct page *page,
800 int low, int high, struct free_area *area,
803 unsigned long size = 1 << high;
809 VM_BUG_ON(bad_range(zone, &page[size]));
811 #ifdef CONFIG_DEBUG_PAGEALLOC
812 if (high < debug_guardpage_minorder()) {
814 * Mark as guard pages (or page), that will allow to
815 * merge back to allocator when buddy will be freed.
816 * Corresponding page table entries will not be touched,
817 * pages will stay not present in virtual address space
819 INIT_LIST_HEAD(&page[size].lru);
820 set_page_guard_flag(&page[size]);
821 set_page_private(&page[size], high);
822 /* Guard pages are not available for any usage */
823 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high));
827 list_add(&page[size].lru, &area->free_list[migratetype]);
829 set_page_order(&page[size], high);
834 * This page is about to be returned from the page allocator
836 static inline int check_new_page(struct page *page)
838 if (unlikely(page_mapcount(page) |
839 (page->mapping != NULL) |
840 (atomic_read(&page->_count) != 0) |
841 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
842 (mem_cgroup_bad_page_check(page)))) {
849 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
853 for (i = 0; i < (1 << order); i++) {
854 struct page *p = page + i;
855 if (unlikely(check_new_page(p)))
859 set_page_private(page, 0);
860 set_page_refcounted(page);
862 arch_alloc_page(page, order);
863 kernel_map_pages(page, 1 << order, 1);
865 if (gfp_flags & __GFP_ZERO)
866 prep_zero_page(page, order, gfp_flags);
868 if (order && (gfp_flags & __GFP_COMP))
869 prep_compound_page(page, order);
875 * Go through the free lists for the given migratetype and remove
876 * the smallest available page from the freelists
879 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
882 unsigned int current_order;
883 struct free_area * area;
886 /* Find a page of the appropriate size in the preferred list */
887 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
888 area = &(zone->free_area[current_order]);
889 if (list_empty(&area->free_list[migratetype]))
892 page = list_entry(area->free_list[migratetype].next,
894 list_del(&page->lru);
895 rmv_page_order(page);
897 expand(zone, page, order, current_order, area, migratetype);
906 * This array describes the order lists are fallen back to when
907 * the free lists for the desirable migrate type are depleted
909 static int fallbacks[MIGRATE_TYPES][3] = {
910 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
911 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
912 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
913 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
914 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
918 * Move the free pages in a range to the free lists of the requested type.
919 * Note that start_page and end_pages are not aligned on a pageblock
920 * boundary. If alignment is required, use move_freepages_block()
922 static int move_freepages(struct zone *zone,
923 struct page *start_page, struct page *end_page,
930 #ifndef CONFIG_HOLES_IN_ZONE
932 * page_zone is not safe to call in this context when
933 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
934 * anyway as we check zone boundaries in move_freepages_block().
935 * Remove at a later date when no bug reports exist related to
936 * grouping pages by mobility
938 BUG_ON(page_zone(start_page) != page_zone(end_page));
941 for (page = start_page; page <= end_page;) {
942 /* Make sure we are not inadvertently changing nodes */
943 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
945 if (!pfn_valid_within(page_to_pfn(page))) {
950 if (!PageBuddy(page)) {
955 order = page_order(page);
956 list_move(&page->lru,
957 &zone->free_area[order].free_list[migratetype]);
959 pages_moved += 1 << order;
965 static int move_freepages_block(struct zone *zone, struct page *page,
968 unsigned long start_pfn, end_pfn;
969 struct page *start_page, *end_page;
971 start_pfn = page_to_pfn(page);
972 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
973 start_page = pfn_to_page(start_pfn);
974 end_page = start_page + pageblock_nr_pages - 1;
975 end_pfn = start_pfn + pageblock_nr_pages - 1;
977 /* Do not cross zone boundaries */
978 if (start_pfn < zone->zone_start_pfn)
980 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
983 return move_freepages(zone, start_page, end_page, migratetype);
986 static void change_pageblock_range(struct page *pageblock_page,
987 int start_order, int migratetype)
989 int nr_pageblocks = 1 << (start_order - pageblock_order);
991 while (nr_pageblocks--) {
992 set_pageblock_migratetype(pageblock_page, migratetype);
993 pageblock_page += pageblock_nr_pages;
997 /* Remove an element from the buddy allocator from the fallback list */
998 static inline struct page *
999 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1001 struct free_area * area;
1006 /* Find the largest possible block of pages in the other list */
1007 for (current_order = MAX_ORDER-1; current_order >= order;
1010 migratetype = fallbacks[start_migratetype][i];
1012 /* MIGRATE_RESERVE handled later if necessary */
1013 if (migratetype == MIGRATE_RESERVE)
1016 area = &(zone->free_area[current_order]);
1017 if (list_empty(&area->free_list[migratetype]))
1020 page = list_entry(area->free_list[migratetype].next,
1025 * If breaking a large block of pages, move all free
1026 * pages to the preferred allocation list. If falling
1027 * back for a reclaimable kernel allocation, be more
1028 * aggressive about taking ownership of free pages
1030 if (unlikely(current_order >= (pageblock_order >> 1)) ||
1031 start_migratetype == MIGRATE_RECLAIMABLE ||
1032 page_group_by_mobility_disabled) {
1033 unsigned long pages;
1034 pages = move_freepages_block(zone, page,
1037 /* Claim the whole block if over half of it is free */
1038 if (pages >= (1 << (pageblock_order-1)) ||
1039 page_group_by_mobility_disabled)
1040 set_pageblock_migratetype(page,
1043 migratetype = start_migratetype;
1046 /* Remove the page from the freelists */
1047 list_del(&page->lru);
1048 rmv_page_order(page);
1050 /* Take ownership for orders >= pageblock_order */
1051 if (current_order >= pageblock_order)
1052 change_pageblock_range(page, current_order,
1055 expand(zone, page, order, current_order, area, migratetype);
1057 trace_mm_page_alloc_extfrag(page, order, current_order,
1058 start_migratetype, migratetype);
1068 * Do the hard work of removing an element from the buddy allocator.
1069 * Call me with the zone->lock already held.
1071 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1077 page = __rmqueue_smallest(zone, order, migratetype);
1079 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1080 page = __rmqueue_fallback(zone, order, migratetype);
1083 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1084 * is used because __rmqueue_smallest is an inline function
1085 * and we want just one call site
1088 migratetype = MIGRATE_RESERVE;
1093 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1098 * Obtain a specified number of elements from the buddy allocator, all under
1099 * a single hold of the lock, for efficiency. Add them to the supplied list.
1100 * Returns the number of new pages which were placed at *list.
1102 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1103 unsigned long count, struct list_head *list,
1104 int migratetype, int cold)
1108 spin_lock(&zone->lock);
1109 for (i = 0; i < count; ++i) {
1110 struct page *page = __rmqueue(zone, order, migratetype);
1111 if (unlikely(page == NULL))
1115 * Split buddy pages returned by expand() are received here
1116 * in physical page order. The page is added to the callers and
1117 * list and the list head then moves forward. From the callers
1118 * perspective, the linked list is ordered by page number in
1119 * some conditions. This is useful for IO devices that can
1120 * merge IO requests if the physical pages are ordered
1123 if (likely(cold == 0))
1124 list_add(&page->lru, list);
1126 list_add_tail(&page->lru, list);
1127 set_page_private(page, migratetype);
1130 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1131 spin_unlock(&zone->lock);
1137 * Called from the vmstat counter updater to drain pagesets of this
1138 * currently executing processor on remote nodes after they have
1141 * Note that this function must be called with the thread pinned to
1142 * a single processor.
1144 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1146 unsigned long flags;
1149 local_irq_save(flags);
1150 if (pcp->count >= pcp->batch)
1151 to_drain = pcp->batch;
1153 to_drain = pcp->count;
1154 free_pcppages_bulk(zone, to_drain, pcp);
1155 pcp->count -= to_drain;
1156 local_irq_restore(flags);
1161 * Drain pages of the indicated processor.
1163 * The processor must either be the current processor and the
1164 * thread pinned to the current processor or a processor that
1167 static void drain_pages(unsigned int cpu)
1169 unsigned long flags;
1172 for_each_populated_zone(zone) {
1173 struct per_cpu_pageset *pset;
1174 struct per_cpu_pages *pcp;
1176 local_irq_save(flags);
1177 pset = per_cpu_ptr(zone->pageset, cpu);
1181 free_pcppages_bulk(zone, pcp->count, pcp);
1184 local_irq_restore(flags);
1189 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1191 void drain_local_pages(void *arg)
1193 drain_pages(smp_processor_id());
1197 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1199 void drain_all_pages(void)
1201 on_each_cpu(drain_local_pages, NULL, 1);
1204 #ifdef CONFIG_HIBERNATION
1206 void mark_free_pages(struct zone *zone)
1208 unsigned long pfn, max_zone_pfn;
1209 unsigned long flags;
1211 struct list_head *curr;
1213 if (!zone->spanned_pages)
1216 spin_lock_irqsave(&zone->lock, flags);
1218 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1219 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1220 if (pfn_valid(pfn)) {
1221 struct page *page = pfn_to_page(pfn);
1223 if (!swsusp_page_is_forbidden(page))
1224 swsusp_unset_page_free(page);
1227 for_each_migratetype_order(order, t) {
1228 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1231 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1232 for (i = 0; i < (1UL << order); i++)
1233 swsusp_set_page_free(pfn_to_page(pfn + i));
1236 spin_unlock_irqrestore(&zone->lock, flags);
1238 #endif /* CONFIG_PM */
1241 * Free a 0-order page
1242 * cold == 1 ? free a cold page : free a hot page
1244 void free_hot_cold_page(struct page *page, int cold)
1246 struct zone *zone = page_zone(page);
1247 struct per_cpu_pages *pcp;
1248 unsigned long flags;
1250 int wasMlocked = __TestClearPageMlocked(page);
1252 if (!free_pages_prepare(page, 0))
1255 migratetype = get_pageblock_migratetype(page);
1256 set_page_private(page, migratetype);
1257 local_irq_save(flags);
1258 if (unlikely(wasMlocked))
1259 free_page_mlock(page);
1260 __count_vm_event(PGFREE);
1263 * We only track unmovable, reclaimable and movable on pcp lists.
1264 * Free ISOLATE pages back to the allocator because they are being
1265 * offlined but treat RESERVE as movable pages so we can get those
1266 * areas back if necessary. Otherwise, we may have to free
1267 * excessively into the page allocator
1269 if (migratetype >= MIGRATE_PCPTYPES) {
1270 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1271 free_one_page(zone, page, 0, migratetype);
1274 migratetype = MIGRATE_MOVABLE;
1277 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1279 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1281 list_add(&page->lru, &pcp->lists[migratetype]);
1283 if (pcp->count >= pcp->high) {
1284 free_pcppages_bulk(zone, pcp->batch, pcp);
1285 pcp->count -= pcp->batch;
1289 local_irq_restore(flags);
1293 * Free a list of 0-order pages
1295 void free_hot_cold_page_list(struct list_head *list, int cold)
1297 struct page *page, *next;
1299 list_for_each_entry_safe(page, next, list, lru) {
1300 trace_mm_page_free_batched(page, cold);
1301 free_hot_cold_page(page, cold);
1306 * split_page takes a non-compound higher-order page, and splits it into
1307 * n (1<<order) sub-pages: page[0..n]
1308 * Each sub-page must be freed individually.
1310 * Note: this is probably too low level an operation for use in drivers.
1311 * Please consult with lkml before using this in your driver.
1313 void split_page(struct page *page, unsigned int order)
1317 VM_BUG_ON(PageCompound(page));
1318 VM_BUG_ON(!page_count(page));
1320 #ifdef CONFIG_KMEMCHECK
1322 * Split shadow pages too, because free(page[0]) would
1323 * otherwise free the whole shadow.
1325 if (kmemcheck_page_is_tracked(page))
1326 split_page(virt_to_page(page[0].shadow), order);
1329 for (i = 1; i < (1 << order); i++)
1330 set_page_refcounted(page + i);
1334 * Similar to split_page except the page is already free. As this is only
1335 * being used for migration, the migratetype of the block also changes.
1336 * As this is called with interrupts disabled, the caller is responsible
1337 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1340 * Note: this is probably too low level an operation for use in drivers.
1341 * Please consult with lkml before using this in your driver.
1343 int split_free_page(struct page *page)
1346 unsigned long watermark;
1349 BUG_ON(!PageBuddy(page));
1351 zone = page_zone(page);
1352 order = page_order(page);
1354 /* Obey watermarks as if the page was being allocated */
1355 watermark = low_wmark_pages(zone) + (1 << order);
1356 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1359 /* Remove page from free list */
1360 list_del(&page->lru);
1361 zone->free_area[order].nr_free--;
1362 rmv_page_order(page);
1363 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1365 /* Split into individual pages */
1366 set_page_refcounted(page);
1367 split_page(page, order);
1369 if (order >= pageblock_order - 1) {
1370 struct page *endpage = page + (1 << order) - 1;
1371 for (; page < endpage; page += pageblock_nr_pages)
1372 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1379 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1380 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1384 struct page *buffered_rmqueue(struct zone *preferred_zone,
1385 struct zone *zone, int order, gfp_t gfp_flags,
1388 unsigned long flags;
1390 int cold = !!(gfp_flags & __GFP_COLD);
1393 if (likely(order == 0)) {
1394 struct per_cpu_pages *pcp;
1395 struct list_head *list;
1397 local_irq_save(flags);
1398 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1399 list = &pcp->lists[migratetype];
1400 if (list_empty(list)) {
1401 pcp->count += rmqueue_bulk(zone, 0,
1404 if (unlikely(list_empty(list)))
1409 page = list_entry(list->prev, struct page, lru);
1411 page = list_entry(list->next, struct page, lru);
1413 list_del(&page->lru);
1416 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1418 * __GFP_NOFAIL is not to be used in new code.
1420 * All __GFP_NOFAIL callers should be fixed so that they
1421 * properly detect and handle allocation failures.
1423 * We most definitely don't want callers attempting to
1424 * allocate greater than order-1 page units with
1427 WARN_ON_ONCE(order > 1);
1429 spin_lock_irqsave(&zone->lock, flags);
1430 page = __rmqueue(zone, order, migratetype);
1431 spin_unlock(&zone->lock);
1434 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1437 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1438 zone_statistics(preferred_zone, zone, gfp_flags);
1439 local_irq_restore(flags);
1441 VM_BUG_ON(bad_range(zone, page));
1442 if (prep_new_page(page, order, gfp_flags))
1447 local_irq_restore(flags);
1451 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1452 #define ALLOC_WMARK_MIN WMARK_MIN
1453 #define ALLOC_WMARK_LOW WMARK_LOW
1454 #define ALLOC_WMARK_HIGH WMARK_HIGH
1455 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1457 /* Mask to get the watermark bits */
1458 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1460 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1461 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1462 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1464 #ifdef CONFIG_FAIL_PAGE_ALLOC
1467 struct fault_attr attr;
1469 u32 ignore_gfp_highmem;
1470 u32 ignore_gfp_wait;
1472 } fail_page_alloc = {
1473 .attr = FAULT_ATTR_INITIALIZER,
1474 .ignore_gfp_wait = 1,
1475 .ignore_gfp_highmem = 1,
1479 static int __init setup_fail_page_alloc(char *str)
1481 return setup_fault_attr(&fail_page_alloc.attr, str);
1483 __setup("fail_page_alloc=", setup_fail_page_alloc);
1485 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1487 if (order < fail_page_alloc.min_order)
1489 if (gfp_mask & __GFP_NOFAIL)
1491 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1493 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1496 return should_fail(&fail_page_alloc.attr, 1 << order);
1499 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1501 static int __init fail_page_alloc_debugfs(void)
1503 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1506 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1507 &fail_page_alloc.attr);
1509 return PTR_ERR(dir);
1511 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1512 &fail_page_alloc.ignore_gfp_wait))
1514 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1515 &fail_page_alloc.ignore_gfp_highmem))
1517 if (!debugfs_create_u32("min-order", mode, dir,
1518 &fail_page_alloc.min_order))
1523 debugfs_remove_recursive(dir);
1528 late_initcall(fail_page_alloc_debugfs);
1530 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1532 #else /* CONFIG_FAIL_PAGE_ALLOC */
1534 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1539 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1542 * Return true if free pages are above 'mark'. This takes into account the order
1543 * of the allocation.
