2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
62 #include <asm/tlbflush.h>
63 #include <asm/div64.h>
66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67 DEFINE_PER_CPU(int, numa_node);
68 EXPORT_PER_CPU_SYMBOL(numa_node);
71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
78 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
83 * Array of node states.
85 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86 [N_POSSIBLE] = NODE_MASK_ALL,
87 [N_ONLINE] = { { [0] = 1UL } },
89 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
91 [N_HIGH_MEMORY] = { { [0] = 1UL } },
93 [N_CPU] = { { [0] = 1UL } },
96 EXPORT_SYMBOL(node_states);
98 unsigned long totalram_pages __read_mostly;
99 unsigned long totalreserve_pages __read_mostly;
101 * When calculating the number of globally allowed dirty pages, there
102 * is a certain number of per-zone reserves that should not be
103 * considered dirtyable memory. This is the sum of those reserves
104 * over all existing zones that contribute dirtyable memory.
106 unsigned long dirty_balance_reserve __read_mostly;
108 int percpu_pagelist_fraction;
109 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
111 #ifdef CONFIG_PM_SLEEP
113 * The following functions are used by the suspend/hibernate code to temporarily
114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
115 * while devices are suspended. To avoid races with the suspend/hibernate code,
116 * they should always be called with pm_mutex held (gfp_allowed_mask also should
117 * only be modified with pm_mutex held, unless the suspend/hibernate code is
118 * guaranteed not to run in parallel with that modification).
121 static gfp_t saved_gfp_mask;
123 void pm_restore_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 if (saved_gfp_mask) {
127 gfp_allowed_mask = saved_gfp_mask;
132 void pm_restrict_gfp_mask(void)
134 WARN_ON(!mutex_is_locked(&pm_mutex));
135 WARN_ON(saved_gfp_mask);
136 saved_gfp_mask = gfp_allowed_mask;
137 gfp_allowed_mask &= ~GFP_IOFS;
140 bool pm_suspended_storage(void)
142 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
146 #endif /* CONFIG_PM_SLEEP */
148 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
149 int pageblock_order __read_mostly;
152 static void __free_pages_ok(struct page *page, unsigned int order);
155 * results with 256, 32 in the lowmem_reserve sysctl:
156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
157 * 1G machine -> (16M dma, 784M normal, 224M high)
158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
162 * TBD: should special case ZONE_DMA32 machines here - in those we normally
163 * don't need any ZONE_NORMAL reservation
165 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
166 #ifdef CONFIG_ZONE_DMA
169 #ifdef CONFIG_ZONE_DMA32
172 #ifdef CONFIG_HIGHMEM
178 EXPORT_SYMBOL(totalram_pages);
180 static char * const zone_names[MAX_NR_ZONES] = {
181 #ifdef CONFIG_ZONE_DMA
184 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 int min_free_kbytes = 1024;
196 static unsigned long __meminitdata nr_kernel_pages;
197 static unsigned long __meminitdata nr_all_pages;
198 static unsigned long __meminitdata dma_reserve;
200 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
201 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
202 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
203 static unsigned long __initdata required_kernelcore;
204 static unsigned long __initdata required_movablecore;
205 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
207 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
209 EXPORT_SYMBOL(movable_zone);
210 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
213 int nr_node_ids __read_mostly = MAX_NUMNODES;
214 int nr_online_nodes __read_mostly = 1;
215 EXPORT_SYMBOL(nr_node_ids);
216 EXPORT_SYMBOL(nr_online_nodes);
219 int page_group_by_mobility_disabled __read_mostly;
223 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
224 * Instead, use {un}set_pageblock_isolate.
226 void set_pageblock_migratetype(struct page *page, int migratetype)
229 if (unlikely(page_group_by_mobility_disabled))
230 migratetype = MIGRATE_UNMOVABLE;
232 set_pageblock_flags_group(page, (unsigned long)migratetype,
233 PB_migrate, PB_migrate_end);
236 bool oom_killer_disabled __read_mostly;
238 #ifdef CONFIG_DEBUG_VM
239 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
243 unsigned long pfn = page_to_pfn(page);
246 seq = zone_span_seqbegin(zone);
247 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
249 else if (pfn < zone->zone_start_pfn)
251 } while (zone_span_seqretry(zone, seq));
256 static int page_is_consistent(struct zone *zone, struct page *page)
258 if (!pfn_valid_within(page_to_pfn(page)))
260 if (zone != page_zone(page))
266 * Temporary debugging check for pages not lying within a given zone.
268 static int bad_range(struct zone *zone, struct page *page)
270 if (page_outside_zone_boundaries(zone, page))
272 if (!page_is_consistent(zone, page))
278 static inline int bad_range(struct zone *zone, struct page *page)
284 static void bad_page(struct page *page)
286 static unsigned long resume;
287 static unsigned long nr_shown;
288 static unsigned long nr_unshown;
290 /* Don't complain about poisoned pages */
291 if (PageHWPoison(page)) {
292 reset_page_mapcount(page); /* remove PageBuddy */
297 * Allow a burst of 60 reports, then keep quiet for that minute;
298 * or allow a steady drip of one report per second.
300 if (nr_shown == 60) {
301 if (time_before(jiffies, resume)) {
307 "BUG: Bad page state: %lu messages suppressed\n",
314 resume = jiffies + 60 * HZ;
316 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
317 current->comm, page_to_pfn(page));
323 /* Leave bad fields for debug, except PageBuddy could make trouble */
324 reset_page_mapcount(page); /* remove PageBuddy */
325 add_taint(TAINT_BAD_PAGE);
329 * Higher-order pages are called "compound pages". They are structured thusly:
331 * The first PAGE_SIZE page is called the "head page".
333 * The remaining PAGE_SIZE pages are called "tail pages".
335 * All pages have PG_compound set. All tail pages have their ->first_page
336 * pointing at the head page.
338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
343 static void free_compound_page(struct page *page)
345 __free_pages_ok(page, compound_order(page));
348 void prep_compound_page(struct page *page, unsigned long order)
351 int nr_pages = 1 << order;
353 set_compound_page_dtor(page, free_compound_page);
354 set_compound_order(page, order);
356 for (i = 1; i < nr_pages; i++) {
357 struct page *p = page + i;
359 set_page_count(p, 0);
360 p->first_page = page;
364 /* update __split_huge_page_refcount if you change this function */
365 static int destroy_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
371 if (unlikely(compound_order(page) != order)) {
376 __ClearPageHead(page);
378 for (i = 1; i < nr_pages; i++) {
379 struct page *p = page + i;
381 if (unlikely(!PageTail(p) || (p->first_page != page))) {
391 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
396 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
397 * and __GFP_HIGHMEM from hard or soft interrupt context.
399 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
400 for (i = 0; i < (1 << order); i++)
401 clear_highpage(page + i);
404 #ifdef CONFIG_DEBUG_PAGEALLOC
405 unsigned int _debug_guardpage_minorder;
407 static int __init debug_guardpage_minorder_setup(char *buf)
411 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
412 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
415 _debug_guardpage_minorder = res;
416 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
419 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
421 static inline void set_page_guard_flag(struct page *page)
423 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
426 static inline void clear_page_guard_flag(struct page *page)
428 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
431 static inline void set_page_guard_flag(struct page *page) { }
432 static inline void clear_page_guard_flag(struct page *page) { }
435 static inline void set_page_order(struct page *page, int order)
437 set_page_private(page, order);
438 __SetPageBuddy(page);
441 static inline void rmv_page_order(struct page *page)
443 __ClearPageBuddy(page);
444 set_page_private(page, 0);
448 * Locate the struct page for both the matching buddy in our
449 * pair (buddy1) and the combined O(n+1) page they form (page).
451 * 1) Any buddy B1 will have an order O twin B2 which satisfies
452 * the following equation:
454 * For example, if the starting buddy (buddy2) is #8 its order
456 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
458 * 2) Any buddy B will have an order O+1 parent P which
459 * satisfies the following equation:
462 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
464 static inline unsigned long
465 __find_buddy_index(unsigned long page_idx, unsigned int order)
467 return page_idx ^ (1 << order);
471 * This function checks whether a page is free && is the buddy
472 * we can do coalesce a page and its buddy if
473 * (a) the buddy is not in a hole &&
474 * (b) the buddy is in the buddy system &&
475 * (c) a page and its buddy have the same order &&
476 * (d) a page and its buddy are in the same zone.
478 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
479 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
481 * For recording page's order, we use page_private(page).
483 static inline int page_is_buddy(struct page *page, struct page *buddy,
486 if (!pfn_valid_within(page_to_pfn(buddy)))
489 if (page_zone_id(page) != page_zone_id(buddy))
492 if (page_is_guard(buddy) && page_order(buddy) == order) {
493 VM_BUG_ON(page_count(buddy) != 0);
497 if (PageBuddy(buddy) && page_order(buddy) == order) {
498 VM_BUG_ON(page_count(buddy) != 0);
505 * Freeing function for a buddy system allocator.
507 * The concept of a buddy system is to maintain direct-mapped table
508 * (containing bit values) for memory blocks of various "orders".
509 * The bottom level table contains the map for the smallest allocatable
510 * units of memory (here, pages), and each level above it describes
511 * pairs of units from the levels below, hence, "buddies".
512 * At a high level, all that happens here is marking the table entry
513 * at the bottom level available, and propagating the changes upward
514 * as necessary, plus some accounting needed to play nicely with other
515 * parts of the VM system.
516 * At each level, we keep a list of pages, which are heads of continuous
517 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
518 * order is recorded in page_private(page) field.
519 * So when we are allocating or freeing one, we can derive the state of the
520 * other. That is, if we allocate a small block, and both were
521 * free, the remainder of the region must be split into blocks.
522 * If a block is freed, and its buddy is also free, then this
523 * triggers coalescing into a block of larger size.
528 static inline void __free_one_page(struct page *page,
529 struct zone *zone, unsigned int order,
532 unsigned long page_idx;
533 unsigned long combined_idx;
534 unsigned long uninitialized_var(buddy_idx);
537 if (unlikely(PageCompound(page)))
538 if (unlikely(destroy_compound_page(page, order)))
541 VM_BUG_ON(migratetype == -1);
543 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
545 VM_BUG_ON(page_idx & ((1 << order) - 1));
546 VM_BUG_ON(bad_range(zone, page));
548 while (order < MAX_ORDER-1) {
549 buddy_idx = __find_buddy_index(page_idx, order);
550 buddy = page + (buddy_idx - page_idx);
551 if (!page_is_buddy(page, buddy, order))
554 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
555 * merge with it and move up one order.
557 if (page_is_guard(buddy)) {
558 clear_page_guard_flag(buddy);
559 set_page_private(page, 0);
560 __mod_zone_freepage_state(zone, 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 = higher_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++;
600 static inline int free_pages_check(struct page *page)
602 if (unlikely(page_mapcount(page) |
603 (page->mapping != NULL) |
604 (atomic_read(&page->_count) != 0) |
605 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
606 (mem_cgroup_bad_page_check(page)))) {
610 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
611 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
616 * Frees a number of pages from the PCP lists
617 * Assumes all pages on list are in same zone, and of same order.
618 * count is the number of pages to free.
620 * If the zone was previously in an "all pages pinned" state then look to
621 * see if this freeing clears that state.
623 * And clear the zone's pages_scanned counter, to hold off the "all pages are
624 * pinned" detection logic.
626 static void free_pcppages_bulk(struct zone *zone, int count,
627 struct per_cpu_pages *pcp)
633 spin_lock(&zone->lock);
634 zone->all_unreclaimable = 0;
635 zone->pages_scanned = 0;
639 struct list_head *list;
642 * Remove pages from lists in a round-robin fashion. A
643 * batch_free count is maintained that is incremented when an
644 * empty list is encountered. This is so more pages are freed
645 * off fuller lists instead of spinning excessively around empty
650 if (++migratetype == MIGRATE_PCPTYPES)
652 list = &pcp->lists[migratetype];
653 } while (list_empty(list));
655 /* This is the only non-empty list. Free them all. */
656 if (batch_free == MIGRATE_PCPTYPES)
657 batch_free = to_free;
660 int mt; /* migratetype of the to-be-freed page */
662 page = list_entry(list->prev, struct page, lru);
663 /* must delete as __free_one_page list manipulates */
664 list_del(&page->lru);
665 mt = get_freepage_migratetype(page);
666 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
667 __free_one_page(page, zone, 0, mt);
668 trace_mm_page_pcpu_drain(page, 0, mt);
669 if (is_migrate_cma(mt))
670 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
671 } while (--to_free && --batch_free && !list_empty(list));
673 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
674 spin_unlock(&zone->lock);
677 static void free_one_page(struct zone *zone, struct page *page, int order,
680 spin_lock(&zone->lock);
681 zone->all_unreclaimable = 0;
682 zone->pages_scanned = 0;
684 __free_one_page(page, zone, order, migratetype);
685 if (unlikely(migratetype != MIGRATE_ISOLATE))
686 __mod_zone_freepage_state(zone, 1 << order, migratetype);
687 spin_unlock(&zone->lock);
690 static bool free_pages_prepare(struct page *page, unsigned int order)
695 trace_mm_page_free(page, order);
696 kmemcheck_free_shadow(page, order);
699 page->mapping = NULL;
700 for (i = 0; i < (1 << order); i++)
701 bad += free_pages_check(page + i);
705 if (!PageHighMem(page)) {
706 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
707 debug_check_no_obj_freed(page_address(page),
710 arch_free_page(page, order);
711 kernel_map_pages(page, 1 << order, 0);
716 static void __free_pages_ok(struct page *page, unsigned int order)
721 if (!free_pages_prepare(page, order))
724 local_irq_save(flags);
725 __count_vm_events(PGFREE, 1 << order);
726 migratetype = get_pageblock_migratetype(page);
727 set_freepage_migratetype(page, migratetype);
728 free_one_page(page_zone(page), page, order, migratetype);
729 local_irq_restore(flags);
732 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
734 unsigned int nr_pages = 1 << order;
738 for (loop = 0; loop < nr_pages; loop++) {
739 struct page *p = &page[loop];
741 if (loop + 1 < nr_pages)
743 __ClearPageReserved(p);
744 set_page_count(p, 0);
747 set_page_refcounted(page);
748 __free_pages(page, order);
752 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
753 void __init init_cma_reserved_pageblock(struct page *page)
755 unsigned i = pageblock_nr_pages;
756 struct page *p = page;
759 __ClearPageReserved(p);
760 set_page_count(p, 0);
763 set_page_refcounted(page);
764 set_pageblock_migratetype(page, MIGRATE_CMA);
765 __free_pages(page, pageblock_order);
766 totalram_pages += pageblock_nr_pages;
771 * The order of subdivision here is critical for the IO subsystem.
772 * Please do not alter this order without good reasons and regression
773 * testing. Specifically, as large blocks of memory are subdivided,
774 * the order in which smaller blocks are delivered depends on the order
775 * they're subdivided in this function. This is the primary factor
776 * influencing the order in which pages are delivered to the IO
777 * subsystem according to empirical testing, and this is also justified
778 * by considering the behavior of a buddy system containing a single
779 * large block of memory acted on by a series of small allocations.
780 * This behavior is a critical factor in sglist merging's success.
784 static inline void expand(struct zone *zone, struct page *page,
785 int low, int high, struct free_area *area,
788 unsigned long size = 1 << high;
794 VM_BUG_ON(bad_range(zone, &page[size]));
796 #ifdef CONFIG_DEBUG_PAGEALLOC
797 if (high < debug_guardpage_minorder()) {
799 * Mark as guard pages (or page), that will allow to
800 * merge back to allocator when buddy will be freed.
801 * Corresponding page table entries will not be touched,
802 * pages will stay not present in virtual address space
804 INIT_LIST_HEAD(&page[size].lru);
805 set_page_guard_flag(&page[size]);
806 set_page_private(&page[size], high);
807 /* Guard pages are not available for any usage */
808 __mod_zone_freepage_state(zone, -(1 << high),
813 list_add(&page[size].lru, &area->free_list[migratetype]);
815 set_page_order(&page[size], high);
820 * This page is about to be returned from the page allocator
822 static inline int check_new_page(struct page *page)
824 if (unlikely(page_mapcount(page) |
825 (page->mapping != NULL) |
826 (atomic_read(&page->_count) != 0) |
827 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
828 (mem_cgroup_bad_page_check(page)))) {
835 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
839 for (i = 0; i < (1 << order); i++) {
840 struct page *p = page + i;
841 if (unlikely(check_new_page(p)))
845 set_page_private(page, 0);
846 set_page_refcounted(page);
848 arch_alloc_page(page, order);
849 kernel_map_pages(page, 1 << order, 1);
851 if (gfp_flags & __GFP_ZERO)
852 prep_zero_page(page, order, gfp_flags);
854 if (order && (gfp_flags & __GFP_COMP))
855 prep_compound_page(page, order);
861 * Go through the free lists for the given migratetype and remove
862 * the smallest available page from the freelists
865 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
868 unsigned int current_order;
869 struct free_area * area;
872 /* Find a page of the appropriate size in the preferred list */
873 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
874 area = &(zone->free_area[current_order]);
875 if (list_empty(&area->free_list[migratetype]))
878 page = list_entry(area->free_list[migratetype].next,
880 list_del(&page->lru);
881 rmv_page_order(page);
883 expand(zone, page, order, current_order, area, migratetype);
892 * This array describes the order lists are fallen back to when
893 * the free lists for the desirable migrate type are depleted
895 static int fallbacks[MIGRATE_TYPES][4] = {
896 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
897 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
899 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
900 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
902 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
904 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
905 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
909 * Move the free pages in a range to the free lists of the requested type.
910 * Note that start_page and end_pages are not aligned on a pageblock
911 * boundary. If alignment is required, use move_freepages_block()
913 int move_freepages(struct zone *zone,
914 struct page *start_page, struct page *end_page,
921 #ifndef CONFIG_HOLES_IN_ZONE
923 * page_zone is not safe to call in this context when
924 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
925 * anyway as we check zone boundaries in move_freepages_block().
