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 = -1;
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, char *reason, unsigned long bad_flags)
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));
332 dump_page_badflags(page, reason, bad_flags);
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)) {
386 bad_page(page, "wrong compound order", 0);
390 __ClearPageHead(page);
392 for (i = 1; i < nr_pages; i++) {
393 struct page *p = page + i;
395 if (unlikely(!PageTail(p))) {
396 bad_page(page, "PageTail not set", 0);
398 } else if (unlikely(p->first_page != page)) {
399 bad_page(page, "first_page not consistent", 0);
408 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
413 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
414 * and __GFP_HIGHMEM from hard or soft interrupt context.
416 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
417 for (i = 0; i < (1 << order); i++)
418 clear_highpage(page + i);
421 #ifdef CONFIG_DEBUG_PAGEALLOC
422 unsigned int _debug_guardpage_minorder;
424 static int __init debug_guardpage_minorder_setup(char *buf)
428 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
429 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
432 _debug_guardpage_minorder = res;
433 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
436 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
438 static inline void set_page_guard_flag(struct page *page)
440 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
443 static inline void clear_page_guard_flag(struct page *page)
445 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
448 static inline void set_page_guard_flag(struct page *page) { }
449 static inline void clear_page_guard_flag(struct page *page) { }
452 static inline void set_page_order(struct page *page, int order)
454 set_page_private(page, order);
455 __SetPageBuddy(page);
458 static inline void rmv_page_order(struct page *page)
460 __ClearPageBuddy(page);
461 set_page_private(page, 0);
465 * Locate the struct page for both the matching buddy in our
466 * pair (buddy1) and the combined O(n+1) page they form (page).
468 * 1) Any buddy B1 will have an order O twin B2 which satisfies
469 * the following equation:
471 * For example, if the starting buddy (buddy2) is #8 its order
473 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
475 * 2) Any buddy B will have an order O+1 parent P which
476 * satisfies the following equation:
479 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
481 static inline unsigned long
482 __find_buddy_index(unsigned long page_idx, unsigned int order)
484 return page_idx ^ (1 << order);
488 * This function checks whether a page is free && is the buddy
489 * we can do coalesce a page and its buddy if
490 * (a) the buddy is not in a hole &&
491 * (b) the buddy is in the buddy system &&
492 * (c) a page and its buddy have the same order &&
493 * (d) a page and its buddy are in the same zone.
495 * For recording whether a page is in the buddy system, we set ->_mapcount
496 * PAGE_BUDDY_MAPCOUNT_VALUE.
497 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
498 * serialized by zone->lock.
500 * For recording page's order, we use page_private(page).
502 static inline int page_is_buddy(struct page *page, struct page *buddy,
505 if (!pfn_valid_within(page_to_pfn(buddy)))
508 if (page_zone_id(page) != page_zone_id(buddy))
511 if (page_is_guard(buddy) && page_order(buddy) == order) {
512 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
516 if (PageBuddy(buddy) && page_order(buddy) == order) {
517 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
524 * Freeing function for a buddy system allocator.
526 * The concept of a buddy system is to maintain direct-mapped table
527 * (containing bit values) for memory blocks of various "orders".
528 * The bottom level table contains the map for the smallest allocatable
529 * units of memory (here, pages), and each level above it describes
530 * pairs of units from the levels below, hence, "buddies".
531 * At a high level, all that happens here is marking the table entry
532 * at the bottom level available, and propagating the changes upward
533 * as necessary, plus some accounting needed to play nicely with other
534 * parts of the VM system.
535 * At each level, we keep a list of pages, which are heads of continuous
536 * free pages of length of (1 << order) and marked with _mapcount
537 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
539 * So when we are allocating or freeing one, we can derive the state of the
540 * other. That is, if we allocate a small block, and both were
541 * free, the remainder of the region must be split into blocks.
542 * If a block is freed, and its buddy is also free, then this
543 * triggers coalescing into a block of larger size.
548 static inline void __free_one_page(struct page *page,
549 struct zone *zone, unsigned int order,
552 unsigned long page_idx;
553 unsigned long combined_idx;
554 unsigned long uninitialized_var(buddy_idx);
557 VM_BUG_ON(!zone_is_initialized(zone));
559 if (unlikely(PageCompound(page)))
560 if (unlikely(destroy_compound_page(page, order)))
563 VM_BUG_ON(migratetype == -1);
565 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
567 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
568 VM_BUG_ON_PAGE(bad_range(zone, page), page);
570 while (order < MAX_ORDER-1) {
571 buddy_idx = __find_buddy_index(page_idx, order);
572 buddy = page + (buddy_idx - page_idx);
573 if (!page_is_buddy(page, buddy, order))
576 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
577 * merge with it and move up one order.
579 if (page_is_guard(buddy)) {
580 clear_page_guard_flag(buddy);
581 set_page_private(page, 0);
582 __mod_zone_freepage_state(zone, 1 << order,
585 list_del(&buddy->lru);
586 zone->free_area[order].nr_free--;
587 rmv_page_order(buddy);
589 combined_idx = buddy_idx & page_idx;
590 page = page + (combined_idx - page_idx);
591 page_idx = combined_idx;
594 set_page_order(page, order);
597 * If this is not the largest possible page, check if the buddy
598 * of the next-highest order is free. If it is, it's possible
599 * that pages are being freed that will coalesce soon. In case,
600 * that is happening, add the free page to the tail of the list
601 * so it's less likely to be used soon and more likely to be merged
602 * as a higher order page
604 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
605 struct page *higher_page, *higher_buddy;
606 combined_idx = buddy_idx & page_idx;
607 higher_page = page + (combined_idx - page_idx);
608 buddy_idx = __find_buddy_index(combined_idx, order + 1);
609 higher_buddy = higher_page + (buddy_idx - combined_idx);
610 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
611 list_add_tail(&page->lru,
612 &zone->free_area[order].free_list[migratetype]);
617 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
619 zone->free_area[order].nr_free++;
622 static inline int free_pages_check(struct page *page)
624 char *bad_reason = NULL;
625 unsigned long bad_flags = 0;
627 if (unlikely(page_mapcount(page)))
628 bad_reason = "nonzero mapcount";
629 if (unlikely(page->mapping != NULL))
630 bad_reason = "non-NULL mapping";
631 if (unlikely(atomic_read(&page->_count) != 0))
632 bad_reason = "nonzero _count";
633 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
634 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
635 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
637 if (unlikely(mem_cgroup_bad_page_check(page)))
638 bad_reason = "cgroup check failed";
639 if (unlikely(bad_reason)) {
640 bad_page(page, bad_reason, bad_flags);
643 page_cpupid_reset_last(page);
644 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
645 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
650 * Frees a number of pages from the PCP lists
651 * Assumes all pages on list are in same zone, and of same order.
652 * count is the number of pages to free.
654 * If the zone was previously in an "all pages pinned" state then look to
655 * see if this freeing clears that state.
657 * And clear the zone's pages_scanned counter, to hold off the "all pages are
658 * pinned" detection logic.
660 static void free_pcppages_bulk(struct zone *zone, int count,
661 struct per_cpu_pages *pcp)
667 spin_lock(&zone->lock);
668 zone->pages_scanned = 0;
672 struct list_head *list;
675 * Remove pages from lists in a round-robin fashion. A
676 * batch_free count is maintained that is incremented when an
677 * empty list is encountered. This is so more pages are freed
678 * off fuller lists instead of spinning excessively around empty
683 if (++migratetype == MIGRATE_PCPTYPES)
685 list = &pcp->lists[migratetype];
686 } while (list_empty(list));
688 /* This is the only non-empty list. Free them all. */
689 if (batch_free == MIGRATE_PCPTYPES)
690 batch_free = to_free;
693 int mt; /* migratetype of the to-be-freed page */
695 page = list_entry(list->prev, struct page, lru);
696 /* must delete as __free_one_page list manipulates */
697 list_del(&page->lru);
698 mt = get_freepage_migratetype(page);
699 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
700 __free_one_page(page, zone, 0, mt);
701 trace_mm_page_pcpu_drain(page, 0, mt);
702 if (likely(!is_migrate_isolate_page(page))) {
703 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
704 if (is_migrate_cma(mt))
705 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
707 } while (--to_free && --batch_free && !list_empty(list));
709 spin_unlock(&zone->lock);
712 static void free_one_page(struct zone *zone, struct page *page, int order,
715 spin_lock(&zone->lock);
716 zone->pages_scanned = 0;
718 __free_one_page(page, zone, order, migratetype);
719 if (unlikely(!is_migrate_isolate(migratetype)))
720 __mod_zone_freepage_state(zone, 1 << order, migratetype);
721 spin_unlock(&zone->lock);
724 static bool free_pages_prepare(struct page *page, unsigned int order)
729 trace_mm_page_free(page, order);
730 kmemcheck_free_shadow(page, order);
733 page->mapping = NULL;
734 for (i = 0; i < (1 << order); i++)
735 bad += free_pages_check(page + i);
739 if (!PageHighMem(page)) {
740 debug_check_no_locks_freed(page_address(page),
742 debug_check_no_obj_freed(page_address(page),
745 arch_free_page(page, order);
746 kernel_map_pages(page, 1 << order, 0);
751 static void __free_pages_ok(struct page *page, unsigned int order)
756 if (!free_pages_prepare(page, order))
759 local_irq_save(flags);
760 __count_vm_events(PGFREE, 1 << order);
761 migratetype = get_pageblock_migratetype(page);
762 set_freepage_migratetype(page, migratetype);
763 free_one_page(page_zone(page), page, order, migratetype);
764 local_irq_restore(flags);
767 void __init __free_pages_bootmem(struct page *page, unsigned int order)
769 unsigned int nr_pages = 1 << order;
770 struct page *p = page;
774 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
776 __ClearPageReserved(p);
777 set_page_count(p, 0);
779 __ClearPageReserved(p);
780 set_page_count(p, 0);
782 page_zone(page)->managed_pages += nr_pages;
783 set_page_refcounted(page);
784 __free_pages(page, order);
788 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
789 void __init init_cma_reserved_pageblock(struct page *page)
791 unsigned i = pageblock_nr_pages;
792 struct page *p = page;
795 __ClearPageReserved(p);
796 set_page_count(p, 0);
799 set_page_refcounted(page);
800 set_pageblock_migratetype(page, MIGRATE_CMA);
801 __free_pages(page, pageblock_order);
802 adjust_managed_page_count(page, pageblock_nr_pages);
807 * The order of subdivision here is critical for the IO subsystem.
808 * Please do not alter this order without good reasons and regression
809 * testing. Specifically, as large blocks of memory are subdivided,
810 * the order in which smaller blocks are delivered depends on the order
811 * they're subdivided in this function. This is the primary factor
812 * influencing the order in which pages are delivered to the IO
813 * subsystem according to empirical testing, and this is also justified
814 * by considering the behavior of a buddy system containing a single
815 * large block of memory acted on by a series of small allocations.
816 * This behavior is a critical factor in sglist merging's success.
820 static inline void expand(struct zone *zone, struct page *page,
821 int low, int high, struct free_area *area,
824 unsigned long size = 1 << high;
830 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
832 #ifdef CONFIG_DEBUG_PAGEALLOC
833 if (high < debug_guardpage_minorder()) {
835 * Mark as guard pages (or page), that will allow to
836 * merge back to allocator when buddy will be freed.
837 * Corresponding page table entries will not be touched,
838 * pages will stay not present in virtual address space
840 INIT_LIST_HEAD(&page[size].lru);
841 set_page_guard_flag(&page[size]);
842 set_page_private(&page[size], high);
843 /* Guard pages are not available for any usage */
844 __mod_zone_freepage_state(zone, -(1 << high),
849 list_add(&page[size].lru, &area->free_list[migratetype]);
851 set_page_order(&page[size], high);
856 * This page is about to be returned from the page allocator
858 static inline int check_new_page(struct page *page)
860 char *bad_reason = NULL;
861 unsigned long bad_flags = 0;
863 if (unlikely(page_mapcount(page)))
864 bad_reason = "nonzero mapcount";
865 if (unlikely(page->mapping != NULL))
866 bad_reason = "non-NULL mapping";
867 if (unlikely(atomic_read(&page->_count) != 0))
868 bad_reason = "nonzero _count";
869 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
870 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
871 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
873 if (unlikely(mem_cgroup_bad_page_check(page)))
874 bad_reason = "cgroup check failed";
875 if (unlikely(bad_reason)) {
876 bad_page(page, bad_reason, bad_flags);
882 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
886 for (i = 0; i < (1 << order); i++) {
887 struct page *p = page + i;
888 if (unlikely(check_new_page(p)))
892 set_page_private(page, 0);
893 arch_alloc_page(page, order);
894 kernel_map_pages(page, 1 << order, 1);
896 if (gfp_flags & __GFP_ZERO)
897 prep_zero_page(page, order, gfp_flags);
899 if (order && (gfp_flags & __GFP_COMP))
900 prep_compound_page(page, order);
903 * Make sure the caller of get_page_unless_zero() will see the
904 * fully initialized page. Say, to ensure that compound_lock()
905 * can't race with the non-atomic __SetPage*() above.
907 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
910 set_page_refcounted(page);
916 * Go through the free lists for the given migratetype and remove
917 * the smallest available page from the freelists
920 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
923 unsigned int current_order;
924 struct free_area *area;
927 /* Find a page of the appropriate size in the preferred list */
928 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
929 area = &(zone->free_area[current_order]);
930 if (list_empty(&area->free_list[migratetype]))
933 page = list_entry(area->free_list[migratetype].next,
935 list_del(&page->lru);
936 rmv_page_order(page);
938 expand(zone, page, order, current_order, area, migratetype);
947 * This array describes the order lists are fallen back to when
948 * the free lists for the desirable migrate type are depleted
950 static int fallbacks[MIGRATE_TYPES][4] = {
951 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
952 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
954 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
955 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
957 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
959 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
960 #ifdef CONFIG_MEMORY_ISOLATION
961 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
966 * Move the free pages in a range to the free lists of the requested type.
967 * Note that start_page and end_pages are not aligned on a pageblock
968 * boundary. If alignment is required, use move_freepages_block()
970 int move_freepages(struct zone *zone,
971 struct page *start_page, struct page *end_page,
978 #ifndef CONFIG_HOLES_IN_ZONE
980 * page_zone is not safe to call in this context when
981 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
982 * anyway as we check zone boundaries in move_freepages_block().
983 * Remove at a later date when no bug reports exist related to
984 * grouping pages by mobility
986 BUG_ON(page_zone(start_page) != page_zone(end_page));
989 for (page = start_page; page <= end_page;) {
990 /* Make sure we are not inadvertently changing nodes */
991 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
993 if (!pfn_valid_within(page_to_pfn(page))) {
998 if (!PageBuddy(page)) {
1003 order = page_order(page);
1004 list_move(&page->lru,
1005 &zone->free_area[order].free_list[migratetype]);
1006 set_freepage_migratetype(page, migratetype);
1008 pages_moved += 1 << order;
1014 int move_freepages_block(struct zone *zone, struct page *page,
1017 unsigned long start_pfn, end_pfn;
1018 struct page *start_page, *end_page;
1020 start_pfn = page_to_pfn(page);
1021 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1022 start_page = pfn_to_page(start_pfn);
1023 end_page = start_page + pageblock_nr_pages - 1;
1024 end_pfn = start_pfn + pageblock_nr_pages - 1;
1026 /* Do not cross zone boundaries */
1027 if (!zone_spans_pfn(zone, start_pfn))
1029 if (!zone_spans_pfn(zone, end_pfn))
1032 return move_freepages(zone, start_page, end_page, migratetype);
1035 static void change_pageblock_range(struct page *pageblock_page,
1036 int start_order, int migratetype)
1038 int nr_pageblocks = 1 << (start_order - pageblock_order);
1040 while (nr_pageblocks--) {
1041 set_pageblock_migratetype(pageblock_page, migratetype);
1042 pageblock_page += pageblock_nr_pages;
1047 * If breaking a large block of pages, move all free pages to the preferred
1048 * allocation list. If falling back for a reclaimable kernel allocation, be
1049 * more aggressive about taking ownership of free pages.
1051 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1052 * nor move CMA pages to different free lists. We don't want unmovable pages
1053 * to be allocated from MIGRATE_CMA areas.
1055 * Returns the new migratetype of the pageblock (or the same old migratetype
1056 * if it was unchanged).
1058 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1059 int start_type, int fallback_type)
1061 int current_order = page_order(page);
1064 * When borrowing from MIGRATE_CMA, we need to release the excess
1065 * buddy pages to CMA itself.
1067 if (is_migrate_cma(fallback_type))
1068 return fallback_type;
1070 /* Take ownership for orders >= pageblock_order */
1071 if (current_order >= pageblock_order) {
1072 change_pageblock_range(page, current_order, start_type);
1076 if (current_order >= pageblock_order / 2 ||
1077 start_type == MIGRATE_RECLAIMABLE ||
1078 page_group_by_mobility_disabled) {
1081 pages = move_freepages_block(zone, page, start_type);
1083 /* Claim the whole block if over half of it is free */
1084 if (pages >= (1 << (pageblock_order-1)) ||
1085 page_group_by_mobility_disabled) {
1087 set_pageblock_migratetype(page, start_type);
1093 return fallback_type;
1096 /* Remove an element from the buddy allocator from the fallback list */
1097 static inline struct page *
1098 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1100 struct free_area *area;
1103 int migratetype, new_type, i;
1105 /* Find the largest possible block of pages in the other list */
1106 for (current_order = MAX_ORDER-1; current_order >= order;
1109 migratetype = fallbacks[start_migratetype][i];
1111 /* MIGRATE_RESERVE handled later if necessary */
1112 if (migratetype == MIGRATE_RESERVE)
1115 area = &(zone->free_area[current_order]);
1116 if (list_empty(&area->free_list[migratetype]))
1119 page = list_entry(area->free_list[migratetype].next,
1123 new_type = try_to_steal_freepages(zone, page,
1127 /* Remove the page from the freelists */
1128 list_del(&page->lru);
1129 rmv_page_order(page);
1131 expand(zone, page, order, current_order, area,
1134 trace_mm_page_alloc_extfrag(page, order, current_order,
1135 start_migratetype, migratetype, new_type);
1145 * Do the hard work of removing an element from the buddy allocator.
1146 * Call me with the zone->lock already held.
1148 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1154 page = __rmqueue_smallest(zone, order, migratetype);
1156 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1157 page = __rmqueue_fallback(zone, order, migratetype);
1160 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1161 * is used because __rmqueue_smallest is an inline function
1162 * and we want just one call site
1165 migratetype = MIGRATE_RESERVE;
1170 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1175 * Obtain a specified number of elements from the buddy allocator, all under
1176 * a single hold of the lock, for efficiency. Add them to the supplied list.
