2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/memory.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/memcontrol.h>
59 #include <linux/prefetch.h>
60 #include <linux/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;
100 int percpu_pagelist_fraction;
101 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
103 #ifdef CONFIG_PM_SLEEP
105 * The following functions are used by the suspend/hibernate code to temporarily
106 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
107 * while devices are suspended. To avoid races with the suspend/hibernate code,
108 * they should always be called with pm_mutex held (gfp_allowed_mask also should
109 * only be modified with pm_mutex held, unless the suspend/hibernate code is
110 * guaranteed not to run in parallel with that modification).
113 static gfp_t saved_gfp_mask;
115 void pm_restore_gfp_mask(void)
117 WARN_ON(!mutex_is_locked(&pm_mutex));
118 if (saved_gfp_mask) {
119 gfp_allowed_mask = saved_gfp_mask;
124 void pm_restrict_gfp_mask(void)
126 WARN_ON(!mutex_is_locked(&pm_mutex));
127 WARN_ON(saved_gfp_mask);
128 saved_gfp_mask = gfp_allowed_mask;
129 gfp_allowed_mask &= ~GFP_IOFS;
132 bool pm_suspended_storage(void)
134 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
138 #endif /* CONFIG_PM_SLEEP */
140 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
141 int pageblock_order __read_mostly;
144 static void __free_pages_ok(struct page *page, unsigned int order);
147 * results with 256, 32 in the lowmem_reserve sysctl:
148 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
149 * 1G machine -> (16M dma, 784M normal, 224M high)
150 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
151 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
152 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
154 * TBD: should special case ZONE_DMA32 machines here - in those we normally
155 * don't need any ZONE_NORMAL reservation
157 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
158 #ifdef CONFIG_ZONE_DMA
161 #ifdef CONFIG_ZONE_DMA32
164 #ifdef CONFIG_HIGHMEM
170 EXPORT_SYMBOL(totalram_pages);
172 static char * const zone_names[MAX_NR_ZONES] = {
173 #ifdef CONFIG_ZONE_DMA
176 #ifdef CONFIG_ZONE_DMA32
180 #ifdef CONFIG_HIGHMEM
187 * Try to keep at least this much lowmem free. Do not allow normal
188 * allocations below this point, only high priority ones. Automatically
189 * tuned according to the amount of memory in the system.
191 int min_free_kbytes = 1024;
194 * Extra memory for the system to try freeing between the min and
195 * low watermarks. Useful for workloads that require low latency
196 * memory allocations in bursts larger than the normal gap between
199 int extra_free_kbytes;
201 static unsigned long __meminitdata nr_kernel_pages;
202 static unsigned long __meminitdata nr_all_pages;
203 static unsigned long __meminitdata dma_reserve;
205 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
206 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
207 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
208 static unsigned long __initdata required_kernelcore;
209 static unsigned long __initdata required_movablecore;
210 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
212 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
214 EXPORT_SYMBOL(movable_zone);
215 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
218 int nr_node_ids __read_mostly = MAX_NUMNODES;
219 int nr_online_nodes __read_mostly = 1;
220 EXPORT_SYMBOL(nr_node_ids);
221 EXPORT_SYMBOL(nr_online_nodes);
224 int page_group_by_mobility_disabled __read_mostly;
226 static 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_flg(struct page *page)
424 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
427 static inline void clear_page_guard_flg(struct page *page)
429 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
432 static inline void set_page_guard_flg(struct page *page) { }
433 static inline void clear_page_guard_flg(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_flg(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);
739 * permit the bootmem allocator to evade page validation on high-order frees
741 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
744 __ClearPageReserved(page);
745 set_page_count(page, 0);
746 set_page_refcounted(page);
752 for (loop = 0; loop < (1 << order); loop++) {
753 struct page *p = &page[loop];
755 if (loop + 1 < (1 << order))
757 __ClearPageReserved(p);
758 set_page_count(p, 0);
761 set_page_refcounted(page);
762 __free_pages(page, order);
768 * The order of subdivision here is critical for the IO subsystem.
769 * Please do not alter this order without good reasons and regression
770 * testing. Specifically, as large blocks of memory are subdivided,
771 * the order in which smaller blocks are delivered depends on the order
772 * they're subdivided in this function. This is the primary factor
773 * influencing the order in which pages are delivered to the IO
774 * subsystem according to empirical testing, and this is also justified
775 * by considering the behavior of a buddy system containing a single
776 * large block of memory acted on by a series of small allocations.
777 * This behavior is a critical factor in sglist merging's success.
781 static inline void expand(struct zone *zone, struct page *page,
782 int low, int high, struct free_area *area,
785 unsigned long size = 1 << high;
791 VM_BUG_ON(bad_range(zone, &page[size]));
793 #ifdef CONFIG_DEBUG_PAGEALLOC
794 if (high < debug_guardpage_minorder()) {
796 * Mark as guard pages (or page), that will allow to
797 * merge back to allocator when buddy will be freed.
798 * Corresponding page table entries will not be touched,
799 * pages will stay not present in virtual address space
801 INIT_LIST_HEAD(&page[size].lru);
802 set_page_guard_flg(&page[size]);
803 set_page_private(&page[size], high);
804 /* Guard pages are not available for any usage */
805 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high));
809 list_add(&page[size].lru, &area->free_list[migratetype]);
811 set_page_order(&page[size], high);
816 * This page is about to be returned from the page allocator
818 static inline int check_new_page(struct page *page)
820 if (unlikely(page_mapcount(page) |
821 (page->mapping != NULL) |
822 (atomic_read(&page->_count) != 0) |
823 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
824 (mem_cgroup_bad_page_check(page)))) {
831 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
835 for (i = 0; i < (1 << order); i++) {
836 struct page *p = page + i;
837 if (unlikely(check_new_page(p)))
841 set_page_private(page, 0);
842 set_page_refcounted(page);
844 arch_alloc_page(page, order);
845 kernel_map_pages(page, 1 << order, 1);
847 if (gfp_flags & __GFP_ZERO)
848 prep_zero_page(page, order, gfp_flags);
850 if (order && (gfp_flags & __GFP_COMP))
851 prep_compound_page(page, order);
857 * Go through the free lists for the given migratetype and remove
858 * the smallest available page from the freelists
861 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
864 unsigned int current_order;
865 struct free_area * area;
868 /* Find a page of the appropriate size in the preferred list */
869 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
870 area = &(zone->free_area[current_order]);
871 if (list_empty(&area->free_list[migratetype]))
874 page = list_entry(area->free_list[migratetype].next,
876 list_del(&page->lru);
877 rmv_page_order(page);
879 expand(zone, page, order, current_order, area, migratetype);
888 * This array describes the order lists are fallen back to when
889 * the free lists for the desirable migrate type are depleted
891 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
892 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
893 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
894 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
895 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
899 * Move the free pages in a range to the free lists of the requested type.
900 * Note that start_page and end_pages are not aligned on a pageblock
901 * boundary. If alignment is required, use move_freepages_block()
903 static int move_freepages(struct zone *zone,
904 struct page *start_page, struct page *end_page,
911 #ifndef CONFIG_HOLES_IN_ZONE
913 * page_zone is not safe to call in this context when
914 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
915 * anyway as we check zone boundaries in move_freepages_block().
916 * Remove at a later date when no bug reports exist related to
917 * grouping pages by mobility
919 BUG_ON(page_zone(start_page) != page_zone(end_page));
922 for (page = start_page; page <= end_page;) {
923 /* Make sure we are not inadvertently changing nodes */
924 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
926 if (!pfn_valid_within(page_to_pfn(page))) {
931 if (!PageBuddy(page)) {
936 order = page_order(page);
937 list_move(&page->lru,
938 &zone->free_area[order].free_list[migratetype]);
940 pages_moved += 1 << order;
946 static int move_freepages_block(struct zone *zone, struct page *page,
949 unsigned long start_pfn, end_pfn;
950 struct page *start_page, *end_page;
952 start_pfn = page_to_pfn(page);
953 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
954 start_page = pfn_to_page(start_pfn);
955 end_page = start_page + pageblock_nr_pages - 1;
956 end_pfn = start_pfn + pageblock_nr_pages - 1;
958 /* Do not cross zone boundaries */
959 if (start_pfn < zone->zone_start_pfn)
961 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
964 return move_freepages(zone, start_page, end_page, migratetype);
967 static void change_pageblock_range(struct page *pageblock_page,
968 int start_order, int migratetype)
970 int nr_pageblocks = 1 << (start_order - pageblock_order);
972 while (nr_pageblocks--) {
973 set_pageblock_migratetype(pageblock_page, migratetype);
974 pageblock_page += pageblock_nr_pages;
978 /* Remove an element from the buddy allocator from the fallback list */
979 static inline struct page *
980 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
982 struct free_area * area;
987 /* Find the largest possible block of pages in the other list */
988 for (current_order = MAX_ORDER-1; current_order >= order;
990 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
991 migratetype = fallbacks[start_migratetype][i];
993 /* MIGRATE_RESERVE handled later if necessary */
994 if (migratetype == MIGRATE_RESERVE)
997 area = &(zone->free_area[current_order]);
998 if (list_empty(&area->free_list[migratetype]))
1001 page = list_entry(area->free_list[migratetype].next,
1006 * If breaking a large block of pages, move all free
1007 * pages to the preferred allocation list. If falling
1008 * back for a reclaimable kernel allocation, be more
1009 * aggressive about taking ownership of free pages
1011 if (unlikely(current_order >= (pageblock_order >> 1)) ||
1012 start_migratetype == MIGRATE_RECLAIMABLE ||
1013 page_group_by_mobility_disabled) {
1014 unsigned long pages;
1015 pages = move_freepages_block(zone, page,
1018 /* Claim the whole block if over half of it is free */
1019 if (pages >= (1 << (pageblock_order-1)) ||
1020 page_group_by_mobility_disabled)
1021 set_pageblock_migratetype(page,
1024 migratetype = start_migratetype;
1027 /* Remove the page from the freelists */
1028 list_del(&page->lru);
1029 rmv_page_order(page);
1031 /* Take ownership for orders >= pageblock_order */
1032 if (current_order >= pageblock_order)
1033 change_pageblock_range(page, current_order,
1036 expand(zone, page, order, current_order, area, migratetype);
1038 trace_mm_page_alloc_extfrag(page, order, current_order,
1039 start_migratetype, migratetype);
1049 * Do the hard work of removing an element from the buddy allocator.
1050 * Call me with the zone->lock already held.
1052 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1058 page = __rmqueue_smallest(zone, order, migratetype);
1060 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1061 page = __rmqueue_fallback(zone, order, migratetype);
1064 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1065 * is used because __rmqueue_smallest is an inline function
1066 * and we want just one call site
1069 migratetype = MIGRATE_RESERVE;
1074 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1079 * Obtain a specified number of elements from the buddy allocator, all under
1080 * a single hold of the lock, for efficiency. Add them to the supplied list.
1081 * Returns the number of new pages which were placed at *list.
1083 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1084 unsigned long count, struct list_head *list,
1085 int migratetype, int cold)
1089 spin_lock(&zone->lock);
1090 for (i = 0; i < count; ++i) {
1091 struct page *page = __rmqueue(zone, order, migratetype);
1092 if (unlikely(page == NULL))
1096 * Split buddy pages returned by expand() are received here
1097 * in physical page order. The page is added to the callers and
1098 * list and the list head then moves forward. From the callers
1099 * perspective, the linked list is ordered by page number in
1100 * some conditions. This is useful for IO devices that can
1101 * merge IO requests if the physical pages are ordered
1104 if (likely(cold == 0))
1105 list_add(&page->lru, list);
1107 list_add_tail(&page->lru, list);
1108 set_page_private(page, migratetype);
1111 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1112 spin_unlock(&zone->lock);
1118 * Called from the vmstat counter updater to drain pagesets of this
1119 * currently executing processor on remote nodes after they have
1122 * Note that this function must be called with the thread pinned to
1123 * a single processor.
1125 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1127 unsigned long flags;
1130 local_irq_save(flags);
1131 if (pcp->count >= pcp->batch)
1132 to_drain = pcp->batch;
1134 to_drain = pcp->count;
1135 free_pcppages_bulk(zone, to_drain, pcp);
1136 pcp->count -= to_drain;
1137 local_irq_restore(flags);
1142 * Drain pages of the indicated processor.
1144 * The processor must either be the current processor and the
1145 * thread pinned to the current processor or a processor that
1148 static void drain_pages(unsigned int cpu)
1150 unsigned long flags;
1153 for_each_populated_zone(zone) {
1154 struct per_cpu_pageset *pset;
1155 struct per_cpu_pages *pcp;
1157 local_irq_save(flags);
1158 pset = per_cpu_ptr(zone->pageset, cpu);
1162 free_pcppages_bulk(zone, pcp->count, pcp);
1165 local_irq_restore(flags);
1170 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1172 void drain_local_pages(void *arg)
1174 drain_pages(smp_processor_id());
1178 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1180 void drain_all_pages(void)
1182 on_each_cpu(drain_local_pages, NULL, 1);
1185 #ifdef CONFIG_HIBERNATION
1187 void mark_free_pages(struct zone *zone)
1189 unsigned long pfn, max_zone_pfn;
1190 unsigned long flags;
1192 struct list_head *curr;
1194 if (!zone->spanned_pages)
1197 spin_lock_irqsave(&zone->lock, flags);
1199 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1200 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1201 if (pfn_valid(pfn)) {
1202 struct page *page = pfn_to_page(pfn);
1204 if (!swsusp_page_is_forbidden(page))
1205 swsusp_unset_page_free(page);
1208 for_each_migratetype_order(order, t) {
1209 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1212 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1213 for (i = 0; i < (1UL << order); i++)
1214 swsusp_set_page_free(pfn_to_page(pfn + i));
1217 spin_unlock_irqrestore(&zone->lock, flags);
1219 #endif /* CONFIG_PM */
1222 * Free a 0-order page
1223 * cold == 1 ? free a cold page : free a hot page
1225 void free_hot_cold_page(struct page *page, int cold)
1227 struct zone *zone = page_zone(page);
1228 struct per_cpu_pages *pcp;
1229 unsigned long flags;
1231 int wasMlocked = __TestClearPageMlocked(page);
1233 if (!free_pages_prepare(page, 0))
1236 migratetype = get_pageblock_migratetype(page);
1237 set_page_private(page, migratetype);
1238 local_irq_save(flags);
1239 if (unlikely(wasMlocked))
1240 free_page_mlock(page);
1241 __count_vm_event(PGFREE);
1244 * We only track unmovable, reclaimable and movable on pcp lists.
