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/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
62 #include <asm/tlbflush.h>
63 #include <asm/div64.h>
66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67 DEFINE_PER_CPU(int, numa_node);
68 EXPORT_PER_CPU_SYMBOL(numa_node);
71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
78 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
83 * Array of node states.
85 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86 [N_POSSIBLE] = NODE_MASK_ALL,
87 [N_ONLINE] = { { [0] = 1UL } },
89 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
91 [N_HIGH_MEMORY] = { { [0] = 1UL } },
93 [N_CPU] = { { [0] = 1UL } },
96 EXPORT_SYMBOL(node_states);
98 unsigned long totalram_pages __read_mostly;
99 unsigned long totalreserve_pages __read_mostly;
101 * When calculating the number of globally allowed dirty pages, there
102 * is a certain number of per-zone reserves that should not be
103 * considered dirtyable memory. This is the sum of those reserves
104 * over all existing zones that contribute dirtyable memory.
106 unsigned long dirty_balance_reserve __read_mostly;
108 int percpu_pagelist_fraction;
109 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
111 #ifdef CONFIG_PM_SLEEP
113 * The following functions are used by the suspend/hibernate code to temporarily
114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
115 * while devices are suspended. To avoid races with the suspend/hibernate code,
116 * they should always be called with pm_mutex held (gfp_allowed_mask also should
117 * only be modified with pm_mutex held, unless the suspend/hibernate code is
118 * guaranteed not to run in parallel with that modification).
121 static gfp_t saved_gfp_mask;
123 void pm_restore_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 if (saved_gfp_mask) {
127 gfp_allowed_mask = saved_gfp_mask;
132 void pm_restrict_gfp_mask(void)
134 WARN_ON(!mutex_is_locked(&pm_mutex));
135 WARN_ON(saved_gfp_mask);
136 saved_gfp_mask = gfp_allowed_mask;
137 gfp_allowed_mask &= ~GFP_IOFS;
140 bool pm_suspended_storage(void)
142 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
146 #endif /* CONFIG_PM_SLEEP */
148 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
149 int pageblock_order __read_mostly;
152 static void __free_pages_ok(struct page *page, unsigned int order);
155 * results with 256, 32 in the lowmem_reserve sysctl:
156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
157 * 1G machine -> (16M dma, 784M normal, 224M high)
158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
162 * TBD: should special case ZONE_DMA32 machines here - in those we normally
163 * don't need any ZONE_NORMAL reservation
165 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
166 #ifdef CONFIG_ZONE_DMA
169 #ifdef CONFIG_ZONE_DMA32
172 #ifdef CONFIG_HIGHMEM
178 EXPORT_SYMBOL(totalram_pages);
180 static char * const zone_names[MAX_NR_ZONES] = {
181 #ifdef CONFIG_ZONE_DMA
184 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 int min_free_kbytes = 1024;
196 static unsigned long __meminitdata nr_kernel_pages;
197 static unsigned long __meminitdata nr_all_pages;
198 static unsigned long __meminitdata dma_reserve;
200 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
201 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
202 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
203 static unsigned long __initdata required_kernelcore;
204 static unsigned long __initdata required_movablecore;
205 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
207 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
209 EXPORT_SYMBOL(movable_zone);
210 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
213 int nr_node_ids __read_mostly = MAX_NUMNODES;
214 int nr_online_nodes __read_mostly = 1;
215 EXPORT_SYMBOL(nr_node_ids);
216 EXPORT_SYMBOL(nr_online_nodes);
219 int page_group_by_mobility_disabled __read_mostly;
223 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
224 * Instead, use {un}set_pageblock_isolate.
226 void set_pageblock_migratetype(struct page *page, int migratetype)
229 if (unlikely(page_group_by_mobility_disabled))
230 migratetype = MIGRATE_UNMOVABLE;
232 set_pageblock_flags_group(page, (unsigned long)migratetype,
233 PB_migrate, PB_migrate_end);
236 bool oom_killer_disabled __read_mostly;
238 #ifdef CONFIG_DEBUG_VM
239 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
243 unsigned long pfn = page_to_pfn(page);
246 seq = zone_span_seqbegin(zone);
247 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
249 else if (pfn < zone->zone_start_pfn)
251 } while (zone_span_seqretry(zone, seq));
256 static int page_is_consistent(struct zone *zone, struct page *page)
258 if (!pfn_valid_within(page_to_pfn(page)))
260 if (zone != page_zone(page))
266 * Temporary debugging check for pages not lying within a given zone.
268 static int bad_range(struct zone *zone, struct page *page)
270 if (page_outside_zone_boundaries(zone, page))
272 if (!page_is_consistent(zone, page))
278 static inline int bad_range(struct zone *zone, struct page *page)
284 static void bad_page(struct page *page)
286 static unsigned long resume;
287 static unsigned long nr_shown;
288 static unsigned long nr_unshown;
290 /* Don't complain about poisoned pages */
291 if (PageHWPoison(page)) {
292 reset_page_mapcount(page); /* remove PageBuddy */
297 * Allow a burst of 60 reports, then keep quiet for that minute;
298 * or allow a steady drip of one report per second.
300 if (nr_shown == 60) {
301 if (time_before(jiffies, resume)) {
307 "BUG: Bad page state: %lu messages suppressed\n",
314 resume = jiffies + 60 * HZ;
316 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
317 current->comm, page_to_pfn(page));
323 /* Leave bad fields for debug, except PageBuddy could make trouble */
324 reset_page_mapcount(page); /* remove PageBuddy */
325 add_taint(TAINT_BAD_PAGE);
329 * Higher-order pages are called "compound pages". They are structured thusly:
331 * The first PAGE_SIZE page is called the "head page".
333 * The remaining PAGE_SIZE pages are called "tail pages".
335 * All pages have PG_compound set. All tail pages have their ->first_page
336 * pointing at the head page.
338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
343 static void free_compound_page(struct page *page)
345 __free_pages_ok(page, compound_order(page));
348 void prep_compound_page(struct page *page, unsigned long order)
351 int nr_pages = 1 << order;
353 set_compound_page_dtor(page, free_compound_page);
354 set_compound_order(page, order);
356 for (i = 1; i < nr_pages; i++) {
357 struct page *p = page + i;
359 set_page_count(p, 0);
360 p->first_page = page;
364 /* update __split_huge_page_refcount if you change this function */
365 static int destroy_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
371 if (unlikely(compound_order(page) != order) ||
372 unlikely(!PageHead(page))) {
377 __ClearPageHead(page);
379 for (i = 1; i < nr_pages; i++) {
380 struct page *p = page + i;
382 if (unlikely(!PageTail(p) || (p->first_page != page))) {
392 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
397 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
398 * and __GFP_HIGHMEM from hard or soft interrupt context.
400 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
401 for (i = 0; i < (1 << order); i++)
402 clear_highpage(page + i);
405 #ifdef CONFIG_DEBUG_PAGEALLOC
406 unsigned int _debug_guardpage_minorder;
408 static int __init debug_guardpage_minorder_setup(char *buf)
412 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
413 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
416 _debug_guardpage_minorder = res;
417 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
420 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
422 static inline void set_page_guard_flag(struct page *page)
424 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
427 static inline void clear_page_guard_flag(struct page *page)
429 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
432 static inline void set_page_guard_flag(struct page *page) { }
433 static inline void clear_page_guard_flag(struct page *page) { }
436 static inline void set_page_order(struct page *page, int order)
438 set_page_private(page, order);
439 __SetPageBuddy(page);
442 static inline void rmv_page_order(struct page *page)
444 __ClearPageBuddy(page);
445 set_page_private(page, 0);
449 * Locate the struct page for both the matching buddy in our
450 * pair (buddy1) and the combined O(n+1) page they form (page).
452 * 1) Any buddy B1 will have an order O twin B2 which satisfies
453 * the following equation:
455 * For example, if the starting buddy (buddy2) is #8 its order
457 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
459 * 2) Any buddy B will have an order O+1 parent P which
460 * satisfies the following equation:
463 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
465 static inline unsigned long
466 __find_buddy_index(unsigned long page_idx, unsigned int order)
468 return page_idx ^ (1 << order);
472 * This function checks whether a page is free && is the buddy
473 * we can do coalesce a page and its buddy if
474 * (a) the buddy is not in a hole &&
475 * (b) the buddy is in the buddy system &&
476 * (c) a page and its buddy have the same order &&
477 * (d) a page and its buddy are in the same zone.
479 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
480 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
482 * For recording page's order, we use page_private(page).
484 static inline int page_is_buddy(struct page *page, struct page *buddy,
487 if (!pfn_valid_within(page_to_pfn(buddy)))
490 if (page_zone_id(page) != page_zone_id(buddy))
493 if (page_is_guard(buddy) && page_order(buddy) == order) {
494 VM_BUG_ON(page_count(buddy) != 0);
498 if (PageBuddy(buddy) && page_order(buddy) == order) {
499 VM_BUG_ON(page_count(buddy) != 0);
506 * Freeing function for a buddy system allocator.
508 * The concept of a buddy system is to maintain direct-mapped table
509 * (containing bit values) for memory blocks of various "orders".
510 * The bottom level table contains the map for the smallest allocatable
511 * units of memory (here, pages), and each level above it describes
512 * pairs of units from the levels below, hence, "buddies".
513 * At a high level, all that happens here is marking the table entry
514 * at the bottom level available, and propagating the changes upward
515 * as necessary, plus some accounting needed to play nicely with other
516 * parts of the VM system.
517 * At each level, we keep a list of pages, which are heads of continuous
518 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
519 * order is recorded in page_private(page) field.
520 * So when we are allocating or freeing one, we can derive the state of the
521 * other. That is, if we allocate a small block, and both were
522 * free, the remainder of the region must be split into blocks.
523 * If a block is freed, and its buddy is also free, then this
524 * triggers coalescing into a block of larger size.
529 static inline void __free_one_page(struct page *page,
530 struct zone *zone, unsigned int order,
533 unsigned long page_idx;
534 unsigned long combined_idx;
535 unsigned long uninitialized_var(buddy_idx);
538 if (unlikely(PageCompound(page)))
539 if (unlikely(destroy_compound_page(page, order)))
542 VM_BUG_ON(migratetype == -1);
544 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
546 VM_BUG_ON(page_idx & ((1 << order) - 1));
547 VM_BUG_ON(bad_range(zone, page));
549 while (order < MAX_ORDER-1) {
550 buddy_idx = __find_buddy_index(page_idx, order);
551 buddy = page + (buddy_idx - page_idx);
552 if (!page_is_buddy(page, buddy, order))
555 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
556 * merge with it and move up one order.
558 if (page_is_guard(buddy)) {
559 clear_page_guard_flag(buddy);
560 set_page_private(page, 0);
561 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
563 list_del(&buddy->lru);
564 zone->free_area[order].nr_free--;
565 rmv_page_order(buddy);
567 combined_idx = buddy_idx & page_idx;
568 page = page + (combined_idx - page_idx);
569 page_idx = combined_idx;
572 set_page_order(page, order);
575 * If this is not the largest possible page, check if the buddy
576 * of the next-highest order is free. If it is, it's possible
577 * that pages are being freed that will coalesce soon. In case,
578 * that is happening, add the free page to the tail of the list
579 * so it's less likely to be used soon and more likely to be merged
580 * as a higher order page
582 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
583 struct page *higher_page, *higher_buddy;
584 combined_idx = buddy_idx & page_idx;
585 higher_page = page + (combined_idx - page_idx);
586 buddy_idx = __find_buddy_index(combined_idx, order + 1);
587 higher_buddy = page + (buddy_idx - combined_idx);
588 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
589 list_add_tail(&page->lru,
590 &zone->free_area[order].free_list[migratetype]);
595 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
597 zone->free_area[order].nr_free++;
601 * free_page_mlock() -- clean up attempts to free and mlocked() page.
602 * Page should not be on lru, so no need to fix that up.
603 * free_pages_check() will verify...
605 static inline void free_page_mlock(struct page *page)
607 __dec_zone_page_state(page, NR_MLOCK);
608 __count_vm_event(UNEVICTABLE_MLOCKFREED);
611 static inline int free_pages_check(struct page *page)
613 if (unlikely(page_mapcount(page) |
614 (page->mapping != NULL) |
615 (atomic_read(&page->_count) != 0) |
616 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
617 (mem_cgroup_bad_page_check(page)))) {
621 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
622 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
627 * Frees a number of pages from the PCP lists
628 * Assumes all pages on list are in same zone, and of same order.
629 * count is the number of pages to free.
631 * If the zone was previously in an "all pages pinned" state then look to
632 * see if this freeing clears that state.
634 * And clear the zone's pages_scanned counter, to hold off the "all pages are
635 * pinned" detection logic.
637 static void free_pcppages_bulk(struct zone *zone, int count,
638 struct per_cpu_pages *pcp)
644 spin_lock(&zone->lock);
645 zone->all_unreclaimable = 0;
646 zone->pages_scanned = 0;
650 struct list_head *list;
653 * Remove pages from lists in a round-robin fashion. A
654 * batch_free count is maintained that is incremented when an
655 * empty list is encountered. This is so more pages are freed
656 * off fuller lists instead of spinning excessively around empty
661 if (++migratetype == MIGRATE_PCPTYPES)
663 list = &pcp->lists[migratetype];
664 } while (list_empty(list));
666 /* This is the only non-empty list. Free them all. */
667 if (batch_free == MIGRATE_PCPTYPES)
668 batch_free = to_free;
671 page = list_entry(list->prev, struct page, lru);
672 /* must delete as __free_one_page list manipulates */
673 list_del(&page->lru);
674 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
675 __free_one_page(page, zone, 0, page_private(page));
676 trace_mm_page_pcpu_drain(page, 0, page_private(page));
677 } while (--to_free && --batch_free && !list_empty(list));
679 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
680 spin_unlock(&zone->lock);
683 static void free_one_page(struct zone *zone, struct page *page, int order,
686 spin_lock(&zone->lock);
687 zone->all_unreclaimable = 0;
688 zone->pages_scanned = 0;
690 __free_one_page(page, zone, order, migratetype);
691 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
692 spin_unlock(&zone->lock);
695 static bool free_pages_prepare(struct page *page, unsigned int order)
700 trace_mm_page_free(page, order);
701 kmemcheck_free_shadow(page, order);
704 page->mapping = NULL;
705 for (i = 0; i < (1 << order); i++)
706 bad += free_pages_check(page + i);
710 if (!PageHighMem(page)) {
711 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
712 debug_check_no_obj_freed(page_address(page),
715 arch_free_page(page, order);
716 kernel_map_pages(page, 1 << order, 0);
721 static void __free_pages_ok(struct page *page, unsigned int order)
724 int wasMlocked = __TestClearPageMlocked(page);
726 if (!free_pages_prepare(page, order))
729 local_irq_save(flags);
730 if (unlikely(wasMlocked))
731 free_page_mlock(page);
732 __count_vm_events(PGFREE, 1 << order);
733 free_one_page(page_zone(page), page, order,
734 get_pageblock_migratetype(page));
735 local_irq_restore(flags);
738 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
740 unsigned int nr_pages = 1 << order;
744 for (loop = 0; loop < nr_pages; loop++) {
745 struct page *p = &page[loop];
747 if (loop + 1 < nr_pages)
749 __ClearPageReserved(p);
750 set_page_count(p, 0);
753 set_page_refcounted(page);
754 __free_pages(page, order);
758 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
759 void __init init_cma_reserved_pageblock(struct page *page)
761 unsigned i = pageblock_nr_pages;
762 struct page *p = page;
765 __ClearPageReserved(p);
766 set_page_count(p, 0);
769 set_page_refcounted(page);
770 set_pageblock_migratetype(page, MIGRATE_CMA);
771 __free_pages(page, pageblock_order);
772 totalram_pages += pageblock_nr_pages;
777 * The order of subdivision here is critical for the IO subsystem.
778 * Please do not alter this order without good reasons and regression
779 * testing. Specifically, as large blocks of memory are subdivided,
780 * the order in which smaller blocks are delivered depends on the order
781 * they're subdivided in this function. This is the primary factor
782 * influencing the order in which pages are delivered to the IO
783 * subsystem according to empirical testing, and this is also justified
784 * by considering the behavior of a buddy system containing a single
785 * large block of memory acted on by a series of small allocations.
786 * This behavior is a critical factor in sglist merging's success.
790 static inline void expand(struct zone *zone, struct page *page,
791 int low, int high, struct free_area *area,
794 unsigned long size = 1 << high;
800 VM_BUG_ON(bad_range(zone, &page[size]));
802 #ifdef CONFIG_DEBUG_PAGEALLOC
803 if (high < debug_guardpage_minorder()) {
805 * Mark as guard pages (or page), that will allow to
806 * merge back to allocator when buddy will be freed.