1177 * Returns the number of new pages which were placed at *list.
1179 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1180 unsigned long count, struct list_head *list,
1181 int migratetype, int cold)
1183 int mt = migratetype, i;
1185 spin_lock(&zone->lock);
1186 for (i = 0; i < count; ++i) {
1187 struct page *page = __rmqueue(zone, order, migratetype);
1188 if (unlikely(page == NULL))
1192 * Split buddy pages returned by expand() are received here
1193 * in physical page order. The page is added to the callers and
1194 * list and the list head then moves forward. From the callers
1195 * perspective, the linked list is ordered by page number in
1196 * some conditions. This is useful for IO devices that can
1197 * merge IO requests if the physical pages are ordered
1200 if (likely(cold == 0))
1201 list_add(&page->lru, list);
1203 list_add_tail(&page->lru, list);
1204 if (IS_ENABLED(CONFIG_CMA)) {
1205 mt = get_pageblock_migratetype(page);
1206 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1209 set_freepage_migratetype(page, mt);
1211 if (is_migrate_cma(mt))
1212 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1215 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1216 spin_unlock(&zone->lock);
1222 * Called from the vmstat counter updater to drain pagesets of this
1223 * currently executing processor on remote nodes after they have
1226 * Note that this function must be called with the thread pinned to
1227 * a single processor.
1229 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1231 unsigned long flags;
1233 unsigned long batch;
1235 local_irq_save(flags);
1236 batch = ACCESS_ONCE(pcp->batch);
1237 if (pcp->count >= batch)
1240 to_drain = pcp->count;
1242 free_pcppages_bulk(zone, to_drain, pcp);
1243 pcp->count -= to_drain;
1245 local_irq_restore(flags);
1250 * Drain pages of the indicated processor.
1252 * The processor must either be the current processor and the
1253 * thread pinned to the current processor or a processor that
1256 static void drain_pages(unsigned int cpu)
1258 unsigned long flags;
1261 for_each_populated_zone(zone) {
1262 struct per_cpu_pageset *pset;
1263 struct per_cpu_pages *pcp;
1265 local_irq_save(flags);
1266 pset = per_cpu_ptr(zone->pageset, cpu);
1270 free_pcppages_bulk(zone, pcp->count, pcp);
1273 local_irq_restore(flags);
1278 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1280 void drain_local_pages(void *arg)
1282 drain_pages(smp_processor_id());
1286 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1288 * Note that this code is protected against sending an IPI to an offline
1289 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1290 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1291 * nothing keeps CPUs from showing up after we populated the cpumask and
1292 * before the call to on_each_cpu_mask().
1294 void drain_all_pages(void)
1297 struct per_cpu_pageset *pcp;
1301 * Allocate in the BSS so we wont require allocation in
1302 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1304 static cpumask_t cpus_with_pcps;
1307 * We don't care about racing with CPU hotplug event
1308 * as offline notification will cause the notified
1309 * cpu to drain that CPU pcps and on_each_cpu_mask
1310 * disables preemption as part of its processing
1312 for_each_online_cpu(cpu) {
1313 bool has_pcps = false;
1314 for_each_populated_zone(zone) {
1315 pcp = per_cpu_ptr(zone->pageset, cpu);
1316 if (pcp->pcp.count) {
1322 cpumask_set_cpu(cpu, &cpus_with_pcps);
1324 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1326 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1329 #ifdef CONFIG_HIBERNATION
1331 void mark_free_pages(struct zone *zone)
1333 unsigned long pfn, max_zone_pfn;
1334 unsigned long flags;
1336 struct list_head *curr;
1338 if (zone_is_empty(zone))
1341 spin_lock_irqsave(&zone->lock, flags);
1343 max_zone_pfn = zone_end_pfn(zone);
1344 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1345 if (pfn_valid(pfn)) {
1346 struct page *page = pfn_to_page(pfn);
1348 if (!swsusp_page_is_forbidden(page))
1349 swsusp_unset_page_free(page);
1352 for_each_migratetype_order(order, t) {
1353 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1356 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1357 for (i = 0; i < (1UL << order); i++)
1358 swsusp_set_page_free(pfn_to_page(pfn + i));
1361 spin_unlock_irqrestore(&zone->lock, flags);
1363 #endif /* CONFIG_PM */
1366 * Free a 0-order page
1367 * cold == 1 ? free a cold page : free a hot page
1369 void free_hot_cold_page(struct page *page, int cold)
1371 struct zone *zone = page_zone(page);
1372 struct per_cpu_pages *pcp;
1373 unsigned long flags;
1376 if (!free_pages_prepare(page, 0))
1379 migratetype = get_pageblock_migratetype(page);
1380 set_freepage_migratetype(page, migratetype);
1381 local_irq_save(flags);
1382 __count_vm_event(PGFREE);
1385 * We only track unmovable, reclaimable and movable on pcp lists.
1386 * Free ISOLATE pages back to the allocator because they are being
1387 * offlined but treat RESERVE as movable pages so we can get those
1388 * areas back if necessary. Otherwise, we may have to free
1389 * excessively into the page allocator
1391 if (migratetype >= MIGRATE_PCPTYPES) {
1392 if (unlikely(is_migrate_isolate(migratetype))) {
1393 free_one_page(zone, page, 0, migratetype);
1396 migratetype = MIGRATE_MOVABLE;
1399 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1401 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1403 list_add(&page->lru, &pcp->lists[migratetype]);
1405 if (pcp->count >= pcp->high) {
1406 unsigned long batch = ACCESS_ONCE(pcp->batch);
1407 free_pcppages_bulk(zone, batch, pcp);
1408 pcp->count -= batch;
1412 local_irq_restore(flags);
1416 * Free a list of 0-order pages
1418 void free_hot_cold_page_list(struct list_head *list, int cold)
1420 struct page *page, *next;
1422 list_for_each_entry_safe(page, next, list, lru) {
1423 trace_mm_page_free_batched(page, cold);
1424 free_hot_cold_page(page, cold);
1429 * split_page takes a non-compound higher-order page, and splits it into
1430 * n (1<<order) sub-pages: page[0..n]
1431 * Each sub-page must be freed individually.
1433 * Note: this is probably too low level an operation for use in drivers.
1434 * Please consult with lkml before using this in your driver.
1436 void split_page(struct page *page, unsigned int order)
1440 VM_BUG_ON_PAGE(PageCompound(page), page);
1441 VM_BUG_ON_PAGE(!page_count(page), page);
1443 #ifdef CONFIG_KMEMCHECK
1445 * Split shadow pages too, because free(page[0]) would
1446 * otherwise free the whole shadow.
1448 if (kmemcheck_page_is_tracked(page))
1449 split_page(virt_to_page(page[0].shadow), order);
1452 for (i = 1; i < (1 << order); i++)
1453 set_page_refcounted(page + i);
1455 EXPORT_SYMBOL_GPL(split_page);
1457 static int __isolate_free_page(struct page *page, unsigned int order)
1459 unsigned long watermark;
1463 BUG_ON(!PageBuddy(page));
1465 zone = page_zone(page);
1466 mt = get_pageblock_migratetype(page);
1468 if (!is_migrate_isolate(mt)) {
1469 /* Obey watermarks as if the page was being allocated */
1470 watermark = low_wmark_pages(zone) + (1 << order);
1471 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1474 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1477 /* Remove page from free list */
1478 list_del(&page->lru);
1479 zone->free_area[order].nr_free--;
1480 rmv_page_order(page);
1482 /* Set the pageblock if the isolated page is at least a pageblock */
1483 if (order >= pageblock_order - 1) {
1484 struct page *endpage = page + (1 << order) - 1;
1485 for (; page < endpage; page += pageblock_nr_pages) {
1486 int mt = get_pageblock_migratetype(page);
1487 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1488 set_pageblock_migratetype(page,
1493 return 1UL << order;
1497 * Similar to split_page except the page is already free. As this is only
1498 * being used for migration, the migratetype of the block also changes.
1499 * As this is called with interrupts disabled, the caller is responsible
1500 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1503 * Note: this is probably too low level an operation for use in drivers.
1504 * Please consult with lkml before using this in your driver.
1506 int split_free_page(struct page *page)
1511 order = page_order(page);
1513 nr_pages = __isolate_free_page(page, order);
1517 /* Split into individual pages */
1518 set_page_refcounted(page);
1519 split_page(page, order);
1524 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1525 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1529 struct page *buffered_rmqueue(struct zone *preferred_zone,
1530 struct zone *zone, int order, gfp_t gfp_flags,
1533 unsigned long flags;
1535 int cold = !!(gfp_flags & __GFP_COLD);
1538 if (likely(order == 0)) {
1539 struct per_cpu_pages *pcp;
1540 struct list_head *list;
1542 local_irq_save(flags);
1543 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1544 list = &pcp->lists[migratetype];
1545 if (list_empty(list)) {
1546 pcp->count += rmqueue_bulk(zone, 0,
1549 if (unlikely(list_empty(list)))
1554 page = list_entry(list->prev, struct page, lru);
1556 page = list_entry(list->next, struct page, lru);
1558 list_del(&page->lru);
1561 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1563 * __GFP_NOFAIL is not to be used in new code.
1565 * All __GFP_NOFAIL callers should be fixed so that they
1566 * properly detect and handle allocation failures.
1568 * We most definitely don't want callers attempting to
1569 * allocate greater than order-1 page units with
1572 WARN_ON_ONCE(order > 1);
1574 spin_lock_irqsave(&zone->lock, flags);
1575 page = __rmqueue(zone, order, migratetype);
1576 spin_unlock(&zone->lock);
1579 __mod_zone_freepage_state(zone, -(1 << order),
1580 get_pageblock_migratetype(page));
1583 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1584 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1585 zone_statistics(preferred_zone, zone, gfp_flags);
1586 local_irq_restore(flags);
1588 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1589 if (prep_new_page(page, order, gfp_flags))
1594 local_irq_restore(flags);
1598 #ifdef CONFIG_FAIL_PAGE_ALLOC
1601 struct fault_attr attr;
1603 u32 ignore_gfp_highmem;
1604 u32 ignore_gfp_wait;
1606 } fail_page_alloc = {
1607 .attr = FAULT_ATTR_INITIALIZER,
1608 .ignore_gfp_wait = 1,
1609 .ignore_gfp_highmem = 1,
1613 static int __init setup_fail_page_alloc(char *str)
1615 return setup_fault_attr(&fail_page_alloc.attr, str);
1617 __setup("fail_page_alloc=", setup_fail_page_alloc);
1619 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1621 if (order < fail_page_alloc.min_order)
1623 if (gfp_mask & __GFP_NOFAIL)
1625 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1627 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1630 return should_fail(&fail_page_alloc.attr, 1 << order);
1633 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1635 static int __init fail_page_alloc_debugfs(void)
1637 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1640 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1641 &fail_page_alloc.attr);
1643 return PTR_ERR(dir);
1645 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1646 &fail_page_alloc.ignore_gfp_wait))
1648 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1649 &fail_page_alloc.ignore_gfp_highmem))
1651 if (!debugfs_create_u32("min-order", mode, dir,
1652 &fail_page_alloc.min_order))
1657 debugfs_remove_recursive(dir);
1662 late_initcall(fail_page_alloc_debugfs);
1664 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1666 #else /* CONFIG_FAIL_PAGE_ALLOC */
1668 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1673 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1676 * Return true if free pages are above 'mark'. This takes into account the order
1677 * of the allocation.
1679 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1680 int classzone_idx, int alloc_flags, long free_pages)
1682 /* free_pages my go negative - that's OK */
1684 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1688 free_pages -= (1 << order) - 1;
1689 if (alloc_flags & ALLOC_HIGH)
1691 if (alloc_flags & ALLOC_HARDER)
1694 /* If allocation can't use CMA areas don't use free CMA pages */
1695 if (!(alloc_flags & ALLOC_CMA))
1696 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1699 if (free_pages - free_cma <= min + lowmem_reserve)
1701 for (o = 0; o < order; o++) {
1702 /* At the next order, this order's pages become unavailable */
1703 free_pages -= z->free_area[o].nr_free << o;
1705 /* Require fewer higher order pages to be free */
1708 if (free_pages <= min)
1714 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1715 int classzone_idx, int alloc_flags)
1717 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1718 zone_page_state(z, NR_FREE_PAGES));
1721 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1722 int classzone_idx, int alloc_flags)
1724 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1726 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1727 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1729 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1735 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1736 * skip over zones that are not allowed by the cpuset, or that have
1737 * been recently (in last second) found to be nearly full. See further
1738 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1739 * that have to skip over a lot of full or unallowed zones.
1741 * If the zonelist cache is present in the passed zonelist, then
1742 * returns a pointer to the allowed node mask (either the current
1743 * tasks mems_allowed, or node_states[N_MEMORY].)
1745 * If the zonelist cache is not available for this zonelist, does
1746 * nothing and returns NULL.
1748 * If the fullzones BITMAP in the zonelist cache is stale (more than
1749 * a second since last zap'd) then we zap it out (clear its bits.)
1751 * We hold off even calling zlc_setup, until after we've checked the
1752 * first zone in the zonelist, on the theory that most allocations will
1753 * be satisfied from that first zone, so best to examine that zone as
1754 * quickly as we can.
1756 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1758 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1759 nodemask_t *allowednodes; /* zonelist_cache approximation */
1761 zlc = zonelist->zlcache_ptr;
1765 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1766 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1767 zlc->last_full_zap = jiffies;
1770 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1771 &cpuset_current_mems_allowed :
1772 &node_states[N_MEMORY];
1773 return allowednodes;
1777 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1778 * if it is worth looking at further for free memory:
1779 * 1) Check that the zone isn't thought to be full (doesn't have its
1780 * bit set in the zonelist_cache fullzones BITMAP).
1781 * 2) Check that the zones node (obtained from the zonelist_cache
1782 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1783 * Return true (non-zero) if zone is worth looking at further, or
1784 * else return false (zero) if it is not.
1786 * This check -ignores- the distinction between various watermarks,
1787 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1788 * found to be full for any variation of these watermarks, it will
1789 * be considered full for up to one second by all requests, unless
1790 * we are so low on memory on all allowed nodes that we are forced
1791 * into the second scan of the zonelist.
1793 * In the second scan we ignore this zonelist cache and exactly
1794 * apply the watermarks to all zones, even it is slower to do so.
1795 * We are low on memory in the second scan, and should leave no stone
1796 * unturned looking for a free page.
1798 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1799 nodemask_t *allowednodes)
1801 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1802 int i; /* index of *z in zonelist zones */
1803 int n; /* node that zone *z is on */
1805 zlc = zonelist->zlcache_ptr;
1809 i = z - zonelist->_zonerefs;
1812 /* This zone is worth trying if it is allowed but not full */
1813 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1817 * Given 'z' scanning a zonelist, set the corresponding bit in
1818 * zlc->fullzones, so that subsequent attempts to allocate a page
1819 * from that zone don't waste time re-examining it.
1821 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1823 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1824 int i; /* index of *z in zonelist zones */
1826 zlc = zonelist->zlcache_ptr;
1830 i = z - zonelist->_zonerefs;
1832 set_bit(i, zlc->fullzones);
1836 * clear all zones full, called after direct reclaim makes progress so that
1837 * a zone that was recently full is not skipped over for up to a second
1839 static void zlc_clear_zones_full(struct zonelist *zonelist)
1841 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1843 zlc = zonelist->zlcache_ptr;
1847 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1850 static bool zone_local(struct zone *local_zone, struct zone *zone)
1852 return local_zone->node == zone->node;
1855 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1857 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1860 static void __paginginit init_zone_allows_reclaim(int nid)
1864 for_each_online_node(i)
1865 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1866 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1868 zone_reclaim_mode = 1;
1871 #else /* CONFIG_NUMA */
1873 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1878 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1879 nodemask_t *allowednodes)
1884 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1888 static void zlc_clear_zones_full(struct zonelist *zonelist)
1892 static bool zone_local(struct zone *local_zone, struct zone *zone)
1897 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1902 static inline void init_zone_allows_reclaim(int nid)
1905 #endif /* CONFIG_NUMA */
1908 * get_page_from_freelist goes through the zonelist trying to allocate
1911 static struct page *
1912 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1913 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1914 struct zone *preferred_zone, int migratetype)
1917 struct page *page = NULL;
1920 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1921 int zlc_active = 0; /* set if using zonelist_cache */
1922 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1924 classzone_idx = zone_idx(preferred_zone);
1927 * Scan zonelist, looking for a zone with enough free.
1928 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1930 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1931 high_zoneidx, nodemask) {
1934 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1935 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1937 if ((alloc_flags & ALLOC_CPUSET) &&
1938 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1940 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1941 if (unlikely(alloc_flags & ALLOC_NO_WATERMARKS))
1944 * Distribute pages in proportion to the individual
1945 * zone size to ensure fair page aging. The zone a
1946 * page was allocated in should have no effect on the
1947 * time the page has in memory before being reclaimed.
1949 * Try to stay in local zones in the fastpath. If
1950 * that fails, the slowpath is entered, which will do
1951 * another pass starting with the local zones, but
1952 * ultimately fall back to remote zones that do not
1953 * partake in the fairness round-robin cycle of this
1956 if (alloc_flags & ALLOC_WMARK_LOW) {
1957 if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1959 if (!zone_local(preferred_zone, zone))
1963 * When allocating a page cache page for writing, we
1964 * want to get it from a zone that is within its dirty
1965 * limit, such that no single zone holds more than its
1966 * proportional share of globally allowed dirty pages.
1967 * The dirty limits take into account the zone's
1968 * lowmem reserves and high watermark so that kswapd
1969 * should be able to balance it without having to
1970 * write pages from its LRU list.
1972 * This may look like it could increase pressure on
1973 * lower zones by failing allocations in higher zones
1974 * before they are full. But the pages that do spill
1975 * over are limited as the lower zones are protected
1976 * by this very same mechanism. It should not become
1977 * a practical burden to them.
1979 * XXX: For now, allow allocations to potentially
1980 * exceed the per-zone dirty limit in the slowpath
1981 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1982 * which is important when on a NUMA setup the allowed
1983 * zones are together not big enough to reach the
1984 * global limit. The proper fix for these situations
1985 * will require awareness of zones in the
1986 * dirty-throttling and the flusher threads.