1245 * Free ISOLATE pages back to the allocator because they are being
1246 * offlined but treat RESERVE as movable pages so we can get those
1247 * areas back if necessary. Otherwise, we may have to free
1248 * excessively into the page allocator
1250 if (migratetype >= MIGRATE_PCPTYPES) {
1251 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1252 free_one_page(zone, page, 0, migratetype);
1255 migratetype = MIGRATE_MOVABLE;
1258 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1260 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1262 list_add(&page->lru, &pcp->lists[migratetype]);
1264 if (pcp->count >= pcp->high) {
1265 free_pcppages_bulk(zone, pcp->batch, pcp);
1266 pcp->count -= pcp->batch;
1270 local_irq_restore(flags);
1274 * Free a list of 0-order pages
1276 void free_hot_cold_page_list(struct list_head *list, int cold)
1278 struct page *page, *next;
1280 list_for_each_entry_safe(page, next, list, lru) {
1281 trace_mm_page_free_batched(page, cold);
1282 free_hot_cold_page(page, cold);
1287 * split_page takes a non-compound higher-order page, and splits it into
1288 * n (1<<order) sub-pages: page[0..n]
1289 * Each sub-page must be freed individually.
1291 * Note: this is probably too low level an operation for use in drivers.
1292 * Please consult with lkml before using this in your driver.
1294 void split_page(struct page *page, unsigned int order)
1298 VM_BUG_ON(PageCompound(page));
1299 VM_BUG_ON(!page_count(page));
1301 #ifdef CONFIG_KMEMCHECK
1303 * Split shadow pages too, because free(page[0]) would
1304 * otherwise free the whole shadow.
1306 if (kmemcheck_page_is_tracked(page))
1307 split_page(virt_to_page(page[0].shadow), order);
1310 for (i = 1; i < (1 << order); i++)
1311 set_page_refcounted(page + i);
1315 * Similar to split_page except the page is already free. As this is only
1316 * being used for migration, the migratetype of the block also changes.
1317 * As this is called with interrupts disabled, the caller is responsible
1318 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1321 * Note: this is probably too low level an operation for use in drivers.
1322 * Please consult with lkml before using this in your driver.
1324 int split_free_page(struct page *page)
1327 unsigned long watermark;
1330 BUG_ON(!PageBuddy(page));
1332 zone = page_zone(page);
1333 order = page_order(page);
1335 /* Obey watermarks as if the page was being allocated */
1336 watermark = low_wmark_pages(zone) + (1 << order);
1337 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1340 /* Remove page from free list */
1341 list_del(&page->lru);
1342 zone->free_area[order].nr_free--;
1343 rmv_page_order(page);
1344 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1346 /* Split into individual pages */
1347 set_page_refcounted(page);
1348 split_page(page, order);
1350 if (order >= pageblock_order - 1) {
1351 struct page *endpage = page + (1 << order) - 1;
1352 for (; page < endpage; page += pageblock_nr_pages)
1353 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1360 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1361 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1365 struct page *buffered_rmqueue(struct zone *preferred_zone,
1366 struct zone *zone, int order, gfp_t gfp_flags,
1369 unsigned long flags;
1371 int cold = !!(gfp_flags & __GFP_COLD);
1374 if (likely(order == 0)) {
1375 struct per_cpu_pages *pcp;
1376 struct list_head *list;
1378 local_irq_save(flags);
1379 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1380 list = &pcp->lists[migratetype];
1381 if (list_empty(list)) {
1382 pcp->count += rmqueue_bulk(zone, 0,
1385 if (unlikely(list_empty(list)))
1390 page = list_entry(list->prev, struct page, lru);
1392 page = list_entry(list->next, struct page, lru);
1394 list_del(&page->lru);
1397 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1399 * __GFP_NOFAIL is not to be used in new code.
1401 * All __GFP_NOFAIL callers should be fixed so that they
1402 * properly detect and handle allocation failures.
1404 * We most definitely don't want callers attempting to
1405 * allocate greater than order-1 page units with
1408 WARN_ON_ONCE(order > 1);
1410 spin_lock_irqsave(&zone->lock, flags);
1411 page = __rmqueue(zone, order, migratetype);
1412 spin_unlock(&zone->lock);
1415 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1418 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1419 zone_statistics(preferred_zone, zone, gfp_flags);
1420 local_irq_restore(flags);
1422 VM_BUG_ON(bad_range(zone, page));
1423 if (prep_new_page(page, order, gfp_flags))
1428 local_irq_restore(flags);
1432 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1433 #define ALLOC_WMARK_MIN WMARK_MIN
1434 #define ALLOC_WMARK_LOW WMARK_LOW
1435 #define ALLOC_WMARK_HIGH WMARK_HIGH
1436 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1438 /* Mask to get the watermark bits */
1439 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1441 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1442 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1443 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1445 #ifdef CONFIG_FAIL_PAGE_ALLOC
1448 struct fault_attr attr;
1450 u32 ignore_gfp_highmem;
1451 u32 ignore_gfp_wait;
1453 } fail_page_alloc = {
1454 .attr = FAULT_ATTR_INITIALIZER,
1455 .ignore_gfp_wait = 1,
1456 .ignore_gfp_highmem = 1,
1460 static int __init setup_fail_page_alloc(char *str)
1462 return setup_fault_attr(&fail_page_alloc.attr, str);
1464 __setup("fail_page_alloc=", setup_fail_page_alloc);
1466 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1468 if (order < fail_page_alloc.min_order)
1470 if (gfp_mask & __GFP_NOFAIL)
1472 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1474 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1477 return should_fail(&fail_page_alloc.attr, 1 << order);
1480 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1482 static int __init fail_page_alloc_debugfs(void)
1484 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1487 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1488 &fail_page_alloc.attr);
1490 return PTR_ERR(dir);
1492 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1493 &fail_page_alloc.ignore_gfp_wait))
1495 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1496 &fail_page_alloc.ignore_gfp_highmem))
1498 if (!debugfs_create_u32("min-order", mode, dir,
1499 &fail_page_alloc.min_order))
1504 debugfs_remove_recursive(dir);
1509 late_initcall(fail_page_alloc_debugfs);
1511 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1513 #else /* CONFIG_FAIL_PAGE_ALLOC */
1515 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1520 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1523 * Return true if free pages are above 'mark'. This takes into account the order
1524 * of the allocation.
1526 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1527 int classzone_idx, int alloc_flags, long free_pages)
1529 /* free_pages my go negative - that's OK */
1533 free_pages -= (1 << order) + 1;
1534 if (alloc_flags & ALLOC_HIGH)
1536 if (alloc_flags & ALLOC_HARDER)
1539 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1541 for (o = 0; o < order; o++) {
1542 /* At the next order, this order's pages become unavailable */
1543 free_pages -= z->free_area[o].nr_free << o;
1545 /* Require fewer higher order pages to be free */
1548 if (free_pages <= min)
1554 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1555 int classzone_idx, int alloc_flags)
1557 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1558 zone_page_state(z, NR_FREE_PAGES));
1561 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1562 int classzone_idx, int alloc_flags)
1564 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1566 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1567 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1569 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1575 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1576 * skip over zones that are not allowed by the cpuset, or that have
1577 * been recently (in last second) found to be nearly full. See further
1578 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1579 * that have to skip over a lot of full or unallowed zones.
1581 * If the zonelist cache is present in the passed in zonelist, then
1582 * returns a pointer to the allowed node mask (either the current
1583 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1585 * If the zonelist cache is not available for this zonelist, does
1586 * nothing and returns NULL.
1588 * If the fullzones BITMAP in the zonelist cache is stale (more than
1589 * a second since last zap'd) then we zap it out (clear its bits.)
1591 * We hold off even calling zlc_setup, until after we've checked the
1592 * first zone in the zonelist, on the theory that most allocations will
1593 * be satisfied from that first zone, so best to examine that zone as
1594 * quickly as we can.
1596 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1598 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1599 nodemask_t *allowednodes; /* zonelist_cache approximation */
1601 zlc = zonelist->zlcache_ptr;
1605 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1606 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1607 zlc->last_full_zap = jiffies;
1610 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1611 &cpuset_current_mems_allowed :
1612 &node_states[N_HIGH_MEMORY];
1613 return allowednodes;
1617 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1618 * if it is worth looking at further for free memory:
1619 * 1) Check that the zone isn't thought to be full (doesn't have its
1620 * bit set in the zonelist_cache fullzones BITMAP).
1621 * 2) Check that the zones node (obtained from the zonelist_cache
1622 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1623 * Return true (non-zero) if zone is worth looking at further, or
1624 * else return false (zero) if it is not.
1626 * This check -ignores- the distinction between various watermarks,
1627 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1628 * found to be full for any variation of these watermarks, it will
1629 * be considered full for up to one second by all requests, unless
1630 * we are so low on memory on all allowed nodes that we are forced
1631 * into the second scan of the zonelist.
1633 * In the second scan we ignore this zonelist cache and exactly
1634 * apply the watermarks to all zones, even it is slower to do so.
1635 * We are low on memory in the second scan, and should leave no stone
1636 * unturned looking for a free page.
1638 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1639 nodemask_t *allowednodes)
1641 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1642 int i; /* index of *z in zonelist zones */
1643 int n; /* node that zone *z is on */
1645 zlc = zonelist->zlcache_ptr;
1649 i = z - zonelist->_zonerefs;
1652 /* This zone is worth trying if it is allowed but not full */
1653 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1657 * Given 'z' scanning a zonelist, set the corresponding bit in
1658 * zlc->fullzones, so that subsequent attempts to allocate a page
1659 * from that zone don't waste time re-examining it.
1661 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1663 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1664 int i; /* index of *z in zonelist zones */
1666 zlc = zonelist->zlcache_ptr;
1670 i = z - zonelist->_zonerefs;
1672 set_bit(i, zlc->fullzones);
1676 * clear all zones full, called after direct reclaim makes progress so that
1677 * a zone that was recently full is not skipped over for up to a second
1679 static void zlc_clear_zones_full(struct zonelist *zonelist)
1681 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1683 zlc = zonelist->zlcache_ptr;
1687 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1690 #else /* CONFIG_NUMA */
1692 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1697 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1698 nodemask_t *allowednodes)
1703 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1707 static void zlc_clear_zones_full(struct zonelist *zonelist)
1710 #endif /* CONFIG_NUMA */
1713 * get_page_from_freelist goes through the zonelist trying to allocate
1716 static struct page *
1717 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1718 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1719 struct zone *preferred_zone, int migratetype)
1722 struct page *page = NULL;
1725 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1726 int zlc_active = 0; /* set if using zonelist_cache */
1727 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1729 classzone_idx = zone_idx(preferred_zone);
1732 * Scan zonelist, looking for a zone with enough free.
1733 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1735 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1736 high_zoneidx, nodemask) {
1737 if (NUMA_BUILD && zlc_active &&
1738 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1740 if ((alloc_flags & ALLOC_CPUSET) &&
1741 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1744 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1745 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1749 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1750 if (zone_watermark_ok(zone, order, mark,
1751 classzone_idx, alloc_flags))
1754 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1756 * we do zlc_setup if there are multiple nodes
1757 * and before considering the first zone allowed
1760 allowednodes = zlc_setup(zonelist, alloc_flags);
1765 if (zone_reclaim_mode == 0)
1766 goto this_zone_full;
1769 * As we may have just activated ZLC, check if the first
1770 * eligible zone has failed zone_reclaim recently.
1772 if (NUMA_BUILD && zlc_active &&
1773 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1776 ret = zone_reclaim(zone, gfp_mask, order);
1778 case ZONE_RECLAIM_NOSCAN:
1781 case ZONE_RECLAIM_FULL:
1782 /* scanned but unreclaimable */
1785 /* did we reclaim enough */
1786 if (!zone_watermark_ok(zone, order, mark,
1787 classzone_idx, alloc_flags))
1788 goto this_zone_full;
1793 page = buffered_rmqueue(preferred_zone, zone, order,
1794 gfp_mask, migratetype);
1799 zlc_mark_zone_full(zonelist, z);
1802 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1803 /* Disable zlc cache for second zonelist scan */
1811 * Large machines with many possible nodes should not always dump per-node
1812 * meminfo in irq context.
1814 static inline bool should_suppress_show_mem(void)
1819 ret = in_interrupt();
1824 static DEFINE_RATELIMIT_STATE(nopage_rs,
1825 DEFAULT_RATELIMIT_INTERVAL,
1826 DEFAULT_RATELIMIT_BURST);
1828 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1830 unsigned int filter = SHOW_MEM_FILTER_NODES;
1832 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1833 debug_guardpage_minorder() > 0)
1837 * This documents exceptions given to allocations in certain
1838 * contexts that are allowed to allocate outside current's set
1841 if (!(gfp_mask & __GFP_NOMEMALLOC))
1842 if (test_thread_flag(TIF_MEMDIE) ||
1843 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1844 filter &= ~SHOW_MEM_FILTER_NODES;
1845 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1846 filter &= ~SHOW_MEM_FILTER_NODES;
1849 struct va_format vaf;
1852 va_start(args, fmt);
1857 pr_warn("%pV", &vaf);
1862 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1863 current->comm, order, gfp_mask);
1866 if (!should_suppress_show_mem())
1871 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1872 unsigned long did_some_progress,
1873 unsigned long pages_reclaimed)
1875 /* Do not loop if specifically requested */
1876 if (gfp_mask & __GFP_NORETRY)
1879 /* Always retry if specifically requested */
1880 if (gfp_mask & __GFP_NOFAIL)
1884 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
1885 * making forward progress without invoking OOM. Suspend also disables
1886 * storage devices so kswapd will not help. Bail if we are suspending.
1888 if (!did_some_progress && pm_suspended_storage())
1892 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1893 * means __GFP_NOFAIL, but that may not be true in other
1896 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1900 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1901 * specified, then we retry until we no longer reclaim any pages
1902 * (above), or we've reclaimed an order of pages at least as
1903 * large as the allocation's order. In both cases, if the
1904 * allocation still fails, we stop retrying.
1906 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1912 static inline struct page *
1913 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1914 struct zonelist *zonelist, enum zone_type high_zoneidx,
1915 nodemask_t *nodemask, struct zone *preferred_zone,
1920 /* Acquire the OOM killer lock for the zones in zonelist */
1921 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1922 schedule_timeout_uninterruptible(1);
1927 * Go through the zonelist yet one more time, keep very high watermark
1928 * here, this is only to catch a parallel oom killing, we must fail if
1929 * we're still under heavy pressure.
1931 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1932 order, zonelist, high_zoneidx,
1933 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1934 preferred_zone, migratetype);
1938 if (!(gfp_mask & __GFP_NOFAIL)) {
1939 /* The OOM killer will not help higher order allocs */
1940 if (order > PAGE_ALLOC_COSTLY_ORDER)
1942 /* The OOM killer does not needlessly kill tasks for lowmem */
1943 if (high_zoneidx < ZONE_NORMAL)
1946 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1947 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1948 * The caller should handle page allocation failure by itself if
1949 * it specifies __GFP_THISNODE.