926 * Remove at a later date when no bug reports exist related to
927 * grouping pages by mobility
929 BUG_ON(page_zone(start_page) != page_zone(end_page));
932 for (page = start_page; page <= end_page;) {
933 /* Make sure we are not inadvertently changing nodes */
934 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
936 if (!pfn_valid_within(page_to_pfn(page))) {
941 if (!PageBuddy(page)) {
946 order = page_order(page);
947 list_move(&page->lru,
948 &zone->free_area[order].free_list[migratetype]);
949 set_freepage_migratetype(page, migratetype);
951 pages_moved += 1 << order;
957 int move_freepages_block(struct zone *zone, struct page *page,
960 unsigned long start_pfn, end_pfn;
961 struct page *start_page, *end_page;
963 start_pfn = page_to_pfn(page);
964 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
965 start_page = pfn_to_page(start_pfn);
966 end_page = start_page + pageblock_nr_pages - 1;
967 end_pfn = start_pfn + pageblock_nr_pages - 1;
969 /* Do not cross zone boundaries */
970 if (start_pfn < zone->zone_start_pfn)
972 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
975 return move_freepages(zone, start_page, end_page, migratetype);
978 static void change_pageblock_range(struct page *pageblock_page,
979 int start_order, int migratetype)
981 int nr_pageblocks = 1 << (start_order - pageblock_order);
983 while (nr_pageblocks--) {
984 set_pageblock_migratetype(pageblock_page, migratetype);
985 pageblock_page += pageblock_nr_pages;
989 /* Remove an element from the buddy allocator from the fallback list */
990 static inline struct page *
991 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
993 struct free_area * area;
998 /* Find the largest possible block of pages in the other list */
999 for (current_order = MAX_ORDER-1; current_order >= order;
1002 migratetype = fallbacks[start_migratetype][i];
1004 /* MIGRATE_RESERVE handled later if necessary */
1005 if (migratetype == MIGRATE_RESERVE)
1008 area = &(zone->free_area[current_order]);
1009 if (list_empty(&area->free_list[migratetype]))
1012 page = list_entry(area->free_list[migratetype].next,
1017 * If breaking a large block of pages, move all free
1018 * pages to the preferred allocation list. If falling
1019 * back for a reclaimable kernel allocation, be more
1020 * aggressive about taking ownership of free pages
1022 * On the other hand, never change migration
1023 * type of MIGRATE_CMA pageblocks nor move CMA
1024 * pages on different free lists. We don't
1025 * want unmovable pages to be allocated from
1026 * MIGRATE_CMA areas.
1028 if (!is_migrate_cma(migratetype) &&
1029 (unlikely(current_order >= pageblock_order / 2) ||
1030 start_migratetype == MIGRATE_RECLAIMABLE ||
1031 page_group_by_mobility_disabled)) {
1033 pages = move_freepages_block(zone, page,
1036 /* Claim the whole block if over half of it is free */
1037 if (pages >= (1 << (pageblock_order-1)) ||
1038 page_group_by_mobility_disabled)
1039 set_pageblock_migratetype(page,
1042 migratetype = start_migratetype;
1045 /* Remove the page from the freelists */
1046 list_del(&page->lru);
1047 rmv_page_order(page);
1049 /* Take ownership for orders >= pageblock_order */
1050 if (current_order >= pageblock_order &&
1051 !is_migrate_cma(migratetype))
1052 change_pageblock_range(page, current_order,
1055 expand(zone, page, order, current_order, area,
1056 is_migrate_cma(migratetype)
1057 ? migratetype : start_migratetype);
1059 trace_mm_page_alloc_extfrag(page, order, current_order,
1060 start_migratetype, migratetype);
1070 * Do the hard work of removing an element from the buddy allocator.
1071 * Call me with the zone->lock already held.
1073 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1079 page = __rmqueue_smallest(zone, order, migratetype);
1081 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1082 page = __rmqueue_fallback(zone, order, migratetype);
1085 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1086 * is used because __rmqueue_smallest is an inline function
1087 * and we want just one call site
1090 migratetype = MIGRATE_RESERVE;
1095 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1100 * Obtain a specified number of elements from the buddy allocator, all under
1101 * a single hold of the lock, for efficiency. Add them to the supplied list.
1102 * Returns the number of new pages which were placed at *list.
1104 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1105 unsigned long count, struct list_head *list,
1106 int migratetype, int cold)
1108 int mt = migratetype, i;
1110 spin_lock(&zone->lock);
1111 for (i = 0; i < count; ++i) {
1112 struct page *page = __rmqueue(zone, order, migratetype);
1113 if (unlikely(page == NULL))
1117 * Split buddy pages returned by expand() are received here
1118 * in physical page order. The page is added to the callers and
1119 * list and the list head then moves forward. From the callers
1120 * perspective, the linked list is ordered by page number in
1121 * some conditions. This is useful for IO devices that can
1122 * merge IO requests if the physical pages are ordered
1125 if (likely(cold == 0))
1126 list_add(&page->lru, list);
1128 list_add_tail(&page->lru, list);
1129 if (IS_ENABLED(CONFIG_CMA)) {
1130 mt = get_pageblock_migratetype(page);
1131 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1134 set_freepage_migratetype(page, mt);
1136 if (is_migrate_cma(mt))
1137 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1140 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1141 spin_unlock(&zone->lock);
1147 * Called from the vmstat counter updater to drain pagesets of this
1148 * currently executing processor on remote nodes after they have
1151 * Note that this function must be called with the thread pinned to
1152 * a single processor.
1154 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1156 unsigned long flags;
1159 local_irq_save(flags);
1160 if (pcp->count >= pcp->batch)
1161 to_drain = pcp->batch;
1163 to_drain = pcp->count;
1165 free_pcppages_bulk(zone, to_drain, pcp);
1166 pcp->count -= to_drain;
1168 local_irq_restore(flags);
1173 * Drain pages of the indicated processor.
1175 * The processor must either be the current processor and the
1176 * thread pinned to the current processor or a processor that
1179 static void drain_pages(unsigned int cpu)
1181 unsigned long flags;
1184 for_each_populated_zone(zone) {
1185 struct per_cpu_pageset *pset;
1186 struct per_cpu_pages *pcp;
1188 local_irq_save(flags);
1189 pset = per_cpu_ptr(zone->pageset, cpu);
1193 free_pcppages_bulk(zone, pcp->count, pcp);
1196 local_irq_restore(flags);
1201 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1203 void drain_local_pages(void *arg)
1205 drain_pages(smp_processor_id());
1209 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1211 * Note that this code is protected against sending an IPI to an offline
1212 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1213 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1214 * nothing keeps CPUs from showing up after we populated the cpumask and
1215 * before the call to on_each_cpu_mask().
1217 void drain_all_pages(void)
1220 struct per_cpu_pageset *pcp;
1224 * Allocate in the BSS so we wont require allocation in
1225 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1227 static cpumask_t cpus_with_pcps;
1230 * We don't care about racing with CPU hotplug event
1231 * as offline notification will cause the notified
1232 * cpu to drain that CPU pcps and on_each_cpu_mask
1233 * disables preemption as part of its processing
1235 for_each_online_cpu(cpu) {
1236 bool has_pcps = false;
1237 for_each_populated_zone(zone) {
1238 pcp = per_cpu_ptr(zone->pageset, cpu);
1239 if (pcp->pcp.count) {
1245 cpumask_set_cpu(cpu, &cpus_with_pcps);
1247 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1249 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1252 #ifdef CONFIG_HIBERNATION
1254 void mark_free_pages(struct zone *zone)
1256 unsigned long pfn, max_zone_pfn;
1257 unsigned long flags;
1259 struct list_head *curr;
1261 if (!zone->spanned_pages)
1264 spin_lock_irqsave(&zone->lock, flags);
1266 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1267 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1268 if (pfn_valid(pfn)) {
1269 struct page *page = pfn_to_page(pfn);
1271 if (!swsusp_page_is_forbidden(page))
1272 swsusp_unset_page_free(page);
1275 for_each_migratetype_order(order, t) {
1276 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1279 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1280 for (i = 0; i < (1UL << order); i++)
1281 swsusp_set_page_free(pfn_to_page(pfn + i));
1284 spin_unlock_irqrestore(&zone->lock, flags);
1286 #endif /* CONFIG_PM */
1289 * Free a 0-order page
1290 * cold == 1 ? free a cold page : free a hot page
1292 void free_hot_cold_page(struct page *page, int cold)
1294 struct zone *zone = page_zone(page);
1295 struct per_cpu_pages *pcp;
1296 unsigned long flags;
1299 if (!free_pages_prepare(page, 0))
1302 migratetype = get_pageblock_migratetype(page);
1303 set_freepage_migratetype(page, migratetype);
1304 local_irq_save(flags);
1305 __count_vm_event(PGFREE);
1308 * We only track unmovable, reclaimable and movable on pcp lists.
1309 * Free ISOLATE pages back to the allocator because they are being
1310 * offlined but treat RESERVE as movable pages so we can get those
1311 * areas back if necessary. Otherwise, we may have to free
1312 * excessively into the page allocator
1314 if (migratetype >= MIGRATE_PCPTYPES) {
1315 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1316 free_one_page(zone, page, 0, migratetype);
1319 migratetype = MIGRATE_MOVABLE;
1322 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1324 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1326 list_add(&page->lru, &pcp->lists[migratetype]);
1328 if (pcp->count >= pcp->high) {
1329 free_pcppages_bulk(zone, pcp->batch, pcp);
1330 pcp->count -= pcp->batch;
1334 local_irq_restore(flags);
1338 * Free a list of 0-order pages
1340 void free_hot_cold_page_list(struct list_head *list, int cold)
1342 struct page *page, *next;
1344 list_for_each_entry_safe(page, next, list, lru) {
1345 trace_mm_page_free_batched(page, cold);
1346 free_hot_cold_page(page, cold);
1351 * split_page takes a non-compound higher-order page, and splits it into
1352 * n (1<<order) sub-pages: page[0..n]
1353 * Each sub-page must be freed individually.
1355 * Note: this is probably too low level an operation for use in drivers.
1356 * Please consult with lkml before using this in your driver.
1358 void split_page(struct page *page, unsigned int order)
1362 VM_BUG_ON(PageCompound(page));
1363 VM_BUG_ON(!page_count(page));
1365 #ifdef CONFIG_KMEMCHECK
1367 * Split shadow pages too, because free(page[0]) would
1368 * otherwise free the whole shadow.
1370 if (kmemcheck_page_is_tracked(page))
1371 split_page(virt_to_page(page[0].shadow), order);
1374 for (i = 1; i < (1 << order); i++)
1375 set_page_refcounted(page + i);
1379 * Similar to the split_page family of functions except that the page
1380 * required at the given order and being isolated now to prevent races
1381 * with parallel allocators
1383 int capture_free_page(struct page *page, int alloc_order, int migratetype)
1386 unsigned long watermark;
1390 BUG_ON(!PageBuddy(page));
1392 zone = page_zone(page);
1393 order = page_order(page);
1395 /* Obey watermarks as if the page was being allocated */
1396 watermark = low_wmark_pages(zone) + (1 << order);
1397 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1400 /* Remove page from free list */
1401 list_del(&page->lru);
1402 zone->free_area[order].nr_free--;
1403 rmv_page_order(page);
1405 mt = get_pageblock_migratetype(page);
1406 if (unlikely(mt != MIGRATE_ISOLATE))
1407 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1409 if (alloc_order != order)
1410 expand(zone, page, alloc_order, order,
1411 &zone->free_area[order], migratetype);
1413 /* Set the pageblock if the captured page is at least a pageblock */
1414 if (order >= pageblock_order - 1) {
1415 struct page *endpage = page + (1 << order) - 1;
1416 for (; page < endpage; page += pageblock_nr_pages) {
1417 int mt = get_pageblock_migratetype(page);
1418 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1419 set_pageblock_migratetype(page,
1424 return 1UL << order;
1428 * Similar to split_page except the page is already free. As this is only
1429 * being used for migration, the migratetype of the block also changes.
1430 * As this is called with interrupts disabled, the caller is responsible
1431 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1434 * Note: this is probably too low level an operation for use in drivers.
1435 * Please consult with lkml before using this in your driver.
1437 int split_free_page(struct page *page)
1442 BUG_ON(!PageBuddy(page));
1443 order = page_order(page);
1445 nr_pages = capture_free_page(page, order, 0);
1449 /* Split into individual pages */
1450 set_page_refcounted(page);
1451 split_page(page, order);
1456 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1457 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1461 struct page *buffered_rmqueue(struct zone *preferred_zone,
1462 struct zone *zone, int order, gfp_t gfp_flags,
1465 unsigned long flags;
1467 int cold = !!(gfp_flags & __GFP_COLD);
1470 if (likely(order == 0)) {
1471 struct per_cpu_pages *pcp;
1472 struct list_head *list;
1474 local_irq_save(flags);
1475 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1476 list = &pcp->lists[migratetype];
1477 if (list_empty(list)) {
1478 pcp->count += rmqueue_bulk(zone, 0,
1481 if (unlikely(list_empty(list)))
1486 page = list_entry(list->prev, struct page, lru);
1488 page = list_entry(list->next, struct page, lru);
1490 list_del(&page->lru);
1493 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1495 * __GFP_NOFAIL is not to be used in new code.
1497 * All __GFP_NOFAIL callers should be fixed so that they
1498 * properly detect and handle allocation failures.
1500 * We most definitely don't want callers attempting to
1501 * allocate greater than order-1 page units with
1504 WARN_ON_ONCE(order > 1);
1506 spin_lock_irqsave(&zone->lock, flags);
1507 page = __rmqueue(zone, order, migratetype);
1508 spin_unlock(&zone->lock);
1511 __mod_zone_freepage_state(zone, -(1 << order),
1512 get_pageblock_migratetype(page));
1515 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1516 zone_statistics(preferred_zone, zone, gfp_flags);
1517 local_irq_restore(flags);
1519 VM_BUG_ON(bad_range(zone, page));
1520 if (prep_new_page(page, order, gfp_flags))
1525 local_irq_restore(flags);
1529 #ifdef CONFIG_FAIL_PAGE_ALLOC
1532 struct fault_attr attr;
1534 u32 ignore_gfp_highmem;
1535 u32 ignore_gfp_wait;
1537 } fail_page_alloc = {
1538 .attr = FAULT_ATTR_INITIALIZER,
1539 .ignore_gfp_wait = 1,
1540 .ignore_gfp_highmem = 1,
1544 static int __init setup_fail_page_alloc(char *str)
1546 return setup_fault_attr(&fail_page_alloc.attr, str);
1548 __setup("fail_page_alloc=", setup_fail_page_alloc);
1550 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1552 if (order < fail_page_alloc.min_order)
1554 if (gfp_mask & __GFP_NOFAIL)
1556 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1558 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1561 return should_fail(&fail_page_alloc.attr, 1 << order);
1564 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1566 static int __init fail_page_alloc_debugfs(void)
1568 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1571 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1572 &fail_page_alloc.attr);
1574 return PTR_ERR(dir);
1576 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1577 &fail_page_alloc.ignore_gfp_wait))
1579 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1580 &fail_page_alloc.ignore_gfp_highmem))
1582 if (!debugfs_create_u32("min-order", mode, dir,
1583 &fail_page_alloc.min_order))
1588 debugfs_remove_recursive(dir);
1593 late_initcall(fail_page_alloc_debugfs);
1595 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1597 #else /* CONFIG_FAIL_PAGE_ALLOC */
1599 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1604 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1607 * Return true if free pages are above 'mark'. This takes into account the order
1608 * of the allocation.
1610 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1611 int classzone_idx, int alloc_flags, long free_pages)
1613 /* free_pages my go negative - that's OK */
1615 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1618 free_pages -= (1 << order) - 1;
1619 if (alloc_flags & ALLOC_HIGH)
1621 if (alloc_flags & ALLOC_HARDER)
1624 /* If allocation can't use CMA areas don't use free CMA pages */
1625 if (!(alloc_flags & ALLOC_CMA))
1626 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1628 if (free_pages <= min + lowmem_reserve)
1630 for (o = 0; o < order; o++) {
1631 /* At the next order, this order's pages become unavailable */
1632 free_pages -= z->free_area[o].nr_free << o;
1634 /* Require fewer higher order pages to be free */
1637 if (free_pages <= min)
1643 #ifdef CONFIG_MEMORY_ISOLATION
1644 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1646 if (unlikely(zone->nr_pageblock_isolate))
1647 return zone->nr_pageblock_isolate * pageblock_nr_pages;
1651 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1657 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1658 int classzone_idx, int alloc_flags)
1660 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1661 zone_page_state(z, NR_FREE_PAGES));
1664 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1665 int classzone_idx, int alloc_flags)
1667 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1669 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1670 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1673 * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1674 * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1675 * sleep although it could do so. But this is more desirable for memory
1676 * hotplug than sleeping which can cause a livelock in the direct
1679 free_pages -= nr_zone_isolate_freepages(z);
1680 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1686 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1687 * skip over zones that are not allowed by the cpuset, or that have
1688 * been recently (in last second) found to be nearly full. See further
1689 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1690 * that have to skip over a lot of full or unallowed zones.
1692 * If the zonelist cache is present in the passed in zonelist, then
1693 * returns a pointer to the allowed node mask (either the current
1694 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1696 * If the zonelist cache is not available for this zonelist, does
1697 * nothing and returns NULL.
1699 * If the fullzones BITMAP in the zonelist cache is stale (more than
1700 * a second since last zap'd) then we zap it out (clear its bits.)
1702 * We hold off even calling zlc_setup, until after we've checked the
1703 * first zone in the zonelist, on the theory that most allocations will
1704 * be satisfied from that first zone, so best to examine that zone as
1705 * quickly as we can.