1988 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1989 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1990 goto this_zone_full;
1992 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1993 if (!zone_watermark_ok(zone, order, mark,
1994 classzone_idx, alloc_flags)) {
1997 if (IS_ENABLED(CONFIG_NUMA) &&
1998 !did_zlc_setup && nr_online_nodes > 1) {
2000 * we do zlc_setup if there are multiple nodes
2001 * and before considering the first zone allowed
2004 allowednodes = zlc_setup(zonelist, alloc_flags);
2009 if (zone_reclaim_mode == 0 ||
2010 !zone_allows_reclaim(preferred_zone, zone))
2011 goto this_zone_full;
2014 * As we may have just activated ZLC, check if the first
2015 * eligible zone has failed zone_reclaim recently.
2017 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2018 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2021 ret = zone_reclaim(zone, gfp_mask, order);
2023 case ZONE_RECLAIM_NOSCAN:
2026 case ZONE_RECLAIM_FULL:
2027 /* scanned but unreclaimable */
2030 /* did we reclaim enough */
2031 if (zone_watermark_ok(zone, order, mark,
2032 classzone_idx, alloc_flags))
2036 * Failed to reclaim enough to meet watermark.
2037 * Only mark the zone full if checking the min
2038 * watermark or if we failed to reclaim just
2039 * 1<<order pages or else the page allocator
2040 * fastpath will prematurely mark zones full
2041 * when the watermark is between the low and
2044 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2045 ret == ZONE_RECLAIM_SOME)
2046 goto this_zone_full;
2053 page = buffered_rmqueue(preferred_zone, zone, order,
2054 gfp_mask, migratetype);
2058 if (IS_ENABLED(CONFIG_NUMA))
2059 zlc_mark_zone_full(zonelist, z);
2062 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2063 /* Disable zlc cache for second zonelist scan */
2070 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2071 * necessary to allocate the page. The expectation is
2072 * that the caller is taking steps that will free more
2073 * memory. The caller should avoid the page being used
2074 * for !PFMEMALLOC purposes.
2076 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2082 * Large machines with many possible nodes should not always dump per-node
2083 * meminfo in irq context.
2085 static inline bool should_suppress_show_mem(void)
2090 ret = in_interrupt();
2095 static DEFINE_RATELIMIT_STATE(nopage_rs,
2096 DEFAULT_RATELIMIT_INTERVAL,
2097 DEFAULT_RATELIMIT_BURST);
2099 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2101 unsigned int filter = SHOW_MEM_FILTER_NODES;
2103 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2104 debug_guardpage_minorder() > 0)
2108 * This documents exceptions given to allocations in certain
2109 * contexts that are allowed to allocate outside current's set
2112 if (!(gfp_mask & __GFP_NOMEMALLOC))
2113 if (test_thread_flag(TIF_MEMDIE) ||
2114 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2115 filter &= ~SHOW_MEM_FILTER_NODES;
2116 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2117 filter &= ~SHOW_MEM_FILTER_NODES;
2120 struct va_format vaf;
2123 va_start(args, fmt);
2128 pr_warn("%pV", &vaf);
2133 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2134 current->comm, order, gfp_mask);
2137 if (!should_suppress_show_mem())
2142 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2143 unsigned long did_some_progress,
2144 unsigned long pages_reclaimed)
2146 /* Do not loop if specifically requested */
2147 if (gfp_mask & __GFP_NORETRY)
2150 /* Always retry if specifically requested */
2151 if (gfp_mask & __GFP_NOFAIL)
2155 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2156 * making forward progress without invoking OOM. Suspend also disables
2157 * storage devices so kswapd will not help. Bail if we are suspending.
2159 if (!did_some_progress && pm_suspended_storage())
2163 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2164 * means __GFP_NOFAIL, but that may not be true in other
2167 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2171 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2172 * specified, then we retry until we no longer reclaim any pages
2173 * (above), or we've reclaimed an order of pages at least as
2174 * large as the allocation's order. In both cases, if the
2175 * allocation still fails, we stop retrying.
2177 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2183 static inline struct page *
2184 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2185 struct zonelist *zonelist, enum zone_type high_zoneidx,
2186 nodemask_t *nodemask, struct zone *preferred_zone,
2191 /* Acquire the OOM killer lock for the zones in zonelist */
2192 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2193 schedule_timeout_uninterruptible(1);
2198 * Go through the zonelist yet one more time, keep very high watermark
2199 * here, this is only to catch a parallel oom killing, we must fail if
2200 * we're still under heavy pressure.
2202 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2203 order, zonelist, high_zoneidx,
2204 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2205 preferred_zone, migratetype);
2209 if (!(gfp_mask & __GFP_NOFAIL)) {
2210 /* The OOM killer will not help higher order allocs */
2211 if (order > PAGE_ALLOC_COSTLY_ORDER)
2213 /* The OOM killer does not needlessly kill tasks for lowmem */
2214 if (high_zoneidx < ZONE_NORMAL)
2217 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2218 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2219 * The caller should handle page allocation failure by itself if
2220 * it specifies __GFP_THISNODE.
2221 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2223 if (gfp_mask & __GFP_THISNODE)
2226 /* Exhausted what can be done so it's blamo time */
2227 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2230 clear_zonelist_oom(zonelist, gfp_mask);
2234 #ifdef CONFIG_COMPACTION
2235 /* Try memory compaction for high-order allocations before reclaim */
2236 static struct page *
2237 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2238 struct zonelist *zonelist, enum zone_type high_zoneidx,
2239 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2240 int migratetype, bool sync_migration,
2241 bool *contended_compaction, bool *deferred_compaction,
2242 unsigned long *did_some_progress)
2247 if (compaction_deferred(preferred_zone, order)) {
2248 *deferred_compaction = true;
2252 current->flags |= PF_MEMALLOC;
2253 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2254 nodemask, sync_migration,
2255 contended_compaction);
2256 current->flags &= ~PF_MEMALLOC;
2258 if (*did_some_progress != COMPACT_SKIPPED) {
2261 /* Page migration frees to the PCP lists but we want merging */
2262 drain_pages(get_cpu());
2265 page = get_page_from_freelist(gfp_mask, nodemask,
2266 order, zonelist, high_zoneidx,
2267 alloc_flags & ~ALLOC_NO_WATERMARKS,
2268 preferred_zone, migratetype);
2270 preferred_zone->compact_blockskip_flush = false;
2271 compaction_defer_reset(preferred_zone, order, true);
2272 count_vm_event(COMPACTSUCCESS);
2277 * It's bad if compaction run occurs and fails.
2278 * The most likely reason is that pages exist,
2279 * but not enough to satisfy watermarks.
2281 count_vm_event(COMPACTFAIL);
2284 * As async compaction considers a subset of pageblocks, only
2285 * defer if the failure was a sync compaction failure.
2288 defer_compaction(preferred_zone, order);
2296 static inline struct page *
2297 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2298 struct zonelist *zonelist, enum zone_type high_zoneidx,
2299 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2300 int migratetype, bool sync_migration,
2301 bool *contended_compaction, bool *deferred_compaction,
2302 unsigned long *did_some_progress)
2306 #endif /* CONFIG_COMPACTION */
2308 /* Perform direct synchronous page reclaim */
2310 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2311 nodemask_t *nodemask)
2313 struct reclaim_state reclaim_state;
2318 /* We now go into synchronous reclaim */
2319 cpuset_memory_pressure_bump();
2320 current->flags |= PF_MEMALLOC;
2321 lockdep_set_current_reclaim_state(gfp_mask);
2322 reclaim_state.reclaimed_slab = 0;
2323 current->reclaim_state = &reclaim_state;
2325 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2327 current->reclaim_state = NULL;
2328 lockdep_clear_current_reclaim_state();
2329 current->flags &= ~PF_MEMALLOC;
2336 /* The really slow allocator path where we enter direct reclaim */
2337 static inline struct page *
2338 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2339 struct zonelist *zonelist, enum zone_type high_zoneidx,
2340 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2341 int migratetype, unsigned long *did_some_progress)
2343 struct page *page = NULL;
2344 bool drained = false;
2346 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2348 if (unlikely(!(*did_some_progress)))
2351 /* After successful reclaim, reconsider all zones for allocation */
2352 if (IS_ENABLED(CONFIG_NUMA))
2353 zlc_clear_zones_full(zonelist);
2356 page = get_page_from_freelist(gfp_mask, nodemask, order,
2357 zonelist, high_zoneidx,
2358 alloc_flags & ~ALLOC_NO_WATERMARKS,
2359 preferred_zone, migratetype);
2362 * If an allocation failed after direct reclaim, it could be because
2363 * pages are pinned on the per-cpu lists. Drain them and try again
2365 if (!page && !drained) {
2375 * This is called in the allocator slow-path if the allocation request is of
2376 * sufficient urgency to ignore watermarks and take other desperate measures
2378 static inline struct page *
2379 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2380 struct zonelist *zonelist, enum zone_type high_zoneidx,
2381 nodemask_t *nodemask, struct zone *preferred_zone,
2387 page = get_page_from_freelist(gfp_mask, nodemask, order,
2388 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2389 preferred_zone, migratetype);
2391 if (!page && gfp_mask & __GFP_NOFAIL)
2392 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2393 } while (!page && (gfp_mask & __GFP_NOFAIL));
2398 static void prepare_slowpath(gfp_t gfp_mask, unsigned int order,
2399 struct zonelist *zonelist,
2400 enum zone_type high_zoneidx,
2401 struct zone *preferred_zone)
2406 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2407 if (!(gfp_mask & __GFP_NO_KSWAPD))
2408 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2410 * Only reset the batches of zones that were actually
2411 * considered in the fast path, we don't want to
2412 * thrash fairness information for zones that are not
2413 * actually part of this zonelist's round-robin cycle.
2415 if (!zone_local(preferred_zone, zone))
2417 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2418 high_wmark_pages(zone) -
2419 low_wmark_pages(zone) -
2420 zone_page_state(zone, NR_ALLOC_BATCH));
2425 gfp_to_alloc_flags(gfp_t gfp_mask)
2427 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2428 const gfp_t wait = gfp_mask & __GFP_WAIT;
2430 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2431 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2434 * The caller may dip into page reserves a bit more if the caller
2435 * cannot run direct reclaim, or if the caller has realtime scheduling
2436 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2437 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2439 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2443 * Not worth trying to allocate harder for
2444 * __GFP_NOMEMALLOC even if it can't schedule.
2446 if (!(gfp_mask & __GFP_NOMEMALLOC))
2447 alloc_flags |= ALLOC_HARDER;
2449 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2450 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2452 alloc_flags &= ~ALLOC_CPUSET;
2453 } else if (unlikely(rt_task(current)) && !in_interrupt())
2454 alloc_flags |= ALLOC_HARDER;
2456 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2457 if (gfp_mask & __GFP_MEMALLOC)
2458 alloc_flags |= ALLOC_NO_WATERMARKS;
2459 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2460 alloc_flags |= ALLOC_NO_WATERMARKS;
2461 else if (!in_interrupt() &&
2462 ((current->flags & PF_MEMALLOC) ||
2463 unlikely(test_thread_flag(TIF_MEMDIE))))
2464 alloc_flags |= ALLOC_NO_WATERMARKS;
2467 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2468 alloc_flags |= ALLOC_CMA;
2473 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2475 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2478 static inline struct page *
2479 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2480 struct zonelist *zonelist, enum zone_type high_zoneidx,
2481 nodemask_t *nodemask, struct zone *preferred_zone,
2484 const gfp_t wait = gfp_mask & __GFP_WAIT;
2485 struct page *page = NULL;
2487 unsigned long pages_reclaimed = 0;
2488 unsigned long did_some_progress;
2489 bool sync_migration = false;
2490 bool deferred_compaction = false;
2491 bool contended_compaction = false;
2494 * In the slowpath, we sanity check order to avoid ever trying to
2495 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2496 * be using allocators in order of preference for an area that is
2499 if (order >= MAX_ORDER) {
2500 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2505 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2506 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2507 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2508 * using a larger set of nodes after it has established that the
2509 * allowed per node queues are empty and that nodes are
2512 if (IS_ENABLED(CONFIG_NUMA) &&
2513 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2517 prepare_slowpath(gfp_mask, order, zonelist,
2518 high_zoneidx, preferred_zone);
2521 * OK, we're below the kswapd watermark and have kicked background
2522 * reclaim. Now things get more complex, so set up alloc_flags according
2523 * to how we want to proceed.
2525 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2528 * Find the true preferred zone if the allocation is unconstrained by
2531 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2532 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2536 /* This is the last chance, in general, before the goto nopage. */
2537 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2538 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2539 preferred_zone, migratetype);
2543 /* Allocate without watermarks if the context allows */
2544 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2546 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2547 * the allocation is high priority and these type of
2548 * allocations are system rather than user orientated
2550 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2552 page = __alloc_pages_high_priority(gfp_mask, order,
2553 zonelist, high_zoneidx, nodemask,
2554 preferred_zone, migratetype);
2560 /* Atomic allocations - we can't balance anything */
2563 * All existing users of the deprecated __GFP_NOFAIL are
2564 * blockable, so warn of any new users that actually allow this
2565 * type of allocation to fail.
2567 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2571 /* Avoid recursion of direct reclaim */
2572 if (current->flags & PF_MEMALLOC)
2575 /* Avoid allocations with no watermarks from looping endlessly */
2576 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2580 * Try direct compaction. The first pass is asynchronous. Subsequent
2581 * attempts after direct reclaim are synchronous
2583 page = __alloc_pages_direct_compact(gfp_mask, order,
2584 zonelist, high_zoneidx,
2586 alloc_flags, preferred_zone,
2587 migratetype, sync_migration,
2588 &contended_compaction,
2589 &deferred_compaction,
2590 &did_some_progress);
2593 sync_migration = true;
2596 * If compaction is deferred for high-order allocations, it is because
2597 * sync compaction recently failed. In this is the case and the caller
2598 * requested a movable allocation that does not heavily disrupt the
2599 * system then fail the allocation instead of entering direct reclaim.
2601 if ((deferred_compaction || contended_compaction) &&
2602 (gfp_mask & __GFP_NO_KSWAPD))
2605 /* Try direct reclaim and then allocating */
2606 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2607 zonelist, high_zoneidx,
2609 alloc_flags, preferred_zone,
2610 migratetype, &did_some_progress);
2615 * If we failed to make any progress reclaiming, then we are
2616 * running out of options and have to consider going OOM
2618 if (!did_some_progress) {
2619 if (oom_gfp_allowed(gfp_mask)) {
2620 if (oom_killer_disabled)
2622 /* Coredumps can quickly deplete all memory reserves */
2623 if ((current->flags & PF_DUMPCORE) &&
2624 !(gfp_mask & __GFP_NOFAIL))
2626 page = __alloc_pages_may_oom(gfp_mask, order,
2627 zonelist, high_zoneidx,
2628 nodemask, preferred_zone,
2633 if (!(gfp_mask & __GFP_NOFAIL)) {
2635 * The oom killer is not called for high-order
2636 * allocations that may fail, so if no progress
2637 * is being made, there are no other options and
2638 * retrying is unlikely to help.
2640 if (order > PAGE_ALLOC_COSTLY_ORDER)
2643 * The oom killer is not called for lowmem
2644 * allocations to prevent needlessly killing
2647 if (high_zoneidx < ZONE_NORMAL)
2655 /* Check if we should retry the allocation */
2656 pages_reclaimed += did_some_progress;
2657 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2659 /* Wait for some write requests to complete then retry */
2660 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2662 /* Allocations that cannot fail must allocate from somewhere */
2663 if (gfp_mask & __GFP_NOFAIL)
2664 alloc_flags |= ALLOC_HARDER;
2669 * High-order allocations do not necessarily loop after
2670 * direct reclaim and reclaim/compaction depends on compaction
2671 * being called after reclaim so call directly if necessary
2673 page = __alloc_pages_direct_compact(gfp_mask, order,
2674 zonelist, high_zoneidx,
2676 alloc_flags, preferred_zone,
2677 migratetype, sync_migration,
2678 &contended_compaction,
2679 &deferred_compaction,
2680 &did_some_progress);
2686 warn_alloc_failed(gfp_mask, order, NULL);
2689 if (kmemcheck_enabled)
2690 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2696 * This is the 'heart' of the zoned buddy allocator.
2699 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2700 struct zonelist *zonelist, nodemask_t *nodemask)
2702 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2703 struct zone *preferred_zone;
2704 struct page *page = NULL;
2705 int migratetype = allocflags_to_migratetype(gfp_mask);
2706 unsigned int cpuset_mems_cookie;
2707 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2708 struct mem_cgroup *memcg = NULL;
2710 gfp_mask &= gfp_allowed_mask;
2712 lockdep_trace_alloc(gfp_mask);
2714 might_sleep_if(gfp_mask & __GFP_WAIT);
2716 if (should_fail_alloc_page(gfp_mask, order))
2720 * Check the zones suitable for the gfp_mask contain at least one
2721 * valid zone. It's possible to have an empty zonelist as a result
2722 * of GFP_THISNODE and a memoryless node
2724 if (unlikely(!zonelist->_zonerefs->zone))
2728 * Will only have any effect when __GFP_KMEMCG is set. This is
2729 * verified in the (always inline) callee
2731 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2735 cpuset_mems_cookie = get_mems_allowed();
2737 /* The preferred zone is used for statistics later */
2738 first_zones_zonelist(zonelist, high_zoneidx,
2739 nodemask ? : &cpuset_current_mems_allowed,
2741 if (!preferred_zone)
2745 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2746 alloc_flags |= ALLOC_CMA;
2748 /* First allocation attempt */
2749 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2750 zonelist, high_zoneidx, alloc_flags,
2751 preferred_zone, migratetype);
2752 if (unlikely(!page)) {
2754 * Runtime PM, block IO and its error handling path
2755 * can deadlock because I/O on the device might not
2758 gfp_mask = memalloc_noio_flags(gfp_mask);
2759 page = __alloc_pages_slowpath(gfp_mask, order,
2760 zonelist, high_zoneidx, nodemask,
2761 preferred_zone, migratetype);
2764 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2768 * When updating a task's mems_allowed, it is possible to race with
2769 * parallel threads in such a way that an allocation can fail while
2770 * the mask is being updated. If a page allocation is about to fail,
2771 * check if the cpuset changed during allocation and if so, retry.
2773 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2776 memcg_kmem_commit_charge(page, memcg, order);
2780 EXPORT_SYMBOL(__alloc_pages_nodemask);
2783 * Common helper functions.