1950 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1952 if (gfp_mask & __GFP_THISNODE)
1955 /* Exhausted what can be done so it's blamo time */
1956 out_of_memory(zonelist, gfp_mask, order, nodemask);
1959 clear_zonelist_oom(zonelist, gfp_mask);
1963 #ifdef CONFIG_COMPACTION
1964 /* Try memory compaction for high-order allocations before reclaim */
1965 static struct page *
1966 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1967 struct zonelist *zonelist, enum zone_type high_zoneidx,
1968 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1969 int migratetype, unsigned long *did_some_progress,
1970 bool sync_migration)
1974 if (!order || compaction_deferred(preferred_zone))
1977 current->flags |= PF_MEMALLOC;
1978 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1979 nodemask, sync_migration);
1980 current->flags &= ~PF_MEMALLOC;
1981 if (*did_some_progress != COMPACT_SKIPPED) {
1983 /* Page migration frees to the PCP lists but we want merging */
1984 drain_pages(get_cpu());
1987 page = get_page_from_freelist(gfp_mask, nodemask,
1988 order, zonelist, high_zoneidx,
1989 alloc_flags, preferred_zone,
1992 preferred_zone->compact_considered = 0;
1993 preferred_zone->compact_defer_shift = 0;
1994 count_vm_event(COMPACTSUCCESS);
1999 * It's bad if compaction run occurs and fails.
2000 * The most likely reason is that pages exist,
2001 * but not enough to satisfy watermarks.
2003 count_vm_event(COMPACTFAIL);
2004 defer_compaction(preferred_zone);
2012 static inline struct page *
2013 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2014 struct zonelist *zonelist, enum zone_type high_zoneidx,
2015 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2016 int migratetype, unsigned long *did_some_progress,
2017 bool sync_migration)
2021 #endif /* CONFIG_COMPACTION */
2023 /* The really slow allocator path where we enter direct reclaim */
2024 static inline struct page *
2025 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2026 struct zonelist *zonelist, enum zone_type high_zoneidx,
2027 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2028 int migratetype, unsigned long *did_some_progress)
2030 struct page *page = NULL;
2031 struct reclaim_state reclaim_state;
2032 bool drained = false;
2036 /* We now go into synchronous reclaim */
2037 cpuset_memory_pressure_bump();
2038 current->flags |= PF_MEMALLOC;
2039 lockdep_set_current_reclaim_state(gfp_mask);
2040 reclaim_state.reclaimed_slab = 0;
2041 current->reclaim_state = &reclaim_state;
2043 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2045 current->reclaim_state = NULL;
2046 lockdep_clear_current_reclaim_state();
2047 current->flags &= ~PF_MEMALLOC;
2051 if (unlikely(!(*did_some_progress)))
2054 /* After successful reclaim, reconsider all zones for allocation */
2056 zlc_clear_zones_full(zonelist);
2059 page = get_page_from_freelist(gfp_mask, nodemask, order,
2060 zonelist, high_zoneidx,
2061 alloc_flags, preferred_zone,
2065 * If an allocation failed after direct reclaim, it could be because
2066 * pages are pinned on the per-cpu lists. Drain them and try again
2068 if (!page && !drained) {
2078 * This is called in the allocator slow-path if the allocation request is of
2079 * sufficient urgency to ignore watermarks and take other desperate measures
2081 static inline struct page *
2082 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2083 struct zonelist *zonelist, enum zone_type high_zoneidx,
2084 nodemask_t *nodemask, struct zone *preferred_zone,
2090 page = get_page_from_freelist(gfp_mask, nodemask, order,
2091 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2092 preferred_zone, migratetype);
2094 if (!page && gfp_mask & __GFP_NOFAIL)
2095 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2096 } while (!page && (gfp_mask & __GFP_NOFAIL));
2102 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2103 enum zone_type high_zoneidx,
2104 enum zone_type classzone_idx)
2109 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2110 wakeup_kswapd(zone, order, classzone_idx);
2114 gfp_to_alloc_flags(gfp_t gfp_mask)
2116 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2117 const gfp_t wait = gfp_mask & __GFP_WAIT;
2119 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2120 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2123 * The caller may dip into page reserves a bit more if the caller
2124 * cannot run direct reclaim, or if the caller has realtime scheduling
2125 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2126 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2128 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2132 * Not worth trying to allocate harder for
2133 * __GFP_NOMEMALLOC even if it can't schedule.
2135 if (!(gfp_mask & __GFP_NOMEMALLOC))
2136 alloc_flags |= ALLOC_HARDER;
2138 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2139 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2141 alloc_flags &= ~ALLOC_CPUSET;
2142 } else if (unlikely(rt_task(current)) && !in_interrupt())
2143 alloc_flags |= ALLOC_HARDER;
2145 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2146 if (!in_interrupt() &&
2147 ((current->flags & PF_MEMALLOC) ||
2148 unlikely(test_thread_flag(TIF_MEMDIE))))
2149 alloc_flags |= ALLOC_NO_WATERMARKS;
2155 static inline struct page *
2156 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2157 struct zonelist *zonelist, enum zone_type high_zoneidx,
2158 nodemask_t *nodemask, struct zone *preferred_zone,
2161 const gfp_t wait = gfp_mask & __GFP_WAIT;
2162 struct page *page = NULL;
2164 unsigned long pages_reclaimed = 0;
2165 unsigned long did_some_progress;
2166 bool sync_migration = false;
2169 * In the slowpath, we sanity check order to avoid ever trying to
2170 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2171 * be using allocators in order of preference for an area that is
2174 if (order >= MAX_ORDER) {
2175 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2180 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2181 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2182 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2183 * using a larger set of nodes after it has established that the
2184 * allowed per node queues are empty and that nodes are
2187 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2191 if (!(gfp_mask & __GFP_NO_KSWAPD))
2192 wake_all_kswapd(order, zonelist, high_zoneidx,
2193 zone_idx(preferred_zone));
2196 * OK, we're below the kswapd watermark and have kicked background
2197 * reclaim. Now things get more complex, so set up alloc_flags according
2198 * to how we want to proceed.
2200 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2203 * Find the true preferred zone if the allocation is unconstrained by
2206 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2207 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2211 /* This is the last chance, in general, before the goto nopage. */
2212 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2213 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2214 preferred_zone, migratetype);
2218 /* Allocate without watermarks if the context allows */
2219 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2220 page = __alloc_pages_high_priority(gfp_mask, order,
2221 zonelist, high_zoneidx, nodemask,
2222 preferred_zone, migratetype);
2227 /* Atomic allocations - we can't balance anything */
2231 /* Avoid recursion of direct reclaim */
2232 if (current->flags & PF_MEMALLOC)
2235 /* Avoid allocations with no watermarks from looping endlessly */
2236 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2240 * Try direct compaction. The first pass is asynchronous. Subsequent
2241 * attempts after direct reclaim are synchronous
2243 page = __alloc_pages_direct_compact(gfp_mask, order,
2244 zonelist, high_zoneidx,
2246 alloc_flags, preferred_zone,
2247 migratetype, &did_some_progress,
2253 * Do not use sync migration if __GFP_NO_KSWAPD is used to indicate
2254 * the system should not be heavily disrupted. In practice, this is
2255 * to avoid THP callers being stalled in writeback during migration
2256 * as it's preferable for the the allocations to fail than to stall
2258 sync_migration = !(gfp_mask & __GFP_NO_KSWAPD);
2260 /* Try direct reclaim and then allocating */
2261 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2262 zonelist, high_zoneidx,
2264 alloc_flags, preferred_zone,
2265 migratetype, &did_some_progress);
2270 * If we failed to make any progress reclaiming, then we are
2271 * running out of options and have to consider going OOM
2273 if (!did_some_progress) {
2274 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2275 if (oom_killer_disabled)
2277 page = __alloc_pages_may_oom(gfp_mask, order,
2278 zonelist, high_zoneidx,
2279 nodemask, preferred_zone,
2284 if (!(gfp_mask & __GFP_NOFAIL)) {
2286 * The oom killer is not called for high-order
2287 * allocations that may fail, so if no progress
2288 * is being made, there are no other options and
2289 * retrying is unlikely to help.
2291 if (order > PAGE_ALLOC_COSTLY_ORDER)
2294 * The oom killer is not called for lowmem
2295 * allocations to prevent needlessly killing
2298 if (high_zoneidx < ZONE_NORMAL)
2306 /* Check if we should retry the allocation */
2307 pages_reclaimed += did_some_progress;
2308 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2310 /* Wait for some write requests to complete then retry */
2311 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2315 * High-order allocations do not necessarily loop after
2316 * direct reclaim and reclaim/compaction depends on compaction
2317 * being called after reclaim so call directly if necessary
2319 page = __alloc_pages_direct_compact(gfp_mask, order,
2320 zonelist, high_zoneidx,
2322 alloc_flags, preferred_zone,
2323 migratetype, &did_some_progress,
2330 warn_alloc_failed(gfp_mask, order, NULL);
2333 if (kmemcheck_enabled)
2334 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2340 * This is the 'heart' of the zoned buddy allocator.
2343 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2344 struct zonelist *zonelist, nodemask_t *nodemask)
2346 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2347 struct zone *preferred_zone;
2349 int migratetype = allocflags_to_migratetype(gfp_mask);
2351 gfp_mask &= gfp_allowed_mask;
2353 lockdep_trace_alloc(gfp_mask);
2355 might_sleep_if(gfp_mask & __GFP_WAIT);
2357 if (should_fail_alloc_page(gfp_mask, order))
2361 * Check the zones suitable for the gfp_mask contain at least one
2362 * valid zone. It's possible to have an empty zonelist as a result
2363 * of GFP_THISNODE and a memoryless node
2365 if (unlikely(!zonelist->_zonerefs->zone))
2369 /* The preferred zone is used for statistics later */
2370 first_zones_zonelist(zonelist, high_zoneidx,
2371 nodemask ? : &cpuset_current_mems_allowed,
2373 if (!preferred_zone) {
2378 /* First allocation attempt */
2379 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2380 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2381 preferred_zone, migratetype);
2382 if (unlikely(!page))
2383 page = __alloc_pages_slowpath(gfp_mask, order,
2384 zonelist, high_zoneidx, nodemask,
2385 preferred_zone, migratetype);
2388 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2391 EXPORT_SYMBOL(__alloc_pages_nodemask);
2394 * Common helper functions.
2396 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2401 * __get_free_pages() returns a 32-bit address, which cannot represent
2404 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2406 page = alloc_pages(gfp_mask, order);
2409 return (unsigned long) page_address(page);
2411 EXPORT_SYMBOL(__get_free_pages);
2413 unsigned long get_zeroed_page(gfp_t gfp_mask)
2415 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2417 EXPORT_SYMBOL(get_zeroed_page);
2419 void __free_pages(struct page *page, unsigned int order)
2421 if (put_page_testzero(page)) {
2423 free_hot_cold_page(page, 0);
2425 __free_pages_ok(page, order);
2429 EXPORT_SYMBOL(__free_pages);
2431 void free_pages(unsigned long addr, unsigned int order)
2434 VM_BUG_ON(!virt_addr_valid((void *)addr));
2435 __free_pages(virt_to_page((void *)addr), order);
2439 EXPORT_SYMBOL(free_pages);
2441 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2444 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2445 unsigned long used = addr + PAGE_ALIGN(size);
2447 split_page(virt_to_page((void *)addr), order);
2448 while (used < alloc_end) {
2453 return (void *)addr;
2457 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2458 * @size: the number of bytes to allocate
2459 * @gfp_mask: GFP flags for the allocation
2461 * This function is similar to alloc_pages(), except that it allocates the
2462 * minimum number of pages to satisfy the request. alloc_pages() can only
2463 * allocate memory in power-of-two pages.
2465 * This function is also limited by MAX_ORDER.
2467 * Memory allocated by this function must be released by free_pages_exact().
2469 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2471 unsigned int order = get_order(size);
2474 addr = __get_free_pages(gfp_mask, order);
2475 return make_alloc_exact(addr, order, size);
2477 EXPORT_SYMBOL(alloc_pages_exact);
2480 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2482 * @nid: the preferred node ID where memory should be allocated
2483 * @size: the number of bytes to allocate
2484 * @gfp_mask: GFP flags for the allocation
2486 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2488 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2491 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2493 unsigned order = get_order(size);
2494 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2497 return make_alloc_exact((unsigned long)page_address(p), order, size);
2499 EXPORT_SYMBOL(alloc_pages_exact_nid);
2502 * free_pages_exact - release memory allocated via alloc_pages_exact()
2503 * @virt: the value returned by alloc_pages_exact.
2504 * @size: size of allocation, same value as passed to alloc_pages_exact().
2506 * Release the memory allocated by a previous call to alloc_pages_exact.
2508 void free_pages_exact(void *virt, size_t size)
2510 unsigned long addr = (unsigned long)virt;
2511 unsigned long end = addr + PAGE_ALIGN(size);
2513 while (addr < end) {
2518 EXPORT_SYMBOL(free_pages_exact);
2520 static unsigned int nr_free_zone_pages(int offset)
2525 /* Just pick one node, since fallback list is circular */
2526 unsigned int sum = 0;
2528 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2530 for_each_zone_zonelist(zone, z, zonelist, offset) {
2531 unsigned long size = zone->present_pages;
2532 unsigned long high = high_wmark_pages(zone);
2541 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2543 unsigned int nr_free_buffer_pages(void)
2545 return nr_free_zone_pages(gfp_zone(GFP_USER));
2547 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2550 * Amount of free RAM allocatable within all zones
2552 unsigned int nr_free_pagecache_pages(void)
2554 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2557 static inline void show_node(struct zone *zone)
2560 printk("Node %d ", zone_to_nid(zone));
2563 void si_meminfo(struct sysinfo *val)
2565 val->totalram = totalram_pages;
2567 val->freeram = global_page_state(NR_FREE_PAGES);
2568 val->bufferram = nr_blockdev_pages();
2569 val->totalhigh = totalhigh_pages;
2570 val->freehigh = nr_free_highpages();
2571 val->mem_unit = PAGE_SIZE;
2574 EXPORT_SYMBOL(si_meminfo);
2577 void si_meminfo_node(struct sysinfo *val, int nid)
2579 pg_data_t *pgdat = NODE_DATA(nid);
2581 val->totalram = pgdat->node_present_pages;
2582 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2583 #ifdef CONFIG_HIGHMEM
2584 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2585 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2591 val->mem_unit = PAGE_SIZE;
2596 * Determine whether the node should be displayed or not, depending on whether
2597 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2599 bool skip_free_areas_node(unsigned int flags, int nid)
2603 if (!(flags & SHOW_MEM_FILTER_NODES))
2607 ret = !node_isset(nid, cpuset_current_mems_allowed);
2613 #define K(x) ((x) << (PAGE_SHIFT-10))
2616 * Show free area list (used inside shift_scroll-lock stuff)
2617 * We also calculate the percentage fragmentation. We do this by counting the
2618 * memory on each free list with the exception of the first item on the list.