1545 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1546 int classzone_idx, int alloc_flags, long free_pages)
1548 /* free_pages my go negative - that's OK */
1552 free_pages -= (1 << order) - 1;
1553 if (alloc_flags & ALLOC_HIGH)
1555 if (alloc_flags & ALLOC_HARDER)
1558 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1560 for (o = 0; o < order; o++) {
1561 /* At the next order, this order's pages become unavailable */
1562 free_pages -= z->free_area[o].nr_free << o;
1564 /* Require fewer higher order pages to be free */
1567 if (free_pages <= min)
1573 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1574 int classzone_idx, int alloc_flags)
1576 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1577 zone_page_state(z, NR_FREE_PAGES));
1580 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1581 int classzone_idx, int alloc_flags)
1583 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1585 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1586 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1588 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1594 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1595 * skip over zones that are not allowed by the cpuset, or that have
1596 * been recently (in last second) found to be nearly full. See further
1597 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1598 * that have to skip over a lot of full or unallowed zones.
1600 * If the zonelist cache is present in the passed in zonelist, then
1601 * returns a pointer to the allowed node mask (either the current
1602 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1604 * If the zonelist cache is not available for this zonelist, does
1605 * nothing and returns NULL.
1607 * If the fullzones BITMAP in the zonelist cache is stale (more than
1608 * a second since last zap'd) then we zap it out (clear its bits.)
1610 * We hold off even calling zlc_setup, until after we've checked the
1611 * first zone in the zonelist, on the theory that most allocations will
1612 * be satisfied from that first zone, so best to examine that zone as
1613 * quickly as we can.
1615 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1617 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1618 nodemask_t *allowednodes; /* zonelist_cache approximation */
1620 zlc = zonelist->zlcache_ptr;
1624 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1625 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1626 zlc->last_full_zap = jiffies;
1629 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1630 &cpuset_current_mems_allowed :
1631 &node_states[N_HIGH_MEMORY];
1632 return allowednodes;
1636 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1637 * if it is worth looking at further for free memory:
1638 * 1) Check that the zone isn't thought to be full (doesn't have its
1639 * bit set in the zonelist_cache fullzones BITMAP).
1640 * 2) Check that the zones node (obtained from the zonelist_cache
1641 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1642 * Return true (non-zero) if zone is worth looking at further, or
1643 * else return false (zero) if it is not.
1645 * This check -ignores- the distinction between various watermarks,
1646 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1647 * found to be full for any variation of these watermarks, it will
1648 * be considered full for up to one second by all requests, unless
1649 * we are so low on memory on all allowed nodes that we are forced
1650 * into the second scan of the zonelist.
1652 * In the second scan we ignore this zonelist cache and exactly
1653 * apply the watermarks to all zones, even it is slower to do so.
1654 * We are low on memory in the second scan, and should leave no stone
1655 * unturned looking for a free page.
1657 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1658 nodemask_t *allowednodes)
1660 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1661 int i; /* index of *z in zonelist zones */
1662 int n; /* node that zone *z is on */
1664 zlc = zonelist->zlcache_ptr;
1668 i = z - zonelist->_zonerefs;
1671 /* This zone is worth trying if it is allowed but not full */
1672 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1676 * Given 'z' scanning a zonelist, set the corresponding bit in
1677 * zlc->fullzones, so that subsequent attempts to allocate a page
1678 * from that zone don't waste time re-examining it.
1680 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1682 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1683 int i; /* index of *z in zonelist zones */
1685 zlc = zonelist->zlcache_ptr;
1689 i = z - zonelist->_zonerefs;
1691 set_bit(i, zlc->fullzones);
1695 * clear all zones full, called after direct reclaim makes progress so that
1696 * a zone that was recently full is not skipped over for up to a second
1698 static void zlc_clear_zones_full(struct zonelist *zonelist)
1700 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1702 zlc = zonelist->zlcache_ptr;
1706 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1709 #else /* CONFIG_NUMA */
1711 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1716 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1717 nodemask_t *allowednodes)
1722 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1726 static void zlc_clear_zones_full(struct zonelist *zonelist)
1729 #endif /* CONFIG_NUMA */
1732 * get_page_from_freelist goes through the zonelist trying to allocate
1735 static struct page *
1736 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1737 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1738 struct zone *preferred_zone, int migratetype)
1741 struct page *page = NULL;
1744 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1745 int zlc_active = 0; /* set if using zonelist_cache */
1746 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1748 classzone_idx = zone_idx(preferred_zone);
1751 * Scan zonelist, looking for a zone with enough free.
1752 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1754 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1755 high_zoneidx, nodemask) {
1756 if (NUMA_BUILD && zlc_active &&
1757 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1759 if ((alloc_flags & ALLOC_CPUSET) &&
1760 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1763 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1764 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1768 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1769 if (zone_watermark_ok(zone, order, mark,
1770 classzone_idx, alloc_flags))
1773 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1775 * we do zlc_setup if there are multiple nodes
1776 * and before considering the first zone allowed
1779 allowednodes = zlc_setup(zonelist, alloc_flags);
1784 if (zone_reclaim_mode == 0)
1785 goto this_zone_full;
1788 * As we may have just activated ZLC, check if the first
1789 * eligible zone has failed zone_reclaim recently.
1791 if (NUMA_BUILD && zlc_active &&
1792 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1795 ret = zone_reclaim(zone, gfp_mask, order);
1797 case ZONE_RECLAIM_NOSCAN:
1800 case ZONE_RECLAIM_FULL:
1801 /* scanned but unreclaimable */
1804 /* did we reclaim enough */
1805 if (!zone_watermark_ok(zone, order, mark,
1806 classzone_idx, alloc_flags))
1807 goto this_zone_full;
1812 page = buffered_rmqueue(preferred_zone, zone, order,
1813 gfp_mask, migratetype);
1818 zlc_mark_zone_full(zonelist, z);
1821 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1822 /* Disable zlc cache for second zonelist scan */
1830 * Large machines with many possible nodes should not always dump per-node
1831 * meminfo in irq context.
1833 static inline bool should_suppress_show_mem(void)
1838 ret = in_interrupt();
1843 static DEFINE_RATELIMIT_STATE(nopage_rs,
1844 DEFAULT_RATELIMIT_INTERVAL,
1845 DEFAULT_RATELIMIT_BURST);
1847 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1849 unsigned int filter = SHOW_MEM_FILTER_NODES;
1851 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1852 debug_guardpage_minorder() > 0)
1856 * This documents exceptions given to allocations in certain
1857 * contexts that are allowed to allocate outside current's set
1860 if (!(gfp_mask & __GFP_NOMEMALLOC))
1861 if (test_thread_flag(TIF_MEMDIE) ||
1862 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1863 filter &= ~SHOW_MEM_FILTER_NODES;
1864 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1865 filter &= ~SHOW_MEM_FILTER_NODES;
1868 struct va_format vaf;
1871 va_start(args, fmt);
1876 pr_warn("%pV", &vaf);
1881 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1882 current->comm, order, gfp_mask);
1885 if (!should_suppress_show_mem())
1890 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1891 unsigned long did_some_progress,
1892 unsigned long pages_reclaimed)
1894 /* Do not loop if specifically requested */
1895 if (gfp_mask & __GFP_NORETRY)
1898 /* Always retry if specifically requested */
1899 if (gfp_mask & __GFP_NOFAIL)
1903 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
1904 * making forward progress without invoking OOM. Suspend also disables
1905 * storage devices so kswapd will not help. Bail if we are suspending.
1907 if (!did_some_progress && pm_suspended_storage())
1911 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1912 * means __GFP_NOFAIL, but that may not be true in other
1915 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1919 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1920 * specified, then we retry until we no longer reclaim any pages
1921 * (above), or we've reclaimed an order of pages at least as
1922 * large as the allocation's order. In both cases, if the
1923 * allocation still fails, we stop retrying.
1925 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1931 static inline struct page *
1932 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1933 struct zonelist *zonelist, enum zone_type high_zoneidx,
1934 nodemask_t *nodemask, struct zone *preferred_zone,
1939 /* Acquire the OOM killer lock for the zones in zonelist */
1940 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1941 schedule_timeout_uninterruptible(1);
1946 * Go through the zonelist yet one more time, keep very high watermark
1947 * here, this is only to catch a parallel oom killing, we must fail if
1948 * we're still under heavy pressure.
1950 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1951 order, zonelist, high_zoneidx,
1952 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1953 preferred_zone, migratetype);
1957 if (!(gfp_mask & __GFP_NOFAIL)) {
1958 /* The OOM killer will not help higher order allocs */
1959 if (order > PAGE_ALLOC_COSTLY_ORDER)
1961 /* The OOM killer does not needlessly kill tasks for lowmem */
1962 if (high_zoneidx < ZONE_NORMAL)
1965 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1966 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1967 * The caller should handle page allocation failure by itself if
1968 * it specifies __GFP_THISNODE.
1969 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1971 if (gfp_mask & __GFP_THISNODE)
1974 /* Exhausted what can be done so it's blamo time */
1975 out_of_memory(zonelist, gfp_mask, order, nodemask);
1978 clear_zonelist_oom(zonelist, gfp_mask);
1982 #ifdef CONFIG_COMPACTION
1983 /* Try memory compaction for high-order allocations before reclaim */
1984 static struct page *
1985 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1986 struct zonelist *zonelist, enum zone_type high_zoneidx,
1987 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1988 int migratetype, bool sync_migration,
1989 bool *deferred_compaction,
1990 unsigned long *did_some_progress)
1997 if (compaction_deferred(preferred_zone)) {
1998 *deferred_compaction = true;
2002 current->flags |= PF_MEMALLOC;
2003 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2004 nodemask, sync_migration);
2005 current->flags &= ~PF_MEMALLOC;
2006 if (*did_some_progress != COMPACT_SKIPPED) {
2008 /* Page migration frees to the PCP lists but we want merging */
2009 drain_pages(get_cpu());
2012 page = get_page_from_freelist(gfp_mask, nodemask,
2013 order, zonelist, high_zoneidx,
2014 alloc_flags, preferred_zone,
2017 preferred_zone->compact_considered = 0;
2018 preferred_zone->compact_defer_shift = 0;
2019 count_vm_event(COMPACTSUCCESS);
2024 * It's bad if compaction run occurs and fails.
2025 * The most likely reason is that pages exist,
2026 * but not enough to satisfy watermarks.
2028 count_vm_event(COMPACTFAIL);
2031 * As async compaction considers a subset of pageblocks, only
2032 * defer if the failure was a sync compaction failure.
2035 defer_compaction(preferred_zone);
2043 static inline struct page *
2044 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2045 struct zonelist *zonelist, enum zone_type high_zoneidx,
2046 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2047 int migratetype, bool sync_migration,
2048 bool *deferred_compaction,
2049 unsigned long *did_some_progress)
2053 #endif /* CONFIG_COMPACTION */
2055 /* The really slow allocator path where we enter direct reclaim */
2056 static inline struct page *
2057 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2058 struct zonelist *zonelist, enum zone_type high_zoneidx,
2059 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2060 int migratetype, unsigned long *did_some_progress)
2062 struct page *page = NULL;
2063 struct reclaim_state reclaim_state;
2064 bool drained = false;
2068 /* We now go into synchronous reclaim */
2069 cpuset_memory_pressure_bump();
2070 current->flags |= PF_MEMALLOC;
2071 lockdep_set_current_reclaim_state(gfp_mask);
2072 reclaim_state.reclaimed_slab = 0;
2073 current->reclaim_state = &reclaim_state;
2075 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2077 current->reclaim_state = NULL;
2078 lockdep_clear_current_reclaim_state();
2079 current->flags &= ~PF_MEMALLOC;
2083 if (unlikely(!(*did_some_progress)))
2086 /* After successful reclaim, reconsider all zones for allocation */
2088 zlc_clear_zones_full(zonelist);
2091 page = get_page_from_freelist(gfp_mask, nodemask, order,
2092 zonelist, high_zoneidx,
2093 alloc_flags, preferred_zone,
2097 * If an allocation failed after direct reclaim, it could be because
2098 * pages are pinned on the per-cpu lists. Drain them and try again
2100 if (!page && !drained) {
2110 * This is called in the allocator slow-path if the allocation request is of
2111 * sufficient urgency to ignore watermarks and take other desperate measures
2113 static inline struct page *
2114 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2115 struct zonelist *zonelist, enum zone_type high_zoneidx,
2116 nodemask_t *nodemask, struct zone *preferred_zone,
2122 page = get_page_from_freelist(gfp_mask, nodemask, order,
2123 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2124 preferred_zone, migratetype);
2126 if (!page && gfp_mask & __GFP_NOFAIL)
2127 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2128 } while (!page && (gfp_mask & __GFP_NOFAIL));
2134 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2135 enum zone_type high_zoneidx,
2136 enum zone_type classzone_idx)
2141 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2142 wakeup_kswapd(zone, order, classzone_idx);
2146 gfp_to_alloc_flags(gfp_t gfp_mask)
2148 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2149 const gfp_t wait = gfp_mask & __GFP_WAIT;
2151 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2152 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2155 * The caller may dip into page reserves a bit more if the caller
2156 * cannot run direct reclaim, or if the caller has realtime scheduling
2157 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2158 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2160 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2164 * Not worth trying to allocate harder for
2165 * __GFP_NOMEMALLOC even if it can't schedule.
2167 if (!(gfp_mask & __GFP_NOMEMALLOC))
2168 alloc_flags |= ALLOC_HARDER;
2170 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2171 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2173 alloc_flags &= ~ALLOC_CPUSET;
2174 } else if (unlikely(rt_task(current)) && !in_interrupt())
2175 alloc_flags |= ALLOC_HARDER;
2177 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2178 if (!in_interrupt() &&
2179 ((current->flags & PF_MEMALLOC) ||
2180 unlikely(test_thread_flag(TIF_MEMDIE))))
2181 alloc_flags |= ALLOC_NO_WATERMARKS;
2187 static inline struct page *
2188 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2189 struct zonelist *zonelist, enum zone_type high_zoneidx,
2190 nodemask_t *nodemask, struct zone *preferred_zone,
2193 const gfp_t wait = gfp_mask & __GFP_WAIT;
2194 struct page *page = NULL;
2196 unsigned long pages_reclaimed = 0;
2197 unsigned long did_some_progress;
2198 bool sync_migration = false;
2199 bool deferred_compaction = false;
2202 * In the slowpath, we sanity check order to avoid ever trying to
2203 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2204 * be using allocators in order of preference for an area that is
2207 if (order >= MAX_ORDER) {
2208 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2213 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2214 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2215 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2216 * using a larger set of nodes after it has established that the
2217 * allowed per node queues are empty and that nodes are
2220 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2224 if (!(gfp_mask & __GFP_NO_KSWAPD))
2225 wake_all_kswapd(order, zonelist, high_zoneidx,
2226 zone_idx(preferred_zone));
2229 * OK, we're below the kswapd watermark and have kicked background
2230 * reclaim. Now things get more complex, so set up alloc_flags according
2231 * to how we want to proceed.