1707 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1709 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1710 nodemask_t *allowednodes; /* zonelist_cache approximation */
1712 zlc = zonelist->zlcache_ptr;
1716 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1717 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1718 zlc->last_full_zap = jiffies;
1721 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1722 &cpuset_current_mems_allowed :
1723 &node_states[N_HIGH_MEMORY];
1724 return allowednodes;
1728 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1729 * if it is worth looking at further for free memory:
1730 * 1) Check that the zone isn't thought to be full (doesn't have its
1731 * bit set in the zonelist_cache fullzones BITMAP).
1732 * 2) Check that the zones node (obtained from the zonelist_cache
1733 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1734 * Return true (non-zero) if zone is worth looking at further, or
1735 * else return false (zero) if it is not.
1737 * This check -ignores- the distinction between various watermarks,
1738 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1739 * found to be full for any variation of these watermarks, it will
1740 * be considered full for up to one second by all requests, unless
1741 * we are so low on memory on all allowed nodes that we are forced
1742 * into the second scan of the zonelist.
1744 * In the second scan we ignore this zonelist cache and exactly
1745 * apply the watermarks to all zones, even it is slower to do so.
1746 * We are low on memory in the second scan, and should leave no stone
1747 * unturned looking for a free page.
1749 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1750 nodemask_t *allowednodes)
1752 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1753 int i; /* index of *z in zonelist zones */
1754 int n; /* node that zone *z is on */
1756 zlc = zonelist->zlcache_ptr;
1760 i = z - zonelist->_zonerefs;
1763 /* This zone is worth trying if it is allowed but not full */
1764 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1768 * Given 'z' scanning a zonelist, set the corresponding bit in
1769 * zlc->fullzones, so that subsequent attempts to allocate a page
1770 * from that zone don't waste time re-examining it.
1772 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1774 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1775 int i; /* index of *z in zonelist zones */
1777 zlc = zonelist->zlcache_ptr;
1781 i = z - zonelist->_zonerefs;
1783 set_bit(i, zlc->fullzones);
1787 * clear all zones full, called after direct reclaim makes progress so that
1788 * a zone that was recently full is not skipped over for up to a second
1790 static void zlc_clear_zones_full(struct zonelist *zonelist)
1792 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1794 zlc = zonelist->zlcache_ptr;
1798 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1801 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1803 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1806 static void __paginginit init_zone_allows_reclaim(int nid)
1810 for_each_online_node(i)
1811 if (node_distance(nid, i) <= RECLAIM_DISTANCE) {
1812 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1813 zone_reclaim_mode = 1;
1817 #else /* CONFIG_NUMA */
1819 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1824 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1825 nodemask_t *allowednodes)
1830 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1834 static void zlc_clear_zones_full(struct zonelist *zonelist)
1838 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1843 static inline void init_zone_allows_reclaim(int nid)
1846 #endif /* CONFIG_NUMA */
1849 * get_page_from_freelist goes through the zonelist trying to allocate
1852 static struct page *
1853 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1854 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1855 struct zone *preferred_zone, int migratetype)
1858 struct page *page = NULL;
1861 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1862 int zlc_active = 0; /* set if using zonelist_cache */
1863 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1865 classzone_idx = zone_idx(preferred_zone);
1868 * Scan zonelist, looking for a zone with enough free.
1869 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1871 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1872 high_zoneidx, nodemask) {
1873 if (NUMA_BUILD && zlc_active &&
1874 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1876 if ((alloc_flags & ALLOC_CPUSET) &&
1877 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1880 * When allocating a page cache page for writing, we
1881 * want to get it from a zone that is within its dirty
1882 * limit, such that no single zone holds more than its
1883 * proportional share of globally allowed dirty pages.
1884 * The dirty limits take into account the zone's
1885 * lowmem reserves and high watermark so that kswapd
1886 * should be able to balance it without having to
1887 * write pages from its LRU list.
1889 * This may look like it could increase pressure on
1890 * lower zones by failing allocations in higher zones
1891 * before they are full. But the pages that do spill
1892 * over are limited as the lower zones are protected
1893 * by this very same mechanism. It should not become
1894 * a practical burden to them.
1896 * XXX: For now, allow allocations to potentially
1897 * exceed the per-zone dirty limit in the slowpath
1898 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1899 * which is important when on a NUMA setup the allowed
1900 * zones are together not big enough to reach the
1901 * global limit. The proper fix for these situations
1902 * will require awareness of zones in the
1903 * dirty-throttling and the flusher threads.
1905 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1906 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1907 goto this_zone_full;
1909 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1910 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1914 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1915 if (zone_watermark_ok(zone, order, mark,
1916 classzone_idx, alloc_flags))
1919 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1921 * we do zlc_setup if there are multiple nodes
1922 * and before considering the first zone allowed
1925 allowednodes = zlc_setup(zonelist, alloc_flags);
1930 if (zone_reclaim_mode == 0 ||
1931 !zone_allows_reclaim(preferred_zone, zone))
1932 goto this_zone_full;
1935 * As we may have just activated ZLC, check if the first
1936 * eligible zone has failed zone_reclaim recently.
1938 if (NUMA_BUILD && zlc_active &&
1939 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1942 ret = zone_reclaim(zone, gfp_mask, order);
1944 case ZONE_RECLAIM_NOSCAN:
1947 case ZONE_RECLAIM_FULL:
1948 /* scanned but unreclaimable */
1951 /* did we reclaim enough */
1952 if (!zone_watermark_ok(zone, order, mark,
1953 classzone_idx, alloc_flags))
1954 goto this_zone_full;
1959 page = buffered_rmqueue(preferred_zone, zone, order,
1960 gfp_mask, migratetype);
1965 zlc_mark_zone_full(zonelist, z);
1968 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1969 /* Disable zlc cache for second zonelist scan */
1976 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1977 * necessary to allocate the page. The expectation is
1978 * that the caller is taking steps that will free more
1979 * memory. The caller should avoid the page being used
1980 * for !PFMEMALLOC purposes.
1982 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1988 * Large machines with many possible nodes should not always dump per-node
1989 * meminfo in irq context.
1991 static inline bool should_suppress_show_mem(void)
1996 ret = in_interrupt();
2001 static DEFINE_RATELIMIT_STATE(nopage_rs,
2002 DEFAULT_RATELIMIT_INTERVAL,
2003 DEFAULT_RATELIMIT_BURST);
2005 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2007 unsigned int filter = SHOW_MEM_FILTER_NODES;
2009 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2010 debug_guardpage_minorder() > 0)
2014 * This documents exceptions given to allocations in certain
2015 * contexts that are allowed to allocate outside current's set
2018 if (!(gfp_mask & __GFP_NOMEMALLOC))
2019 if (test_thread_flag(TIF_MEMDIE) ||
2020 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2021 filter &= ~SHOW_MEM_FILTER_NODES;
2022 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2023 filter &= ~SHOW_MEM_FILTER_NODES;
2026 struct va_format vaf;
2029 va_start(args, fmt);
2034 pr_warn("%pV", &vaf);
2039 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2040 current->comm, order, gfp_mask);
2043 if (!should_suppress_show_mem())
2048 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2049 unsigned long did_some_progress,
2050 unsigned long pages_reclaimed)
2052 /* Do not loop if specifically requested */
2053 if (gfp_mask & __GFP_NORETRY)
2056 /* Always retry if specifically requested */
2057 if (gfp_mask & __GFP_NOFAIL)
2061 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2062 * making forward progress without invoking OOM. Suspend also disables
2063 * storage devices so kswapd will not help. Bail if we are suspending.
2065 if (!did_some_progress && pm_suspended_storage())
2069 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2070 * means __GFP_NOFAIL, but that may not be true in other
2073 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2077 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2078 * specified, then we retry until we no longer reclaim any pages
2079 * (above), or we've reclaimed an order of pages at least as
2080 * large as the allocation's order. In both cases, if the
2081 * allocation still fails, we stop retrying.
2083 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2089 static inline struct page *
2090 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2091 struct zonelist *zonelist, enum zone_type high_zoneidx,
2092 nodemask_t *nodemask, struct zone *preferred_zone,
2097 /* Acquire the OOM killer lock for the zones in zonelist */
2098 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2099 schedule_timeout_uninterruptible(1);
2104 * Go through the zonelist yet one more time, keep very high watermark
2105 * here, this is only to catch a parallel oom killing, we must fail if
2106 * we're still under heavy pressure.
2108 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2109 order, zonelist, high_zoneidx,
2110 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2111 preferred_zone, migratetype);
2115 if (!(gfp_mask & __GFP_NOFAIL)) {
2116 /* The OOM killer will not help higher order allocs */
2117 if (order > PAGE_ALLOC_COSTLY_ORDER)
2119 /* The OOM killer does not needlessly kill tasks for lowmem */
2120 if (high_zoneidx < ZONE_NORMAL)
2123 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2124 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2125 * The caller should handle page allocation failure by itself if
2126 * it specifies __GFP_THISNODE.
2127 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2129 if (gfp_mask & __GFP_THISNODE)
2132 /* Exhausted what can be done so it's blamo time */
2133 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2136 clear_zonelist_oom(zonelist, gfp_mask);
2140 #ifdef CONFIG_COMPACTION
2141 /* Try memory compaction for high-order allocations before reclaim */
2142 static struct page *
2143 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2144 struct zonelist *zonelist, enum zone_type high_zoneidx,
2145 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2146 int migratetype, bool sync_migration,
2147 bool *contended_compaction, bool *deferred_compaction,
2148 unsigned long *did_some_progress)
2150 struct page *page = NULL;
2155 if (compaction_deferred(preferred_zone, order)) {
2156 *deferred_compaction = true;
2160 current->flags |= PF_MEMALLOC;
2161 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2162 nodemask, sync_migration,
2163 contended_compaction, &page);
2164 current->flags &= ~PF_MEMALLOC;
2166 /* If compaction captured a page, prep and use it */
2168 prep_new_page(page, order, gfp_mask);
2172 if (*did_some_progress != COMPACT_SKIPPED) {
2173 /* Page migration frees to the PCP lists but we want merging */
2174 drain_pages(get_cpu());
2177 page = get_page_from_freelist(gfp_mask, nodemask,
2178 order, zonelist, high_zoneidx,
2179 alloc_flags & ~ALLOC_NO_WATERMARKS,
2180 preferred_zone, migratetype);
2183 preferred_zone->compact_blockskip_flush = false;
2184 preferred_zone->compact_considered = 0;
2185 preferred_zone->compact_defer_shift = 0;
2186 if (order >= preferred_zone->compact_order_failed)
2187 preferred_zone->compact_order_failed = order + 1;
2188 count_vm_event(COMPACTSUCCESS);
2193 * It's bad if compaction run occurs and fails.
2194 * The most likely reason is that pages exist,
2195 * but not enough to satisfy watermarks.
2197 count_vm_event(COMPACTFAIL);
2200 * As async compaction considers a subset of pageblocks, only
2201 * defer if the failure was a sync compaction failure.
2204 defer_compaction(preferred_zone, order);
2212 static inline struct page *
2213 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2214 struct zonelist *zonelist, enum zone_type high_zoneidx,
2215 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2216 int migratetype, bool sync_migration,
2217 bool *contended_compaction, bool *deferred_compaction,
2218 unsigned long *did_some_progress)
2222 #endif /* CONFIG_COMPACTION */
2224 /* Perform direct synchronous page reclaim */
2226 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2227 nodemask_t *nodemask)
2229 struct reclaim_state reclaim_state;
2234 /* We now go into synchronous reclaim */
2235 cpuset_memory_pressure_bump();
2236 current->flags |= PF_MEMALLOC;
2237 lockdep_set_current_reclaim_state(gfp_mask);
2238 reclaim_state.reclaimed_slab = 0;
2239 current->reclaim_state = &reclaim_state;
2241 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2243 current->reclaim_state = NULL;
2244 lockdep_clear_current_reclaim_state();
2245 current->flags &= ~PF_MEMALLOC;
2252 /* The really slow allocator path where we enter direct reclaim */
2253 static inline struct page *
2254 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2255 struct zonelist *zonelist, enum zone_type high_zoneidx,
2256 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2257 int migratetype, unsigned long *did_some_progress)
2259 struct page *page = NULL;
2260 bool drained = false;
2262 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2264 if (unlikely(!(*did_some_progress)))
2267 /* After successful reclaim, reconsider all zones for allocation */
2269 zlc_clear_zones_full(zonelist);
2272 page = get_page_from_freelist(gfp_mask, nodemask, order,
2273 zonelist, high_zoneidx,
2274 alloc_flags & ~ALLOC_NO_WATERMARKS,
2275 preferred_zone, migratetype);
2278 * If an allocation failed after direct reclaim, it could be because
2279 * pages are pinned on the per-cpu lists. Drain them and try again
2281 if (!page && !drained) {
2291 * This is called in the allocator slow-path if the allocation request is of
2292 * sufficient urgency to ignore watermarks and take other desperate measures
2294 static inline struct page *
2295 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2296 struct zonelist *zonelist, enum zone_type high_zoneidx,
2297 nodemask_t *nodemask, struct zone *preferred_zone,
2303 page = get_page_from_freelist(gfp_mask, nodemask, order,
2304 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2305 preferred_zone, migratetype);
2307 if (!page && gfp_mask & __GFP_NOFAIL)
2308 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2309 } while (!page && (gfp_mask & __GFP_NOFAIL));
2315 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2316 enum zone_type high_zoneidx,
2317 enum zone_type classzone_idx)
2322 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2323 wakeup_kswapd(zone, order, classzone_idx);
2327 gfp_to_alloc_flags(gfp_t gfp_mask)
2329 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2330 const gfp_t wait = gfp_mask & __GFP_WAIT;
2332 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2333 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2336 * The caller may dip into page reserves a bit more if the caller
2337 * cannot run direct reclaim, or if the caller has realtime scheduling
2338 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2339 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2341 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2345 * Not worth trying to allocate harder for
2346 * __GFP_NOMEMALLOC even if it can't schedule.
2348 if (!(gfp_mask & __GFP_NOMEMALLOC))
2349 alloc_flags |= ALLOC_HARDER;
2351 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2352 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2354 alloc_flags &= ~ALLOC_CPUSET;
2355 } else if (unlikely(rt_task(current)) && !in_interrupt())
2356 alloc_flags |= ALLOC_HARDER;
2358 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2359 if (gfp_mask & __GFP_MEMALLOC)
2360 alloc_flags |= ALLOC_NO_WATERMARKS;
2361 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2362 alloc_flags |= ALLOC_NO_WATERMARKS;
2363 else if (!in_interrupt() &&
2364 ((current->flags & PF_MEMALLOC) ||
2365 unlikely(test_thread_flag(TIF_MEMDIE))))
2366 alloc_flags |= ALLOC_NO_WATERMARKS;
2369 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2370 alloc_flags |= ALLOC_CMA;
2375 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2377 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2380 static inline struct page *
2381 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2382 struct zonelist *zonelist, enum zone_type high_zoneidx,
2383 nodemask_t *nodemask, struct zone *preferred_zone,
2386 const gfp_t wait = gfp_mask & __GFP_WAIT;
2387 struct page *page = NULL;
2389 unsigned long pages_reclaimed = 0;
2390 unsigned long did_some_progress;
2391 bool sync_migration = false;
2392 bool deferred_compaction = false;
2393 bool contended_compaction = false;
2396 * In the slowpath, we sanity check order to avoid ever trying to
2397 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2398 * be using allocators in order of preference for an area that is
2401 if (order >= MAX_ORDER) {
2402 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2407 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2408 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2409 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2410 * using a larger set of nodes after it has established that the
2411 * allowed per node queues are empty and that nodes are
2414 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2418 wake_all_kswapd(order, zonelist, high_zoneidx,
2419 zone_idx(preferred_zone));
2422 * OK, we're below the kswapd watermark and have kicked background
2423 * reclaim. Now things get more complex, so set up alloc_flags according
2424 * to how we want to proceed.
2426 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2429 * Find the true preferred zone if the allocation is unconstrained by
2432 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2433 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2437 /* This is the last chance, in general, before the goto nopage. */
2438 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2439 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2440 preferred_zone, migratetype);
2444 /* Allocate without watermarks if the context allows */
2445 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2447 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2448 * the allocation is high priority and these type of
2449 * allocations are system rather than user orientated
2451 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2453 page = __alloc_pages_high_priority(gfp_mask, order,
2454 zonelist, high_zoneidx, nodemask,
2455 preferred_zone, migratetype);
2461 /* Atomic allocations - we can't balance anything */
2465 /* Avoid recursion of direct reclaim */
2466 if (current->flags & PF_MEMALLOC)
2469 /* Avoid allocations with no watermarks from looping endlessly */
2470 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2474 * Try direct compaction. The first pass is asynchronous. Subsequent
2475 * attempts after direct reclaim are synchronous
2477 page = __alloc_pages_direct_compact(gfp_mask, order,
2478 zonelist, high_zoneidx,
2480 alloc_flags, preferred_zone,
2481 migratetype, sync_migration,
2482 &contended_compaction,
2483 &deferred_compaction,
2484 &did_some_progress);
2487 sync_migration = true;
2490 * If compaction is deferred for high-order allocations, it is because
2491 * sync compaction recently failed. In this is the case and the caller
2492 * requested a movable allocation that does not heavily disrupt the
2493 * system then fail the allocation instead of entering direct reclaim.
2495 if ((deferred_compaction || contended_compaction) &&
2496 (gfp_mask & (__GFP_MOVABLE|__GFP_REPEAT)) == __GFP_MOVABLE)
2499 /* Try direct reclaim and then allocating */
2500 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2501 zonelist, high_zoneidx,
2503 alloc_flags, preferred_zone,
2504 migratetype, &did_some_progress);
2509 * If we failed to make any progress reclaiming, then we are
2510 * running out of options and have to consider going OOM
2512 if (!did_some_progress) {
2513 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2514 if (oom_killer_disabled)
2516 /* Coredumps can quickly deplete all memory reserves */
2517 if ((current->flags & PF_DUMPCORE) &&
2518 !(gfp_mask & __GFP_NOFAIL))
2520 page = __alloc_pages_may_oom(gfp_mask, order,
2521 zonelist, high_zoneidx,
2522 nodemask, preferred_zone,
2527 if (!(gfp_mask & __GFP_NOFAIL)) {
2529 * The oom killer is not called for high-order
2530 * allocations that may fail, so if no progress
2531 * is being made, there are no other options and
2532 * retrying is unlikely to help.
