2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
62 #include <asm/tlbflush.h>
63 #include <asm/div64.h>
66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67 DEFINE_PER_CPU(int, numa_node);
68 EXPORT_PER_CPU_SYMBOL(numa_node);
71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
78 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
83 * Array of node states.
85 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86 [N_POSSIBLE] = NODE_MASK_ALL,
87 [N_ONLINE] = { { [0] = 1UL } },
89 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
91 [N_HIGH_MEMORY] = { { [0] = 1UL } },
93 [N_CPU] = { { [0] = 1UL } },
96 EXPORT_SYMBOL(node_states);
98 unsigned long totalram_pages __read_mostly;
99 unsigned long totalreserve_pages __read_mostly;
101 * When calculating the number of globally allowed dirty pages, there
102 * is a certain number of per-zone reserves that should not be
103 * considered dirtyable memory. This is the sum of those reserves
104 * over all existing zones that contribute dirtyable memory.
106 unsigned long dirty_balance_reserve __read_mostly;
108 int percpu_pagelist_fraction;
109 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
111 #ifdef CONFIG_PM_SLEEP
113 * The following functions are used by the suspend/hibernate code to temporarily
114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
115 * while devices are suspended. To avoid races with the suspend/hibernate code,
116 * they should always be called with pm_mutex held (gfp_allowed_mask also should
117 * only be modified with pm_mutex held, unless the suspend/hibernate code is
118 * guaranteed not to run in parallel with that modification).
121 static gfp_t saved_gfp_mask;
123 void pm_restore_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 if (saved_gfp_mask) {
127 gfp_allowed_mask = saved_gfp_mask;
132 void pm_restrict_gfp_mask(void)
134 WARN_ON(!mutex_is_locked(&pm_mutex));
135 WARN_ON(saved_gfp_mask);
136 saved_gfp_mask = gfp_allowed_mask;
137 gfp_allowed_mask &= ~GFP_IOFS;
140 bool pm_suspended_storage(void)
142 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
146 #endif /* CONFIG_PM_SLEEP */
148 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
149 int pageblock_order __read_mostly;
152 static void __free_pages_ok(struct page *page, unsigned int order);
155 * results with 256, 32 in the lowmem_reserve sysctl:
156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
157 * 1G machine -> (16M dma, 784M normal, 224M high)
158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
162 * TBD: should special case ZONE_DMA32 machines here - in those we normally
163 * don't need any ZONE_NORMAL reservation
165 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
166 #ifdef CONFIG_ZONE_DMA
169 #ifdef CONFIG_ZONE_DMA32
172 #ifdef CONFIG_HIGHMEM
178 EXPORT_SYMBOL(totalram_pages);
180 static char * const zone_names[MAX_NR_ZONES] = {
181 #ifdef CONFIG_ZONE_DMA
184 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 int min_free_kbytes = 1024;
196 static unsigned long __meminitdata nr_kernel_pages;
197 static unsigned long __meminitdata nr_all_pages;
198 static unsigned long __meminitdata dma_reserve;
200 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
201 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
202 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
203 static unsigned long __initdata required_kernelcore;
204 static unsigned long __initdata required_movablecore;
205 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
207 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
209 EXPORT_SYMBOL(movable_zone);
210 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
213 int nr_node_ids __read_mostly = MAX_NUMNODES;
214 int nr_online_nodes __read_mostly = 1;
215 EXPORT_SYMBOL(nr_node_ids);
216 EXPORT_SYMBOL(nr_online_nodes);
219 int page_group_by_mobility_disabled __read_mostly;
223 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
224 * Instead, use {un}set_pageblock_isolate.
226 void set_pageblock_migratetype(struct page *page, int migratetype)
229 if (unlikely(page_group_by_mobility_disabled))
230 migratetype = MIGRATE_UNMOVABLE;
232 set_pageblock_flags_group(page, (unsigned long)migratetype,
233 PB_migrate, PB_migrate_end);
236 bool oom_killer_disabled __read_mostly;
238 #ifdef CONFIG_DEBUG_VM
239 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
243 unsigned long pfn = page_to_pfn(page);
246 seq = zone_span_seqbegin(zone);
247 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
249 else if (pfn < zone->zone_start_pfn)
251 } while (zone_span_seqretry(zone, seq));
256 static int page_is_consistent(struct zone *zone, struct page *page)
258 if (!pfn_valid_within(page_to_pfn(page)))
260 if (zone != page_zone(page))
266 * Temporary debugging check for pages not lying within a given zone.
268 static int bad_range(struct zone *zone, struct page *page)
270 if (page_outside_zone_boundaries(zone, page))
272 if (!page_is_consistent(zone, page))
278 static inline int bad_range(struct zone *zone, struct page *page)
284 static void bad_page(struct page *page)
286 static unsigned long resume;
287 static unsigned long nr_shown;
288 static unsigned long nr_unshown;
290 /* Don't complain about poisoned pages */
291 if (PageHWPoison(page)) {
292 reset_page_mapcount(page); /* remove PageBuddy */
297 * Allow a burst of 60 reports, then keep quiet for that minute;
298 * or allow a steady drip of one report per second.
300 if (nr_shown == 60) {
301 if (time_before(jiffies, resume)) {
307 "BUG: Bad page state: %lu messages suppressed\n",
314 resume = jiffies + 60 * HZ;
316 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
317 current->comm, page_to_pfn(page));
323 /* Leave bad fields for debug, except PageBuddy could make trouble */
324 reset_page_mapcount(page); /* remove PageBuddy */
325 add_taint(TAINT_BAD_PAGE);
329 * Higher-order pages are called "compound pages". They are structured thusly:
331 * The first PAGE_SIZE page is called the "head page".
333 * The remaining PAGE_SIZE pages are called "tail pages".
335 * All pages have PG_compound set. All tail pages have their ->first_page
336 * pointing at the head page.
338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
343 static void free_compound_page(struct page *page)
345 __free_pages_ok(page, compound_order(page));
348 void prep_compound_page(struct page *page, unsigned long order)
351 int nr_pages = 1 << order;
353 set_compound_page_dtor(page, free_compound_page);
354 set_compound_order(page, order);
356 for (i = 1; i < nr_pages; i++) {
357 struct page *p = page + i;
359 set_page_count(p, 0);
360 p->first_page = page;
364 /* update __split_huge_page_refcount if you change this function */
365 static int destroy_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
371 if (unlikely(compound_order(page) != order)) {
376 __ClearPageHead(page);
378 for (i = 1; i < nr_pages; i++) {
379 struct page *p = page + i;
381 if (unlikely(!PageTail(p) || (p->first_page != page))) {
391 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
396 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
397 * and __GFP_HIGHMEM from hard or soft interrupt context.
399 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
400 for (i = 0; i < (1 << order); i++)
401 clear_highpage(page + i);
404 #ifdef CONFIG_DEBUG_PAGEALLOC
405 unsigned int _debug_guardpage_minorder;
407 static int __init debug_guardpage_minorder_setup(char *buf)
411 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
412 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
415 _debug_guardpage_minorder = res;
416 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
419 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
421 static inline void set_page_guard_flag(struct page *page)
423 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
426 static inline void clear_page_guard_flag(struct page *page)
428 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
431 static inline void set_page_guard_flag(struct page *page) { }
432 static inline void clear_page_guard_flag(struct page *page) { }
435 static inline void set_page_order(struct page *page, int order)
437 set_page_private(page, order);
438 __SetPageBuddy(page);
441 static inline void rmv_page_order(struct page *page)
443 __ClearPageBuddy(page);
444 set_page_private(page, 0);
448 * Locate the struct page for both the matching buddy in our
449 * pair (buddy1) and the combined O(n+1) page they form (page).
451 * 1) Any buddy B1 will have an order O twin B2 which satisfies
452 * the following equation:
454 * For example, if the starting buddy (buddy2) is #8 its order
456 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
458 * 2) Any buddy B will have an order O+1 parent P which
459 * satisfies the following equation:
462 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
464 static inline unsigned long
465 __find_buddy_index(unsigned long page_idx, unsigned int order)
467 return page_idx ^ (1 << order);
471 * This function checks whether a page is free && is the buddy
472 * we can do coalesce a page and its buddy if
473 * (a) the buddy is not in a hole &&
474 * (b) the buddy is in the buddy system &&
475 * (c) a page and its buddy have the same order &&
476 * (d) a page and its buddy are in the same zone.
478 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
479 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
481 * For recording page's order, we use page_private(page).
483 static inline int page_is_buddy(struct page *page, struct page *buddy,
486 if (!pfn_valid_within(page_to_pfn(buddy)))
489 if (page_zone_id(page) != page_zone_id(buddy))
492 if (page_is_guard(buddy) && page_order(buddy) == order) {
493 VM_BUG_ON(page_count(buddy) != 0);
497 if (PageBuddy(buddy) && page_order(buddy) == order) {
498 VM_BUG_ON(page_count(buddy) != 0);
505 * Freeing function for a buddy system allocator.
507 * The concept of a buddy system is to maintain direct-mapped table
508 * (containing bit values) for memory blocks of various "orders".
509 * The bottom level table contains the map for the smallest allocatable
510 * units of memory (here, pages), and each level above it describes
511 * pairs of units from the levels below, hence, "buddies".
512 * At a high level, all that happens here is marking the table entry
513 * at the bottom level available, and propagating the changes upward
514 * as necessary, plus some accounting needed to play nicely with other
515 * parts of the VM system.
516 * At each level, we keep a list of pages, which are heads of continuous
517 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
518 * order is recorded in page_private(page) field.
519 * So when we are allocating or freeing one, we can derive the state of the
520 * other. That is, if we allocate a small block, and both were
521 * free, the remainder of the region must be split into blocks.
522 * If a block is freed, and its buddy is also free, then this
523 * triggers coalescing into a block of larger size.
528 static inline void __free_one_page(struct page *page,
529 struct zone *zone, unsigned int order,
532 unsigned long page_idx;
533 unsigned long combined_idx;
534 unsigned long uninitialized_var(buddy_idx);
537 if (unlikely(PageCompound(page)))
538 if (unlikely(destroy_compound_page(page, order)))
541 VM_BUG_ON(migratetype == -1);
543 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
545 VM_BUG_ON(page_idx & ((1 << order) - 1));
546 VM_BUG_ON(bad_range(zone, page));
548 while (order < MAX_ORDER-1) {
549 buddy_idx = __find_buddy_index(page_idx, order);
550 buddy = page + (buddy_idx - page_idx);
551 if (!page_is_buddy(page, buddy, order))
554 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
555 * merge with it and move up one order.
557 if (page_is_guard(buddy)) {
558 clear_page_guard_flag(buddy);
559 set_page_private(page, 0);
560 __mod_zone_freepage_state(zone, 1 << order,
563 list_del(&buddy->lru);
564 zone->free_area[order].nr_free--;
565 rmv_page_order(buddy);
567 combined_idx = buddy_idx & page_idx;
568 page = page + (combined_idx - page_idx);
569 page_idx = combined_idx;
572 set_page_order(page, order);
575 * If this is not the largest possible page, check if the buddy
576 * of the next-highest order is free. If it is, it's possible
577 * that pages are being freed that will coalesce soon. In case,
578 * that is happening, add the free page to the tail of the list
579 * so it's less likely to be used soon and more likely to be merged
580 * as a higher order page
582 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
583 struct page *higher_page, *higher_buddy;
584 combined_idx = buddy_idx & page_idx;
585 higher_page = page + (combined_idx - page_idx);
586 buddy_idx = __find_buddy_index(combined_idx, order + 1);
587 higher_buddy = higher_page + (buddy_idx - combined_idx);
588 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
589 list_add_tail(&page->lru,
590 &zone->free_area[order].free_list[migratetype]);
595 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
597 zone->free_area[order].nr_free++;
601 * free_page_mlock() -- clean up attempts to free and mlocked() page.
602 * Page should not be on lru, so no need to fix that up.
603 * free_pages_check() will verify...
605 static inline void free_page_mlock(struct page *page)
607 __dec_zone_page_state(page, NR_MLOCK);
608 __count_vm_event(UNEVICTABLE_MLOCKFREED);
611 static inline int free_pages_check(struct page *page)
613 if (unlikely(page_mapcount(page) |
614 (page->mapping != NULL) |
615 (atomic_read(&page->_count) != 0) |
616 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
617 (mem_cgroup_bad_page_check(page)))) {
621 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
622 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
627 * Frees a number of pages from the PCP lists
628 * Assumes all pages on list are in same zone, and of same order.
629 * count is the number of pages to free.
631 * If the zone was previously in an "all pages pinned" state then look to
632 * see if this freeing clears that state.
634 * And clear the zone's pages_scanned counter, to hold off the "all pages are
635 * pinned" detection logic.
637 static void free_pcppages_bulk(struct zone *zone, int count,
638 struct per_cpu_pages *pcp)
644 spin_lock(&zone->lock);
645 zone->all_unreclaimable = 0;
646 zone->pages_scanned = 0;
650 struct list_head *list;
653 * Remove pages from lists in a round-robin fashion. A
654 * batch_free count is maintained that is incremented when an
655 * empty list is encountered. This is so more pages are freed
656 * off fuller lists instead of spinning excessively around empty
661 if (++migratetype == MIGRATE_PCPTYPES)
663 list = &pcp->lists[migratetype];
664 } while (list_empty(list));
666 /* This is the only non-empty list. Free them all. */
667 if (batch_free == MIGRATE_PCPTYPES)
668 batch_free = to_free;
671 int mt; /* migratetype of the to-be-freed page */
673 page = list_entry(list->prev, struct page, lru);
674 /* must delete as __free_one_page list manipulates */
675 list_del(&page->lru);
676 mt = page_private(page);
677 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
678 __free_one_page(page, zone, 0, mt);
679 trace_mm_page_pcpu_drain(page, 0, mt);
680 if (is_migrate_cma(mt))
681 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
682 } while (--to_free && --batch_free && !list_empty(list));
684 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
685 spin_unlock(&zone->lock);
688 static void free_one_page(struct zone *zone, struct page *page, int order,
691 spin_lock(&zone->lock);
692 zone->all_unreclaimable = 0;
693 zone->pages_scanned = 0;
695 __free_one_page(page, zone, order, migratetype);
696 if (unlikely(migratetype != MIGRATE_ISOLATE))
697 __mod_zone_freepage_state(zone, 1 << order, migratetype);
698 spin_unlock(&zone->lock);
701 static bool free_pages_prepare(struct page *page, unsigned int order)
706 trace_mm_page_free(page, order);
707 kmemcheck_free_shadow(page, order);
710 page->mapping = NULL;
711 for (i = 0; i < (1 << order); i++)
712 bad += free_pages_check(page + i);
716 if (!PageHighMem(page)) {
717 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
718 debug_check_no_obj_freed(page_address(page),
721 arch_free_page(page, order);
722 kernel_map_pages(page, 1 << order, 0);
727 static void __free_pages_ok(struct page *page, unsigned int order)
730 int wasMlocked = __TestClearPageMlocked(page);
732 if (!free_pages_prepare(page, order))
735 local_irq_save(flags);
736 if (unlikely(wasMlocked))
737 free_page_mlock(page);
738 __count_vm_events(PGFREE, 1 << order);
739 free_one_page(page_zone(page), page, order,
740 get_pageblock_migratetype(page));
741 local_irq_restore(flags);
744 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
746 unsigned int nr_pages = 1 << order;
750 for (loop = 0; loop < nr_pages; loop++) {
751 struct page *p = &page[loop];
753 if (loop + 1 < nr_pages)
755 __ClearPageReserved(p);
756 set_page_count(p, 0);
759 set_page_refcounted(page);
760 __free_pages(page, order);
764 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
765 void __init init_cma_reserved_pageblock(struct page *page)
767 unsigned i = pageblock_nr_pages;
768 struct page *p = page;
771 __ClearPageReserved(p);
772 set_page_count(p, 0);
775 set_page_refcounted(page);
776 set_pageblock_migratetype(page, MIGRATE_CMA);
777 __free_pages(page, pageblock_order);
778 totalram_pages += pageblock_nr_pages;
783 * The order of subdivision here is critical for the IO subsystem.
784 * Please do not alter this order without good reasons and regression
785 * testing. Specifically, as large blocks of memory are subdivided,
786 * the order in which smaller blocks are delivered depends on the order
787 * they're subdivided in this function. This is the primary factor
788 * influencing the order in which pages are delivered to the IO
789 * subsystem according to empirical testing, and this is also justified
790 * by considering the behavior of a buddy system containing a single
791 * large block of memory acted on by a series of small allocations.
792 * This behavior is a critical factor in sglist merging's success.
796 static inline void expand(struct zone *zone, struct page *page,
797 int low, int high, struct free_area *area,
800 unsigned long size = 1 << high;
806 VM_BUG_ON(bad_range(zone, &page[size]));
808 #ifdef CONFIG_DEBUG_PAGEALLOC
809 if (high < debug_guardpage_minorder()) {
811 * Mark as guard pages (or page), that will allow to
812 * merge back to allocator when buddy will be freed.
