2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
62 #include <asm/tlbflush.h>
63 #include <asm/div64.h>
66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67 DEFINE_PER_CPU(int, numa_node);
68 EXPORT_PER_CPU_SYMBOL(numa_node);
71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
78 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
83 * Array of node states.
85 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86 [N_POSSIBLE] = NODE_MASK_ALL,
87 [N_ONLINE] = { { [0] = 1UL } },
89 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
91 [N_HIGH_MEMORY] = { { [0] = 1UL } },
93 [N_CPU] = { { [0] = 1UL } },
96 EXPORT_SYMBOL(node_states);
98 unsigned long totalram_pages __read_mostly;
99 unsigned long totalreserve_pages __read_mostly;
101 * When calculating the number of globally allowed dirty pages, there
102 * is a certain number of per-zone reserves that should not be
103 * considered dirtyable memory. This is the sum of those reserves
104 * over all existing zones that contribute dirtyable memory.
106 unsigned long dirty_balance_reserve __read_mostly;
108 int percpu_pagelist_fraction;
109 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
111 #ifdef CONFIG_PM_SLEEP
113 * The following functions are used by the suspend/hibernate code to temporarily
114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
115 * while devices are suspended. To avoid races with the suspend/hibernate code,
116 * they should always be called with pm_mutex held (gfp_allowed_mask also should
117 * only be modified with pm_mutex held, unless the suspend/hibernate code is
118 * guaranteed not to run in parallel with that modification).
121 static gfp_t saved_gfp_mask;
123 void pm_restore_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 if (saved_gfp_mask) {
127 gfp_allowed_mask = saved_gfp_mask;
132 void pm_restrict_gfp_mask(void)
134 WARN_ON(!mutex_is_locked(&pm_mutex));
135 WARN_ON(saved_gfp_mask);
136 saved_gfp_mask = gfp_allowed_mask;
137 gfp_allowed_mask &= ~GFP_IOFS;
140 bool pm_suspended_storage(void)
142 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
146 #endif /* CONFIG_PM_SLEEP */
148 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
149 int pageblock_order __read_mostly;
152 static void __free_pages_ok(struct page *page, unsigned int order);
155 * results with 256, 32 in the lowmem_reserve sysctl:
156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
157 * 1G machine -> (16M dma, 784M normal, 224M high)
158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
162 * TBD: should special case ZONE_DMA32 machines here - in those we normally
163 * don't need any ZONE_NORMAL reservation
165 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
166 #ifdef CONFIG_ZONE_DMA
169 #ifdef CONFIG_ZONE_DMA32
172 #ifdef CONFIG_HIGHMEM
178 EXPORT_SYMBOL(totalram_pages);
180 static char * const zone_names[MAX_NR_ZONES] = {
181 #ifdef CONFIG_ZONE_DMA
184 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 int min_free_kbytes = 1024;
196 static unsigned long __meminitdata nr_kernel_pages;
197 static unsigned long __meminitdata nr_all_pages;
198 static unsigned long __meminitdata dma_reserve;
200 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
201 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
202 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
203 static unsigned long __initdata required_kernelcore;
204 static unsigned long __initdata required_movablecore;
205 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
207 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
209 EXPORT_SYMBOL(movable_zone);
210 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
213 int nr_node_ids __read_mostly = MAX_NUMNODES;
214 int nr_online_nodes __read_mostly = 1;
215 EXPORT_SYMBOL(nr_node_ids);
216 EXPORT_SYMBOL(nr_online_nodes);
219 int page_group_by_mobility_disabled __read_mostly;
223 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
224 * Instead, use {un}set_pageblock_isolate.
226 void set_pageblock_migratetype(struct page *page, int migratetype)
229 if (unlikely(page_group_by_mobility_disabled))
230 migratetype = MIGRATE_UNMOVABLE;
232 set_pageblock_flags_group(page, (unsigned long)migratetype,
233 PB_migrate, PB_migrate_end);
236 bool oom_killer_disabled __read_mostly;
238 #ifdef CONFIG_DEBUG_VM
239 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
243 unsigned long pfn = page_to_pfn(page);
246 seq = zone_span_seqbegin(zone);
247 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
249 else if (pfn < zone->zone_start_pfn)
251 } while (zone_span_seqretry(zone, seq));
256 static int page_is_consistent(struct zone *zone, struct page *page)
258 if (!pfn_valid_within(page_to_pfn(page)))
260 if (zone != page_zone(page))
266 * Temporary debugging check for pages not lying within a given zone.
268 static int bad_range(struct zone *zone, struct page *page)
270 if (page_outside_zone_boundaries(zone, page))
272 if (!page_is_consistent(zone, page))
278 static inline int bad_range(struct zone *zone, struct page *page)
284 static void bad_page(struct page *page)
286 static unsigned long resume;
287 static unsigned long nr_shown;
288 static unsigned long nr_unshown;
290 /* Don't complain about poisoned pages */
291 if (PageHWPoison(page)) {
292 reset_page_mapcount(page); /* remove PageBuddy */
297 * Allow a burst of 60 reports, then keep quiet for that minute;
298 * or allow a steady drip of one report per second.
300 if (nr_shown == 60) {
301 if (time_before(jiffies, resume)) {
307 "BUG: Bad page state: %lu messages suppressed\n",
314 resume = jiffies + 60 * HZ;
316 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
317 current->comm, page_to_pfn(page));
323 /* Leave bad fields for debug, except PageBuddy could make trouble */
324 reset_page_mapcount(page); /* remove PageBuddy */
325 add_taint(TAINT_BAD_PAGE);
329 * Higher-order pages are called "compound pages". They are structured thusly:
331 * The first PAGE_SIZE page is called the "head page".
333 * The remaining PAGE_SIZE pages are called "tail pages".
335 * All pages have PG_compound set. All tail pages have their ->first_page
336 * pointing at the head page.
338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
343 static void free_compound_page(struct page *page)
345 __free_pages_ok(page, compound_order(page));
348 void prep_compound_page(struct page *page, unsigned long order)
351 int nr_pages = 1 << order;
353 set_compound_page_dtor(page, free_compound_page);
354 set_compound_order(page, order);
356 for (i = 1; i < nr_pages; i++) {
357 struct page *p = page + i;
359 set_page_count(p, 0);
360 p->first_page = page;
364 /* update __split_huge_page_refcount if you change this function */
365 static int destroy_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
371 if (unlikely(compound_order(page) != order)) {
376 __ClearPageHead(page);
378 for (i = 1; i < nr_pages; i++) {
379 struct page *p = page + i;
381 if (unlikely(!PageTail(p) || (p->first_page != page))) {
391 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
396 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
397 * and __GFP_HIGHMEM from hard or soft interrupt context.
399 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
400 for (i = 0; i < (1 << order); i++)
401 clear_highpage(page + i);
404 #ifdef CONFIG_DEBUG_PAGEALLOC
405 unsigned int _debug_guardpage_minorder;
407 static int __init debug_guardpage_minorder_setup(char *buf)
411 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
412 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
415 _debug_guardpage_minorder = res;
416 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
419 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
421 static inline void set_page_guard_flag(struct page *page)
423 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
426 static inline void clear_page_guard_flag(struct page *page)
428 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
431 static inline void set_page_guard_flag(struct page *page) { }
432 static inline void clear_page_guard_flag(struct page *page) { }
435 static inline void set_page_order(struct page *page, int order)
437 set_page_private(page, order);
438 __SetPageBuddy(page);
441 static inline void rmv_page_order(struct page *page)
443 __ClearPageBuddy(page);
444 set_page_private(page, 0);
448 * Locate the struct page for both the matching buddy in our
449 * pair (buddy1) and the combined O(n+1) page they form (page).
451 * 1) Any buddy B1 will have an order O twin B2 which satisfies
452 * the following equation:
454 * For example, if the starting buddy (buddy2) is #8 its order
456 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
458 * 2) Any buddy B will have an order O+1 parent P which
459 * satisfies the following equation:
462 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
464 static inline unsigned long
465 __find_buddy_index(unsigned long page_idx, unsigned int order)
467 return page_idx ^ (1 << order);
471 * This function checks whether a page is free && is the buddy
472 * we can do coalesce a page and its buddy if
473 * (a) the buddy is not in a hole &&
474 * (b) the buddy is in the buddy system &&
475 * (c) a page and its buddy have the same order &&
476 * (d) a page and its buddy are in the same zone.
478 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
479 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
481 * For recording page's order, we use page_private(page).
483 static inline int page_is_buddy(struct page *page, struct page *buddy,
486 if (!pfn_valid_within(page_to_pfn(buddy)))
489 if (page_zone_id(page) != page_zone_id(buddy))
492 if (page_is_guard(buddy) && page_order(buddy) == order) {
493 VM_BUG_ON(page_count(buddy) != 0);
497 if (PageBuddy(buddy) && page_order(buddy) == order) {
498 VM_BUG_ON(page_count(buddy) != 0);
505 * Freeing function for a buddy system allocator.
507 * The concept of a buddy system is to maintain direct-mapped table
508 * (containing bit values) for memory blocks of various "orders".
509 * The bottom level table contains the map for the smallest allocatable
510 * units of memory (here, pages), and each level above it describes
511 * pairs of units from the levels below, hence, "buddies".
512 * At a high level, all that happens here is marking the table entry
513 * at the bottom level available, and propagating the changes upward
514 * as necessary, plus some accounting needed to play nicely with other
515 * parts of the VM system.
516 * At each level, we keep a list of pages, which are heads of continuous
517 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
518 * order is recorded in page_private(page) field.
519 * So when we are allocating or freeing one, we can derive the state of the
520 * other. That is, if we allocate a small block, and both were
521 * free, the remainder of the region must be split into blocks.
522 * If a block is freed, and its buddy is also free, then this
523 * triggers coalescing into a block of larger size.
528 static inline void __free_one_page(struct page *page,
529 struct zone *zone, unsigned int order,
532 unsigned long page_idx;
533 unsigned long combined_idx;
534 unsigned long uninitialized_var(buddy_idx);
537 if (unlikely(PageCompound(page)))
538 if (unlikely(destroy_compound_page(page, order)))
541 VM_BUG_ON(migratetype == -1);
543 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
545 VM_BUG_ON(page_idx & ((1 << order) - 1));
546 VM_BUG_ON(bad_range(zone, page));
548 while (order < MAX_ORDER-1) {
549 buddy_idx = __find_buddy_index(page_idx, order);
550 buddy = page + (buddy_idx - page_idx);
551 if (!page_is_buddy(page, buddy, order))
554 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
555 * merge with it and move up one order.
557 if (page_is_guard(buddy)) {
558 clear_page_guard_flag(buddy);
559 set_page_private(page, 0);
560 __mod_zone_freepage_state(zone, 1 << order,
563 list_del(&buddy->lru);
564 zone->free_area[order].nr_free--;
565 rmv_page_order(buddy);
567 combined_idx = buddy_idx & page_idx;
568 page = page + (combined_idx - page_idx);
569 page_idx = combined_idx;
572 set_page_order(page, order);
575 * If this is not the largest possible page, check if the buddy
576 * of the next-highest order is free. If it is, it's possible
577 * that pages are being freed that will coalesce soon. In case,
578 * that is happening, add the free page to the tail of the list
579 * so it's less likely to be used soon and more likely to be merged
580 * as a higher order page
582 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
583 struct page *higher_page, *higher_buddy;
584 combined_idx = buddy_idx & page_idx;
585 higher_page = page + (combined_idx - page_idx);
586 buddy_idx = __find_buddy_index(combined_idx, order + 1);
587 higher_buddy = higher_page + (buddy_idx - combined_idx);
588 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
589 list_add_tail(&page->lru,
590 &zone->free_area[order].free_list[migratetype]);
595 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
597 zone->free_area[order].nr_free++;
600 static inline int free_pages_check(struct page *page)
602 if (unlikely(page_mapcount(page) |
603 (page->mapping != NULL) |
604 (atomic_read(&page->_count) != 0) |
605 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
606 (mem_cgroup_bad_page_check(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);
734 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
736 unsigned int nr_pages = 1 << order;
740 for (loop = 0; loop < nr_pages; loop++) {
741 struct page *p = &page[loop];
743 if (loop + 1 < nr_pages)
745 __ClearPageReserved(p);
746 set_page_count(p, 0);
749 set_page_refcounted(page);
750 __free_pages(page, order);
754 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
755 void __init init_cma_reserved_pageblock(struct page *page)
757 unsigned i = pageblock_nr_pages;
758 struct page *p = page;
761 __ClearPageReserved(p);
762 set_page_count(p, 0);
765 set_page_refcounted(page);
766 set_pageblock_migratetype(page, MIGRATE_CMA);
767 __free_pages(page, pageblock_order);
768 totalram_pages += pageblock_nr_pages;
773 * The order of subdivision here is critical for the IO subsystem.
774 * Please do not alter this order without good reasons and regression
775 * testing. Specifically, as large blocks of memory are subdivided,
776 * the order in which smaller blocks are delivered depends on the order
777 * they're subdivided in this function. This is the primary factor
778 * influencing the order in which pages are delivered to the IO
779 * subsystem according to empirical testing, and this is also justified
780 * by considering the behavior of a buddy system containing a single
781 * large block of memory acted on by a series of small allocations.
782 * This behavior is a critical factor in sglist merging's success.
786 static inline void expand(struct zone *zone, struct page *page,
787 int low, int high, struct free_area *area,
790 unsigned long size = 1 << high;
796 VM_BUG_ON(bad_range(zone, &page[size]));
798 #ifdef CONFIG_DEBUG_PAGEALLOC
799 if (high < debug_guardpage_minorder()) {
801 * Mark as guard pages (or page), that will allow to
802 * merge back to allocator when buddy will be freed.
803 * Corresponding page table entries will not be touched,
804 * pages will stay not present in virtual address space
806 INIT_LIST_HEAD(&page[size].lru);
807 set_page_guard_flag(&page[size]);
808 set_page_private(&page[size], high);
809 /* Guard pages are not available for any usage */
810 __mod_zone_freepage_state(zone, -(1 << high),
815 list_add(&page[size].lru, &area->free_list[migratetype]);
817 set_page_order(&page[size], high);
822 * This page is about to be returned from the page allocator
824 static inline int check_new_page(struct page *page)
826 if (unlikely(page_mapcount(page) |
827 (page->mapping != NULL) |
828 (atomic_read(&page->_count) != 0) |
829 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
830 (mem_cgroup_bad_page_check(page)))) {
837 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
841 for (i = 0; i < (1 << order); i++) {
842 struct page *p = page + i;
843 if (unlikely(check_new_page(p)))
847 set_page_private(page, 0);
848 set_page_refcounted(page);
850 arch_alloc_page(page, order);
851 kernel_map_pages(page, 1 << order, 1);
853 if (gfp_flags & __GFP_ZERO)
854 prep_zero_page(page, order, gfp_flags);
856 if (order && (gfp_flags & __GFP_COMP))
857 prep_compound_page(page, order);
863 * Go through the free lists for the given migratetype and remove
864 * the smallest available page from the freelists
867 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
870 unsigned int current_order;
871 struct free_area * area;
874 /* Find a page of the appropriate size in the preferred list */
875 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
876 area = &(zone->free_area[current_order]);
877 if (list_empty(&area->free_list[migratetype]))
880 page = list_entry(area->free_list[migratetype].next,
882 list_del(&page->lru);
883 rmv_page_order(page);
885 expand(zone, page, order, current_order, area, migratetype);
894 * This array describes the order lists are fallen back to when
895 * the free lists for the desirable migrate type are depleted
897 static int fallbacks[MIGRATE_TYPES][4] = {
898 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
899 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
901 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
902 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
904 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
906 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
907 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
911 * Move the free pages in a range to the free lists of the requested type.
912 * Note that start_page and end_pages are not aligned on a pageblock
913 * boundary. If alignment is required, use move_freepages_block()
915 int move_freepages(struct zone *zone,
916 struct page *start_page, struct page *end_page,
923 #ifndef CONFIG_HOLES_IN_ZONE
925 * page_zone is not safe to call in this context when
926 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
927 * anyway as we check zone boundaries in move_freepages_block().
928 * Remove at a later date when no bug reports exist related to
929 * grouping pages by mobility
931 BUG_ON(page_zone(start_page) != page_zone(end_page));
934 for (page = start_page; page <= end_page;) {
935 /* Make sure we are not inadvertently changing nodes */
936 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
938 if (!pfn_valid_within(page_to_pfn(page))) {
943 if (!PageBuddy(page)) {
948 order = page_order(page);
949 list_move(&page->lru,
950 &zone->free_area[order].free_list[migratetype]);
951 set_freepage_migratetype(page, migratetype);
953 pages_moved += 1 << order;
959 int move_freepages_block(struct zone *zone, struct page *page,
962 unsigned long start_pfn, end_pfn;
963 struct page *start_page, *end_page;
965 start_pfn = page_to_pfn(page);
966 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
967 start_page = pfn_to_page(start_pfn);
968 end_page = start_page + pageblock_nr_pages - 1;
969 end_pfn = start_pfn + pageblock_nr_pages - 1;
971 /* Do not cross zone boundaries */
972 if (start_pfn < zone->zone_start_pfn)
974 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
977 return move_freepages(zone, start_page, end_page, migratetype);
980 static void change_pageblock_range(struct page *pageblock_page,
981 int start_order, int migratetype)
983 int nr_pageblocks = 1 << (start_order - pageblock_order);
985 while (nr_pageblocks--) {
986 set_pageblock_migratetype(pageblock_page, migratetype);
987 pageblock_page += pageblock_nr_pages;
991 /* Remove an element from the buddy allocator from the fallback list */
992 static inline struct page *
993 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
995 struct free_area * area;
1000 /* Find the largest possible block of pages in the other list */
1001 for (current_order = MAX_ORDER-1; current_order >= order;
1004 migratetype = fallbacks[start_migratetype][i];
1006 /* MIGRATE_RESERVE handled later if necessary */
1007 if (migratetype == MIGRATE_RESERVE)
1010 area = &(zone->free_area[current_order]);
1011 if (list_empty(&area->free_list[migratetype]))
1014 page = list_entry(area->free_list[migratetype].next,
1019 * If breaking a large block of pages, move all free
1020 * pages to the preferred allocation list. If falling
1021 * back for a reclaimable kernel allocation, be more
1022 * aggressive about taking ownership of free pages
1024 * On the other hand, never change migration
1025 * type of MIGRATE_CMA pageblocks nor move CMA
1026 * pages on different free lists. We don't
1027 * want unmovable pages to be allocated from
1028 * MIGRATE_CMA areas.
