2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
61 #include <linux/sched/rt.h>
63 #include <asm/tlbflush.h>
64 #include <asm/div64.h>
67 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
68 DEFINE_PER_CPU(int, numa_node);
69 EXPORT_PER_CPU_SYMBOL(numa_node);
72 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
74 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
75 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
76 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
77 * defined in <linux/topology.h>.
79 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
80 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
84 * Array of node states.
86 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
87 [N_POSSIBLE] = NODE_MASK_ALL,
88 [N_ONLINE] = { { [0] = 1UL } },
90 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
92 [N_HIGH_MEMORY] = { { [0] = 1UL } },
94 #ifdef CONFIG_MOVABLE_NODE
95 [N_MEMORY] = { { [0] = 1UL } },
97 [N_CPU] = { { [0] = 1UL } },
100 EXPORT_SYMBOL(node_states);
102 unsigned long totalram_pages __read_mostly;
103 unsigned long totalreserve_pages __read_mostly;
105 * When calculating the number of globally allowed dirty pages, there
106 * is a certain number of per-zone reserves that should not be
107 * considered dirtyable memory. This is the sum of those reserves
108 * over all existing zones that contribute dirtyable memory.
110 unsigned long dirty_balance_reserve __read_mostly;
112 int percpu_pagelist_fraction;
113 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
115 #ifdef CONFIG_PM_SLEEP
117 * The following functions are used by the suspend/hibernate code to temporarily
118 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
119 * while devices are suspended. To avoid races with the suspend/hibernate code,
120 * they should always be called with pm_mutex held (gfp_allowed_mask also should
121 * only be modified with pm_mutex held, unless the suspend/hibernate code is
122 * guaranteed not to run in parallel with that modification).
125 static gfp_t saved_gfp_mask;
127 void pm_restore_gfp_mask(void)
129 WARN_ON(!mutex_is_locked(&pm_mutex));
130 if (saved_gfp_mask) {
131 gfp_allowed_mask = saved_gfp_mask;
136 void pm_restrict_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex));
139 WARN_ON(saved_gfp_mask);
140 saved_gfp_mask = gfp_allowed_mask;
141 gfp_allowed_mask &= ~GFP_IOFS;
144 bool pm_suspended_storage(void)
146 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
150 #endif /* CONFIG_PM_SLEEP */
152 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
153 int pageblock_order __read_mostly;
156 static void __free_pages_ok(struct page *page, unsigned int order);
159 * results with 256, 32 in the lowmem_reserve sysctl:
160 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
161 * 1G machine -> (16M dma, 784M normal, 224M high)
162 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
163 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
164 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
166 * TBD: should special case ZONE_DMA32 machines here - in those we normally
167 * don't need any ZONE_NORMAL reservation
169 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
170 #ifdef CONFIG_ZONE_DMA
173 #ifdef CONFIG_ZONE_DMA32
176 #ifdef CONFIG_HIGHMEM
182 EXPORT_SYMBOL(totalram_pages);
184 static char * const zone_names[MAX_NR_ZONES] = {
185 #ifdef CONFIG_ZONE_DMA
188 #ifdef CONFIG_ZONE_DMA32
192 #ifdef CONFIG_HIGHMEM
198 int min_free_kbytes = 1024;
200 static unsigned long __meminitdata nr_kernel_pages;
201 static unsigned long __meminitdata nr_all_pages;
202 static unsigned long __meminitdata dma_reserve;
204 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
205 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
206 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
207 static unsigned long __initdata required_kernelcore;
208 static unsigned long __initdata required_movablecore;
209 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
211 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
213 EXPORT_SYMBOL(movable_zone);
214 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
217 int nr_node_ids __read_mostly = MAX_NUMNODES;
218 int nr_online_nodes __read_mostly = 1;
219 EXPORT_SYMBOL(nr_node_ids);
220 EXPORT_SYMBOL(nr_online_nodes);
223 int page_group_by_mobility_disabled __read_mostly;
225 void set_pageblock_migratetype(struct page *page, int migratetype)
228 if (unlikely(page_group_by_mobility_disabled))
229 migratetype = MIGRATE_UNMOVABLE;
231 set_pageblock_flags_group(page, (unsigned long)migratetype,
232 PB_migrate, PB_migrate_end);
235 bool oom_killer_disabled __read_mostly;
237 #ifdef CONFIG_DEBUG_VM
238 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
242 unsigned long pfn = page_to_pfn(page);
245 seq = zone_span_seqbegin(zone);
246 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
248 else if (pfn < zone->zone_start_pfn)
250 } while (zone_span_seqretry(zone, seq));
255 static int page_is_consistent(struct zone *zone, struct page *page)
257 if (!pfn_valid_within(page_to_pfn(page)))
259 if (zone != page_zone(page))
265 * Temporary debugging check for pages not lying within a given zone.
267 static int bad_range(struct zone *zone, struct page *page)
269 if (page_outside_zone_boundaries(zone, page))
271 if (!page_is_consistent(zone, page))
277 static inline int bad_range(struct zone *zone, struct page *page)
283 static void bad_page(struct page *page)
285 static unsigned long resume;
286 static unsigned long nr_shown;
287 static unsigned long nr_unshown;
289 /* Don't complain about poisoned pages */
290 if (PageHWPoison(page)) {
291 reset_page_mapcount(page); /* remove PageBuddy */
296 * Allow a burst of 60 reports, then keep quiet for that minute;
297 * or allow a steady drip of one report per second.
299 if (nr_shown == 60) {
300 if (time_before(jiffies, resume)) {
306 "BUG: Bad page state: %lu messages suppressed\n",
313 resume = jiffies + 60 * HZ;
315 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
316 current->comm, page_to_pfn(page));
322 /* Leave bad fields for debug, except PageBuddy could make trouble */
323 reset_page_mapcount(page); /* remove PageBuddy */
324 add_taint(TAINT_BAD_PAGE);
328 * Higher-order pages are called "compound pages". They are structured thusly:
330 * The first PAGE_SIZE page is called the "head page".
332 * The remaining PAGE_SIZE pages are called "tail pages".
334 * All pages have PG_compound set. All tail pages have their ->first_page
335 * pointing at the head page.
337 * The first tail page's ->lru.next holds the address of the compound page's
338 * put_page() function. Its ->lru.prev holds the order of allocation.
339 * This usage means that zero-order pages may not be compound.
342 static void free_compound_page(struct page *page)
344 __free_pages_ok(page, compound_order(page));
347 void prep_compound_page(struct page *page, unsigned long order)
350 int nr_pages = 1 << order;
352 set_compound_page_dtor(page, free_compound_page);
353 set_compound_order(page, order);
355 for (i = 1; i < nr_pages; i++) {
356 struct page *p = page + i;
358 set_page_count(p, 0);
359 p->first_page = page;
363 /* update __split_huge_page_refcount if you change this function */
364 static int destroy_compound_page(struct page *page, unsigned long order)
367 int nr_pages = 1 << order;
370 if (unlikely(compound_order(page) != order)) {
375 __ClearPageHead(page);
377 for (i = 1; i < nr_pages; i++) {
378 struct page *p = page + i;
380 if (unlikely(!PageTail(p) || (p->first_page != page))) {
390 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
395 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
396 * and __GFP_HIGHMEM from hard or soft interrupt context.
398 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
399 for (i = 0; i < (1 << order); i++)
400 clear_highpage(page + i);
403 #ifdef CONFIG_DEBUG_PAGEALLOC
404 unsigned int _debug_guardpage_minorder;
406 static int __init debug_guardpage_minorder_setup(char *buf)
410 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
411 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
414 _debug_guardpage_minorder = res;
415 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
418 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
420 static inline void set_page_guard_flag(struct page *page)
422 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
425 static inline void clear_page_guard_flag(struct page *page)
427 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
430 static inline void set_page_guard_flag(struct page *page) { }
431 static inline void clear_page_guard_flag(struct page *page) { }
434 static inline void set_page_order(struct page *page, int order)
436 set_page_private(page, order);
437 __SetPageBuddy(page);
440 static inline void rmv_page_order(struct page *page)
442 __ClearPageBuddy(page);
443 set_page_private(page, 0);
447 * Locate the struct page for both the matching buddy in our
448 * pair (buddy1) and the combined O(n+1) page they form (page).
450 * 1) Any buddy B1 will have an order O twin B2 which satisfies
451 * the following equation:
453 * For example, if the starting buddy (buddy2) is #8 its order
455 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
457 * 2) Any buddy B will have an order O+1 parent P which
458 * satisfies the following equation:
461 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
463 static inline unsigned long
464 __find_buddy_index(unsigned long page_idx, unsigned int order)
466 return page_idx ^ (1 << order);
470 * This function checks whether a page is free && is the buddy
471 * we can do coalesce a page and its buddy if
472 * (a) the buddy is not in a hole &&
473 * (b) the buddy is in the buddy system &&
474 * (c) a page and its buddy have the same order &&
475 * (d) a page and its buddy are in the same zone.
477 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
478 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
480 * For recording page's order, we use page_private(page).
482 static inline int page_is_buddy(struct page *page, struct page *buddy,
485 if (!pfn_valid_within(page_to_pfn(buddy)))
488 if (page_zone_id(page) != page_zone_id(buddy))
491 if (page_is_guard(buddy) && page_order(buddy) == order) {
492 VM_BUG_ON(page_count(buddy) != 0);
496 if (PageBuddy(buddy) && page_order(buddy) == order) {
497 VM_BUG_ON(page_count(buddy) != 0);
504 * Freeing function for a buddy system allocator.
506 * The concept of a buddy system is to maintain direct-mapped table
507 * (containing bit values) for memory blocks of various "orders".
508 * The bottom level table contains the map for the smallest allocatable
509 * units of memory (here, pages), and each level above it describes
510 * pairs of units from the levels below, hence, "buddies".
511 * At a high level, all that happens here is marking the table entry
512 * at the bottom level available, and propagating the changes upward
513 * as necessary, plus some accounting needed to play nicely with other
514 * parts of the VM system.
515 * At each level, we keep a list of pages, which are heads of continuous
516 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
517 * order is recorded in page_private(page) field.
518 * So when we are allocating or freeing one, we can derive the state of the
519 * other. That is, if we allocate a small block, and both were
520 * free, the remainder of the region must be split into blocks.
521 * If a block is freed, and its buddy is also free, then this
522 * triggers coalescing into a block of larger size.
527 static inline void __free_one_page(struct page *page,
528 struct zone *zone, unsigned int order,
531 unsigned long page_idx;
532 unsigned long combined_idx;
533 unsigned long uninitialized_var(buddy_idx);
536 if (unlikely(PageCompound(page)))
537 if (unlikely(destroy_compound_page(page, order)))
540 VM_BUG_ON(migratetype == -1);
542 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
544 VM_BUG_ON(page_idx & ((1 << order) - 1));
545 VM_BUG_ON(bad_range(zone, page));
547 while (order < MAX_ORDER-1) {
548 buddy_idx = __find_buddy_index(page_idx, order);
549 buddy = page + (buddy_idx - page_idx);
550 if (!page_is_buddy(page, buddy, order))
553 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
554 * merge with it and move up one order.
556 if (page_is_guard(buddy)) {
557 clear_page_guard_flag(buddy);
558 set_page_private(page, 0);
559 __mod_zone_freepage_state(zone, 1 << order,
562 list_del(&buddy->lru);
563 zone->free_area[order].nr_free--;
564 rmv_page_order(buddy);
566 combined_idx = buddy_idx & page_idx;
567 page = page + (combined_idx - page_idx);
568 page_idx = combined_idx;
571 set_page_order(page, order);
574 * If this is not the largest possible page, check if the buddy
575 * of the next-highest order is free. If it is, it's possible
576 * that pages are being freed that will coalesce soon. In case,
577 * that is happening, add the free page to the tail of the list
578 * so it's less likely to be used soon and more likely to be merged
579 * as a higher order page
581 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
582 struct page *higher_page, *higher_buddy;
583 combined_idx = buddy_idx & page_idx;
584 higher_page = page + (combined_idx - page_idx);
585 buddy_idx = __find_buddy_index(combined_idx, order + 1);
586 higher_buddy = higher_page + (buddy_idx - combined_idx);
587 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
588 list_add_tail(&page->lru,
589 &zone->free_area[order].free_list[migratetype]);
594 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
596 zone->free_area[order].nr_free++;
599 static inline int free_pages_check(struct page *page)
601 if (unlikely(page_mapcount(page) |
602 (page->mapping != NULL) |
603 (atomic_read(&page->_count) != 0) |
604 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
605 (mem_cgroup_bad_page_check(page)))) {
609 reset_page_last_nid(page);
610 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
611 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
616 * Frees a number of pages from the PCP lists
617 * Assumes all pages on list are in same zone, and of same order.
618 * count is the number of pages to free.
620 * If the zone was previously in an "all pages pinned" state then look to
621 * see if this freeing clears that state.
623 * And clear the zone's pages_scanned counter, to hold off the "all pages are
624 * pinned" detection logic.
626 static void free_pcppages_bulk(struct zone *zone, int count,
627 struct per_cpu_pages *pcp)
633 spin_lock(&zone->lock);
634 zone->all_unreclaimable = 0;
635 zone->pages_scanned = 0;
639 struct list_head *list;
642 * Remove pages from lists in a round-robin fashion. A
643 * batch_free count is maintained that is incremented when an
644 * empty list is encountered. This is so more pages are freed
645 * off fuller lists instead of spinning excessively around empty
650 if (++migratetype == MIGRATE_PCPTYPES)
652 list = &pcp->lists[migratetype];
653 } while (list_empty(list));
655 /* This is the only non-empty list. Free them all. */
656 if (batch_free == MIGRATE_PCPTYPES)
657 batch_free = to_free;
660 int mt; /* migratetype of the to-be-freed page */
662 page = list_entry(list->prev, struct page, lru);
663 /* must delete as __free_one_page list manipulates */
664 list_del(&page->lru);
665 mt = get_freepage_migratetype(page);
666 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
667 __free_one_page(page, zone, 0, mt);
668 trace_mm_page_pcpu_drain(page, 0, mt);
669 if (likely(get_pageblock_migratetype(page) != MIGRATE_ISOLATE)) {
670 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
671 if (is_migrate_cma(mt))
672 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
674 } while (--to_free && --batch_free && !list_empty(list));
676 spin_unlock(&zone->lock);
679 static void free_one_page(struct zone *zone, struct page *page, int order,
682 spin_lock(&zone->lock);
683 zone->all_unreclaimable = 0;
684 zone->pages_scanned = 0;
686 __free_one_page(page, zone, order, migratetype);
687 if (unlikely(migratetype != MIGRATE_ISOLATE))
688 __mod_zone_freepage_state(zone, 1 << order, migratetype);
689 spin_unlock(&zone->lock);
692 static bool free_pages_prepare(struct page *page, unsigned int order)
697 trace_mm_page_free(page, order);
698 kmemcheck_free_shadow(page, order);
701 page->mapping = NULL;
702 for (i = 0; i < (1 << order); i++)
703 bad += free_pages_check(page + i);
707 if (!PageHighMem(page)) {
708 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
709 debug_check_no_obj_freed(page_address(page),
712 arch_free_page(page, order);
713 kernel_map_pages(page, 1 << order, 0);
718 static void __free_pages_ok(struct page *page, unsigned int order)
723 if (!free_pages_prepare(page, order))
726 local_irq_save(flags);
727 __count_vm_events(PGFREE, 1 << order);
728 migratetype = get_pageblock_migratetype(page);
729 set_freepage_migratetype(page, migratetype);
730 free_one_page(page_zone(page), page, order, migratetype);
731 local_irq_restore(flags);
735 * Read access to zone->managed_pages is safe because it's unsigned long,
736 * but we still need to serialize writers. Currently all callers of
737 * __free_pages_bootmem() except put_page_bootmem() should only be used
738 * at boot time. So for shorter boot time, we shift the burden to
739 * put_page_bootmem() to serialize writers.
741 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
743 unsigned int nr_pages = 1 << order;
747 for (loop = 0; loop < nr_pages; loop++) {
748 struct page *p = &page[loop];
750 if (loop + 1 < nr_pages)
752 __ClearPageReserved(p);
753 set_page_count(p, 0);
756 page_zone(page)->managed_pages += 1 << order;
757 set_page_refcounted(page);
758 __free_pages(page, order);
762 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
763 void __init init_cma_reserved_pageblock(struct page *page)
765 unsigned i = pageblock_nr_pages;
766 struct page *p = page;
769 __ClearPageReserved(p);
770 set_page_count(p, 0);
773 set_page_refcounted(page);
774 set_pageblock_migratetype(page, MIGRATE_CMA);
775 __free_pages(page, pageblock_order);
776 totalram_pages += pageblock_nr_pages;
777 #ifdef CONFIG_HIGHMEM
778 if (PageHighMem(page))
779 totalhigh_pages += pageblock_nr_pages;
785 * The order of subdivision here is critical for the IO subsystem.
786 * Please do not alter this order without good reasons and regression
787 * testing. Specifically, as large blocks of memory are subdivided,
788 * the order in which smaller blocks are delivered depends on the order
789 * they're subdivided in this function. This is the primary factor
790 * influencing the order in which pages are delivered to the IO
791 * subsystem according to empirical testing, and this is also justified
792 * by considering the behavior of a buddy system containing a single
793 * large block of memory acted on by a series of small allocations.
794 * This behavior is a critical factor in sglist merging's success.
798 static inline void expand(struct zone *zone, struct page *page,
799 int low, int high, struct free_area *area,
802 unsigned long size = 1 << high;
808 VM_BUG_ON(bad_range(zone, &page[size]));
810 #ifdef CONFIG_DEBUG_PAGEALLOC
811 if (high < debug_guardpage_minorder()) {
813 * Mark as guard pages (or page), that will allow to
814 * merge back to allocator when buddy will be freed.