813 * Corresponding page table entries will not be touched,
814 * pages will stay not present in virtual address space
816 INIT_LIST_HEAD(&page[size].lru);
817 set_page_guard_flag(&page[size]);
818 set_page_private(&page[size], high);
819 /* Guard pages are not available for any usage */
820 __mod_zone_freepage_state(zone, -(1 << high),
825 list_add(&page[size].lru, &area->free_list[migratetype]);
827 set_page_order(&page[size], high);
832 * This page is about to be returned from the page allocator
834 static inline int check_new_page(struct page *page)
836 if (unlikely(page_mapcount(page) |
837 (page->mapping != NULL) |
838 (atomic_read(&page->_count) != 0) |
839 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
840 (mem_cgroup_bad_page_check(page)))) {
847 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
851 for (i = 0; i < (1 << order); i++) {
852 struct page *p = page + i;
853 if (unlikely(check_new_page(p)))
857 set_page_private(page, 0);
858 set_page_refcounted(page);
860 arch_alloc_page(page, order);
861 kernel_map_pages(page, 1 << order, 1);
863 if (gfp_flags & __GFP_ZERO)
864 prep_zero_page(page, order, gfp_flags);
866 if (order && (gfp_flags & __GFP_COMP))
867 prep_compound_page(page, order);
873 * Go through the free lists for the given migratetype and remove
874 * the smallest available page from the freelists
877 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
880 unsigned int current_order;
881 struct free_area * area;
884 /* Find a page of the appropriate size in the preferred list */
885 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
886 area = &(zone->free_area[current_order]);
887 if (list_empty(&area->free_list[migratetype]))
890 page = list_entry(area->free_list[migratetype].next,
892 list_del(&page->lru);
893 rmv_page_order(page);
895 expand(zone, page, order, current_order, area, migratetype);
904 * This array describes the order lists are fallen back to when
905 * the free lists for the desirable migrate type are depleted
907 static int fallbacks[MIGRATE_TYPES][4] = {
908 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
909 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
911 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
912 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
914 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
916 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
917 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
921 * Move the free pages in a range to the free lists of the requested type.
922 * Note that start_page and end_pages are not aligned on a pageblock
923 * boundary. If alignment is required, use move_freepages_block()
925 static int move_freepages(struct zone *zone,
926 struct page *start_page, struct page *end_page,
933 #ifndef CONFIG_HOLES_IN_ZONE
935 * page_zone is not safe to call in this context when
936 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
937 * anyway as we check zone boundaries in move_freepages_block().
938 * Remove at a later date when no bug reports exist related to
939 * grouping pages by mobility
941 BUG_ON(page_zone(start_page) != page_zone(end_page));
944 for (page = start_page; page <= end_page;) {
945 /* Make sure we are not inadvertently changing nodes */
946 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
948 if (!pfn_valid_within(page_to_pfn(page))) {
953 if (!PageBuddy(page)) {
958 order = page_order(page);
959 list_move(&page->lru,
960 &zone->free_area[order].free_list[migratetype]);
962 pages_moved += 1 << order;
968 int move_freepages_block(struct zone *zone, struct page *page,
971 unsigned long start_pfn, end_pfn;
972 struct page *start_page, *end_page;
974 start_pfn = page_to_pfn(page);
975 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
976 start_page = pfn_to_page(start_pfn);
977 end_page = start_page + pageblock_nr_pages - 1;
978 end_pfn = start_pfn + pageblock_nr_pages - 1;
980 /* Do not cross zone boundaries */
981 if (start_pfn < zone->zone_start_pfn)
983 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
986 return move_freepages(zone, start_page, end_page, migratetype);
989 static void change_pageblock_range(struct page *pageblock_page,
990 int start_order, int migratetype)
992 int nr_pageblocks = 1 << (start_order - pageblock_order);
994 while (nr_pageblocks--) {
995 set_pageblock_migratetype(pageblock_page, migratetype);
996 pageblock_page += pageblock_nr_pages;
1000 /* Remove an element from the buddy allocator from the fallback list */
1001 static inline struct page *
1002 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1004 struct free_area * area;
1009 /* Find the largest possible block of pages in the other list */
1010 for (current_order = MAX_ORDER-1; current_order >= order;
1013 migratetype = fallbacks[start_migratetype][i];
1015 /* MIGRATE_RESERVE handled later if necessary */
1016 if (migratetype == MIGRATE_RESERVE)
1019 area = &(zone->free_area[current_order]);
1020 if (list_empty(&area->free_list[migratetype]))
1023 page = list_entry(area->free_list[migratetype].next,
1028 * If breaking a large block of pages, move all free
1029 * pages to the preferred allocation list. If falling
1030 * back for a reclaimable kernel allocation, be more
1031 * aggressive about taking ownership of free pages
1033 * On the other hand, never change migration
1034 * type of MIGRATE_CMA pageblocks nor move CMA
1035 * pages on different free lists. We don't
1036 * want unmovable pages to be allocated from
1037 * MIGRATE_CMA areas.
1039 if (!is_migrate_cma(migratetype) &&
1040 (unlikely(current_order >= pageblock_order / 2) ||
1041 start_migratetype == MIGRATE_RECLAIMABLE ||
1042 page_group_by_mobility_disabled)) {
1044 pages = move_freepages_block(zone, page,
1047 /* Claim the whole block if over half of it is free */
1048 if (pages >= (1 << (pageblock_order-1)) ||
1049 page_group_by_mobility_disabled)
1050 set_pageblock_migratetype(page,
1053 migratetype = start_migratetype;
1056 /* Remove the page from the freelists */
1057 list_del(&page->lru);
1058 rmv_page_order(page);
1060 /* Take ownership for orders >= pageblock_order */
1061 if (current_order >= pageblock_order &&
1062 !is_migrate_cma(migratetype))
1063 change_pageblock_range(page, current_order,
1066 expand(zone, page, order, current_order, area,
1067 is_migrate_cma(migratetype)
1068 ? migratetype : start_migratetype);
1070 trace_mm_page_alloc_extfrag(page, order, current_order,
1071 start_migratetype, migratetype);
1081 * Do the hard work of removing an element from the buddy allocator.
1082 * Call me with the zone->lock already held.
1084 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1090 page = __rmqueue_smallest(zone, order, migratetype);
1092 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1093 page = __rmqueue_fallback(zone, order, migratetype);
1096 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1097 * is used because __rmqueue_smallest is an inline function
1098 * and we want just one call site
1101 migratetype = MIGRATE_RESERVE;
1106 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1111 * Obtain a specified number of elements from the buddy allocator, all under
1112 * a single hold of the lock, for efficiency. Add them to the supplied list.
1113 * Returns the number of new pages which were placed at *list.
1115 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1116 unsigned long count, struct list_head *list,
1117 int migratetype, int cold)
1119 int mt = migratetype, i;
1121 spin_lock(&zone->lock);
1122 for (i = 0; i < count; ++i) {
1123 struct page *page = __rmqueue(zone, order, migratetype);
1124 if (unlikely(page == NULL))
1128 * Split buddy pages returned by expand() are received here
1129 * in physical page order. The page is added to the callers and
1130 * list and the list head then moves forward. From the callers
1131 * perspective, the linked list is ordered by page number in
1132 * some conditions. This is useful for IO devices that can
1133 * merge IO requests if the physical pages are ordered
1136 if (likely(cold == 0))
1137 list_add(&page->lru, list);
1139 list_add_tail(&page->lru, list);
1140 if (IS_ENABLED(CONFIG_CMA)) {
1141 mt = get_pageblock_migratetype(page);
1142 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1145 set_page_private(page, mt);
1147 if (is_migrate_cma(mt))
1148 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1151 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1152 spin_unlock(&zone->lock);
1158 * Called from the vmstat counter updater to drain pagesets of this
1159 * currently executing processor on remote nodes after they have
1162 * Note that this function must be called with the thread pinned to
1163 * a single processor.
1165 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1167 unsigned long flags;
1170 local_irq_save(flags);
1171 if (pcp->count >= pcp->batch)
1172 to_drain = pcp->batch;
1174 to_drain = pcp->count;
1176 free_pcppages_bulk(zone, to_drain, pcp);
1177 pcp->count -= to_drain;
1179 local_irq_restore(flags);
1184 * Drain pages of the indicated processor.
1186 * The processor must either be the current processor and the
1187 * thread pinned to the current processor or a processor that
1190 static void drain_pages(unsigned int cpu)
1192 unsigned long flags;
1195 for_each_populated_zone(zone) {
1196 struct per_cpu_pageset *pset;
1197 struct per_cpu_pages *pcp;
1199 local_irq_save(flags);
1200 pset = per_cpu_ptr(zone->pageset, cpu);
1204 free_pcppages_bulk(zone, pcp->count, pcp);
1207 local_irq_restore(flags);
1212 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1214 void drain_local_pages(void *arg)
1216 drain_pages(smp_processor_id());
1220 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1222 * Note that this code is protected against sending an IPI to an offline
1223 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1224 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1225 * nothing keeps CPUs from showing up after we populated the cpumask and
1226 * before the call to on_each_cpu_mask().
1228 void drain_all_pages(void)
1231 struct per_cpu_pageset *pcp;
1235 * Allocate in the BSS so we wont require allocation in
1236 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1238 static cpumask_t cpus_with_pcps;
1241 * We don't care about racing with CPU hotplug event
1242 * as offline notification will cause the notified
1243 * cpu to drain that CPU pcps and on_each_cpu_mask
1244 * disables preemption as part of its processing
1246 for_each_online_cpu(cpu) {
1247 bool has_pcps = false;
1248 for_each_populated_zone(zone) {
1249 pcp = per_cpu_ptr(zone->pageset, cpu);
1250 if (pcp->pcp.count) {
1256 cpumask_set_cpu(cpu, &cpus_with_pcps);
1258 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1260 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1263 #ifdef CONFIG_HIBERNATION
1265 void mark_free_pages(struct zone *zone)
1267 unsigned long pfn, max_zone_pfn;
1268 unsigned long flags;
1270 struct list_head *curr;
1272 if (!zone->spanned_pages)
1275 spin_lock_irqsave(&zone->lock, flags);
1277 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1278 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1279 if (pfn_valid(pfn)) {
1280 struct page *page = pfn_to_page(pfn);
1282 if (!swsusp_page_is_forbidden(page))
1283 swsusp_unset_page_free(page);
1286 for_each_migratetype_order(order, t) {
1287 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1290 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1291 for (i = 0; i < (1UL << order); i++)
1292 swsusp_set_page_free(pfn_to_page(pfn + i));
1295 spin_unlock_irqrestore(&zone->lock, flags);
1297 #endif /* CONFIG_PM */
1300 * Free a 0-order page
1301 * cold == 1 ? free a cold page : free a hot page
1303 void free_hot_cold_page(struct page *page, int cold)
1305 struct zone *zone = page_zone(page);
1306 struct per_cpu_pages *pcp;
1307 unsigned long flags;
1309 int wasMlocked = __TestClearPageMlocked(page);
1311 if (!free_pages_prepare(page, 0))
1314 migratetype = get_pageblock_migratetype(page);
1315 set_page_private(page, migratetype);
1316 local_irq_save(flags);
1317 if (unlikely(wasMlocked))
1318 free_page_mlock(page);
1319 __count_vm_event(PGFREE);
1322 * We only track unmovable, reclaimable and movable on pcp lists.
1323 * Free ISOLATE pages back to the allocator because they are being
1324 * offlined but treat RESERVE as movable pages so we can get those
1325 * areas back if necessary. Otherwise, we may have to free
1326 * excessively into the page allocator
1328 if (migratetype >= MIGRATE_PCPTYPES) {
1329 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1330 free_one_page(zone, page, 0, migratetype);
1333 migratetype = MIGRATE_MOVABLE;
1336 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1338 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1340 list_add(&page->lru, &pcp->lists[migratetype]);
1342 if (pcp->count >= pcp->high) {
1343 free_pcppages_bulk(zone, pcp->batch, pcp);
1344 pcp->count -= pcp->batch;
1348 local_irq_restore(flags);
1352 * Free a list of 0-order pages
1354 void free_hot_cold_page_list(struct list_head *list, int cold)
1356 struct page *page, *next;
1358 list_for_each_entry_safe(page, next, list, lru) {
1359 trace_mm_page_free_batched(page, cold);
1360 free_hot_cold_page(page, cold);
1365 * split_page takes a non-compound higher-order page, and splits it into
1366 * n (1<<order) sub-pages: page[0..n]
1367 * Each sub-page must be freed individually.
1369 * Note: this is probably too low level an operation for use in drivers.
1370 * Please consult with lkml before using this in your driver.
1372 void split_page(struct page *page, unsigned int order)
1376 VM_BUG_ON(PageCompound(page));
1377 VM_BUG_ON(!page_count(page));
1379 #ifdef CONFIG_KMEMCHECK
1381 * Split shadow pages too, because free(page[0]) would
1382 * otherwise free the whole shadow.
1384 if (kmemcheck_page_is_tracked(page))
1385 split_page(virt_to_page(page[0].shadow), order);
1388 for (i = 1; i < (1 << order); i++)
1389 set_page_refcounted(page + i);
1393 * Similar to the split_page family of functions except that the page
1394 * required at the given order and being isolated now to prevent races
1395 * with parallel allocators
1397 int capture_free_page(struct page *page, int alloc_order, int migratetype)
1400 unsigned long watermark;
1404 BUG_ON(!PageBuddy(page));
1406 zone = page_zone(page);
1407 order = page_order(page);
1409 /* Obey watermarks as if the page was being allocated */
1410 watermark = low_wmark_pages(zone) + (1 << order);
1411 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1414 /* Remove page from free list */
1415 list_del(&page->lru);
1416 zone->free_area[order].nr_free--;
1417 rmv_page_order(page);
1419 mt = get_pageblock_migratetype(page);
1420 if (unlikely(mt != MIGRATE_ISOLATE))
1421 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1423 if (alloc_order != order)
1424 expand(zone, page, alloc_order, order,
1425 &zone->free_area[order], migratetype);
1427 /* Set the pageblock if the captured page is at least a pageblock */
1428 if (order >= pageblock_order - 1) {
1429 struct page *endpage = page + (1 << order) - 1;
1430 for (; page < endpage; page += pageblock_nr_pages) {
1431 int mt = get_pageblock_migratetype(page);
1432 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1433 set_pageblock_migratetype(page,
1438 return 1UL << order;
1442 * Similar to split_page except the page is already free. As this is only
1443 * being used for migration, the migratetype of the block also changes.
1444 * As this is called with interrupts disabled, the caller is responsible
1445 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1448 * Note: this is probably too low level an operation for use in drivers.
1449 * Please consult with lkml before using this in your driver.
1451 int split_free_page(struct page *page)
1456 BUG_ON(!PageBuddy(page));
1457 order = page_order(page);
1459 nr_pages = capture_free_page(page, order, 0);
1463 /* Split into individual pages */
1464 set_page_refcounted(page);
1465 split_page(page, order);
1470 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1471 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1475 struct page *buffered_rmqueue(struct zone *preferred_zone,
1476 struct zone *zone, int order, gfp_t gfp_flags,
1479 unsigned long flags;
1481 int cold = !!(gfp_flags & __GFP_COLD);
1484 if (likely(order == 0)) {
1485 struct per_cpu_pages *pcp;
1486 struct list_head *list;
1488 local_irq_save(flags);
1489 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1490 list = &pcp->lists[migratetype];
1491 if (list_empty(list)) {
1492 pcp->count += rmqueue_bulk(zone, 0,
1495 if (unlikely(list_empty(list)))
1500 page = list_entry(list->prev, struct page, lru);
1502 page = list_entry(list->next, struct page, lru);
1504 list_del(&page->lru);
1507 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1509 * __GFP_NOFAIL is not to be used in new code.
1511 * All __GFP_NOFAIL callers should be fixed so that they
1512 * properly detect and handle allocation failures.
1514 * We most definitely don't want callers attempting to
1515 * allocate greater than order-1 page units with
1518 WARN_ON_ONCE(order > 1);
1520 spin_lock_irqsave(&zone->lock, flags);
1521 page = __rmqueue(zone, order, migratetype);
1522 spin_unlock(&zone->lock);
1525 __mod_zone_freepage_state(zone, -(1 << order),
1526 get_pageblock_migratetype(page));
1529 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1530 zone_statistics(preferred_zone, zone, gfp_flags);
1531 local_irq_restore(flags);
1533 VM_BUG_ON(bad_range(zone, page));
1534 if (prep_new_page(page, order, gfp_flags))
1539 local_irq_restore(flags);
1543 #ifdef CONFIG_FAIL_PAGE_ALLOC
1546 struct fault_attr attr;
1548 u32 ignore_gfp_highmem;
1549 u32 ignore_gfp_wait;
1551 } fail_page_alloc = {
1552 .attr = FAULT_ATTR_INITIALIZER,
1553 .ignore_gfp_wait = 1,
1554 .ignore_gfp_highmem = 1,
1558 static int __init setup_fail_page_alloc(char *str)
1560 return setup_fault_attr(&fail_page_alloc.attr, str);
1562 __setup("fail_page_alloc=", setup_fail_page_alloc);
1564 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1566 if (order < fail_page_alloc.min_order)
1568 if (gfp_mask & __GFP_NOFAIL)
1570 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1572 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1575 return should_fail(&fail_page_alloc.attr, 1 << order);
1578 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1580 static int __init fail_page_alloc_debugfs(void)
1582 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1585 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1586 &fail_page_alloc.attr);
1588 return PTR_ERR(dir);
1590 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1591 &fail_page_alloc.ignore_gfp_wait))
1593 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1594 &fail_page_alloc.ignore_gfp_highmem))
1596 if (!debugfs_create_u32("min-order", mode, dir,
1597 &fail_page_alloc.min_order))
1602 debugfs_remove_recursive(dir);
1607 late_initcall(fail_page_alloc_debugfs);
1609 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1611 #else /* CONFIG_FAIL_PAGE_ALLOC */
1613 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1618 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1621 * Return true if free pages are above 'mark'. This takes into account the order
1622 * of the allocation.
