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/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74 DEFINE_PER_CPU(int, numa_node);
75 EXPORT_PER_CPU_SYMBOL(numa_node);
78 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
85 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
86 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
90 * Array of node states.
92 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
93 [N_POSSIBLE] = NODE_MASK_ALL,
94 [N_ONLINE] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY] = { { [0] = 1UL } },
103 [N_CPU] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock);
111 unsigned long totalram_pages __read_mostly;
112 unsigned long totalreserve_pages __read_mostly;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly;
121 int percpu_pagelist_fraction;
122 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex));
139 if (saved_gfp_mask) {
140 gfp_allowed_mask = saved_gfp_mask;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex));
148 WARN_ON(saved_gfp_mask);
149 saved_gfp_mask = gfp_allowed_mask;
150 gfp_allowed_mask &= ~GFP_IOFS;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly;
165 static void __free_pages_ok(struct page *page, unsigned int order);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages);
193 static char * const zone_names[MAX_NR_ZONES] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes = 1024;
208 int user_min_free_kbytes;
210 static unsigned long __meminitdata nr_kernel_pages;
211 static unsigned long __meminitdata nr_all_pages;
212 static unsigned long __meminitdata dma_reserve;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
217 static unsigned long __initdata required_kernelcore;
218 static unsigned long __initdata required_movablecore;
219 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly = MAX_NUMNODES;
228 int nr_online_nodes __read_mostly = 1;
229 EXPORT_SYMBOL(nr_node_ids);
230 EXPORT_SYMBOL(nr_online_nodes);
233 int page_group_by_mobility_disabled __read_mostly;
235 void set_pageblock_migratetype(struct page *page, int migratetype)
237 if (unlikely(page_group_by_mobility_disabled &&
238 migratetype < MIGRATE_PCPTYPES))
239 migratetype = MIGRATE_UNMOVABLE;
241 set_pageblock_flags_group(page, (unsigned long)migratetype,
242 PB_migrate, PB_migrate_end);
245 bool oom_killer_disabled __read_mostly;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
252 unsigned long pfn = page_to_pfn(page);
253 unsigned long sp, start_pfn;
256 seq = zone_span_seqbegin(zone);
257 start_pfn = zone->zone_start_pfn;
258 sp = zone->spanned_pages;
259 if (!zone_spans_pfn(zone, pfn))
261 } while (zone_span_seqretry(zone, seq));
264 pr_err("page %lu outside zone [ %lu - %lu ]\n",
265 pfn, start_pfn, start_pfn + sp);
270 static int page_is_consistent(struct zone *zone, struct page *page)
272 if (!pfn_valid_within(page_to_pfn(page)))
274 if (zone != page_zone(page))
280 * Temporary debugging check for pages not lying within a given zone.
282 static int bad_range(struct zone *zone, struct page *page)
284 if (page_outside_zone_boundaries(zone, page))
286 if (!page_is_consistent(zone, page))
292 static inline int bad_range(struct zone *zone, struct page *page)
298 static void bad_page(struct page *page)
300 static unsigned long resume;
301 static unsigned long nr_shown;
302 static unsigned long nr_unshown;
304 /* Don't complain about poisoned pages */
305 if (PageHWPoison(page)) {
306 page_mapcount_reset(page); /* remove PageBuddy */
311 * Allow a burst of 60 reports, then keep quiet for that minute;
312 * or allow a steady drip of one report per second.
314 if (nr_shown == 60) {
315 if (time_before(jiffies, resume)) {
321 "BUG: Bad page state: %lu messages suppressed\n",
328 resume = jiffies + 60 * HZ;
330 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
331 current->comm, page_to_pfn(page));
337 /* Leave bad fields for debug, except PageBuddy could make trouble */
338 page_mapcount_reset(page); /* remove PageBuddy */
339 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
343 * Higher-order pages are called "compound pages". They are structured thusly:
345 * The first PAGE_SIZE page is called the "head page".
347 * The remaining PAGE_SIZE pages are called "tail pages".
349 * All pages have PG_compound set. All tail pages have their ->first_page
350 * pointing at the head page.
352 * The first tail page's ->lru.next holds the address of the compound page's
353 * put_page() function. Its ->lru.prev holds the order of allocation.
354 * This usage means that zero-order pages may not be compound.
357 static void free_compound_page(struct page *page)
359 __free_pages_ok(page, compound_order(page));
362 void prep_compound_page(struct page *page, unsigned long order)
365 int nr_pages = 1 << order;
367 set_compound_page_dtor(page, free_compound_page);
368 set_compound_order(page, order);
370 for (i = 1; i < nr_pages; i++) {
371 struct page *p = page + i;
373 set_page_count(p, 0);
374 p->first_page = page;
378 /* update __split_huge_page_refcount if you change this function */
379 static int destroy_compound_page(struct page *page, unsigned long order)
382 int nr_pages = 1 << order;
385 if (unlikely(compound_order(page) != order)) {
390 __ClearPageHead(page);
392 for (i = 1; i < nr_pages; i++) {
393 struct page *p = page + i;
395 if (unlikely(!PageTail(p) || (p->first_page != page))) {
405 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
410 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
411 * and __GFP_HIGHMEM from hard or soft interrupt context.
413 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
414 for (i = 0; i < (1 << order); i++)
415 clear_highpage(page + i);
418 #ifdef CONFIG_DEBUG_PAGEALLOC
419 unsigned int _debug_guardpage_minorder;
421 static int __init debug_guardpage_minorder_setup(char *buf)
425 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
426 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
429 _debug_guardpage_minorder = res;
430 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
433 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
435 static inline void set_page_guard_flag(struct page *page)
437 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
440 static inline void clear_page_guard_flag(struct page *page)
442 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
445 static inline void set_page_guard_flag(struct page *page) { }
446 static inline void clear_page_guard_flag(struct page *page) { }
449 static inline void set_page_order(struct page *page, int order)
451 set_page_private(page, order);
452 __SetPageBuddy(page);
455 static inline void rmv_page_order(struct page *page)
457 __ClearPageBuddy(page);
458 set_page_private(page, 0);
462 * Locate the struct page for both the matching buddy in our
463 * pair (buddy1) and the combined O(n+1) page they form (page).
465 * 1) Any buddy B1 will have an order O twin B2 which satisfies
466 * the following equation:
468 * For example, if the starting buddy (buddy2) is #8 its order
470 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
472 * 2) Any buddy B will have an order O+1 parent P which
473 * satisfies the following equation:
476 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
478 static inline unsigned long
479 __find_buddy_index(unsigned long page_idx, unsigned int order)
481 return page_idx ^ (1 << order);
485 * This function checks whether a page is free && is the buddy
486 * we can do coalesce a page and its buddy if
487 * (a) the buddy is not in a hole &&
488 * (b) the buddy is in the buddy system &&
489 * (c) a page and its buddy have the same order &&
490 * (d) a page and its buddy are in the same zone.
492 * For recording whether a page is in the buddy system, we set ->_mapcount
493 * PAGE_BUDDY_MAPCOUNT_VALUE.
494 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
495 * serialized by zone->lock.
497 * For recording page's order, we use page_private(page).
499 static inline int page_is_buddy(struct page *page, struct page *buddy,
502 if (!pfn_valid_within(page_to_pfn(buddy)))
505 if (page_zone_id(page) != page_zone_id(buddy))
508 if (page_is_guard(buddy) && page_order(buddy) == order) {
509 VM_BUG_ON(page_count(buddy) != 0);
513 if (PageBuddy(buddy) && page_order(buddy) == order) {
514 VM_BUG_ON(page_count(buddy) != 0);
521 * Freeing function for a buddy system allocator.
523 * The concept of a buddy system is to maintain direct-mapped table
524 * (containing bit values) for memory blocks of various "orders".
525 * The bottom level table contains the map for the smallest allocatable
526 * units of memory (here, pages), and each level above it describes
527 * pairs of units from the levels below, hence, "buddies".
528 * At a high level, all that happens here is marking the table entry
529 * at the bottom level available, and propagating the changes upward
530 * as necessary, plus some accounting needed to play nicely with other
531 * parts of the VM system.
532 * At each level, we keep a list of pages, which are heads of continuous
533 * free pages of length of (1 << order) and marked with _mapcount
534 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
536 * So when we are allocating or freeing one, we can derive the state of the
537 * other. That is, if we allocate a small block, and both were
538 * free, the remainder of the region must be split into blocks.
539 * If a block is freed, and its buddy is also free, then this
540 * triggers coalescing into a block of larger size.
545 static inline void __free_one_page(struct page *page,
546 struct zone *zone, unsigned int order,
549 unsigned long page_idx;
550 unsigned long combined_idx;
551 unsigned long uninitialized_var(buddy_idx);
554 VM_BUG_ON(!zone_is_initialized(zone));
556 if (unlikely(PageCompound(page)))
557 if (unlikely(destroy_compound_page(page, order)))
560 VM_BUG_ON(migratetype == -1);
562 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
564 VM_BUG_ON(page_idx & ((1 << order) - 1));
565 VM_BUG_ON(bad_range(zone, page));
567 while (order < MAX_ORDER-1) {
568 buddy_idx = __find_buddy_index(page_idx, order);
569 buddy = page + (buddy_idx - page_idx);
570 if (!page_is_buddy(page, buddy, order))
573 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
574 * merge with it and move up one order.
576 if (page_is_guard(buddy)) {
577 clear_page_guard_flag(buddy);
578 set_page_private(page, 0);
579 __mod_zone_freepage_state(zone, 1 << order,
582 list_del(&buddy->lru);
583 zone->free_area[order].nr_free--;
584 rmv_page_order(buddy);
586 combined_idx = buddy_idx & page_idx;
587 page = page + (combined_idx - page_idx);
588 page_idx = combined_idx;
591 set_page_order(page, order);
594 * If this is not the largest possible page, check if the buddy
595 * of the next-highest order is free. If it is, it's possible
596 * that pages are being freed that will coalesce soon. In case,
597 * that is happening, add the free page to the tail of the list
598 * so it's less likely to be used soon and more likely to be merged
599 * as a higher order page
601 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
602 struct page *higher_page, *higher_buddy;
603 combined_idx = buddy_idx & page_idx;
604 higher_page = page + (combined_idx - page_idx);
605 buddy_idx = __find_buddy_index(combined_idx, order + 1);
606 higher_buddy = higher_page + (buddy_idx - combined_idx);
607 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
608 list_add_tail(&page->lru,
609 &zone->free_area[order].free_list[migratetype]);
614 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
616 zone->free_area[order].nr_free++;
619 static inline int free_pages_check(struct page *page)
621 if (unlikely(page_mapcount(page) |
622 (page->mapping != NULL) |
623 (atomic_read(&page->_count) != 0) |
624 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
625 (mem_cgroup_bad_page_check(page)))) {
629 page_cpupid_reset_last(page);
630 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
631 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
636 * Frees a number of pages from the PCP lists
637 * Assumes all pages on list are in same zone, and of same order.
638 * count is the number of pages to free.
640 * If the zone was previously in an "all pages pinned" state then look to
641 * see if this freeing clears that state.
643 * And clear the zone's pages_scanned counter, to hold off the "all pages are
644 * pinned" detection logic.
646 static void free_pcppages_bulk(struct zone *zone, int count,
647 struct per_cpu_pages *pcp)
653 spin_lock(&zone->lock);
654 zone->pages_scanned = 0;
658 struct list_head *list;
661 * Remove pages from lists in a round-robin fashion. A
662 * batch_free count is maintained that is incremented when an
663 * empty list is encountered. This is so more pages are freed
664 * off fuller lists instead of spinning excessively around empty
669 if (++migratetype == MIGRATE_PCPTYPES)
671 list = &pcp->lists[migratetype];
672 } while (list_empty(list));
674 /* This is the only non-empty list. Free them all. */
675 if (batch_free == MIGRATE_PCPTYPES)
676 batch_free = to_free;
679 int mt; /* migratetype of the to-be-freed page */
681 page = list_entry(list->prev, struct page, lru);
682 /* must delete as __free_one_page list manipulates */
683 list_del(&page->lru);
684 mt = get_freepage_migratetype(page);
685 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
686 __free_one_page(page, zone, 0, mt);
687 trace_mm_page_pcpu_drain(page, 0, mt);
688 if (likely(!is_migrate_isolate_page(page))) {
689 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
690 if (is_migrate_cma(mt))
691 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
693 } while (--to_free && --batch_free && !list_empty(list));
695 spin_unlock(&zone->lock);
698 static void free_one_page(struct zone *zone, struct page *page, int order,
701 spin_lock(&zone->lock);
702 zone->pages_scanned = 0;
704 __free_one_page(page, zone, order, migratetype);
705 if (unlikely(!is_migrate_isolate(migratetype)))
706 __mod_zone_freepage_state(zone, 1 << order, migratetype);
707 spin_unlock(&zone->lock);
710 static bool free_pages_prepare(struct page *page, unsigned int order)
715 trace_mm_page_free(page, order);
716 kmemcheck_free_shadow(page, order);
719 page->mapping = NULL;
720 for (i = 0; i < (1 << order); i++)
721 bad += free_pages_check(page + i);
725 if (!PageHighMem(page)) {
726 debug_check_no_locks_freed(page_address(page),
728 debug_check_no_obj_freed(page_address(page),
731 arch_free_page(page, order);
732 kernel_map_pages(page, 1 << order, 0);
737 static void __free_pages_ok(struct page *page, unsigned int order)
742 if (!free_pages_prepare(page, order))
745 local_irq_save(flags);
746 __count_vm_events(PGFREE, 1 << order);
747 migratetype = get_pageblock_migratetype(page);
748 set_freepage_migratetype(page, migratetype);
749 free_one_page(page_zone(page), page, order, migratetype);
750 local_irq_restore(flags);
753 void __init __free_pages_bootmem(struct page *page, unsigned int order)
755 unsigned int nr_pages = 1 << order;
756 struct page *p = page;
760 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
762 __ClearPageReserved(p);
763 set_page_count(p, 0);
765 __ClearPageReserved(p);
766 set_page_count(p, 0);
768 page_zone(page)->managed_pages += nr_pages;
769 set_page_refcounted(page);
770 __free_pages(page, order);
774 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
775 void __init init_cma_reserved_pageblock(struct page *page)
777 unsigned i = pageblock_nr_pages;
778 struct page *p = page;
781 __ClearPageReserved(p);
782 set_page_count(p, 0);
785 set_page_refcounted(page);
786 set_pageblock_migratetype(page, MIGRATE_CMA);
787 __free_pages(page, pageblock_order);
788 adjust_managed_page_count(page, pageblock_nr_pages);
793 * The order of subdivision here is critical for the IO subsystem.
794 * Please do not alter this order without good reasons and regression
795 * testing. Specifically, as large blocks of memory are subdivided,
796 * the order in which smaller blocks are delivered depends on the order
797 * they're subdivided in this function. This is the primary factor
798 * influencing the order in which pages are delivered to the IO
799 * subsystem according to empirical testing, and this is also justified
800 * by considering the behavior of a buddy system containing a single
801 * large block of memory acted on by a series of small allocations.
802 * This behavior is a critical factor in sglist merging's success.
806 static inline void expand(struct zone *zone, struct page *page,
807 int low, int high, struct free_area *area,
810 unsigned long size = 1 << high;
816 VM_BUG_ON(bad_range(zone, &page[size]));
818 #ifdef CONFIG_DEBUG_PAGEALLOC
819 if (high < debug_guardpage_minorder()) {
821 * Mark as guard pages (or page), that will allow to
822 * merge back to allocator when buddy will be freed.
823 * Corresponding page table entries will not be touched,
824 * pages will stay not present in virtual address space
826 INIT_LIST_HEAD(&page[size].lru);
827 set_page_guard_flag(&page[size]);
828 set_page_private(&page[size], high);
829 /* Guard pages are not available for any usage */
830 __mod_zone_freepage_state(zone, -(1 << high),
835 list_add(&page[size].lru, &area->free_list[migratetype]);
837 set_page_order(&page[size], high);
842 * This page is about to be returned from the page allocator
844 static inline int check_new_page(struct page *page)
846 if (unlikely(page_mapcount(page) |
847 (page->mapping != NULL) |
848 (atomic_read(&page->_count) != 0) |
849 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
850 (mem_cgroup_bad_page_check(page)))) {
857 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
861 for (i = 0; i < (1 << order); i++) {
862 struct page *p = page + i;
863 if (unlikely(check_new_page(p)))
867 set_page_private(page, 0);
868 set_page_refcounted(page);
870 arch_alloc_page(page, order);
871 kernel_map_pages(page, 1 << order, 1);
873 if (gfp_flags & __GFP_ZERO)
874 prep_zero_page(page, order, gfp_flags);
876 if (order && (gfp_flags & __GFP_COMP))
877 prep_compound_page(page, order);
883 * Go through the free lists for the given migratetype and remove
884 * the smallest available page from the freelists
887 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
890 unsigned int current_order;
891 struct free_area *area;
894 /* Find a page of the appropriate size in the preferred list */
895 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
896 area = &(zone->free_area[current_order]);
897 if (list_empty(&area->free_list[migratetype]))
900 page = list_entry(area->free_list[migratetype].next,
902 list_del(&page->lru);
903 rmv_page_order(page);
905 expand(zone, page, order, current_order, area, migratetype);
914 * This array describes the order lists are fallen back to when
915 * the free lists for the desirable migrate type are depleted
917 static int fallbacks[MIGRATE_TYPES][4] = {
918 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
919 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
921 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
922 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
924 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
926 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
927 #ifdef CONFIG_MEMORY_ISOLATION
928 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
933 * Move the free pages in a range to the free lists of the requested type.
934 * Note that start_page and end_pages are not aligned on a pageblock
935 * boundary. If alignment is required, use move_freepages_block()
937 int move_freepages(struct zone *zone,
938 struct page *start_page, struct page *end_page,
945 #ifndef CONFIG_HOLES_IN_ZONE
947 * page_zone is not safe to call in this context when
948 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
949 * anyway as we check zone boundaries in move_freepages_block().
950 * Remove at a later date when no bug reports exist related to
951 * grouping pages by mobility
953 BUG_ON(page_zone(start_page) != page_zone(end_page));
956 for (page = start_page; page <= end_page;) {
957 /* Make sure we are not inadvertently changing nodes */
958 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
960 if (!pfn_valid_within(page_to_pfn(page))) {
965 if (!PageBuddy(page)) {
970 order = page_order(page);
971 list_move(&page->lru,
972 &zone->free_area[order].free_list[migratetype]);
973 set_freepage_migratetype(page, migratetype);
975 pages_moved += 1 << order;
981 int move_freepages_block(struct zone *zone, struct page *page,
984 unsigned long start_pfn, end_pfn;
985 struct page *start_page, *end_page;
987 start_pfn = page_to_pfn(page);
988 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
989 start_page = pfn_to_page(start_pfn);
990 end_page = start_page + pageblock_nr_pages - 1;
991 end_pfn = start_pfn + pageblock_nr_pages - 1;
993 /* Do not cross zone boundaries */
994 if (!zone_spans_pfn(zone, start_pfn))
996 if (!zone_spans_pfn(zone, end_pfn))
999 return move_freepages(zone, start_page, end_page, migratetype);
1002 static void change_pageblock_range(struct page *pageblock_page,
1003 int start_order, int migratetype)
1005 int nr_pageblocks = 1 << (start_order - pageblock_order);
1007 while (nr_pageblocks--) {
1008 set_pageblock_migratetype(pageblock_page, migratetype);
1009 pageblock_page += pageblock_nr_pages;
1014 * If breaking a large block of pages, move all free pages to the preferred
1015 * allocation list. If falling back for a reclaimable kernel allocation, be
1016 * more aggressive about taking ownership of free pages.