2785 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2790 * __get_free_pages() returns a 32-bit address, which cannot represent
2793 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2795 page = alloc_pages(gfp_mask, order);
2798 return (unsigned long) page_address(page);
2800 EXPORT_SYMBOL(__get_free_pages);
2802 unsigned long get_zeroed_page(gfp_t gfp_mask)
2804 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2806 EXPORT_SYMBOL(get_zeroed_page);
2808 void __free_pages(struct page *page, unsigned int order)
2810 if (put_page_testzero(page)) {
2812 free_hot_cold_page(page, 0);
2814 __free_pages_ok(page, order);
2818 EXPORT_SYMBOL(__free_pages);
2820 void free_pages(unsigned long addr, unsigned int order)
2823 VM_BUG_ON(!virt_addr_valid((void *)addr));
2824 __free_pages(virt_to_page((void *)addr), order);
2828 EXPORT_SYMBOL(free_pages);
2831 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2832 * pages allocated with __GFP_KMEMCG.
2834 * Those pages are accounted to a particular memcg, embedded in the
2835 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2836 * for that information only to find out that it is NULL for users who have no
2837 * interest in that whatsoever, we provide these functions.
2839 * The caller knows better which flags it relies on.
2841 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2843 memcg_kmem_uncharge_pages(page, order);
2844 __free_pages(page, order);
2847 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2850 VM_BUG_ON(!virt_addr_valid((void *)addr));
2851 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2855 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2858 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2859 unsigned long used = addr + PAGE_ALIGN(size);
2861 split_page(virt_to_page((void *)addr), order);
2862 while (used < alloc_end) {
2867 return (void *)addr;
2871 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2872 * @size: the number of bytes to allocate
2873 * @gfp_mask: GFP flags for the allocation
2875 * This function is similar to alloc_pages(), except that it allocates the
2876 * minimum number of pages to satisfy the request. alloc_pages() can only
2877 * allocate memory in power-of-two pages.
2879 * This function is also limited by MAX_ORDER.
2881 * Memory allocated by this function must be released by free_pages_exact().
2883 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2885 unsigned int order = get_order(size);
2888 addr = __get_free_pages(gfp_mask, order);
2889 return make_alloc_exact(addr, order, size);
2891 EXPORT_SYMBOL(alloc_pages_exact);
2894 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2896 * @nid: the preferred node ID where memory should be allocated
2897 * @size: the number of bytes to allocate
2898 * @gfp_mask: GFP flags for the allocation
2900 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2902 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2905 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2907 unsigned order = get_order(size);
2908 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2911 return make_alloc_exact((unsigned long)page_address(p), order, size);
2913 EXPORT_SYMBOL(alloc_pages_exact_nid);
2916 * free_pages_exact - release memory allocated via alloc_pages_exact()
2917 * @virt: the value returned by alloc_pages_exact.
2918 * @size: size of allocation, same value as passed to alloc_pages_exact().
2920 * Release the memory allocated by a previous call to alloc_pages_exact.
2922 void free_pages_exact(void *virt, size_t size)
2924 unsigned long addr = (unsigned long)virt;
2925 unsigned long end = addr + PAGE_ALIGN(size);
2927 while (addr < end) {
2932 EXPORT_SYMBOL(free_pages_exact);
2935 * nr_free_zone_pages - count number of pages beyond high watermark
2936 * @offset: The zone index of the highest zone
2938 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2939 * high watermark within all zones at or below a given zone index. For each
2940 * zone, the number of pages is calculated as:
2941 * managed_pages - high_pages
2943 static unsigned long nr_free_zone_pages(int offset)
2948 /* Just pick one node, since fallback list is circular */
2949 unsigned long sum = 0;
2951 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2953 for_each_zone_zonelist(zone, z, zonelist, offset) {
2954 unsigned long size = zone->managed_pages;
2955 unsigned long high = high_wmark_pages(zone);
2964 * nr_free_buffer_pages - count number of pages beyond high watermark
2966 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2967 * watermark within ZONE_DMA and ZONE_NORMAL.
2969 unsigned long nr_free_buffer_pages(void)
2971 return nr_free_zone_pages(gfp_zone(GFP_USER));
2973 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2976 * nr_free_pagecache_pages - count number of pages beyond high watermark
2978 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2979 * high watermark within all zones.
2981 unsigned long nr_free_pagecache_pages(void)
2983 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2986 static inline void show_node(struct zone *zone)
2988 if (IS_ENABLED(CONFIG_NUMA))
2989 printk("Node %d ", zone_to_nid(zone));
2992 void si_meminfo(struct sysinfo *val)
2994 val->totalram = totalram_pages;
2996 val->freeram = global_page_state(NR_FREE_PAGES);
2997 val->bufferram = nr_blockdev_pages();
2998 val->totalhigh = totalhigh_pages;
2999 val->freehigh = nr_free_highpages();
3000 val->mem_unit = PAGE_SIZE;
3003 EXPORT_SYMBOL(si_meminfo);
3006 void si_meminfo_node(struct sysinfo *val, int nid)
3008 int zone_type; /* needs to be signed */
3009 unsigned long managed_pages = 0;
3010 pg_data_t *pgdat = NODE_DATA(nid);
3012 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3013 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3014 val->totalram = managed_pages;
3015 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3016 #ifdef CONFIG_HIGHMEM
3017 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3018 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3024 val->mem_unit = PAGE_SIZE;
3029 * Determine whether the node should be displayed or not, depending on whether
3030 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3032 bool skip_free_areas_node(unsigned int flags, int nid)
3035 unsigned int cpuset_mems_cookie;
3037 if (!(flags & SHOW_MEM_FILTER_NODES))
3041 cpuset_mems_cookie = get_mems_allowed();
3042 ret = !node_isset(nid, cpuset_current_mems_allowed);
3043 } while (!put_mems_allowed(cpuset_mems_cookie));
3048 #define K(x) ((x) << (PAGE_SHIFT-10))
3050 static void show_migration_types(unsigned char type)
3052 static const char types[MIGRATE_TYPES] = {
3053 [MIGRATE_UNMOVABLE] = 'U',
3054 [MIGRATE_RECLAIMABLE] = 'E',
3055 [MIGRATE_MOVABLE] = 'M',
3056 [MIGRATE_RESERVE] = 'R',
3058 [MIGRATE_CMA] = 'C',
3060 #ifdef CONFIG_MEMORY_ISOLATION
3061 [MIGRATE_ISOLATE] = 'I',
3064 char tmp[MIGRATE_TYPES + 1];
3068 for (i = 0; i < MIGRATE_TYPES; i++) {
3069 if (type & (1 << i))
3074 printk("(%s) ", tmp);
3078 * Show free area list (used inside shift_scroll-lock stuff)
3079 * We also calculate the percentage fragmentation. We do this by counting the
3080 * memory on each free list with the exception of the first item on the list.
3081 * Suppresses nodes that are not allowed by current's cpuset if
3082 * SHOW_MEM_FILTER_NODES is passed.
3084 void show_free_areas(unsigned int filter)
3089 for_each_populated_zone(zone) {
3090 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3093 printk("%s per-cpu:\n", zone->name);
3095 for_each_online_cpu(cpu) {
3096 struct per_cpu_pageset *pageset;
3098 pageset = per_cpu_ptr(zone->pageset, cpu);
3100 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3101 cpu, pageset->pcp.high,
3102 pageset->pcp.batch, pageset->pcp.count);
3106 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3107 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3109 " dirty:%lu writeback:%lu unstable:%lu\n"
3110 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3111 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3113 global_page_state(NR_ACTIVE_ANON),
3114 global_page_state(NR_INACTIVE_ANON),
3115 global_page_state(NR_ISOLATED_ANON),
3116 global_page_state(NR_ACTIVE_FILE),
3117 global_page_state(NR_INACTIVE_FILE),
3118 global_page_state(NR_ISOLATED_FILE),
3119 global_page_state(NR_UNEVICTABLE),
3120 global_page_state(NR_FILE_DIRTY),
3121 global_page_state(NR_WRITEBACK),
3122 global_page_state(NR_UNSTABLE_NFS),
3123 global_page_state(NR_FREE_PAGES),
3124 global_page_state(NR_SLAB_RECLAIMABLE),
3125 global_page_state(NR_SLAB_UNRECLAIMABLE),
3126 global_page_state(NR_FILE_MAPPED),
3127 global_page_state(NR_SHMEM),
3128 global_page_state(NR_PAGETABLE),
3129 global_page_state(NR_BOUNCE),
3130 global_page_state(NR_FREE_CMA_PAGES));
3132 for_each_populated_zone(zone) {
3135 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3143 " active_anon:%lukB"
3144 " inactive_anon:%lukB"
3145 " active_file:%lukB"
3146 " inactive_file:%lukB"
3147 " unevictable:%lukB"
3148 " isolated(anon):%lukB"
3149 " isolated(file):%lukB"
3157 " slab_reclaimable:%lukB"
3158 " slab_unreclaimable:%lukB"
3159 " kernel_stack:%lukB"
3164 " writeback_tmp:%lukB"
3165 " pages_scanned:%lu"
3166 " all_unreclaimable? %s"
3169 K(zone_page_state(zone, NR_FREE_PAGES)),
3170 K(min_wmark_pages(zone)),
3171 K(low_wmark_pages(zone)),
3172 K(high_wmark_pages(zone)),
3173 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3174 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3175 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3176 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3177 K(zone_page_state(zone, NR_UNEVICTABLE)),
3178 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3179 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3180 K(zone->present_pages),
3181 K(zone->managed_pages),
3182 K(zone_page_state(zone, NR_MLOCK)),
3183 K(zone_page_state(zone, NR_FILE_DIRTY)),
3184 K(zone_page_state(zone, NR_WRITEBACK)),
3185 K(zone_page_state(zone, NR_FILE_MAPPED)),
3186 K(zone_page_state(zone, NR_SHMEM)),
3187 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3188 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3189 zone_page_state(zone, NR_KERNEL_STACK) *
3191 K(zone_page_state(zone, NR_PAGETABLE)),
3192 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3193 K(zone_page_state(zone, NR_BOUNCE)),
3194 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3195 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3196 zone->pages_scanned,
3197 (!zone_reclaimable(zone) ? "yes" : "no")
3199 printk("lowmem_reserve[]:");
3200 for (i = 0; i < MAX_NR_ZONES; i++)
3201 printk(" %lu", zone->lowmem_reserve[i]);
3205 for_each_populated_zone(zone) {
3206 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3207 unsigned char types[MAX_ORDER];
3209 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3212 printk("%s: ", zone->name);
3214 spin_lock_irqsave(&zone->lock, flags);
3215 for (order = 0; order < MAX_ORDER; order++) {
3216 struct free_area *area = &zone->free_area[order];
3219 nr[order] = area->nr_free;
3220 total += nr[order] << order;
3223 for (type = 0; type < MIGRATE_TYPES; type++) {
3224 if (!list_empty(&area->free_list[type]))
3225 types[order] |= 1 << type;
3228 spin_unlock_irqrestore(&zone->lock, flags);
3229 for (order = 0; order < MAX_ORDER; order++) {
3230 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3232 show_migration_types(types[order]);
3234 printk("= %lukB\n", K(total));
3237 hugetlb_show_meminfo();
3239 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3241 show_swap_cache_info();
3244 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3246 zoneref->zone = zone;
3247 zoneref->zone_idx = zone_idx(zone);
3251 * Builds allocation fallback zone lists.
3253 * Add all populated zones of a node to the zonelist.
3255 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3259 enum zone_type zone_type = MAX_NR_ZONES;
3263 zone = pgdat->node_zones + zone_type;
3264 if (populated_zone(zone)) {
3265 zoneref_set_zone(zone,
3266 &zonelist->_zonerefs[nr_zones++]);
3267 check_highest_zone(zone_type);
3269 } while (zone_type);
3277 * 0 = automatic detection of better ordering.
3278 * 1 = order by ([node] distance, -zonetype)
3279 * 2 = order by (-zonetype, [node] distance)
3281 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3282 * the same zonelist. So only NUMA can configure this param.
3284 #define ZONELIST_ORDER_DEFAULT 0
3285 #define ZONELIST_ORDER_NODE 1
3286 #define ZONELIST_ORDER_ZONE 2
3288 /* zonelist order in the kernel.
3289 * set_zonelist_order() will set this to NODE or ZONE.
3291 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3292 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3296 /* The value user specified ....changed by config */
3297 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3298 /* string for sysctl */
3299 #define NUMA_ZONELIST_ORDER_LEN 16
3300 char numa_zonelist_order[16] = "default";
3303 * interface for configure zonelist ordering.
3304 * command line option "numa_zonelist_order"
3305 * = "[dD]efault - default, automatic configuration.
3306 * = "[nN]ode - order by node locality, then by zone within node
3307 * = "[zZ]one - order by zone, then by locality within zone
3310 static int __parse_numa_zonelist_order(char *s)
3312 if (*s == 'd' || *s == 'D') {
3313 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3314 } else if (*s == 'n' || *s == 'N') {
3315 user_zonelist_order = ZONELIST_ORDER_NODE;
3316 } else if (*s == 'z' || *s == 'Z') {
3317 user_zonelist_order = ZONELIST_ORDER_ZONE;
3320 "Ignoring invalid numa_zonelist_order value: "
3327 static __init int setup_numa_zonelist_order(char *s)
3334 ret = __parse_numa_zonelist_order(s);
3336 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3340 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3343 * sysctl handler for numa_zonelist_order
3345 int numa_zonelist_order_handler(ctl_table *table, int write,
3346 void __user *buffer, size_t *length,
3349 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3351 static DEFINE_MUTEX(zl_order_mutex);
3353 mutex_lock(&zl_order_mutex);
3355 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3359 strcpy(saved_string, (char *)table->data);
3361 ret = proc_dostring(table, write, buffer, length, ppos);
3365 int oldval = user_zonelist_order;
3367 ret = __parse_numa_zonelist_order((char *)table->data);
3370 * bogus value. restore saved string
3372 strncpy((char *)table->data, saved_string,
3373 NUMA_ZONELIST_ORDER_LEN);
3374 user_zonelist_order = oldval;
3375 } else if (oldval != user_zonelist_order) {
3376 mutex_lock(&zonelists_mutex);
3377 build_all_zonelists(NULL, NULL);
3378 mutex_unlock(&zonelists_mutex);
3382 mutex_unlock(&zl_order_mutex);
3387 #define MAX_NODE_LOAD (nr_online_nodes)
3388 static int node_load[MAX_NUMNODES];
3391 * find_next_best_node - find the next node that should appear in a given node's fallback list
3392 * @node: node whose fallback list we're appending
3393 * @used_node_mask: nodemask_t of already used nodes
3395 * We use a number of factors to determine which is the next node that should
3396 * appear on a given node's fallback list. The node should not have appeared
3397 * already in @node's fallback list, and it should be the next closest node
3398 * according to the distance array (which contains arbitrary distance values
3399 * from each node to each node in the system), and should also prefer nodes
3400 * with no CPUs, since presumably they'll have very little allocation pressure
3401 * on them otherwise.
3402 * It returns -1 if no node is found.
3404 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3407 int min_val = INT_MAX;
3408 int best_node = NUMA_NO_NODE;
3409 const struct cpumask *tmp = cpumask_of_node(0);
3411 /* Use the local node if we haven't already */
3412 if (!node_isset(node, *used_node_mask)) {
3413 node_set(node, *used_node_mask);
3417 for_each_node_state(n, N_MEMORY) {
3419 /* Don't want a node to appear more than once */
3420 if (node_isset(n, *used_node_mask))
3423 /* Use the distance array to find the distance */
3424 val = node_distance(node, n);
3426 /* Penalize nodes under us ("prefer the next node") */
3429 /* Give preference to headless and unused nodes */
3430 tmp = cpumask_of_node(n);
3431 if (!cpumask_empty(tmp))
3432 val += PENALTY_FOR_NODE_WITH_CPUS;
3434 /* Slight preference for less loaded node */
3435 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3436 val += node_load[n];
3438 if (val < min_val) {
3445 node_set(best_node, *used_node_mask);
3452 * Build zonelists ordered by node and zones within node.
3453 * This results in maximum locality--normal zone overflows into local
3454 * DMA zone, if any--but risks exhausting DMA zone.
3456 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3459 struct zonelist *zonelist;
3461 zonelist = &pgdat->node_zonelists[0];
3462 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3464 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3465 zonelist->_zonerefs[j].zone = NULL;
3466 zonelist->_zonerefs[j].zone_idx = 0;
3470 * Build gfp_thisnode zonelists
3472 static void build_thisnode_zonelists(pg_data_t *pgdat)
3475 struct zonelist *zonelist;
3477 zonelist = &pgdat->node_zonelists[1];
3478 j = build_zonelists_node(pgdat, zonelist, 0);
3479 zonelist->_zonerefs[j].zone = NULL;
3480 zonelist->_zonerefs[j].zone_idx = 0;
3484 * Build zonelists ordered by zone and nodes within zones.
3485 * This results in conserving DMA zone[s] until all Normal memory is
3486 * exhausted, but results in overflowing to remote node while memory
3487 * may still exist in local DMA zone.
3489 static int node_order[MAX_NUMNODES];
3491 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3494 int zone_type; /* needs to be signed */
3496 struct zonelist *zonelist;
3498 zonelist = &pgdat->node_zonelists[0];
3500 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3501 for (j = 0; j < nr_nodes; j++) {
3502 node = node_order[j];
3503 z = &NODE_DATA(node)->node_zones[zone_type];
3504 if (populated_zone(z)) {
3506 &zonelist->_zonerefs[pos++]);
3507 check_highest_zone(zone_type);
3511 zonelist->_zonerefs[pos].zone = NULL;
3512 zonelist->_zonerefs[pos].zone_idx = 0;
3515 static int default_zonelist_order(void)
3518 unsigned long low_kmem_size, total_size;
3522 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3523 * If they are really small and used heavily, the system can fall
3524 * into OOM very easily.
3525 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3527 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3530 for_each_online_node(nid) {
3531 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3532 z = &NODE_DATA(nid)->node_zones[zone_type];
3533 if (populated_zone(z)) {
3534 if (zone_type < ZONE_NORMAL)
3535 low_kmem_size += z->managed_pages;
3536 total_size += z->managed_pages;
3537 } else if (zone_type == ZONE_NORMAL) {
3539 * If any node has only lowmem, then node order
3540 * is preferred to allow kernel allocations
3541 * locally; otherwise, they can easily infringe
3542 * on other nodes when there is an abundance of
3543 * lowmem available to allocate from.
3545 return ZONELIST_ORDER_NODE;
3549 if (!low_kmem_size || /* there are no DMA area. */
3550 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3551 return ZONELIST_ORDER_NODE;
3553 * look into each node's config.