1030 if (!is_migrate_cma(migratetype) &&
1031 (unlikely(current_order >= pageblock_order / 2) ||
1032 start_migratetype == MIGRATE_RECLAIMABLE ||
1033 page_group_by_mobility_disabled)) {
1035 pages = move_freepages_block(zone, page,
1038 /* Claim the whole block if over half of it is free */
1039 if (pages >= (1 << (pageblock_order-1)) ||
1040 page_group_by_mobility_disabled)
1041 set_pageblock_migratetype(page,
1044 migratetype = start_migratetype;
1047 /* Remove the page from the freelists */
1048 list_del(&page->lru);
1049 rmv_page_order(page);
1051 /* Take ownership for orders >= pageblock_order */
1052 if (current_order >= pageblock_order &&
1053 !is_migrate_cma(migratetype))
1054 change_pageblock_range(page, current_order,
1057 expand(zone, page, order, current_order, area,
1058 is_migrate_cma(migratetype)
1059 ? migratetype : start_migratetype);
1061 trace_mm_page_alloc_extfrag(page, order, current_order,
1062 start_migratetype, migratetype);
1072 * Do the hard work of removing an element from the buddy allocator.
1073 * Call me with the zone->lock already held.
1075 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1081 page = __rmqueue_smallest(zone, order, migratetype);
1083 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1084 page = __rmqueue_fallback(zone, order, migratetype);
1087 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1088 * is used because __rmqueue_smallest is an inline function
1089 * and we want just one call site
1092 migratetype = MIGRATE_RESERVE;
1097 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1102 * Obtain a specified number of elements from the buddy allocator, all under
1103 * a single hold of the lock, for efficiency. Add them to the supplied list.
1104 * Returns the number of new pages which were placed at *list.
1106 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1107 unsigned long count, struct list_head *list,
1108 int migratetype, int cold)
1110 int mt = migratetype, i;
1112 spin_lock(&zone->lock);
1113 for (i = 0; i < count; ++i) {
1114 struct page *page = __rmqueue(zone, order, migratetype);
1115 if (unlikely(page == NULL))
1119 * Split buddy pages returned by expand() are received here
1120 * in physical page order. The page is added to the callers and
1121 * list and the list head then moves forward. From the callers
1122 * perspective, the linked list is ordered by page number in
1123 * some conditions. This is useful for IO devices that can
1124 * merge IO requests if the physical pages are ordered
1127 if (likely(cold == 0))
1128 list_add(&page->lru, list);
1130 list_add_tail(&page->lru, list);
1131 if (IS_ENABLED(CONFIG_CMA)) {
1132 mt = get_pageblock_migratetype(page);
1133 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1136 set_freepage_migratetype(page, mt);
1138 if (is_migrate_cma(mt))
1139 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1142 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1143 spin_unlock(&zone->lock);
1149 * Called from the vmstat counter updater to drain pagesets of this
1150 * currently executing processor on remote nodes after they have
1153 * Note that this function must be called with the thread pinned to
1154 * a single processor.
1156 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1158 unsigned long flags;
1161 local_irq_save(flags);
1162 if (pcp->count >= pcp->batch)
1163 to_drain = pcp->batch;
1165 to_drain = pcp->count;
1167 free_pcppages_bulk(zone, to_drain, pcp);
1168 pcp->count -= to_drain;
1170 local_irq_restore(flags);
1175 * Drain pages of the indicated processor.
1177 * The processor must either be the current processor and the
1178 * thread pinned to the current processor or a processor that
1181 static void drain_pages(unsigned int cpu)
1183 unsigned long flags;
1186 for_each_populated_zone(zone) {
1187 struct per_cpu_pageset *pset;
1188 struct per_cpu_pages *pcp;
1190 local_irq_save(flags);
1191 pset = per_cpu_ptr(zone->pageset, cpu);
1195 free_pcppages_bulk(zone, pcp->count, pcp);
1198 local_irq_restore(flags);
1203 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1205 void drain_local_pages(void *arg)
1207 drain_pages(smp_processor_id());
1211 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1213 * Note that this code is protected against sending an IPI to an offline
1214 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1215 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1216 * nothing keeps CPUs from showing up after we populated the cpumask and
1217 * before the call to on_each_cpu_mask().
1219 void drain_all_pages(void)
1222 struct per_cpu_pageset *pcp;
1226 * Allocate in the BSS so we wont require allocation in
1227 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1229 static cpumask_t cpus_with_pcps;
1232 * We don't care about racing with CPU hotplug event
1233 * as offline notification will cause the notified
1234 * cpu to drain that CPU pcps and on_each_cpu_mask
1235 * disables preemption as part of its processing
1237 for_each_online_cpu(cpu) {
1238 bool has_pcps = false;
1239 for_each_populated_zone(zone) {
1240 pcp = per_cpu_ptr(zone->pageset, cpu);
1241 if (pcp->pcp.count) {
1247 cpumask_set_cpu(cpu, &cpus_with_pcps);
1249 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1251 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1254 #ifdef CONFIG_HIBERNATION
1256 void mark_free_pages(struct zone *zone)
1258 unsigned long pfn, max_zone_pfn;
1259 unsigned long flags;
1261 struct list_head *curr;
1263 if (!zone->spanned_pages)
1266 spin_lock_irqsave(&zone->lock, flags);
1268 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1269 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1270 if (pfn_valid(pfn)) {
1271 struct page *page = pfn_to_page(pfn);
1273 if (!swsusp_page_is_forbidden(page))
1274 swsusp_unset_page_free(page);
1277 for_each_migratetype_order(order, t) {
1278 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1281 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1282 for (i = 0; i < (1UL << order); i++)
1283 swsusp_set_page_free(pfn_to_page(pfn + i));
1286 spin_unlock_irqrestore(&zone->lock, flags);
1288 #endif /* CONFIG_PM */
1291 * Free a 0-order page
1292 * cold == 1 ? free a cold page : free a hot page
1294 void free_hot_cold_page(struct page *page, int cold)
1296 struct zone *zone = page_zone(page);
1297 struct per_cpu_pages *pcp;
1298 unsigned long flags;
1301 if (!free_pages_prepare(page, 0))
1304 migratetype = get_pageblock_migratetype(page);
1305 set_freepage_migratetype(page, migratetype);
1306 local_irq_save(flags);
1307 __count_vm_event(PGFREE);
1310 * We only track unmovable, reclaimable and movable on pcp lists.
1311 * Free ISOLATE pages back to the allocator because they are being
1312 * offlined but treat RESERVE as movable pages so we can get those
1313 * areas back if necessary. Otherwise, we may have to free
1314 * excessively into the page allocator
1316 if (migratetype >= MIGRATE_PCPTYPES) {
1317 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1318 free_one_page(zone, page, 0, migratetype);
1321 migratetype = MIGRATE_MOVABLE;
1324 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1326 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1328 list_add(&page->lru, &pcp->lists[migratetype]);
1330 if (pcp->count >= pcp->high) {
1331 free_pcppages_bulk(zone, pcp->batch, pcp);
1332 pcp->count -= pcp->batch;
1336 local_irq_restore(flags);
1340 * Free a list of 0-order pages
1342 void free_hot_cold_page_list(struct list_head *list, int cold)
1344 struct page *page, *next;
1346 list_for_each_entry_safe(page, next, list, lru) {
1347 trace_mm_page_free_batched(page, cold);
1348 free_hot_cold_page(page, cold);
1353 * split_page takes a non-compound higher-order page, and splits it into
1354 * n (1<<order) sub-pages: page[0..n]
1355 * Each sub-page must be freed individually.
1357 * Note: this is probably too low level an operation for use in drivers.
1358 * Please consult with lkml before using this in your driver.
1360 void split_page(struct page *page, unsigned int order)
1364 VM_BUG_ON(PageCompound(page));
1365 VM_BUG_ON(!page_count(page));
1367 #ifdef CONFIG_KMEMCHECK
1369 * Split shadow pages too, because free(page[0]) would
1370 * otherwise free the whole shadow.
1372 if (kmemcheck_page_is_tracked(page))
1373 split_page(virt_to_page(page[0].shadow), order);
1376 for (i = 1; i < (1 << order); i++)
1377 set_page_refcounted(page + i);
1381 * Similar to the split_page family of functions except that the page
1382 * required at the given order and being isolated now to prevent races
1383 * with parallel allocators
1385 int capture_free_page(struct page *page, int alloc_order, int migratetype)
1388 unsigned long watermark;
1392 BUG_ON(!PageBuddy(page));
1394 zone = page_zone(page);
1395 order = page_order(page);
1397 /* Obey watermarks as if the page was being allocated */
1398 watermark = low_wmark_pages(zone) + (1 << order);
1399 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1402 /* Remove page from free list */
1403 list_del(&page->lru);
1404 zone->free_area[order].nr_free--;
1405 rmv_page_order(page);
1407 mt = get_pageblock_migratetype(page);
1408 if (unlikely(mt != MIGRATE_ISOLATE))
1409 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1411 if (alloc_order != order)
1412 expand(zone, page, alloc_order, order,
1413 &zone->free_area[order], migratetype);
1415 /* Set the pageblock if the captured page is at least a pageblock */
1416 if (order >= pageblock_order - 1) {
1417 struct page *endpage = page + (1 << order) - 1;
1418 for (; page < endpage; page += pageblock_nr_pages) {
1419 int mt = get_pageblock_migratetype(page);
1420 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1421 set_pageblock_migratetype(page,
1426 return 1UL << order;
1430 * Similar to split_page except the page is already free. As this is only
1431 * being used for migration, the migratetype of the block also changes.
1432 * As this is called with interrupts disabled, the caller is responsible
1433 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1436 * Note: this is probably too low level an operation for use in drivers.
1437 * Please consult with lkml before using this in your driver.
1439 int split_free_page(struct page *page)
1444 BUG_ON(!PageBuddy(page));
1445 order = page_order(page);
1447 nr_pages = capture_free_page(page, order, 0);
1451 /* Split into individual pages */
1452 set_page_refcounted(page);
1453 split_page(page, order);
1458 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1459 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1463 struct page *buffered_rmqueue(struct zone *preferred_zone,
1464 struct zone *zone, int order, gfp_t gfp_flags,
1467 unsigned long flags;
1469 int cold = !!(gfp_flags & __GFP_COLD);
1472 if (likely(order == 0)) {
1473 struct per_cpu_pages *pcp;
1474 struct list_head *list;
1476 local_irq_save(flags);
1477 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1478 list = &pcp->lists[migratetype];
1479 if (list_empty(list)) {
1480 pcp->count += rmqueue_bulk(zone, 0,
1483 if (unlikely(list_empty(list)))
1488 page = list_entry(list->prev, struct page, lru);
1490 page = list_entry(list->next, struct page, lru);
1492 list_del(&page->lru);
1495 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1497 * __GFP_NOFAIL is not to be used in new code.
1499 * All __GFP_NOFAIL callers should be fixed so that they
1500 * properly detect and handle allocation failures.
1502 * We most definitely don't want callers attempting to
1503 * allocate greater than order-1 page units with
1506 WARN_ON_ONCE(order > 1);
1508 spin_lock_irqsave(&zone->lock, flags);
1509 page = __rmqueue(zone, order, migratetype);
1510 spin_unlock(&zone->lock);
1513 __mod_zone_freepage_state(zone, -(1 << order),
1514 get_pageblock_migratetype(page));
1517 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1518 zone_statistics(preferred_zone, zone, gfp_flags);
1519 local_irq_restore(flags);
1521 VM_BUG_ON(bad_range(zone, page));
1522 if (prep_new_page(page, order, gfp_flags))
1527 local_irq_restore(flags);
1531 #ifdef CONFIG_FAIL_PAGE_ALLOC
1534 struct fault_attr attr;
1536 u32 ignore_gfp_highmem;
1537 u32 ignore_gfp_wait;
1539 } fail_page_alloc = {
1540 .attr = FAULT_ATTR_INITIALIZER,
1541 .ignore_gfp_wait = 1,
1542 .ignore_gfp_highmem = 1,
1546 static int __init setup_fail_page_alloc(char *str)
1548 return setup_fault_attr(&fail_page_alloc.attr, str);
1550 __setup("fail_page_alloc=", setup_fail_page_alloc);
1552 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1554 if (order < fail_page_alloc.min_order)
1556 if (gfp_mask & __GFP_NOFAIL)
1558 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1560 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1563 return should_fail(&fail_page_alloc.attr, 1 << order);
1566 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1568 static int __init fail_page_alloc_debugfs(void)
1570 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1573 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1574 &fail_page_alloc.attr);
1576 return PTR_ERR(dir);
1578 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1579 &fail_page_alloc.ignore_gfp_wait))
1581 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1582 &fail_page_alloc.ignore_gfp_highmem))
1584 if (!debugfs_create_u32("min-order", mode, dir,
1585 &fail_page_alloc.min_order))
1590 debugfs_remove_recursive(dir);
1595 late_initcall(fail_page_alloc_debugfs);
1597 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1599 #else /* CONFIG_FAIL_PAGE_ALLOC */
1601 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1606 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1609 * Return true if free pages are above 'mark'. This takes into account the order
1610 * of the allocation.
1612 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1613 int classzone_idx, int alloc_flags, long free_pages)
1615 /* free_pages my go negative - that's OK */
1617 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1620 free_pages -= (1 << order) - 1;
1621 if (alloc_flags & ALLOC_HIGH)
1623 if (alloc_flags & ALLOC_HARDER)
1626 /* If allocation can't use CMA areas don't use free CMA pages */
1627 if (!(alloc_flags & ALLOC_CMA))
1628 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1630 if (free_pages <= min + lowmem_reserve)
1632 for (o = 0; o < order; o++) {
1633 /* At the next order, this order's pages become unavailable */
1634 free_pages -= z->free_area[o].nr_free << o;
1636 /* Require fewer higher order pages to be free */
1639 if (free_pages <= min)
1645 #ifdef CONFIG_MEMORY_ISOLATION
1646 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1648 if (unlikely(zone->nr_pageblock_isolate))
1649 return zone->nr_pageblock_isolate * pageblock_nr_pages;
1653 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1659 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1660 int classzone_idx, int alloc_flags)
1662 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1663 zone_page_state(z, NR_FREE_PAGES));
1666 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1667 int classzone_idx, int alloc_flags)
1669 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1671 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1672 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1675 * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1676 * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1677 * sleep although it could do so. But this is more desirable for memory
1678 * hotplug than sleeping which can cause a livelock in the direct
1681 free_pages -= nr_zone_isolate_freepages(z);
1682 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1688 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1689 * skip over zones that are not allowed by the cpuset, or that have
1690 * been recently (in last second) found to be nearly full. See further
1691 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1692 * that have to skip over a lot of full or unallowed zones.
1694 * If the zonelist cache is present in the passed in zonelist, then
1695 * returns a pointer to the allowed node mask (either the current
1696 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1698 * If the zonelist cache is not available for this zonelist, does
1699 * nothing and returns NULL.
1701 * If the fullzones BITMAP in the zonelist cache is stale (more than
1702 * a second since last zap'd) then we zap it out (clear its bits.)
1704 * We hold off even calling zlc_setup, until after we've checked the
1705 * first zone in the zonelist, on the theory that most allocations will
1706 * be satisfied from that first zone, so best to examine that zone as
1707 * quickly as we can.
1709 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1711 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1712 nodemask_t *allowednodes; /* zonelist_cache approximation */
1714 zlc = zonelist->zlcache_ptr;
1718 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1719 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1720 zlc->last_full_zap = jiffies;
1723 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1724 &cpuset_current_mems_allowed :
1725 &node_states[N_HIGH_MEMORY];
1726 return allowednodes;
1730 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1731 * if it is worth looking at further for free memory:
1732 * 1) Check that the zone isn't thought to be full (doesn't have its
1733 * bit set in the zonelist_cache fullzones BITMAP).
1734 * 2) Check that the zones node (obtained from the zonelist_cache
1735 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1736 * Return true (non-zero) if zone is worth looking at further, or
1737 * else return false (zero) if it is not.
1739 * This check -ignores- the distinction between various watermarks,
1740 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1741 * found to be full for any variation of these watermarks, it will
1742 * be considered full for up to one second by all requests, unless
1743 * we are so low on memory on all allowed nodes that we are forced
1744 * into the second scan of the zonelist.
1746 * In the second scan we ignore this zonelist cache and exactly
1747 * apply the watermarks to all zones, even it is slower to do so.
1748 * We are low on memory in the second scan, and should leave no stone
1749 * unturned looking for a free page.
1751 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1752 nodemask_t *allowednodes)
1754 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1755 int i; /* index of *z in zonelist zones */
1756 int n; /* node that zone *z is on */
1758 zlc = zonelist->zlcache_ptr;
1762 i = z - zonelist->_zonerefs;
1765 /* This zone is worth trying if it is allowed but not full */
1766 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1770 * Given 'z' scanning a zonelist, set the corresponding bit in
1771 * zlc->fullzones, so that subsequent attempts to allocate a page
1772 * from that zone don't waste time re-examining it.