1018 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1019 * nor move CMA pages to different free lists. We don't want unmovable pages
1020 * to be allocated from MIGRATE_CMA areas.
1022 * Returns the new migratetype of the pageblock (or the same old migratetype
1023 * if it was unchanged).
1025 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1026 int start_type, int fallback_type)
1028 int current_order = page_order(page);
1031 * When borrowing from MIGRATE_CMA, we need to release the excess
1032 * buddy pages to CMA itself.
1034 if (is_migrate_cma(fallback_type))
1035 return fallback_type;
1037 /* Take ownership for orders >= pageblock_order */
1038 if (current_order >= pageblock_order) {
1039 change_pageblock_range(page, current_order, start_type);
1043 if (current_order >= pageblock_order / 2 ||
1044 start_type == MIGRATE_RECLAIMABLE ||
1045 page_group_by_mobility_disabled) {
1048 pages = move_freepages_block(zone, page, start_type);
1050 /* Claim the whole block if over half of it is free */
1051 if (pages >= (1 << (pageblock_order-1)) ||
1052 page_group_by_mobility_disabled) {
1054 set_pageblock_migratetype(page, start_type);
1060 return fallback_type;
1063 /* Remove an element from the buddy allocator from the fallback list */
1064 static inline struct page *
1065 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1067 struct free_area *area;
1070 int migratetype, new_type, i;
1072 /* Find the largest possible block of pages in the other list */
1073 for (current_order = MAX_ORDER-1; current_order >= order;
1076 migratetype = fallbacks[start_migratetype][i];
1078 /* MIGRATE_RESERVE handled later if necessary */
1079 if (migratetype == MIGRATE_RESERVE)
1082 area = &(zone->free_area[current_order]);
1083 if (list_empty(&area->free_list[migratetype]))
1086 page = list_entry(area->free_list[migratetype].next,
1090 new_type = try_to_steal_freepages(zone, page,
1094 /* Remove the page from the freelists */
1095 list_del(&page->lru);
1096 rmv_page_order(page);
1098 expand(zone, page, order, current_order, area,
1101 trace_mm_page_alloc_extfrag(page, order, current_order,
1102 start_migratetype, migratetype, new_type);
1112 * Do the hard work of removing an element from the buddy allocator.
1113 * Call me with the zone->lock already held.
1115 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1121 page = __rmqueue_smallest(zone, order, migratetype);
1123 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1124 page = __rmqueue_fallback(zone, order, migratetype);
1127 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1128 * is used because __rmqueue_smallest is an inline function
1129 * and we want just one call site
1132 migratetype = MIGRATE_RESERVE;
1137 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1142 * Obtain a specified number of elements from the buddy allocator, all under
1143 * a single hold of the lock, for efficiency. Add them to the supplied list.
1144 * Returns the number of new pages which were placed at *list.
1146 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1147 unsigned long count, struct list_head *list,
1148 int migratetype, int cold)
1150 int mt = migratetype, i;
1152 spin_lock(&zone->lock);
1153 for (i = 0; i < count; ++i) {
1154 struct page *page = __rmqueue(zone, order, migratetype);
1155 if (unlikely(page == NULL))
1159 * Split buddy pages returned by expand() are received here
1160 * in physical page order. The page is added to the callers and
1161 * list and the list head then moves forward. From the callers
1162 * perspective, the linked list is ordered by page number in
1163 * some conditions. This is useful for IO devices that can
1164 * merge IO requests if the physical pages are ordered
1167 if (likely(cold == 0))
1168 list_add(&page->lru, list);
1170 list_add_tail(&page->lru, list);
1171 if (IS_ENABLED(CONFIG_CMA)) {
1172 mt = get_pageblock_migratetype(page);
1173 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1176 set_freepage_migratetype(page, mt);
1178 if (is_migrate_cma(mt))
1179 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1182 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1183 spin_unlock(&zone->lock);
1189 * Called from the vmstat counter updater to drain pagesets of this
1190 * currently executing processor on remote nodes after they have
1193 * Note that this function must be called with the thread pinned to
1194 * a single processor.
1196 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1198 unsigned long flags;
1200 unsigned long batch;
1202 local_irq_save(flags);
1203 batch = ACCESS_ONCE(pcp->batch);
1204 if (pcp->count >= batch)
1207 to_drain = pcp->count;
1209 free_pcppages_bulk(zone, to_drain, pcp);
1210 pcp->count -= to_drain;
1212 local_irq_restore(flags);
1217 * Drain pages of the indicated processor.
1219 * The processor must either be the current processor and the
1220 * thread pinned to the current processor or a processor that
1223 static void drain_pages(unsigned int cpu)
1225 unsigned long flags;
1228 for_each_populated_zone(zone) {
1229 struct per_cpu_pageset *pset;
1230 struct per_cpu_pages *pcp;
1232 local_irq_save(flags);
1233 pset = per_cpu_ptr(zone->pageset, cpu);
1237 free_pcppages_bulk(zone, pcp->count, pcp);
1240 local_irq_restore(flags);
1245 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1247 void drain_local_pages(void *arg)
1249 drain_pages(smp_processor_id());
1253 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1255 * Note that this code is protected against sending an IPI to an offline
1256 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1257 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1258 * nothing keeps CPUs from showing up after we populated the cpumask and
1259 * before the call to on_each_cpu_mask().
1261 void drain_all_pages(void)
1264 struct per_cpu_pageset *pcp;
1268 * Allocate in the BSS so we wont require allocation in
1269 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1271 static cpumask_t cpus_with_pcps;
1274 * We don't care about racing with CPU hotplug event
1275 * as offline notification will cause the notified
1276 * cpu to drain that CPU pcps and on_each_cpu_mask
1277 * disables preemption as part of its processing
1279 for_each_online_cpu(cpu) {
1280 bool has_pcps = false;
1281 for_each_populated_zone(zone) {
1282 pcp = per_cpu_ptr(zone->pageset, cpu);
1283 if (pcp->pcp.count) {
1289 cpumask_set_cpu(cpu, &cpus_with_pcps);
1291 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1293 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1296 #ifdef CONFIG_HIBERNATION
1298 void mark_free_pages(struct zone *zone)
1300 unsigned long pfn, max_zone_pfn;
1301 unsigned long flags;
1303 struct list_head *curr;
1305 if (zone_is_empty(zone))
1308 spin_lock_irqsave(&zone->lock, flags);
1310 max_zone_pfn = zone_end_pfn(zone);
1311 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1312 if (pfn_valid(pfn)) {
1313 struct page *page = pfn_to_page(pfn);
1315 if (!swsusp_page_is_forbidden(page))
1316 swsusp_unset_page_free(page);
1319 for_each_migratetype_order(order, t) {
1320 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1323 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1324 for (i = 0; i < (1UL << order); i++)
1325 swsusp_set_page_free(pfn_to_page(pfn + i));
1328 spin_unlock_irqrestore(&zone->lock, flags);
1330 #endif /* CONFIG_PM */
1333 * Free a 0-order page
1334 * cold == 1 ? free a cold page : free a hot page
1336 void free_hot_cold_page(struct page *page, int cold)
1338 struct zone *zone = page_zone(page);
1339 struct per_cpu_pages *pcp;
1340 unsigned long flags;
1343 if (!free_pages_prepare(page, 0))
1346 migratetype = get_pageblock_migratetype(page);
1347 set_freepage_migratetype(page, migratetype);
1348 local_irq_save(flags);
1349 __count_vm_event(PGFREE);
1352 * We only track unmovable, reclaimable and movable on pcp lists.
1353 * Free ISOLATE pages back to the allocator because they are being
1354 * offlined but treat RESERVE as movable pages so we can get those
1355 * areas back if necessary. Otherwise, we may have to free
1356 * excessively into the page allocator
1358 if (migratetype >= MIGRATE_PCPTYPES) {
1359 if (unlikely(is_migrate_isolate(migratetype))) {
1360 free_one_page(zone, page, 0, migratetype);
1363 migratetype = MIGRATE_MOVABLE;
1366 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1368 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1370 list_add(&page->lru, &pcp->lists[migratetype]);
1372 if (pcp->count >= pcp->high) {
1373 unsigned long batch = ACCESS_ONCE(pcp->batch);
1374 free_pcppages_bulk(zone, batch, pcp);
1375 pcp->count -= batch;
1379 local_irq_restore(flags);
1383 * Free a list of 0-order pages
1385 void free_hot_cold_page_list(struct list_head *list, int cold)
1387 struct page *page, *next;
1389 list_for_each_entry_safe(page, next, list, lru) {
1390 trace_mm_page_free_batched(page, cold);
1391 free_hot_cold_page(page, cold);
1396 * split_page takes a non-compound higher-order page, and splits it into
1397 * n (1<<order) sub-pages: page[0..n]
1398 * Each sub-page must be freed individually.
1400 * Note: this is probably too low level an operation for use in drivers.
1401 * Please consult with lkml before using this in your driver.
1403 void split_page(struct page *page, unsigned int order)
1407 VM_BUG_ON(PageCompound(page));
1408 VM_BUG_ON(!page_count(page));
1410 #ifdef CONFIG_KMEMCHECK
1412 * Split shadow pages too, because free(page[0]) would
1413 * otherwise free the whole shadow.
1415 if (kmemcheck_page_is_tracked(page))
1416 split_page(virt_to_page(page[0].shadow), order);
1419 for (i = 1; i < (1 << order); i++)
1420 set_page_refcounted(page + i);
1422 EXPORT_SYMBOL_GPL(split_page);
1424 static int __isolate_free_page(struct page *page, unsigned int order)
1426 unsigned long watermark;
1430 BUG_ON(!PageBuddy(page));
1432 zone = page_zone(page);
1433 mt = get_pageblock_migratetype(page);
1435 if (!is_migrate_isolate(mt)) {
1436 /* Obey watermarks as if the page was being allocated */
1437 watermark = low_wmark_pages(zone) + (1 << order);
1438 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1441 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1444 /* Remove page from free list */
1445 list_del(&page->lru);
1446 zone->free_area[order].nr_free--;
1447 rmv_page_order(page);
1449 /* Set the pageblock if the isolated page is at least a pageblock */
1450 if (order >= pageblock_order - 1) {
1451 struct page *endpage = page + (1 << order) - 1;
1452 for (; page < endpage; page += pageblock_nr_pages) {
1453 int mt = get_pageblock_migratetype(page);
1454 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1455 set_pageblock_migratetype(page,
1460 return 1UL << order;
1464 * Similar to split_page except the page is already free. As this is only
1465 * being used for migration, the migratetype of the block also changes.
1466 * As this is called with interrupts disabled, the caller is responsible
1467 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1470 * Note: this is probably too low level an operation for use in drivers.
1471 * Please consult with lkml before using this in your driver.
1473 int split_free_page(struct page *page)
1478 order = page_order(page);
1480 nr_pages = __isolate_free_page(page, order);
1484 /* Split into individual pages */
1485 set_page_refcounted(page);
1486 split_page(page, order);
1491 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1492 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1496 struct page *buffered_rmqueue(struct zone *preferred_zone,
1497 struct zone *zone, int order, gfp_t gfp_flags,
1500 unsigned long flags;
1502 int cold = !!(gfp_flags & __GFP_COLD);
1505 if (likely(order == 0)) {
1506 struct per_cpu_pages *pcp;
1507 struct list_head *list;
1509 local_irq_save(flags);
1510 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1511 list = &pcp->lists[migratetype];
1512 if (list_empty(list)) {
1513 pcp->count += rmqueue_bulk(zone, 0,
1516 if (unlikely(list_empty(list)))
1521 page = list_entry(list->prev, struct page, lru);
1523 page = list_entry(list->next, struct page, lru);
1525 list_del(&page->lru);
1528 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1530 * __GFP_NOFAIL is not to be used in new code.
1532 * All __GFP_NOFAIL callers should be fixed so that they
1533 * properly detect and handle allocation failures.
1535 * We most definitely don't want callers attempting to
1536 * allocate greater than order-1 page units with
1539 WARN_ON_ONCE(order > 1);
1541 spin_lock_irqsave(&zone->lock, flags);
1542 page = __rmqueue(zone, order, migratetype);
1543 spin_unlock(&zone->lock);
1546 __mod_zone_freepage_state(zone, -(1 << order),
1547 get_pageblock_migratetype(page));
1550 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1551 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1552 zone_statistics(preferred_zone, zone, gfp_flags);
1553 local_irq_restore(flags);
1555 VM_BUG_ON(bad_range(zone, page));
1556 if (prep_new_page(page, order, gfp_flags))
1561 local_irq_restore(flags);
1565 #ifdef CONFIG_FAIL_PAGE_ALLOC
1568 struct fault_attr attr;
1570 u32 ignore_gfp_highmem;
1571 u32 ignore_gfp_wait;
1573 } fail_page_alloc = {
1574 .attr = FAULT_ATTR_INITIALIZER,
1575 .ignore_gfp_wait = 1,
1576 .ignore_gfp_highmem = 1,
1580 static int __init setup_fail_page_alloc(char *str)
1582 return setup_fault_attr(&fail_page_alloc.attr, str);
1584 __setup("fail_page_alloc=", setup_fail_page_alloc);
1586 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1588 if (order < fail_page_alloc.min_order)
1590 if (gfp_mask & __GFP_NOFAIL)
1592 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1594 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1597 return should_fail(&fail_page_alloc.attr, 1 << order);
1600 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1602 static int __init fail_page_alloc_debugfs(void)
1604 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1607 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1608 &fail_page_alloc.attr);
1610 return PTR_ERR(dir);
1612 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1613 &fail_page_alloc.ignore_gfp_wait))
1615 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1616 &fail_page_alloc.ignore_gfp_highmem))
1618 if (!debugfs_create_u32("min-order", mode, dir,
1619 &fail_page_alloc.min_order))
1624 debugfs_remove_recursive(dir);
1629 late_initcall(fail_page_alloc_debugfs);
1631 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1633 #else /* CONFIG_FAIL_PAGE_ALLOC */
1635 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1640 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1643 * Return true if free pages are above 'mark'. This takes into account the order
1644 * of the allocation.
1646 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1647 int classzone_idx, int alloc_flags, long free_pages)
1649 /* free_pages my go negative - that's OK */
1651 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1655 free_pages -= (1 << order) - 1;
1656 if (alloc_flags & ALLOC_HIGH)
1658 if (alloc_flags & ALLOC_HARDER)
1661 /* If allocation can't use CMA areas don't use free CMA pages */
1662 if (!(alloc_flags & ALLOC_CMA))
1663 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1666 if (free_pages - free_cma <= min + lowmem_reserve)
1668 for (o = 0; o < order; o++) {
1669 /* At the next order, this order's pages become unavailable */
1670 free_pages -= z->free_area[o].nr_free << o;
1672 /* Require fewer higher order pages to be free */
1675 if (free_pages <= min)
1681 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1682 int classzone_idx, int alloc_flags)
1684 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1685 zone_page_state(z, NR_FREE_PAGES));
1688 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1689 int classzone_idx, int alloc_flags)
1691 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1693 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1694 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1696 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1702 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1703 * skip over zones that are not allowed by the cpuset, or that have
1704 * been recently (in last second) found to be nearly full. See further
1705 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1706 * that have to skip over a lot of full or unallowed zones.
1708 * If the zonelist cache is present in the passed zonelist, then
1709 * returns a pointer to the allowed node mask (either the current
1710 * tasks mems_allowed, or node_states[N_MEMORY].)
1712 * If the zonelist cache is not available for this zonelist, does
1713 * nothing and returns NULL.
1715 * If the fullzones BITMAP in the zonelist cache is stale (more than
1716 * a second since last zap'd) then we zap it out (clear its bits.)
1718 * We hold off even calling zlc_setup, until after we've checked the
1719 * first zone in the zonelist, on the theory that most allocations will
1720 * be satisfied from that first zone, so best to examine that zone as
1721 * quickly as we can.
1723 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1725 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1726 nodemask_t *allowednodes; /* zonelist_cache approximation */
1728 zlc = zonelist->zlcache_ptr;
1732 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1733 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1734 zlc->last_full_zap = jiffies;
1737 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1738 &cpuset_current_mems_allowed :
1739 &node_states[N_MEMORY];
1740 return allowednodes;
1744 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1745 * if it is worth looking at further for free memory:
1746 * 1) Check that the zone isn't thought to be full (doesn't have its
1747 * bit set in the zonelist_cache fullzones BITMAP).
1748 * 2) Check that the zones node (obtained from the zonelist_cache
1749 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1750 * Return true (non-zero) if zone is worth looking at further, or
1751 * else return false (zero) if it is not.
1753 * This check -ignores- the distinction between various watermarks,
1754 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1755 * found to be full for any variation of these watermarks, it will
1756 * be considered full for up to one second by all requests, unless
1757 * we are so low on memory on all allowed nodes that we are forced
1758 * into the second scan of the zonelist.
1760 * In the second scan we ignore this zonelist cache and exactly
1761 * apply the watermarks to all zones, even it is slower to do so.
1762 * We are low on memory in the second scan, and should leave no stone
1763 * unturned looking for a free page.
1765 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1766 nodemask_t *allowednodes)
1768 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1769 int i; /* index of *z in zonelist zones */
1770 int n; /* node that zone *z is on */
1772 zlc = zonelist->zlcache_ptr;
1776 i = z - zonelist->_zonerefs;
1779 /* This zone is worth trying if it is allowed but not full */
1780 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1784 * Given 'z' scanning a zonelist, set the corresponding bit in
1785 * zlc->fullzones, so that subsequent attempts to allocate a page
1786 * from that zone don't waste time re-examining it.
