2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
62 #include <asm/tlbflush.h>
63 #include <asm/div64.h>
66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67 DEFINE_PER_CPU(int, numa_node);
68 EXPORT_PER_CPU_SYMBOL(numa_node);
71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
78 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
83 * Array of node states.
85 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86 [N_POSSIBLE] = NODE_MASK_ALL,
87 [N_ONLINE] = { { [0] = 1UL } },
89 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
91 [N_HIGH_MEMORY] = { { [0] = 1UL } },
93 [N_CPU] = { { [0] = 1UL } },
96 EXPORT_SYMBOL(node_states);
98 unsigned long totalram_pages __read_mostly;
99 unsigned long totalreserve_pages __read_mostly;
101 * When calculating the number of globally allowed dirty pages, there
102 * is a certain number of per-zone reserves that should not be
103 * considered dirtyable memory. This is the sum of those reserves
104 * over all existing zones that contribute dirtyable memory.
106 unsigned long dirty_balance_reserve __read_mostly;
108 int percpu_pagelist_fraction;
109 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
111 #ifdef CONFIG_PM_SLEEP
113 * The following functions are used by the suspend/hibernate code to temporarily
114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
115 * while devices are suspended. To avoid races with the suspend/hibernate code,
116 * they should always be called with pm_mutex held (gfp_allowed_mask also should
117 * only be modified with pm_mutex held, unless the suspend/hibernate code is
118 * guaranteed not to run in parallel with that modification).
121 static gfp_t saved_gfp_mask;
123 void pm_restore_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 if (saved_gfp_mask) {
127 gfp_allowed_mask = saved_gfp_mask;
132 void pm_restrict_gfp_mask(void)
134 WARN_ON(!mutex_is_locked(&pm_mutex));
135 WARN_ON(saved_gfp_mask);
136 saved_gfp_mask = gfp_allowed_mask;
137 gfp_allowed_mask &= ~GFP_IOFS;
140 bool pm_suspended_storage(void)
142 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
146 #endif /* CONFIG_PM_SLEEP */
148 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
149 int pageblock_order __read_mostly;
152 static void __free_pages_ok(struct page *page, unsigned int order);
155 * results with 256, 32 in the lowmem_reserve sysctl:
156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
157 * 1G machine -> (16M dma, 784M normal, 224M high)
158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
162 * TBD: should special case ZONE_DMA32 machines here - in those we normally
163 * don't need any ZONE_NORMAL reservation
165 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
166 #ifdef CONFIG_ZONE_DMA
169 #ifdef CONFIG_ZONE_DMA32
172 #ifdef CONFIG_HIGHMEM
178 EXPORT_SYMBOL(totalram_pages);
180 static char * const zone_names[MAX_NR_ZONES] = {
181 #ifdef CONFIG_ZONE_DMA
184 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 int min_free_kbytes = 1024;
196 static unsigned long __meminitdata nr_kernel_pages;
197 static unsigned long __meminitdata nr_all_pages;
198 static unsigned long __meminitdata dma_reserve;
200 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
201 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
202 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
203 static unsigned long __initdata required_kernelcore;
204 static unsigned long __initdata required_movablecore;
205 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
207 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
209 EXPORT_SYMBOL(movable_zone);
210 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
213 int nr_node_ids __read_mostly = MAX_NUMNODES;
214 int nr_online_nodes __read_mostly = 1;
215 EXPORT_SYMBOL(nr_node_ids);
216 EXPORT_SYMBOL(nr_online_nodes);
219 int page_group_by_mobility_disabled __read_mostly;
223 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
224 * Instead, use {un}set_pageblock_isolate.
226 void set_pageblock_migratetype(struct page *page, int migratetype)
229 if (unlikely(page_group_by_mobility_disabled))
230 migratetype = MIGRATE_UNMOVABLE;
232 set_pageblock_flags_group(page, (unsigned long)migratetype,
233 PB_migrate, PB_migrate_end);
236 bool oom_killer_disabled __read_mostly;
238 #ifdef CONFIG_DEBUG_VM
239 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
243 unsigned long pfn = page_to_pfn(page);
246 seq = zone_span_seqbegin(zone);
247 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
249 else if (pfn < zone->zone_start_pfn)
251 } while (zone_span_seqretry(zone, seq));
256 static int page_is_consistent(struct zone *zone, struct page *page)
258 if (!pfn_valid_within(page_to_pfn(page)))
260 if (zone != page_zone(page))
266 * Temporary debugging check for pages not lying within a given zone.
268 static int bad_range(struct zone *zone, struct page *page)
270 if (page_outside_zone_boundaries(zone, page))
272 if (!page_is_consistent(zone, page))
278 static inline int bad_range(struct zone *zone, struct page *page)
284 static void bad_page(struct page *page)
286 static unsigned long resume;
287 static unsigned long nr_shown;
288 static unsigned long nr_unshown;
290 /* Don't complain about poisoned pages */
291 if (PageHWPoison(page)) {
292 reset_page_mapcount(page); /* remove PageBuddy */
297 * Allow a burst of 60 reports, then keep quiet for that minute;
298 * or allow a steady drip of one report per second.
300 if (nr_shown == 60) {
301 if (time_before(jiffies, resume)) {
307 "BUG: Bad page state: %lu messages suppressed\n",
314 resume = jiffies + 60 * HZ;
316 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
317 current->comm, page_to_pfn(page));
323 /* Leave bad fields for debug, except PageBuddy could make trouble */
324 reset_page_mapcount(page); /* remove PageBuddy */
325 add_taint(TAINT_BAD_PAGE);
329 * Higher-order pages are called "compound pages". They are structured thusly:
331 * The first PAGE_SIZE page is called the "head page".
333 * The remaining PAGE_SIZE pages are called "tail pages".
335 * All pages have PG_compound set. All tail pages have their ->first_page
336 * pointing at the head page.
338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
343 static void free_compound_page(struct page *page)
345 __free_pages_ok(page, compound_order(page));
348 void prep_compound_page(struct page *page, unsigned long order)
351 int nr_pages = 1 << order;
353 set_compound_page_dtor(page, free_compound_page);
354 set_compound_order(page, order);
356 for (i = 1; i < nr_pages; i++) {
357 struct page *p = page + i;
359 set_page_count(p, 0);
360 p->first_page = page;
364 /* update __split_huge_page_refcount if you change this function */
365 static int destroy_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
371 if (unlikely(compound_order(page) != order) ||
372 unlikely(!PageHead(page))) {
377 __ClearPageHead(page);
379 for (i = 1; i < nr_pages; i++) {
380 struct page *p = page + i;
382 if (unlikely(!PageTail(p) || (p->first_page != page))) {
392 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
397 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
398 * and __GFP_HIGHMEM from hard or soft interrupt context.
400 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
401 for (i = 0; i < (1 << order); i++)
402 clear_highpage(page + i);
405 #ifdef CONFIG_DEBUG_PAGEALLOC
406 unsigned int _debug_guardpage_minorder;
408 static int __init debug_guardpage_minorder_setup(char *buf)
412 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
413 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
416 _debug_guardpage_minorder = res;
417 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
420 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
422 static inline void set_page_guard_flag(struct page *page)
424 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
427 static inline void clear_page_guard_flag(struct page *page)
429 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
432 static inline void set_page_guard_flag(struct page *page) { }
433 static inline void clear_page_guard_flag(struct page *page) { }
436 static inline void set_page_order(struct page *page, int order)
438 set_page_private(page, order);
439 __SetPageBuddy(page);
442 static inline void rmv_page_order(struct page *page)
444 __ClearPageBuddy(page);
445 set_page_private(page, 0);
449 * Locate the struct page for both the matching buddy in our
450 * pair (buddy1) and the combined O(n+1) page they form (page).
452 * 1) Any buddy B1 will have an order O twin B2 which satisfies
453 * the following equation:
455 * For example, if the starting buddy (buddy2) is #8 its order
457 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
459 * 2) Any buddy B will have an order O+1 parent P which
460 * satisfies the following equation:
463 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
465 static inline unsigned long
466 __find_buddy_index(unsigned long page_idx, unsigned int order)
468 return page_idx ^ (1 << order);
472 * This function checks whether a page is free && is the buddy
473 * we can do coalesce a page and its buddy if
474 * (a) the buddy is not in a hole &&
475 * (b) the buddy is in the buddy system &&
476 * (c) a page and its buddy have the same order &&
477 * (d) a page and its buddy are in the same zone.
479 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
480 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
482 * For recording page's order, we use page_private(page).
484 static inline int page_is_buddy(struct page *page, struct page *buddy,
487 if (!pfn_valid_within(page_to_pfn(buddy)))
490 if (page_zone_id(page) != page_zone_id(buddy))
493 if (page_is_guard(buddy) && page_order(buddy) == order) {
494 VM_BUG_ON(page_count(buddy) != 0);
498 if (PageBuddy(buddy) && page_order(buddy) == order) {
499 VM_BUG_ON(page_count(buddy) != 0);
506 * Freeing function for a buddy system allocator.
508 * The concept of a buddy system is to maintain direct-mapped table
509 * (containing bit values) for memory blocks of various "orders".
510 * The bottom level table contains the map for the smallest allocatable
511 * units of memory (here, pages), and each level above it describes
512 * pairs of units from the levels below, hence, "buddies".
513 * At a high level, all that happens here is marking the table entry
514 * at the bottom level available, and propagating the changes upward
515 * as necessary, plus some accounting needed to play nicely with other
516 * parts of the VM system.
517 * At each level, we keep a list of pages, which are heads of continuous
518 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
519 * order is recorded in page_private(page) field.
520 * So when we are allocating or freeing one, we can derive the state of the
521 * other. That is, if we allocate a small block, and both were
522 * free, the remainder of the region must be split into blocks.
523 * If a block is freed, and its buddy is also free, then this
524 * triggers coalescing into a block of larger size.
529 static inline void __free_one_page(struct page *page,
530 struct zone *zone, unsigned int order,
533 unsigned long page_idx;
534 unsigned long combined_idx;
535 unsigned long uninitialized_var(buddy_idx);
538 if (unlikely(PageCompound(page)))
539 if (unlikely(destroy_compound_page(page, order)))
542 VM_BUG_ON(migratetype == -1);
544 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
546 VM_BUG_ON(page_idx & ((1 << order) - 1));
547 VM_BUG_ON(bad_range(zone, page));
549 while (order < MAX_ORDER-1) {
550 buddy_idx = __find_buddy_index(page_idx, order);
551 buddy = page + (buddy_idx - page_idx);
552 if (!page_is_buddy(page, buddy, order))
555 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
556 * merge with it and move up one order.
558 if (page_is_guard(buddy)) {
559 clear_page_guard_flag(buddy);
560 set_page_private(page, 0);
561 __mod_zone_freepage_state(zone, 1 << order,
564 list_del(&buddy->lru);
565 zone->free_area[order].nr_free--;
566 rmv_page_order(buddy);
568 combined_idx = buddy_idx & page_idx;
569 page = page + (combined_idx - page_idx);
570 page_idx = combined_idx;
573 set_page_order(page, order);
576 * If this is not the largest possible page, check if the buddy
577 * of the next-highest order is free. If it is, it's possible
578 * that pages are being freed that will coalesce soon. In case,
579 * that is happening, add the free page to the tail of the list
580 * so it's less likely to be used soon and more likely to be merged
581 * as a higher order page
583 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
584 struct page *higher_page, *higher_buddy;
585 combined_idx = buddy_idx & page_idx;
586 higher_page = page + (combined_idx - page_idx);
587 buddy_idx = __find_buddy_index(combined_idx, order + 1);
588 higher_buddy = higher_page + (buddy_idx - combined_idx);
589 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
590 list_add_tail(&page->lru,
591 &zone->free_area[order].free_list[migratetype]);
596 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
598 zone->free_area[order].nr_free++;
602 * free_page_mlock() -- clean up attempts to free and mlocked() page.
603 * Page should not be on lru, so no need to fix that up.
604 * free_pages_check() will verify...
606 static inline void free_page_mlock(struct page *page)
608 __dec_zone_page_state(page, NR_MLOCK);
609 __count_vm_event(UNEVICTABLE_MLOCKFREED);
612 static inline int free_pages_check(struct page *page)
614 if (unlikely(page_mapcount(page) |
615 (page->mapping != NULL) |
616 (atomic_read(&page->_count) != 0) |
617 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
618 (mem_cgroup_bad_page_check(page)))) {
622 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
623 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
628 * Frees a number of pages from the PCP lists
629 * Assumes all pages on list are in same zone, and of same order.
630 * count is the number of pages to free.
632 * If the zone was previously in an "all pages pinned" state then look to
633 * see if this freeing clears that state.
635 * And clear the zone's pages_scanned counter, to hold off the "all pages are
636 * pinned" detection logic.
638 static void free_pcppages_bulk(struct zone *zone, int count,
639 struct per_cpu_pages *pcp)
645 spin_lock(&zone->lock);
646 zone->all_unreclaimable = 0;
647 zone->pages_scanned = 0;
651 struct list_head *list;
654 * Remove pages from lists in a round-robin fashion. A
655 * batch_free count is maintained that is incremented when an
656 * empty list is encountered. This is so more pages are freed
657 * off fuller lists instead of spinning excessively around empty
662 if (++migratetype == MIGRATE_PCPTYPES)
664 list = &pcp->lists[migratetype];
665 } while (list_empty(list));
667 /* This is the only non-empty list. Free them all. */
668 if (batch_free == MIGRATE_PCPTYPES)
669 batch_free = to_free;
672 int mt; /* migratetype of the to-be-freed page */
674 page = list_entry(list->prev, struct page, lru);
675 /* must delete as __free_one_page list manipulates */
676 list_del(&page->lru);
677 mt = get_freepage_migratetype(page);
678 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
679 __free_one_page(page, zone, 0, mt);
680 trace_mm_page_pcpu_drain(page, 0, mt);
681 if (is_migrate_cma(mt))
682 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
683 } while (--to_free && --batch_free && !list_empty(list));
685 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
686 spin_unlock(&zone->lock);
689 static void free_one_page(struct zone *zone, struct page *page, int order,
692 spin_lock(&zone->lock);
693 zone->all_unreclaimable = 0;
694 zone->pages_scanned = 0;
696 __free_one_page(page, zone, order, migratetype);
697 if (unlikely(migratetype != MIGRATE_ISOLATE))
698 __mod_zone_freepage_state(zone, 1 << order, migratetype);
699 spin_unlock(&zone->lock);
702 static bool free_pages_prepare(struct page *page, unsigned int order)
707 trace_mm_page_free(page, order);
708 kmemcheck_free_shadow(page, order);
711 page->mapping = NULL;
712 for (i = 0; i < (1 << order); i++)
713 bad += free_pages_check(page + i);
717 if (!PageHighMem(page)) {
718 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
719 debug_check_no_obj_freed(page_address(page),
722 arch_free_page(page, order);
723 kernel_map_pages(page, 1 << order, 0);
728 static void __free_pages_ok(struct page *page, unsigned int order)
731 int wasMlocked = __TestClearPageMlocked(page);
734 if (!free_pages_prepare(page, order))
737 local_irq_save(flags);
738 if (unlikely(wasMlocked))
739 free_page_mlock(page);
740 __count_vm_events(PGFREE, 1 << order);
741 migratetype = get_pageblock_migratetype(page);
742 set_freepage_migratetype(page, migratetype);
743 free_one_page(page_zone(page), page, order, migratetype);
744 local_irq_restore(flags);
747 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
749 unsigned int nr_pages = 1 << order;
753 for (loop = 0; loop < nr_pages; loop++) {
754 struct page *p = &page[loop];
756 if (loop + 1 < nr_pages)
758 __ClearPageReserved(p);
759 set_page_count(p, 0);
762 set_page_refcounted(page);
763 __free_pages(page, order);
767 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
768 void __init init_cma_reserved_pageblock(struct page *page)
770 unsigned i = pageblock_nr_pages;
771 struct page *p = page;
774 __ClearPageReserved(p);
775 set_page_count(p, 0);
778 set_page_refcounted(page);
779 set_pageblock_migratetype(page, MIGRATE_CMA);
780 __free_pages(page, pageblock_order);
781 totalram_pages += pageblock_nr_pages;
786 * The order of subdivision here is critical for the IO subsystem.
787 * Please do not alter this order without good reasons and regression
788 * testing. Specifically, as large blocks of memory are subdivided,
789 * the order in which smaller blocks are delivered depends on the order
790 * they're subdivided in this function. This is the primary factor
791 * influencing the order in which pages are delivered to the IO
792 * subsystem according to empirical testing, and this is also justified
793 * by considering the behavior of a buddy system containing a single
794 * large block of memory acted on by a series of small allocations.
795 * This behavior is a critical factor in sglist merging's success.
799 static inline void expand(struct zone *zone, struct page *page,
800 int low, int high, struct free_area *area,
803 unsigned long size = 1 << high;
809 VM_BUG_ON(bad_range(zone, &page[size]));
811 #ifdef CONFIG_DEBUG_PAGEALLOC
812 if (high < debug_guardpage_minorder()) {
814 * Mark as guard pages (or page), that will allow to
815 * merge back to allocator when buddy will be freed.