1774 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1776 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1777 int i; /* index of *z in zonelist zones */
1779 zlc = zonelist->zlcache_ptr;
1783 i = z - zonelist->_zonerefs;
1785 set_bit(i, zlc->fullzones);
1789 * clear all zones full, called after direct reclaim makes progress so that
1790 * a zone that was recently full is not skipped over for up to a second
1792 static void zlc_clear_zones_full(struct zonelist *zonelist)
1794 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1796 zlc = zonelist->zlcache_ptr;
1800 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1803 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1805 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1808 static void __paginginit init_zone_allows_reclaim(int nid)
1812 for_each_online_node(i)
1813 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1814 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1816 zone_reclaim_mode = 1;
1819 #else /* CONFIG_NUMA */
1821 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1826 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1827 nodemask_t *allowednodes)
1832 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1836 static void zlc_clear_zones_full(struct zonelist *zonelist)
1840 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1845 static inline void init_zone_allows_reclaim(int nid)
1848 #endif /* CONFIG_NUMA */
1851 * get_page_from_freelist goes through the zonelist trying to allocate
1854 static struct page *
1855 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1856 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1857 struct zone *preferred_zone, int migratetype)
1860 struct page *page = NULL;
1863 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1864 int zlc_active = 0; /* set if using zonelist_cache */
1865 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1867 classzone_idx = zone_idx(preferred_zone);
1870 * Scan zonelist, looking for a zone with enough free.
1871 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1873 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1874 high_zoneidx, nodemask) {
1875 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1876 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1878 if ((alloc_flags & ALLOC_CPUSET) &&
1879 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1882 * When allocating a page cache page for writing, we
1883 * want to get it from a zone that is within its dirty
1884 * limit, such that no single zone holds more than its
1885 * proportional share of globally allowed dirty pages.
1886 * The dirty limits take into account the zone's
1887 * lowmem reserves and high watermark so that kswapd
1888 * should be able to balance it without having to
1889 * write pages from its LRU list.
1891 * This may look like it could increase pressure on
1892 * lower zones by failing allocations in higher zones
1893 * before they are full. But the pages that do spill
1894 * over are limited as the lower zones are protected
1895 * by this very same mechanism. It should not become
1896 * a practical burden to them.
1898 * XXX: For now, allow allocations to potentially
1899 * exceed the per-zone dirty limit in the slowpath
1900 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1901 * which is important when on a NUMA setup the allowed
1902 * zones are together not big enough to reach the
1903 * global limit. The proper fix for these situations
1904 * will require awareness of zones in the
1905 * dirty-throttling and the flusher threads.
1907 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1908 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1909 goto this_zone_full;
1911 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1912 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1916 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1917 if (zone_watermark_ok(zone, order, mark,
1918 classzone_idx, alloc_flags))
1921 if (IS_ENABLED(CONFIG_NUMA) &&
1922 !did_zlc_setup && nr_online_nodes > 1) {
1924 * we do zlc_setup if there are multiple nodes
1925 * and before considering the first zone allowed
1928 allowednodes = zlc_setup(zonelist, alloc_flags);
1933 if (zone_reclaim_mode == 0 ||
1934 !zone_allows_reclaim(preferred_zone, zone))
1935 goto this_zone_full;
1938 * As we may have just activated ZLC, check if the first
1939 * eligible zone has failed zone_reclaim recently.
1941 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1942 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1945 ret = zone_reclaim(zone, gfp_mask, order);
1947 case ZONE_RECLAIM_NOSCAN:
1950 case ZONE_RECLAIM_FULL:
1951 /* scanned but unreclaimable */
1954 /* did we reclaim enough */
1955 if (!zone_watermark_ok(zone, order, mark,
1956 classzone_idx, alloc_flags))
1957 goto this_zone_full;
1962 page = buffered_rmqueue(preferred_zone, zone, order,
1963 gfp_mask, migratetype);
1967 if (IS_ENABLED(CONFIG_NUMA))
1968 zlc_mark_zone_full(zonelist, z);
1971 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1972 /* Disable zlc cache for second zonelist scan */
1979 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1980 * necessary to allocate the page. The expectation is
1981 * that the caller is taking steps that will free more
1982 * memory. The caller should avoid the page being used
1983 * for !PFMEMALLOC purposes.
1985 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1991 * Large machines with many possible nodes should not always dump per-node
1992 * meminfo in irq context.
1994 static inline bool should_suppress_show_mem(void)
1999 ret = in_interrupt();
2004 static DEFINE_RATELIMIT_STATE(nopage_rs,
2005 DEFAULT_RATELIMIT_INTERVAL,
2006 DEFAULT_RATELIMIT_BURST);
2008 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2010 unsigned int filter = SHOW_MEM_FILTER_NODES;
2012 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2013 debug_guardpage_minorder() > 0)
2017 * This documents exceptions given to allocations in certain
2018 * contexts that are allowed to allocate outside current's set
2021 if (!(gfp_mask & __GFP_NOMEMALLOC))
2022 if (test_thread_flag(TIF_MEMDIE) ||
2023 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2024 filter &= ~SHOW_MEM_FILTER_NODES;
2025 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2026 filter &= ~SHOW_MEM_FILTER_NODES;
2029 struct va_format vaf;
2032 va_start(args, fmt);
2037 pr_warn("%pV", &vaf);
2042 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2043 current->comm, order, gfp_mask);
2046 if (!should_suppress_show_mem())
2051 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2052 unsigned long did_some_progress,
2053 unsigned long pages_reclaimed)
2055 /* Do not loop if specifically requested */
2056 if (gfp_mask & __GFP_NORETRY)
2059 /* Always retry if specifically requested */
2060 if (gfp_mask & __GFP_NOFAIL)
2064 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2065 * making forward progress without invoking OOM. Suspend also disables
2066 * storage devices so kswapd will not help. Bail if we are suspending.
2068 if (!did_some_progress && pm_suspended_storage())
2072 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2073 * means __GFP_NOFAIL, but that may not be true in other
2076 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2080 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2081 * specified, then we retry until we no longer reclaim any pages
2082 * (above), or we've reclaimed an order of pages at least as
2083 * large as the allocation's order. In both cases, if the
2084 * allocation still fails, we stop retrying.
2086 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2092 static inline struct page *
2093 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2094 struct zonelist *zonelist, enum zone_type high_zoneidx,
2095 nodemask_t *nodemask, struct zone *preferred_zone,
2100 /* Acquire the OOM killer lock for the zones in zonelist */
2101 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2102 schedule_timeout_uninterruptible(1);
2107 * Go through the zonelist yet one more time, keep very high watermark
2108 * here, this is only to catch a parallel oom killing, we must fail if
2109 * we're still under heavy pressure.
2111 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2112 order, zonelist, high_zoneidx,
2113 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2114 preferred_zone, migratetype);
2118 if (!(gfp_mask & __GFP_NOFAIL)) {
2119 /* The OOM killer will not help higher order allocs */
2120 if (order > PAGE_ALLOC_COSTLY_ORDER)
2122 /* The OOM killer does not needlessly kill tasks for lowmem */
2123 if (high_zoneidx < ZONE_NORMAL)
2126 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2127 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2128 * The caller should handle page allocation failure by itself if
2129 * it specifies __GFP_THISNODE.
2130 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2132 if (gfp_mask & __GFP_THISNODE)
2135 /* Exhausted what can be done so it's blamo time */
2136 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2139 clear_zonelist_oom(zonelist, gfp_mask);
2143 #ifdef CONFIG_COMPACTION
2144 /* Try memory compaction for high-order allocations before reclaim */
2145 static struct page *
2146 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2147 struct zonelist *zonelist, enum zone_type high_zoneidx,
2148 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2149 int migratetype, bool sync_migration,
2150 bool *contended_compaction, bool *deferred_compaction,
2151 unsigned long *did_some_progress)
2153 struct page *page = NULL;
2158 if (compaction_deferred(preferred_zone, order)) {
2159 *deferred_compaction = true;
2163 current->flags |= PF_MEMALLOC;
2164 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2165 nodemask, sync_migration,
2166 contended_compaction, &page);
2167 current->flags &= ~PF_MEMALLOC;
2169 /* If compaction captured a page, prep and use it */
2171 prep_new_page(page, order, gfp_mask);
2175 if (*did_some_progress != COMPACT_SKIPPED) {
2176 /* Page migration frees to the PCP lists but we want merging */
2177 drain_pages(get_cpu());
2180 page = get_page_from_freelist(gfp_mask, nodemask,
2181 order, zonelist, high_zoneidx,
2182 alloc_flags & ~ALLOC_NO_WATERMARKS,
2183 preferred_zone, migratetype);
2186 preferred_zone->compact_blockskip_flush = false;
2187 preferred_zone->compact_considered = 0;
2188 preferred_zone->compact_defer_shift = 0;
2189 if (order >= preferred_zone->compact_order_failed)
2190 preferred_zone->compact_order_failed = order + 1;
2191 count_vm_event(COMPACTSUCCESS);
2196 * It's bad if compaction run occurs and fails.
2197 * The most likely reason is that pages exist,
2198 * but not enough to satisfy watermarks.
2200 count_vm_event(COMPACTFAIL);
2203 * As async compaction considers a subset of pageblocks, only
2204 * defer if the failure was a sync compaction failure.
2207 defer_compaction(preferred_zone, order);
2215 static inline struct page *
2216 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2217 struct zonelist *zonelist, enum zone_type high_zoneidx,
2218 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2219 int migratetype, bool sync_migration,
2220 bool *contended_compaction, bool *deferred_compaction,
2221 unsigned long *did_some_progress)
2225 #endif /* CONFIG_COMPACTION */
2227 /* Perform direct synchronous page reclaim */
2229 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2230 nodemask_t *nodemask)
2232 struct reclaim_state reclaim_state;
2237 /* We now go into synchronous reclaim */
2238 cpuset_memory_pressure_bump();
2239 current->flags |= PF_MEMALLOC;
2240 lockdep_set_current_reclaim_state(gfp_mask);
2241 reclaim_state.reclaimed_slab = 0;
2242 current->reclaim_state = &reclaim_state;
2244 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2246 current->reclaim_state = NULL;
2247 lockdep_clear_current_reclaim_state();
2248 current->flags &= ~PF_MEMALLOC;
2255 /* The really slow allocator path where we enter direct reclaim */
2256 static inline struct page *
2257 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2258 struct zonelist *zonelist, enum zone_type high_zoneidx,
2259 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2260 int migratetype, unsigned long *did_some_progress)
2262 struct page *page = NULL;
2263 bool drained = false;
2265 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2267 if (unlikely(!(*did_some_progress)))
2270 /* After successful reclaim, reconsider all zones for allocation */
2271 if (IS_ENABLED(CONFIG_NUMA))
2272 zlc_clear_zones_full(zonelist);
2275 page = get_page_from_freelist(gfp_mask, nodemask, order,
2276 zonelist, high_zoneidx,
2277 alloc_flags & ~ALLOC_NO_WATERMARKS,
2278 preferred_zone, migratetype);
2281 * If an allocation failed after direct reclaim, it could be because
2282 * pages are pinned on the per-cpu lists. Drain them and try again
2284 if (!page && !drained) {
2294 * This is called in the allocator slow-path if the allocation request is of
2295 * sufficient urgency to ignore watermarks and take other desperate measures
2297 static inline struct page *
2298 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2299 struct zonelist *zonelist, enum zone_type high_zoneidx,
2300 nodemask_t *nodemask, struct zone *preferred_zone,
2306 page = get_page_from_freelist(gfp_mask, nodemask, order,
2307 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2308 preferred_zone, migratetype);
2310 if (!page && gfp_mask & __GFP_NOFAIL)
2311 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2312 } while (!page && (gfp_mask & __GFP_NOFAIL));
2318 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2319 enum zone_type high_zoneidx,
2320 enum zone_type classzone_idx)
2325 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2326 wakeup_kswapd(zone, order, classzone_idx);
2330 gfp_to_alloc_flags(gfp_t gfp_mask)
2332 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2333 const gfp_t wait = gfp_mask & __GFP_WAIT;
2335 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2336 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2339 * The caller may dip into page reserves a bit more if the caller
2340 * cannot run direct reclaim, or if the caller has realtime scheduling
2341 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2342 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2344 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2348 * Not worth trying to allocate harder for
2349 * __GFP_NOMEMALLOC even if it can't schedule.
2351 if (!(gfp_mask & __GFP_NOMEMALLOC))
2352 alloc_flags |= ALLOC_HARDER;
2354 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2355 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2357 alloc_flags &= ~ALLOC_CPUSET;
2358 } else if (unlikely(rt_task(current)) && !in_interrupt())
2359 alloc_flags |= ALLOC_HARDER;
2361 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2362 if (gfp_mask & __GFP_MEMALLOC)
2363 alloc_flags |= ALLOC_NO_WATERMARKS;
2364 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2365 alloc_flags |= ALLOC_NO_WATERMARKS;
2366 else if (!in_interrupt() &&
2367 ((current->flags & PF_MEMALLOC) ||
2368 unlikely(test_thread_flag(TIF_MEMDIE))))
2369 alloc_flags |= ALLOC_NO_WATERMARKS;
2372 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2373 alloc_flags |= ALLOC_CMA;
2378 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2380 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2383 static inline struct page *
2384 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2385 struct zonelist *zonelist, enum zone_type high_zoneidx,
2386 nodemask_t *nodemask, struct zone *preferred_zone,
2389 const gfp_t wait = gfp_mask & __GFP_WAIT;
2390 struct page *page = NULL;
2392 unsigned long pages_reclaimed = 0;
2393 unsigned long did_some_progress;
2394 bool sync_migration = false;
2395 bool deferred_compaction = false;
2396 bool contended_compaction = false;
2399 * In the slowpath, we sanity check order to avoid ever trying to
2400 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2401 * be using allocators in order of preference for an area that is
2404 if (order >= MAX_ORDER) {
2405 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2410 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2411 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2412 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2413 * using a larger set of nodes after it has established that the
2414 * allowed per node queues are empty and that nodes are
2417 if (IS_ENABLED(CONFIG_NUMA) &&
2418 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2422 wake_all_kswapd(order, zonelist, high_zoneidx,
2423 zone_idx(preferred_zone));
2426 * OK, we're below the kswapd watermark and have kicked background
2427 * reclaim. Now things get more complex, so set up alloc_flags according
2428 * to how we want to proceed.
2430 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2433 * Find the true preferred zone if the allocation is unconstrained by
2436 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2437 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2441 /* This is the last chance, in general, before the goto nopage. */
2442 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2443 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2444 preferred_zone, migratetype);
2448 /* Allocate without watermarks if the context allows */
2449 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2451 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2452 * the allocation is high priority and these type of
2453 * allocations are system rather than user orientated
2455 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2457 page = __alloc_pages_high_priority(gfp_mask, order,
2458 zonelist, high_zoneidx, nodemask,
2459 preferred_zone, migratetype);
2465 /* Atomic allocations - we can't balance anything */
2469 /* Avoid recursion of direct reclaim */
2470 if (current->flags & PF_MEMALLOC)
2473 /* Avoid allocations with no watermarks from looping endlessly */
2474 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2478 * Try direct compaction. The first pass is asynchronous. Subsequent
2479 * attempts after direct reclaim are synchronous
2481 page = __alloc_pages_direct_compact(gfp_mask, order,
2482 zonelist, high_zoneidx,
2484 alloc_flags, preferred_zone,
2485 migratetype, sync_migration,
2486 &contended_compaction,
2487 &deferred_compaction,
2488 &did_some_progress);
2491 sync_migration = true;
2494 * If compaction is deferred for high-order allocations, it is because
2495 * sync compaction recently failed. In this is the case and the caller
2496 * requested a movable allocation that does not heavily disrupt the
2497 * system then fail the allocation instead of entering direct reclaim.
2499 if ((deferred_compaction || contended_compaction) &&
2500 (gfp_mask & (__GFP_MOVABLE|__GFP_REPEAT)) == __GFP_MOVABLE)
2503 /* Try direct reclaim and then allocating */
2504 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2505 zonelist, high_zoneidx,
2507 alloc_flags, preferred_zone,
2508 migratetype, &did_some_progress);
2513 * If we failed to make any progress reclaiming, then we are
2514 * running out of options and have to consider going OOM
2516 if (!did_some_progress) {
2517 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2518 if (oom_killer_disabled)
2520 /* Coredumps can quickly deplete all memory reserves */
2521 if ((current->flags & PF_DUMPCORE) &&
2522 !(gfp_mask & __GFP_NOFAIL))
2524 page = __alloc_pages_may_oom(gfp_mask, order,
2525 zonelist, high_zoneidx,
2526 nodemask, preferred_zone,
2531 if (!(gfp_mask & __GFP_NOFAIL)) {
2533 * The oom killer is not called for high-order
2534 * allocations that may fail, so if no progress
2535 * is being made, there are no other options and
2536 * retrying is unlikely to help.
2538 if (order > PAGE_ALLOC_COSTLY_ORDER)
2541 * The oom killer is not called for lowmem
2542 * allocations to prevent needlessly killing
2545 if (high_zoneidx < ZONE_NORMAL)
2553 /* Check if we should retry the allocation */
2554 pages_reclaimed += did_some_progress;
2555 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2557 /* Wait for some write requests to complete then retry */
2558 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2562 * High-order allocations do not necessarily loop after
2563 * direct reclaim and reclaim/compaction depends on compaction
2564 * being called after reclaim so call directly if necessary
2566 page = __alloc_pages_direct_compact(gfp_mask, order,
2567 zonelist, high_zoneidx,
2569 alloc_flags, preferred_zone,
2570 migratetype, sync_migration,
2571 &contended_compaction,
2572 &deferred_compaction,
2573 &did_some_progress);
2579 warn_alloc_failed(gfp_mask, order, NULL);
2582 if (kmemcheck_enabled)
2583 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2589 * This is the 'heart' of the zoned buddy allocator.