1788 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1790 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1791 int i; /* index of *z in zonelist zones */
1793 zlc = zonelist->zlcache_ptr;
1797 i = z - zonelist->_zonerefs;
1799 set_bit(i, zlc->fullzones);
1803 * clear all zones full, called after direct reclaim makes progress so that
1804 * a zone that was recently full is not skipped over for up to a second
1806 static void zlc_clear_zones_full(struct zonelist *zonelist)
1808 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1810 zlc = zonelist->zlcache_ptr;
1814 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1817 static bool zone_local(struct zone *local_zone, struct zone *zone)
1819 return node_distance(local_zone->node, zone->node) == LOCAL_DISTANCE;
1822 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1824 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1827 static void __paginginit init_zone_allows_reclaim(int nid)
1831 for_each_online_node(i)
1832 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1833 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1835 zone_reclaim_mode = 1;
1838 #else /* CONFIG_NUMA */
1840 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1845 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1846 nodemask_t *allowednodes)
1851 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1855 static void zlc_clear_zones_full(struct zonelist *zonelist)
1859 static bool zone_local(struct zone *local_zone, struct zone *zone)
1864 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1869 static inline void init_zone_allows_reclaim(int nid)
1872 #endif /* CONFIG_NUMA */
1875 * get_page_from_freelist goes through the zonelist trying to allocate
1878 static struct page *
1879 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1880 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1881 struct zone *preferred_zone, int migratetype)
1884 struct page *page = NULL;
1887 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1888 int zlc_active = 0; /* set if using zonelist_cache */
1889 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1891 classzone_idx = zone_idx(preferred_zone);
1894 * Scan zonelist, looking for a zone with enough free.
1895 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1897 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1898 high_zoneidx, nodemask) {
1901 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1902 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1904 if ((alloc_flags & ALLOC_CPUSET) &&
1905 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1907 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1908 if (unlikely(alloc_flags & ALLOC_NO_WATERMARKS))
1911 * Distribute pages in proportion to the individual
1912 * zone size to ensure fair page aging. The zone a
1913 * page was allocated in should have no effect on the
1914 * time the page has in memory before being reclaimed.
1916 * When zone_reclaim_mode is enabled, try to stay in
1917 * local zones in the fastpath. If that fails, the
1918 * slowpath is entered, which will do another pass
1919 * starting with the local zones, but ultimately fall
1920 * back to remote zones that do not partake in the
1921 * fairness round-robin cycle of this zonelist.
1923 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1924 (gfp_mask & GFP_MOVABLE_MASK)) {
1925 if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1927 if (zone_reclaim_mode &&
1928 !zone_local(preferred_zone, zone))
1932 * When allocating a page cache page for writing, we
1933 * want to get it from a zone that is within its dirty
1934 * limit, such that no single zone holds more than its
1935 * proportional share of globally allowed dirty pages.
1936 * The dirty limits take into account the zone's
1937 * lowmem reserves and high watermark so that kswapd
1938 * should be able to balance it without having to
1939 * write pages from its LRU list.
1941 * This may look like it could increase pressure on
1942 * lower zones by failing allocations in higher zones
1943 * before they are full. But the pages that do spill
1944 * over are limited as the lower zones are protected
1945 * by this very same mechanism. It should not become
1946 * a practical burden to them.
1948 * XXX: For now, allow allocations to potentially
1949 * exceed the per-zone dirty limit in the slowpath
1950 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1951 * which is important when on a NUMA setup the allowed
1952 * zones are together not big enough to reach the
1953 * global limit. The proper fix for these situations
1954 * will require awareness of zones in the
1955 * dirty-throttling and the flusher threads.
1957 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1958 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1959 goto this_zone_full;
1961 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1962 if (!zone_watermark_ok(zone, order, mark,
1963 classzone_idx, alloc_flags)) {
1966 if (IS_ENABLED(CONFIG_NUMA) &&
1967 !did_zlc_setup && nr_online_nodes > 1) {
1969 * we do zlc_setup if there are multiple nodes
1970 * and before considering the first zone allowed
1973 allowednodes = zlc_setup(zonelist, alloc_flags);
1978 if (zone_reclaim_mode == 0 ||
1979 !zone_allows_reclaim(preferred_zone, zone))
1980 goto this_zone_full;
1983 * As we may have just activated ZLC, check if the first
1984 * eligible zone has failed zone_reclaim recently.
1986 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1987 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1990 ret = zone_reclaim(zone, gfp_mask, order);
1992 case ZONE_RECLAIM_NOSCAN:
1995 case ZONE_RECLAIM_FULL:
1996 /* scanned but unreclaimable */
1999 /* did we reclaim enough */
2000 if (zone_watermark_ok(zone, order, mark,
2001 classzone_idx, alloc_flags))
2005 * Failed to reclaim enough to meet watermark.
2006 * Only mark the zone full if checking the min
2007 * watermark or if we failed to reclaim just
2008 * 1<<order pages or else the page allocator
2009 * fastpath will prematurely mark zones full
2010 * when the watermark is between the low and
2013 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2014 ret == ZONE_RECLAIM_SOME)
2015 goto this_zone_full;
2022 page = buffered_rmqueue(preferred_zone, zone, order,
2023 gfp_mask, migratetype);
2027 if (IS_ENABLED(CONFIG_NUMA))
2028 zlc_mark_zone_full(zonelist, z);
2031 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2032 /* Disable zlc cache for second zonelist scan */
2039 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2040 * necessary to allocate the page. The expectation is
2041 * that the caller is taking steps that will free more
2042 * memory. The caller should avoid the page being used
2043 * for !PFMEMALLOC purposes.
2045 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2051 * Large machines with many possible nodes should not always dump per-node
2052 * meminfo in irq context.
2054 static inline bool should_suppress_show_mem(void)
2059 ret = in_interrupt();
2064 static DEFINE_RATELIMIT_STATE(nopage_rs,
2065 DEFAULT_RATELIMIT_INTERVAL,
2066 DEFAULT_RATELIMIT_BURST);
2068 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2070 unsigned int filter = SHOW_MEM_FILTER_NODES;
2072 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2073 debug_guardpage_minorder() > 0)
2077 * Walking all memory to count page types is very expensive and should
2078 * be inhibited in non-blockable contexts.
2080 if (!(gfp_mask & __GFP_WAIT))
2081 filter |= SHOW_MEM_FILTER_PAGE_COUNT;
2084 * This documents exceptions given to allocations in certain
2085 * contexts that are allowed to allocate outside current's set
2088 if (!(gfp_mask & __GFP_NOMEMALLOC))
2089 if (test_thread_flag(TIF_MEMDIE) ||
2090 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2091 filter &= ~SHOW_MEM_FILTER_NODES;
2092 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2093 filter &= ~SHOW_MEM_FILTER_NODES;
2096 struct va_format vaf;
2099 va_start(args, fmt);
2104 pr_warn("%pV", &vaf);
2109 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2110 current->comm, order, gfp_mask);
2113 if (!should_suppress_show_mem())
2118 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2119 unsigned long did_some_progress,
2120 unsigned long pages_reclaimed)
2122 /* Do not loop if specifically requested */
2123 if (gfp_mask & __GFP_NORETRY)
2126 /* Always retry if specifically requested */
2127 if (gfp_mask & __GFP_NOFAIL)
2131 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2132 * making forward progress without invoking OOM. Suspend also disables
2133 * storage devices so kswapd will not help. Bail if we are suspending.
2135 if (!did_some_progress && pm_suspended_storage())
2139 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2140 * means __GFP_NOFAIL, but that may not be true in other
2143 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2147 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2148 * specified, then we retry until we no longer reclaim any pages
2149 * (above), or we've reclaimed an order of pages at least as
2150 * large as the allocation's order. In both cases, if the
2151 * allocation still fails, we stop retrying.
2153 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2159 static inline struct page *
2160 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2161 struct zonelist *zonelist, enum zone_type high_zoneidx,
2162 nodemask_t *nodemask, struct zone *preferred_zone,
2167 /* Acquire the OOM killer lock for the zones in zonelist */
2168 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2169 schedule_timeout_uninterruptible(1);
2174 * Go through the zonelist yet one more time, keep very high watermark
2175 * here, this is only to catch a parallel oom killing, we must fail if
2176 * we're still under heavy pressure.
2178 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2179 order, zonelist, high_zoneidx,
2180 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2181 preferred_zone, migratetype);
2185 if (!(gfp_mask & __GFP_NOFAIL)) {
2186 /* The OOM killer will not help higher order allocs */
2187 if (order > PAGE_ALLOC_COSTLY_ORDER)
2189 /* The OOM killer does not needlessly kill tasks for lowmem */
2190 if (high_zoneidx < ZONE_NORMAL)
2193 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2194 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2195 * The caller should handle page allocation failure by itself if
2196 * it specifies __GFP_THISNODE.
2197 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2199 if (gfp_mask & __GFP_THISNODE)
2202 /* Exhausted what can be done so it's blamo time */
2203 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2206 clear_zonelist_oom(zonelist, gfp_mask);
2210 #ifdef CONFIG_COMPACTION
2211 /* Try memory compaction for high-order allocations before reclaim */
2212 static struct page *
2213 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2214 struct zonelist *zonelist, enum zone_type high_zoneidx,
2215 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2216 int migratetype, bool sync_migration,
2217 bool *contended_compaction, bool *deferred_compaction,
2218 unsigned long *did_some_progress)
2223 if (compaction_deferred(preferred_zone, order)) {
2224 *deferred_compaction = true;
2228 current->flags |= PF_MEMALLOC;
2229 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2230 nodemask, sync_migration,
2231 contended_compaction);
2232 current->flags &= ~PF_MEMALLOC;
2234 if (*did_some_progress != COMPACT_SKIPPED) {
2237 /* Page migration frees to the PCP lists but we want merging */
2238 drain_pages(get_cpu());
2241 page = get_page_from_freelist(gfp_mask, nodemask,
2242 order, zonelist, high_zoneidx,
2243 alloc_flags & ~ALLOC_NO_WATERMARKS,
2244 preferred_zone, migratetype);
2246 preferred_zone->compact_blockskip_flush = false;
2247 preferred_zone->compact_considered = 0;
2248 preferred_zone->compact_defer_shift = 0;
2249 if (order >= preferred_zone->compact_order_failed)
2250 preferred_zone->compact_order_failed = order + 1;
2251 count_vm_event(COMPACTSUCCESS);
2256 * It's bad if compaction run occurs and fails.
2257 * The most likely reason is that pages exist,
2258 * but not enough to satisfy watermarks.
2260 count_vm_event(COMPACTFAIL);
2263 * As async compaction considers a subset of pageblocks, only
2264 * defer if the failure was a sync compaction failure.
2267 defer_compaction(preferred_zone, order);
2275 static inline struct page *
2276 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2277 struct zonelist *zonelist, enum zone_type high_zoneidx,
2278 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2279 int migratetype, bool sync_migration,
2280 bool *contended_compaction, bool *deferred_compaction,
2281 unsigned long *did_some_progress)
2285 #endif /* CONFIG_COMPACTION */
2287 /* Perform direct synchronous page reclaim */
2289 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2290 nodemask_t *nodemask)
2292 struct reclaim_state reclaim_state;
2297 /* We now go into synchronous reclaim */
2298 cpuset_memory_pressure_bump();
2299 current->flags |= PF_MEMALLOC;
2300 lockdep_set_current_reclaim_state(gfp_mask);
2301 reclaim_state.reclaimed_slab = 0;
2302 current->reclaim_state = &reclaim_state;
2304 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2306 current->reclaim_state = NULL;
2307 lockdep_clear_current_reclaim_state();
2308 current->flags &= ~PF_MEMALLOC;
2315 /* The really slow allocator path where we enter direct reclaim */
2316 static inline struct page *
2317 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2318 struct zonelist *zonelist, enum zone_type high_zoneidx,
2319 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2320 int migratetype, unsigned long *did_some_progress)
2322 struct page *page = NULL;
2323 bool drained = false;
2325 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2327 if (unlikely(!(*did_some_progress)))
2330 /* After successful reclaim, reconsider all zones for allocation */
2331 if (IS_ENABLED(CONFIG_NUMA))
2332 zlc_clear_zones_full(zonelist);
2335 page = get_page_from_freelist(gfp_mask, nodemask, order,
2336 zonelist, high_zoneidx,
2337 alloc_flags & ~ALLOC_NO_WATERMARKS,
2338 preferred_zone, migratetype);
2341 * If an allocation failed after direct reclaim, it could be because
2342 * pages are pinned on the per-cpu lists. Drain them and try again
2344 if (!page && !drained) {
2354 * This is called in the allocator slow-path if the allocation request is of
2355 * sufficient urgency to ignore watermarks and take other desperate measures
2357 static inline struct page *
2358 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2359 struct zonelist *zonelist, enum zone_type high_zoneidx,
2360 nodemask_t *nodemask, struct zone *preferred_zone,
2366 page = get_page_from_freelist(gfp_mask, nodemask, order,
2367 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2368 preferred_zone, migratetype);
2370 if (!page && gfp_mask & __GFP_NOFAIL)
2371 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2372 } while (!page && (gfp_mask & __GFP_NOFAIL));
2377 static void prepare_slowpath(gfp_t gfp_mask, unsigned int order,
2378 struct zonelist *zonelist,
2379 enum zone_type high_zoneidx,
2380 struct zone *preferred_zone)
2385 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2386 if (!(gfp_mask & __GFP_NO_KSWAPD))
2387 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2389 * Only reset the batches of zones that were actually
2390 * considered in the fast path, we don't want to
2391 * thrash fairness information for zones that are not
2392 * actually part of this zonelist's round-robin cycle.
2394 if (zone_reclaim_mode && !zone_local(preferred_zone, zone))
2396 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2397 high_wmark_pages(zone) -
2398 low_wmark_pages(zone) -
2399 zone_page_state(zone, NR_ALLOC_BATCH));
2404 gfp_to_alloc_flags(gfp_t gfp_mask)
2406 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2407 const gfp_t wait = gfp_mask & __GFP_WAIT;
2409 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2410 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2413 * The caller may dip into page reserves a bit more if the caller
2414 * cannot run direct reclaim, or if the caller has realtime scheduling
2415 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2416 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2418 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2422 * Not worth trying to allocate harder for
2423 * __GFP_NOMEMALLOC even if it can't schedule.
2425 if (!(gfp_mask & __GFP_NOMEMALLOC))
2426 alloc_flags |= ALLOC_HARDER;
2428 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2429 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2431 alloc_flags &= ~ALLOC_CPUSET;
2432 } else if (unlikely(rt_task(current)) && !in_interrupt())
2433 alloc_flags |= ALLOC_HARDER;
2435 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2436 if (gfp_mask & __GFP_MEMALLOC)
2437 alloc_flags |= ALLOC_NO_WATERMARKS;
2438 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2439 alloc_flags |= ALLOC_NO_WATERMARKS;
2440 else if (!in_interrupt() &&
2441 ((current->flags & PF_MEMALLOC) ||
2442 unlikely(test_thread_flag(TIF_MEMDIE))))
2443 alloc_flags |= ALLOC_NO_WATERMARKS;
2446 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2447 alloc_flags |= ALLOC_CMA;
2452 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2454 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2457 static inline struct page *
2458 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2459 struct zonelist *zonelist, enum zone_type high_zoneidx,
2460 nodemask_t *nodemask, struct zone *preferred_zone,
2463 const gfp_t wait = gfp_mask & __GFP_WAIT;
2464 struct page *page = NULL;
2466 unsigned long pages_reclaimed = 0;
2467 unsigned long did_some_progress;
2468 bool sync_migration = false;
2469 bool deferred_compaction = false;
2470 bool contended_compaction = false;
2473 * In the slowpath, we sanity check order to avoid ever trying to
2474 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2475 * be using allocators in order of preference for an area that is
2478 if (order >= MAX_ORDER) {
2479 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2484 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2485 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2486 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2487 * using a larger set of nodes after it has established that the
2488 * allowed per node queues are empty and that nodes are
2491 if (IS_ENABLED(CONFIG_NUMA) &&
2492 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2496 prepare_slowpath(gfp_mask, order, zonelist,
2497 high_zoneidx, preferred_zone);
2500 * OK, we're below the kswapd watermark and have kicked background
2501 * reclaim. Now things get more complex, so set up alloc_flags according
2502 * to how we want to proceed.
2504 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2507 * Find the true preferred zone if the allocation is unconstrained by
2510 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2511 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2515 /* This is the last chance, in general, before the goto nopage. */
2516 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2517 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2518 preferred_zone, migratetype);
2522 /* Allocate without watermarks if the context allows */
2523 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2525 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2526 * the allocation is high priority and these type of
2527 * allocations are system rather than user orientated
2529 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2531 page = __alloc_pages_high_priority(gfp_mask, order,
2532 zonelist, high_zoneidx, nodemask,
2533 preferred_zone, migratetype);
2539 /* Atomic allocations - we can't balance anything */
2543 /* Avoid recursion of direct reclaim */
2544 if (current->flags & PF_MEMALLOC)
2547 /* Avoid allocations with no watermarks from looping endlessly */
2548 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2552 * Try direct compaction. The first pass is asynchronous. Subsequent
2553 * attempts after direct reclaim are synchronous
2555 page = __alloc_pages_direct_compact(gfp_mask, order,
2556 zonelist, high_zoneidx,
2558 alloc_flags, preferred_zone,
2559 migratetype, sync_migration,
2560 &contended_compaction,
2561 &deferred_compaction,
2562 &did_some_progress);
2565 sync_migration = true;
2568 * If compaction is deferred for high-order allocations, it is because
2569 * sync compaction recently failed. In this is the case and the caller
2570 * requested a movable allocation that does not heavily disrupt the
2571 * system then fail the allocation instead of entering direct reclaim.
2573 if ((deferred_compaction || contended_compaction) &&
2574 (gfp_mask & __GFP_NO_KSWAPD))
2577 /* Try direct reclaim and then allocating */
2578 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2579 zonelist, high_zoneidx,
2581 alloc_flags, preferred_zone,
2582 migratetype, &did_some_progress);
2587 * If we failed to make any progress reclaiming, then we are
2588 * running out of options and have to consider going OOM
2590 if (!did_some_progress) {
2591 if (oom_gfp_allowed(gfp_mask)) {
2592 if (oom_killer_disabled)
2594 /* Coredumps can quickly deplete all memory reserves */
2595 if ((current->flags & PF_DUMPCORE) &&
2596 !(gfp_mask & __GFP_NOFAIL))
2598 page = __alloc_pages_may_oom(gfp_mask, order,
2599 zonelist, high_zoneidx,
2600 nodemask, preferred_zone,
2605 if (!(gfp_mask & __GFP_NOFAIL)) {
2607 * The oom killer is not called for high-order
2608 * allocations that may fail, so if no progress
2609 * is being made, there are no other options and
2610 * retrying is unlikely to help.
2612 if (order > PAGE_ALLOC_COSTLY_ORDER)
2615 * The oom killer is not called for lowmem
2616 * allocations to prevent needlessly killing
2619 if (high_zoneidx < ZONE_NORMAL)
2627 /* Check if we should retry the allocation */
2628 pages_reclaimed += did_some_progress;
2629 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2631 /* Wait for some write requests to complete then retry */
2632 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2636 * High-order allocations do not necessarily loop after
2637 * direct reclaim and reclaim/compaction depends on compaction
2638 * being called after reclaim so call directly if necessary
2640 page = __alloc_pages_direct_compact(gfp_mask, order,
2641 zonelist, high_zoneidx,
2643 alloc_flags, preferred_zone,
2644 migratetype, sync_migration,
2645 &contended_compaction,
2646 &deferred_compaction,
2647 &did_some_progress);
2653 warn_alloc_failed(gfp_mask, order, NULL);
2656 if (kmemcheck_enabled)
2657 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2663 * This is the 'heart' of the zoned buddy allocator.