2534 if (order > PAGE_ALLOC_COSTLY_ORDER)
2537 * The oom killer is not called for lowmem
2538 * allocations to prevent needlessly killing
2541 if (high_zoneidx < ZONE_NORMAL)
2549 /* Check if we should retry the allocation */
2550 pages_reclaimed += did_some_progress;
2551 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2553 /* Wait for some write requests to complete then retry */
2554 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2558 * High-order allocations do not necessarily loop after
2559 * direct reclaim and reclaim/compaction depends on compaction
2560 * being called after reclaim so call directly if necessary
2562 page = __alloc_pages_direct_compact(gfp_mask, order,
2563 zonelist, high_zoneidx,
2565 alloc_flags, preferred_zone,
2566 migratetype, sync_migration,
2567 &contended_compaction,
2568 &deferred_compaction,
2569 &did_some_progress);
2575 warn_alloc_failed(gfp_mask, order, NULL);
2578 if (kmemcheck_enabled)
2579 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2585 * This is the 'heart' of the zoned buddy allocator.
2588 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2589 struct zonelist *zonelist, nodemask_t *nodemask)
2591 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2592 struct zone *preferred_zone;
2593 struct page *page = NULL;
2594 int migratetype = allocflags_to_migratetype(gfp_mask);
2595 unsigned int cpuset_mems_cookie;
2596 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2598 gfp_mask &= gfp_allowed_mask;
2600 lockdep_trace_alloc(gfp_mask);
2602 might_sleep_if(gfp_mask & __GFP_WAIT);
2604 if (should_fail_alloc_page(gfp_mask, order))
2608 * Check the zones suitable for the gfp_mask contain at least one
2609 * valid zone. It's possible to have an empty zonelist as a result
2610 * of GFP_THISNODE and a memoryless node
2612 if (unlikely(!zonelist->_zonerefs->zone))
2616 cpuset_mems_cookie = get_mems_allowed();
2618 /* The preferred zone is used for statistics later */
2619 first_zones_zonelist(zonelist, high_zoneidx,
2620 nodemask ? : &cpuset_current_mems_allowed,
2622 if (!preferred_zone)
2626 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2627 alloc_flags |= ALLOC_CMA;
2629 /* First allocation attempt */
2630 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2631 zonelist, high_zoneidx, alloc_flags,
2632 preferred_zone, migratetype);
2633 if (unlikely(!page))
2634 page = __alloc_pages_slowpath(gfp_mask, order,
2635 zonelist, high_zoneidx, nodemask,
2636 preferred_zone, migratetype);
2638 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2642 * When updating a task's mems_allowed, it is possible to race with
2643 * parallel threads in such a way that an allocation can fail while
2644 * the mask is being updated. If a page allocation is about to fail,
2645 * check if the cpuset changed during allocation and if so, retry.
2647 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2652 EXPORT_SYMBOL(__alloc_pages_nodemask);
2655 * Common helper functions.
2657 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2662 * __get_free_pages() returns a 32-bit address, which cannot represent
2665 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2667 page = alloc_pages(gfp_mask, order);
2670 return (unsigned long) page_address(page);
2672 EXPORT_SYMBOL(__get_free_pages);
2674 unsigned long get_zeroed_page(gfp_t gfp_mask)
2676 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2678 EXPORT_SYMBOL(get_zeroed_page);
2680 void __free_pages(struct page *page, unsigned int order)
2682 if (put_page_testzero(page)) {
2684 free_hot_cold_page(page, 0);
2686 __free_pages_ok(page, order);
2690 EXPORT_SYMBOL(__free_pages);
2692 void free_pages(unsigned long addr, unsigned int order)
2695 VM_BUG_ON(!virt_addr_valid((void *)addr));
2696 __free_pages(virt_to_page((void *)addr), order);
2700 EXPORT_SYMBOL(free_pages);
2702 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2705 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2706 unsigned long used = addr + PAGE_ALIGN(size);
2708 split_page(virt_to_page((void *)addr), order);
2709 while (used < alloc_end) {
2714 return (void *)addr;
2718 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2719 * @size: the number of bytes to allocate
2720 * @gfp_mask: GFP flags for the allocation
2722 * This function is similar to alloc_pages(), except that it allocates the
2723 * minimum number of pages to satisfy the request. alloc_pages() can only
2724 * allocate memory in power-of-two pages.
2726 * This function is also limited by MAX_ORDER.
2728 * Memory allocated by this function must be released by free_pages_exact().
2730 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2732 unsigned int order = get_order(size);
2735 addr = __get_free_pages(gfp_mask, order);
2736 return make_alloc_exact(addr, order, size);
2738 EXPORT_SYMBOL(alloc_pages_exact);
2741 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2743 * @nid: the preferred node ID where memory should be allocated
2744 * @size: the number of bytes to allocate
2745 * @gfp_mask: GFP flags for the allocation
2747 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2749 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2752 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2754 unsigned order = get_order(size);
2755 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2758 return make_alloc_exact((unsigned long)page_address(p), order, size);
2760 EXPORT_SYMBOL(alloc_pages_exact_nid);
2763 * free_pages_exact - release memory allocated via alloc_pages_exact()
2764 * @virt: the value returned by alloc_pages_exact.
2765 * @size: size of allocation, same value as passed to alloc_pages_exact().
2767 * Release the memory allocated by a previous call to alloc_pages_exact.
2769 void free_pages_exact(void *virt, size_t size)
2771 unsigned long addr = (unsigned long)virt;
2772 unsigned long end = addr + PAGE_ALIGN(size);
2774 while (addr < end) {
2779 EXPORT_SYMBOL(free_pages_exact);
2781 static unsigned int nr_free_zone_pages(int offset)
2786 /* Just pick one node, since fallback list is circular */
2787 unsigned int sum = 0;
2789 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2791 for_each_zone_zonelist(zone, z, zonelist, offset) {
2792 unsigned long size = zone->present_pages;
2793 unsigned long high = high_wmark_pages(zone);
2802 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2804 unsigned int nr_free_buffer_pages(void)
2806 return nr_free_zone_pages(gfp_zone(GFP_USER));
2808 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2811 * Amount of free RAM allocatable within all zones
2813 unsigned int nr_free_pagecache_pages(void)
2815 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2818 static inline void show_node(struct zone *zone)
2821 printk("Node %d ", zone_to_nid(zone));
2824 void si_meminfo(struct sysinfo *val)
2826 val->totalram = totalram_pages;
2828 val->freeram = global_page_state(NR_FREE_PAGES);
2829 val->bufferram = nr_blockdev_pages();
2830 val->totalhigh = totalhigh_pages;
2831 val->freehigh = nr_free_highpages();
2832 val->mem_unit = PAGE_SIZE;
2835 EXPORT_SYMBOL(si_meminfo);
2838 void si_meminfo_node(struct sysinfo *val, int nid)
2840 pg_data_t *pgdat = NODE_DATA(nid);
2842 val->totalram = pgdat->node_present_pages;
2843 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2844 #ifdef CONFIG_HIGHMEM
2845 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2846 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2852 val->mem_unit = PAGE_SIZE;
2857 * Determine whether the node should be displayed or not, depending on whether
2858 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2860 bool skip_free_areas_node(unsigned int flags, int nid)
2863 unsigned int cpuset_mems_cookie;
2865 if (!(flags & SHOW_MEM_FILTER_NODES))
2869 cpuset_mems_cookie = get_mems_allowed();
2870 ret = !node_isset(nid, cpuset_current_mems_allowed);
2871 } while (!put_mems_allowed(cpuset_mems_cookie));
2876 #define K(x) ((x) << (PAGE_SHIFT-10))
2879 * Show free area list (used inside shift_scroll-lock stuff)
2880 * We also calculate the percentage fragmentation. We do this by counting the
2881 * memory on each free list with the exception of the first item on the list.
2882 * Suppresses nodes that are not allowed by current's cpuset if
2883 * SHOW_MEM_FILTER_NODES is passed.
2885 void show_free_areas(unsigned int filter)
2890 for_each_populated_zone(zone) {
2891 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2894 printk("%s per-cpu:\n", zone->name);
2896 for_each_online_cpu(cpu) {
2897 struct per_cpu_pageset *pageset;
2899 pageset = per_cpu_ptr(zone->pageset, cpu);
2901 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2902 cpu, pageset->pcp.high,
2903 pageset->pcp.batch, pageset->pcp.count);
2907 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2908 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2910 " dirty:%lu writeback:%lu unstable:%lu\n"
2911 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2912 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2914 global_page_state(NR_ACTIVE_ANON),
2915 global_page_state(NR_INACTIVE_ANON),
2916 global_page_state(NR_ISOLATED_ANON),
2917 global_page_state(NR_ACTIVE_FILE),
2918 global_page_state(NR_INACTIVE_FILE),
2919 global_page_state(NR_ISOLATED_FILE),
2920 global_page_state(NR_UNEVICTABLE),
2921 global_page_state(NR_FILE_DIRTY),
2922 global_page_state(NR_WRITEBACK),
2923 global_page_state(NR_UNSTABLE_NFS),
2924 global_page_state(NR_FREE_PAGES),
2925 global_page_state(NR_SLAB_RECLAIMABLE),
2926 global_page_state(NR_SLAB_UNRECLAIMABLE),
2927 global_page_state(NR_FILE_MAPPED),
2928 global_page_state(NR_SHMEM),
2929 global_page_state(NR_PAGETABLE),
2930 global_page_state(NR_BOUNCE),
2931 global_page_state(NR_FREE_CMA_PAGES));
2933 for_each_populated_zone(zone) {
2936 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2944 " active_anon:%lukB"
2945 " inactive_anon:%lukB"
2946 " active_file:%lukB"
2947 " inactive_file:%lukB"
2948 " unevictable:%lukB"
2949 " isolated(anon):%lukB"
2950 " isolated(file):%lukB"
2957 " slab_reclaimable:%lukB"
2958 " slab_unreclaimable:%lukB"
2959 " kernel_stack:%lukB"
2964 " writeback_tmp:%lukB"
2965 " pages_scanned:%lu"
2966 " all_unreclaimable? %s"
2969 K(zone_page_state(zone, NR_FREE_PAGES)),
2970 K(min_wmark_pages(zone)),
2971 K(low_wmark_pages(zone)),
2972 K(high_wmark_pages(zone)),
2973 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2974 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2975 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2976 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2977 K(zone_page_state(zone, NR_UNEVICTABLE)),
2978 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2979 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2980 K(zone->present_pages),
2981 K(zone_page_state(zone, NR_MLOCK)),
2982 K(zone_page_state(zone, NR_FILE_DIRTY)),
2983 K(zone_page_state(zone, NR_WRITEBACK)),
2984 K(zone_page_state(zone, NR_FILE_MAPPED)),
2985 K(zone_page_state(zone, NR_SHMEM)),
2986 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2987 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2988 zone_page_state(zone, NR_KERNEL_STACK) *
2990 K(zone_page_state(zone, NR_PAGETABLE)),
2991 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2992 K(zone_page_state(zone, NR_BOUNCE)),
2993 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
2994 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2995 zone->pages_scanned,
2996 (zone->all_unreclaimable ? "yes" : "no")
2998 printk("lowmem_reserve[]:");
2999 for (i = 0; i < MAX_NR_ZONES; i++)
3000 printk(" %lu", zone->lowmem_reserve[i]);
3004 for_each_populated_zone(zone) {
3005 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3007 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3010 printk("%s: ", zone->name);
3012 spin_lock_irqsave(&zone->lock, flags);
3013 for (order = 0; order < MAX_ORDER; order++) {
3014 nr[order] = zone->free_area[order].nr_free;
3015 total += nr[order] << order;
3017 spin_unlock_irqrestore(&zone->lock, flags);
3018 for (order = 0; order < MAX_ORDER; order++)
3019 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3020 printk("= %lukB\n", K(total));
3023 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3025 show_swap_cache_info();
3028 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3030 zoneref->zone = zone;
3031 zoneref->zone_idx = zone_idx(zone);
3035 * Builds allocation fallback zone lists.
3037 * Add all populated zones of a node to the zonelist.
3039 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3040 int nr_zones, enum zone_type zone_type)
3044 BUG_ON(zone_type >= MAX_NR_ZONES);
3049 zone = pgdat->node_zones + zone_type;
3050 if (populated_zone(zone)) {
3051 zoneref_set_zone(zone,
3052 &zonelist->_zonerefs[nr_zones++]);
3053 check_highest_zone(zone_type);
3056 } while (zone_type);
3063 * 0 = automatic detection of better ordering.
3064 * 1 = order by ([node] distance, -zonetype)
3065 * 2 = order by (-zonetype, [node] distance)
3067 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3068 * the same zonelist. So only NUMA can configure this param.
3070 #define ZONELIST_ORDER_DEFAULT 0
3071 #define ZONELIST_ORDER_NODE 1
3072 #define ZONELIST_ORDER_ZONE 2
3074 /* zonelist order in the kernel.
3075 * set_zonelist_order() will set this to NODE or ZONE.
3077 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3078 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3082 /* The value user specified ....changed by config */
3083 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3084 /* string for sysctl */
3085 #define NUMA_ZONELIST_ORDER_LEN 16
3086 char numa_zonelist_order[16] = "default";
3089 * interface for configure zonelist ordering.
3090 * command line option "numa_zonelist_order"
3091 * = "[dD]efault - default, automatic configuration.
3092 * = "[nN]ode - order by node locality, then by zone within node
3093 * = "[zZ]one - order by zone, then by locality within zone
3096 static int __parse_numa_zonelist_order(char *s)
3098 if (*s == 'd' || *s == 'D') {
3099 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3100 } else if (*s == 'n' || *s == 'N') {
3101 user_zonelist_order = ZONELIST_ORDER_NODE;
3102 } else if (*s == 'z' || *s == 'Z') {
3103 user_zonelist_order = ZONELIST_ORDER_ZONE;
3106 "Ignoring invalid numa_zonelist_order value: "
3113 static __init int setup_numa_zonelist_order(char *s)
3120 ret = __parse_numa_zonelist_order(s);
3122 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3126 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3129 * sysctl handler for numa_zonelist_order
3131 int numa_zonelist_order_handler(ctl_table *table, int write,
3132 void __user *buffer, size_t *length,
3135 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3137 static DEFINE_MUTEX(zl_order_mutex);
3139 mutex_lock(&zl_order_mutex);
3141 strcpy(saved_string, (char*)table->data);
3142 ret = proc_dostring(table, write, buffer, length, ppos);
3146 int oldval = user_zonelist_order;
3147 if (__parse_numa_zonelist_order((char*)table->data)) {
3149 * bogus value. restore saved string
3151 strncpy((char*)table->data, saved_string,
3152 NUMA_ZONELIST_ORDER_LEN);
3153 user_zonelist_order = oldval;
3154 } else if (oldval != user_zonelist_order) {
3155 mutex_lock(&zonelists_mutex);
3156 build_all_zonelists(NULL, NULL);
3157 mutex_unlock(&zonelists_mutex);
3161 mutex_unlock(&zl_order_mutex);
3166 #define MAX_NODE_LOAD (nr_online_nodes)
3167 static int node_load[MAX_NUMNODES];
3170 * find_next_best_node - find the next node that should appear in a given node's fallback list
3171 * @node: node whose fallback list we're appending
3172 * @used_node_mask: nodemask_t of already used nodes
3174 * We use a number of factors to determine which is the next node that should
3175 * appear on a given node's fallback list. The node should not have appeared
3176 * already in @node's fallback list, and it should be the next closest node
3177 * according to the distance array (which contains arbitrary distance values
3178 * from each node to each node in the system), and should also prefer nodes
3179 * with no CPUs, since presumably they'll have very little allocation pressure
3180 * on them otherwise.
3181 * It returns -1 if no node is found.
3183 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3186 int min_val = INT_MAX;
3188 const struct cpumask *tmp = cpumask_of_node(0);
3190 /* Use the local node if we haven't already */
3191 if (!node_isset(node, *used_node_mask)) {
3192 node_set(node, *used_node_mask);
3196 for_each_node_state(n, N_HIGH_MEMORY) {
3198 /* Don't want a node to appear more than once */
3199 if (node_isset(n, *used_node_mask))
3202 /* Use the distance array to find the distance */
3203 val = node_distance(node, n);
3205 /* Penalize nodes under us ("prefer the next node") */
3208 /* Give preference to headless and unused nodes */
3209 tmp = cpumask_of_node(n);
3210 if (!cpumask_empty(tmp))
3211 val += PENALTY_FOR_NODE_WITH_CPUS;
3213 /* Slight preference for less loaded node */
3214 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3215 val += node_load[n];
3217 if (val < min_val) {
3224 node_set(best_node, *used_node_mask);
3231 * Build zonelists ordered by node and zones within node.
3232 * This results in maximum locality--normal zone overflows into local
3233 * DMA zone, if any--but risks exhausting DMA zone.
3235 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3238 struct zonelist *zonelist;
3240 zonelist = &pgdat->node_zonelists[0];
3241 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3243 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3245 zonelist->_zonerefs[j].zone = NULL;
3246 zonelist->_zonerefs[j].zone_idx = 0;
3250 * Build gfp_thisnode zonelists
3252 static void build_thisnode_zonelists(pg_data_t *pgdat)
3255 struct zonelist *zonelist;
3257 zonelist = &pgdat->node_zonelists[1];
3258 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3259 zonelist->_zonerefs[j].zone = NULL;
3260 zonelist->_zonerefs[j].zone_idx = 0;
3264 * Build zonelists ordered by zone and nodes within zones.