816 * Corresponding page table entries will not be touched,
817 * pages will stay not present in virtual address space
819 INIT_LIST_HEAD(&page[size].lru);
820 set_page_guard_flag(&page[size]);
821 set_page_private(&page[size], high);
822 /* Guard pages are not available for any usage */
823 __mod_zone_freepage_state(zone, -(1 << high),
828 list_add(&page[size].lru, &area->free_list[migratetype]);
830 set_page_order(&page[size], high);
835 * This page is about to be returned from the page allocator
837 static inline int check_new_page(struct page *page)
839 if (unlikely(page_mapcount(page) |
840 (page->mapping != NULL) |
841 (atomic_read(&page->_count) != 0) |
842 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
843 (mem_cgroup_bad_page_check(page)))) {
850 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
854 for (i = 0; i < (1 << order); i++) {
855 struct page *p = page + i;
856 if (unlikely(check_new_page(p)))
860 set_page_private(page, 0);
861 set_page_refcounted(page);
863 arch_alloc_page(page, order);
864 kernel_map_pages(page, 1 << order, 1);
866 if (gfp_flags & __GFP_ZERO)
867 prep_zero_page(page, order, gfp_flags);
869 if (order && (gfp_flags & __GFP_COMP))
870 prep_compound_page(page, order);
876 * Go through the free lists for the given migratetype and remove
877 * the smallest available page from the freelists
880 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
883 unsigned int current_order;
884 struct free_area * area;
887 /* Find a page of the appropriate size in the preferred list */
888 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
889 area = &(zone->free_area[current_order]);
890 if (list_empty(&area->free_list[migratetype]))
893 page = list_entry(area->free_list[migratetype].next,
895 list_del(&page->lru);
896 rmv_page_order(page);
898 expand(zone, page, order, current_order, area, migratetype);
907 * This array describes the order lists are fallen back to when
908 * the free lists for the desirable migrate type are depleted
910 static int fallbacks[MIGRATE_TYPES][4] = {
911 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
912 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
914 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
915 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
917 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
919 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
920 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
924 * Move the free pages in a range to the free lists of the requested type.
925 * Note that start_page and end_pages are not aligned on a pageblock
926 * boundary. If alignment is required, use move_freepages_block()
928 static int move_freepages(struct zone *zone,
929 struct page *start_page, struct page *end_page,
936 #ifndef CONFIG_HOLES_IN_ZONE
938 * page_zone is not safe to call in this context when
939 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
940 * anyway as we check zone boundaries in move_freepages_block().
941 * Remove at a later date when no bug reports exist related to
942 * grouping pages by mobility
944 BUG_ON(page_zone(start_page) != page_zone(end_page));
947 for (page = start_page; page <= end_page;) {
948 /* Make sure we are not inadvertently changing nodes */
949 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
951 if (!pfn_valid_within(page_to_pfn(page))) {
956 if (!PageBuddy(page)) {
961 order = page_order(page);
962 list_move(&page->lru,
963 &zone->free_area[order].free_list[migratetype]);
964 set_freepage_migratetype(page, migratetype);
966 pages_moved += 1 << order;
972 int move_freepages_block(struct zone *zone, struct page *page,
975 unsigned long start_pfn, end_pfn;
976 struct page *start_page, *end_page;
978 start_pfn = page_to_pfn(page);
979 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
980 start_page = pfn_to_page(start_pfn);
981 end_page = start_page + pageblock_nr_pages - 1;
982 end_pfn = start_pfn + pageblock_nr_pages - 1;
984 /* Do not cross zone boundaries */
985 if (start_pfn < zone->zone_start_pfn)
987 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
990 return move_freepages(zone, start_page, end_page, migratetype);
993 static void change_pageblock_range(struct page *pageblock_page,
994 int start_order, int migratetype)
996 int nr_pageblocks = 1 << (start_order - pageblock_order);
998 while (nr_pageblocks--) {
999 set_pageblock_migratetype(pageblock_page, migratetype);
1000 pageblock_page += pageblock_nr_pages;
1004 /* Remove an element from the buddy allocator from the fallback list */
1005 static inline struct page *
1006 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1008 struct free_area * area;
1013 /* Find the largest possible block of pages in the other list */
1014 for (current_order = MAX_ORDER-1; current_order >= order;
1017 migratetype = fallbacks[start_migratetype][i];
1019 /* MIGRATE_RESERVE handled later if necessary */
1020 if (migratetype == MIGRATE_RESERVE)
1023 area = &(zone->free_area[current_order]);
1024 if (list_empty(&area->free_list[migratetype]))
1027 page = list_entry(area->free_list[migratetype].next,
1032 * If breaking a large block of pages, move all free
1033 * pages to the preferred allocation list. If falling
1034 * back for a reclaimable kernel allocation, be more
1035 * aggressive about taking ownership of free pages
1037 * On the other hand, never change migration
1038 * type of MIGRATE_CMA pageblocks nor move CMA
1039 * pages on different free lists. We don't
1040 * want unmovable pages to be allocated from
1041 * MIGRATE_CMA areas.
1043 if (!is_migrate_cma(migratetype) &&
1044 (unlikely(current_order >= pageblock_order / 2) ||
1045 start_migratetype == MIGRATE_RECLAIMABLE ||
1046 page_group_by_mobility_disabled)) {
1048 pages = move_freepages_block(zone, page,
1051 /* Claim the whole block if over half of it is free */
1052 if (pages >= (1 << (pageblock_order-1)) ||
1053 page_group_by_mobility_disabled)
1054 set_pageblock_migratetype(page,
1057 migratetype = start_migratetype;
1060 /* Remove the page from the freelists */
1061 list_del(&page->lru);
1062 rmv_page_order(page);
1064 /* Take ownership for orders >= pageblock_order */
1065 if (current_order >= pageblock_order &&
1066 !is_migrate_cma(migratetype))
1067 change_pageblock_range(page, current_order,
1070 expand(zone, page, order, current_order, area,
1071 is_migrate_cma(migratetype)
1072 ? migratetype : start_migratetype);
1074 trace_mm_page_alloc_extfrag(page, order, current_order,
1075 start_migratetype, migratetype);
1085 * Do the hard work of removing an element from the buddy allocator.
1086 * Call me with the zone->lock already held.
1088 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1094 page = __rmqueue_smallest(zone, order, migratetype);
1096 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1097 page = __rmqueue_fallback(zone, order, migratetype);
1100 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1101 * is used because __rmqueue_smallest is an inline function
1102 * and we want just one call site
1105 migratetype = MIGRATE_RESERVE;
1110 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1115 * Obtain a specified number of elements from the buddy allocator, all under
1116 * a single hold of the lock, for efficiency. Add them to the supplied list.
1117 * Returns the number of new pages which were placed at *list.
1119 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1120 unsigned long count, struct list_head *list,
1121 int migratetype, int cold)
1123 int mt = migratetype, i;
1125 spin_lock(&zone->lock);
1126 for (i = 0; i < count; ++i) {
1127 struct page *page = __rmqueue(zone, order, migratetype);
1128 if (unlikely(page == NULL))
1132 * Split buddy pages returned by expand() are received here
1133 * in physical page order. The page is added to the callers and
1134 * list and the list head then moves forward. From the callers
1135 * perspective, the linked list is ordered by page number in
1136 * some conditions. This is useful for IO devices that can
1137 * merge IO requests if the physical pages are ordered
1140 if (likely(cold == 0))
1141 list_add(&page->lru, list);
1143 list_add_tail(&page->lru, list);
1144 if (IS_ENABLED(CONFIG_CMA)) {
1145 mt = get_pageblock_migratetype(page);
1146 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1149 set_freepage_migratetype(page, mt);
1151 if (is_migrate_cma(mt))
1152 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1155 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1156 spin_unlock(&zone->lock);
1162 * Called from the vmstat counter updater to drain pagesets of this
1163 * currently executing processor on remote nodes after they have
1166 * Note that this function must be called with the thread pinned to
1167 * a single processor.
1169 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1171 unsigned long flags;
1174 local_irq_save(flags);
1175 if (pcp->count >= pcp->batch)
1176 to_drain = pcp->batch;
1178 to_drain = pcp->count;
1180 free_pcppages_bulk(zone, to_drain, pcp);
1181 pcp->count -= to_drain;
1183 local_irq_restore(flags);
1188 * Drain pages of the indicated processor.
1190 * The processor must either be the current processor and the
1191 * thread pinned to the current processor or a processor that
1194 static void drain_pages(unsigned int cpu)
1196 unsigned long flags;
1199 for_each_populated_zone(zone) {
1200 struct per_cpu_pageset *pset;
1201 struct per_cpu_pages *pcp;
1203 local_irq_save(flags);
1204 pset = per_cpu_ptr(zone->pageset, cpu);
1208 free_pcppages_bulk(zone, pcp->count, pcp);
1211 local_irq_restore(flags);
1216 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1218 void drain_local_pages(void *arg)
1220 drain_pages(smp_processor_id());
1224 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1226 * Note that this code is protected against sending an IPI to an offline
1227 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1228 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1229 * nothing keeps CPUs from showing up after we populated the cpumask and
1230 * before the call to on_each_cpu_mask().
1232 void drain_all_pages(void)
1235 struct per_cpu_pageset *pcp;
1239 * Allocate in the BSS so we wont require allocation in
1240 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1242 static cpumask_t cpus_with_pcps;
1245 * We don't care about racing with CPU hotplug event
1246 * as offline notification will cause the notified
1247 * cpu to drain that CPU pcps and on_each_cpu_mask
1248 * disables preemption as part of its processing
1250 for_each_online_cpu(cpu) {
1251 bool has_pcps = false;
1252 for_each_populated_zone(zone) {
1253 pcp = per_cpu_ptr(zone->pageset, cpu);
1254 if (pcp->pcp.count) {
1260 cpumask_set_cpu(cpu, &cpus_with_pcps);
1262 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1264 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1267 #ifdef CONFIG_HIBERNATION
1269 void mark_free_pages(struct zone *zone)
1271 unsigned long pfn, max_zone_pfn;
1272 unsigned long flags;
1274 struct list_head *curr;
1276 if (!zone->spanned_pages)
1279 spin_lock_irqsave(&zone->lock, flags);
1281 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1282 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1283 if (pfn_valid(pfn)) {
1284 struct page *page = pfn_to_page(pfn);
1286 if (!swsusp_page_is_forbidden(page))
1287 swsusp_unset_page_free(page);
1290 for_each_migratetype_order(order, t) {
1291 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1294 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1295 for (i = 0; i < (1UL << order); i++)
1296 swsusp_set_page_free(pfn_to_page(pfn + i));
1299 spin_unlock_irqrestore(&zone->lock, flags);
1301 #endif /* CONFIG_PM */
1304 * Free a 0-order page
1305 * cold == 1 ? free a cold page : free a hot page
1307 void free_hot_cold_page(struct page *page, int cold)
1309 struct zone *zone = page_zone(page);
1310 struct per_cpu_pages *pcp;
1311 unsigned long flags;
1313 int wasMlocked = __TestClearPageMlocked(page);
1315 if (!free_pages_prepare(page, 0))
1318 migratetype = get_pageblock_migratetype(page);
1319 set_freepage_migratetype(page, migratetype);
1320 local_irq_save(flags);
1321 if (unlikely(wasMlocked))
1322 free_page_mlock(page);
1323 __count_vm_event(PGFREE);
1326 * We only track unmovable, reclaimable and movable on pcp lists.
1327 * Free ISOLATE pages back to the allocator because they are being
1328 * offlined but treat RESERVE as movable pages so we can get those
1329 * areas back if necessary. Otherwise, we may have to free
1330 * excessively into the page allocator
1332 if (migratetype >= MIGRATE_PCPTYPES) {
1333 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1334 free_one_page(zone, page, 0, migratetype);
1337 migratetype = MIGRATE_MOVABLE;
1340 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1342 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1344 list_add(&page->lru, &pcp->lists[migratetype]);
1346 if (pcp->count >= pcp->high) {
1347 free_pcppages_bulk(zone, pcp->batch, pcp);
1348 pcp->count -= pcp->batch;
1352 local_irq_restore(flags);
1356 * Free a list of 0-order pages
1358 void free_hot_cold_page_list(struct list_head *list, int cold)
1360 struct page *page, *next;
1362 list_for_each_entry_safe(page, next, list, lru) {
1363 trace_mm_page_free_batched(page, cold);
1364 free_hot_cold_page(page, cold);
1369 * split_page takes a non-compound higher-order page, and splits it into
1370 * n (1<<order) sub-pages: page[0..n]
1371 * Each sub-page must be freed individually.
1373 * Note: this is probably too low level an operation for use in drivers.
1374 * Please consult with lkml before using this in your driver.
1376 void split_page(struct page *page, unsigned int order)
1380 VM_BUG_ON(PageCompound(page));
1381 VM_BUG_ON(!page_count(page));
1383 #ifdef CONFIG_KMEMCHECK
1385 * Split shadow pages too, because free(page[0]) would
1386 * otherwise free the whole shadow.
1388 if (kmemcheck_page_is_tracked(page))
1389 split_page(virt_to_page(page[0].shadow), order);
1392 for (i = 1; i < (1 << order); i++)
1393 set_page_refcounted(page + i);
1397 * Similar to the split_page family of functions except that the page
1398 * required at the given order and being isolated now to prevent races
1399 * with parallel allocators
1401 int capture_free_page(struct page *page, int alloc_order, int migratetype)
1404 unsigned long watermark;
1408 BUG_ON(!PageBuddy(page));
1410 zone = page_zone(page);
1411 order = page_order(page);
1413 /* Obey watermarks as if the page was being allocated */
1414 watermark = low_wmark_pages(zone) + (1 << order);
1415 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1418 /* Remove page from free list */
1419 list_del(&page->lru);
1420 zone->free_area[order].nr_free--;
1421 rmv_page_order(page);
1423 mt = get_pageblock_migratetype(page);
1424 if (unlikely(mt != MIGRATE_ISOLATE))
1425 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1427 if (alloc_order != order)
1428 expand(zone, page, alloc_order, order,
1429 &zone->free_area[order], migratetype);
1431 /* Set the pageblock if the captured page is at least a pageblock */
1432 if (order >= pageblock_order - 1) {
1433 struct page *endpage = page + (1 << order) - 1;
1434 for (; page < endpage; page += pageblock_nr_pages) {
1435 int mt = get_pageblock_migratetype(page);
1436 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1437 set_pageblock_migratetype(page,
1442 return 1UL << order;
1446 * Similar to split_page except the page is already free. As this is only
1447 * being used for migration, the migratetype of the block also changes.
1448 * As this is called with interrupts disabled, the caller is responsible
1449 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1452 * Note: this is probably too low level an operation for use in drivers.
1453 * Please consult with lkml before using this in your driver.
1455 int split_free_page(struct page *page)
1460 BUG_ON(!PageBuddy(page));
1461 order = page_order(page);
1463 nr_pages = capture_free_page(page, order, 0);
1467 /* Split into individual pages */
1468 set_page_refcounted(page);
1469 split_page(page, order);
1474 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1475 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1479 struct page *buffered_rmqueue(struct zone *preferred_zone,
1480 struct zone *zone, int order, gfp_t gfp_flags,
1483 unsigned long flags;
1485 int cold = !!(gfp_flags & __GFP_COLD);
1488 if (likely(order == 0)) {
1489 struct per_cpu_pages *pcp;
1490 struct list_head *list;
1492 local_irq_save(flags);
1493 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1494 list = &pcp->lists[migratetype];
1495 if (list_empty(list)) {
1496 pcp->count += rmqueue_bulk(zone, 0,
1499 if (unlikely(list_empty(list)))
1504 page = list_entry(list->prev, struct page, lru);
1506 page = list_entry(list->next, struct page, lru);
1508 list_del(&page->lru);
1511 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1513 * __GFP_NOFAIL is not to be used in new code.
1515 * All __GFP_NOFAIL callers should be fixed so that they
1516 * properly detect and handle allocation failures.
1518 * We most definitely don't want callers attempting to
1519 * allocate greater than order-1 page units with
1522 WARN_ON_ONCE(order > 1);
1524 spin_lock_irqsave(&zone->lock, flags);
1525 page = __rmqueue(zone, order, migratetype);
1526 spin_unlock(&zone->lock);
1529 __mod_zone_freepage_state(zone, -(1 << order),
1530 get_pageblock_migratetype(page));
1533 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1534 zone_statistics(preferred_zone, zone, gfp_flags);
1535 local_irq_restore(flags);
1537 VM_BUG_ON(bad_range(zone, page));
1538 if (prep_new_page(page, order, gfp_flags))
1543 local_irq_restore(flags);
1547 #ifdef CONFIG_FAIL_PAGE_ALLOC
1550 struct fault_attr attr;
1552 u32 ignore_gfp_highmem;
1553 u32 ignore_gfp_wait;
1555 } fail_page_alloc = {
1556 .attr = FAULT_ATTR_INITIALIZER,
1557 .ignore_gfp_wait = 1,
1558 .ignore_gfp_highmem = 1,
1562 static int __init setup_fail_page_alloc(char *str)
1564 return setup_fault_attr(&fail_page_alloc.attr, str);
1566 __setup("fail_page_alloc=", setup_fail_page_alloc);
1568 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1570 if (order < fail_page_alloc.min_order)
1572 if (gfp_mask & __GFP_NOFAIL)
1574 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1576 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1579 return should_fail(&fail_page_alloc.attr, 1 << order);
1582 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1584 static int __init fail_page_alloc_debugfs(void)
1586 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1589 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1590 &fail_page_alloc.attr);
1592 return PTR_ERR(dir);
1594 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1595 &fail_page_alloc.ignore_gfp_wait))
1597 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1598 &fail_page_alloc.ignore_gfp_highmem))
1600 if (!debugfs_create_u32("min-order", mode, dir,
1601 &fail_page_alloc.min_order))
1606 debugfs_remove_recursive(dir);
1611 late_initcall(fail_page_alloc_debugfs);
1613 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1615 #else /* CONFIG_FAIL_PAGE_ALLOC */
1617 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1622 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1625 * Return true if free pages are above 'mark'. This takes into account the order
1626 * of the allocation.