2666 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2667 struct zonelist *zonelist, nodemask_t *nodemask)
2669 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2670 struct zone *preferred_zone;
2671 struct page *page = NULL;
2672 int migratetype = allocflags_to_migratetype(gfp_mask);
2673 unsigned int cpuset_mems_cookie;
2674 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2675 struct mem_cgroup *memcg = NULL;
2677 gfp_mask &= gfp_allowed_mask;
2679 lockdep_trace_alloc(gfp_mask);
2681 might_sleep_if(gfp_mask & __GFP_WAIT);
2683 if (should_fail_alloc_page(gfp_mask, order))
2687 * Check the zones suitable for the gfp_mask contain at least one
2688 * valid zone. It's possible to have an empty zonelist as a result
2689 * of GFP_THISNODE and a memoryless node
2691 if (unlikely(!zonelist->_zonerefs->zone))
2695 * Will only have any effect when __GFP_KMEMCG is set. This is
2696 * verified in the (always inline) callee
2698 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2702 cpuset_mems_cookie = get_mems_allowed();
2704 /* The preferred zone is used for statistics later */
2705 first_zones_zonelist(zonelist, high_zoneidx,
2706 nodemask ? : &cpuset_current_mems_allowed,
2708 if (!preferred_zone)
2712 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2713 alloc_flags |= ALLOC_CMA;
2715 /* First allocation attempt */
2716 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2717 zonelist, high_zoneidx, alloc_flags,
2718 preferred_zone, migratetype);
2719 if (unlikely(!page)) {
2721 * Runtime PM, block IO and its error handling path
2722 * can deadlock because I/O on the device might not
2725 gfp_mask = memalloc_noio_flags(gfp_mask);
2726 page = __alloc_pages_slowpath(gfp_mask, order,
2727 zonelist, high_zoneidx, nodemask,
2728 preferred_zone, migratetype);
2731 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2735 * When updating a task's mems_allowed, it is possible to race with
2736 * parallel threads in such a way that an allocation can fail while
2737 * the mask is being updated. If a page allocation is about to fail,
2738 * check if the cpuset changed during allocation and if so, retry.
2740 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2743 memcg_kmem_commit_charge(page, memcg, order);
2747 EXPORT_SYMBOL(__alloc_pages_nodemask);
2750 * Common helper functions.
2752 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2757 * __get_free_pages() returns a 32-bit address, which cannot represent
2760 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2762 page = alloc_pages(gfp_mask, order);
2765 return (unsigned long) page_address(page);
2767 EXPORT_SYMBOL(__get_free_pages);
2769 unsigned long get_zeroed_page(gfp_t gfp_mask)
2771 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2773 EXPORT_SYMBOL(get_zeroed_page);
2775 void __free_pages(struct page *page, unsigned int order)
2777 if (put_page_testzero(page)) {
2779 free_hot_cold_page(page, 0);
2781 __free_pages_ok(page, order);
2785 EXPORT_SYMBOL(__free_pages);
2787 void free_pages(unsigned long addr, unsigned int order)
2790 VM_BUG_ON(!virt_addr_valid((void *)addr));
2791 __free_pages(virt_to_page((void *)addr), order);
2795 EXPORT_SYMBOL(free_pages);
2798 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2799 * pages allocated with __GFP_KMEMCG.
2801 * Those pages are accounted to a particular memcg, embedded in the
2802 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2803 * for that information only to find out that it is NULL for users who have no
2804 * interest in that whatsoever, we provide these functions.
2806 * The caller knows better which flags it relies on.
2808 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2810 memcg_kmem_uncharge_pages(page, order);
2811 __free_pages(page, order);
2814 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2817 VM_BUG_ON(!virt_addr_valid((void *)addr));
2818 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2822 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2825 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2826 unsigned long used = addr + PAGE_ALIGN(size);
2828 split_page(virt_to_page((void *)addr), order);
2829 while (used < alloc_end) {
2834 return (void *)addr;
2838 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2839 * @size: the number of bytes to allocate
2840 * @gfp_mask: GFP flags for the allocation
2842 * This function is similar to alloc_pages(), except that it allocates the
2843 * minimum number of pages to satisfy the request. alloc_pages() can only
2844 * allocate memory in power-of-two pages.
2846 * This function is also limited by MAX_ORDER.
2848 * Memory allocated by this function must be released by free_pages_exact().
2850 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2852 unsigned int order = get_order(size);
2855 addr = __get_free_pages(gfp_mask, order);
2856 return make_alloc_exact(addr, order, size);
2858 EXPORT_SYMBOL(alloc_pages_exact);
2861 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2863 * @nid: the preferred node ID where memory should be allocated
2864 * @size: the number of bytes to allocate
2865 * @gfp_mask: GFP flags for the allocation
2867 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2869 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2872 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2874 unsigned order = get_order(size);
2875 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2878 return make_alloc_exact((unsigned long)page_address(p), order, size);
2880 EXPORT_SYMBOL(alloc_pages_exact_nid);
2883 * free_pages_exact - release memory allocated via alloc_pages_exact()
2884 * @virt: the value returned by alloc_pages_exact.
2885 * @size: size of allocation, same value as passed to alloc_pages_exact().
2887 * Release the memory allocated by a previous call to alloc_pages_exact.
2889 void free_pages_exact(void *virt, size_t size)
2891 unsigned long addr = (unsigned long)virt;
2892 unsigned long end = addr + PAGE_ALIGN(size);
2894 while (addr < end) {
2899 EXPORT_SYMBOL(free_pages_exact);
2902 * nr_free_zone_pages - count number of pages beyond high watermark
2903 * @offset: The zone index of the highest zone
2905 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2906 * high watermark within all zones at or below a given zone index. For each
2907 * zone, the number of pages is calculated as:
2908 * managed_pages - high_pages
2910 static unsigned long nr_free_zone_pages(int offset)
2915 /* Just pick one node, since fallback list is circular */
2916 unsigned long sum = 0;
2918 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2920 for_each_zone_zonelist(zone, z, zonelist, offset) {
2921 unsigned long size = zone->managed_pages;
2922 unsigned long high = high_wmark_pages(zone);
2931 * nr_free_buffer_pages - count number of pages beyond high watermark
2933 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2934 * watermark within ZONE_DMA and ZONE_NORMAL.
2936 unsigned long nr_free_buffer_pages(void)
2938 return nr_free_zone_pages(gfp_zone(GFP_USER));
2940 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2943 * nr_free_pagecache_pages - count number of pages beyond high watermark
2945 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2946 * high watermark within all zones.
2948 unsigned long nr_free_pagecache_pages(void)
2950 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2953 static inline void show_node(struct zone *zone)
2955 if (IS_ENABLED(CONFIG_NUMA))
2956 printk("Node %d ", zone_to_nid(zone));
2959 void si_meminfo(struct sysinfo *val)
2961 val->totalram = totalram_pages;
2963 val->freeram = global_page_state(NR_FREE_PAGES);
2964 val->bufferram = nr_blockdev_pages();
2965 val->totalhigh = totalhigh_pages;
2966 val->freehigh = nr_free_highpages();
2967 val->mem_unit = PAGE_SIZE;
2970 EXPORT_SYMBOL(si_meminfo);
2973 void si_meminfo_node(struct sysinfo *val, int nid)
2975 int zone_type; /* needs to be signed */
2976 unsigned long managed_pages = 0;
2977 pg_data_t *pgdat = NODE_DATA(nid);
2979 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
2980 managed_pages += pgdat->node_zones[zone_type].managed_pages;
2981 val->totalram = managed_pages;
2982 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2983 #ifdef CONFIG_HIGHMEM
2984 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
2985 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2991 val->mem_unit = PAGE_SIZE;
2996 * Determine whether the node should be displayed or not, depending on whether
2997 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2999 bool skip_free_areas_node(unsigned int flags, int nid)
3002 unsigned int cpuset_mems_cookie;
3004 if (!(flags & SHOW_MEM_FILTER_NODES))
3008 cpuset_mems_cookie = get_mems_allowed();
3009 ret = !node_isset(nid, cpuset_current_mems_allowed);
3010 } while (!put_mems_allowed(cpuset_mems_cookie));
3015 #define K(x) ((x) << (PAGE_SHIFT-10))
3017 static void show_migration_types(unsigned char type)
3019 static const char types[MIGRATE_TYPES] = {
3020 [MIGRATE_UNMOVABLE] = 'U',
3021 [MIGRATE_RECLAIMABLE] = 'E',
3022 [MIGRATE_MOVABLE] = 'M',
3023 [MIGRATE_RESERVE] = 'R',
3025 [MIGRATE_CMA] = 'C',
3027 #ifdef CONFIG_MEMORY_ISOLATION
3028 [MIGRATE_ISOLATE] = 'I',
3031 char tmp[MIGRATE_TYPES + 1];
3035 for (i = 0; i < MIGRATE_TYPES; i++) {
3036 if (type & (1 << i))
3041 printk("(%s) ", tmp);
3045 * Show free area list (used inside shift_scroll-lock stuff)
3046 * We also calculate the percentage fragmentation. We do this by counting the
3047 * memory on each free list with the exception of the first item on the list.
3048 * Suppresses nodes that are not allowed by current's cpuset if
3049 * SHOW_MEM_FILTER_NODES is passed.
3051 void show_free_areas(unsigned int filter)
3056 for_each_populated_zone(zone) {
3057 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3060 printk("%s per-cpu:\n", zone->name);
3062 for_each_online_cpu(cpu) {
3063 struct per_cpu_pageset *pageset;
3065 pageset = per_cpu_ptr(zone->pageset, cpu);
3067 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3068 cpu, pageset->pcp.high,
3069 pageset->pcp.batch, pageset->pcp.count);
3073 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3074 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3076 " dirty:%lu writeback:%lu unstable:%lu\n"
3077 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3078 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3080 global_page_state(NR_ACTIVE_ANON),
3081 global_page_state(NR_INACTIVE_ANON),
3082 global_page_state(NR_ISOLATED_ANON),
3083 global_page_state(NR_ACTIVE_FILE),
3084 global_page_state(NR_INACTIVE_FILE),
3085 global_page_state(NR_ISOLATED_FILE),
3086 global_page_state(NR_UNEVICTABLE),
3087 global_page_state(NR_FILE_DIRTY),
3088 global_page_state(NR_WRITEBACK),
3089 global_page_state(NR_UNSTABLE_NFS),
3090 global_page_state(NR_FREE_PAGES),
3091 global_page_state(NR_SLAB_RECLAIMABLE),
3092 global_page_state(NR_SLAB_UNRECLAIMABLE),
3093 global_page_state(NR_FILE_MAPPED),
3094 global_page_state(NR_SHMEM),
3095 global_page_state(NR_PAGETABLE),
3096 global_page_state(NR_BOUNCE),
3097 global_page_state(NR_FREE_CMA_PAGES));
3099 for_each_populated_zone(zone) {
3102 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3110 " active_anon:%lukB"
3111 " inactive_anon:%lukB"
3112 " active_file:%lukB"
3113 " inactive_file:%lukB"
3114 " unevictable:%lukB"
3115 " isolated(anon):%lukB"
3116 " isolated(file):%lukB"
3124 " slab_reclaimable:%lukB"
3125 " slab_unreclaimable:%lukB"
3126 " kernel_stack:%lukB"
3131 " writeback_tmp:%lukB"
3132 " pages_scanned:%lu"
3133 " all_unreclaimable? %s"
3136 K(zone_page_state(zone, NR_FREE_PAGES)),
3137 K(min_wmark_pages(zone)),
3138 K(low_wmark_pages(zone)),
3139 K(high_wmark_pages(zone)),
3140 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3141 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3142 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3143 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3144 K(zone_page_state(zone, NR_UNEVICTABLE)),
3145 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3146 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3147 K(zone->present_pages),
3148 K(zone->managed_pages),
3149 K(zone_page_state(zone, NR_MLOCK)),
3150 K(zone_page_state(zone, NR_FILE_DIRTY)),
3151 K(zone_page_state(zone, NR_WRITEBACK)),
3152 K(zone_page_state(zone, NR_FILE_MAPPED)),
3153 K(zone_page_state(zone, NR_SHMEM)),
3154 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3155 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3156 zone_page_state(zone, NR_KERNEL_STACK) *
3158 K(zone_page_state(zone, NR_PAGETABLE)),
3159 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3160 K(zone_page_state(zone, NR_BOUNCE)),
3161 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3162 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3163 zone->pages_scanned,
3164 (!zone_reclaimable(zone) ? "yes" : "no")
3166 printk("lowmem_reserve[]:");
3167 for (i = 0; i < MAX_NR_ZONES; i++)
3168 printk(" %lu", zone->lowmem_reserve[i]);
3172 for_each_populated_zone(zone) {
3173 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3174 unsigned char types[MAX_ORDER];
3176 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3179 printk("%s: ", zone->name);
3181 spin_lock_irqsave(&zone->lock, flags);
3182 for (order = 0; order < MAX_ORDER; order++) {
3183 struct free_area *area = &zone->free_area[order];
3186 nr[order] = area->nr_free;
3187 total += nr[order] << order;
3190 for (type = 0; type < MIGRATE_TYPES; type++) {
3191 if (!list_empty(&area->free_list[type]))
3192 types[order] |= 1 << type;
3195 spin_unlock_irqrestore(&zone->lock, flags);
3196 for (order = 0; order < MAX_ORDER; order++) {
3197 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3199 show_migration_types(types[order]);
3201 printk("= %lukB\n", K(total));
3204 hugetlb_show_meminfo();
3206 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3208 show_swap_cache_info();
3211 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3213 zoneref->zone = zone;
3214 zoneref->zone_idx = zone_idx(zone);
3218 * Builds allocation fallback zone lists.
3220 * Add all populated zones of a node to the zonelist.
3222 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3226 enum zone_type zone_type = MAX_NR_ZONES;
3230 zone = pgdat->node_zones + zone_type;
3231 if (populated_zone(zone)) {
3232 zoneref_set_zone(zone,
3233 &zonelist->_zonerefs[nr_zones++]);
3234 check_highest_zone(zone_type);
3236 } while (zone_type);
3244 * 0 = automatic detection of better ordering.
3245 * 1 = order by ([node] distance, -zonetype)
3246 * 2 = order by (-zonetype, [node] distance)
3248 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3249 * the same zonelist. So only NUMA can configure this param.
3251 #define ZONELIST_ORDER_DEFAULT 0
3252 #define ZONELIST_ORDER_NODE 1
3253 #define ZONELIST_ORDER_ZONE 2
3255 /* zonelist order in the kernel.
3256 * set_zonelist_order() will set this to NODE or ZONE.
3258 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3259 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3263 /* The value user specified ....changed by config */
3264 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3265 /* string for sysctl */
3266 #define NUMA_ZONELIST_ORDER_LEN 16
3267 char numa_zonelist_order[16] = "default";
3270 * interface for configure zonelist ordering.
3271 * command line option "numa_zonelist_order"
3272 * = "[dD]efault - default, automatic configuration.
3273 * = "[nN]ode - order by node locality, then by zone within node
3274 * = "[zZ]one - order by zone, then by locality within zone
3277 static int __parse_numa_zonelist_order(char *s)
3279 if (*s == 'd' || *s == 'D') {
3280 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3281 } else if (*s == 'n' || *s == 'N') {
3282 user_zonelist_order = ZONELIST_ORDER_NODE;
3283 } else if (*s == 'z' || *s == 'Z') {
3284 user_zonelist_order = ZONELIST_ORDER_ZONE;
3287 "Ignoring invalid numa_zonelist_order value: "
3294 static __init int setup_numa_zonelist_order(char *s)
3301 ret = __parse_numa_zonelist_order(s);
3303 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3307 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3310 * sysctl handler for numa_zonelist_order
3312 int numa_zonelist_order_handler(ctl_table *table, int write,
3313 void __user *buffer, size_t *length,
3316 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3318 static DEFINE_MUTEX(zl_order_mutex);
3320 mutex_lock(&zl_order_mutex);
3322 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3326 strcpy(saved_string, (char *)table->data);
3328 ret = proc_dostring(table, write, buffer, length, ppos);
3332 int oldval = user_zonelist_order;
3334 ret = __parse_numa_zonelist_order((char *)table->data);
3337 * bogus value. restore saved string
3339 strncpy((char *)table->data, saved_string,
3340 NUMA_ZONELIST_ORDER_LEN);
3341 user_zonelist_order = oldval;
3342 } else if (oldval != user_zonelist_order) {
3343 mutex_lock(&zonelists_mutex);
3344 build_all_zonelists(NULL, NULL);
3345 mutex_unlock(&zonelists_mutex);
3349 mutex_unlock(&zl_order_mutex);
3354 #define MAX_NODE_LOAD (nr_online_nodes)
3355 static int node_load[MAX_NUMNODES];
3358 * find_next_best_node - find the next node that should appear in a given node's fallback list
3359 * @node: node whose fallback list we're appending
3360 * @used_node_mask: nodemask_t of already used nodes
3362 * We use a number of factors to determine which is the next node that should
3363 * appear on a given node's fallback list. The node should not have appeared
3364 * already in @node's fallback list, and it should be the next closest node
3365 * according to the distance array (which contains arbitrary distance values
3366 * from each node to each node in the system), and should also prefer nodes
3367 * with no CPUs, since presumably they'll have very little allocation pressure
3368 * on them otherwise.
3369 * It returns -1 if no node is found.
3371 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3374 int min_val = INT_MAX;
3375 int best_node = NUMA_NO_NODE;
3376 const struct cpumask *tmp = cpumask_of_node(0);
3378 /* Use the local node if we haven't already */
3379 if (!node_isset(node, *used_node_mask)) {
3380 node_set(node, *used_node_mask);
3384 for_each_node_state(n, N_MEMORY) {
3386 /* Don't want a node to appear more than once */
3387 if (node_isset(n, *used_node_mask))
3390 /* Use the distance array to find the distance */
3391 val = node_distance(node, n);
3393 /* Penalize nodes under us ("prefer the next node") */
3396 /* Give preference to headless and unused nodes */
3397 tmp = cpumask_of_node(n);
3398 if (!cpumask_empty(tmp))
3399 val += PENALTY_FOR_NODE_WITH_CPUS;
3401 /* Slight preference for less loaded node */
3402 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3403 val += node_load[n];
3405 if (val < min_val) {
3412 node_set(best_node, *used_node_mask);
3419 * Build zonelists ordered by node and zones within node.