2592 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2593 struct zonelist *zonelist, nodemask_t *nodemask)
2595 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2596 struct zone *preferred_zone;
2597 struct page *page = NULL;
2598 int migratetype = allocflags_to_migratetype(gfp_mask);
2599 unsigned int cpuset_mems_cookie;
2600 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2601 struct mem_cgroup *memcg = NULL;
2603 gfp_mask &= gfp_allowed_mask;
2605 lockdep_trace_alloc(gfp_mask);
2607 might_sleep_if(gfp_mask & __GFP_WAIT);
2609 if (should_fail_alloc_page(gfp_mask, order))
2613 * Check the zones suitable for the gfp_mask contain at least one
2614 * valid zone. It's possible to have an empty zonelist as a result
2615 * of GFP_THISNODE and a memoryless node
2617 if (unlikely(!zonelist->_zonerefs->zone))
2621 * Will only have any effect when __GFP_KMEMCG is set. This is
2622 * verified in the (always inline) callee
2624 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2628 cpuset_mems_cookie = get_mems_allowed();
2630 /* The preferred zone is used for statistics later */
2631 first_zones_zonelist(zonelist, high_zoneidx,
2632 nodemask ? : &cpuset_current_mems_allowed,
2634 if (!preferred_zone)
2638 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2639 alloc_flags |= ALLOC_CMA;
2641 /* First allocation attempt */
2642 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2643 zonelist, high_zoneidx, alloc_flags,
2644 preferred_zone, migratetype);
2645 if (unlikely(!page))
2646 page = __alloc_pages_slowpath(gfp_mask, order,
2647 zonelist, high_zoneidx, nodemask,
2648 preferred_zone, migratetype);
2650 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2654 * When updating a task's mems_allowed, it is possible to race with
2655 * parallel threads in such a way that an allocation can fail while
2656 * the mask is being updated. If a page allocation is about to fail,
2657 * check if the cpuset changed during allocation and if so, retry.
2659 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2662 memcg_kmem_commit_charge(page, memcg, order);
2666 EXPORT_SYMBOL(__alloc_pages_nodemask);
2669 * Common helper functions.
2671 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2676 * __get_free_pages() returns a 32-bit address, which cannot represent
2679 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2681 page = alloc_pages(gfp_mask, order);
2684 return (unsigned long) page_address(page);
2686 EXPORT_SYMBOL(__get_free_pages);
2688 unsigned long get_zeroed_page(gfp_t gfp_mask)
2690 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2692 EXPORT_SYMBOL(get_zeroed_page);
2694 void __free_pages(struct page *page, unsigned int order)
2696 if (put_page_testzero(page)) {
2698 free_hot_cold_page(page, 0);
2700 __free_pages_ok(page, order);
2704 EXPORT_SYMBOL(__free_pages);
2706 void free_pages(unsigned long addr, unsigned int order)
2709 VM_BUG_ON(!virt_addr_valid((void *)addr));
2710 __free_pages(virt_to_page((void *)addr), order);
2714 EXPORT_SYMBOL(free_pages);
2717 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2718 * pages allocated with __GFP_KMEMCG.
2720 * Those pages are accounted to a particular memcg, embedded in the
2721 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2722 * for that information only to find out that it is NULL for users who have no
2723 * interest in that whatsoever, we provide these functions.
2725 * The caller knows better which flags it relies on.
2727 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2729 memcg_kmem_uncharge_pages(page, order);
2730 __free_pages(page, order);
2733 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2736 VM_BUG_ON(!virt_addr_valid((void *)addr));
2737 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2741 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2744 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2745 unsigned long used = addr + PAGE_ALIGN(size);
2747 split_page(virt_to_page((void *)addr), order);
2748 while (used < alloc_end) {
2753 return (void *)addr;
2757 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2758 * @size: the number of bytes to allocate
2759 * @gfp_mask: GFP flags for the allocation
2761 * This function is similar to alloc_pages(), except that it allocates the
2762 * minimum number of pages to satisfy the request. alloc_pages() can only
2763 * allocate memory in power-of-two pages.
2765 * This function is also limited by MAX_ORDER.
2767 * Memory allocated by this function must be released by free_pages_exact().
2769 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2771 unsigned int order = get_order(size);
2774 addr = __get_free_pages(gfp_mask, order);
2775 return make_alloc_exact(addr, order, size);
2777 EXPORT_SYMBOL(alloc_pages_exact);
2780 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2782 * @nid: the preferred node ID where memory should be allocated
2783 * @size: the number of bytes to allocate
2784 * @gfp_mask: GFP flags for the allocation
2786 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2788 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2791 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2793 unsigned order = get_order(size);
2794 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2797 return make_alloc_exact((unsigned long)page_address(p), order, size);
2799 EXPORT_SYMBOL(alloc_pages_exact_nid);
2802 * free_pages_exact - release memory allocated via alloc_pages_exact()
2803 * @virt: the value returned by alloc_pages_exact.
2804 * @size: size of allocation, same value as passed to alloc_pages_exact().
2806 * Release the memory allocated by a previous call to alloc_pages_exact.
2808 void free_pages_exact(void *virt, size_t size)
2810 unsigned long addr = (unsigned long)virt;
2811 unsigned long end = addr + PAGE_ALIGN(size);
2813 while (addr < end) {
2818 EXPORT_SYMBOL(free_pages_exact);
2820 static unsigned int nr_free_zone_pages(int offset)
2825 /* Just pick one node, since fallback list is circular */
2826 unsigned int sum = 0;
2828 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2830 for_each_zone_zonelist(zone, z, zonelist, offset) {
2831 unsigned long size = zone->present_pages;
2832 unsigned long high = high_wmark_pages(zone);
2841 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2843 unsigned int nr_free_buffer_pages(void)
2845 return nr_free_zone_pages(gfp_zone(GFP_USER));
2847 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2850 * Amount of free RAM allocatable within all zones
2852 unsigned int nr_free_pagecache_pages(void)
2854 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2857 static inline void show_node(struct zone *zone)
2859 if (IS_ENABLED(CONFIG_NUMA))
2860 printk("Node %d ", zone_to_nid(zone));
2863 void si_meminfo(struct sysinfo *val)
2865 val->totalram = totalram_pages;
2867 val->freeram = global_page_state(NR_FREE_PAGES);
2868 val->bufferram = nr_blockdev_pages();
2869 val->totalhigh = totalhigh_pages;
2870 val->freehigh = nr_free_highpages();
2871 val->mem_unit = PAGE_SIZE;
2874 EXPORT_SYMBOL(si_meminfo);
2877 void si_meminfo_node(struct sysinfo *val, int nid)
2879 pg_data_t *pgdat = NODE_DATA(nid);
2881 val->totalram = pgdat->node_present_pages;
2882 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2883 #ifdef CONFIG_HIGHMEM
2884 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2885 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2891 val->mem_unit = PAGE_SIZE;
2896 * Determine whether the node should be displayed or not, depending on whether
2897 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2899 bool skip_free_areas_node(unsigned int flags, int nid)
2902 unsigned int cpuset_mems_cookie;
2904 if (!(flags & SHOW_MEM_FILTER_NODES))
2908 cpuset_mems_cookie = get_mems_allowed();
2909 ret = !node_isset(nid, cpuset_current_mems_allowed);
2910 } while (!put_mems_allowed(cpuset_mems_cookie));
2915 #define K(x) ((x) << (PAGE_SHIFT-10))
2917 static void show_migration_types(unsigned char type)
2919 static const char types[MIGRATE_TYPES] = {
2920 [MIGRATE_UNMOVABLE] = 'U',
2921 [MIGRATE_RECLAIMABLE] = 'E',
2922 [MIGRATE_MOVABLE] = 'M',
2923 [MIGRATE_RESERVE] = 'R',
2925 [MIGRATE_CMA] = 'C',
2927 [MIGRATE_ISOLATE] = 'I',
2929 char tmp[MIGRATE_TYPES + 1];
2933 for (i = 0; i < MIGRATE_TYPES; i++) {
2934 if (type & (1 << i))
2939 printk("(%s) ", tmp);
2943 * Show free area list (used inside shift_scroll-lock stuff)
2944 * We also calculate the percentage fragmentation. We do this by counting the
2945 * memory on each free list with the exception of the first item on the list.
2946 * Suppresses nodes that are not allowed by current's cpuset if
2947 * SHOW_MEM_FILTER_NODES is passed.
2949 void show_free_areas(unsigned int filter)
2954 for_each_populated_zone(zone) {
2955 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2958 printk("%s per-cpu:\n", zone->name);
2960 for_each_online_cpu(cpu) {
2961 struct per_cpu_pageset *pageset;
2963 pageset = per_cpu_ptr(zone->pageset, cpu);
2965 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2966 cpu, pageset->pcp.high,
2967 pageset->pcp.batch, pageset->pcp.count);
2971 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2972 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2974 " dirty:%lu writeback:%lu unstable:%lu\n"
2975 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2976 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2978 global_page_state(NR_ACTIVE_ANON),
2979 global_page_state(NR_INACTIVE_ANON),
2980 global_page_state(NR_ISOLATED_ANON),
2981 global_page_state(NR_ACTIVE_FILE),
2982 global_page_state(NR_INACTIVE_FILE),
2983 global_page_state(NR_ISOLATED_FILE),
2984 global_page_state(NR_UNEVICTABLE),
2985 global_page_state(NR_FILE_DIRTY),
2986 global_page_state(NR_WRITEBACK),
2987 global_page_state(NR_UNSTABLE_NFS),
2988 global_page_state(NR_FREE_PAGES),
2989 global_page_state(NR_SLAB_RECLAIMABLE),
2990 global_page_state(NR_SLAB_UNRECLAIMABLE),
2991 global_page_state(NR_FILE_MAPPED),
2992 global_page_state(NR_SHMEM),
2993 global_page_state(NR_PAGETABLE),
2994 global_page_state(NR_BOUNCE),
2995 global_page_state(NR_FREE_CMA_PAGES));
2997 for_each_populated_zone(zone) {
3000 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3008 " active_anon:%lukB"
3009 " inactive_anon:%lukB"
3010 " active_file:%lukB"
3011 " inactive_file:%lukB"
3012 " unevictable:%lukB"
3013 " isolated(anon):%lukB"
3014 " isolated(file):%lukB"
3021 " slab_reclaimable:%lukB"
3022 " slab_unreclaimable:%lukB"
3023 " kernel_stack:%lukB"
3028 " writeback_tmp:%lukB"
3029 " pages_scanned:%lu"
3030 " all_unreclaimable? %s"
3033 K(zone_page_state(zone, NR_FREE_PAGES)),
3034 K(min_wmark_pages(zone)),
3035 K(low_wmark_pages(zone)),
3036 K(high_wmark_pages(zone)),
3037 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3038 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3039 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3040 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3041 K(zone_page_state(zone, NR_UNEVICTABLE)),
3042 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3043 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3044 K(zone->present_pages),
3045 K(zone_page_state(zone, NR_MLOCK)),
3046 K(zone_page_state(zone, NR_FILE_DIRTY)),
3047 K(zone_page_state(zone, NR_WRITEBACK)),
3048 K(zone_page_state(zone, NR_FILE_MAPPED)),
3049 K(zone_page_state(zone, NR_SHMEM)),
3050 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3051 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3052 zone_page_state(zone, NR_KERNEL_STACK) *
3054 K(zone_page_state(zone, NR_PAGETABLE)),
3055 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3056 K(zone_page_state(zone, NR_BOUNCE)),
3057 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3058 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3059 zone->pages_scanned,
3060 (zone->all_unreclaimable ? "yes" : "no")
3062 printk("lowmem_reserve[]:");
3063 for (i = 0; i < MAX_NR_ZONES; i++)
3064 printk(" %lu", zone->lowmem_reserve[i]);
3068 for_each_populated_zone(zone) {
3069 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3070 unsigned char types[MAX_ORDER];
3072 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3075 printk("%s: ", zone->name);
3077 spin_lock_irqsave(&zone->lock, flags);
3078 for (order = 0; order < MAX_ORDER; order++) {
3079 struct free_area *area = &zone->free_area[order];
3082 nr[order] = area->nr_free;
3083 total += nr[order] << order;
3086 for (type = 0; type < MIGRATE_TYPES; type++) {
3087 if (!list_empty(&area->free_list[type]))
3088 types[order] |= 1 << type;
3091 spin_unlock_irqrestore(&zone->lock, flags);
3092 for (order = 0; order < MAX_ORDER; order++) {
3093 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3095 show_migration_types(types[order]);
3097 printk("= %lukB\n", K(total));
3100 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3102 show_swap_cache_info();
3105 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3107 zoneref->zone = zone;
3108 zoneref->zone_idx = zone_idx(zone);
3112 * Builds allocation fallback zone lists.
3114 * Add all populated zones of a node to the zonelist.
3116 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3117 int nr_zones, enum zone_type zone_type)
3121 BUG_ON(zone_type >= MAX_NR_ZONES);
3126 zone = pgdat->node_zones + zone_type;
3127 if (populated_zone(zone)) {
3128 zoneref_set_zone(zone,
3129 &zonelist->_zonerefs[nr_zones++]);
3130 check_highest_zone(zone_type);
3133 } while (zone_type);
3140 * 0 = automatic detection of better ordering.
3141 * 1 = order by ([node] distance, -zonetype)
3142 * 2 = order by (-zonetype, [node] distance)
3144 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3145 * the same zonelist. So only NUMA can configure this param.
3147 #define ZONELIST_ORDER_DEFAULT 0
3148 #define ZONELIST_ORDER_NODE 1
3149 #define ZONELIST_ORDER_ZONE 2
3151 /* zonelist order in the kernel.
3152 * set_zonelist_order() will set this to NODE or ZONE.
3154 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3155 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3159 /* The value user specified ....changed by config */
3160 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3161 /* string for sysctl */
3162 #define NUMA_ZONELIST_ORDER_LEN 16
3163 char numa_zonelist_order[16] = "default";
3166 * interface for configure zonelist ordering.
3167 * command line option "numa_zonelist_order"
3168 * = "[dD]efault - default, automatic configuration.
3169 * = "[nN]ode - order by node locality, then by zone within node
3170 * = "[zZ]one - order by zone, then by locality within zone
3173 static int __parse_numa_zonelist_order(char *s)
3175 if (*s == 'd' || *s == 'D') {
3176 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3177 } else if (*s == 'n' || *s == 'N') {
3178 user_zonelist_order = ZONELIST_ORDER_NODE;
3179 } else if (*s == 'z' || *s == 'Z') {
3180 user_zonelist_order = ZONELIST_ORDER_ZONE;
3183 "Ignoring invalid numa_zonelist_order value: "
3190 static __init int setup_numa_zonelist_order(char *s)
3197 ret = __parse_numa_zonelist_order(s);
3199 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3203 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3206 * sysctl handler for numa_zonelist_order
3208 int numa_zonelist_order_handler(ctl_table *table, int write,
3209 void __user *buffer, size_t *length,
3212 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3214 static DEFINE_MUTEX(zl_order_mutex);
3216 mutex_lock(&zl_order_mutex);
3218 strcpy(saved_string, (char*)table->data);
3219 ret = proc_dostring(table, write, buffer, length, ppos);
3223 int oldval = user_zonelist_order;
3224 if (__parse_numa_zonelist_order((char*)table->data)) {
3226 * bogus value. restore saved string
3228 strncpy((char*)table->data, saved_string,
3229 NUMA_ZONELIST_ORDER_LEN);
3230 user_zonelist_order = oldval;
3231 } else if (oldval != user_zonelist_order) {
3232 mutex_lock(&zonelists_mutex);
3233 build_all_zonelists(NULL, NULL);
3234 mutex_unlock(&zonelists_mutex);
3238 mutex_unlock(&zl_order_mutex);
3243 #define MAX_NODE_LOAD (nr_online_nodes)
3244 static int node_load[MAX_NUMNODES];
3247 * find_next_best_node - find the next node that should appear in a given node's fallback list
3248 * @node: node whose fallback list we're appending
3249 * @used_node_mask: nodemask_t of already used nodes
3251 * We use a number of factors to determine which is the next node that should
3252 * appear on a given node's fallback list. The node should not have appeared
3253 * already in @node's fallback list, and it should be the next closest node
3254 * according to the distance array (which contains arbitrary distance values
3255 * from each node to each node in the system), and should also prefer nodes
3256 * with no CPUs, since presumably they'll have very little allocation pressure
3257 * on them otherwise.
3258 * It returns -1 if no node is found.
3260 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3263 int min_val = INT_MAX;
3265 const struct cpumask *tmp = cpumask_of_node(0);
3267 /* Use the local node if we haven't already */
3268 if (!node_isset(node, *used_node_mask)) {
3269 node_set(node, *used_node_mask);
3273 for_each_node_state(n, N_HIGH_MEMORY) {
3275 /* Don't want a node to appear more than once */
3276 if (node_isset(n, *used_node_mask))
3279 /* Use the distance array to find the distance */
3280 val = node_distance(node, n);
3282 /* Penalize nodes under us ("prefer the next node") */
3285 /* Give preference to headless and unused nodes */
3286 tmp = cpumask_of_node(n);
3287 if (!cpumask_empty(tmp))
3288 val += PENALTY_FOR_NODE_WITH_CPUS;
3290 /* Slight preference for less loaded node */
3291 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3292 val += node_load[n];
3294 if (val < min_val) {
3301 node_set(best_node, *used_node_mask);
3308 * Build zonelists ordered by node and zones within node.
3309 * This results in maximum locality--normal zone overflows into local
3310 * DMA zone, if any--but risks exhausting DMA zone.
3312 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3315 struct zonelist *zonelist;
3317 zonelist = &pgdat->node_zonelists[0];
3318 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3320 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3322 zonelist->_zonerefs[j].zone = NULL;
3323 zonelist->_zonerefs[j].zone_idx = 0;
3327 * Build gfp_thisnode zonelists
3329 static void build_thisnode_zonelists(pg_data_t *pgdat)
3332 struct zonelist *zonelist;
3334 zonelist = &pgdat->node_zonelists[1];
3335 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3336 zonelist->_zonerefs[j].zone = NULL;
3337 zonelist->_zonerefs[j].zone_idx = 0;
3341 * Build zonelists ordered by zone and nodes within zones.