807 * Corresponding page table entries will not be touched,
808 * pages will stay not present in virtual address space
810 INIT_LIST_HEAD(&page[size].lru);
811 set_page_guard_flag(&page[size]);
812 set_page_private(&page[size], high);
813 /* Guard pages are not available for any usage */
814 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high));
818 list_add(&page[size].lru, &area->free_list[migratetype]);
820 set_page_order(&page[size], high);
825 * This page is about to be returned from the page allocator
827 static inline int check_new_page(struct page *page)
829 if (unlikely(page_mapcount(page) |
830 (page->mapping != NULL) |
831 (atomic_read(&page->_count) != 0) |
832 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
833 (mem_cgroup_bad_page_check(page)))) {
840 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
844 for (i = 0; i < (1 << order); i++) {
845 struct page *p = page + i;
846 if (unlikely(check_new_page(p)))
850 set_page_private(page, 0);
851 set_page_refcounted(page);
853 arch_alloc_page(page, order);
854 kernel_map_pages(page, 1 << order, 1);
856 if (gfp_flags & __GFP_ZERO)
857 prep_zero_page(page, order, gfp_flags);
859 if (order && (gfp_flags & __GFP_COMP))
860 prep_compound_page(page, order);
866 * Go through the free lists for the given migratetype and remove
867 * the smallest available page from the freelists
870 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
873 unsigned int current_order;
874 struct free_area * area;
877 /* Find a page of the appropriate size in the preferred list */
878 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
879 area = &(zone->free_area[current_order]);
880 if (list_empty(&area->free_list[migratetype]))
883 page = list_entry(area->free_list[migratetype].next,
885 list_del(&page->lru);
886 rmv_page_order(page);
888 expand(zone, page, order, current_order, area, migratetype);
897 * This array describes the order lists are fallen back to when
898 * the free lists for the desirable migrate type are depleted
900 static int fallbacks[MIGRATE_TYPES][4] = {
901 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
902 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
904 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
905 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
907 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
909 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
910 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
914 * Move the free pages in a range to the free lists of the requested type.
915 * Note that start_page and end_pages are not aligned on a pageblock
916 * boundary. If alignment is required, use move_freepages_block()
918 static int move_freepages(struct zone *zone,
919 struct page *start_page, struct page *end_page,
926 #ifndef CONFIG_HOLES_IN_ZONE
928 * page_zone is not safe to call in this context when
929 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
930 * anyway as we check zone boundaries in move_freepages_block().
931 * Remove at a later date when no bug reports exist related to
932 * grouping pages by mobility
934 BUG_ON(page_zone(start_page) != page_zone(end_page));
937 for (page = start_page; page <= end_page;) {
938 /* Make sure we are not inadvertently changing nodes */
939 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
941 if (!pfn_valid_within(page_to_pfn(page))) {
946 if (!PageBuddy(page)) {
951 order = page_order(page);
952 list_move(&page->lru,
953 &zone->free_area[order].free_list[migratetype]);
955 pages_moved += 1 << order;
961 int move_freepages_block(struct zone *zone, struct page *page,
964 unsigned long start_pfn, end_pfn;
965 struct page *start_page, *end_page;
967 start_pfn = page_to_pfn(page);
968 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
969 start_page = pfn_to_page(start_pfn);
970 end_page = start_page + pageblock_nr_pages - 1;
971 end_pfn = start_pfn + pageblock_nr_pages - 1;
973 /* Do not cross zone boundaries */
974 if (start_pfn < zone->zone_start_pfn)
976 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
979 return move_freepages(zone, start_page, end_page, migratetype);
982 static void change_pageblock_range(struct page *pageblock_page,
983 int start_order, int migratetype)
985 int nr_pageblocks = 1 << (start_order - pageblock_order);
987 while (nr_pageblocks--) {
988 set_pageblock_migratetype(pageblock_page, migratetype);
989 pageblock_page += pageblock_nr_pages;
993 /* Remove an element from the buddy allocator from the fallback list */
994 static inline struct page *
995 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
997 struct free_area * area;
1002 /* Find the largest possible block of pages in the other list */
1003 for (current_order = MAX_ORDER-1; current_order >= order;
1006 migratetype = fallbacks[start_migratetype][i];
1008 /* MIGRATE_RESERVE handled later if necessary */
1009 if (migratetype == MIGRATE_RESERVE)
1012 area = &(zone->free_area[current_order]);
1013 if (list_empty(&area->free_list[migratetype]))
1016 page = list_entry(area->free_list[migratetype].next,
1021 * If breaking a large block of pages, move all free
1022 * pages to the preferred allocation list. If falling
1023 * back for a reclaimable kernel allocation, be more
1024 * aggressive about taking ownership of free pages
1026 * On the other hand, never change migration
1027 * type of MIGRATE_CMA pageblocks nor move CMA
1028 * pages on different free lists. We don't
1029 * want unmovable pages to be allocated from
1030 * MIGRATE_CMA areas.
1032 if (!is_migrate_cma(migratetype) &&
1033 (unlikely(current_order >= pageblock_order / 2) ||
1034 start_migratetype == MIGRATE_RECLAIMABLE ||
1035 page_group_by_mobility_disabled)) {
1037 pages = move_freepages_block(zone, page,
1040 /* Claim the whole block if over half of it is free */
1041 if (pages >= (1 << (pageblock_order-1)) ||
1042 page_group_by_mobility_disabled)
1043 set_pageblock_migratetype(page,
1046 migratetype = start_migratetype;
1049 /* Remove the page from the freelists */
1050 list_del(&page->lru);
1051 rmv_page_order(page);
1053 /* Take ownership for orders >= pageblock_order */
1054 if (current_order >= pageblock_order &&
1055 !is_migrate_cma(migratetype))
1056 change_pageblock_range(page, current_order,
1059 expand(zone, page, order, current_order, area,
1060 is_migrate_cma(migratetype)
1061 ? migratetype : start_migratetype);
1063 trace_mm_page_alloc_extfrag(page, order, current_order,
1064 start_migratetype, migratetype);
1074 * Do the hard work of removing an element from the buddy allocator.
1075 * Call me with the zone->lock already held.
1077 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1083 page = __rmqueue_smallest(zone, order, migratetype);
1085 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1086 page = __rmqueue_fallback(zone, order, migratetype);
1089 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1090 * is used because __rmqueue_smallest is an inline function
1091 * and we want just one call site
1094 migratetype = MIGRATE_RESERVE;
1099 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1104 * Obtain a specified number of elements from the buddy allocator, all under
1105 * a single hold of the lock, for efficiency. Add them to the supplied list.
1106 * Returns the number of new pages which were placed at *list.
1108 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1109 unsigned long count, struct list_head *list,
1110 int migratetype, int cold)
1112 int mt = migratetype, i;
1114 spin_lock(&zone->lock);
1115 for (i = 0; i < count; ++i) {
1116 struct page *page = __rmqueue(zone, order, migratetype);
1117 if (unlikely(page == NULL))
1121 * Split buddy pages returned by expand() are received here
1122 * in physical page order. The page is added to the callers and
1123 * list and the list head then moves forward. From the callers
1124 * perspective, the linked list is ordered by page number in
1125 * some conditions. This is useful for IO devices that can
1126 * merge IO requests if the physical pages are ordered
1129 if (likely(cold == 0))
1130 list_add(&page->lru, list);
1132 list_add_tail(&page->lru, list);
1133 if (IS_ENABLED(CONFIG_CMA)) {
1134 mt = get_pageblock_migratetype(page);
1135 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1138 set_page_private(page, mt);
1141 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1142 spin_unlock(&zone->lock);
1148 * Called from the vmstat counter updater to drain pagesets of this
1149 * currently executing processor on remote nodes after they have
1152 * Note that this function must be called with the thread pinned to
1153 * a single processor.
1155 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1157 unsigned long flags;
1160 local_irq_save(flags);
1161 if (pcp->count >= pcp->batch)
1162 to_drain = pcp->batch;
1164 to_drain = pcp->count;
1166 free_pcppages_bulk(zone, to_drain, pcp);
1167 pcp->count -= to_drain;
1169 local_irq_restore(flags);
1174 * Drain pages of the indicated processor.
1176 * The processor must either be the current processor and the
1177 * thread pinned to the current processor or a processor that
1180 static void drain_pages(unsigned int cpu)
1182 unsigned long flags;
1185 for_each_populated_zone(zone) {
1186 struct per_cpu_pageset *pset;
1187 struct per_cpu_pages *pcp;
1189 local_irq_save(flags);
1190 pset = per_cpu_ptr(zone->pageset, cpu);
1194 free_pcppages_bulk(zone, pcp->count, pcp);
1197 local_irq_restore(flags);
1202 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1204 void drain_local_pages(void *arg)
1206 drain_pages(smp_processor_id());
1210 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1212 * Note that this code is protected against sending an IPI to an offline
1213 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1214 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1215 * nothing keeps CPUs from showing up after we populated the cpumask and
1216 * before the call to on_each_cpu_mask().
1218 void drain_all_pages(void)
1221 struct per_cpu_pageset *pcp;
1225 * Allocate in the BSS so we wont require allocation in
1226 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1228 static cpumask_t cpus_with_pcps;
1231 * We don't care about racing with CPU hotplug event
1232 * as offline notification will cause the notified
1233 * cpu to drain that CPU pcps and on_each_cpu_mask
1234 * disables preemption as part of its processing
1236 for_each_online_cpu(cpu) {
1237 bool has_pcps = false;
1238 for_each_populated_zone(zone) {
1239 pcp = per_cpu_ptr(zone->pageset, cpu);
1240 if (pcp->pcp.count) {
1246 cpumask_set_cpu(cpu, &cpus_with_pcps);
1248 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1250 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1253 #ifdef CONFIG_HIBERNATION
1255 void mark_free_pages(struct zone *zone)
1257 unsigned long pfn, max_zone_pfn;
1258 unsigned long flags;
1260 struct list_head *curr;
1262 if (!zone->spanned_pages)
1265 spin_lock_irqsave(&zone->lock, flags);
1267 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1268 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1269 if (pfn_valid(pfn)) {
1270 struct page *page = pfn_to_page(pfn);
1272 if (!swsusp_page_is_forbidden(page))
1273 swsusp_unset_page_free(page);
1276 for_each_migratetype_order(order, t) {
1277 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1280 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1281 for (i = 0; i < (1UL << order); i++)
1282 swsusp_set_page_free(pfn_to_page(pfn + i));
1285 spin_unlock_irqrestore(&zone->lock, flags);
1287 #endif /* CONFIG_PM */
1290 * Free a 0-order page
1291 * cold == 1 ? free a cold page : free a hot page
1293 void free_hot_cold_page(struct page *page, int cold)
1295 struct zone *zone = page_zone(page);
1296 struct per_cpu_pages *pcp;
1297 unsigned long flags;
1299 int wasMlocked = __TestClearPageMlocked(page);
1301 if (!free_pages_prepare(page, 0))
1304 migratetype = get_pageblock_migratetype(page);
1305 set_page_private(page, migratetype);
1306 local_irq_save(flags);
1307 if (unlikely(wasMlocked))
1308 free_page_mlock(page);
1309 __count_vm_event(PGFREE);
1312 * We only track unmovable, reclaimable and movable on pcp lists.
1313 * Free ISOLATE pages back to the allocator because they are being
1314 * offlined but treat RESERVE as movable pages so we can get those
1315 * areas back if necessary. Otherwise, we may have to free
1316 * excessively into the page allocator
1318 if (migratetype >= MIGRATE_PCPTYPES) {
1319 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1320 free_one_page(zone, page, 0, migratetype);
1323 migratetype = MIGRATE_MOVABLE;
1326 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1328 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1330 list_add(&page->lru, &pcp->lists[migratetype]);
1332 if (pcp->count >= pcp->high) {
1333 free_pcppages_bulk(zone, pcp->batch, pcp);
1334 pcp->count -= pcp->batch;
1338 local_irq_restore(flags);
1342 * Free a list of 0-order pages
1344 void free_hot_cold_page_list(struct list_head *list, int cold)
1346 struct page *page, *next;
1348 list_for_each_entry_safe(page, next, list, lru) {
1349 trace_mm_page_free_batched(page, cold);
1350 free_hot_cold_page(page, cold);
1355 * split_page takes a non-compound higher-order page, and splits it into
1356 * n (1<<order) sub-pages: page[0..n]
1357 * Each sub-page must be freed individually.
1359 * Note: this is probably too low level an operation for use in drivers.
1360 * Please consult with lkml before using this in your driver.
1362 void split_page(struct page *page, unsigned int order)
1366 VM_BUG_ON(PageCompound(page));
1367 VM_BUG_ON(!page_count(page));
1369 #ifdef CONFIG_KMEMCHECK
1371 * Split shadow pages too, because free(page[0]) would
1372 * otherwise free the whole shadow.
1374 if (kmemcheck_page_is_tracked(page))
1375 split_page(virt_to_page(page[0].shadow), order);
1378 for (i = 1; i < (1 << order); i++)
1379 set_page_refcounted(page + i);
1383 * Similar to split_page except the page is already free. As this is only
1384 * being used for migration, the migratetype of the block also changes.
1385 * As this is called with interrupts disabled, the caller is responsible
1386 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1389 * Note: this is probably too low level an operation for use in drivers.
1390 * Please consult with lkml before using this in your driver.
1392 int split_free_page(struct page *page)
1395 unsigned long watermark;
1398 BUG_ON(!PageBuddy(page));
1400 zone = page_zone(page);
1401 order = page_order(page);
1403 /* Obey watermarks as if the page was being allocated */
1404 watermark = low_wmark_pages(zone) + (1 << order);
1405 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1408 /* Remove page from free list */
1409 list_del(&page->lru);
1410 zone->free_area[order].nr_free--;
1411 rmv_page_order(page);
1412 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1414 /* Split into individual pages */
1415 set_page_refcounted(page);
1416 split_page(page, order);
1418 if (order >= pageblock_order - 1) {
1419 struct page *endpage = page + (1 << order) - 1;
1420 for (; page < endpage; page += pageblock_nr_pages) {
1421 int mt = get_pageblock_migratetype(page);
1422 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1423 set_pageblock_migratetype(page,
1432 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1433 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1437 struct page *buffered_rmqueue(struct zone *preferred_zone,
1438 struct zone *zone, int order, gfp_t gfp_flags,
1441 unsigned long flags;
1443 int cold = !!(gfp_flags & __GFP_COLD);
1446 if (likely(order == 0)) {
1447 struct per_cpu_pages *pcp;
1448 struct list_head *list;
1450 local_irq_save(flags);
1451 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1452 list = &pcp->lists[migratetype];
1453 if (list_empty(list)) {
1454 pcp->count += rmqueue_bulk(zone, 0,
1457 if (unlikely(list_empty(list)))
1462 page = list_entry(list->prev, struct page, lru);
1464 page = list_entry(list->next, struct page, lru);
1466 list_del(&page->lru);
1469 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1471 * __GFP_NOFAIL is not to be used in new code.
1473 * All __GFP_NOFAIL callers should be fixed so that they
1474 * properly detect and handle allocation failures.
1476 * We most definitely don't want callers attempting to
1477 * allocate greater than order-1 page units with
1480 WARN_ON_ONCE(order > 1);
1482 spin_lock_irqsave(&zone->lock, flags);
1483 page = __rmqueue(zone, order, migratetype);
1484 spin_unlock(&zone->lock);
1487 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1490 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1491 zone_statistics(preferred_zone, zone, gfp_flags);
1492 local_irq_restore(flags);
1494 VM_BUG_ON(bad_range(zone, page));
1495 if (prep_new_page(page, order, gfp_flags))
1500 local_irq_restore(flags);
1504 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1505 #define ALLOC_WMARK_MIN WMARK_MIN
1506 #define ALLOC_WMARK_LOW WMARK_LOW
1507 #define ALLOC_WMARK_HIGH WMARK_HIGH
1508 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1510 /* Mask to get the watermark bits */
1511 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1513 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1514 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1515 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1517 #ifdef CONFIG_FAIL_PAGE_ALLOC
1520 struct fault_attr attr;
1522 u32 ignore_gfp_highmem;
1523 u32 ignore_gfp_wait;
1525 } fail_page_alloc = {
1526 .attr = FAULT_ATTR_INITIALIZER,
1527 .ignore_gfp_wait = 1,
1528 .ignore_gfp_highmem = 1,
1532 static int __init setup_fail_page_alloc(char *str)
1534 return setup_fault_attr(&fail_page_alloc.attr, str);
1536 __setup("fail_page_alloc=", setup_fail_page_alloc);
1538 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1540 if (order < fail_page_alloc.min_order)
1542 if (gfp_mask & __GFP_NOFAIL)
1544 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1546 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1549 return should_fail(&fail_page_alloc.attr, 1 << order);
1552 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1554 static int __init fail_page_alloc_debugfs(void)
1556 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1559 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1560 &fail_page_alloc.attr);
1562 return PTR_ERR(dir);
1564 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1565 &fail_page_alloc.ignore_gfp_wait))
1567 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1568 &fail_page_alloc.ignore_gfp_highmem))
1570 if (!debugfs_create_u32("min-order", mode, dir,
1571 &fail_page_alloc.min_order))
1576 debugfs_remove_recursive(dir);
1581 late_initcall(fail_page_alloc_debugfs);
1583 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1585 #else /* CONFIG_FAIL_PAGE_ALLOC */
1587 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1592 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1595 * Return true if free pages are above 'mark'. This takes into account the order
1596 * of the allocation.
1598 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1599 int classzone_idx, int alloc_flags, long free_pages)
1601 /* free_pages my go negative - that's OK */
1603 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1606 free_pages -= (1 << order) - 1;
1607 if (alloc_flags & ALLOC_HIGH)
1609 if (alloc_flags & ALLOC_HARDER)
1612 if (free_pages <= min + lowmem_reserve)
1614 for (o = 0; o < order; o++) {
1615 /* At the next order, this order's pages become unavailable */
1616 free_pages -= z->free_area[o].nr_free << o;
1618 /* Require fewer higher order pages to be free */
1621 if (free_pages <= min)
1627 #ifdef CONFIG_MEMORY_ISOLATION
1628 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1630 if (unlikely(zone->nr_pageblock_isolate))
1631 return zone->nr_pageblock_isolate * pageblock_nr_pages;
1635 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1641 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1642 int classzone_idx, int alloc_flags)
1644 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1645 zone_page_state(z, NR_FREE_PAGES));
1648 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1649 int classzone_idx, int alloc_flags)
1651 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1653 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1654 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1657 * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1658 * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1659 * sleep although it could do so. But this is more desirable for memory
1660 * hotplug than sleeping which can cause a livelock in the direct
1663 free_pages -= nr_zone_isolate_freepages(z);
1664 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1670 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1671 * skip over zones that are not allowed by the cpuset, or that have
1672 * been recently (in last second) found to be nearly full. See further
1673 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1674 * that have to skip over a lot of full or unallowed zones.