3420 * This results in maximum locality--normal zone overflows into local
3421 * DMA zone, if any--but risks exhausting DMA zone.
3423 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3426 struct zonelist *zonelist;
3428 zonelist = &pgdat->node_zonelists[0];
3429 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3431 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3432 zonelist->_zonerefs[j].zone = NULL;
3433 zonelist->_zonerefs[j].zone_idx = 0;
3437 * Build gfp_thisnode zonelists
3439 static void build_thisnode_zonelists(pg_data_t *pgdat)
3442 struct zonelist *zonelist;
3444 zonelist = &pgdat->node_zonelists[1];
3445 j = build_zonelists_node(pgdat, zonelist, 0);
3446 zonelist->_zonerefs[j].zone = NULL;
3447 zonelist->_zonerefs[j].zone_idx = 0;
3451 * Build zonelists ordered by zone and nodes within zones.
3452 * This results in conserving DMA zone[s] until all Normal memory is
3453 * exhausted, but results in overflowing to remote node while memory
3454 * may still exist in local DMA zone.
3456 static int node_order[MAX_NUMNODES];
3458 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3461 int zone_type; /* needs to be signed */
3463 struct zonelist *zonelist;
3465 zonelist = &pgdat->node_zonelists[0];
3467 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3468 for (j = 0; j < nr_nodes; j++) {
3469 node = node_order[j];
3470 z = &NODE_DATA(node)->node_zones[zone_type];
3471 if (populated_zone(z)) {
3473 &zonelist->_zonerefs[pos++]);
3474 check_highest_zone(zone_type);
3478 zonelist->_zonerefs[pos].zone = NULL;
3479 zonelist->_zonerefs[pos].zone_idx = 0;
3482 static int default_zonelist_order(void)
3485 unsigned long low_kmem_size, total_size;
3489 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3490 * If they are really small and used heavily, the system can fall
3491 * into OOM very easily.
3492 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3494 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3497 for_each_online_node(nid) {
3498 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3499 z = &NODE_DATA(nid)->node_zones[zone_type];
3500 if (populated_zone(z)) {
3501 if (zone_type < ZONE_NORMAL)
3502 low_kmem_size += z->managed_pages;
3503 total_size += z->managed_pages;
3504 } else if (zone_type == ZONE_NORMAL) {
3506 * If any node has only lowmem, then node order
3507 * is preferred to allow kernel allocations
3508 * locally; otherwise, they can easily infringe
3509 * on other nodes when there is an abundance of
3510 * lowmem available to allocate from.
3512 return ZONELIST_ORDER_NODE;
3516 if (!low_kmem_size || /* there are no DMA area. */
3517 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3518 return ZONELIST_ORDER_NODE;
3520 * look into each node's config.
3521 * If there is a node whose DMA/DMA32 memory is very big area on
3522 * local memory, NODE_ORDER may be suitable.
3524 average_size = total_size /
3525 (nodes_weight(node_states[N_MEMORY]) + 1);
3526 for_each_online_node(nid) {
3529 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3530 z = &NODE_DATA(nid)->node_zones[zone_type];
3531 if (populated_zone(z)) {
3532 if (zone_type < ZONE_NORMAL)
3533 low_kmem_size += z->present_pages;
3534 total_size += z->present_pages;
3537 if (low_kmem_size &&
3538 total_size > average_size && /* ignore small node */
3539 low_kmem_size > total_size * 70/100)
3540 return ZONELIST_ORDER_NODE;
3542 return ZONELIST_ORDER_ZONE;
3545 static void set_zonelist_order(void)
3547 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3548 current_zonelist_order = default_zonelist_order();
3550 current_zonelist_order = user_zonelist_order;
3553 static void build_zonelists(pg_data_t *pgdat)
3557 nodemask_t used_mask;
3558 int local_node, prev_node;
3559 struct zonelist *zonelist;
3560 int order = current_zonelist_order;
3562 /* initialize zonelists */
3563 for (i = 0; i < MAX_ZONELISTS; i++) {
3564 zonelist = pgdat->node_zonelists + i;
3565 zonelist->_zonerefs[0].zone = NULL;
3566 zonelist->_zonerefs[0].zone_idx = 0;
3569 /* NUMA-aware ordering of nodes */
3570 local_node = pgdat->node_id;
3571 load = nr_online_nodes;
3572 prev_node = local_node;
3573 nodes_clear(used_mask);
3575 memset(node_order, 0, sizeof(node_order));
3578 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3580 * We don't want to pressure a particular node.
3581 * So adding penalty to the first node in same
3582 * distance group to make it round-robin.
3584 if (node_distance(local_node, node) !=
3585 node_distance(local_node, prev_node))
3586 node_load[node] = load;
3590 if (order == ZONELIST_ORDER_NODE)
3591 build_zonelists_in_node_order(pgdat, node);
3593 node_order[j++] = node; /* remember order */
3596 if (order == ZONELIST_ORDER_ZONE) {
3597 /* calculate node order -- i.e., DMA last! */
3598 build_zonelists_in_zone_order(pgdat, j);
3601 build_thisnode_zonelists(pgdat);
3604 /* Construct the zonelist performance cache - see further mmzone.h */
3605 static void build_zonelist_cache(pg_data_t *pgdat)
3607 struct zonelist *zonelist;
3608 struct zonelist_cache *zlc;
3611 zonelist = &pgdat->node_zonelists[0];
3612 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3613 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3614 for (z = zonelist->_zonerefs; z->zone; z++)
3615 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3618 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3620 * Return node id of node used for "local" allocations.
3621 * I.e., first node id of first zone in arg node's generic zonelist.
3622 * Used for initializing percpu 'numa_mem', which is used primarily
3623 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3625 int local_memory_node(int node)
3629 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3630 gfp_zone(GFP_KERNEL),
3637 #else /* CONFIG_NUMA */
3639 static void set_zonelist_order(void)
3641 current_zonelist_order = ZONELIST_ORDER_ZONE;
3644 static void build_zonelists(pg_data_t *pgdat)
3646 int node, local_node;
3648 struct zonelist *zonelist;
3650 local_node = pgdat->node_id;
3652 zonelist = &pgdat->node_zonelists[0];
3653 j = build_zonelists_node(pgdat, zonelist, 0);
3656 * Now we build the zonelist so that it contains the zones
3657 * of all the other nodes.
3658 * We don't want to pressure a particular node, so when
3659 * building the zones for node N, we make sure that the
3660 * zones coming right after the local ones are those from
3661 * node N+1 (modulo N)
3663 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3664 if (!node_online(node))
3666 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3668 for (node = 0; node < local_node; node++) {
3669 if (!node_online(node))
3671 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3674 zonelist->_zonerefs[j].zone = NULL;
3675 zonelist->_zonerefs[j].zone_idx = 0;
3678 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3679 static void build_zonelist_cache(pg_data_t *pgdat)
3681 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3684 #endif /* CONFIG_NUMA */
3687 * Boot pageset table. One per cpu which is going to be used for all
3688 * zones and all nodes. The parameters will be set in such a way
3689 * that an item put on a list will immediately be handed over to
3690 * the buddy list. This is safe since pageset manipulation is done
3691 * with interrupts disabled.
3693 * The boot_pagesets must be kept even after bootup is complete for
3694 * unused processors and/or zones. They do play a role for bootstrapping
3695 * hotplugged processors.
3697 * zoneinfo_show() and maybe other functions do
3698 * not check if the processor is online before following the pageset pointer.
3699 * Other parts of the kernel may not check if the zone is available.
3701 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3702 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3703 static void setup_zone_pageset(struct zone *zone);
3706 * Global mutex to protect against size modification of zonelists
3707 * as well as to serialize pageset setup for the new populated zone.
3709 DEFINE_MUTEX(zonelists_mutex);
3711 /* return values int ....just for stop_machine() */
3712 static int __build_all_zonelists(void *data)
3716 pg_data_t *self = data;
3719 memset(node_load, 0, sizeof(node_load));
3722 if (self && !node_online(self->node_id)) {
3723 build_zonelists(self);
3724 build_zonelist_cache(self);
3727 for_each_online_node(nid) {
3728 pg_data_t *pgdat = NODE_DATA(nid);
3730 build_zonelists(pgdat);
3731 build_zonelist_cache(pgdat);
3735 * Initialize the boot_pagesets that are going to be used
3736 * for bootstrapping processors. The real pagesets for
3737 * each zone will be allocated later when the per cpu
3738 * allocator is available.
3740 * boot_pagesets are used also for bootstrapping offline
3741 * cpus if the system is already booted because the pagesets
3742 * are needed to initialize allocators on a specific cpu too.
3743 * F.e. the percpu allocator needs the page allocator which
3744 * needs the percpu allocator in order to allocate its pagesets
3745 * (a chicken-egg dilemma).
3747 for_each_possible_cpu(cpu) {
3748 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3750 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3752 * We now know the "local memory node" for each node--
3753 * i.e., the node of the first zone in the generic zonelist.
3754 * Set up numa_mem percpu variable for on-line cpus. During
3755 * boot, only the boot cpu should be on-line; we'll init the
3756 * secondary cpus' numa_mem as they come on-line. During
3757 * node/memory hotplug, we'll fixup all on-line cpus.
3759 if (cpu_online(cpu))
3760 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3768 * Called with zonelists_mutex held always
3769 * unless system_state == SYSTEM_BOOTING.
3771 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3773 set_zonelist_order();
3775 if (system_state == SYSTEM_BOOTING) {
3776 __build_all_zonelists(NULL);
3777 mminit_verify_zonelist();
3778 cpuset_init_current_mems_allowed();
3780 #ifdef CONFIG_MEMORY_HOTPLUG
3782 setup_zone_pageset(zone);
3784 /* we have to stop all cpus to guarantee there is no user
3786 stop_machine(__build_all_zonelists, pgdat, NULL);
3787 /* cpuset refresh routine should be here */
3789 vm_total_pages = nr_free_pagecache_pages();
3791 * Disable grouping by mobility if the number of pages in the
3792 * system is too low to allow the mechanism to work. It would be
3793 * more accurate, but expensive to check per-zone. This check is
3794 * made on memory-hotadd so a system can start with mobility
3795 * disabled and enable it later
3797 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3798 page_group_by_mobility_disabled = 1;
3800 page_group_by_mobility_disabled = 0;
3802 printk("Built %i zonelists in %s order, mobility grouping %s. "
3803 "Total pages: %ld\n",
3805 zonelist_order_name[current_zonelist_order],
3806 page_group_by_mobility_disabled ? "off" : "on",
3809 printk("Policy zone: %s\n", zone_names[policy_zone]);
3814 * Helper functions to size the waitqueue hash table.
3815 * Essentially these want to choose hash table sizes sufficiently
3816 * large so that collisions trying to wait on pages are rare.
3817 * But in fact, the number of active page waitqueues on typical
3818 * systems is ridiculously low, less than 200. So this is even
3819 * conservative, even though it seems large.
3821 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3822 * waitqueues, i.e. the size of the waitq table given the number of pages.
3824 #define PAGES_PER_WAITQUEUE 256
3826 #ifndef CONFIG_MEMORY_HOTPLUG
3827 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3829 unsigned long size = 1;
3831 pages /= PAGES_PER_WAITQUEUE;
3833 while (size < pages)
3837 * Once we have dozens or even hundreds of threads sleeping
3838 * on IO we've got bigger problems than wait queue collision.
3839 * Limit the size of the wait table to a reasonable size.
3841 size = min(size, 4096UL);
3843 return max(size, 4UL);
3847 * A zone's size might be changed by hot-add, so it is not possible to determine
3848 * a suitable size for its wait_table. So we use the maximum size now.
3850 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3852 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3853 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3854 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3856 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3857 * or more by the traditional way. (See above). It equals:
3859 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3860 * ia64(16K page size) : = ( 8G + 4M)byte.
3861 * powerpc (64K page size) : = (32G +16M)byte.
3863 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3870 * This is an integer logarithm so that shifts can be used later
3871 * to extract the more random high bits from the multiplicative
3872 * hash function before the remainder is taken.
3874 static inline unsigned long wait_table_bits(unsigned long size)
3880 * Check if a pageblock contains reserved pages
3882 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3886 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3887 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3894 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3895 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3896 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3897 * higher will lead to a bigger reserve which will get freed as contiguous
3898 * blocks as reclaim kicks in
3900 static void setup_zone_migrate_reserve(struct zone *zone)
3902 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3904 unsigned long block_migratetype;
3908 * Get the start pfn, end pfn and the number of blocks to reserve
3909 * We have to be careful to be aligned to pageblock_nr_pages to
3910 * make sure that we always check pfn_valid for the first page in
3913 start_pfn = zone->zone_start_pfn;
3914 end_pfn = zone_end_pfn(zone);
3915 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3916 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3920 * Reserve blocks are generally in place to help high-order atomic
3921 * allocations that are short-lived. A min_free_kbytes value that
3922 * would result in more than 2 reserve blocks for atomic allocations
3923 * is assumed to be in place to help anti-fragmentation for the
3924 * future allocation of hugepages at runtime.
3926 reserve = min(2, reserve);
3928 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3929 if (!pfn_valid(pfn))
3931 page = pfn_to_page(pfn);
3933 /* Watch out for overlapping nodes */
3934 if (page_to_nid(page) != zone_to_nid(zone))
3937 block_migratetype = get_pageblock_migratetype(page);
3939 /* Only test what is necessary when the reserves are not met */
3942 * Blocks with reserved pages will never free, skip
3945 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3946 if (pageblock_is_reserved(pfn, block_end_pfn))
3949 /* If this block is reserved, account for it */
3950 if (block_migratetype == MIGRATE_RESERVE) {
3955 /* Suitable for reserving if this block is movable */
3956 if (block_migratetype == MIGRATE_MOVABLE) {
3957 set_pageblock_migratetype(page,
3959 move_freepages_block(zone, page,
3967 * If the reserve is met and this is a previous reserved block,
3970 if (block_migratetype == MIGRATE_RESERVE) {
3971 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3972 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3978 * Initially all pages are reserved - free ones are freed
3979 * up by free_all_bootmem() once the early boot process is
3980 * done. Non-atomic initialization, single-pass.
3982 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3983 unsigned long start_pfn, enum memmap_context context)
3986 unsigned long end_pfn = start_pfn + size;
3990 if (highest_memmap_pfn < end_pfn - 1)
3991 highest_memmap_pfn = end_pfn - 1;
3993 z = &NODE_DATA(nid)->node_zones[zone];
3994 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3996 * There can be holes in boot-time mem_map[]s
3997 * handed to this function. They do not
3998 * exist on hotplugged memory.
4000 if (context == MEMMAP_EARLY) {
4001 if (!early_pfn_valid(pfn))
4003 if (!early_pfn_in_nid(pfn, nid))
4006 page = pfn_to_page(pfn);
4007 set_page_links(page, zone, nid, pfn);
4008 mminit_verify_page_links(page, zone, nid, pfn);
4009 init_page_count(page);
4010 page_mapcount_reset(page);
4011 page_cpupid_reset_last(page);
4012 SetPageReserved(page);
4014 * Mark the block movable so that blocks are reserved for
4015 * movable at startup. This will force kernel allocations
4016 * to reserve their blocks rather than leaking throughout
4017 * the address space during boot when many long-lived
4018 * kernel allocations are made. Later some blocks near
4019 * the start are marked MIGRATE_RESERVE by
4020 * setup_zone_migrate_reserve()
4022 * bitmap is created for zone's valid pfn range. but memmap
4023 * can be created for invalid pages (for alignment)
4024 * check here not to call set_pageblock_migratetype() against
4027 if ((z->zone_start_pfn <= pfn)
4028 && (pfn < zone_end_pfn(z))
4029 && !(pfn & (pageblock_nr_pages - 1)))
4030 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4032 INIT_LIST_HEAD(&page->lru);
4033 #ifdef WANT_PAGE_VIRTUAL
4034 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4035 if (!is_highmem_idx(zone))
4036 set_page_address(page, __va(pfn << PAGE_SHIFT));
4041 static void __meminit zone_init_free_lists(struct zone *zone)
4044 for_each_migratetype_order(order, t) {
4045 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4046 zone->free_area[order].nr_free = 0;
4050 #ifndef __HAVE_ARCH_MEMMAP_INIT
4051 #define memmap_init(size, nid, zone, start_pfn) \
4052 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4055 static int __meminit zone_batchsize(struct zone *zone)
4061 * The per-cpu-pages pools are set to around 1000th of the
4062 * size of the zone. But no more than 1/2 of a meg.
4064 * OK, so we don't know how big the cache is. So guess.
4066 batch = zone->managed_pages / 1024;
4067 if (batch * PAGE_SIZE > 512 * 1024)
4068 batch = (512 * 1024) / PAGE_SIZE;
4069 batch /= 4; /* We effectively *= 4 below */
4074 * Clamp the batch to a 2^n - 1 value. Having a power
4075 * of 2 value was found to be more likely to have
4076 * suboptimal cache aliasing properties in some cases.
4078 * For example if 2 tasks are alternately allocating
4079 * batches of pages, one task can end up with a lot
4080 * of pages of one half of the possible page colors
4081 * and the other with pages of the other colors.
4083 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4088 /* The deferral and batching of frees should be suppressed under NOMMU
4091 * The problem is that NOMMU needs to be able to allocate large chunks
4092 * of contiguous memory as there's no hardware page translation to
4093 * assemble apparent contiguous memory from discontiguous pages.
4095 * Queueing large contiguous runs of pages for batching, however,
4096 * causes the pages to actually be freed in smaller chunks. As there
4097 * can be a significant delay between the individual batches being
4098 * recycled, this leads to the once large chunks of space being
4099 * fragmented and becoming unavailable for high-order allocations.
4106 * pcp->high and pcp->batch values are related and dependent on one another:
4107 * ->batch must never be higher then ->high.
4108 * The following function updates them in a safe manner without read side
4111 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4112 * those fields changing asynchronously (acording the the above rule).
4114 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4115 * outside of boot time (or some other assurance that no concurrent updaters
4118 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4119 unsigned long batch)
4121 /* start with a fail safe value for batch */
4125 /* Update high, then batch, in order */
4132 /* a companion to pageset_set_high() */
4133 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4135 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4138 static void pageset_init(struct per_cpu_pageset *p)
4140 struct per_cpu_pages *pcp;
4143 memset(p, 0, sizeof(*p));
4147 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4148 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4151 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4154 pageset_set_batch(p, batch);
4158 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4159 * to the value high for the pageset p.