1624 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1625 int classzone_idx, int alloc_flags, long free_pages)
1627 /* free_pages my go negative - that's OK */
1629 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1632 free_pages -= (1 << order) - 1;
1633 if (alloc_flags & ALLOC_HIGH)
1635 if (alloc_flags & ALLOC_HARDER)
1638 /* If allocation can't use CMA areas don't use free CMA pages */
1639 if (!(alloc_flags & ALLOC_CMA))
1640 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1642 if (free_pages <= min + lowmem_reserve)
1644 for (o = 0; o < order; o++) {
1645 /* At the next order, this order's pages become unavailable */
1646 free_pages -= z->free_area[o].nr_free << o;
1648 /* Require fewer higher order pages to be free */
1651 if (free_pages <= min)
1657 #ifdef CONFIG_MEMORY_ISOLATION
1658 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1660 if (unlikely(zone->nr_pageblock_isolate))
1661 return zone->nr_pageblock_isolate * pageblock_nr_pages;
1665 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1671 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1672 int classzone_idx, int alloc_flags)
1674 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1675 zone_page_state(z, NR_FREE_PAGES));
1678 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1679 int classzone_idx, int alloc_flags)
1681 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1683 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1684 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1687 * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1688 * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1689 * sleep although it could do so. But this is more desirable for memory
1690 * hotplug than sleeping which can cause a livelock in the direct
1693 free_pages -= nr_zone_isolate_freepages(z);
1694 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1700 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1701 * skip over zones that are not allowed by the cpuset, or that have
1702 * been recently (in last second) found to be nearly full. See further
1703 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1704 * that have to skip over a lot of full or unallowed zones.
1706 * If the zonelist cache is present in the passed in zonelist, then
1707 * returns a pointer to the allowed node mask (either the current
1708 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1710 * If the zonelist cache is not available for this zonelist, does
1711 * nothing and returns NULL.
1713 * If the fullzones BITMAP in the zonelist cache is stale (more than
1714 * a second since last zap'd) then we zap it out (clear its bits.)
1716 * We hold off even calling zlc_setup, until after we've checked the
1717 * first zone in the zonelist, on the theory that most allocations will
1718 * be satisfied from that first zone, so best to examine that zone as
1719 * quickly as we can.
1721 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1723 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1724 nodemask_t *allowednodes; /* zonelist_cache approximation */
1726 zlc = zonelist->zlcache_ptr;
1730 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1731 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1732 zlc->last_full_zap = jiffies;
1735 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1736 &cpuset_current_mems_allowed :
1737 &node_states[N_HIGH_MEMORY];
1738 return allowednodes;
1742 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1743 * if it is worth looking at further for free memory:
1744 * 1) Check that the zone isn't thought to be full (doesn't have its
1745 * bit set in the zonelist_cache fullzones BITMAP).
1746 * 2) Check that the zones node (obtained from the zonelist_cache
1747 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1748 * Return true (non-zero) if zone is worth looking at further, or
1749 * else return false (zero) if it is not.
1751 * This check -ignores- the distinction between various watermarks,
1752 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1753 * found to be full for any variation of these watermarks, it will
1754 * be considered full for up to one second by all requests, unless
1755 * we are so low on memory on all allowed nodes that we are forced
1756 * into the second scan of the zonelist.
1758 * In the second scan we ignore this zonelist cache and exactly
1759 * apply the watermarks to all zones, even it is slower to do so.
1760 * We are low on memory in the second scan, and should leave no stone
1761 * unturned looking for a free page.
1763 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1764 nodemask_t *allowednodes)
1766 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1767 int i; /* index of *z in zonelist zones */
1768 int n; /* node that zone *z is on */
1770 zlc = zonelist->zlcache_ptr;
1774 i = z - zonelist->_zonerefs;
1777 /* This zone is worth trying if it is allowed but not full */
1778 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1782 * Given 'z' scanning a zonelist, set the corresponding bit in
1783 * zlc->fullzones, so that subsequent attempts to allocate a page
1784 * from that zone don't waste time re-examining it.
1786 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1788 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1789 int i; /* index of *z in zonelist zones */
1791 zlc = zonelist->zlcache_ptr;
1795 i = z - zonelist->_zonerefs;
1797 set_bit(i, zlc->fullzones);
1801 * clear all zones full, called after direct reclaim makes progress so that
1802 * a zone that was recently full is not skipped over for up to a second
1804 static void zlc_clear_zones_full(struct zonelist *zonelist)
1806 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1808 zlc = zonelist->zlcache_ptr;
1812 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1815 #else /* CONFIG_NUMA */
1817 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1822 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1823 nodemask_t *allowednodes)
1828 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1832 static void zlc_clear_zones_full(struct zonelist *zonelist)
1835 #endif /* CONFIG_NUMA */
1838 * get_page_from_freelist goes through the zonelist trying to allocate
1841 static struct page *
1842 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1843 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1844 struct zone *preferred_zone, int migratetype)
1847 struct page *page = NULL;
1850 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1851 int zlc_active = 0; /* set if using zonelist_cache */
1852 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1854 classzone_idx = zone_idx(preferred_zone);
1857 * Scan zonelist, looking for a zone with enough free.
1858 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1860 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1861 high_zoneidx, nodemask) {
1862 if (NUMA_BUILD && zlc_active &&
1863 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1865 if ((alloc_flags & ALLOC_CPUSET) &&
1866 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1869 * When allocating a page cache page for writing, we
1870 * want to get it from a zone that is within its dirty
1871 * limit, such that no single zone holds more than its
1872 * proportional share of globally allowed dirty pages.
1873 * The dirty limits take into account the zone's
1874 * lowmem reserves and high watermark so that kswapd
1875 * should be able to balance it without having to
1876 * write pages from its LRU list.
1878 * This may look like it could increase pressure on
1879 * lower zones by failing allocations in higher zones
1880 * before they are full. But the pages that do spill
1881 * over are limited as the lower zones are protected
1882 * by this very same mechanism. It should not become
1883 * a practical burden to them.
1885 * XXX: For now, allow allocations to potentially
1886 * exceed the per-zone dirty limit in the slowpath
1887 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1888 * which is important when on a NUMA setup the allowed
1889 * zones are together not big enough to reach the
1890 * global limit. The proper fix for these situations
1891 * will require awareness of zones in the
1892 * dirty-throttling and the flusher threads.
1894 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1895 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1896 goto this_zone_full;
1898 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1899 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1903 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1904 if (zone_watermark_ok(zone, order, mark,
1905 classzone_idx, alloc_flags))
1908 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1910 * we do zlc_setup if there are multiple nodes
1911 * and before considering the first zone allowed
1914 allowednodes = zlc_setup(zonelist, alloc_flags);
1919 if (zone_reclaim_mode == 0)
1920 goto this_zone_full;
1923 * As we may have just activated ZLC, check if the first
1924 * eligible zone has failed zone_reclaim recently.
1926 if (NUMA_BUILD && zlc_active &&
1927 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1930 ret = zone_reclaim(zone, gfp_mask, order);
1932 case ZONE_RECLAIM_NOSCAN:
1935 case ZONE_RECLAIM_FULL:
1936 /* scanned but unreclaimable */
1939 /* did we reclaim enough */
1940 if (!zone_watermark_ok(zone, order, mark,
1941 classzone_idx, alloc_flags))
1942 goto this_zone_full;
1947 page = buffered_rmqueue(preferred_zone, zone, order,
1948 gfp_mask, migratetype);
1953 zlc_mark_zone_full(zonelist, z);
1956 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1957 /* Disable zlc cache for second zonelist scan */
1964 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1965 * necessary to allocate the page. The expectation is
1966 * that the caller is taking steps that will free more
1967 * memory. The caller should avoid the page being used
1968 * for !PFMEMALLOC purposes.
1970 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1976 * Large machines with many possible nodes should not always dump per-node
1977 * meminfo in irq context.
1979 static inline bool should_suppress_show_mem(void)
1984 ret = in_interrupt();
1989 static DEFINE_RATELIMIT_STATE(nopage_rs,
1990 DEFAULT_RATELIMIT_INTERVAL,
1991 DEFAULT_RATELIMIT_BURST);
1993 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1995 unsigned int filter = SHOW_MEM_FILTER_NODES;
1997 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1998 debug_guardpage_minorder() > 0)
2002 * This documents exceptions given to allocations in certain
2003 * contexts that are allowed to allocate outside current's set
2006 if (!(gfp_mask & __GFP_NOMEMALLOC))
2007 if (test_thread_flag(TIF_MEMDIE) ||
2008 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2009 filter &= ~SHOW_MEM_FILTER_NODES;
2010 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2011 filter &= ~SHOW_MEM_FILTER_NODES;
2014 struct va_format vaf;
2017 va_start(args, fmt);
2022 pr_warn("%pV", &vaf);
2027 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2028 current->comm, order, gfp_mask);
2031 if (!should_suppress_show_mem())
2036 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2037 unsigned long did_some_progress,
2038 unsigned long pages_reclaimed)
2040 /* Do not loop if specifically requested */
2041 if (gfp_mask & __GFP_NORETRY)
2044 /* Always retry if specifically requested */
2045 if (gfp_mask & __GFP_NOFAIL)
2049 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2050 * making forward progress without invoking OOM. Suspend also disables
2051 * storage devices so kswapd will not help. Bail if we are suspending.
2053 if (!did_some_progress && pm_suspended_storage())
2057 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2058 * means __GFP_NOFAIL, but that may not be true in other
2061 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2065 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2066 * specified, then we retry until we no longer reclaim any pages
2067 * (above), or we've reclaimed an order of pages at least as
2068 * large as the allocation's order. In both cases, if the
2069 * allocation still fails, we stop retrying.
2071 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2077 static inline struct page *
2078 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2079 struct zonelist *zonelist, enum zone_type high_zoneidx,
2080 nodemask_t *nodemask, struct zone *preferred_zone,
2085 /* Acquire the OOM killer lock for the zones in zonelist */
2086 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2087 schedule_timeout_uninterruptible(1);
2092 * Go through the zonelist yet one more time, keep very high watermark
2093 * here, this is only to catch a parallel oom killing, we must fail if
2094 * we're still under heavy pressure.
2096 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2097 order, zonelist, high_zoneidx,
2098 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2099 preferred_zone, migratetype);
2103 if (!(gfp_mask & __GFP_NOFAIL)) {
2104 /* The OOM killer will not help higher order allocs */
2105 if (order > PAGE_ALLOC_COSTLY_ORDER)
2107 /* The OOM killer does not needlessly kill tasks for lowmem */
2108 if (high_zoneidx < ZONE_NORMAL)
2111 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2112 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2113 * The caller should handle page allocation failure by itself if
2114 * it specifies __GFP_THISNODE.
2115 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2117 if (gfp_mask & __GFP_THISNODE)
2120 /* Exhausted what can be done so it's blamo time */
2121 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2124 clear_zonelist_oom(zonelist, gfp_mask);
2128 #ifdef CONFIG_COMPACTION
2129 /* Try memory compaction for high-order allocations before reclaim */
2130 static struct page *
2131 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2132 struct zonelist *zonelist, enum zone_type high_zoneidx,
2133 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2134 int migratetype, bool sync_migration,
2135 bool *contended_compaction, bool *deferred_compaction,
2136 unsigned long *did_some_progress)
2138 struct page *page = NULL;
2143 if (compaction_deferred(preferred_zone, order)) {
2144 *deferred_compaction = true;
2148 current->flags |= PF_MEMALLOC;
2149 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2150 nodemask, sync_migration,
2151 contended_compaction, &page);
2152 current->flags &= ~PF_MEMALLOC;
2154 /* If compaction captured a page, prep and use it */
2156 prep_new_page(page, order, gfp_mask);
2160 if (*did_some_progress != COMPACT_SKIPPED) {
2161 /* Page migration frees to the PCP lists but we want merging */
2162 drain_pages(get_cpu());
2165 page = get_page_from_freelist(gfp_mask, nodemask,
2166 order, zonelist, high_zoneidx,
2167 alloc_flags & ~ALLOC_NO_WATERMARKS,
2168 preferred_zone, migratetype);
2171 preferred_zone->compact_considered = 0;
2172 preferred_zone->compact_defer_shift = 0;
2173 if (order >= preferred_zone->compact_order_failed)
2174 preferred_zone->compact_order_failed = order + 1;
2175 count_vm_event(COMPACTSUCCESS);
2180 * It's bad if compaction run occurs and fails.
2181 * The most likely reason is that pages exist,
2182 * but not enough to satisfy watermarks.
2184 count_vm_event(COMPACTFAIL);
2187 * As async compaction considers a subset of pageblocks, only
2188 * defer if the failure was a sync compaction failure.
2191 defer_compaction(preferred_zone, order);
2199 static inline struct page *
2200 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2201 struct zonelist *zonelist, enum zone_type high_zoneidx,
2202 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2203 int migratetype, bool sync_migration,
2204 bool *contended_compaction, bool *deferred_compaction,
2205 unsigned long *did_some_progress)
2209 #endif /* CONFIG_COMPACTION */
2211 /* Perform direct synchronous page reclaim */
2213 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2214 nodemask_t *nodemask)
2216 struct reclaim_state reclaim_state;
2221 /* We now go into synchronous reclaim */
2222 cpuset_memory_pressure_bump();
2223 current->flags |= PF_MEMALLOC;
2224 lockdep_set_current_reclaim_state(gfp_mask);
2225 reclaim_state.reclaimed_slab = 0;
2226 current->reclaim_state = &reclaim_state;
2228 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2230 current->reclaim_state = NULL;
2231 lockdep_clear_current_reclaim_state();
2232 current->flags &= ~PF_MEMALLOC;
2239 /* The really slow allocator path where we enter direct reclaim */
2240 static inline struct page *
2241 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2242 struct zonelist *zonelist, enum zone_type high_zoneidx,
2243 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2244 int migratetype, unsigned long *did_some_progress)
2246 struct page *page = NULL;
2247 bool drained = false;
2249 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2251 if (unlikely(!(*did_some_progress)))
2254 /* After successful reclaim, reconsider all zones for allocation */
2256 zlc_clear_zones_full(zonelist);
2259 page = get_page_from_freelist(gfp_mask, nodemask, order,
2260 zonelist, high_zoneidx,
2261 alloc_flags & ~ALLOC_NO_WATERMARKS,
2262 preferred_zone, migratetype);
2265 * If an allocation failed after direct reclaim, it could be because
2266 * pages are pinned on the per-cpu lists. Drain them and try again
2268 if (!page && !drained) {
2278 * This is called in the allocator slow-path if the allocation request is of
2279 * sufficient urgency to ignore watermarks and take other desperate measures
2281 static inline struct page *
2282 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2283 struct zonelist *zonelist, enum zone_type high_zoneidx,
2284 nodemask_t *nodemask, struct zone *preferred_zone,
2290 page = get_page_from_freelist(gfp_mask, nodemask, order,
2291 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2292 preferred_zone, migratetype);
2294 if (!page && gfp_mask & __GFP_NOFAIL)
2295 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2296 } while (!page && (gfp_mask & __GFP_NOFAIL));
2302 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2303 enum zone_type high_zoneidx,
2304 enum zone_type classzone_idx)
2309 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2310 wakeup_kswapd(zone, order, classzone_idx);
2314 gfp_to_alloc_flags(gfp_t gfp_mask)
2316 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2317 const gfp_t wait = gfp_mask & __GFP_WAIT;
2319 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2320 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2323 * The caller may dip into page reserves a bit more if the caller
2324 * cannot run direct reclaim, or if the caller has realtime scheduling
2325 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2326 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2328 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2332 * Not worth trying to allocate harder for
2333 * __GFP_NOMEMALLOC even if it can't schedule.
2335 if (!(gfp_mask & __GFP_NOMEMALLOC))
2336 alloc_flags |= ALLOC_HARDER;
2338 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2339 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2341 alloc_flags &= ~ALLOC_CPUSET;
2342 } else if (unlikely(rt_task(current)) && !in_interrupt())
2343 alloc_flags |= ALLOC_HARDER;
2345 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2346 if (gfp_mask & __GFP_MEMALLOC)
2347 alloc_flags |= ALLOC_NO_WATERMARKS;
2348 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2349 alloc_flags |= ALLOC_NO_WATERMARKS;
2350 else if (!in_interrupt() &&
2351 ((current->flags & PF_MEMALLOC) ||
2352 unlikely(test_thread_flag(TIF_MEMDIE))))
2353 alloc_flags |= ALLOC_NO_WATERMARKS;
2356 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2357 alloc_flags |= ALLOC_CMA;
2362 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2364 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2367 static inline struct page *
2368 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2369 struct zonelist *zonelist, enum zone_type high_zoneidx,
2370 nodemask_t *nodemask, struct zone *preferred_zone,
2373 const gfp_t wait = gfp_mask & __GFP_WAIT;
2374 struct page *page = NULL;
2376 unsigned long pages_reclaimed = 0;
2377 unsigned long did_some_progress;
2378 bool sync_migration = false;
2379 bool deferred_compaction = false;
2380 bool contended_compaction = false;
2383 * In the slowpath, we sanity check order to avoid ever trying to
2384 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2385 * be using allocators in order of preference for an area that is
2388 if (order >= MAX_ORDER) {
2389 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2394 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2395 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2396 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2397 * using a larger set of nodes after it has established that the
2398 * allowed per node queues are empty and that nodes are
2401 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2405 wake_all_kswapd(order, zonelist, high_zoneidx,
2406 zone_idx(preferred_zone));
2409 * OK, we're below the kswapd watermark and have kicked background
2410 * reclaim. Now things get more complex, so set up alloc_flags according
2411 * to how we want to proceed.
2413 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2416 * Find the true preferred zone if the allocation is unconstrained by
2419 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2420 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2424 /* This is the last chance, in general, before the goto nopage. */
2425 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2426 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2427 preferred_zone, migratetype);
2431 /* Allocate without watermarks if the context allows */
2432 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2434 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2435 * the allocation is high priority and these type of
2436 * allocations are system rather than user orientated
2438 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2440 page = __alloc_pages_high_priority(gfp_mask, order,
2441 zonelist, high_zoneidx, nodemask,
2442 preferred_zone, migratetype);
2448 /* Atomic allocations - we can't balance anything */
2452 /* Avoid recursion of direct reclaim */
2453 if (current->flags & PF_MEMALLOC)
2456 /* Avoid allocations with no watermarks from looping endlessly */
2457 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2461 * Try direct compaction. The first pass is asynchronous. Subsequent
2462 * attempts after direct reclaim are synchronous
2464 page = __alloc_pages_direct_compact(gfp_mask, order,
2465 zonelist, high_zoneidx,
2467 alloc_flags, preferred_zone,
2468 migratetype, sync_migration,
2469 &contended_compaction,
2470 &deferred_compaction,
2471 &did_some_progress);
2474 sync_migration = true;
2477 * If compaction is deferred for high-order allocations, it is because
2478 * sync compaction recently failed. In this is the case and the caller
2479 * requested a movable allocation that does not heavily disrupt the
2480 * system then fail the allocation instead of entering direct reclaim.