2233 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2236 * Find the true preferred zone if the allocation is unconstrained by
2239 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2240 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2244 /* This is the last chance, in general, before the goto nopage. */
2245 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2246 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2247 preferred_zone, migratetype);
2251 /* Allocate without watermarks if the context allows */
2252 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2253 page = __alloc_pages_high_priority(gfp_mask, order,
2254 zonelist, high_zoneidx, nodemask,
2255 preferred_zone, migratetype);
2260 /* Atomic allocations - we can't balance anything */
2264 /* Avoid recursion of direct reclaim */
2265 if (current->flags & PF_MEMALLOC)
2268 /* Avoid allocations with no watermarks from looping endlessly */
2269 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2273 * Try direct compaction. The first pass is asynchronous. Subsequent
2274 * attempts after direct reclaim are synchronous
2276 page = __alloc_pages_direct_compact(gfp_mask, order,
2277 zonelist, high_zoneidx,
2279 alloc_flags, preferred_zone,
2280 migratetype, sync_migration,
2281 &deferred_compaction,
2282 &did_some_progress);
2285 sync_migration = true;
2288 * If compaction is deferred for high-order allocations, it is because
2289 * sync compaction recently failed. In this is the case and the caller
2290 * has requested the system not be heavily disrupted, fail the
2291 * allocation now instead of entering direct reclaim
2293 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2296 /* Try direct reclaim and then allocating */
2297 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2298 zonelist, high_zoneidx,
2300 alloc_flags, preferred_zone,
2301 migratetype, &did_some_progress);
2306 * If we failed to make any progress reclaiming, then we are
2307 * running out of options and have to consider going OOM
2309 if (!did_some_progress) {
2310 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2311 if (oom_killer_disabled)
2313 page = __alloc_pages_may_oom(gfp_mask, order,
2314 zonelist, high_zoneidx,
2315 nodemask, preferred_zone,
2320 if (!(gfp_mask & __GFP_NOFAIL)) {
2322 * The oom killer is not called for high-order
2323 * allocations that may fail, so if no progress
2324 * is being made, there are no other options and
2325 * retrying is unlikely to help.
2327 if (order > PAGE_ALLOC_COSTLY_ORDER)
2330 * The oom killer is not called for lowmem
2331 * allocations to prevent needlessly killing
2334 if (high_zoneidx < ZONE_NORMAL)
2342 /* Check if we should retry the allocation */
2343 pages_reclaimed += did_some_progress;
2344 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2346 /* Wait for some write requests to complete then retry */
2347 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2351 * High-order allocations do not necessarily loop after
2352 * direct reclaim and reclaim/compaction depends on compaction
2353 * being called after reclaim so call directly if necessary
2355 page = __alloc_pages_direct_compact(gfp_mask, order,
2356 zonelist, high_zoneidx,
2358 alloc_flags, preferred_zone,
2359 migratetype, sync_migration,
2360 &deferred_compaction,
2361 &did_some_progress);
2367 warn_alloc_failed(gfp_mask, order, NULL);
2370 if (kmemcheck_enabled)
2371 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2377 * This is the 'heart' of the zoned buddy allocator.
2380 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2381 struct zonelist *zonelist, nodemask_t *nodemask)
2383 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2384 struct zone *preferred_zone;
2385 struct page *page = NULL;
2386 int migratetype = allocflags_to_migratetype(gfp_mask);
2387 unsigned int cpuset_mems_cookie;
2389 gfp_mask &= gfp_allowed_mask;
2391 lockdep_trace_alloc(gfp_mask);
2393 might_sleep_if(gfp_mask & __GFP_WAIT);
2395 if (should_fail_alloc_page(gfp_mask, order))
2399 * Check the zones suitable for the gfp_mask contain at least one
2400 * valid zone. It's possible to have an empty zonelist as a result
2401 * of GFP_THISNODE and a memoryless node
2403 if (unlikely(!zonelist->_zonerefs->zone))
2407 cpuset_mems_cookie = get_mems_allowed();
2409 /* The preferred zone is used for statistics later */
2410 first_zones_zonelist(zonelist, high_zoneidx,
2411 nodemask ? : &cpuset_current_mems_allowed,
2413 if (!preferred_zone)
2416 /* First allocation attempt */
2417 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2418 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2419 preferred_zone, migratetype);
2420 if (unlikely(!page))
2421 page = __alloc_pages_slowpath(gfp_mask, order,
2422 zonelist, high_zoneidx, nodemask,
2423 preferred_zone, migratetype);
2425 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2429 * When updating a task's mems_allowed, it is possible to race with
2430 * parallel threads in such a way that an allocation can fail while
2431 * the mask is being updated. If a page allocation is about to fail,
2432 * check if the cpuset changed during allocation and if so, retry.
2434 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2439 EXPORT_SYMBOL(__alloc_pages_nodemask);
2442 * Common helper functions.
2444 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2449 * __get_free_pages() returns a 32-bit address, which cannot represent
2452 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2454 page = alloc_pages(gfp_mask, order);
2457 return (unsigned long) page_address(page);
2459 EXPORT_SYMBOL(__get_free_pages);
2461 unsigned long get_zeroed_page(gfp_t gfp_mask)
2463 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2465 EXPORT_SYMBOL(get_zeroed_page);
2467 void __free_pages(struct page *page, unsigned int order)
2469 if (put_page_testzero(page)) {
2471 free_hot_cold_page(page, 0);
2473 __free_pages_ok(page, order);
2477 EXPORT_SYMBOL(__free_pages);
2479 void free_pages(unsigned long addr, unsigned int order)
2482 VM_BUG_ON(!virt_addr_valid((void *)addr));
2483 __free_pages(virt_to_page((void *)addr), order);
2487 EXPORT_SYMBOL(free_pages);
2489 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2492 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2493 unsigned long used = addr + PAGE_ALIGN(size);
2495 split_page(virt_to_page((void *)addr), order);
2496 while (used < alloc_end) {
2501 return (void *)addr;
2505 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2506 * @size: the number of bytes to allocate
2507 * @gfp_mask: GFP flags for the allocation
2509 * This function is similar to alloc_pages(), except that it allocates the
2510 * minimum number of pages to satisfy the request. alloc_pages() can only
2511 * allocate memory in power-of-two pages.
2513 * This function is also limited by MAX_ORDER.
2515 * Memory allocated by this function must be released by free_pages_exact().
2517 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2519 unsigned int order = get_order(size);
2522 addr = __get_free_pages(gfp_mask, order);
2523 return make_alloc_exact(addr, order, size);
2525 EXPORT_SYMBOL(alloc_pages_exact);
2528 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2530 * @nid: the preferred node ID where memory should be allocated
2531 * @size: the number of bytes to allocate
2532 * @gfp_mask: GFP flags for the allocation
2534 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2536 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2539 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2541 unsigned order = get_order(size);
2542 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2545 return make_alloc_exact((unsigned long)page_address(p), order, size);
2547 EXPORT_SYMBOL(alloc_pages_exact_nid);
2550 * free_pages_exact - release memory allocated via alloc_pages_exact()
2551 * @virt: the value returned by alloc_pages_exact.
2552 * @size: size of allocation, same value as passed to alloc_pages_exact().
2554 * Release the memory allocated by a previous call to alloc_pages_exact.
2556 void free_pages_exact(void *virt, size_t size)
2558 unsigned long addr = (unsigned long)virt;
2559 unsigned long end = addr + PAGE_ALIGN(size);
2561 while (addr < end) {
2566 EXPORT_SYMBOL(free_pages_exact);
2568 static unsigned int nr_free_zone_pages(int offset)
2573 /* Just pick one node, since fallback list is circular */
2574 unsigned int sum = 0;
2576 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2578 for_each_zone_zonelist(zone, z, zonelist, offset) {
2579 unsigned long size = zone->present_pages;
2580 unsigned long high = high_wmark_pages(zone);
2589 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2591 unsigned int nr_free_buffer_pages(void)
2593 return nr_free_zone_pages(gfp_zone(GFP_USER));
2595 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2598 * Amount of free RAM allocatable within all zones
2600 unsigned int nr_free_pagecache_pages(void)
2602 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2605 static inline void show_node(struct zone *zone)
2608 printk("Node %d ", zone_to_nid(zone));
2611 void si_meminfo(struct sysinfo *val)
2613 val->totalram = totalram_pages;
2615 val->freeram = global_page_state(NR_FREE_PAGES);
2616 val->bufferram = nr_blockdev_pages();
2617 val->totalhigh = totalhigh_pages;
2618 val->freehigh = nr_free_highpages();
2619 val->mem_unit = PAGE_SIZE;
2622 EXPORT_SYMBOL(si_meminfo);
2625 void si_meminfo_node(struct sysinfo *val, int nid)
2627 pg_data_t *pgdat = NODE_DATA(nid);
2629 val->totalram = pgdat->node_present_pages;
2630 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2631 #ifdef CONFIG_HIGHMEM
2632 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2633 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2639 val->mem_unit = PAGE_SIZE;
2644 * Determine whether the node should be displayed or not, depending on whether
2645 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2647 bool skip_free_areas_node(unsigned int flags, int nid)
2650 unsigned int cpuset_mems_cookie;
2652 if (!(flags & SHOW_MEM_FILTER_NODES))
2656 cpuset_mems_cookie = get_mems_allowed();
2657 ret = !node_isset(nid, cpuset_current_mems_allowed);
2658 } while (!put_mems_allowed(cpuset_mems_cookie));
2663 #define K(x) ((x) << (PAGE_SHIFT-10))
2666 * Show free area list (used inside shift_scroll-lock stuff)
2667 * We also calculate the percentage fragmentation. We do this by counting the
2668 * memory on each free list with the exception of the first item on the list.
2669 * Suppresses nodes that are not allowed by current's cpuset if
2670 * SHOW_MEM_FILTER_NODES is passed.
2672 void show_free_areas(unsigned int filter)
2677 for_each_populated_zone(zone) {
2678 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2681 printk("%s per-cpu:\n", zone->name);
2683 for_each_online_cpu(cpu) {
2684 struct per_cpu_pageset *pageset;
2686 pageset = per_cpu_ptr(zone->pageset, cpu);
2688 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2689 cpu, pageset->pcp.high,
2690 pageset->pcp.batch, pageset->pcp.count);
2694 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2695 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2697 " dirty:%lu writeback:%lu unstable:%lu\n"
2698 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2699 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2700 global_page_state(NR_ACTIVE_ANON),
2701 global_page_state(NR_INACTIVE_ANON),
2702 global_page_state(NR_ISOLATED_ANON),
2703 global_page_state(NR_ACTIVE_FILE),
2704 global_page_state(NR_INACTIVE_FILE),
2705 global_page_state(NR_ISOLATED_FILE),
2706 global_page_state(NR_UNEVICTABLE),
2707 global_page_state(NR_FILE_DIRTY),
2708 global_page_state(NR_WRITEBACK),
2709 global_page_state(NR_UNSTABLE_NFS),
2710 global_page_state(NR_FREE_PAGES),
2711 global_page_state(NR_SLAB_RECLAIMABLE),
2712 global_page_state(NR_SLAB_UNRECLAIMABLE),
2713 global_page_state(NR_FILE_MAPPED),
2714 global_page_state(NR_SHMEM),
2715 global_page_state(NR_PAGETABLE),
2716 global_page_state(NR_BOUNCE));
2718 for_each_populated_zone(zone) {
2721 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2729 " active_anon:%lukB"
2730 " inactive_anon:%lukB"
2731 " active_file:%lukB"
2732 " inactive_file:%lukB"
2733 " unevictable:%lukB"
2734 " isolated(anon):%lukB"
2735 " isolated(file):%lukB"
2742 " slab_reclaimable:%lukB"
2743 " slab_unreclaimable:%lukB"
2744 " kernel_stack:%lukB"
2748 " writeback_tmp:%lukB"
2749 " pages_scanned:%lu"
2750 " all_unreclaimable? %s"
2753 K(zone_page_state(zone, NR_FREE_PAGES)),
2754 K(min_wmark_pages(zone)),
2755 K(low_wmark_pages(zone)),
2756 K(high_wmark_pages(zone)),
2757 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2758 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2759 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2760 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2761 K(zone_page_state(zone, NR_UNEVICTABLE)),
2762 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2763 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2764 K(zone->present_pages),
2765 K(zone_page_state(zone, NR_MLOCK)),
2766 K(zone_page_state(zone, NR_FILE_DIRTY)),
2767 K(zone_page_state(zone, NR_WRITEBACK)),
2768 K(zone_page_state(zone, NR_FILE_MAPPED)),
2769 K(zone_page_state(zone, NR_SHMEM)),
2770 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2771 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2772 zone_page_state(zone, NR_KERNEL_STACK) *
2774 K(zone_page_state(zone, NR_PAGETABLE)),
2775 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2776 K(zone_page_state(zone, NR_BOUNCE)),
2777 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2778 zone->pages_scanned,
2779 (zone->all_unreclaimable ? "yes" : "no")
2781 printk("lowmem_reserve[]:");
2782 for (i = 0; i < MAX_NR_ZONES; i++)
2783 printk(" %lu", zone->lowmem_reserve[i]);
2787 for_each_populated_zone(zone) {
2788 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2790 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2793 printk("%s: ", zone->name);
2795 spin_lock_irqsave(&zone->lock, flags);
2796 for (order = 0; order < MAX_ORDER; order++) {
2797 nr[order] = zone->free_area[order].nr_free;
2798 total += nr[order] << order;
2800 spin_unlock_irqrestore(&zone->lock, flags);
2801 for (order = 0; order < MAX_ORDER; order++)
2802 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2803 printk("= %lukB\n", K(total));
2806 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2808 show_swap_cache_info();
2811 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2813 zoneref->zone = zone;
2814 zoneref->zone_idx = zone_idx(zone);
2818 * Builds allocation fallback zone lists.
2820 * Add all populated zones of a node to the zonelist.
2822 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2823 int nr_zones, enum zone_type zone_type)
2827 BUG_ON(zone_type >= MAX_NR_ZONES);
2832 zone = pgdat->node_zones + zone_type;
2833 if (populated_zone(zone)) {
2834 zoneref_set_zone(zone,
2835 &zonelist->_zonerefs[nr_zones++]);
2836 check_highest_zone(zone_type);
2839 } while (zone_type);
2846 * 0 = automatic detection of better ordering.
2847 * 1 = order by ([node] distance, -zonetype)
2848 * 2 = order by (-zonetype, [node] distance)
2850 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2851 * the same zonelist. So only NUMA can configure this param.
2853 #define ZONELIST_ORDER_DEFAULT 0
2854 #define ZONELIST_ORDER_NODE 1
2855 #define ZONELIST_ORDER_ZONE 2
2857 /* zonelist order in the kernel.
2858 * set_zonelist_order() will set this to NODE or ZONE.
2860 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2861 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2865 /* The value user specified ....changed by config */
2866 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2867 /* string for sysctl */
2868 #define NUMA_ZONELIST_ORDER_LEN 16
2869 char numa_zonelist_order[16] = "default";
2872 * interface for configure zonelist ordering.
2873 * command line option "numa_zonelist_order"
2874 * = "[dD]efault - default, automatic configuration.
2875 * = "[nN]ode - order by node locality, then by zone within node
2876 * = "[zZ]one - order by zone, then by locality within zone
2879 static int __parse_numa_zonelist_order(char *s)
2881 if (*s == 'd' || *s == 'D') {
2882 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2883 } else if (*s == 'n' || *s == 'N') {
2884 user_zonelist_order = ZONELIST_ORDER_NODE;
2885 } else if (*s == 'z' || *s == 'Z') {
2886 user_zonelist_order = ZONELIST_ORDER_ZONE;
2889 "Ignoring invalid numa_zonelist_order value: "
2896 static __init int setup_numa_zonelist_order(char *s)
2903 ret = __parse_numa_zonelist_order(s);
2905 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2909 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2912 * sysctl handler for numa_zonelist_order
2914 int numa_zonelist_order_handler(ctl_table *table, int write,
2915 void __user *buffer, size_t *length,
2918 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2920 static DEFINE_MUTEX(zl_order_mutex);
2922 mutex_lock(&zl_order_mutex);
2924 strcpy(saved_string, (char*)table->data);
2925 ret = proc_dostring(table, write, buffer, length, ppos);
2929 int oldval = user_zonelist_order;
2930 if (__parse_numa_zonelist_order((char*)table->data)) {
2932 * bogus value. restore saved string
2934 strncpy((char*)table->data, saved_string,
2935 NUMA_ZONELIST_ORDER_LEN);
2936 user_zonelist_order = oldval;
2937 } else if (oldval != user_zonelist_order) {
2938 mutex_lock(&zonelists_mutex);
2939 build_all_zonelists(NULL);
2940 mutex_unlock(&zonelists_mutex);
2944 mutex_unlock(&zl_order_mutex);
2949 #define MAX_NODE_LOAD (nr_online_nodes)
2950 static int node_load[MAX_NUMNODES];
2953 * find_next_best_node - find the next node that should appear in a given node's fallback list
2954 * @node: node whose fallback list we're appending
2955 * @used_node_mask: nodemask_t of already used nodes
2957 * We use a number of factors to determine which is the next node that should
2958 * appear on a given node's fallback list. The node should not have appeared
2959 * already in @node's fallback list, and it should be the next closest node
2960 * according to the distance array (which contains arbitrary distance values
2961 * from each node to each node in the system), and should also prefer nodes
2962 * with no CPUs, since presumably they'll have very little allocation pressure
2963 * on them otherwise.