4161 static void pageset_set_high(struct per_cpu_pageset *p,
4164 unsigned long batch = max(1UL, high / 4);
4165 if ((high / 4) > (PAGE_SHIFT * 8))
4166 batch = PAGE_SHIFT * 8;
4168 pageset_update(&p->pcp, high, batch);
4171 static void __meminit pageset_set_high_and_batch(struct zone *zone,
4172 struct per_cpu_pageset *pcp)
4174 if (percpu_pagelist_fraction)
4175 pageset_set_high(pcp,
4176 (zone->managed_pages /
4177 percpu_pagelist_fraction));
4179 pageset_set_batch(pcp, zone_batchsize(zone));
4182 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4184 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4187 pageset_set_high_and_batch(zone, pcp);
4190 static void __meminit setup_zone_pageset(struct zone *zone)
4193 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4194 for_each_possible_cpu(cpu)
4195 zone_pageset_init(zone, cpu);
4199 * Allocate per cpu pagesets and initialize them.
4200 * Before this call only boot pagesets were available.
4202 void __init setup_per_cpu_pageset(void)
4206 for_each_populated_zone(zone)
4207 setup_zone_pageset(zone);
4210 static noinline __init_refok
4211 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4214 struct pglist_data *pgdat = zone->zone_pgdat;
4218 * The per-page waitqueue mechanism uses hashed waitqueues
4221 zone->wait_table_hash_nr_entries =
4222 wait_table_hash_nr_entries(zone_size_pages);
4223 zone->wait_table_bits =
4224 wait_table_bits(zone->wait_table_hash_nr_entries);
4225 alloc_size = zone->wait_table_hash_nr_entries
4226 * sizeof(wait_queue_head_t);
4228 if (!slab_is_available()) {
4229 zone->wait_table = (wait_queue_head_t *)
4230 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4233 * This case means that a zone whose size was 0 gets new memory
4234 * via memory hot-add.
4235 * But it may be the case that a new node was hot-added. In
4236 * this case vmalloc() will not be able to use this new node's
4237 * memory - this wait_table must be initialized to use this new
4238 * node itself as well.
4239 * To use this new node's memory, further consideration will be
4242 zone->wait_table = vmalloc(alloc_size);
4244 if (!zone->wait_table)
4247 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4248 init_waitqueue_head(zone->wait_table + i);
4253 static __meminit void zone_pcp_init(struct zone *zone)
4256 * per cpu subsystem is not up at this point. The following code
4257 * relies on the ability of the linker to provide the
4258 * offset of a (static) per cpu variable into the per cpu area.
4260 zone->pageset = &boot_pageset;
4262 if (populated_zone(zone))
4263 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4264 zone->name, zone->present_pages,
4265 zone_batchsize(zone));
4268 int __meminit init_currently_empty_zone(struct zone *zone,
4269 unsigned long zone_start_pfn,
4271 enum memmap_context context)
4273 struct pglist_data *pgdat = zone->zone_pgdat;
4275 ret = zone_wait_table_init(zone, size);
4278 pgdat->nr_zones = zone_idx(zone) + 1;
4280 zone->zone_start_pfn = zone_start_pfn;
4282 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4283 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4285 (unsigned long)zone_idx(zone),
4286 zone_start_pfn, (zone_start_pfn + size));
4288 zone_init_free_lists(zone);
4293 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4294 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4296 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4297 * Architectures may implement their own version but if add_active_range()
4298 * was used and there are no special requirements, this is a convenient
4301 int __meminit __early_pfn_to_nid(unsigned long pfn)
4303 unsigned long start_pfn, end_pfn;
4306 * NOTE: The following SMP-unsafe globals are only used early in boot
4307 * when the kernel is running single-threaded.
4309 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4310 static int __meminitdata last_nid;
4312 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4315 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4317 last_start_pfn = start_pfn;
4318 last_end_pfn = end_pfn;
4324 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4326 int __meminit early_pfn_to_nid(unsigned long pfn)
4330 nid = __early_pfn_to_nid(pfn);
4333 /* just returns 0 */
4337 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4338 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4342 nid = __early_pfn_to_nid(pfn);
4343 if (nid >= 0 && nid != node)
4350 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4351 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4352 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4354 * If an architecture guarantees that all ranges registered with
4355 * add_active_ranges() contain no holes and may be freed, this
4356 * this function may be used instead of calling free_bootmem() manually.
4358 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4360 unsigned long start_pfn, end_pfn;
4363 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4364 start_pfn = min(start_pfn, max_low_pfn);
4365 end_pfn = min(end_pfn, max_low_pfn);
4367 if (start_pfn < end_pfn)
4368 free_bootmem_node(NODE_DATA(this_nid),
4369 PFN_PHYS(start_pfn),
4370 (end_pfn - start_pfn) << PAGE_SHIFT);
4375 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4376 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4378 * If an architecture guarantees that all ranges registered with
4379 * add_active_ranges() contain no holes and may be freed, this
4380 * function may be used instead of calling memory_present() manually.
4382 void __init sparse_memory_present_with_active_regions(int nid)
4384 unsigned long start_pfn, end_pfn;
4387 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4388 memory_present(this_nid, start_pfn, end_pfn);
4392 * get_pfn_range_for_nid - Return the start and end page frames for a node
4393 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4394 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4395 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4397 * It returns the start and end page frame of a node based on information
4398 * provided by an arch calling add_active_range(). If called for a node
4399 * with no available memory, a warning is printed and the start and end
4402 void __meminit get_pfn_range_for_nid(unsigned int nid,
4403 unsigned long *start_pfn, unsigned long *end_pfn)
4405 unsigned long this_start_pfn, this_end_pfn;
4411 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4412 *start_pfn = min(*start_pfn, this_start_pfn);
4413 *end_pfn = max(*end_pfn, this_end_pfn);
4416 if (*start_pfn == -1UL)
4421 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4422 * assumption is made that zones within a node are ordered in monotonic
4423 * increasing memory addresses so that the "highest" populated zone is used
4425 static void __init find_usable_zone_for_movable(void)
4428 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4429 if (zone_index == ZONE_MOVABLE)
4432 if (arch_zone_highest_possible_pfn[zone_index] >
4433 arch_zone_lowest_possible_pfn[zone_index])
4437 VM_BUG_ON(zone_index == -1);
4438 movable_zone = zone_index;
4442 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4443 * because it is sized independent of architecture. Unlike the other zones,
4444 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4445 * in each node depending on the size of each node and how evenly kernelcore
4446 * is distributed. This helper function adjusts the zone ranges
4447 * provided by the architecture for a given node by using the end of the
4448 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4449 * zones within a node are in order of monotonic increases memory addresses
4451 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4452 unsigned long zone_type,
4453 unsigned long node_start_pfn,
4454 unsigned long node_end_pfn,
4455 unsigned long *zone_start_pfn,
4456 unsigned long *zone_end_pfn)
4458 /* Only adjust if ZONE_MOVABLE is on this node */
4459 if (zone_movable_pfn[nid]) {
4460 /* Size ZONE_MOVABLE */
4461 if (zone_type == ZONE_MOVABLE) {
4462 *zone_start_pfn = zone_movable_pfn[nid];
4463 *zone_end_pfn = min(node_end_pfn,
4464 arch_zone_highest_possible_pfn[movable_zone]);
4466 /* Adjust for ZONE_MOVABLE starting within this range */
4467 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4468 *zone_end_pfn > zone_movable_pfn[nid]) {
4469 *zone_end_pfn = zone_movable_pfn[nid];
4471 /* Check if this whole range is within ZONE_MOVABLE */
4472 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4473 *zone_start_pfn = *zone_end_pfn;
4478 * Return the number of pages a zone spans in a node, including holes
4479 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4481 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4482 unsigned long zone_type,
4483 unsigned long node_start_pfn,
4484 unsigned long node_end_pfn,
4485 unsigned long *ignored)
4487 unsigned long zone_start_pfn, zone_end_pfn;
4489 /* Get the start and end of the zone */
4490 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4491 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4492 adjust_zone_range_for_zone_movable(nid, zone_type,
4493 node_start_pfn, node_end_pfn,
4494 &zone_start_pfn, &zone_end_pfn);
4496 /* Check that this node has pages within the zone's required range */
4497 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4500 /* Move the zone boundaries inside the node if necessary */
4501 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4502 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4504 /* Return the spanned pages */
4505 return zone_end_pfn - zone_start_pfn;
4509 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4510 * then all holes in the requested range will be accounted for.
4512 unsigned long __meminit __absent_pages_in_range(int nid,
4513 unsigned long range_start_pfn,
4514 unsigned long range_end_pfn)
4516 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4517 unsigned long start_pfn, end_pfn;
4520 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4521 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4522 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4523 nr_absent -= end_pfn - start_pfn;
4529 * absent_pages_in_range - Return number of page frames in holes within a range
4530 * @start_pfn: The start PFN to start searching for holes
4531 * @end_pfn: The end PFN to stop searching for holes
4533 * It returns the number of pages frames in memory holes within a range.
4535 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4536 unsigned long end_pfn)
4538 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4541 /* Return the number of page frames in holes in a zone on a node */
4542 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4543 unsigned long zone_type,
4544 unsigned long node_start_pfn,
4545 unsigned long node_end_pfn,
4546 unsigned long *ignored)
4548 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4549 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4550 unsigned long zone_start_pfn, zone_end_pfn;
4552 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4553 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4555 adjust_zone_range_for_zone_movable(nid, zone_type,
4556 node_start_pfn, node_end_pfn,
4557 &zone_start_pfn, &zone_end_pfn);
4558 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4561 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4562 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4563 unsigned long zone_type,
4564 unsigned long node_start_pfn,
4565 unsigned long node_end_pfn,
4566 unsigned long *zones_size)
4568 return zones_size[zone_type];
4571 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4572 unsigned long zone_type,
4573 unsigned long node_start_pfn,
4574 unsigned long node_end_pfn,
4575 unsigned long *zholes_size)
4580 return zholes_size[zone_type];
4583 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4585 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4586 unsigned long node_start_pfn,
4587 unsigned long node_end_pfn,
4588 unsigned long *zones_size,
4589 unsigned long *zholes_size)
4591 unsigned long realtotalpages, totalpages = 0;
4594 for (i = 0; i < MAX_NR_ZONES; i++)
4595 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4599 pgdat->node_spanned_pages = totalpages;
4601 realtotalpages = totalpages;
4602 for (i = 0; i < MAX_NR_ZONES; i++)
4604 zone_absent_pages_in_node(pgdat->node_id, i,
4605 node_start_pfn, node_end_pfn,
4607 pgdat->node_present_pages = realtotalpages;
4608 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4612 #ifndef CONFIG_SPARSEMEM
4614 * Calculate the size of the zone->blockflags rounded to an unsigned long
4615 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4616 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4617 * round what is now in bits to nearest long in bits, then return it in
4620 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4622 unsigned long usemapsize;
4624 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4625 usemapsize = roundup(zonesize, pageblock_nr_pages);
4626 usemapsize = usemapsize >> pageblock_order;
4627 usemapsize *= NR_PAGEBLOCK_BITS;
4628 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4630 return usemapsize / 8;
4633 static void __init setup_usemap(struct pglist_data *pgdat,
4635 unsigned long zone_start_pfn,
4636 unsigned long zonesize)
4638 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4639 zone->pageblock_flags = NULL;
4641 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4645 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4646 unsigned long zone_start_pfn, unsigned long zonesize) {}
4647 #endif /* CONFIG_SPARSEMEM */
4649 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4651 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4652 void __paginginit set_pageblock_order(void)
4656 /* Check that pageblock_nr_pages has not already been setup */
4657 if (pageblock_order)
4660 if (HPAGE_SHIFT > PAGE_SHIFT)
4661 order = HUGETLB_PAGE_ORDER;
4663 order = MAX_ORDER - 1;
4666 * Assume the largest contiguous order of interest is a huge page.
4667 * This value may be variable depending on boot parameters on IA64 and
4670 pageblock_order = order;
4672 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4675 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4676 * is unused as pageblock_order is set at compile-time. See
4677 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4680 void __paginginit set_pageblock_order(void)
4684 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4686 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4687 unsigned long present_pages)
4689 unsigned long pages = spanned_pages;
4692 * Provide a more accurate estimation if there are holes within
4693 * the zone and SPARSEMEM is in use. If there are holes within the
4694 * zone, each populated memory region may cost us one or two extra
4695 * memmap pages due to alignment because memmap pages for each
4696 * populated regions may not naturally algined on page boundary.
4697 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4699 if (spanned_pages > present_pages + (present_pages >> 4) &&
4700 IS_ENABLED(CONFIG_SPARSEMEM))
4701 pages = present_pages;
4703 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4707 * Set up the zone data structures:
4708 * - mark all pages reserved
4709 * - mark all memory queues empty
4710 * - clear the memory bitmaps
4712 * NOTE: pgdat should get zeroed by caller.
4714 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4715 unsigned long node_start_pfn, unsigned long node_end_pfn,
4716 unsigned long *zones_size, unsigned long *zholes_size)
4719 int nid = pgdat->node_id;
4720 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4723 pgdat_resize_init(pgdat);
4724 #ifdef CONFIG_NUMA_BALANCING
4725 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4726 pgdat->numabalancing_migrate_nr_pages = 0;
4727 pgdat->numabalancing_migrate_next_window = jiffies;
4729 init_waitqueue_head(&pgdat->kswapd_wait);
4730 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4731 pgdat_page_cgroup_init(pgdat);
4733 for (j = 0; j < MAX_NR_ZONES; j++) {
4734 struct zone *zone = pgdat->node_zones + j;
4735 unsigned long size, realsize, freesize, memmap_pages;
4737 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4738 node_end_pfn, zones_size);
4739 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4745 * Adjust freesize so that it accounts for how much memory
4746 * is used by this zone for memmap. This affects the watermark
4747 * and per-cpu initialisations
4749 memmap_pages = calc_memmap_size(size, realsize);
4750 if (freesize >= memmap_pages) {
4751 freesize -= memmap_pages;
4754 " %s zone: %lu pages used for memmap\n",
4755 zone_names[j], memmap_pages);
4758 " %s zone: %lu pages exceeds freesize %lu\n",
4759 zone_names[j], memmap_pages, freesize);
4761 /* Account for reserved pages */
4762 if (j == 0 && freesize > dma_reserve) {
4763 freesize -= dma_reserve;
4764 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4765 zone_names[0], dma_reserve);
4768 if (!is_highmem_idx(j))
4769 nr_kernel_pages += freesize;
4770 /* Charge for highmem memmap if there are enough kernel pages */
4771 else if (nr_kernel_pages > memmap_pages * 2)
4772 nr_kernel_pages -= memmap_pages;
4773 nr_all_pages += freesize;
4775 zone->spanned_pages = size;
4776 zone->present_pages = realsize;
4778 * Set an approximate value for lowmem here, it will be adjusted
4779 * when the bootmem allocator frees pages into the buddy system.
4780 * And all highmem pages will be managed by the buddy system.
4782 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4785 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4787 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4789 zone->name = zone_names[j];
4790 spin_lock_init(&zone->lock);
4791 spin_lock_init(&zone->lru_lock);
4792 zone_seqlock_init(zone);
4793 zone->zone_pgdat = pgdat;
4794 zone_pcp_init(zone);
4796 /* For bootup, initialized properly in watermark setup */
4797 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4799 lruvec_init(&zone->lruvec);
4803 set_pageblock_order();
4804 setup_usemap(pgdat, zone, zone_start_pfn, size);
4805 ret = init_currently_empty_zone(zone, zone_start_pfn,
4806 size, MEMMAP_EARLY);
4808 memmap_init(size, nid, j, zone_start_pfn);
4809 zone_start_pfn += size;
4813 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4815 /* Skip empty nodes */
4816 if (!pgdat->node_spanned_pages)
4819 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4820 /* ia64 gets its own node_mem_map, before this, without bootmem */
4821 if (!pgdat->node_mem_map) {
4822 unsigned long size, start, end;
4826 * The zone's endpoints aren't required to be MAX_ORDER
4827 * aligned but the node_mem_map endpoints must be in order
4828 * for the buddy allocator to function correctly.
4830 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4831 end = pgdat_end_pfn(pgdat);
4832 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4833 size = (end - start) * sizeof(struct page);
4834 map = alloc_remap(pgdat->node_id, size);
4836 map = alloc_bootmem_node_nopanic(pgdat, size);
4837 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4839 #ifndef CONFIG_NEED_MULTIPLE_NODES
4841 * With no DISCONTIG, the global mem_map is just set as node 0's
4843 if (pgdat == NODE_DATA(0)) {
4844 mem_map = NODE_DATA(0)->node_mem_map;
4845 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4846 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4847 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4848 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4851 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4854 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4855 unsigned long node_start_pfn, unsigned long *zholes_size)
4857 pg_data_t *pgdat = NODE_DATA(nid);
4858 unsigned long start_pfn = 0;
4859 unsigned long end_pfn = 0;
4861 /* pg_data_t should be reset to zero when it's allocated */
4862 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4864 pgdat->node_id = nid;
4865 pgdat->node_start_pfn = node_start_pfn;
4866 init_zone_allows_reclaim(nid);
4867 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4868 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4870 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4871 zones_size, zholes_size);
4873 alloc_node_mem_map(pgdat);
4874 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4875 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4876 nid, (unsigned long)pgdat,
4877 (unsigned long)pgdat->node_mem_map);
4880 free_area_init_core(pgdat, start_pfn, end_pfn,
4881 zones_size, zholes_size);
4884 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4886 #if MAX_NUMNODES > 1
4888 * Figure out the number of possible node ids.
4890 void __init setup_nr_node_ids(void)
4893 unsigned int highest = 0;
4895 for_each_node_mask(node, node_possible_map)
4897 nr_node_ids = highest + 1;
4902 * node_map_pfn_alignment - determine the maximum internode alignment
4904 * This function should be called after node map is populated and sorted.
4905 * It calculates the maximum power of two alignment which can distinguish
4908 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4909 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4910 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4911 * shifted, 1GiB is enough and this function will indicate so.
4913 * This is used to test whether pfn -> nid mapping of the chosen memory
4914 * model has fine enough granularity to avoid incorrect mapping for the
4915 * populated node map.
4917 * Returns the determined alignment in pfn's. 0 if there is no alignment
4918 * requirement (single node).
4920 unsigned long __init node_map_pfn_alignment(void)
4922 unsigned long accl_mask = 0, last_end = 0;
4923 unsigned long start, end, mask;
4927 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4928 if (!start || last_nid < 0 || last_nid == nid) {
4935 * Start with a mask granular enough to pin-point to the
4936 * start pfn and tick off bits one-by-one until it becomes
4937 * too coarse to separate the current node from the last.
4939 mask = ~((1 << __ffs(start)) - 1);
4940 while (mask && last_end <= (start & (mask << 1)))
4943 /* accumulate all internode masks */
4947 /* convert mask to number of pages */
4948 return ~accl_mask + 1;
4951 /* Find the lowest pfn for a node */
4952 static unsigned long __init find_min_pfn_for_node(int nid)
4954 unsigned long min_pfn = ULONG_MAX;
4955 unsigned long start_pfn;
4958 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4959 min_pfn = min(min_pfn, start_pfn);
4961 if (min_pfn == ULONG_MAX) {
4963 "Could not find start_pfn for node %d\n", nid);
4971 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4973 * It returns the minimum PFN based on information provided via
4974 * add_active_range().