1676 * If the zonelist cache is present in the passed in zonelist, then
1677 * returns a pointer to the allowed node mask (either the current
1678 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1680 * If the zonelist cache is not available for this zonelist, does
1681 * nothing and returns NULL.
1683 * If the fullzones BITMAP in the zonelist cache is stale (more than
1684 * a second since last zap'd) then we zap it out (clear its bits.)
1686 * We hold off even calling zlc_setup, until after we've checked the
1687 * first zone in the zonelist, on the theory that most allocations will
1688 * be satisfied from that first zone, so best to examine that zone as
1689 * quickly as we can.
1691 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1693 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1694 nodemask_t *allowednodes; /* zonelist_cache approximation */
1696 zlc = zonelist->zlcache_ptr;
1700 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1701 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1702 zlc->last_full_zap = jiffies;
1705 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1706 &cpuset_current_mems_allowed :
1707 &node_states[N_HIGH_MEMORY];
1708 return allowednodes;
1712 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1713 * if it is worth looking at further for free memory:
1714 * 1) Check that the zone isn't thought to be full (doesn't have its
1715 * bit set in the zonelist_cache fullzones BITMAP).
1716 * 2) Check that the zones node (obtained from the zonelist_cache
1717 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1718 * Return true (non-zero) if zone is worth looking at further, or
1719 * else return false (zero) if it is not.
1721 * This check -ignores- the distinction between various watermarks,
1722 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1723 * found to be full for any variation of these watermarks, it will
1724 * be considered full for up to one second by all requests, unless
1725 * we are so low on memory on all allowed nodes that we are forced
1726 * into the second scan of the zonelist.
1728 * In the second scan we ignore this zonelist cache and exactly
1729 * apply the watermarks to all zones, even it is slower to do so.
1730 * We are low on memory in the second scan, and should leave no stone
1731 * unturned looking for a free page.
1733 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1734 nodemask_t *allowednodes)
1736 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1737 int i; /* index of *z in zonelist zones */
1738 int n; /* node that zone *z is on */
1740 zlc = zonelist->zlcache_ptr;
1744 i = z - zonelist->_zonerefs;
1747 /* This zone is worth trying if it is allowed but not full */
1748 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1752 * Given 'z' scanning a zonelist, set the corresponding bit in
1753 * zlc->fullzones, so that subsequent attempts to allocate a page
1754 * from that zone don't waste time re-examining it.
1756 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1758 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1759 int i; /* index of *z in zonelist zones */
1761 zlc = zonelist->zlcache_ptr;
1765 i = z - zonelist->_zonerefs;
1767 set_bit(i, zlc->fullzones);
1771 * clear all zones full, called after direct reclaim makes progress so that
1772 * a zone that was recently full is not skipped over for up to a second
1774 static void zlc_clear_zones_full(struct zonelist *zonelist)
1776 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1778 zlc = zonelist->zlcache_ptr;
1782 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1785 #else /* CONFIG_NUMA */
1787 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1792 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1793 nodemask_t *allowednodes)
1798 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1802 static void zlc_clear_zones_full(struct zonelist *zonelist)
1805 #endif /* CONFIG_NUMA */
1808 * get_page_from_freelist goes through the zonelist trying to allocate
1811 static struct page *
1812 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1813 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1814 struct zone *preferred_zone, int migratetype)
1817 struct page *page = NULL;
1820 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1821 int zlc_active = 0; /* set if using zonelist_cache */
1822 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1824 classzone_idx = zone_idx(preferred_zone);
1827 * Scan zonelist, looking for a zone with enough free.
1828 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1830 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1831 high_zoneidx, nodemask) {
1832 if (NUMA_BUILD && zlc_active &&
1833 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1835 if ((alloc_flags & ALLOC_CPUSET) &&
1836 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1839 * When allocating a page cache page for writing, we
1840 * want to get it from a zone that is within its dirty
1841 * limit, such that no single zone holds more than its
1842 * proportional share of globally allowed dirty pages.
1843 * The dirty limits take into account the zone's
1844 * lowmem reserves and high watermark so that kswapd
1845 * should be able to balance it without having to
1846 * write pages from its LRU list.
1848 * This may look like it could increase pressure on
1849 * lower zones by failing allocations in higher zones
1850 * before they are full. But the pages that do spill
1851 * over are limited as the lower zones are protected
1852 * by this very same mechanism. It should not become
1853 * a practical burden to them.
1855 * XXX: For now, allow allocations to potentially
1856 * exceed the per-zone dirty limit in the slowpath
1857 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1858 * which is important when on a NUMA setup the allowed
1859 * zones are together not big enough to reach the
1860 * global limit. The proper fix for these situations
1861 * will require awareness of zones in the
1862 * dirty-throttling and the flusher threads.
1864 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1865 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1866 goto this_zone_full;
1868 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1869 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1873 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1874 if (zone_watermark_ok(zone, order, mark,
1875 classzone_idx, alloc_flags))
1878 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1880 * we do zlc_setup if there are multiple nodes
1881 * and before considering the first zone allowed
1884 allowednodes = zlc_setup(zonelist, alloc_flags);
1889 if (zone_reclaim_mode == 0)
1890 goto this_zone_full;
1893 * As we may have just activated ZLC, check if the first
1894 * eligible zone has failed zone_reclaim recently.
1896 if (NUMA_BUILD && zlc_active &&
1897 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1900 ret = zone_reclaim(zone, gfp_mask, order);
1902 case ZONE_RECLAIM_NOSCAN:
1905 case ZONE_RECLAIM_FULL:
1906 /* scanned but unreclaimable */
1909 /* did we reclaim enough */
1910 if (!zone_watermark_ok(zone, order, mark,
1911 classzone_idx, alloc_flags))
1912 goto this_zone_full;
1917 page = buffered_rmqueue(preferred_zone, zone, order,
1918 gfp_mask, migratetype);
1923 zlc_mark_zone_full(zonelist, z);
1926 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1927 /* Disable zlc cache for second zonelist scan */
1935 * Large machines with many possible nodes should not always dump per-node
1936 * meminfo in irq context.
1938 static inline bool should_suppress_show_mem(void)
1943 ret = in_interrupt();
1948 static DEFINE_RATELIMIT_STATE(nopage_rs,
1949 DEFAULT_RATELIMIT_INTERVAL,
1950 DEFAULT_RATELIMIT_BURST);
1952 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1954 unsigned int filter = SHOW_MEM_FILTER_NODES;
1956 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1957 debug_guardpage_minorder() > 0)
1961 * This documents exceptions given to allocations in certain
1962 * contexts that are allowed to allocate outside current's set
1965 if (!(gfp_mask & __GFP_NOMEMALLOC))
1966 if (test_thread_flag(TIF_MEMDIE) ||
1967 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1968 filter &= ~SHOW_MEM_FILTER_NODES;
1969 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1970 filter &= ~SHOW_MEM_FILTER_NODES;
1973 struct va_format vaf;
1976 va_start(args, fmt);
1981 pr_warn("%pV", &vaf);
1986 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1987 current->comm, order, gfp_mask);
1990 if (!should_suppress_show_mem())
1995 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1996 unsigned long did_some_progress,
1997 unsigned long pages_reclaimed)
1999 /* Do not loop if specifically requested */
2000 if (gfp_mask & __GFP_NORETRY)
2003 /* Always retry if specifically requested */
2004 if (gfp_mask & __GFP_NOFAIL)
2008 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2009 * making forward progress without invoking OOM. Suspend also disables
2010 * storage devices so kswapd will not help. Bail if we are suspending.
2012 if (!did_some_progress && pm_suspended_storage())
2016 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2017 * means __GFP_NOFAIL, but that may not be true in other
2020 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2024 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2025 * specified, then we retry until we no longer reclaim any pages
2026 * (above), or we've reclaimed an order of pages at least as
2027 * large as the allocation's order. In both cases, if the
2028 * allocation still fails, we stop retrying.
2030 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2036 static inline struct page *
2037 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2038 struct zonelist *zonelist, enum zone_type high_zoneidx,
2039 nodemask_t *nodemask, struct zone *preferred_zone,
2044 /* Acquire the OOM killer lock for the zones in zonelist */
2045 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2046 schedule_timeout_uninterruptible(1);
2051 * Go through the zonelist yet one more time, keep very high watermark
2052 * here, this is only to catch a parallel oom killing, we must fail if
2053 * we're still under heavy pressure.
2055 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2056 order, zonelist, high_zoneidx,
2057 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2058 preferred_zone, migratetype);
2062 if (!(gfp_mask & __GFP_NOFAIL)) {
2063 /* The OOM killer will not help higher order allocs */
2064 if (order > PAGE_ALLOC_COSTLY_ORDER)
2066 /* The OOM killer does not needlessly kill tasks for lowmem */
2067 if (high_zoneidx < ZONE_NORMAL)
2070 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2071 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2072 * The caller should handle page allocation failure by itself if
2073 * it specifies __GFP_THISNODE.
2074 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2076 if (gfp_mask & __GFP_THISNODE)
2079 /* Exhausted what can be done so it's blamo time */
2080 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2083 clear_zonelist_oom(zonelist, gfp_mask);
2087 #ifdef CONFIG_COMPACTION
2088 /* Try memory compaction for high-order allocations before reclaim */
2089 static struct page *
2090 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2091 struct zonelist *zonelist, enum zone_type high_zoneidx,
2092 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2093 int migratetype, bool sync_migration,
2094 bool *deferred_compaction,
2095 unsigned long *did_some_progress)
2102 if (compaction_deferred(preferred_zone, order)) {
2103 *deferred_compaction = true;
2107 current->flags |= PF_MEMALLOC;
2108 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2109 nodemask, sync_migration);
2110 current->flags &= ~PF_MEMALLOC;
2111 if (*did_some_progress != COMPACT_SKIPPED) {
2113 /* Page migration frees to the PCP lists but we want merging */
2114 drain_pages(get_cpu());
2117 page = get_page_from_freelist(gfp_mask, nodemask,
2118 order, zonelist, high_zoneidx,
2119 alloc_flags, preferred_zone,
2122 preferred_zone->compact_considered = 0;
2123 preferred_zone->compact_defer_shift = 0;
2124 if (order >= preferred_zone->compact_order_failed)
2125 preferred_zone->compact_order_failed = order + 1;
2126 count_vm_event(COMPACTSUCCESS);
2131 * It's bad if compaction run occurs and fails.
2132 * The most likely reason is that pages exist,
2133 * but not enough to satisfy watermarks.
2135 count_vm_event(COMPACTFAIL);
2138 * As async compaction considers a subset of pageblocks, only
2139 * defer if the failure was a sync compaction failure.
2142 defer_compaction(preferred_zone, order);
2150 static inline struct page *
2151 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2152 struct zonelist *zonelist, enum zone_type high_zoneidx,
2153 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2154 int migratetype, bool sync_migration,
2155 bool *deferred_compaction,
2156 unsigned long *did_some_progress)
2160 #endif /* CONFIG_COMPACTION */
2162 /* Perform direct synchronous page reclaim */
2164 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2165 nodemask_t *nodemask)
2167 struct reclaim_state reclaim_state;
2172 /* We now go into synchronous reclaim */
2173 cpuset_memory_pressure_bump();
2174 current->flags |= PF_MEMALLOC;
2175 lockdep_set_current_reclaim_state(gfp_mask);
2176 reclaim_state.reclaimed_slab = 0;
2177 current->reclaim_state = &reclaim_state;
2179 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2181 current->reclaim_state = NULL;
2182 lockdep_clear_current_reclaim_state();
2183 current->flags &= ~PF_MEMALLOC;
2190 /* The really slow allocator path where we enter direct reclaim */
2191 static inline struct page *
2192 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2193 struct zonelist *zonelist, enum zone_type high_zoneidx,
2194 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2195 int migratetype, unsigned long *did_some_progress)
2197 struct page *page = NULL;
2198 bool drained = false;
2200 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2202 if (unlikely(!(*did_some_progress)))
2205 /* After successful reclaim, reconsider all zones for allocation */
2207 zlc_clear_zones_full(zonelist);
2210 page = get_page_from_freelist(gfp_mask, nodemask, order,
2211 zonelist, high_zoneidx,
2212 alloc_flags, preferred_zone,
2216 * If an allocation failed after direct reclaim, it could be because
2217 * pages are pinned on the per-cpu lists. Drain them and try again
2219 if (!page && !drained) {
2229 * This is called in the allocator slow-path if the allocation request is of
2230 * sufficient urgency to ignore watermarks and take other desperate measures
2232 static inline struct page *
2233 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2234 struct zonelist *zonelist, enum zone_type high_zoneidx,
2235 nodemask_t *nodemask, struct zone *preferred_zone,
2241 page = get_page_from_freelist(gfp_mask, nodemask, order,
2242 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2243 preferred_zone, migratetype);
2245 if (!page && gfp_mask & __GFP_NOFAIL)
2246 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2247 } while (!page && (gfp_mask & __GFP_NOFAIL));
2253 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2254 enum zone_type high_zoneidx,
2255 enum zone_type classzone_idx)
2260 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2261 wakeup_kswapd(zone, order, classzone_idx);
2265 gfp_to_alloc_flags(gfp_t gfp_mask)
2267 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2268 const gfp_t wait = gfp_mask & __GFP_WAIT;
2270 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2271 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2274 * The caller may dip into page reserves a bit more if the caller
2275 * cannot run direct reclaim, or if the caller has realtime scheduling
2276 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2277 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2279 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2283 * Not worth trying to allocate harder for
2284 * __GFP_NOMEMALLOC even if it can't schedule.
2286 if (!(gfp_mask & __GFP_NOMEMALLOC))
2287 alloc_flags |= ALLOC_HARDER;
2289 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2290 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2292 alloc_flags &= ~ALLOC_CPUSET;
2293 } else if (unlikely(rt_task(current)) && !in_interrupt())
2294 alloc_flags |= ALLOC_HARDER;
2296 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2297 if (!in_interrupt() &&
2298 ((current->flags & PF_MEMALLOC) ||
2299 unlikely(test_thread_flag(TIF_MEMDIE))))
2300 alloc_flags |= ALLOC_NO_WATERMARKS;
2306 static inline struct page *
2307 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2308 struct zonelist *zonelist, enum zone_type high_zoneidx,
2309 nodemask_t *nodemask, struct zone *preferred_zone,
2312 const gfp_t wait = gfp_mask & __GFP_WAIT;
2313 struct page *page = NULL;
2315 unsigned long pages_reclaimed = 0;
2316 unsigned long did_some_progress;
2317 bool sync_migration = false;
2318 bool deferred_compaction = false;
2321 * In the slowpath, we sanity check order to avoid ever trying to
2322 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2323 * be using allocators in order of preference for an area that is
2326 if (order >= MAX_ORDER) {
2327 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2332 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2333 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2334 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2335 * using a larger set of nodes after it has established that the
2336 * allowed per node queues are empty and that nodes are
2339 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2343 if (!(gfp_mask & __GFP_NO_KSWAPD))
2344 wake_all_kswapd(order, zonelist, high_zoneidx,
2345 zone_idx(preferred_zone));
2348 * OK, we're below the kswapd watermark and have kicked background
2349 * reclaim. Now things get more complex, so set up alloc_flags according
2350 * to how we want to proceed.
2352 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2355 * Find the true preferred zone if the allocation is unconstrained by
2358 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2359 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2363 /* This is the last chance, in general, before the goto nopage. */
2364 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2365 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2366 preferred_zone, migratetype);
2370 /* Allocate without watermarks if the context allows */
2371 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2372 page = __alloc_pages_high_priority(gfp_mask, order,
2373 zonelist, high_zoneidx, nodemask,
2374 preferred_zone, migratetype);
2379 /* Atomic allocations - we can't balance anything */
2383 /* Avoid recursion of direct reclaim */
2384 if (current->flags & PF_MEMALLOC)
2387 /* Avoid allocations with no watermarks from looping endlessly */
2388 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2392 * Try direct compaction. The first pass is asynchronous. Subsequent
2393 * attempts after direct reclaim are synchronous
2395 page = __alloc_pages_direct_compact(gfp_mask, order,
2396 zonelist, high_zoneidx,
2398 alloc_flags, preferred_zone,
2399 migratetype, sync_migration,
2400 &deferred_compaction,
2401 &did_some_progress);
2404 sync_migration = true;
2407 * If compaction is deferred for high-order allocations, it is because
2408 * sync compaction recently failed. In this is the case and the caller
2409 * has requested the system not be heavily disrupted, fail the
2410 * allocation now instead of entering direct reclaim
2412 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2415 /* Try direct reclaim and then allocating */
2416 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2417 zonelist, high_zoneidx,
2419 alloc_flags, preferred_zone,
2420 migratetype, &did_some_progress);
2425 * If we failed to make any progress reclaiming, then we are
2426 * running out of options and have to consider going OOM
2428 if (!did_some_progress) {
2429 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2430 if (oom_killer_disabled)
2432 /* Coredumps can quickly deplete all memory reserves */
2433 if ((current->flags & PF_DUMPCORE) &&
2434 !(gfp_mask & __GFP_NOFAIL))
2436 page = __alloc_pages_may_oom(gfp_mask, order,
2437 zonelist, high_zoneidx,
2438 nodemask, preferred_zone,
2443 if (!(gfp_mask & __GFP_NOFAIL)) {
2445 * The oom killer is not called for high-order
2446 * allocations that may fail, so if no progress
2447 * is being made, there are no other options and
2448 * retrying is unlikely to help.