2482 if ((deferred_compaction || contended_compaction) &&
2483 (gfp_mask & (__GFP_MOVABLE|__GFP_REPEAT)) == __GFP_MOVABLE)
2486 /* Try direct reclaim and then allocating */
2487 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2488 zonelist, high_zoneidx,
2490 alloc_flags, preferred_zone,
2491 migratetype, &did_some_progress);
2496 * If we failed to make any progress reclaiming, then we are
2497 * running out of options and have to consider going OOM
2499 if (!did_some_progress) {
2500 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2501 if (oom_killer_disabled)
2503 /* Coredumps can quickly deplete all memory reserves */
2504 if ((current->flags & PF_DUMPCORE) &&
2505 !(gfp_mask & __GFP_NOFAIL))
2507 page = __alloc_pages_may_oom(gfp_mask, order,
2508 zonelist, high_zoneidx,
2509 nodemask, preferred_zone,
2514 if (!(gfp_mask & __GFP_NOFAIL)) {
2516 * The oom killer is not called for high-order
2517 * allocations that may fail, so if no progress
2518 * is being made, there are no other options and
2519 * retrying is unlikely to help.
2521 if (order > PAGE_ALLOC_COSTLY_ORDER)
2524 * The oom killer is not called for lowmem
2525 * allocations to prevent needlessly killing
2528 if (high_zoneidx < ZONE_NORMAL)
2536 /* Check if we should retry the allocation */
2537 pages_reclaimed += did_some_progress;
2538 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2540 /* Wait for some write requests to complete then retry */
2541 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2545 * High-order allocations do not necessarily loop after
2546 * direct reclaim and reclaim/compaction depends on compaction
2547 * being called after reclaim so call directly if necessary
2549 page = __alloc_pages_direct_compact(gfp_mask, order,
2550 zonelist, high_zoneidx,
2552 alloc_flags, preferred_zone,
2553 migratetype, sync_migration,
2554 &contended_compaction,
2555 &deferred_compaction,
2556 &did_some_progress);
2562 warn_alloc_failed(gfp_mask, order, NULL);
2565 if (kmemcheck_enabled)
2566 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2572 * This is the 'heart' of the zoned buddy allocator.
2575 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2576 struct zonelist *zonelist, nodemask_t *nodemask)
2578 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2579 struct zone *preferred_zone;
2580 struct page *page = NULL;
2581 int migratetype = allocflags_to_migratetype(gfp_mask);
2582 unsigned int cpuset_mems_cookie;
2583 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2585 gfp_mask &= gfp_allowed_mask;
2587 lockdep_trace_alloc(gfp_mask);
2589 might_sleep_if(gfp_mask & __GFP_WAIT);
2591 if (should_fail_alloc_page(gfp_mask, order))
2595 * Check the zones suitable for the gfp_mask contain at least one
2596 * valid zone. It's possible to have an empty zonelist as a result
2597 * of GFP_THISNODE and a memoryless node
2599 if (unlikely(!zonelist->_zonerefs->zone))
2603 cpuset_mems_cookie = get_mems_allowed();
2605 /* The preferred zone is used for statistics later */
2606 first_zones_zonelist(zonelist, high_zoneidx,
2607 nodemask ? : &cpuset_current_mems_allowed,
2609 if (!preferred_zone)
2613 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2614 alloc_flags |= ALLOC_CMA;
2616 /* First allocation attempt */
2617 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2618 zonelist, high_zoneidx, alloc_flags,
2619 preferred_zone, migratetype);
2620 if (unlikely(!page))
2621 page = __alloc_pages_slowpath(gfp_mask, order,
2622 zonelist, high_zoneidx, nodemask,
2623 preferred_zone, migratetype);
2625 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2629 * When updating a task's mems_allowed, it is possible to race with
2630 * parallel threads in such a way that an allocation can fail while
2631 * the mask is being updated. If a page allocation is about to fail,
2632 * check if the cpuset changed during allocation and if so, retry.
2634 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2639 EXPORT_SYMBOL(__alloc_pages_nodemask);
2642 * Common helper functions.
2644 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2649 * __get_free_pages() returns a 32-bit address, which cannot represent
2652 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2654 page = alloc_pages(gfp_mask, order);
2657 return (unsigned long) page_address(page);
2659 EXPORT_SYMBOL(__get_free_pages);
2661 unsigned long get_zeroed_page(gfp_t gfp_mask)
2663 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2665 EXPORT_SYMBOL(get_zeroed_page);
2667 void __free_pages(struct page *page, unsigned int order)
2669 if (put_page_testzero(page)) {
2671 free_hot_cold_page(page, 0);
2673 __free_pages_ok(page, order);
2677 EXPORT_SYMBOL(__free_pages);
2679 void free_pages(unsigned long addr, unsigned int order)
2682 VM_BUG_ON(!virt_addr_valid((void *)addr));
2683 __free_pages(virt_to_page((void *)addr), order);
2687 EXPORT_SYMBOL(free_pages);
2689 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2692 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2693 unsigned long used = addr + PAGE_ALIGN(size);
2695 split_page(virt_to_page((void *)addr), order);
2696 while (used < alloc_end) {
2701 return (void *)addr;
2705 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2706 * @size: the number of bytes to allocate
2707 * @gfp_mask: GFP flags for the allocation
2709 * This function is similar to alloc_pages(), except that it allocates the
2710 * minimum number of pages to satisfy the request. alloc_pages() can only
2711 * allocate memory in power-of-two pages.
2713 * This function is also limited by MAX_ORDER.
2715 * Memory allocated by this function must be released by free_pages_exact().
2717 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2719 unsigned int order = get_order(size);
2722 addr = __get_free_pages(gfp_mask, order);
2723 return make_alloc_exact(addr, order, size);
2725 EXPORT_SYMBOL(alloc_pages_exact);
2728 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2730 * @nid: the preferred node ID where memory should be allocated
2731 * @size: the number of bytes to allocate
2732 * @gfp_mask: GFP flags for the allocation
2734 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2736 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2739 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2741 unsigned order = get_order(size);
2742 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2745 return make_alloc_exact((unsigned long)page_address(p), order, size);
2747 EXPORT_SYMBOL(alloc_pages_exact_nid);
2750 * free_pages_exact - release memory allocated via alloc_pages_exact()
2751 * @virt: the value returned by alloc_pages_exact.
2752 * @size: size of allocation, same value as passed to alloc_pages_exact().
2754 * Release the memory allocated by a previous call to alloc_pages_exact.
2756 void free_pages_exact(void *virt, size_t size)
2758 unsigned long addr = (unsigned long)virt;
2759 unsigned long end = addr + PAGE_ALIGN(size);
2761 while (addr < end) {
2766 EXPORT_SYMBOL(free_pages_exact);
2768 static unsigned int nr_free_zone_pages(int offset)
2773 /* Just pick one node, since fallback list is circular */
2774 unsigned int sum = 0;
2776 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2778 for_each_zone_zonelist(zone, z, zonelist, offset) {
2779 unsigned long size = zone->present_pages;
2780 unsigned long high = high_wmark_pages(zone);
2789 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2791 unsigned int nr_free_buffer_pages(void)
2793 return nr_free_zone_pages(gfp_zone(GFP_USER));
2795 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2798 * Amount of free RAM allocatable within all zones
2800 unsigned int nr_free_pagecache_pages(void)
2802 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2805 static inline void show_node(struct zone *zone)
2808 printk("Node %d ", zone_to_nid(zone));
2811 void si_meminfo(struct sysinfo *val)
2813 val->totalram = totalram_pages;
2815 val->freeram = global_page_state(NR_FREE_PAGES);
2816 val->bufferram = nr_blockdev_pages();
2817 val->totalhigh = totalhigh_pages;
2818 val->freehigh = nr_free_highpages();
2819 val->mem_unit = PAGE_SIZE;
2822 EXPORT_SYMBOL(si_meminfo);
2825 void si_meminfo_node(struct sysinfo *val, int nid)
2827 pg_data_t *pgdat = NODE_DATA(nid);
2829 val->totalram = pgdat->node_present_pages;
2830 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2831 #ifdef CONFIG_HIGHMEM
2832 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2833 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2839 val->mem_unit = PAGE_SIZE;
2844 * Determine whether the node should be displayed or not, depending on whether
2845 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2847 bool skip_free_areas_node(unsigned int flags, int nid)
2850 unsigned int cpuset_mems_cookie;
2852 if (!(flags & SHOW_MEM_FILTER_NODES))
2856 cpuset_mems_cookie = get_mems_allowed();
2857 ret = !node_isset(nid, cpuset_current_mems_allowed);
2858 } while (!put_mems_allowed(cpuset_mems_cookie));
2863 #define K(x) ((x) << (PAGE_SHIFT-10))
2866 * Show free area list (used inside shift_scroll-lock stuff)
2867 * We also calculate the percentage fragmentation. We do this by counting the
2868 * memory on each free list with the exception of the first item on the list.
2869 * Suppresses nodes that are not allowed by current's cpuset if
2870 * SHOW_MEM_FILTER_NODES is passed.
2872 void show_free_areas(unsigned int filter)
2877 for_each_populated_zone(zone) {
2878 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2881 printk("%s per-cpu:\n", zone->name);
2883 for_each_online_cpu(cpu) {
2884 struct per_cpu_pageset *pageset;
2886 pageset = per_cpu_ptr(zone->pageset, cpu);
2888 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2889 cpu, pageset->pcp.high,
2890 pageset->pcp.batch, pageset->pcp.count);
2894 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2895 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2897 " dirty:%lu writeback:%lu unstable:%lu\n"
2898 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2899 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2901 global_page_state(NR_ACTIVE_ANON),
2902 global_page_state(NR_INACTIVE_ANON),
2903 global_page_state(NR_ISOLATED_ANON),
2904 global_page_state(NR_ACTIVE_FILE),
2905 global_page_state(NR_INACTIVE_FILE),
2906 global_page_state(NR_ISOLATED_FILE),
2907 global_page_state(NR_UNEVICTABLE),
2908 global_page_state(NR_FILE_DIRTY),
2909 global_page_state(NR_WRITEBACK),
2910 global_page_state(NR_UNSTABLE_NFS),
2911 global_page_state(NR_FREE_PAGES),
2912 global_page_state(NR_SLAB_RECLAIMABLE),
2913 global_page_state(NR_SLAB_UNRECLAIMABLE),
2914 global_page_state(NR_FILE_MAPPED),
2915 global_page_state(NR_SHMEM),
2916 global_page_state(NR_PAGETABLE),
2917 global_page_state(NR_BOUNCE),
2918 global_page_state(NR_FREE_CMA_PAGES));
2920 for_each_populated_zone(zone) {
2923 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2931 " active_anon:%lukB"
2932 " inactive_anon:%lukB"
2933 " active_file:%lukB"
2934 " inactive_file:%lukB"
2935 " unevictable:%lukB"
2936 " isolated(anon):%lukB"
2937 " isolated(file):%lukB"
2944 " slab_reclaimable:%lukB"
2945 " slab_unreclaimable:%lukB"
2946 " kernel_stack:%lukB"
2951 " writeback_tmp:%lukB"
2952 " pages_scanned:%lu"
2953 " all_unreclaimable? %s"
2956 K(zone_page_state(zone, NR_FREE_PAGES)),
2957 K(min_wmark_pages(zone)),
2958 K(low_wmark_pages(zone)),
2959 K(high_wmark_pages(zone)),
2960 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2961 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2962 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2963 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2964 K(zone_page_state(zone, NR_UNEVICTABLE)),
2965 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2966 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2967 K(zone->present_pages),
2968 K(zone_page_state(zone, NR_MLOCK)),
2969 K(zone_page_state(zone, NR_FILE_DIRTY)),
2970 K(zone_page_state(zone, NR_WRITEBACK)),
2971 K(zone_page_state(zone, NR_FILE_MAPPED)),
2972 K(zone_page_state(zone, NR_SHMEM)),
2973 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2974 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2975 zone_page_state(zone, NR_KERNEL_STACK) *
2977 K(zone_page_state(zone, NR_PAGETABLE)),
2978 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2979 K(zone_page_state(zone, NR_BOUNCE)),
2980 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
2981 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2982 zone->pages_scanned,
2983 (zone->all_unreclaimable ? "yes" : "no")
2985 printk("lowmem_reserve[]:");
2986 for (i = 0; i < MAX_NR_ZONES; i++)
2987 printk(" %lu", zone->lowmem_reserve[i]);
2991 for_each_populated_zone(zone) {
2992 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2994 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2997 printk("%s: ", zone->name);
2999 spin_lock_irqsave(&zone->lock, flags);
3000 for (order = 0; order < MAX_ORDER; order++) {
3001 nr[order] = zone->free_area[order].nr_free;
3002 total += nr[order] << order;
3004 spin_unlock_irqrestore(&zone->lock, flags);
3005 for (order = 0; order < MAX_ORDER; order++)
3006 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3007 printk("= %lukB\n", K(total));
3010 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3012 show_swap_cache_info();
3015 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3017 zoneref->zone = zone;
3018 zoneref->zone_idx = zone_idx(zone);
3022 * Builds allocation fallback zone lists.
3024 * Add all populated zones of a node to the zonelist.
3026 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3027 int nr_zones, enum zone_type zone_type)
3031 BUG_ON(zone_type >= MAX_NR_ZONES);
3036 zone = pgdat->node_zones + zone_type;
3037 if (populated_zone(zone)) {
3038 zoneref_set_zone(zone,
3039 &zonelist->_zonerefs[nr_zones++]);
3040 check_highest_zone(zone_type);
3043 } while (zone_type);
3050 * 0 = automatic detection of better ordering.
3051 * 1 = order by ([node] distance, -zonetype)
3052 * 2 = order by (-zonetype, [node] distance)
3054 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3055 * the same zonelist. So only NUMA can configure this param.
3057 #define ZONELIST_ORDER_DEFAULT 0
3058 #define ZONELIST_ORDER_NODE 1
3059 #define ZONELIST_ORDER_ZONE 2
3061 /* zonelist order in the kernel.
3062 * set_zonelist_order() will set this to NODE or ZONE.
3064 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3065 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3069 /* The value user specified ....changed by config */
3070 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3071 /* string for sysctl */
3072 #define NUMA_ZONELIST_ORDER_LEN 16
3073 char numa_zonelist_order[16] = "default";
3076 * interface for configure zonelist ordering.
3077 * command line option "numa_zonelist_order"
3078 * = "[dD]efault - default, automatic configuration.
3079 * = "[nN]ode - order by node locality, then by zone within node
3080 * = "[zZ]one - order by zone, then by locality within zone
3083 static int __parse_numa_zonelist_order(char *s)
3085 if (*s == 'd' || *s == 'D') {
3086 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3087 } else if (*s == 'n' || *s == 'N') {
3088 user_zonelist_order = ZONELIST_ORDER_NODE;
3089 } else if (*s == 'z' || *s == 'Z') {
3090 user_zonelist_order = ZONELIST_ORDER_ZONE;
3093 "Ignoring invalid numa_zonelist_order value: "
3100 static __init int setup_numa_zonelist_order(char *s)
3107 ret = __parse_numa_zonelist_order(s);
3109 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3113 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3116 * sysctl handler for numa_zonelist_order
3118 int numa_zonelist_order_handler(ctl_table *table, int write,
3119 void __user *buffer, size_t *length,
3122 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3124 static DEFINE_MUTEX(zl_order_mutex);
3126 mutex_lock(&zl_order_mutex);
3128 strcpy(saved_string, (char*)table->data);
3129 ret = proc_dostring(table, write, buffer, length, ppos);
3133 int oldval = user_zonelist_order;
3134 if (__parse_numa_zonelist_order((char*)table->data)) {
3136 * bogus value. restore saved string
3138 strncpy((char*)table->data, saved_string,
3139 NUMA_ZONELIST_ORDER_LEN);
3140 user_zonelist_order = oldval;
3141 } else if (oldval != user_zonelist_order) {
3142 mutex_lock(&zonelists_mutex);
3143 build_all_zonelists(NULL, NULL);
3144 mutex_unlock(&zonelists_mutex);
3148 mutex_unlock(&zl_order_mutex);
3153 #define MAX_NODE_LOAD (nr_online_nodes)
3154 static int node_load[MAX_NUMNODES];
3157 * find_next_best_node - find the next node that should appear in a given node's fallback list
3158 * @node: node whose fallback list we're appending
3159 * @used_node_mask: nodemask_t of already used nodes
3161 * We use a number of factors to determine which is the next node that should
3162 * appear on a given node's fallback list. The node should not have appeared
3163 * already in @node's fallback list, and it should be the next closest node
3164 * according to the distance array (which contains arbitrary distance values
3165 * from each node to each node in the system), and should also prefer nodes
3166 * with no CPUs, since presumably they'll have very little allocation pressure
3167 * on them otherwise.
3168 * It returns -1 if no node is found.
3170 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3173 int min_val = INT_MAX;
3175 const struct cpumask *tmp = cpumask_of_node(0);
3177 /* Use the local node if we haven't already */
3178 if (!node_isset(node, *used_node_mask)) {
3179 node_set(node, *used_node_mask);
3183 for_each_node_state(n, N_HIGH_MEMORY) {
3185 /* Don't want a node to appear more than once */
3186 if (node_isset(n, *used_node_mask))
3189 /* Use the distance array to find the distance */
3190 val = node_distance(node, n);
3192 /* Penalize nodes under us ("prefer the next node") */
3195 /* Give preference to headless and unused nodes */
3196 tmp = cpumask_of_node(n);
3197 if (!cpumask_empty(tmp))
3198 val += PENALTY_FOR_NODE_WITH_CPUS;
3200 /* Slight preference for less loaded node */
3201 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3202 val += node_load[n];
3204 if (val < min_val) {
3211 node_set(best_node, *used_node_mask);
3218 * Build zonelists ordered by node and zones within node.