3554 * If there is a node whose DMA/DMA32 memory is very big area on
3555 * local memory, NODE_ORDER may be suitable.
3557 average_size = total_size /
3558 (nodes_weight(node_states[N_MEMORY]) + 1);
3559 for_each_online_node(nid) {
3562 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3563 z = &NODE_DATA(nid)->node_zones[zone_type];
3564 if (populated_zone(z)) {
3565 if (zone_type < ZONE_NORMAL)
3566 low_kmem_size += z->present_pages;
3567 total_size += z->present_pages;
3570 if (low_kmem_size &&
3571 total_size > average_size && /* ignore small node */
3572 low_kmem_size > total_size * 70/100)
3573 return ZONELIST_ORDER_NODE;
3575 return ZONELIST_ORDER_ZONE;
3578 static void set_zonelist_order(void)
3580 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3581 current_zonelist_order = default_zonelist_order();
3583 current_zonelist_order = user_zonelist_order;
3586 static void build_zonelists(pg_data_t *pgdat)
3590 nodemask_t used_mask;
3591 int local_node, prev_node;
3592 struct zonelist *zonelist;
3593 int order = current_zonelist_order;
3595 /* initialize zonelists */
3596 for (i = 0; i < MAX_ZONELISTS; i++) {
3597 zonelist = pgdat->node_zonelists + i;
3598 zonelist->_zonerefs[0].zone = NULL;
3599 zonelist->_zonerefs[0].zone_idx = 0;
3602 /* NUMA-aware ordering of nodes */
3603 local_node = pgdat->node_id;
3604 load = nr_online_nodes;
3605 prev_node = local_node;
3606 nodes_clear(used_mask);
3608 memset(node_order, 0, sizeof(node_order));
3611 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3613 * We don't want to pressure a particular node.
3614 * So adding penalty to the first node in same
3615 * distance group to make it round-robin.
3617 if (node_distance(local_node, node) !=
3618 node_distance(local_node, prev_node))
3619 node_load[node] = load;
3623 if (order == ZONELIST_ORDER_NODE)
3624 build_zonelists_in_node_order(pgdat, node);
3626 node_order[j++] = node; /* remember order */
3629 if (order == ZONELIST_ORDER_ZONE) {
3630 /* calculate node order -- i.e., DMA last! */
3631 build_zonelists_in_zone_order(pgdat, j);
3634 build_thisnode_zonelists(pgdat);
3637 /* Construct the zonelist performance cache - see further mmzone.h */
3638 static void build_zonelist_cache(pg_data_t *pgdat)
3640 struct zonelist *zonelist;
3641 struct zonelist_cache *zlc;
3644 zonelist = &pgdat->node_zonelists[0];
3645 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3646 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3647 for (z = zonelist->_zonerefs; z->zone; z++)
3648 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3651 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3653 * Return node id of node used for "local" allocations.
3654 * I.e., first node id of first zone in arg node's generic zonelist.
3655 * Used for initializing percpu 'numa_mem', which is used primarily
3656 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3658 int local_memory_node(int node)
3662 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3663 gfp_zone(GFP_KERNEL),
3670 #else /* CONFIG_NUMA */
3672 static void set_zonelist_order(void)
3674 current_zonelist_order = ZONELIST_ORDER_ZONE;
3677 static void build_zonelists(pg_data_t *pgdat)
3679 int node, local_node;
3681 struct zonelist *zonelist;
3683 local_node = pgdat->node_id;
3685 zonelist = &pgdat->node_zonelists[0];
3686 j = build_zonelists_node(pgdat, zonelist, 0);
3689 * Now we build the zonelist so that it contains the zones
3690 * of all the other nodes.
3691 * We don't want to pressure a particular node, so when
3692 * building the zones for node N, we make sure that the
3693 * zones coming right after the local ones are those from
3694 * node N+1 (modulo N)
3696 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3697 if (!node_online(node))
3699 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3701 for (node = 0; node < local_node; node++) {
3702 if (!node_online(node))
3704 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3707 zonelist->_zonerefs[j].zone = NULL;
3708 zonelist->_zonerefs[j].zone_idx = 0;
3711 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3712 static void build_zonelist_cache(pg_data_t *pgdat)
3714 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3717 #endif /* CONFIG_NUMA */
3720 * Boot pageset table. One per cpu which is going to be used for all
3721 * zones and all nodes. The parameters will be set in such a way
3722 * that an item put on a list will immediately be handed over to
3723 * the buddy list. This is safe since pageset manipulation is done
3724 * with interrupts disabled.
3726 * The boot_pagesets must be kept even after bootup is complete for
3727 * unused processors and/or zones. They do play a role for bootstrapping
3728 * hotplugged processors.
3730 * zoneinfo_show() and maybe other functions do
3731 * not check if the processor is online before following the pageset pointer.
3732 * Other parts of the kernel may not check if the zone is available.
3734 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3735 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3736 static void setup_zone_pageset(struct zone *zone);
3739 * Global mutex to protect against size modification of zonelists
3740 * as well as to serialize pageset setup for the new populated zone.
3742 DEFINE_MUTEX(zonelists_mutex);
3744 /* return values int ....just for stop_machine() */
3745 static int __build_all_zonelists(void *data)
3749 pg_data_t *self = data;
3752 memset(node_load, 0, sizeof(node_load));
3755 if (self && !node_online(self->node_id)) {
3756 build_zonelists(self);
3757 build_zonelist_cache(self);
3760 for_each_online_node(nid) {
3761 pg_data_t *pgdat = NODE_DATA(nid);
3763 build_zonelists(pgdat);
3764 build_zonelist_cache(pgdat);
3768 * Initialize the boot_pagesets that are going to be used
3769 * for bootstrapping processors. The real pagesets for
3770 * each zone will be allocated later when the per cpu
3771 * allocator is available.
3773 * boot_pagesets are used also for bootstrapping offline
3774 * cpus if the system is already booted because the pagesets
3775 * are needed to initialize allocators on a specific cpu too.
3776 * F.e. the percpu allocator needs the page allocator which
3777 * needs the percpu allocator in order to allocate its pagesets
3778 * (a chicken-egg dilemma).
3780 for_each_possible_cpu(cpu) {
3781 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3783 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3785 * We now know the "local memory node" for each node--
3786 * i.e., the node of the first zone in the generic zonelist.
3787 * Set up numa_mem percpu variable for on-line cpus. During
3788 * boot, only the boot cpu should be on-line; we'll init the
3789 * secondary cpus' numa_mem as they come on-line. During
3790 * node/memory hotplug, we'll fixup all on-line cpus.
3792 if (cpu_online(cpu))
3793 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3801 * Called with zonelists_mutex held always
3802 * unless system_state == SYSTEM_BOOTING.
3804 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3806 set_zonelist_order();
3808 if (system_state == SYSTEM_BOOTING) {
3809 __build_all_zonelists(NULL);
3810 mminit_verify_zonelist();
3811 cpuset_init_current_mems_allowed();
3813 #ifdef CONFIG_MEMORY_HOTPLUG
3815 setup_zone_pageset(zone);
3817 /* we have to stop all cpus to guarantee there is no user
3819 stop_machine(__build_all_zonelists, pgdat, NULL);
3820 /* cpuset refresh routine should be here */
3822 vm_total_pages = nr_free_pagecache_pages();
3824 * Disable grouping by mobility if the number of pages in the
3825 * system is too low to allow the mechanism to work. It would be
3826 * more accurate, but expensive to check per-zone. This check is
3827 * made on memory-hotadd so a system can start with mobility
3828 * disabled and enable it later
3830 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3831 page_group_by_mobility_disabled = 1;
3833 page_group_by_mobility_disabled = 0;
3835 printk("Built %i zonelists in %s order, mobility grouping %s. "
3836 "Total pages: %ld\n",
3838 zonelist_order_name[current_zonelist_order],
3839 page_group_by_mobility_disabled ? "off" : "on",
3842 printk("Policy zone: %s\n", zone_names[policy_zone]);
3847 * Helper functions to size the waitqueue hash table.
3848 * Essentially these want to choose hash table sizes sufficiently
3849 * large so that collisions trying to wait on pages are rare.
3850 * But in fact, the number of active page waitqueues on typical
3851 * systems is ridiculously low, less than 200. So this is even
3852 * conservative, even though it seems large.
3854 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3855 * waitqueues, i.e. the size of the waitq table given the number of pages.
3857 #define PAGES_PER_WAITQUEUE 256
3859 #ifndef CONFIG_MEMORY_HOTPLUG
3860 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3862 unsigned long size = 1;
3864 pages /= PAGES_PER_WAITQUEUE;
3866 while (size < pages)
3870 * Once we have dozens or even hundreds of threads sleeping
3871 * on IO we've got bigger problems than wait queue collision.
3872 * Limit the size of the wait table to a reasonable size.
3874 size = min(size, 4096UL);
3876 return max(size, 4UL);
3880 * A zone's size might be changed by hot-add, so it is not possible to determine
3881 * a suitable size for its wait_table. So we use the maximum size now.
3883 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3885 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3886 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3887 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3889 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3890 * or more by the traditional way. (See above). It equals:
3892 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3893 * ia64(16K page size) : = ( 8G + 4M)byte.
3894 * powerpc (64K page size) : = (32G +16M)byte.
3896 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3903 * This is an integer logarithm so that shifts can be used later
3904 * to extract the more random high bits from the multiplicative
3905 * hash function before the remainder is taken.
3907 static inline unsigned long wait_table_bits(unsigned long size)
3913 * Check if a pageblock contains reserved pages
3915 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3919 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3920 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3927 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3928 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3929 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3930 * higher will lead to a bigger reserve which will get freed as contiguous
3931 * blocks as reclaim kicks in
3933 static void setup_zone_migrate_reserve(struct zone *zone)
3935 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3937 unsigned long block_migratetype;
3942 * Get the start pfn, end pfn and the number of blocks to reserve
3943 * We have to be careful to be aligned to pageblock_nr_pages to
3944 * make sure that we always check pfn_valid for the first page in
3947 start_pfn = zone->zone_start_pfn;
3948 end_pfn = zone_end_pfn(zone);
3949 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3950 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3954 * Reserve blocks are generally in place to help high-order atomic
3955 * allocations that are short-lived. A min_free_kbytes value that
3956 * would result in more than 2 reserve blocks for atomic allocations
3957 * is assumed to be in place to help anti-fragmentation for the
3958 * future allocation of hugepages at runtime.
3960 reserve = min(2, reserve);
3961 old_reserve = zone->nr_migrate_reserve_block;
3963 /* When memory hot-add, we almost always need to do nothing */
3964 if (reserve == old_reserve)
3966 zone->nr_migrate_reserve_block = reserve;
3968 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3969 if (!pfn_valid(pfn))
3971 page = pfn_to_page(pfn);
3973 /* Watch out for overlapping nodes */
3974 if (page_to_nid(page) != zone_to_nid(zone))
3977 block_migratetype = get_pageblock_migratetype(page);
3979 /* Only test what is necessary when the reserves are not met */
3982 * Blocks with reserved pages will never free, skip
3985 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3986 if (pageblock_is_reserved(pfn, block_end_pfn))
3989 /* If this block is reserved, account for it */
3990 if (block_migratetype == MIGRATE_RESERVE) {
3995 /* Suitable for reserving if this block is movable */
3996 if (block_migratetype == MIGRATE_MOVABLE) {
3997 set_pageblock_migratetype(page,
3999 move_freepages_block(zone, page,
4004 } else if (!old_reserve) {
4006 * At boot time we don't need to scan the whole zone
4007 * for turning off MIGRATE_RESERVE.
4013 * If the reserve is met and this is a previous reserved block,
4016 if (block_migratetype == MIGRATE_RESERVE) {
4017 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4018 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4024 * Initially all pages are reserved - free ones are freed
4025 * up by free_all_bootmem() once the early boot process is
4026 * done. Non-atomic initialization, single-pass.
4028 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4029 unsigned long start_pfn, enum memmap_context context)
4032 unsigned long end_pfn = start_pfn + size;
4036 if (highest_memmap_pfn < end_pfn - 1)
4037 highest_memmap_pfn = end_pfn - 1;
4039 z = &NODE_DATA(nid)->node_zones[zone];
4040 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4042 * There can be holes in boot-time mem_map[]s
4043 * handed to this function. They do not
4044 * exist on hotplugged memory.
4046 if (context == MEMMAP_EARLY) {
4047 if (!early_pfn_valid(pfn))
4049 if (!early_pfn_in_nid(pfn, nid))
4052 page = pfn_to_page(pfn);
4053 set_page_links(page, zone, nid, pfn);
4054 mminit_verify_page_links(page, zone, nid, pfn);
4055 init_page_count(page);
4056 page_mapcount_reset(page);
4057 page_cpupid_reset_last(page);
4058 SetPageReserved(page);
4060 * Mark the block movable so that blocks are reserved for
4061 * movable at startup. This will force kernel allocations
4062 * to reserve their blocks rather than leaking throughout
4063 * the address space during boot when many long-lived
4064 * kernel allocations are made. Later some blocks near
4065 * the start are marked MIGRATE_RESERVE by
4066 * setup_zone_migrate_reserve()
4068 * bitmap is created for zone's valid pfn range. but memmap
4069 * can be created for invalid pages (for alignment)
4070 * check here not to call set_pageblock_migratetype() against
4073 if ((z->zone_start_pfn <= pfn)
4074 && (pfn < zone_end_pfn(z))
4075 && !(pfn & (pageblock_nr_pages - 1)))
4076 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4078 INIT_LIST_HEAD(&page->lru);
4079 #ifdef WANT_PAGE_VIRTUAL
4080 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4081 if (!is_highmem_idx(zone))
4082 set_page_address(page, __va(pfn << PAGE_SHIFT));
4087 static void __meminit zone_init_free_lists(struct zone *zone)
4090 for_each_migratetype_order(order, t) {
4091 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4092 zone->free_area[order].nr_free = 0;
4096 #ifndef __HAVE_ARCH_MEMMAP_INIT
4097 #define memmap_init(size, nid, zone, start_pfn) \
4098 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4101 static int __meminit zone_batchsize(struct zone *zone)
4107 * The per-cpu-pages pools are set to around 1000th of the
4108 * size of the zone. But no more than 1/2 of a meg.
4110 * OK, so we don't know how big the cache is. So guess.
4112 batch = zone->managed_pages / 1024;
4113 if (batch * PAGE_SIZE > 512 * 1024)
4114 batch = (512 * 1024) / PAGE_SIZE;
4115 batch /= 4; /* We effectively *= 4 below */
4120 * Clamp the batch to a 2^n - 1 value. Having a power
4121 * of 2 value was found to be more likely to have
4122 * suboptimal cache aliasing properties in some cases.
4124 * For example if 2 tasks are alternately allocating
4125 * batches of pages, one task can end up with a lot
4126 * of pages of one half of the possible page colors
4127 * and the other with pages of the other colors.
4129 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4134 /* The deferral and batching of frees should be suppressed under NOMMU
4137 * The problem is that NOMMU needs to be able to allocate large chunks
4138 * of contiguous memory as there's no hardware page translation to
4139 * assemble apparent contiguous memory from discontiguous pages.
4141 * Queueing large contiguous runs of pages for batching, however,
4142 * causes the pages to actually be freed in smaller chunks. As there
4143 * can be a significant delay between the individual batches being
4144 * recycled, this leads to the once large chunks of space being
4145 * fragmented and becoming unavailable for high-order allocations.
4152 * pcp->high and pcp->batch values are related and dependent on one another:
4153 * ->batch must never be higher then ->high.
4154 * The following function updates them in a safe manner without read side
4157 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4158 * those fields changing asynchronously (acording the the above rule).
4160 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4161 * outside of boot time (or some other assurance that no concurrent updaters
4164 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4165 unsigned long batch)
4167 /* start with a fail safe value for batch */
4171 /* Update high, then batch, in order */
4178 /* a companion to pageset_set_high() */
4179 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4181 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4184 static void pageset_init(struct per_cpu_pageset *p)
4186 struct per_cpu_pages *pcp;
4189 memset(p, 0, sizeof(*p));
4193 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4194 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4197 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4200 pageset_set_batch(p, batch);
4204 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4205 * to the value high for the pageset p.
4207 static void pageset_set_high(struct per_cpu_pageset *p,
4210 unsigned long batch = max(1UL, high / 4);
4211 if ((high / 4) > (PAGE_SHIFT * 8))
4212 batch = PAGE_SHIFT * 8;
4214 pageset_update(&p->pcp, high, batch);
4217 static void __meminit pageset_set_high_and_batch(struct zone *zone,
4218 struct per_cpu_pageset *pcp)
4220 if (percpu_pagelist_fraction)
4221 pageset_set_high(pcp,
4222 (zone->managed_pages /
4223 percpu_pagelist_fraction));
4225 pageset_set_batch(pcp, zone_batchsize(zone));
4228 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4230 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4233 pageset_set_high_and_batch(zone, pcp);
4236 static void __meminit setup_zone_pageset(struct zone *zone)
4239 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4240 for_each_possible_cpu(cpu)
4241 zone_pageset_init(zone, cpu);
4245 * Allocate per cpu pagesets and initialize them.
4246 * Before this call only boot pagesets were available.
4248 void __init setup_per_cpu_pageset(void)
4252 for_each_populated_zone(zone)
4253 setup_zone_pageset(zone);
4256 static noinline __init_refok
4257 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4263 * The per-page waitqueue mechanism uses hashed waitqueues
4266 zone->wait_table_hash_nr_entries =
4267 wait_table_hash_nr_entries(zone_size_pages);
4268 zone->wait_table_bits =
4269 wait_table_bits(zone->wait_table_hash_nr_entries);
4270 alloc_size = zone->wait_table_hash_nr_entries
4271 * sizeof(wait_queue_head_t);
4273 if (!slab_is_available()) {
4274 zone->wait_table = (wait_queue_head_t *)
4275 memblock_virt_alloc_node_nopanic(
4276 alloc_size, zone->zone_pgdat->node_id);
4279 * This case means that a zone whose size was 0 gets new memory
4280 * via memory hot-add.
4281 * But it may be the case that a new node was hot-added. In
4282 * this case vmalloc() will not be able to use this new node's
4283 * memory - this wait_table must be initialized to use this new
4284 * node itself as well.
4285 * To use this new node's memory, further consideration will be
4288 zone->wait_table = vmalloc(alloc_size);
4290 if (!zone->wait_table)
4293 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4294 init_waitqueue_head(zone->wait_table + i);
4299 static __meminit void zone_pcp_init(struct zone *zone)
4302 * per cpu subsystem is not up at this point. The following code
4303 * relies on the ability of the linker to provide the
4304 * offset of a (static) per cpu variable into the per cpu area.