2450 if (order > PAGE_ALLOC_COSTLY_ORDER)
2453 * The oom killer is not called for lowmem
2454 * allocations to prevent needlessly killing
2457 if (high_zoneidx < ZONE_NORMAL)
2465 /* Check if we should retry the allocation */
2466 pages_reclaimed += did_some_progress;
2467 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2469 /* Wait for some write requests to complete then retry */
2470 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2474 * High-order allocations do not necessarily loop after
2475 * direct reclaim and reclaim/compaction depends on compaction
2476 * being called after reclaim so call directly if necessary
2478 page = __alloc_pages_direct_compact(gfp_mask, order,
2479 zonelist, high_zoneidx,
2481 alloc_flags, preferred_zone,
2482 migratetype, sync_migration,
2483 &deferred_compaction,
2484 &did_some_progress);
2490 warn_alloc_failed(gfp_mask, order, NULL);
2493 if (kmemcheck_enabled)
2494 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2500 * This is the 'heart' of the zoned buddy allocator.
2503 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2504 struct zonelist *zonelist, nodemask_t *nodemask)
2506 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2507 struct zone *preferred_zone;
2508 struct page *page = NULL;
2509 int migratetype = allocflags_to_migratetype(gfp_mask);
2510 unsigned int cpuset_mems_cookie;
2512 gfp_mask &= gfp_allowed_mask;
2514 lockdep_trace_alloc(gfp_mask);
2516 might_sleep_if(gfp_mask & __GFP_WAIT);
2518 if (should_fail_alloc_page(gfp_mask, order))
2522 * Check the zones suitable for the gfp_mask contain at least one
2523 * valid zone. It's possible to have an empty zonelist as a result
2524 * of GFP_THISNODE and a memoryless node
2526 if (unlikely(!zonelist->_zonerefs->zone))
2530 cpuset_mems_cookie = get_mems_allowed();
2532 /* The preferred zone is used for statistics later */
2533 first_zones_zonelist(zonelist, high_zoneidx,
2534 nodemask ? : &cpuset_current_mems_allowed,
2536 if (!preferred_zone)
2539 /* First allocation attempt */
2540 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2541 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2542 preferred_zone, migratetype);
2543 if (unlikely(!page))
2544 page = __alloc_pages_slowpath(gfp_mask, order,
2545 zonelist, high_zoneidx, nodemask,
2546 preferred_zone, migratetype);
2548 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2552 * When updating a task's mems_allowed, it is possible to race with
2553 * parallel threads in such a way that an allocation can fail while
2554 * the mask is being updated. If a page allocation is about to fail,
2555 * check if the cpuset changed during allocation and if so, retry.
2557 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2562 EXPORT_SYMBOL(__alloc_pages_nodemask);
2565 * Common helper functions.
2567 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2572 * __get_free_pages() returns a 32-bit address, which cannot represent
2575 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2577 page = alloc_pages(gfp_mask, order);
2580 return (unsigned long) page_address(page);
2582 EXPORT_SYMBOL(__get_free_pages);
2584 unsigned long get_zeroed_page(gfp_t gfp_mask)
2586 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2588 EXPORT_SYMBOL(get_zeroed_page);
2590 void __free_pages(struct page *page, unsigned int order)
2592 if (put_page_testzero(page)) {
2594 free_hot_cold_page(page, 0);
2596 __free_pages_ok(page, order);
2600 EXPORT_SYMBOL(__free_pages);
2602 void free_pages(unsigned long addr, unsigned int order)
2605 VM_BUG_ON(!virt_addr_valid((void *)addr));
2606 __free_pages(virt_to_page((void *)addr), order);
2610 EXPORT_SYMBOL(free_pages);
2612 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2615 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2616 unsigned long used = addr + PAGE_ALIGN(size);
2618 split_page(virt_to_page((void *)addr), order);
2619 while (used < alloc_end) {
2624 return (void *)addr;
2628 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2629 * @size: the number of bytes to allocate
2630 * @gfp_mask: GFP flags for the allocation
2632 * This function is similar to alloc_pages(), except that it allocates the
2633 * minimum number of pages to satisfy the request. alloc_pages() can only
2634 * allocate memory in power-of-two pages.
2636 * This function is also limited by MAX_ORDER.
2638 * Memory allocated by this function must be released by free_pages_exact().
2640 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2642 unsigned int order = get_order(size);
2645 addr = __get_free_pages(gfp_mask, order);
2646 return make_alloc_exact(addr, order, size);
2648 EXPORT_SYMBOL(alloc_pages_exact);
2651 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2653 * @nid: the preferred node ID where memory should be allocated
2654 * @size: the number of bytes to allocate
2655 * @gfp_mask: GFP flags for the allocation
2657 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2659 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2662 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2664 unsigned order = get_order(size);
2665 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2668 return make_alloc_exact((unsigned long)page_address(p), order, size);
2670 EXPORT_SYMBOL(alloc_pages_exact_nid);
2673 * free_pages_exact - release memory allocated via alloc_pages_exact()
2674 * @virt: the value returned by alloc_pages_exact.
2675 * @size: size of allocation, same value as passed to alloc_pages_exact().
2677 * Release the memory allocated by a previous call to alloc_pages_exact.
2679 void free_pages_exact(void *virt, size_t size)
2681 unsigned long addr = (unsigned long)virt;
2682 unsigned long end = addr + PAGE_ALIGN(size);
2684 while (addr < end) {
2689 EXPORT_SYMBOL(free_pages_exact);
2691 static unsigned int nr_free_zone_pages(int offset)
2696 /* Just pick one node, since fallback list is circular */
2697 unsigned int sum = 0;
2699 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2701 for_each_zone_zonelist(zone, z, zonelist, offset) {
2702 unsigned long size = zone->present_pages;
2703 unsigned long high = high_wmark_pages(zone);
2712 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2714 unsigned int nr_free_buffer_pages(void)
2716 return nr_free_zone_pages(gfp_zone(GFP_USER));
2718 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2721 * Amount of free RAM allocatable within all zones
2723 unsigned int nr_free_pagecache_pages(void)
2725 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2728 static inline void show_node(struct zone *zone)
2731 printk("Node %d ", zone_to_nid(zone));
2734 void si_meminfo(struct sysinfo *val)
2736 val->totalram = totalram_pages;
2738 val->freeram = global_page_state(NR_FREE_PAGES);
2739 val->bufferram = nr_blockdev_pages();
2740 val->totalhigh = totalhigh_pages;
2741 val->freehigh = nr_free_highpages();
2742 val->mem_unit = PAGE_SIZE;
2745 EXPORT_SYMBOL(si_meminfo);
2748 void si_meminfo_node(struct sysinfo *val, int nid)
2750 pg_data_t *pgdat = NODE_DATA(nid);
2752 val->totalram = pgdat->node_present_pages;
2753 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2754 #ifdef CONFIG_HIGHMEM
2755 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2756 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2762 val->mem_unit = PAGE_SIZE;
2767 * Determine whether the node should be displayed or not, depending on whether
2768 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2770 bool skip_free_areas_node(unsigned int flags, int nid)
2773 unsigned int cpuset_mems_cookie;
2775 if (!(flags & SHOW_MEM_FILTER_NODES))
2779 cpuset_mems_cookie = get_mems_allowed();
2780 ret = !node_isset(nid, cpuset_current_mems_allowed);
2781 } while (!put_mems_allowed(cpuset_mems_cookie));
2786 #define K(x) ((x) << (PAGE_SHIFT-10))
2789 * Show free area list (used inside shift_scroll-lock stuff)
2790 * We also calculate the percentage fragmentation. We do this by counting the
2791 * memory on each free list with the exception of the first item on the list.
2792 * Suppresses nodes that are not allowed by current's cpuset if
2793 * SHOW_MEM_FILTER_NODES is passed.
2795 void show_free_areas(unsigned int filter)
2800 for_each_populated_zone(zone) {
2801 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2804 printk("%s per-cpu:\n", zone->name);
2806 for_each_online_cpu(cpu) {
2807 struct per_cpu_pageset *pageset;
2809 pageset = per_cpu_ptr(zone->pageset, cpu);
2811 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2812 cpu, pageset->pcp.high,
2813 pageset->pcp.batch, pageset->pcp.count);
2817 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2818 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2820 " dirty:%lu writeback:%lu unstable:%lu\n"
2821 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2822 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2823 global_page_state(NR_ACTIVE_ANON),
2824 global_page_state(NR_INACTIVE_ANON),
2825 global_page_state(NR_ISOLATED_ANON),
2826 global_page_state(NR_ACTIVE_FILE),
2827 global_page_state(NR_INACTIVE_FILE),
2828 global_page_state(NR_ISOLATED_FILE),
2829 global_page_state(NR_UNEVICTABLE),
2830 global_page_state(NR_FILE_DIRTY),
2831 global_page_state(NR_WRITEBACK),
2832 global_page_state(NR_UNSTABLE_NFS),
2833 global_page_state(NR_FREE_PAGES),
2834 global_page_state(NR_SLAB_RECLAIMABLE),
2835 global_page_state(NR_SLAB_UNRECLAIMABLE),
2836 global_page_state(NR_FILE_MAPPED),
2837 global_page_state(NR_SHMEM),
2838 global_page_state(NR_PAGETABLE),
2839 global_page_state(NR_BOUNCE));
2841 for_each_populated_zone(zone) {
2844 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2852 " active_anon:%lukB"
2853 " inactive_anon:%lukB"
2854 " active_file:%lukB"
2855 " inactive_file:%lukB"
2856 " unevictable:%lukB"
2857 " isolated(anon):%lukB"
2858 " isolated(file):%lukB"
2865 " slab_reclaimable:%lukB"
2866 " slab_unreclaimable:%lukB"
2867 " kernel_stack:%lukB"
2871 " writeback_tmp:%lukB"
2872 " pages_scanned:%lu"
2873 " all_unreclaimable? %s"
2876 K(zone_page_state(zone, NR_FREE_PAGES)),
2877 K(min_wmark_pages(zone)),
2878 K(low_wmark_pages(zone)),
2879 K(high_wmark_pages(zone)),
2880 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2881 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2882 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2883 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2884 K(zone_page_state(zone, NR_UNEVICTABLE)),
2885 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2886 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2887 K(zone->present_pages),
2888 K(zone_page_state(zone, NR_MLOCK)),
2889 K(zone_page_state(zone, NR_FILE_DIRTY)),
2890 K(zone_page_state(zone, NR_WRITEBACK)),
2891 K(zone_page_state(zone, NR_FILE_MAPPED)),
2892 K(zone_page_state(zone, NR_SHMEM)),
2893 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2894 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2895 zone_page_state(zone, NR_KERNEL_STACK) *
2897 K(zone_page_state(zone, NR_PAGETABLE)),
2898 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2899 K(zone_page_state(zone, NR_BOUNCE)),
2900 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2901 zone->pages_scanned,
2902 (zone->all_unreclaimable ? "yes" : "no")
2904 printk("lowmem_reserve[]:");
2905 for (i = 0; i < MAX_NR_ZONES; i++)
2906 printk(" %lu", zone->lowmem_reserve[i]);
2910 for_each_populated_zone(zone) {
2911 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2913 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2916 printk("%s: ", zone->name);
2918 spin_lock_irqsave(&zone->lock, flags);
2919 for (order = 0; order < MAX_ORDER; order++) {
2920 nr[order] = zone->free_area[order].nr_free;
2921 total += nr[order] << order;
2923 spin_unlock_irqrestore(&zone->lock, flags);
2924 for (order = 0; order < MAX_ORDER; order++)
2925 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2926 printk("= %lukB\n", K(total));
2929 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2931 show_swap_cache_info();
2934 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2936 zoneref->zone = zone;
2937 zoneref->zone_idx = zone_idx(zone);
2941 * Builds allocation fallback zone lists.
2943 * Add all populated zones of a node to the zonelist.
2945 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2946 int nr_zones, enum zone_type zone_type)
2950 BUG_ON(zone_type >= MAX_NR_ZONES);
2955 zone = pgdat->node_zones + zone_type;
2956 if (populated_zone(zone)) {
2957 zoneref_set_zone(zone,
2958 &zonelist->_zonerefs[nr_zones++]);
2959 check_highest_zone(zone_type);
2962 } while (zone_type);
2969 * 0 = automatic detection of better ordering.
2970 * 1 = order by ([node] distance, -zonetype)
2971 * 2 = order by (-zonetype, [node] distance)
2973 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2974 * the same zonelist. So only NUMA can configure this param.
2976 #define ZONELIST_ORDER_DEFAULT 0
2977 #define ZONELIST_ORDER_NODE 1
2978 #define ZONELIST_ORDER_ZONE 2
2980 /* zonelist order in the kernel.
2981 * set_zonelist_order() will set this to NODE or ZONE.
2983 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2984 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2988 /* The value user specified ....changed by config */
2989 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2990 /* string for sysctl */
2991 #define NUMA_ZONELIST_ORDER_LEN 16
2992 char numa_zonelist_order[16] = "default";
2995 * interface for configure zonelist ordering.
2996 * command line option "numa_zonelist_order"
2997 * = "[dD]efault - default, automatic configuration.
2998 * = "[nN]ode - order by node locality, then by zone within node
2999 * = "[zZ]one - order by zone, then by locality within zone
3002 static int __parse_numa_zonelist_order(char *s)
3004 if (*s == 'd' || *s == 'D') {
3005 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3006 } else if (*s == 'n' || *s == 'N') {
3007 user_zonelist_order = ZONELIST_ORDER_NODE;
3008 } else if (*s == 'z' || *s == 'Z') {
3009 user_zonelist_order = ZONELIST_ORDER_ZONE;
3012 "Ignoring invalid numa_zonelist_order value: "
3019 static __init int setup_numa_zonelist_order(char *s)
3026 ret = __parse_numa_zonelist_order(s);
3028 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3032 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3035 * sysctl handler for numa_zonelist_order
3037 int numa_zonelist_order_handler(ctl_table *table, int write,
3038 void __user *buffer, size_t *length,
3041 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3043 static DEFINE_MUTEX(zl_order_mutex);
3045 mutex_lock(&zl_order_mutex);
3047 strcpy(saved_string, (char*)table->data);
3048 ret = proc_dostring(table, write, buffer, length, ppos);
3052 int oldval = user_zonelist_order;
3053 if (__parse_numa_zonelist_order((char*)table->data)) {
3055 * bogus value. restore saved string
3057 strncpy((char*)table->data, saved_string,
3058 NUMA_ZONELIST_ORDER_LEN);
3059 user_zonelist_order = oldval;
3060 } else if (oldval != user_zonelist_order) {
3061 mutex_lock(&zonelists_mutex);
3062 build_all_zonelists(NULL, NULL);
3063 mutex_unlock(&zonelists_mutex);
3067 mutex_unlock(&zl_order_mutex);
3072 #define MAX_NODE_LOAD (nr_online_nodes)
3073 static int node_load[MAX_NUMNODES];
3076 * find_next_best_node - find the next node that should appear in a given node's fallback list
3077 * @node: node whose fallback list we're appending
3078 * @used_node_mask: nodemask_t of already used nodes
3080 * We use a number of factors to determine which is the next node that should
3081 * appear on a given node's fallback list. The node should not have appeared
3082 * already in @node's fallback list, and it should be the next closest node
3083 * according to the distance array (which contains arbitrary distance values
3084 * from each node to each node in the system), and should also prefer nodes
3085 * with no CPUs, since presumably they'll have very little allocation pressure
3086 * on them otherwise.
3087 * It returns -1 if no node is found.
3089 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3092 int min_val = INT_MAX;
3094 const struct cpumask *tmp = cpumask_of_node(0);
3096 /* Use the local node if we haven't already */
3097 if (!node_isset(node, *used_node_mask)) {
3098 node_set(node, *used_node_mask);
3102 for_each_node_state(n, N_HIGH_MEMORY) {
3104 /* Don't want a node to appear more than once */
3105 if (node_isset(n, *used_node_mask))
3108 /* Use the distance array to find the distance */
3109 val = node_distance(node, n);
3111 /* Penalize nodes under us ("prefer the next node") */
3114 /* Give preference to headless and unused nodes */
3115 tmp = cpumask_of_node(n);
3116 if (!cpumask_empty(tmp))
3117 val += PENALTY_FOR_NODE_WITH_CPUS;
3119 /* Slight preference for less loaded node */
3120 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3121 val += node_load[n];
3123 if (val < min_val) {
3130 node_set(best_node, *used_node_mask);
3137 * Build zonelists ordered by node and zones within node.
3138 * This results in maximum locality--normal zone overflows into local
3139 * DMA zone, if any--but risks exhausting DMA zone.
3141 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3144 struct zonelist *zonelist;
3146 zonelist = &pgdat->node_zonelists[0];
3147 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3149 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3151 zonelist->_zonerefs[j].zone = NULL;
3152 zonelist->_zonerefs[j].zone_idx = 0;
3156 * Build gfp_thisnode zonelists
3158 static void build_thisnode_zonelists(pg_data_t *pgdat)
3161 struct zonelist *zonelist;
3163 zonelist = &pgdat->node_zonelists[1];
3164 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3165 zonelist->_zonerefs[j].zone = NULL;
3166 zonelist->_zonerefs[j].zone_idx = 0;
3170 * Build zonelists ordered by zone and nodes within zones.
3171 * This results in conserving DMA zone[s] until all Normal memory is
3172 * exhausted, but results in overflowing to remote node while memory
3173 * may still exist in local DMA zone.