3219 * This results in maximum locality--normal zone overflows into local
3220 * DMA zone, if any--but risks exhausting DMA zone.
3222 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3225 struct zonelist *zonelist;
3227 zonelist = &pgdat->node_zonelists[0];
3228 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3230 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3232 zonelist->_zonerefs[j].zone = NULL;
3233 zonelist->_zonerefs[j].zone_idx = 0;
3237 * Build gfp_thisnode zonelists
3239 static void build_thisnode_zonelists(pg_data_t *pgdat)
3242 struct zonelist *zonelist;
3244 zonelist = &pgdat->node_zonelists[1];
3245 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3246 zonelist->_zonerefs[j].zone = NULL;
3247 zonelist->_zonerefs[j].zone_idx = 0;
3251 * Build zonelists ordered by zone and nodes within zones.
3252 * This results in conserving DMA zone[s] until all Normal memory is
3253 * exhausted, but results in overflowing to remote node while memory
3254 * may still exist in local DMA zone.
3256 static int node_order[MAX_NUMNODES];
3258 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3261 int zone_type; /* needs to be signed */
3263 struct zonelist *zonelist;
3265 zonelist = &pgdat->node_zonelists[0];
3267 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3268 for (j = 0; j < nr_nodes; j++) {
3269 node = node_order[j];
3270 z = &NODE_DATA(node)->node_zones[zone_type];
3271 if (populated_zone(z)) {
3273 &zonelist->_zonerefs[pos++]);
3274 check_highest_zone(zone_type);
3278 zonelist->_zonerefs[pos].zone = NULL;
3279 zonelist->_zonerefs[pos].zone_idx = 0;
3282 static int default_zonelist_order(void)
3285 unsigned long low_kmem_size,total_size;
3289 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3290 * If they are really small and used heavily, the system can fall
3291 * into OOM very easily.
3292 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3294 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3297 for_each_online_node(nid) {
3298 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3299 z = &NODE_DATA(nid)->node_zones[zone_type];
3300 if (populated_zone(z)) {
3301 if (zone_type < ZONE_NORMAL)
3302 low_kmem_size += z->present_pages;
3303 total_size += z->present_pages;
3304 } else if (zone_type == ZONE_NORMAL) {
3306 * If any node has only lowmem, then node order
3307 * is preferred to allow kernel allocations
3308 * locally; otherwise, they can easily infringe
3309 * on other nodes when there is an abundance of
3310 * lowmem available to allocate from.
3312 return ZONELIST_ORDER_NODE;
3316 if (!low_kmem_size || /* there are no DMA area. */
3317 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3318 return ZONELIST_ORDER_NODE;
3320 * look into each node's config.
3321 * If there is a node whose DMA/DMA32 memory is very big area on
3322 * local memory, NODE_ORDER may be suitable.
3324 average_size = total_size /
3325 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3326 for_each_online_node(nid) {
3329 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3330 z = &NODE_DATA(nid)->node_zones[zone_type];
3331 if (populated_zone(z)) {
3332 if (zone_type < ZONE_NORMAL)
3333 low_kmem_size += z->present_pages;
3334 total_size += z->present_pages;
3337 if (low_kmem_size &&
3338 total_size > average_size && /* ignore small node */
3339 low_kmem_size > total_size * 70/100)
3340 return ZONELIST_ORDER_NODE;
3342 return ZONELIST_ORDER_ZONE;
3345 static void set_zonelist_order(void)
3347 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3348 current_zonelist_order = default_zonelist_order();
3350 current_zonelist_order = user_zonelist_order;
3353 static void build_zonelists(pg_data_t *pgdat)
3357 nodemask_t used_mask;
3358 int local_node, prev_node;
3359 struct zonelist *zonelist;
3360 int order = current_zonelist_order;
3362 /* initialize zonelists */
3363 for (i = 0; i < MAX_ZONELISTS; i++) {
3364 zonelist = pgdat->node_zonelists + i;
3365 zonelist->_zonerefs[0].zone = NULL;
3366 zonelist->_zonerefs[0].zone_idx = 0;
3369 /* NUMA-aware ordering of nodes */
3370 local_node = pgdat->node_id;
3371 load = nr_online_nodes;
3372 prev_node = local_node;
3373 nodes_clear(used_mask);
3375 memset(node_order, 0, sizeof(node_order));
3378 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3379 int distance = node_distance(local_node, node);
3382 * If another node is sufficiently far away then it is better
3383 * to reclaim pages in a zone before going off node.
3385 if (distance > RECLAIM_DISTANCE)
3386 zone_reclaim_mode = 1;
3389 * We don't want to pressure a particular node.
3390 * So adding penalty to the first node in same
3391 * distance group to make it round-robin.
3393 if (distance != node_distance(local_node, prev_node))
3394 node_load[node] = load;
3398 if (order == ZONELIST_ORDER_NODE)
3399 build_zonelists_in_node_order(pgdat, node);
3401 node_order[j++] = node; /* remember order */
3404 if (order == ZONELIST_ORDER_ZONE) {
3405 /* calculate node order -- i.e., DMA last! */
3406 build_zonelists_in_zone_order(pgdat, j);
3409 build_thisnode_zonelists(pgdat);
3412 /* Construct the zonelist performance cache - see further mmzone.h */
3413 static void build_zonelist_cache(pg_data_t *pgdat)
3415 struct zonelist *zonelist;
3416 struct zonelist_cache *zlc;
3419 zonelist = &pgdat->node_zonelists[0];
3420 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3421 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3422 for (z = zonelist->_zonerefs; z->zone; z++)
3423 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3426 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3428 * Return node id of node used for "local" allocations.
3429 * I.e., first node id of first zone in arg node's generic zonelist.
3430 * Used for initializing percpu 'numa_mem', which is used primarily
3431 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3433 int local_memory_node(int node)
3437 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3438 gfp_zone(GFP_KERNEL),
3445 #else /* CONFIG_NUMA */
3447 static void set_zonelist_order(void)
3449 current_zonelist_order = ZONELIST_ORDER_ZONE;
3452 static void build_zonelists(pg_data_t *pgdat)
3454 int node, local_node;
3456 struct zonelist *zonelist;
3458 local_node = pgdat->node_id;
3460 zonelist = &pgdat->node_zonelists[0];
3461 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3464 * Now we build the zonelist so that it contains the zones
3465 * of all the other nodes.
3466 * We don't want to pressure a particular node, so when
3467 * building the zones for node N, we make sure that the
3468 * zones coming right after the local ones are those from
3469 * node N+1 (modulo N)
3471 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3472 if (!node_online(node))
3474 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3477 for (node = 0; node < local_node; node++) {
3478 if (!node_online(node))
3480 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3484 zonelist->_zonerefs[j].zone = NULL;
3485 zonelist->_zonerefs[j].zone_idx = 0;
3488 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3489 static void build_zonelist_cache(pg_data_t *pgdat)
3491 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3494 #endif /* CONFIG_NUMA */
3497 * Boot pageset table. One per cpu which is going to be used for all
3498 * zones and all nodes. The parameters will be set in such a way
3499 * that an item put on a list will immediately be handed over to
3500 * the buddy list. This is safe since pageset manipulation is done
3501 * with interrupts disabled.
3503 * The boot_pagesets must be kept even after bootup is complete for
3504 * unused processors and/or zones. They do play a role for bootstrapping
3505 * hotplugged processors.
3507 * zoneinfo_show() and maybe other functions do
3508 * not check if the processor is online before following the pageset pointer.
3509 * Other parts of the kernel may not check if the zone is available.
3511 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3512 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3513 static void setup_zone_pageset(struct zone *zone);
3516 * Global mutex to protect against size modification of zonelists
3517 * as well as to serialize pageset setup for the new populated zone.
3519 DEFINE_MUTEX(zonelists_mutex);
3521 /* return values int ....just for stop_machine() */
3522 static int __build_all_zonelists(void *data)
3526 pg_data_t *self = data;
3529 memset(node_load, 0, sizeof(node_load));
3532 if (self && !node_online(self->node_id)) {
3533 build_zonelists(self);
3534 build_zonelist_cache(self);
3537 for_each_online_node(nid) {
3538 pg_data_t *pgdat = NODE_DATA(nid);
3540 build_zonelists(pgdat);
3541 build_zonelist_cache(pgdat);
3545 * Initialize the boot_pagesets that are going to be used
3546 * for bootstrapping processors. The real pagesets for
3547 * each zone will be allocated later when the per cpu
3548 * allocator is available.
3550 * boot_pagesets are used also for bootstrapping offline
3551 * cpus if the system is already booted because the pagesets
3552 * are needed to initialize allocators on a specific cpu too.
3553 * F.e. the percpu allocator needs the page allocator which
3554 * needs the percpu allocator in order to allocate its pagesets
3555 * (a chicken-egg dilemma).
3557 for_each_possible_cpu(cpu) {
3558 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3560 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3562 * We now know the "local memory node" for each node--
3563 * i.e., the node of the first zone in the generic zonelist.
3564 * Set up numa_mem percpu variable for on-line cpus. During
3565 * boot, only the boot cpu should be on-line; we'll init the
3566 * secondary cpus' numa_mem as they come on-line. During
3567 * node/memory hotplug, we'll fixup all on-line cpus.
3569 if (cpu_online(cpu))
3570 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3578 * Called with zonelists_mutex held always
3579 * unless system_state == SYSTEM_BOOTING.
3581 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3583 set_zonelist_order();
3585 if (system_state == SYSTEM_BOOTING) {
3586 __build_all_zonelists(NULL);
3587 mminit_verify_zonelist();
3588 cpuset_init_current_mems_allowed();
3590 /* we have to stop all cpus to guarantee there is no user
3592 #ifdef CONFIG_MEMORY_HOTPLUG
3594 setup_zone_pageset(zone);
3596 stop_machine(__build_all_zonelists, pgdat, NULL);
3597 /* cpuset refresh routine should be here */
3599 vm_total_pages = nr_free_pagecache_pages();
3601 * Disable grouping by mobility if the number of pages in the
3602 * system is too low to allow the mechanism to work. It would be
3603 * more accurate, but expensive to check per-zone. This check is
3604 * made on memory-hotadd so a system can start with mobility
3605 * disabled and enable it later
3607 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3608 page_group_by_mobility_disabled = 1;
3610 page_group_by_mobility_disabled = 0;
3612 printk("Built %i zonelists in %s order, mobility grouping %s. "
3613 "Total pages: %ld\n",
3615 zonelist_order_name[current_zonelist_order],
3616 page_group_by_mobility_disabled ? "off" : "on",
3619 printk("Policy zone: %s\n", zone_names[policy_zone]);
3624 * Helper functions to size the waitqueue hash table.
3625 * Essentially these want to choose hash table sizes sufficiently
3626 * large so that collisions trying to wait on pages are rare.
3627 * But in fact, the number of active page waitqueues on typical
3628 * systems is ridiculously low, less than 200. So this is even
3629 * conservative, even though it seems large.
3631 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3632 * waitqueues, i.e. the size of the waitq table given the number of pages.
3634 #define PAGES_PER_WAITQUEUE 256
3636 #ifndef CONFIG_MEMORY_HOTPLUG
3637 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3639 unsigned long size = 1;
3641 pages /= PAGES_PER_WAITQUEUE;
3643 while (size < pages)
3647 * Once we have dozens or even hundreds of threads sleeping
3648 * on IO we've got bigger problems than wait queue collision.
3649 * Limit the size of the wait table to a reasonable size.
3651 size = min(size, 4096UL);
3653 return max(size, 4UL);
3657 * A zone's size might be changed by hot-add, so it is not possible to determine
3658 * a suitable size for its wait_table. So we use the maximum size now.
3660 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3662 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3663 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3664 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3666 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3667 * or more by the traditional way. (See above). It equals:
3669 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3670 * ia64(16K page size) : = ( 8G + 4M)byte.
3671 * powerpc (64K page size) : = (32G +16M)byte.
3673 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3680 * This is an integer logarithm so that shifts can be used later
3681 * to extract the more random high bits from the multiplicative
3682 * hash function before the remainder is taken.
3684 static inline unsigned long wait_table_bits(unsigned long size)
3689 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3692 * Check if a pageblock contains reserved pages
3694 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3698 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3699 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3706 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3707 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3708 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3709 * higher will lead to a bigger reserve which will get freed as contiguous
3710 * blocks as reclaim kicks in
3712 static void setup_zone_migrate_reserve(struct zone *zone)
3714 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3716 unsigned long block_migratetype;
3720 * Get the start pfn, end pfn and the number of blocks to reserve
3721 * We have to be careful to be aligned to pageblock_nr_pages to
3722 * make sure that we always check pfn_valid for the first page in
3725 start_pfn = zone->zone_start_pfn;
3726 end_pfn = start_pfn + zone->spanned_pages;
3727 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3728 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3732 * Reserve blocks are generally in place to help high-order atomic
3733 * allocations that are short-lived. A min_free_kbytes value that
3734 * would result in more than 2 reserve blocks for atomic allocations
3735 * is assumed to be in place to help anti-fragmentation for the
3736 * future allocation of hugepages at runtime.
3738 reserve = min(2, reserve);
3740 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3741 if (!pfn_valid(pfn))
3743 page = pfn_to_page(pfn);
3745 /* Watch out for overlapping nodes */
3746 if (page_to_nid(page) != zone_to_nid(zone))
3749 block_migratetype = get_pageblock_migratetype(page);
3751 /* Only test what is necessary when the reserves are not met */
3754 * Blocks with reserved pages will never free, skip
3757 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3758 if (pageblock_is_reserved(pfn, block_end_pfn))
3761 /* If this block is reserved, account for it */
3762 if (block_migratetype == MIGRATE_RESERVE) {
3767 /* Suitable for reserving if this block is movable */
3768 if (block_migratetype == MIGRATE_MOVABLE) {
3769 set_pageblock_migratetype(page,
3771 move_freepages_block(zone, page,
3779 * If the reserve is met and this is a previous reserved block,
3782 if (block_migratetype == MIGRATE_RESERVE) {
3783 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3784 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3790 * Initially all pages are reserved - free ones are freed
3791 * up by free_all_bootmem() once the early boot process is
3792 * done. Non-atomic initialization, single-pass.
3794 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3795 unsigned long start_pfn, enum memmap_context context)
3798 unsigned long end_pfn = start_pfn + size;
3802 if (highest_memmap_pfn < end_pfn - 1)
3803 highest_memmap_pfn = end_pfn - 1;
3805 z = &NODE_DATA(nid)->node_zones[zone];
3806 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3808 * There can be holes in boot-time mem_map[]s
3809 * handed to this function. They do not
3810 * exist on hotplugged memory.
3812 if (context == MEMMAP_EARLY) {
3813 if (!early_pfn_valid(pfn))
3815 if (!early_pfn_in_nid(pfn, nid))
3818 page = pfn_to_page(pfn);
3819 set_page_links(page, zone, nid, pfn);
3820 mminit_verify_page_links(page, zone, nid, pfn);
3821 init_page_count(page);
3822 reset_page_mapcount(page);
3823 SetPageReserved(page);
3825 * Mark the block movable so that blocks are reserved for
3826 * movable at startup. This will force kernel allocations
3827 * to reserve their blocks rather than leaking throughout
3828 * the address space during boot when many long-lived
3829 * kernel allocations are made. Later some blocks near
3830 * the start are marked MIGRATE_RESERVE by
3831 * setup_zone_migrate_reserve()
3833 * bitmap is created for zone's valid pfn range. but memmap
3834 * can be created for invalid pages (for alignment)
3835 * check here not to call set_pageblock_migratetype() against
3838 if ((z->zone_start_pfn <= pfn)
3839 && (pfn < z->zone_start_pfn + z->spanned_pages)
3840 && !(pfn & (pageblock_nr_pages - 1)))
3841 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3843 INIT_LIST_HEAD(&page->lru);
3844 #ifdef WANT_PAGE_VIRTUAL
3845 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3846 if (!is_highmem_idx(zone))
3847 set_page_address(page, __va(pfn << PAGE_SHIFT));
3852 static void __meminit zone_init_free_lists(struct zone *zone)
3855 for_each_migratetype_order(order, t) {
3856 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3857 zone->free_area[order].nr_free = 0;
3861 #ifndef __HAVE_ARCH_MEMMAP_INIT
3862 #define memmap_init(size, nid, zone, start_pfn) \
3863 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3866 static int __meminit zone_batchsize(struct zone *zone)
3872 * The per-cpu-pages pools are set to around 1000th of the
3873 * size of the zone. But no more than 1/2 of a meg.
3875 * OK, so we don't know how big the cache is. So guess.
3877 batch = zone->present_pages / 1024;
3878 if (batch * PAGE_SIZE > 512 * 1024)
3879 batch = (512 * 1024) / PAGE_SIZE;
3880 batch /= 4; /* We effectively *= 4 below */
3885 * Clamp the batch to a 2^n - 1 value. Having a power
3886 * of 2 value was found to be more likely to have
3887 * suboptimal cache aliasing properties in some cases.
3889 * For example if 2 tasks are alternately allocating
3890 * batches of pages, one task can end up with a lot
3891 * of pages of one half of the possible page colors
3892 * and the other with pages of the other colors.
3894 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3899 /* The deferral and batching of frees should be suppressed under NOMMU
3902 * The problem is that NOMMU needs to be able to allocate large chunks
3903 * of contiguous memory as there's no hardware page translation to
3904 * assemble apparent contiguous memory from discontiguous pages.
3906 * Queueing large contiguous runs of pages for batching, however,
3907 * causes the pages to actually be freed in smaller chunks. As there
3908 * can be a significant delay between the individual batches being
3909 * recycled, this leads to the once large chunks of space being
3910 * fragmented and becoming unavailable for high-order allocations.