2964 * It returns -1 if no node is found.
2966 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2969 int min_val = INT_MAX;
2971 const struct cpumask *tmp = cpumask_of_node(0);
2973 /* Use the local node if we haven't already */
2974 if (!node_isset(node, *used_node_mask)) {
2975 node_set(node, *used_node_mask);
2979 for_each_node_state(n, N_HIGH_MEMORY) {
2981 /* Don't want a node to appear more than once */
2982 if (node_isset(n, *used_node_mask))
2985 /* Use the distance array to find the distance */
2986 val = node_distance(node, n);
2988 /* Penalize nodes under us ("prefer the next node") */
2991 /* Give preference to headless and unused nodes */
2992 tmp = cpumask_of_node(n);
2993 if (!cpumask_empty(tmp))
2994 val += PENALTY_FOR_NODE_WITH_CPUS;
2996 /* Slight preference for less loaded node */
2997 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2998 val += node_load[n];
3000 if (val < min_val) {
3007 node_set(best_node, *used_node_mask);
3014 * Build zonelists ordered by node and zones within node.
3015 * This results in maximum locality--normal zone overflows into local
3016 * DMA zone, if any--but risks exhausting DMA zone.
3018 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3021 struct zonelist *zonelist;
3023 zonelist = &pgdat->node_zonelists[0];
3024 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3026 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3028 zonelist->_zonerefs[j].zone = NULL;
3029 zonelist->_zonerefs[j].zone_idx = 0;
3033 * Build gfp_thisnode zonelists
3035 static void build_thisnode_zonelists(pg_data_t *pgdat)
3038 struct zonelist *zonelist;
3040 zonelist = &pgdat->node_zonelists[1];
3041 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3042 zonelist->_zonerefs[j].zone = NULL;
3043 zonelist->_zonerefs[j].zone_idx = 0;
3047 * Build zonelists ordered by zone and nodes within zones.
3048 * This results in conserving DMA zone[s] until all Normal memory is
3049 * exhausted, but results in overflowing to remote node while memory
3050 * may still exist in local DMA zone.
3052 static int node_order[MAX_NUMNODES];
3054 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3057 int zone_type; /* needs to be signed */
3059 struct zonelist *zonelist;
3061 zonelist = &pgdat->node_zonelists[0];
3063 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3064 for (j = 0; j < nr_nodes; j++) {
3065 node = node_order[j];
3066 z = &NODE_DATA(node)->node_zones[zone_type];
3067 if (populated_zone(z)) {
3069 &zonelist->_zonerefs[pos++]);
3070 check_highest_zone(zone_type);
3074 zonelist->_zonerefs[pos].zone = NULL;
3075 zonelist->_zonerefs[pos].zone_idx = 0;
3078 static int default_zonelist_order(void)
3081 unsigned long low_kmem_size,total_size;
3085 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3086 * If they are really small and used heavily, the system can fall
3087 * into OOM very easily.
3088 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3090 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3093 for_each_online_node(nid) {
3094 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3095 z = &NODE_DATA(nid)->node_zones[zone_type];
3096 if (populated_zone(z)) {
3097 if (zone_type < ZONE_NORMAL)
3098 low_kmem_size += z->present_pages;
3099 total_size += z->present_pages;
3100 } else if (zone_type == ZONE_NORMAL) {
3102 * If any node has only lowmem, then node order
3103 * is preferred to allow kernel allocations
3104 * locally; otherwise, they can easily infringe
3105 * on other nodes when there is an abundance of
3106 * lowmem available to allocate from.
3108 return ZONELIST_ORDER_NODE;
3112 if (!low_kmem_size || /* there are no DMA area. */
3113 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3114 return ZONELIST_ORDER_NODE;
3116 * look into each node's config.
3117 * If there is a node whose DMA/DMA32 memory is very big area on
3118 * local memory, NODE_ORDER may be suitable.
3120 average_size = total_size /
3121 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3122 for_each_online_node(nid) {
3125 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3126 z = &NODE_DATA(nid)->node_zones[zone_type];
3127 if (populated_zone(z)) {
3128 if (zone_type < ZONE_NORMAL)
3129 low_kmem_size += z->present_pages;
3130 total_size += z->present_pages;
3133 if (low_kmem_size &&
3134 total_size > average_size && /* ignore small node */
3135 low_kmem_size > total_size * 70/100)
3136 return ZONELIST_ORDER_NODE;
3138 return ZONELIST_ORDER_ZONE;
3141 static void set_zonelist_order(void)
3143 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3144 current_zonelist_order = default_zonelist_order();
3146 current_zonelist_order = user_zonelist_order;
3149 static void build_zonelists(pg_data_t *pgdat)
3153 nodemask_t used_mask;
3154 int local_node, prev_node;
3155 struct zonelist *zonelist;
3156 int order = current_zonelist_order;
3158 /* initialize zonelists */
3159 for (i = 0; i < MAX_ZONELISTS; i++) {
3160 zonelist = pgdat->node_zonelists + i;
3161 zonelist->_zonerefs[0].zone = NULL;
3162 zonelist->_zonerefs[0].zone_idx = 0;
3165 /* NUMA-aware ordering of nodes */
3166 local_node = pgdat->node_id;
3167 load = nr_online_nodes;
3168 prev_node = local_node;
3169 nodes_clear(used_mask);
3171 memset(node_order, 0, sizeof(node_order));
3174 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3175 int distance = node_distance(local_node, node);
3178 * If another node is sufficiently far away then it is better
3179 * to reclaim pages in a zone before going off node.
3181 if (distance > RECLAIM_DISTANCE)
3182 zone_reclaim_mode = 1;
3185 * We don't want to pressure a particular node.
3186 * So adding penalty to the first node in same
3187 * distance group to make it round-robin.
3189 if (distance != node_distance(local_node, prev_node))
3190 node_load[node] = load;
3194 if (order == ZONELIST_ORDER_NODE)
3195 build_zonelists_in_node_order(pgdat, node);
3197 node_order[j++] = node; /* remember order */
3200 if (order == ZONELIST_ORDER_ZONE) {
3201 /* calculate node order -- i.e., DMA last! */
3202 build_zonelists_in_zone_order(pgdat, j);
3205 build_thisnode_zonelists(pgdat);
3208 /* Construct the zonelist performance cache - see further mmzone.h */
3209 static void build_zonelist_cache(pg_data_t *pgdat)
3211 struct zonelist *zonelist;
3212 struct zonelist_cache *zlc;
3215 zonelist = &pgdat->node_zonelists[0];
3216 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3217 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3218 for (z = zonelist->_zonerefs; z->zone; z++)
3219 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3222 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3224 * Return node id of node used for "local" allocations.
3225 * I.e., first node id of first zone in arg node's generic zonelist.
3226 * Used for initializing percpu 'numa_mem', which is used primarily
3227 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3229 int local_memory_node(int node)
3233 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3234 gfp_zone(GFP_KERNEL),
3241 #else /* CONFIG_NUMA */
3243 static void set_zonelist_order(void)
3245 current_zonelist_order = ZONELIST_ORDER_ZONE;
3248 static void build_zonelists(pg_data_t *pgdat)
3250 int node, local_node;
3252 struct zonelist *zonelist;
3254 local_node = pgdat->node_id;
3256 zonelist = &pgdat->node_zonelists[0];
3257 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3260 * Now we build the zonelist so that it contains the zones
3261 * of all the other nodes.
3262 * We don't want to pressure a particular node, so when
3263 * building the zones for node N, we make sure that the
3264 * zones coming right after the local ones are those from
3265 * node N+1 (modulo N)
3267 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3268 if (!node_online(node))
3270 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3273 for (node = 0; node < local_node; node++) {
3274 if (!node_online(node))
3276 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3280 zonelist->_zonerefs[j].zone = NULL;
3281 zonelist->_zonerefs[j].zone_idx = 0;
3284 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3285 static void build_zonelist_cache(pg_data_t *pgdat)
3287 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3290 #endif /* CONFIG_NUMA */
3293 * Boot pageset table. One per cpu which is going to be used for all
3294 * zones and all nodes. The parameters will be set in such a way
3295 * that an item put on a list will immediately be handed over to
3296 * the buddy list. This is safe since pageset manipulation is done
3297 * with interrupts disabled.
3299 * The boot_pagesets must be kept even after bootup is complete for
3300 * unused processors and/or zones. They do play a role for bootstrapping
3301 * hotplugged processors.
3303 * zoneinfo_show() and maybe other functions do
3304 * not check if the processor is online before following the pageset pointer.
3305 * Other parts of the kernel may not check if the zone is available.
3307 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3308 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3309 static void setup_zone_pageset(struct zone *zone);
3312 * Global mutex to protect against size modification of zonelists
3313 * as well as to serialize pageset setup for the new populated zone.
3315 DEFINE_MUTEX(zonelists_mutex);
3317 /* return values int ....just for stop_machine() */
3318 static __init_refok int __build_all_zonelists(void *data)
3324 memset(node_load, 0, sizeof(node_load));
3326 for_each_online_node(nid) {
3327 pg_data_t *pgdat = NODE_DATA(nid);
3329 build_zonelists(pgdat);
3330 build_zonelist_cache(pgdat);
3334 * Initialize the boot_pagesets that are going to be used
3335 * for bootstrapping processors. The real pagesets for
3336 * each zone will be allocated later when the per cpu
3337 * allocator is available.
3339 * boot_pagesets are used also for bootstrapping offline
3340 * cpus if the system is already booted because the pagesets
3341 * are needed to initialize allocators on a specific cpu too.
3342 * F.e. the percpu allocator needs the page allocator which
3343 * needs the percpu allocator in order to allocate its pagesets
3344 * (a chicken-egg dilemma).
3346 for_each_possible_cpu(cpu) {
3347 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3349 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3351 * We now know the "local memory node" for each node--
3352 * i.e., the node of the first zone in the generic zonelist.
3353 * Set up numa_mem percpu variable for on-line cpus. During
3354 * boot, only the boot cpu should be on-line; we'll init the
3355 * secondary cpus' numa_mem as they come on-line. During
3356 * node/memory hotplug, we'll fixup all on-line cpus.
3358 if (cpu_online(cpu))
3359 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3367 * Called with zonelists_mutex held always
3368 * unless system_state == SYSTEM_BOOTING.
3370 void __ref build_all_zonelists(void *data)
3372 set_zonelist_order();
3374 if (system_state == SYSTEM_BOOTING) {
3375 __build_all_zonelists(NULL);
3376 mminit_verify_zonelist();
3377 cpuset_init_current_mems_allowed();
3379 /* we have to stop all cpus to guarantee there is no user
3381 #ifdef CONFIG_MEMORY_HOTPLUG
3383 setup_zone_pageset((struct zone *)data);
3385 stop_machine(__build_all_zonelists, NULL, NULL);
3386 /* cpuset refresh routine should be here */
3388 vm_total_pages = nr_free_pagecache_pages();
3390 * Disable grouping by mobility if the number of pages in the
3391 * system is too low to allow the mechanism to work. It would be
3392 * more accurate, but expensive to check per-zone. This check is
3393 * made on memory-hotadd so a system can start with mobility
3394 * disabled and enable it later
3396 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3397 page_group_by_mobility_disabled = 1;
3399 page_group_by_mobility_disabled = 0;
3401 printk("Built %i zonelists in %s order, mobility grouping %s. "
3402 "Total pages: %ld\n",
3404 zonelist_order_name[current_zonelist_order],
3405 page_group_by_mobility_disabled ? "off" : "on",
3408 printk("Policy zone: %s\n", zone_names[policy_zone]);
3413 * Helper functions to size the waitqueue hash table.
3414 * Essentially these want to choose hash table sizes sufficiently
3415 * large so that collisions trying to wait on pages are rare.
3416 * But in fact, the number of active page waitqueues on typical
3417 * systems is ridiculously low, less than 200. So this is even
3418 * conservative, even though it seems large.
3420 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3421 * waitqueues, i.e. the size of the waitq table given the number of pages.
3423 #define PAGES_PER_WAITQUEUE 256
3425 #ifndef CONFIG_MEMORY_HOTPLUG
3426 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3428 unsigned long size = 1;
3430 pages /= PAGES_PER_WAITQUEUE;
3432 while (size < pages)
3436 * Once we have dozens or even hundreds of threads sleeping
3437 * on IO we've got bigger problems than wait queue collision.
3438 * Limit the size of the wait table to a reasonable size.
3440 size = min(size, 4096UL);
3442 return max(size, 4UL);
3446 * A zone's size might be changed by hot-add, so it is not possible to determine
3447 * a suitable size for its wait_table. So we use the maximum size now.
3449 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3451 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3452 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3453 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3455 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3456 * or more by the traditional way. (See above). It equals:
3458 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3459 * ia64(16K page size) : = ( 8G + 4M)byte.
3460 * powerpc (64K page size) : = (32G +16M)byte.
3462 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3469 * This is an integer logarithm so that shifts can be used later
3470 * to extract the more random high bits from the multiplicative
3471 * hash function before the remainder is taken.
3473 static inline unsigned long wait_table_bits(unsigned long size)
3478 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3481 * Check if a pageblock contains reserved pages
3483 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3487 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3488 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3495 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3496 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3497 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3498 * higher will lead to a bigger reserve which will get freed as contiguous
3499 * blocks as reclaim kicks in
3501 static void setup_zone_migrate_reserve(struct zone *zone)
3503 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3505 unsigned long block_migratetype;
3509 * Get the start pfn, end pfn and the number of blocks to reserve
3510 * We have to be careful to be aligned to pageblock_nr_pages to
3511 * make sure that we always check pfn_valid for the first page in
3514 start_pfn = zone->zone_start_pfn;
3515 end_pfn = start_pfn + zone->spanned_pages;
3516 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3517 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3521 * Reserve blocks are generally in place to help high-order atomic
3522 * allocations that are short-lived. A min_free_kbytes value that
3523 * would result in more than 2 reserve blocks for atomic allocations
3524 * is assumed to be in place to help anti-fragmentation for the
3525 * future allocation of hugepages at runtime.
3527 reserve = min(2, reserve);
3529 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3530 if (!pfn_valid(pfn))
3532 page = pfn_to_page(pfn);
3534 /* Watch out for overlapping nodes */
3535 if (page_to_nid(page) != zone_to_nid(zone))
3538 block_migratetype = get_pageblock_migratetype(page);
3540 /* Only test what is necessary when the reserves are not met */
3543 * Blocks with reserved pages will never free, skip
3546 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3547 if (pageblock_is_reserved(pfn, block_end_pfn))
3550 /* If this block is reserved, account for it */
3551 if (block_migratetype == MIGRATE_RESERVE) {
3556 /* Suitable for reserving if this block is movable */
3557 if (block_migratetype == MIGRATE_MOVABLE) {
3558 set_pageblock_migratetype(page,
3560 move_freepages_block(zone, page,
3568 * If the reserve is met and this is a previous reserved block,
3571 if (block_migratetype == MIGRATE_RESERVE) {
3572 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3573 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3579 * Initially all pages are reserved - free ones are freed
3580 * up by free_all_bootmem() once the early boot process is
3581 * done. Non-atomic initialization, single-pass.
3583 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3584 unsigned long start_pfn, enum memmap_context context)
3587 unsigned long end_pfn = start_pfn + size;
3591 if (highest_memmap_pfn < end_pfn - 1)
3592 highest_memmap_pfn = end_pfn - 1;
3594 z = &NODE_DATA(nid)->node_zones[zone];
3595 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3597 * There can be holes in boot-time mem_map[]s
3598 * handed to this function. They do not
3599 * exist on hotplugged memory.