3265 * This results in conserving DMA zone[s] until all Normal memory is
3266 * exhausted, but results in overflowing to remote node while memory
3267 * may still exist in local DMA zone.
3269 static int node_order[MAX_NUMNODES];
3271 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3274 int zone_type; /* needs to be signed */
3276 struct zonelist *zonelist;
3278 zonelist = &pgdat->node_zonelists[0];
3280 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3281 for (j = 0; j < nr_nodes; j++) {
3282 node = node_order[j];
3283 z = &NODE_DATA(node)->node_zones[zone_type];
3284 if (populated_zone(z)) {
3286 &zonelist->_zonerefs[pos++]);
3287 check_highest_zone(zone_type);
3291 zonelist->_zonerefs[pos].zone = NULL;
3292 zonelist->_zonerefs[pos].zone_idx = 0;
3295 static int default_zonelist_order(void)
3298 unsigned long low_kmem_size,total_size;
3302 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3303 * If they are really small and used heavily, the system can fall
3304 * into OOM very easily.
3305 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3307 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3310 for_each_online_node(nid) {
3311 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3312 z = &NODE_DATA(nid)->node_zones[zone_type];
3313 if (populated_zone(z)) {
3314 if (zone_type < ZONE_NORMAL)
3315 low_kmem_size += z->present_pages;
3316 total_size += z->present_pages;
3317 } else if (zone_type == ZONE_NORMAL) {
3319 * If any node has only lowmem, then node order
3320 * is preferred to allow kernel allocations
3321 * locally; otherwise, they can easily infringe
3322 * on other nodes when there is an abundance of
3323 * lowmem available to allocate from.
3325 return ZONELIST_ORDER_NODE;
3329 if (!low_kmem_size || /* there are no DMA area. */
3330 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3331 return ZONELIST_ORDER_NODE;
3333 * look into each node's config.
3334 * If there is a node whose DMA/DMA32 memory is very big area on
3335 * local memory, NODE_ORDER may be suitable.
3337 average_size = total_size /
3338 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3339 for_each_online_node(nid) {
3342 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3343 z = &NODE_DATA(nid)->node_zones[zone_type];
3344 if (populated_zone(z)) {
3345 if (zone_type < ZONE_NORMAL)
3346 low_kmem_size += z->present_pages;
3347 total_size += z->present_pages;
3350 if (low_kmem_size &&
3351 total_size > average_size && /* ignore small node */
3352 low_kmem_size > total_size * 70/100)
3353 return ZONELIST_ORDER_NODE;
3355 return ZONELIST_ORDER_ZONE;
3358 static void set_zonelist_order(void)
3360 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3361 current_zonelist_order = default_zonelist_order();
3363 current_zonelist_order = user_zonelist_order;
3366 static void build_zonelists(pg_data_t *pgdat)
3370 nodemask_t used_mask;
3371 int local_node, prev_node;
3372 struct zonelist *zonelist;
3373 int order = current_zonelist_order;
3375 /* initialize zonelists */
3376 for (i = 0; i < MAX_ZONELISTS; i++) {
3377 zonelist = pgdat->node_zonelists + i;
3378 zonelist->_zonerefs[0].zone = NULL;
3379 zonelist->_zonerefs[0].zone_idx = 0;
3382 /* NUMA-aware ordering of nodes */
3383 local_node = pgdat->node_id;
3384 load = nr_online_nodes;
3385 prev_node = local_node;
3386 nodes_clear(used_mask);
3388 memset(node_order, 0, sizeof(node_order));
3391 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3393 * We don't want to pressure a particular node.
3394 * So adding penalty to the first node in same
3395 * distance group to make it round-robin.
3397 if (node_distance(local_node, node) !=
3398 node_distance(local_node, prev_node))
3399 node_load[node] = load;
3403 if (order == ZONELIST_ORDER_NODE)
3404 build_zonelists_in_node_order(pgdat, node);
3406 node_order[j++] = node; /* remember order */
3409 if (order == ZONELIST_ORDER_ZONE) {
3410 /* calculate node order -- i.e., DMA last! */
3411 build_zonelists_in_zone_order(pgdat, j);
3414 build_thisnode_zonelists(pgdat);
3417 /* Construct the zonelist performance cache - see further mmzone.h */
3418 static void build_zonelist_cache(pg_data_t *pgdat)
3420 struct zonelist *zonelist;
3421 struct zonelist_cache *zlc;
3424 zonelist = &pgdat->node_zonelists[0];
3425 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3426 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3427 for (z = zonelist->_zonerefs; z->zone; z++)
3428 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3431 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3433 * Return node id of node used for "local" allocations.
3434 * I.e., first node id of first zone in arg node's generic zonelist.
3435 * Used for initializing percpu 'numa_mem', which is used primarily
3436 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3438 int local_memory_node(int node)
3442 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3443 gfp_zone(GFP_KERNEL),
3450 #else /* CONFIG_NUMA */
3452 static void set_zonelist_order(void)
3454 current_zonelist_order = ZONELIST_ORDER_ZONE;
3457 static void build_zonelists(pg_data_t *pgdat)
3459 int node, local_node;
3461 struct zonelist *zonelist;
3463 local_node = pgdat->node_id;
3465 zonelist = &pgdat->node_zonelists[0];
3466 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3469 * Now we build the zonelist so that it contains the zones
3470 * of all the other nodes.
3471 * We don't want to pressure a particular node, so when
3472 * building the zones for node N, we make sure that the
3473 * zones coming right after the local ones are those from
3474 * node N+1 (modulo N)
3476 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3477 if (!node_online(node))
3479 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3482 for (node = 0; node < local_node; node++) {
3483 if (!node_online(node))
3485 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3489 zonelist->_zonerefs[j].zone = NULL;
3490 zonelist->_zonerefs[j].zone_idx = 0;
3493 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3494 static void build_zonelist_cache(pg_data_t *pgdat)
3496 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3499 #endif /* CONFIG_NUMA */
3502 * Boot pageset table. One per cpu which is going to be used for all
3503 * zones and all nodes. The parameters will be set in such a way
3504 * that an item put on a list will immediately be handed over to
3505 * the buddy list. This is safe since pageset manipulation is done
3506 * with interrupts disabled.
3508 * The boot_pagesets must be kept even after bootup is complete for
3509 * unused processors and/or zones. They do play a role for bootstrapping
3510 * hotplugged processors.
3512 * zoneinfo_show() and maybe other functions do
3513 * not check if the processor is online before following the pageset pointer.
3514 * Other parts of the kernel may not check if the zone is available.
3516 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3517 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3518 static void setup_zone_pageset(struct zone *zone);
3521 * Global mutex to protect against size modification of zonelists
3522 * as well as to serialize pageset setup for the new populated zone.
3524 DEFINE_MUTEX(zonelists_mutex);
3526 /* return values int ....just for stop_machine() */
3527 static int __build_all_zonelists(void *data)
3531 pg_data_t *self = data;
3534 memset(node_load, 0, sizeof(node_load));
3537 if (self && !node_online(self->node_id)) {
3538 build_zonelists(self);
3539 build_zonelist_cache(self);
3542 for_each_online_node(nid) {
3543 pg_data_t *pgdat = NODE_DATA(nid);
3545 build_zonelists(pgdat);
3546 build_zonelist_cache(pgdat);
3550 * Initialize the boot_pagesets that are going to be used
3551 * for bootstrapping processors. The real pagesets for
3552 * each zone will be allocated later when the per cpu
3553 * allocator is available.
3555 * boot_pagesets are used also for bootstrapping offline
3556 * cpus if the system is already booted because the pagesets
3557 * are needed to initialize allocators on a specific cpu too.
3558 * F.e. the percpu allocator needs the page allocator which
3559 * needs the percpu allocator in order to allocate its pagesets
3560 * (a chicken-egg dilemma).
3562 for_each_possible_cpu(cpu) {
3563 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3565 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3567 * We now know the "local memory node" for each node--
3568 * i.e., the node of the first zone in the generic zonelist.
3569 * Set up numa_mem percpu variable for on-line cpus. During
3570 * boot, only the boot cpu should be on-line; we'll init the
3571 * secondary cpus' numa_mem as they come on-line. During
3572 * node/memory hotplug, we'll fixup all on-line cpus.
3574 if (cpu_online(cpu))
3575 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3583 * Called with zonelists_mutex held always
3584 * unless system_state == SYSTEM_BOOTING.
3586 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3588 set_zonelist_order();
3590 if (system_state == SYSTEM_BOOTING) {
3591 __build_all_zonelists(NULL);
3592 mminit_verify_zonelist();
3593 cpuset_init_current_mems_allowed();
3595 /* we have to stop all cpus to guarantee there is no user
3597 #ifdef CONFIG_MEMORY_HOTPLUG
3599 setup_zone_pageset(zone);
3601 stop_machine(__build_all_zonelists, pgdat, NULL);
3602 /* cpuset refresh routine should be here */
3604 vm_total_pages = nr_free_pagecache_pages();
3606 * Disable grouping by mobility if the number of pages in the
3607 * system is too low to allow the mechanism to work. It would be
3608 * more accurate, but expensive to check per-zone. This check is
3609 * made on memory-hotadd so a system can start with mobility
3610 * disabled and enable it later
3612 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3613 page_group_by_mobility_disabled = 1;
3615 page_group_by_mobility_disabled = 0;
3617 printk("Built %i zonelists in %s order, mobility grouping %s. "
3618 "Total pages: %ld\n",
3620 zonelist_order_name[current_zonelist_order],
3621 page_group_by_mobility_disabled ? "off" : "on",
3624 printk("Policy zone: %s\n", zone_names[policy_zone]);
3629 * Helper functions to size the waitqueue hash table.
3630 * Essentially these want to choose hash table sizes sufficiently
3631 * large so that collisions trying to wait on pages are rare.
3632 * But in fact, the number of active page waitqueues on typical
3633 * systems is ridiculously low, less than 200. So this is even
3634 * conservative, even though it seems large.
3636 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3637 * waitqueues, i.e. the size of the waitq table given the number of pages.
3639 #define PAGES_PER_WAITQUEUE 256
3641 #ifndef CONFIG_MEMORY_HOTPLUG
3642 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3644 unsigned long size = 1;
3646 pages /= PAGES_PER_WAITQUEUE;
3648 while (size < pages)
3652 * Once we have dozens or even hundreds of threads sleeping
3653 * on IO we've got bigger problems than wait queue collision.
3654 * Limit the size of the wait table to a reasonable size.
3656 size = min(size, 4096UL);
3658 return max(size, 4UL);
3662 * A zone's size might be changed by hot-add, so it is not possible to determine
3663 * a suitable size for its wait_table. So we use the maximum size now.
3665 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3667 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3668 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3669 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3671 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3672 * or more by the traditional way. (See above). It equals:
3674 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3675 * ia64(16K page size) : = ( 8G + 4M)byte.
3676 * powerpc (64K page size) : = (32G +16M)byte.
3678 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3685 * This is an integer logarithm so that shifts can be used later
3686 * to extract the more random high bits from the multiplicative
3687 * hash function before the remainder is taken.
3689 static inline unsigned long wait_table_bits(unsigned long size)
3694 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3697 * Check if a pageblock contains reserved pages
3699 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3703 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3704 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3711 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3712 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3713 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3714 * higher will lead to a bigger reserve which will get freed as contiguous
3715 * blocks as reclaim kicks in
3717 static void setup_zone_migrate_reserve(struct zone *zone)
3719 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3721 unsigned long block_migratetype;
3725 * Get the start pfn, end pfn and the number of blocks to reserve
3726 * We have to be careful to be aligned to pageblock_nr_pages to
3727 * make sure that we always check pfn_valid for the first page in
3730 start_pfn = zone->zone_start_pfn;
3731 end_pfn = start_pfn + zone->spanned_pages;
3732 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3733 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3737 * Reserve blocks are generally in place to help high-order atomic
3738 * allocations that are short-lived. A min_free_kbytes value that
3739 * would result in more than 2 reserve blocks for atomic allocations
3740 * is assumed to be in place to help anti-fragmentation for the
3741 * future allocation of hugepages at runtime.
3743 reserve = min(2, reserve);
3745 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3746 if (!pfn_valid(pfn))
3748 page = pfn_to_page(pfn);
3750 /* Watch out for overlapping nodes */
3751 if (page_to_nid(page) != zone_to_nid(zone))
3754 block_migratetype = get_pageblock_migratetype(page);
3756 /* Only test what is necessary when the reserves are not met */
3759 * Blocks with reserved pages will never free, skip
3762 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3763 if (pageblock_is_reserved(pfn, block_end_pfn))
3766 /* If this block is reserved, account for it */
3767 if (block_migratetype == MIGRATE_RESERVE) {
3772 /* Suitable for reserving if this block is movable */
3773 if (block_migratetype == MIGRATE_MOVABLE) {
3774 set_pageblock_migratetype(page,
3776 move_freepages_block(zone, page,
3784 * If the reserve is met and this is a previous reserved block,
3787 if (block_migratetype == MIGRATE_RESERVE) {
3788 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3789 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3795 * Initially all pages are reserved - free ones are freed
3796 * up by free_all_bootmem() once the early boot process is
3797 * done. Non-atomic initialization, single-pass.
3799 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3800 unsigned long start_pfn, enum memmap_context context)
3803 unsigned long end_pfn = start_pfn + size;
3807 if (highest_memmap_pfn < end_pfn - 1)
3808 highest_memmap_pfn = end_pfn - 1;
3810 z = &NODE_DATA(nid)->node_zones[zone];
3811 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3813 * There can be holes in boot-time mem_map[]s
3814 * handed to this function. They do not
3815 * exist on hotplugged memory.
3817 if (context == MEMMAP_EARLY) {
3818 if (!early_pfn_valid(pfn))
3820 if (!early_pfn_in_nid(pfn, nid))
3823 page = pfn_to_page(pfn);
3824 set_page_links(page, zone, nid, pfn);
3825 mminit_verify_page_links(page, zone, nid, pfn);
3826 init_page_count(page);
3827 reset_page_mapcount(page);
3828 SetPageReserved(page);
3830 * Mark the block movable so that blocks are reserved for
3831 * movable at startup. This will force kernel allocations
3832 * to reserve their blocks rather than leaking throughout
3833 * the address space during boot when many long-lived
3834 * kernel allocations are made. Later some blocks near
3835 * the start are marked MIGRATE_RESERVE by
3836 * setup_zone_migrate_reserve()
3838 * bitmap is created for zone's valid pfn range. but memmap
3839 * can be created for invalid pages (for alignment)
3840 * check here not to call set_pageblock_migratetype() against
3843 if ((z->zone_start_pfn <= pfn)
3844 && (pfn < z->zone_start_pfn + z->spanned_pages)
3845 && !(pfn & (pageblock_nr_pages - 1)))
3846 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3848 INIT_LIST_HEAD(&page->lru);
3849 #ifdef WANT_PAGE_VIRTUAL
3850 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3851 if (!is_highmem_idx(zone))
3852 set_page_address(page, __va(pfn << PAGE_SHIFT));
3857 static void __meminit zone_init_free_lists(struct zone *zone)
3860 for_each_migratetype_order(order, t) {
3861 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3862 zone->free_area[order].nr_free = 0;
3866 #ifndef __HAVE_ARCH_MEMMAP_INIT
3867 #define memmap_init(size, nid, zone, start_pfn) \
3868 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3871 static int __meminit zone_batchsize(struct zone *zone)
3877 * The per-cpu-pages pools are set to around 1000th of the
3878 * size of the zone. But no more than 1/2 of a meg.
3880 * OK, so we don't know how big the cache is. So guess.
3882 batch = zone->present_pages / 1024;
3883 if (batch * PAGE_SIZE > 512 * 1024)
3884 batch = (512 * 1024) / PAGE_SIZE;
3885 batch /= 4; /* We effectively *= 4 below */
3890 * Clamp the batch to a 2^n - 1 value. Having a power
3891 * of 2 value was found to be more likely to have
3892 * suboptimal cache aliasing properties in some cases.
3894 * For example if 2 tasks are alternately allocating
3895 * batches of pages, one task can end up with a lot
3896 * of pages of one half of the possible page colors
3897 * and the other with pages of the other colors.
3899 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3904 /* The deferral and batching of frees should be suppressed under NOMMU
3907 * The problem is that NOMMU needs to be able to allocate large chunks
3908 * of contiguous memory as there's no hardware page translation to
3909 * assemble apparent contiguous memory from discontiguous pages.
3911 * Queueing large contiguous runs of pages for batching, however,
3912 * causes the pages to actually be freed in smaller chunks. As there
3913 * can be a significant delay between the individual batches being
3914 * recycled, this leads to the once large chunks of space being
3915 * fragmented and becoming unavailable for high-order allocations.
3921 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3923 struct per_cpu_pages *pcp;
3926 memset(p, 0, sizeof(*p));
3930 pcp->high = 6 * batch;
3931 pcp->batch = max(1UL, 1 * batch);
3932 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3933 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3937 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3938 * to the value high for the pageset p.
3941 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3944 struct per_cpu_pages *pcp;
3948 pcp->batch = max(1UL, high/4);
3949 if ((high/4) > (PAGE_SHIFT * 8))
3950 pcp->batch = PAGE_SHIFT * 8;
3953 static void __meminit setup_zone_pageset(struct zone *zone)
3957 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3959 for_each_possible_cpu(cpu) {
3960 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3962 setup_pageset(pcp, zone_batchsize(zone));
3964 if (percpu_pagelist_fraction)
3965 setup_pagelist_highmark(pcp,
3966 (zone->present_pages /
3967 percpu_pagelist_fraction));
3972 * Allocate per cpu pagesets and initialize them.
3973 * Before this call only boot pagesets were available.