815 * Corresponding page table entries will not be touched,
816 * pages will stay not present in virtual address space
818 INIT_LIST_HEAD(&page[size].lru);
819 set_page_guard_flag(&page[size]);
820 set_page_private(&page[size], high);
821 /* Guard pages are not available for any usage */
822 __mod_zone_freepage_state(zone, -(1 << high),
827 list_add(&page[size].lru, &area->free_list[migratetype]);
829 set_page_order(&page[size], high);
834 * This page is about to be returned from the page allocator
836 static inline int check_new_page(struct page *page)
838 if (unlikely(page_mapcount(page) |
839 (page->mapping != NULL) |
840 (atomic_read(&page->_count) != 0) |
841 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
842 (mem_cgroup_bad_page_check(page)))) {
849 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
853 for (i = 0; i < (1 << order); i++) {
854 struct page *p = page + i;
855 if (unlikely(check_new_page(p)))
859 set_page_private(page, 0);
860 set_page_refcounted(page);
862 arch_alloc_page(page, order);
863 kernel_map_pages(page, 1 << order, 1);
865 if (gfp_flags & __GFP_ZERO)
866 prep_zero_page(page, order, gfp_flags);
868 if (order && (gfp_flags & __GFP_COMP))
869 prep_compound_page(page, order);
875 * Go through the free lists for the given migratetype and remove
876 * the smallest available page from the freelists
879 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
882 unsigned int current_order;
883 struct free_area * area;
886 /* Find a page of the appropriate size in the preferred list */
887 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
888 area = &(zone->free_area[current_order]);
889 if (list_empty(&area->free_list[migratetype]))
892 page = list_entry(area->free_list[migratetype].next,
894 list_del(&page->lru);
895 rmv_page_order(page);
897 expand(zone, page, order, current_order, area, migratetype);
906 * This array describes the order lists are fallen back to when
907 * the free lists for the desirable migrate type are depleted
909 static int fallbacks[MIGRATE_TYPES][4] = {
910 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
911 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
913 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
914 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
916 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
918 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
919 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
923 * Move the free pages in a range to the free lists of the requested type.
924 * Note that start_page and end_pages are not aligned on a pageblock
925 * boundary. If alignment is required, use move_freepages_block()
927 int move_freepages(struct zone *zone,
928 struct page *start_page, struct page *end_page,
935 #ifndef CONFIG_HOLES_IN_ZONE
937 * page_zone is not safe to call in this context when
938 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
939 * anyway as we check zone boundaries in move_freepages_block().
940 * Remove at a later date when no bug reports exist related to
941 * grouping pages by mobility
943 BUG_ON(page_zone(start_page) != page_zone(end_page));
946 for (page = start_page; page <= end_page;) {
947 /* Make sure we are not inadvertently changing nodes */
948 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
950 if (!pfn_valid_within(page_to_pfn(page))) {
955 if (!PageBuddy(page)) {
960 order = page_order(page);
961 list_move(&page->lru,
962 &zone->free_area[order].free_list[migratetype]);
963 set_freepage_migratetype(page, migratetype);
965 pages_moved += 1 << order;
971 int move_freepages_block(struct zone *zone, struct page *page,
974 unsigned long start_pfn, end_pfn;
975 struct page *start_page, *end_page;
977 start_pfn = page_to_pfn(page);
978 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
979 start_page = pfn_to_page(start_pfn);
980 end_page = start_page + pageblock_nr_pages - 1;
981 end_pfn = start_pfn + pageblock_nr_pages - 1;
983 /* Do not cross zone boundaries */
984 if (start_pfn < zone->zone_start_pfn)
986 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
989 return move_freepages(zone, start_page, end_page, migratetype);
992 static void change_pageblock_range(struct page *pageblock_page,
993 int start_order, int migratetype)
995 int nr_pageblocks = 1 << (start_order - pageblock_order);
997 while (nr_pageblocks--) {
998 set_pageblock_migratetype(pageblock_page, migratetype);
999 pageblock_page += pageblock_nr_pages;
1003 /* Remove an element from the buddy allocator from the fallback list */
1004 static inline struct page *
1005 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1007 struct free_area * area;
1012 /* Find the largest possible block of pages in the other list */
1013 for (current_order = MAX_ORDER-1; current_order >= order;
1016 migratetype = fallbacks[start_migratetype][i];
1018 /* MIGRATE_RESERVE handled later if necessary */
1019 if (migratetype == MIGRATE_RESERVE)
1022 area = &(zone->free_area[current_order]);
1023 if (list_empty(&area->free_list[migratetype]))
1026 page = list_entry(area->free_list[migratetype].next,
1031 * If breaking a large block of pages, move all free
1032 * pages to the preferred allocation list. If falling
1033 * back for a reclaimable kernel allocation, be more
1034 * aggressive about taking ownership of free pages
1036 * On the other hand, never change migration
1037 * type of MIGRATE_CMA pageblocks nor move CMA
1038 * pages on different free lists. We don't
1039 * want unmovable pages to be allocated from
1040 * MIGRATE_CMA areas.
1042 if (!is_migrate_cma(migratetype) &&
1043 (unlikely(current_order >= pageblock_order / 2) ||
1044 start_migratetype == MIGRATE_RECLAIMABLE ||
1045 page_group_by_mobility_disabled)) {
1047 pages = move_freepages_block(zone, page,
1050 /* Claim the whole block if over half of it is free */
1051 if (pages >= (1 << (pageblock_order-1)) ||
1052 page_group_by_mobility_disabled)
1053 set_pageblock_migratetype(page,
1056 migratetype = start_migratetype;
1059 /* Remove the page from the freelists */
1060 list_del(&page->lru);
1061 rmv_page_order(page);
1063 /* Take ownership for orders >= pageblock_order */
1064 if (current_order >= pageblock_order &&
1065 !is_migrate_cma(migratetype))
1066 change_pageblock_range(page, current_order,
1069 expand(zone, page, order, current_order, area,
1070 is_migrate_cma(migratetype)
1071 ? migratetype : start_migratetype);
1073 trace_mm_page_alloc_extfrag(page, order, current_order,
1074 start_migratetype, migratetype);
1084 * Do the hard work of removing an element from the buddy allocator.
1085 * Call me with the zone->lock already held.
1087 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1093 page = __rmqueue_smallest(zone, order, migratetype);
1095 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1096 page = __rmqueue_fallback(zone, order, migratetype);
1099 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1100 * is used because __rmqueue_smallest is an inline function
1101 * and we want just one call site
1104 migratetype = MIGRATE_RESERVE;
1109 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1114 * Obtain a specified number of elements from the buddy allocator, all under
1115 * a single hold of the lock, for efficiency. Add them to the supplied list.
1116 * Returns the number of new pages which were placed at *list.
1118 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1119 unsigned long count, struct list_head *list,
1120 int migratetype, int cold)
1122 int mt = migratetype, i;
1124 spin_lock(&zone->lock);
1125 for (i = 0; i < count; ++i) {
1126 struct page *page = __rmqueue(zone, order, migratetype);
1127 if (unlikely(page == NULL))
1131 * Split buddy pages returned by expand() are received here
1132 * in physical page order. The page is added to the callers and
1133 * list and the list head then moves forward. From the callers
1134 * perspective, the linked list is ordered by page number in
1135 * some conditions. This is useful for IO devices that can
1136 * merge IO requests if the physical pages are ordered
1139 if (likely(cold == 0))
1140 list_add(&page->lru, list);
1142 list_add_tail(&page->lru, list);
1143 if (IS_ENABLED(CONFIG_CMA)) {
1144 mt = get_pageblock_migratetype(page);
1145 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1148 set_freepage_migratetype(page, mt);
1150 if (is_migrate_cma(mt))
1151 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1154 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1155 spin_unlock(&zone->lock);
1161 * Called from the vmstat counter updater to drain pagesets of this
1162 * currently executing processor on remote nodes after they have
1165 * Note that this function must be called with the thread pinned to
1166 * a single processor.
1168 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1170 unsigned long flags;
1173 local_irq_save(flags);
1174 if (pcp->count >= pcp->batch)
1175 to_drain = pcp->batch;
1177 to_drain = pcp->count;
1179 free_pcppages_bulk(zone, to_drain, pcp);
1180 pcp->count -= to_drain;
1182 local_irq_restore(flags);
1187 * Drain pages of the indicated processor.
1189 * The processor must either be the current processor and the
1190 * thread pinned to the current processor or a processor that
1193 static void drain_pages(unsigned int cpu)
1195 unsigned long flags;
1198 for_each_populated_zone(zone) {
1199 struct per_cpu_pageset *pset;
1200 struct per_cpu_pages *pcp;
1202 local_irq_save(flags);
1203 pset = per_cpu_ptr(zone->pageset, cpu);
1207 free_pcppages_bulk(zone, pcp->count, pcp);
1210 local_irq_restore(flags);
1215 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1217 void drain_local_pages(void *arg)
1219 drain_pages(smp_processor_id());
1223 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1225 * Note that this code is protected against sending an IPI to an offline
1226 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1227 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1228 * nothing keeps CPUs from showing up after we populated the cpumask and
1229 * before the call to on_each_cpu_mask().
1231 void drain_all_pages(void)
1234 struct per_cpu_pageset *pcp;
1238 * Allocate in the BSS so we wont require allocation in
1239 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1241 static cpumask_t cpus_with_pcps;
1244 * We don't care about racing with CPU hotplug event
1245 * as offline notification will cause the notified
1246 * cpu to drain that CPU pcps and on_each_cpu_mask
1247 * disables preemption as part of its processing
1249 for_each_online_cpu(cpu) {
1250 bool has_pcps = false;
1251 for_each_populated_zone(zone) {
1252 pcp = per_cpu_ptr(zone->pageset, cpu);
1253 if (pcp->pcp.count) {
1259 cpumask_set_cpu(cpu, &cpus_with_pcps);
1261 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1263 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1266 #ifdef CONFIG_HIBERNATION
1268 void mark_free_pages(struct zone *zone)
1270 unsigned long pfn, max_zone_pfn;
1271 unsigned long flags;
1273 struct list_head *curr;
1275 if (!zone->spanned_pages)
1278 spin_lock_irqsave(&zone->lock, flags);
1280 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1281 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1282 if (pfn_valid(pfn)) {
1283 struct page *page = pfn_to_page(pfn);
1285 if (!swsusp_page_is_forbidden(page))
1286 swsusp_unset_page_free(page);
1289 for_each_migratetype_order(order, t) {
1290 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1293 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1294 for (i = 0; i < (1UL << order); i++)
1295 swsusp_set_page_free(pfn_to_page(pfn + i));
1298 spin_unlock_irqrestore(&zone->lock, flags);
1300 #endif /* CONFIG_PM */
1303 * Free a 0-order page
1304 * cold == 1 ? free a cold page : free a hot page
1306 void free_hot_cold_page(struct page *page, int cold)
1308 struct zone *zone = page_zone(page);
1309 struct per_cpu_pages *pcp;
1310 unsigned long flags;
1313 if (!free_pages_prepare(page, 0))
1316 migratetype = get_pageblock_migratetype(page);
1317 set_freepage_migratetype(page, migratetype);
1318 local_irq_save(flags);
1319 __count_vm_event(PGFREE);
1322 * We only track unmovable, reclaimable and movable on pcp lists.
1323 * Free ISOLATE pages back to the allocator because they are being
1324 * offlined but treat RESERVE as movable pages so we can get those
1325 * areas back if necessary. Otherwise, we may have to free
1326 * excessively into the page allocator
1328 if (migratetype >= MIGRATE_PCPTYPES) {
1329 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1330 free_one_page(zone, page, 0, migratetype);
1333 migratetype = MIGRATE_MOVABLE;
1336 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1338 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1340 list_add(&page->lru, &pcp->lists[migratetype]);
1342 if (pcp->count >= pcp->high) {
1343 free_pcppages_bulk(zone, pcp->batch, pcp);
1344 pcp->count -= pcp->batch;
1348 local_irq_restore(flags);
1352 * Free a list of 0-order pages
1354 void free_hot_cold_page_list(struct list_head *list, int cold)
1356 struct page *page, *next;
1358 list_for_each_entry_safe(page, next, list, lru) {
1359 trace_mm_page_free_batched(page, cold);
1360 free_hot_cold_page(page, cold);
1365 * split_page takes a non-compound higher-order page, and splits it into
1366 * n (1<<order) sub-pages: page[0..n]
1367 * Each sub-page must be freed individually.
1369 * Note: this is probably too low level an operation for use in drivers.
1370 * Please consult with lkml before using this in your driver.
1372 void split_page(struct page *page, unsigned int order)
1376 VM_BUG_ON(PageCompound(page));
1377 VM_BUG_ON(!page_count(page));
1379 #ifdef CONFIG_KMEMCHECK
1381 * Split shadow pages too, because free(page[0]) would
1382 * otherwise free the whole shadow.
1384 if (kmemcheck_page_is_tracked(page))
1385 split_page(virt_to_page(page[0].shadow), order);
1388 for (i = 1; i < (1 << order); i++)
1389 set_page_refcounted(page + i);
1392 static int __isolate_free_page(struct page *page, unsigned int order)
1394 unsigned long watermark;
1398 BUG_ON(!PageBuddy(page));
1400 zone = page_zone(page);
1401 mt = get_pageblock_migratetype(page);
1403 if (mt != MIGRATE_ISOLATE) {
1404 /* Obey watermarks as if the page was being allocated */
1405 watermark = low_wmark_pages(zone) + (1 << order);
1406 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1409 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1412 /* Remove page from free list */
1413 list_del(&page->lru);
1414 zone->free_area[order].nr_free--;
1415 rmv_page_order(page);
1417 /* Set the pageblock if the isolated page is at least a pageblock */
1418 if (order >= pageblock_order - 1) {
1419 struct page *endpage = page + (1 << order) - 1;
1420 for (; page < endpage; page += pageblock_nr_pages) {
1421 int mt = get_pageblock_migratetype(page);
1422 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1423 set_pageblock_migratetype(page,
1428 return 1UL << order;
1432 * Similar to split_page except the page is already free. As this is only
1433 * being used for migration, the migratetype of the block also changes.
1434 * As this is called with interrupts disabled, the caller is responsible
1435 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1438 * Note: this is probably too low level an operation for use in drivers.
1439 * Please consult with lkml before using this in your driver.
1441 int split_free_page(struct page *page)
1446 order = page_order(page);
1448 nr_pages = __isolate_free_page(page, order);
1452 /* Split into individual pages */
1453 set_page_refcounted(page);
1454 split_page(page, order);
1459 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1460 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1464 struct page *buffered_rmqueue(struct zone *preferred_zone,
1465 struct zone *zone, int order, gfp_t gfp_flags,
1468 unsigned long flags;
1470 int cold = !!(gfp_flags & __GFP_COLD);
1473 if (likely(order == 0)) {
1474 struct per_cpu_pages *pcp;
1475 struct list_head *list;
1477 local_irq_save(flags);
1478 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1479 list = &pcp->lists[migratetype];
1480 if (list_empty(list)) {
1481 pcp->count += rmqueue_bulk(zone, 0,
1484 if (unlikely(list_empty(list)))
1489 page = list_entry(list->prev, struct page, lru);
1491 page = list_entry(list->next, struct page, lru);
1493 list_del(&page->lru);
1496 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1498 * __GFP_NOFAIL is not to be used in new code.
1500 * All __GFP_NOFAIL callers should be fixed so that they
1501 * properly detect and handle allocation failures.
1503 * We most definitely don't want callers attempting to
1504 * allocate greater than order-1 page units with
1507 WARN_ON_ONCE(order > 1);
1509 spin_lock_irqsave(&zone->lock, flags);
1510 page = __rmqueue(zone, order, migratetype);
1511 spin_unlock(&zone->lock);
1514 __mod_zone_freepage_state(zone, -(1 << order),
1515 get_pageblock_migratetype(page));
1518 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1519 zone_statistics(preferred_zone, zone, gfp_flags);
1520 local_irq_restore(flags);
1522 VM_BUG_ON(bad_range(zone, page));
1523 if (prep_new_page(page, order, gfp_flags))
1528 local_irq_restore(flags);
1532 #ifdef CONFIG_FAIL_PAGE_ALLOC
1535 struct fault_attr attr;
1537 u32 ignore_gfp_highmem;
1538 u32 ignore_gfp_wait;
1540 } fail_page_alloc = {
1541 .attr = FAULT_ATTR_INITIALIZER,
1542 .ignore_gfp_wait = 1,
1543 .ignore_gfp_highmem = 1,
1547 static int __init setup_fail_page_alloc(char *str)
1549 return setup_fault_attr(&fail_page_alloc.attr, str);
1551 __setup("fail_page_alloc=", setup_fail_page_alloc);
1553 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1555 if (order < fail_page_alloc.min_order)
1557 if (gfp_mask & __GFP_NOFAIL)
1559 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1561 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1564 return should_fail(&fail_page_alloc.attr, 1 << order);
1567 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1569 static int __init fail_page_alloc_debugfs(void)
1571 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1574 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1575 &fail_page_alloc.attr);
1577 return PTR_ERR(dir);
1579 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1580 &fail_page_alloc.ignore_gfp_wait))
1582 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1583 &fail_page_alloc.ignore_gfp_highmem))
1585 if (!debugfs_create_u32("min-order", mode, dir,
1586 &fail_page_alloc.min_order))
1591 debugfs_remove_recursive(dir);
1596 late_initcall(fail_page_alloc_debugfs);
1598 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1600 #else /* CONFIG_FAIL_PAGE_ALLOC */
1602 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1607 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1610 * Return true if free pages are above 'mark'. This takes into account the order
1611 * of the allocation.
1613 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1614 int classzone_idx, int alloc_flags, long free_pages)
1616 /* free_pages my go negative - that's OK */
1618 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1621 free_pages -= (1 << order) - 1;
1622 if (alloc_flags & ALLOC_HIGH)
1624 if (alloc_flags & ALLOC_HARDER)
1627 /* If allocation can't use CMA areas don't use free CMA pages */
1628 if (!(alloc_flags & ALLOC_CMA))
1629 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1631 if (free_pages <= min + lowmem_reserve)
1633 for (o = 0; o < order; o++) {
1634 /* At the next order, this order's pages become unavailable */
1635 free_pages -= z->free_area[o].nr_free << o;
1637 /* Require fewer higher order pages to be free */
1640 if (free_pages <= min)
1646 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1647 int classzone_idx, int alloc_flags)
1649 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1650 zone_page_state(z, NR_FREE_PAGES));
1653 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1654 int classzone_idx, int alloc_flags)
1656 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1658 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1659 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1661 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1667 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1668 * skip over zones that are not allowed by the cpuset, or that have
1669 * been recently (in last second) found to be nearly full. See further
1670 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1671 * that have to skip over a lot of full or unallowed zones.