3342 * This results in conserving DMA zone[s] until all Normal memory is
3343 * exhausted, but results in overflowing to remote node while memory
3344 * may still exist in local DMA zone.
3346 static int node_order[MAX_NUMNODES];
3348 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3351 int zone_type; /* needs to be signed */
3353 struct zonelist *zonelist;
3355 zonelist = &pgdat->node_zonelists[0];
3357 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3358 for (j = 0; j < nr_nodes; j++) {
3359 node = node_order[j];
3360 z = &NODE_DATA(node)->node_zones[zone_type];
3361 if (populated_zone(z)) {
3363 &zonelist->_zonerefs[pos++]);
3364 check_highest_zone(zone_type);
3368 zonelist->_zonerefs[pos].zone = NULL;
3369 zonelist->_zonerefs[pos].zone_idx = 0;
3372 static int default_zonelist_order(void)
3375 unsigned long low_kmem_size,total_size;
3379 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3380 * If they are really small and used heavily, the system can fall
3381 * into OOM very easily.
3382 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3384 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3387 for_each_online_node(nid) {
3388 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3389 z = &NODE_DATA(nid)->node_zones[zone_type];
3390 if (populated_zone(z)) {
3391 if (zone_type < ZONE_NORMAL)
3392 low_kmem_size += z->present_pages;
3393 total_size += z->present_pages;
3394 } else if (zone_type == ZONE_NORMAL) {
3396 * If any node has only lowmem, then node order
3397 * is preferred to allow kernel allocations
3398 * locally; otherwise, they can easily infringe
3399 * on other nodes when there is an abundance of
3400 * lowmem available to allocate from.
3402 return ZONELIST_ORDER_NODE;
3406 if (!low_kmem_size || /* there are no DMA area. */
3407 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3408 return ZONELIST_ORDER_NODE;
3410 * look into each node's config.
3411 * If there is a node whose DMA/DMA32 memory is very big area on
3412 * local memory, NODE_ORDER may be suitable.
3414 average_size = total_size /
3415 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3416 for_each_online_node(nid) {
3419 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3420 z = &NODE_DATA(nid)->node_zones[zone_type];
3421 if (populated_zone(z)) {
3422 if (zone_type < ZONE_NORMAL)
3423 low_kmem_size += z->present_pages;
3424 total_size += z->present_pages;
3427 if (low_kmem_size &&
3428 total_size > average_size && /* ignore small node */
3429 low_kmem_size > total_size * 70/100)
3430 return ZONELIST_ORDER_NODE;
3432 return ZONELIST_ORDER_ZONE;
3435 static void set_zonelist_order(void)
3437 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3438 current_zonelist_order = default_zonelist_order();
3440 current_zonelist_order = user_zonelist_order;
3443 static void build_zonelists(pg_data_t *pgdat)
3447 nodemask_t used_mask;
3448 int local_node, prev_node;
3449 struct zonelist *zonelist;
3450 int order = current_zonelist_order;
3452 /* initialize zonelists */
3453 for (i = 0; i < MAX_ZONELISTS; i++) {
3454 zonelist = pgdat->node_zonelists + i;
3455 zonelist->_zonerefs[0].zone = NULL;
3456 zonelist->_zonerefs[0].zone_idx = 0;
3459 /* NUMA-aware ordering of nodes */
3460 local_node = pgdat->node_id;
3461 load = nr_online_nodes;
3462 prev_node = local_node;
3463 nodes_clear(used_mask);
3465 memset(node_order, 0, sizeof(node_order));
3468 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3470 * We don't want to pressure a particular node.
3471 * So adding penalty to the first node in same
3472 * distance group to make it round-robin.
3474 if (node_distance(local_node, node) !=
3475 node_distance(local_node, prev_node))
3476 node_load[node] = load;
3480 if (order == ZONELIST_ORDER_NODE)
3481 build_zonelists_in_node_order(pgdat, node);
3483 node_order[j++] = node; /* remember order */
3486 if (order == ZONELIST_ORDER_ZONE) {
3487 /* calculate node order -- i.e., DMA last! */
3488 build_zonelists_in_zone_order(pgdat, j);
3491 build_thisnode_zonelists(pgdat);
3494 /* Construct the zonelist performance cache - see further mmzone.h */
3495 static void build_zonelist_cache(pg_data_t *pgdat)
3497 struct zonelist *zonelist;
3498 struct zonelist_cache *zlc;
3501 zonelist = &pgdat->node_zonelists[0];
3502 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3503 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3504 for (z = zonelist->_zonerefs; z->zone; z++)
3505 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3508 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3510 * Return node id of node used for "local" allocations.
3511 * I.e., first node id of first zone in arg node's generic zonelist.
3512 * Used for initializing percpu 'numa_mem', which is used primarily
3513 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3515 int local_memory_node(int node)
3519 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3520 gfp_zone(GFP_KERNEL),
3527 #else /* CONFIG_NUMA */
3529 static void set_zonelist_order(void)
3531 current_zonelist_order = ZONELIST_ORDER_ZONE;
3534 static void build_zonelists(pg_data_t *pgdat)
3536 int node, local_node;
3538 struct zonelist *zonelist;
3540 local_node = pgdat->node_id;
3542 zonelist = &pgdat->node_zonelists[0];
3543 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3546 * Now we build the zonelist so that it contains the zones
3547 * of all the other nodes.
3548 * We don't want to pressure a particular node, so when
3549 * building the zones for node N, we make sure that the
3550 * zones coming right after the local ones are those from
3551 * node N+1 (modulo N)
3553 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3554 if (!node_online(node))
3556 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3559 for (node = 0; node < local_node; node++) {
3560 if (!node_online(node))
3562 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3566 zonelist->_zonerefs[j].zone = NULL;
3567 zonelist->_zonerefs[j].zone_idx = 0;
3570 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3571 static void build_zonelist_cache(pg_data_t *pgdat)
3573 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3576 #endif /* CONFIG_NUMA */
3579 * Boot pageset table. One per cpu which is going to be used for all
3580 * zones and all nodes. The parameters will be set in such a way
3581 * that an item put on a list will immediately be handed over to
3582 * the buddy list. This is safe since pageset manipulation is done
3583 * with interrupts disabled.
3585 * The boot_pagesets must be kept even after bootup is complete for
3586 * unused processors and/or zones. They do play a role for bootstrapping
3587 * hotplugged processors.
3589 * zoneinfo_show() and maybe other functions do
3590 * not check if the processor is online before following the pageset pointer.
3591 * Other parts of the kernel may not check if the zone is available.
3593 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3594 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3595 static void setup_zone_pageset(struct zone *zone);
3598 * Global mutex to protect against size modification of zonelists
3599 * as well as to serialize pageset setup for the new populated zone.
3601 DEFINE_MUTEX(zonelists_mutex);
3603 /* return values int ....just for stop_machine() */
3604 static int __build_all_zonelists(void *data)
3608 pg_data_t *self = data;
3611 memset(node_load, 0, sizeof(node_load));
3614 if (self && !node_online(self->node_id)) {
3615 build_zonelists(self);
3616 build_zonelist_cache(self);
3619 for_each_online_node(nid) {
3620 pg_data_t *pgdat = NODE_DATA(nid);
3622 build_zonelists(pgdat);
3623 build_zonelist_cache(pgdat);
3627 * Initialize the boot_pagesets that are going to be used
3628 * for bootstrapping processors. The real pagesets for
3629 * each zone will be allocated later when the per cpu
3630 * allocator is available.
3632 * boot_pagesets are used also for bootstrapping offline
3633 * cpus if the system is already booted because the pagesets
3634 * are needed to initialize allocators on a specific cpu too.
3635 * F.e. the percpu allocator needs the page allocator which
3636 * needs the percpu allocator in order to allocate its pagesets
3637 * (a chicken-egg dilemma).
3639 for_each_possible_cpu(cpu) {
3640 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3642 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3644 * We now know the "local memory node" for each node--
3645 * i.e., the node of the first zone in the generic zonelist.
3646 * Set up numa_mem percpu variable for on-line cpus. During
3647 * boot, only the boot cpu should be on-line; we'll init the
3648 * secondary cpus' numa_mem as they come on-line. During
3649 * node/memory hotplug, we'll fixup all on-line cpus.
3651 if (cpu_online(cpu))
3652 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3660 * Called with zonelists_mutex held always
3661 * unless system_state == SYSTEM_BOOTING.
3663 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3665 set_zonelist_order();
3667 if (system_state == SYSTEM_BOOTING) {
3668 __build_all_zonelists(NULL);
3669 mminit_verify_zonelist();
3670 cpuset_init_current_mems_allowed();
3672 /* we have to stop all cpus to guarantee there is no user
3674 #ifdef CONFIG_MEMORY_HOTPLUG
3676 setup_zone_pageset(zone);
3678 stop_machine(__build_all_zonelists, pgdat, NULL);
3679 /* cpuset refresh routine should be here */
3681 vm_total_pages = nr_free_pagecache_pages();
3683 * Disable grouping by mobility if the number of pages in the
3684 * system is too low to allow the mechanism to work. It would be
3685 * more accurate, but expensive to check per-zone. This check is
3686 * made on memory-hotadd so a system can start with mobility
3687 * disabled and enable it later
3689 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3690 page_group_by_mobility_disabled = 1;
3692 page_group_by_mobility_disabled = 0;
3694 printk("Built %i zonelists in %s order, mobility grouping %s. "
3695 "Total pages: %ld\n",
3697 zonelist_order_name[current_zonelist_order],
3698 page_group_by_mobility_disabled ? "off" : "on",
3701 printk("Policy zone: %s\n", zone_names[policy_zone]);
3706 * Helper functions to size the waitqueue hash table.
3707 * Essentially these want to choose hash table sizes sufficiently
3708 * large so that collisions trying to wait on pages are rare.
3709 * But in fact, the number of active page waitqueues on typical
3710 * systems is ridiculously low, less than 200. So this is even
3711 * conservative, even though it seems large.
3713 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3714 * waitqueues, i.e. the size of the waitq table given the number of pages.
3716 #define PAGES_PER_WAITQUEUE 256
3718 #ifndef CONFIG_MEMORY_HOTPLUG
3719 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3721 unsigned long size = 1;
3723 pages /= PAGES_PER_WAITQUEUE;
3725 while (size < pages)
3729 * Once we have dozens or even hundreds of threads sleeping
3730 * on IO we've got bigger problems than wait queue collision.
3731 * Limit the size of the wait table to a reasonable size.
3733 size = min(size, 4096UL);
3735 return max(size, 4UL);
3739 * A zone's size might be changed by hot-add, so it is not possible to determine
3740 * a suitable size for its wait_table. So we use the maximum size now.
3742 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3744 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3745 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3746 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3748 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3749 * or more by the traditional way. (See above). It equals:
3751 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3752 * ia64(16K page size) : = ( 8G + 4M)byte.
3753 * powerpc (64K page size) : = (32G +16M)byte.
3755 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3762 * This is an integer logarithm so that shifts can be used later
3763 * to extract the more random high bits from the multiplicative
3764 * hash function before the remainder is taken.
3766 static inline unsigned long wait_table_bits(unsigned long size)
3771 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3774 * Check if a pageblock contains reserved pages
3776 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3780 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3781 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3788 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3789 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3790 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3791 * higher will lead to a bigger reserve which will get freed as contiguous
3792 * blocks as reclaim kicks in
3794 static void setup_zone_migrate_reserve(struct zone *zone)
3796 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3798 unsigned long block_migratetype;
3802 * Get the start pfn, end pfn and the number of blocks to reserve
3803 * We have to be careful to be aligned to pageblock_nr_pages to
3804 * make sure that we always check pfn_valid for the first page in
3807 start_pfn = zone->zone_start_pfn;
3808 end_pfn = start_pfn + zone->spanned_pages;
3809 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3810 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3814 * Reserve blocks are generally in place to help high-order atomic
3815 * allocations that are short-lived. A min_free_kbytes value that
3816 * would result in more than 2 reserve blocks for atomic allocations
3817 * is assumed to be in place to help anti-fragmentation for the
3818 * future allocation of hugepages at runtime.
3820 reserve = min(2, reserve);
3822 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3823 if (!pfn_valid(pfn))
3825 page = pfn_to_page(pfn);
3827 /* Watch out for overlapping nodes */
3828 if (page_to_nid(page) != zone_to_nid(zone))
3831 block_migratetype = get_pageblock_migratetype(page);
3833 /* Only test what is necessary when the reserves are not met */
3836 * Blocks with reserved pages will never free, skip
3839 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3840 if (pageblock_is_reserved(pfn, block_end_pfn))
3843 /* If this block is reserved, account for it */
3844 if (block_migratetype == MIGRATE_RESERVE) {
3849 /* Suitable for reserving if this block is movable */
3850 if (block_migratetype == MIGRATE_MOVABLE) {
3851 set_pageblock_migratetype(page,
3853 move_freepages_block(zone, page,
3861 * If the reserve is met and this is a previous reserved block,
3864 if (block_migratetype == MIGRATE_RESERVE) {
3865 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3866 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3872 * Initially all pages are reserved - free ones are freed
3873 * up by free_all_bootmem() once the early boot process is
3874 * done. Non-atomic initialization, single-pass.
3876 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3877 unsigned long start_pfn, enum memmap_context context)
3880 unsigned long end_pfn = start_pfn + size;
3884 if (highest_memmap_pfn < end_pfn - 1)
3885 highest_memmap_pfn = end_pfn - 1;
3887 z = &NODE_DATA(nid)->node_zones[zone];
3888 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3890 * There can be holes in boot-time mem_map[]s
3891 * handed to this function. They do not
3892 * exist on hotplugged memory.
3894 if (context == MEMMAP_EARLY) {
3895 if (!early_pfn_valid(pfn)) {
3896 pfn = ALIGN(pfn + MAX_ORDER_NR_PAGES,
3897 MAX_ORDER_NR_PAGES) - 1;
3900 if (!early_pfn_in_nid(pfn, nid))
3903 page = pfn_to_page(pfn);
3904 set_page_links(page, zone, nid, pfn);
3905 mminit_verify_page_links(page, zone, nid, pfn);
3906 init_page_count(page);
3907 reset_page_mapcount(page);
3908 SetPageReserved(page);
3910 * Mark the block movable so that blocks are reserved for
3911 * movable at startup. This will force kernel allocations
3912 * to reserve their blocks rather than leaking throughout
3913 * the address space during boot when many long-lived
3914 * kernel allocations are made. Later some blocks near
3915 * the start are marked MIGRATE_RESERVE by
3916 * setup_zone_migrate_reserve()
3918 * bitmap is created for zone's valid pfn range. but memmap
3919 * can be created for invalid pages (for alignment)
3920 * check here not to call set_pageblock_migratetype() against
3923 if ((z->zone_start_pfn <= pfn)
3924 && (pfn < z->zone_start_pfn + z->spanned_pages)
3925 && !(pfn & (pageblock_nr_pages - 1)))
3926 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3928 INIT_LIST_HEAD(&page->lru);
3929 #ifdef WANT_PAGE_VIRTUAL
3930 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3931 if (!is_highmem_idx(zone))
3932 set_page_address(page, __va(pfn << PAGE_SHIFT));
3937 static void __meminit zone_init_free_lists(struct zone *zone)
3940 for_each_migratetype_order(order, t) {
3941 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3942 zone->free_area[order].nr_free = 0;
3946 #ifndef __HAVE_ARCH_MEMMAP_INIT
3947 #define memmap_init(size, nid, zone, start_pfn) \
3948 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3951 static int __meminit zone_batchsize(struct zone *zone)
3957 * The per-cpu-pages pools are set to around 1000th of the
3958 * size of the zone. But no more than 1/2 of a meg.
3960 * OK, so we don't know how big the cache is. So guess.
3962 batch = zone->present_pages / 1024;
3963 if (batch * PAGE_SIZE > 512 * 1024)
3964 batch = (512 * 1024) / PAGE_SIZE;
3965 batch /= 4; /* We effectively *= 4 below */
3970 * Clamp the batch to a 2^n - 1 value. Having a power
3971 * of 2 value was found to be more likely to have
3972 * suboptimal cache aliasing properties in some cases.
3974 * For example if 2 tasks are alternately allocating
3975 * batches of pages, one task can end up with a lot
3976 * of pages of one half of the possible page colors
3977 * and the other with pages of the other colors.
3979 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3984 /* The deferral and batching of frees should be suppressed under NOMMU
3987 * The problem is that NOMMU needs to be able to allocate large chunks
3988 * of contiguous memory as there's no hardware page translation to
3989 * assemble apparent contiguous memory from discontiguous pages.
3991 * Queueing large contiguous runs of pages for batching, however,
3992 * causes the pages to actually be freed in smaller chunks. As there
3993 * can be a significant delay between the individual batches being
3994 * recycled, this leads to the once large chunks of space being
3995 * fragmented and becoming unavailable for high-order allocations.
4001 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4003 struct per_cpu_pages *pcp;
4006 memset(p, 0, sizeof(*p));
4010 pcp->high = 6 * batch;
4011 pcp->batch = max(1UL, 1 * batch);
4012 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4013 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4017 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4018 * to the value high for the pageset p.
4021 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
4024 struct per_cpu_pages *pcp;
4028 pcp->batch = max(1UL, high/4);
4029 if ((high/4) > (PAGE_SHIFT * 8))
4030 pcp->batch = PAGE_SHIFT * 8;
4033 static void __meminit setup_zone_pageset(struct zone *zone)
4037 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4039 for_each_possible_cpu(cpu) {
4040 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4042 setup_pageset(pcp, zone_batchsize(zone));
4044 if (percpu_pagelist_fraction)
4045 setup_pagelist_highmark(pcp,
4046 (zone->present_pages /
4047 percpu_pagelist_fraction));
4052 * Allocate per cpu pagesets and initialize them.
4053 * Before this call only boot pagesets were available.