3975 void __init setup_per_cpu_pageset(void)
3979 for_each_populated_zone(zone)
3980 setup_zone_pageset(zone);
3983 static noinline __init_refok
3984 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3987 struct pglist_data *pgdat = zone->zone_pgdat;
3991 * The per-page waitqueue mechanism uses hashed waitqueues
3994 zone->wait_table_hash_nr_entries =
3995 wait_table_hash_nr_entries(zone_size_pages);
3996 zone->wait_table_bits =
3997 wait_table_bits(zone->wait_table_hash_nr_entries);
3998 alloc_size = zone->wait_table_hash_nr_entries
3999 * sizeof(wait_queue_head_t);
4001 if (!slab_is_available()) {
4002 zone->wait_table = (wait_queue_head_t *)
4003 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4006 * This case means that a zone whose size was 0 gets new memory
4007 * via memory hot-add.
4008 * But it may be the case that a new node was hot-added. In
4009 * this case vmalloc() will not be able to use this new node's
4010 * memory - this wait_table must be initialized to use this new
4011 * node itself as well.
4012 * To use this new node's memory, further consideration will be
4015 zone->wait_table = vmalloc(alloc_size);
4017 if (!zone->wait_table)
4020 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4021 init_waitqueue_head(zone->wait_table + i);
4026 static __meminit void zone_pcp_init(struct zone *zone)
4029 * per cpu subsystem is not up at this point. The following code
4030 * relies on the ability of the linker to provide the
4031 * offset of a (static) per cpu variable into the per cpu area.
4033 zone->pageset = &boot_pageset;
4035 if (zone->present_pages)
4036 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4037 zone->name, zone->present_pages,
4038 zone_batchsize(zone));
4041 int __meminit init_currently_empty_zone(struct zone *zone,
4042 unsigned long zone_start_pfn,
4044 enum memmap_context context)
4046 struct pglist_data *pgdat = zone->zone_pgdat;
4048 ret = zone_wait_table_init(zone, size);
4051 pgdat->nr_zones = zone_idx(zone) + 1;
4053 zone->zone_start_pfn = zone_start_pfn;
4055 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4056 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4058 (unsigned long)zone_idx(zone),
4059 zone_start_pfn, (zone_start_pfn + size));
4061 zone_init_free_lists(zone);
4066 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4067 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4069 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4070 * Architectures may implement their own version but if add_active_range()
4071 * was used and there are no special requirements, this is a convenient
4074 int __meminit __early_pfn_to_nid(unsigned long pfn)
4076 unsigned long start_pfn, end_pfn;
4079 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4080 if (start_pfn <= pfn && pfn < end_pfn)
4082 /* This is a memory hole */
4085 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4087 int __meminit early_pfn_to_nid(unsigned long pfn)
4091 nid = __early_pfn_to_nid(pfn);
4094 /* just returns 0 */
4098 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4099 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4103 nid = __early_pfn_to_nid(pfn);
4104 if (nid >= 0 && nid != node)
4111 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4112 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4113 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4115 * If an architecture guarantees that all ranges registered with
4116 * add_active_ranges() contain no holes and may be freed, this
4117 * this function may be used instead of calling free_bootmem() manually.
4119 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4121 unsigned long start_pfn, end_pfn;
4124 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4125 start_pfn = min(start_pfn, max_low_pfn);
4126 end_pfn = min(end_pfn, max_low_pfn);
4128 if (start_pfn < end_pfn)
4129 free_bootmem_node(NODE_DATA(this_nid),
4130 PFN_PHYS(start_pfn),
4131 (end_pfn - start_pfn) << PAGE_SHIFT);
4136 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4137 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4139 * If an architecture guarantees that all ranges registered with
4140 * add_active_ranges() contain no holes and may be freed, this
4141 * function may be used instead of calling memory_present() manually.
4143 void __init sparse_memory_present_with_active_regions(int nid)
4145 unsigned long start_pfn, end_pfn;
4148 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4149 memory_present(this_nid, start_pfn, end_pfn);
4153 * get_pfn_range_for_nid - Return the start and end page frames for a node
4154 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4155 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4156 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4158 * It returns the start and end page frame of a node based on information
4159 * provided by an arch calling add_active_range(). If called for a node
4160 * with no available memory, a warning is printed and the start and end
4163 void __meminit get_pfn_range_for_nid(unsigned int nid,
4164 unsigned long *start_pfn, unsigned long *end_pfn)
4166 unsigned long this_start_pfn, this_end_pfn;
4172 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4173 *start_pfn = min(*start_pfn, this_start_pfn);
4174 *end_pfn = max(*end_pfn, this_end_pfn);
4177 if (*start_pfn == -1UL)
4182 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4183 * assumption is made that zones within a node are ordered in monotonic
4184 * increasing memory addresses so that the "highest" populated zone is used
4186 static void __init find_usable_zone_for_movable(void)
4189 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4190 if (zone_index == ZONE_MOVABLE)
4193 if (arch_zone_highest_possible_pfn[zone_index] >
4194 arch_zone_lowest_possible_pfn[zone_index])
4198 VM_BUG_ON(zone_index == -1);
4199 movable_zone = zone_index;
4203 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4204 * because it is sized independent of architecture. Unlike the other zones,
4205 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4206 * in each node depending on the size of each node and how evenly kernelcore
4207 * is distributed. This helper function adjusts the zone ranges
4208 * provided by the architecture for a given node by using the end of the
4209 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4210 * zones within a node are in order of monotonic increases memory addresses
4212 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4213 unsigned long zone_type,
4214 unsigned long node_start_pfn,
4215 unsigned long node_end_pfn,
4216 unsigned long *zone_start_pfn,
4217 unsigned long *zone_end_pfn)
4219 /* Only adjust if ZONE_MOVABLE is on this node */
4220 if (zone_movable_pfn[nid]) {
4221 /* Size ZONE_MOVABLE */
4222 if (zone_type == ZONE_MOVABLE) {
4223 *zone_start_pfn = zone_movable_pfn[nid];
4224 *zone_end_pfn = min(node_end_pfn,
4225 arch_zone_highest_possible_pfn[movable_zone]);
4227 /* Adjust for ZONE_MOVABLE starting within this range */
4228 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4229 *zone_end_pfn > zone_movable_pfn[nid]) {
4230 *zone_end_pfn = zone_movable_pfn[nid];
4232 /* Check if this whole range is within ZONE_MOVABLE */
4233 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4234 *zone_start_pfn = *zone_end_pfn;
4239 * Return the number of pages a zone spans in a node, including holes
4240 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4242 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4243 unsigned long zone_type,
4244 unsigned long *ignored)
4246 unsigned long node_start_pfn, node_end_pfn;
4247 unsigned long zone_start_pfn, zone_end_pfn;
4249 /* Get the start and end of the node and zone */
4250 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4251 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4252 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4253 adjust_zone_range_for_zone_movable(nid, zone_type,
4254 node_start_pfn, node_end_pfn,
4255 &zone_start_pfn, &zone_end_pfn);
4257 /* Check that this node has pages within the zone's required range */
4258 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4261 /* Move the zone boundaries inside the node if necessary */
4262 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4263 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4265 /* Return the spanned pages */
4266 return zone_end_pfn - zone_start_pfn;
4270 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4271 * then all holes in the requested range will be accounted for.
4273 unsigned long __meminit __absent_pages_in_range(int nid,
4274 unsigned long range_start_pfn,
4275 unsigned long range_end_pfn)
4277 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4278 unsigned long start_pfn, end_pfn;
4281 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4282 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4283 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4284 nr_absent -= end_pfn - start_pfn;
4290 * absent_pages_in_range - Return number of page frames in holes within a range
4291 * @start_pfn: The start PFN to start searching for holes
4292 * @end_pfn: The end PFN to stop searching for holes
4294 * It returns the number of pages frames in memory holes within a range.
4296 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4297 unsigned long end_pfn)
4299 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4302 /* Return the number of page frames in holes in a zone on a node */
4303 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4304 unsigned long zone_type,
4305 unsigned long *ignored)
4307 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4308 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4309 unsigned long node_start_pfn, node_end_pfn;
4310 unsigned long zone_start_pfn, zone_end_pfn;
4312 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4313 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4314 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4316 adjust_zone_range_for_zone_movable(nid, zone_type,
4317 node_start_pfn, node_end_pfn,
4318 &zone_start_pfn, &zone_end_pfn);
4319 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4322 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4323 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4324 unsigned long zone_type,
4325 unsigned long *zones_size)
4327 return zones_size[zone_type];
4330 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4331 unsigned long zone_type,
4332 unsigned long *zholes_size)
4337 return zholes_size[zone_type];
4340 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4342 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4343 unsigned long *zones_size, unsigned long *zholes_size)
4345 unsigned long realtotalpages, totalpages = 0;
4348 for (i = 0; i < MAX_NR_ZONES; i++)
4349 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4351 pgdat->node_spanned_pages = totalpages;
4353 realtotalpages = totalpages;
4354 for (i = 0; i < MAX_NR_ZONES; i++)
4356 zone_absent_pages_in_node(pgdat->node_id, i,
4358 pgdat->node_present_pages = realtotalpages;
4359 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4363 #ifndef CONFIG_SPARSEMEM
4365 * Calculate the size of the zone->blockflags rounded to an unsigned long
4366 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4367 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4368 * round what is now in bits to nearest long in bits, then return it in
4371 static unsigned long __init usemap_size(unsigned long zonesize)
4373 unsigned long usemapsize;
4375 usemapsize = roundup(zonesize, pageblock_nr_pages);
4376 usemapsize = usemapsize >> pageblock_order;
4377 usemapsize *= NR_PAGEBLOCK_BITS;
4378 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4380 return usemapsize / 8;
4383 static void __init setup_usemap(struct pglist_data *pgdat,
4384 struct zone *zone, unsigned long zonesize)
4386 unsigned long usemapsize = usemap_size(zonesize);
4387 zone->pageblock_flags = NULL;
4389 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4393 static inline void setup_usemap(struct pglist_data *pgdat,
4394 struct zone *zone, unsigned long zonesize) {}
4395 #endif /* CONFIG_SPARSEMEM */
4397 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4399 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4400 void __init set_pageblock_order(void)
4404 /* Check that pageblock_nr_pages has not already been setup */
4405 if (pageblock_order)
4408 if (HPAGE_SHIFT > PAGE_SHIFT)
4409 order = HUGETLB_PAGE_ORDER;
4411 order = MAX_ORDER - 1;
4414 * Assume the largest contiguous order of interest is a huge page.
4415 * This value may be variable depending on boot parameters on IA64 and
4418 pageblock_order = order;
4420 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4423 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4424 * is unused as pageblock_order is set at compile-time. See
4425 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4428 void __init set_pageblock_order(void)
4432 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4435 * Set up the zone data structures:
4436 * - mark all pages reserved
4437 * - mark all memory queues empty
4438 * - clear the memory bitmaps
4440 * NOTE: pgdat should get zeroed by caller.
4442 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4443 unsigned long *zones_size, unsigned long *zholes_size)
4446 int nid = pgdat->node_id;
4447 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4450 pgdat_resize_init(pgdat);
4451 init_waitqueue_head(&pgdat->kswapd_wait);
4452 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4453 pgdat_page_cgroup_init(pgdat);
4455 for (j = 0; j < MAX_NR_ZONES; j++) {
4456 struct zone *zone = pgdat->node_zones + j;
4457 unsigned long size, realsize, memmap_pages;
4459 size = zone_spanned_pages_in_node(nid, j, zones_size);
4460 realsize = size - zone_absent_pages_in_node(nid, j,
4464 * Adjust realsize so that it accounts for how much memory
4465 * is used by this zone for memmap. This affects the watermark
4466 * and per-cpu initialisations
4469 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4470 if (realsize >= memmap_pages) {
4471 realsize -= memmap_pages;
4474 " %s zone: %lu pages used for memmap\n",
4475 zone_names[j], memmap_pages);
4478 " %s zone: %lu pages exceeds realsize %lu\n",
4479 zone_names[j], memmap_pages, realsize);
4481 /* Account for reserved pages */
4482 if (j == 0 && realsize > dma_reserve) {
4483 realsize -= dma_reserve;
4484 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4485 zone_names[0], dma_reserve);
4488 if (!is_highmem_idx(j))
4489 nr_kernel_pages += realsize;
4490 nr_all_pages += realsize;
4492 zone->spanned_pages = size;
4493 zone->present_pages = realsize;
4496 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4498 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4500 zone->name = zone_names[j];
4501 spin_lock_init(&zone->lock);
4502 spin_lock_init(&zone->lru_lock);
4503 zone_seqlock_init(zone);
4504 zone->zone_pgdat = pgdat;
4506 zone_pcp_init(zone);
4507 lruvec_init(&zone->lruvec, zone);
4511 set_pageblock_order();
4512 setup_usemap(pgdat, zone, size);
4513 ret = init_currently_empty_zone(zone, zone_start_pfn,
4514 size, MEMMAP_EARLY);
4516 memmap_init(size, nid, j, zone_start_pfn);
4517 zone_start_pfn += size;
4521 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4523 /* Skip empty nodes */
4524 if (!pgdat->node_spanned_pages)
4527 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4528 /* ia64 gets its own node_mem_map, before this, without bootmem */
4529 if (!pgdat->node_mem_map) {
4530 unsigned long size, start, end;
4534 * The zone's endpoints aren't required to be MAX_ORDER
4535 * aligned but the node_mem_map endpoints must be in order
4536 * for the buddy allocator to function correctly.
4538 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4539 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4540 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4541 size = (end - start) * sizeof(struct page);
4542 map = alloc_remap(pgdat->node_id, size);
4544 map = alloc_bootmem_node_nopanic(pgdat, size);
4545 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4547 #ifndef CONFIG_NEED_MULTIPLE_NODES
4549 * With no DISCONTIG, the global mem_map is just set as node 0's
4551 if (pgdat == NODE_DATA(0)) {
4552 mem_map = NODE_DATA(0)->node_mem_map;
4553 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4554 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4555 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4556 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4559 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4562 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4563 unsigned long node_start_pfn, unsigned long *zholes_size)
4565 pg_data_t *pgdat = NODE_DATA(nid);
4567 /* pg_data_t should be reset to zero when it's allocated */
4568 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4570 pgdat->node_id = nid;
4571 pgdat->node_start_pfn = node_start_pfn;
4572 init_zone_allows_reclaim(nid);
4573 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4575 alloc_node_mem_map(pgdat);
4576 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4577 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4578 nid, (unsigned long)pgdat,
4579 (unsigned long)pgdat->node_mem_map);
4582 free_area_init_core(pgdat, zones_size, zholes_size);
4585 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4587 #if MAX_NUMNODES > 1
4589 * Figure out the number of possible node ids.
4591 static void __init setup_nr_node_ids(void)
4594 unsigned int highest = 0;
4596 for_each_node_mask(node, node_possible_map)
4598 nr_node_ids = highest + 1;
4601 static inline void setup_nr_node_ids(void)
4607 * node_map_pfn_alignment - determine the maximum internode alignment
4609 * This function should be called after node map is populated and sorted.
4610 * It calculates the maximum power of two alignment which can distinguish
4613 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4614 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4615 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4616 * shifted, 1GiB is enough and this function will indicate so.
4618 * This is used to test whether pfn -> nid mapping of the chosen memory
4619 * model has fine enough granularity to avoid incorrect mapping for the
4620 * populated node map.
4622 * Returns the determined alignment in pfn's. 0 if there is no alignment
4623 * requirement (single node).
4625 unsigned long __init node_map_pfn_alignment(void)
4627 unsigned long accl_mask = 0, last_end = 0;
4628 unsigned long start, end, mask;
4632 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4633 if (!start || last_nid < 0 || last_nid == nid) {
4640 * Start with a mask granular enough to pin-point to the
4641 * start pfn and tick off bits one-by-one until it becomes
4642 * too coarse to separate the current node from the last.
4644 mask = ~((1 << __ffs(start)) - 1);
4645 while (mask && last_end <= (start & (mask << 1)))
4648 /* accumulate all internode masks */
4652 /* convert mask to number of pages */
4653 return ~accl_mask + 1;
4656 /* Find the lowest pfn for a node */
4657 static unsigned long __init find_min_pfn_for_node(int nid)
4659 unsigned long min_pfn = ULONG_MAX;
4660 unsigned long start_pfn;
4663 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4664 min_pfn = min(min_pfn, start_pfn);
4666 if (min_pfn == ULONG_MAX) {
4668 "Could not find start_pfn for node %d\n", nid);
4676 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4678 * It returns the minimum PFN based on information provided via
4679 * add_active_range().
4681 unsigned long __init find_min_pfn_with_active_regions(void)
4683 return find_min_pfn_for_node(MAX_NUMNODES);
4687 * early_calculate_totalpages()
4688 * Sum pages in active regions for movable zone.
4689 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4691 static unsigned long __init early_calculate_totalpages(void)
4693 unsigned long totalpages = 0;
4694 unsigned long start_pfn, end_pfn;
4697 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4698 unsigned long pages = end_pfn - start_pfn;
4700 totalpages += pages;
4702 node_set_state(nid, N_HIGH_MEMORY);
4708 * Find the PFN the Movable zone begins in each node. Kernel memory
4709 * is spread evenly between nodes as long as the nodes have enough
4710 * memory. When they don't, some nodes will have more kernelcore than
4713 static void __init find_zone_movable_pfns_for_nodes(void)
4716 unsigned long usable_startpfn;
4717 unsigned long kernelcore_node, kernelcore_remaining;
4718 /* save the state before borrow the nodemask */
4719 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4720 unsigned long totalpages = early_calculate_totalpages();
4721 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4724 * If movablecore was specified, calculate what size of
4725 * kernelcore that corresponds so that memory usable for
4726 * any allocation type is evenly spread. If both kernelcore
4727 * and movablecore are specified, then the value of kernelcore
4728 * will be used for required_kernelcore if it's greater than
4729 * what movablecore would have allowed.