1628 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1629 int classzone_idx, int alloc_flags, long free_pages)
1631 /* free_pages my go negative - that's OK */
1633 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1636 free_pages -= (1 << order) - 1;
1637 if (alloc_flags & ALLOC_HIGH)
1639 if (alloc_flags & ALLOC_HARDER)
1642 /* If allocation can't use CMA areas don't use free CMA pages */
1643 if (!(alloc_flags & ALLOC_CMA))
1644 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1646 if (free_pages <= min + lowmem_reserve)
1648 for (o = 0; o < order; o++) {
1649 /* At the next order, this order's pages become unavailable */
1650 free_pages -= z->free_area[o].nr_free << o;
1652 /* Require fewer higher order pages to be free */
1655 if (free_pages <= min)
1661 #ifdef CONFIG_MEMORY_ISOLATION
1662 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1664 if (unlikely(zone->nr_pageblock_isolate))
1665 return zone->nr_pageblock_isolate * pageblock_nr_pages;
1669 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1675 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1676 int classzone_idx, int alloc_flags)
1678 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1679 zone_page_state(z, NR_FREE_PAGES));
1682 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1683 int classzone_idx, int alloc_flags)
1685 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1687 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1688 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1691 * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1692 * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1693 * sleep although it could do so. But this is more desirable for memory
1694 * hotplug than sleeping which can cause a livelock in the direct
1697 free_pages -= nr_zone_isolate_freepages(z);
1698 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1704 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1705 * skip over zones that are not allowed by the cpuset, or that have
1706 * been recently (in last second) found to be nearly full. See further
1707 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1708 * that have to skip over a lot of full or unallowed zones.
1710 * If the zonelist cache is present in the passed in zonelist, then
1711 * returns a pointer to the allowed node mask (either the current
1712 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1714 * If the zonelist cache is not available for this zonelist, does
1715 * nothing and returns NULL.
1717 * If the fullzones BITMAP in the zonelist cache is stale (more than
1718 * a second since last zap'd) then we zap it out (clear its bits.)
1720 * We hold off even calling zlc_setup, until after we've checked the
1721 * first zone in the zonelist, on the theory that most allocations will
1722 * be satisfied from that first zone, so best to examine that zone as
1723 * quickly as we can.
1725 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1727 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1728 nodemask_t *allowednodes; /* zonelist_cache approximation */
1730 zlc = zonelist->zlcache_ptr;
1734 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1735 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1736 zlc->last_full_zap = jiffies;
1739 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1740 &cpuset_current_mems_allowed :
1741 &node_states[N_HIGH_MEMORY];
1742 return allowednodes;
1746 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1747 * if it is worth looking at further for free memory:
1748 * 1) Check that the zone isn't thought to be full (doesn't have its
1749 * bit set in the zonelist_cache fullzones BITMAP).
1750 * 2) Check that the zones node (obtained from the zonelist_cache
1751 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1752 * Return true (non-zero) if zone is worth looking at further, or
1753 * else return false (zero) if it is not.
1755 * This check -ignores- the distinction between various watermarks,
1756 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1757 * found to be full for any variation of these watermarks, it will
1758 * be considered full for up to one second by all requests, unless
1759 * we are so low on memory on all allowed nodes that we are forced
1760 * into the second scan of the zonelist.
1762 * In the second scan we ignore this zonelist cache and exactly
1763 * apply the watermarks to all zones, even it is slower to do so.
1764 * We are low on memory in the second scan, and should leave no stone
1765 * unturned looking for a free page.
1767 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1768 nodemask_t *allowednodes)
1770 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1771 int i; /* index of *z in zonelist zones */
1772 int n; /* node that zone *z is on */
1774 zlc = zonelist->zlcache_ptr;
1778 i = z - zonelist->_zonerefs;
1781 /* This zone is worth trying if it is allowed but not full */
1782 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1786 * Given 'z' scanning a zonelist, set the corresponding bit in
1787 * zlc->fullzones, so that subsequent attempts to allocate a page
1788 * from that zone don't waste time re-examining it.
1790 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1792 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1793 int i; /* index of *z in zonelist zones */
1795 zlc = zonelist->zlcache_ptr;
1799 i = z - zonelist->_zonerefs;
1801 set_bit(i, zlc->fullzones);
1805 * clear all zones full, called after direct reclaim makes progress so that
1806 * a zone that was recently full is not skipped over for up to a second
1808 static void zlc_clear_zones_full(struct zonelist *zonelist)
1810 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1812 zlc = zonelist->zlcache_ptr;
1816 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1819 #else /* CONFIG_NUMA */
1821 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1826 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1827 nodemask_t *allowednodes)
1832 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1836 static void zlc_clear_zones_full(struct zonelist *zonelist)
1839 #endif /* CONFIG_NUMA */
1842 * get_page_from_freelist goes through the zonelist trying to allocate
1845 static struct page *
1846 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1847 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1848 struct zone *preferred_zone, int migratetype)
1851 struct page *page = NULL;
1854 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1855 int zlc_active = 0; /* set if using zonelist_cache */
1856 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1858 classzone_idx = zone_idx(preferred_zone);
1861 * Scan zonelist, looking for a zone with enough free.
1862 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1864 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1865 high_zoneidx, nodemask) {
1866 if (NUMA_BUILD && zlc_active &&
1867 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1869 if ((alloc_flags & ALLOC_CPUSET) &&
1870 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1873 * When allocating a page cache page for writing, we
1874 * want to get it from a zone that is within its dirty
1875 * limit, such that no single zone holds more than its
1876 * proportional share of globally allowed dirty pages.
1877 * The dirty limits take into account the zone's
1878 * lowmem reserves and high watermark so that kswapd
1879 * should be able to balance it without having to
1880 * write pages from its LRU list.
1882 * This may look like it could increase pressure on
1883 * lower zones by failing allocations in higher zones
1884 * before they are full. But the pages that do spill
1885 * over are limited as the lower zones are protected
1886 * by this very same mechanism. It should not become
1887 * a practical burden to them.
1889 * XXX: For now, allow allocations to potentially
1890 * exceed the per-zone dirty limit in the slowpath
1891 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1892 * which is important when on a NUMA setup the allowed
1893 * zones are together not big enough to reach the
1894 * global limit. The proper fix for these situations
1895 * will require awareness of zones in the
1896 * dirty-throttling and the flusher threads.
1898 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1899 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1900 goto this_zone_full;
1902 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1903 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1907 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1908 if (zone_watermark_ok(zone, order, mark,
1909 classzone_idx, alloc_flags))
1912 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1914 * we do zlc_setup if there are multiple nodes
1915 * and before considering the first zone allowed
1918 allowednodes = zlc_setup(zonelist, alloc_flags);
1923 if (zone_reclaim_mode == 0)
1924 goto this_zone_full;
1927 * As we may have just activated ZLC, check if the first
1928 * eligible zone has failed zone_reclaim recently.
1930 if (NUMA_BUILD && zlc_active &&
1931 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1934 ret = zone_reclaim(zone, gfp_mask, order);
1936 case ZONE_RECLAIM_NOSCAN:
1939 case ZONE_RECLAIM_FULL:
1940 /* scanned but unreclaimable */
1943 /* did we reclaim enough */
1944 if (!zone_watermark_ok(zone, order, mark,
1945 classzone_idx, alloc_flags))
1946 goto this_zone_full;
1951 page = buffered_rmqueue(preferred_zone, zone, order,
1952 gfp_mask, migratetype);
1957 zlc_mark_zone_full(zonelist, z);
1960 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1961 /* Disable zlc cache for second zonelist scan */
1968 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1969 * necessary to allocate the page. The expectation is
1970 * that the caller is taking steps that will free more
1971 * memory. The caller should avoid the page being used
1972 * for !PFMEMALLOC purposes.
1974 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1980 * Large machines with many possible nodes should not always dump per-node
1981 * meminfo in irq context.
1983 static inline bool should_suppress_show_mem(void)
1988 ret = in_interrupt();
1993 static DEFINE_RATELIMIT_STATE(nopage_rs,
1994 DEFAULT_RATELIMIT_INTERVAL,
1995 DEFAULT_RATELIMIT_BURST);
1997 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1999 unsigned int filter = SHOW_MEM_FILTER_NODES;
2001 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2002 debug_guardpage_minorder() > 0)
2006 * This documents exceptions given to allocations in certain
2007 * contexts that are allowed to allocate outside current's set
2010 if (!(gfp_mask & __GFP_NOMEMALLOC))
2011 if (test_thread_flag(TIF_MEMDIE) ||
2012 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2013 filter &= ~SHOW_MEM_FILTER_NODES;
2014 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2015 filter &= ~SHOW_MEM_FILTER_NODES;
2018 struct va_format vaf;
2021 va_start(args, fmt);
2026 pr_warn("%pV", &vaf);
2031 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2032 current->comm, order, gfp_mask);
2035 if (!should_suppress_show_mem())
2040 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2041 unsigned long did_some_progress,
2042 unsigned long pages_reclaimed)
2044 /* Do not loop if specifically requested */
2045 if (gfp_mask & __GFP_NORETRY)
2048 /* Always retry if specifically requested */
2049 if (gfp_mask & __GFP_NOFAIL)
2053 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2054 * making forward progress without invoking OOM. Suspend also disables
2055 * storage devices so kswapd will not help. Bail if we are suspending.
2057 if (!did_some_progress && pm_suspended_storage())
2061 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2062 * means __GFP_NOFAIL, but that may not be true in other
2065 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2069 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2070 * specified, then we retry until we no longer reclaim any pages
2071 * (above), or we've reclaimed an order of pages at least as
2072 * large as the allocation's order. In both cases, if the
2073 * allocation still fails, we stop retrying.
2075 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2081 static inline struct page *
2082 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2083 struct zonelist *zonelist, enum zone_type high_zoneidx,
2084 nodemask_t *nodemask, struct zone *preferred_zone,
2089 /* Acquire the OOM killer lock for the zones in zonelist */
2090 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2091 schedule_timeout_uninterruptible(1);
2096 * Go through the zonelist yet one more time, keep very high watermark
2097 * here, this is only to catch a parallel oom killing, we must fail if
2098 * we're still under heavy pressure.
2100 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2101 order, zonelist, high_zoneidx,
2102 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2103 preferred_zone, migratetype);
2107 if (!(gfp_mask & __GFP_NOFAIL)) {
2108 /* The OOM killer will not help higher order allocs */
2109 if (order > PAGE_ALLOC_COSTLY_ORDER)
2111 /* The OOM killer does not needlessly kill tasks for lowmem */
2112 if (high_zoneidx < ZONE_NORMAL)
2115 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2116 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2117 * The caller should handle page allocation failure by itself if
2118 * it specifies __GFP_THISNODE.
2119 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2121 if (gfp_mask & __GFP_THISNODE)
2124 /* Exhausted what can be done so it's blamo time */
2125 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2128 clear_zonelist_oom(zonelist, gfp_mask);
2132 #ifdef CONFIG_COMPACTION
2133 /* Try memory compaction for high-order allocations before reclaim */
2134 static struct page *
2135 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2136 struct zonelist *zonelist, enum zone_type high_zoneidx,
2137 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2138 int migratetype, bool sync_migration,
2139 bool *contended_compaction, bool *deferred_compaction,
2140 unsigned long *did_some_progress)
2142 struct page *page = NULL;
2147 if (compaction_deferred(preferred_zone, order)) {
2148 *deferred_compaction = true;
2152 current->flags |= PF_MEMALLOC;
2153 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2154 nodemask, sync_migration,
2155 contended_compaction, &page);
2156 current->flags &= ~PF_MEMALLOC;
2158 /* If compaction captured a page, prep and use it */
2160 prep_new_page(page, order, gfp_mask);
2164 if (*did_some_progress != COMPACT_SKIPPED) {
2165 /* Page migration frees to the PCP lists but we want merging */
2166 drain_pages(get_cpu());
2169 page = get_page_from_freelist(gfp_mask, nodemask,
2170 order, zonelist, high_zoneidx,
2171 alloc_flags & ~ALLOC_NO_WATERMARKS,
2172 preferred_zone, migratetype);
2175 preferred_zone->compact_considered = 0;
2176 preferred_zone->compact_defer_shift = 0;
2177 if (order >= preferred_zone->compact_order_failed)
2178 preferred_zone->compact_order_failed = order + 1;
2179 count_vm_event(COMPACTSUCCESS);
2184 * It's bad if compaction run occurs and fails.
2185 * The most likely reason is that pages exist,
2186 * but not enough to satisfy watermarks.
2188 count_vm_event(COMPACTFAIL);
2191 * As async compaction considers a subset of pageblocks, only
2192 * defer if the failure was a sync compaction failure.
2195 defer_compaction(preferred_zone, order);
2203 static inline struct page *
2204 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2205 struct zonelist *zonelist, enum zone_type high_zoneidx,
2206 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2207 int migratetype, bool sync_migration,
2208 bool *contended_compaction, bool *deferred_compaction,
2209 unsigned long *did_some_progress)
2213 #endif /* CONFIG_COMPACTION */
2215 /* Perform direct synchronous page reclaim */
2217 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2218 nodemask_t *nodemask)
2220 struct reclaim_state reclaim_state;
2225 /* We now go into synchronous reclaim */
2226 cpuset_memory_pressure_bump();
2227 current->flags |= PF_MEMALLOC;
2228 lockdep_set_current_reclaim_state(gfp_mask);
2229 reclaim_state.reclaimed_slab = 0;
2230 current->reclaim_state = &reclaim_state;
2232 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2234 current->reclaim_state = NULL;
2235 lockdep_clear_current_reclaim_state();
2236 current->flags &= ~PF_MEMALLOC;
2243 /* The really slow allocator path where we enter direct reclaim */
2244 static inline struct page *
2245 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2246 struct zonelist *zonelist, enum zone_type high_zoneidx,
2247 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2248 int migratetype, unsigned long *did_some_progress)
2250 struct page *page = NULL;
2251 bool drained = false;
2253 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2255 if (unlikely(!(*did_some_progress)))
2258 /* After successful reclaim, reconsider all zones for allocation */
2260 zlc_clear_zones_full(zonelist);
2263 page = get_page_from_freelist(gfp_mask, nodemask, order,
2264 zonelist, high_zoneidx,
2265 alloc_flags & ~ALLOC_NO_WATERMARKS,
2266 preferred_zone, migratetype);
2269 * If an allocation failed after direct reclaim, it could be because
2270 * pages are pinned on the per-cpu lists. Drain them and try again
2272 if (!page && !drained) {
2282 * This is called in the allocator slow-path if the allocation request is of
2283 * sufficient urgency to ignore watermarks and take other desperate measures
2285 static inline struct page *
2286 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2287 struct zonelist *zonelist, enum zone_type high_zoneidx,
2288 nodemask_t *nodemask, struct zone *preferred_zone,
2294 page = get_page_from_freelist(gfp_mask, nodemask, order,
2295 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2296 preferred_zone, migratetype);
2298 if (!page && gfp_mask & __GFP_NOFAIL)
2299 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2300 } while (!page && (gfp_mask & __GFP_NOFAIL));
2306 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2307 enum zone_type high_zoneidx,
2308 enum zone_type classzone_idx)
2313 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2314 wakeup_kswapd(zone, order, classzone_idx);
2318 gfp_to_alloc_flags(gfp_t gfp_mask)
2320 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2321 const gfp_t wait = gfp_mask & __GFP_WAIT;
2323 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2324 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2327 * The caller may dip into page reserves a bit more if the caller
2328 * cannot run direct reclaim, or if the caller has realtime scheduling
2329 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2330 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2332 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2336 * Not worth trying to allocate harder for
2337 * __GFP_NOMEMALLOC even if it can't schedule.
2339 if (!(gfp_mask & __GFP_NOMEMALLOC))
2340 alloc_flags |= ALLOC_HARDER;
2342 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2343 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2345 alloc_flags &= ~ALLOC_CPUSET;
2346 } else if (unlikely(rt_task(current)) && !in_interrupt())
2347 alloc_flags |= ALLOC_HARDER;
2349 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2350 if (gfp_mask & __GFP_MEMALLOC)
2351 alloc_flags |= ALLOC_NO_WATERMARKS;
2352 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2353 alloc_flags |= ALLOC_NO_WATERMARKS;
2354 else if (!in_interrupt() &&
2355 ((current->flags & PF_MEMALLOC) ||
2356 unlikely(test_thread_flag(TIF_MEMDIE))))
2357 alloc_flags |= ALLOC_NO_WATERMARKS;
2360 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2361 alloc_flags |= ALLOC_CMA;
2366 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2368 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2371 static inline struct page *
2372 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2373 struct zonelist *zonelist, enum zone_type high_zoneidx,
2374 nodemask_t *nodemask, struct zone *preferred_zone,
2377 const gfp_t wait = gfp_mask & __GFP_WAIT;
2378 struct page *page = NULL;
2380 unsigned long pages_reclaimed = 0;
2381 unsigned long did_some_progress;
2382 bool sync_migration = false;
2383 bool deferred_compaction = false;
2384 bool contended_compaction = false;
2387 * In the slowpath, we sanity check order to avoid ever trying to
2388 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2389 * be using allocators in order of preference for an area that is
2392 if (order >= MAX_ORDER) {
2393 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2398 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2399 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2400 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2401 * using a larger set of nodes after it has established that the
2402 * allowed per node queues are empty and that nodes are
2405 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2409 wake_all_kswapd(order, zonelist, high_zoneidx,
2410 zone_idx(preferred_zone));
2413 * OK, we're below the kswapd watermark and have kicked background
2414 * reclaim. Now things get more complex, so set up alloc_flags according
2415 * to how we want to proceed.
2417 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2420 * Find the true preferred zone if the allocation is unconstrained by
2423 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2424 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2428 /* This is the last chance, in general, before the goto nopage. */
2429 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2430 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2431 preferred_zone, migratetype);
2435 /* Allocate without watermarks if the context allows */
2436 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2438 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2439 * the allocation is high priority and these type of
2440 * allocations are system rather than user orientated
2442 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2444 page = __alloc_pages_high_priority(gfp_mask, order,
2445 zonelist, high_zoneidx, nodemask,
2446 preferred_zone, migratetype);
2452 /* Atomic allocations - we can't balance anything */
2456 /* Avoid recursion of direct reclaim */
2457 if (current->flags & PF_MEMALLOC)
2460 /* Avoid allocations with no watermarks from looping endlessly */
2461 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2465 * Try direct compaction. The first pass is asynchronous. Subsequent
2466 * attempts after direct reclaim are synchronous
2468 page = __alloc_pages_direct_compact(gfp_mask, order,
2469 zonelist, high_zoneidx,
2471 alloc_flags, preferred_zone,
2472 migratetype, sync_migration,
2473 &contended_compaction,
2474 &deferred_compaction,
2475 &did_some_progress);
2478 sync_migration = true;
2481 * If compaction is deferred for high-order allocations, it is because
2482 * sync compaction recently failed. In this is the case and the caller
2483 * requested a movable allocation that does not heavily disrupt the
2484 * system then fail the allocation instead of entering direct reclaim.