3175 static int node_order[MAX_NUMNODES];
3177 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3180 int zone_type; /* needs to be signed */
3182 struct zonelist *zonelist;
3184 zonelist = &pgdat->node_zonelists[0];
3186 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3187 for (j = 0; j < nr_nodes; j++) {
3188 node = node_order[j];
3189 z = &NODE_DATA(node)->node_zones[zone_type];
3190 if (populated_zone(z)) {
3192 &zonelist->_zonerefs[pos++]);
3193 check_highest_zone(zone_type);
3197 zonelist->_zonerefs[pos].zone = NULL;
3198 zonelist->_zonerefs[pos].zone_idx = 0;
3201 static int default_zonelist_order(void)
3204 unsigned long low_kmem_size,total_size;
3208 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3209 * If they are really small and used heavily, the system can fall
3210 * into OOM very easily.
3211 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3213 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3216 for_each_online_node(nid) {
3217 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3218 z = &NODE_DATA(nid)->node_zones[zone_type];
3219 if (populated_zone(z)) {
3220 if (zone_type < ZONE_NORMAL)
3221 low_kmem_size += z->present_pages;
3222 total_size += z->present_pages;
3223 } else if (zone_type == ZONE_NORMAL) {
3225 * If any node has only lowmem, then node order
3226 * is preferred to allow kernel allocations
3227 * locally; otherwise, they can easily infringe
3228 * on other nodes when there is an abundance of
3229 * lowmem available to allocate from.
3231 return ZONELIST_ORDER_NODE;
3235 if (!low_kmem_size || /* there are no DMA area. */
3236 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3237 return ZONELIST_ORDER_NODE;
3239 * look into each node's config.
3240 * If there is a node whose DMA/DMA32 memory is very big area on
3241 * local memory, NODE_ORDER may be suitable.
3243 average_size = total_size /
3244 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3245 for_each_online_node(nid) {
3248 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3249 z = &NODE_DATA(nid)->node_zones[zone_type];
3250 if (populated_zone(z)) {
3251 if (zone_type < ZONE_NORMAL)
3252 low_kmem_size += z->present_pages;
3253 total_size += z->present_pages;
3256 if (low_kmem_size &&
3257 total_size > average_size && /* ignore small node */
3258 low_kmem_size > total_size * 70/100)
3259 return ZONELIST_ORDER_NODE;
3261 return ZONELIST_ORDER_ZONE;
3264 static void set_zonelist_order(void)
3266 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3267 current_zonelist_order = default_zonelist_order();
3269 current_zonelist_order = user_zonelist_order;
3272 static void build_zonelists(pg_data_t *pgdat)
3276 nodemask_t used_mask;
3277 int local_node, prev_node;
3278 struct zonelist *zonelist;
3279 int order = current_zonelist_order;
3281 /* initialize zonelists */
3282 for (i = 0; i < MAX_ZONELISTS; i++) {
3283 zonelist = pgdat->node_zonelists + i;
3284 zonelist->_zonerefs[0].zone = NULL;
3285 zonelist->_zonerefs[0].zone_idx = 0;
3288 /* NUMA-aware ordering of nodes */
3289 local_node = pgdat->node_id;
3290 load = nr_online_nodes;
3291 prev_node = local_node;
3292 nodes_clear(used_mask);
3294 memset(node_order, 0, sizeof(node_order));
3297 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3298 int distance = node_distance(local_node, node);
3301 * If another node is sufficiently far away then it is better
3302 * to reclaim pages in a zone before going off node.
3304 if (distance > RECLAIM_DISTANCE)
3305 zone_reclaim_mode = 1;
3308 * We don't want to pressure a particular node.
3309 * So adding penalty to the first node in same
3310 * distance group to make it round-robin.
3312 if (distance != node_distance(local_node, prev_node))
3313 node_load[node] = load;
3317 if (order == ZONELIST_ORDER_NODE)
3318 build_zonelists_in_node_order(pgdat, node);
3320 node_order[j++] = node; /* remember order */
3323 if (order == ZONELIST_ORDER_ZONE) {
3324 /* calculate node order -- i.e., DMA last! */
3325 build_zonelists_in_zone_order(pgdat, j);
3328 build_thisnode_zonelists(pgdat);
3331 /* Construct the zonelist performance cache - see further mmzone.h */
3332 static void build_zonelist_cache(pg_data_t *pgdat)
3334 struct zonelist *zonelist;
3335 struct zonelist_cache *zlc;
3338 zonelist = &pgdat->node_zonelists[0];
3339 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3340 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3341 for (z = zonelist->_zonerefs; z->zone; z++)
3342 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3345 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3347 * Return node id of node used for "local" allocations.
3348 * I.e., first node id of first zone in arg node's generic zonelist.
3349 * Used for initializing percpu 'numa_mem', which is used primarily
3350 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3352 int local_memory_node(int node)
3356 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3357 gfp_zone(GFP_KERNEL),
3364 #else /* CONFIG_NUMA */
3366 static void set_zonelist_order(void)
3368 current_zonelist_order = ZONELIST_ORDER_ZONE;
3371 static void build_zonelists(pg_data_t *pgdat)
3373 int node, local_node;
3375 struct zonelist *zonelist;
3377 local_node = pgdat->node_id;
3379 zonelist = &pgdat->node_zonelists[0];
3380 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3383 * Now we build the zonelist so that it contains the zones
3384 * of all the other nodes.
3385 * We don't want to pressure a particular node, so when
3386 * building the zones for node N, we make sure that the
3387 * zones coming right after the local ones are those from
3388 * node N+1 (modulo N)
3390 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3391 if (!node_online(node))
3393 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3396 for (node = 0; node < local_node; node++) {
3397 if (!node_online(node))
3399 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3403 zonelist->_zonerefs[j].zone = NULL;
3404 zonelist->_zonerefs[j].zone_idx = 0;
3407 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3408 static void build_zonelist_cache(pg_data_t *pgdat)
3410 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3413 #endif /* CONFIG_NUMA */
3416 * Boot pageset table. One per cpu which is going to be used for all
3417 * zones and all nodes. The parameters will be set in such a way
3418 * that an item put on a list will immediately be handed over to
3419 * the buddy list. This is safe since pageset manipulation is done
3420 * with interrupts disabled.
3422 * The boot_pagesets must be kept even after bootup is complete for
3423 * unused processors and/or zones. They do play a role for bootstrapping
3424 * hotplugged processors.
3426 * zoneinfo_show() and maybe other functions do
3427 * not check if the processor is online before following the pageset pointer.
3428 * Other parts of the kernel may not check if the zone is available.
3430 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3431 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3432 static void setup_zone_pageset(struct zone *zone);
3435 * Global mutex to protect against size modification of zonelists
3436 * as well as to serialize pageset setup for the new populated zone.
3438 DEFINE_MUTEX(zonelists_mutex);
3440 /* return values int ....just for stop_machine() */
3441 static int __build_all_zonelists(void *data)
3445 pg_data_t *self = data;
3448 memset(node_load, 0, sizeof(node_load));
3451 if (self && !node_online(self->node_id)) {
3452 build_zonelists(self);
3453 build_zonelist_cache(self);
3456 for_each_online_node(nid) {
3457 pg_data_t *pgdat = NODE_DATA(nid);
3459 build_zonelists(pgdat);
3460 build_zonelist_cache(pgdat);
3464 * Initialize the boot_pagesets that are going to be used
3465 * for bootstrapping processors. The real pagesets for
3466 * each zone will be allocated later when the per cpu
3467 * allocator is available.
3469 * boot_pagesets are used also for bootstrapping offline
3470 * cpus if the system is already booted because the pagesets
3471 * are needed to initialize allocators on a specific cpu too.
3472 * F.e. the percpu allocator needs the page allocator which
3473 * needs the percpu allocator in order to allocate its pagesets
3474 * (a chicken-egg dilemma).
3476 for_each_possible_cpu(cpu) {
3477 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3479 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3481 * We now know the "local memory node" for each node--
3482 * i.e., the node of the first zone in the generic zonelist.
3483 * Set up numa_mem percpu variable for on-line cpus. During
3484 * boot, only the boot cpu should be on-line; we'll init the
3485 * secondary cpus' numa_mem as they come on-line. During
3486 * node/memory hotplug, we'll fixup all on-line cpus.
3488 if (cpu_online(cpu))
3489 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3497 * Called with zonelists_mutex held always
3498 * unless system_state == SYSTEM_BOOTING.
3500 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3502 set_zonelist_order();
3504 if (system_state == SYSTEM_BOOTING) {
3505 __build_all_zonelists(NULL);
3506 mminit_verify_zonelist();
3507 cpuset_init_current_mems_allowed();
3509 /* we have to stop all cpus to guarantee there is no user
3511 #ifdef CONFIG_MEMORY_HOTPLUG
3513 setup_zone_pageset(zone);
3515 stop_machine(__build_all_zonelists, pgdat, NULL);
3516 /* cpuset refresh routine should be here */
3518 vm_total_pages = nr_free_pagecache_pages();
3520 * Disable grouping by mobility if the number of pages in the
3521 * system is too low to allow the mechanism to work. It would be
3522 * more accurate, but expensive to check per-zone. This check is
3523 * made on memory-hotadd so a system can start with mobility
3524 * disabled and enable it later
3526 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3527 page_group_by_mobility_disabled = 1;
3529 page_group_by_mobility_disabled = 0;
3531 printk("Built %i zonelists in %s order, mobility grouping %s. "
3532 "Total pages: %ld\n",
3534 zonelist_order_name[current_zonelist_order],
3535 page_group_by_mobility_disabled ? "off" : "on",
3538 printk("Policy zone: %s\n", zone_names[policy_zone]);
3543 * Helper functions to size the waitqueue hash table.
3544 * Essentially these want to choose hash table sizes sufficiently
3545 * large so that collisions trying to wait on pages are rare.
3546 * But in fact, the number of active page waitqueues on typical
3547 * systems is ridiculously low, less than 200. So this is even
3548 * conservative, even though it seems large.
3550 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3551 * waitqueues, i.e. the size of the waitq table given the number of pages.
3553 #define PAGES_PER_WAITQUEUE 256
3555 #ifndef CONFIG_MEMORY_HOTPLUG
3556 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3558 unsigned long size = 1;
3560 pages /= PAGES_PER_WAITQUEUE;
3562 while (size < pages)
3566 * Once we have dozens or even hundreds of threads sleeping
3567 * on IO we've got bigger problems than wait queue collision.
3568 * Limit the size of the wait table to a reasonable size.
3570 size = min(size, 4096UL);
3572 return max(size, 4UL);
3576 * A zone's size might be changed by hot-add, so it is not possible to determine
3577 * a suitable size for its wait_table. So we use the maximum size now.
3579 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3581 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3582 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3583 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3585 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3586 * or more by the traditional way. (See above). It equals:
3588 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3589 * ia64(16K page size) : = ( 8G + 4M)byte.
3590 * powerpc (64K page size) : = (32G +16M)byte.
3592 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3599 * This is an integer logarithm so that shifts can be used later
3600 * to extract the more random high bits from the multiplicative
3601 * hash function before the remainder is taken.
3603 static inline unsigned long wait_table_bits(unsigned long size)
3608 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3611 * Check if a pageblock contains reserved pages
3613 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3617 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3618 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3625 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3626 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3627 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3628 * higher will lead to a bigger reserve which will get freed as contiguous
3629 * blocks as reclaim kicks in
3631 static void setup_zone_migrate_reserve(struct zone *zone)
3633 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3635 unsigned long block_migratetype;
3639 * Get the start pfn, end pfn and the number of blocks to reserve
3640 * We have to be careful to be aligned to pageblock_nr_pages to
3641 * make sure that we always check pfn_valid for the first page in
3644 start_pfn = zone->zone_start_pfn;
3645 end_pfn = start_pfn + zone->spanned_pages;
3646 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3647 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3651 * Reserve blocks are generally in place to help high-order atomic
3652 * allocations that are short-lived. A min_free_kbytes value that
3653 * would result in more than 2 reserve blocks for atomic allocations
3654 * is assumed to be in place to help anti-fragmentation for the
3655 * future allocation of hugepages at runtime.
3657 reserve = min(2, reserve);
3659 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3660 if (!pfn_valid(pfn))
3662 page = pfn_to_page(pfn);
3664 /* Watch out for overlapping nodes */
3665 if (page_to_nid(page) != zone_to_nid(zone))
3668 block_migratetype = get_pageblock_migratetype(page);
3670 /* Only test what is necessary when the reserves are not met */
3673 * Blocks with reserved pages will never free, skip
3676 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3677 if (pageblock_is_reserved(pfn, block_end_pfn))
3680 /* If this block is reserved, account for it */
3681 if (block_migratetype == MIGRATE_RESERVE) {
3686 /* Suitable for reserving if this block is movable */
3687 if (block_migratetype == MIGRATE_MOVABLE) {
3688 set_pageblock_migratetype(page,
3690 move_freepages_block(zone, page,
3698 * If the reserve is met and this is a previous reserved block,
3701 if (block_migratetype == MIGRATE_RESERVE) {
3702 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3703 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3709 * Initially all pages are reserved - free ones are freed
3710 * up by free_all_bootmem() once the early boot process is
3711 * done. Non-atomic initialization, single-pass.
3713 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3714 unsigned long start_pfn, enum memmap_context context)
3717 unsigned long end_pfn = start_pfn + size;
3721 if (highest_memmap_pfn < end_pfn - 1)
3722 highest_memmap_pfn = end_pfn - 1;
3724 z = &NODE_DATA(nid)->node_zones[zone];
3725 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3727 * There can be holes in boot-time mem_map[]s
3728 * handed to this function. They do not
3729 * exist on hotplugged memory.
3731 if (context == MEMMAP_EARLY) {
3732 if (!early_pfn_valid(pfn))
3734 if (!early_pfn_in_nid(pfn, nid))
3737 page = pfn_to_page(pfn);
3738 set_page_links(page, zone, nid, pfn);
3739 mminit_verify_page_links(page, zone, nid, pfn);
3740 init_page_count(page);
3741 reset_page_mapcount(page);
3742 SetPageReserved(page);
3744 * Mark the block movable so that blocks are reserved for
3745 * movable at startup. This will force kernel allocations
3746 * to reserve their blocks rather than leaking throughout
3747 * the address space during boot when many long-lived
3748 * kernel allocations are made. Later some blocks near
3749 * the start are marked MIGRATE_RESERVE by
3750 * setup_zone_migrate_reserve()
3752 * bitmap is created for zone's valid pfn range. but memmap
3753 * can be created for invalid pages (for alignment)
3754 * check here not to call set_pageblock_migratetype() against
3757 if ((z->zone_start_pfn <= pfn)
3758 && (pfn < z->zone_start_pfn + z->spanned_pages)
3759 && !(pfn & (pageblock_nr_pages - 1)))
3760 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3762 INIT_LIST_HEAD(&page->lru);
3763 #ifdef WANT_PAGE_VIRTUAL
3764 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3765 if (!is_highmem_idx(zone))
3766 set_page_address(page, __va(pfn << PAGE_SHIFT));
3771 static void __meminit zone_init_free_lists(struct zone *zone)
3774 for_each_migratetype_order(order, t) {
3775 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3776 zone->free_area[order].nr_free = 0;
3780 #ifndef __HAVE_ARCH_MEMMAP_INIT
3781 #define memmap_init(size, nid, zone, start_pfn) \
3782 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3785 static int __meminit zone_batchsize(struct zone *zone)
3791 * The per-cpu-pages pools are set to around 1000th of the
3792 * size of the zone. But no more than 1/2 of a meg.
3794 * OK, so we don't know how big the cache is. So guess.
3796 batch = zone->present_pages / 1024;
3797 if (batch * PAGE_SIZE > 512 * 1024)
3798 batch = (512 * 1024) / PAGE_SIZE;
3799 batch /= 4; /* We effectively *= 4 below */
3804 * Clamp the batch to a 2^n - 1 value. Having a power
3805 * of 2 value was found to be more likely to have
3806 * suboptimal cache aliasing properties in some cases.
3808 * For example if 2 tasks are alternately allocating
3809 * batches of pages, one task can end up with a lot
3810 * of pages of one half of the possible page colors
3811 * and the other with pages of the other colors.
3813 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3818 /* The deferral and batching of frees should be suppressed under NOMMU
3821 * The problem is that NOMMU needs to be able to allocate large chunks
3822 * of contiguous memory as there's no hardware page translation to
3823 * assemble apparent contiguous memory from discontiguous pages.
3825 * Queueing large contiguous runs of pages for batching, however,
3826 * causes the pages to actually be freed in smaller chunks. As there
3827 * can be a significant delay between the individual batches being
3828 * recycled, this leads to the once large chunks of space being
3829 * fragmented and becoming unavailable for high-order allocations.
3835 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3837 struct per_cpu_pages *pcp;
3840 memset(p, 0, sizeof(*p));
3844 pcp->high = 6 * batch;
3845 pcp->batch = max(1UL, 1 * batch);
3846 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3847 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3851 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3852 * to the value high for the pageset p.
3855 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3858 struct per_cpu_pages *pcp;
3862 pcp->batch = max(1UL, high/4);
3863 if ((high/4) > (PAGE_SHIFT * 8))
3864 pcp->batch = PAGE_SHIFT * 8;
3867 static void __meminit setup_zone_pageset(struct zone *zone)
3871 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3873 for_each_possible_cpu(cpu) {
3874 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3876 setup_pageset(pcp, zone_batchsize(zone));
3878 if (percpu_pagelist_fraction)
3879 setup_pagelist_highmark(pcp,
3880 (zone->present_pages /
3881 percpu_pagelist_fraction));
3886 * Allocate per cpu pagesets and initialize them.