1673 * If the zonelist cache is present in the passed in zonelist, then
1674 * returns a pointer to the allowed node mask (either the current
1675 * tasks mems_allowed, or node_states[N_MEMORY].)
1677 * If the zonelist cache is not available for this zonelist, does
1678 * nothing and returns NULL.
1680 * If the fullzones BITMAP in the zonelist cache is stale (more than
1681 * a second since last zap'd) then we zap it out (clear its bits.)
1683 * We hold off even calling zlc_setup, until after we've checked the
1684 * first zone in the zonelist, on the theory that most allocations will
1685 * be satisfied from that first zone, so best to examine that zone as
1686 * quickly as we can.
1688 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1690 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1691 nodemask_t *allowednodes; /* zonelist_cache approximation */
1693 zlc = zonelist->zlcache_ptr;
1697 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1698 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1699 zlc->last_full_zap = jiffies;
1702 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1703 &cpuset_current_mems_allowed :
1704 &node_states[N_MEMORY];
1705 return allowednodes;
1709 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1710 * if it is worth looking at further for free memory:
1711 * 1) Check that the zone isn't thought to be full (doesn't have its
1712 * bit set in the zonelist_cache fullzones BITMAP).
1713 * 2) Check that the zones node (obtained from the zonelist_cache
1714 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1715 * Return true (non-zero) if zone is worth looking at further, or
1716 * else return false (zero) if it is not.
1718 * This check -ignores- the distinction between various watermarks,
1719 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1720 * found to be full for any variation of these watermarks, it will
1721 * be considered full for up to one second by all requests, unless
1722 * we are so low on memory on all allowed nodes that we are forced
1723 * into the second scan of the zonelist.
1725 * In the second scan we ignore this zonelist cache and exactly
1726 * apply the watermarks to all zones, even it is slower to do so.
1727 * We are low on memory in the second scan, and should leave no stone
1728 * unturned looking for a free page.
1730 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1731 nodemask_t *allowednodes)
1733 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1734 int i; /* index of *z in zonelist zones */
1735 int n; /* node that zone *z is on */
1737 zlc = zonelist->zlcache_ptr;
1741 i = z - zonelist->_zonerefs;
1744 /* This zone is worth trying if it is allowed but not full */
1745 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1749 * Given 'z' scanning a zonelist, set the corresponding bit in
1750 * zlc->fullzones, so that subsequent attempts to allocate a page
1751 * from that zone don't waste time re-examining it.
1753 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1755 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1756 int i; /* index of *z in zonelist zones */
1758 zlc = zonelist->zlcache_ptr;
1762 i = z - zonelist->_zonerefs;
1764 set_bit(i, zlc->fullzones);
1768 * clear all zones full, called after direct reclaim makes progress so that
1769 * a zone that was recently full is not skipped over for up to a second
1771 static void zlc_clear_zones_full(struct zonelist *zonelist)
1773 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1775 zlc = zonelist->zlcache_ptr;
1779 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1782 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1784 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1787 static void __paginginit init_zone_allows_reclaim(int nid)
1791 for_each_online_node(i)
1792 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1793 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1795 zone_reclaim_mode = 1;
1798 #else /* CONFIG_NUMA */
1800 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1805 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1806 nodemask_t *allowednodes)
1811 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1815 static void zlc_clear_zones_full(struct zonelist *zonelist)
1819 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1824 static inline void init_zone_allows_reclaim(int nid)
1827 #endif /* CONFIG_NUMA */
1830 * get_page_from_freelist goes through the zonelist trying to allocate
1833 static struct page *
1834 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1835 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1836 struct zone *preferred_zone, int migratetype)
1839 struct page *page = NULL;
1842 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1843 int zlc_active = 0; /* set if using zonelist_cache */
1844 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1846 classzone_idx = zone_idx(preferred_zone);
1849 * Scan zonelist, looking for a zone with enough free.
1850 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1852 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1853 high_zoneidx, nodemask) {
1854 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1855 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1857 if ((alloc_flags & ALLOC_CPUSET) &&
1858 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1861 * When allocating a page cache page for writing, we
1862 * want to get it from a zone that is within its dirty
1863 * limit, such that no single zone holds more than its
1864 * proportional share of globally allowed dirty pages.
1865 * The dirty limits take into account the zone's
1866 * lowmem reserves and high watermark so that kswapd
1867 * should be able to balance it without having to
1868 * write pages from its LRU list.
1870 * This may look like it could increase pressure on
1871 * lower zones by failing allocations in higher zones
1872 * before they are full. But the pages that do spill
1873 * over are limited as the lower zones are protected
1874 * by this very same mechanism. It should not become
1875 * a practical burden to them.
1877 * XXX: For now, allow allocations to potentially
1878 * exceed the per-zone dirty limit in the slowpath
1879 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1880 * which is important when on a NUMA setup the allowed
1881 * zones are together not big enough to reach the
1882 * global limit. The proper fix for these situations
1883 * will require awareness of zones in the
1884 * dirty-throttling and the flusher threads.
1886 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1887 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1888 goto this_zone_full;
1890 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1891 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1895 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1896 if (zone_watermark_ok(zone, order, mark,
1897 classzone_idx, alloc_flags))
1900 if (IS_ENABLED(CONFIG_NUMA) &&
1901 !did_zlc_setup && nr_online_nodes > 1) {
1903 * we do zlc_setup if there are multiple nodes
1904 * and before considering the first zone allowed
1907 allowednodes = zlc_setup(zonelist, alloc_flags);
1912 if (zone_reclaim_mode == 0 ||
1913 !zone_allows_reclaim(preferred_zone, zone))
1914 goto this_zone_full;
1917 * As we may have just activated ZLC, check if the first
1918 * eligible zone has failed zone_reclaim recently.
1920 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1921 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1924 ret = zone_reclaim(zone, gfp_mask, order);
1926 case ZONE_RECLAIM_NOSCAN:
1929 case ZONE_RECLAIM_FULL:
1930 /* scanned but unreclaimable */
1933 /* did we reclaim enough */
1934 if (!zone_watermark_ok(zone, order, mark,
1935 classzone_idx, alloc_flags))
1936 goto this_zone_full;
1941 page = buffered_rmqueue(preferred_zone, zone, order,
1942 gfp_mask, migratetype);
1946 if (IS_ENABLED(CONFIG_NUMA))
1947 zlc_mark_zone_full(zonelist, z);
1950 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1951 /* Disable zlc cache for second zonelist scan */
1958 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1959 * necessary to allocate the page. The expectation is
1960 * that the caller is taking steps that will free more
1961 * memory. The caller should avoid the page being used
1962 * for !PFMEMALLOC purposes.
1964 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1970 * Large machines with many possible nodes should not always dump per-node
1971 * meminfo in irq context.
1973 static inline bool should_suppress_show_mem(void)
1978 ret = in_interrupt();
1983 static DEFINE_RATELIMIT_STATE(nopage_rs,
1984 DEFAULT_RATELIMIT_INTERVAL,
1985 DEFAULT_RATELIMIT_BURST);
1987 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1989 unsigned int filter = SHOW_MEM_FILTER_NODES;
1991 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1992 debug_guardpage_minorder() > 0)
1996 * This documents exceptions given to allocations in certain
1997 * contexts that are allowed to allocate outside current's set
2000 if (!(gfp_mask & __GFP_NOMEMALLOC))
2001 if (test_thread_flag(TIF_MEMDIE) ||
2002 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2003 filter &= ~SHOW_MEM_FILTER_NODES;
2004 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2005 filter &= ~SHOW_MEM_FILTER_NODES;
2008 struct va_format vaf;
2011 va_start(args, fmt);
2016 pr_warn("%pV", &vaf);
2021 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2022 current->comm, order, gfp_mask);
2025 if (!should_suppress_show_mem())
2030 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2031 unsigned long did_some_progress,
2032 unsigned long pages_reclaimed)
2034 /* Do not loop if specifically requested */
2035 if (gfp_mask & __GFP_NORETRY)
2038 /* Always retry if specifically requested */
2039 if (gfp_mask & __GFP_NOFAIL)
2043 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2044 * making forward progress without invoking OOM. Suspend also disables
2045 * storage devices so kswapd will not help. Bail if we are suspending.
2047 if (!did_some_progress && pm_suspended_storage())
2051 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2052 * means __GFP_NOFAIL, but that may not be true in other
2055 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2059 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2060 * specified, then we retry until we no longer reclaim any pages
2061 * (above), or we've reclaimed an order of pages at least as
2062 * large as the allocation's order. In both cases, if the
2063 * allocation still fails, we stop retrying.
2065 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2071 static inline struct page *
2072 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2073 struct zonelist *zonelist, enum zone_type high_zoneidx,
2074 nodemask_t *nodemask, struct zone *preferred_zone,
2079 /* Acquire the OOM killer lock for the zones in zonelist */
2080 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2081 schedule_timeout_uninterruptible(1);
2086 * Go through the zonelist yet one more time, keep very high watermark
2087 * here, this is only to catch a parallel oom killing, we must fail if
2088 * we're still under heavy pressure.
2090 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2091 order, zonelist, high_zoneidx,
2092 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2093 preferred_zone, migratetype);
2097 if (!(gfp_mask & __GFP_NOFAIL)) {
2098 /* The OOM killer will not help higher order allocs */
2099 if (order > PAGE_ALLOC_COSTLY_ORDER)
2101 /* The OOM killer does not needlessly kill tasks for lowmem */
2102 if (high_zoneidx < ZONE_NORMAL)
2105 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2106 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2107 * The caller should handle page allocation failure by itself if
2108 * it specifies __GFP_THISNODE.
2109 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2111 if (gfp_mask & __GFP_THISNODE)
2114 /* Exhausted what can be done so it's blamo time */
2115 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2118 clear_zonelist_oom(zonelist, gfp_mask);
2122 #ifdef CONFIG_COMPACTION
2123 /* Try memory compaction for high-order allocations before reclaim */
2124 static struct page *
2125 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2126 struct zonelist *zonelist, enum zone_type high_zoneidx,
2127 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2128 int migratetype, bool sync_migration,
2129 bool *contended_compaction, bool *deferred_compaction,
2130 unsigned long *did_some_progress)
2135 if (compaction_deferred(preferred_zone, order)) {
2136 *deferred_compaction = true;
2140 current->flags |= PF_MEMALLOC;
2141 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2142 nodemask, sync_migration,
2143 contended_compaction);
2144 current->flags &= ~PF_MEMALLOC;
2146 if (*did_some_progress != COMPACT_SKIPPED) {
2149 /* Page migration frees to the PCP lists but we want merging */
2150 drain_pages(get_cpu());
2153 page = get_page_from_freelist(gfp_mask, nodemask,
2154 order, zonelist, high_zoneidx,
2155 alloc_flags & ~ALLOC_NO_WATERMARKS,
2156 preferred_zone, migratetype);
2158 preferred_zone->compact_blockskip_flush = false;
2159 preferred_zone->compact_considered = 0;
2160 preferred_zone->compact_defer_shift = 0;
2161 if (order >= preferred_zone->compact_order_failed)
2162 preferred_zone->compact_order_failed = order + 1;
2163 count_vm_event(COMPACTSUCCESS);
2168 * It's bad if compaction run occurs and fails.
2169 * The most likely reason is that pages exist,
2170 * but not enough to satisfy watermarks.
2172 count_vm_event(COMPACTFAIL);
2175 * As async compaction considers a subset of pageblocks, only
2176 * defer if the failure was a sync compaction failure.
2179 defer_compaction(preferred_zone, order);
2187 static inline struct page *
2188 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2189 struct zonelist *zonelist, enum zone_type high_zoneidx,
2190 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2191 int migratetype, bool sync_migration,
2192 bool *contended_compaction, bool *deferred_compaction,
2193 unsigned long *did_some_progress)
2197 #endif /* CONFIG_COMPACTION */
2199 /* Perform direct synchronous page reclaim */
2201 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2202 nodemask_t *nodemask)
2204 struct reclaim_state reclaim_state;
2209 /* We now go into synchronous reclaim */
2210 cpuset_memory_pressure_bump();
2211 current->flags |= PF_MEMALLOC;
2212 lockdep_set_current_reclaim_state(gfp_mask);
2213 reclaim_state.reclaimed_slab = 0;
2214 current->reclaim_state = &reclaim_state;
2216 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2218 current->reclaim_state = NULL;
2219 lockdep_clear_current_reclaim_state();
2220 current->flags &= ~PF_MEMALLOC;
2227 /* The really slow allocator path where we enter direct reclaim */
2228 static inline struct page *
2229 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2230 struct zonelist *zonelist, enum zone_type high_zoneidx,
2231 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2232 int migratetype, unsigned long *did_some_progress)
2234 struct page *page = NULL;
2235 bool drained = false;
2237 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2239 if (unlikely(!(*did_some_progress)))
2242 /* After successful reclaim, reconsider all zones for allocation */
2243 if (IS_ENABLED(CONFIG_NUMA))
2244 zlc_clear_zones_full(zonelist);
2247 page = get_page_from_freelist(gfp_mask, nodemask, order,
2248 zonelist, high_zoneidx,
2249 alloc_flags & ~ALLOC_NO_WATERMARKS,
2250 preferred_zone, migratetype);
2253 * If an allocation failed after direct reclaim, it could be because
2254 * pages are pinned on the per-cpu lists. Drain them and try again
2256 if (!page && !drained) {
2266 * This is called in the allocator slow-path if the allocation request is of
2267 * sufficient urgency to ignore watermarks and take other desperate measures
2269 static inline struct page *
2270 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2271 struct zonelist *zonelist, enum zone_type high_zoneidx,
2272 nodemask_t *nodemask, struct zone *preferred_zone,
2278 page = get_page_from_freelist(gfp_mask, nodemask, order,
2279 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2280 preferred_zone, migratetype);
2282 if (!page && gfp_mask & __GFP_NOFAIL)
2283 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2284 } while (!page && (gfp_mask & __GFP_NOFAIL));
2290 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2291 enum zone_type high_zoneidx,
2292 enum zone_type classzone_idx)
2297 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2298 wakeup_kswapd(zone, order, classzone_idx);
2302 gfp_to_alloc_flags(gfp_t gfp_mask)
2304 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2305 const gfp_t wait = gfp_mask & __GFP_WAIT;
2307 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2308 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2311 * The caller may dip into page reserves a bit more if the caller
2312 * cannot run direct reclaim, or if the caller has realtime scheduling
2313 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2314 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2316 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2320 * Not worth trying to allocate harder for
2321 * __GFP_NOMEMALLOC even if it can't schedule.
2323 if (!(gfp_mask & __GFP_NOMEMALLOC))
2324 alloc_flags |= ALLOC_HARDER;
2326 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2327 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2329 alloc_flags &= ~ALLOC_CPUSET;
2330 } else if (unlikely(rt_task(current)) && !in_interrupt())
2331 alloc_flags |= ALLOC_HARDER;
2333 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2334 if (gfp_mask & __GFP_MEMALLOC)
2335 alloc_flags |= ALLOC_NO_WATERMARKS;
2336 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2337 alloc_flags |= ALLOC_NO_WATERMARKS;
2338 else if (!in_interrupt() &&
2339 ((current->flags & PF_MEMALLOC) ||
2340 unlikely(test_thread_flag(TIF_MEMDIE))))
2341 alloc_flags |= ALLOC_NO_WATERMARKS;
2344 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2345 alloc_flags |= ALLOC_CMA;
2350 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2352 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2355 static inline struct page *
2356 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2357 struct zonelist *zonelist, enum zone_type high_zoneidx,
2358 nodemask_t *nodemask, struct zone *preferred_zone,
2361 const gfp_t wait = gfp_mask & __GFP_WAIT;
2362 struct page *page = NULL;
2364 unsigned long pages_reclaimed = 0;
2365 unsigned long did_some_progress;
2366 bool sync_migration = false;
2367 bool deferred_compaction = false;
2368 bool contended_compaction = false;
2371 * In the slowpath, we sanity check order to avoid ever trying to
2372 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2373 * be using allocators in order of preference for an area that is
2376 if (order >= MAX_ORDER) {
2377 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2382 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2383 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2384 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2385 * using a larger set of nodes after it has established that the
2386 * allowed per node queues are empty and that nodes are
2389 if (IS_ENABLED(CONFIG_NUMA) &&
2390 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2394 if (!(gfp_mask & __GFP_NO_KSWAPD))
2395 wake_all_kswapd(order, zonelist, high_zoneidx,
2396 zone_idx(preferred_zone));
2399 * OK, we're below the kswapd watermark and have kicked background
2400 * reclaim. Now things get more complex, so set up alloc_flags according
2401 * to how we want to proceed.
2403 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2406 * Find the true preferred zone if the allocation is unconstrained by
2409 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2410 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2414 /* This is the last chance, in general, before the goto nopage. */
2415 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2416 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2417 preferred_zone, migratetype);
2421 /* Allocate without watermarks if the context allows */
2422 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2424 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2425 * the allocation is high priority and these type of
2426 * allocations are system rather than user orientated
2428 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2430 page = __alloc_pages_high_priority(gfp_mask, order,
2431 zonelist, high_zoneidx, nodemask,
2432 preferred_zone, migratetype);
2438 /* Atomic allocations - we can't balance anything */
2442 /* Avoid recursion of direct reclaim */
2443 if (current->flags & PF_MEMALLOC)
2446 /* Avoid allocations with no watermarks from looping endlessly */
2447 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2451 * Try direct compaction. The first pass is asynchronous. Subsequent
2452 * attempts after direct reclaim are synchronous
2454 page = __alloc_pages_direct_compact(gfp_mask, order,
2455 zonelist, high_zoneidx,
2457 alloc_flags, preferred_zone,
2458 migratetype, sync_migration,
2459 &contended_compaction,
2460 &deferred_compaction,
2461 &did_some_progress);
2464 sync_migration = true;
2467 * If compaction is deferred for high-order allocations, it is because
2468 * sync compaction recently failed. In this is the case and the caller
2469 * requested a movable allocation that does not heavily disrupt the
2470 * system then fail the allocation instead of entering direct reclaim.