2486 if ((deferred_compaction || contended_compaction) &&
2487 (gfp_mask & (__GFP_MOVABLE|__GFP_REPEAT)) == __GFP_MOVABLE)
2490 /* Try direct reclaim and then allocating */
2491 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2492 zonelist, high_zoneidx,
2494 alloc_flags, preferred_zone,
2495 migratetype, &did_some_progress);
2500 * If we failed to make any progress reclaiming, then we are
2501 * running out of options and have to consider going OOM
2503 if (!did_some_progress) {
2504 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2505 if (oom_killer_disabled)
2507 /* Coredumps can quickly deplete all memory reserves */
2508 if ((current->flags & PF_DUMPCORE) &&
2509 !(gfp_mask & __GFP_NOFAIL))
2511 page = __alloc_pages_may_oom(gfp_mask, order,
2512 zonelist, high_zoneidx,
2513 nodemask, preferred_zone,
2518 if (!(gfp_mask & __GFP_NOFAIL)) {
2520 * The oom killer is not called for high-order
2521 * allocations that may fail, so if no progress
2522 * is being made, there are no other options and
2523 * retrying is unlikely to help.
2525 if (order > PAGE_ALLOC_COSTLY_ORDER)
2528 * The oom killer is not called for lowmem
2529 * allocations to prevent needlessly killing
2532 if (high_zoneidx < ZONE_NORMAL)
2540 /* Check if we should retry the allocation */
2541 pages_reclaimed += did_some_progress;
2542 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2544 /* Wait for some write requests to complete then retry */
2545 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2549 * High-order allocations do not necessarily loop after
2550 * direct reclaim and reclaim/compaction depends on compaction
2551 * being called after reclaim so call directly if necessary
2553 page = __alloc_pages_direct_compact(gfp_mask, order,
2554 zonelist, high_zoneidx,
2556 alloc_flags, preferred_zone,
2557 migratetype, sync_migration,
2558 &contended_compaction,
2559 &deferred_compaction,
2560 &did_some_progress);
2566 warn_alloc_failed(gfp_mask, order, NULL);
2569 if (kmemcheck_enabled)
2570 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2576 * This is the 'heart' of the zoned buddy allocator.
2579 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2580 struct zonelist *zonelist, nodemask_t *nodemask)
2582 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2583 struct zone *preferred_zone;
2584 struct page *page = NULL;
2585 int migratetype = allocflags_to_migratetype(gfp_mask);
2586 unsigned int cpuset_mems_cookie;
2587 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2589 gfp_mask &= gfp_allowed_mask;
2591 lockdep_trace_alloc(gfp_mask);
2593 might_sleep_if(gfp_mask & __GFP_WAIT);
2595 if (should_fail_alloc_page(gfp_mask, order))
2599 * Check the zones suitable for the gfp_mask contain at least one
2600 * valid zone. It's possible to have an empty zonelist as a result
2601 * of GFP_THISNODE and a memoryless node
2603 if (unlikely(!zonelist->_zonerefs->zone))
2607 cpuset_mems_cookie = get_mems_allowed();
2609 /* The preferred zone is used for statistics later */
2610 first_zones_zonelist(zonelist, high_zoneidx,
2611 nodemask ? : &cpuset_current_mems_allowed,
2613 if (!preferred_zone)
2617 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2618 alloc_flags |= ALLOC_CMA;
2620 /* First allocation attempt */
2621 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2622 zonelist, high_zoneidx, alloc_flags,
2623 preferred_zone, migratetype);
2624 if (unlikely(!page))
2625 page = __alloc_pages_slowpath(gfp_mask, order,
2626 zonelist, high_zoneidx, nodemask,
2627 preferred_zone, migratetype);
2629 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2633 * When updating a task's mems_allowed, it is possible to race with
2634 * parallel threads in such a way that an allocation can fail while
2635 * the mask is being updated. If a page allocation is about to fail,
2636 * check if the cpuset changed during allocation and if so, retry.
2638 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2643 EXPORT_SYMBOL(__alloc_pages_nodemask);
2646 * Common helper functions.
2648 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2653 * __get_free_pages() returns a 32-bit address, which cannot represent
2656 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2658 page = alloc_pages(gfp_mask, order);
2661 return (unsigned long) page_address(page);
2663 EXPORT_SYMBOL(__get_free_pages);
2665 unsigned long get_zeroed_page(gfp_t gfp_mask)
2667 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2669 EXPORT_SYMBOL(get_zeroed_page);
2671 void __free_pages(struct page *page, unsigned int order)
2673 if (put_page_testzero(page)) {
2675 free_hot_cold_page(page, 0);
2677 __free_pages_ok(page, order);
2681 EXPORT_SYMBOL(__free_pages);
2683 void free_pages(unsigned long addr, unsigned int order)
2686 VM_BUG_ON(!virt_addr_valid((void *)addr));
2687 __free_pages(virt_to_page((void *)addr), order);
2691 EXPORT_SYMBOL(free_pages);
2693 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2696 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2697 unsigned long used = addr + PAGE_ALIGN(size);
2699 split_page(virt_to_page((void *)addr), order);
2700 while (used < alloc_end) {
2705 return (void *)addr;
2709 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2710 * @size: the number of bytes to allocate
2711 * @gfp_mask: GFP flags for the allocation
2713 * This function is similar to alloc_pages(), except that it allocates the
2714 * minimum number of pages to satisfy the request. alloc_pages() can only
2715 * allocate memory in power-of-two pages.
2717 * This function is also limited by MAX_ORDER.
2719 * Memory allocated by this function must be released by free_pages_exact().
2721 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2723 unsigned int order = get_order(size);
2726 addr = __get_free_pages(gfp_mask, order);
2727 return make_alloc_exact(addr, order, size);
2729 EXPORT_SYMBOL(alloc_pages_exact);
2732 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2734 * @nid: the preferred node ID where memory should be allocated
2735 * @size: the number of bytes to allocate
2736 * @gfp_mask: GFP flags for the allocation
2738 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2740 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2743 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2745 unsigned order = get_order(size);
2746 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2749 return make_alloc_exact((unsigned long)page_address(p), order, size);
2751 EXPORT_SYMBOL(alloc_pages_exact_nid);
2754 * free_pages_exact - release memory allocated via alloc_pages_exact()
2755 * @virt: the value returned by alloc_pages_exact.
2756 * @size: size of allocation, same value as passed to alloc_pages_exact().
2758 * Release the memory allocated by a previous call to alloc_pages_exact.
2760 void free_pages_exact(void *virt, size_t size)
2762 unsigned long addr = (unsigned long)virt;
2763 unsigned long end = addr + PAGE_ALIGN(size);
2765 while (addr < end) {
2770 EXPORT_SYMBOL(free_pages_exact);
2772 static unsigned int nr_free_zone_pages(int offset)
2777 /* Just pick one node, since fallback list is circular */
2778 unsigned int sum = 0;
2780 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2782 for_each_zone_zonelist(zone, z, zonelist, offset) {
2783 unsigned long size = zone->present_pages;
2784 unsigned long high = high_wmark_pages(zone);
2793 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2795 unsigned int nr_free_buffer_pages(void)
2797 return nr_free_zone_pages(gfp_zone(GFP_USER));
2799 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2802 * Amount of free RAM allocatable within all zones
2804 unsigned int nr_free_pagecache_pages(void)
2806 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2809 static inline void show_node(struct zone *zone)
2812 printk("Node %d ", zone_to_nid(zone));
2815 void si_meminfo(struct sysinfo *val)
2817 val->totalram = totalram_pages;
2819 val->freeram = global_page_state(NR_FREE_PAGES);
2820 val->bufferram = nr_blockdev_pages();
2821 val->totalhigh = totalhigh_pages;
2822 val->freehigh = nr_free_highpages();
2823 val->mem_unit = PAGE_SIZE;
2826 EXPORT_SYMBOL(si_meminfo);
2829 void si_meminfo_node(struct sysinfo *val, int nid)
2831 pg_data_t *pgdat = NODE_DATA(nid);
2833 val->totalram = pgdat->node_present_pages;
2834 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2835 #ifdef CONFIG_HIGHMEM
2836 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2837 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2843 val->mem_unit = PAGE_SIZE;
2848 * Determine whether the node should be displayed or not, depending on whether
2849 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2851 bool skip_free_areas_node(unsigned int flags, int nid)
2854 unsigned int cpuset_mems_cookie;
2856 if (!(flags & SHOW_MEM_FILTER_NODES))
2860 cpuset_mems_cookie = get_mems_allowed();
2861 ret = !node_isset(nid, cpuset_current_mems_allowed);
2862 } while (!put_mems_allowed(cpuset_mems_cookie));
2867 #define K(x) ((x) << (PAGE_SHIFT-10))
2870 * Show free area list (used inside shift_scroll-lock stuff)
2871 * We also calculate the percentage fragmentation. We do this by counting the
2872 * memory on each free list with the exception of the first item on the list.
2873 * Suppresses nodes that are not allowed by current's cpuset if
2874 * SHOW_MEM_FILTER_NODES is passed.
2876 void show_free_areas(unsigned int filter)
2881 for_each_populated_zone(zone) {
2882 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2885 printk("%s per-cpu:\n", zone->name);
2887 for_each_online_cpu(cpu) {
2888 struct per_cpu_pageset *pageset;
2890 pageset = per_cpu_ptr(zone->pageset, cpu);
2892 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2893 cpu, pageset->pcp.high,
2894 pageset->pcp.batch, pageset->pcp.count);
2898 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2899 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2901 " dirty:%lu writeback:%lu unstable:%lu\n"
2902 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2903 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2905 global_page_state(NR_ACTIVE_ANON),
2906 global_page_state(NR_INACTIVE_ANON),
2907 global_page_state(NR_ISOLATED_ANON),
2908 global_page_state(NR_ACTIVE_FILE),
2909 global_page_state(NR_INACTIVE_FILE),
2910 global_page_state(NR_ISOLATED_FILE),
2911 global_page_state(NR_UNEVICTABLE),
2912 global_page_state(NR_FILE_DIRTY),
2913 global_page_state(NR_WRITEBACK),
2914 global_page_state(NR_UNSTABLE_NFS),
2915 global_page_state(NR_FREE_PAGES),
2916 global_page_state(NR_SLAB_RECLAIMABLE),
2917 global_page_state(NR_SLAB_UNRECLAIMABLE),
2918 global_page_state(NR_FILE_MAPPED),
2919 global_page_state(NR_SHMEM),
2920 global_page_state(NR_PAGETABLE),
2921 global_page_state(NR_BOUNCE),
2922 global_page_state(NR_FREE_CMA_PAGES));
2924 for_each_populated_zone(zone) {
2927 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2935 " active_anon:%lukB"
2936 " inactive_anon:%lukB"
2937 " active_file:%lukB"
2938 " inactive_file:%lukB"
2939 " unevictable:%lukB"
2940 " isolated(anon):%lukB"
2941 " isolated(file):%lukB"
2948 " slab_reclaimable:%lukB"
2949 " slab_unreclaimable:%lukB"
2950 " kernel_stack:%lukB"
2955 " writeback_tmp:%lukB"
2956 " pages_scanned:%lu"
2957 " all_unreclaimable? %s"
2960 K(zone_page_state(zone, NR_FREE_PAGES)),
2961 K(min_wmark_pages(zone)),
2962 K(low_wmark_pages(zone)),
2963 K(high_wmark_pages(zone)),
2964 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2965 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2966 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2967 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2968 K(zone_page_state(zone, NR_UNEVICTABLE)),
2969 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2970 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2971 K(zone->present_pages),
2972 K(zone_page_state(zone, NR_MLOCK)),
2973 K(zone_page_state(zone, NR_FILE_DIRTY)),
2974 K(zone_page_state(zone, NR_WRITEBACK)),
2975 K(zone_page_state(zone, NR_FILE_MAPPED)),
2976 K(zone_page_state(zone, NR_SHMEM)),
2977 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2978 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2979 zone_page_state(zone, NR_KERNEL_STACK) *
2981 K(zone_page_state(zone, NR_PAGETABLE)),
2982 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2983 K(zone_page_state(zone, NR_BOUNCE)),
2984 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
2985 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2986 zone->pages_scanned,
2987 (zone->all_unreclaimable ? "yes" : "no")
2989 printk("lowmem_reserve[]:");
2990 for (i = 0; i < MAX_NR_ZONES; i++)
2991 printk(" %lu", zone->lowmem_reserve[i]);
2995 for_each_populated_zone(zone) {
2996 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2998 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3001 printk("%s: ", zone->name);
3003 spin_lock_irqsave(&zone->lock, flags);
3004 for (order = 0; order < MAX_ORDER; order++) {
3005 nr[order] = zone->free_area[order].nr_free;
3006 total += nr[order] << order;
3008 spin_unlock_irqrestore(&zone->lock, flags);
3009 for (order = 0; order < MAX_ORDER; order++)
3010 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3011 printk("= %lukB\n", K(total));
3014 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3016 show_swap_cache_info();
3019 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3021 zoneref->zone = zone;
3022 zoneref->zone_idx = zone_idx(zone);
3026 * Builds allocation fallback zone lists.
3028 * Add all populated zones of a node to the zonelist.
3030 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3031 int nr_zones, enum zone_type zone_type)
3035 BUG_ON(zone_type >= MAX_NR_ZONES);
3040 zone = pgdat->node_zones + zone_type;
3041 if (populated_zone(zone)) {
3042 zoneref_set_zone(zone,
3043 &zonelist->_zonerefs[nr_zones++]);
3044 check_highest_zone(zone_type);
3047 } while (zone_type);
3054 * 0 = automatic detection of better ordering.
3055 * 1 = order by ([node] distance, -zonetype)
3056 * 2 = order by (-zonetype, [node] distance)
3058 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3059 * the same zonelist. So only NUMA can configure this param.
3061 #define ZONELIST_ORDER_DEFAULT 0
3062 #define ZONELIST_ORDER_NODE 1
3063 #define ZONELIST_ORDER_ZONE 2
3065 /* zonelist order in the kernel.
3066 * set_zonelist_order() will set this to NODE or ZONE.
3068 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3069 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3073 /* The value user specified ....changed by config */
3074 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3075 /* string for sysctl */
3076 #define NUMA_ZONELIST_ORDER_LEN 16
3077 char numa_zonelist_order[16] = "default";
3080 * interface for configure zonelist ordering.
3081 * command line option "numa_zonelist_order"
3082 * = "[dD]efault - default, automatic configuration.
3083 * = "[nN]ode - order by node locality, then by zone within node
3084 * = "[zZ]one - order by zone, then by locality within zone
3087 static int __parse_numa_zonelist_order(char *s)
3089 if (*s == 'd' || *s == 'D') {
3090 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3091 } else if (*s == 'n' || *s == 'N') {
3092 user_zonelist_order = ZONELIST_ORDER_NODE;
3093 } else if (*s == 'z' || *s == 'Z') {
3094 user_zonelist_order = ZONELIST_ORDER_ZONE;
3097 "Ignoring invalid numa_zonelist_order value: "
3104 static __init int setup_numa_zonelist_order(char *s)
3111 ret = __parse_numa_zonelist_order(s);
3113 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3117 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3120 * sysctl handler for numa_zonelist_order
3122 int numa_zonelist_order_handler(ctl_table *table, int write,
3123 void __user *buffer, size_t *length,
3126 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3128 static DEFINE_MUTEX(zl_order_mutex);
3130 mutex_lock(&zl_order_mutex);
3132 strcpy(saved_string, (char*)table->data);
3133 ret = proc_dostring(table, write, buffer, length, ppos);
3137 int oldval = user_zonelist_order;
3138 if (__parse_numa_zonelist_order((char*)table->data)) {
3140 * bogus value. restore saved string
3142 strncpy((char*)table->data, saved_string,
3143 NUMA_ZONELIST_ORDER_LEN);
3144 user_zonelist_order = oldval;
3145 } else if (oldval != user_zonelist_order) {
3146 mutex_lock(&zonelists_mutex);
3147 build_all_zonelists(NULL, NULL);
3148 mutex_unlock(&zonelists_mutex);
3152 mutex_unlock(&zl_order_mutex);
3157 #define MAX_NODE_LOAD (nr_online_nodes)
3158 static int node_load[MAX_NUMNODES];
3161 * find_next_best_node - find the next node that should appear in a given node's fallback list
3162 * @node: node whose fallback list we're appending
3163 * @used_node_mask: nodemask_t of already used nodes
3165 * We use a number of factors to determine which is the next node that should
3166 * appear on a given node's fallback list. The node should not have appeared
3167 * already in @node's fallback list, and it should be the next closest node
3168 * according to the distance array (which contains arbitrary distance values
3169 * from each node to each node in the system), and should also prefer nodes
3170 * with no CPUs, since presumably they'll have very little allocation pressure
3171 * on them otherwise.
3172 * It returns -1 if no node is found.
3174 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3177 int min_val = INT_MAX;
3179 const struct cpumask *tmp = cpumask_of_node(0);
3181 /* Use the local node if we haven't already */
3182 if (!node_isset(node, *used_node_mask)) {
3183 node_set(node, *used_node_mask);
3187 for_each_node_state(n, N_HIGH_MEMORY) {
3189 /* Don't want a node to appear more than once */
3190 if (node_isset(n, *used_node_mask))
3193 /* Use the distance array to find the distance */
3194 val = node_distance(node, n);
3196 /* Penalize nodes under us ("prefer the next node") */
3199 /* Give preference to headless and unused nodes */
3200 tmp = cpumask_of_node(n);
3201 if (!cpumask_empty(tmp))
3202 val += PENALTY_FOR_NODE_WITH_CPUS;
3204 /* Slight preference for less loaded node */
3205 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3206 val += node_load[n];
3208 if (val < min_val) {
3215 node_set(best_node, *used_node_mask);
3222 * Build zonelists ordered by node and zones within node.