3916 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3918 struct per_cpu_pages *pcp;
3921 memset(p, 0, sizeof(*p));
3925 pcp->high = 6 * batch;
3926 pcp->batch = max(1UL, 1 * batch);
3927 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3928 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3932 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3933 * to the value high for the pageset p.
3936 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3939 struct per_cpu_pages *pcp;
3943 pcp->batch = max(1UL, high/4);
3944 if ((high/4) > (PAGE_SHIFT * 8))
3945 pcp->batch = PAGE_SHIFT * 8;
3948 static void __meminit setup_zone_pageset(struct zone *zone)
3952 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3954 for_each_possible_cpu(cpu) {
3955 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3957 setup_pageset(pcp, zone_batchsize(zone));
3959 if (percpu_pagelist_fraction)
3960 setup_pagelist_highmark(pcp,
3961 (zone->present_pages /
3962 percpu_pagelist_fraction));
3967 * Allocate per cpu pagesets and initialize them.
3968 * Before this call only boot pagesets were available.
3970 void __init setup_per_cpu_pageset(void)
3974 for_each_populated_zone(zone)
3975 setup_zone_pageset(zone);
3978 static noinline __init_refok
3979 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3982 struct pglist_data *pgdat = zone->zone_pgdat;
3986 * The per-page waitqueue mechanism uses hashed waitqueues
3989 zone->wait_table_hash_nr_entries =
3990 wait_table_hash_nr_entries(zone_size_pages);
3991 zone->wait_table_bits =
3992 wait_table_bits(zone->wait_table_hash_nr_entries);
3993 alloc_size = zone->wait_table_hash_nr_entries
3994 * sizeof(wait_queue_head_t);
3996 if (!slab_is_available()) {
3997 zone->wait_table = (wait_queue_head_t *)
3998 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4001 * This case means that a zone whose size was 0 gets new memory
4002 * via memory hot-add.
4003 * But it may be the case that a new node was hot-added. In
4004 * this case vmalloc() will not be able to use this new node's
4005 * memory - this wait_table must be initialized to use this new
4006 * node itself as well.
4007 * To use this new node's memory, further consideration will be
4010 zone->wait_table = vmalloc(alloc_size);
4012 if (!zone->wait_table)
4015 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4016 init_waitqueue_head(zone->wait_table + i);
4021 static __meminit void zone_pcp_init(struct zone *zone)
4024 * per cpu subsystem is not up at this point. The following code
4025 * relies on the ability of the linker to provide the
4026 * offset of a (static) per cpu variable into the per cpu area.
4028 zone->pageset = &boot_pageset;
4030 if (zone->present_pages)
4031 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4032 zone->name, zone->present_pages,
4033 zone_batchsize(zone));
4036 int __meminit init_currently_empty_zone(struct zone *zone,
4037 unsigned long zone_start_pfn,
4039 enum memmap_context context)
4041 struct pglist_data *pgdat = zone->zone_pgdat;
4043 ret = zone_wait_table_init(zone, size);
4046 pgdat->nr_zones = zone_idx(zone) + 1;
4048 zone->zone_start_pfn = zone_start_pfn;
4050 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4051 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4053 (unsigned long)zone_idx(zone),
4054 zone_start_pfn, (zone_start_pfn + size));
4056 zone_init_free_lists(zone);
4061 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4062 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4064 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4065 * Architectures may implement their own version but if add_active_range()
4066 * was used and there are no special requirements, this is a convenient
4069 int __meminit __early_pfn_to_nid(unsigned long pfn)
4071 unsigned long start_pfn, end_pfn;
4074 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4075 if (start_pfn <= pfn && pfn < end_pfn)
4077 /* This is a memory hole */
4080 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4082 int __meminit early_pfn_to_nid(unsigned long pfn)
4086 nid = __early_pfn_to_nid(pfn);
4089 /* just returns 0 */
4093 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4094 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4098 nid = __early_pfn_to_nid(pfn);
4099 if (nid >= 0 && nid != node)
4106 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4107 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4108 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4110 * If an architecture guarantees that all ranges registered with
4111 * add_active_ranges() contain no holes and may be freed, this
4112 * this function may be used instead of calling free_bootmem() manually.
4114 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4116 unsigned long start_pfn, end_pfn;
4119 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4120 start_pfn = min(start_pfn, max_low_pfn);
4121 end_pfn = min(end_pfn, max_low_pfn);
4123 if (start_pfn < end_pfn)
4124 free_bootmem_node(NODE_DATA(this_nid),
4125 PFN_PHYS(start_pfn),
4126 (end_pfn - start_pfn) << PAGE_SHIFT);
4131 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4132 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4134 * If an architecture guarantees that all ranges registered with
4135 * add_active_ranges() contain no holes and may be freed, this
4136 * function may be used instead of calling memory_present() manually.
4138 void __init sparse_memory_present_with_active_regions(int nid)
4140 unsigned long start_pfn, end_pfn;
4143 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4144 memory_present(this_nid, start_pfn, end_pfn);
4148 * get_pfn_range_for_nid - Return the start and end page frames for a node
4149 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4150 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4151 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4153 * It returns the start and end page frame of a node based on information
4154 * provided by an arch calling add_active_range(). If called for a node
4155 * with no available memory, a warning is printed and the start and end
4158 void __meminit get_pfn_range_for_nid(unsigned int nid,
4159 unsigned long *start_pfn, unsigned long *end_pfn)
4161 unsigned long this_start_pfn, this_end_pfn;
4167 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4168 *start_pfn = min(*start_pfn, this_start_pfn);
4169 *end_pfn = max(*end_pfn, this_end_pfn);
4172 if (*start_pfn == -1UL)
4177 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4178 * assumption is made that zones within a node are ordered in monotonic
4179 * increasing memory addresses so that the "highest" populated zone is used
4181 static void __init find_usable_zone_for_movable(void)
4184 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4185 if (zone_index == ZONE_MOVABLE)
4188 if (arch_zone_highest_possible_pfn[zone_index] >
4189 arch_zone_lowest_possible_pfn[zone_index])
4193 VM_BUG_ON(zone_index == -1);
4194 movable_zone = zone_index;
4198 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4199 * because it is sized independent of architecture. Unlike the other zones,
4200 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4201 * in each node depending on the size of each node and how evenly kernelcore
4202 * is distributed. This helper function adjusts the zone ranges
4203 * provided by the architecture for a given node by using the end of the
4204 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4205 * zones within a node are in order of monotonic increases memory addresses
4207 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4208 unsigned long zone_type,
4209 unsigned long node_start_pfn,
4210 unsigned long node_end_pfn,
4211 unsigned long *zone_start_pfn,
4212 unsigned long *zone_end_pfn)
4214 /* Only adjust if ZONE_MOVABLE is on this node */
4215 if (zone_movable_pfn[nid]) {
4216 /* Size ZONE_MOVABLE */
4217 if (zone_type == ZONE_MOVABLE) {
4218 *zone_start_pfn = zone_movable_pfn[nid];
4219 *zone_end_pfn = min(node_end_pfn,
4220 arch_zone_highest_possible_pfn[movable_zone]);
4222 /* Adjust for ZONE_MOVABLE starting within this range */
4223 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4224 *zone_end_pfn > zone_movable_pfn[nid]) {
4225 *zone_end_pfn = zone_movable_pfn[nid];
4227 /* Check if this whole range is within ZONE_MOVABLE */
4228 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4229 *zone_start_pfn = *zone_end_pfn;
4234 * Return the number of pages a zone spans in a node, including holes
4235 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4237 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4238 unsigned long zone_type,
4239 unsigned long *ignored)
4241 unsigned long node_start_pfn, node_end_pfn;
4242 unsigned long zone_start_pfn, zone_end_pfn;
4244 /* Get the start and end of the node and zone */
4245 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4246 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4247 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4248 adjust_zone_range_for_zone_movable(nid, zone_type,
4249 node_start_pfn, node_end_pfn,
4250 &zone_start_pfn, &zone_end_pfn);
4252 /* Check that this node has pages within the zone's required range */
4253 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4256 /* Move the zone boundaries inside the node if necessary */
4257 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4258 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4260 /* Return the spanned pages */
4261 return zone_end_pfn - zone_start_pfn;
4265 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4266 * then all holes in the requested range will be accounted for.
4268 unsigned long __meminit __absent_pages_in_range(int nid,
4269 unsigned long range_start_pfn,
4270 unsigned long range_end_pfn)
4272 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4273 unsigned long start_pfn, end_pfn;
4276 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4277 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4278 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4279 nr_absent -= end_pfn - start_pfn;
4285 * absent_pages_in_range - Return number of page frames in holes within a range
4286 * @start_pfn: The start PFN to start searching for holes
4287 * @end_pfn: The end PFN to stop searching for holes
4289 * It returns the number of pages frames in memory holes within a range.
4291 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4292 unsigned long end_pfn)
4294 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4297 /* Return the number of page frames in holes in a zone on a node */
4298 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4299 unsigned long zone_type,
4300 unsigned long *ignored)
4302 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4303 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4304 unsigned long node_start_pfn, node_end_pfn;
4305 unsigned long zone_start_pfn, zone_end_pfn;
4307 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4308 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4309 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4311 adjust_zone_range_for_zone_movable(nid, zone_type,
4312 node_start_pfn, node_end_pfn,
4313 &zone_start_pfn, &zone_end_pfn);
4314 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4317 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4318 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4319 unsigned long zone_type,
4320 unsigned long *zones_size)
4322 return zones_size[zone_type];
4325 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4326 unsigned long zone_type,
4327 unsigned long *zholes_size)
4332 return zholes_size[zone_type];
4335 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4337 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4338 unsigned long *zones_size, unsigned long *zholes_size)
4340 unsigned long realtotalpages, totalpages = 0;
4343 for (i = 0; i < MAX_NR_ZONES; i++)
4344 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4346 pgdat->node_spanned_pages = totalpages;
4348 realtotalpages = totalpages;
4349 for (i = 0; i < MAX_NR_ZONES; i++)
4351 zone_absent_pages_in_node(pgdat->node_id, i,
4353 pgdat->node_present_pages = realtotalpages;
4354 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4358 #ifndef CONFIG_SPARSEMEM
4360 * Calculate the size of the zone->blockflags rounded to an unsigned long
4361 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4362 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4363 * round what is now in bits to nearest long in bits, then return it in
4366 static unsigned long __init usemap_size(unsigned long zonesize)
4368 unsigned long usemapsize;
4370 usemapsize = roundup(zonesize, pageblock_nr_pages);
4371 usemapsize = usemapsize >> pageblock_order;
4372 usemapsize *= NR_PAGEBLOCK_BITS;
4373 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4375 return usemapsize / 8;
4378 static void __init setup_usemap(struct pglist_data *pgdat,
4379 struct zone *zone, unsigned long zonesize)
4381 unsigned long usemapsize = usemap_size(zonesize);
4382 zone->pageblock_flags = NULL;
4384 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4388 static inline void setup_usemap(struct pglist_data *pgdat,
4389 struct zone *zone, unsigned long zonesize) {}
4390 #endif /* CONFIG_SPARSEMEM */
4392 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4394 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4395 void __init set_pageblock_order(void)
4399 /* Check that pageblock_nr_pages has not already been setup */
4400 if (pageblock_order)
4403 if (HPAGE_SHIFT > PAGE_SHIFT)
4404 order = HUGETLB_PAGE_ORDER;
4406 order = MAX_ORDER - 1;
4409 * Assume the largest contiguous order of interest is a huge page.
4410 * This value may be variable depending on boot parameters on IA64 and
4413 pageblock_order = order;
4415 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4418 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4419 * is unused as pageblock_order is set at compile-time. See
4420 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4423 void __init set_pageblock_order(void)
4427 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4430 * Set up the zone data structures:
4431 * - mark all pages reserved
4432 * - mark all memory queues empty
4433 * - clear the memory bitmaps
4435 * NOTE: pgdat should get zeroed by caller.
4437 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4438 unsigned long *zones_size, unsigned long *zholes_size)
4441 int nid = pgdat->node_id;
4442 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4445 pgdat_resize_init(pgdat);
4446 init_waitqueue_head(&pgdat->kswapd_wait);
4447 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4448 pgdat_page_cgroup_init(pgdat);
4450 for (j = 0; j < MAX_NR_ZONES; j++) {
4451 struct zone *zone = pgdat->node_zones + j;
4452 unsigned long size, realsize, memmap_pages;
4454 size = zone_spanned_pages_in_node(nid, j, zones_size);
4455 realsize = size - zone_absent_pages_in_node(nid, j,
4459 * Adjust realsize so that it accounts for how much memory
4460 * is used by this zone for memmap. This affects the watermark
4461 * and per-cpu initialisations
4464 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4465 if (realsize >= memmap_pages) {
4466 realsize -= memmap_pages;
4469 " %s zone: %lu pages used for memmap\n",
4470 zone_names[j], memmap_pages);
4473 " %s zone: %lu pages exceeds realsize %lu\n",
4474 zone_names[j], memmap_pages, realsize);
4476 /* Account for reserved pages */
4477 if (j == 0 && realsize > dma_reserve) {
4478 realsize -= dma_reserve;
4479 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4480 zone_names[0], dma_reserve);
4483 if (!is_highmem_idx(j))
4484 nr_kernel_pages += realsize;
4485 nr_all_pages += realsize;
4487 zone->spanned_pages = size;
4488 zone->present_pages = realsize;
4489 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
4490 zone->compact_cached_free_pfn = zone->zone_start_pfn +
4491 zone->spanned_pages;
4492 zone->compact_cached_free_pfn &= ~(pageblock_nr_pages-1);
4496 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4498 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4500 zone->name = zone_names[j];
4501 spin_lock_init(&zone->lock);
4502 spin_lock_init(&zone->lru_lock);
4503 zone_seqlock_init(zone);
4504 zone->zone_pgdat = pgdat;
4506 zone_pcp_init(zone);
4507 lruvec_init(&zone->lruvec, zone);
4511 set_pageblock_order();
4512 setup_usemap(pgdat, zone, size);
4513 ret = init_currently_empty_zone(zone, zone_start_pfn,
4514 size, MEMMAP_EARLY);
4516 memmap_init(size, nid, j, zone_start_pfn);
4517 zone_start_pfn += size;
4521 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4523 /* Skip empty nodes */
4524 if (!pgdat->node_spanned_pages)
4527 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4528 /* ia64 gets its own node_mem_map, before this, without bootmem */
4529 if (!pgdat->node_mem_map) {
4530 unsigned long size, start, end;
4534 * The zone's endpoints aren't required to be MAX_ORDER
4535 * aligned but the node_mem_map endpoints must be in order
4536 * for the buddy allocator to function correctly.
4538 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4539 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4540 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4541 size = (end - start) * sizeof(struct page);
4542 map = alloc_remap(pgdat->node_id, size);
4544 map = alloc_bootmem_node_nopanic(pgdat, size);
4545 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4547 #ifndef CONFIG_NEED_MULTIPLE_NODES
4549 * With no DISCONTIG, the global mem_map is just set as node 0's
4551 if (pgdat == NODE_DATA(0)) {
4552 mem_map = NODE_DATA(0)->node_mem_map;
4553 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4554 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4555 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4556 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4559 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4562 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4563 unsigned long node_start_pfn, unsigned long *zholes_size)
4565 pg_data_t *pgdat = NODE_DATA(nid);
4567 /* pg_data_t should be reset to zero when it's allocated */
4568 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4570 pgdat->node_id = nid;
4571 pgdat->node_start_pfn = node_start_pfn;
4572 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4574 alloc_node_mem_map(pgdat);
4575 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4576 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4577 nid, (unsigned long)pgdat,
4578 (unsigned long)pgdat->node_mem_map);
4581 free_area_init_core(pgdat, zones_size, zholes_size);
4584 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4586 #if MAX_NUMNODES > 1
4588 * Figure out the number of possible node ids.
4590 static void __init setup_nr_node_ids(void)
4593 unsigned int highest = 0;
4595 for_each_node_mask(node, node_possible_map)
4597 nr_node_ids = highest + 1;
4600 static inline void setup_nr_node_ids(void)
4606 * node_map_pfn_alignment - determine the maximum internode alignment
4608 * This function should be called after node map is populated and sorted.
4609 * It calculates the maximum power of two alignment which can distinguish
4612 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4613 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4614 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4615 * shifted, 1GiB is enough and this function will indicate so.
4617 * This is used to test whether pfn -> nid mapping of the chosen memory
4618 * model has fine enough granularity to avoid incorrect mapping for the
4619 * populated node map.
4621 * Returns the determined alignment in pfn's. 0 if there is no alignment
4622 * requirement (single node).
4624 unsigned long __init node_map_pfn_alignment(void)
4626 unsigned long accl_mask = 0, last_end = 0;
4627 unsigned long start, end, mask;
4631 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4632 if (!start || last_nid < 0 || last_nid == nid) {
4639 * Start with a mask granular enough to pin-point to the
4640 * start pfn and tick off bits one-by-one until it becomes
4641 * too coarse to separate the current node from the last.
4643 mask = ~((1 << __ffs(start)) - 1);
4644 while (mask && last_end <= (start & (mask << 1)))
4647 /* accumulate all internode masks */
4651 /* convert mask to number of pages */
4652 return ~accl_mask + 1;
4655 /* Find the lowest pfn for a node */
4656 static unsigned long __init find_min_pfn_for_node(int nid)
4658 unsigned long min_pfn = ULONG_MAX;
4659 unsigned long start_pfn;
4662 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4663 min_pfn = min(min_pfn, start_pfn);
4665 if (min_pfn == ULONG_MAX) {
4667 "Could not find start_pfn for node %d\n", nid);
4675 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4677 * It returns the minimum PFN based on information provided via
4678 * add_active_range().