2619 * Suppresses nodes that are not allowed by current's cpuset if
2620 * SHOW_MEM_FILTER_NODES is passed.
2622 void show_free_areas(unsigned int filter)
2627 for_each_populated_zone(zone) {
2628 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2631 printk("%s per-cpu:\n", zone->name);
2633 for_each_online_cpu(cpu) {
2634 struct per_cpu_pageset *pageset;
2636 pageset = per_cpu_ptr(zone->pageset, cpu);
2638 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2639 cpu, pageset->pcp.high,
2640 pageset->pcp.batch, pageset->pcp.count);
2644 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2645 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2647 " dirty:%lu writeback:%lu unstable:%lu\n"
2648 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2649 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2650 global_page_state(NR_ACTIVE_ANON),
2651 global_page_state(NR_INACTIVE_ANON),
2652 global_page_state(NR_ISOLATED_ANON),
2653 global_page_state(NR_ACTIVE_FILE),
2654 global_page_state(NR_INACTIVE_FILE),
2655 global_page_state(NR_ISOLATED_FILE),
2656 global_page_state(NR_UNEVICTABLE),
2657 global_page_state(NR_FILE_DIRTY),
2658 global_page_state(NR_WRITEBACK),
2659 global_page_state(NR_UNSTABLE_NFS),
2660 global_page_state(NR_FREE_PAGES),
2661 global_page_state(NR_SLAB_RECLAIMABLE),
2662 global_page_state(NR_SLAB_UNRECLAIMABLE),
2663 global_page_state(NR_FILE_MAPPED),
2664 global_page_state(NR_SHMEM),
2665 global_page_state(NR_PAGETABLE),
2666 global_page_state(NR_BOUNCE));
2668 for_each_populated_zone(zone) {
2671 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2679 " active_anon:%lukB"
2680 " inactive_anon:%lukB"
2681 " active_file:%lukB"
2682 " inactive_file:%lukB"
2683 " unevictable:%lukB"
2684 " isolated(anon):%lukB"
2685 " isolated(file):%lukB"
2692 " slab_reclaimable:%lukB"
2693 " slab_unreclaimable:%lukB"
2694 " kernel_stack:%lukB"
2698 " writeback_tmp:%lukB"
2699 " pages_scanned:%lu"
2700 " all_unreclaimable? %s"
2703 K(zone_page_state(zone, NR_FREE_PAGES)),
2704 K(min_wmark_pages(zone)),
2705 K(low_wmark_pages(zone)),
2706 K(high_wmark_pages(zone)),
2707 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2708 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2709 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2710 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2711 K(zone_page_state(zone, NR_UNEVICTABLE)),
2712 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2713 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2714 K(zone->present_pages),
2715 K(zone_page_state(zone, NR_MLOCK)),
2716 K(zone_page_state(zone, NR_FILE_DIRTY)),
2717 K(zone_page_state(zone, NR_WRITEBACK)),
2718 K(zone_page_state(zone, NR_FILE_MAPPED)),
2719 K(zone_page_state(zone, NR_SHMEM)),
2720 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2721 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2722 zone_page_state(zone, NR_KERNEL_STACK) *
2724 K(zone_page_state(zone, NR_PAGETABLE)),
2725 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2726 K(zone_page_state(zone, NR_BOUNCE)),
2727 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2728 zone->pages_scanned,
2729 (zone->all_unreclaimable ? "yes" : "no")
2731 printk("lowmem_reserve[]:");
2732 for (i = 0; i < MAX_NR_ZONES; i++)
2733 printk(" %lu", zone->lowmem_reserve[i]);
2737 for_each_populated_zone(zone) {
2738 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2740 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2743 printk("%s: ", zone->name);
2745 spin_lock_irqsave(&zone->lock, flags);
2746 for (order = 0; order < MAX_ORDER; order++) {
2747 nr[order] = zone->free_area[order].nr_free;
2748 total += nr[order] << order;
2750 spin_unlock_irqrestore(&zone->lock, flags);
2751 for (order = 0; order < MAX_ORDER; order++)
2752 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2753 printk("= %lukB\n", K(total));
2756 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2758 show_swap_cache_info();
2761 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2763 zoneref->zone = zone;
2764 zoneref->zone_idx = zone_idx(zone);
2768 * Builds allocation fallback zone lists.
2770 * Add all populated zones of a node to the zonelist.
2772 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2773 int nr_zones, enum zone_type zone_type)
2777 BUG_ON(zone_type >= MAX_NR_ZONES);
2782 zone = pgdat->node_zones + zone_type;
2783 if (populated_zone(zone)) {
2784 zoneref_set_zone(zone,
2785 &zonelist->_zonerefs[nr_zones++]);
2786 check_highest_zone(zone_type);
2789 } while (zone_type);
2796 * 0 = automatic detection of better ordering.
2797 * 1 = order by ([node] distance, -zonetype)
2798 * 2 = order by (-zonetype, [node] distance)
2800 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2801 * the same zonelist. So only NUMA can configure this param.
2803 #define ZONELIST_ORDER_DEFAULT 0
2804 #define ZONELIST_ORDER_NODE 1
2805 #define ZONELIST_ORDER_ZONE 2
2807 /* zonelist order in the kernel.
2808 * set_zonelist_order() will set this to NODE or ZONE.
2810 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2811 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2815 /* The value user specified ....changed by config */
2816 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2817 /* string for sysctl */
2818 #define NUMA_ZONELIST_ORDER_LEN 16
2819 char numa_zonelist_order[16] = "default";
2822 * interface for configure zonelist ordering.
2823 * command line option "numa_zonelist_order"
2824 * = "[dD]efault - default, automatic configuration.
2825 * = "[nN]ode - order by node locality, then by zone within node
2826 * = "[zZ]one - order by zone, then by locality within zone
2829 static int __parse_numa_zonelist_order(char *s)
2831 if (*s == 'd' || *s == 'D') {
2832 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2833 } else if (*s == 'n' || *s == 'N') {
2834 user_zonelist_order = ZONELIST_ORDER_NODE;
2835 } else if (*s == 'z' || *s == 'Z') {
2836 user_zonelist_order = ZONELIST_ORDER_ZONE;
2839 "Ignoring invalid numa_zonelist_order value: "
2846 static __init int setup_numa_zonelist_order(char *s)
2853 ret = __parse_numa_zonelist_order(s);
2855 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2859 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2862 * sysctl handler for numa_zonelist_order
2864 int numa_zonelist_order_handler(ctl_table *table, int write,
2865 void __user *buffer, size_t *length,
2868 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2870 static DEFINE_MUTEX(zl_order_mutex);
2872 mutex_lock(&zl_order_mutex);
2874 strcpy(saved_string, (char*)table->data);
2875 ret = proc_dostring(table, write, buffer, length, ppos);
2879 int oldval = user_zonelist_order;
2880 if (__parse_numa_zonelist_order((char*)table->data)) {
2882 * bogus value. restore saved string
2884 strncpy((char*)table->data, saved_string,
2885 NUMA_ZONELIST_ORDER_LEN);
2886 user_zonelist_order = oldval;
2887 } else if (oldval != user_zonelist_order) {
2888 mutex_lock(&zonelists_mutex);
2889 build_all_zonelists(NULL);
2890 mutex_unlock(&zonelists_mutex);
2894 mutex_unlock(&zl_order_mutex);
2899 #define MAX_NODE_LOAD (nr_online_nodes)
2900 static int node_load[MAX_NUMNODES];
2903 * find_next_best_node - find the next node that should appear in a given node's fallback list
2904 * @node: node whose fallback list we're appending
2905 * @used_node_mask: nodemask_t of already used nodes
2907 * We use a number of factors to determine which is the next node that should
2908 * appear on a given node's fallback list. The node should not have appeared
2909 * already in @node's fallback list, and it should be the next closest node
2910 * according to the distance array (which contains arbitrary distance values
2911 * from each node to each node in the system), and should also prefer nodes
2912 * with no CPUs, since presumably they'll have very little allocation pressure
2913 * on them otherwise.
2914 * It returns -1 if no node is found.
2916 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2919 int min_val = INT_MAX;
2921 const struct cpumask *tmp = cpumask_of_node(0);
2923 /* Use the local node if we haven't already */
2924 if (!node_isset(node, *used_node_mask)) {
2925 node_set(node, *used_node_mask);
2929 for_each_node_state(n, N_HIGH_MEMORY) {
2931 /* Don't want a node to appear more than once */
2932 if (node_isset(n, *used_node_mask))
2935 /* Use the distance array to find the distance */
2936 val = node_distance(node, n);
2938 /* Penalize nodes under us ("prefer the next node") */
2941 /* Give preference to headless and unused nodes */
2942 tmp = cpumask_of_node(n);
2943 if (!cpumask_empty(tmp))
2944 val += PENALTY_FOR_NODE_WITH_CPUS;
2946 /* Slight preference for less loaded node */
2947 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2948 val += node_load[n];
2950 if (val < min_val) {
2957 node_set(best_node, *used_node_mask);
2964 * Build zonelists ordered by node and zones within node.
2965 * This results in maximum locality--normal zone overflows into local
2966 * DMA zone, if any--but risks exhausting DMA zone.
2968 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2971 struct zonelist *zonelist;
2973 zonelist = &pgdat->node_zonelists[0];
2974 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2976 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2978 zonelist->_zonerefs[j].zone = NULL;
2979 zonelist->_zonerefs[j].zone_idx = 0;
2983 * Build gfp_thisnode zonelists
2985 static void build_thisnode_zonelists(pg_data_t *pgdat)
2988 struct zonelist *zonelist;
2990 zonelist = &pgdat->node_zonelists[1];
2991 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2992 zonelist->_zonerefs[j].zone = NULL;
2993 zonelist->_zonerefs[j].zone_idx = 0;
2997 * Build zonelists ordered by zone and nodes within zones.
2998 * This results in conserving DMA zone[s] until all Normal memory is
2999 * exhausted, but results in overflowing to remote node while memory
3000 * may still exist in local DMA zone.
3002 static int node_order[MAX_NUMNODES];
3004 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3007 int zone_type; /* needs to be signed */
3009 struct zonelist *zonelist;
3011 zonelist = &pgdat->node_zonelists[0];
3013 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3014 for (j = 0; j < nr_nodes; j++) {
3015 node = node_order[j];
3016 z = &NODE_DATA(node)->node_zones[zone_type];
3017 if (populated_zone(z)) {
3019 &zonelist->_zonerefs[pos++]);
3020 check_highest_zone(zone_type);
3024 zonelist->_zonerefs[pos].zone = NULL;
3025 zonelist->_zonerefs[pos].zone_idx = 0;
3028 static int default_zonelist_order(void)
3031 unsigned long low_kmem_size,total_size;
3035 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3036 * If they are really small and used heavily, the system can fall
3037 * into OOM very easily.
3038 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3040 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3043 for_each_online_node(nid) {
3044 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3045 z = &NODE_DATA(nid)->node_zones[zone_type];
3046 if (populated_zone(z)) {
3047 if (zone_type < ZONE_NORMAL)
3048 low_kmem_size += z->present_pages;
3049 total_size += z->present_pages;
3050 } else if (zone_type == ZONE_NORMAL) {
3052 * If any node has only lowmem, then node order
3053 * is preferred to allow kernel allocations
3054 * locally; otherwise, they can easily infringe
3055 * on other nodes when there is an abundance of
3056 * lowmem available to allocate from.
3058 return ZONELIST_ORDER_NODE;
3062 if (!low_kmem_size || /* there are no DMA area. */
3063 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3064 return ZONELIST_ORDER_NODE;
3066 * look into each node's config.
3067 * If there is a node whose DMA/DMA32 memory is very big area on
3068 * local memory, NODE_ORDER may be suitable.
3070 average_size = total_size /
3071 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3072 for_each_online_node(nid) {
3075 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3076 z = &NODE_DATA(nid)->node_zones[zone_type];
3077 if (populated_zone(z)) {
3078 if (zone_type < ZONE_NORMAL)
3079 low_kmem_size += z->present_pages;
3080 total_size += z->present_pages;
3083 if (low_kmem_size &&
3084 total_size > average_size && /* ignore small node */
3085 low_kmem_size > total_size * 70/100)
3086 return ZONELIST_ORDER_NODE;
3088 return ZONELIST_ORDER_ZONE;
3091 static void set_zonelist_order(void)
3093 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3094 current_zonelist_order = default_zonelist_order();
3096 current_zonelist_order = user_zonelist_order;
3099 static void build_zonelists(pg_data_t *pgdat)
3103 nodemask_t used_mask;
3104 int local_node, prev_node;
3105 struct zonelist *zonelist;
3106 int order = current_zonelist_order;
3108 /* initialize zonelists */
3109 for (i = 0; i < MAX_ZONELISTS; i++) {
3110 zonelist = pgdat->node_zonelists + i;
3111 zonelist->_zonerefs[0].zone = NULL;
3112 zonelist->_zonerefs[0].zone_idx = 0;
3115 /* NUMA-aware ordering of nodes */
3116 local_node = pgdat->node_id;
3117 load = nr_online_nodes;
3118 prev_node = local_node;
3119 nodes_clear(used_mask);
3121 memset(node_order, 0, sizeof(node_order));
3124 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3125 int distance = node_distance(local_node, node);
3128 * If another node is sufficiently far away then it is better
3129 * to reclaim pages in a zone before going off node.
3131 if (distance > RECLAIM_DISTANCE)
3132 zone_reclaim_mode = 1;
3135 * We don't want to pressure a particular node.
3136 * So adding penalty to the first node in same
3137 * distance group to make it round-robin.
3139 if (distance != node_distance(local_node, prev_node))
3140 node_load[node] = load;
3144 if (order == ZONELIST_ORDER_NODE)
3145 build_zonelists_in_node_order(pgdat, node);
3147 node_order[j++] = node; /* remember order */
3150 if (order == ZONELIST_ORDER_ZONE) {
3151 /* calculate node order -- i.e., DMA last! */
3152 build_zonelists_in_zone_order(pgdat, j);
3155 build_thisnode_zonelists(pgdat);
3158 /* Construct the zonelist performance cache - see further mmzone.h */
3159 static void build_zonelist_cache(pg_data_t *pgdat)
3161 struct zonelist *zonelist;
3162 struct zonelist_cache *zlc;
3165 zonelist = &pgdat->node_zonelists[0];
3166 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3167 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3168 for (z = zonelist->_zonerefs; z->zone; z++)
3169 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3172 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3174 * Return node id of node used for "local" allocations.
3175 * I.e., first node id of first zone in arg node's generic zonelist.