3887 * Before this call only boot pagesets were available.
3889 void __init setup_per_cpu_pageset(void)
3893 for_each_populated_zone(zone)
3894 setup_zone_pageset(zone);
3897 static noinline __init_refok
3898 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3901 struct pglist_data *pgdat = zone->zone_pgdat;
3905 * The per-page waitqueue mechanism uses hashed waitqueues
3908 zone->wait_table_hash_nr_entries =
3909 wait_table_hash_nr_entries(zone_size_pages);
3910 zone->wait_table_bits =
3911 wait_table_bits(zone->wait_table_hash_nr_entries);
3912 alloc_size = zone->wait_table_hash_nr_entries
3913 * sizeof(wait_queue_head_t);
3915 if (!slab_is_available()) {
3916 zone->wait_table = (wait_queue_head_t *)
3917 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3920 * This case means that a zone whose size was 0 gets new memory
3921 * via memory hot-add.
3922 * But it may be the case that a new node was hot-added. In
3923 * this case vmalloc() will not be able to use this new node's
3924 * memory - this wait_table must be initialized to use this new
3925 * node itself as well.
3926 * To use this new node's memory, further consideration will be
3929 zone->wait_table = vmalloc(alloc_size);
3931 if (!zone->wait_table)
3934 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3935 init_waitqueue_head(zone->wait_table + i);
3940 static __meminit void zone_pcp_init(struct zone *zone)
3943 * per cpu subsystem is not up at this point. The following code
3944 * relies on the ability of the linker to provide the
3945 * offset of a (static) per cpu variable into the per cpu area.
3947 zone->pageset = &boot_pageset;
3949 if (zone->present_pages)
3950 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3951 zone->name, zone->present_pages,
3952 zone_batchsize(zone));
3955 int __meminit init_currently_empty_zone(struct zone *zone,
3956 unsigned long zone_start_pfn,
3958 enum memmap_context context)
3960 struct pglist_data *pgdat = zone->zone_pgdat;
3962 ret = zone_wait_table_init(zone, size);
3965 pgdat->nr_zones = zone_idx(zone) + 1;
3967 zone->zone_start_pfn = zone_start_pfn;
3969 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3970 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3972 (unsigned long)zone_idx(zone),
3973 zone_start_pfn, (zone_start_pfn + size));
3975 zone_init_free_lists(zone);
3980 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
3981 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3983 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3984 * Architectures may implement their own version but if add_active_range()
3985 * was used and there are no special requirements, this is a convenient
3988 int __meminit __early_pfn_to_nid(unsigned long pfn)
3990 unsigned long start_pfn, end_pfn;
3993 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
3994 if (start_pfn <= pfn && pfn < end_pfn)
3996 /* This is a memory hole */
3999 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4001 int __meminit early_pfn_to_nid(unsigned long pfn)
4005 nid = __early_pfn_to_nid(pfn);
4008 /* just returns 0 */
4012 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4013 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4017 nid = __early_pfn_to_nid(pfn);
4018 if (nid >= 0 && nid != node)
4025 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4026 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4027 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4029 * If an architecture guarantees that all ranges registered with
4030 * add_active_ranges() contain no holes and may be freed, this
4031 * this function may be used instead of calling free_bootmem() manually.
4033 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4035 unsigned long start_pfn, end_pfn;
4038 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4039 start_pfn = min(start_pfn, max_low_pfn);
4040 end_pfn = min(end_pfn, max_low_pfn);
4042 if (start_pfn < end_pfn)
4043 free_bootmem_node(NODE_DATA(this_nid),
4044 PFN_PHYS(start_pfn),
4045 (end_pfn - start_pfn) << PAGE_SHIFT);
4050 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4051 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4053 * If an architecture guarantees that all ranges registered with
4054 * add_active_ranges() contain no holes and may be freed, this
4055 * function may be used instead of calling memory_present() manually.
4057 void __init sparse_memory_present_with_active_regions(int nid)
4059 unsigned long start_pfn, end_pfn;
4062 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4063 memory_present(this_nid, start_pfn, end_pfn);
4067 * get_pfn_range_for_nid - Return the start and end page frames for a node
4068 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4069 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4070 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4072 * It returns the start and end page frame of a node based on information
4073 * provided by an arch calling add_active_range(). If called for a node
4074 * with no available memory, a warning is printed and the start and end
4077 void __meminit get_pfn_range_for_nid(unsigned int nid,
4078 unsigned long *start_pfn, unsigned long *end_pfn)
4080 unsigned long this_start_pfn, this_end_pfn;
4086 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4087 *start_pfn = min(*start_pfn, this_start_pfn);
4088 *end_pfn = max(*end_pfn, this_end_pfn);
4091 if (*start_pfn == -1UL)
4096 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4097 * assumption is made that zones within a node are ordered in monotonic
4098 * increasing memory addresses so that the "highest" populated zone is used
4100 static void __init find_usable_zone_for_movable(void)
4103 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4104 if (zone_index == ZONE_MOVABLE)
4107 if (arch_zone_highest_possible_pfn[zone_index] >
4108 arch_zone_lowest_possible_pfn[zone_index])
4112 VM_BUG_ON(zone_index == -1);
4113 movable_zone = zone_index;
4117 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4118 * because it is sized independent of architecture. Unlike the other zones,
4119 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4120 * in each node depending on the size of each node and how evenly kernelcore
4121 * is distributed. This helper function adjusts the zone ranges
4122 * provided by the architecture for a given node by using the end of the
4123 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4124 * zones within a node are in order of monotonic increases memory addresses
4126 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4127 unsigned long zone_type,
4128 unsigned long node_start_pfn,
4129 unsigned long node_end_pfn,
4130 unsigned long *zone_start_pfn,
4131 unsigned long *zone_end_pfn)
4133 /* Only adjust if ZONE_MOVABLE is on this node */
4134 if (zone_movable_pfn[nid]) {
4135 /* Size ZONE_MOVABLE */
4136 if (zone_type == ZONE_MOVABLE) {
4137 *zone_start_pfn = zone_movable_pfn[nid];
4138 *zone_end_pfn = min(node_end_pfn,
4139 arch_zone_highest_possible_pfn[movable_zone]);
4141 /* Adjust for ZONE_MOVABLE starting within this range */
4142 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4143 *zone_end_pfn > zone_movable_pfn[nid]) {
4144 *zone_end_pfn = zone_movable_pfn[nid];
4146 /* Check if this whole range is within ZONE_MOVABLE */
4147 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4148 *zone_start_pfn = *zone_end_pfn;
4153 * Return the number of pages a zone spans in a node, including holes
4154 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4156 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4157 unsigned long zone_type,
4158 unsigned long *ignored)
4160 unsigned long node_start_pfn, node_end_pfn;
4161 unsigned long zone_start_pfn, zone_end_pfn;
4163 /* Get the start and end of the node and zone */
4164 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4165 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4166 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4167 adjust_zone_range_for_zone_movable(nid, zone_type,
4168 node_start_pfn, node_end_pfn,
4169 &zone_start_pfn, &zone_end_pfn);
4171 /* Check that this node has pages within the zone's required range */
4172 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4175 /* Move the zone boundaries inside the node if necessary */
4176 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4177 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4179 /* Return the spanned pages */
4180 return zone_end_pfn - zone_start_pfn;
4184 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4185 * then all holes in the requested range will be accounted for.
4187 unsigned long __meminit __absent_pages_in_range(int nid,
4188 unsigned long range_start_pfn,
4189 unsigned long range_end_pfn)
4191 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4192 unsigned long start_pfn, end_pfn;
4195 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4196 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4197 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4198 nr_absent -= end_pfn - start_pfn;
4204 * absent_pages_in_range - Return number of page frames in holes within a range
4205 * @start_pfn: The start PFN to start searching for holes
4206 * @end_pfn: The end PFN to stop searching for holes
4208 * It returns the number of pages frames in memory holes within a range.
4210 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4211 unsigned long end_pfn)
4213 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4216 /* Return the number of page frames in holes in a zone on a node */
4217 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4218 unsigned long zone_type,
4219 unsigned long *ignored)
4221 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4222 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4223 unsigned long node_start_pfn, node_end_pfn;
4224 unsigned long zone_start_pfn, zone_end_pfn;
4226 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4227 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4228 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4230 adjust_zone_range_for_zone_movable(nid, zone_type,
4231 node_start_pfn, node_end_pfn,
4232 &zone_start_pfn, &zone_end_pfn);
4233 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4236 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4237 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4238 unsigned long zone_type,
4239 unsigned long *zones_size)
4241 return zones_size[zone_type];
4244 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4245 unsigned long zone_type,
4246 unsigned long *zholes_size)
4251 return zholes_size[zone_type];
4254 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4256 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4257 unsigned long *zones_size, unsigned long *zholes_size)
4259 unsigned long realtotalpages, totalpages = 0;
4262 for (i = 0; i < MAX_NR_ZONES; i++)
4263 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4265 pgdat->node_spanned_pages = totalpages;
4267 realtotalpages = totalpages;
4268 for (i = 0; i < MAX_NR_ZONES; i++)
4270 zone_absent_pages_in_node(pgdat->node_id, i,
4272 pgdat->node_present_pages = realtotalpages;
4273 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4277 #ifndef CONFIG_SPARSEMEM
4279 * Calculate the size of the zone->blockflags rounded to an unsigned long
4280 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4281 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4282 * round what is now in bits to nearest long in bits, then return it in
4285 static unsigned long __init usemap_size(unsigned long zonesize)
4287 unsigned long usemapsize;
4289 usemapsize = roundup(zonesize, pageblock_nr_pages);
4290 usemapsize = usemapsize >> pageblock_order;
4291 usemapsize *= NR_PAGEBLOCK_BITS;
4292 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4294 return usemapsize / 8;
4297 static void __init setup_usemap(struct pglist_data *pgdat,
4298 struct zone *zone, unsigned long zonesize)
4300 unsigned long usemapsize = usemap_size(zonesize);
4301 zone->pageblock_flags = NULL;
4303 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4307 static inline void setup_usemap(struct pglist_data *pgdat,
4308 struct zone *zone, unsigned long zonesize) {}
4309 #endif /* CONFIG_SPARSEMEM */
4311 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4313 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4314 void __init set_pageblock_order(void)
4318 /* Check that pageblock_nr_pages has not already been setup */
4319 if (pageblock_order)
4322 if (HPAGE_SHIFT > PAGE_SHIFT)
4323 order = HUGETLB_PAGE_ORDER;
4325 order = MAX_ORDER - 1;
4328 * Assume the largest contiguous order of interest is a huge page.
4329 * This value may be variable depending on boot parameters on IA64 and
4332 pageblock_order = order;
4334 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4337 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4338 * is unused as pageblock_order is set at compile-time. See
4339 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4342 void __init set_pageblock_order(void)
4346 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4349 * Set up the zone data structures:
4350 * - mark all pages reserved
4351 * - mark all memory queues empty
4352 * - clear the memory bitmaps
4354 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4355 unsigned long *zones_size, unsigned long *zholes_size)
4358 int nid = pgdat->node_id;
4359 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4362 pgdat_resize_init(pgdat);
4363 pgdat->nr_zones = 0;
4364 init_waitqueue_head(&pgdat->kswapd_wait);
4365 pgdat->kswapd_max_order = 0;
4366 pgdat_page_cgroup_init(pgdat);
4368 for (j = 0; j < MAX_NR_ZONES; j++) {
4369 struct zone *zone = pgdat->node_zones + j;
4370 unsigned long size, realsize, memmap_pages;
4372 size = zone_spanned_pages_in_node(nid, j, zones_size);
4373 realsize = size - zone_absent_pages_in_node(nid, j,
4377 * Adjust realsize so that it accounts for how much memory
4378 * is used by this zone for memmap. This affects the watermark
4379 * and per-cpu initialisations
4382 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4383 if (realsize >= memmap_pages) {
4384 realsize -= memmap_pages;
4387 " %s zone: %lu pages used for memmap\n",
4388 zone_names[j], memmap_pages);
4391 " %s zone: %lu pages exceeds realsize %lu\n",
4392 zone_names[j], memmap_pages, realsize);
4394 /* Account for reserved pages */
4395 if (j == 0 && realsize > dma_reserve) {
4396 realsize -= dma_reserve;
4397 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4398 zone_names[0], dma_reserve);
4401 if (!is_highmem_idx(j))
4402 nr_kernel_pages += realsize;
4403 nr_all_pages += realsize;
4405 zone->spanned_pages = size;
4406 zone->present_pages = realsize;
4407 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
4408 zone->compact_cached_free_pfn = zone->zone_start_pfn +
4409 zone->spanned_pages;
4410 zone->compact_cached_free_pfn &= ~(pageblock_nr_pages-1);
4414 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4416 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4418 zone->name = zone_names[j];
4419 spin_lock_init(&zone->lock);
4420 spin_lock_init(&zone->lru_lock);
4421 zone_seqlock_init(zone);
4422 zone->zone_pgdat = pgdat;
4424 zone_pcp_init(zone);
4425 lruvec_init(&zone->lruvec, zone);
4426 zap_zone_vm_stats(zone);
4428 #ifdef CONFIG_MEMORY_ISOLATION
4429 zone->nr_pageblock_isolate = 0;
4434 set_pageblock_order();
4435 setup_usemap(pgdat, zone, size);
4436 ret = init_currently_empty_zone(zone, zone_start_pfn,
4437 size, MEMMAP_EARLY);
4439 memmap_init(size, nid, j, zone_start_pfn);
4440 zone_start_pfn += size;
4444 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4446 /* Skip empty nodes */
4447 if (!pgdat->node_spanned_pages)
4450 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4451 /* ia64 gets its own node_mem_map, before this, without bootmem */
4452 if (!pgdat->node_mem_map) {
4453 unsigned long size, start, end;
4457 * The zone's endpoints aren't required to be MAX_ORDER
4458 * aligned but the node_mem_map endpoints must be in order
4459 * for the buddy allocator to function correctly.
4461 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4462 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4463 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4464 size = (end - start) * sizeof(struct page);
4465 map = alloc_remap(pgdat->node_id, size);
4467 map = alloc_bootmem_node_nopanic(pgdat, size);
4468 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4470 #ifndef CONFIG_NEED_MULTIPLE_NODES
4472 * With no DISCONTIG, the global mem_map is just set as node 0's
4474 if (pgdat == NODE_DATA(0)) {
4475 mem_map = NODE_DATA(0)->node_mem_map;
4476 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4477 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4478 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4479 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4482 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4485 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4486 unsigned long node_start_pfn, unsigned long *zholes_size)
4488 pg_data_t *pgdat = NODE_DATA(nid);
4490 pgdat->node_id = nid;
4491 pgdat->node_start_pfn = node_start_pfn;
4492 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4494 alloc_node_mem_map(pgdat);
4495 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4496 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4497 nid, (unsigned long)pgdat,
4498 (unsigned long)pgdat->node_mem_map);
4501 free_area_init_core(pgdat, zones_size, zholes_size);
4504 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4506 #if MAX_NUMNODES > 1
4508 * Figure out the number of possible node ids.
4510 static void __init setup_nr_node_ids(void)
4513 unsigned int highest = 0;
4515 for_each_node_mask(node, node_possible_map)
4517 nr_node_ids = highest + 1;
4520 static inline void setup_nr_node_ids(void)
4526 * node_map_pfn_alignment - determine the maximum internode alignment
4528 * This function should be called after node map is populated and sorted.
4529 * It calculates the maximum power of two alignment which can distinguish
4532 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4533 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4534 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4535 * shifted, 1GiB is enough and this function will indicate so.
4537 * This is used to test whether pfn -> nid mapping of the chosen memory
4538 * model has fine enough granularity to avoid incorrect mapping for the
4539 * populated node map.
4541 * Returns the determined alignment in pfn's. 0 if there is no alignment
4542 * requirement (single node).
4544 unsigned long __init node_map_pfn_alignment(void)
4546 unsigned long accl_mask = 0, last_end = 0;
4547 unsigned long start, end, mask;
4551 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4552 if (!start || last_nid < 0 || last_nid == nid) {
4559 * Start with a mask granular enough to pin-point to the
4560 * start pfn and tick off bits one-by-one until it becomes
4561 * too coarse to separate the current node from the last.
4563 mask = ~((1 << __ffs(start)) - 1);
4564 while (mask && last_end <= (start & (mask << 1)))
4567 /* accumulate all internode masks */
4571 /* convert mask to number of pages */
4572 return ~accl_mask + 1;
4575 /* Find the lowest pfn for a node */
4576 static unsigned long __init find_min_pfn_for_node(int nid)
4578 unsigned long min_pfn = ULONG_MAX;
4579 unsigned long start_pfn;
4582 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4583 min_pfn = min(min_pfn, start_pfn);
4585 if (min_pfn == ULONG_MAX) {
4587 "Could not find start_pfn for node %d\n", nid);
4595 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4597 * It returns the minimum PFN based on information provided via
4598 * add_active_range().
4600 unsigned long __init find_min_pfn_with_active_regions(void)
4602 return find_min_pfn_for_node(MAX_NUMNODES);
4606 * early_calculate_totalpages()
4607 * Sum pages in active regions for movable zone.