4680 unsigned long __init find_min_pfn_with_active_regions(void)
4682 return find_min_pfn_for_node(MAX_NUMNODES);
4686 * early_calculate_totalpages()
4687 * Sum pages in active regions for movable zone.
4688 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4690 static unsigned long __init early_calculate_totalpages(void)
4692 unsigned long totalpages = 0;
4693 unsigned long start_pfn, end_pfn;
4696 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4697 unsigned long pages = end_pfn - start_pfn;
4699 totalpages += pages;
4701 node_set_state(nid, N_HIGH_MEMORY);
4707 * Find the PFN the Movable zone begins in each node. Kernel memory
4708 * is spread evenly between nodes as long as the nodes have enough
4709 * memory. When they don't, some nodes will have more kernelcore than
4712 static void __init find_zone_movable_pfns_for_nodes(void)
4715 unsigned long usable_startpfn;
4716 unsigned long kernelcore_node, kernelcore_remaining;
4717 /* save the state before borrow the nodemask */
4718 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4719 unsigned long totalpages = early_calculate_totalpages();
4720 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4723 * If movablecore was specified, calculate what size of
4724 * kernelcore that corresponds so that memory usable for
4725 * any allocation type is evenly spread. If both kernelcore
4726 * and movablecore are specified, then the value of kernelcore
4727 * will be used for required_kernelcore if it's greater than
4728 * what movablecore would have allowed.
4730 if (required_movablecore) {
4731 unsigned long corepages;
4734 * Round-up so that ZONE_MOVABLE is at least as large as what
4735 * was requested by the user
4737 required_movablecore =
4738 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4739 corepages = totalpages - required_movablecore;
4741 required_kernelcore = max(required_kernelcore, corepages);
4744 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4745 if (!required_kernelcore)
4748 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4749 find_usable_zone_for_movable();
4750 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4753 /* Spread kernelcore memory as evenly as possible throughout nodes */
4754 kernelcore_node = required_kernelcore / usable_nodes;
4755 for_each_node_state(nid, N_HIGH_MEMORY) {
4756 unsigned long start_pfn, end_pfn;
4759 * Recalculate kernelcore_node if the division per node
4760 * now exceeds what is necessary to satisfy the requested
4761 * amount of memory for the kernel
4763 if (required_kernelcore < kernelcore_node)
4764 kernelcore_node = required_kernelcore / usable_nodes;
4767 * As the map is walked, we track how much memory is usable
4768 * by the kernel using kernelcore_remaining. When it is
4769 * 0, the rest of the node is usable by ZONE_MOVABLE
4771 kernelcore_remaining = kernelcore_node;
4773 /* Go through each range of PFNs within this node */
4774 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4775 unsigned long size_pages;
4777 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4778 if (start_pfn >= end_pfn)
4781 /* Account for what is only usable for kernelcore */
4782 if (start_pfn < usable_startpfn) {
4783 unsigned long kernel_pages;
4784 kernel_pages = min(end_pfn, usable_startpfn)
4787 kernelcore_remaining -= min(kernel_pages,
4788 kernelcore_remaining);
4789 required_kernelcore -= min(kernel_pages,
4790 required_kernelcore);
4792 /* Continue if range is now fully accounted */
4793 if (end_pfn <= usable_startpfn) {
4796 * Push zone_movable_pfn to the end so
4797 * that if we have to rebalance
4798 * kernelcore across nodes, we will
4799 * not double account here
4801 zone_movable_pfn[nid] = end_pfn;
4804 start_pfn = usable_startpfn;
4808 * The usable PFN range for ZONE_MOVABLE is from
4809 * start_pfn->end_pfn. Calculate size_pages as the
4810 * number of pages used as kernelcore
4812 size_pages = end_pfn - start_pfn;
4813 if (size_pages > kernelcore_remaining)
4814 size_pages = kernelcore_remaining;
4815 zone_movable_pfn[nid] = start_pfn + size_pages;
4818 * Some kernelcore has been met, update counts and
4819 * break if the kernelcore for this node has been
4822 required_kernelcore -= min(required_kernelcore,
4824 kernelcore_remaining -= size_pages;
4825 if (!kernelcore_remaining)
4831 * If there is still required_kernelcore, we do another pass with one
4832 * less node in the count. This will push zone_movable_pfn[nid] further
4833 * along on the nodes that still have memory until kernelcore is
4837 if (usable_nodes && required_kernelcore > usable_nodes)
4840 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4841 for (nid = 0; nid < MAX_NUMNODES; nid++)
4842 zone_movable_pfn[nid] =
4843 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4846 /* restore the node_state */
4847 node_states[N_HIGH_MEMORY] = saved_node_state;
4850 /* Any regular memory on that node ? */
4851 static void __init check_for_regular_memory(pg_data_t *pgdat)
4853 #ifdef CONFIG_HIGHMEM
4854 enum zone_type zone_type;
4856 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4857 struct zone *zone = &pgdat->node_zones[zone_type];
4858 if (zone->present_pages) {
4859 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4867 * free_area_init_nodes - Initialise all pg_data_t and zone data
4868 * @max_zone_pfn: an array of max PFNs for each zone
4870 * This will call free_area_init_node() for each active node in the system.
4871 * Using the page ranges provided by add_active_range(), the size of each
4872 * zone in each node and their holes is calculated. If the maximum PFN
4873 * between two adjacent zones match, it is assumed that the zone is empty.
4874 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4875 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4876 * starts where the previous one ended. For example, ZONE_DMA32 starts
4877 * at arch_max_dma_pfn.
4879 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4881 unsigned long start_pfn, end_pfn;
4884 /* Record where the zone boundaries are */
4885 memset(arch_zone_lowest_possible_pfn, 0,
4886 sizeof(arch_zone_lowest_possible_pfn));
4887 memset(arch_zone_highest_possible_pfn, 0,
4888 sizeof(arch_zone_highest_possible_pfn));
4889 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4890 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4891 for (i = 1; i < MAX_NR_ZONES; i++) {
4892 if (i == ZONE_MOVABLE)
4894 arch_zone_lowest_possible_pfn[i] =
4895 arch_zone_highest_possible_pfn[i-1];
4896 arch_zone_highest_possible_pfn[i] =
4897 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4899 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4900 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4902 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4903 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4904 find_zone_movable_pfns_for_nodes();
4906 /* Print out the zone ranges */
4907 printk("Zone ranges:\n");
4908 for (i = 0; i < MAX_NR_ZONES; i++) {
4909 if (i == ZONE_MOVABLE)
4911 printk(KERN_CONT " %-8s ", zone_names[i]);
4912 if (arch_zone_lowest_possible_pfn[i] ==
4913 arch_zone_highest_possible_pfn[i])
4914 printk(KERN_CONT "empty\n");
4916 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4917 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4918 (arch_zone_highest_possible_pfn[i]
4919 << PAGE_SHIFT) - 1);
4922 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4923 printk("Movable zone start for each node\n");
4924 for (i = 0; i < MAX_NUMNODES; i++) {
4925 if (zone_movable_pfn[i])
4926 printk(" Node %d: %#010lx\n", i,
4927 zone_movable_pfn[i] << PAGE_SHIFT);
4930 /* Print out the early_node_map[] */
4931 printk("Early memory node ranges\n");
4932 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4933 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
4934 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
4936 /* Initialise every node */
4937 mminit_verify_pageflags_layout();
4938 setup_nr_node_ids();
4939 for_each_online_node(nid) {
4940 pg_data_t *pgdat = NODE_DATA(nid);
4941 free_area_init_node(nid, NULL,
4942 find_min_pfn_for_node(nid), NULL);
4944 /* Any memory on that node */
4945 if (pgdat->node_present_pages)
4946 node_set_state(nid, N_HIGH_MEMORY);
4947 check_for_regular_memory(pgdat);
4951 static int __init cmdline_parse_core(char *p, unsigned long *core)
4953 unsigned long long coremem;
4957 coremem = memparse(p, &p);
4958 *core = coremem >> PAGE_SHIFT;
4960 /* Paranoid check that UL is enough for the coremem value */
4961 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4967 * kernelcore=size sets the amount of memory for use for allocations that
4968 * cannot be reclaimed or migrated.
4970 static int __init cmdline_parse_kernelcore(char *p)
4972 return cmdline_parse_core(p, &required_kernelcore);
4976 * movablecore=size sets the amount of memory for use for allocations that
4977 * can be reclaimed or migrated.
4979 static int __init cmdline_parse_movablecore(char *p)
4981 return cmdline_parse_core(p, &required_movablecore);
4984 early_param("kernelcore", cmdline_parse_kernelcore);
4985 early_param("movablecore", cmdline_parse_movablecore);
4987 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4990 * set_dma_reserve - set the specified number of pages reserved in the first zone
4991 * @new_dma_reserve: The number of pages to mark reserved
4993 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4994 * In the DMA zone, a significant percentage may be consumed by kernel image
4995 * and other unfreeable allocations which can skew the watermarks badly. This
4996 * function may optionally be used to account for unfreeable pages in the
4997 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4998 * smaller per-cpu batchsize.
5000 void __init set_dma_reserve(unsigned long new_dma_reserve)
5002 dma_reserve = new_dma_reserve;
5005 void __init free_area_init(unsigned long *zones_size)
5007 free_area_init_node(0, zones_size,
5008 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5011 static int page_alloc_cpu_notify(struct notifier_block *self,
5012 unsigned long action, void *hcpu)
5014 int cpu = (unsigned long)hcpu;
5016 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5017 lru_add_drain_cpu(cpu);
5021 * Spill the event counters of the dead processor
5022 * into the current processors event counters.
5023 * This artificially elevates the count of the current
5026 vm_events_fold_cpu(cpu);
5029 * Zero the differential counters of the dead processor
5030 * so that the vm statistics are consistent.
5032 * This is only okay since the processor is dead and cannot
5033 * race with what we are doing.
5035 refresh_cpu_vm_stats(cpu);
5040 void __init page_alloc_init(void)
5042 hotcpu_notifier(page_alloc_cpu_notify, 0);
5046 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5047 * or min_free_kbytes changes.
5049 static void calculate_totalreserve_pages(void)
5051 struct pglist_data *pgdat;
5052 unsigned long reserve_pages = 0;
5053 enum zone_type i, j;
5055 for_each_online_pgdat(pgdat) {
5056 for (i = 0; i < MAX_NR_ZONES; i++) {
5057 struct zone *zone = pgdat->node_zones + i;
5058 unsigned long max = 0;
5060 /* Find valid and maximum lowmem_reserve in the zone */
5061 for (j = i; j < MAX_NR_ZONES; j++) {
5062 if (zone->lowmem_reserve[j] > max)
5063 max = zone->lowmem_reserve[j];
5066 /* we treat the high watermark as reserved pages. */
5067 max += high_wmark_pages(zone);
5069 if (max > zone->present_pages)
5070 max = zone->present_pages;
5071 reserve_pages += max;
5073 * Lowmem reserves are not available to
5074 * GFP_HIGHUSER page cache allocations and
5075 * kswapd tries to balance zones to their high
5076 * watermark. As a result, neither should be
5077 * regarded as dirtyable memory, to prevent a
5078 * situation where reclaim has to clean pages
5079 * in order to balance the zones.
5081 zone->dirty_balance_reserve = max;
5084 dirty_balance_reserve = reserve_pages;
5085 totalreserve_pages = reserve_pages;
5089 * setup_per_zone_lowmem_reserve - called whenever
5090 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5091 * has a correct pages reserved value, so an adequate number of
5092 * pages are left in the zone after a successful __alloc_pages().
5094 static void setup_per_zone_lowmem_reserve(void)
5096 struct pglist_data *pgdat;
5097 enum zone_type j, idx;
5099 for_each_online_pgdat(pgdat) {
5100 for (j = 0; j < MAX_NR_ZONES; j++) {
5101 struct zone *zone = pgdat->node_zones + j;
5102 unsigned long present_pages = zone->present_pages;
5104 zone->lowmem_reserve[j] = 0;
5108 struct zone *lower_zone;
5112 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5113 sysctl_lowmem_reserve_ratio[idx] = 1;
5115 lower_zone = pgdat->node_zones + idx;
5116 lower_zone->lowmem_reserve[j] = present_pages /
5117 sysctl_lowmem_reserve_ratio[idx];
5118 present_pages += lower_zone->present_pages;
5123 /* update totalreserve_pages */
5124 calculate_totalreserve_pages();
5127 static void __setup_per_zone_wmarks(void)
5129 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5130 unsigned long lowmem_pages = 0;
5132 unsigned long flags;
5134 /* Calculate total number of !ZONE_HIGHMEM pages */
5135 for_each_zone(zone) {
5136 if (!is_highmem(zone))
5137 lowmem_pages += zone->present_pages;
5140 for_each_zone(zone) {
5143 spin_lock_irqsave(&zone->lock, flags);
5144 tmp = (u64)pages_min * zone->present_pages;
5145 do_div(tmp, lowmem_pages);
5146 if (is_highmem(zone)) {
5148 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5149 * need highmem pages, so cap pages_min to a small
5152 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5153 * deltas controls asynch page reclaim, and so should
5154 * not be capped for highmem.
5158 min_pages = zone->present_pages / 1024;
5159 if (min_pages < SWAP_CLUSTER_MAX)
5160 min_pages = SWAP_CLUSTER_MAX;
5161 if (min_pages > 128)
5163 zone->watermark[WMARK_MIN] = min_pages;
5166 * If it's a lowmem zone, reserve a number of pages
5167 * proportionate to the zone's size.
5169 zone->watermark[WMARK_MIN] = tmp;
5172 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5173 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5175 zone->watermark[WMARK_MIN] += cma_wmark_pages(zone);
5176 zone->watermark[WMARK_LOW] += cma_wmark_pages(zone);
5177 zone->watermark[WMARK_HIGH] += cma_wmark_pages(zone);
5179 setup_zone_migrate_reserve(zone);
5180 spin_unlock_irqrestore(&zone->lock, flags);
5183 /* update totalreserve_pages */
5184 calculate_totalreserve_pages();
5188 * setup_per_zone_wmarks - called when min_free_kbytes changes
5189 * or when memory is hot-{added|removed}
5191 * Ensures that the watermark[min,low,high] values for each zone are set
5192 * correctly with respect to min_free_kbytes.
5194 void setup_per_zone_wmarks(void)
5196 mutex_lock(&zonelists_mutex);
5197 __setup_per_zone_wmarks();
5198 mutex_unlock(&zonelists_mutex);
5202 * The inactive anon list should be small enough that the VM never has to
5203 * do too much work, but large enough that each inactive page has a chance
5204 * to be referenced again before it is swapped out.
5206 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5207 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5208 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5209 * the anonymous pages are kept on the inactive list.
5212 * memory ratio inactive anon
5213 * -------------------------------------
5222 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5224 unsigned int gb, ratio;
5226 /* Zone size in gigabytes */
5227 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5229 ratio = int_sqrt(10 * gb);
5233 zone->inactive_ratio = ratio;
5236 static void __meminit setup_per_zone_inactive_ratio(void)
5241 calculate_zone_inactive_ratio(zone);
5245 * Initialise min_free_kbytes.
5247 * For small machines we want it small (128k min). For large machines
5248 * we want it large (64MB max). But it is not linear, because network
5249 * bandwidth does not increase linearly with machine size. We use
5251 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5252 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5268 int __meminit init_per_zone_wmark_min(void)
5270 unsigned long lowmem_kbytes;
5272 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5274 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5275 if (min_free_kbytes < 128)
5276 min_free_kbytes = 128;
5277 if (min_free_kbytes > 65536)
5278 min_free_kbytes = 65536;
5279 setup_per_zone_wmarks();
5280 refresh_zone_stat_thresholds();
5281 setup_per_zone_lowmem_reserve();
5282 setup_per_zone_inactive_ratio();
5285 module_init(init_per_zone_wmark_min)
5288 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5289 * that we can call two helper functions whenever min_free_kbytes
5292 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5293 void __user *buffer, size_t *length, loff_t *ppos)
5295 proc_dointvec(table, write, buffer, length, ppos);
5297 setup_per_zone_wmarks();
5302 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5303 void __user *buffer, size_t *length, loff_t *ppos)
5308 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5313 zone->min_unmapped_pages = (zone->present_pages *
5314 sysctl_min_unmapped_ratio) / 100;
5318 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5319 void __user *buffer, size_t *length, loff_t *ppos)
5324 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5329 zone->min_slab_pages = (zone->present_pages *
5330 sysctl_min_slab_ratio) / 100;
5336 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5337 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5338 * whenever sysctl_lowmem_reserve_ratio changes.
5340 * The reserve ratio obviously has absolutely no relation with the
5341 * minimum watermarks. The lowmem reserve ratio can only make sense
5342 * if in function of the boot time zone sizes.
5344 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5345 void __user *buffer, size_t *length, loff_t *ppos)
5347 proc_dointvec_minmax(table, write, buffer, length, ppos);
5348 setup_per_zone_lowmem_reserve();
5353 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5354 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5355 * can have before it gets flushed back to buddy allocator.