4976 unsigned long __init find_min_pfn_with_active_regions(void)
4978 return find_min_pfn_for_node(MAX_NUMNODES);
4982 * early_calculate_totalpages()
4983 * Sum pages in active regions for movable zone.
4984 * Populate N_MEMORY for calculating usable_nodes.
4986 static unsigned long __init early_calculate_totalpages(void)
4988 unsigned long totalpages = 0;
4989 unsigned long start_pfn, end_pfn;
4992 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4993 unsigned long pages = end_pfn - start_pfn;
4995 totalpages += pages;
4997 node_set_state(nid, N_MEMORY);
5003 * Find the PFN the Movable zone begins in each node. Kernel memory
5004 * is spread evenly between nodes as long as the nodes have enough
5005 * memory. When they don't, some nodes will have more kernelcore than
5008 static void __init find_zone_movable_pfns_for_nodes(void)
5011 unsigned long usable_startpfn;
5012 unsigned long kernelcore_node, kernelcore_remaining;
5013 /* save the state before borrow the nodemask */
5014 nodemask_t saved_node_state = node_states[N_MEMORY];
5015 unsigned long totalpages = early_calculate_totalpages();
5016 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5019 * If movablecore was specified, calculate what size of
5020 * kernelcore that corresponds so that memory usable for
5021 * any allocation type is evenly spread. If both kernelcore
5022 * and movablecore are specified, then the value of kernelcore
5023 * will be used for required_kernelcore if it's greater than
5024 * what movablecore would have allowed.
5026 if (required_movablecore) {
5027 unsigned long corepages;
5030 * Round-up so that ZONE_MOVABLE is at least as large as what
5031 * was requested by the user
5033 required_movablecore =
5034 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5035 corepages = totalpages - required_movablecore;
5037 required_kernelcore = max(required_kernelcore, corepages);
5040 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5041 if (!required_kernelcore)
5044 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5045 find_usable_zone_for_movable();
5046 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5049 /* Spread kernelcore memory as evenly as possible throughout nodes */
5050 kernelcore_node = required_kernelcore / usable_nodes;
5051 for_each_node_state(nid, N_MEMORY) {
5052 unsigned long start_pfn, end_pfn;
5055 * Recalculate kernelcore_node if the division per node
5056 * now exceeds what is necessary to satisfy the requested
5057 * amount of memory for the kernel
5059 if (required_kernelcore < kernelcore_node)
5060 kernelcore_node = required_kernelcore / usable_nodes;
5063 * As the map is walked, we track how much memory is usable
5064 * by the kernel using kernelcore_remaining. When it is
5065 * 0, the rest of the node is usable by ZONE_MOVABLE
5067 kernelcore_remaining = kernelcore_node;
5069 /* Go through each range of PFNs within this node */
5070 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5071 unsigned long size_pages;
5073 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5074 if (start_pfn >= end_pfn)
5077 /* Account for what is only usable for kernelcore */
5078 if (start_pfn < usable_startpfn) {
5079 unsigned long kernel_pages;
5080 kernel_pages = min(end_pfn, usable_startpfn)
5083 kernelcore_remaining -= min(kernel_pages,
5084 kernelcore_remaining);
5085 required_kernelcore -= min(kernel_pages,
5086 required_kernelcore);
5088 /* Continue if range is now fully accounted */
5089 if (end_pfn <= usable_startpfn) {
5092 * Push zone_movable_pfn to the end so
5093 * that if we have to rebalance
5094 * kernelcore across nodes, we will
5095 * not double account here
5097 zone_movable_pfn[nid] = end_pfn;
5100 start_pfn = usable_startpfn;
5104 * The usable PFN range for ZONE_MOVABLE is from
5105 * start_pfn->end_pfn. Calculate size_pages as the
5106 * number of pages used as kernelcore
5108 size_pages = end_pfn - start_pfn;
5109 if (size_pages > kernelcore_remaining)
5110 size_pages = kernelcore_remaining;
5111 zone_movable_pfn[nid] = start_pfn + size_pages;
5114 * Some kernelcore has been met, update counts and
5115 * break if the kernelcore for this node has been
5118 required_kernelcore -= min(required_kernelcore,
5120 kernelcore_remaining -= size_pages;
5121 if (!kernelcore_remaining)
5127 * If there is still required_kernelcore, we do another pass with one
5128 * less node in the count. This will push zone_movable_pfn[nid] further
5129 * along on the nodes that still have memory until kernelcore is
5133 if (usable_nodes && required_kernelcore > usable_nodes)
5136 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5137 for (nid = 0; nid < MAX_NUMNODES; nid++)
5138 zone_movable_pfn[nid] =
5139 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5142 /* restore the node_state */
5143 node_states[N_MEMORY] = saved_node_state;
5146 /* Any regular or high memory on that node ? */
5147 static void check_for_memory(pg_data_t *pgdat, int nid)
5149 enum zone_type zone_type;
5151 if (N_MEMORY == N_NORMAL_MEMORY)
5154 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5155 struct zone *zone = &pgdat->node_zones[zone_type];
5156 if (populated_zone(zone)) {
5157 node_set_state(nid, N_HIGH_MEMORY);
5158 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5159 zone_type <= ZONE_NORMAL)
5160 node_set_state(nid, N_NORMAL_MEMORY);
5167 * free_area_init_nodes - Initialise all pg_data_t and zone data
5168 * @max_zone_pfn: an array of max PFNs for each zone
5170 * This will call free_area_init_node() for each active node in the system.
5171 * Using the page ranges provided by add_active_range(), the size of each
5172 * zone in each node and their holes is calculated. If the maximum PFN
5173 * between two adjacent zones match, it is assumed that the zone is empty.
5174 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5175 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5176 * starts where the previous one ended. For example, ZONE_DMA32 starts
5177 * at arch_max_dma_pfn.
5179 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5181 unsigned long start_pfn, end_pfn;
5184 /* Record where the zone boundaries are */
5185 memset(arch_zone_lowest_possible_pfn, 0,
5186 sizeof(arch_zone_lowest_possible_pfn));
5187 memset(arch_zone_highest_possible_pfn, 0,
5188 sizeof(arch_zone_highest_possible_pfn));
5189 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5190 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5191 for (i = 1; i < MAX_NR_ZONES; i++) {
5192 if (i == ZONE_MOVABLE)
5194 arch_zone_lowest_possible_pfn[i] =
5195 arch_zone_highest_possible_pfn[i-1];
5196 arch_zone_highest_possible_pfn[i] =
5197 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5199 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5200 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5202 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5203 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5204 find_zone_movable_pfns_for_nodes();
5206 /* Print out the zone ranges */
5207 printk("Zone ranges:\n");
5208 for (i = 0; i < MAX_NR_ZONES; i++) {
5209 if (i == ZONE_MOVABLE)
5211 printk(KERN_CONT " %-8s ", zone_names[i]);
5212 if (arch_zone_lowest_possible_pfn[i] ==
5213 arch_zone_highest_possible_pfn[i])
5214 printk(KERN_CONT "empty\n");
5216 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5217 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5218 (arch_zone_highest_possible_pfn[i]
5219 << PAGE_SHIFT) - 1);
5222 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5223 printk("Movable zone start for each node\n");
5224 for (i = 0; i < MAX_NUMNODES; i++) {
5225 if (zone_movable_pfn[i])
5226 printk(" Node %d: %#010lx\n", i,
5227 zone_movable_pfn[i] << PAGE_SHIFT);
5230 /* Print out the early node map */
5231 printk("Early memory node ranges\n");
5232 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5233 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5234 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5236 /* Initialise every node */
5237 mminit_verify_pageflags_layout();
5238 setup_nr_node_ids();
5239 for_each_online_node(nid) {
5240 pg_data_t *pgdat = NODE_DATA(nid);
5241 free_area_init_node(nid, NULL,
5242 find_min_pfn_for_node(nid), NULL);
5244 /* Any memory on that node */
5245 if (pgdat->node_present_pages)
5246 node_set_state(nid, N_MEMORY);
5247 check_for_memory(pgdat, nid);
5251 static int __init cmdline_parse_core(char *p, unsigned long *core)
5253 unsigned long long coremem;
5257 coremem = memparse(p, &p);
5258 *core = coremem >> PAGE_SHIFT;
5260 /* Paranoid check that UL is enough for the coremem value */
5261 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5267 * kernelcore=size sets the amount of memory for use for allocations that
5268 * cannot be reclaimed or migrated.
5270 static int __init cmdline_parse_kernelcore(char *p)
5272 return cmdline_parse_core(p, &required_kernelcore);
5276 * movablecore=size sets the amount of memory for use for allocations that
5277 * can be reclaimed or migrated.
5279 static int __init cmdline_parse_movablecore(char *p)
5281 return cmdline_parse_core(p, &required_movablecore);
5284 early_param("kernelcore", cmdline_parse_kernelcore);
5285 early_param("movablecore", cmdline_parse_movablecore);
5287 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5289 void adjust_managed_page_count(struct page *page, long count)
5291 spin_lock(&managed_page_count_lock);
5292 page_zone(page)->managed_pages += count;
5293 totalram_pages += count;
5294 #ifdef CONFIG_HIGHMEM
5295 if (PageHighMem(page))
5296 totalhigh_pages += count;
5298 spin_unlock(&managed_page_count_lock);
5300 EXPORT_SYMBOL(adjust_managed_page_count);
5302 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5305 unsigned long pages = 0;
5307 start = (void *)PAGE_ALIGN((unsigned long)start);
5308 end = (void *)((unsigned long)end & PAGE_MASK);
5309 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5310 if ((unsigned int)poison <= 0xFF)
5311 memset(pos, poison, PAGE_SIZE);
5312 free_reserved_page(virt_to_page(pos));
5316 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5317 s, pages << (PAGE_SHIFT - 10), start, end);
5321 EXPORT_SYMBOL(free_reserved_area);
5323 #ifdef CONFIG_HIGHMEM
5324 void free_highmem_page(struct page *page)
5326 __free_reserved_page(page);
5328 page_zone(page)->managed_pages++;
5334 void __init mem_init_print_info(const char *str)
5336 unsigned long physpages, codesize, datasize, rosize, bss_size;
5337 unsigned long init_code_size, init_data_size;
5339 physpages = get_num_physpages();
5340 codesize = _etext - _stext;
5341 datasize = _edata - _sdata;
5342 rosize = __end_rodata - __start_rodata;
5343 bss_size = __bss_stop - __bss_start;
5344 init_data_size = __init_end - __init_begin;
5345 init_code_size = _einittext - _sinittext;
5348 * Detect special cases and adjust section sizes accordingly:
5349 * 1) .init.* may be embedded into .data sections
5350 * 2) .init.text.* may be out of [__init_begin, __init_end],
5351 * please refer to arch/tile/kernel/vmlinux.lds.S.
5352 * 3) .rodata.* may be embedded into .text or .data sections.
5354 #define adj_init_size(start, end, size, pos, adj) \
5356 if (start <= pos && pos < end && size > adj) \
5360 adj_init_size(__init_begin, __init_end, init_data_size,
5361 _sinittext, init_code_size);
5362 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5363 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5364 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5365 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5367 #undef adj_init_size
5369 printk("Memory: %luK/%luK available "
5370 "(%luK kernel code, %luK rwdata, %luK rodata, "
5371 "%luK init, %luK bss, %luK reserved"
5372 #ifdef CONFIG_HIGHMEM
5376 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5377 codesize >> 10, datasize >> 10, rosize >> 10,
5378 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5379 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5380 #ifdef CONFIG_HIGHMEM
5381 totalhigh_pages << (PAGE_SHIFT-10),
5383 str ? ", " : "", str ? str : "");
5387 * set_dma_reserve - set the specified number of pages reserved in the first zone
5388 * @new_dma_reserve: The number of pages to mark reserved
5390 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5391 * In the DMA zone, a significant percentage may be consumed by kernel image
5392 * and other unfreeable allocations which can skew the watermarks badly. This
5393 * function may optionally be used to account for unfreeable pages in the
5394 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5395 * smaller per-cpu batchsize.
5397 void __init set_dma_reserve(unsigned long new_dma_reserve)
5399 dma_reserve = new_dma_reserve;
5402 void __init free_area_init(unsigned long *zones_size)
5404 free_area_init_node(0, zones_size,
5405 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5408 static int page_alloc_cpu_notify(struct notifier_block *self,
5409 unsigned long action, void *hcpu)
5411 int cpu = (unsigned long)hcpu;
5413 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5414 lru_add_drain_cpu(cpu);
5418 * Spill the event counters of the dead processor
5419 * into the current processors event counters.
5420 * This artificially elevates the count of the current
5423 vm_events_fold_cpu(cpu);
5426 * Zero the differential counters of the dead processor
5427 * so that the vm statistics are consistent.
5429 * This is only okay since the processor is dead and cannot
5430 * race with what we are doing.
5432 cpu_vm_stats_fold(cpu);
5437 void __init page_alloc_init(void)
5439 hotcpu_notifier(page_alloc_cpu_notify, 0);
5443 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5444 * or min_free_kbytes changes.
5446 static void calculate_totalreserve_pages(void)
5448 struct pglist_data *pgdat;
5449 unsigned long reserve_pages = 0;
5450 enum zone_type i, j;
5452 for_each_online_pgdat(pgdat) {
5453 for (i = 0; i < MAX_NR_ZONES; i++) {
5454 struct zone *zone = pgdat->node_zones + i;
5455 unsigned long max = 0;
5457 /* Find valid and maximum lowmem_reserve in the zone */
5458 for (j = i; j < MAX_NR_ZONES; j++) {
5459 if (zone->lowmem_reserve[j] > max)
5460 max = zone->lowmem_reserve[j];
5463 /* we treat the high watermark as reserved pages. */
5464 max += high_wmark_pages(zone);
5466 if (max > zone->managed_pages)
5467 max = zone->managed_pages;
5468 reserve_pages += max;
5470 * Lowmem reserves are not available to
5471 * GFP_HIGHUSER page cache allocations and
5472 * kswapd tries to balance zones to their high
5473 * watermark. As a result, neither should be
5474 * regarded as dirtyable memory, to prevent a
5475 * situation where reclaim has to clean pages
5476 * in order to balance the zones.
5478 zone->dirty_balance_reserve = max;
5481 dirty_balance_reserve = reserve_pages;
5482 totalreserve_pages = reserve_pages;
5486 * setup_per_zone_lowmem_reserve - called whenever
5487 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5488 * has a correct pages reserved value, so an adequate number of
5489 * pages are left in the zone after a successful __alloc_pages().
5491 static void setup_per_zone_lowmem_reserve(void)
5493 struct pglist_data *pgdat;
5494 enum zone_type j, idx;
5496 for_each_online_pgdat(pgdat) {
5497 for (j = 0; j < MAX_NR_ZONES; j++) {
5498 struct zone *zone = pgdat->node_zones + j;
5499 unsigned long managed_pages = zone->managed_pages;
5501 zone->lowmem_reserve[j] = 0;
5505 struct zone *lower_zone;
5509 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5510 sysctl_lowmem_reserve_ratio[idx] = 1;
5512 lower_zone = pgdat->node_zones + idx;
5513 lower_zone->lowmem_reserve[j] = managed_pages /
5514 sysctl_lowmem_reserve_ratio[idx];
5515 managed_pages += lower_zone->managed_pages;
5520 /* update totalreserve_pages */
5521 calculate_totalreserve_pages();
5524 static void __setup_per_zone_wmarks(void)
5526 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5527 unsigned long lowmem_pages = 0;
5529 unsigned long flags;
5531 /* Calculate total number of !ZONE_HIGHMEM pages */
5532 for_each_zone(zone) {
5533 if (!is_highmem(zone))
5534 lowmem_pages += zone->managed_pages;
5537 for_each_zone(zone) {
5540 spin_lock_irqsave(&zone->lock, flags);
5541 tmp = (u64)pages_min * zone->managed_pages;
5542 do_div(tmp, lowmem_pages);
5543 if (is_highmem(zone)) {
5545 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5546 * need highmem pages, so cap pages_min to a small
5549 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5550 * deltas controls asynch page reclaim, and so should
5551 * not be capped for highmem.
5553 unsigned long min_pages;
5555 min_pages = zone->managed_pages / 1024;
5556 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5557 zone->watermark[WMARK_MIN] = min_pages;
5560 * If it's a lowmem zone, reserve a number of pages
5561 * proportionate to the zone's size.
5563 zone->watermark[WMARK_MIN] = tmp;
5566 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5567 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5569 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5570 high_wmark_pages(zone) -
5571 low_wmark_pages(zone) -
5572 zone_page_state(zone, NR_ALLOC_BATCH));
5574 setup_zone_migrate_reserve(zone);
5575 spin_unlock_irqrestore(&zone->lock, flags);
5578 /* update totalreserve_pages */
5579 calculate_totalreserve_pages();
5583 * setup_per_zone_wmarks - called when min_free_kbytes changes
5584 * or when memory is hot-{added|removed}
5586 * Ensures that the watermark[min,low,high] values for each zone are set
5587 * correctly with respect to min_free_kbytes.
5589 void setup_per_zone_wmarks(void)
5591 mutex_lock(&zonelists_mutex);
5592 __setup_per_zone_wmarks();
5593 mutex_unlock(&zonelists_mutex);
5597 * The inactive anon list should be small enough that the VM never has to
5598 * do too much work, but large enough that each inactive page has a chance
5599 * to be referenced again before it is swapped out.
5601 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5602 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5603 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5604 * the anonymous pages are kept on the inactive list.
5607 * memory ratio inactive anon
5608 * -------------------------------------
5617 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5619 unsigned int gb, ratio;
5621 /* Zone size in gigabytes */
5622 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5624 ratio = int_sqrt(10 * gb);
5628 zone->inactive_ratio = ratio;
5631 static void __meminit setup_per_zone_inactive_ratio(void)
5636 calculate_zone_inactive_ratio(zone);
5640 * Initialise min_free_kbytes.
5642 * For small machines we want it small (128k min). For large machines
5643 * we want it large (64MB max). But it is not linear, because network
5644 * bandwidth does not increase linearly with machine size. We use
5646 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5647 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5663 int __meminit init_per_zone_wmark_min(void)
5665 unsigned long lowmem_kbytes;
5666 int new_min_free_kbytes;
5668 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5669 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5671 if (new_min_free_kbytes > user_min_free_kbytes) {
5672 min_free_kbytes = new_min_free_kbytes;
5673 if (min_free_kbytes < 128)
5674 min_free_kbytes = 128;
5675 if (min_free_kbytes > 65536)
5676 min_free_kbytes = 65536;
5678 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5679 new_min_free_kbytes, user_min_free_kbytes);
5681 setup_per_zone_wmarks();
5682 refresh_zone_stat_thresholds();
5683 setup_per_zone_lowmem_reserve();
5684 setup_per_zone_inactive_ratio();
5687 module_init(init_per_zone_wmark_min)
5690 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5691 * that we can call two helper functions whenever min_free_kbytes
5694 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5695 void __user *buffer, size_t *length, loff_t *ppos)
5697 proc_dointvec(table, write, buffer, length, ppos);
5699 user_min_free_kbytes = min_free_kbytes;
5700 setup_per_zone_wmarks();
5706 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5707 void __user *buffer, size_t *length, loff_t *ppos)
5712 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5717 zone->min_unmapped_pages = (zone->managed_pages *
5718 sysctl_min_unmapped_ratio) / 100;
5722 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5723 void __user *buffer, size_t *length, loff_t *ppos)
5728 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5733 zone->min_slab_pages = (zone->managed_pages *
5734 sysctl_min_slab_ratio) / 100;
5740 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5741 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5742 * whenever sysctl_lowmem_reserve_ratio changes.