4731 if (required_movablecore) {
4732 unsigned long corepages;
4735 * Round-up so that ZONE_MOVABLE is at least as large as what
4736 * was requested by the user
4738 required_movablecore =
4739 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4740 corepages = totalpages - required_movablecore;
4742 required_kernelcore = max(required_kernelcore, corepages);
4745 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4746 if (!required_kernelcore)
4749 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4750 find_usable_zone_for_movable();
4751 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4754 /* Spread kernelcore memory as evenly as possible throughout nodes */
4755 kernelcore_node = required_kernelcore / usable_nodes;
4756 for_each_node_state(nid, N_HIGH_MEMORY) {
4757 unsigned long start_pfn, end_pfn;
4760 * Recalculate kernelcore_node if the division per node
4761 * now exceeds what is necessary to satisfy the requested
4762 * amount of memory for the kernel
4764 if (required_kernelcore < kernelcore_node)
4765 kernelcore_node = required_kernelcore / usable_nodes;
4768 * As the map is walked, we track how much memory is usable
4769 * by the kernel using kernelcore_remaining. When it is
4770 * 0, the rest of the node is usable by ZONE_MOVABLE
4772 kernelcore_remaining = kernelcore_node;
4774 /* Go through each range of PFNs within this node */
4775 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4776 unsigned long size_pages;
4778 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4779 if (start_pfn >= end_pfn)
4782 /* Account for what is only usable for kernelcore */
4783 if (start_pfn < usable_startpfn) {
4784 unsigned long kernel_pages;
4785 kernel_pages = min(end_pfn, usable_startpfn)
4788 kernelcore_remaining -= min(kernel_pages,
4789 kernelcore_remaining);
4790 required_kernelcore -= min(kernel_pages,
4791 required_kernelcore);
4793 /* Continue if range is now fully accounted */
4794 if (end_pfn <= usable_startpfn) {
4797 * Push zone_movable_pfn to the end so
4798 * that if we have to rebalance
4799 * kernelcore across nodes, we will
4800 * not double account here
4802 zone_movable_pfn[nid] = end_pfn;
4805 start_pfn = usable_startpfn;
4809 * The usable PFN range for ZONE_MOVABLE is from
4810 * start_pfn->end_pfn. Calculate size_pages as the
4811 * number of pages used as kernelcore
4813 size_pages = end_pfn - start_pfn;
4814 if (size_pages > kernelcore_remaining)
4815 size_pages = kernelcore_remaining;
4816 zone_movable_pfn[nid] = start_pfn + size_pages;
4819 * Some kernelcore has been met, update counts and
4820 * break if the kernelcore for this node has been
4823 required_kernelcore -= min(required_kernelcore,
4825 kernelcore_remaining -= size_pages;
4826 if (!kernelcore_remaining)
4832 * If there is still required_kernelcore, we do another pass with one
4833 * less node in the count. This will push zone_movable_pfn[nid] further
4834 * along on the nodes that still have memory until kernelcore is
4838 if (usable_nodes && required_kernelcore > usable_nodes)
4841 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4842 for (nid = 0; nid < MAX_NUMNODES; nid++)
4843 zone_movable_pfn[nid] =
4844 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4847 /* restore the node_state */
4848 node_states[N_HIGH_MEMORY] = saved_node_state;
4851 /* Any regular memory on that node ? */
4852 static void __init check_for_regular_memory(pg_data_t *pgdat)
4854 #ifdef CONFIG_HIGHMEM
4855 enum zone_type zone_type;
4857 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4858 struct zone *zone = &pgdat->node_zones[zone_type];
4859 if (zone->present_pages) {
4860 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4868 * free_area_init_nodes - Initialise all pg_data_t and zone data
4869 * @max_zone_pfn: an array of max PFNs for each zone
4871 * This will call free_area_init_node() for each active node in the system.
4872 * Using the page ranges provided by add_active_range(), the size of each
4873 * zone in each node and their holes is calculated. If the maximum PFN
4874 * between two adjacent zones match, it is assumed that the zone is empty.
4875 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4876 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4877 * starts where the previous one ended. For example, ZONE_DMA32 starts
4878 * at arch_max_dma_pfn.
4880 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4882 unsigned long start_pfn, end_pfn;
4885 /* Record where the zone boundaries are */
4886 memset(arch_zone_lowest_possible_pfn, 0,
4887 sizeof(arch_zone_lowest_possible_pfn));
4888 memset(arch_zone_highest_possible_pfn, 0,
4889 sizeof(arch_zone_highest_possible_pfn));
4890 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4891 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4892 for (i = 1; i < MAX_NR_ZONES; i++) {
4893 if (i == ZONE_MOVABLE)
4895 arch_zone_lowest_possible_pfn[i] =
4896 arch_zone_highest_possible_pfn[i-1];
4897 arch_zone_highest_possible_pfn[i] =
4898 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4900 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4901 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4903 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4904 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4905 find_zone_movable_pfns_for_nodes();
4907 /* Print out the zone ranges */
4908 printk("Zone ranges:\n");
4909 for (i = 0; i < MAX_NR_ZONES; i++) {
4910 if (i == ZONE_MOVABLE)
4912 printk(KERN_CONT " %-8s ", zone_names[i]);
4913 if (arch_zone_lowest_possible_pfn[i] ==
4914 arch_zone_highest_possible_pfn[i])
4915 printk(KERN_CONT "empty\n");
4917 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4918 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4919 (arch_zone_highest_possible_pfn[i]
4920 << PAGE_SHIFT) - 1);
4923 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4924 printk("Movable zone start for each node\n");
4925 for (i = 0; i < MAX_NUMNODES; i++) {
4926 if (zone_movable_pfn[i])
4927 printk(" Node %d: %#010lx\n", i,
4928 zone_movable_pfn[i] << PAGE_SHIFT);
4931 /* Print out the early node map */
4932 printk("Early memory node ranges\n");
4933 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4934 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
4935 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
4937 /* Initialise every node */
4938 mminit_verify_pageflags_layout();
4939 setup_nr_node_ids();
4940 for_each_online_node(nid) {
4941 pg_data_t *pgdat = NODE_DATA(nid);
4942 free_area_init_node(nid, NULL,
4943 find_min_pfn_for_node(nid), NULL);
4945 /* Any memory on that node */
4946 if (pgdat->node_present_pages)
4947 node_set_state(nid, N_HIGH_MEMORY);
4948 check_for_regular_memory(pgdat);
4952 static int __init cmdline_parse_core(char *p, unsigned long *core)
4954 unsigned long long coremem;
4958 coremem = memparse(p, &p);
4959 *core = coremem >> PAGE_SHIFT;
4961 /* Paranoid check that UL is enough for the coremem value */
4962 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4968 * kernelcore=size sets the amount of memory for use for allocations that
4969 * cannot be reclaimed or migrated.
4971 static int __init cmdline_parse_kernelcore(char *p)
4973 return cmdline_parse_core(p, &required_kernelcore);
4977 * movablecore=size sets the amount of memory for use for allocations that
4978 * can be reclaimed or migrated.
4980 static int __init cmdline_parse_movablecore(char *p)
4982 return cmdline_parse_core(p, &required_movablecore);
4985 early_param("kernelcore", cmdline_parse_kernelcore);
4986 early_param("movablecore", cmdline_parse_movablecore);
4988 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4991 * set_dma_reserve - set the specified number of pages reserved in the first zone
4992 * @new_dma_reserve: The number of pages to mark reserved
4994 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4995 * In the DMA zone, a significant percentage may be consumed by kernel image
4996 * and other unfreeable allocations which can skew the watermarks badly. This
4997 * function may optionally be used to account for unfreeable pages in the
4998 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4999 * smaller per-cpu batchsize.
5001 void __init set_dma_reserve(unsigned long new_dma_reserve)
5003 dma_reserve = new_dma_reserve;
5006 void __init free_area_init(unsigned long *zones_size)
5008 free_area_init_node(0, zones_size,
5009 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5012 static int page_alloc_cpu_notify(struct notifier_block *self,
5013 unsigned long action, void *hcpu)
5015 int cpu = (unsigned long)hcpu;
5017 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5018 lru_add_drain_cpu(cpu);
5022 * Spill the event counters of the dead processor
5023 * into the current processors event counters.
5024 * This artificially elevates the count of the current
5027 vm_events_fold_cpu(cpu);
5030 * Zero the differential counters of the dead processor
5031 * so that the vm statistics are consistent.
5033 * This is only okay since the processor is dead and cannot
5034 * race with what we are doing.
5036 refresh_cpu_vm_stats(cpu);
5041 void __init page_alloc_init(void)
5043 hotcpu_notifier(page_alloc_cpu_notify, 0);
5047 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5048 * or min_free_kbytes changes.
5050 static void calculate_totalreserve_pages(void)
5052 struct pglist_data *pgdat;
5053 unsigned long reserve_pages = 0;
5054 enum zone_type i, j;
5056 for_each_online_pgdat(pgdat) {
5057 for (i = 0; i < MAX_NR_ZONES; i++) {
5058 struct zone *zone = pgdat->node_zones + i;
5059 unsigned long max = 0;
5061 /* Find valid and maximum lowmem_reserve in the zone */
5062 for (j = i; j < MAX_NR_ZONES; j++) {
5063 if (zone->lowmem_reserve[j] > max)
5064 max = zone->lowmem_reserve[j];
5067 /* we treat the high watermark as reserved pages. */
5068 max += high_wmark_pages(zone);
5070 if (max > zone->present_pages)
5071 max = zone->present_pages;
5072 reserve_pages += max;
5074 * Lowmem reserves are not available to
5075 * GFP_HIGHUSER page cache allocations and
5076 * kswapd tries to balance zones to their high
5077 * watermark. As a result, neither should be
5078 * regarded as dirtyable memory, to prevent a
5079 * situation where reclaim has to clean pages
5080 * in order to balance the zones.
5082 zone->dirty_balance_reserve = max;
5085 dirty_balance_reserve = reserve_pages;
5086 totalreserve_pages = reserve_pages;
5090 * setup_per_zone_lowmem_reserve - called whenever
5091 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5092 * has a correct pages reserved value, so an adequate number of
5093 * pages are left in the zone after a successful __alloc_pages().
5095 static void setup_per_zone_lowmem_reserve(void)
5097 struct pglist_data *pgdat;
5098 enum zone_type j, idx;
5100 for_each_online_pgdat(pgdat) {
5101 for (j = 0; j < MAX_NR_ZONES; j++) {
5102 struct zone *zone = pgdat->node_zones + j;
5103 unsigned long present_pages = zone->present_pages;
5105 zone->lowmem_reserve[j] = 0;
5109 struct zone *lower_zone;
5113 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5114 sysctl_lowmem_reserve_ratio[idx] = 1;
5116 lower_zone = pgdat->node_zones + idx;
5117 lower_zone->lowmem_reserve[j] = present_pages /
5118 sysctl_lowmem_reserve_ratio[idx];
5119 present_pages += lower_zone->present_pages;
5124 /* update totalreserve_pages */
5125 calculate_totalreserve_pages();
5128 static void __setup_per_zone_wmarks(void)
5130 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5131 unsigned long lowmem_pages = 0;
5133 unsigned long flags;
5135 /* Calculate total number of !ZONE_HIGHMEM pages */
5136 for_each_zone(zone) {
5137 if (!is_highmem(zone))
5138 lowmem_pages += zone->present_pages;
5141 for_each_zone(zone) {
5144 spin_lock_irqsave(&zone->lock, flags);
5145 tmp = (u64)pages_min * zone->present_pages;
5146 do_div(tmp, lowmem_pages);
5147 if (is_highmem(zone)) {
5149 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5150 * need highmem pages, so cap pages_min to a small
5153 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5154 * deltas controls asynch page reclaim, and so should
5155 * not be capped for highmem.
5159 min_pages = zone->present_pages / 1024;
5160 if (min_pages < SWAP_CLUSTER_MAX)
5161 min_pages = SWAP_CLUSTER_MAX;
5162 if (min_pages > 128)
5164 zone->watermark[WMARK_MIN] = min_pages;
5167 * If it's a lowmem zone, reserve a number of pages
5168 * proportionate to the zone's size.
5170 zone->watermark[WMARK_MIN] = tmp;
5173 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5174 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5176 zone->watermark[WMARK_MIN] += cma_wmark_pages(zone);
5177 zone->watermark[WMARK_LOW] += cma_wmark_pages(zone);
5178 zone->watermark[WMARK_HIGH] += cma_wmark_pages(zone);
5180 setup_zone_migrate_reserve(zone);
5181 spin_unlock_irqrestore(&zone->lock, flags);
5184 /* update totalreserve_pages */
5185 calculate_totalreserve_pages();
5189 * setup_per_zone_wmarks - called when min_free_kbytes changes
5190 * or when memory is hot-{added|removed}
5192 * Ensures that the watermark[min,low,high] values for each zone are set
5193 * correctly with respect to min_free_kbytes.
5195 void setup_per_zone_wmarks(void)
5197 mutex_lock(&zonelists_mutex);
5198 __setup_per_zone_wmarks();
5199 mutex_unlock(&zonelists_mutex);
5203 * The inactive anon list should be small enough that the VM never has to
5204 * do too much work, but large enough that each inactive page has a chance
5205 * to be referenced again before it is swapped out.
5207 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5208 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5209 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5210 * the anonymous pages are kept on the inactive list.
5213 * memory ratio inactive anon
5214 * -------------------------------------
5223 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5225 unsigned int gb, ratio;
5227 /* Zone size in gigabytes */
5228 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5230 ratio = int_sqrt(10 * gb);
5234 zone->inactive_ratio = ratio;
5237 static void __meminit setup_per_zone_inactive_ratio(void)
5242 calculate_zone_inactive_ratio(zone);
5246 * Initialise min_free_kbytes.
5248 * For small machines we want it small (128k min). For large machines
5249 * we want it large (64MB max). But it is not linear, because network
5250 * bandwidth does not increase linearly with machine size. We use
5252 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5253 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5269 int __meminit init_per_zone_wmark_min(void)
5271 unsigned long lowmem_kbytes;
5273 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5275 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5276 if (min_free_kbytes < 128)
5277 min_free_kbytes = 128;
5278 if (min_free_kbytes > 65536)
5279 min_free_kbytes = 65536;
5280 setup_per_zone_wmarks();
5281 refresh_zone_stat_thresholds();
5282 setup_per_zone_lowmem_reserve();
5283 setup_per_zone_inactive_ratio();
5286 module_init(init_per_zone_wmark_min)
5289 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5290 * that we can call two helper functions whenever min_free_kbytes
5293 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5294 void __user *buffer, size_t *length, loff_t *ppos)
5296 proc_dointvec(table, write, buffer, length, ppos);
5298 setup_per_zone_wmarks();
5303 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5304 void __user *buffer, size_t *length, loff_t *ppos)
5309 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5314 zone->min_unmapped_pages = (zone->present_pages *
5315 sysctl_min_unmapped_ratio) / 100;
5319 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5320 void __user *buffer, size_t *length, loff_t *ppos)
5325 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5330 zone->min_slab_pages = (zone->present_pages *
5331 sysctl_min_slab_ratio) / 100;
5337 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5338 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5339 * whenever sysctl_lowmem_reserve_ratio changes.
5341 * The reserve ratio obviously has absolutely no relation with the
5342 * minimum watermarks. The lowmem reserve ratio can only make sense
5343 * if in function of the boot time zone sizes.
5345 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5346 void __user *buffer, size_t *length, loff_t *ppos)
5348 proc_dointvec_minmax(table, write, buffer, length, ppos);
5349 setup_per_zone_lowmem_reserve();
5354 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5355 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5356 * can have before it gets flushed back to buddy allocator.