3601 if (context == MEMMAP_EARLY) {
3602 if (!early_pfn_valid(pfn))
3604 if (!early_pfn_in_nid(pfn, nid))
3607 page = pfn_to_page(pfn);
3608 set_page_links(page, zone, nid, pfn);
3609 mminit_verify_page_links(page, zone, nid, pfn);
3610 init_page_count(page);
3611 reset_page_mapcount(page);
3612 SetPageReserved(page);
3614 * Mark the block movable so that blocks are reserved for
3615 * movable at startup. This will force kernel allocations
3616 * to reserve their blocks rather than leaking throughout
3617 * the address space during boot when many long-lived
3618 * kernel allocations are made. Later some blocks near
3619 * the start are marked MIGRATE_RESERVE by
3620 * setup_zone_migrate_reserve()
3622 * bitmap is created for zone's valid pfn range. but memmap
3623 * can be created for invalid pages (for alignment)
3624 * check here not to call set_pageblock_migratetype() against
3627 if ((z->zone_start_pfn <= pfn)
3628 && (pfn < z->zone_start_pfn + z->spanned_pages)
3629 && !(pfn & (pageblock_nr_pages - 1)))
3630 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3632 INIT_LIST_HEAD(&page->lru);
3633 #ifdef WANT_PAGE_VIRTUAL
3634 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3635 if (!is_highmem_idx(zone))
3636 set_page_address(page, __va(pfn << PAGE_SHIFT));
3641 static void __meminit zone_init_free_lists(struct zone *zone)
3644 for_each_migratetype_order(order, t) {
3645 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3646 zone->free_area[order].nr_free = 0;
3650 #ifndef __HAVE_ARCH_MEMMAP_INIT
3651 #define memmap_init(size, nid, zone, start_pfn) \
3652 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3655 static int zone_batchsize(struct zone *zone)
3661 * The per-cpu-pages pools are set to around 1000th of the
3662 * size of the zone. But no more than 1/2 of a meg.
3664 * OK, so we don't know how big the cache is. So guess.
3666 batch = zone->present_pages / 1024;
3667 if (batch * PAGE_SIZE > 512 * 1024)
3668 batch = (512 * 1024) / PAGE_SIZE;
3669 batch /= 4; /* We effectively *= 4 below */
3674 * Clamp the batch to a 2^n - 1 value. Having a power
3675 * of 2 value was found to be more likely to have
3676 * suboptimal cache aliasing properties in some cases.
3678 * For example if 2 tasks are alternately allocating
3679 * batches of pages, one task can end up with a lot
3680 * of pages of one half of the possible page colors
3681 * and the other with pages of the other colors.
3683 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3688 /* The deferral and batching of frees should be suppressed under NOMMU
3691 * The problem is that NOMMU needs to be able to allocate large chunks
3692 * of contiguous memory as there's no hardware page translation to
3693 * assemble apparent contiguous memory from discontiguous pages.
3695 * Queueing large contiguous runs of pages for batching, however,
3696 * causes the pages to actually be freed in smaller chunks. As there
3697 * can be a significant delay between the individual batches being
3698 * recycled, this leads to the once large chunks of space being
3699 * fragmented and becoming unavailable for high-order allocations.
3705 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3707 struct per_cpu_pages *pcp;
3710 memset(p, 0, sizeof(*p));
3714 pcp->high = 6 * batch;
3715 pcp->batch = max(1UL, 1 * batch);
3716 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3717 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3721 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3722 * to the value high for the pageset p.
3725 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3728 struct per_cpu_pages *pcp;
3732 pcp->batch = max(1UL, high/4);
3733 if ((high/4) > (PAGE_SHIFT * 8))
3734 pcp->batch = PAGE_SHIFT * 8;
3737 static void setup_zone_pageset(struct zone *zone)
3741 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3743 for_each_possible_cpu(cpu) {
3744 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3746 setup_pageset(pcp, zone_batchsize(zone));
3748 if (percpu_pagelist_fraction)
3749 setup_pagelist_highmark(pcp,
3750 (zone->present_pages /
3751 percpu_pagelist_fraction));
3756 * Allocate per cpu pagesets and initialize them.
3757 * Before this call only boot pagesets were available.
3759 void __init setup_per_cpu_pageset(void)
3763 for_each_populated_zone(zone)
3764 setup_zone_pageset(zone);
3767 static noinline __init_refok
3768 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3771 struct pglist_data *pgdat = zone->zone_pgdat;
3775 * The per-page waitqueue mechanism uses hashed waitqueues
3778 zone->wait_table_hash_nr_entries =
3779 wait_table_hash_nr_entries(zone_size_pages);
3780 zone->wait_table_bits =
3781 wait_table_bits(zone->wait_table_hash_nr_entries);
3782 alloc_size = zone->wait_table_hash_nr_entries
3783 * sizeof(wait_queue_head_t);
3785 if (!slab_is_available()) {
3786 zone->wait_table = (wait_queue_head_t *)
3787 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3790 * This case means that a zone whose size was 0 gets new memory
3791 * via memory hot-add.
3792 * But it may be the case that a new node was hot-added. In
3793 * this case vmalloc() will not be able to use this new node's
3794 * memory - this wait_table must be initialized to use this new
3795 * node itself as well.
3796 * To use this new node's memory, further consideration will be
3799 zone->wait_table = vmalloc(alloc_size);
3801 if (!zone->wait_table)
3804 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3805 init_waitqueue_head(zone->wait_table + i);
3810 static int __zone_pcp_update(void *data)
3812 struct zone *zone = data;
3814 unsigned long batch = zone_batchsize(zone), flags;
3816 for_each_possible_cpu(cpu) {
3817 struct per_cpu_pageset *pset;
3818 struct per_cpu_pages *pcp;
3820 pset = per_cpu_ptr(zone->pageset, cpu);
3823 local_irq_save(flags);
3824 free_pcppages_bulk(zone, pcp->count, pcp);
3825 setup_pageset(pset, batch);
3826 local_irq_restore(flags);
3831 void zone_pcp_update(struct zone *zone)
3833 stop_machine(__zone_pcp_update, zone, NULL);
3836 static __meminit void zone_pcp_init(struct zone *zone)
3839 * per cpu subsystem is not up at this point. The following code
3840 * relies on the ability of the linker to provide the
3841 * offset of a (static) per cpu variable into the per cpu area.
3843 zone->pageset = &boot_pageset;
3845 if (zone->present_pages)
3846 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3847 zone->name, zone->present_pages,
3848 zone_batchsize(zone));
3851 __meminit int init_currently_empty_zone(struct zone *zone,
3852 unsigned long zone_start_pfn,
3854 enum memmap_context context)
3856 struct pglist_data *pgdat = zone->zone_pgdat;
3858 ret = zone_wait_table_init(zone, size);
3861 pgdat->nr_zones = zone_idx(zone) + 1;
3863 zone->zone_start_pfn = zone_start_pfn;
3865 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3866 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3868 (unsigned long)zone_idx(zone),
3869 zone_start_pfn, (zone_start_pfn + size));
3871 zone_init_free_lists(zone);
3876 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3878 * Basic iterator support. Return the first range of PFNs for a node
3879 * Note: nid == MAX_NUMNODES returns first region regardless of node
3881 static int __meminit first_active_region_index_in_nid(int nid)
3885 for (i = 0; i < nr_nodemap_entries; i++)
3886 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3893 * Basic iterator support. Return the next active range of PFNs for a node
3894 * Note: nid == MAX_NUMNODES returns next region regardless of node
3896 static int __meminit next_active_region_index_in_nid(int index, int nid)
3898 for (index = index + 1; index < nr_nodemap_entries; index++)
3899 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3905 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3907 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3908 * Architectures may implement their own version but if add_active_range()
3909 * was used and there are no special requirements, this is a convenient
3912 int __meminit __early_pfn_to_nid(unsigned long pfn)
3916 for (i = 0; i < nr_nodemap_entries; i++) {
3917 unsigned long start_pfn = early_node_map[i].start_pfn;
3918 unsigned long end_pfn = early_node_map[i].end_pfn;
3920 if (start_pfn <= pfn && pfn < end_pfn)
3921 return early_node_map[i].nid;
3923 /* This is a memory hole */
3926 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3928 int __meminit early_pfn_to_nid(unsigned long pfn)
3932 nid = __early_pfn_to_nid(pfn);
3935 /* just returns 0 */
3939 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3940 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3944 nid = __early_pfn_to_nid(pfn);
3945 if (nid >= 0 && nid != node)
3951 /* Basic iterator support to walk early_node_map[] */
3952 #define for_each_active_range_index_in_nid(i, nid) \
3953 for (i = first_active_region_index_in_nid(nid); i != -1; \
3954 i = next_active_region_index_in_nid(i, nid))
3957 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3958 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3959 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3961 * If an architecture guarantees that all ranges registered with
3962 * add_active_ranges() contain no holes and may be freed, this
3963 * this function may be used instead of calling free_bootmem() manually.
3965 void __init free_bootmem_with_active_regions(int nid,
3966 unsigned long max_low_pfn)
3970 for_each_active_range_index_in_nid(i, nid) {
3971 unsigned long size_pages = 0;
3972 unsigned long end_pfn = early_node_map[i].end_pfn;
3974 if (early_node_map[i].start_pfn >= max_low_pfn)
3977 if (end_pfn > max_low_pfn)
3978 end_pfn = max_low_pfn;
3980 size_pages = end_pfn - early_node_map[i].start_pfn;
3981 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3982 PFN_PHYS(early_node_map[i].start_pfn),
3983 size_pages << PAGE_SHIFT);
3987 #ifdef CONFIG_HAVE_MEMBLOCK
3989 * Basic iterator support. Return the last range of PFNs for a node
3990 * Note: nid == MAX_NUMNODES returns last region regardless of node
3992 static int __meminit last_active_region_index_in_nid(int nid)
3996 for (i = nr_nodemap_entries - 1; i >= 0; i--)
3997 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
4004 * Basic iterator support. Return the previous active range of PFNs for a node
4005 * Note: nid == MAX_NUMNODES returns next region regardless of node
4007 static int __meminit previous_active_region_index_in_nid(int index, int nid)
4009 for (index = index - 1; index >= 0; index--)
4010 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
4016 #define for_each_active_range_index_in_nid_reverse(i, nid) \
4017 for (i = last_active_region_index_in_nid(nid); i != -1; \
4018 i = previous_active_region_index_in_nid(i, nid))
4020 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
4021 u64 goal, u64 limit)
4025 /* Need to go over early_node_map to find out good range for node */
4026 for_each_active_range_index_in_nid_reverse(i, nid) {
4028 u64 ei_start, ei_last;
4029 u64 final_start, final_end;
4031 ei_last = early_node_map[i].end_pfn;
4032 ei_last <<= PAGE_SHIFT;
4033 ei_start = early_node_map[i].start_pfn;
4034 ei_start <<= PAGE_SHIFT;
4036 final_start = max(ei_start, goal);
4037 final_end = min(ei_last, limit);
4039 if (final_start >= final_end)
4042 addr = memblock_find_in_range(final_start, final_end, size, align);
4044 if (addr == MEMBLOCK_ERROR)
4050 return MEMBLOCK_ERROR;
4054 int __init add_from_early_node_map(struct range *range, int az,
4055 int nr_range, int nid)
4060 /* need to go over early_node_map to find out good range for node */
4061 for_each_active_range_index_in_nid(i, nid) {
4062 start = early_node_map[i].start_pfn;
4063 end = early_node_map[i].end_pfn;
4064 nr_range = add_range(range, az, nr_range, start, end);
4069 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
4074 for_each_active_range_index_in_nid(i, nid) {
4075 ret = work_fn(early_node_map[i].start_pfn,
4076 early_node_map[i].end_pfn, data);
4082 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4083 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4085 * If an architecture guarantees that all ranges registered with
4086 * add_active_ranges() contain no holes and may be freed, this
4087 * function may be used instead of calling memory_present() manually.
4089 void __init sparse_memory_present_with_active_regions(int nid)
4093 for_each_active_range_index_in_nid(i, nid)
4094 memory_present(early_node_map[i].nid,
4095 early_node_map[i].start_pfn,
4096 early_node_map[i].end_pfn);
4100 * get_pfn_range_for_nid - Return the start and end page frames for a node
4101 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4102 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4103 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4105 * It returns the start and end page frame of a node based on information
4106 * provided by an arch calling add_active_range(). If called for a node
4107 * with no available memory, a warning is printed and the start and end
4110 void __meminit get_pfn_range_for_nid(unsigned int nid,
4111 unsigned long *start_pfn, unsigned long *end_pfn)
4117 for_each_active_range_index_in_nid(i, nid) {
4118 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
4119 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
4122 if (*start_pfn == -1UL)
4127 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4128 * assumption is made that zones within a node are ordered in monotonic
4129 * increasing memory addresses so that the "highest" populated zone is used
4131 static void __init find_usable_zone_for_movable(void)
4134 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4135 if (zone_index == ZONE_MOVABLE)
4138 if (arch_zone_highest_possible_pfn[zone_index] >
4139 arch_zone_lowest_possible_pfn[zone_index])
4143 VM_BUG_ON(zone_index == -1);
4144 movable_zone = zone_index;
4148 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4149 * because it is sized independent of architecture. Unlike the other zones,
4150 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4151 * in each node depending on the size of each node and how evenly kernelcore
4152 * is distributed. This helper function adjusts the zone ranges
4153 * provided by the architecture for a given node by using the end of the
4154 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4155 * zones within a node are in order of monotonic increases memory addresses
4157 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4158 unsigned long zone_type,
4159 unsigned long node_start_pfn,
4160 unsigned long node_end_pfn,
4161 unsigned long *zone_start_pfn,
4162 unsigned long *zone_end_pfn)
4164 /* Only adjust if ZONE_MOVABLE is on this node */
4165 if (zone_movable_pfn[nid]) {
4166 /* Size ZONE_MOVABLE */
4167 if (zone_type == ZONE_MOVABLE) {
4168 *zone_start_pfn = zone_movable_pfn[nid];
4169 *zone_end_pfn = min(node_end_pfn,
4170 arch_zone_highest_possible_pfn[movable_zone]);
4172 /* Adjust for ZONE_MOVABLE starting within this range */
4173 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4174 *zone_end_pfn > zone_movable_pfn[nid]) {
4175 *zone_end_pfn = zone_movable_pfn[nid];
4177 /* Check if this whole range is within ZONE_MOVABLE */
4178 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4179 *zone_start_pfn = *zone_end_pfn;
4184 * Return the number of pages a zone spans in a node, including holes
4185 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4187 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4188 unsigned long zone_type,
4189 unsigned long *ignored)
4191 unsigned long node_start_pfn, node_end_pfn;
4192 unsigned long zone_start_pfn, zone_end_pfn;
4194 /* Get the start and end of the node and zone */
4195 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4196 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4197 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4198 adjust_zone_range_for_zone_movable(nid, zone_type,
4199 node_start_pfn, node_end_pfn,
4200 &zone_start_pfn, &zone_end_pfn);
4202 /* Check that this node has pages within the zone's required range */
4203 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4206 /* Move the zone boundaries inside the node if necessary */
4207 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4208 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4210 /* Return the spanned pages */
4211 return zone_end_pfn - zone_start_pfn;
4215 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4216 * then all holes in the requested range will be accounted for.