3223 * This results in maximum locality--normal zone overflows into local
3224 * DMA zone, if any--but risks exhausting DMA zone.
3226 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3229 struct zonelist *zonelist;
3231 zonelist = &pgdat->node_zonelists[0];
3232 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3234 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3236 zonelist->_zonerefs[j].zone = NULL;
3237 zonelist->_zonerefs[j].zone_idx = 0;
3241 * Build gfp_thisnode zonelists
3243 static void build_thisnode_zonelists(pg_data_t *pgdat)
3246 struct zonelist *zonelist;
3248 zonelist = &pgdat->node_zonelists[1];
3249 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3250 zonelist->_zonerefs[j].zone = NULL;
3251 zonelist->_zonerefs[j].zone_idx = 0;
3255 * Build zonelists ordered by zone and nodes within zones.
3256 * This results in conserving DMA zone[s] until all Normal memory is
3257 * exhausted, but results in overflowing to remote node while memory
3258 * may still exist in local DMA zone.
3260 static int node_order[MAX_NUMNODES];
3262 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3265 int zone_type; /* needs to be signed */
3267 struct zonelist *zonelist;
3269 zonelist = &pgdat->node_zonelists[0];
3271 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3272 for (j = 0; j < nr_nodes; j++) {
3273 node = node_order[j];
3274 z = &NODE_DATA(node)->node_zones[zone_type];
3275 if (populated_zone(z)) {
3277 &zonelist->_zonerefs[pos++]);
3278 check_highest_zone(zone_type);
3282 zonelist->_zonerefs[pos].zone = NULL;
3283 zonelist->_zonerefs[pos].zone_idx = 0;
3286 static int default_zonelist_order(void)
3289 unsigned long low_kmem_size,total_size;
3293 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3294 * If they are really small and used heavily, the system can fall
3295 * into OOM very easily.
3296 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3298 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3301 for_each_online_node(nid) {
3302 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3303 z = &NODE_DATA(nid)->node_zones[zone_type];
3304 if (populated_zone(z)) {
3305 if (zone_type < ZONE_NORMAL)
3306 low_kmem_size += z->present_pages;
3307 total_size += z->present_pages;
3308 } else if (zone_type == ZONE_NORMAL) {
3310 * If any node has only lowmem, then node order
3311 * is preferred to allow kernel allocations
3312 * locally; otherwise, they can easily infringe
3313 * on other nodes when there is an abundance of
3314 * lowmem available to allocate from.
3316 return ZONELIST_ORDER_NODE;
3320 if (!low_kmem_size || /* there are no DMA area. */
3321 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3322 return ZONELIST_ORDER_NODE;
3324 * look into each node's config.
3325 * If there is a node whose DMA/DMA32 memory is very big area on
3326 * local memory, NODE_ORDER may be suitable.
3328 average_size = total_size /
3329 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3330 for_each_online_node(nid) {
3333 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3334 z = &NODE_DATA(nid)->node_zones[zone_type];
3335 if (populated_zone(z)) {
3336 if (zone_type < ZONE_NORMAL)
3337 low_kmem_size += z->present_pages;
3338 total_size += z->present_pages;
3341 if (low_kmem_size &&
3342 total_size > average_size && /* ignore small node */
3343 low_kmem_size > total_size * 70/100)
3344 return ZONELIST_ORDER_NODE;
3346 return ZONELIST_ORDER_ZONE;
3349 static void set_zonelist_order(void)
3351 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3352 current_zonelist_order = default_zonelist_order();
3354 current_zonelist_order = user_zonelist_order;
3357 static void build_zonelists(pg_data_t *pgdat)
3361 nodemask_t used_mask;
3362 int local_node, prev_node;
3363 struct zonelist *zonelist;
3364 int order = current_zonelist_order;
3366 /* initialize zonelists */
3367 for (i = 0; i < MAX_ZONELISTS; i++) {
3368 zonelist = pgdat->node_zonelists + i;
3369 zonelist->_zonerefs[0].zone = NULL;
3370 zonelist->_zonerefs[0].zone_idx = 0;
3373 /* NUMA-aware ordering of nodes */
3374 local_node = pgdat->node_id;
3375 load = nr_online_nodes;
3376 prev_node = local_node;
3377 nodes_clear(used_mask);
3379 memset(node_order, 0, sizeof(node_order));
3382 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3383 int distance = node_distance(local_node, node);
3386 * If another node is sufficiently far away then it is better
3387 * to reclaim pages in a zone before going off node.
3389 if (distance > RECLAIM_DISTANCE)
3390 zone_reclaim_mode = 1;
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 (distance != node_distance(local_node, prev_node))
3398 node_load[node] = load;
3402 if (order == ZONELIST_ORDER_NODE)
3403 build_zonelists_in_node_order(pgdat, node);
3405 node_order[j++] = node; /* remember order */
3408 if (order == ZONELIST_ORDER_ZONE) {
3409 /* calculate node order -- i.e., DMA last! */
3410 build_zonelists_in_zone_order(pgdat, j);
3413 build_thisnode_zonelists(pgdat);
3416 /* Construct the zonelist performance cache - see further mmzone.h */
3417 static void build_zonelist_cache(pg_data_t *pgdat)
3419 struct zonelist *zonelist;
3420 struct zonelist_cache *zlc;
3423 zonelist = &pgdat->node_zonelists[0];
3424 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3425 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3426 for (z = zonelist->_zonerefs; z->zone; z++)
3427 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3430 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3432 * Return node id of node used for "local" allocations.
3433 * I.e., first node id of first zone in arg node's generic zonelist.
3434 * Used for initializing percpu 'numa_mem', which is used primarily
3435 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3437 int local_memory_node(int node)
3441 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3442 gfp_zone(GFP_KERNEL),
3449 #else /* CONFIG_NUMA */
3451 static void set_zonelist_order(void)
3453 current_zonelist_order = ZONELIST_ORDER_ZONE;
3456 static void build_zonelists(pg_data_t *pgdat)
3458 int node, local_node;
3460 struct zonelist *zonelist;
3462 local_node = pgdat->node_id;
3464 zonelist = &pgdat->node_zonelists[0];
3465 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3468 * Now we build the zonelist so that it contains the zones
3469 * of all the other nodes.
3470 * We don't want to pressure a particular node, so when
3471 * building the zones for node N, we make sure that the
3472 * zones coming right after the local ones are those from
3473 * node N+1 (modulo N)
3475 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3476 if (!node_online(node))
3478 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3481 for (node = 0; node < local_node; node++) {
3482 if (!node_online(node))
3484 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3488 zonelist->_zonerefs[j].zone = NULL;
3489 zonelist->_zonerefs[j].zone_idx = 0;
3492 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3493 static void build_zonelist_cache(pg_data_t *pgdat)
3495 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3498 #endif /* CONFIG_NUMA */
3501 * Boot pageset table. One per cpu which is going to be used for all
3502 * zones and all nodes. The parameters will be set in such a way
3503 * that an item put on a list will immediately be handed over to
3504 * the buddy list. This is safe since pageset manipulation is done
3505 * with interrupts disabled.
3507 * The boot_pagesets must be kept even after bootup is complete for
3508 * unused processors and/or zones. They do play a role for bootstrapping
3509 * hotplugged processors.
3511 * zoneinfo_show() and maybe other functions do
3512 * not check if the processor is online before following the pageset pointer.
3513 * Other parts of the kernel may not check if the zone is available.
3515 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3516 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3517 static void setup_zone_pageset(struct zone *zone);
3520 * Global mutex to protect against size modification of zonelists
3521 * as well as to serialize pageset setup for the new populated zone.
3523 DEFINE_MUTEX(zonelists_mutex);
3525 /* return values int ....just for stop_machine() */
3526 static int __build_all_zonelists(void *data)
3530 pg_data_t *self = data;
3533 memset(node_load, 0, sizeof(node_load));
3536 if (self && !node_online(self->node_id)) {
3537 build_zonelists(self);
3538 build_zonelist_cache(self);
3541 for_each_online_node(nid) {
3542 pg_data_t *pgdat = NODE_DATA(nid);
3544 build_zonelists(pgdat);
3545 build_zonelist_cache(pgdat);
3549 * Initialize the boot_pagesets that are going to be used
3550 * for bootstrapping processors. The real pagesets for
3551 * each zone will be allocated later when the per cpu
3552 * allocator is available.
3554 * boot_pagesets are used also for bootstrapping offline
3555 * cpus if the system is already booted because the pagesets
3556 * are needed to initialize allocators on a specific cpu too.
3557 * F.e. the percpu allocator needs the page allocator which
3558 * needs the percpu allocator in order to allocate its pagesets
3559 * (a chicken-egg dilemma).
3561 for_each_possible_cpu(cpu) {
3562 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3564 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3566 * We now know the "local memory node" for each node--
3567 * i.e., the node of the first zone in the generic zonelist.
3568 * Set up numa_mem percpu variable for on-line cpus. During
3569 * boot, only the boot cpu should be on-line; we'll init the
3570 * secondary cpus' numa_mem as they come on-line. During
3571 * node/memory hotplug, we'll fixup all on-line cpus.
3573 if (cpu_online(cpu))
3574 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3582 * Called with zonelists_mutex held always
3583 * unless system_state == SYSTEM_BOOTING.
3585 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3587 set_zonelist_order();
3589 if (system_state == SYSTEM_BOOTING) {
3590 __build_all_zonelists(NULL);
3591 mminit_verify_zonelist();
3592 cpuset_init_current_mems_allowed();
3594 /* we have to stop all cpus to guarantee there is no user
3596 #ifdef CONFIG_MEMORY_HOTPLUG
3598 setup_zone_pageset(zone);
3600 stop_machine(__build_all_zonelists, pgdat, NULL);
3601 /* cpuset refresh routine should be here */
3603 vm_total_pages = nr_free_pagecache_pages();
3605 * Disable grouping by mobility if the number of pages in the
3606 * system is too low to allow the mechanism to work. It would be
3607 * more accurate, but expensive to check per-zone. This check is
3608 * made on memory-hotadd so a system can start with mobility
3609 * disabled and enable it later
3611 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3612 page_group_by_mobility_disabled = 1;
3614 page_group_by_mobility_disabled = 0;
3616 printk("Built %i zonelists in %s order, mobility grouping %s. "
3617 "Total pages: %ld\n",
3619 zonelist_order_name[current_zonelist_order],
3620 page_group_by_mobility_disabled ? "off" : "on",
3623 printk("Policy zone: %s\n", zone_names[policy_zone]);
3628 * Helper functions to size the waitqueue hash table.
3629 * Essentially these want to choose hash table sizes sufficiently
3630 * large so that collisions trying to wait on pages are rare.
3631 * But in fact, the number of active page waitqueues on typical
3632 * systems is ridiculously low, less than 200. So this is even
3633 * conservative, even though it seems large.
3635 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3636 * waitqueues, i.e. the size of the waitq table given the number of pages.
3638 #define PAGES_PER_WAITQUEUE 256
3640 #ifndef CONFIG_MEMORY_HOTPLUG
3641 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3643 unsigned long size = 1;
3645 pages /= PAGES_PER_WAITQUEUE;
3647 while (size < pages)
3651 * Once we have dozens or even hundreds of threads sleeping
3652 * on IO we've got bigger problems than wait queue collision.
3653 * Limit the size of the wait table to a reasonable size.
3655 size = min(size, 4096UL);
3657 return max(size, 4UL);
3661 * A zone's size might be changed by hot-add, so it is not possible to determine
3662 * a suitable size for its wait_table. So we use the maximum size now.
3664 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3666 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3667 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3668 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3670 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3671 * or more by the traditional way. (See above). It equals:
3673 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3674 * ia64(16K page size) : = ( 8G + 4M)byte.
3675 * powerpc (64K page size) : = (32G +16M)byte.
3677 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3684 * This is an integer logarithm so that shifts can be used later
3685 * to extract the more random high bits from the multiplicative
3686 * hash function before the remainder is taken.
3688 static inline unsigned long wait_table_bits(unsigned long size)
3693 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3696 * Check if a pageblock contains reserved pages
3698 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3702 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3703 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3710 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3711 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3712 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3713 * higher will lead to a bigger reserve which will get freed as contiguous
3714 * blocks as reclaim kicks in
3716 static void setup_zone_migrate_reserve(struct zone *zone)
3718 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3720 unsigned long block_migratetype;
3724 * Get the start pfn, end pfn and the number of blocks to reserve
3725 * We have to be careful to be aligned to pageblock_nr_pages to
3726 * make sure that we always check pfn_valid for the first page in
3729 start_pfn = zone->zone_start_pfn;
3730 end_pfn = start_pfn + zone->spanned_pages;
3731 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3732 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3736 * Reserve blocks are generally in place to help high-order atomic
3737 * allocations that are short-lived. A min_free_kbytes value that
3738 * would result in more than 2 reserve blocks for atomic allocations
3739 * is assumed to be in place to help anti-fragmentation for the
3740 * future allocation of hugepages at runtime.
3742 reserve = min(2, reserve);
3744 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3745 if (!pfn_valid(pfn))
3747 page = pfn_to_page(pfn);
3749 /* Watch out for overlapping nodes */
3750 if (page_to_nid(page) != zone_to_nid(zone))
3753 block_migratetype = get_pageblock_migratetype(page);
3755 /* Only test what is necessary when the reserves are not met */
3758 * Blocks with reserved pages will never free, skip
3761 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3762 if (pageblock_is_reserved(pfn, block_end_pfn))
3765 /* If this block is reserved, account for it */
3766 if (block_migratetype == MIGRATE_RESERVE) {
3771 /* Suitable for reserving if this block is movable */
3772 if (block_migratetype == MIGRATE_MOVABLE) {
3773 set_pageblock_migratetype(page,
3775 move_freepages_block(zone, page,
3783 * If the reserve is met and this is a previous reserved block,
3786 if (block_migratetype == MIGRATE_RESERVE) {
3787 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3788 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3794 * Initially all pages are reserved - free ones are freed
3795 * up by free_all_bootmem() once the early boot process is
3796 * done. Non-atomic initialization, single-pass.
3798 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3799 unsigned long start_pfn, enum memmap_context context)
3802 unsigned long end_pfn = start_pfn + size;
3806 if (highest_memmap_pfn < end_pfn - 1)
3807 highest_memmap_pfn = end_pfn - 1;
3809 z = &NODE_DATA(nid)->node_zones[zone];
3810 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3812 * There can be holes in boot-time mem_map[]s
3813 * handed to this function. They do not
3814 * exist on hotplugged memory.
3816 if (context == MEMMAP_EARLY) {
3817 if (!early_pfn_valid(pfn))
3819 if (!early_pfn_in_nid(pfn, nid))
3822 page = pfn_to_page(pfn);
3823 set_page_links(page, zone, nid, pfn);
3824 mminit_verify_page_links(page, zone, nid, pfn);
3825 init_page_count(page);
3826 reset_page_mapcount(page);
3827 SetPageReserved(page);
3829 * Mark the block movable so that blocks are reserved for
3830 * movable at startup. This will force kernel allocations
3831 * to reserve their blocks rather than leaking throughout
3832 * the address space during boot when many long-lived
3833 * kernel allocations are made. Later some blocks near
3834 * the start are marked MIGRATE_RESERVE by
3835 * setup_zone_migrate_reserve()
3837 * bitmap is created for zone's valid pfn range. but memmap
3838 * can be created for invalid pages (for alignment)
3839 * check here not to call set_pageblock_migratetype() against
3842 if ((z->zone_start_pfn <= pfn)
3843 && (pfn < z->zone_start_pfn + z->spanned_pages)
3844 && !(pfn & (pageblock_nr_pages - 1)))
3845 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3847 INIT_LIST_HEAD(&page->lru);
3848 #ifdef WANT_PAGE_VIRTUAL
3849 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3850 if (!is_highmem_idx(zone))
3851 set_page_address(page, __va(pfn << PAGE_SHIFT));
3856 static void __meminit zone_init_free_lists(struct zone *zone)
3859 for_each_migratetype_order(order, t) {
3860 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3861 zone->free_area[order].nr_free = 0;
3865 #ifndef __HAVE_ARCH_MEMMAP_INIT
3866 #define memmap_init(size, nid, zone, start_pfn) \
3867 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3870 static int __meminit zone_batchsize(struct zone *zone)
3876 * The per-cpu-pages pools are set to around 1000th of the
3877 * size of the zone. But no more than 1/2 of a meg.
3879 * OK, so we don't know how big the cache is. So guess.
3881 batch = zone->present_pages / 1024;
3882 if (batch * PAGE_SIZE > 512 * 1024)
3883 batch = (512 * 1024) / PAGE_SIZE;
3884 batch /= 4; /* We effectively *= 4 below */
3889 * Clamp the batch to a 2^n - 1 value. Having a power
3890 * of 2 value was found to be more likely to have
3891 * suboptimal cache aliasing properties in some cases.
3893 * For example if 2 tasks are alternately allocating
3894 * batches of pages, one task can end up with a lot
3895 * of pages of one half of the possible page colors
3896 * and the other with pages of the other colors.
3898 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3903 /* The deferral and batching of frees should be suppressed under NOMMU
3906 * The problem is that NOMMU needs to be able to allocate large chunks
3907 * of contiguous memory as there's no hardware page translation to
3908 * assemble apparent contiguous memory from discontiguous pages.
3910 * Queueing large contiguous runs of pages for batching, however,
3911 * causes the pages to actually be freed in smaller chunks. As there
3912 * can be a significant delay between the individual batches being
3913 * recycled, this leads to the once large chunks of space being
3914 * fragmented and becoming unavailable for high-order allocations.