2472 if ((deferred_compaction || contended_compaction) &&
2473 (gfp_mask & __GFP_NO_KSWAPD))
2476 /* Try direct reclaim and then allocating */
2477 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2478 zonelist, high_zoneidx,
2480 alloc_flags, preferred_zone,
2481 migratetype, &did_some_progress);
2486 * If we failed to make any progress reclaiming, then we are
2487 * running out of options and have to consider going OOM
2489 if (!did_some_progress) {
2490 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2491 if (oom_killer_disabled)
2493 /* Coredumps can quickly deplete all memory reserves */
2494 if ((current->flags & PF_DUMPCORE) &&
2495 !(gfp_mask & __GFP_NOFAIL))
2497 page = __alloc_pages_may_oom(gfp_mask, order,
2498 zonelist, high_zoneidx,
2499 nodemask, preferred_zone,
2504 if (!(gfp_mask & __GFP_NOFAIL)) {
2506 * The oom killer is not called for high-order
2507 * allocations that may fail, so if no progress
2508 * is being made, there are no other options and
2509 * retrying is unlikely to help.
2511 if (order > PAGE_ALLOC_COSTLY_ORDER)
2514 * The oom killer is not called for lowmem
2515 * allocations to prevent needlessly killing
2518 if (high_zoneidx < ZONE_NORMAL)
2526 /* Check if we should retry the allocation */
2527 pages_reclaimed += did_some_progress;
2528 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2530 /* Wait for some write requests to complete then retry */
2531 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2535 * High-order allocations do not necessarily loop after
2536 * direct reclaim and reclaim/compaction depends on compaction
2537 * being called after reclaim so call directly if necessary
2539 page = __alloc_pages_direct_compact(gfp_mask, order,
2540 zonelist, high_zoneidx,
2542 alloc_flags, preferred_zone,
2543 migratetype, sync_migration,
2544 &contended_compaction,
2545 &deferred_compaction,
2546 &did_some_progress);
2552 warn_alloc_failed(gfp_mask, order, NULL);
2555 if (kmemcheck_enabled)
2556 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2562 * This is the 'heart' of the zoned buddy allocator.
2565 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2566 struct zonelist *zonelist, nodemask_t *nodemask)
2568 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2569 struct zone *preferred_zone;
2570 struct page *page = NULL;
2571 int migratetype = allocflags_to_migratetype(gfp_mask);
2572 unsigned int cpuset_mems_cookie;
2573 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2574 struct mem_cgroup *memcg = NULL;
2576 gfp_mask &= gfp_allowed_mask;
2578 lockdep_trace_alloc(gfp_mask);
2580 might_sleep_if(gfp_mask & __GFP_WAIT);
2582 if (should_fail_alloc_page(gfp_mask, order))
2586 * Check the zones suitable for the gfp_mask contain at least one
2587 * valid zone. It's possible to have an empty zonelist as a result
2588 * of GFP_THISNODE and a memoryless node
2590 if (unlikely(!zonelist->_zonerefs->zone))
2594 * Will only have any effect when __GFP_KMEMCG is set. This is
2595 * verified in the (always inline) callee
2597 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2601 cpuset_mems_cookie = get_mems_allowed();
2603 /* The preferred zone is used for statistics later */
2604 first_zones_zonelist(zonelist, high_zoneidx,
2605 nodemask ? : &cpuset_current_mems_allowed,
2607 if (!preferred_zone)
2611 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2612 alloc_flags |= ALLOC_CMA;
2614 /* First allocation attempt */
2615 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2616 zonelist, high_zoneidx, alloc_flags,
2617 preferred_zone, migratetype);
2618 if (unlikely(!page))
2619 page = __alloc_pages_slowpath(gfp_mask, order,
2620 zonelist, high_zoneidx, nodemask,
2621 preferred_zone, migratetype);
2623 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2627 * When updating a task's mems_allowed, it is possible to race with
2628 * parallel threads in such a way that an allocation can fail while
2629 * the mask is being updated. If a page allocation is about to fail,
2630 * check if the cpuset changed during allocation and if so, retry.
2632 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2635 memcg_kmem_commit_charge(page, memcg, order);
2639 EXPORT_SYMBOL(__alloc_pages_nodemask);
2642 * Common helper functions.
2644 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2649 * __get_free_pages() returns a 32-bit address, which cannot represent
2652 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2654 page = alloc_pages(gfp_mask, order);
2657 return (unsigned long) page_address(page);
2659 EXPORT_SYMBOL(__get_free_pages);
2661 unsigned long get_zeroed_page(gfp_t gfp_mask)
2663 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2665 EXPORT_SYMBOL(get_zeroed_page);
2667 void __free_pages(struct page *page, unsigned int order)
2669 if (put_page_testzero(page)) {
2671 free_hot_cold_page(page, 0);
2673 __free_pages_ok(page, order);
2677 EXPORT_SYMBOL(__free_pages);
2679 void free_pages(unsigned long addr, unsigned int order)
2682 VM_BUG_ON(!virt_addr_valid((void *)addr));
2683 __free_pages(virt_to_page((void *)addr), order);
2687 EXPORT_SYMBOL(free_pages);
2690 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2691 * pages allocated with __GFP_KMEMCG.
2693 * Those pages are accounted to a particular memcg, embedded in the
2694 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2695 * for that information only to find out that it is NULL for users who have no
2696 * interest in that whatsoever, we provide these functions.
2698 * The caller knows better which flags it relies on.
2700 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2702 memcg_kmem_uncharge_pages(page, order);
2703 __free_pages(page, order);
2706 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2709 VM_BUG_ON(!virt_addr_valid((void *)addr));
2710 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2714 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2717 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2718 unsigned long used = addr + PAGE_ALIGN(size);
2720 split_page(virt_to_page((void *)addr), order);
2721 while (used < alloc_end) {
2726 return (void *)addr;
2730 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2731 * @size: the number of bytes to allocate
2732 * @gfp_mask: GFP flags for the allocation
2734 * This function is similar to alloc_pages(), except that it allocates the
2735 * minimum number of pages to satisfy the request. alloc_pages() can only
2736 * allocate memory in power-of-two pages.
2738 * This function is also limited by MAX_ORDER.
2740 * Memory allocated by this function must be released by free_pages_exact().
2742 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2744 unsigned int order = get_order(size);
2747 addr = __get_free_pages(gfp_mask, order);
2748 return make_alloc_exact(addr, order, size);
2750 EXPORT_SYMBOL(alloc_pages_exact);
2753 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2755 * @nid: the preferred node ID where memory should be allocated
2756 * @size: the number of bytes to allocate
2757 * @gfp_mask: GFP flags for the allocation
2759 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2761 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2764 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2766 unsigned order = get_order(size);
2767 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2770 return make_alloc_exact((unsigned long)page_address(p), order, size);
2772 EXPORT_SYMBOL(alloc_pages_exact_nid);
2775 * free_pages_exact - release memory allocated via alloc_pages_exact()
2776 * @virt: the value returned by alloc_pages_exact.
2777 * @size: size of allocation, same value as passed to alloc_pages_exact().
2779 * Release the memory allocated by a previous call to alloc_pages_exact.
2781 void free_pages_exact(void *virt, size_t size)
2783 unsigned long addr = (unsigned long)virt;
2784 unsigned long end = addr + PAGE_ALIGN(size);
2786 while (addr < end) {
2791 EXPORT_SYMBOL(free_pages_exact);
2793 static unsigned int nr_free_zone_pages(int offset)
2798 /* Just pick one node, since fallback list is circular */
2799 unsigned int sum = 0;
2801 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2803 for_each_zone_zonelist(zone, z, zonelist, offset) {
2804 unsigned long size = zone->present_pages;
2805 unsigned long high = high_wmark_pages(zone);
2814 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2816 unsigned int nr_free_buffer_pages(void)
2818 return nr_free_zone_pages(gfp_zone(GFP_USER));
2820 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2823 * Amount of free RAM allocatable within all zones
2825 unsigned int nr_free_pagecache_pages(void)
2827 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2830 static inline void show_node(struct zone *zone)
2832 if (IS_ENABLED(CONFIG_NUMA))
2833 printk("Node %d ", zone_to_nid(zone));
2836 void si_meminfo(struct sysinfo *val)
2838 val->totalram = totalram_pages;
2840 val->freeram = global_page_state(NR_FREE_PAGES);
2841 val->bufferram = nr_blockdev_pages();
2842 val->totalhigh = totalhigh_pages;
2843 val->freehigh = nr_free_highpages();
2844 val->mem_unit = PAGE_SIZE;
2847 EXPORT_SYMBOL(si_meminfo);
2850 void si_meminfo_node(struct sysinfo *val, int nid)
2852 pg_data_t *pgdat = NODE_DATA(nid);
2854 val->totalram = pgdat->node_present_pages;
2855 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2856 #ifdef CONFIG_HIGHMEM
2857 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2858 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2864 val->mem_unit = PAGE_SIZE;
2869 * Determine whether the node should be displayed or not, depending on whether
2870 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2872 bool skip_free_areas_node(unsigned int flags, int nid)
2875 unsigned int cpuset_mems_cookie;
2877 if (!(flags & SHOW_MEM_FILTER_NODES))
2881 cpuset_mems_cookie = get_mems_allowed();
2882 ret = !node_isset(nid, cpuset_current_mems_allowed);
2883 } while (!put_mems_allowed(cpuset_mems_cookie));
2888 #define K(x) ((x) << (PAGE_SHIFT-10))
2890 static void show_migration_types(unsigned char type)
2892 static const char types[MIGRATE_TYPES] = {
2893 [MIGRATE_UNMOVABLE] = 'U',
2894 [MIGRATE_RECLAIMABLE] = 'E',
2895 [MIGRATE_MOVABLE] = 'M',
2896 [MIGRATE_RESERVE] = 'R',
2898 [MIGRATE_CMA] = 'C',
2900 [MIGRATE_ISOLATE] = 'I',
2902 char tmp[MIGRATE_TYPES + 1];
2906 for (i = 0; i < MIGRATE_TYPES; i++) {
2907 if (type & (1 << i))
2912 printk("(%s) ", tmp);
2916 * Show free area list (used inside shift_scroll-lock stuff)
2917 * We also calculate the percentage fragmentation. We do this by counting the
2918 * memory on each free list with the exception of the first item on the list.
2919 * Suppresses nodes that are not allowed by current's cpuset if
2920 * SHOW_MEM_FILTER_NODES is passed.
2922 void show_free_areas(unsigned int filter)
2927 for_each_populated_zone(zone) {
2928 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2931 printk("%s per-cpu:\n", zone->name);
2933 for_each_online_cpu(cpu) {
2934 struct per_cpu_pageset *pageset;
2936 pageset = per_cpu_ptr(zone->pageset, cpu);
2938 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2939 cpu, pageset->pcp.high,
2940 pageset->pcp.batch, pageset->pcp.count);
2944 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2945 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2947 " dirty:%lu writeback:%lu unstable:%lu\n"
2948 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2949 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2951 global_page_state(NR_ACTIVE_ANON),
2952 global_page_state(NR_INACTIVE_ANON),
2953 global_page_state(NR_ISOLATED_ANON),
2954 global_page_state(NR_ACTIVE_FILE),
2955 global_page_state(NR_INACTIVE_FILE),
2956 global_page_state(NR_ISOLATED_FILE),
2957 global_page_state(NR_UNEVICTABLE),
2958 global_page_state(NR_FILE_DIRTY),
2959 global_page_state(NR_WRITEBACK),
2960 global_page_state(NR_UNSTABLE_NFS),
2961 global_page_state(NR_FREE_PAGES),
2962 global_page_state(NR_SLAB_RECLAIMABLE),
2963 global_page_state(NR_SLAB_UNRECLAIMABLE),
2964 global_page_state(NR_FILE_MAPPED),
2965 global_page_state(NR_SHMEM),
2966 global_page_state(NR_PAGETABLE),
2967 global_page_state(NR_BOUNCE),
2968 global_page_state(NR_FREE_CMA_PAGES));
2970 for_each_populated_zone(zone) {
2973 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2981 " active_anon:%lukB"
2982 " inactive_anon:%lukB"
2983 " active_file:%lukB"
2984 " inactive_file:%lukB"
2985 " unevictable:%lukB"
2986 " isolated(anon):%lukB"
2987 " isolated(file):%lukB"
2995 " slab_reclaimable:%lukB"
2996 " slab_unreclaimable:%lukB"
2997 " kernel_stack:%lukB"
3002 " writeback_tmp:%lukB"
3003 " pages_scanned:%lu"
3004 " all_unreclaimable? %s"
3007 K(zone_page_state(zone, NR_FREE_PAGES)),
3008 K(min_wmark_pages(zone)),
3009 K(low_wmark_pages(zone)),
3010 K(high_wmark_pages(zone)),
3011 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3012 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3013 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3014 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3015 K(zone_page_state(zone, NR_UNEVICTABLE)),
3016 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3017 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3018 K(zone->present_pages),
3019 K(zone->managed_pages),
3020 K(zone_page_state(zone, NR_MLOCK)),
3021 K(zone_page_state(zone, NR_FILE_DIRTY)),
3022 K(zone_page_state(zone, NR_WRITEBACK)),
3023 K(zone_page_state(zone, NR_FILE_MAPPED)),
3024 K(zone_page_state(zone, NR_SHMEM)),
3025 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3026 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3027 zone_page_state(zone, NR_KERNEL_STACK) *
3029 K(zone_page_state(zone, NR_PAGETABLE)),
3030 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3031 K(zone_page_state(zone, NR_BOUNCE)),
3032 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3033 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3034 zone->pages_scanned,
3035 (zone->all_unreclaimable ? "yes" : "no")
3037 printk("lowmem_reserve[]:");
3038 for (i = 0; i < MAX_NR_ZONES; i++)
3039 printk(" %lu", zone->lowmem_reserve[i]);
3043 for_each_populated_zone(zone) {
3044 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3045 unsigned char types[MAX_ORDER];
3047 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3050 printk("%s: ", zone->name);
3052 spin_lock_irqsave(&zone->lock, flags);
3053 for (order = 0; order < MAX_ORDER; order++) {
3054 struct free_area *area = &zone->free_area[order];
3057 nr[order] = area->nr_free;
3058 total += nr[order] << order;
3061 for (type = 0; type < MIGRATE_TYPES; type++) {
3062 if (!list_empty(&area->free_list[type]))
3063 types[order] |= 1 << type;
3066 spin_unlock_irqrestore(&zone->lock, flags);
3067 for (order = 0; order < MAX_ORDER; order++) {
3068 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3070 show_migration_types(types[order]);
3072 printk("= %lukB\n", K(total));
3075 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3077 show_swap_cache_info();
3080 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3082 zoneref->zone = zone;
3083 zoneref->zone_idx = zone_idx(zone);
3087 * Builds allocation fallback zone lists.
3089 * Add all populated zones of a node to the zonelist.
3091 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3092 int nr_zones, enum zone_type zone_type)
3096 BUG_ON(zone_type >= MAX_NR_ZONES);
3101 zone = pgdat->node_zones + zone_type;
3102 if (populated_zone(zone)) {
3103 zoneref_set_zone(zone,
3104 &zonelist->_zonerefs[nr_zones++]);
3105 check_highest_zone(zone_type);
3108 } while (zone_type);
3115 * 0 = automatic detection of better ordering.
3116 * 1 = order by ([node] distance, -zonetype)
3117 * 2 = order by (-zonetype, [node] distance)
3119 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3120 * the same zonelist. So only NUMA can configure this param.
3122 #define ZONELIST_ORDER_DEFAULT 0
3123 #define ZONELIST_ORDER_NODE 1
3124 #define ZONELIST_ORDER_ZONE 2
3126 /* zonelist order in the kernel.
3127 * set_zonelist_order() will set this to NODE or ZONE.
3129 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3130 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3134 /* The value user specified ....changed by config */
3135 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3136 /* string for sysctl */
3137 #define NUMA_ZONELIST_ORDER_LEN 16
3138 char numa_zonelist_order[16] = "default";
3141 * interface for configure zonelist ordering.
3142 * command line option "numa_zonelist_order"
3143 * = "[dD]efault - default, automatic configuration.
3144 * = "[nN]ode - order by node locality, then by zone within node
3145 * = "[zZ]one - order by zone, then by locality within zone
3148 static int __parse_numa_zonelist_order(char *s)
3150 if (*s == 'd' || *s == 'D') {
3151 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3152 } else if (*s == 'n' || *s == 'N') {
3153 user_zonelist_order = ZONELIST_ORDER_NODE;
3154 } else if (*s == 'z' || *s == 'Z') {
3155 user_zonelist_order = ZONELIST_ORDER_ZONE;
3158 "Ignoring invalid numa_zonelist_order value: "
3165 static __init int setup_numa_zonelist_order(char *s)
3172 ret = __parse_numa_zonelist_order(s);
3174 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3178 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3181 * sysctl handler for numa_zonelist_order
3183 int numa_zonelist_order_handler(ctl_table *table, int write,
3184 void __user *buffer, size_t *length,
3187 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3189 static DEFINE_MUTEX(zl_order_mutex);
3191 mutex_lock(&zl_order_mutex);
3193 strcpy(saved_string, (char*)table->data);
3194 ret = proc_dostring(table, write, buffer, length, ppos);
3198 int oldval = user_zonelist_order;
3199 if (__parse_numa_zonelist_order((char*)table->data)) {
3201 * bogus value. restore saved string
3203 strncpy((char*)table->data, saved_string,
3204 NUMA_ZONELIST_ORDER_LEN);
3205 user_zonelist_order = oldval;
3206 } else if (oldval != user_zonelist_order) {
3207 mutex_lock(&zonelists_mutex);
3208 build_all_zonelists(NULL, NULL);
3209 mutex_unlock(&zonelists_mutex);
3213 mutex_unlock(&zl_order_mutex);
3218 #define MAX_NODE_LOAD (nr_online_nodes)
3219 static int node_load[MAX_NUMNODES];
3222 * find_next_best_node - find the next node that should appear in a given node's fallback list
3223 * @node: node whose fallback list we're appending
3224 * @used_node_mask: nodemask_t of already used nodes
3226 * We use a number of factors to determine which is the next node that should
3227 * appear on a given node's fallback list. The node should not have appeared
3228 * already in @node's fallback list, and it should be the next closest node
3229 * according to the distance array (which contains arbitrary distance values
3230 * from each node to each node in the system), and should also prefer nodes
3231 * with no CPUs, since presumably they'll have very little allocation pressure
3232 * on them otherwise.