4055 void __init setup_per_cpu_pageset(void)
4059 for_each_populated_zone(zone)
4060 setup_zone_pageset(zone);
4063 static noinline __init_refok
4064 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4067 struct pglist_data *pgdat = zone->zone_pgdat;
4071 * The per-page waitqueue mechanism uses hashed waitqueues
4074 zone->wait_table_hash_nr_entries =
4075 wait_table_hash_nr_entries(zone_size_pages);
4076 zone->wait_table_bits =
4077 wait_table_bits(zone->wait_table_hash_nr_entries);
4078 alloc_size = zone->wait_table_hash_nr_entries
4079 * sizeof(wait_queue_head_t);
4081 if (!slab_is_available()) {
4082 zone->wait_table = (wait_queue_head_t *)
4083 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4086 * This case means that a zone whose size was 0 gets new memory
4087 * via memory hot-add.
4088 * But it may be the case that a new node was hot-added. In
4089 * this case vmalloc() will not be able to use this new node's
4090 * memory - this wait_table must be initialized to use this new
4091 * node itself as well.
4092 * To use this new node's memory, further consideration will be
4095 zone->wait_table = vmalloc(alloc_size);
4097 if (!zone->wait_table)
4100 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4101 init_waitqueue_head(zone->wait_table + i);
4106 static __meminit void zone_pcp_init(struct zone *zone)
4109 * per cpu subsystem is not up at this point. The following code
4110 * relies on the ability of the linker to provide the
4111 * offset of a (static) per cpu variable into the per cpu area.
4113 zone->pageset = &boot_pageset;
4115 if (zone->present_pages)
4116 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4117 zone->name, zone->present_pages,
4118 zone_batchsize(zone));
4121 int __meminit init_currently_empty_zone(struct zone *zone,
4122 unsigned long zone_start_pfn,
4124 enum memmap_context context)
4126 struct pglist_data *pgdat = zone->zone_pgdat;
4128 ret = zone_wait_table_init(zone, size);
4131 pgdat->nr_zones = zone_idx(zone) + 1;
4133 zone->zone_start_pfn = zone_start_pfn;
4135 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4136 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4138 (unsigned long)zone_idx(zone),
4139 zone_start_pfn, (zone_start_pfn + size));
4141 zone_init_free_lists(zone);
4146 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4147 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4149 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4150 * Architectures may implement their own version but if add_active_range()
4151 * was used and there are no special requirements, this is a convenient
4154 int __meminit __early_pfn_to_nid(unsigned long pfn)
4156 unsigned long start_pfn, end_pfn;
4159 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4160 if (start_pfn <= pfn && pfn < end_pfn)
4162 /* This is a memory hole */
4165 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4167 int __meminit early_pfn_to_nid(unsigned long pfn)
4171 nid = __early_pfn_to_nid(pfn);
4174 /* just returns 0 */
4178 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4179 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4183 nid = __early_pfn_to_nid(pfn);
4184 if (nid >= 0 && nid != node)
4191 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4192 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4193 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4195 * If an architecture guarantees that all ranges registered with
4196 * add_active_ranges() contain no holes and may be freed, this
4197 * this function may be used instead of calling free_bootmem() manually.
4199 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4201 unsigned long start_pfn, end_pfn;
4204 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4205 start_pfn = min(start_pfn, max_low_pfn);
4206 end_pfn = min(end_pfn, max_low_pfn);
4208 if (start_pfn < end_pfn)
4209 free_bootmem_node(NODE_DATA(this_nid),
4210 PFN_PHYS(start_pfn),
4211 (end_pfn - start_pfn) << PAGE_SHIFT);
4216 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4217 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4219 * If an architecture guarantees that all ranges registered with
4220 * add_active_ranges() contain no holes and may be freed, this
4221 * function may be used instead of calling memory_present() manually.
4223 void __init sparse_memory_present_with_active_regions(int nid)
4225 unsigned long start_pfn, end_pfn;
4228 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4229 memory_present(this_nid, start_pfn, end_pfn);
4233 * get_pfn_range_for_nid - Return the start and end page frames for a node
4234 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4235 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4236 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4238 * It returns the start and end page frame of a node based on information
4239 * provided by an arch calling add_active_range(). If called for a node
4240 * with no available memory, a warning is printed and the start and end
4243 void __meminit get_pfn_range_for_nid(unsigned int nid,
4244 unsigned long *start_pfn, unsigned long *end_pfn)
4246 unsigned long this_start_pfn, this_end_pfn;
4252 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4253 *start_pfn = min(*start_pfn, this_start_pfn);
4254 *end_pfn = max(*end_pfn, this_end_pfn);
4257 if (*start_pfn == -1UL)
4262 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4263 * assumption is made that zones within a node are ordered in monotonic
4264 * increasing memory addresses so that the "highest" populated zone is used
4266 static void __init find_usable_zone_for_movable(void)
4269 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4270 if (zone_index == ZONE_MOVABLE)
4273 if (arch_zone_highest_possible_pfn[zone_index] >
4274 arch_zone_lowest_possible_pfn[zone_index])
4278 VM_BUG_ON(zone_index == -1);
4279 movable_zone = zone_index;
4283 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4284 * because it is sized independent of architecture. Unlike the other zones,
4285 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4286 * in each node depending on the size of each node and how evenly kernelcore
4287 * is distributed. This helper function adjusts the zone ranges
4288 * provided by the architecture for a given node by using the end of the
4289 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4290 * zones within a node are in order of monotonic increases memory addresses
4292 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4293 unsigned long zone_type,
4294 unsigned long node_start_pfn,
4295 unsigned long node_end_pfn,
4296 unsigned long *zone_start_pfn,
4297 unsigned long *zone_end_pfn)
4299 /* Only adjust if ZONE_MOVABLE is on this node */
4300 if (zone_movable_pfn[nid]) {
4301 /* Size ZONE_MOVABLE */
4302 if (zone_type == ZONE_MOVABLE) {
4303 *zone_start_pfn = zone_movable_pfn[nid];
4304 *zone_end_pfn = min(node_end_pfn,
4305 arch_zone_highest_possible_pfn[movable_zone]);
4307 /* Adjust for ZONE_MOVABLE starting within this range */
4308 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4309 *zone_end_pfn > zone_movable_pfn[nid]) {
4310 *zone_end_pfn = zone_movable_pfn[nid];
4312 /* Check if this whole range is within ZONE_MOVABLE */
4313 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4314 *zone_start_pfn = *zone_end_pfn;
4319 * Return the number of pages a zone spans in a node, including holes
4320 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4322 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4323 unsigned long zone_type,
4324 unsigned long *ignored)
4326 unsigned long node_start_pfn, node_end_pfn;
4327 unsigned long zone_start_pfn, zone_end_pfn;
4329 /* Get the start and end of the node and zone */
4330 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4331 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4332 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4333 adjust_zone_range_for_zone_movable(nid, zone_type,
4334 node_start_pfn, node_end_pfn,
4335 &zone_start_pfn, &zone_end_pfn);
4337 /* Check that this node has pages within the zone's required range */
4338 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4341 /* Move the zone boundaries inside the node if necessary */
4342 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4343 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4345 /* Return the spanned pages */
4346 return zone_end_pfn - zone_start_pfn;
4350 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4351 * then all holes in the requested range will be accounted for.
4353 unsigned long __meminit __absent_pages_in_range(int nid,
4354 unsigned long range_start_pfn,
4355 unsigned long range_end_pfn)
4357 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4358 unsigned long start_pfn, end_pfn;
4361 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4362 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4363 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4364 nr_absent -= end_pfn - start_pfn;
4370 * absent_pages_in_range - Return number of page frames in holes within a range
4371 * @start_pfn: The start PFN to start searching for holes
4372 * @end_pfn: The end PFN to stop searching for holes
4374 * It returns the number of pages frames in memory holes within a range.
4376 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4377 unsigned long end_pfn)
4379 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4382 /* Return the number of page frames in holes in a zone on a node */
4383 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4384 unsigned long zone_type,
4385 unsigned long *ignored)
4387 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4388 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4389 unsigned long node_start_pfn, node_end_pfn;
4390 unsigned long zone_start_pfn, zone_end_pfn;
4392 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4393 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4394 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4396 adjust_zone_range_for_zone_movable(nid, zone_type,
4397 node_start_pfn, node_end_pfn,
4398 &zone_start_pfn, &zone_end_pfn);
4399 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4402 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4403 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4404 unsigned long zone_type,
4405 unsigned long *zones_size)
4407 return zones_size[zone_type];
4410 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4411 unsigned long zone_type,
4412 unsigned long *zholes_size)
4417 return zholes_size[zone_type];
4420 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4422 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4423 unsigned long *zones_size, unsigned long *zholes_size)
4425 unsigned long realtotalpages, totalpages = 0;
4428 for (i = 0; i < MAX_NR_ZONES; i++)
4429 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4431 pgdat->node_spanned_pages = totalpages;
4433 realtotalpages = totalpages;
4434 for (i = 0; i < MAX_NR_ZONES; i++)
4436 zone_absent_pages_in_node(pgdat->node_id, i,
4438 pgdat->node_present_pages = realtotalpages;
4439 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4443 #ifndef CONFIG_SPARSEMEM
4445 * Calculate the size of the zone->blockflags rounded to an unsigned long
4446 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4447 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4448 * round what is now in bits to nearest long in bits, then return it in
4451 static unsigned long __init usemap_size(unsigned long zonesize)
4453 unsigned long usemapsize;
4455 usemapsize = roundup(zonesize, pageblock_nr_pages);
4456 usemapsize = usemapsize >> pageblock_order;
4457 usemapsize *= NR_PAGEBLOCK_BITS;
4458 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4460 return usemapsize / 8;
4463 static void __init setup_usemap(struct pglist_data *pgdat,
4464 struct zone *zone, unsigned long zonesize)
4466 unsigned long usemapsize = usemap_size(zonesize);
4467 zone->pageblock_flags = NULL;
4469 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4473 static inline void setup_usemap(struct pglist_data *pgdat,
4474 struct zone *zone, unsigned long zonesize) {}
4475 #endif /* CONFIG_SPARSEMEM */
4477 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4479 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4480 void __init set_pageblock_order(void)
4484 /* Check that pageblock_nr_pages has not already been setup */
4485 if (pageblock_order)
4488 if (HPAGE_SHIFT > PAGE_SHIFT)
4489 order = HUGETLB_PAGE_ORDER;
4491 order = MAX_ORDER - 1;
4494 * Assume the largest contiguous order of interest is a huge page.
4495 * This value may be variable depending on boot parameters on IA64 and
4498 pageblock_order = order;
4500 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4503 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4504 * is unused as pageblock_order is set at compile-time. See
4505 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4508 void __init set_pageblock_order(void)
4512 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4515 * Set up the zone data structures:
4516 * - mark all pages reserved
4517 * - mark all memory queues empty
4518 * - clear the memory bitmaps
4520 * NOTE: pgdat should get zeroed by caller.
4522 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4523 unsigned long *zones_size, unsigned long *zholes_size)
4526 int nid = pgdat->node_id;
4527 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4530 pgdat_resize_init(pgdat);
4531 init_waitqueue_head(&pgdat->kswapd_wait);
4532 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4533 pgdat_page_cgroup_init(pgdat);
4535 for (j = 0; j < MAX_NR_ZONES; j++) {
4536 struct zone *zone = pgdat->node_zones + j;
4537 unsigned long size, realsize, memmap_pages;
4539 size = zone_spanned_pages_in_node(nid, j, zones_size);
4540 realsize = size - zone_absent_pages_in_node(nid, j,
4544 * Adjust realsize so that it accounts for how much memory
4545 * is used by this zone for memmap. This affects the watermark
4546 * and per-cpu initialisations
4549 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4550 if (realsize >= memmap_pages) {
4551 realsize -= memmap_pages;
4554 " %s zone: %lu pages used for memmap\n",
4555 zone_names[j], memmap_pages);
4558 " %s zone: %lu pages exceeds realsize %lu\n",
4559 zone_names[j], memmap_pages, realsize);
4561 /* Account for reserved pages */
4562 if (j == 0 && realsize > dma_reserve) {
4563 realsize -= 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 += realsize;
4570 nr_all_pages += realsize;
4572 zone->spanned_pages = size;
4573 zone->present_pages = realsize;
4576 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4578 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4580 zone->name = zone_names[j];
4581 spin_lock_init(&zone->lock);
4582 spin_lock_init(&zone->lru_lock);
4583 zone_seqlock_init(zone);
4584 zone->zone_pgdat = pgdat;
4586 zone_pcp_init(zone);
4587 lruvec_init(&zone->lruvec, zone);
4591 set_pageblock_order();
4592 setup_usemap(pgdat, zone, size);
4593 ret = init_currently_empty_zone(zone, zone_start_pfn,
4594 size, MEMMAP_EARLY);
4596 memmap_init(size, nid, j, zone_start_pfn);
4597 zone_start_pfn += size;
4601 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4603 /* Skip empty nodes */
4604 if (!pgdat->node_spanned_pages)
4607 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4608 /* ia64 gets its own node_mem_map, before this, without bootmem */
4609 if (!pgdat->node_mem_map) {
4610 unsigned long size, start, end;
4614 * The zone's endpoints aren't required to be MAX_ORDER
4615 * aligned but the node_mem_map endpoints must be in order
4616 * for the buddy allocator to function correctly.
4618 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4619 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4620 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4621 size = (end - start) * sizeof(struct page);
4622 map = alloc_remap(pgdat->node_id, size);
4624 map = alloc_bootmem_node_nopanic(pgdat, size);
4625 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4627 #ifndef CONFIG_NEED_MULTIPLE_NODES
4629 * With no DISCONTIG, the global mem_map is just set as node 0's
4631 if (pgdat == NODE_DATA(0)) {
4632 mem_map = NODE_DATA(0)->node_mem_map;
4633 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4634 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4635 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4636 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4639 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4642 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4643 unsigned long node_start_pfn, unsigned long *zholes_size)
4645 pg_data_t *pgdat = NODE_DATA(nid);
4647 /* pg_data_t should be reset to zero when it's allocated */
4648 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4650 pgdat->node_id = nid;
4651 pgdat->node_start_pfn = node_start_pfn;
4652 init_zone_allows_reclaim(nid);
4653 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4655 alloc_node_mem_map(pgdat);
4656 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4657 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4658 nid, (unsigned long)pgdat,
4659 (unsigned long)pgdat->node_mem_map);
4662 free_area_init_core(pgdat, zones_size, zholes_size);
4665 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4667 #if MAX_NUMNODES > 1
4669 * Figure out the number of possible node ids.
4671 static void __init setup_nr_node_ids(void)
4674 unsigned int highest = 0;
4676 for_each_node_mask(node, node_possible_map)
4678 nr_node_ids = highest + 1;
4681 static inline void setup_nr_node_ids(void)
4687 * node_map_pfn_alignment - determine the maximum internode alignment
4689 * This function should be called after node map is populated and sorted.
4690 * It calculates the maximum power of two alignment which can distinguish
4693 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4694 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4695 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4696 * shifted, 1GiB is enough and this function will indicate so.
4698 * This is used to test whether pfn -> nid mapping of the chosen memory
4699 * model has fine enough granularity to avoid incorrect mapping for the
4700 * populated node map.
4702 * Returns the determined alignment in pfn's. 0 if there is no alignment
4703 * requirement (single node).
4705 unsigned long __init node_map_pfn_alignment(void)
4707 unsigned long accl_mask = 0, last_end = 0;
4708 unsigned long start, end, mask;
4712 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4713 if (!start || last_nid < 0 || last_nid == nid) {
4720 * Start with a mask granular enough to pin-point to the
4721 * start pfn and tick off bits one-by-one until it becomes
4722 * too coarse to separate the current node from the last.
4724 mask = ~((1 << __ffs(start)) - 1);
4725 while (mask && last_end <= (start & (mask << 1)))
4728 /* accumulate all internode masks */
4732 /* convert mask to number of pages */
4733 return ~accl_mask + 1;
4736 /* Find the lowest pfn for a node */
4737 static unsigned long __init find_min_pfn_for_node(int nid)
4739 unsigned long min_pfn = ULONG_MAX;
4740 unsigned long start_pfn;
4743 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4744 min_pfn = min(min_pfn, start_pfn);
4746 if (min_pfn == ULONG_MAX) {
4748 "Could not find start_pfn for node %d\n", nid);
4756 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4758 * It returns the minimum PFN based on information provided via
4759 * add_active_range().
4761 unsigned long __init find_min_pfn_with_active_regions(void)
4763 return find_min_pfn_for_node(MAX_NUMNODES);
4767 * early_calculate_totalpages()
4768 * Sum pages in active regions for movable zone.
4769 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4771 static unsigned long __init early_calculate_totalpages(void)
4773 unsigned long totalpages = 0;
4774 unsigned long start_pfn, end_pfn;
4777 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4778 unsigned long pages = end_pfn - start_pfn;
4780 totalpages += pages;
4782 node_set_state(nid, N_HIGH_MEMORY);
4788 * Find the PFN the Movable zone begins in each node. Kernel memory
4789 * is spread evenly between nodes as long as the nodes have enough
4790 * memory. When they don't, some nodes will have more kernelcore than
4793 static void __init find_zone_movable_pfns_for_nodes(void)
4796 unsigned long usable_startpfn;
4797 unsigned long kernelcore_node, kernelcore_remaining;
4798 /* save the state before borrow the nodemask */
4799 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4800 unsigned long totalpages = early_calculate_totalpages();
4801 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4804 * If movablecore was specified, calculate what size of
4805 * kernelcore that corresponds so that memory usable for
4806 * any allocation type is evenly spread. If both kernelcore
4807 * and movablecore are specified, then the value of kernelcore
4808 * will be used for required_kernelcore if it's greater than
4809 * what movablecore would have allowed.