4306 zone->pageset = &boot_pageset;
4308 if (populated_zone(zone))
4309 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4310 zone->name, zone->present_pages,
4311 zone_batchsize(zone));
4314 int __meminit init_currently_empty_zone(struct zone *zone,
4315 unsigned long zone_start_pfn,
4317 enum memmap_context context)
4319 struct pglist_data *pgdat = zone->zone_pgdat;
4321 ret = zone_wait_table_init(zone, size);
4324 pgdat->nr_zones = zone_idx(zone) + 1;
4326 zone->zone_start_pfn = zone_start_pfn;
4328 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4329 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4331 (unsigned long)zone_idx(zone),
4332 zone_start_pfn, (zone_start_pfn + size));
4334 zone_init_free_lists(zone);
4339 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4340 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4342 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4343 * Architectures may implement their own version but if add_active_range()
4344 * was used and there are no special requirements, this is a convenient
4347 int __meminit __early_pfn_to_nid(unsigned long pfn)
4349 unsigned long start_pfn, end_pfn;
4352 * NOTE: The following SMP-unsafe globals are only used early in boot
4353 * when the kernel is running single-threaded.
4355 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4356 static int __meminitdata last_nid;
4358 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4361 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4363 last_start_pfn = start_pfn;
4364 last_end_pfn = end_pfn;
4370 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4372 int __meminit early_pfn_to_nid(unsigned long pfn)
4376 nid = __early_pfn_to_nid(pfn);
4379 /* just returns 0 */
4383 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4384 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4388 nid = __early_pfn_to_nid(pfn);
4389 if (nid >= 0 && nid != node)
4396 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4397 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4398 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4400 * If an architecture guarantees that all ranges registered with
4401 * add_active_ranges() contain no holes and may be freed, this
4402 * this function may be used instead of calling memblock_free_early_nid()
4405 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4407 unsigned long start_pfn, end_pfn;
4410 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4411 start_pfn = min(start_pfn, max_low_pfn);
4412 end_pfn = min(end_pfn, max_low_pfn);
4414 if (start_pfn < end_pfn)
4415 memblock_free_early_nid(PFN_PHYS(start_pfn),
4416 (end_pfn - start_pfn) << PAGE_SHIFT,
4422 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4423 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4425 * If an architecture guarantees that all ranges registered with
4426 * add_active_ranges() contain no holes and may be freed, this
4427 * function may be used instead of calling memory_present() manually.
4429 void __init sparse_memory_present_with_active_regions(int nid)
4431 unsigned long start_pfn, end_pfn;
4434 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4435 memory_present(this_nid, start_pfn, end_pfn);
4439 * get_pfn_range_for_nid - Return the start and end page frames for a node
4440 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4441 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4442 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4444 * It returns the start and end page frame of a node based on information
4445 * provided by an arch calling add_active_range(). If called for a node
4446 * with no available memory, a warning is printed and the start and end
4449 void __meminit get_pfn_range_for_nid(unsigned int nid,
4450 unsigned long *start_pfn, unsigned long *end_pfn)
4452 unsigned long this_start_pfn, this_end_pfn;
4458 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4459 *start_pfn = min(*start_pfn, this_start_pfn);
4460 *end_pfn = max(*end_pfn, this_end_pfn);
4463 if (*start_pfn == -1UL)
4468 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4469 * assumption is made that zones within a node are ordered in monotonic
4470 * increasing memory addresses so that the "highest" populated zone is used
4472 static void __init find_usable_zone_for_movable(void)
4475 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4476 if (zone_index == ZONE_MOVABLE)
4479 if (arch_zone_highest_possible_pfn[zone_index] >
4480 arch_zone_lowest_possible_pfn[zone_index])
4484 VM_BUG_ON(zone_index == -1);
4485 movable_zone = zone_index;
4489 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4490 * because it is sized independent of architecture. Unlike the other zones,
4491 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4492 * in each node depending on the size of each node and how evenly kernelcore
4493 * is distributed. This helper function adjusts the zone ranges
4494 * provided by the architecture for a given node by using the end of the
4495 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4496 * zones within a node are in order of monotonic increases memory addresses
4498 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4499 unsigned long zone_type,
4500 unsigned long node_start_pfn,
4501 unsigned long node_end_pfn,
4502 unsigned long *zone_start_pfn,
4503 unsigned long *zone_end_pfn)
4505 /* Only adjust if ZONE_MOVABLE is on this node */
4506 if (zone_movable_pfn[nid]) {
4507 /* Size ZONE_MOVABLE */
4508 if (zone_type == ZONE_MOVABLE) {
4509 *zone_start_pfn = zone_movable_pfn[nid];
4510 *zone_end_pfn = min(node_end_pfn,
4511 arch_zone_highest_possible_pfn[movable_zone]);
4513 /* Adjust for ZONE_MOVABLE starting within this range */
4514 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4515 *zone_end_pfn > zone_movable_pfn[nid]) {
4516 *zone_end_pfn = zone_movable_pfn[nid];
4518 /* Check if this whole range is within ZONE_MOVABLE */
4519 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4520 *zone_start_pfn = *zone_end_pfn;
4525 * Return the number of pages a zone spans in a node, including holes
4526 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4528 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4529 unsigned long zone_type,
4530 unsigned long node_start_pfn,
4531 unsigned long node_end_pfn,
4532 unsigned long *ignored)
4534 unsigned long zone_start_pfn, zone_end_pfn;
4536 /* Get the start and end of the zone */
4537 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4538 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4539 adjust_zone_range_for_zone_movable(nid, zone_type,
4540 node_start_pfn, node_end_pfn,
4541 &zone_start_pfn, &zone_end_pfn);
4543 /* Check that this node has pages within the zone's required range */
4544 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4547 /* Move the zone boundaries inside the node if necessary */
4548 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4549 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4551 /* Return the spanned pages */
4552 return zone_end_pfn - zone_start_pfn;
4556 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4557 * then all holes in the requested range will be accounted for.
4559 unsigned long __meminit __absent_pages_in_range(int nid,
4560 unsigned long range_start_pfn,
4561 unsigned long range_end_pfn)
4563 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4564 unsigned long start_pfn, end_pfn;
4567 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4568 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4569 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4570 nr_absent -= end_pfn - start_pfn;
4576 * absent_pages_in_range - Return number of page frames in holes within a range
4577 * @start_pfn: The start PFN to start searching for holes
4578 * @end_pfn: The end PFN to stop searching for holes
4580 * It returns the number of pages frames in memory holes within a range.
4582 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4583 unsigned long end_pfn)
4585 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4588 /* Return the number of page frames in holes in a zone on a node */
4589 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4590 unsigned long zone_type,
4591 unsigned long node_start_pfn,
4592 unsigned long node_end_pfn,
4593 unsigned long *ignored)
4595 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4596 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4597 unsigned long zone_start_pfn, zone_end_pfn;
4599 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4600 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4602 adjust_zone_range_for_zone_movable(nid, zone_type,
4603 node_start_pfn, node_end_pfn,
4604 &zone_start_pfn, &zone_end_pfn);
4605 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4608 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4609 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4610 unsigned long zone_type,
4611 unsigned long node_start_pfn,
4612 unsigned long node_end_pfn,
4613 unsigned long *zones_size)
4615 return zones_size[zone_type];
4618 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4619 unsigned long zone_type,
4620 unsigned long node_start_pfn,
4621 unsigned long node_end_pfn,
4622 unsigned long *zholes_size)
4627 return zholes_size[zone_type];
4630 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4632 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4633 unsigned long node_start_pfn,
4634 unsigned long node_end_pfn,
4635 unsigned long *zones_size,
4636 unsigned long *zholes_size)
4638 unsigned long realtotalpages, totalpages = 0;
4641 for (i = 0; i < MAX_NR_ZONES; i++)
4642 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4646 pgdat->node_spanned_pages = totalpages;
4648 realtotalpages = totalpages;
4649 for (i = 0; i < MAX_NR_ZONES; i++)
4651 zone_absent_pages_in_node(pgdat->node_id, i,
4652 node_start_pfn, node_end_pfn,
4654 pgdat->node_present_pages = realtotalpages;
4655 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4659 #ifndef CONFIG_SPARSEMEM
4661 * Calculate the size of the zone->blockflags rounded to an unsigned long
4662 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4663 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4664 * round what is now in bits to nearest long in bits, then return it in
4667 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4669 unsigned long usemapsize;
4671 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4672 usemapsize = roundup(zonesize, pageblock_nr_pages);
4673 usemapsize = usemapsize >> pageblock_order;
4674 usemapsize *= NR_PAGEBLOCK_BITS;
4675 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4677 return usemapsize / 8;
4680 static void __init setup_usemap(struct pglist_data *pgdat,
4682 unsigned long zone_start_pfn,
4683 unsigned long zonesize)
4685 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4686 zone->pageblock_flags = NULL;
4688 zone->pageblock_flags =
4689 memblock_virt_alloc_node_nopanic(usemapsize,
4693 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4694 unsigned long zone_start_pfn, unsigned long zonesize) {}
4695 #endif /* CONFIG_SPARSEMEM */
4697 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4699 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4700 void __paginginit set_pageblock_order(void)
4704 /* Check that pageblock_nr_pages has not already been setup */
4705 if (pageblock_order)
4708 if (HPAGE_SHIFT > PAGE_SHIFT)
4709 order = HUGETLB_PAGE_ORDER;
4711 order = MAX_ORDER - 1;
4714 * Assume the largest contiguous order of interest is a huge page.
4715 * This value may be variable depending on boot parameters on IA64 and
4718 pageblock_order = order;
4720 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4723 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4724 * is unused as pageblock_order is set at compile-time. See
4725 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4728 void __paginginit set_pageblock_order(void)
4732 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4734 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4735 unsigned long present_pages)
4737 unsigned long pages = spanned_pages;
4740 * Provide a more accurate estimation if there are holes within
4741 * the zone and SPARSEMEM is in use. If there are holes within the
4742 * zone, each populated memory region may cost us one or two extra
4743 * memmap pages due to alignment because memmap pages for each
4744 * populated regions may not naturally algined on page boundary.
4745 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4747 if (spanned_pages > present_pages + (present_pages >> 4) &&
4748 IS_ENABLED(CONFIG_SPARSEMEM))
4749 pages = present_pages;
4751 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4755 * Set up the zone data structures:
4756 * - mark all pages reserved
4757 * - mark all memory queues empty
4758 * - clear the memory bitmaps
4760 * NOTE: pgdat should get zeroed by caller.
4762 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4763 unsigned long node_start_pfn, unsigned long node_end_pfn,
4764 unsigned long *zones_size, unsigned long *zholes_size)
4767 int nid = pgdat->node_id;
4768 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4771 pgdat_resize_init(pgdat);
4772 #ifdef CONFIG_NUMA_BALANCING
4773 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4774 pgdat->numabalancing_migrate_nr_pages = 0;
4775 pgdat->numabalancing_migrate_next_window = jiffies;
4777 init_waitqueue_head(&pgdat->kswapd_wait);
4778 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4779 pgdat_page_cgroup_init(pgdat);
4781 for (j = 0; j < MAX_NR_ZONES; j++) {
4782 struct zone *zone = pgdat->node_zones + j;
4783 unsigned long size, realsize, freesize, memmap_pages;
4785 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4786 node_end_pfn, zones_size);
4787 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4793 * Adjust freesize so that it accounts for how much memory
4794 * is used by this zone for memmap. This affects the watermark
4795 * and per-cpu initialisations
4797 memmap_pages = calc_memmap_size(size, realsize);
4798 if (freesize >= memmap_pages) {
4799 freesize -= memmap_pages;
4802 " %s zone: %lu pages used for memmap\n",
4803 zone_names[j], memmap_pages);
4806 " %s zone: %lu pages exceeds freesize %lu\n",
4807 zone_names[j], memmap_pages, freesize);
4809 /* Account for reserved pages */
4810 if (j == 0 && freesize > dma_reserve) {
4811 freesize -= dma_reserve;
4812 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4813 zone_names[0], dma_reserve);
4816 if (!is_highmem_idx(j))
4817 nr_kernel_pages += freesize;
4818 /* Charge for highmem memmap if there are enough kernel pages */
4819 else if (nr_kernel_pages > memmap_pages * 2)
4820 nr_kernel_pages -= memmap_pages;
4821 nr_all_pages += freesize;
4823 zone->spanned_pages = size;
4824 zone->present_pages = realsize;
4826 * Set an approximate value for lowmem here, it will be adjusted
4827 * when the bootmem allocator frees pages into the buddy system.
4828 * And all highmem pages will be managed by the buddy system.
4830 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4833 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4835 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4837 zone->name = zone_names[j];
4838 spin_lock_init(&zone->lock);
4839 spin_lock_init(&zone->lru_lock);
4840 zone_seqlock_init(zone);
4841 zone->zone_pgdat = pgdat;
4842 zone_pcp_init(zone);
4844 /* For bootup, initialized properly in watermark setup */
4845 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4847 lruvec_init(&zone->lruvec);
4851 set_pageblock_order();
4852 setup_usemap(pgdat, zone, zone_start_pfn, size);
4853 ret = init_currently_empty_zone(zone, zone_start_pfn,
4854 size, MEMMAP_EARLY);
4856 memmap_init(size, nid, j, zone_start_pfn);
4857 zone_start_pfn += size;
4861 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4863 /* Skip empty nodes */
4864 if (!pgdat->node_spanned_pages)
4867 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4868 /* ia64 gets its own node_mem_map, before this, without bootmem */
4869 if (!pgdat->node_mem_map) {
4870 unsigned long size, start, end;
4874 * The zone's endpoints aren't required to be MAX_ORDER
4875 * aligned but the node_mem_map endpoints must be in order
4876 * for the buddy allocator to function correctly.
4878 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4879 end = pgdat_end_pfn(pgdat);
4880 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4881 size = (end - start) * sizeof(struct page);
4882 map = alloc_remap(pgdat->node_id, size);
4884 map = memblock_virt_alloc_node_nopanic(size,
4886 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4888 #ifndef CONFIG_NEED_MULTIPLE_NODES
4890 * With no DISCONTIG, the global mem_map is just set as node 0's
4892 if (pgdat == NODE_DATA(0)) {
4893 mem_map = NODE_DATA(0)->node_mem_map;
4894 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4895 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4896 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4897 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4900 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4903 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4904 unsigned long node_start_pfn, unsigned long *zholes_size)
4906 pg_data_t *pgdat = NODE_DATA(nid);
4907 unsigned long start_pfn = 0;
4908 unsigned long end_pfn = 0;
4910 /* pg_data_t should be reset to zero when it's allocated */
4911 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4913 pgdat->node_id = nid;
4914 pgdat->node_start_pfn = node_start_pfn;
4915 init_zone_allows_reclaim(nid);
4916 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4917 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4919 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4920 zones_size, zholes_size);
4922 alloc_node_mem_map(pgdat);
4923 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4924 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4925 nid, (unsigned long)pgdat,
4926 (unsigned long)pgdat->node_mem_map);
4929 free_area_init_core(pgdat, start_pfn, end_pfn,
4930 zones_size, zholes_size);
4933 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4935 #if MAX_NUMNODES > 1
4937 * Figure out the number of possible node ids.
4939 void __init setup_nr_node_ids(void)
4942 unsigned int highest = 0;
4944 for_each_node_mask(node, node_possible_map)
4946 nr_node_ids = highest + 1;
4951 * node_map_pfn_alignment - determine the maximum internode alignment
4953 * This function should be called after node map is populated and sorted.
4954 * It calculates the maximum power of two alignment which can distinguish
4957 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4958 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4959 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4960 * shifted, 1GiB is enough and this function will indicate so.
4962 * This is used to test whether pfn -> nid mapping of the chosen memory
4963 * model has fine enough granularity to avoid incorrect mapping for the
4964 * populated node map.
4966 * Returns the determined alignment in pfn's. 0 if there is no alignment
4967 * requirement (single node).
4969 unsigned long __init node_map_pfn_alignment(void)
4971 unsigned long accl_mask = 0, last_end = 0;
4972 unsigned long start, end, mask;
4976 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4977 if (!start || last_nid < 0 || last_nid == nid) {
4984 * Start with a mask granular enough to pin-point to the
4985 * start pfn and tick off bits one-by-one until it becomes
4986 * too coarse to separate the current node from the last.
4988 mask = ~((1 << __ffs(start)) - 1);
4989 while (mask && last_end <= (start & (mask << 1)))
4992 /* accumulate all internode masks */
4996 /* convert mask to number of pages */
4997 return ~accl_mask + 1;
5000 /* Find the lowest pfn for a node */
5001 static unsigned long __init find_min_pfn_for_node(int nid)
5003 unsigned long min_pfn = ULONG_MAX;
5004 unsigned long start_pfn;
5007 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5008 min_pfn = min(min_pfn, start_pfn);
5010 if (min_pfn == ULONG_MAX) {
5012 "Could not find start_pfn for node %d\n", nid);
5020 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5022 * It returns the minimum PFN based on information provided via
5023 * add_active_range().
5025 unsigned long __init find_min_pfn_with_active_regions(void)
5027 return find_min_pfn_for_node(MAX_NUMNODES);
5031 * early_calculate_totalpages()
5032 * Sum pages in active regions for movable zone.
5033 * Populate N_MEMORY for calculating usable_nodes.
5035 static unsigned long __init early_calculate_totalpages(void)
5037 unsigned long totalpages = 0;
5038 unsigned long start_pfn, end_pfn;
5041 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5042 unsigned long pages = end_pfn - start_pfn;
5044 totalpages += pages;
5046 node_set_state(nid, N_MEMORY);
5052 * Find the PFN the Movable zone begins in each node. Kernel memory
5053 * is spread evenly between nodes as long as the nodes have enough
5054 * memory. When they don't, some nodes will have more kernelcore than
5057 static void __init find_zone_movable_pfns_for_nodes(void)
5060 unsigned long usable_startpfn;
5061 unsigned long kernelcore_node, kernelcore_remaining;
5062 /* save the state before borrow the nodemask */
5063 nodemask_t saved_node_state = node_states[N_MEMORY];
5064 unsigned long totalpages = early_calculate_totalpages();
5065 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5066 struct memblock_type *type = &memblock.memory;
5068 /* Need to find movable_zone earlier when movable_node is specified. */
5069 find_usable_zone_for_movable();
5072 * If movable_node is specified, ignore kernelcore and movablecore
5075 if (movable_node_is_enabled()) {
5076 for (i = 0; i < type->cnt; i++) {
5077 if (!memblock_is_hotpluggable(&type->regions[i]))
5080 nid = type->regions[i].nid;
5082 usable_startpfn = PFN_DOWN(type->regions[i].base);
5083 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5084 min(usable_startpfn, zone_movable_pfn[nid]) :
5092 * If movablecore=nn[KMG] was specified, calculate what size of
5093 * kernelcore that corresponds so that memory usable for
5094 * any allocation type is evenly spread. If both kernelcore
5095 * and movablecore are specified, then the value of kernelcore
5096 * will be used for required_kernelcore if it's greater than
5097 * what movablecore would have allowed.