3176 * Used for initializing percpu 'numa_mem', which is used primarily
3177 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3179 int local_memory_node(int node)
3183 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3184 gfp_zone(GFP_KERNEL),
3191 #else /* CONFIG_NUMA */
3193 static void set_zonelist_order(void)
3195 current_zonelist_order = ZONELIST_ORDER_ZONE;
3198 static void build_zonelists(pg_data_t *pgdat)
3200 int node, local_node;
3202 struct zonelist *zonelist;
3204 local_node = pgdat->node_id;
3206 zonelist = &pgdat->node_zonelists[0];
3207 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3210 * Now we build the zonelist so that it contains the zones
3211 * of all the other nodes.
3212 * We don't want to pressure a particular node, so when
3213 * building the zones for node N, we make sure that the
3214 * zones coming right after the local ones are those from
3215 * node N+1 (modulo N)
3217 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3218 if (!node_online(node))
3220 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3223 for (node = 0; node < local_node; node++) {
3224 if (!node_online(node))
3226 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3230 zonelist->_zonerefs[j].zone = NULL;
3231 zonelist->_zonerefs[j].zone_idx = 0;
3234 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3235 static void build_zonelist_cache(pg_data_t *pgdat)
3237 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3240 #endif /* CONFIG_NUMA */
3243 * Boot pageset table. One per cpu which is going to be used for all
3244 * zones and all nodes. The parameters will be set in such a way
3245 * that an item put on a list will immediately be handed over to
3246 * the buddy list. This is safe since pageset manipulation is done
3247 * with interrupts disabled.
3249 * The boot_pagesets must be kept even after bootup is complete for
3250 * unused processors and/or zones. They do play a role for bootstrapping
3251 * hotplugged processors.
3253 * zoneinfo_show() and maybe other functions do
3254 * not check if the processor is online before following the pageset pointer.
3255 * Other parts of the kernel may not check if the zone is available.
3257 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3258 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3259 static void setup_zone_pageset(struct zone *zone);
3262 * Global mutex to protect against size modification of zonelists
3263 * as well as to serialize pageset setup for the new populated zone.
3265 DEFINE_MUTEX(zonelists_mutex);
3267 /* return values int ....just for stop_machine() */
3268 static __init_refok int __build_all_zonelists(void *data)
3274 memset(node_load, 0, sizeof(node_load));
3276 for_each_online_node(nid) {
3277 pg_data_t *pgdat = NODE_DATA(nid);
3279 build_zonelists(pgdat);
3280 build_zonelist_cache(pgdat);
3284 * Initialize the boot_pagesets that are going to be used
3285 * for bootstrapping processors. The real pagesets for
3286 * each zone will be allocated later when the per cpu
3287 * allocator is available.
3289 * boot_pagesets are used also for bootstrapping offline
3290 * cpus if the system is already booted because the pagesets
3291 * are needed to initialize allocators on a specific cpu too.
3292 * F.e. the percpu allocator needs the page allocator which
3293 * needs the percpu allocator in order to allocate its pagesets
3294 * (a chicken-egg dilemma).
3296 for_each_possible_cpu(cpu) {
3297 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3299 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3301 * We now know the "local memory node" for each node--
3302 * i.e., the node of the first zone in the generic zonelist.
3303 * Set up numa_mem percpu variable for on-line cpus. During
3304 * boot, only the boot cpu should be on-line; we'll init the
3305 * secondary cpus' numa_mem as they come on-line. During
3306 * node/memory hotplug, we'll fixup all on-line cpus.
3308 if (cpu_online(cpu))
3309 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3317 * Called with zonelists_mutex held always
3318 * unless system_state == SYSTEM_BOOTING.
3320 void __ref build_all_zonelists(void *data)
3322 set_zonelist_order();
3324 if (system_state == SYSTEM_BOOTING) {
3325 __build_all_zonelists(NULL);
3326 mminit_verify_zonelist();
3327 cpuset_init_current_mems_allowed();
3329 /* we have to stop all cpus to guarantee there is no user
3331 #ifdef CONFIG_MEMORY_HOTPLUG
3333 setup_zone_pageset((struct zone *)data);
3335 stop_machine(__build_all_zonelists, NULL, NULL);
3336 /* cpuset refresh routine should be here */
3338 vm_total_pages = nr_free_pagecache_pages();
3340 * Disable grouping by mobility if the number of pages in the
3341 * system is too low to allow the mechanism to work. It would be
3342 * more accurate, but expensive to check per-zone. This check is
3343 * made on memory-hotadd so a system can start with mobility
3344 * disabled and enable it later
3346 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3347 page_group_by_mobility_disabled = 1;
3349 page_group_by_mobility_disabled = 0;
3351 printk("Built %i zonelists in %s order, mobility grouping %s. "
3352 "Total pages: %ld\n",
3354 zonelist_order_name[current_zonelist_order],
3355 page_group_by_mobility_disabled ? "off" : "on",
3358 printk("Policy zone: %s\n", zone_names[policy_zone]);
3363 * Helper functions to size the waitqueue hash table.
3364 * Essentially these want to choose hash table sizes sufficiently
3365 * large so that collisions trying to wait on pages are rare.
3366 * But in fact, the number of active page waitqueues on typical
3367 * systems is ridiculously low, less than 200. So this is even
3368 * conservative, even though it seems large.
3370 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3371 * waitqueues, i.e. the size of the waitq table given the number of pages.
3373 #define PAGES_PER_WAITQUEUE 256
3375 #ifndef CONFIG_MEMORY_HOTPLUG
3376 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3378 unsigned long size = 1;
3380 pages /= PAGES_PER_WAITQUEUE;
3382 while (size < pages)
3386 * Once we have dozens or even hundreds of threads sleeping
3387 * on IO we've got bigger problems than wait queue collision.
3388 * Limit the size of the wait table to a reasonable size.
3390 size = min(size, 4096UL);
3392 return max(size, 4UL);
3396 * A zone's size might be changed by hot-add, so it is not possible to determine
3397 * a suitable size for its wait_table. So we use the maximum size now.
3399 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3401 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3402 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3403 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3405 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3406 * or more by the traditional way. (See above). It equals:
3408 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3409 * ia64(16K page size) : = ( 8G + 4M)byte.
3410 * powerpc (64K page size) : = (32G +16M)byte.
3412 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3419 * This is an integer logarithm so that shifts can be used later
3420 * to extract the more random high bits from the multiplicative
3421 * hash function before the remainder is taken.
3423 static inline unsigned long wait_table_bits(unsigned long size)
3428 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3431 * Check if a pageblock contains reserved pages
3433 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3437 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3438 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3445 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3446 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3447 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3448 * higher will lead to a bigger reserve which will get freed as contiguous
3449 * blocks as reclaim kicks in
3451 static void setup_zone_migrate_reserve(struct zone *zone)
3453 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3455 unsigned long block_migratetype;
3459 * Get the start pfn, end pfn and the number of blocks to reserve
3460 * We have to be careful to be aligned to pageblock_nr_pages to
3461 * make sure that we always check pfn_valid for the first page in
3464 start_pfn = zone->zone_start_pfn;
3465 end_pfn = start_pfn + zone->spanned_pages;
3466 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3467 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3471 * Reserve blocks are generally in place to help high-order atomic
3472 * allocations that are short-lived. A min_free_kbytes value that
3473 * would result in more than 2 reserve blocks for atomic allocations
3474 * is assumed to be in place to help anti-fragmentation for the
3475 * future allocation of hugepages at runtime.
3477 reserve = min(2, reserve);
3479 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3480 if (!pfn_valid(pfn))
3482 page = pfn_to_page(pfn);
3484 /* Watch out for overlapping nodes */
3485 if (page_to_nid(page) != zone_to_nid(zone))
3488 block_migratetype = get_pageblock_migratetype(page);
3490 /* Only test what is necessary when the reserves are not met */
3493 * Blocks with reserved pages will never free, skip
3496 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3497 if (pageblock_is_reserved(pfn, block_end_pfn))
3500 /* If this block is reserved, account for it */
3501 if (block_migratetype == MIGRATE_RESERVE) {
3506 /* Suitable for reserving if this block is movable */
3507 if (block_migratetype == MIGRATE_MOVABLE) {
3508 set_pageblock_migratetype(page,
3510 move_freepages_block(zone, page,
3518 * If the reserve is met and this is a previous reserved block,
3521 if (block_migratetype == MIGRATE_RESERVE) {
3522 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3523 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3529 * Initially all pages are reserved - free ones are freed
3530 * up by free_all_bootmem() once the early boot process is
3531 * done. Non-atomic initialization, single-pass.
3533 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3534 unsigned long start_pfn, enum memmap_context context)
3537 unsigned long end_pfn = start_pfn + size;
3541 if (highest_memmap_pfn < end_pfn - 1)
3542 highest_memmap_pfn = end_pfn - 1;
3544 z = &NODE_DATA(nid)->node_zones[zone];
3545 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3547 * There can be holes in boot-time mem_map[]s
3548 * handed to this function. They do not
3549 * exist on hotplugged memory.
3551 if (context == MEMMAP_EARLY) {
3552 if (!early_pfn_valid(pfn))
3554 if (!early_pfn_in_nid(pfn, nid))
3557 page = pfn_to_page(pfn);
3558 set_page_links(page, zone, nid, pfn);
3559 mminit_verify_page_links(page, zone, nid, pfn);
3560 init_page_count(page);
3561 reset_page_mapcount(page);
3562 SetPageReserved(page);
3564 * Mark the block movable so that blocks are reserved for
3565 * movable at startup. This will force kernel allocations
3566 * to reserve their blocks rather than leaking throughout
3567 * the address space during boot when many long-lived
3568 * kernel allocations are made. Later some blocks near
3569 * the start are marked MIGRATE_RESERVE by
3570 * setup_zone_migrate_reserve()
3572 * bitmap is created for zone's valid pfn range. but memmap
3573 * can be created for invalid pages (for alignment)
3574 * check here not to call set_pageblock_migratetype() against
3577 if ((z->zone_start_pfn <= pfn)
3578 && (pfn < z->zone_start_pfn + z->spanned_pages)
3579 && !(pfn & (pageblock_nr_pages - 1)))
3580 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3582 INIT_LIST_HEAD(&page->lru);
3583 #ifdef WANT_PAGE_VIRTUAL
3584 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3585 if (!is_highmem_idx(zone))
3586 set_page_address(page, __va(pfn << PAGE_SHIFT));
3591 static void __meminit zone_init_free_lists(struct zone *zone)
3594 for_each_migratetype_order(order, t) {
3595 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3596 zone->free_area[order].nr_free = 0;
3600 #ifndef __HAVE_ARCH_MEMMAP_INIT
3601 #define memmap_init(size, nid, zone, start_pfn) \
3602 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3605 static int zone_batchsize(struct zone *zone)
3611 * The per-cpu-pages pools are set to around 1000th of the
3612 * size of the zone. But no more than 1/2 of a meg.
3614 * OK, so we don't know how big the cache is. So guess.
3616 batch = zone->present_pages / 1024;
3617 if (batch * PAGE_SIZE > 512 * 1024)
3618 batch = (512 * 1024) / PAGE_SIZE;
3619 batch /= 4; /* We effectively *= 4 below */
3624 * Clamp the batch to a 2^n - 1 value. Having a power
3625 * of 2 value was found to be more likely to have
3626 * suboptimal cache aliasing properties in some cases.
3628 * For example if 2 tasks are alternately allocating
3629 * batches of pages, one task can end up with a lot
3630 * of pages of one half of the possible page colors
3631 * and the other with pages of the other colors.
3633 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3638 /* The deferral and batching of frees should be suppressed under NOMMU
3641 * The problem is that NOMMU needs to be able to allocate large chunks
3642 * of contiguous memory as there's no hardware page translation to
3643 * assemble apparent contiguous memory from discontiguous pages.
3645 * Queueing large contiguous runs of pages for batching, however,
3646 * causes the pages to actually be freed in smaller chunks. As there
3647 * can be a significant delay between the individual batches being
3648 * recycled, this leads to the once large chunks of space being
3649 * fragmented and becoming unavailable for high-order allocations.
3655 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3657 struct per_cpu_pages *pcp;
3660 memset(p, 0, sizeof(*p));
3664 pcp->high = 6 * batch;
3665 pcp->batch = max(1UL, 1 * batch);
3666 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3667 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3671 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3672 * to the value high for the pageset p.
3675 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3678 struct per_cpu_pages *pcp;
3682 pcp->batch = max(1UL, high/4);
3683 if ((high/4) > (PAGE_SHIFT * 8))
3684 pcp->batch = PAGE_SHIFT * 8;
3687 static void setup_zone_pageset(struct zone *zone)
3691 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3693 for_each_possible_cpu(cpu) {
3694 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3696 setup_pageset(pcp, zone_batchsize(zone));
3698 if (percpu_pagelist_fraction)
3699 setup_pagelist_highmark(pcp,
3700 (zone->present_pages /
3701 percpu_pagelist_fraction));
3706 * Allocate per cpu pagesets and initialize them.
3707 * Before this call only boot pagesets were available.
3709 void __init setup_per_cpu_pageset(void)
3713 for_each_populated_zone(zone)
3714 setup_zone_pageset(zone);
3717 static noinline __init_refok
3718 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3721 struct pglist_data *pgdat = zone->zone_pgdat;
3725 * The per-page waitqueue mechanism uses hashed waitqueues
3728 zone->wait_table_hash_nr_entries =
3729 wait_table_hash_nr_entries(zone_size_pages);
3730 zone->wait_table_bits =
3731 wait_table_bits(zone->wait_table_hash_nr_entries);
3732 alloc_size = zone->wait_table_hash_nr_entries
3733 * sizeof(wait_queue_head_t);
3735 if (!slab_is_available()) {
3736 zone->wait_table = (wait_queue_head_t *)
3737 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3740 * This case means that a zone whose size was 0 gets new memory
3741 * via memory hot-add.
3742 * But it may be the case that a new node was hot-added. In
3743 * this case vmalloc() will not be able to use this new node's
3744 * memory - this wait_table must be initialized to use this new
3745 * node itself as well.
3746 * To use this new node's memory, further consideration will be
3749 zone->wait_table = vmalloc(alloc_size);
3751 if (!zone->wait_table)
3754 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3755 init_waitqueue_head(zone->wait_table + i);
3760 static int __zone_pcp_update(void *data)
3762 struct zone *zone = data;
3764 unsigned long batch = zone_batchsize(zone), flags;
3766 for_each_possible_cpu(cpu) {
3767 struct per_cpu_pageset *pset;
3768 struct per_cpu_pages *pcp;
3770 pset = per_cpu_ptr(zone->pageset, cpu);
3773 local_irq_save(flags);
3774 free_pcppages_bulk(zone, pcp->count, pcp);
3775 setup_pageset(pset, batch);
3776 local_irq_restore(flags);
3781 void zone_pcp_update(struct zone *zone)
3783 stop_machine(__zone_pcp_update, zone, NULL);
3786 static __meminit void zone_pcp_init(struct zone *zone)
3789 * per cpu subsystem is not up at this point. The following code
3790 * relies on the ability of the linker to provide the
3791 * offset of a (static) per cpu variable into the per cpu area.