4218 unsigned long __meminit __absent_pages_in_range(int nid,
4219 unsigned long range_start_pfn,
4220 unsigned long range_end_pfn)
4223 unsigned long prev_end_pfn = 0, hole_pages = 0;
4224 unsigned long start_pfn;
4226 /* Find the end_pfn of the first active range of pfns in the node */
4227 i = first_active_region_index_in_nid(nid);
4231 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4233 /* Account for ranges before physical memory on this node */
4234 if (early_node_map[i].start_pfn > range_start_pfn)
4235 hole_pages = prev_end_pfn - range_start_pfn;
4237 /* Find all holes for the zone within the node */
4238 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4240 /* No need to continue if prev_end_pfn is outside the zone */
4241 if (prev_end_pfn >= range_end_pfn)
4244 /* Make sure the end of the zone is not within the hole */
4245 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4246 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4248 /* Update the hole size cound and move on */
4249 if (start_pfn > range_start_pfn) {
4250 BUG_ON(prev_end_pfn > start_pfn);
4251 hole_pages += start_pfn - prev_end_pfn;
4253 prev_end_pfn = early_node_map[i].end_pfn;
4256 /* Account for ranges past physical memory on this node */
4257 if (range_end_pfn > prev_end_pfn)
4258 hole_pages += range_end_pfn -
4259 max(range_start_pfn, prev_end_pfn);
4265 * absent_pages_in_range - Return number of page frames in holes within a range
4266 * @start_pfn: The start PFN to start searching for holes
4267 * @end_pfn: The end PFN to stop searching for holes
4269 * It returns the number of pages frames in memory holes within a range.
4271 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4272 unsigned long end_pfn)
4274 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4277 /* Return the number of page frames in holes in a zone on a node */
4278 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4279 unsigned long zone_type,
4280 unsigned long *ignored)
4282 unsigned long node_start_pfn, node_end_pfn;
4283 unsigned long zone_start_pfn, zone_end_pfn;
4285 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4286 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4288 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4291 adjust_zone_range_for_zone_movable(nid, zone_type,
4292 node_start_pfn, node_end_pfn,
4293 &zone_start_pfn, &zone_end_pfn);
4294 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4298 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4299 unsigned long zone_type,
4300 unsigned long *zones_size)
4302 return zones_size[zone_type];
4305 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4306 unsigned long zone_type,
4307 unsigned long *zholes_size)
4312 return zholes_size[zone_type];
4317 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4318 unsigned long *zones_size, unsigned long *zholes_size)
4320 unsigned long realtotalpages, totalpages = 0;
4323 for (i = 0; i < MAX_NR_ZONES; i++)
4324 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4326 pgdat->node_spanned_pages = totalpages;
4328 realtotalpages = totalpages;
4329 for (i = 0; i < MAX_NR_ZONES; i++)
4331 zone_absent_pages_in_node(pgdat->node_id, i,
4333 pgdat->node_present_pages = realtotalpages;
4334 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4338 #ifndef CONFIG_SPARSEMEM
4340 * Calculate the size of the zone->blockflags rounded to an unsigned long
4341 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4342 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4343 * round what is now in bits to nearest long in bits, then return it in
4346 static unsigned long __init usemap_size(unsigned long zonesize)
4348 unsigned long usemapsize;
4350 usemapsize = roundup(zonesize, pageblock_nr_pages);
4351 usemapsize = usemapsize >> pageblock_order;
4352 usemapsize *= NR_PAGEBLOCK_BITS;
4353 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4355 return usemapsize / 8;
4358 static void __init setup_usemap(struct pglist_data *pgdat,
4359 struct zone *zone, unsigned long zonesize)
4361 unsigned long usemapsize = usemap_size(zonesize);
4362 zone->pageblock_flags = NULL;
4364 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4368 static inline void setup_usemap(struct pglist_data *pgdat,
4369 struct zone *zone, unsigned long zonesize) {}
4370 #endif /* CONFIG_SPARSEMEM */
4372 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4374 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4375 void __init set_pageblock_order(void)
4379 /* Check that pageblock_nr_pages has not already been setup */
4380 if (pageblock_order)
4383 if (HPAGE_SHIFT > PAGE_SHIFT)
4384 order = HUGETLB_PAGE_ORDER;
4386 order = MAX_ORDER - 1;
4389 * Assume the largest contiguous order of interest is a huge page.
4390 * This value may be variable depending on boot parameters on IA64 and
4393 pageblock_order = order;
4395 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4398 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4399 * is unused as pageblock_order is set at compile-time. See
4400 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4403 void __init set_pageblock_order(void)
4407 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4410 * Set up the zone data structures:
4411 * - mark all pages reserved
4412 * - mark all memory queues empty
4413 * - clear the memory bitmaps
4415 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4416 unsigned long *zones_size, unsigned long *zholes_size)
4419 int nid = pgdat->node_id;
4420 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4423 pgdat_resize_init(pgdat);
4424 pgdat->nr_zones = 0;
4425 init_waitqueue_head(&pgdat->kswapd_wait);
4426 pgdat->kswapd_max_order = 0;
4427 pgdat_page_cgroup_init(pgdat);
4429 for (j = 0; j < MAX_NR_ZONES; j++) {
4430 struct zone *zone = pgdat->node_zones + j;
4431 unsigned long size, realsize, memmap_pages;
4434 size = zone_spanned_pages_in_node(nid, j, zones_size);
4435 realsize = size - zone_absent_pages_in_node(nid, j,
4439 * Adjust realsize so that it accounts for how much memory
4440 * is used by this zone for memmap. This affects the watermark
4441 * and per-cpu initialisations
4444 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4445 if (realsize >= memmap_pages) {
4446 realsize -= memmap_pages;
4449 " %s zone: %lu pages used for memmap\n",
4450 zone_names[j], memmap_pages);
4453 " %s zone: %lu pages exceeds realsize %lu\n",
4454 zone_names[j], memmap_pages, realsize);
4456 /* Account for reserved pages */
4457 if (j == 0 && realsize > dma_reserve) {
4458 realsize -= dma_reserve;
4459 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4460 zone_names[0], dma_reserve);
4463 if (!is_highmem_idx(j))
4464 nr_kernel_pages += realsize;
4465 nr_all_pages += realsize;
4467 zone->spanned_pages = size;
4468 zone->present_pages = realsize;
4471 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4473 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4475 zone->name = zone_names[j];
4476 spin_lock_init(&zone->lock);
4477 spin_lock_init(&zone->lru_lock);
4478 zone_seqlock_init(zone);
4479 zone->zone_pgdat = pgdat;
4481 zone_pcp_init(zone);
4483 INIT_LIST_HEAD(&zone->lru[l].list);
4484 zone->reclaim_stat.recent_rotated[0] = 0;
4485 zone->reclaim_stat.recent_rotated[1] = 0;
4486 zone->reclaim_stat.recent_scanned[0] = 0;
4487 zone->reclaim_stat.recent_scanned[1] = 0;
4488 zap_zone_vm_stats(zone);
4493 set_pageblock_order();
4494 setup_usemap(pgdat, zone, size);
4495 ret = init_currently_empty_zone(zone, zone_start_pfn,
4496 size, MEMMAP_EARLY);
4498 memmap_init(size, nid, j, zone_start_pfn);
4499 zone_start_pfn += size;
4503 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4505 /* Skip empty nodes */
4506 if (!pgdat->node_spanned_pages)
4509 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4510 /* ia64 gets its own node_mem_map, before this, without bootmem */
4511 if (!pgdat->node_mem_map) {
4512 unsigned long size, start, end;
4516 * The zone's endpoints aren't required to be MAX_ORDER
4517 * aligned but the node_mem_map endpoints must be in order
4518 * for the buddy allocator to function correctly.
4520 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4521 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4522 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4523 size = (end - start) * sizeof(struct page);
4524 map = alloc_remap(pgdat->node_id, size);
4526 map = alloc_bootmem_node_nopanic(pgdat, size);
4527 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4529 #ifndef CONFIG_NEED_MULTIPLE_NODES
4531 * With no DISCONTIG, the global mem_map is just set as node 0's
4533 if (pgdat == NODE_DATA(0)) {
4534 mem_map = NODE_DATA(0)->node_mem_map;
4535 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4536 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4537 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4538 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4541 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4544 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4545 unsigned long node_start_pfn, unsigned long *zholes_size)
4547 pg_data_t *pgdat = NODE_DATA(nid);
4549 pgdat->node_id = nid;
4550 pgdat->node_start_pfn = node_start_pfn;
4551 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4553 alloc_node_mem_map(pgdat);
4554 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4555 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4556 nid, (unsigned long)pgdat,
4557 (unsigned long)pgdat->node_mem_map);
4560 free_area_init_core(pgdat, zones_size, zholes_size);
4563 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4565 #if MAX_NUMNODES > 1
4567 * Figure out the number of possible node ids.
4569 static void __init setup_nr_node_ids(void)
4572 unsigned int highest = 0;
4574 for_each_node_mask(node, node_possible_map)
4576 nr_node_ids = highest + 1;
4579 static inline void setup_nr_node_ids(void)
4585 * add_active_range - Register a range of PFNs backed by physical memory
4586 * @nid: The node ID the range resides on
4587 * @start_pfn: The start PFN of the available physical memory
4588 * @end_pfn: The end PFN of the available physical memory
4590 * These ranges are stored in an early_node_map[] and later used by
4591 * free_area_init_nodes() to calculate zone sizes and holes. If the
4592 * range spans a memory hole, it is up to the architecture to ensure
4593 * the memory is not freed by the bootmem allocator. If possible
4594 * the range being registered will be merged with existing ranges.
4596 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4597 unsigned long end_pfn)
4601 mminit_dprintk(MMINIT_TRACE, "memory_register",
4602 "Entering add_active_range(%d, %#lx, %#lx) "
4603 "%d entries of %d used\n",
4604 nid, start_pfn, end_pfn,
4605 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4607 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4609 /* Merge with existing active regions if possible */
4610 for (i = 0; i < nr_nodemap_entries; i++) {
4611 if (early_node_map[i].nid != nid)
4614 /* Skip if an existing region covers this new one */
4615 if (start_pfn >= early_node_map[i].start_pfn &&
4616 end_pfn <= early_node_map[i].end_pfn)
4619 /* Merge forward if suitable */
4620 if (start_pfn <= early_node_map[i].end_pfn &&
4621 end_pfn > early_node_map[i].end_pfn) {
4622 early_node_map[i].end_pfn = end_pfn;
4626 /* Merge backward if suitable */
4627 if (start_pfn < early_node_map[i].start_pfn &&
4628 end_pfn >= early_node_map[i].start_pfn) {
4629 early_node_map[i].start_pfn = start_pfn;
4634 /* Check that early_node_map is large enough */
4635 if (i >= MAX_ACTIVE_REGIONS) {
4636 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4637 MAX_ACTIVE_REGIONS);
4641 early_node_map[i].nid = nid;
4642 early_node_map[i].start_pfn = start_pfn;
4643 early_node_map[i].end_pfn = end_pfn;
4644 nr_nodemap_entries = i + 1;
4648 * remove_active_range - Shrink an existing registered range of PFNs
4649 * @nid: The node id the range is on that should be shrunk
4650 * @start_pfn: The new PFN of the range
4651 * @end_pfn: The new PFN of the range
4653 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4654 * The map is kept near the end physical page range that has already been
4655 * registered. This function allows an arch to shrink an existing registered
4658 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4659 unsigned long end_pfn)
4664 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4665 nid, start_pfn, end_pfn);
4667 /* Find the old active region end and shrink */
4668 for_each_active_range_index_in_nid(i, nid) {
4669 if (early_node_map[i].start_pfn >= start_pfn &&
4670 early_node_map[i].end_pfn <= end_pfn) {
4672 early_node_map[i].start_pfn = 0;
4673 early_node_map[i].end_pfn = 0;
4677 if (early_node_map[i].start_pfn < start_pfn &&
4678 early_node_map[i].end_pfn > start_pfn) {
4679 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4680 early_node_map[i].end_pfn = start_pfn;
4681 if (temp_end_pfn > end_pfn)
4682 add_active_range(nid, end_pfn, temp_end_pfn);
4685 if (early_node_map[i].start_pfn >= start_pfn &&
4686 early_node_map[i].end_pfn > end_pfn &&
4687 early_node_map[i].start_pfn < end_pfn) {
4688 early_node_map[i].start_pfn = end_pfn;
4696 /* remove the blank ones */
4697 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4698 if (early_node_map[i].nid != nid)
4700 if (early_node_map[i].end_pfn)
4702 /* we found it, get rid of it */
4703 for (j = i; j < nr_nodemap_entries - 1; j++)
4704 memcpy(&early_node_map[j], &early_node_map[j+1],
4705 sizeof(early_node_map[j]));
4706 j = nr_nodemap_entries - 1;
4707 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4708 nr_nodemap_entries--;
4713 * remove_all_active_ranges - Remove all currently registered regions
4715 * During discovery, it may be found that a table like SRAT is invalid
4716 * and an alternative discovery method must be used. This function removes
4717 * all currently registered regions.
4719 void __init remove_all_active_ranges(void)
4721 memset(early_node_map, 0, sizeof(early_node_map));
4722 nr_nodemap_entries = 0;
4725 /* Compare two active node_active_regions */
4726 static int __init cmp_node_active_region(const void *a, const void *b)
4728 struct node_active_region *arange = (struct node_active_region *)a;
4729 struct node_active_region *brange = (struct node_active_region *)b;
4731 /* Done this way to avoid overflows */
4732 if (arange->start_pfn > brange->start_pfn)
4734 if (arange->start_pfn < brange->start_pfn)
4740 /* sort the node_map by start_pfn */
4741 void __init sort_node_map(void)
4743 sort(early_node_map, (size_t)nr_nodemap_entries,
4744 sizeof(struct node_active_region),
4745 cmp_node_active_region, NULL);
4749 * node_map_pfn_alignment - determine the maximum internode alignment
4751 * This function should be called after node map is populated and sorted.
4752 * It calculates the maximum power of two alignment which can distinguish
4755 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4756 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4757 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4758 * shifted, 1GiB is enough and this function will indicate so.
4760 * This is used to test whether pfn -> nid mapping of the chosen memory
4761 * model has fine enough granularity to avoid incorrect mapping for the
4762 * populated node map.
4764 * Returns the determined alignment in pfn's. 0 if there is no alignment
4765 * requirement (single node).
4767 unsigned long __init node_map_pfn_alignment(void)
4769 unsigned long accl_mask = 0, last_end = 0;
4773 for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4774 int nid = early_node_map[i].nid;
4775 unsigned long start = early_node_map[i].start_pfn;
4776 unsigned long end = early_node_map[i].end_pfn;
4779 if (!start || last_nid < 0 || last_nid == nid) {
4786 * Start with a mask granular enough to pin-point to the
4787 * start pfn and tick off bits one-by-one until it becomes
4788 * too coarse to separate the current node from the last.
4790 mask = ~((1 << __ffs(start)) - 1);
4791 while (mask && last_end <= (start & (mask << 1)))
4794 /* accumulate all internode masks */
4798 /* convert mask to number of pages */
4799 return ~accl_mask + 1;
4802 /* Find the lowest pfn for a node */
4803 static unsigned long __init find_min_pfn_for_node(int nid)
4806 unsigned long min_pfn = ULONG_MAX;
4808 /* Assuming a sorted map, the first range found has the starting pfn */
4809 for_each_active_range_index_in_nid(i, nid)
4810 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4812 if (min_pfn == ULONG_MAX) {
4814 "Could not find start_pfn for node %d\n", nid);
4822 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4824 * It returns the minimum PFN based on information provided via
4825 * add_active_range().
4827 unsigned long __init find_min_pfn_with_active_regions(void)
4829 return find_min_pfn_for_node(MAX_NUMNODES);
4833 * early_calculate_totalpages()
4834 * Sum pages in active regions for movable zone.
4835 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4837 static unsigned long __init early_calculate_totalpages(void)
4840 unsigned long totalpages = 0;
4842 for (i = 0; i < nr_nodemap_entries; i++) {
4843 unsigned long pages = early_node_map[i].end_pfn -
4844 early_node_map[i].start_pfn;
4845 totalpages += pages;
4847 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4853 * Find the PFN the Movable zone begins in each node. Kernel memory
4854 * is spread evenly between nodes as long as the nodes have enough
4855 * memory. When they don't, some nodes will have more kernelcore than
4858 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4861 unsigned long usable_startpfn;
4862 unsigned long kernelcore_node, kernelcore_remaining;
4863 /* save the state before borrow the nodemask */
4864 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4865 unsigned long totalpages = early_calculate_totalpages();
4866 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4869 * If movablecore was specified, calculate what size of
4870 * kernelcore that corresponds so that memory usable for
4871 * any allocation type is evenly spread. If both kernelcore
4872 * and movablecore are specified, then the value of kernelcore
4873 * will be used for required_kernelcore if it's greater than
4874 * what movablecore would have allowed.