5099 if (required_movablecore) {
5100 unsigned long corepages;
5103 * Round-up so that ZONE_MOVABLE is at least as large as what
5104 * was requested by the user
5106 required_movablecore =
5107 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5108 corepages = totalpages - required_movablecore;
5110 required_kernelcore = max(required_kernelcore, corepages);
5113 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5114 if (!required_kernelcore)
5117 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5118 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5121 /* Spread kernelcore memory as evenly as possible throughout nodes */
5122 kernelcore_node = required_kernelcore / usable_nodes;
5123 for_each_node_state(nid, N_MEMORY) {
5124 unsigned long start_pfn, end_pfn;
5127 * Recalculate kernelcore_node if the division per node
5128 * now exceeds what is necessary to satisfy the requested
5129 * amount of memory for the kernel
5131 if (required_kernelcore < kernelcore_node)
5132 kernelcore_node = required_kernelcore / usable_nodes;
5135 * As the map is walked, we track how much memory is usable
5136 * by the kernel using kernelcore_remaining. When it is
5137 * 0, the rest of the node is usable by ZONE_MOVABLE
5139 kernelcore_remaining = kernelcore_node;
5141 /* Go through each range of PFNs within this node */
5142 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5143 unsigned long size_pages;
5145 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5146 if (start_pfn >= end_pfn)
5149 /* Account for what is only usable for kernelcore */
5150 if (start_pfn < usable_startpfn) {
5151 unsigned long kernel_pages;
5152 kernel_pages = min(end_pfn, usable_startpfn)
5155 kernelcore_remaining -= min(kernel_pages,
5156 kernelcore_remaining);
5157 required_kernelcore -= min(kernel_pages,
5158 required_kernelcore);
5160 /* Continue if range is now fully accounted */
5161 if (end_pfn <= usable_startpfn) {
5164 * Push zone_movable_pfn to the end so
5165 * that if we have to rebalance
5166 * kernelcore across nodes, we will
5167 * not double account here
5169 zone_movable_pfn[nid] = end_pfn;
5172 start_pfn = usable_startpfn;
5176 * The usable PFN range for ZONE_MOVABLE is from
5177 * start_pfn->end_pfn. Calculate size_pages as the
5178 * number of pages used as kernelcore
5180 size_pages = end_pfn - start_pfn;
5181 if (size_pages > kernelcore_remaining)
5182 size_pages = kernelcore_remaining;
5183 zone_movable_pfn[nid] = start_pfn + size_pages;
5186 * Some kernelcore has been met, update counts and
5187 * break if the kernelcore for this node has been
5190 required_kernelcore -= min(required_kernelcore,
5192 kernelcore_remaining -= size_pages;
5193 if (!kernelcore_remaining)
5199 * If there is still required_kernelcore, we do another pass with one
5200 * less node in the count. This will push zone_movable_pfn[nid] further
5201 * along on the nodes that still have memory until kernelcore is
5205 if (usable_nodes && required_kernelcore > usable_nodes)
5209 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5210 for (nid = 0; nid < MAX_NUMNODES; nid++)
5211 zone_movable_pfn[nid] =
5212 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5215 /* restore the node_state */
5216 node_states[N_MEMORY] = saved_node_state;
5219 /* Any regular or high memory on that node ? */
5220 static void check_for_memory(pg_data_t *pgdat, int nid)
5222 enum zone_type zone_type;
5224 if (N_MEMORY == N_NORMAL_MEMORY)
5227 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5228 struct zone *zone = &pgdat->node_zones[zone_type];
5229 if (populated_zone(zone)) {
5230 node_set_state(nid, N_HIGH_MEMORY);
5231 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5232 zone_type <= ZONE_NORMAL)
5233 node_set_state(nid, N_NORMAL_MEMORY);
5240 * free_area_init_nodes - Initialise all pg_data_t and zone data
5241 * @max_zone_pfn: an array of max PFNs for each zone
5243 * This will call free_area_init_node() for each active node in the system.
5244 * Using the page ranges provided by add_active_range(), the size of each
5245 * zone in each node and their holes is calculated. If the maximum PFN
5246 * between two adjacent zones match, it is assumed that the zone is empty.
5247 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5248 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5249 * starts where the previous one ended. For example, ZONE_DMA32 starts
5250 * at arch_max_dma_pfn.
5252 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5254 unsigned long start_pfn, end_pfn;
5257 /* Record where the zone boundaries are */
5258 memset(arch_zone_lowest_possible_pfn, 0,
5259 sizeof(arch_zone_lowest_possible_pfn));
5260 memset(arch_zone_highest_possible_pfn, 0,
5261 sizeof(arch_zone_highest_possible_pfn));
5262 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5263 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5264 for (i = 1; i < MAX_NR_ZONES; i++) {
5265 if (i == ZONE_MOVABLE)
5267 arch_zone_lowest_possible_pfn[i] =
5268 arch_zone_highest_possible_pfn[i-1];
5269 arch_zone_highest_possible_pfn[i] =
5270 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5272 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5273 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5275 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5276 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5277 find_zone_movable_pfns_for_nodes();
5279 /* Print out the zone ranges */
5280 printk("Zone ranges:\n");
5281 for (i = 0; i < MAX_NR_ZONES; i++) {
5282 if (i == ZONE_MOVABLE)
5284 printk(KERN_CONT " %-8s ", zone_names[i]);
5285 if (arch_zone_lowest_possible_pfn[i] ==
5286 arch_zone_highest_possible_pfn[i])
5287 printk(KERN_CONT "empty\n");
5289 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5290 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5291 (arch_zone_highest_possible_pfn[i]
5292 << PAGE_SHIFT) - 1);
5295 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5296 printk("Movable zone start for each node\n");
5297 for (i = 0; i < MAX_NUMNODES; i++) {
5298 if (zone_movable_pfn[i])
5299 printk(" Node %d: %#010lx\n", i,
5300 zone_movable_pfn[i] << PAGE_SHIFT);
5303 /* Print out the early node map */
5304 printk("Early memory node ranges\n");
5305 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5306 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5307 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5309 /* Initialise every node */
5310 mminit_verify_pageflags_layout();
5311 setup_nr_node_ids();
5312 for_each_online_node(nid) {
5313 pg_data_t *pgdat = NODE_DATA(nid);
5314 free_area_init_node(nid, NULL,
5315 find_min_pfn_for_node(nid), NULL);
5317 /* Any memory on that node */
5318 if (pgdat->node_present_pages)
5319 node_set_state(nid, N_MEMORY);
5320 check_for_memory(pgdat, nid);
5324 static int __init cmdline_parse_core(char *p, unsigned long *core)
5326 unsigned long long coremem;
5330 coremem = memparse(p, &p);
5331 *core = coremem >> PAGE_SHIFT;
5333 /* Paranoid check that UL is enough for the coremem value */
5334 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5340 * kernelcore=size sets the amount of memory for use for allocations that
5341 * cannot be reclaimed or migrated.
5343 static int __init cmdline_parse_kernelcore(char *p)
5345 return cmdline_parse_core(p, &required_kernelcore);
5349 * movablecore=size sets the amount of memory for use for allocations that
5350 * can be reclaimed or migrated.
5352 static int __init cmdline_parse_movablecore(char *p)
5354 return cmdline_parse_core(p, &required_movablecore);
5357 early_param("kernelcore", cmdline_parse_kernelcore);
5358 early_param("movablecore", cmdline_parse_movablecore);
5360 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5362 void adjust_managed_page_count(struct page *page, long count)
5364 spin_lock(&managed_page_count_lock);
5365 page_zone(page)->managed_pages += count;
5366 totalram_pages += count;
5367 #ifdef CONFIG_HIGHMEM
5368 if (PageHighMem(page))
5369 totalhigh_pages += count;
5371 spin_unlock(&managed_page_count_lock);
5373 EXPORT_SYMBOL(adjust_managed_page_count);
5375 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5378 unsigned long pages = 0;
5380 start = (void *)PAGE_ALIGN((unsigned long)start);
5381 end = (void *)((unsigned long)end & PAGE_MASK);
5382 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5383 if ((unsigned int)poison <= 0xFF)
5384 memset(pos, poison, PAGE_SIZE);
5385 free_reserved_page(virt_to_page(pos));
5389 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5390 s, pages << (PAGE_SHIFT - 10), start, end);
5394 EXPORT_SYMBOL(free_reserved_area);
5396 #ifdef CONFIG_HIGHMEM
5397 void free_highmem_page(struct page *page)
5399 __free_reserved_page(page);
5401 page_zone(page)->managed_pages++;
5407 void __init mem_init_print_info(const char *str)
5409 unsigned long physpages, codesize, datasize, rosize, bss_size;
5410 unsigned long init_code_size, init_data_size;
5412 physpages = get_num_physpages();
5413 codesize = _etext - _stext;
5414 datasize = _edata - _sdata;
5415 rosize = __end_rodata - __start_rodata;
5416 bss_size = __bss_stop - __bss_start;
5417 init_data_size = __init_end - __init_begin;
5418 init_code_size = _einittext - _sinittext;
5421 * Detect special cases and adjust section sizes accordingly:
5422 * 1) .init.* may be embedded into .data sections
5423 * 2) .init.text.* may be out of [__init_begin, __init_end],
5424 * please refer to arch/tile/kernel/vmlinux.lds.S.
5425 * 3) .rodata.* may be embedded into .text or .data sections.
5427 #define adj_init_size(start, end, size, pos, adj) \
5429 if (start <= pos && pos < end && size > adj) \
5433 adj_init_size(__init_begin, __init_end, init_data_size,
5434 _sinittext, init_code_size);
5435 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5436 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5437 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5438 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5440 #undef adj_init_size
5442 printk("Memory: %luK/%luK available "
5443 "(%luK kernel code, %luK rwdata, %luK rodata, "
5444 "%luK init, %luK bss, %luK reserved"
5445 #ifdef CONFIG_HIGHMEM
5449 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5450 codesize >> 10, datasize >> 10, rosize >> 10,
5451 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5452 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5453 #ifdef CONFIG_HIGHMEM
5454 totalhigh_pages << (PAGE_SHIFT-10),
5456 str ? ", " : "", str ? str : "");
5460 * set_dma_reserve - set the specified number of pages reserved in the first zone
5461 * @new_dma_reserve: The number of pages to mark reserved
5463 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5464 * In the DMA zone, a significant percentage may be consumed by kernel image
5465 * and other unfreeable allocations which can skew the watermarks badly. This
5466 * function may optionally be used to account for unfreeable pages in the
5467 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5468 * smaller per-cpu batchsize.
5470 void __init set_dma_reserve(unsigned long new_dma_reserve)
5472 dma_reserve = new_dma_reserve;
5475 void __init free_area_init(unsigned long *zones_size)
5477 free_area_init_node(0, zones_size,
5478 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5481 static int page_alloc_cpu_notify(struct notifier_block *self,
5482 unsigned long action, void *hcpu)
5484 int cpu = (unsigned long)hcpu;
5486 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5487 lru_add_drain_cpu(cpu);
5491 * Spill the event counters of the dead processor
5492 * into the current processors event counters.
5493 * This artificially elevates the count of the current
5496 vm_events_fold_cpu(cpu);
5499 * Zero the differential counters of the dead processor
5500 * so that the vm statistics are consistent.
5502 * This is only okay since the processor is dead and cannot
5503 * race with what we are doing.
5505 cpu_vm_stats_fold(cpu);
5510 void __init page_alloc_init(void)
5512 hotcpu_notifier(page_alloc_cpu_notify, 0);
5516 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5517 * or min_free_kbytes changes.
5519 static void calculate_totalreserve_pages(void)
5521 struct pglist_data *pgdat;
5522 unsigned long reserve_pages = 0;
5523 enum zone_type i, j;
5525 for_each_online_pgdat(pgdat) {
5526 for (i = 0; i < MAX_NR_ZONES; i++) {
5527 struct zone *zone = pgdat->node_zones + i;
5528 unsigned long max = 0;
5530 /* Find valid and maximum lowmem_reserve in the zone */
5531 for (j = i; j < MAX_NR_ZONES; j++) {
5532 if (zone->lowmem_reserve[j] > max)
5533 max = zone->lowmem_reserve[j];
5536 /* we treat the high watermark as reserved pages. */
5537 max += high_wmark_pages(zone);
5539 if (max > zone->managed_pages)
5540 max = zone->managed_pages;
5541 reserve_pages += max;
5543 * Lowmem reserves are not available to
5544 * GFP_HIGHUSER page cache allocations and
5545 * kswapd tries to balance zones to their high
5546 * watermark. As a result, neither should be
5547 * regarded as dirtyable memory, to prevent a
5548 * situation where reclaim has to clean pages
5549 * in order to balance the zones.
5551 zone->dirty_balance_reserve = max;
5554 dirty_balance_reserve = reserve_pages;
5555 totalreserve_pages = reserve_pages;
5559 * setup_per_zone_lowmem_reserve - called whenever
5560 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5561 * has a correct pages reserved value, so an adequate number of
5562 * pages are left in the zone after a successful __alloc_pages().
5564 static void setup_per_zone_lowmem_reserve(void)
5566 struct pglist_data *pgdat;
5567 enum zone_type j, idx;
5569 for_each_online_pgdat(pgdat) {
5570 for (j = 0; j < MAX_NR_ZONES; j++) {
5571 struct zone *zone = pgdat->node_zones + j;
5572 unsigned long managed_pages = zone->managed_pages;
5574 zone->lowmem_reserve[j] = 0;
5578 struct zone *lower_zone;
5582 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5583 sysctl_lowmem_reserve_ratio[idx] = 1;
5585 lower_zone = pgdat->node_zones + idx;
5586 lower_zone->lowmem_reserve[j] = managed_pages /
5587 sysctl_lowmem_reserve_ratio[idx];
5588 managed_pages += lower_zone->managed_pages;
5593 /* update totalreserve_pages */
5594 calculate_totalreserve_pages();
5597 static void __setup_per_zone_wmarks(void)
5599 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5600 unsigned long lowmem_pages = 0;
5602 unsigned long flags;
5604 /* Calculate total number of !ZONE_HIGHMEM pages */
5605 for_each_zone(zone) {
5606 if (!is_highmem(zone))
5607 lowmem_pages += zone->managed_pages;
5610 for_each_zone(zone) {
5613 spin_lock_irqsave(&zone->lock, flags);
5614 tmp = (u64)pages_min * zone->managed_pages;
5615 do_div(tmp, lowmem_pages);
5616 if (is_highmem(zone)) {
5618 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5619 * need highmem pages, so cap pages_min to a small
5622 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5623 * deltas controls asynch page reclaim, and so should
5624 * not be capped for highmem.
5626 unsigned long min_pages;
5628 min_pages = zone->managed_pages / 1024;
5629 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5630 zone->watermark[WMARK_MIN] = min_pages;
5633 * If it's a lowmem zone, reserve a number of pages
5634 * proportionate to the zone's size.
5636 zone->watermark[WMARK_MIN] = tmp;
5639 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5640 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5642 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5643 high_wmark_pages(zone) -
5644 low_wmark_pages(zone) -
5645 zone_page_state(zone, NR_ALLOC_BATCH));
5647 setup_zone_migrate_reserve(zone);
5648 spin_unlock_irqrestore(&zone->lock, flags);
5651 /* update totalreserve_pages */
5652 calculate_totalreserve_pages();
5656 * setup_per_zone_wmarks - called when min_free_kbytes changes
5657 * or when memory is hot-{added|removed}
5659 * Ensures that the watermark[min,low,high] values for each zone are set
5660 * correctly with respect to min_free_kbytes.
5662 void setup_per_zone_wmarks(void)
5664 mutex_lock(&zonelists_mutex);
5665 __setup_per_zone_wmarks();
5666 mutex_unlock(&zonelists_mutex);
5670 * The inactive anon list should be small enough that the VM never has to
5671 * do too much work, but large enough that each inactive page has a chance
5672 * to be referenced again before it is swapped out.
5674 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5675 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5676 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5677 * the anonymous pages are kept on the inactive list.
5680 * memory ratio inactive anon
5681 * -------------------------------------
5690 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5692 unsigned int gb, ratio;
5694 /* Zone size in gigabytes */
5695 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5697 ratio = int_sqrt(10 * gb);
5701 zone->inactive_ratio = ratio;
5704 static void __meminit setup_per_zone_inactive_ratio(void)
5709 calculate_zone_inactive_ratio(zone);
5713 * Initialise min_free_kbytes.
5715 * For small machines we want it small (128k min). For large machines
5716 * we want it large (64MB max). But it is not linear, because network
5717 * bandwidth does not increase linearly with machine size. We use
5719 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5720 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5736 int __meminit init_per_zone_wmark_min(void)
5738 unsigned long lowmem_kbytes;
5739 int new_min_free_kbytes;
5741 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5742 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5744 if (new_min_free_kbytes > user_min_free_kbytes) {
5745 min_free_kbytes = new_min_free_kbytes;
5746 if (min_free_kbytes < 128)
5747 min_free_kbytes = 128;
5748 if (min_free_kbytes > 65536)
5749 min_free_kbytes = 65536;
5751 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5752 new_min_free_kbytes, user_min_free_kbytes);
5754 setup_per_zone_wmarks();
5755 refresh_zone_stat_thresholds();
5756 setup_per_zone_lowmem_reserve();
5757 setup_per_zone_inactive_ratio();
5760 module_init(init_per_zone_wmark_min)
5763 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5764 * that we can call two helper functions whenever min_free_kbytes
5767 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5768 void __user *buffer, size_t *length, loff_t *ppos)
5772 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5777 user_min_free_kbytes = min_free_kbytes;
5778 setup_per_zone_wmarks();
5784 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5785 void __user *buffer, size_t *length, loff_t *ppos)
5790 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5795 zone->min_unmapped_pages = (zone->managed_pages *
5796 sysctl_min_unmapped_ratio) / 100;
5800 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5801 void __user *buffer, size_t *length, loff_t *ppos)
5806 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5811 zone->min_slab_pages = (zone->managed_pages *
5812 sysctl_min_slab_ratio) / 100;
5818 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5819 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5820 * whenever sysctl_lowmem_reserve_ratio changes.