5358 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5359 void __user *buffer, size_t *length, loff_t *ppos)
5365 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5366 if (!write || (ret < 0))
5368 for_each_populated_zone(zone) {
5369 for_each_possible_cpu(cpu) {
5371 high = zone->present_pages / percpu_pagelist_fraction;
5372 setup_pagelist_highmark(
5373 per_cpu_ptr(zone->pageset, cpu), high);
5379 int hashdist = HASHDIST_DEFAULT;
5382 static int __init set_hashdist(char *str)
5386 hashdist = simple_strtoul(str, &str, 0);
5389 __setup("hashdist=", set_hashdist);
5393 * allocate a large system hash table from bootmem
5394 * - it is assumed that the hash table must contain an exact power-of-2
5395 * quantity of entries
5396 * - limit is the number of hash buckets, not the total allocation size
5398 void *__init alloc_large_system_hash(const char *tablename,
5399 unsigned long bucketsize,
5400 unsigned long numentries,
5403 unsigned int *_hash_shift,
5404 unsigned int *_hash_mask,
5405 unsigned long low_limit,
5406 unsigned long high_limit)
5408 unsigned long long max = high_limit;
5409 unsigned long log2qty, size;
5412 /* allow the kernel cmdline to have a say */
5414 /* round applicable memory size up to nearest megabyte */
5415 numentries = nr_kernel_pages;
5416 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5417 numentries >>= 20 - PAGE_SHIFT;
5418 numentries <<= 20 - PAGE_SHIFT;
5420 /* limit to 1 bucket per 2^scale bytes of low memory */
5421 if (scale > PAGE_SHIFT)
5422 numentries >>= (scale - PAGE_SHIFT);
5424 numentries <<= (PAGE_SHIFT - scale);
5426 /* Make sure we've got at least a 0-order allocation.. */
5427 if (unlikely(flags & HASH_SMALL)) {
5428 /* Makes no sense without HASH_EARLY */
5429 WARN_ON(!(flags & HASH_EARLY));
5430 if (!(numentries >> *_hash_shift)) {
5431 numentries = 1UL << *_hash_shift;
5432 BUG_ON(!numentries);
5434 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5435 numentries = PAGE_SIZE / bucketsize;
5437 numentries = roundup_pow_of_two(numentries);
5439 /* limit allocation size to 1/16 total memory by default */
5441 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5442 do_div(max, bucketsize);
5444 max = min(max, 0x80000000ULL);
5446 if (numentries < low_limit)
5447 numentries = low_limit;
5448 if (numentries > max)
5451 log2qty = ilog2(numentries);
5454 size = bucketsize << log2qty;
5455 if (flags & HASH_EARLY)
5456 table = alloc_bootmem_nopanic(size);
5458 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5461 * If bucketsize is not a power-of-two, we may free
5462 * some pages at the end of hash table which
5463 * alloc_pages_exact() automatically does
5465 if (get_order(size) < MAX_ORDER) {
5466 table = alloc_pages_exact(size, GFP_ATOMIC);
5467 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5470 } while (!table && size > PAGE_SIZE && --log2qty);
5473 panic("Failed to allocate %s hash table\n", tablename);
5475 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5478 ilog2(size) - PAGE_SHIFT,
5482 *_hash_shift = log2qty;
5484 *_hash_mask = (1 << log2qty) - 1;
5489 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5490 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5493 #ifdef CONFIG_SPARSEMEM
5494 return __pfn_to_section(pfn)->pageblock_flags;
5496 return zone->pageblock_flags;
5497 #endif /* CONFIG_SPARSEMEM */
5500 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5502 #ifdef CONFIG_SPARSEMEM
5503 pfn &= (PAGES_PER_SECTION-1);
5504 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5506 pfn = pfn - zone->zone_start_pfn;
5507 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5508 #endif /* CONFIG_SPARSEMEM */
5512 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5513 * @page: The page within the block of interest
5514 * @start_bitidx: The first bit of interest to retrieve
5515 * @end_bitidx: The last bit of interest
5516 * returns pageblock_bits flags
5518 unsigned long get_pageblock_flags_group(struct page *page,
5519 int start_bitidx, int end_bitidx)
5522 unsigned long *bitmap;
5523 unsigned long pfn, bitidx;
5524 unsigned long flags = 0;
5525 unsigned long value = 1;
5527 zone = page_zone(page);
5528 pfn = page_to_pfn(page);
5529 bitmap = get_pageblock_bitmap(zone, pfn);
5530 bitidx = pfn_to_bitidx(zone, pfn);
5532 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5533 if (test_bit(bitidx + start_bitidx, bitmap))
5540 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5541 * @page: The page within the block of interest
5542 * @start_bitidx: The first bit of interest
5543 * @end_bitidx: The last bit of interest
5544 * @flags: The flags to set
5546 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5547 int start_bitidx, int end_bitidx)
5550 unsigned long *bitmap;
5551 unsigned long pfn, bitidx;
5552 unsigned long value = 1;
5554 zone = page_zone(page);
5555 pfn = page_to_pfn(page);
5556 bitmap = get_pageblock_bitmap(zone, pfn);
5557 bitidx = pfn_to_bitidx(zone, pfn);
5558 VM_BUG_ON(pfn < zone->zone_start_pfn);
5559 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5561 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5563 __set_bit(bitidx + start_bitidx, bitmap);
5565 __clear_bit(bitidx + start_bitidx, bitmap);
5569 * This function checks whether pageblock includes unmovable pages or not.
5570 * If @count is not zero, it is okay to include less @count unmovable pages
5572 * PageLRU check wihtout isolation or lru_lock could race so that
5573 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5574 * expect this function should be exact.
5576 bool has_unmovable_pages(struct zone *zone, struct page *page, int count)
5578 unsigned long pfn, iter, found;
5582 * For avoiding noise data, lru_add_drain_all() should be called
5583 * If ZONE_MOVABLE, the zone never contains unmovable pages
5585 if (zone_idx(zone) == ZONE_MOVABLE)
5587 mt = get_pageblock_migratetype(page);
5588 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5591 pfn = page_to_pfn(page);
5592 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5593 unsigned long check = pfn + iter;
5595 if (!pfn_valid_within(check))
5598 page = pfn_to_page(check);
5600 * We can't use page_count without pin a page
5601 * because another CPU can free compound page.
5602 * This check already skips compound tails of THP
5603 * because their page->_count is zero at all time.
5605 if (!atomic_read(&page->_count)) {
5606 if (PageBuddy(page))
5607 iter += (1 << page_order(page)) - 1;
5614 * If there are RECLAIMABLE pages, we need to check it.
5615 * But now, memory offline itself doesn't call shrink_slab()
5616 * and it still to be fixed.
5619 * If the page is not RAM, page_count()should be 0.
5620 * we don't need more check. This is an _used_ not-movable page.
5622 * The problematic thing here is PG_reserved pages. PG_reserved
5623 * is set to both of a memory hole page and a _used_ kernel
5632 bool is_pageblock_removable_nolock(struct page *page)
5638 * We have to be careful here because we are iterating over memory
5639 * sections which are not zone aware so we might end up outside of
5640 * the zone but still within the section.
5641 * We have to take care about the node as well. If the node is offline
5642 * its NODE_DATA will be NULL - see page_zone.
5644 if (!node_online(page_to_nid(page)))
5647 zone = page_zone(page);
5648 pfn = page_to_pfn(page);
5649 if (zone->zone_start_pfn > pfn ||
5650 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5653 return !has_unmovable_pages(zone, page, 0);
5658 static unsigned long pfn_max_align_down(unsigned long pfn)
5660 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5661 pageblock_nr_pages) - 1);
5664 static unsigned long pfn_max_align_up(unsigned long pfn)
5666 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5667 pageblock_nr_pages));
5670 static struct page *
5671 __alloc_contig_migrate_alloc(struct page *page, unsigned long private,
5674 gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
5676 if (PageHighMem(page))
5677 gfp_mask |= __GFP_HIGHMEM;
5679 return alloc_page(gfp_mask);
5682 /* [start, end) must belong to a single zone. */
5683 static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5685 /* This function is based on compact_zone() from compaction.c. */
5687 unsigned long pfn = start;
5688 unsigned int tries = 0;
5691 struct compact_control cc = {
5692 .nr_migratepages = 0,
5694 .zone = page_zone(pfn_to_page(start)),
5697 INIT_LIST_HEAD(&cc.migratepages);
5699 migrate_prep_local();
5701 while (pfn < end || !list_empty(&cc.migratepages)) {
5702 if (fatal_signal_pending(current)) {
5707 if (list_empty(&cc.migratepages)) {
5708 cc.nr_migratepages = 0;
5709 pfn = isolate_migratepages_range(cc.zone, &cc,
5716 } else if (++tries == 5) {
5717 ret = ret < 0 ? ret : -EBUSY;
5721 reclaim_clean_pages_from_list(cc.zone, &cc.migratepages);
5723 ret = migrate_pages(&cc.migratepages,
5724 __alloc_contig_migrate_alloc,
5725 0, false, MIGRATE_SYNC);
5728 putback_lru_pages(&cc.migratepages);
5729 return ret > 0 ? 0 : ret;
5733 * Update zone's cma pages counter used for watermark level calculation.
5735 static inline void __update_cma_watermarks(struct zone *zone, int count)
5737 unsigned long flags;
5738 spin_lock_irqsave(&zone->lock, flags);
5739 zone->min_cma_pages += count;
5740 spin_unlock_irqrestore(&zone->lock, flags);
5741 setup_per_zone_wmarks();
5745 * Trigger memory pressure bump to reclaim some pages in order to be able to
5746 * allocate 'count' pages in single page units. Does similar work as
5747 *__alloc_pages_slowpath() function.
5749 static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
5751 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
5752 struct zonelist *zonelist = node_zonelist(0, gfp_mask);
5753 int did_some_progress = 0;
5757 * Increase level of watermarks to force kswapd do his job
5758 * to stabilise at new watermark level.
5760 __update_cma_watermarks(zone, count);
5762 /* Obey watermarks as if the page was being allocated */
5763 while (!zone_watermark_ok(zone, 0, low_wmark_pages(zone), 0, 0)) {
5764 wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));
5766 did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
5768 if (!did_some_progress) {
5769 /* Exhausted what can be done so it's blamo time */
5770 out_of_memory(zonelist, gfp_mask, order, NULL, false);
5774 /* Restore original watermark levels. */
5775 __update_cma_watermarks(zone, -count);
5781 * alloc_contig_range() -- tries to allocate given range of pages
5782 * @start: start PFN to allocate
5783 * @end: one-past-the-last PFN to allocate
5784 * @migratetype: migratetype of the underlaying pageblocks (either
5785 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5786 * in range must have the same migratetype and it must
5787 * be either of the two.
5789 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5790 * aligned, however it's the caller's responsibility to guarantee that
5791 * we are the only thread that changes migrate type of pageblocks the
5794 * The PFN range must belong to a single zone.
5796 * Returns zero on success or negative error code. On success all
5797 * pages which PFN is in [start, end) are allocated for the caller and
5798 * need to be freed with free_contig_range().
5800 int alloc_contig_range(unsigned long start, unsigned long end,
5801 unsigned migratetype)
5803 struct zone *zone = page_zone(pfn_to_page(start));
5804 unsigned long outer_start, outer_end;
5808 * What we do here is we mark all pageblocks in range as
5809 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5810 * have different sizes, and due to the way page allocator
5811 * work, we align the range to biggest of the two pages so
5812 * that page allocator won't try to merge buddies from
5813 * different pageblocks and change MIGRATE_ISOLATE to some
5814 * other migration type.
5816 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5817 * migrate the pages from an unaligned range (ie. pages that
5818 * we are interested in). This will put all the pages in
5819 * range back to page allocator as MIGRATE_ISOLATE.
5821 * When this is done, we take the pages in range from page
5822 * allocator removing them from the buddy system. This way
5823 * page allocator will never consider using them.
5825 * This lets us mark the pageblocks back as
5826 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5827 * aligned range but not in the unaligned, original range are
5828 * put back to page allocator so that buddy can use them.
5831 ret = start_isolate_page_range(pfn_max_align_down(start),
5832 pfn_max_align_up(end), migratetype);
5836 ret = __alloc_contig_migrate_range(start, end);
5841 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5842 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5843 * more, all pages in [start, end) are free in page allocator.
5844 * What we are going to do is to allocate all pages from
5845 * [start, end) (that is remove them from page allocator).
5847 * The only problem is that pages at the beginning and at the
5848 * end of interesting range may be not aligned with pages that
5849 * page allocator holds, ie. they can be part of higher order
5850 * pages. Because of this, we reserve the bigger range and
5851 * once this is done free the pages we are not interested in.
5853 * We don't have to hold zone->lock here because the pages are
5854 * isolated thus they won't get removed from buddy.
5857 lru_add_drain_all();
5861 outer_start = start;
5862 while (!PageBuddy(pfn_to_page(outer_start))) {
5863 if (++order >= MAX_ORDER) {
5867 outer_start &= ~0UL << order;
5870 /* Make sure the range is really isolated. */
5871 if (test_pages_isolated(outer_start, end)) {
5872 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5879 * Reclaim enough pages to make sure that contiguous allocation
5880 * will not starve the system.
5882 __reclaim_pages(zone, GFP_HIGHUSER_MOVABLE, end-start);
5884 /* Grab isolated pages from freelists. */
5885 outer_end = isolate_freepages_range(outer_start, end);
5891 /* Free head and tail (if any) */
5892 if (start != outer_start)
5893 free_contig_range(outer_start, start - outer_start);
5894 if (end != outer_end)
5895 free_contig_range(end, outer_end - end);
5898 undo_isolate_page_range(pfn_max_align_down(start),
5899 pfn_max_align_up(end), migratetype);
5903 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5905 for (; nr_pages--; ++pfn)
5906 __free_page(pfn_to_page(pfn));
5910 #ifdef CONFIG_MEMORY_HOTPLUG
5911 static int __meminit __zone_pcp_update(void *data)
5913 struct zone *zone = data;
5915 unsigned long batch = zone_batchsize(zone), flags;
5917 for_each_possible_cpu(cpu) {
5918 struct per_cpu_pageset *pset;
5919 struct per_cpu_pages *pcp;
5921 pset = per_cpu_ptr(zone->pageset, cpu);
5924 local_irq_save(flags);
5926 free_pcppages_bulk(zone, pcp->count, pcp);
5927 setup_pageset(pset, batch);
5928 local_irq_restore(flags);
5933 void __meminit zone_pcp_update(struct zone *zone)
5935 stop_machine(__zone_pcp_update, zone, NULL);
5939 #ifdef CONFIG_MEMORY_HOTREMOVE
5940 void zone_pcp_reset(struct zone *zone)
5942 unsigned long flags;
5944 /* avoid races with drain_pages() */
5945 local_irq_save(flags);
5946 if (zone->pageset != &boot_pageset) {
5947 free_percpu(zone->pageset);
5948 zone->pageset = &boot_pageset;
5950 local_irq_restore(flags);
5954 * All pages in the range must be isolated before calling this.
5957 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5963 unsigned long flags;
5964 /* find the first valid pfn */
5965 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5970 zone = page_zone(pfn_to_page(pfn));
5971 spin_lock_irqsave(&zone->lock, flags);
5973 while (pfn < end_pfn) {
5974 if (!pfn_valid(pfn)) {
5978 page = pfn_to_page(pfn);
5979 BUG_ON(page_count(page));
5980 BUG_ON(!PageBuddy(page));
5981 order = page_order(page);
5982 #ifdef CONFIG_DEBUG_VM
5983 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5984 pfn, 1 << order, end_pfn);
5986 list_del(&page->lru);
5987 rmv_page_order(page);
5988 zone->free_area[order].nr_free--;
5989 __mod_zone_page_state(zone, NR_FREE_PAGES,
5991 for (i = 0; i < (1 << order); i++)
5992 SetPageReserved((page+i));
5993 pfn += (1 << order);
5995 spin_unlock_irqrestore(&zone->lock, flags);
5999 #ifdef CONFIG_MEMORY_FAILURE
6000 bool is_free_buddy_page(struct page *page)
6002 struct zone *zone = page_zone(page);
6003 unsigned long pfn = page_to_pfn(page);
6004 unsigned long flags;
6007 spin_lock_irqsave(&zone->lock, flags);
6008 for (order = 0; order < MAX_ORDER; order++) {
6009 struct page *page_head = page - (pfn & ((1 << order) - 1));
6011 if (PageBuddy(page_head) && page_order(page_head) >= order)
6014 spin_unlock_irqrestore(&zone->lock, flags);
6016 return order < MAX_ORDER;
6020 static const struct trace_print_flags pageflag_names[] = {
6021 {1UL << PG_locked, "locked" },
6022 {1UL << PG_error, "error" },
6023 {1UL << PG_referenced, "referenced" },
6024 {1UL << PG_uptodate, "uptodate" },
6025 {1UL << PG_dirty, "dirty" },
6026 {1UL << PG_lru, "lru" },
6027 {1UL << PG_active, "active" },
6028 {1UL << PG_slab, "slab" },
6029 {1UL << PG_owner_priv_1, "owner_priv_1" },
6030 {1UL << PG_arch_1, "arch_1" },
6031 {1UL << PG_reserved, "reserved" },
6032 {1UL << PG_private, "private" },
6033 {1UL << PG_private_2, "private_2" },
6034 {1UL << PG_writeback, "writeback" },
6035 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6036 {1UL << PG_head, "head" },
6037 {1UL << PG_tail, "tail" },
6039 {1UL << PG_compound, "compound" },
6041 {1UL << PG_swapcache, "swapcache" },
6042 {1UL << PG_mappedtodisk, "mappedtodisk" },
6043 {1UL << PG_reclaim, "reclaim" },
6044 {1UL << PG_swapbacked, "swapbacked" },
6045 {1UL << PG_unevictable, "unevictable" },
6047 {1UL << PG_mlocked, "mlocked" },
6049 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6050 {1UL << PG_uncached, "uncached" },
6052 #ifdef CONFIG_MEMORY_FAILURE
6053 {1UL << PG_hwpoison, "hwpoison" },
6055 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6056 {1UL << PG_compound_lock, "compound_lock" },
6060 static void dump_page_flags(unsigned long flags)
6062 const char *delim = "";
6066 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6068 printk(KERN_ALERT "page flags: %#lx(", flags);
6070 /* remove zone id */
6071 flags &= (1UL << NR_PAGEFLAGS) - 1;
6073 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6075 mask = pageflag_names[i].mask;
6076 if ((flags & mask) != mask)
6080 printk("%s%s", delim, pageflag_names[i].name);
6084 /* check for left over flags */
6086 printk("%s%#lx", delim, flags);
6091 void dump_page(struct page *page)
6094 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6095 page, atomic_read(&page->_count), page_mapcount(page),
6096 page->mapping, page->index);
6097 dump_page_flags(page->flags);
6098 mem_cgroup_print_bad_page(page);