4811 if (required_movablecore) {
4812 unsigned long corepages;
4815 * Round-up so that ZONE_MOVABLE is at least as large as what
4816 * was requested by the user
4818 required_movablecore =
4819 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4820 corepages = totalpages - required_movablecore;
4822 required_kernelcore = max(required_kernelcore, corepages);
4825 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4826 if (!required_kernelcore)
4829 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4830 find_usable_zone_for_movable();
4831 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4834 /* Spread kernelcore memory as evenly as possible throughout nodes */
4835 kernelcore_node = required_kernelcore / usable_nodes;
4836 for_each_node_state(nid, N_HIGH_MEMORY) {
4837 unsigned long start_pfn, end_pfn;
4840 * Recalculate kernelcore_node if the division per node
4841 * now exceeds what is necessary to satisfy the requested
4842 * amount of memory for the kernel
4844 if (required_kernelcore < kernelcore_node)
4845 kernelcore_node = required_kernelcore / usable_nodes;
4848 * As the map is walked, we track how much memory is usable
4849 * by the kernel using kernelcore_remaining. When it is
4850 * 0, the rest of the node is usable by ZONE_MOVABLE
4852 kernelcore_remaining = kernelcore_node;
4854 /* Go through each range of PFNs within this node */
4855 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4856 unsigned long size_pages;
4858 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4859 if (start_pfn >= end_pfn)
4862 /* Account for what is only usable for kernelcore */
4863 if (start_pfn < usable_startpfn) {
4864 unsigned long kernel_pages;
4865 kernel_pages = min(end_pfn, usable_startpfn)
4868 kernelcore_remaining -= min(kernel_pages,
4869 kernelcore_remaining);
4870 required_kernelcore -= min(kernel_pages,
4871 required_kernelcore);
4873 /* Continue if range is now fully accounted */
4874 if (end_pfn <= usable_startpfn) {
4877 * Push zone_movable_pfn to the end so
4878 * that if we have to rebalance
4879 * kernelcore across nodes, we will
4880 * not double account here
4882 zone_movable_pfn[nid] = end_pfn;
4885 start_pfn = usable_startpfn;
4889 * The usable PFN range for ZONE_MOVABLE is from
4890 * start_pfn->end_pfn. Calculate size_pages as the
4891 * number of pages used as kernelcore
4893 size_pages = end_pfn - start_pfn;
4894 if (size_pages > kernelcore_remaining)
4895 size_pages = kernelcore_remaining;
4896 zone_movable_pfn[nid] = start_pfn + size_pages;
4899 * Some kernelcore has been met, update counts and
4900 * break if the kernelcore for this node has been
4903 required_kernelcore -= min(required_kernelcore,
4905 kernelcore_remaining -= size_pages;
4906 if (!kernelcore_remaining)
4912 * If there is still required_kernelcore, we do another pass with one
4913 * less node in the count. This will push zone_movable_pfn[nid] further
4914 * along on the nodes that still have memory until kernelcore is
4918 if (usable_nodes && required_kernelcore > usable_nodes)
4921 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4922 for (nid = 0; nid < MAX_NUMNODES; nid++)
4923 zone_movable_pfn[nid] =
4924 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4927 /* restore the node_state */
4928 node_states[N_HIGH_MEMORY] = saved_node_state;
4931 /* Any regular memory on that node ? */
4932 static void __init check_for_regular_memory(pg_data_t *pgdat)
4934 #ifdef CONFIG_HIGHMEM
4935 enum zone_type zone_type;
4937 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4938 struct zone *zone = &pgdat->node_zones[zone_type];
4939 if (zone->present_pages) {
4940 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4948 * free_area_init_nodes - Initialise all pg_data_t and zone data
4949 * @max_zone_pfn: an array of max PFNs for each zone
4951 * This will call free_area_init_node() for each active node in the system.
4952 * Using the page ranges provided by add_active_range(), the size of each
4953 * zone in each node and their holes is calculated. If the maximum PFN
4954 * between two adjacent zones match, it is assumed that the zone is empty.
4955 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4956 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4957 * starts where the previous one ended. For example, ZONE_DMA32 starts
4958 * at arch_max_dma_pfn.
4960 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4962 unsigned long start_pfn, end_pfn;
4965 /* Record where the zone boundaries are */
4966 memset(arch_zone_lowest_possible_pfn, 0,
4967 sizeof(arch_zone_lowest_possible_pfn));
4968 memset(arch_zone_highest_possible_pfn, 0,
4969 sizeof(arch_zone_highest_possible_pfn));
4970 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4971 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4972 for (i = 1; i < MAX_NR_ZONES; i++) {
4973 if (i == ZONE_MOVABLE)
4975 arch_zone_lowest_possible_pfn[i] =
4976 arch_zone_highest_possible_pfn[i-1];
4977 arch_zone_highest_possible_pfn[i] =
4978 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4980 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4981 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4983 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4984 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4985 find_zone_movable_pfns_for_nodes();
4987 /* Print out the zone ranges */
4988 printk("Zone ranges:\n");
4989 for (i = 0; i < MAX_NR_ZONES; i++) {
4990 if (i == ZONE_MOVABLE)
4992 printk(KERN_CONT " %-8s ", zone_names[i]);
4993 if (arch_zone_lowest_possible_pfn[i] ==
4994 arch_zone_highest_possible_pfn[i])
4995 printk(KERN_CONT "empty\n");
4997 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4998 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4999 (arch_zone_highest_possible_pfn[i]
5000 << PAGE_SHIFT) - 1);
5003 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5004 printk("Movable zone start for each node\n");
5005 for (i = 0; i < MAX_NUMNODES; i++) {
5006 if (zone_movable_pfn[i])
5007 printk(" Node %d: %#010lx\n", i,
5008 zone_movable_pfn[i] << PAGE_SHIFT);
5011 /* Print out the early node map */
5012 printk("Early memory node ranges\n");
5013 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5014 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5015 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5017 /* Initialise every node */
5018 mminit_verify_pageflags_layout();
5019 setup_nr_node_ids();
5020 for_each_online_node(nid) {
5021 pg_data_t *pgdat = NODE_DATA(nid);
5022 free_area_init_node(nid, NULL,
5023 find_min_pfn_for_node(nid), NULL);
5025 /* Any memory on that node */
5026 if (pgdat->node_present_pages)
5027 node_set_state(nid, N_HIGH_MEMORY);
5028 check_for_regular_memory(pgdat);
5032 static int __init cmdline_parse_core(char *p, unsigned long *core)
5034 unsigned long long coremem;
5038 coremem = memparse(p, &p);
5039 *core = coremem >> PAGE_SHIFT;
5041 /* Paranoid check that UL is enough for the coremem value */
5042 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5048 * kernelcore=size sets the amount of memory for use for allocations that
5049 * cannot be reclaimed or migrated.
5051 static int __init cmdline_parse_kernelcore(char *p)
5053 return cmdline_parse_core(p, &required_kernelcore);
5057 * movablecore=size sets the amount of memory for use for allocations that
5058 * can be reclaimed or migrated.
5060 static int __init cmdline_parse_movablecore(char *p)
5062 return cmdline_parse_core(p, &required_movablecore);
5065 early_param("kernelcore", cmdline_parse_kernelcore);
5066 early_param("movablecore", cmdline_parse_movablecore);
5068 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5071 * set_dma_reserve - set the specified number of pages reserved in the first zone
5072 * @new_dma_reserve: The number of pages to mark reserved
5074 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5075 * In the DMA zone, a significant percentage may be consumed by kernel image
5076 * and other unfreeable allocations which can skew the watermarks badly. This
5077 * function may optionally be used to account for unfreeable pages in the
5078 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5079 * smaller per-cpu batchsize.
5081 void __init set_dma_reserve(unsigned long new_dma_reserve)
5083 dma_reserve = new_dma_reserve;
5086 void __init free_area_init(unsigned long *zones_size)
5088 free_area_init_node(0, zones_size,
5089 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5092 static int page_alloc_cpu_notify(struct notifier_block *self,
5093 unsigned long action, void *hcpu)
5095 int cpu = (unsigned long)hcpu;
5097 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5098 lru_add_drain_cpu(cpu);
5102 * Spill the event counters of the dead processor
5103 * into the current processors event counters.
5104 * This artificially elevates the count of the current
5107 vm_events_fold_cpu(cpu);
5110 * Zero the differential counters of the dead processor
5111 * so that the vm statistics are consistent.
5113 * This is only okay since the processor is dead and cannot
5114 * race with what we are doing.
5116 refresh_cpu_vm_stats(cpu);
5121 void __init page_alloc_init(void)
5123 hotcpu_notifier(page_alloc_cpu_notify, 0);
5127 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5128 * or min_free_kbytes changes.
5130 static void calculate_totalreserve_pages(void)
5132 struct pglist_data *pgdat;
5133 unsigned long reserve_pages = 0;
5134 enum zone_type i, j;
5136 for_each_online_pgdat(pgdat) {
5137 for (i = 0; i < MAX_NR_ZONES; i++) {
5138 struct zone *zone = pgdat->node_zones + i;
5139 unsigned long max = 0;
5141 /* Find valid and maximum lowmem_reserve in the zone */
5142 for (j = i; j < MAX_NR_ZONES; j++) {
5143 if (zone->lowmem_reserve[j] > max)
5144 max = zone->lowmem_reserve[j];
5147 /* we treat the high watermark as reserved pages. */
5148 max += high_wmark_pages(zone);
5150 if (max > zone->present_pages)
5151 max = zone->present_pages;
5152 reserve_pages += max;
5154 * Lowmem reserves are not available to
5155 * GFP_HIGHUSER page cache allocations and
5156 * kswapd tries to balance zones to their high
5157 * watermark. As a result, neither should be
5158 * regarded as dirtyable memory, to prevent a
5159 * situation where reclaim has to clean pages
5160 * in order to balance the zones.
5162 zone->dirty_balance_reserve = max;
5165 dirty_balance_reserve = reserve_pages;
5166 totalreserve_pages = reserve_pages;
5170 * setup_per_zone_lowmem_reserve - called whenever
5171 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5172 * has a correct pages reserved value, so an adequate number of
5173 * pages are left in the zone after a successful __alloc_pages().
5175 static void setup_per_zone_lowmem_reserve(void)
5177 struct pglist_data *pgdat;
5178 enum zone_type j, idx;
5180 for_each_online_pgdat(pgdat) {
5181 for (j = 0; j < MAX_NR_ZONES; j++) {
5182 struct zone *zone = pgdat->node_zones + j;
5183 unsigned long present_pages = zone->present_pages;
5185 zone->lowmem_reserve[j] = 0;
5189 struct zone *lower_zone;
5193 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5194 sysctl_lowmem_reserve_ratio[idx] = 1;
5196 lower_zone = pgdat->node_zones + idx;
5197 lower_zone->lowmem_reserve[j] = present_pages /
5198 sysctl_lowmem_reserve_ratio[idx];
5199 present_pages += lower_zone->present_pages;
5204 /* update totalreserve_pages */
5205 calculate_totalreserve_pages();
5208 static void __setup_per_zone_wmarks(void)
5210 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5211 unsigned long lowmem_pages = 0;
5213 unsigned long flags;
5215 /* Calculate total number of !ZONE_HIGHMEM pages */
5216 for_each_zone(zone) {
5217 if (!is_highmem(zone))
5218 lowmem_pages += zone->present_pages;
5221 for_each_zone(zone) {
5224 spin_lock_irqsave(&zone->lock, flags);
5225 tmp = (u64)pages_min * zone->present_pages;
5226 do_div(tmp, lowmem_pages);
5227 if (is_highmem(zone)) {
5229 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5230 * need highmem pages, so cap pages_min to a small
5233 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5234 * deltas controls asynch page reclaim, and so should
5235 * not be capped for highmem.
5239 min_pages = zone->present_pages / 1024;
5240 if (min_pages < SWAP_CLUSTER_MAX)
5241 min_pages = SWAP_CLUSTER_MAX;
5242 if (min_pages > 128)
5244 zone->watermark[WMARK_MIN] = min_pages;
5247 * If it's a lowmem zone, reserve a number of pages
5248 * proportionate to the zone's size.
5250 zone->watermark[WMARK_MIN] = tmp;
5253 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5254 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5256 zone->watermark[WMARK_MIN] += cma_wmark_pages(zone);
5257 zone->watermark[WMARK_LOW] += cma_wmark_pages(zone);
5258 zone->watermark[WMARK_HIGH] += cma_wmark_pages(zone);
5260 setup_zone_migrate_reserve(zone);
5261 spin_unlock_irqrestore(&zone->lock, flags);
5264 /* update totalreserve_pages */
5265 calculate_totalreserve_pages();
5269 * setup_per_zone_wmarks - called when min_free_kbytes changes
5270 * or when memory is hot-{added|removed}
5272 * Ensures that the watermark[min,low,high] values for each zone are set
5273 * correctly with respect to min_free_kbytes.
5275 void setup_per_zone_wmarks(void)
5277 mutex_lock(&zonelists_mutex);
5278 __setup_per_zone_wmarks();
5279 mutex_unlock(&zonelists_mutex);
5283 * The inactive anon list should be small enough that the VM never has to
5284 * do too much work, but large enough that each inactive page has a chance
5285 * to be referenced again before it is swapped out.
5287 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5288 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5289 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5290 * the anonymous pages are kept on the inactive list.
5293 * memory ratio inactive anon
5294 * -------------------------------------
5303 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5305 unsigned int gb, ratio;
5307 /* Zone size in gigabytes */
5308 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5310 ratio = int_sqrt(10 * gb);
5314 zone->inactive_ratio = ratio;
5317 static void __meminit setup_per_zone_inactive_ratio(void)
5322 calculate_zone_inactive_ratio(zone);
5326 * Initialise min_free_kbytes.
5328 * For small machines we want it small (128k min). For large machines
5329 * we want it large (64MB max). But it is not linear, because network
5330 * bandwidth does not increase linearly with machine size. We use
5332 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5333 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5349 int __meminit init_per_zone_wmark_min(void)
5351 unsigned long lowmem_kbytes;
5353 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5355 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5356 if (min_free_kbytes < 128)
5357 min_free_kbytes = 128;
5358 if (min_free_kbytes > 65536)
5359 min_free_kbytes = 65536;
5360 setup_per_zone_wmarks();
5361 refresh_zone_stat_thresholds();
5362 setup_per_zone_lowmem_reserve();
5363 setup_per_zone_inactive_ratio();
5366 module_init(init_per_zone_wmark_min)
5369 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5370 * that we can call two helper functions whenever min_free_kbytes
5373 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5374 void __user *buffer, size_t *length, loff_t *ppos)
5376 proc_dointvec(table, write, buffer, length, ppos);
5378 setup_per_zone_wmarks();
5383 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5384 void __user *buffer, size_t *length, loff_t *ppos)
5389 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5394 zone->min_unmapped_pages = (zone->present_pages *
5395 sysctl_min_unmapped_ratio) / 100;
5399 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5400 void __user *buffer, size_t *length, loff_t *ppos)
5405 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5410 zone->min_slab_pages = (zone->present_pages *
5411 sysctl_min_slab_ratio) / 100;
5417 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5418 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5419 * whenever sysctl_lowmem_reserve_ratio changes.
5421 * The reserve ratio obviously has absolutely no relation with the
5422 * minimum watermarks. The lowmem reserve ratio can only make sense
5423 * if in function of the boot time zone sizes.
5425 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5426 void __user *buffer, size_t *length, loff_t *ppos)
5428 proc_dointvec_minmax(table, write, buffer, length, ppos);
5429 setup_per_zone_lowmem_reserve();
5434 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5435 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5436 * can have before it gets flushed back to buddy allocator.