4608 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4610 static unsigned long __init early_calculate_totalpages(void)
4612 unsigned long totalpages = 0;
4613 unsigned long start_pfn, end_pfn;
4616 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4617 unsigned long pages = end_pfn - start_pfn;
4619 totalpages += pages;
4621 node_set_state(nid, N_HIGH_MEMORY);
4627 * Find the PFN the Movable zone begins in each node. Kernel memory
4628 * is spread evenly between nodes as long as the nodes have enough
4629 * memory. When they don't, some nodes will have more kernelcore than
4632 static void __init find_zone_movable_pfns_for_nodes(void)
4635 unsigned long usable_startpfn;
4636 unsigned long kernelcore_node, kernelcore_remaining;
4637 /* save the state before borrow the nodemask */
4638 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4639 unsigned long totalpages = early_calculate_totalpages();
4640 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4643 * If movablecore was specified, calculate what size of
4644 * kernelcore that corresponds so that memory usable for
4645 * any allocation type is evenly spread. If both kernelcore
4646 * and movablecore are specified, then the value of kernelcore
4647 * will be used for required_kernelcore if it's greater than
4648 * what movablecore would have allowed.
4650 if (required_movablecore) {
4651 unsigned long corepages;
4654 * Round-up so that ZONE_MOVABLE is at least as large as what
4655 * was requested by the user
4657 required_movablecore =
4658 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4659 corepages = totalpages - required_movablecore;
4661 required_kernelcore = max(required_kernelcore, corepages);
4664 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4665 if (!required_kernelcore)
4668 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4669 find_usable_zone_for_movable();
4670 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4673 /* Spread kernelcore memory as evenly as possible throughout nodes */
4674 kernelcore_node = required_kernelcore / usable_nodes;
4675 for_each_node_state(nid, N_HIGH_MEMORY) {
4676 unsigned long start_pfn, end_pfn;
4679 * Recalculate kernelcore_node if the division per node
4680 * now exceeds what is necessary to satisfy the requested
4681 * amount of memory for the kernel
4683 if (required_kernelcore < kernelcore_node)
4684 kernelcore_node = required_kernelcore / usable_nodes;
4687 * As the map is walked, we track how much memory is usable
4688 * by the kernel using kernelcore_remaining. When it is
4689 * 0, the rest of the node is usable by ZONE_MOVABLE
4691 kernelcore_remaining = kernelcore_node;
4693 /* Go through each range of PFNs within this node */
4694 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4695 unsigned long size_pages;
4697 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4698 if (start_pfn >= end_pfn)
4701 /* Account for what is only usable for kernelcore */
4702 if (start_pfn < usable_startpfn) {
4703 unsigned long kernel_pages;
4704 kernel_pages = min(end_pfn, usable_startpfn)
4707 kernelcore_remaining -= min(kernel_pages,
4708 kernelcore_remaining);
4709 required_kernelcore -= min(kernel_pages,
4710 required_kernelcore);
4712 /* Continue if range is now fully accounted */
4713 if (end_pfn <= usable_startpfn) {
4716 * Push zone_movable_pfn to the end so
4717 * that if we have to rebalance
4718 * kernelcore across nodes, we will
4719 * not double account here
4721 zone_movable_pfn[nid] = end_pfn;
4724 start_pfn = usable_startpfn;
4728 * The usable PFN range for ZONE_MOVABLE is from
4729 * start_pfn->end_pfn. Calculate size_pages as the
4730 * number of pages used as kernelcore
4732 size_pages = end_pfn - start_pfn;
4733 if (size_pages > kernelcore_remaining)
4734 size_pages = kernelcore_remaining;
4735 zone_movable_pfn[nid] = start_pfn + size_pages;
4738 * Some kernelcore has been met, update counts and
4739 * break if the kernelcore for this node has been
4742 required_kernelcore -= min(required_kernelcore,
4744 kernelcore_remaining -= size_pages;
4745 if (!kernelcore_remaining)
4751 * If there is still required_kernelcore, we do another pass with one
4752 * less node in the count. This will push zone_movable_pfn[nid] further
4753 * along on the nodes that still have memory until kernelcore is
4757 if (usable_nodes && required_kernelcore > usable_nodes)
4760 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4761 for (nid = 0; nid < MAX_NUMNODES; nid++)
4762 zone_movable_pfn[nid] =
4763 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4766 /* restore the node_state */
4767 node_states[N_HIGH_MEMORY] = saved_node_state;
4770 /* Any regular memory on that node ? */
4771 static void __init check_for_regular_memory(pg_data_t *pgdat)
4773 #ifdef CONFIG_HIGHMEM
4774 enum zone_type zone_type;
4776 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4777 struct zone *zone = &pgdat->node_zones[zone_type];
4778 if (zone->present_pages) {
4779 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4787 * free_area_init_nodes - Initialise all pg_data_t and zone data
4788 * @max_zone_pfn: an array of max PFNs for each zone
4790 * This will call free_area_init_node() for each active node in the system.
4791 * Using the page ranges provided by add_active_range(), the size of each
4792 * zone in each node and their holes is calculated. If the maximum PFN
4793 * between two adjacent zones match, it is assumed that the zone is empty.
4794 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4795 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4796 * starts where the previous one ended. For example, ZONE_DMA32 starts
4797 * at arch_max_dma_pfn.
4799 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4801 unsigned long start_pfn, end_pfn;
4804 /* Record where the zone boundaries are */
4805 memset(arch_zone_lowest_possible_pfn, 0,
4806 sizeof(arch_zone_lowest_possible_pfn));
4807 memset(arch_zone_highest_possible_pfn, 0,
4808 sizeof(arch_zone_highest_possible_pfn));
4809 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4810 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4811 for (i = 1; i < MAX_NR_ZONES; i++) {
4812 if (i == ZONE_MOVABLE)
4814 arch_zone_lowest_possible_pfn[i] =
4815 arch_zone_highest_possible_pfn[i-1];
4816 arch_zone_highest_possible_pfn[i] =
4817 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4819 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4820 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4822 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4823 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4824 find_zone_movable_pfns_for_nodes();
4826 /* Print out the zone ranges */
4827 printk("Zone ranges:\n");
4828 for (i = 0; i < MAX_NR_ZONES; i++) {
4829 if (i == ZONE_MOVABLE)
4831 printk(KERN_CONT " %-8s ", zone_names[i]);
4832 if (arch_zone_lowest_possible_pfn[i] ==
4833 arch_zone_highest_possible_pfn[i])
4834 printk(KERN_CONT "empty\n");
4836 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4837 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4838 (arch_zone_highest_possible_pfn[i]
4839 << PAGE_SHIFT) - 1);
4842 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4843 printk("Movable zone start for each node\n");
4844 for (i = 0; i < MAX_NUMNODES; i++) {
4845 if (zone_movable_pfn[i])
4846 printk(" Node %d: %#010lx\n", i,
4847 zone_movable_pfn[i] << PAGE_SHIFT);
4850 /* Print out the early_node_map[] */
4851 printk("Early memory node ranges\n");
4852 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4853 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
4854 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
4856 /* Initialise every node */
4857 mminit_verify_pageflags_layout();
4858 setup_nr_node_ids();
4859 for_each_online_node(nid) {
4860 pg_data_t *pgdat = NODE_DATA(nid);
4861 free_area_init_node(nid, NULL,
4862 find_min_pfn_for_node(nid), NULL);
4864 /* Any memory on that node */
4865 if (pgdat->node_present_pages)
4866 node_set_state(nid, N_HIGH_MEMORY);
4867 check_for_regular_memory(pgdat);
4871 static int __init cmdline_parse_core(char *p, unsigned long *core)
4873 unsigned long long coremem;
4877 coremem = memparse(p, &p);
4878 *core = coremem >> PAGE_SHIFT;
4880 /* Paranoid check that UL is enough for the coremem value */
4881 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4887 * kernelcore=size sets the amount of memory for use for allocations that
4888 * cannot be reclaimed or migrated.
4890 static int __init cmdline_parse_kernelcore(char *p)
4892 return cmdline_parse_core(p, &required_kernelcore);
4896 * movablecore=size sets the amount of memory for use for allocations that
4897 * can be reclaimed or migrated.
4899 static int __init cmdline_parse_movablecore(char *p)
4901 return cmdline_parse_core(p, &required_movablecore);
4904 early_param("kernelcore", cmdline_parse_kernelcore);
4905 early_param("movablecore", cmdline_parse_movablecore);
4907 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4910 * set_dma_reserve - set the specified number of pages reserved in the first zone
4911 * @new_dma_reserve: The number of pages to mark reserved
4913 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4914 * In the DMA zone, a significant percentage may be consumed by kernel image
4915 * and other unfreeable allocations which can skew the watermarks badly. This
4916 * function may optionally be used to account for unfreeable pages in the
4917 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4918 * smaller per-cpu batchsize.
4920 void __init set_dma_reserve(unsigned long new_dma_reserve)
4922 dma_reserve = new_dma_reserve;
4925 void __init free_area_init(unsigned long *zones_size)
4927 free_area_init_node(0, zones_size,
4928 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4931 static int page_alloc_cpu_notify(struct notifier_block *self,
4932 unsigned long action, void *hcpu)
4934 int cpu = (unsigned long)hcpu;
4936 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4937 lru_add_drain_cpu(cpu);
4941 * Spill the event counters of the dead processor
4942 * into the current processors event counters.
4943 * This artificially elevates the count of the current
4946 vm_events_fold_cpu(cpu);
4949 * Zero the differential counters of the dead processor
4950 * so that the vm statistics are consistent.
4952 * This is only okay since the processor is dead and cannot
4953 * race with what we are doing.
4955 refresh_cpu_vm_stats(cpu);
4960 void __init page_alloc_init(void)
4962 hotcpu_notifier(page_alloc_cpu_notify, 0);
4966 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4967 * or min_free_kbytes changes.
4969 static void calculate_totalreserve_pages(void)
4971 struct pglist_data *pgdat;
4972 unsigned long reserve_pages = 0;
4973 enum zone_type i, j;
4975 for_each_online_pgdat(pgdat) {
4976 for (i = 0; i < MAX_NR_ZONES; i++) {
4977 struct zone *zone = pgdat->node_zones + i;
4978 unsigned long max = 0;
4980 /* Find valid and maximum lowmem_reserve in the zone */
4981 for (j = i; j < MAX_NR_ZONES; j++) {
4982 if (zone->lowmem_reserve[j] > max)
4983 max = zone->lowmem_reserve[j];
4986 /* we treat the high watermark as reserved pages. */
4987 max += high_wmark_pages(zone);
4989 if (max > zone->present_pages)
4990 max = zone->present_pages;
4991 reserve_pages += max;
4993 * Lowmem reserves are not available to
4994 * GFP_HIGHUSER page cache allocations and
4995 * kswapd tries to balance zones to their high
4996 * watermark. As a result, neither should be
4997 * regarded as dirtyable memory, to prevent a
4998 * situation where reclaim has to clean pages
4999 * in order to balance the zones.
5001 zone->dirty_balance_reserve = max;
5004 dirty_balance_reserve = reserve_pages;
5005 totalreserve_pages = reserve_pages;
5009 * setup_per_zone_lowmem_reserve - called whenever
5010 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5011 * has a correct pages reserved value, so an adequate number of
5012 * pages are left in the zone after a successful __alloc_pages().
5014 static void setup_per_zone_lowmem_reserve(void)
5016 struct pglist_data *pgdat;
5017 enum zone_type j, idx;
5019 for_each_online_pgdat(pgdat) {
5020 for (j = 0; j < MAX_NR_ZONES; j++) {
5021 struct zone *zone = pgdat->node_zones + j;
5022 unsigned long present_pages = zone->present_pages;
5024 zone->lowmem_reserve[j] = 0;
5028 struct zone *lower_zone;
5032 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5033 sysctl_lowmem_reserve_ratio[idx] = 1;
5035 lower_zone = pgdat->node_zones + idx;
5036 lower_zone->lowmem_reserve[j] = present_pages /
5037 sysctl_lowmem_reserve_ratio[idx];
5038 present_pages += lower_zone->present_pages;
5043 /* update totalreserve_pages */
5044 calculate_totalreserve_pages();
5047 static void __setup_per_zone_wmarks(void)
5049 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5050 unsigned long lowmem_pages = 0;
5052 unsigned long flags;
5054 /* Calculate total number of !ZONE_HIGHMEM pages */
5055 for_each_zone(zone) {
5056 if (!is_highmem(zone))
5057 lowmem_pages += zone->present_pages;
5060 for_each_zone(zone) {
5063 spin_lock_irqsave(&zone->lock, flags);
5064 tmp = (u64)pages_min * zone->present_pages;
5065 do_div(tmp, lowmem_pages);
5066 if (is_highmem(zone)) {
5068 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5069 * need highmem pages, so cap pages_min to a small
5072 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5073 * deltas controls asynch page reclaim, and so should
5074 * not be capped for highmem.
5078 min_pages = zone->present_pages / 1024;
5079 if (min_pages < SWAP_CLUSTER_MAX)
5080 min_pages = SWAP_CLUSTER_MAX;
5081 if (min_pages > 128)
5083 zone->watermark[WMARK_MIN] = min_pages;
5086 * If it's a lowmem zone, reserve a number of pages
5087 * proportionate to the zone's size.
5089 zone->watermark[WMARK_MIN] = tmp;
5092 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5093 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5095 zone->watermark[WMARK_MIN] += cma_wmark_pages(zone);
5096 zone->watermark[WMARK_LOW] += cma_wmark_pages(zone);
5097 zone->watermark[WMARK_HIGH] += cma_wmark_pages(zone);
5099 setup_zone_migrate_reserve(zone);
5100 spin_unlock_irqrestore(&zone->lock, flags);
5103 /* update totalreserve_pages */
5104 calculate_totalreserve_pages();
5108 * setup_per_zone_wmarks - called when min_free_kbytes changes
5109 * or when memory is hot-{added|removed}
5111 * Ensures that the watermark[min,low,high] values for each zone are set
5112 * correctly with respect to min_free_kbytes.
5114 void setup_per_zone_wmarks(void)
5116 mutex_lock(&zonelists_mutex);
5117 __setup_per_zone_wmarks();
5118 mutex_unlock(&zonelists_mutex);
5122 * The inactive anon list should be small enough that the VM never has to
5123 * do too much work, but large enough that each inactive page has a chance
5124 * to be referenced again before it is swapped out.
5126 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5127 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5128 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5129 * the anonymous pages are kept on the inactive list.
5132 * memory ratio inactive anon
5133 * -------------------------------------
5142 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5144 unsigned int gb, ratio;
5146 /* Zone size in gigabytes */
5147 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5149 ratio = int_sqrt(10 * gb);
5153 zone->inactive_ratio = ratio;
5156 static void __meminit setup_per_zone_inactive_ratio(void)
5161 calculate_zone_inactive_ratio(zone);
5165 * Initialise min_free_kbytes.
5167 * For small machines we want it small (128k min). For large machines
5168 * we want it large (64MB max). But it is not linear, because network
5169 * bandwidth does not increase linearly with machine size. We use
5171 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5172 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5188 int __meminit init_per_zone_wmark_min(void)
5190 unsigned long lowmem_kbytes;
5192 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5194 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5195 if (min_free_kbytes < 128)
5196 min_free_kbytes = 128;
5197 if (min_free_kbytes > 65536)
5198 min_free_kbytes = 65536;
5199 setup_per_zone_wmarks();
5200 refresh_zone_stat_thresholds();
5201 setup_per_zone_lowmem_reserve();
5202 setup_per_zone_inactive_ratio();
5205 module_init(init_per_zone_wmark_min)
5208 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5209 * that we can call two helper functions whenever min_free_kbytes
5212 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5213 void __user *buffer, size_t *length, loff_t *ppos)
5215 proc_dointvec(table, write, buffer, length, ppos);
5217 setup_per_zone_wmarks();
5222 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5223 void __user *buffer, size_t *length, loff_t *ppos)
5228 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5233 zone->min_unmapped_pages = (zone->present_pages *
5234 sysctl_min_unmapped_ratio) / 100;
5238 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5239 void __user *buffer, size_t *length, loff_t *ppos)
5244 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5249 zone->min_slab_pages = (zone->present_pages *
5250 sysctl_min_slab_ratio) / 100;
5256 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5257 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5258 * whenever sysctl_lowmem_reserve_ratio changes.
5260 * The reserve ratio obviously has absolutely no relation with the
5261 * minimum watermarks. The lowmem reserve ratio can only make sense
5262 * if in function of the boot time zone sizes.
5264 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5265 void __user *buffer, size_t *length, loff_t *ppos)
5267 proc_dointvec_minmax(table, write, buffer, length, ppos);
5268 setup_per_zone_lowmem_reserve();
5273 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5274 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5275 * can have before it gets flushed back to buddy allocator.