3793 zone->pageset = &boot_pageset;
3795 if (zone->present_pages)
3796 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3797 zone->name, zone->present_pages,
3798 zone_batchsize(zone));
3801 __meminit int init_currently_empty_zone(struct zone *zone,
3802 unsigned long zone_start_pfn,
3804 enum memmap_context context)
3806 struct pglist_data *pgdat = zone->zone_pgdat;
3808 ret = zone_wait_table_init(zone, size);
3811 pgdat->nr_zones = zone_idx(zone) + 1;
3813 zone->zone_start_pfn = zone_start_pfn;
3815 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3816 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3818 (unsigned long)zone_idx(zone),
3819 zone_start_pfn, (zone_start_pfn + size));
3821 zone_init_free_lists(zone);
3826 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
3827 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3829 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3830 * Architectures may implement their own version but if add_active_range()
3831 * was used and there are no special requirements, this is a convenient
3834 int __meminit __early_pfn_to_nid(unsigned long pfn)
3836 unsigned long start_pfn, end_pfn;
3839 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
3840 if (start_pfn <= pfn && pfn < end_pfn)
3842 /* This is a memory hole */
3845 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3847 int __meminit early_pfn_to_nid(unsigned long pfn)
3851 nid = __early_pfn_to_nid(pfn);
3854 /* just returns 0 */
3858 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3859 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3863 nid = __early_pfn_to_nid(pfn);
3864 if (nid >= 0 && nid != node)
3871 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3872 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3873 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3875 * If an architecture guarantees that all ranges registered with
3876 * add_active_ranges() contain no holes and may be freed, this
3877 * this function may be used instead of calling free_bootmem() manually.
3879 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
3881 unsigned long start_pfn, end_pfn;
3884 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
3885 start_pfn = min(start_pfn, max_low_pfn);
3886 end_pfn = min(end_pfn, max_low_pfn);
3888 if (start_pfn < end_pfn)
3889 free_bootmem_node(NODE_DATA(this_nid),
3890 PFN_PHYS(start_pfn),
3891 (end_pfn - start_pfn) << PAGE_SHIFT);
3895 int __init add_from_early_node_map(struct range *range, int az,
3896 int nr_range, int nid)
3898 unsigned long start_pfn, end_pfn;
3901 /* need to go over early_node_map to find out good range for node */
3902 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL)
3903 nr_range = add_range(range, az, nr_range, start_pfn, end_pfn);
3908 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3909 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3911 * If an architecture guarantees that all ranges registered with
3912 * add_active_ranges() contain no holes and may be freed, this
3913 * function may be used instead of calling memory_present() manually.
3915 void __init sparse_memory_present_with_active_regions(int nid)
3917 unsigned long start_pfn, end_pfn;
3920 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
3921 memory_present(this_nid, start_pfn, end_pfn);
3925 * get_pfn_range_for_nid - Return the start and end page frames for a node
3926 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3927 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3928 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3930 * It returns the start and end page frame of a node based on information
3931 * provided by an arch calling add_active_range(). If called for a node
3932 * with no available memory, a warning is printed and the start and end
3935 void __meminit get_pfn_range_for_nid(unsigned int nid,
3936 unsigned long *start_pfn, unsigned long *end_pfn)
3938 unsigned long this_start_pfn, this_end_pfn;
3944 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
3945 *start_pfn = min(*start_pfn, this_start_pfn);
3946 *end_pfn = max(*end_pfn, this_end_pfn);
3949 if (*start_pfn == -1UL)
3954 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3955 * assumption is made that zones within a node are ordered in monotonic
3956 * increasing memory addresses so that the "highest" populated zone is used
3958 static void __init find_usable_zone_for_movable(void)
3961 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3962 if (zone_index == ZONE_MOVABLE)
3965 if (arch_zone_highest_possible_pfn[zone_index] >
3966 arch_zone_lowest_possible_pfn[zone_index])
3970 VM_BUG_ON(zone_index == -1);
3971 movable_zone = zone_index;
3975 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3976 * because it is sized independent of architecture. Unlike the other zones,
3977 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3978 * in each node depending on the size of each node and how evenly kernelcore
3979 * is distributed. This helper function adjusts the zone ranges
3980 * provided by the architecture for a given node by using the end of the
3981 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3982 * zones within a node are in order of monotonic increases memory addresses
3984 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3985 unsigned long zone_type,
3986 unsigned long node_start_pfn,
3987 unsigned long node_end_pfn,
3988 unsigned long *zone_start_pfn,
3989 unsigned long *zone_end_pfn)
3991 /* Only adjust if ZONE_MOVABLE is on this node */
3992 if (zone_movable_pfn[nid]) {
3993 /* Size ZONE_MOVABLE */
3994 if (zone_type == ZONE_MOVABLE) {
3995 *zone_start_pfn = zone_movable_pfn[nid];
3996 *zone_end_pfn = min(node_end_pfn,
3997 arch_zone_highest_possible_pfn[movable_zone]);
3999 /* Adjust for ZONE_MOVABLE starting within this range */
4000 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4001 *zone_end_pfn > zone_movable_pfn[nid]) {
4002 *zone_end_pfn = zone_movable_pfn[nid];
4004 /* Check if this whole range is within ZONE_MOVABLE */
4005 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4006 *zone_start_pfn = *zone_end_pfn;
4011 * Return the number of pages a zone spans in a node, including holes
4012 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4014 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4015 unsigned long zone_type,
4016 unsigned long *ignored)
4018 unsigned long node_start_pfn, node_end_pfn;
4019 unsigned long zone_start_pfn, zone_end_pfn;
4021 /* Get the start and end of the node and zone */
4022 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4023 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4024 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4025 adjust_zone_range_for_zone_movable(nid, zone_type,
4026 node_start_pfn, node_end_pfn,
4027 &zone_start_pfn, &zone_end_pfn);
4029 /* Check that this node has pages within the zone's required range */
4030 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4033 /* Move the zone boundaries inside the node if necessary */
4034 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4035 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4037 /* Return the spanned pages */
4038 return zone_end_pfn - zone_start_pfn;
4042 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4043 * then all holes in the requested range will be accounted for.
4045 unsigned long __meminit __absent_pages_in_range(int nid,
4046 unsigned long range_start_pfn,
4047 unsigned long range_end_pfn)
4049 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4050 unsigned long start_pfn, end_pfn;
4053 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4054 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4055 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4056 nr_absent -= end_pfn - start_pfn;
4062 * absent_pages_in_range - Return number of page frames in holes within a range
4063 * @start_pfn: The start PFN to start searching for holes
4064 * @end_pfn: The end PFN to stop searching for holes
4066 * It returns the number of pages frames in memory holes within a range.
4068 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4069 unsigned long end_pfn)
4071 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4074 /* Return the number of page frames in holes in a zone on a node */
4075 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4076 unsigned long zone_type,
4077 unsigned long *ignored)
4079 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4080 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4081 unsigned long node_start_pfn, node_end_pfn;
4082 unsigned long zone_start_pfn, zone_end_pfn;
4084 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4085 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4086 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4088 adjust_zone_range_for_zone_movable(nid, zone_type,
4089 node_start_pfn, node_end_pfn,
4090 &zone_start_pfn, &zone_end_pfn);
4091 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4094 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4095 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4096 unsigned long zone_type,
4097 unsigned long *zones_size)
4099 return zones_size[zone_type];
4102 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4103 unsigned long zone_type,
4104 unsigned long *zholes_size)
4109 return zholes_size[zone_type];
4112 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4114 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4115 unsigned long *zones_size, unsigned long *zholes_size)
4117 unsigned long realtotalpages, totalpages = 0;
4120 for (i = 0; i < MAX_NR_ZONES; i++)
4121 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4123 pgdat->node_spanned_pages = totalpages;
4125 realtotalpages = totalpages;
4126 for (i = 0; i < MAX_NR_ZONES; i++)
4128 zone_absent_pages_in_node(pgdat->node_id, i,
4130 pgdat->node_present_pages = realtotalpages;
4131 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4135 #ifndef CONFIG_SPARSEMEM
4137 * Calculate the size of the zone->blockflags rounded to an unsigned long
4138 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4139 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4140 * round what is now in bits to nearest long in bits, then return it in
4143 static unsigned long __init usemap_size(unsigned long zonesize)
4145 unsigned long usemapsize;
4147 usemapsize = roundup(zonesize, pageblock_nr_pages);
4148 usemapsize = usemapsize >> pageblock_order;
4149 usemapsize *= NR_PAGEBLOCK_BITS;
4150 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4152 return usemapsize / 8;
4155 static void __init setup_usemap(struct pglist_data *pgdat,
4156 struct zone *zone, unsigned long zonesize)
4158 unsigned long usemapsize = usemap_size(zonesize);
4159 zone->pageblock_flags = NULL;
4161 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4165 static inline void setup_usemap(struct pglist_data *pgdat,
4166 struct zone *zone, unsigned long zonesize) {}
4167 #endif /* CONFIG_SPARSEMEM */
4169 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4171 /* Return a sensible default order for the pageblock size. */
4172 static inline int pageblock_default_order(void)
4174 if (HPAGE_SHIFT > PAGE_SHIFT)
4175 return HUGETLB_PAGE_ORDER;
4180 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4181 static inline void __init set_pageblock_order(unsigned int order)
4183 /* Check that pageblock_nr_pages has not already been setup */
4184 if (pageblock_order)
4188 * Assume the largest contiguous order of interest is a huge page.
4189 * This value may be variable depending on boot parameters on IA64
4191 pageblock_order = order;
4193 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4196 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4197 * and pageblock_default_order() are unused as pageblock_order is set
4198 * at compile-time. See include/linux/pageblock-flags.h for the values of
4199 * pageblock_order based on the kernel config
4201 static inline int pageblock_default_order(unsigned int order)
4205 #define set_pageblock_order(x) do {} while (0)
4207 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4210 * Set up the zone data structures:
4211 * - mark all pages reserved
4212 * - mark all memory queues empty
4213 * - clear the memory bitmaps
4215 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4216 unsigned long *zones_size, unsigned long *zholes_size)
4219 int nid = pgdat->node_id;
4220 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4223 pgdat_resize_init(pgdat);
4224 pgdat->nr_zones = 0;
4225 init_waitqueue_head(&pgdat->kswapd_wait);
4226 pgdat->kswapd_max_order = 0;
4227 pgdat_page_cgroup_init(pgdat);
4229 for (j = 0; j < MAX_NR_ZONES; j++) {
4230 struct zone *zone = pgdat->node_zones + j;
4231 unsigned long size, realsize, memmap_pages;
4234 size = zone_spanned_pages_in_node(nid, j, zones_size);
4235 realsize = size - zone_absent_pages_in_node(nid, j,
4239 * Adjust realsize so that it accounts for how much memory
4240 * is used by this zone for memmap. This affects the watermark
4241 * and per-cpu initialisations
4244 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4245 if (realsize >= memmap_pages) {
4246 realsize -= memmap_pages;
4249 " %s zone: %lu pages used for memmap\n",
4250 zone_names[j], memmap_pages);
4253 " %s zone: %lu pages exceeds realsize %lu\n",
4254 zone_names[j], memmap_pages, realsize);
4256 /* Account for reserved pages */
4257 if (j == 0 && realsize > dma_reserve) {
4258 realsize -= dma_reserve;
4259 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4260 zone_names[0], dma_reserve);
4263 if (!is_highmem_idx(j))
4264 nr_kernel_pages += realsize;
4265 nr_all_pages += realsize;
4267 zone->spanned_pages = size;
4268 zone->present_pages = realsize;
4271 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4273 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4275 zone->name = zone_names[j];
4276 spin_lock_init(&zone->lock);
4277 spin_lock_init(&zone->lru_lock);
4278 zone_seqlock_init(zone);
4279 zone->zone_pgdat = pgdat;
4281 zone_pcp_init(zone);
4283 INIT_LIST_HEAD(&zone->lru[l].list);
4284 zone->reclaim_stat.recent_rotated[0] = 0;
4285 zone->reclaim_stat.recent_rotated[1] = 0;
4286 zone->reclaim_stat.recent_scanned[0] = 0;
4287 zone->reclaim_stat.recent_scanned[1] = 0;
4288 zap_zone_vm_stats(zone);
4293 set_pageblock_order(pageblock_default_order());
4294 setup_usemap(pgdat, zone, size);
4295 ret = init_currently_empty_zone(zone, zone_start_pfn,
4296 size, MEMMAP_EARLY);
4298 memmap_init(size, nid, j, zone_start_pfn);
4299 zone_start_pfn += size;
4303 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4305 /* Skip empty nodes */
4306 if (!pgdat->node_spanned_pages)
4309 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4310 /* ia64 gets its own node_mem_map, before this, without bootmem */
4311 if (!pgdat->node_mem_map) {
4312 unsigned long size, start, end;
4316 * The zone's endpoints aren't required to be MAX_ORDER
4317 * aligned but the node_mem_map endpoints must be in order
4318 * for the buddy allocator to function correctly.
4320 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4321 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4322 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4323 size = (end - start) * sizeof(struct page);
4324 map = alloc_remap(pgdat->node_id, size);
4326 map = alloc_bootmem_node_nopanic(pgdat, size);
4327 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4329 #ifndef CONFIG_NEED_MULTIPLE_NODES
4331 * With no DISCONTIG, the global mem_map is just set as node 0's
4333 if (pgdat == NODE_DATA(0)) {
4334 mem_map = NODE_DATA(0)->node_mem_map;
4335 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4336 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4337 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4338 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4341 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4344 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4345 unsigned long node_start_pfn, unsigned long *zholes_size)
4347 pg_data_t *pgdat = NODE_DATA(nid);
4349 pgdat->node_id = nid;
4350 pgdat->node_start_pfn = node_start_pfn;
4351 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4353 alloc_node_mem_map(pgdat);
4354 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4355 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4356 nid, (unsigned long)pgdat,
4357 (unsigned long)pgdat->node_mem_map);
4360 free_area_init_core(pgdat, zones_size, zholes_size);
4363 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4365 #if MAX_NUMNODES > 1
4367 * Figure out the number of possible node ids.
4369 static void __init setup_nr_node_ids(void)
4372 unsigned int highest = 0;
4374 for_each_node_mask(node, node_possible_map)
4376 nr_node_ids = highest + 1;
4379 static inline void setup_nr_node_ids(void)
4385 * node_map_pfn_alignment - determine the maximum internode alignment
4387 * This function should be called after node map is populated and sorted.
4388 * It calculates the maximum power of two alignment which can distinguish
4391 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4392 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4393 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4394 * shifted, 1GiB is enough and this function will indicate so.
4396 * This is used to test whether pfn -> nid mapping of the chosen memory
4397 * model has fine enough granularity to avoid incorrect mapping for the
4398 * populated node map.