4876 if (required_movablecore) {
4877 unsigned long corepages;
4880 * Round-up so that ZONE_MOVABLE is at least as large as what
4881 * was requested by the user
4883 required_movablecore =
4884 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4885 corepages = totalpages - required_movablecore;
4887 required_kernelcore = max(required_kernelcore, corepages);
4890 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4891 if (!required_kernelcore)
4894 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4895 find_usable_zone_for_movable();
4896 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4899 /* Spread kernelcore memory as evenly as possible throughout nodes */
4900 kernelcore_node = required_kernelcore / usable_nodes;
4901 for_each_node_state(nid, N_HIGH_MEMORY) {
4903 * Recalculate kernelcore_node if the division per node
4904 * now exceeds what is necessary to satisfy the requested
4905 * amount of memory for the kernel
4907 if (required_kernelcore < kernelcore_node)
4908 kernelcore_node = required_kernelcore / usable_nodes;
4911 * As the map is walked, we track how much memory is usable
4912 * by the kernel using kernelcore_remaining. When it is
4913 * 0, the rest of the node is usable by ZONE_MOVABLE
4915 kernelcore_remaining = kernelcore_node;
4917 /* Go through each range of PFNs within this node */
4918 for_each_active_range_index_in_nid(i, nid) {
4919 unsigned long start_pfn, end_pfn;
4920 unsigned long size_pages;
4922 start_pfn = max(early_node_map[i].start_pfn,
4923 zone_movable_pfn[nid]);
4924 end_pfn = early_node_map[i].end_pfn;
4925 if (start_pfn >= end_pfn)
4928 /* Account for what is only usable for kernelcore */
4929 if (start_pfn < usable_startpfn) {
4930 unsigned long kernel_pages;
4931 kernel_pages = min(end_pfn, usable_startpfn)
4934 kernelcore_remaining -= min(kernel_pages,
4935 kernelcore_remaining);
4936 required_kernelcore -= min(kernel_pages,
4937 required_kernelcore);
4939 /* Continue if range is now fully accounted */
4940 if (end_pfn <= usable_startpfn) {
4943 * Push zone_movable_pfn to the end so
4944 * that if we have to rebalance
4945 * kernelcore across nodes, we will
4946 * not double account here
4948 zone_movable_pfn[nid] = end_pfn;
4951 start_pfn = usable_startpfn;
4955 * The usable PFN range for ZONE_MOVABLE is from
4956 * start_pfn->end_pfn. Calculate size_pages as the
4957 * number of pages used as kernelcore
4959 size_pages = end_pfn - start_pfn;
4960 if (size_pages > kernelcore_remaining)
4961 size_pages = kernelcore_remaining;
4962 zone_movable_pfn[nid] = start_pfn + size_pages;
4965 * Some kernelcore has been met, update counts and
4966 * break if the kernelcore for this node has been
4969 required_kernelcore -= min(required_kernelcore,
4971 kernelcore_remaining -= size_pages;
4972 if (!kernelcore_remaining)
4978 * If there is still required_kernelcore, we do another pass with one
4979 * less node in the count. This will push zone_movable_pfn[nid] further
4980 * along on the nodes that still have memory until kernelcore is
4984 if (usable_nodes && required_kernelcore > usable_nodes)
4987 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4988 for (nid = 0; nid < MAX_NUMNODES; nid++)
4989 zone_movable_pfn[nid] =
4990 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4993 /* restore the node_state */
4994 node_states[N_HIGH_MEMORY] = saved_node_state;
4997 /* Any regular memory on that node ? */
4998 static void check_for_regular_memory(pg_data_t *pgdat)
5000 #ifdef CONFIG_HIGHMEM
5001 enum zone_type zone_type;
5003 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
5004 struct zone *zone = &pgdat->node_zones[zone_type];
5005 if (zone->present_pages)
5006 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
5012 * free_area_init_nodes - Initialise all pg_data_t and zone data
5013 * @max_zone_pfn: an array of max PFNs for each zone
5015 * This will call free_area_init_node() for each active node in the system.
5016 * Using the page ranges provided by add_active_range(), the size of each
5017 * zone in each node and their holes is calculated. If the maximum PFN
5018 * between two adjacent zones match, it is assumed that the zone is empty.
5019 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5020 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5021 * starts where the previous one ended. For example, ZONE_DMA32 starts
5022 * at arch_max_dma_pfn.
5024 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5029 /* Sort early_node_map as initialisation assumes it is sorted */
5032 /* Record where the zone boundaries are */
5033 memset(arch_zone_lowest_possible_pfn, 0,
5034 sizeof(arch_zone_lowest_possible_pfn));
5035 memset(arch_zone_highest_possible_pfn, 0,
5036 sizeof(arch_zone_highest_possible_pfn));
5037 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5038 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5039 for (i = 1; i < MAX_NR_ZONES; i++) {
5040 if (i == ZONE_MOVABLE)
5042 arch_zone_lowest_possible_pfn[i] =
5043 arch_zone_highest_possible_pfn[i-1];
5044 arch_zone_highest_possible_pfn[i] =
5045 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5047 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5048 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5050 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5051 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5052 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
5054 /* Print out the zone ranges */
5055 printk("Zone PFN ranges:\n");
5056 for (i = 0; i < MAX_NR_ZONES; i++) {
5057 if (i == ZONE_MOVABLE)
5059 printk(" %-8s ", zone_names[i]);
5060 if (arch_zone_lowest_possible_pfn[i] ==
5061 arch_zone_highest_possible_pfn[i])
5064 printk("%0#10lx -> %0#10lx\n",
5065 arch_zone_lowest_possible_pfn[i],
5066 arch_zone_highest_possible_pfn[i]);
5069 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5070 printk("Movable zone start PFN for each node\n");
5071 for (i = 0; i < MAX_NUMNODES; i++) {
5072 if (zone_movable_pfn[i])
5073 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
5076 /* Print out the early_node_map[] */
5077 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
5078 for (i = 0; i < nr_nodemap_entries; i++)
5079 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
5080 early_node_map[i].start_pfn,
5081 early_node_map[i].end_pfn);
5083 /* Initialise every node */
5084 mminit_verify_pageflags_layout();
5085 setup_nr_node_ids();
5086 for_each_online_node(nid) {
5087 pg_data_t *pgdat = NODE_DATA(nid);
5088 free_area_init_node(nid, NULL,
5089 find_min_pfn_for_node(nid), NULL);
5091 /* Any memory on that node */
5092 if (pgdat->node_present_pages)
5093 node_set_state(nid, N_HIGH_MEMORY);
5094 check_for_regular_memory(pgdat);
5098 static int __init cmdline_parse_core(char *p, unsigned long *core)
5100 unsigned long long coremem;
5104 coremem = memparse(p, &p);
5105 *core = coremem >> PAGE_SHIFT;
5107 /* Paranoid check that UL is enough for the coremem value */
5108 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5114 * kernelcore=size sets the amount of memory for use for allocations that
5115 * cannot be reclaimed or migrated.
5117 static int __init cmdline_parse_kernelcore(char *p)
5119 return cmdline_parse_core(p, &required_kernelcore);
5123 * movablecore=size sets the amount of memory for use for allocations that
5124 * can be reclaimed or migrated.
5126 static int __init cmdline_parse_movablecore(char *p)
5128 return cmdline_parse_core(p, &required_movablecore);
5131 early_param("kernelcore", cmdline_parse_kernelcore);
5132 early_param("movablecore", cmdline_parse_movablecore);
5134 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
5137 * set_dma_reserve - set the specified number of pages reserved in the first zone
5138 * @new_dma_reserve: The number of pages to mark reserved
5140 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5141 * In the DMA zone, a significant percentage may be consumed by kernel image
5142 * and other unfreeable allocations which can skew the watermarks badly. This
5143 * function may optionally be used to account for unfreeable pages in the
5144 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5145 * smaller per-cpu batchsize.
5147 void __init set_dma_reserve(unsigned long new_dma_reserve)
5149 dma_reserve = new_dma_reserve;
5152 void __init free_area_init(unsigned long *zones_size)
5154 free_area_init_node(0, zones_size,
5155 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5158 static int page_alloc_cpu_notify(struct notifier_block *self,
5159 unsigned long action, void *hcpu)
5161 int cpu = (unsigned long)hcpu;
5163 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5167 * Spill the event counters of the dead processor
5168 * into the current processors event counters.
5169 * This artificially elevates the count of the current
5172 vm_events_fold_cpu(cpu);
5175 * Zero the differential counters of the dead processor
5176 * so that the vm statistics are consistent.
5178 * This is only okay since the processor is dead and cannot
5179 * race with what we are doing.
5181 refresh_cpu_vm_stats(cpu);
5186 void __init page_alloc_init(void)
5188 hotcpu_notifier(page_alloc_cpu_notify, 0);
5192 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5193 * or min_free_kbytes changes.
5195 static void calculate_totalreserve_pages(void)
5197 struct pglist_data *pgdat;
5198 unsigned long reserve_pages = 0;
5199 enum zone_type i, j;
5201 for_each_online_pgdat(pgdat) {
5202 for (i = 0; i < MAX_NR_ZONES; i++) {
5203 struct zone *zone = pgdat->node_zones + i;
5204 unsigned long max = 0;
5206 /* Find valid and maximum lowmem_reserve in the zone */
5207 for (j = i; j < MAX_NR_ZONES; j++) {
5208 if (zone->lowmem_reserve[j] > max)
5209 max = zone->lowmem_reserve[j];
5212 /* we treat the high watermark as reserved pages. */
5213 max += high_wmark_pages(zone);
5215 if (max > zone->present_pages)
5216 max = zone->present_pages;
5217 reserve_pages += max;
5219 * Lowmem reserves are not available to
5220 * GFP_HIGHUSER page cache allocations and
5221 * kswapd tries to balance zones to their high
5222 * watermark. As a result, neither should be
5223 * regarded as dirtyable memory, to prevent a
5224 * situation where reclaim has to clean pages
5225 * in order to balance the zones.
5227 zone->dirty_balance_reserve = max;
5230 dirty_balance_reserve = reserve_pages;
5231 totalreserve_pages = reserve_pages;
5235 * setup_per_zone_lowmem_reserve - called whenever
5236 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5237 * has a correct pages reserved value, so an adequate number of
5238 * pages are left in the zone after a successful __alloc_pages().
5240 static void setup_per_zone_lowmem_reserve(void)
5242 struct pglist_data *pgdat;
5243 enum zone_type j, idx;
5245 for_each_online_pgdat(pgdat) {
5246 for (j = 0; j < MAX_NR_ZONES; j++) {
5247 struct zone *zone = pgdat->node_zones + j;
5248 unsigned long present_pages = zone->present_pages;
5250 zone->lowmem_reserve[j] = 0;
5254 struct zone *lower_zone;
5258 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5259 sysctl_lowmem_reserve_ratio[idx] = 1;
5261 lower_zone = pgdat->node_zones + idx;
5262 lower_zone->lowmem_reserve[j] = present_pages /
5263 sysctl_lowmem_reserve_ratio[idx];
5264 present_pages += lower_zone->present_pages;
5269 /* update totalreserve_pages */
5270 calculate_totalreserve_pages();
5274 * setup_per_zone_wmarks - called when min_free_kbytes changes
5275 * or when memory is hot-{added|removed}
5277 * Ensures that the watermark[min,low,high] values for each zone are set
5278 * correctly with respect to min_free_kbytes.
5280 void setup_per_zone_wmarks(void)
5282 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5283 unsigned long lowmem_pages = 0;
5285 unsigned long flags;
5287 /* Calculate total number of !ZONE_HIGHMEM pages */
5288 for_each_zone(zone) {
5289 if (!is_highmem(zone))
5290 lowmem_pages += zone->present_pages;
5293 for_each_zone(zone) {
5296 spin_lock_irqsave(&zone->lock, flags);
5297 tmp = (u64)pages_min * zone->present_pages;
5298 do_div(tmp, lowmem_pages);
5299 if (is_highmem(zone)) {
5301 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5302 * need highmem pages, so cap pages_min to a small
5305 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5306 * deltas controls asynch page reclaim, and so should
5307 * not be capped for highmem.
5311 min_pages = zone->present_pages / 1024;
5312 if (min_pages < SWAP_CLUSTER_MAX)
5313 min_pages = SWAP_CLUSTER_MAX;
5314 if (min_pages > 128)
5316 zone->watermark[WMARK_MIN] = min_pages;
5319 * If it's a lowmem zone, reserve a number of pages
5320 * proportionate to the zone's size.
5322 zone->watermark[WMARK_MIN] = tmp;
5325 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5326 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5327 setup_zone_migrate_reserve(zone);
5328 spin_unlock_irqrestore(&zone->lock, flags);
5331 /* update totalreserve_pages */
5332 calculate_totalreserve_pages();
5336 * The inactive anon list should be small enough that the VM never has to
5337 * do too much work, but large enough that each inactive page has a chance
5338 * to be referenced again before it is swapped out.
5340 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5341 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5342 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5343 * the anonymous pages are kept on the inactive list.
5346 * memory ratio inactive anon
5347 * -------------------------------------
5356 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5358 unsigned int gb, ratio;
5360 /* Zone size in gigabytes */
5361 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5363 ratio = int_sqrt(10 * gb);
5367 zone->inactive_ratio = ratio;
5370 static void __meminit setup_per_zone_inactive_ratio(void)
5375 calculate_zone_inactive_ratio(zone);
5379 * Initialise min_free_kbytes.
5381 * For small machines we want it small (128k min). For large machines
5382 * we want it large (64MB max). But it is not linear, because network
5383 * bandwidth does not increase linearly with machine size. We use
5385 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5386 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5402 int __meminit init_per_zone_wmark_min(void)
5404 unsigned long lowmem_kbytes;
5406 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5408 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5409 if (min_free_kbytes < 128)
5410 min_free_kbytes = 128;
5411 if (min_free_kbytes > 65536)
5412 min_free_kbytes = 65536;
5413 setup_per_zone_wmarks();
5414 refresh_zone_stat_thresholds();
5415 setup_per_zone_lowmem_reserve();
5416 setup_per_zone_inactive_ratio();
5419 module_init(init_per_zone_wmark_min)
5422 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5423 * that we can call two helper functions whenever min_free_kbytes
5426 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5427 void __user *buffer, size_t *length, loff_t *ppos)
5429 proc_dointvec(table, write, buffer, length, ppos);
5431 setup_per_zone_wmarks();
5436 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5437 void __user *buffer, size_t *length, loff_t *ppos)
5442 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5447 zone->min_unmapped_pages = (zone->present_pages *
5448 sysctl_min_unmapped_ratio) / 100;
5452 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5453 void __user *buffer, size_t *length, loff_t *ppos)
5458 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5463 zone->min_slab_pages = (zone->present_pages *
5464 sysctl_min_slab_ratio) / 100;
5470 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5471 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5472 * whenever sysctl_lowmem_reserve_ratio changes.
5474 * The reserve ratio obviously has absolutely no relation with the
5475 * minimum watermarks. The lowmem reserve ratio can only make sense
5476 * if in function of the boot time zone sizes.
5478 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5479 void __user *buffer, size_t *length, loff_t *ppos)
5481 proc_dointvec_minmax(table, write, buffer, length, ppos);
5482 setup_per_zone_lowmem_reserve();
5487 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5488 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5489 * can have before it gets flushed back to buddy allocator.