3233 * It returns -1 if no node is found.
3235 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3238 int min_val = INT_MAX;
3240 const struct cpumask *tmp = cpumask_of_node(0);
3242 /* Use the local node if we haven't already */
3243 if (!node_isset(node, *used_node_mask)) {
3244 node_set(node, *used_node_mask);
3248 for_each_node_state(n, N_MEMORY) {
3250 /* Don't want a node to appear more than once */
3251 if (node_isset(n, *used_node_mask))
3254 /* Use the distance array to find the distance */
3255 val = node_distance(node, n);
3257 /* Penalize nodes under us ("prefer the next node") */
3260 /* Give preference to headless and unused nodes */
3261 tmp = cpumask_of_node(n);
3262 if (!cpumask_empty(tmp))
3263 val += PENALTY_FOR_NODE_WITH_CPUS;
3265 /* Slight preference for less loaded node */
3266 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3267 val += node_load[n];
3269 if (val < min_val) {
3276 node_set(best_node, *used_node_mask);
3283 * Build zonelists ordered by node and zones within node.
3284 * This results in maximum locality--normal zone overflows into local
3285 * DMA zone, if any--but risks exhausting DMA zone.
3287 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3290 struct zonelist *zonelist;
3292 zonelist = &pgdat->node_zonelists[0];
3293 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3295 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3297 zonelist->_zonerefs[j].zone = NULL;
3298 zonelist->_zonerefs[j].zone_idx = 0;
3302 * Build gfp_thisnode zonelists
3304 static void build_thisnode_zonelists(pg_data_t *pgdat)
3307 struct zonelist *zonelist;
3309 zonelist = &pgdat->node_zonelists[1];
3310 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3311 zonelist->_zonerefs[j].zone = NULL;
3312 zonelist->_zonerefs[j].zone_idx = 0;
3316 * Build zonelists ordered by zone and nodes within zones.
3317 * This results in conserving DMA zone[s] until all Normal memory is
3318 * exhausted, but results in overflowing to remote node while memory
3319 * may still exist in local DMA zone.
3321 static int node_order[MAX_NUMNODES];
3323 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3326 int zone_type; /* needs to be signed */
3328 struct zonelist *zonelist;
3330 zonelist = &pgdat->node_zonelists[0];
3332 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3333 for (j = 0; j < nr_nodes; j++) {
3334 node = node_order[j];
3335 z = &NODE_DATA(node)->node_zones[zone_type];
3336 if (populated_zone(z)) {
3338 &zonelist->_zonerefs[pos++]);
3339 check_highest_zone(zone_type);
3343 zonelist->_zonerefs[pos].zone = NULL;
3344 zonelist->_zonerefs[pos].zone_idx = 0;
3347 static int default_zonelist_order(void)
3350 unsigned long low_kmem_size,total_size;
3354 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3355 * If they are really small and used heavily, the system can fall
3356 * into OOM very easily.
3357 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3359 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3362 for_each_online_node(nid) {
3363 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3364 z = &NODE_DATA(nid)->node_zones[zone_type];
3365 if (populated_zone(z)) {
3366 if (zone_type < ZONE_NORMAL)
3367 low_kmem_size += z->present_pages;
3368 total_size += z->present_pages;
3369 } else if (zone_type == ZONE_NORMAL) {
3371 * If any node has only lowmem, then node order
3372 * is preferred to allow kernel allocations
3373 * locally; otherwise, they can easily infringe
3374 * on other nodes when there is an abundance of
3375 * lowmem available to allocate from.
3377 return ZONELIST_ORDER_NODE;
3381 if (!low_kmem_size || /* there are no DMA area. */
3382 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3383 return ZONELIST_ORDER_NODE;
3385 * look into each node's config.
3386 * If there is a node whose DMA/DMA32 memory is very big area on
3387 * local memory, NODE_ORDER may be suitable.
3389 average_size = total_size /
3390 (nodes_weight(node_states[N_MEMORY]) + 1);
3391 for_each_online_node(nid) {
3394 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3395 z = &NODE_DATA(nid)->node_zones[zone_type];
3396 if (populated_zone(z)) {
3397 if (zone_type < ZONE_NORMAL)
3398 low_kmem_size += z->present_pages;
3399 total_size += z->present_pages;
3402 if (low_kmem_size &&
3403 total_size > average_size && /* ignore small node */
3404 low_kmem_size > total_size * 70/100)
3405 return ZONELIST_ORDER_NODE;
3407 return ZONELIST_ORDER_ZONE;
3410 static void set_zonelist_order(void)
3412 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3413 current_zonelist_order = default_zonelist_order();
3415 current_zonelist_order = user_zonelist_order;
3418 static void build_zonelists(pg_data_t *pgdat)
3422 nodemask_t used_mask;
3423 int local_node, prev_node;
3424 struct zonelist *zonelist;
3425 int order = current_zonelist_order;
3427 /* initialize zonelists */
3428 for (i = 0; i < MAX_ZONELISTS; i++) {
3429 zonelist = pgdat->node_zonelists + i;
3430 zonelist->_zonerefs[0].zone = NULL;
3431 zonelist->_zonerefs[0].zone_idx = 0;
3434 /* NUMA-aware ordering of nodes */
3435 local_node = pgdat->node_id;
3436 load = nr_online_nodes;
3437 prev_node = local_node;
3438 nodes_clear(used_mask);
3440 memset(node_order, 0, sizeof(node_order));
3443 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3445 * We don't want to pressure a particular node.
3446 * So adding penalty to the first node in same
3447 * distance group to make it round-robin.
3449 if (node_distance(local_node, node) !=
3450 node_distance(local_node, prev_node))
3451 node_load[node] = load;
3455 if (order == ZONELIST_ORDER_NODE)
3456 build_zonelists_in_node_order(pgdat, node);
3458 node_order[j++] = node; /* remember order */
3461 if (order == ZONELIST_ORDER_ZONE) {
3462 /* calculate node order -- i.e., DMA last! */
3463 build_zonelists_in_zone_order(pgdat, j);
3466 build_thisnode_zonelists(pgdat);
3469 /* Construct the zonelist performance cache - see further mmzone.h */
3470 static void build_zonelist_cache(pg_data_t *pgdat)
3472 struct zonelist *zonelist;
3473 struct zonelist_cache *zlc;
3476 zonelist = &pgdat->node_zonelists[0];
3477 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3478 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3479 for (z = zonelist->_zonerefs; z->zone; z++)
3480 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3483 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3485 * Return node id of node used for "local" allocations.
3486 * I.e., first node id of first zone in arg node's generic zonelist.
3487 * Used for initializing percpu 'numa_mem', which is used primarily
3488 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3490 int local_memory_node(int node)
3494 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3495 gfp_zone(GFP_KERNEL),
3502 #else /* CONFIG_NUMA */
3504 static void set_zonelist_order(void)
3506 current_zonelist_order = ZONELIST_ORDER_ZONE;
3509 static void build_zonelists(pg_data_t *pgdat)
3511 int node, local_node;
3513 struct zonelist *zonelist;
3515 local_node = pgdat->node_id;
3517 zonelist = &pgdat->node_zonelists[0];
3518 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3521 * Now we build the zonelist so that it contains the zones
3522 * of all the other nodes.
3523 * We don't want to pressure a particular node, so when
3524 * building the zones for node N, we make sure that the
3525 * zones coming right after the local ones are those from
3526 * node N+1 (modulo N)
3528 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3529 if (!node_online(node))
3531 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3534 for (node = 0; node < local_node; node++) {
3535 if (!node_online(node))
3537 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3541 zonelist->_zonerefs[j].zone = NULL;
3542 zonelist->_zonerefs[j].zone_idx = 0;
3545 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3546 static void build_zonelist_cache(pg_data_t *pgdat)
3548 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3551 #endif /* CONFIG_NUMA */
3554 * Boot pageset table. One per cpu which is going to be used for all
3555 * zones and all nodes. The parameters will be set in such a way
3556 * that an item put on a list will immediately be handed over to
3557 * the buddy list. This is safe since pageset manipulation is done
3558 * with interrupts disabled.
3560 * The boot_pagesets must be kept even after bootup is complete for
3561 * unused processors and/or zones. They do play a role for bootstrapping
3562 * hotplugged processors.
3564 * zoneinfo_show() and maybe other functions do
3565 * not check if the processor is online before following the pageset pointer.
3566 * Other parts of the kernel may not check if the zone is available.
3568 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3569 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3570 static void setup_zone_pageset(struct zone *zone);
3573 * Global mutex to protect against size modification of zonelists
3574 * as well as to serialize pageset setup for the new populated zone.
3576 DEFINE_MUTEX(zonelists_mutex);
3578 /* return values int ....just for stop_machine() */
3579 static int __build_all_zonelists(void *data)
3583 pg_data_t *self = data;
3586 memset(node_load, 0, sizeof(node_load));
3589 if (self && !node_online(self->node_id)) {
3590 build_zonelists(self);
3591 build_zonelist_cache(self);
3594 for_each_online_node(nid) {
3595 pg_data_t *pgdat = NODE_DATA(nid);
3597 build_zonelists(pgdat);
3598 build_zonelist_cache(pgdat);
3602 * Initialize the boot_pagesets that are going to be used
3603 * for bootstrapping processors. The real pagesets for
3604 * each zone will be allocated later when the per cpu
3605 * allocator is available.
3607 * boot_pagesets are used also for bootstrapping offline
3608 * cpus if the system is already booted because the pagesets
3609 * are needed to initialize allocators on a specific cpu too.
3610 * F.e. the percpu allocator needs the page allocator which
3611 * needs the percpu allocator in order to allocate its pagesets
3612 * (a chicken-egg dilemma).
3614 for_each_possible_cpu(cpu) {
3615 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3617 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3619 * We now know the "local memory node" for each node--
3620 * i.e., the node of the first zone in the generic zonelist.
3621 * Set up numa_mem percpu variable for on-line cpus. During
3622 * boot, only the boot cpu should be on-line; we'll init the
3623 * secondary cpus' numa_mem as they come on-line. During
3624 * node/memory hotplug, we'll fixup all on-line cpus.
3626 if (cpu_online(cpu))
3627 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3635 * Called with zonelists_mutex held always
3636 * unless system_state == SYSTEM_BOOTING.
3638 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3640 set_zonelist_order();
3642 if (system_state == SYSTEM_BOOTING) {
3643 __build_all_zonelists(NULL);
3644 mminit_verify_zonelist();
3645 cpuset_init_current_mems_allowed();
3647 /* we have to stop all cpus to guarantee there is no user
3649 #ifdef CONFIG_MEMORY_HOTPLUG
3651 setup_zone_pageset(zone);
3653 stop_machine(__build_all_zonelists, pgdat, NULL);
3654 /* cpuset refresh routine should be here */
3656 vm_total_pages = nr_free_pagecache_pages();
3658 * Disable grouping by mobility if the number of pages in the
3659 * system is too low to allow the mechanism to work. It would be
3660 * more accurate, but expensive to check per-zone. This check is
3661 * made on memory-hotadd so a system can start with mobility
3662 * disabled and enable it later
3664 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3665 page_group_by_mobility_disabled = 1;
3667 page_group_by_mobility_disabled = 0;
3669 printk("Built %i zonelists in %s order, mobility grouping %s. "
3670 "Total pages: %ld\n",
3672 zonelist_order_name[current_zonelist_order],
3673 page_group_by_mobility_disabled ? "off" : "on",
3676 printk("Policy zone: %s\n", zone_names[policy_zone]);
3681 * Helper functions to size the waitqueue hash table.
3682 * Essentially these want to choose hash table sizes sufficiently
3683 * large so that collisions trying to wait on pages are rare.
3684 * But in fact, the number of active page waitqueues on typical
3685 * systems is ridiculously low, less than 200. So this is even
3686 * conservative, even though it seems large.
3688 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3689 * waitqueues, i.e. the size of the waitq table given the number of pages.
3691 #define PAGES_PER_WAITQUEUE 256
3693 #ifndef CONFIG_MEMORY_HOTPLUG
3694 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3696 unsigned long size = 1;
3698 pages /= PAGES_PER_WAITQUEUE;
3700 while (size < pages)
3704 * Once we have dozens or even hundreds of threads sleeping
3705 * on IO we've got bigger problems than wait queue collision.
3706 * Limit the size of the wait table to a reasonable size.
3708 size = min(size, 4096UL);
3710 return max(size, 4UL);
3714 * A zone's size might be changed by hot-add, so it is not possible to determine
3715 * a suitable size for its wait_table. So we use the maximum size now.
3717 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3719 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3720 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3721 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3723 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3724 * or more by the traditional way. (See above). It equals:
3726 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3727 * ia64(16K page size) : = ( 8G + 4M)byte.
3728 * powerpc (64K page size) : = (32G +16M)byte.
3730 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3737 * This is an integer logarithm so that shifts can be used later
3738 * to extract the more random high bits from the multiplicative
3739 * hash function before the remainder is taken.
3741 static inline unsigned long wait_table_bits(unsigned long size)
3746 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3749 * Check if a pageblock contains reserved pages
3751 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3755 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3756 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3763 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3764 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3765 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3766 * higher will lead to a bigger reserve which will get freed as contiguous
3767 * blocks as reclaim kicks in
3769 static void setup_zone_migrate_reserve(struct zone *zone)
3771 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3773 unsigned long block_migratetype;
3777 * Get the start pfn, end pfn and the number of blocks to reserve
3778 * We have to be careful to be aligned to pageblock_nr_pages to
3779 * make sure that we always check pfn_valid for the first page in
3782 start_pfn = zone->zone_start_pfn;
3783 end_pfn = start_pfn + zone->spanned_pages;
3784 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3785 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3789 * Reserve blocks are generally in place to help high-order atomic
3790 * allocations that are short-lived. A min_free_kbytes value that
3791 * would result in more than 2 reserve blocks for atomic allocations
3792 * is assumed to be in place to help anti-fragmentation for the
3793 * future allocation of hugepages at runtime.
3795 reserve = min(2, reserve);
3797 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3798 if (!pfn_valid(pfn))
3800 page = pfn_to_page(pfn);
3802 /* Watch out for overlapping nodes */
3803 if (page_to_nid(page) != zone_to_nid(zone))
3806 block_migratetype = get_pageblock_migratetype(page);
3808 /* Only test what is necessary when the reserves are not met */
3811 * Blocks with reserved pages will never free, skip
3814 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3815 if (pageblock_is_reserved(pfn, block_end_pfn))
3818 /* If this block is reserved, account for it */
3819 if (block_migratetype == MIGRATE_RESERVE) {
3824 /* Suitable for reserving if this block is movable */
3825 if (block_migratetype == MIGRATE_MOVABLE) {
3826 set_pageblock_migratetype(page,
3828 move_freepages_block(zone, page,
3836 * If the reserve is met and this is a previous reserved block,
3839 if (block_migratetype == MIGRATE_RESERVE) {
3840 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3841 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3847 * Initially all pages are reserved - free ones are freed
3848 * up by free_all_bootmem() once the early boot process is
3849 * done. Non-atomic initialization, single-pass.
3851 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3852 unsigned long start_pfn, enum memmap_context context)
3855 unsigned long end_pfn = start_pfn + size;
3859 if (highest_memmap_pfn < end_pfn - 1)
3860 highest_memmap_pfn = end_pfn - 1;
3862 z = &NODE_DATA(nid)->node_zones[zone];
3863 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3865 * There can be holes in boot-time mem_map[]s
3866 * handed to this function. They do not
3867 * exist on hotplugged memory.
3869 if (context == MEMMAP_EARLY) {
3870 if (!early_pfn_valid(pfn))
3872 if (!early_pfn_in_nid(pfn, nid))
3875 page = pfn_to_page(pfn);
3876 set_page_links(page, zone, nid, pfn);
3877 mminit_verify_page_links(page, zone, nid, pfn);
3878 init_page_count(page);
3879 reset_page_mapcount(page);
3880 reset_page_last_nid(page);
3881 SetPageReserved(page);
3883 * Mark the block movable so that blocks are reserved for
3884 * movable at startup. This will force kernel allocations
3885 * to reserve their blocks rather than leaking throughout
3886 * the address space during boot when many long-lived
3887 * kernel allocations are made. Later some blocks near
3888 * the start are marked MIGRATE_RESERVE by
3889 * setup_zone_migrate_reserve()
3891 * bitmap is created for zone's valid pfn range. but memmap
3892 * can be created for invalid pages (for alignment)
3893 * check here not to call set_pageblock_migratetype() against
3896 if ((z->zone_start_pfn <= pfn)
3897 && (pfn < z->zone_start_pfn + z->spanned_pages)
3898 && !(pfn & (pageblock_nr_pages - 1)))
3899 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3901 INIT_LIST_HEAD(&page->lru);
3902 #ifdef WANT_PAGE_VIRTUAL
3903 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3904 if (!is_highmem_idx(zone))
3905 set_page_address(page, __va(pfn << PAGE_SHIFT));
3910 static void __meminit zone_init_free_lists(struct zone *zone)
3913 for_each_migratetype_order(order, t) {
3914 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3915 zone->free_area[order].nr_free = 0;
3919 #ifndef __HAVE_ARCH_MEMMAP_INIT
3920 #define memmap_init(size, nid, zone, start_pfn) \
3921 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3924 static int __meminit zone_batchsize(struct zone *zone)
3930 * The per-cpu-pages pools are set to around 1000th of the
3931 * size of the zone. But no more than 1/2 of a meg.
3933 * OK, so we don't know how big the cache is. So guess.