3920 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3922 struct per_cpu_pages *pcp;
3925 memset(p, 0, sizeof(*p));
3929 pcp->high = 6 * batch;
3930 pcp->batch = max(1UL, 1 * batch);
3931 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3932 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3936 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3937 * to the value high for the pageset p.
3940 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3943 struct per_cpu_pages *pcp;
3947 pcp->batch = max(1UL, high/4);
3948 if ((high/4) > (PAGE_SHIFT * 8))
3949 pcp->batch = PAGE_SHIFT * 8;
3952 static void __meminit setup_zone_pageset(struct zone *zone)
3956 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3958 for_each_possible_cpu(cpu) {
3959 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3961 setup_pageset(pcp, zone_batchsize(zone));
3963 if (percpu_pagelist_fraction)
3964 setup_pagelist_highmark(pcp,
3965 (zone->present_pages /
3966 percpu_pagelist_fraction));
3971 * Allocate per cpu pagesets and initialize them.
3972 * Before this call only boot pagesets were available.
3974 void __init setup_per_cpu_pageset(void)
3978 for_each_populated_zone(zone)
3979 setup_zone_pageset(zone);
3982 static noinline __init_refok
3983 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3986 struct pglist_data *pgdat = zone->zone_pgdat;
3990 * The per-page waitqueue mechanism uses hashed waitqueues
3993 zone->wait_table_hash_nr_entries =
3994 wait_table_hash_nr_entries(zone_size_pages);
3995 zone->wait_table_bits =
3996 wait_table_bits(zone->wait_table_hash_nr_entries);
3997 alloc_size = zone->wait_table_hash_nr_entries
3998 * sizeof(wait_queue_head_t);
4000 if (!slab_is_available()) {
4001 zone->wait_table = (wait_queue_head_t *)
4002 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4005 * This case means that a zone whose size was 0 gets new memory
4006 * via memory hot-add.
4007 * But it may be the case that a new node was hot-added. In
4008 * this case vmalloc() will not be able to use this new node's
4009 * memory - this wait_table must be initialized to use this new
4010 * node itself as well.
4011 * To use this new node's memory, further consideration will be
4014 zone->wait_table = vmalloc(alloc_size);
4016 if (!zone->wait_table)
4019 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4020 init_waitqueue_head(zone->wait_table + i);
4025 static __meminit void zone_pcp_init(struct zone *zone)
4028 * per cpu subsystem is not up at this point. The following code
4029 * relies on the ability of the linker to provide the
4030 * offset of a (static) per cpu variable into the per cpu area.
4032 zone->pageset = &boot_pageset;
4034 if (zone->present_pages)
4035 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4036 zone->name, zone->present_pages,
4037 zone_batchsize(zone));
4040 int __meminit init_currently_empty_zone(struct zone *zone,
4041 unsigned long zone_start_pfn,
4043 enum memmap_context context)
4045 struct pglist_data *pgdat = zone->zone_pgdat;
4047 ret = zone_wait_table_init(zone, size);
4050 pgdat->nr_zones = zone_idx(zone) + 1;
4052 zone->zone_start_pfn = zone_start_pfn;
4054 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4055 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4057 (unsigned long)zone_idx(zone),
4058 zone_start_pfn, (zone_start_pfn + size));
4060 zone_init_free_lists(zone);
4065 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4066 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4068 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4069 * Architectures may implement their own version but if add_active_range()
4070 * was used and there are no special requirements, this is a convenient
4073 int __meminit __early_pfn_to_nid(unsigned long pfn)
4075 unsigned long start_pfn, end_pfn;
4078 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4079 if (start_pfn <= pfn && pfn < end_pfn)
4081 /* This is a memory hole */
4084 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4086 int __meminit early_pfn_to_nid(unsigned long pfn)
4090 nid = __early_pfn_to_nid(pfn);
4093 /* just returns 0 */
4097 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4098 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4102 nid = __early_pfn_to_nid(pfn);
4103 if (nid >= 0 && nid != node)
4110 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4111 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4112 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4114 * If an architecture guarantees that all ranges registered with
4115 * add_active_ranges() contain no holes and may be freed, this
4116 * this function may be used instead of calling free_bootmem() manually.
4118 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4120 unsigned long start_pfn, end_pfn;
4123 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4124 start_pfn = min(start_pfn, max_low_pfn);
4125 end_pfn = min(end_pfn, max_low_pfn);
4127 if (start_pfn < end_pfn)
4128 free_bootmem_node(NODE_DATA(this_nid),
4129 PFN_PHYS(start_pfn),
4130 (end_pfn - start_pfn) << PAGE_SHIFT);
4135 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4136 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4138 * If an architecture guarantees that all ranges registered with
4139 * add_active_ranges() contain no holes and may be freed, this
4140 * function may be used instead of calling memory_present() manually.
4142 void __init sparse_memory_present_with_active_regions(int nid)
4144 unsigned long start_pfn, end_pfn;
4147 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4148 memory_present(this_nid, start_pfn, end_pfn);
4152 * get_pfn_range_for_nid - Return the start and end page frames for a node
4153 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4154 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4155 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4157 * It returns the start and end page frame of a node based on information
4158 * provided by an arch calling add_active_range(). If called for a node
4159 * with no available memory, a warning is printed and the start and end
4162 void __meminit get_pfn_range_for_nid(unsigned int nid,
4163 unsigned long *start_pfn, unsigned long *end_pfn)
4165 unsigned long this_start_pfn, this_end_pfn;
4171 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4172 *start_pfn = min(*start_pfn, this_start_pfn);
4173 *end_pfn = max(*end_pfn, this_end_pfn);
4176 if (*start_pfn == -1UL)
4181 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4182 * assumption is made that zones within a node are ordered in monotonic
4183 * increasing memory addresses so that the "highest" populated zone is used
4185 static void __init find_usable_zone_for_movable(void)
4188 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4189 if (zone_index == ZONE_MOVABLE)
4192 if (arch_zone_highest_possible_pfn[zone_index] >
4193 arch_zone_lowest_possible_pfn[zone_index])
4197 VM_BUG_ON(zone_index == -1);
4198 movable_zone = zone_index;
4202 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4203 * because it is sized independent of architecture. Unlike the other zones,
4204 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4205 * in each node depending on the size of each node and how evenly kernelcore
4206 * is distributed. This helper function adjusts the zone ranges
4207 * provided by the architecture for a given node by using the end of the
4208 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4209 * zones within a node are in order of monotonic increases memory addresses
4211 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4212 unsigned long zone_type,
4213 unsigned long node_start_pfn,
4214 unsigned long node_end_pfn,
4215 unsigned long *zone_start_pfn,
4216 unsigned long *zone_end_pfn)
4218 /* Only adjust if ZONE_MOVABLE is on this node */
4219 if (zone_movable_pfn[nid]) {
4220 /* Size ZONE_MOVABLE */
4221 if (zone_type == ZONE_MOVABLE) {
4222 *zone_start_pfn = zone_movable_pfn[nid];
4223 *zone_end_pfn = min(node_end_pfn,
4224 arch_zone_highest_possible_pfn[movable_zone]);
4226 /* Adjust for ZONE_MOVABLE starting within this range */
4227 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4228 *zone_end_pfn > zone_movable_pfn[nid]) {
4229 *zone_end_pfn = zone_movable_pfn[nid];
4231 /* Check if this whole range is within ZONE_MOVABLE */
4232 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4233 *zone_start_pfn = *zone_end_pfn;
4238 * Return the number of pages a zone spans in a node, including holes
4239 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4241 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4242 unsigned long zone_type,
4243 unsigned long *ignored)
4245 unsigned long node_start_pfn, node_end_pfn;
4246 unsigned long zone_start_pfn, zone_end_pfn;
4248 /* Get the start and end of the node and zone */
4249 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4250 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4251 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4252 adjust_zone_range_for_zone_movable(nid, zone_type,
4253 node_start_pfn, node_end_pfn,
4254 &zone_start_pfn, &zone_end_pfn);
4256 /* Check that this node has pages within the zone's required range */
4257 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4260 /* Move the zone boundaries inside the node if necessary */
4261 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4262 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4264 /* Return the spanned pages */
4265 return zone_end_pfn - zone_start_pfn;
4269 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4270 * then all holes in the requested range will be accounted for.
4272 unsigned long __meminit __absent_pages_in_range(int nid,
4273 unsigned long range_start_pfn,
4274 unsigned long range_end_pfn)
4276 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4277 unsigned long start_pfn, end_pfn;
4280 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4281 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4282 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4283 nr_absent -= end_pfn - start_pfn;
4289 * absent_pages_in_range - Return number of page frames in holes within a range
4290 * @start_pfn: The start PFN to start searching for holes
4291 * @end_pfn: The end PFN to stop searching for holes
4293 * It returns the number of pages frames in memory holes within a range.
4295 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4296 unsigned long end_pfn)
4298 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4301 /* Return the number of page frames in holes in a zone on a node */
4302 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4303 unsigned long zone_type,
4304 unsigned long *ignored)
4306 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4307 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4308 unsigned long node_start_pfn, node_end_pfn;
4309 unsigned long zone_start_pfn, zone_end_pfn;
4311 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4312 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4313 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4315 adjust_zone_range_for_zone_movable(nid, zone_type,
4316 node_start_pfn, node_end_pfn,
4317 &zone_start_pfn, &zone_end_pfn);
4318 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4321 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4322 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4323 unsigned long zone_type,
4324 unsigned long *zones_size)
4326 return zones_size[zone_type];
4329 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4330 unsigned long zone_type,
4331 unsigned long *zholes_size)
4336 return zholes_size[zone_type];
4339 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4341 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4342 unsigned long *zones_size, unsigned long *zholes_size)
4344 unsigned long realtotalpages, totalpages = 0;
4347 for (i = 0; i < MAX_NR_ZONES; i++)
4348 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4350 pgdat->node_spanned_pages = totalpages;
4352 realtotalpages = totalpages;
4353 for (i = 0; i < MAX_NR_ZONES; i++)
4355 zone_absent_pages_in_node(pgdat->node_id, i,
4357 pgdat->node_present_pages = realtotalpages;
4358 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4362 #ifndef CONFIG_SPARSEMEM
4364 * Calculate the size of the zone->blockflags rounded to an unsigned long
4365 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4366 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4367 * round what is now in bits to nearest long in bits, then return it in
4370 static unsigned long __init usemap_size(unsigned long zonesize)
4372 unsigned long usemapsize;
4374 usemapsize = roundup(zonesize, pageblock_nr_pages);
4375 usemapsize = usemapsize >> pageblock_order;
4376 usemapsize *= NR_PAGEBLOCK_BITS;
4377 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4379 return usemapsize / 8;
4382 static void __init setup_usemap(struct pglist_data *pgdat,
4383 struct zone *zone, unsigned long zonesize)
4385 unsigned long usemapsize = usemap_size(zonesize);
4386 zone->pageblock_flags = NULL;
4388 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4392 static inline void setup_usemap(struct pglist_data *pgdat,
4393 struct zone *zone, unsigned long zonesize) {}
4394 #endif /* CONFIG_SPARSEMEM */
4396 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4398 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4399 void __init set_pageblock_order(void)
4403 /* Check that pageblock_nr_pages has not already been setup */
4404 if (pageblock_order)
4407 if (HPAGE_SHIFT > PAGE_SHIFT)
4408 order = HUGETLB_PAGE_ORDER;
4410 order = MAX_ORDER - 1;
4413 * Assume the largest contiguous order of interest is a huge page.
4414 * This value may be variable depending on boot parameters on IA64 and
4417 pageblock_order = order;
4419 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4422 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4423 * is unused as pageblock_order is set at compile-time. See
4424 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4427 void __init set_pageblock_order(void)
4431 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4434 * Set up the zone data structures:
4435 * - mark all pages reserved
4436 * - mark all memory queues empty
4437 * - clear the memory bitmaps
4439 * NOTE: pgdat should get zeroed by caller.
4441 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4442 unsigned long *zones_size, unsigned long *zholes_size)
4445 int nid = pgdat->node_id;
4446 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4449 pgdat_resize_init(pgdat);
4450 init_waitqueue_head(&pgdat->kswapd_wait);
4451 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4452 pgdat_page_cgroup_init(pgdat);
4454 for (j = 0; j < MAX_NR_ZONES; j++) {
4455 struct zone *zone = pgdat->node_zones + j;
4456 unsigned long size, realsize, memmap_pages;
4458 size = zone_spanned_pages_in_node(nid, j, zones_size);
4459 realsize = size - zone_absent_pages_in_node(nid, j,
4463 * Adjust realsize so that it accounts for how much memory
4464 * is used by this zone for memmap. This affects the watermark
4465 * and per-cpu initialisations
4468 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4469 if (realsize >= memmap_pages) {
4470 realsize -= memmap_pages;
4473 " %s zone: %lu pages used for memmap\n",
4474 zone_names[j], memmap_pages);
4477 " %s zone: %lu pages exceeds realsize %lu\n",
4478 zone_names[j], memmap_pages, realsize);
4480 /* Account for reserved pages */
4481 if (j == 0 && realsize > dma_reserve) {
4482 realsize -= dma_reserve;
4483 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4484 zone_names[0], dma_reserve);
4487 if (!is_highmem_idx(j))
4488 nr_kernel_pages += realsize;
4489 nr_all_pages += realsize;
4491 zone->spanned_pages = size;
4492 zone->present_pages = realsize;
4493 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
4494 zone->compact_cached_free_pfn = zone->zone_start_pfn +
4495 zone->spanned_pages;
4496 zone->compact_cached_free_pfn &= ~(pageblock_nr_pages-1);
4500 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4502 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4504 zone->name = zone_names[j];
4505 spin_lock_init(&zone->lock);
4506 spin_lock_init(&zone->lru_lock);
4507 zone_seqlock_init(zone);
4508 zone->zone_pgdat = pgdat;
4510 zone_pcp_init(zone);
4511 lruvec_init(&zone->lruvec, zone);
4515 set_pageblock_order();
4516 setup_usemap(pgdat, zone, size);
4517 ret = init_currently_empty_zone(zone, zone_start_pfn,
4518 size, MEMMAP_EARLY);
4520 memmap_init(size, nid, j, zone_start_pfn);
4521 zone_start_pfn += size;
4525 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4527 /* Skip empty nodes */
4528 if (!pgdat->node_spanned_pages)
4531 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4532 /* ia64 gets its own node_mem_map, before this, without bootmem */
4533 if (!pgdat->node_mem_map) {
4534 unsigned long size, start, end;
4538 * The zone's endpoints aren't required to be MAX_ORDER
4539 * aligned but the node_mem_map endpoints must be in order
4540 * for the buddy allocator to function correctly.
4542 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4543 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4544 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4545 size = (end - start) * sizeof(struct page);
4546 map = alloc_remap(pgdat->node_id, size);
4548 map = alloc_bootmem_node_nopanic(pgdat, size);
4549 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4551 #ifndef CONFIG_NEED_MULTIPLE_NODES
4553 * With no DISCONTIG, the global mem_map is just set as node 0's
4555 if (pgdat == NODE_DATA(0)) {
4556 mem_map = NODE_DATA(0)->node_mem_map;
4557 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4558 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4559 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4560 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4563 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4566 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4567 unsigned long node_start_pfn, unsigned long *zholes_size)
4569 pg_data_t *pgdat = NODE_DATA(nid);
4571 /* pg_data_t should be reset to zero when it's allocated */
4572 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4574 pgdat->node_id = nid;
4575 pgdat->node_start_pfn = node_start_pfn;
4576 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4578 alloc_node_mem_map(pgdat);
4579 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4580 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4581 nid, (unsigned long)pgdat,
4582 (unsigned long)pgdat->node_mem_map);
4585 free_area_init_core(pgdat, zones_size, zholes_size);
4588 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4590 #if MAX_NUMNODES > 1
4592 * Figure out the number of possible node ids.
4594 static void __init setup_nr_node_ids(void)
4597 unsigned int highest = 0;
4599 for_each_node_mask(node, node_possible_map)
4601 nr_node_ids = highest + 1;
4604 static inline void setup_nr_node_ids(void)
4610 * node_map_pfn_alignment - determine the maximum internode alignment
4612 * This function should be called after node map is populated and sorted.
4613 * It calculates the maximum power of two alignment which can distinguish
4616 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4617 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4618 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4619 * shifted, 1GiB is enough and this function will indicate so.
4621 * This is used to test whether pfn -> nid mapping of the chosen memory
4622 * model has fine enough granularity to avoid incorrect mapping for the
4623 * populated node map.
4625 * Returns the determined alignment in pfn's. 0 if there is no alignment
4626 * requirement (single node).
4628 unsigned long __init node_map_pfn_alignment(void)
4630 unsigned long accl_mask = 0, last_end = 0;
4631 unsigned long start, end, mask;
4635 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4636 if (!start || last_nid < 0 || last_nid == nid) {
4643 * Start with a mask granular enough to pin-point to the
4644 * start pfn and tick off bits one-by-one until it becomes
4645 * too coarse to separate the current node from the last.
4647 mask = ~((1 << __ffs(start)) - 1);
4648 while (mask && last_end <= (start & (mask << 1)))
4651 /* accumulate all internode masks */
4655 /* convert mask to number of pages */
4656 return ~accl_mask + 1;
4659 /* Find the lowest pfn for a node */
4660 static unsigned long __init find_min_pfn_for_node(int nid)
4662 unsigned long min_pfn = ULONG_MAX;
4663 unsigned long start_pfn;
4666 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4667 min_pfn = min(min_pfn, start_pfn);
4669 if (min_pfn == ULONG_MAX) {
4671 "Could not find start_pfn for node %d\n", nid);
4679 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4681 * It returns the minimum PFN based on information provided via
4682 * add_active_range().