5744 * The reserve ratio obviously has absolutely no relation with the
5745 * minimum watermarks. The lowmem reserve ratio can only make sense
5746 * if in function of the boot time zone sizes.
5748 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5749 void __user *buffer, size_t *length, loff_t *ppos)
5751 proc_dointvec_minmax(table, write, buffer, length, ppos);
5752 setup_per_zone_lowmem_reserve();
5757 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5758 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5759 * pagelist can have before it gets flushed back to buddy allocator.
5761 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5762 void __user *buffer, size_t *length, loff_t *ppos)
5768 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5769 if (!write || (ret < 0))
5772 mutex_lock(&pcp_batch_high_lock);
5773 for_each_populated_zone(zone) {
5775 high = zone->managed_pages / percpu_pagelist_fraction;
5776 for_each_possible_cpu(cpu)
5777 pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5780 mutex_unlock(&pcp_batch_high_lock);
5784 int hashdist = HASHDIST_DEFAULT;
5787 static int __init set_hashdist(char *str)
5791 hashdist = simple_strtoul(str, &str, 0);
5794 __setup("hashdist=", set_hashdist);
5798 * allocate a large system hash table from bootmem
5799 * - it is assumed that the hash table must contain an exact power-of-2
5800 * quantity of entries
5801 * - limit is the number of hash buckets, not the total allocation size
5803 void *__init alloc_large_system_hash(const char *tablename,
5804 unsigned long bucketsize,
5805 unsigned long numentries,
5808 unsigned int *_hash_shift,
5809 unsigned int *_hash_mask,
5810 unsigned long low_limit,
5811 unsigned long high_limit)
5813 unsigned long long max = high_limit;
5814 unsigned long log2qty, size;
5817 /* allow the kernel cmdline to have a say */
5819 /* round applicable memory size up to nearest megabyte */
5820 numentries = nr_kernel_pages;
5822 /* It isn't necessary when PAGE_SIZE >= 1MB */
5823 if (PAGE_SHIFT < 20)
5824 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5826 /* limit to 1 bucket per 2^scale bytes of low memory */
5827 if (scale > PAGE_SHIFT)
5828 numentries >>= (scale - PAGE_SHIFT);
5830 numentries <<= (PAGE_SHIFT - scale);
5832 /* Make sure we've got at least a 0-order allocation.. */
5833 if (unlikely(flags & HASH_SMALL)) {
5834 /* Makes no sense without HASH_EARLY */
5835 WARN_ON(!(flags & HASH_EARLY));
5836 if (!(numentries >> *_hash_shift)) {
5837 numentries = 1UL << *_hash_shift;
5838 BUG_ON(!numentries);
5840 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5841 numentries = PAGE_SIZE / bucketsize;
5843 numentries = roundup_pow_of_two(numentries);
5845 /* limit allocation size to 1/16 total memory by default */
5847 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5848 do_div(max, bucketsize);
5850 max = min(max, 0x80000000ULL);
5852 if (numentries < low_limit)
5853 numentries = low_limit;
5854 if (numentries > max)
5857 log2qty = ilog2(numentries);
5860 size = bucketsize << log2qty;
5861 if (flags & HASH_EARLY)
5862 table = alloc_bootmem_nopanic(size);
5864 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5867 * If bucketsize is not a power-of-two, we may free
5868 * some pages at the end of hash table which
5869 * alloc_pages_exact() automatically does
5871 if (get_order(size) < MAX_ORDER) {
5872 table = alloc_pages_exact(size, GFP_ATOMIC);
5873 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5876 } while (!table && size > PAGE_SIZE && --log2qty);
5879 panic("Failed to allocate %s hash table\n", tablename);
5881 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5884 ilog2(size) - PAGE_SHIFT,
5888 *_hash_shift = log2qty;
5890 *_hash_mask = (1 << log2qty) - 1;
5895 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5896 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5899 #ifdef CONFIG_SPARSEMEM
5900 return __pfn_to_section(pfn)->pageblock_flags;
5902 return zone->pageblock_flags;
5903 #endif /* CONFIG_SPARSEMEM */
5906 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5908 #ifdef CONFIG_SPARSEMEM
5909 pfn &= (PAGES_PER_SECTION-1);
5910 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5912 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5913 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5914 #endif /* CONFIG_SPARSEMEM */
5918 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5919 * @page: The page within the block of interest
5920 * @start_bitidx: The first bit of interest to retrieve
5921 * @end_bitidx: The last bit of interest
5922 * returns pageblock_bits flags
5924 unsigned long get_pageblock_flags_group(struct page *page,
5925 int start_bitidx, int end_bitidx)
5928 unsigned long *bitmap;
5929 unsigned long pfn, bitidx;
5930 unsigned long flags = 0;
5931 unsigned long value = 1;
5933 zone = page_zone(page);
5934 pfn = page_to_pfn(page);
5935 bitmap = get_pageblock_bitmap(zone, pfn);
5936 bitidx = pfn_to_bitidx(zone, pfn);
5938 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5939 if (test_bit(bitidx + start_bitidx, bitmap))
5946 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5947 * @page: The page within the block of interest
5948 * @start_bitidx: The first bit of interest
5949 * @end_bitidx: The last bit of interest
5950 * @flags: The flags to set
5952 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5953 int start_bitidx, int end_bitidx)
5956 unsigned long *bitmap;
5957 unsigned long pfn, bitidx;
5958 unsigned long value = 1;
5960 zone = page_zone(page);
5961 pfn = page_to_pfn(page);
5962 bitmap = get_pageblock_bitmap(zone, pfn);
5963 bitidx = pfn_to_bitidx(zone, pfn);
5964 VM_BUG_ON(!zone_spans_pfn(zone, pfn));
5966 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5968 __set_bit(bitidx + start_bitidx, bitmap);
5970 __clear_bit(bitidx + start_bitidx, bitmap);
5974 * This function checks whether pageblock includes unmovable pages or not.
5975 * If @count is not zero, it is okay to include less @count unmovable pages
5977 * PageLRU check without isolation or lru_lock could race so that
5978 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5979 * expect this function should be exact.
5981 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5982 bool skip_hwpoisoned_pages)
5984 unsigned long pfn, iter, found;
5988 * For avoiding noise data, lru_add_drain_all() should be called
5989 * If ZONE_MOVABLE, the zone never contains unmovable pages
5991 if (zone_idx(zone) == ZONE_MOVABLE)
5993 mt = get_pageblock_migratetype(page);
5994 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5997 pfn = page_to_pfn(page);
5998 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5999 unsigned long check = pfn + iter;
6001 if (!pfn_valid_within(check))
6004 page = pfn_to_page(check);
6007 * Hugepages are not in LRU lists, but they're movable.
6008 * We need not scan over tail pages bacause we don't
6009 * handle each tail page individually in migration.
6011 if (PageHuge(page)) {
6012 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6017 * We can't use page_count without pin a page
6018 * because another CPU can free compound page.
6019 * This check already skips compound tails of THP
6020 * because their page->_count is zero at all time.
6022 if (!atomic_read(&page->_count)) {
6023 if (PageBuddy(page))
6024 iter += (1 << page_order(page)) - 1;
6029 * The HWPoisoned page may be not in buddy system, and
6030 * page_count() is not 0.
6032 if (skip_hwpoisoned_pages && PageHWPoison(page))
6038 * If there are RECLAIMABLE pages, we need to check it.
6039 * But now, memory offline itself doesn't call shrink_slab()
6040 * and it still to be fixed.
6043 * If the page is not RAM, page_count()should be 0.
6044 * we don't need more check. This is an _used_ not-movable page.
6046 * The problematic thing here is PG_reserved pages. PG_reserved
6047 * is set to both of a memory hole page and a _used_ kernel
6056 bool is_pageblock_removable_nolock(struct page *page)
6062 * We have to be careful here because we are iterating over memory
6063 * sections which are not zone aware so we might end up outside of
6064 * the zone but still within the section.
6065 * We have to take care about the node as well. If the node is offline
6066 * its NODE_DATA will be NULL - see page_zone.
6068 if (!node_online(page_to_nid(page)))
6071 zone = page_zone(page);
6072 pfn = page_to_pfn(page);
6073 if (!zone_spans_pfn(zone, pfn))
6076 return !has_unmovable_pages(zone, page, 0, true);
6081 static unsigned long pfn_max_align_down(unsigned long pfn)
6083 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6084 pageblock_nr_pages) - 1);
6087 static unsigned long pfn_max_align_up(unsigned long pfn)
6089 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6090 pageblock_nr_pages));
6093 /* [start, end) must belong to a single zone. */
6094 static int __alloc_contig_migrate_range(struct compact_control *cc,
6095 unsigned long start, unsigned long end)
6097 /* This function is based on compact_zone() from compaction.c. */
6098 unsigned long nr_reclaimed;
6099 unsigned long pfn = start;
6100 unsigned int tries = 0;
6105 while (pfn < end || !list_empty(&cc->migratepages)) {
6106 if (fatal_signal_pending(current)) {
6111 if (list_empty(&cc->migratepages)) {
6112 cc->nr_migratepages = 0;
6113 pfn = isolate_migratepages_range(cc->zone, cc,
6120 } else if (++tries == 5) {
6121 ret = ret < 0 ? ret : -EBUSY;
6125 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6127 cc->nr_migratepages -= nr_reclaimed;
6129 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6130 0, MIGRATE_SYNC, MR_CMA);
6133 putback_movable_pages(&cc->migratepages);
6140 * alloc_contig_range() -- tries to allocate given range of pages
6141 * @start: start PFN to allocate
6142 * @end: one-past-the-last PFN to allocate
6143 * @migratetype: migratetype of the underlaying pageblocks (either
6144 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6145 * in range must have the same migratetype and it must
6146 * be either of the two.
6148 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6149 * aligned, however it's the caller's responsibility to guarantee that
6150 * we are the only thread that changes migrate type of pageblocks the
6153 * The PFN range must belong to a single zone.
6155 * Returns zero on success or negative error code. On success all
6156 * pages which PFN is in [start, end) are allocated for the caller and
6157 * need to be freed with free_contig_range().
6159 int alloc_contig_range(unsigned long start, unsigned long end,
6160 unsigned migratetype)
6162 unsigned long outer_start, outer_end;
6165 struct compact_control cc = {
6166 .nr_migratepages = 0,
6168 .zone = page_zone(pfn_to_page(start)),
6170 .ignore_skip_hint = true,
6172 INIT_LIST_HEAD(&cc.migratepages);
6175 * What we do here is we mark all pageblocks in range as
6176 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6177 * have different sizes, and due to the way page allocator
6178 * work, we align the range to biggest of the two pages so
6179 * that page allocator won't try to merge buddies from
6180 * different pageblocks and change MIGRATE_ISOLATE to some
6181 * other migration type.
6183 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6184 * migrate the pages from an unaligned range (ie. pages that
6185 * we are interested in). This will put all the pages in
6186 * range back to page allocator as MIGRATE_ISOLATE.
6188 * When this is done, we take the pages in range from page
6189 * allocator removing them from the buddy system. This way
6190 * page allocator will never consider using them.
6192 * This lets us mark the pageblocks back as
6193 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6194 * aligned range but not in the unaligned, original range are
6195 * put back to page allocator so that buddy can use them.
6198 ret = start_isolate_page_range(pfn_max_align_down(start),
6199 pfn_max_align_up(end), migratetype,
6204 ret = __alloc_contig_migrate_range(&cc, start, end);
6209 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6210 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6211 * more, all pages in [start, end) are free in page allocator.
6212 * What we are going to do is to allocate all pages from
6213 * [start, end) (that is remove them from page allocator).
6215 * The only problem is that pages at the beginning and at the
6216 * end of interesting range may be not aligned with pages that
6217 * page allocator holds, ie. they can be part of higher order
6218 * pages. Because of this, we reserve the bigger range and
6219 * once this is done free the pages we are not interested in.
6221 * We don't have to hold zone->lock here because the pages are
6222 * isolated thus they won't get removed from buddy.
6225 lru_add_drain_all();
6229 outer_start = start;
6230 while (!PageBuddy(pfn_to_page(outer_start))) {
6231 if (++order >= MAX_ORDER) {
6235 outer_start &= ~0UL << order;
6238 /* Make sure the range is really isolated. */
6239 if (test_pages_isolated(outer_start, end, false)) {
6240 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6247 /* Grab isolated pages from freelists. */
6248 outer_end = isolate_freepages_range(&cc, outer_start, end);
6254 /* Free head and tail (if any) */
6255 if (start != outer_start)
6256 free_contig_range(outer_start, start - outer_start);
6257 if (end != outer_end)
6258 free_contig_range(end, outer_end - end);
6261 undo_isolate_page_range(pfn_max_align_down(start),
6262 pfn_max_align_up(end), migratetype);
6266 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6268 unsigned int count = 0;
6270 for (; nr_pages--; pfn++) {
6271 struct page *page = pfn_to_page(pfn);
6273 count += page_count(page) != 1;
6276 WARN(count != 0, "%d pages are still in use!\n", count);
6280 #ifdef CONFIG_MEMORY_HOTPLUG
6282 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6283 * page high values need to be recalulated.
6285 void __meminit zone_pcp_update(struct zone *zone)
6288 mutex_lock(&pcp_batch_high_lock);
6289 for_each_possible_cpu(cpu)
6290 pageset_set_high_and_batch(zone,
6291 per_cpu_ptr(zone->pageset, cpu));
6292 mutex_unlock(&pcp_batch_high_lock);
6296 void zone_pcp_reset(struct zone *zone)
6298 unsigned long flags;
6300 struct per_cpu_pageset *pset;
6302 /* avoid races with drain_pages() */
6303 local_irq_save(flags);
6304 if (zone->pageset != &boot_pageset) {
6305 for_each_online_cpu(cpu) {
6306 pset = per_cpu_ptr(zone->pageset, cpu);
6307 drain_zonestat(zone, pset);
6309 free_percpu(zone->pageset);
6310 zone->pageset = &boot_pageset;
6312 local_irq_restore(flags);
6315 #ifdef CONFIG_MEMORY_HOTREMOVE
6317 * All pages in the range must be isolated before calling this.
6320 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6326 unsigned long flags;
6327 /* find the first valid pfn */
6328 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6333 zone = page_zone(pfn_to_page(pfn));
6334 spin_lock_irqsave(&zone->lock, flags);
6336 while (pfn < end_pfn) {
6337 if (!pfn_valid(pfn)) {
6341 page = pfn_to_page(pfn);
6343 * The HWPoisoned page may be not in buddy system, and
6344 * page_count() is not 0.
6346 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6348 SetPageReserved(page);
6352 BUG_ON(page_count(page));
6353 BUG_ON(!PageBuddy(page));
6354 order = page_order(page);
6355 #ifdef CONFIG_DEBUG_VM
6356 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6357 pfn, 1 << order, end_pfn);
6359 list_del(&page->lru);
6360 rmv_page_order(page);
6361 zone->free_area[order].nr_free--;
6362 for (i = 0; i < (1 << order); i++)
6363 SetPageReserved((page+i));
6364 pfn += (1 << order);
6366 spin_unlock_irqrestore(&zone->lock, flags);
6370 #ifdef CONFIG_MEMORY_FAILURE
6371 bool is_free_buddy_page(struct page *page)
6373 struct zone *zone = page_zone(page);
6374 unsigned long pfn = page_to_pfn(page);
6375 unsigned long flags;
6378 spin_lock_irqsave(&zone->lock, flags);
6379 for (order = 0; order < MAX_ORDER; order++) {
6380 struct page *page_head = page - (pfn & ((1 << order) - 1));
6382 if (PageBuddy(page_head) && page_order(page_head) >= order)
6385 spin_unlock_irqrestore(&zone->lock, flags);
6387 return order < MAX_ORDER;
6391 static const struct trace_print_flags pageflag_names[] = {
6392 {1UL << PG_locked, "locked" },
6393 {1UL << PG_error, "error" },
6394 {1UL << PG_referenced, "referenced" },
6395 {1UL << PG_uptodate, "uptodate" },
6396 {1UL << PG_dirty, "dirty" },
6397 {1UL << PG_lru, "lru" },
6398 {1UL << PG_active, "active" },
6399 {1UL << PG_slab, "slab" },
6400 {1UL << PG_owner_priv_1, "owner_priv_1" },
6401 {1UL << PG_arch_1, "arch_1" },
6402 {1UL << PG_reserved, "reserved" },
6403 {1UL << PG_private, "private" },
6404 {1UL << PG_private_2, "private_2" },
6405 {1UL << PG_writeback, "writeback" },
6406 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6407 {1UL << PG_head, "head" },
6408 {1UL << PG_tail, "tail" },
6410 {1UL << PG_compound, "compound" },
6412 {1UL << PG_swapcache, "swapcache" },
6413 {1UL << PG_mappedtodisk, "mappedtodisk" },
6414 {1UL << PG_reclaim, "reclaim" },
6415 {1UL << PG_swapbacked, "swapbacked" },
6416 {1UL << PG_unevictable, "unevictable" },
6418 {1UL << PG_mlocked, "mlocked" },
6420 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6421 {1UL << PG_uncached, "uncached" },
6423 #ifdef CONFIG_MEMORY_FAILURE
6424 {1UL << PG_hwpoison, "hwpoison" },
6426 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6427 {1UL << PG_compound_lock, "compound_lock" },
6431 static void dump_page_flags(unsigned long flags)
6433 const char *delim = "";
6437 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6439 printk(KERN_ALERT "page flags: %#lx(", flags);
6441 /* remove zone id */
6442 flags &= (1UL << NR_PAGEFLAGS) - 1;
6444 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6446 mask = pageflag_names[i].mask;
6447 if ((flags & mask) != mask)
6451 printk("%s%s", delim, pageflag_names[i].name);
6455 /* check for left over flags */
6457 printk("%s%#lx", delim, flags);
6462 void dump_page(struct page *page)
6465 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6466 page, atomic_read(&page->_count), page_mapcount(page),
6467 page->mapping, page->index);
6468 dump_page_flags(page->flags);
6469 mem_cgroup_print_bad_page(page);