5822 * The reserve ratio obviously has absolutely no relation with the
5823 * minimum watermarks. The lowmem reserve ratio can only make sense
5824 * if in function of the boot time zone sizes.
5826 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5827 void __user *buffer, size_t *length, loff_t *ppos)
5829 proc_dointvec_minmax(table, write, buffer, length, ppos);
5830 setup_per_zone_lowmem_reserve();
5835 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5836 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5837 * pagelist can have before it gets flushed back to buddy allocator.
5839 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5840 void __user *buffer, size_t *length, loff_t *ppos)
5846 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5847 if (!write || (ret < 0))
5850 mutex_lock(&pcp_batch_high_lock);
5851 for_each_populated_zone(zone) {
5853 high = zone->managed_pages / percpu_pagelist_fraction;
5854 for_each_possible_cpu(cpu)
5855 pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5858 mutex_unlock(&pcp_batch_high_lock);
5862 int hashdist = HASHDIST_DEFAULT;
5865 static int __init set_hashdist(char *str)
5869 hashdist = simple_strtoul(str, &str, 0);
5872 __setup("hashdist=", set_hashdist);
5876 * allocate a large system hash table from bootmem
5877 * - it is assumed that the hash table must contain an exact power-of-2
5878 * quantity of entries
5879 * - limit is the number of hash buckets, not the total allocation size
5881 void *__init alloc_large_system_hash(const char *tablename,
5882 unsigned long bucketsize,
5883 unsigned long numentries,
5886 unsigned int *_hash_shift,
5887 unsigned int *_hash_mask,
5888 unsigned long low_limit,
5889 unsigned long high_limit)
5891 unsigned long long max = high_limit;
5892 unsigned long log2qty, size;
5895 /* allow the kernel cmdline to have a say */
5897 /* round applicable memory size up to nearest megabyte */
5898 numentries = nr_kernel_pages;
5900 /* It isn't necessary when PAGE_SIZE >= 1MB */
5901 if (PAGE_SHIFT < 20)
5902 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5904 /* limit to 1 bucket per 2^scale bytes of low memory */
5905 if (scale > PAGE_SHIFT)
5906 numentries >>= (scale - PAGE_SHIFT);
5908 numentries <<= (PAGE_SHIFT - scale);
5910 /* Make sure we've got at least a 0-order allocation.. */
5911 if (unlikely(flags & HASH_SMALL)) {
5912 /* Makes no sense without HASH_EARLY */
5913 WARN_ON(!(flags & HASH_EARLY));
5914 if (!(numentries >> *_hash_shift)) {
5915 numentries = 1UL << *_hash_shift;
5916 BUG_ON(!numentries);
5918 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5919 numentries = PAGE_SIZE / bucketsize;
5921 numentries = roundup_pow_of_two(numentries);
5923 /* limit allocation size to 1/16 total memory by default */
5925 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5926 do_div(max, bucketsize);
5928 max = min(max, 0x80000000ULL);
5930 if (numentries < low_limit)
5931 numentries = low_limit;
5932 if (numentries > max)
5935 log2qty = ilog2(numentries);
5938 size = bucketsize << log2qty;
5939 if (flags & HASH_EARLY)
5940 table = memblock_virt_alloc_nopanic(size, 0);
5942 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5945 * If bucketsize is not a power-of-two, we may free
5946 * some pages at the end of hash table which
5947 * alloc_pages_exact() automatically does
5949 if (get_order(size) < MAX_ORDER) {
5950 table = alloc_pages_exact(size, GFP_ATOMIC);
5951 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5954 } while (!table && size > PAGE_SIZE && --log2qty);
5957 panic("Failed to allocate %s hash table\n", tablename);
5959 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5962 ilog2(size) - PAGE_SHIFT,
5966 *_hash_shift = log2qty;
5968 *_hash_mask = (1 << log2qty) - 1;
5973 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5974 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5977 #ifdef CONFIG_SPARSEMEM
5978 return __pfn_to_section(pfn)->pageblock_flags;
5980 return zone->pageblock_flags;
5981 #endif /* CONFIG_SPARSEMEM */
5984 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5986 #ifdef CONFIG_SPARSEMEM
5987 pfn &= (PAGES_PER_SECTION-1);
5988 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5990 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5991 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5992 #endif /* CONFIG_SPARSEMEM */
5996 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5997 * @page: The page within the block of interest
5998 * @start_bitidx: The first bit of interest to retrieve
5999 * @end_bitidx: The last bit of interest
6000 * returns pageblock_bits flags
6002 unsigned long get_pageblock_flags_group(struct page *page,
6003 int start_bitidx, int end_bitidx)
6006 unsigned long *bitmap;
6007 unsigned long pfn, bitidx;
6008 unsigned long flags = 0;
6009 unsigned long value = 1;
6011 zone = page_zone(page);
6012 pfn = page_to_pfn(page);
6013 bitmap = get_pageblock_bitmap(zone, pfn);
6014 bitidx = pfn_to_bitidx(zone, pfn);
6016 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6017 if (test_bit(bitidx + start_bitidx, bitmap))
6024 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
6025 * @page: The page within the block of interest
6026 * @start_bitidx: The first bit of interest
6027 * @end_bitidx: The last bit of interest
6028 * @flags: The flags to set
6030 void set_pageblock_flags_group(struct page *page, unsigned long flags,
6031 int start_bitidx, int end_bitidx)
6034 unsigned long *bitmap;
6035 unsigned long pfn, bitidx;
6036 unsigned long value = 1;
6038 zone = page_zone(page);
6039 pfn = page_to_pfn(page);
6040 bitmap = get_pageblock_bitmap(zone, pfn);
6041 bitidx = pfn_to_bitidx(zone, pfn);
6042 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6044 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6046 __set_bit(bitidx + start_bitidx, bitmap);
6048 __clear_bit(bitidx + start_bitidx, bitmap);
6052 * This function checks whether pageblock includes unmovable pages or not.
6053 * If @count is not zero, it is okay to include less @count unmovable pages
6055 * PageLRU check without isolation or lru_lock could race so that
6056 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6057 * expect this function should be exact.
6059 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6060 bool skip_hwpoisoned_pages)
6062 unsigned long pfn, iter, found;
6066 * For avoiding noise data, lru_add_drain_all() should be called
6067 * If ZONE_MOVABLE, the zone never contains unmovable pages
6069 if (zone_idx(zone) == ZONE_MOVABLE)
6071 mt = get_pageblock_migratetype(page);
6072 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6075 pfn = page_to_pfn(page);
6076 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6077 unsigned long check = pfn + iter;
6079 if (!pfn_valid_within(check))
6082 page = pfn_to_page(check);
6085 * Hugepages are not in LRU lists, but they're movable.
6086 * We need not scan over tail pages bacause we don't
6087 * handle each tail page individually in migration.
6089 if (PageHuge(page)) {
6090 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6095 * We can't use page_count without pin a page
6096 * because another CPU can free compound page.
6097 * This check already skips compound tails of THP
6098 * because their page->_count is zero at all time.
6100 if (!atomic_read(&page->_count)) {
6101 if (PageBuddy(page))
6102 iter += (1 << page_order(page)) - 1;
6107 * The HWPoisoned page may be not in buddy system, and
6108 * page_count() is not 0.
6110 if (skip_hwpoisoned_pages && PageHWPoison(page))
6116 * If there are RECLAIMABLE pages, we need to check it.
6117 * But now, memory offline itself doesn't call shrink_slab()
6118 * and it still to be fixed.
6121 * If the page is not RAM, page_count()should be 0.
6122 * we don't need more check. This is an _used_ not-movable page.
6124 * The problematic thing here is PG_reserved pages. PG_reserved
6125 * is set to both of a memory hole page and a _used_ kernel
6134 bool is_pageblock_removable_nolock(struct page *page)
6140 * We have to be careful here because we are iterating over memory
6141 * sections which are not zone aware so we might end up outside of
6142 * the zone but still within the section.
6143 * We have to take care about the node as well. If the node is offline
6144 * its NODE_DATA will be NULL - see page_zone.
6146 if (!node_online(page_to_nid(page)))
6149 zone = page_zone(page);
6150 pfn = page_to_pfn(page);
6151 if (!zone_spans_pfn(zone, pfn))
6154 return !has_unmovable_pages(zone, page, 0, true);
6159 static unsigned long pfn_max_align_down(unsigned long pfn)
6161 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6162 pageblock_nr_pages) - 1);
6165 static unsigned long pfn_max_align_up(unsigned long pfn)
6167 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6168 pageblock_nr_pages));
6171 /* [start, end) must belong to a single zone. */
6172 static int __alloc_contig_migrate_range(struct compact_control *cc,
6173 unsigned long start, unsigned long end)
6175 /* This function is based on compact_zone() from compaction.c. */
6176 unsigned long nr_reclaimed;
6177 unsigned long pfn = start;
6178 unsigned int tries = 0;
6183 while (pfn < end || !list_empty(&cc->migratepages)) {
6184 if (fatal_signal_pending(current)) {
6189 if (list_empty(&cc->migratepages)) {
6190 cc->nr_migratepages = 0;
6191 pfn = isolate_migratepages_range(cc->zone, cc,
6198 } else if (++tries == 5) {
6199 ret = ret < 0 ? ret : -EBUSY;
6203 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6205 cc->nr_migratepages -= nr_reclaimed;
6207 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6208 0, MIGRATE_SYNC, MR_CMA);
6211 putback_movable_pages(&cc->migratepages);
6218 * alloc_contig_range() -- tries to allocate given range of pages
6219 * @start: start PFN to allocate
6220 * @end: one-past-the-last PFN to allocate
6221 * @migratetype: migratetype of the underlaying pageblocks (either
6222 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6223 * in range must have the same migratetype and it must
6224 * be either of the two.
6226 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6227 * aligned, however it's the caller's responsibility to guarantee that
6228 * we are the only thread that changes migrate type of pageblocks the
6231 * The PFN range must belong to a single zone.
6233 * Returns zero on success or negative error code. On success all
6234 * pages which PFN is in [start, end) are allocated for the caller and
6235 * need to be freed with free_contig_range().
6237 int alloc_contig_range(unsigned long start, unsigned long end,
6238 unsigned migratetype)
6240 unsigned long outer_start, outer_end;
6243 struct compact_control cc = {
6244 .nr_migratepages = 0,
6246 .zone = page_zone(pfn_to_page(start)),
6248 .ignore_skip_hint = true,
6250 INIT_LIST_HEAD(&cc.migratepages);
6253 * What we do here is we mark all pageblocks in range as
6254 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6255 * have different sizes, and due to the way page allocator
6256 * work, we align the range to biggest of the two pages so
6257 * that page allocator won't try to merge buddies from
6258 * different pageblocks and change MIGRATE_ISOLATE to some
6259 * other migration type.
6261 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6262 * migrate the pages from an unaligned range (ie. pages that
6263 * we are interested in). This will put all the pages in
6264 * range back to page allocator as MIGRATE_ISOLATE.
6266 * When this is done, we take the pages in range from page
6267 * allocator removing them from the buddy system. This way
6268 * page allocator will never consider using them.
6270 * This lets us mark the pageblocks back as
6271 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6272 * aligned range but not in the unaligned, original range are
6273 * put back to page allocator so that buddy can use them.
6276 ret = start_isolate_page_range(pfn_max_align_down(start),
6277 pfn_max_align_up(end), migratetype,
6282 ret = __alloc_contig_migrate_range(&cc, start, end);
6287 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6288 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6289 * more, all pages in [start, end) are free in page allocator.
6290 * What we are going to do is to allocate all pages from
6291 * [start, end) (that is remove them from page allocator).
6293 * The only problem is that pages at the beginning and at the
6294 * end of interesting range may be not aligned with pages that
6295 * page allocator holds, ie. they can be part of higher order
6296 * pages. Because of this, we reserve the bigger range and
6297 * once this is done free the pages we are not interested in.
6299 * We don't have to hold zone->lock here because the pages are
6300 * isolated thus they won't get removed from buddy.
6303 lru_add_drain_all();
6307 outer_start = start;
6308 while (!PageBuddy(pfn_to_page(outer_start))) {
6309 if (++order >= MAX_ORDER) {
6313 outer_start &= ~0UL << order;
6316 /* Make sure the range is really isolated. */
6317 if (test_pages_isolated(outer_start, end, false)) {
6318 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6325 /* Grab isolated pages from freelists. */
6326 outer_end = isolate_freepages_range(&cc, outer_start, end);
6332 /* Free head and tail (if any) */
6333 if (start != outer_start)
6334 free_contig_range(outer_start, start - outer_start);
6335 if (end != outer_end)
6336 free_contig_range(end, outer_end - end);
6339 undo_isolate_page_range(pfn_max_align_down(start),
6340 pfn_max_align_up(end), migratetype);
6344 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6346 unsigned int count = 0;
6348 for (; nr_pages--; pfn++) {
6349 struct page *page = pfn_to_page(pfn);
6351 count += page_count(page) != 1;
6354 WARN(count != 0, "%d pages are still in use!\n", count);
6358 #ifdef CONFIG_MEMORY_HOTPLUG
6360 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6361 * page high values need to be recalulated.
6363 void __meminit zone_pcp_update(struct zone *zone)
6366 mutex_lock(&pcp_batch_high_lock);
6367 for_each_possible_cpu(cpu)
6368 pageset_set_high_and_batch(zone,
6369 per_cpu_ptr(zone->pageset, cpu));
6370 mutex_unlock(&pcp_batch_high_lock);
6374 void zone_pcp_reset(struct zone *zone)
6376 unsigned long flags;
6378 struct per_cpu_pageset *pset;
6380 /* avoid races with drain_pages() */
6381 local_irq_save(flags);
6382 if (zone->pageset != &boot_pageset) {
6383 for_each_online_cpu(cpu) {
6384 pset = per_cpu_ptr(zone->pageset, cpu);
6385 drain_zonestat(zone, pset);
6387 free_percpu(zone->pageset);
6388 zone->pageset = &boot_pageset;
6390 local_irq_restore(flags);
6393 #ifdef CONFIG_MEMORY_HOTREMOVE
6395 * All pages in the range must be isolated before calling this.
6398 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6404 unsigned long flags;
6405 /* find the first valid pfn */
6406 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6411 zone = page_zone(pfn_to_page(pfn));
6412 spin_lock_irqsave(&zone->lock, flags);
6414 while (pfn < end_pfn) {
6415 if (!pfn_valid(pfn)) {
6419 page = pfn_to_page(pfn);
6421 * The HWPoisoned page may be not in buddy system, and
6422 * page_count() is not 0.
6424 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6426 SetPageReserved(page);
6430 BUG_ON(page_count(page));
6431 BUG_ON(!PageBuddy(page));
6432 order = page_order(page);
6433 #ifdef CONFIG_DEBUG_VM
6434 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6435 pfn, 1 << order, end_pfn);
6437 list_del(&page->lru);
6438 rmv_page_order(page);
6439 zone->free_area[order].nr_free--;
6440 for (i = 0; i < (1 << order); i++)
6441 SetPageReserved((page+i));
6442 pfn += (1 << order);
6444 spin_unlock_irqrestore(&zone->lock, flags);
6448 #ifdef CONFIG_MEMORY_FAILURE
6449 bool is_free_buddy_page(struct page *page)
6451 struct zone *zone = page_zone(page);
6452 unsigned long pfn = page_to_pfn(page);
6453 unsigned long flags;
6456 spin_lock_irqsave(&zone->lock, flags);
6457 for (order = 0; order < MAX_ORDER; order++) {
6458 struct page *page_head = page - (pfn & ((1 << order) - 1));
6460 if (PageBuddy(page_head) && page_order(page_head) >= order)
6463 spin_unlock_irqrestore(&zone->lock, flags);
6465 return order < MAX_ORDER;
6469 static const struct trace_print_flags pageflag_names[] = {
6470 {1UL << PG_locked, "locked" },
6471 {1UL << PG_error, "error" },
6472 {1UL << PG_referenced, "referenced" },
6473 {1UL << PG_uptodate, "uptodate" },
6474 {1UL << PG_dirty, "dirty" },
6475 {1UL << PG_lru, "lru" },
6476 {1UL << PG_active, "active" },
6477 {1UL << PG_slab, "slab" },
6478 {1UL << PG_owner_priv_1, "owner_priv_1" },
6479 {1UL << PG_arch_1, "arch_1" },
6480 {1UL << PG_reserved, "reserved" },
6481 {1UL << PG_private, "private" },
6482 {1UL << PG_private_2, "private_2" },
6483 {1UL << PG_writeback, "writeback" },
6484 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6485 {1UL << PG_head, "head" },
6486 {1UL << PG_tail, "tail" },
6488 {1UL << PG_compound, "compound" },
6490 {1UL << PG_swapcache, "swapcache" },
6491 {1UL << PG_mappedtodisk, "mappedtodisk" },
6492 {1UL << PG_reclaim, "reclaim" },
6493 {1UL << PG_swapbacked, "swapbacked" },
6494 {1UL << PG_unevictable, "unevictable" },
6496 {1UL << PG_mlocked, "mlocked" },
6498 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6499 {1UL << PG_uncached, "uncached" },
6501 #ifdef CONFIG_MEMORY_FAILURE
6502 {1UL << PG_hwpoison, "hwpoison" },
6504 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6505 {1UL << PG_compound_lock, "compound_lock" },
6509 static void dump_page_flags(unsigned long flags)
6511 const char *delim = "";
6515 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6517 printk(KERN_ALERT "page flags: %#lx(", flags);
6519 /* remove zone id */
6520 flags &= (1UL << NR_PAGEFLAGS) - 1;
6522 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6524 mask = pageflag_names[i].mask;
6525 if ((flags & mask) != mask)
6529 printk("%s%s", delim, pageflag_names[i].name);
6533 /* check for left over flags */
6535 printk("%s%#lx", delim, flags);
6540 void dump_page_badflags(struct page *page, char *reason, unsigned long badflags)
6543 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6544 page, atomic_read(&page->_count), page_mapcount(page),
6545 page->mapping, page->index);
6546 dump_page_flags(page->flags);
6548 pr_alert("page dumped because: %s\n", reason);
6549 if (page->flags & badflags) {
6550 pr_alert("bad because of flags:\n");
6551 dump_page_flags(page->flags & badflags);
6553 mem_cgroup_print_bad_page(page);
6556 void dump_page(struct page *page, char *reason)
6558 dump_page_badflags(page, reason, 0);
6560 EXPORT_SYMBOL_GPL(dump_page);