4684 unsigned long __init find_min_pfn_with_active_regions(void)
4686 return find_min_pfn_for_node(MAX_NUMNODES);
4690 * early_calculate_totalpages()
4691 * Sum pages in active regions for movable zone.
4692 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4694 static unsigned long __init early_calculate_totalpages(void)
4696 unsigned long totalpages = 0;
4697 unsigned long start_pfn, end_pfn;
4700 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4701 unsigned long pages = end_pfn - start_pfn;
4703 totalpages += pages;
4705 node_set_state(nid, N_HIGH_MEMORY);
4711 * Find the PFN the Movable zone begins in each node. Kernel memory
4712 * is spread evenly between nodes as long as the nodes have enough
4713 * memory. When they don't, some nodes will have more kernelcore than
4716 static void __init find_zone_movable_pfns_for_nodes(void)
4719 unsigned long usable_startpfn;
4720 unsigned long kernelcore_node, kernelcore_remaining;
4721 /* save the state before borrow the nodemask */
4722 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4723 unsigned long totalpages = early_calculate_totalpages();
4724 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4727 * If movablecore was specified, calculate what size of
4728 * kernelcore that corresponds so that memory usable for
4729 * any allocation type is evenly spread. If both kernelcore
4730 * and movablecore are specified, then the value of kernelcore
4731 * will be used for required_kernelcore if it's greater than
4732 * what movablecore would have allowed.
4734 if (required_movablecore) {
4735 unsigned long corepages;
4738 * Round-up so that ZONE_MOVABLE is at least as large as what
4739 * was requested by the user
4741 required_movablecore =
4742 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4743 corepages = totalpages - required_movablecore;
4745 required_kernelcore = max(required_kernelcore, corepages);
4748 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4749 if (!required_kernelcore)
4752 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4753 find_usable_zone_for_movable();
4754 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4757 /* Spread kernelcore memory as evenly as possible throughout nodes */
4758 kernelcore_node = required_kernelcore / usable_nodes;
4759 for_each_node_state(nid, N_HIGH_MEMORY) {
4760 unsigned long start_pfn, end_pfn;
4763 * Recalculate kernelcore_node if the division per node
4764 * now exceeds what is necessary to satisfy the requested
4765 * amount of memory for the kernel
4767 if (required_kernelcore < kernelcore_node)
4768 kernelcore_node = required_kernelcore / usable_nodes;
4771 * As the map is walked, we track how much memory is usable
4772 * by the kernel using kernelcore_remaining. When it is
4773 * 0, the rest of the node is usable by ZONE_MOVABLE
4775 kernelcore_remaining = kernelcore_node;
4777 /* Go through each range of PFNs within this node */
4778 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4779 unsigned long size_pages;
4781 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4782 if (start_pfn >= end_pfn)
4785 /* Account for what is only usable for kernelcore */
4786 if (start_pfn < usable_startpfn) {
4787 unsigned long kernel_pages;
4788 kernel_pages = min(end_pfn, usable_startpfn)
4791 kernelcore_remaining -= min(kernel_pages,
4792 kernelcore_remaining);
4793 required_kernelcore -= min(kernel_pages,
4794 required_kernelcore);
4796 /* Continue if range is now fully accounted */
4797 if (end_pfn <= usable_startpfn) {
4800 * Push zone_movable_pfn to the end so
4801 * that if we have to rebalance
4802 * kernelcore across nodes, we will
4803 * not double account here
4805 zone_movable_pfn[nid] = end_pfn;
4808 start_pfn = usable_startpfn;
4812 * The usable PFN range for ZONE_MOVABLE is from
4813 * start_pfn->end_pfn. Calculate size_pages as the
4814 * number of pages used as kernelcore
4816 size_pages = end_pfn - start_pfn;
4817 if (size_pages > kernelcore_remaining)
4818 size_pages = kernelcore_remaining;
4819 zone_movable_pfn[nid] = start_pfn + size_pages;
4822 * Some kernelcore has been met, update counts and
4823 * break if the kernelcore for this node has been
4826 required_kernelcore -= min(required_kernelcore,
4828 kernelcore_remaining -= size_pages;
4829 if (!kernelcore_remaining)
4835 * If there is still required_kernelcore, we do another pass with one
4836 * less node in the count. This will push zone_movable_pfn[nid] further
4837 * along on the nodes that still have memory until kernelcore is
4841 if (usable_nodes && required_kernelcore > usable_nodes)
4844 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4845 for (nid = 0; nid < MAX_NUMNODES; nid++)
4846 zone_movable_pfn[nid] =
4847 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4850 /* restore the node_state */
4851 node_states[N_HIGH_MEMORY] = saved_node_state;
4854 /* Any regular memory on that node ? */
4855 static void __init check_for_regular_memory(pg_data_t *pgdat)
4857 #ifdef CONFIG_HIGHMEM
4858 enum zone_type zone_type;
4860 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4861 struct zone *zone = &pgdat->node_zones[zone_type];
4862 if (zone->present_pages) {
4863 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4871 * free_area_init_nodes - Initialise all pg_data_t and zone data
4872 * @max_zone_pfn: an array of max PFNs for each zone
4874 * This will call free_area_init_node() for each active node in the system.
4875 * Using the page ranges provided by add_active_range(), the size of each
4876 * zone in each node and their holes is calculated. If the maximum PFN
4877 * between two adjacent zones match, it is assumed that the zone is empty.
4878 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4879 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4880 * starts where the previous one ended. For example, ZONE_DMA32 starts
4881 * at arch_max_dma_pfn.
4883 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4885 unsigned long start_pfn, end_pfn;
4888 /* Record where the zone boundaries are */
4889 memset(arch_zone_lowest_possible_pfn, 0,
4890 sizeof(arch_zone_lowest_possible_pfn));
4891 memset(arch_zone_highest_possible_pfn, 0,
4892 sizeof(arch_zone_highest_possible_pfn));
4893 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4894 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4895 for (i = 1; i < MAX_NR_ZONES; i++) {
4896 if (i == ZONE_MOVABLE)
4898 arch_zone_lowest_possible_pfn[i] =
4899 arch_zone_highest_possible_pfn[i-1];
4900 arch_zone_highest_possible_pfn[i] =
4901 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4903 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4904 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4906 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4907 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4908 find_zone_movable_pfns_for_nodes();
4910 /* Print out the zone ranges */
4911 printk("Zone ranges:\n");
4912 for (i = 0; i < MAX_NR_ZONES; i++) {
4913 if (i == ZONE_MOVABLE)
4915 printk(KERN_CONT " %-8s ", zone_names[i]);
4916 if (arch_zone_lowest_possible_pfn[i] ==
4917 arch_zone_highest_possible_pfn[i])
4918 printk(KERN_CONT "empty\n");
4920 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4921 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4922 (arch_zone_highest_possible_pfn[i]
4923 << PAGE_SHIFT) - 1);
4926 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4927 printk("Movable zone start for each node\n");
4928 for (i = 0; i < MAX_NUMNODES; i++) {
4929 if (zone_movable_pfn[i])
4930 printk(" Node %d: %#010lx\n", i,
4931 zone_movable_pfn[i] << PAGE_SHIFT);
4934 /* Print out the early_node_map[] */
4935 printk("Early memory node ranges\n");
4936 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4937 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
4938 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
4940 /* Initialise every node */
4941 mminit_verify_pageflags_layout();
4942 setup_nr_node_ids();
4943 for_each_online_node(nid) {
4944 pg_data_t *pgdat = NODE_DATA(nid);
4945 free_area_init_node(nid, NULL,
4946 find_min_pfn_for_node(nid), NULL);
4948 /* Any memory on that node */
4949 if (pgdat->node_present_pages)
4950 node_set_state(nid, N_HIGH_MEMORY);
4951 check_for_regular_memory(pgdat);
4955 static int __init cmdline_parse_core(char *p, unsigned long *core)
4957 unsigned long long coremem;
4961 coremem = memparse(p, &p);
4962 *core = coremem >> PAGE_SHIFT;
4964 /* Paranoid check that UL is enough for the coremem value */
4965 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4971 * kernelcore=size sets the amount of memory for use for allocations that
4972 * cannot be reclaimed or migrated.
4974 static int __init cmdline_parse_kernelcore(char *p)
4976 return cmdline_parse_core(p, &required_kernelcore);
4980 * movablecore=size sets the amount of memory for use for allocations that
4981 * can be reclaimed or migrated.
4983 static int __init cmdline_parse_movablecore(char *p)
4985 return cmdline_parse_core(p, &required_movablecore);
4988 early_param("kernelcore", cmdline_parse_kernelcore);
4989 early_param("movablecore", cmdline_parse_movablecore);
4991 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4994 * set_dma_reserve - set the specified number of pages reserved in the first zone
4995 * @new_dma_reserve: The number of pages to mark reserved
4997 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4998 * In the DMA zone, a significant percentage may be consumed by kernel image
4999 * and other unfreeable allocations which can skew the watermarks badly. This
5000 * function may optionally be used to account for unfreeable pages in the
5001 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5002 * smaller per-cpu batchsize.
5004 void __init set_dma_reserve(unsigned long new_dma_reserve)
5006 dma_reserve = new_dma_reserve;
5009 void __init free_area_init(unsigned long *zones_size)
5011 free_area_init_node(0, zones_size,
5012 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5015 static int page_alloc_cpu_notify(struct notifier_block *self,
5016 unsigned long action, void *hcpu)
5018 int cpu = (unsigned long)hcpu;
5020 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5021 lru_add_drain_cpu(cpu);
5025 * Spill the event counters of the dead processor
5026 * into the current processors event counters.
5027 * This artificially elevates the count of the current
5030 vm_events_fold_cpu(cpu);
5033 * Zero the differential counters of the dead processor
5034 * so that the vm statistics are consistent.
5036 * This is only okay since the processor is dead and cannot
5037 * race with what we are doing.
5039 refresh_cpu_vm_stats(cpu);
5044 void __init page_alloc_init(void)
5046 hotcpu_notifier(page_alloc_cpu_notify, 0);
5050 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5051 * or min_free_kbytes changes.
5053 static void calculate_totalreserve_pages(void)
5055 struct pglist_data *pgdat;
5056 unsigned long reserve_pages = 0;
5057 enum zone_type i, j;
5059 for_each_online_pgdat(pgdat) {
5060 for (i = 0; i < MAX_NR_ZONES; i++) {
5061 struct zone *zone = pgdat->node_zones + i;
5062 unsigned long max = 0;
5064 /* Find valid and maximum lowmem_reserve in the zone */
5065 for (j = i; j < MAX_NR_ZONES; j++) {
5066 if (zone->lowmem_reserve[j] > max)
5067 max = zone->lowmem_reserve[j];
5070 /* we treat the high watermark as reserved pages. */
5071 max += high_wmark_pages(zone);
5073 if (max > zone->present_pages)
5074 max = zone->present_pages;
5075 reserve_pages += max;
5077 * Lowmem reserves are not available to
5078 * GFP_HIGHUSER page cache allocations and
5079 * kswapd tries to balance zones to their high
5080 * watermark. As a result, neither should be
5081 * regarded as dirtyable memory, to prevent a
5082 * situation where reclaim has to clean pages
5083 * in order to balance the zones.
5085 zone->dirty_balance_reserve = max;
5088 dirty_balance_reserve = reserve_pages;
5089 totalreserve_pages = reserve_pages;
5093 * setup_per_zone_lowmem_reserve - called whenever
5094 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5095 * has a correct pages reserved value, so an adequate number of
5096 * pages are left in the zone after a successful __alloc_pages().
5098 static void setup_per_zone_lowmem_reserve(void)
5100 struct pglist_data *pgdat;
5101 enum zone_type j, idx;
5103 for_each_online_pgdat(pgdat) {
5104 for (j = 0; j < MAX_NR_ZONES; j++) {
5105 struct zone *zone = pgdat->node_zones + j;
5106 unsigned long present_pages = zone->present_pages;
5108 zone->lowmem_reserve[j] = 0;
5112 struct zone *lower_zone;
5116 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5117 sysctl_lowmem_reserve_ratio[idx] = 1;
5119 lower_zone = pgdat->node_zones + idx;
5120 lower_zone->lowmem_reserve[j] = present_pages /
5121 sysctl_lowmem_reserve_ratio[idx];
5122 present_pages += lower_zone->present_pages;
5127 /* update totalreserve_pages */
5128 calculate_totalreserve_pages();
5131 static void __setup_per_zone_wmarks(void)
5133 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5134 unsigned long lowmem_pages = 0;
5136 unsigned long flags;
5138 /* Calculate total number of !ZONE_HIGHMEM pages */
5139 for_each_zone(zone) {
5140 if (!is_highmem(zone))
5141 lowmem_pages += zone->present_pages;
5144 for_each_zone(zone) {
5147 spin_lock_irqsave(&zone->lock, flags);
5148 tmp = (u64)pages_min * zone->present_pages;
5149 do_div(tmp, lowmem_pages);
5150 if (is_highmem(zone)) {
5152 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5153 * need highmem pages, so cap pages_min to a small
5156 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5157 * deltas controls asynch page reclaim, and so should
5158 * not be capped for highmem.
5162 min_pages = zone->present_pages / 1024;
5163 if (min_pages < SWAP_CLUSTER_MAX)
5164 min_pages = SWAP_CLUSTER_MAX;
5165 if (min_pages > 128)
5167 zone->watermark[WMARK_MIN] = min_pages;
5170 * If it's a lowmem zone, reserve a number of pages
5171 * proportionate to the zone's size.
5173 zone->watermark[WMARK_MIN] = tmp;
5176 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5177 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5179 zone->watermark[WMARK_MIN] += cma_wmark_pages(zone);
5180 zone->watermark[WMARK_LOW] += cma_wmark_pages(zone);
5181 zone->watermark[WMARK_HIGH] += cma_wmark_pages(zone);
5183 setup_zone_migrate_reserve(zone);
5184 spin_unlock_irqrestore(&zone->lock, flags);
5187 /* update totalreserve_pages */
5188 calculate_totalreserve_pages();
5192 * setup_per_zone_wmarks - called when min_free_kbytes changes
5193 * or when memory is hot-{added|removed}
5195 * Ensures that the watermark[min,low,high] values for each zone are set
5196 * correctly with respect to min_free_kbytes.
5198 void setup_per_zone_wmarks(void)
5200 mutex_lock(&zonelists_mutex);
5201 __setup_per_zone_wmarks();
5202 mutex_unlock(&zonelists_mutex);
5206 * The inactive anon list should be small enough that the VM never has to
5207 * do too much work, but large enough that each inactive page has a chance
5208 * to be referenced again before it is swapped out.
5210 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5211 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5212 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5213 * the anonymous pages are kept on the inactive list.
5216 * memory ratio inactive anon
5217 * -------------------------------------
5226 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5228 unsigned int gb, ratio;
5230 /* Zone size in gigabytes */
5231 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5233 ratio = int_sqrt(10 * gb);
5237 zone->inactive_ratio = ratio;
5240 static void __meminit setup_per_zone_inactive_ratio(void)
5245 calculate_zone_inactive_ratio(zone);
5249 * Initialise min_free_kbytes.
5251 * For small machines we want it small (128k min). For large machines
5252 * we want it large (64MB max). But it is not linear, because network
5253 * bandwidth does not increase linearly with machine size. We use
5255 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5256 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5272 int __meminit init_per_zone_wmark_min(void)
5274 unsigned long lowmem_kbytes;
5276 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5278 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5279 if (min_free_kbytes < 128)
5280 min_free_kbytes = 128;
5281 if (min_free_kbytes > 65536)
5282 min_free_kbytes = 65536;
5283 setup_per_zone_wmarks();
5284 refresh_zone_stat_thresholds();
5285 setup_per_zone_lowmem_reserve();
5286 setup_per_zone_inactive_ratio();
5289 module_init(init_per_zone_wmark_min)
5292 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5293 * that we can call two helper functions whenever min_free_kbytes
5296 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5297 void __user *buffer, size_t *length, loff_t *ppos)
5299 proc_dointvec(table, write, buffer, length, ppos);
5301 setup_per_zone_wmarks();
5306 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5307 void __user *buffer, size_t *length, loff_t *ppos)
5312 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5317 zone->min_unmapped_pages = (zone->present_pages *
5318 sysctl_min_unmapped_ratio) / 100;
5322 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5323 void __user *buffer, size_t *length, loff_t *ppos)
5328 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5333 zone->min_slab_pages = (zone->present_pages *
5334 sysctl_min_slab_ratio) / 100;
5340 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5341 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5342 * whenever sysctl_lowmem_reserve_ratio changes.
5344 * The reserve ratio obviously has absolutely no relation with the
5345 * minimum watermarks. The lowmem reserve ratio can only make sense
5346 * if in function of the boot time zone sizes.
5348 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5349 void __user *buffer, size_t *length, loff_t *ppos)
5351 proc_dointvec_minmax(table, write, buffer, length, ppos);
5352 setup_per_zone_lowmem_reserve();
5357 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5358 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5359 * can have before it gets flushed back to buddy allocator.