4400 * Returns the determined alignment in pfn's. 0 if there is no alignment
4401 * requirement (single node).
4403 unsigned long __init node_map_pfn_alignment(void)
4405 unsigned long accl_mask = 0, last_end = 0;
4406 unsigned long start, end, mask;
4410 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4411 if (!start || last_nid < 0 || last_nid == nid) {
4418 * Start with a mask granular enough to pin-point to the
4419 * start pfn and tick off bits one-by-one until it becomes
4420 * too coarse to separate the current node from the last.
4422 mask = ~((1 << __ffs(start)) - 1);
4423 while (mask && last_end <= (start & (mask << 1)))
4426 /* accumulate all internode masks */
4430 /* convert mask to number of pages */
4431 return ~accl_mask + 1;
4434 /* Find the lowest pfn for a node */
4435 static unsigned long __init find_min_pfn_for_node(int nid)
4437 unsigned long min_pfn = ULONG_MAX;
4438 unsigned long start_pfn;
4441 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4442 min_pfn = min(min_pfn, start_pfn);
4444 if (min_pfn == ULONG_MAX) {
4446 "Could not find start_pfn for node %d\n", nid);
4454 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4456 * It returns the minimum PFN based on information provided via
4457 * add_active_range().
4459 unsigned long __init find_min_pfn_with_active_regions(void)
4461 return find_min_pfn_for_node(MAX_NUMNODES);
4465 * early_calculate_totalpages()
4466 * Sum pages in active regions for movable zone.
4467 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4469 static unsigned long __init early_calculate_totalpages(void)
4471 unsigned long totalpages = 0;
4472 unsigned long start_pfn, end_pfn;
4475 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4476 unsigned long pages = end_pfn - start_pfn;
4478 totalpages += pages;
4480 node_set_state(nid, N_HIGH_MEMORY);
4486 * Find the PFN the Movable zone begins in each node. Kernel memory
4487 * is spread evenly between nodes as long as the nodes have enough
4488 * memory. When they don't, some nodes will have more kernelcore than
4491 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4494 unsigned long usable_startpfn;
4495 unsigned long kernelcore_node, kernelcore_remaining;
4496 /* save the state before borrow the nodemask */
4497 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4498 unsigned long totalpages = early_calculate_totalpages();
4499 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4502 * If movablecore was specified, calculate what size of
4503 * kernelcore that corresponds so that memory usable for
4504 * any allocation type is evenly spread. If both kernelcore
4505 * and movablecore are specified, then the value of kernelcore
4506 * will be used for required_kernelcore if it's greater than
4507 * what movablecore would have allowed.
4509 if (required_movablecore) {
4510 unsigned long corepages;
4513 * Round-up so that ZONE_MOVABLE is at least as large as what
4514 * was requested by the user
4516 required_movablecore =
4517 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4518 corepages = totalpages - required_movablecore;
4520 required_kernelcore = max(required_kernelcore, corepages);
4523 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4524 if (!required_kernelcore)
4527 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4528 find_usable_zone_for_movable();
4529 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4532 /* Spread kernelcore memory as evenly as possible throughout nodes */
4533 kernelcore_node = required_kernelcore / usable_nodes;
4534 for_each_node_state(nid, N_HIGH_MEMORY) {
4535 unsigned long start_pfn, end_pfn;
4538 * Recalculate kernelcore_node if the division per node
4539 * now exceeds what is necessary to satisfy the requested
4540 * amount of memory for the kernel
4542 if (required_kernelcore < kernelcore_node)
4543 kernelcore_node = required_kernelcore / usable_nodes;
4546 * As the map is walked, we track how much memory is usable
4547 * by the kernel using kernelcore_remaining. When it is
4548 * 0, the rest of the node is usable by ZONE_MOVABLE
4550 kernelcore_remaining = kernelcore_node;
4552 /* Go through each range of PFNs within this node */
4553 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4554 unsigned long size_pages;
4556 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4557 if (start_pfn >= end_pfn)
4560 /* Account for what is only usable for kernelcore */
4561 if (start_pfn < usable_startpfn) {
4562 unsigned long kernel_pages;
4563 kernel_pages = min(end_pfn, usable_startpfn)
4566 kernelcore_remaining -= min(kernel_pages,
4567 kernelcore_remaining);
4568 required_kernelcore -= min(kernel_pages,
4569 required_kernelcore);
4571 /* Continue if range is now fully accounted */
4572 if (end_pfn <= usable_startpfn) {
4575 * Push zone_movable_pfn to the end so
4576 * that if we have to rebalance
4577 * kernelcore across nodes, we will
4578 * not double account here
4580 zone_movable_pfn[nid] = end_pfn;
4583 start_pfn = usable_startpfn;
4587 * The usable PFN range for ZONE_MOVABLE is from
4588 * start_pfn->end_pfn. Calculate size_pages as the
4589 * number of pages used as kernelcore
4591 size_pages = end_pfn - start_pfn;
4592 if (size_pages > kernelcore_remaining)
4593 size_pages = kernelcore_remaining;
4594 zone_movable_pfn[nid] = start_pfn + size_pages;
4597 * Some kernelcore has been met, update counts and
4598 * break if the kernelcore for this node has been
4601 required_kernelcore -= min(required_kernelcore,
4603 kernelcore_remaining -= size_pages;
4604 if (!kernelcore_remaining)
4610 * If there is still required_kernelcore, we do another pass with one
4611 * less node in the count. This will push zone_movable_pfn[nid] further
4612 * along on the nodes that still have memory until kernelcore is
4616 if (usable_nodes && required_kernelcore > usable_nodes)
4619 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4620 for (nid = 0; nid < MAX_NUMNODES; nid++)
4621 zone_movable_pfn[nid] =
4622 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4625 /* restore the node_state */
4626 node_states[N_HIGH_MEMORY] = saved_node_state;
4629 /* Any regular memory on that node ? */
4630 static void check_for_regular_memory(pg_data_t *pgdat)
4632 #ifdef CONFIG_HIGHMEM
4633 enum zone_type zone_type;
4635 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4636 struct zone *zone = &pgdat->node_zones[zone_type];
4637 if (zone->present_pages)
4638 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4644 * free_area_init_nodes - Initialise all pg_data_t and zone data
4645 * @max_zone_pfn: an array of max PFNs for each zone
4647 * This will call free_area_init_node() for each active node in the system.
4648 * Using the page ranges provided by add_active_range(), the size of each
4649 * zone in each node and their holes is calculated. If the maximum PFN
4650 * between two adjacent zones match, it is assumed that the zone is empty.
4651 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4652 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4653 * starts where the previous one ended. For example, ZONE_DMA32 starts
4654 * at arch_max_dma_pfn.
4656 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4658 unsigned long start_pfn, end_pfn;
4661 /* Record where the zone boundaries are */
4662 memset(arch_zone_lowest_possible_pfn, 0,
4663 sizeof(arch_zone_lowest_possible_pfn));
4664 memset(arch_zone_highest_possible_pfn, 0,
4665 sizeof(arch_zone_highest_possible_pfn));
4666 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4667 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4668 for (i = 1; i < MAX_NR_ZONES; i++) {
4669 if (i == ZONE_MOVABLE)
4671 arch_zone_lowest_possible_pfn[i] =
4672 arch_zone_highest_possible_pfn[i-1];
4673 arch_zone_highest_possible_pfn[i] =
4674 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4676 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4677 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4679 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4680 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4681 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4683 /* Print out the zone ranges */
4684 printk("Zone PFN ranges:\n");
4685 for (i = 0; i < MAX_NR_ZONES; i++) {
4686 if (i == ZONE_MOVABLE)
4688 printk(" %-8s ", zone_names[i]);
4689 if (arch_zone_lowest_possible_pfn[i] ==
4690 arch_zone_highest_possible_pfn[i])
4693 printk("%0#10lx -> %0#10lx\n",
4694 arch_zone_lowest_possible_pfn[i],
4695 arch_zone_highest_possible_pfn[i]);
4698 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4699 printk("Movable zone start PFN for each node\n");
4700 for (i = 0; i < MAX_NUMNODES; i++) {
4701 if (zone_movable_pfn[i])
4702 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4705 /* Print out the early_node_map[] */
4706 printk("Early memory PFN ranges\n");
4707 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4708 printk(" %3d: %0#10lx -> %0#10lx\n", nid, start_pfn, end_pfn);
4710 /* Initialise every node */
4711 mminit_verify_pageflags_layout();
4712 setup_nr_node_ids();
4713 for_each_online_node(nid) {
4714 pg_data_t *pgdat = NODE_DATA(nid);
4715 free_area_init_node(nid, NULL,
4716 find_min_pfn_for_node(nid), NULL);
4718 /* Any memory on that node */
4719 if (pgdat->node_present_pages)
4720 node_set_state(nid, N_HIGH_MEMORY);
4721 check_for_regular_memory(pgdat);
4725 static int __init cmdline_parse_core(char *p, unsigned long *core)
4727 unsigned long long coremem;
4731 coremem = memparse(p, &p);
4732 *core = coremem >> PAGE_SHIFT;
4734 /* Paranoid check that UL is enough for the coremem value */
4735 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4741 * kernelcore=size sets the amount of memory for use for allocations that
4742 * cannot be reclaimed or migrated.
4744 static int __init cmdline_parse_kernelcore(char *p)
4746 return cmdline_parse_core(p, &required_kernelcore);
4750 * movablecore=size sets the amount of memory for use for allocations that
4751 * can be reclaimed or migrated.
4753 static int __init cmdline_parse_movablecore(char *p)
4755 return cmdline_parse_core(p, &required_movablecore);
4758 early_param("kernelcore", cmdline_parse_kernelcore);
4759 early_param("movablecore", cmdline_parse_movablecore);
4761 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4764 * set_dma_reserve - set the specified number of pages reserved in the first zone
4765 * @new_dma_reserve: The number of pages to mark reserved
4767 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4768 * In the DMA zone, a significant percentage may be consumed by kernel image
4769 * and other unfreeable allocations which can skew the watermarks badly. This
4770 * function may optionally be used to account for unfreeable pages in the
4771 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4772 * smaller per-cpu batchsize.
4774 void __init set_dma_reserve(unsigned long new_dma_reserve)
4776 dma_reserve = new_dma_reserve;
4779 void __init free_area_init(unsigned long *zones_size)
4781 free_area_init_node(0, zones_size,
4782 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4785 static int page_alloc_cpu_notify(struct notifier_block *self,
4786 unsigned long action, void *hcpu)
4788 int cpu = (unsigned long)hcpu;
4790 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4794 * Spill the event counters of the dead processor
4795 * into the current processors event counters.
4796 * This artificially elevates the count of the current
4799 vm_events_fold_cpu(cpu);
4802 * Zero the differential counters of the dead processor
4803 * so that the vm statistics are consistent.
4805 * This is only okay since the processor is dead and cannot
4806 * race with what we are doing.
4808 refresh_cpu_vm_stats(cpu);
4813 void __init page_alloc_init(void)
4815 hotcpu_notifier(page_alloc_cpu_notify, 0);
4819 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4820 * or min_free_kbytes changes.
4822 static void calculate_totalreserve_pages(void)
4824 struct pglist_data *pgdat;
4825 unsigned long reserve_pages = 0;
4826 enum zone_type i, j;
4828 for_each_online_pgdat(pgdat) {
4829 for (i = 0; i < MAX_NR_ZONES; i++) {
4830 struct zone *zone = pgdat->node_zones + i;
4831 unsigned long max = 0;
4833 /* Find valid and maximum lowmem_reserve in the zone */
4834 for (j = i; j < MAX_NR_ZONES; j++) {
4835 if (zone->lowmem_reserve[j] > max)
4836 max = zone->lowmem_reserve[j];
4839 /* we treat the high watermark as reserved pages. */
4840 max += high_wmark_pages(zone);
4842 if (max > zone->present_pages)
4843 max = zone->present_pages;
4844 reserve_pages += max;
4847 totalreserve_pages = reserve_pages;
4851 * setup_per_zone_lowmem_reserve - called whenever
4852 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4853 * has a correct pages reserved value, so an adequate number of
4854 * pages are left in the zone after a successful __alloc_pages().
4856 static void setup_per_zone_lowmem_reserve(void)
4858 struct pglist_data *pgdat;
4859 enum zone_type j, idx;
4861 for_each_online_pgdat(pgdat) {
4862 for (j = 0; j < MAX_NR_ZONES; j++) {
4863 struct zone *zone = pgdat->node_zones + j;
4864 unsigned long present_pages = zone->present_pages;
4866 zone->lowmem_reserve[j] = 0;
4870 struct zone *lower_zone;
4874 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4875 sysctl_lowmem_reserve_ratio[idx] = 1;
4877 lower_zone = pgdat->node_zones + idx;
4878 lower_zone->lowmem_reserve[j] = present_pages /
4879 sysctl_lowmem_reserve_ratio[idx];
4880 present_pages += lower_zone->present_pages;
4885 /* update totalreserve_pages */
4886 calculate_totalreserve_pages();
4890 * setup_per_zone_wmarks - called when min_free_kbytes changes
4891 * or when memory is hot-{added|removed}
4893 * Ensures that the watermark[min,low,high] values for each zone are set
4894 * correctly with respect to min_free_kbytes.
4896 void setup_per_zone_wmarks(void)
4898 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4899 unsigned long pages_low = extra_free_kbytes >> (PAGE_SHIFT - 10);
4900 unsigned long lowmem_pages = 0;
4902 unsigned long flags;
4904 /* Calculate total number of !ZONE_HIGHMEM pages */
4905 for_each_zone(zone) {
4906 if (!is_highmem(zone))
4907 lowmem_pages += zone->present_pages;
4910 for_each_zone(zone) {
4913 spin_lock_irqsave(&zone->lock, flags);
4914 min = (u64)pages_min * zone->present_pages;
4915 do_div(min, lowmem_pages);
4916 low = (u64)pages_low * zone->present_pages;
4917 do_div(low, vm_total_pages);
4919 if (is_highmem(zone)) {
4921 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4922 * need highmem pages, so cap pages_min to a small
4925 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4926 * deltas controls asynch page reclaim, and so should
4927 * not be capped for highmem.
4931 min_pages = zone->present_pages / 1024;
4932 if (min_pages < SWAP_CLUSTER_MAX)
4933 min_pages = SWAP_CLUSTER_MAX;
4934 if (min_pages > 128)
4936 zone->watermark[WMARK_MIN] = min_pages;
4939 * If it's a lowmem zone, reserve a number of pages
4940 * proportionate to the zone's size.
4942 zone->watermark[WMARK_MIN] = min;
4945 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) +
4947 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) +
4949 setup_zone_migrate_reserve(zone);
4950 spin_unlock_irqrestore(&zone->lock, flags);
4953 /* update totalreserve_pages */
4954 calculate_totalreserve_pages();
4958 * The inactive anon list should be small enough that the VM never has to
4959 * do too much work, but large enough that each inactive page has a chance
4960 * to be referenced again before it is swapped out.