5439 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5440 void __user *buffer, size_t *length, loff_t *ppos)
5446 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5447 if (!write || (ret < 0))
5449 for_each_populated_zone(zone) {
5450 for_each_possible_cpu(cpu) {
5452 high = zone->present_pages / percpu_pagelist_fraction;
5453 setup_pagelist_highmark(
5454 per_cpu_ptr(zone->pageset, cpu), high);
5460 int hashdist = HASHDIST_DEFAULT;
5463 static int __init set_hashdist(char *str)
5467 hashdist = simple_strtoul(str, &str, 0);
5470 __setup("hashdist=", set_hashdist);
5474 * allocate a large system hash table from bootmem
5475 * - it is assumed that the hash table must contain an exact power-of-2
5476 * quantity of entries
5477 * - limit is the number of hash buckets, not the total allocation size
5479 void *__init alloc_large_system_hash(const char *tablename,
5480 unsigned long bucketsize,
5481 unsigned long numentries,
5484 unsigned int *_hash_shift,
5485 unsigned int *_hash_mask,
5486 unsigned long low_limit,
5487 unsigned long high_limit)
5489 unsigned long long max = high_limit;
5490 unsigned long log2qty, size;
5493 /* allow the kernel cmdline to have a say */
5495 /* round applicable memory size up to nearest megabyte */
5496 numentries = nr_kernel_pages;
5497 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5498 numentries >>= 20 - PAGE_SHIFT;
5499 numentries <<= 20 - PAGE_SHIFT;
5501 /* limit to 1 bucket per 2^scale bytes of low memory */
5502 if (scale > PAGE_SHIFT)
5503 numentries >>= (scale - PAGE_SHIFT);
5505 numentries <<= (PAGE_SHIFT - scale);
5507 /* Make sure we've got at least a 0-order allocation.. */
5508 if (unlikely(flags & HASH_SMALL)) {
5509 /* Makes no sense without HASH_EARLY */
5510 WARN_ON(!(flags & HASH_EARLY));
5511 if (!(numentries >> *_hash_shift)) {
5512 numentries = 1UL << *_hash_shift;
5513 BUG_ON(!numentries);
5515 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5516 numentries = PAGE_SIZE / bucketsize;
5518 numentries = roundup_pow_of_two(numentries);
5520 /* limit allocation size to 1/16 total memory by default */
5522 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5523 do_div(max, bucketsize);
5525 max = min(max, 0x80000000ULL);
5527 if (numentries < low_limit)
5528 numentries = low_limit;
5529 if (numentries > max)
5532 log2qty = ilog2(numentries);
5535 size = bucketsize << log2qty;
5536 if (flags & HASH_EARLY)
5537 table = alloc_bootmem_nopanic(size);
5539 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5542 * If bucketsize is not a power-of-two, we may free
5543 * some pages at the end of hash table which
5544 * alloc_pages_exact() automatically does
5546 if (get_order(size) < MAX_ORDER) {
5547 table = alloc_pages_exact(size, GFP_ATOMIC);
5548 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5551 } while (!table && size > PAGE_SIZE && --log2qty);
5554 panic("Failed to allocate %s hash table\n", tablename);
5556 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5559 ilog2(size) - PAGE_SHIFT,
5563 *_hash_shift = log2qty;
5565 *_hash_mask = (1 << log2qty) - 1;
5570 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5571 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5574 #ifdef CONFIG_SPARSEMEM
5575 return __pfn_to_section(pfn)->pageblock_flags;
5577 return zone->pageblock_flags;
5578 #endif /* CONFIG_SPARSEMEM */
5581 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5583 #ifdef CONFIG_SPARSEMEM
5584 pfn &= (PAGES_PER_SECTION-1);
5585 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5587 pfn = pfn - zone->zone_start_pfn;
5588 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5589 #endif /* CONFIG_SPARSEMEM */
5593 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5594 * @page: The page within the block of interest
5595 * @start_bitidx: The first bit of interest to retrieve
5596 * @end_bitidx: The last bit of interest
5597 * returns pageblock_bits flags
5599 unsigned long get_pageblock_flags_group(struct page *page,
5600 int start_bitidx, int end_bitidx)
5603 unsigned long *bitmap;
5604 unsigned long pfn, bitidx;
5605 unsigned long flags = 0;
5606 unsigned long value = 1;
5608 zone = page_zone(page);
5609 pfn = page_to_pfn(page);
5610 bitmap = get_pageblock_bitmap(zone, pfn);
5611 bitidx = pfn_to_bitidx(zone, pfn);
5613 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5614 if (test_bit(bitidx + start_bitidx, bitmap))
5621 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5622 * @page: The page within the block of interest
5623 * @start_bitidx: The first bit of interest
5624 * @end_bitidx: The last bit of interest
5625 * @flags: The flags to set
5627 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5628 int start_bitidx, int end_bitidx)
5631 unsigned long *bitmap;
5632 unsigned long pfn, bitidx;
5633 unsigned long value = 1;
5635 zone = page_zone(page);
5636 pfn = page_to_pfn(page);
5637 bitmap = get_pageblock_bitmap(zone, pfn);
5638 bitidx = pfn_to_bitidx(zone, pfn);
5639 VM_BUG_ON(pfn < zone->zone_start_pfn);
5640 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5642 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5644 __set_bit(bitidx + start_bitidx, bitmap);
5646 __clear_bit(bitidx + start_bitidx, bitmap);
5650 * This function checks whether pageblock includes unmovable pages or not.
5651 * If @count is not zero, it is okay to include less @count unmovable pages
5653 * PageLRU check wihtout isolation or lru_lock could race so that
5654 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5655 * expect this function should be exact.
5657 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5658 bool skip_hwpoisoned_pages)
5660 unsigned long pfn, iter, found;
5664 * For avoiding noise data, lru_add_drain_all() should be called
5665 * If ZONE_MOVABLE, the zone never contains unmovable pages
5667 if (zone_idx(zone) == ZONE_MOVABLE)
5669 mt = get_pageblock_migratetype(page);
5670 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5673 pfn = page_to_pfn(page);
5674 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5675 unsigned long check = pfn + iter;
5677 if (!pfn_valid_within(check))
5680 page = pfn_to_page(check);
5682 * We can't use page_count without pin a page
5683 * because another CPU can free compound page.
5684 * This check already skips compound tails of THP
5685 * because their page->_count is zero at all time.
5687 if (!atomic_read(&page->_count)) {
5688 if (PageBuddy(page))
5689 iter += (1 << page_order(page)) - 1;
5694 * The HWPoisoned page may be not in buddy system, and
5695 * page_count() is not 0.
5697 if (skip_hwpoisoned_pages && PageHWPoison(page))
5703 * If there are RECLAIMABLE pages, we need to check it.
5704 * But now, memory offline itself doesn't call shrink_slab()
5705 * and it still to be fixed.
5708 * If the page is not RAM, page_count()should be 0.
5709 * we don't need more check. This is an _used_ not-movable page.
5711 * The problematic thing here is PG_reserved pages. PG_reserved
5712 * is set to both of a memory hole page and a _used_ kernel
5721 bool is_pageblock_removable_nolock(struct page *page)
5727 * We have to be careful here because we are iterating over memory
5728 * sections which are not zone aware so we might end up outside of
5729 * the zone but still within the section.
5730 * We have to take care about the node as well. If the node is offline
5731 * its NODE_DATA will be NULL - see page_zone.
5733 if (!node_online(page_to_nid(page)))
5736 zone = page_zone(page);
5737 pfn = page_to_pfn(page);
5738 if (zone->zone_start_pfn > pfn ||
5739 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5742 return !has_unmovable_pages(zone, page, 0, true);
5747 static unsigned long pfn_max_align_down(unsigned long pfn)
5749 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5750 pageblock_nr_pages) - 1);
5753 static unsigned long pfn_max_align_up(unsigned long pfn)
5755 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5756 pageblock_nr_pages));
5759 /* [start, end) must belong to a single zone. */
5760 static int __alloc_contig_migrate_range(struct compact_control *cc,
5761 unsigned long start, unsigned long end)
5763 /* This function is based on compact_zone() from compaction.c. */
5764 unsigned long nr_reclaimed;
5765 unsigned long pfn = start;
5766 unsigned int tries = 0;
5769 migrate_prep_local();
5771 while (pfn < end || !list_empty(&cc->migratepages)) {
5772 if (fatal_signal_pending(current)) {
5777 if (list_empty(&cc->migratepages)) {
5778 cc->nr_migratepages = 0;
5779 pfn = isolate_migratepages_range(cc->zone, cc,
5786 } else if (++tries == 5) {
5787 ret = ret < 0 ? ret : -EBUSY;
5791 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5793 cc->nr_migratepages -= nr_reclaimed;
5795 ret = migrate_pages(&cc->migratepages,
5796 alloc_migrate_target,
5797 0, false, MIGRATE_SYNC);
5800 putback_movable_pages(&cc->migratepages);
5801 return ret > 0 ? 0 : ret;
5805 * Update zone's cma pages counter used for watermark level calculation.
5807 static inline void __update_cma_watermarks(struct zone *zone, int count)
5809 unsigned long flags;
5810 spin_lock_irqsave(&zone->lock, flags);
5811 zone->min_cma_pages += count;
5812 spin_unlock_irqrestore(&zone->lock, flags);
5813 setup_per_zone_wmarks();
5817 * Trigger memory pressure bump to reclaim some pages in order to be able to
5818 * allocate 'count' pages in single page units. Does similar work as
5819 *__alloc_pages_slowpath() function.
5821 static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
5823 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
5824 struct zonelist *zonelist = node_zonelist(0, gfp_mask);
5825 int did_some_progress = 0;
5829 * Increase level of watermarks to force kswapd do his job
5830 * to stabilise at new watermark level.
5832 __update_cma_watermarks(zone, count);
5834 /* Obey watermarks as if the page was being allocated */
5835 while (!zone_watermark_ok(zone, 0, low_wmark_pages(zone), 0, 0)) {
5836 wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));
5838 did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
5840 if (!did_some_progress) {
5841 /* Exhausted what can be done so it's blamo time */
5842 out_of_memory(zonelist, gfp_mask, order, NULL, false);
5846 /* Restore original watermark levels. */
5847 __update_cma_watermarks(zone, -count);
5853 * alloc_contig_range() -- tries to allocate given range of pages
5854 * @start: start PFN to allocate
5855 * @end: one-past-the-last PFN to allocate
5856 * @migratetype: migratetype of the underlaying pageblocks (either
5857 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5858 * in range must have the same migratetype and it must
5859 * be either of the two.
5861 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5862 * aligned, however it's the caller's responsibility to guarantee that
5863 * we are the only thread that changes migrate type of pageblocks the
5866 * The PFN range must belong to a single zone.
5868 * Returns zero on success or negative error code. On success all
5869 * pages which PFN is in [start, end) are allocated for the caller and
5870 * need to be freed with free_contig_range().
5872 int alloc_contig_range(unsigned long start, unsigned long end,
5873 unsigned migratetype)
5875 struct zone *zone = page_zone(pfn_to_page(start));
5876 unsigned long outer_start, outer_end;
5879 struct compact_control cc = {
5880 .nr_migratepages = 0,
5882 .zone = page_zone(pfn_to_page(start)),
5884 .ignore_skip_hint = true,
5886 INIT_LIST_HEAD(&cc.migratepages);
5889 * What we do here is we mark all pageblocks in range as
5890 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5891 * have different sizes, and due to the way page allocator
5892 * work, we align the range to biggest of the two pages so
5893 * that page allocator won't try to merge buddies from
5894 * different pageblocks and change MIGRATE_ISOLATE to some
5895 * other migration type.
5897 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5898 * migrate the pages from an unaligned range (ie. pages that
5899 * we are interested in). This will put all the pages in
5900 * range back to page allocator as MIGRATE_ISOLATE.
5902 * When this is done, we take the pages in range from page
5903 * allocator removing them from the buddy system. This way
5904 * page allocator will never consider using them.
5906 * This lets us mark the pageblocks back as
5907 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5908 * aligned range but not in the unaligned, original range are
5909 * put back to page allocator so that buddy can use them.
5912 ret = start_isolate_page_range(pfn_max_align_down(start),
5913 pfn_max_align_up(end), migratetype,
5918 ret = __alloc_contig_migrate_range(&cc, start, end);
5923 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5924 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5925 * more, all pages in [start, end) are free in page allocator.
5926 * What we are going to do is to allocate all pages from
5927 * [start, end) (that is remove them from page allocator).
5929 * The only problem is that pages at the beginning and at the
5930 * end of interesting range may be not aligned with pages that
5931 * page allocator holds, ie. they can be part of higher order
5932 * pages. Because of this, we reserve the bigger range and
5933 * once this is done free the pages we are not interested in.
5935 * We don't have to hold zone->lock here because the pages are
5936 * isolated thus they won't get removed from buddy.
5939 lru_add_drain_all();
5943 outer_start = start;
5944 while (!PageBuddy(pfn_to_page(outer_start))) {
5945 if (++order >= MAX_ORDER) {
5949 outer_start &= ~0UL << order;
5952 /* Make sure the range is really isolated. */
5953 if (test_pages_isolated(outer_start, end, false)) {
5954 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5961 * Reclaim enough pages to make sure that contiguous allocation
5962 * will not starve the system.
5964 __reclaim_pages(zone, GFP_HIGHUSER_MOVABLE, end-start);
5966 /* Grab isolated pages from freelists. */
5967 outer_end = isolate_freepages_range(&cc, outer_start, end);
5973 /* Free head and tail (if any) */
5974 if (start != outer_start)
5975 free_contig_range(outer_start, start - outer_start);
5976 if (end != outer_end)
5977 free_contig_range(end, outer_end - end);
5980 undo_isolate_page_range(pfn_max_align_down(start),
5981 pfn_max_align_up(end), migratetype);
5985 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5987 for (; nr_pages--; ++pfn)
5988 __free_page(pfn_to_page(pfn));
5992 #ifdef CONFIG_MEMORY_HOTPLUG
5993 static int __meminit __zone_pcp_update(void *data)
5995 struct zone *zone = data;
5997 unsigned long batch = zone_batchsize(zone), flags;
5999 for_each_possible_cpu(cpu) {
6000 struct per_cpu_pageset *pset;
6001 struct per_cpu_pages *pcp;
6003 pset = per_cpu_ptr(zone->pageset, cpu);
6006 local_irq_save(flags);
6008 free_pcppages_bulk(zone, pcp->count, pcp);
6009 drain_zonestat(zone, pset);
6010 setup_pageset(pset, batch);
6011 local_irq_restore(flags);
6016 void __meminit zone_pcp_update(struct zone *zone)
6018 stop_machine(__zone_pcp_update, zone, NULL);
6022 void zone_pcp_reset(struct zone *zone)
6024 unsigned long flags;
6026 struct per_cpu_pageset *pset;
6028 /* avoid races with drain_pages() */
6029 local_irq_save(flags);
6030 if (zone->pageset != &boot_pageset) {
6031 for_each_online_cpu(cpu) {
6032 pset = per_cpu_ptr(zone->pageset, cpu);
6033 drain_zonestat(zone, pset);
6035 free_percpu(zone->pageset);
6036 zone->pageset = &boot_pageset;
6038 local_irq_restore(flags);
6041 #ifdef CONFIG_MEMORY_HOTREMOVE
6043 * All pages in the range must be isolated before calling this.
6046 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6052 unsigned long flags;
6053 /* find the first valid pfn */
6054 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6059 zone = page_zone(pfn_to_page(pfn));
6060 spin_lock_irqsave(&zone->lock, flags);
6062 while (pfn < end_pfn) {
6063 if (!pfn_valid(pfn)) {
6067 page = pfn_to_page(pfn);
6069 * The HWPoisoned page may be not in buddy system, and
6070 * page_count() is not 0.
6072 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6074 SetPageReserved(page);
6078 BUG_ON(page_count(page));
6079 BUG_ON(!PageBuddy(page));
6080 order = page_order(page);
6081 #ifdef CONFIG_DEBUG_VM
6082 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6083 pfn, 1 << order, end_pfn);
6085 list_del(&page->lru);
6086 rmv_page_order(page);
6087 zone->free_area[order].nr_free--;
6088 for (i = 0; i < (1 << order); i++)
6089 SetPageReserved((page+i));
6090 pfn += (1 << order);
6092 spin_unlock_irqrestore(&zone->lock, flags);
6096 #ifdef CONFIG_MEMORY_FAILURE
6097 bool is_free_buddy_page(struct page *page)
6099 struct zone *zone = page_zone(page);
6100 unsigned long pfn = page_to_pfn(page);
6101 unsigned long flags;
6104 spin_lock_irqsave(&zone->lock, flags);
6105 for (order = 0; order < MAX_ORDER; order++) {
6106 struct page *page_head = page - (pfn & ((1 << order) - 1));
6108 if (PageBuddy(page_head) && page_order(page_head) >= order)
6111 spin_unlock_irqrestore(&zone->lock, flags);
6113 return order < MAX_ORDER;
6117 static const struct trace_print_flags pageflag_names[] = {
6118 {1UL << PG_locked, "locked" },
6119 {1UL << PG_error, "error" },
6120 {1UL << PG_referenced, "referenced" },
6121 {1UL << PG_uptodate, "uptodate" },
6122 {1UL << PG_dirty, "dirty" },
6123 {1UL << PG_lru, "lru" },
6124 {1UL << PG_active, "active" },
6125 {1UL << PG_slab, "slab" },
6126 {1UL << PG_owner_priv_1, "owner_priv_1" },
6127 {1UL << PG_arch_1, "arch_1" },
6128 {1UL << PG_reserved, "reserved" },
6129 {1UL << PG_private, "private" },
6130 {1UL << PG_private_2, "private_2" },
6131 {1UL << PG_writeback, "writeback" },
6132 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6133 {1UL << PG_head, "head" },
6134 {1UL << PG_tail, "tail" },
6136 {1UL << PG_compound, "compound" },
6138 {1UL << PG_swapcache, "swapcache" },
6139 {1UL << PG_mappedtodisk, "mappedtodisk" },
6140 {1UL << PG_reclaim, "reclaim" },
6141 {1UL << PG_swapbacked, "swapbacked" },
6142 {1UL << PG_unevictable, "unevictable" },
6144 {1UL << PG_mlocked, "mlocked" },
6146 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6147 {1UL << PG_uncached, "uncached" },
6149 #ifdef CONFIG_MEMORY_FAILURE
6150 {1UL << PG_hwpoison, "hwpoison" },
6152 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6153 {1UL << PG_compound_lock, "compound_lock" },
6157 static void dump_page_flags(unsigned long flags)
6159 const char *delim = "";
6163 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6165 printk(KERN_ALERT "page flags: %#lx(", flags);
6167 /* remove zone id */
6168 flags &= (1UL << NR_PAGEFLAGS) - 1;
6170 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6172 mask = pageflag_names[i].mask;
6173 if ((flags & mask) != mask)
6177 printk("%s%s", delim, pageflag_names[i].name);
6181 /* check for left over flags */
6183 printk("%s%#lx", delim, flags);
6188 void dump_page(struct page *page)
6191 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6192 page, atomic_read(&page->_count), page_mapcount(page),
6193 page->mapping, page->index);
6194 dump_page_flags(page->flags);
6195 mem_cgroup_print_bad_page(page);
6198 /* reset zone->present_pages */
6199 void reset_zone_present_pages(void)
6204 for_each_node_state(nid, N_HIGH_MEMORY) {
6205 for (i = 0; i < MAX_NR_ZONES; i++) {
6206 z = NODE_DATA(nid)->node_zones + i;
6208 z->present_pages = 0;
6213 /* calculate zone's present pages in buddy system */
6214 void fixup_zone_present_pages(int nid, unsigned long start_pfn,
6215 unsigned long end_pfn)
6218 unsigned long zone_start_pfn, zone_end_pfn;
6221 for (i = 0; i < MAX_NR_ZONES; i++) {
6222 z = NODE_DATA(nid)->node_zones + i;
6226 zone_start_pfn = z->zone_start_pfn;
6227 zone_end_pfn = zone_start_pfn + z->spanned_pages;
6228 if (!(zone_start_pfn >= end_pfn || zone_end_pfn <= start_pfn))
6229 z->present_pages += min(end_pfn, zone_end_pfn) -
6230 max(start_pfn, zone_start_pfn);