3935 batch = zone->present_pages / 1024;
3936 if (batch * PAGE_SIZE > 512 * 1024)
3937 batch = (512 * 1024) / PAGE_SIZE;
3938 batch /= 4; /* We effectively *= 4 below */
3943 * Clamp the batch to a 2^n - 1 value. Having a power
3944 * of 2 value was found to be more likely to have
3945 * suboptimal cache aliasing properties in some cases.
3947 * For example if 2 tasks are alternately allocating
3948 * batches of pages, one task can end up with a lot
3949 * of pages of one half of the possible page colors
3950 * and the other with pages of the other colors.
3952 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3957 /* The deferral and batching of frees should be suppressed under NOMMU
3960 * The problem is that NOMMU needs to be able to allocate large chunks
3961 * of contiguous memory as there's no hardware page translation to
3962 * assemble apparent contiguous memory from discontiguous pages.
3964 * Queueing large contiguous runs of pages for batching, however,
3965 * causes the pages to actually be freed in smaller chunks. As there
3966 * can be a significant delay between the individual batches being
3967 * recycled, this leads to the once large chunks of space being
3968 * fragmented and becoming unavailable for high-order allocations.
3974 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3976 struct per_cpu_pages *pcp;
3979 memset(p, 0, sizeof(*p));
3983 pcp->high = 6 * batch;
3984 pcp->batch = max(1UL, 1 * batch);
3985 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3986 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3990 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3991 * to the value high for the pageset p.
3994 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3997 struct per_cpu_pages *pcp;
4001 pcp->batch = max(1UL, high/4);
4002 if ((high/4) > (PAGE_SHIFT * 8))
4003 pcp->batch = PAGE_SHIFT * 8;
4006 static void __meminit setup_zone_pageset(struct zone *zone)
4010 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4012 for_each_possible_cpu(cpu) {
4013 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4015 setup_pageset(pcp, zone_batchsize(zone));
4017 if (percpu_pagelist_fraction)
4018 setup_pagelist_highmark(pcp,
4019 (zone->present_pages /
4020 percpu_pagelist_fraction));
4025 * Allocate per cpu pagesets and initialize them.
4026 * Before this call only boot pagesets were available.
4028 void __init setup_per_cpu_pageset(void)
4032 for_each_populated_zone(zone)
4033 setup_zone_pageset(zone);
4036 static noinline __init_refok
4037 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4040 struct pglist_data *pgdat = zone->zone_pgdat;
4044 * The per-page waitqueue mechanism uses hashed waitqueues
4047 zone->wait_table_hash_nr_entries =
4048 wait_table_hash_nr_entries(zone_size_pages);
4049 zone->wait_table_bits =
4050 wait_table_bits(zone->wait_table_hash_nr_entries);
4051 alloc_size = zone->wait_table_hash_nr_entries
4052 * sizeof(wait_queue_head_t);
4054 if (!slab_is_available()) {
4055 zone->wait_table = (wait_queue_head_t *)
4056 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4059 * This case means that a zone whose size was 0 gets new memory
4060 * via memory hot-add.
4061 * But it may be the case that a new node was hot-added. In
4062 * this case vmalloc() will not be able to use this new node's
4063 * memory - this wait_table must be initialized to use this new
4064 * node itself as well.
4065 * To use this new node's memory, further consideration will be
4068 zone->wait_table = vmalloc(alloc_size);
4070 if (!zone->wait_table)
4073 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4074 init_waitqueue_head(zone->wait_table + i);
4079 static __meminit void zone_pcp_init(struct zone *zone)
4082 * per cpu subsystem is not up at this point. The following code
4083 * relies on the ability of the linker to provide the
4084 * offset of a (static) per cpu variable into the per cpu area.
4086 zone->pageset = &boot_pageset;
4088 if (zone->present_pages)
4089 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4090 zone->name, zone->present_pages,
4091 zone_batchsize(zone));
4094 int __meminit init_currently_empty_zone(struct zone *zone,
4095 unsigned long zone_start_pfn,
4097 enum memmap_context context)
4099 struct pglist_data *pgdat = zone->zone_pgdat;
4101 ret = zone_wait_table_init(zone, size);
4104 pgdat->nr_zones = zone_idx(zone) + 1;
4106 zone->zone_start_pfn = zone_start_pfn;
4108 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4109 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4111 (unsigned long)zone_idx(zone),
4112 zone_start_pfn, (zone_start_pfn + size));
4114 zone_init_free_lists(zone);
4119 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4120 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4122 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4123 * Architectures may implement their own version but if add_active_range()
4124 * was used and there are no special requirements, this is a convenient
4127 int __meminit __early_pfn_to_nid(unsigned long pfn)
4129 unsigned long start_pfn, end_pfn;
4132 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4133 if (start_pfn <= pfn && pfn < end_pfn)
4135 /* This is a memory hole */
4138 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4140 int __meminit early_pfn_to_nid(unsigned long pfn)
4144 nid = __early_pfn_to_nid(pfn);
4147 /* just returns 0 */
4151 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4152 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4156 nid = __early_pfn_to_nid(pfn);
4157 if (nid >= 0 && nid != node)
4164 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4165 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4166 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4168 * If an architecture guarantees that all ranges registered with
4169 * add_active_ranges() contain no holes and may be freed, this
4170 * this function may be used instead of calling free_bootmem() manually.
4172 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4174 unsigned long start_pfn, end_pfn;
4177 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4178 start_pfn = min(start_pfn, max_low_pfn);
4179 end_pfn = min(end_pfn, max_low_pfn);
4181 if (start_pfn < end_pfn)
4182 free_bootmem_node(NODE_DATA(this_nid),
4183 PFN_PHYS(start_pfn),
4184 (end_pfn - start_pfn) << PAGE_SHIFT);
4189 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4190 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4192 * If an architecture guarantees that all ranges registered with
4193 * add_active_ranges() contain no holes and may be freed, this
4194 * function may be used instead of calling memory_present() manually.
4196 void __init sparse_memory_present_with_active_regions(int nid)
4198 unsigned long start_pfn, end_pfn;
4201 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4202 memory_present(this_nid, start_pfn, end_pfn);
4206 * get_pfn_range_for_nid - Return the start and end page frames for a node
4207 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4208 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4209 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4211 * It returns the start and end page frame of a node based on information
4212 * provided by an arch calling add_active_range(). If called for a node
4213 * with no available memory, a warning is printed and the start and end
4216 void __meminit get_pfn_range_for_nid(unsigned int nid,
4217 unsigned long *start_pfn, unsigned long *end_pfn)
4219 unsigned long this_start_pfn, this_end_pfn;
4225 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4226 *start_pfn = min(*start_pfn, this_start_pfn);
4227 *end_pfn = max(*end_pfn, this_end_pfn);
4230 if (*start_pfn == -1UL)
4235 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4236 * assumption is made that zones within a node are ordered in monotonic
4237 * increasing memory addresses so that the "highest" populated zone is used
4239 static void __init find_usable_zone_for_movable(void)
4242 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4243 if (zone_index == ZONE_MOVABLE)
4246 if (arch_zone_highest_possible_pfn[zone_index] >
4247 arch_zone_lowest_possible_pfn[zone_index])
4251 VM_BUG_ON(zone_index == -1);
4252 movable_zone = zone_index;
4256 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4257 * because it is sized independent of architecture. Unlike the other zones,
4258 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4259 * in each node depending on the size of each node and how evenly kernelcore
4260 * is distributed. This helper function adjusts the zone ranges
4261 * provided by the architecture for a given node by using the end of the
4262 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4263 * zones within a node are in order of monotonic increases memory addresses
4265 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4266 unsigned long zone_type,
4267 unsigned long node_start_pfn,
4268 unsigned long node_end_pfn,
4269 unsigned long *zone_start_pfn,
4270 unsigned long *zone_end_pfn)
4272 /* Only adjust if ZONE_MOVABLE is on this node */
4273 if (zone_movable_pfn[nid]) {
4274 /* Size ZONE_MOVABLE */
4275 if (zone_type == ZONE_MOVABLE) {
4276 *zone_start_pfn = zone_movable_pfn[nid];
4277 *zone_end_pfn = min(node_end_pfn,
4278 arch_zone_highest_possible_pfn[movable_zone]);
4280 /* Adjust for ZONE_MOVABLE starting within this range */
4281 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4282 *zone_end_pfn > zone_movable_pfn[nid]) {
4283 *zone_end_pfn = zone_movable_pfn[nid];
4285 /* Check if this whole range is within ZONE_MOVABLE */
4286 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4287 *zone_start_pfn = *zone_end_pfn;
4292 * Return the number of pages a zone spans in a node, including holes
4293 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4295 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4296 unsigned long zone_type,
4297 unsigned long *ignored)
4299 unsigned long node_start_pfn, node_end_pfn;
4300 unsigned long zone_start_pfn, zone_end_pfn;
4302 /* Get the start and end of the node and zone */
4303 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4304 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4305 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4306 adjust_zone_range_for_zone_movable(nid, zone_type,
4307 node_start_pfn, node_end_pfn,
4308 &zone_start_pfn, &zone_end_pfn);
4310 /* Check that this node has pages within the zone's required range */
4311 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4314 /* Move the zone boundaries inside the node if necessary */
4315 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4316 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4318 /* Return the spanned pages */
4319 return zone_end_pfn - zone_start_pfn;
4323 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4324 * then all holes in the requested range will be accounted for.
4326 unsigned long __meminit __absent_pages_in_range(int nid,
4327 unsigned long range_start_pfn,
4328 unsigned long range_end_pfn)
4330 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4331 unsigned long start_pfn, end_pfn;
4334 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4335 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4336 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4337 nr_absent -= end_pfn - start_pfn;
4343 * absent_pages_in_range - Return number of page frames in holes within a range
4344 * @start_pfn: The start PFN to start searching for holes
4345 * @end_pfn: The end PFN to stop searching for holes
4347 * It returns the number of pages frames in memory holes within a range.
4349 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4350 unsigned long end_pfn)
4352 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4355 /* Return the number of page frames in holes in a zone on a node */
4356 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4357 unsigned long zone_type,
4358 unsigned long *ignored)
4360 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4361 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4362 unsigned long node_start_pfn, node_end_pfn;
4363 unsigned long zone_start_pfn, zone_end_pfn;
4365 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4366 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4367 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4369 adjust_zone_range_for_zone_movable(nid, zone_type,
4370 node_start_pfn, node_end_pfn,
4371 &zone_start_pfn, &zone_end_pfn);
4372 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4375 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4376 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4377 unsigned long zone_type,
4378 unsigned long *zones_size)
4380 return zones_size[zone_type];
4383 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4384 unsigned long zone_type,
4385 unsigned long *zholes_size)
4390 return zholes_size[zone_type];
4393 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4395 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4396 unsigned long *zones_size, unsigned long *zholes_size)
4398 unsigned long realtotalpages, totalpages = 0;
4401 for (i = 0; i < MAX_NR_ZONES; i++)
4402 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4404 pgdat->node_spanned_pages = totalpages;
4406 realtotalpages = totalpages;
4407 for (i = 0; i < MAX_NR_ZONES; i++)
4409 zone_absent_pages_in_node(pgdat->node_id, i,
4411 pgdat->node_present_pages = realtotalpages;
4412 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4416 #ifndef CONFIG_SPARSEMEM
4418 * Calculate the size of the zone->blockflags rounded to an unsigned long
4419 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4420 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4421 * round what is now in bits to nearest long in bits, then return it in
4424 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4426 unsigned long usemapsize;
4428 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4429 usemapsize = roundup(zonesize, pageblock_nr_pages);
4430 usemapsize = usemapsize >> pageblock_order;
4431 usemapsize *= NR_PAGEBLOCK_BITS;
4432 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4434 return usemapsize / 8;
4437 static void __init setup_usemap(struct pglist_data *pgdat,
4439 unsigned long zone_start_pfn,
4440 unsigned long zonesize)
4442 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4443 zone->pageblock_flags = NULL;
4445 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4449 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4450 unsigned long zone_start_pfn, unsigned long zonesize) {}
4451 #endif /* CONFIG_SPARSEMEM */
4453 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4455 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4456 void __init set_pageblock_order(void)
4460 /* Check that pageblock_nr_pages has not already been setup */
4461 if (pageblock_order)
4464 if (HPAGE_SHIFT > PAGE_SHIFT)
4465 order = HUGETLB_PAGE_ORDER;
4467 order = MAX_ORDER - 1;
4470 * Assume the largest contiguous order of interest is a huge page.
4471 * This value may be variable depending on boot parameters on IA64 and
4474 pageblock_order = order;
4476 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4479 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4480 * is unused as pageblock_order is set at compile-time. See
4481 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4484 void __init set_pageblock_order(void)
4488 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4490 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4491 unsigned long present_pages)
4493 unsigned long pages = spanned_pages;
4496 * Provide a more accurate estimation if there are holes within
4497 * the zone and SPARSEMEM is in use. If there are holes within the
4498 * zone, each populated memory region may cost us one or two extra
4499 * memmap pages due to alignment because memmap pages for each
4500 * populated regions may not naturally algined on page boundary.
4501 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4503 if (spanned_pages > present_pages + (present_pages >> 4) &&
4504 IS_ENABLED(CONFIG_SPARSEMEM))
4505 pages = present_pages;
4507 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4511 * Set up the zone data structures:
4512 * - mark all pages reserved
4513 * - mark all memory queues empty
4514 * - clear the memory bitmaps
4516 * NOTE: pgdat should get zeroed by caller.
4518 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4519 unsigned long *zones_size, unsigned long *zholes_size)
4522 int nid = pgdat->node_id;
4523 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4526 pgdat_resize_init(pgdat);
4527 #ifdef CONFIG_NUMA_BALANCING
4528 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4529 pgdat->numabalancing_migrate_nr_pages = 0;
4530 pgdat->numabalancing_migrate_next_window = jiffies;
4532 init_waitqueue_head(&pgdat->kswapd_wait);
4533 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4534 pgdat_page_cgroup_init(pgdat);
4536 for (j = 0; j < MAX_NR_ZONES; j++) {
4537 struct zone *zone = pgdat->node_zones + j;
4538 unsigned long size, realsize, freesize, memmap_pages;
4540 size = zone_spanned_pages_in_node(nid, j, zones_size);
4541 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4545 * Adjust freesize so that it accounts for how much memory
4546 * is used by this zone for memmap. This affects the watermark
4547 * and per-cpu initialisations
4549 memmap_pages = calc_memmap_size(size, realsize);
4550 if (freesize >= memmap_pages) {
4551 freesize -= memmap_pages;
4554 " %s zone: %lu pages used for memmap\n",
4555 zone_names[j], memmap_pages);
4558 " %s zone: %lu pages exceeds freesize %lu\n",
4559 zone_names[j], memmap_pages, freesize);
4561 /* Account for reserved pages */
4562 if (j == 0 && freesize > dma_reserve) {
4563 freesize -= dma_reserve;
4564 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4565 zone_names[0], dma_reserve);
4568 if (!is_highmem_idx(j))
4569 nr_kernel_pages += freesize;
4570 /* Charge for highmem memmap if there are enough kernel pages */
4571 else if (nr_kernel_pages > memmap_pages * 2)
4572 nr_kernel_pages -= memmap_pages;
4573 nr_all_pages += freesize;
4575 zone->spanned_pages = size;
4576 zone->present_pages = freesize;
4578 * Set an approximate value for lowmem here, it will be adjusted
4579 * when the bootmem allocator frees pages into the buddy system.
4580 * And all highmem pages will be managed by the buddy system.
4582 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4585 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4587 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4589 zone->name = zone_names[j];
4590 spin_lock_init(&zone->lock);
4591 spin_lock_init(&zone->lru_lock);
4592 zone_seqlock_init(zone);
4593 zone->zone_pgdat = pgdat;
4595 zone_pcp_init(zone);
4596 lruvec_init(&zone->lruvec);
4600 set_pageblock_order();
4601 setup_usemap(pgdat, zone, zone_start_pfn, size);
4602 ret = init_currently_empty_zone(zone, zone_start_pfn,
4603 size, MEMMAP_EARLY);
4605 memmap_init(size, nid, j, zone_start_pfn);
4606 zone_start_pfn += size;
4610 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4612 /* Skip empty nodes */
4613 if (!pgdat->node_spanned_pages)
4616 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4617 /* ia64 gets its own node_mem_map, before this, without bootmem */
4618 if (!pgdat->node_mem_map) {
4619 unsigned long size, start, end;
4623 * The zone's endpoints aren't required to be MAX_ORDER
4624 * aligned but the node_mem_map endpoints must be in order
4625 * for the buddy allocator to function correctly.
4627 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4628 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4629 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4630 size = (end - start) * sizeof(struct page);
4631 map = alloc_remap(pgdat->node_id, size);
4633 map = alloc_bootmem_node_nopanic(pgdat, size);
4634 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4636 #ifndef CONFIG_NEED_MULTIPLE_NODES
4638 * With no DISCONTIG, the global mem_map is just set as node 0's
4640 if (pgdat == NODE_DATA(0)) {
4641 mem_map = NODE_DATA(0)->node_mem_map;
4642 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4643 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4644 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4645 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4648 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4651 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4652 unsigned long node_start_pfn, unsigned long *zholes_size)
4654 pg_data_t *pgdat = NODE_DATA(nid);
4656 /* pg_data_t should be reset to zero when it's allocated */
4657 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4659 pgdat->node_id = nid;
4660 pgdat->node_start_pfn = node_start_pfn;
4661 init_zone_allows_reclaim(nid);
4662 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4664 alloc_node_mem_map(pgdat);
4665 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4666 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4667 nid, (unsigned long)pgdat,
4668 (unsigned long)pgdat->node_mem_map);
4671 free_area_init_core(pgdat, zones_size, zholes_size);
4674 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4676 #if MAX_NUMNODES > 1
4678 * Figure out the number of possible node ids.
4680 static void __init setup_nr_node_ids(void)
4683 unsigned int highest = 0;
4685 for_each_node_mask(node, node_possible_map)
4687 nr_node_ids = highest + 1;
4690 static inline void setup_nr_node_ids(void)
4696 * node_map_pfn_alignment - determine the maximum internode alignment
4698 * This function should be called after node map is populated and sorted.