5362 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5363 void __user *buffer, size_t *length, loff_t *ppos)
5369 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5370 if (!write || (ret < 0))
5372 for_each_populated_zone(zone) {
5373 for_each_possible_cpu(cpu) {
5375 high = zone->present_pages / percpu_pagelist_fraction;
5376 setup_pagelist_highmark(
5377 per_cpu_ptr(zone->pageset, cpu), high);
5383 int hashdist = HASHDIST_DEFAULT;
5386 static int __init set_hashdist(char *str)
5390 hashdist = simple_strtoul(str, &str, 0);
5393 __setup("hashdist=", set_hashdist);
5397 * allocate a large system hash table from bootmem
5398 * - it is assumed that the hash table must contain an exact power-of-2
5399 * quantity of entries
5400 * - limit is the number of hash buckets, not the total allocation size
5402 void *__init alloc_large_system_hash(const char *tablename,
5403 unsigned long bucketsize,
5404 unsigned long numentries,
5407 unsigned int *_hash_shift,
5408 unsigned int *_hash_mask,
5409 unsigned long low_limit,
5410 unsigned long high_limit)
5412 unsigned long long max = high_limit;
5413 unsigned long log2qty, size;
5416 /* allow the kernel cmdline to have a say */
5418 /* round applicable memory size up to nearest megabyte */
5419 numentries = nr_kernel_pages;
5420 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5421 numentries >>= 20 - PAGE_SHIFT;
5422 numentries <<= 20 - PAGE_SHIFT;
5424 /* limit to 1 bucket per 2^scale bytes of low memory */
5425 if (scale > PAGE_SHIFT)
5426 numentries >>= (scale - PAGE_SHIFT);
5428 numentries <<= (PAGE_SHIFT - scale);
5430 /* Make sure we've got at least a 0-order allocation.. */
5431 if (unlikely(flags & HASH_SMALL)) {
5432 /* Makes no sense without HASH_EARLY */
5433 WARN_ON(!(flags & HASH_EARLY));
5434 if (!(numentries >> *_hash_shift)) {
5435 numentries = 1UL << *_hash_shift;
5436 BUG_ON(!numentries);
5438 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5439 numentries = PAGE_SIZE / bucketsize;
5441 numentries = roundup_pow_of_two(numentries);
5443 /* limit allocation size to 1/16 total memory by default */
5445 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5446 do_div(max, bucketsize);
5448 max = min(max, 0x80000000ULL);
5450 if (numentries < low_limit)
5451 numentries = low_limit;
5452 if (numentries > max)
5455 log2qty = ilog2(numentries);
5458 size = bucketsize << log2qty;
5459 if (flags & HASH_EARLY)
5460 table = alloc_bootmem_nopanic(size);
5462 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5465 * If bucketsize is not a power-of-two, we may free
5466 * some pages at the end of hash table which
5467 * alloc_pages_exact() automatically does
5469 if (get_order(size) < MAX_ORDER) {
5470 table = alloc_pages_exact(size, GFP_ATOMIC);
5471 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5474 } while (!table && size > PAGE_SIZE && --log2qty);
5477 panic("Failed to allocate %s hash table\n", tablename);
5479 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5482 ilog2(size) - PAGE_SHIFT,
5486 *_hash_shift = log2qty;
5488 *_hash_mask = (1 << log2qty) - 1;
5493 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5494 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5497 #ifdef CONFIG_SPARSEMEM
5498 return __pfn_to_section(pfn)->pageblock_flags;
5500 return zone->pageblock_flags;
5501 #endif /* CONFIG_SPARSEMEM */
5504 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5506 #ifdef CONFIG_SPARSEMEM
5507 pfn &= (PAGES_PER_SECTION-1);
5508 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5510 pfn = pfn - zone->zone_start_pfn;
5511 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5512 #endif /* CONFIG_SPARSEMEM */
5516 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5517 * @page: The page within the block of interest
5518 * @start_bitidx: The first bit of interest to retrieve
5519 * @end_bitidx: The last bit of interest
5520 * returns pageblock_bits flags
5522 unsigned long get_pageblock_flags_group(struct page *page,
5523 int start_bitidx, int end_bitidx)
5526 unsigned long *bitmap;
5527 unsigned long pfn, bitidx;
5528 unsigned long flags = 0;
5529 unsigned long value = 1;
5531 zone = page_zone(page);
5532 pfn = page_to_pfn(page);
5533 bitmap = get_pageblock_bitmap(zone, pfn);
5534 bitidx = pfn_to_bitidx(zone, pfn);
5536 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5537 if (test_bit(bitidx + start_bitidx, bitmap))
5544 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5545 * @page: The page within the block of interest
5546 * @start_bitidx: The first bit of interest
5547 * @end_bitidx: The last bit of interest
5548 * @flags: The flags to set
5550 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5551 int start_bitidx, int end_bitidx)
5554 unsigned long *bitmap;
5555 unsigned long pfn, bitidx;
5556 unsigned long value = 1;
5558 zone = page_zone(page);
5559 pfn = page_to_pfn(page);
5560 bitmap = get_pageblock_bitmap(zone, pfn);
5561 bitidx = pfn_to_bitidx(zone, pfn);
5562 VM_BUG_ON(pfn < zone->zone_start_pfn);
5563 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5565 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5567 __set_bit(bitidx + start_bitidx, bitmap);
5569 __clear_bit(bitidx + start_bitidx, bitmap);
5573 * This function checks whether pageblock includes unmovable pages or not.
5574 * If @count is not zero, it is okay to include less @count unmovable pages
5576 * PageLRU check wihtout isolation or lru_lock could race so that
5577 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5578 * expect this function should be exact.
5580 bool has_unmovable_pages(struct zone *zone, struct page *page, int count)
5582 unsigned long pfn, iter, found;
5586 * For avoiding noise data, lru_add_drain_all() should be called
5587 * If ZONE_MOVABLE, the zone never contains unmovable pages
5589 if (zone_idx(zone) == ZONE_MOVABLE)
5591 mt = get_pageblock_migratetype(page);
5592 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5595 pfn = page_to_pfn(page);
5596 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5597 unsigned long check = pfn + iter;
5599 if (!pfn_valid_within(check))
5602 page = pfn_to_page(check);
5604 * We can't use page_count without pin a page
5605 * because another CPU can free compound page.
5606 * This check already skips compound tails of THP
5607 * because their page->_count is zero at all time.
5609 if (!atomic_read(&page->_count)) {
5610 if (PageBuddy(page))
5611 iter += (1 << page_order(page)) - 1;
5618 * If there are RECLAIMABLE pages, we need to check it.
5619 * But now, memory offline itself doesn't call shrink_slab()
5620 * and it still to be fixed.
5623 * If the page is not RAM, page_count()should be 0.
5624 * we don't need more check. This is an _used_ not-movable page.
5626 * The problematic thing here is PG_reserved pages. PG_reserved
5627 * is set to both of a memory hole page and a _used_ kernel
5636 bool is_pageblock_removable_nolock(struct page *page)
5642 * We have to be careful here because we are iterating over memory
5643 * sections which are not zone aware so we might end up outside of
5644 * the zone but still within the section.
5645 * We have to take care about the node as well. If the node is offline
5646 * its NODE_DATA will be NULL - see page_zone.
5648 if (!node_online(page_to_nid(page)))
5651 zone = page_zone(page);
5652 pfn = page_to_pfn(page);
5653 if (zone->zone_start_pfn > pfn ||
5654 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5657 return !has_unmovable_pages(zone, page, 0);
5662 static unsigned long pfn_max_align_down(unsigned long pfn)
5664 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5665 pageblock_nr_pages) - 1);
5668 static unsigned long pfn_max_align_up(unsigned long pfn)
5670 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5671 pageblock_nr_pages));
5674 static struct page *
5675 __alloc_contig_migrate_alloc(struct page *page, unsigned long private,
5678 gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
5680 if (PageHighMem(page))
5681 gfp_mask |= __GFP_HIGHMEM;
5683 return alloc_page(gfp_mask);
5686 /* [start, end) must belong to a single zone. */
5687 static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5689 /* This function is based on compact_zone() from compaction.c. */
5691 unsigned long pfn = start;
5692 unsigned int tries = 0;
5695 struct compact_control cc = {
5696 .nr_migratepages = 0,
5698 .zone = page_zone(pfn_to_page(start)),
5701 INIT_LIST_HEAD(&cc.migratepages);
5703 migrate_prep_local();
5705 while (pfn < end || !list_empty(&cc.migratepages)) {
5706 if (fatal_signal_pending(current)) {
5711 if (list_empty(&cc.migratepages)) {
5712 cc.nr_migratepages = 0;
5713 pfn = isolate_migratepages_range(cc.zone, &cc,
5720 } else if (++tries == 5) {
5721 ret = ret < 0 ? ret : -EBUSY;
5725 reclaim_clean_pages_from_list(cc.zone, &cc.migratepages);
5727 ret = migrate_pages(&cc.migratepages,
5728 __alloc_contig_migrate_alloc,
5729 0, false, MIGRATE_SYNC);
5732 putback_lru_pages(&cc.migratepages);
5733 return ret > 0 ? 0 : ret;
5737 * Update zone's cma pages counter used for watermark level calculation.
5739 static inline void __update_cma_watermarks(struct zone *zone, int count)
5741 unsigned long flags;
5742 spin_lock_irqsave(&zone->lock, flags);
5743 zone->min_cma_pages += count;
5744 spin_unlock_irqrestore(&zone->lock, flags);
5745 setup_per_zone_wmarks();
5749 * Trigger memory pressure bump to reclaim some pages in order to be able to
5750 * allocate 'count' pages in single page units. Does similar work as
5751 *__alloc_pages_slowpath() function.
5753 static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
5755 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
5756 struct zonelist *zonelist = node_zonelist(0, gfp_mask);
5757 int did_some_progress = 0;
5761 * Increase level of watermarks to force kswapd do his job
5762 * to stabilise at new watermark level.
5764 __update_cma_watermarks(zone, count);
5766 /* Obey watermarks as if the page was being allocated */
5767 while (!zone_watermark_ok(zone, 0, low_wmark_pages(zone), 0, 0)) {
5768 wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));
5770 did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
5772 if (!did_some_progress) {
5773 /* Exhausted what can be done so it's blamo time */
5774 out_of_memory(zonelist, gfp_mask, order, NULL, false);
5778 /* Restore original watermark levels. */
5779 __update_cma_watermarks(zone, -count);
5785 * alloc_contig_range() -- tries to allocate given range of pages
5786 * @start: start PFN to allocate
5787 * @end: one-past-the-last PFN to allocate
5788 * @migratetype: migratetype of the underlaying pageblocks (either
5789 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5790 * in range must have the same migratetype and it must
5791 * be either of the two.
5793 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5794 * aligned, however it's the caller's responsibility to guarantee that
5795 * we are the only thread that changes migrate type of pageblocks the
5798 * The PFN range must belong to a single zone.
5800 * Returns zero on success or negative error code. On success all
5801 * pages which PFN is in [start, end) are allocated for the caller and
5802 * need to be freed with free_contig_range().
5804 int alloc_contig_range(unsigned long start, unsigned long end,
5805 unsigned migratetype)
5807 struct zone *zone = page_zone(pfn_to_page(start));
5808 unsigned long outer_start, outer_end;
5812 * What we do here is we mark all pageblocks in range as
5813 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5814 * have different sizes, and due to the way page allocator
5815 * work, we align the range to biggest of the two pages so
5816 * that page allocator won't try to merge buddies from
5817 * different pageblocks and change MIGRATE_ISOLATE to some
5818 * other migration type.
5820 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5821 * migrate the pages from an unaligned range (ie. pages that
5822 * we are interested in). This will put all the pages in
5823 * range back to page allocator as MIGRATE_ISOLATE.
5825 * When this is done, we take the pages in range from page
5826 * allocator removing them from the buddy system. This way
5827 * page allocator will never consider using them.
5829 * This lets us mark the pageblocks back as
5830 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5831 * aligned range but not in the unaligned, original range are
5832 * put back to page allocator so that buddy can use them.
5835 ret = start_isolate_page_range(pfn_max_align_down(start),
5836 pfn_max_align_up(end), migratetype);
5840 ret = __alloc_contig_migrate_range(start, end);
5845 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5846 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5847 * more, all pages in [start, end) are free in page allocator.
5848 * What we are going to do is to allocate all pages from
5849 * [start, end) (that is remove them from page allocator).
5851 * The only problem is that pages at the beginning and at the
5852 * end of interesting range may be not aligned with pages that
5853 * page allocator holds, ie. they can be part of higher order
5854 * pages. Because of this, we reserve the bigger range and
5855 * once this is done free the pages we are not interested in.
5857 * We don't have to hold zone->lock here because the pages are
5858 * isolated thus they won't get removed from buddy.
5861 lru_add_drain_all();
5865 outer_start = start;
5866 while (!PageBuddy(pfn_to_page(outer_start))) {
5867 if (++order >= MAX_ORDER) {
5871 outer_start &= ~0UL << order;
5874 /* Make sure the range is really isolated. */
5875 if (test_pages_isolated(outer_start, end)) {
5876 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5883 * Reclaim enough pages to make sure that contiguous allocation
5884 * will not starve the system.
5886 __reclaim_pages(zone, GFP_HIGHUSER_MOVABLE, end-start);
5888 /* Grab isolated pages from freelists. */
5889 outer_end = isolate_freepages_range(outer_start, end);
5895 /* Free head and tail (if any) */
5896 if (start != outer_start)
5897 free_contig_range(outer_start, start - outer_start);
5898 if (end != outer_end)
5899 free_contig_range(end, outer_end - end);
5902 undo_isolate_page_range(pfn_max_align_down(start),
5903 pfn_max_align_up(end), migratetype);
5907 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5909 for (; nr_pages--; ++pfn)
5910 __free_page(pfn_to_page(pfn));
5914 #ifdef CONFIG_MEMORY_HOTPLUG
5915 static int __meminit __zone_pcp_update(void *data)
5917 struct zone *zone = data;
5919 unsigned long batch = zone_batchsize(zone), flags;
5921 for_each_possible_cpu(cpu) {
5922 struct per_cpu_pageset *pset;
5923 struct per_cpu_pages *pcp;
5925 pset = per_cpu_ptr(zone->pageset, cpu);
5928 local_irq_save(flags);
5930 free_pcppages_bulk(zone, pcp->count, pcp);
5931 setup_pageset(pset, batch);
5932 local_irq_restore(flags);
5937 void __meminit zone_pcp_update(struct zone *zone)
5939 stop_machine(__zone_pcp_update, zone, NULL);
5943 #ifdef CONFIG_MEMORY_HOTREMOVE
5944 void zone_pcp_reset(struct zone *zone)
5946 unsigned long flags;
5948 /* avoid races with drain_pages() */
5949 local_irq_save(flags);
5950 if (zone->pageset != &boot_pageset) {
5951 free_percpu(zone->pageset);
5952 zone->pageset = &boot_pageset;
5954 local_irq_restore(flags);
5958 * All pages in the range must be isolated before calling this.
5961 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5967 unsigned long flags;
5968 /* find the first valid pfn */
5969 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5974 zone = page_zone(pfn_to_page(pfn));
5975 spin_lock_irqsave(&zone->lock, flags);
5977 while (pfn < end_pfn) {
5978 if (!pfn_valid(pfn)) {
5982 page = pfn_to_page(pfn);
5983 BUG_ON(page_count(page));
5984 BUG_ON(!PageBuddy(page));
5985 order = page_order(page);
5986 #ifdef CONFIG_DEBUG_VM
5987 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5988 pfn, 1 << order, end_pfn);
5990 list_del(&page->lru);
5991 rmv_page_order(page);
5992 zone->free_area[order].nr_free--;
5993 __mod_zone_page_state(zone, NR_FREE_PAGES,
5995 for (i = 0; i < (1 << order); i++)
5996 SetPageReserved((page+i));
5997 pfn += (1 << order);
5999 spin_unlock_irqrestore(&zone->lock, flags);
6003 #ifdef CONFIG_MEMORY_FAILURE
6004 bool is_free_buddy_page(struct page *page)
6006 struct zone *zone = page_zone(page);
6007 unsigned long pfn = page_to_pfn(page);
6008 unsigned long flags;
6011 spin_lock_irqsave(&zone->lock, flags);
6012 for (order = 0; order < MAX_ORDER; order++) {
6013 struct page *page_head = page - (pfn & ((1 << order) - 1));
6015 if (PageBuddy(page_head) && page_order(page_head) >= order)
6018 spin_unlock_irqrestore(&zone->lock, flags);
6020 return order < MAX_ORDER;
6024 static const struct trace_print_flags pageflag_names[] = {
6025 {1UL << PG_locked, "locked" },
6026 {1UL << PG_error, "error" },
6027 {1UL << PG_referenced, "referenced" },
6028 {1UL << PG_uptodate, "uptodate" },
6029 {1UL << PG_dirty, "dirty" },
6030 {1UL << PG_lru, "lru" },
6031 {1UL << PG_active, "active" },
6032 {1UL << PG_slab, "slab" },
6033 {1UL << PG_owner_priv_1, "owner_priv_1" },
6034 {1UL << PG_arch_1, "arch_1" },
6035 {1UL << PG_reserved, "reserved" },
6036 {1UL << PG_private, "private" },
6037 {1UL << PG_private_2, "private_2" },
6038 {1UL << PG_writeback, "writeback" },
6039 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6040 {1UL << PG_head, "head" },
6041 {1UL << PG_tail, "tail" },
6043 {1UL << PG_compound, "compound" },
6045 {1UL << PG_swapcache, "swapcache" },
6046 {1UL << PG_mappedtodisk, "mappedtodisk" },
6047 {1UL << PG_reclaim, "reclaim" },
6048 {1UL << PG_swapbacked, "swapbacked" },
6049 {1UL << PG_unevictable, "unevictable" },
6051 {1UL << PG_mlocked, "mlocked" },
6053 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6054 {1UL << PG_uncached, "uncached" },
6056 #ifdef CONFIG_MEMORY_FAILURE
6057 {1UL << PG_hwpoison, "hwpoison" },
6059 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6060 {1UL << PG_compound_lock, "compound_lock" },
6064 static void dump_page_flags(unsigned long flags)
6066 const char *delim = "";
6070 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6072 printk(KERN_ALERT "page flags: %#lx(", flags);
6074 /* remove zone id */
6075 flags &= (1UL << NR_PAGEFLAGS) - 1;
6077 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6079 mask = pageflag_names[i].mask;
6080 if ((flags & mask) != mask)
6084 printk("%s%s", delim, pageflag_names[i].name);
6088 /* check for left over flags */
6090 printk("%s%#lx", delim, flags);
6095 void dump_page(struct page *page)
6098 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6099 page, atomic_read(&page->_count), page_mapcount(page),
6100 page->mapping, page->index);
6101 dump_page_flags(page->flags);
6102 mem_cgroup_print_bad_page(page);