5492 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5493 void __user *buffer, size_t *length, loff_t *ppos)
5499 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5500 if (!write || (ret == -EINVAL))
5502 for_each_populated_zone(zone) {
5503 for_each_possible_cpu(cpu) {
5505 high = zone->present_pages / percpu_pagelist_fraction;
5506 setup_pagelist_highmark(
5507 per_cpu_ptr(zone->pageset, cpu), high);
5513 int hashdist = HASHDIST_DEFAULT;
5516 static int __init set_hashdist(char *str)
5520 hashdist = simple_strtoul(str, &str, 0);
5523 __setup("hashdist=", set_hashdist);
5527 * allocate a large system hash table from bootmem
5528 * - it is assumed that the hash table must contain an exact power-of-2
5529 * quantity of entries
5530 * - limit is the number of hash buckets, not the total allocation size
5532 void *__init alloc_large_system_hash(const char *tablename,
5533 unsigned long bucketsize,
5534 unsigned long numentries,
5537 unsigned int *_hash_shift,
5538 unsigned int *_hash_mask,
5539 unsigned long limit)
5541 unsigned long long max = limit;
5542 unsigned long log2qty, size;
5545 /* allow the kernel cmdline to have a say */
5547 /* round applicable memory size up to nearest megabyte */
5548 numentries = nr_kernel_pages;
5549 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5550 numentries >>= 20 - PAGE_SHIFT;
5551 numentries <<= 20 - PAGE_SHIFT;
5553 /* limit to 1 bucket per 2^scale bytes of low memory */
5554 if (scale > PAGE_SHIFT)
5555 numentries >>= (scale - PAGE_SHIFT);
5557 numentries <<= (PAGE_SHIFT - scale);
5559 /* Make sure we've got at least a 0-order allocation.. */
5560 if (unlikely(flags & HASH_SMALL)) {
5561 /* Makes no sense without HASH_EARLY */
5562 WARN_ON(!(flags & HASH_EARLY));
5563 if (!(numentries >> *_hash_shift)) {
5564 numentries = 1UL << *_hash_shift;
5565 BUG_ON(!numentries);
5567 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5568 numentries = PAGE_SIZE / bucketsize;
5570 numentries = roundup_pow_of_two(numentries);
5572 /* limit allocation size to 1/16 total memory by default */
5574 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5575 do_div(max, bucketsize);
5578 if (numentries > max)
5581 log2qty = ilog2(numentries);
5584 size = bucketsize << log2qty;
5585 if (flags & HASH_EARLY)
5586 table = alloc_bootmem_nopanic(size);
5588 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5591 * If bucketsize is not a power-of-two, we may free
5592 * some pages at the end of hash table which
5593 * alloc_pages_exact() automatically does
5595 if (get_order(size) < MAX_ORDER) {
5596 table = alloc_pages_exact(size, GFP_ATOMIC);
5597 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5600 } while (!table && size > PAGE_SIZE && --log2qty);
5603 panic("Failed to allocate %s hash table\n", tablename);
5605 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5608 ilog2(size) - PAGE_SHIFT,
5612 *_hash_shift = log2qty;
5614 *_hash_mask = (1 << log2qty) - 1;
5619 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5620 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5623 #ifdef CONFIG_SPARSEMEM
5624 return __pfn_to_section(pfn)->pageblock_flags;
5626 return zone->pageblock_flags;
5627 #endif /* CONFIG_SPARSEMEM */
5630 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5632 #ifdef CONFIG_SPARSEMEM
5633 pfn &= (PAGES_PER_SECTION-1);
5634 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5636 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5637 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5638 #endif /* CONFIG_SPARSEMEM */
5642 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5643 * @page: The page within the block of interest
5644 * @start_bitidx: The first bit of interest to retrieve
5645 * @end_bitidx: The last bit of interest
5646 * returns pageblock_bits flags
5648 unsigned long get_pageblock_flags_group(struct page *page,
5649 int start_bitidx, int end_bitidx)
5652 unsigned long *bitmap;
5653 unsigned long pfn, bitidx;
5654 unsigned long flags = 0;
5655 unsigned long value = 1;
5657 zone = page_zone(page);
5658 pfn = page_to_pfn(page);
5659 bitmap = get_pageblock_bitmap(zone, pfn);
5660 bitidx = pfn_to_bitidx(zone, pfn);
5662 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5663 if (test_bit(bitidx + start_bitidx, bitmap))
5670 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5671 * @page: The page within the block of interest
5672 * @start_bitidx: The first bit of interest
5673 * @end_bitidx: The last bit of interest
5674 * @flags: The flags to set
5676 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5677 int start_bitidx, int end_bitidx)
5680 unsigned long *bitmap;
5681 unsigned long pfn, bitidx;
5682 unsigned long value = 1;
5684 zone = page_zone(page);
5685 pfn = page_to_pfn(page);
5686 bitmap = get_pageblock_bitmap(zone, pfn);
5687 bitidx = pfn_to_bitidx(zone, pfn);
5688 VM_BUG_ON(pfn < zone->zone_start_pfn);
5689 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5691 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5693 __set_bit(bitidx + start_bitidx, bitmap);
5695 __clear_bit(bitidx + start_bitidx, bitmap);
5699 * This is designed as sub function...plz see page_isolation.c also.
5700 * set/clear page block's type to be ISOLATE.
5701 * page allocater never alloc memory from ISOLATE block.
5705 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5707 unsigned long pfn, iter, found;
5709 * For avoiding noise data, lru_add_drain_all() should be called
5710 * If ZONE_MOVABLE, the zone never contains immobile pages
5712 if (zone_idx(zone) == ZONE_MOVABLE)
5715 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5718 pfn = page_to_pfn(page);
5719 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5720 unsigned long check = pfn + iter;
5722 if (!pfn_valid_within(check))
5725 page = pfn_to_page(check);
5726 if (!page_count(page)) {
5727 if (PageBuddy(page))
5728 iter += (1 << page_order(page)) - 1;
5734 * If there are RECLAIMABLE pages, we need to check it.
5735 * But now, memory offline itself doesn't call shrink_slab()
5736 * and it still to be fixed.
5739 * If the page is not RAM, page_count()should be 0.
5740 * we don't need more check. This is an _used_ not-movable page.
5742 * The problematic thing here is PG_reserved pages. PG_reserved
5743 * is set to both of a memory hole page and a _used_ kernel
5752 bool is_pageblock_removable_nolock(struct page *page)
5754 struct zone *zone = page_zone(page);
5755 unsigned long pfn = page_to_pfn(page);
5758 * We have to be careful here because we are iterating over memory
5759 * sections which are not zone aware so we might end up outside of
5760 * the zone but still within the section.
5762 if (!zone || zone->zone_start_pfn > pfn ||
5763 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5766 return __count_immobile_pages(zone, page, 0);
5769 int set_migratetype_isolate(struct page *page)
5772 unsigned long flags, pfn;
5773 struct memory_isolate_notify arg;
5777 zone = page_zone(page);
5779 spin_lock_irqsave(&zone->lock, flags);
5781 pfn = page_to_pfn(page);
5782 arg.start_pfn = pfn;
5783 arg.nr_pages = pageblock_nr_pages;
5784 arg.pages_found = 0;
5787 * It may be possible to isolate a pageblock even if the
5788 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5789 * notifier chain is used by balloon drivers to return the
5790 * number of pages in a range that are held by the balloon
5791 * driver to shrink memory. If all the pages are accounted for
5792 * by balloons, are free, or on the LRU, isolation can continue.
5793 * Later, for example, when memory hotplug notifier runs, these
5794 * pages reported as "can be isolated" should be isolated(freed)
5795 * by the balloon driver through the memory notifier chain.
5797 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5798 notifier_ret = notifier_to_errno(notifier_ret);
5802 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5803 * We just check MOVABLE pages.
5805 if (__count_immobile_pages(zone, page, arg.pages_found))
5809 * immobile means "not-on-lru" paes. If immobile is larger than
5810 * removable-by-driver pages reported by notifier, we'll fail.
5815 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5816 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5819 spin_unlock_irqrestore(&zone->lock, flags);
5825 void unset_migratetype_isolate(struct page *page)
5828 unsigned long flags;
5829 zone = page_zone(page);
5830 spin_lock_irqsave(&zone->lock, flags);
5831 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5833 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5834 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5836 spin_unlock_irqrestore(&zone->lock, flags);
5841 static unsigned long pfn_max_align_down(unsigned long pfn)
5843 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5844 pageblock_nr_pages) - 1);
5847 static unsigned long pfn_max_align_up(unsigned long pfn)
5849 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5850 pageblock_nr_pages));
5853 static struct page *
5854 __alloc_contig_migrate_alloc(struct page *page, unsigned long private,
5857 return alloc_page(GFP_HIGHUSER_MOVABLE);
5860 /* [start, end) must belong to a single zone. */
5861 static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5863 /* This function is based on compact_zone() from compaction.c. */
5865 unsigned long pfn = start;
5866 unsigned int tries = 0;
5869 struct compact_control cc = {
5870 .nr_migratepages = 0,
5872 .zone = page_zone(pfn_to_page(start)),
5875 INIT_LIST_HEAD(&cc.migratepages);
5877 migrate_prep_local();
5879 while (pfn < end || !list_empty(&cc.migratepages)) {
5880 if (fatal_signal_pending(current)) {
5885 if (list_empty(&cc.migratepages)) {
5886 cc.nr_migratepages = 0;
5887 pfn = isolate_migratepages_range(cc.zone, &cc,
5894 } else if (++tries == 5) {
5895 ret = ret < 0 ? ret : -EBUSY;
5899 ret = migrate_pages(&cc.migratepages,
5900 __alloc_contig_migrate_alloc,
5904 putback_lru_pages(&cc.migratepages);
5905 return ret > 0 ? 0 : ret;
5909 * alloc_contig_range() -- tries to allocate given range of pages
5910 * @start: start PFN to allocate
5911 * @end: one-past-the-last PFN to allocate
5913 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5914 * aligned, however it's the caller's responsibility to guarantee that
5915 * we are the only thread that changes migrate type of pageblocks the
5918 * The PFN range must belong to a single zone.
5920 * Returns zero on success or negative error code. On success all
5921 * pages which PFN is in [start, end) are allocated for the caller and
5922 * need to be freed with free_contig_range().
5924 int alloc_contig_range(unsigned long start, unsigned long end)
5926 struct zone *zone = page_zone(pfn_to_page(start));
5927 unsigned long outer_start, outer_end;
5931 * What we do here is we mark all pageblocks in range as
5932 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5933 * have different sizes, and due to the way page allocator
5934 * work, we align the range to biggest of the two pages so
5935 * that page allocator won't try to merge buddies from
5936 * different pageblocks and change MIGRATE_ISOLATE to some
5937 * other migration type.
5939 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5940 * migrate the pages from an unaligned range (ie. pages that
5941 * we are interested in). This will put all the pages in
5942 * range back to page allocator as MIGRATE_ISOLATE.
5944 * When this is done, we take the pages in range from page
5945 * allocator removing them from the buddy system. This way
5946 * page allocator will never consider using them.
5948 * This lets us mark the pageblocks back as
5949 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5950 * aligned range but not in the unaligned, original range are
5951 * put back to page allocator so that buddy can use them.
5954 ret = start_isolate_page_range(pfn_max_align_down(start),
5955 pfn_max_align_up(end));
5959 ret = __alloc_contig_migrate_range(start, end);
5964 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5965 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5966 * more, all pages in [start, end) are free in page allocator.
5967 * What we are going to do is to allocate all pages from
5968 * [start, end) (that is remove them from page allocator).
5970 * The only problem is that pages at the beginning and at the
5971 * end of interesting range may be not aligned with pages that
5972 * page allocator holds, ie. they can be part of higher order
5973 * pages. Because of this, we reserve the bigger range and
5974 * once this is done free the pages we are not interested in.
5976 * We don't have to hold zone->lock here because the pages are
5977 * isolated thus they won't get removed from buddy.
5980 lru_add_drain_all();
5984 outer_start = start;
5985 while (!PageBuddy(pfn_to_page(outer_start))) {
5986 if (++order >= MAX_ORDER) {
5990 outer_start &= ~0UL << order;
5993 /* Make sure the range is really isolated. */
5994 if (test_pages_isolated(outer_start, end)) {
5995 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6001 outer_end = isolate_freepages_range(outer_start, end);
6007 /* Free head and tail (if any) */
6008 if (start != outer_start)
6009 free_contig_range(outer_start, start - outer_start);
6010 if (end != outer_end)
6011 free_contig_range(end, outer_end - end);
6014 undo_isolate_page_range(pfn_max_align_down(start),
6015 pfn_max_align_up(end));
6019 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6021 for (; nr_pages--; ++pfn)
6022 __free_page(pfn_to_page(pfn));
6026 #ifdef CONFIG_MEMORY_HOTREMOVE
6028 * All pages in the range must be isolated before calling this.
6031 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6037 unsigned long flags;
6038 /* find the first valid pfn */
6039 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6044 zone = page_zone(pfn_to_page(pfn));
6045 spin_lock_irqsave(&zone->lock, flags);
6047 while (pfn < end_pfn) {
6048 if (!pfn_valid(pfn)) {
6052 page = pfn_to_page(pfn);
6053 BUG_ON(page_count(page));
6054 BUG_ON(!PageBuddy(page));
6055 order = page_order(page);
6056 #ifdef CONFIG_DEBUG_VM
6057 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6058 pfn, 1 << order, end_pfn);
6060 list_del(&page->lru);
6061 rmv_page_order(page);
6062 zone->free_area[order].nr_free--;
6063 __mod_zone_page_state(zone, NR_FREE_PAGES,
6065 for (i = 0; i < (1 << order); i++)
6066 SetPageReserved((page+i));
6067 pfn += (1 << order);
6069 spin_unlock_irqrestore(&zone->lock, flags);
6073 #ifdef CONFIG_MEMORY_FAILURE
6074 bool is_free_buddy_page(struct page *page)
6076 struct zone *zone = page_zone(page);
6077 unsigned long pfn = page_to_pfn(page);
6078 unsigned long flags;
6081 spin_lock_irqsave(&zone->lock, flags);
6082 for (order = 0; order < MAX_ORDER; order++) {
6083 struct page *page_head = page - (pfn & ((1 << order) - 1));
6085 if (PageBuddy(page_head) && page_order(page_head) >= order)
6088 spin_unlock_irqrestore(&zone->lock, flags);
6090 return order < MAX_ORDER;
6094 static struct trace_print_flags pageflag_names[] = {
6095 {1UL << PG_locked, "locked" },
6096 {1UL << PG_error, "error" },
6097 {1UL << PG_referenced, "referenced" },
6098 {1UL << PG_uptodate, "uptodate" },
6099 {1UL << PG_dirty, "dirty" },
6100 {1UL << PG_lru, "lru" },
6101 {1UL << PG_active, "active" },
6102 {1UL << PG_slab, "slab" },
6103 {1UL << PG_owner_priv_1, "owner_priv_1" },
6104 {1UL << PG_arch_1, "arch_1" },
6105 {1UL << PG_reserved, "reserved" },
6106 {1UL << PG_private, "private" },
6107 {1UL << PG_private_2, "private_2" },
6108 {1UL << PG_writeback, "writeback" },
6109 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6110 {1UL << PG_head, "head" },
6111 {1UL << PG_tail, "tail" },
6113 {1UL << PG_compound, "compound" },
6115 {1UL << PG_swapcache, "swapcache" },
6116 {1UL << PG_mappedtodisk, "mappedtodisk" },
6117 {1UL << PG_reclaim, "reclaim" },
6118 {1UL << PG_swapbacked, "swapbacked" },
6119 {1UL << PG_unevictable, "unevictable" },
6121 {1UL << PG_mlocked, "mlocked" },
6123 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6124 {1UL << PG_uncached, "uncached" },
6126 #ifdef CONFIG_MEMORY_FAILURE
6127 {1UL << PG_hwpoison, "hwpoison" },
6132 static void dump_page_flags(unsigned long flags)
6134 const char *delim = "";
6138 printk(KERN_ALERT "page flags: %#lx(", flags);
6140 /* remove zone id */
6141 flags &= (1UL << NR_PAGEFLAGS) - 1;
6143 for (i = 0; pageflag_names[i].name && flags; i++) {
6145 mask = pageflag_names[i].mask;
6146 if ((flags & mask) != mask)
6150 printk("%s%s", delim, pageflag_names[i].name);
6154 /* check for left over flags */
6156 printk("%s%#lx", delim, flags);
6161 void dump_page(struct page *page)
6164 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6165 page, atomic_read(&page->_count), page_mapcount(page),
6166 page->mapping, page->index);
6167 dump_page_flags(page->flags);
6168 mem_cgroup_print_bad_page(page);