5359 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5360 void __user *buffer, size_t *length, loff_t *ppos)
5366 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5367 if (!write || (ret < 0))
5369 for_each_populated_zone(zone) {
5370 for_each_possible_cpu(cpu) {
5372 high = zone->present_pages / percpu_pagelist_fraction;
5373 setup_pagelist_highmark(
5374 per_cpu_ptr(zone->pageset, cpu), high);
5380 int hashdist = HASHDIST_DEFAULT;
5383 static int __init set_hashdist(char *str)
5387 hashdist = simple_strtoul(str, &str, 0);
5390 __setup("hashdist=", set_hashdist);
5394 * allocate a large system hash table from bootmem
5395 * - it is assumed that the hash table must contain an exact power-of-2
5396 * quantity of entries
5397 * - limit is the number of hash buckets, not the total allocation size
5399 void *__init alloc_large_system_hash(const char *tablename,
5400 unsigned long bucketsize,
5401 unsigned long numentries,
5404 unsigned int *_hash_shift,
5405 unsigned int *_hash_mask,
5406 unsigned long low_limit,
5407 unsigned long high_limit)
5409 unsigned long long max = high_limit;
5410 unsigned long log2qty, size;
5413 /* allow the kernel cmdline to have a say */
5415 /* round applicable memory size up to nearest megabyte */
5416 numentries = nr_kernel_pages;
5417 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5418 numentries >>= 20 - PAGE_SHIFT;
5419 numentries <<= 20 - PAGE_SHIFT;
5421 /* limit to 1 bucket per 2^scale bytes of low memory */
5422 if (scale > PAGE_SHIFT)
5423 numentries >>= (scale - PAGE_SHIFT);
5425 numentries <<= (PAGE_SHIFT - scale);
5427 /* Make sure we've got at least a 0-order allocation.. */
5428 if (unlikely(flags & HASH_SMALL)) {
5429 /* Makes no sense without HASH_EARLY */
5430 WARN_ON(!(flags & HASH_EARLY));
5431 if (!(numentries >> *_hash_shift)) {
5432 numentries = 1UL << *_hash_shift;
5433 BUG_ON(!numentries);
5435 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5436 numentries = PAGE_SIZE / bucketsize;
5438 numentries = roundup_pow_of_two(numentries);
5440 /* limit allocation size to 1/16 total memory by default */
5442 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5443 do_div(max, bucketsize);
5445 max = min(max, 0x80000000ULL);
5447 if (numentries < low_limit)
5448 numentries = low_limit;
5449 if (numentries > max)
5452 log2qty = ilog2(numentries);
5455 size = bucketsize << log2qty;
5456 if (flags & HASH_EARLY)
5457 table = alloc_bootmem_nopanic(size);
5459 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5462 * If bucketsize is not a power-of-two, we may free
5463 * some pages at the end of hash table which
5464 * alloc_pages_exact() automatically does
5466 if (get_order(size) < MAX_ORDER) {
5467 table = alloc_pages_exact(size, GFP_ATOMIC);
5468 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5471 } while (!table && size > PAGE_SIZE && --log2qty);
5474 panic("Failed to allocate %s hash table\n", tablename);
5476 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5479 ilog2(size) - PAGE_SHIFT,
5483 *_hash_shift = log2qty;
5485 *_hash_mask = (1 << log2qty) - 1;
5490 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5491 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5494 #ifdef CONFIG_SPARSEMEM
5495 return __pfn_to_section(pfn)->pageblock_flags;
5497 return zone->pageblock_flags;
5498 #endif /* CONFIG_SPARSEMEM */
5501 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5503 #ifdef CONFIG_SPARSEMEM
5504 pfn &= (PAGES_PER_SECTION-1);
5505 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5507 pfn = pfn - zone->zone_start_pfn;
5508 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5509 #endif /* CONFIG_SPARSEMEM */
5513 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5514 * @page: The page within the block of interest
5515 * @start_bitidx: The first bit of interest to retrieve
5516 * @end_bitidx: The last bit of interest
5517 * returns pageblock_bits flags
5519 unsigned long get_pageblock_flags_group(struct page *page,
5520 int start_bitidx, int end_bitidx)
5523 unsigned long *bitmap;
5524 unsigned long pfn, bitidx;
5525 unsigned long flags = 0;
5526 unsigned long value = 1;
5528 zone = page_zone(page);
5529 pfn = page_to_pfn(page);
5530 bitmap = get_pageblock_bitmap(zone, pfn);
5531 bitidx = pfn_to_bitidx(zone, pfn);
5533 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5534 if (test_bit(bitidx + start_bitidx, bitmap))
5541 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5542 * @page: The page within the block of interest
5543 * @start_bitidx: The first bit of interest
5544 * @end_bitidx: The last bit of interest
5545 * @flags: The flags to set
5547 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5548 int start_bitidx, int end_bitidx)
5551 unsigned long *bitmap;
5552 unsigned long pfn, bitidx;
5553 unsigned long value = 1;
5555 zone = page_zone(page);
5556 pfn = page_to_pfn(page);
5557 bitmap = get_pageblock_bitmap(zone, pfn);
5558 bitidx = pfn_to_bitidx(zone, pfn);
5559 VM_BUG_ON(pfn < zone->zone_start_pfn);
5560 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5562 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5564 __set_bit(bitidx + start_bitidx, bitmap);
5566 __clear_bit(bitidx + start_bitidx, bitmap);
5570 * This function checks whether pageblock includes unmovable pages or not.
5571 * If @count is not zero, it is okay to include less @count unmovable pages
5573 * PageLRU check wihtout isolation or lru_lock could race so that
5574 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5575 * expect this function should be exact.
5577 bool has_unmovable_pages(struct zone *zone, struct page *page, int count)
5579 unsigned long pfn, iter, found;
5583 * For avoiding noise data, lru_add_drain_all() should be called
5584 * If ZONE_MOVABLE, the zone never contains unmovable pages
5586 if (zone_idx(zone) == ZONE_MOVABLE)
5588 mt = get_pageblock_migratetype(page);
5589 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5592 pfn = page_to_pfn(page);
5593 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5594 unsigned long check = pfn + iter;
5596 if (!pfn_valid_within(check))
5599 page = pfn_to_page(check);
5601 * We can't use page_count without pin a page
5602 * because another CPU can free compound page.
5603 * This check already skips compound tails of THP
5604 * because their page->_count is zero at all time.
5606 if (!atomic_read(&page->_count)) {
5607 if (PageBuddy(page))
5608 iter += (1 << page_order(page)) - 1;
5615 * If there are RECLAIMABLE pages, we need to check it.
5616 * But now, memory offline itself doesn't call shrink_slab()
5617 * and it still to be fixed.
5620 * If the page is not RAM, page_count()should be 0.
5621 * we don't need more check. This is an _used_ not-movable page.
5623 * The problematic thing here is PG_reserved pages. PG_reserved
5624 * is set to both of a memory hole page and a _used_ kernel
5633 bool is_pageblock_removable_nolock(struct page *page)
5639 * We have to be careful here because we are iterating over memory
5640 * sections which are not zone aware so we might end up outside of
5641 * the zone but still within the section.
5642 * We have to take care about the node as well. If the node is offline
5643 * its NODE_DATA will be NULL - see page_zone.
5645 if (!node_online(page_to_nid(page)))
5648 zone = page_zone(page);
5649 pfn = page_to_pfn(page);
5650 if (zone->zone_start_pfn > pfn ||
5651 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5654 return !has_unmovable_pages(zone, page, 0);
5659 static unsigned long pfn_max_align_down(unsigned long pfn)
5661 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5662 pageblock_nr_pages) - 1);
5665 static unsigned long pfn_max_align_up(unsigned long pfn)
5667 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5668 pageblock_nr_pages));
5671 /* [start, end) must belong to a single zone. */
5672 static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5674 /* This function is based on compact_zone() from compaction.c. */
5675 unsigned long nr_reclaimed;
5676 unsigned long pfn = start;
5677 unsigned int tries = 0;
5680 struct compact_control cc = {
5681 .nr_migratepages = 0,
5683 .zone = page_zone(pfn_to_page(start)),
5685 .ignore_skip_hint = true,
5687 INIT_LIST_HEAD(&cc.migratepages);
5689 migrate_prep_local();
5691 while (pfn < end || !list_empty(&cc.migratepages)) {
5692 if (fatal_signal_pending(current)) {
5697 if (list_empty(&cc.migratepages)) {
5698 cc.nr_migratepages = 0;
5699 pfn = isolate_migratepages_range(cc.zone, &cc,
5706 } else if (++tries == 5) {
5707 ret = ret < 0 ? ret : -EBUSY;
5711 nr_reclaimed = reclaim_clean_pages_from_list(cc.zone,
5713 cc.nr_migratepages -= nr_reclaimed;
5715 ret = migrate_pages(&cc.migratepages,
5716 alloc_migrate_target,
5717 0, false, MIGRATE_SYNC);
5720 putback_lru_pages(&cc.migratepages);
5721 return ret > 0 ? 0 : ret;
5725 * Update zone's cma pages counter used for watermark level calculation.
5727 static inline void __update_cma_watermarks(struct zone *zone, int count)
5729 unsigned long flags;
5730 spin_lock_irqsave(&zone->lock, flags);
5731 zone->min_cma_pages += count;
5732 spin_unlock_irqrestore(&zone->lock, flags);
5733 setup_per_zone_wmarks();
5737 * Trigger memory pressure bump to reclaim some pages in order to be able to
5738 * allocate 'count' pages in single page units. Does similar work as
5739 *__alloc_pages_slowpath() function.
5741 static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
5743 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
5744 struct zonelist *zonelist = node_zonelist(0, gfp_mask);
5745 int did_some_progress = 0;
5749 * Increase level of watermarks to force kswapd do his job
5750 * to stabilise at new watermark level.
5752 __update_cma_watermarks(zone, count);
5754 /* Obey watermarks as if the page was being allocated */
5755 while (!zone_watermark_ok(zone, 0, low_wmark_pages(zone), 0, 0)) {
5756 wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));
5758 did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
5760 if (!did_some_progress) {
5761 /* Exhausted what can be done so it's blamo time */
5762 out_of_memory(zonelist, gfp_mask, order, NULL, false);
5766 /* Restore original watermark levels. */
5767 __update_cma_watermarks(zone, -count);
5773 * alloc_contig_range() -- tries to allocate given range of pages
5774 * @start: start PFN to allocate
5775 * @end: one-past-the-last PFN to allocate
5776 * @migratetype: migratetype of the underlaying pageblocks (either
5777 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5778 * in range must have the same migratetype and it must
5779 * be either of the two.
5781 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5782 * aligned, however it's the caller's responsibility to guarantee that
5783 * we are the only thread that changes migrate type of pageblocks the
5786 * The PFN range must belong to a single zone.
5788 * Returns zero on success or negative error code. On success all
5789 * pages which PFN is in [start, end) are allocated for the caller and
5790 * need to be freed with free_contig_range().
5792 int alloc_contig_range(unsigned long start, unsigned long end,
5793 unsigned migratetype)
5795 struct zone *zone = page_zone(pfn_to_page(start));
5796 unsigned long outer_start, outer_end;
5800 * What we do here is we mark all pageblocks in range as
5801 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5802 * have different sizes, and due to the way page allocator
5803 * work, we align the range to biggest of the two pages so
5804 * that page allocator won't try to merge buddies from
5805 * different pageblocks and change MIGRATE_ISOLATE to some
5806 * other migration type.
5808 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5809 * migrate the pages from an unaligned range (ie. pages that
5810 * we are interested in). This will put all the pages in
5811 * range back to page allocator as MIGRATE_ISOLATE.
5813 * When this is done, we take the pages in range from page
5814 * allocator removing them from the buddy system. This way
5815 * page allocator will never consider using them.
5817 * This lets us mark the pageblocks back as
5818 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5819 * aligned range but not in the unaligned, original range are
5820 * put back to page allocator so that buddy can use them.
5823 ret = start_isolate_page_range(pfn_max_align_down(start),
5824 pfn_max_align_up(end), migratetype);
5828 ret = __alloc_contig_migrate_range(start, end);
5833 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5834 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5835 * more, all pages in [start, end) are free in page allocator.
5836 * What we are going to do is to allocate all pages from
5837 * [start, end) (that is remove them from page allocator).
5839 * The only problem is that pages at the beginning and at the
5840 * end of interesting range may be not aligned with pages that
5841 * page allocator holds, ie. they can be part of higher order
5842 * pages. Because of this, we reserve the bigger range and
5843 * once this is done free the pages we are not interested in.
5845 * We don't have to hold zone->lock here because the pages are
5846 * isolated thus they won't get removed from buddy.
5849 lru_add_drain_all();
5853 outer_start = start;
5854 while (!PageBuddy(pfn_to_page(outer_start))) {
5855 if (++order >= MAX_ORDER) {
5859 outer_start &= ~0UL << order;
5862 /* Make sure the range is really isolated. */
5863 if (test_pages_isolated(outer_start, end)) {
5864 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5871 * Reclaim enough pages to make sure that contiguous allocation
5872 * will not starve the system.
5874 __reclaim_pages(zone, GFP_HIGHUSER_MOVABLE, end-start);
5876 /* Grab isolated pages from freelists. */
5877 outer_end = isolate_freepages_range(outer_start, end);
5883 /* Free head and tail (if any) */
5884 if (start != outer_start)
5885 free_contig_range(outer_start, start - outer_start);
5886 if (end != outer_end)
5887 free_contig_range(end, outer_end - end);
5890 undo_isolate_page_range(pfn_max_align_down(start),
5891 pfn_max_align_up(end), migratetype);
5895 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5897 for (; nr_pages--; ++pfn)
5898 __free_page(pfn_to_page(pfn));
5902 #ifdef CONFIG_MEMORY_HOTPLUG
5903 static int __meminit __zone_pcp_update(void *data)
5905 struct zone *zone = data;
5907 unsigned long batch = zone_batchsize(zone), flags;
5909 for_each_possible_cpu(cpu) {
5910 struct per_cpu_pageset *pset;
5911 struct per_cpu_pages *pcp;
5913 pset = per_cpu_ptr(zone->pageset, cpu);
5916 local_irq_save(flags);
5918 free_pcppages_bulk(zone, pcp->count, pcp);
5919 drain_zonestat(zone, pset);
5920 setup_pageset(pset, batch);
5921 local_irq_restore(flags);
5926 void __meminit zone_pcp_update(struct zone *zone)
5928 stop_machine(__zone_pcp_update, zone, NULL);
5932 #ifdef CONFIG_MEMORY_HOTREMOVE
5933 void zone_pcp_reset(struct zone *zone)
5935 unsigned long flags;
5937 struct per_cpu_pageset *pset;
5939 /* avoid races with drain_pages() */
5940 local_irq_save(flags);
5941 if (zone->pageset != &boot_pageset) {
5942 for_each_online_cpu(cpu) {
5943 pset = per_cpu_ptr(zone->pageset, cpu);
5944 drain_zonestat(zone, pset);
5946 free_percpu(zone->pageset);
5947 zone->pageset = &boot_pageset;
5949 local_irq_restore(flags);
5953 * All pages in the range must be isolated before calling this.
5956 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5962 unsigned long flags;
5963 /* find the first valid pfn */
5964 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5969 zone = page_zone(pfn_to_page(pfn));
5970 spin_lock_irqsave(&zone->lock, flags);
5972 while (pfn < end_pfn) {
5973 if (!pfn_valid(pfn)) {
5977 page = pfn_to_page(pfn);
5978 BUG_ON(page_count(page));
5979 BUG_ON(!PageBuddy(page));
5980 order = page_order(page);
5981 #ifdef CONFIG_DEBUG_VM
5982 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5983 pfn, 1 << order, end_pfn);
5985 list_del(&page->lru);
5986 rmv_page_order(page);
5987 zone->free_area[order].nr_free--;
5988 __mod_zone_page_state(zone, NR_FREE_PAGES,
5990 for (i = 0; i < (1 << order); i++)
5991 SetPageReserved((page+i));
5992 pfn += (1 << order);
5994 spin_unlock_irqrestore(&zone->lock, flags);
5998 #ifdef CONFIG_MEMORY_FAILURE
5999 bool is_free_buddy_page(struct page *page)
6001 struct zone *zone = page_zone(page);
6002 unsigned long pfn = page_to_pfn(page);
6003 unsigned long flags;
6006 spin_lock_irqsave(&zone->lock, flags);
6007 for (order = 0; order < MAX_ORDER; order++) {
6008 struct page *page_head = page - (pfn & ((1 << order) - 1));
6010 if (PageBuddy(page_head) && page_order(page_head) >= order)
6013 spin_unlock_irqrestore(&zone->lock, flags);
6015 return order < MAX_ORDER;
6019 static const struct trace_print_flags pageflag_names[] = {
6020 {1UL << PG_locked, "locked" },
6021 {1UL << PG_error, "error" },
6022 {1UL << PG_referenced, "referenced" },
6023 {1UL << PG_uptodate, "uptodate" },
6024 {1UL << PG_dirty, "dirty" },
6025 {1UL << PG_lru, "lru" },
6026 {1UL << PG_active, "active" },
6027 {1UL << PG_slab, "slab" },
6028 {1UL << PG_owner_priv_1, "owner_priv_1" },
6029 {1UL << PG_arch_1, "arch_1" },
6030 {1UL << PG_reserved, "reserved" },
6031 {1UL << PG_private, "private" },
6032 {1UL << PG_private_2, "private_2" },
6033 {1UL << PG_writeback, "writeback" },
6034 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6035 {1UL << PG_head, "head" },
6036 {1UL << PG_tail, "tail" },
6038 {1UL << PG_compound, "compound" },
6040 {1UL << PG_swapcache, "swapcache" },
6041 {1UL << PG_mappedtodisk, "mappedtodisk" },
6042 {1UL << PG_reclaim, "reclaim" },
6043 {1UL << PG_swapbacked, "swapbacked" },
6044 {1UL << PG_unevictable, "unevictable" },
6046 {1UL << PG_mlocked, "mlocked" },
6048 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6049 {1UL << PG_uncached, "uncached" },
6051 #ifdef CONFIG_MEMORY_FAILURE
6052 {1UL << PG_hwpoison, "hwpoison" },
6054 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6055 {1UL << PG_compound_lock, "compound_lock" },
6059 static void dump_page_flags(unsigned long flags)
6061 const char *delim = "";
6065 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6067 printk(KERN_ALERT "page flags: %#lx(", flags);
6069 /* remove zone id */
6070 flags &= (1UL << NR_PAGEFLAGS) - 1;
6072 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6074 mask = pageflag_names[i].mask;
6075 if ((flags & mask) != mask)
6079 printk("%s%s", delim, pageflag_names[i].name);
6083 /* check for left over flags */
6085 printk("%s%#lx", delim, flags);
6090 void dump_page(struct page *page)
6093 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6094 page, atomic_read(&page->_count), page_mapcount(page),
6095 page->mapping, page->index);
6096 dump_page_flags(page->flags);
6097 mem_cgroup_print_bad_page(page);
6100 /* reset zone->present_pages */
6101 void reset_zone_present_pages(void)
6106 for_each_node_state(nid, N_HIGH_MEMORY) {
6107 for (i = 0; i < MAX_NR_ZONES; i++) {
6108 z = NODE_DATA(nid)->node_zones + i;
6109 z->present_pages = 0;
6114 /* calculate zone's present pages in buddy system */
6115 void fixup_zone_present_pages(int nid, unsigned long start_pfn,
6116 unsigned long end_pfn)
6119 unsigned long zone_start_pfn, zone_end_pfn;
6122 for (i = 0; i < MAX_NR_ZONES; i++) {
6123 z = NODE_DATA(nid)->node_zones + i;
6124 zone_start_pfn = z->zone_start_pfn;
6125 zone_end_pfn = zone_start_pfn + z->spanned_pages;
6127 /* if the two regions intersect */
6128 if (!(zone_start_pfn >= end_pfn || zone_end_pfn <= start_pfn))
6129 z->present_pages += min(end_pfn, zone_end_pfn) -
6130 max(start_pfn, zone_start_pfn);