5278 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5279 void __user *buffer, size_t *length, loff_t *ppos)
5285 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5286 if (!write || (ret < 0))
5288 for_each_populated_zone(zone) {
5289 for_each_possible_cpu(cpu) {
5291 high = zone->present_pages / percpu_pagelist_fraction;
5292 setup_pagelist_highmark(
5293 per_cpu_ptr(zone->pageset, cpu), high);
5299 int hashdist = HASHDIST_DEFAULT;
5302 static int __init set_hashdist(char *str)
5306 hashdist = simple_strtoul(str, &str, 0);
5309 __setup("hashdist=", set_hashdist);
5313 * allocate a large system hash table from bootmem
5314 * - it is assumed that the hash table must contain an exact power-of-2
5315 * quantity of entries
5316 * - limit is the number of hash buckets, not the total allocation size
5318 void *__init alloc_large_system_hash(const char *tablename,
5319 unsigned long bucketsize,
5320 unsigned long numentries,
5323 unsigned int *_hash_shift,
5324 unsigned int *_hash_mask,
5325 unsigned long low_limit,
5326 unsigned long high_limit)
5328 unsigned long long max = high_limit;
5329 unsigned long log2qty, size;
5332 /* allow the kernel cmdline to have a say */
5334 /* round applicable memory size up to nearest megabyte */
5335 numentries = nr_kernel_pages;
5336 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5337 numentries >>= 20 - PAGE_SHIFT;
5338 numentries <<= 20 - PAGE_SHIFT;
5340 /* limit to 1 bucket per 2^scale bytes of low memory */
5341 if (scale > PAGE_SHIFT)
5342 numentries >>= (scale - PAGE_SHIFT);
5344 numentries <<= (PAGE_SHIFT - scale);
5346 /* Make sure we've got at least a 0-order allocation.. */
5347 if (unlikely(flags & HASH_SMALL)) {
5348 /* Makes no sense without HASH_EARLY */
5349 WARN_ON(!(flags & HASH_EARLY));
5350 if (!(numentries >> *_hash_shift)) {
5351 numentries = 1UL << *_hash_shift;
5352 BUG_ON(!numentries);
5354 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5355 numentries = PAGE_SIZE / bucketsize;
5357 numentries = roundup_pow_of_two(numentries);
5359 /* limit allocation size to 1/16 total memory by default */
5361 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5362 do_div(max, bucketsize);
5364 max = min(max, 0x80000000ULL);
5366 if (numentries < low_limit)
5367 numentries = low_limit;
5368 if (numentries > max)
5371 log2qty = ilog2(numentries);
5374 size = bucketsize << log2qty;
5375 if (flags & HASH_EARLY)
5376 table = alloc_bootmem_nopanic(size);
5378 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5381 * If bucketsize is not a power-of-two, we may free
5382 * some pages at the end of hash table which
5383 * alloc_pages_exact() automatically does
5385 if (get_order(size) < MAX_ORDER) {
5386 table = alloc_pages_exact(size, GFP_ATOMIC);
5387 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5390 } while (!table && size > PAGE_SIZE && --log2qty);
5393 panic("Failed to allocate %s hash table\n", tablename);
5395 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5398 ilog2(size) - PAGE_SHIFT,
5402 *_hash_shift = log2qty;
5404 *_hash_mask = (1 << log2qty) - 1;
5409 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5410 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5413 #ifdef CONFIG_SPARSEMEM
5414 return __pfn_to_section(pfn)->pageblock_flags;
5416 return zone->pageblock_flags;
5417 #endif /* CONFIG_SPARSEMEM */
5420 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5422 #ifdef CONFIG_SPARSEMEM
5423 pfn &= (PAGES_PER_SECTION-1);
5424 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5426 pfn = pfn - zone->zone_start_pfn;
5427 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5428 #endif /* CONFIG_SPARSEMEM */
5432 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5433 * @page: The page within the block of interest
5434 * @start_bitidx: The first bit of interest to retrieve
5435 * @end_bitidx: The last bit of interest
5436 * returns pageblock_bits flags
5438 unsigned long get_pageblock_flags_group(struct page *page,
5439 int start_bitidx, int end_bitidx)
5442 unsigned long *bitmap;
5443 unsigned long pfn, bitidx;
5444 unsigned long flags = 0;
5445 unsigned long value = 1;
5447 zone = page_zone(page);
5448 pfn = page_to_pfn(page);
5449 bitmap = get_pageblock_bitmap(zone, pfn);
5450 bitidx = pfn_to_bitidx(zone, pfn);
5452 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5453 if (test_bit(bitidx + start_bitidx, bitmap))
5460 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5461 * @page: The page within the block of interest
5462 * @start_bitidx: The first bit of interest
5463 * @end_bitidx: The last bit of interest
5464 * @flags: The flags to set
5466 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5467 int start_bitidx, int end_bitidx)
5470 unsigned long *bitmap;
5471 unsigned long pfn, bitidx;
5472 unsigned long value = 1;
5474 zone = page_zone(page);
5475 pfn = page_to_pfn(page);
5476 bitmap = get_pageblock_bitmap(zone, pfn);
5477 bitidx = pfn_to_bitidx(zone, pfn);
5478 VM_BUG_ON(pfn < zone->zone_start_pfn);
5479 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5481 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5483 __set_bit(bitidx + start_bitidx, bitmap);
5485 __clear_bit(bitidx + start_bitidx, bitmap);
5489 * This function checks whether pageblock includes unmovable pages or not.
5490 * If @count is not zero, it is okay to include less @count unmovable pages
5492 * PageLRU check wihtout isolation or lru_lock could race so that
5493 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5494 * expect this function should be exact.
5496 bool has_unmovable_pages(struct zone *zone, struct page *page, int count)
5498 unsigned long pfn, iter, found;
5502 * For avoiding noise data, lru_add_drain_all() should be called
5503 * If ZONE_MOVABLE, the zone never contains unmovable pages
5505 if (zone_idx(zone) == ZONE_MOVABLE)
5507 mt = get_pageblock_migratetype(page);
5508 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5511 pfn = page_to_pfn(page);
5512 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5513 unsigned long check = pfn + iter;
5515 if (!pfn_valid_within(check))
5518 page = pfn_to_page(check);
5520 * We can't use page_count without pin a page
5521 * because another CPU can free compound page.
5522 * This check already skips compound tails of THP
5523 * because their page->_count is zero at all time.
5525 if (!atomic_read(&page->_count)) {
5526 if (PageBuddy(page))
5527 iter += (1 << page_order(page)) - 1;
5534 * If there are RECLAIMABLE pages, we need to check it.
5535 * But now, memory offline itself doesn't call shrink_slab()
5536 * and it still to be fixed.
5539 * If the page is not RAM, page_count()should be 0.
5540 * we don't need more check. This is an _used_ not-movable page.
5542 * The problematic thing here is PG_reserved pages. PG_reserved
5543 * is set to both of a memory hole page and a _used_ kernel
5552 bool is_pageblock_removable_nolock(struct page *page)
5558 * We have to be careful here because we are iterating over memory
5559 * sections which are not zone aware so we might end up outside of
5560 * the zone but still within the section.
5561 * We have to take care about the node as well. If the node is offline
5562 * its NODE_DATA will be NULL - see page_zone.
5564 if (!node_online(page_to_nid(page)))
5567 zone = page_zone(page);
5568 pfn = page_to_pfn(page);
5569 if (zone->zone_start_pfn > pfn ||
5570 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5573 return !has_unmovable_pages(zone, page, 0);
5578 static unsigned long pfn_max_align_down(unsigned long pfn)
5580 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5581 pageblock_nr_pages) - 1);
5584 static unsigned long pfn_max_align_up(unsigned long pfn)
5586 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5587 pageblock_nr_pages));
5590 static struct page *
5591 __alloc_contig_migrate_alloc(struct page *page, unsigned long private,
5594 gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
5596 if (PageHighMem(page))
5597 gfp_mask |= __GFP_HIGHMEM;
5599 return alloc_page(gfp_mask);
5602 /* [start, end) must belong to a single zone. */
5603 static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5605 /* This function is based on compact_zone() from compaction.c. */
5607 unsigned long pfn = start;
5608 unsigned int tries = 0;
5611 struct compact_control cc = {
5612 .nr_migratepages = 0,
5614 .zone = page_zone(pfn_to_page(start)),
5617 INIT_LIST_HEAD(&cc.migratepages);
5619 migrate_prep_local();
5621 while (pfn < end || !list_empty(&cc.migratepages)) {
5622 if (fatal_signal_pending(current)) {
5627 if (list_empty(&cc.migratepages)) {
5628 cc.nr_migratepages = 0;
5629 pfn = isolate_migratepages_range(cc.zone, &cc,
5636 } else if (++tries == 5) {
5637 ret = ret < 0 ? ret : -EBUSY;
5641 ret = migrate_pages(&cc.migratepages,
5642 __alloc_contig_migrate_alloc,
5643 0, false, MIGRATE_SYNC);
5646 putback_lru_pages(&cc.migratepages);
5647 return ret > 0 ? 0 : ret;
5651 * Update zone's cma pages counter used for watermark level calculation.
5653 static inline void __update_cma_watermarks(struct zone *zone, int count)
5655 unsigned long flags;
5656 spin_lock_irqsave(&zone->lock, flags);
5657 zone->min_cma_pages += count;
5658 spin_unlock_irqrestore(&zone->lock, flags);
5659 setup_per_zone_wmarks();
5663 * Trigger memory pressure bump to reclaim some pages in order to be able to
5664 * allocate 'count' pages in single page units. Does similar work as
5665 *__alloc_pages_slowpath() function.
5667 static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
5669 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
5670 struct zonelist *zonelist = node_zonelist(0, gfp_mask);
5671 int did_some_progress = 0;
5675 * Increase level of watermarks to force kswapd do his job
5676 * to stabilise at new watermark level.
5678 __update_cma_watermarks(zone, count);
5680 /* Obey watermarks as if the page was being allocated */
5681 while (!zone_watermark_ok(zone, 0, low_wmark_pages(zone), 0, 0)) {
5682 wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));
5684 did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
5686 if (!did_some_progress) {
5687 /* Exhausted what can be done so it's blamo time */
5688 out_of_memory(zonelist, gfp_mask, order, NULL, false);
5692 /* Restore original watermark levels. */
5693 __update_cma_watermarks(zone, -count);
5699 * alloc_contig_range() -- tries to allocate given range of pages
5700 * @start: start PFN to allocate
5701 * @end: one-past-the-last PFN to allocate
5702 * @migratetype: migratetype of the underlaying pageblocks (either
5703 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5704 * in range must have the same migratetype and it must
5705 * be either of the two.
5707 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5708 * aligned, however it's the caller's responsibility to guarantee that
5709 * we are the only thread that changes migrate type of pageblocks the
5712 * The PFN range must belong to a single zone.
5714 * Returns zero on success or negative error code. On success all
5715 * pages which PFN is in [start, end) are allocated for the caller and
5716 * need to be freed with free_contig_range().
5718 int alloc_contig_range(unsigned long start, unsigned long end,
5719 unsigned migratetype)
5721 struct zone *zone = page_zone(pfn_to_page(start));
5722 unsigned long outer_start, outer_end;
5726 * What we do here is we mark all pageblocks in range as
5727 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5728 * have different sizes, and due to the way page allocator
5729 * work, we align the range to biggest of the two pages so
5730 * that page allocator won't try to merge buddies from
5731 * different pageblocks and change MIGRATE_ISOLATE to some
5732 * other migration type.
5734 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5735 * migrate the pages from an unaligned range (ie. pages that
5736 * we are interested in). This will put all the pages in
5737 * range back to page allocator as MIGRATE_ISOLATE.
5739 * When this is done, we take the pages in range from page
5740 * allocator removing them from the buddy system. This way
5741 * page allocator will never consider using them.
5743 * This lets us mark the pageblocks back as
5744 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5745 * aligned range but not in the unaligned, original range are
5746 * put back to page allocator so that buddy can use them.
5749 ret = start_isolate_page_range(pfn_max_align_down(start),
5750 pfn_max_align_up(end), migratetype);
5754 ret = __alloc_contig_migrate_range(start, end);
5759 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5760 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5761 * more, all pages in [start, end) are free in page allocator.
5762 * What we are going to do is to allocate all pages from
5763 * [start, end) (that is remove them from page allocator).
5765 * The only problem is that pages at the beginning and at the
5766 * end of interesting range may be not aligned with pages that
5767 * page allocator holds, ie. they can be part of higher order
5768 * pages. Because of this, we reserve the bigger range and
5769 * once this is done free the pages we are not interested in.
5771 * We don't have to hold zone->lock here because the pages are
5772 * isolated thus they won't get removed from buddy.
5775 lru_add_drain_all();
5779 outer_start = start;
5780 while (!PageBuddy(pfn_to_page(outer_start))) {
5781 if (++order >= MAX_ORDER) {
5785 outer_start &= ~0UL << order;
5788 /* Make sure the range is really isolated. */
5789 if (test_pages_isolated(outer_start, end)) {
5790 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5797 * Reclaim enough pages to make sure that contiguous allocation
5798 * will not starve the system.
5800 __reclaim_pages(zone, GFP_HIGHUSER_MOVABLE, end-start);
5802 /* Grab isolated pages from freelists. */
5803 outer_end = isolate_freepages_range(outer_start, end);
5809 /* Free head and tail (if any) */
5810 if (start != outer_start)
5811 free_contig_range(outer_start, start - outer_start);
5812 if (end != outer_end)
5813 free_contig_range(end, outer_end - end);
5816 undo_isolate_page_range(pfn_max_align_down(start),
5817 pfn_max_align_up(end), migratetype);
5821 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5823 for (; nr_pages--; ++pfn)
5824 __free_page(pfn_to_page(pfn));
5828 #ifdef CONFIG_MEMORY_HOTPLUG
5829 static int __meminit __zone_pcp_update(void *data)
5831 struct zone *zone = data;
5833 unsigned long batch = zone_batchsize(zone), flags;
5835 for_each_possible_cpu(cpu) {
5836 struct per_cpu_pageset *pset;
5837 struct per_cpu_pages *pcp;
5839 pset = per_cpu_ptr(zone->pageset, cpu);
5842 local_irq_save(flags);
5844 free_pcppages_bulk(zone, pcp->count, pcp);
5845 setup_pageset(pset, batch);
5846 local_irq_restore(flags);
5851 void __meminit zone_pcp_update(struct zone *zone)
5853 stop_machine(__zone_pcp_update, zone, NULL);
5857 #ifdef CONFIG_MEMORY_HOTREMOVE
5858 void zone_pcp_reset(struct zone *zone)
5860 unsigned long flags;
5862 /* avoid races with drain_pages() */
5863 local_irq_save(flags);
5864 if (zone->pageset != &boot_pageset) {
5865 free_percpu(zone->pageset);
5866 zone->pageset = &boot_pageset;
5868 local_irq_restore(flags);
5872 * All pages in the range must be isolated before calling this.
5875 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5881 unsigned long flags;
5882 /* find the first valid pfn */
5883 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5888 zone = page_zone(pfn_to_page(pfn));
5889 spin_lock_irqsave(&zone->lock, flags);
5891 while (pfn < end_pfn) {
5892 if (!pfn_valid(pfn)) {
5896 page = pfn_to_page(pfn);
5897 BUG_ON(page_count(page));
5898 BUG_ON(!PageBuddy(page));
5899 order = page_order(page);
5900 #ifdef CONFIG_DEBUG_VM
5901 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5902 pfn, 1 << order, end_pfn);
5904 list_del(&page->lru);
5905 rmv_page_order(page);
5906 zone->free_area[order].nr_free--;
5907 __mod_zone_page_state(zone, NR_FREE_PAGES,
5909 for (i = 0; i < (1 << order); i++)
5910 SetPageReserved((page+i));
5911 pfn += (1 << order);
5913 spin_unlock_irqrestore(&zone->lock, flags);
5917 #ifdef CONFIG_MEMORY_FAILURE
5918 bool is_free_buddy_page(struct page *page)
5920 struct zone *zone = page_zone(page);
5921 unsigned long pfn = page_to_pfn(page);
5922 unsigned long flags;
5925 spin_lock_irqsave(&zone->lock, flags);
5926 for (order = 0; order < MAX_ORDER; order++) {
5927 struct page *page_head = page - (pfn & ((1 << order) - 1));
5929 if (PageBuddy(page_head) && page_order(page_head) >= order)
5932 spin_unlock_irqrestore(&zone->lock, flags);
5934 return order < MAX_ORDER;
5938 static const struct trace_print_flags pageflag_names[] = {
5939 {1UL << PG_locked, "locked" },
5940 {1UL << PG_error, "error" },
5941 {1UL << PG_referenced, "referenced" },
5942 {1UL << PG_uptodate, "uptodate" },
5943 {1UL << PG_dirty, "dirty" },
5944 {1UL << PG_lru, "lru" },
5945 {1UL << PG_active, "active" },
5946 {1UL << PG_slab, "slab" },
5947 {1UL << PG_owner_priv_1, "owner_priv_1" },
5948 {1UL << PG_arch_1, "arch_1" },
5949 {1UL << PG_reserved, "reserved" },
5950 {1UL << PG_private, "private" },
5951 {1UL << PG_private_2, "private_2" },
5952 {1UL << PG_writeback, "writeback" },
5953 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5954 {1UL << PG_head, "head" },
5955 {1UL << PG_tail, "tail" },
5957 {1UL << PG_compound, "compound" },
5959 {1UL << PG_swapcache, "swapcache" },
5960 {1UL << PG_mappedtodisk, "mappedtodisk" },
5961 {1UL << PG_reclaim, "reclaim" },
5962 {1UL << PG_swapbacked, "swapbacked" },
5963 {1UL << PG_unevictable, "unevictable" },
5965 {1UL << PG_mlocked, "mlocked" },
5967 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5968 {1UL << PG_uncached, "uncached" },
5970 #ifdef CONFIG_MEMORY_FAILURE
5971 {1UL << PG_hwpoison, "hwpoison" },
5973 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5974 {1UL << PG_compound_lock, "compound_lock" },
5978 static void dump_page_flags(unsigned long flags)
5980 const char *delim = "";
5984 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
5986 printk(KERN_ALERT "page flags: %#lx(", flags);
5988 /* remove zone id */
5989 flags &= (1UL << NR_PAGEFLAGS) - 1;
5991 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
5993 mask = pageflag_names[i].mask;
5994 if ((flags & mask) != mask)
5998 printk("%s%s", delim, pageflag_names[i].name);
6002 /* check for left over flags */
6004 printk("%s%#lx", delim, flags);
6009 void dump_page(struct page *page)
6012 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6013 page, atomic_read(&page->_count), page_mapcount(page),
6014 page->mapping, page->index);
6015 dump_page_flags(page->flags);
6016 mem_cgroup_print_bad_page(page);