4962 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4963 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4964 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4965 * the anonymous pages are kept on the inactive list.
4968 * memory ratio inactive anon
4969 * -------------------------------------
4978 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
4980 unsigned int gb, ratio;
4982 /* Zone size in gigabytes */
4983 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4985 ratio = int_sqrt(10 * gb);
4989 zone->inactive_ratio = ratio;
4992 static void __meminit setup_per_zone_inactive_ratio(void)
4997 calculate_zone_inactive_ratio(zone);
5001 * Initialise min_free_kbytes.
5003 * For small machines we want it small (128k min). For large machines
5004 * we want it large (64MB max). But it is not linear, because network
5005 * bandwidth does not increase linearly with machine size. We use
5007 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5008 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5024 int __meminit init_per_zone_wmark_min(void)
5026 unsigned long lowmem_kbytes;
5028 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5030 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5031 if (min_free_kbytes < 128)
5032 min_free_kbytes = 128;
5033 if (min_free_kbytes > 65536)
5034 min_free_kbytes = 65536;
5035 setup_per_zone_wmarks();
5036 refresh_zone_stat_thresholds();
5037 setup_per_zone_lowmem_reserve();
5038 setup_per_zone_inactive_ratio();
5041 module_init(init_per_zone_wmark_min)
5044 * free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5045 * that we can call two helper functions whenever min_free_kbytes
5046 * or extra_free_kbytes changes.
5048 int free_kbytes_sysctl_handler(ctl_table *table, int write,
5049 void __user *buffer, size_t *length, loff_t *ppos)
5051 proc_dointvec(table, write, buffer, length, ppos);
5053 setup_per_zone_wmarks();
5058 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5059 void __user *buffer, size_t *length, loff_t *ppos)
5064 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5069 zone->min_unmapped_pages = (zone->present_pages *
5070 sysctl_min_unmapped_ratio) / 100;
5074 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5075 void __user *buffer, size_t *length, loff_t *ppos)
5080 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5085 zone->min_slab_pages = (zone->present_pages *
5086 sysctl_min_slab_ratio) / 100;
5092 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5093 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5094 * whenever sysctl_lowmem_reserve_ratio changes.
5096 * The reserve ratio obviously has absolutely no relation with the
5097 * minimum watermarks. The lowmem reserve ratio can only make sense
5098 * if in function of the boot time zone sizes.
5100 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5101 void __user *buffer, size_t *length, loff_t *ppos)
5103 proc_dointvec_minmax(table, write, buffer, length, ppos);
5104 setup_per_zone_lowmem_reserve();
5109 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5110 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5111 * can have before it gets flushed back to buddy allocator.
5114 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5115 void __user *buffer, size_t *length, loff_t *ppos)
5121 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5122 if (!write || (ret == -EINVAL))
5124 for_each_populated_zone(zone) {
5125 for_each_possible_cpu(cpu) {
5127 high = zone->present_pages / percpu_pagelist_fraction;
5128 setup_pagelist_highmark(
5129 per_cpu_ptr(zone->pageset, cpu), high);
5135 int hashdist = HASHDIST_DEFAULT;
5138 static int __init set_hashdist(char *str)
5142 hashdist = simple_strtoul(str, &str, 0);
5145 __setup("hashdist=", set_hashdist);
5149 * allocate a large system hash table from bootmem
5150 * - it is assumed that the hash table must contain an exact power-of-2
5151 * quantity of entries
5152 * - limit is the number of hash buckets, not the total allocation size
5154 void *__init alloc_large_system_hash(const char *tablename,
5155 unsigned long bucketsize,
5156 unsigned long numentries,
5159 unsigned int *_hash_shift,
5160 unsigned int *_hash_mask,
5161 unsigned long limit)
5163 unsigned long long max = limit;
5164 unsigned long log2qty, size;
5167 /* allow the kernel cmdline to have a say */
5169 /* round applicable memory size up to nearest megabyte */
5170 numentries = nr_kernel_pages;
5171 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5172 numentries >>= 20 - PAGE_SHIFT;
5173 numentries <<= 20 - PAGE_SHIFT;
5175 /* limit to 1 bucket per 2^scale bytes of low memory */
5176 if (scale > PAGE_SHIFT)
5177 numentries >>= (scale - PAGE_SHIFT);
5179 numentries <<= (PAGE_SHIFT - scale);
5181 /* Make sure we've got at least a 0-order allocation.. */
5182 if (unlikely(flags & HASH_SMALL)) {
5183 /* Makes no sense without HASH_EARLY */
5184 WARN_ON(!(flags & HASH_EARLY));
5185 if (!(numentries >> *_hash_shift)) {
5186 numentries = 1UL << *_hash_shift;
5187 BUG_ON(!numentries);
5189 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5190 numentries = PAGE_SIZE / bucketsize;
5192 numentries = roundup_pow_of_two(numentries);
5194 /* limit allocation size to 1/16 total memory by default */
5196 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5197 do_div(max, bucketsize);
5200 if (numentries > max)
5203 log2qty = ilog2(numentries);
5206 size = bucketsize << log2qty;
5207 if (flags & HASH_EARLY)
5208 table = alloc_bootmem_nopanic(size);
5210 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5213 * If bucketsize is not a power-of-two, we may free
5214 * some pages at the end of hash table which
5215 * alloc_pages_exact() automatically does
5217 if (get_order(size) < MAX_ORDER) {
5218 table = alloc_pages_exact(size, GFP_ATOMIC);
5219 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5222 } while (!table && size > PAGE_SIZE && --log2qty);
5225 panic("Failed to allocate %s hash table\n", tablename);
5227 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5230 ilog2(size) - PAGE_SHIFT,
5234 *_hash_shift = log2qty;
5236 *_hash_mask = (1 << log2qty) - 1;
5241 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5242 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5245 #ifdef CONFIG_SPARSEMEM
5246 return __pfn_to_section(pfn)->pageblock_flags;
5248 return zone->pageblock_flags;
5249 #endif /* CONFIG_SPARSEMEM */
5252 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5254 #ifdef CONFIG_SPARSEMEM
5255 pfn &= (PAGES_PER_SECTION-1);
5256 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5258 pfn = pfn - zone->zone_start_pfn;
5259 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5260 #endif /* CONFIG_SPARSEMEM */
5264 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5265 * @page: The page within the block of interest
5266 * @start_bitidx: The first bit of interest to retrieve
5267 * @end_bitidx: The last bit of interest
5268 * returns pageblock_bits flags
5270 unsigned long get_pageblock_flags_group(struct page *page,
5271 int start_bitidx, int end_bitidx)
5274 unsigned long *bitmap;
5275 unsigned long pfn, bitidx;
5276 unsigned long flags = 0;
5277 unsigned long value = 1;
5279 zone = page_zone(page);
5280 pfn = page_to_pfn(page);
5281 bitmap = get_pageblock_bitmap(zone, pfn);
5282 bitidx = pfn_to_bitidx(zone, pfn);
5284 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5285 if (test_bit(bitidx + start_bitidx, bitmap))
5292 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5293 * @page: The page within the block of interest
5294 * @start_bitidx: The first bit of interest
5295 * @end_bitidx: The last bit of interest
5296 * @flags: The flags to set
5298 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5299 int start_bitidx, int end_bitidx)
5302 unsigned long *bitmap;
5303 unsigned long pfn, bitidx;
5304 unsigned long value = 1;
5306 zone = page_zone(page);
5307 pfn = page_to_pfn(page);
5308 bitmap = get_pageblock_bitmap(zone, pfn);
5309 bitidx = pfn_to_bitidx(zone, pfn);
5310 VM_BUG_ON(pfn < zone->zone_start_pfn);
5311 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5313 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5315 __set_bit(bitidx + start_bitidx, bitmap);
5317 __clear_bit(bitidx + start_bitidx, bitmap);
5321 * This is designed as sub function...plz see page_isolation.c also.
5322 * set/clear page block's type to be ISOLATE.
5323 * page allocater never alloc memory from ISOLATE block.
5327 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5329 unsigned long pfn, iter, found;
5331 * For avoiding noise data, lru_add_drain_all() should be called
5332 * If ZONE_MOVABLE, the zone never contains immobile pages
5334 if (zone_idx(zone) == ZONE_MOVABLE)
5337 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5340 pfn = page_to_pfn(page);
5341 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5342 unsigned long check = pfn + iter;
5344 if (!pfn_valid_within(check))
5347 page = pfn_to_page(check);
5348 if (!page_count(page)) {
5349 if (PageBuddy(page))
5350 iter += (1 << page_order(page)) - 1;
5356 * If there are RECLAIMABLE pages, we need to check it.
5357 * But now, memory offline itself doesn't call shrink_slab()
5358 * and it still to be fixed.
5361 * If the page is not RAM, page_count()should be 0.
5362 * we don't need more check. This is an _used_ not-movable page.
5364 * The problematic thing here is PG_reserved pages. PG_reserved
5365 * is set to both of a memory hole page and a _used_ kernel
5374 bool is_pageblock_removable_nolock(struct page *page)
5376 struct zone *zone = page_zone(page);
5377 return __count_immobile_pages(zone, page, 0);
5380 int set_migratetype_isolate(struct page *page)
5383 unsigned long flags, pfn;
5384 struct memory_isolate_notify arg;
5388 zone = page_zone(page);
5390 spin_lock_irqsave(&zone->lock, flags);
5392 pfn = page_to_pfn(page);
5393 arg.start_pfn = pfn;
5394 arg.nr_pages = pageblock_nr_pages;
5395 arg.pages_found = 0;
5398 * It may be possible to isolate a pageblock even if the
5399 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5400 * notifier chain is used by balloon drivers to return the
5401 * number of pages in a range that are held by the balloon
5402 * driver to shrink memory. If all the pages are accounted for
5403 * by balloons, are free, or on the LRU, isolation can continue.
5404 * Later, for example, when memory hotplug notifier runs, these
5405 * pages reported as "can be isolated" should be isolated(freed)
5406 * by the balloon driver through the memory notifier chain.
5408 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5409 notifier_ret = notifier_to_errno(notifier_ret);
5413 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5414 * We just check MOVABLE pages.
5416 if (__count_immobile_pages(zone, page, arg.pages_found))
5420 * immobile means "not-on-lru" paes. If immobile is larger than
5421 * removable-by-driver pages reported by notifier, we'll fail.
5426 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5427 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5430 spin_unlock_irqrestore(&zone->lock, flags);
5436 void unset_migratetype_isolate(struct page *page)
5439 unsigned long flags;
5440 zone = page_zone(page);
5441 spin_lock_irqsave(&zone->lock, flags);
5442 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5444 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5445 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5447 spin_unlock_irqrestore(&zone->lock, flags);
5450 #ifdef CONFIG_MEMORY_HOTREMOVE
5452 * All pages in the range must be isolated before calling this.
5455 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5461 unsigned long flags;
5462 /* find the first valid pfn */
5463 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5468 zone = page_zone(pfn_to_page(pfn));
5469 spin_lock_irqsave(&zone->lock, flags);
5471 while (pfn < end_pfn) {
5472 if (!pfn_valid(pfn)) {
5476 page = pfn_to_page(pfn);
5477 BUG_ON(page_count(page));
5478 BUG_ON(!PageBuddy(page));
5479 order = page_order(page);
5480 #ifdef CONFIG_DEBUG_VM
5481 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5482 pfn, 1 << order, end_pfn);
5484 list_del(&page->lru);
5485 rmv_page_order(page);
5486 zone->free_area[order].nr_free--;
5487 __mod_zone_page_state(zone, NR_FREE_PAGES,
5489 for (i = 0; i < (1 << order); i++)
5490 SetPageReserved((page+i));
5491 pfn += (1 << order);
5493 spin_unlock_irqrestore(&zone->lock, flags);
5497 #ifdef CONFIG_MEMORY_FAILURE
5498 bool is_free_buddy_page(struct page *page)
5500 struct zone *zone = page_zone(page);
5501 unsigned long pfn = page_to_pfn(page);
5502 unsigned long flags;
5505 spin_lock_irqsave(&zone->lock, flags);
5506 for (order = 0; order < MAX_ORDER; order++) {
5507 struct page *page_head = page - (pfn & ((1 << order) - 1));
5509 if (PageBuddy(page_head) && page_order(page_head) >= order)
5512 spin_unlock_irqrestore(&zone->lock, flags);
5514 return order < MAX_ORDER;
5518 static struct trace_print_flags pageflag_names[] = {
5519 {1UL << PG_locked, "locked" },
5520 {1UL << PG_error, "error" },
5521 {1UL << PG_referenced, "referenced" },
5522 {1UL << PG_uptodate, "uptodate" },
5523 {1UL << PG_dirty, "dirty" },
5524 {1UL << PG_lru, "lru" },
5525 {1UL << PG_active, "active" },
5526 {1UL << PG_slab, "slab" },
5527 {1UL << PG_owner_priv_1, "owner_priv_1" },
5528 {1UL << PG_arch_1, "arch_1" },
5529 {1UL << PG_reserved, "reserved" },
5530 {1UL << PG_private, "private" },
5531 {1UL << PG_private_2, "private_2" },
5532 {1UL << PG_writeback, "writeback" },
5533 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5534 {1UL << PG_head, "head" },
5535 {1UL << PG_tail, "tail" },
5537 {1UL << PG_compound, "compound" },
5539 {1UL << PG_swapcache, "swapcache" },
5540 {1UL << PG_mappedtodisk, "mappedtodisk" },
5541 {1UL << PG_reclaim, "reclaim" },
5542 {1UL << PG_swapbacked, "swapbacked" },
5543 {1UL << PG_unevictable, "unevictable" },
5545 {1UL << PG_mlocked, "mlocked" },
5547 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5548 {1UL << PG_uncached, "uncached" },
5550 #ifdef CONFIG_MEMORY_FAILURE
5551 {1UL << PG_hwpoison, "hwpoison" },
5556 static void dump_page_flags(unsigned long flags)
5558 const char *delim = "";
5562 printk(KERN_ALERT "page flags: %#lx(", flags);
5564 /* remove zone id */
5565 flags &= (1UL << NR_PAGEFLAGS) - 1;
5567 for (i = 0; pageflag_names[i].name && flags; i++) {
5569 mask = pageflag_names[i].mask;
5570 if ((flags & mask) != mask)
5574 printk("%s%s", delim, pageflag_names[i].name);
5578 /* check for left over flags */
5580 printk("%s%#lx", delim, flags);
5585 void dump_page(struct page *page)
5588 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5589 page, atomic_read(&page->_count), page_mapcount(page),
5590 page->mapping, page->index);
5591 dump_page_flags(page->flags);
5592 mem_cgroup_print_bad_page(page);