4699 * It calculates the maximum power of two alignment which can distinguish
4702 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4703 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4704 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4705 * shifted, 1GiB is enough and this function will indicate so.
4707 * This is used to test whether pfn -> nid mapping of the chosen memory
4708 * model has fine enough granularity to avoid incorrect mapping for the
4709 * populated node map.
4711 * Returns the determined alignment in pfn's. 0 if there is no alignment
4712 * requirement (single node).
4714 unsigned long __init node_map_pfn_alignment(void)
4716 unsigned long accl_mask = 0, last_end = 0;
4717 unsigned long start, end, mask;
4721 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4722 if (!start || last_nid < 0 || last_nid == nid) {
4729 * Start with a mask granular enough to pin-point to the
4730 * start pfn and tick off bits one-by-one until it becomes
4731 * too coarse to separate the current node from the last.
4733 mask = ~((1 << __ffs(start)) - 1);
4734 while (mask && last_end <= (start & (mask << 1)))
4737 /* accumulate all internode masks */
4741 /* convert mask to number of pages */
4742 return ~accl_mask + 1;
4745 /* Find the lowest pfn for a node */
4746 static unsigned long __init find_min_pfn_for_node(int nid)
4748 unsigned long min_pfn = ULONG_MAX;
4749 unsigned long start_pfn;
4752 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4753 min_pfn = min(min_pfn, start_pfn);
4755 if (min_pfn == ULONG_MAX) {
4757 "Could not find start_pfn for node %d\n", nid);
4765 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4767 * It returns the minimum PFN based on information provided via
4768 * add_active_range().
4770 unsigned long __init find_min_pfn_with_active_regions(void)
4772 return find_min_pfn_for_node(MAX_NUMNODES);
4776 * early_calculate_totalpages()
4777 * Sum pages in active regions for movable zone.
4778 * Populate N_MEMORY for calculating usable_nodes.
4780 static unsigned long __init early_calculate_totalpages(void)
4782 unsigned long totalpages = 0;
4783 unsigned long start_pfn, end_pfn;
4786 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4787 unsigned long pages = end_pfn - start_pfn;
4789 totalpages += pages;
4791 node_set_state(nid, N_MEMORY);
4797 * Find the PFN the Movable zone begins in each node. Kernel memory
4798 * is spread evenly between nodes as long as the nodes have enough
4799 * memory. When they don't, some nodes will have more kernelcore than
4802 static void __init find_zone_movable_pfns_for_nodes(void)
4805 unsigned long usable_startpfn;
4806 unsigned long kernelcore_node, kernelcore_remaining;
4807 /* save the state before borrow the nodemask */
4808 nodemask_t saved_node_state = node_states[N_MEMORY];
4809 unsigned long totalpages = early_calculate_totalpages();
4810 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4813 * If movablecore was specified, calculate what size of
4814 * kernelcore that corresponds so that memory usable for
4815 * any allocation type is evenly spread. If both kernelcore
4816 * and movablecore are specified, then the value of kernelcore
4817 * will be used for required_kernelcore if it's greater than
4818 * what movablecore would have allowed.
4820 if (required_movablecore) {
4821 unsigned long corepages;
4824 * Round-up so that ZONE_MOVABLE is at least as large as what
4825 * was requested by the user
4827 required_movablecore =
4828 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4829 corepages = totalpages - required_movablecore;
4831 required_kernelcore = max(required_kernelcore, corepages);
4834 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4835 if (!required_kernelcore)
4838 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4839 find_usable_zone_for_movable();
4840 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4843 /* Spread kernelcore memory as evenly as possible throughout nodes */
4844 kernelcore_node = required_kernelcore / usable_nodes;
4845 for_each_node_state(nid, N_MEMORY) {
4846 unsigned long start_pfn, end_pfn;
4849 * Recalculate kernelcore_node if the division per node
4850 * now exceeds what is necessary to satisfy the requested
4851 * amount of memory for the kernel
4853 if (required_kernelcore < kernelcore_node)
4854 kernelcore_node = required_kernelcore / usable_nodes;
4857 * As the map is walked, we track how much memory is usable
4858 * by the kernel using kernelcore_remaining. When it is
4859 * 0, the rest of the node is usable by ZONE_MOVABLE
4861 kernelcore_remaining = kernelcore_node;
4863 /* Go through each range of PFNs within this node */
4864 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4865 unsigned long size_pages;
4867 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4868 if (start_pfn >= end_pfn)
4871 /* Account for what is only usable for kernelcore */
4872 if (start_pfn < usable_startpfn) {
4873 unsigned long kernel_pages;
4874 kernel_pages = min(end_pfn, usable_startpfn)
4877 kernelcore_remaining -= min(kernel_pages,
4878 kernelcore_remaining);
4879 required_kernelcore -= min(kernel_pages,
4880 required_kernelcore);
4882 /* Continue if range is now fully accounted */
4883 if (end_pfn <= usable_startpfn) {
4886 * Push zone_movable_pfn to the end so
4887 * that if we have to rebalance
4888 * kernelcore across nodes, we will
4889 * not double account here
4891 zone_movable_pfn[nid] = end_pfn;
4894 start_pfn = usable_startpfn;
4898 * The usable PFN range for ZONE_MOVABLE is from
4899 * start_pfn->end_pfn. Calculate size_pages as the
4900 * number of pages used as kernelcore
4902 size_pages = end_pfn - start_pfn;
4903 if (size_pages > kernelcore_remaining)
4904 size_pages = kernelcore_remaining;
4905 zone_movable_pfn[nid] = start_pfn + size_pages;
4908 * Some kernelcore has been met, update counts and
4909 * break if the kernelcore for this node has been
4912 required_kernelcore -= min(required_kernelcore,
4914 kernelcore_remaining -= size_pages;
4915 if (!kernelcore_remaining)
4921 * If there is still required_kernelcore, we do another pass with one
4922 * less node in the count. This will push zone_movable_pfn[nid] further
4923 * along on the nodes that still have memory until kernelcore is
4927 if (usable_nodes && required_kernelcore > usable_nodes)
4930 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4931 for (nid = 0; nid < MAX_NUMNODES; nid++)
4932 zone_movable_pfn[nid] =
4933 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4936 /* restore the node_state */
4937 node_states[N_MEMORY] = saved_node_state;
4940 /* Any regular or high memory on that node ? */
4941 static void check_for_memory(pg_data_t *pgdat, int nid)
4943 enum zone_type zone_type;
4945 if (N_MEMORY == N_NORMAL_MEMORY)
4948 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
4949 struct zone *zone = &pgdat->node_zones[zone_type];
4950 if (zone->present_pages) {
4951 node_set_state(nid, N_HIGH_MEMORY);
4952 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
4953 zone_type <= ZONE_NORMAL)
4954 node_set_state(nid, N_NORMAL_MEMORY);
4961 * free_area_init_nodes - Initialise all pg_data_t and zone data
4962 * @max_zone_pfn: an array of max PFNs for each zone
4964 * This will call free_area_init_node() for each active node in the system.
4965 * Using the page ranges provided by add_active_range(), the size of each
4966 * zone in each node and their holes is calculated. If the maximum PFN
4967 * between two adjacent zones match, it is assumed that the zone is empty.
4968 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4969 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4970 * starts where the previous one ended. For example, ZONE_DMA32 starts
4971 * at arch_max_dma_pfn.
4973 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4975 unsigned long start_pfn, end_pfn;
4978 /* Record where the zone boundaries are */
4979 memset(arch_zone_lowest_possible_pfn, 0,
4980 sizeof(arch_zone_lowest_possible_pfn));
4981 memset(arch_zone_highest_possible_pfn, 0,
4982 sizeof(arch_zone_highest_possible_pfn));
4983 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4984 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4985 for (i = 1; i < MAX_NR_ZONES; i++) {
4986 if (i == ZONE_MOVABLE)
4988 arch_zone_lowest_possible_pfn[i] =
4989 arch_zone_highest_possible_pfn[i-1];
4990 arch_zone_highest_possible_pfn[i] =
4991 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4993 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4994 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4996 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4997 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4998 find_zone_movable_pfns_for_nodes();
5000 /* Print out the zone ranges */
5001 printk("Zone ranges:\n");
5002 for (i = 0; i < MAX_NR_ZONES; i++) {
5003 if (i == ZONE_MOVABLE)
5005 printk(KERN_CONT " %-8s ", zone_names[i]);
5006 if (arch_zone_lowest_possible_pfn[i] ==
5007 arch_zone_highest_possible_pfn[i])
5008 printk(KERN_CONT "empty\n");
5010 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5011 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5012 (arch_zone_highest_possible_pfn[i]
5013 << PAGE_SHIFT) - 1);
5016 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5017 printk("Movable zone start for each node\n");
5018 for (i = 0; i < MAX_NUMNODES; i++) {
5019 if (zone_movable_pfn[i])
5020 printk(" Node %d: %#010lx\n", i,
5021 zone_movable_pfn[i] << PAGE_SHIFT);
5024 /* Print out the early node map */
5025 printk("Early memory node ranges\n");
5026 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5027 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5028 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5030 /* Initialise every node */
5031 mminit_verify_pageflags_layout();
5032 setup_nr_node_ids();
5033 for_each_online_node(nid) {
5034 pg_data_t *pgdat = NODE_DATA(nid);
5035 free_area_init_node(nid, NULL,
5036 find_min_pfn_for_node(nid), NULL);
5038 /* Any memory on that node */
5039 if (pgdat->node_present_pages)
5040 node_set_state(nid, N_MEMORY);
5041 check_for_memory(pgdat, nid);
5045 static int __init cmdline_parse_core(char *p, unsigned long *core)
5047 unsigned long long coremem;
5051 coremem = memparse(p, &p);
5052 *core = coremem >> PAGE_SHIFT;
5054 /* Paranoid check that UL is enough for the coremem value */
5055 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5061 * kernelcore=size sets the amount of memory for use for allocations that
5062 * cannot be reclaimed or migrated.
5064 static int __init cmdline_parse_kernelcore(char *p)
5066 return cmdline_parse_core(p, &required_kernelcore);
5070 * movablecore=size sets the amount of memory for use for allocations that
5071 * can be reclaimed or migrated.
5073 static int __init cmdline_parse_movablecore(char *p)
5075 return cmdline_parse_core(p, &required_movablecore);
5078 early_param("kernelcore", cmdline_parse_kernelcore);
5079 early_param("movablecore", cmdline_parse_movablecore);
5081 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5084 * set_dma_reserve - set the specified number of pages reserved in the first zone
5085 * @new_dma_reserve: The number of pages to mark reserved
5087 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5088 * In the DMA zone, a significant percentage may be consumed by kernel image
5089 * and other unfreeable allocations which can skew the watermarks badly. This
5090 * function may optionally be used to account for unfreeable pages in the
5091 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5092 * smaller per-cpu batchsize.
5094 void __init set_dma_reserve(unsigned long new_dma_reserve)
5096 dma_reserve = new_dma_reserve;
5099 void __init free_area_init(unsigned long *zones_size)
5101 free_area_init_node(0, zones_size,
5102 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5105 static int page_alloc_cpu_notify(struct notifier_block *self,
5106 unsigned long action, void *hcpu)
5108 int cpu = (unsigned long)hcpu;
5110 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5111 lru_add_drain_cpu(cpu);
5115 * Spill the event counters of the dead processor
5116 * into the current processors event counters.
5117 * This artificially elevates the count of the current
5120 vm_events_fold_cpu(cpu);
5123 * Zero the differential counters of the dead processor
5124 * so that the vm statistics are consistent.
5126 * This is only okay since the processor is dead and cannot
5127 * race with what we are doing.
5129 refresh_cpu_vm_stats(cpu);
5134 void __init page_alloc_init(void)
5136 hotcpu_notifier(page_alloc_cpu_notify, 0);
5140 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5141 * or min_free_kbytes changes.
5143 static void calculate_totalreserve_pages(void)
5145 struct pglist_data *pgdat;
5146 unsigned long reserve_pages = 0;
5147 enum zone_type i, j;
5149 for_each_online_pgdat(pgdat) {
5150 for (i = 0; i < MAX_NR_ZONES; i++) {
5151 struct zone *zone = pgdat->node_zones + i;
5152 unsigned long max = 0;
5154 /* Find valid and maximum lowmem_reserve in the zone */
5155 for (j = i; j < MAX_NR_ZONES; j++) {
5156 if (zone->lowmem_reserve[j] > max)
5157 max = zone->lowmem_reserve[j];
5160 /* we treat the high watermark as reserved pages. */
5161 max += high_wmark_pages(zone);
5163 if (max > zone->present_pages)
5164 max = zone->present_pages;
5165 reserve_pages += max;
5167 * Lowmem reserves are not available to
5168 * GFP_HIGHUSER page cache allocations and
5169 * kswapd tries to balance zones to their high
5170 * watermark. As a result, neither should be
5171 * regarded as dirtyable memory, to prevent a
5172 * situation where reclaim has to clean pages
5173 * in order to balance the zones.
5175 zone->dirty_balance_reserve = max;
5178 dirty_balance_reserve = reserve_pages;
5179 totalreserve_pages = reserve_pages;
5183 * setup_per_zone_lowmem_reserve - called whenever
5184 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5185 * has a correct pages reserved value, so an adequate number of
5186 * pages are left in the zone after a successful __alloc_pages().
5188 static void setup_per_zone_lowmem_reserve(void)
5190 struct pglist_data *pgdat;
5191 enum zone_type j, idx;
5193 for_each_online_pgdat(pgdat) {
5194 for (j = 0; j < MAX_NR_ZONES; j++) {
5195 struct zone *zone = pgdat->node_zones + j;
5196 unsigned long present_pages = zone->present_pages;
5198 zone->lowmem_reserve[j] = 0;
5202 struct zone *lower_zone;
5206 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5207 sysctl_lowmem_reserve_ratio[idx] = 1;
5209 lower_zone = pgdat->node_zones + idx;
5210 lower_zone->lowmem_reserve[j] = present_pages /
5211 sysctl_lowmem_reserve_ratio[idx];
5212 present_pages += lower_zone->present_pages;
5217 /* update totalreserve_pages */
5218 calculate_totalreserve_pages();
5221 static void __setup_per_zone_wmarks(void)
5223 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5224 unsigned long lowmem_pages = 0;
5226 unsigned long flags;
5228 /* Calculate total number of !ZONE_HIGHMEM pages */
5229 for_each_zone(zone) {
5230 if (!is_highmem(zone))
5231 lowmem_pages += zone->present_pages;
5234 for_each_zone(zone) {
5237 spin_lock_irqsave(&zone->lock, flags);
5238 tmp = (u64)pages_min * zone->present_pages;
5239 do_div(tmp, lowmem_pages);
5240 if (is_highmem(zone)) {
5242 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5243 * need highmem pages, so cap pages_min to a small
5246 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5247 * deltas controls asynch page reclaim, and so should
5248 * not be capped for highmem.
5250 unsigned long min_pages;
5252 min_pages = zone->present_pages / 1024;
5253 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5254 zone->watermark[WMARK_MIN] = min_pages;
5257 * If it's a lowmem zone, reserve a number of pages
5258 * proportionate to the zone's size.
5260 zone->watermark[WMARK_MIN] = tmp;
5263 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5264 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5266 setup_zone_migrate_reserve(zone);
5267 spin_unlock_irqrestore(&zone->lock, flags);
5270 /* update totalreserve_pages */
5271 calculate_totalreserve_pages();
5275 * setup_per_zone_wmarks - called when min_free_kbytes changes
5276 * or when memory is hot-{added|removed}
5278 * Ensures that the watermark[min,low,high] values for each zone are set
5279 * correctly with respect to min_free_kbytes.
5281 void setup_per_zone_wmarks(void)
5283 mutex_lock(&zonelists_mutex);
5284 __setup_per_zone_wmarks();
5285 mutex_unlock(&zonelists_mutex);
5289 * The inactive anon list should be small enough that the VM never has to
5290 * do too much work, but large enough that each inactive page has a chance
5291 * to be referenced again before it is swapped out.
5293 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5294 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5295 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5296 * the anonymous pages are kept on the inactive list.
5299 * memory ratio inactive anon
5300 * -------------------------------------
5309 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5311 unsigned int gb, ratio;
5313 /* Zone size in gigabytes */
5314 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5316 ratio = int_sqrt(10 * gb);
5320 zone->inactive_ratio = ratio;
5323 static void __meminit setup_per_zone_inactive_ratio(void)
5328 calculate_zone_inactive_ratio(zone);
5332 * Initialise min_free_kbytes.
5334 * For small machines we want it small (128k min). For large machines
5335 * we want it large (64MB max). But it is not linear, because network
5336 * bandwidth does not increase linearly with machine size. We use
5338 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5339 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5355 int __meminit init_per_zone_wmark_min(void)
5357 unsigned long lowmem_kbytes;
5359 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5361 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5362 if (min_free_kbytes < 128)
5363 min_free_kbytes = 128;
5364 if (min_free_kbytes > 65536)
5365 min_free_kbytes = 65536;
5366 setup_per_zone_wmarks();
5367 refresh_zone_stat_thresholds();
5368 setup_per_zone_lowmem_reserve();
5369 setup_per_zone_inactive_ratio();
5372 module_init(init_per_zone_wmark_min)
5375 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5376 * that we can call two helper functions whenever min_free_kbytes
5379 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5380 void __user *buffer, size_t *length, loff_t *ppos)
5382 proc_dointvec(table, write, buffer, length, ppos);
5384 setup_per_zone_wmarks();
5389 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5390 void __user *buffer, size_t *length, loff_t *ppos)
5395 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5400 zone->min_unmapped_pages = (zone->present_pages *
5401 sysctl_min_unmapped_ratio) / 100;
5405 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5406 void __user *buffer, size_t *length, loff_t *ppos)
5411 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5416 zone->min_slab_pages = (zone->present_pages *
5417 sysctl_min_slab_ratio) / 100;
5423 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5424 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5425 * whenever sysctl_lowmem_reserve_ratio changes.
5427 * The reserve ratio obviously has absolutely no relation with the
5428 * minimum watermarks. The lowmem reserve ratio can only make sense
5429 * if in function of the boot time zone sizes.
5431 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5432 void __user *buffer, size_t *length, loff_t *ppos)
5434 proc_dointvec_minmax(table, write, buffer, length, ppos);
5435 setup_per_zone_lowmem_reserve();
5440 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5441 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5442 * can have before it gets flushed back to buddy allocator.
5445 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5446 void __user *buffer, size_t *length, loff_t *ppos)
5452 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5453 if (!write || (ret < 0))
5455 for_each_populated_zone(zone) {
5456 for_each_possible_cpu(cpu) {
5458 high = zone->present_pages / percpu_pagelist_fraction;
5459 setup_pagelist_highmark(
5460 per_cpu_ptr(zone->pageset, cpu), high);
5466 int hashdist = HASHDIST_DEFAULT;
5469 static int __init set_hashdist(char *str)
5473 hashdist = simple_strtoul(str, &str, 0);
5476 __setup("hashdist=", set_hashdist);
5480 * allocate a large system hash table from bootmem
5481 * - it is assumed that the hash table must contain an exact power-of-2
5482 * quantity of entries
5483 * - limit is the number of hash buckets, not the total allocation size
5485 void *__init alloc_large_system_hash(const char *tablename,
5486 unsigned long bucketsize,
5487 unsigned long numentries,
5490 unsigned int *_hash_shift,
5491 unsigned int *_hash_mask,
5492 unsigned long low_limit,
5493 unsigned long high_limit)
5495 unsigned long long max = high_limit;
5496 unsigned long log2qty, size;
5499 /* allow the kernel cmdline to have a say */
5501 /* round applicable memory size up to nearest megabyte */
5502 numentries = nr_kernel_pages;
5503 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5504 numentries >>= 20 - PAGE_SHIFT;
5505 numentries <<= 20 - PAGE_SHIFT;
5507 /* limit to 1 bucket per 2^scale bytes of low memory */
5508 if (scale > PAGE_SHIFT)
5509 numentries >>= (scale - PAGE_SHIFT);
5511 numentries <<= (PAGE_SHIFT - scale);
5513 /* Make sure we've got at least a 0-order allocation.. */
5514 if (unlikely(flags & HASH_SMALL)) {
5515 /* Makes no sense without HASH_EARLY */
5516 WARN_ON(!(flags & HASH_EARLY));
5517 if (!(numentries >> *_hash_shift)) {
5518 numentries = 1UL << *_hash_shift;
5519 BUG_ON(!numentries);
5521 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5522 numentries = PAGE_SIZE / bucketsize;
5524 numentries = roundup_pow_of_two(numentries);
5526 /* limit allocation size to 1/16 total memory by default */
5528 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5529 do_div(max, bucketsize);
5531 max = min(max, 0x80000000ULL);
5533 if (numentries < low_limit)
5534 numentries = low_limit;
5535 if (numentries > max)
5538 log2qty = ilog2(numentries);
5541 size = bucketsize << log2qty;
5542 if (flags & HASH_EARLY)
5543 table = alloc_bootmem_nopanic(size);
5545 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5548 * If bucketsize is not a power-of-two, we may free
5549 * some pages at the end of hash table which
5550 * alloc_pages_exact() automatically does
5552 if (get_order(size) < MAX_ORDER) {
5553 table = alloc_pages_exact(size, GFP_ATOMIC);
5554 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5557 } while (!table && size > PAGE_SIZE && --log2qty);
5560 panic("Failed to allocate %s hash table\n", tablename);
5562 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5565 ilog2(size) - PAGE_SHIFT,
5569 *_hash_shift = log2qty;
5571 *_hash_mask = (1 << log2qty) - 1;
5576 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5577 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5580 #ifdef CONFIG_SPARSEMEM
5581 return __pfn_to_section(pfn)->pageblock_flags;
5583 return zone->pageblock_flags;
5584 #endif /* CONFIG_SPARSEMEM */
5587 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5589 #ifdef CONFIG_SPARSEMEM
5590 pfn &= (PAGES_PER_SECTION-1);
5591 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5593 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5594 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5595 #endif /* CONFIG_SPARSEMEM */
5599 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5600 * @page: The page within the block of interest
5601 * @start_bitidx: The first bit of interest to retrieve
5602 * @end_bitidx: The last bit of interest
5603 * returns pageblock_bits flags
5605 unsigned long get_pageblock_flags_group(struct page *page,
5606 int start_bitidx, int end_bitidx)
5609 unsigned long *bitmap;
5610 unsigned long pfn, bitidx;
5611 unsigned long flags = 0;
5612 unsigned long value = 1;
5614 zone = page_zone(page);
5615 pfn = page_to_pfn(page);
5616 bitmap = get_pageblock_bitmap(zone, pfn);
5617 bitidx = pfn_to_bitidx(zone, pfn);
5619 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5620 if (test_bit(bitidx + start_bitidx, bitmap))
5627 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5628 * @page: The page within the block of interest
5629 * @start_bitidx: The first bit of interest
5630 * @end_bitidx: The last bit of interest
5631 * @flags: The flags to set
5633 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5634 int start_bitidx, int end_bitidx)
5637 unsigned long *bitmap;
5638 unsigned long pfn, bitidx;
5639 unsigned long value = 1;
5641 zone = page_zone(page);
5642 pfn = page_to_pfn(page);
5643 bitmap = get_pageblock_bitmap(zone, pfn);
5644 bitidx = pfn_to_bitidx(zone, pfn);
5645 VM_BUG_ON(pfn < zone->zone_start_pfn);
5646 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5648 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5650 __set_bit(bitidx + start_bitidx, bitmap);
5652 __clear_bit(bitidx + start_bitidx, bitmap);
5656 * This function checks whether pageblock includes unmovable pages or not.
5657 * If @count is not zero, it is okay to include less @count unmovable pages
5659 * PageLRU check wihtout isolation or lru_lock could race so that
5660 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5661 * expect this function should be exact.
5663 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5664 bool skip_hwpoisoned_pages)
5666 unsigned long pfn, iter, found;
5670 * For avoiding noise data, lru_add_drain_all() should be called
5671 * If ZONE_MOVABLE, the zone never contains unmovable pages
5673 if (zone_idx(zone) == ZONE_MOVABLE)
5675 mt = get_pageblock_migratetype(page);
5676 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5679 pfn = page_to_pfn(page);
5680 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5681 unsigned long check = pfn + iter;
5683 if (!pfn_valid_within(check))
5686 page = pfn_to_page(check);
5688 * We can't use page_count without pin a page
5689 * because another CPU can free compound page.
5690 * This check already skips compound tails of THP
5691 * because their page->_count is zero at all time.
5693 if (!atomic_read(&page->_count)) {
5694 if (PageBuddy(page))
5695 iter += (1 << page_order(page)) - 1;
5700 * The HWPoisoned page may be not in buddy system, and
5701 * page_count() is not 0.
5703 if (skip_hwpoisoned_pages && PageHWPoison(page))
5709 * If there are RECLAIMABLE pages, we need to check it.
5710 * But now, memory offline itself doesn't call shrink_slab()
5711 * and it still to be fixed.
5714 * If the page is not RAM, page_count()should be 0.
5715 * we don't need more check. This is an _used_ not-movable page.
5717 * The problematic thing here is PG_reserved pages. PG_reserved
5718 * is set to both of a memory hole page and a _used_ kernel
5727 bool is_pageblock_removable_nolock(struct page *page)
5733 * We have to be careful here because we are iterating over memory
5734 * sections which are not zone aware so we might end up outside of
5735 * the zone but still within the section.
5736 * We have to take care about the node as well. If the node is offline
5737 * its NODE_DATA will be NULL - see page_zone.
5739 if (!node_online(page_to_nid(page)))
5742 zone = page_zone(page);
5743 pfn = page_to_pfn(page);
5744 if (zone->zone_start_pfn > pfn ||
5745 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5748 return !has_unmovable_pages(zone, page, 0, true);
5753 static unsigned long pfn_max_align_down(unsigned long pfn)
5755 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5756 pageblock_nr_pages) - 1);
5759 static unsigned long pfn_max_align_up(unsigned long pfn)
5761 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5762 pageblock_nr_pages));
5765 /* [start, end) must belong to a single zone. */
5766 static int __alloc_contig_migrate_range(struct compact_control *cc,
5767 unsigned long start, unsigned long end)
5769 /* This function is based on compact_zone() from compaction.c. */
5770 unsigned long nr_reclaimed;
5771 unsigned long pfn = start;
5772 unsigned int tries = 0;
5777 while (pfn < end || !list_empty(&cc->migratepages)) {
5778 if (fatal_signal_pending(current)) {
5783 if (list_empty(&cc->migratepages)) {
5784 cc->nr_migratepages = 0;
5785 pfn = isolate_migratepages_range(cc->zone, cc,
5792 } else if (++tries == 5) {
5793 ret = ret < 0 ? ret : -EBUSY;
5797 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5799 cc->nr_migratepages -= nr_reclaimed;
5801 ret = migrate_pages(&cc->migratepages,
5802 alloc_migrate_target,
5803 0, false, MIGRATE_SYNC,
5807 putback_movable_pages(&cc->migratepages);
5814 * alloc_contig_range() -- tries to allocate given range of pages
5815 * @start: start PFN to allocate
5816 * @end: one-past-the-last PFN to allocate
5817 * @migratetype: migratetype of the underlaying pageblocks (either
5818 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5819 * in range must have the same migratetype and it must
5820 * be either of the two.
5822 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5823 * aligned, however it's the caller's responsibility to guarantee that
5824 * we are the only thread that changes migrate type of pageblocks the
5827 * The PFN range must belong to a single zone.
5829 * Returns zero on success or negative error code. On success all
5830 * pages which PFN is in [start, end) are allocated for the caller and
5831 * need to be freed with free_contig_range().
5833 int alloc_contig_range(unsigned long start, unsigned long end,
5834 unsigned migratetype)
5836 unsigned long outer_start, outer_end;
5839 struct compact_control cc = {
5840 .nr_migratepages = 0,
5842 .zone = page_zone(pfn_to_page(start)),
5844 .ignore_skip_hint = true,
5846 INIT_LIST_HEAD(&cc.migratepages);
5849 * What we do here is we mark all pageblocks in range as
5850 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5851 * have different sizes, and due to the way page allocator
5852 * work, we align the range to biggest of the two pages so
5853 * that page allocator won't try to merge buddies from
5854 * different pageblocks and change MIGRATE_ISOLATE to some
5855 * other migration type.
5857 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5858 * migrate the pages from an unaligned range (ie. pages that
5859 * we are interested in). This will put all the pages in
5860 * range back to page allocator as MIGRATE_ISOLATE.
5862 * When this is done, we take the pages in range from page
5863 * allocator removing them from the buddy system. This way
5864 * page allocator will never consider using them.
5866 * This lets us mark the pageblocks back as
5867 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5868 * aligned range but not in the unaligned, original range are
5869 * put back to page allocator so that buddy can use them.
5872 ret = start_isolate_page_range(pfn_max_align_down(start),
5873 pfn_max_align_up(end), migratetype,
5878 ret = __alloc_contig_migrate_range(&cc, start, end);
5883 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5884 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5885 * more, all pages in [start, end) are free in page allocator.
5886 * What we are going to do is to allocate all pages from
5887 * [start, end) (that is remove them from page allocator).
5889 * The only problem is that pages at the beginning and at the
5890 * end of interesting range may be not aligned with pages that
5891 * page allocator holds, ie. they can be part of higher order
5892 * pages. Because of this, we reserve the bigger range and
5893 * once this is done free the pages we are not interested in.
5895 * We don't have to hold zone->lock here because the pages are
5896 * isolated thus they won't get removed from buddy.
5899 lru_add_drain_all();
5903 outer_start = start;
5904 while (!PageBuddy(pfn_to_page(outer_start))) {
5905 if (++order >= MAX_ORDER) {
5909 outer_start &= ~0UL << order;
5912 /* Make sure the range is really isolated. */
5913 if (test_pages_isolated(outer_start, end, false)) {
5914 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5921 /* Grab isolated pages from freelists. */
5922 outer_end = isolate_freepages_range(&cc, outer_start, end);
5928 /* Free head and tail (if any) */
5929 if (start != outer_start)
5930 free_contig_range(outer_start, start - outer_start);
5931 if (end != outer_end)
5932 free_contig_range(end, outer_end - end);
5935 undo_isolate_page_range(pfn_max_align_down(start),
5936 pfn_max_align_up(end), migratetype);
5940 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5942 unsigned int count = 0;
5944 for (; nr_pages--; pfn++) {
5945 struct page *page = pfn_to_page(pfn);
5947 count += page_count(page) != 1;
5950 WARN(count != 0, "%d pages are still in use!\n", count);
5954 #ifdef CONFIG_MEMORY_HOTPLUG
5955 static int __meminit __zone_pcp_update(void *data)
5957 struct zone *zone = data;
5959 unsigned long batch = zone_batchsize(zone), flags;
5961 for_each_possible_cpu(cpu) {
5962 struct per_cpu_pageset *pset;
5963 struct per_cpu_pages *pcp;
5965 pset = per_cpu_ptr(zone->pageset, cpu);
5968 local_irq_save(flags);
5970 free_pcppages_bulk(zone, pcp->count, pcp);
5971 drain_zonestat(zone, pset);
5972 setup_pageset(pset, batch);
5973 local_irq_restore(flags);
5978 void __meminit zone_pcp_update(struct zone *zone)
5980 stop_machine(__zone_pcp_update, zone, NULL);
5984 void zone_pcp_reset(struct zone *zone)
5986 unsigned long flags;
5988 struct per_cpu_pageset *pset;
5990 /* avoid races with drain_pages() */
5991 local_irq_save(flags);
5992 if (zone->pageset != &boot_pageset) {
5993 for_each_online_cpu(cpu) {
5994 pset = per_cpu_ptr(zone->pageset, cpu);
5995 drain_zonestat(zone, pset);
5997 free_percpu(zone->pageset);
5998 zone->pageset = &boot_pageset;
6000 local_irq_restore(flags);
6003 #ifdef CONFIG_MEMORY_HOTREMOVE
6005 * All pages in the range must be isolated before calling this.
6008 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6014 unsigned long flags;
6015 /* find the first valid pfn */
6016 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6021 zone = page_zone(pfn_to_page(pfn));
6022 spin_lock_irqsave(&zone->lock, flags);
6024 while (pfn < end_pfn) {
6025 if (!pfn_valid(pfn)) {
6029 page = pfn_to_page(pfn);
6031 * The HWPoisoned page may be not in buddy system, and
6032 * page_count() is not 0.
6034 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6036 SetPageReserved(page);
6040 BUG_ON(page_count(page));
6041 BUG_ON(!PageBuddy(page));
6042 order = page_order(page);
6043 #ifdef CONFIG_DEBUG_VM
6044 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6045 pfn, 1 << order, end_pfn);
6047 list_del(&page->lru);
6048 rmv_page_order(page);
6049 zone->free_area[order].nr_free--;
6050 for (i = 0; i < (1 << order); i++)
6051 SetPageReserved((page+i));
6052 pfn += (1 << order);
6054 spin_unlock_irqrestore(&zone->lock, flags);
6058 #ifdef CONFIG_MEMORY_FAILURE
6059 bool is_free_buddy_page(struct page *page)
6061 struct zone *zone = page_zone(page);
6062 unsigned long pfn = page_to_pfn(page);
6063 unsigned long flags;
6066 spin_lock_irqsave(&zone->lock, flags);
6067 for (order = 0; order < MAX_ORDER; order++) {
6068 struct page *page_head = page - (pfn & ((1 << order) - 1));
6070 if (PageBuddy(page_head) && page_order(page_head) >= order)
6073 spin_unlock_irqrestore(&zone->lock, flags);
6075 return order < MAX_ORDER;
6079 static const struct trace_print_flags pageflag_names[] = {
6080 {1UL << PG_locked, "locked" },
6081 {1UL << PG_error, "error" },
6082 {1UL << PG_referenced, "referenced" },
6083 {1UL << PG_uptodate, "uptodate" },
6084 {1UL << PG_dirty, "dirty" },
6085 {1UL << PG_lru, "lru" },
6086 {1UL << PG_active, "active" },
6087 {1UL << PG_slab, "slab" },
6088 {1UL << PG_owner_priv_1, "owner_priv_1" },
6089 {1UL << PG_arch_1, "arch_1" },
6090 {1UL << PG_reserved, "reserved" },
6091 {1UL << PG_private, "private" },
6092 {1UL << PG_private_2, "private_2" },
6093 {1UL << PG_writeback, "writeback" },
6094 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6095 {1UL << PG_head, "head" },
6096 {1UL << PG_tail, "tail" },
6098 {1UL << PG_compound, "compound" },
6100 {1UL << PG_swapcache, "swapcache" },
6101 {1UL << PG_mappedtodisk, "mappedtodisk" },
6102 {1UL << PG_reclaim, "reclaim" },
6103 {1UL << PG_swapbacked, "swapbacked" },
6104 {1UL << PG_unevictable, "unevictable" },
6106 {1UL << PG_mlocked, "mlocked" },
6108 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6109 {1UL << PG_uncached, "uncached" },
6111 #ifdef CONFIG_MEMORY_FAILURE
6112 {1UL << PG_hwpoison, "hwpoison" },
6114 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6115 {1UL << PG_compound_lock, "compound_lock" },
6119 static void dump_page_flags(unsigned long flags)
6121 const char *delim = "";
6125 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6127 printk(KERN_ALERT "page flags: %#lx(", flags);
6129 /* remove zone id */
6130 flags &= (1UL << NR_PAGEFLAGS) - 1;
6132 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6134 mask = pageflag_names[i].mask;
6135 if ((flags & mask) != mask)
6139 printk("%s%s", delim, pageflag_names[i].name);
6143 /* check for left over flags */
6145 printk("%s%#lx", delim, flags);
6150 void dump_page(struct page *page)
6153 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6154 page, atomic_read(&page->_count), page_mapcount(page),
6155 page->mapping, page->index);
6156 dump_page_flags(page->flags);
6157 mem_cgroup_print_bad_page(page);