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 #ifdef CONFIG_MOVABLE_NODE
94 [N_MEMORY] = { { [0] = 1UL } },
96 [N_CPU] = { { [0] = 1UL } },
99 EXPORT_SYMBOL(node_states);
101 unsigned long totalram_pages __read_mostly;
102 unsigned long totalreserve_pages __read_mostly;
104 * When calculating the number of globally allowed dirty pages, there
105 * is a certain number of per-zone reserves that should not be
106 * considered dirtyable memory. This is the sum of those reserves
107 * over all existing zones that contribute dirtyable memory.
109 unsigned long dirty_balance_reserve __read_mostly;
111 int percpu_pagelist_fraction;
112 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
114 #ifdef CONFIG_PM_SLEEP
116 * The following functions are used by the suspend/hibernate code to temporarily
117 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
118 * while devices are suspended. To avoid races with the suspend/hibernate code,
119 * they should always be called with pm_mutex held (gfp_allowed_mask also should
120 * only be modified with pm_mutex held, unless the suspend/hibernate code is
121 * guaranteed not to run in parallel with that modification).
124 static gfp_t saved_gfp_mask;
126 void pm_restore_gfp_mask(void)
128 WARN_ON(!mutex_is_locked(&pm_mutex));
129 if (saved_gfp_mask) {
130 gfp_allowed_mask = saved_gfp_mask;
135 void pm_restrict_gfp_mask(void)
137 WARN_ON(!mutex_is_locked(&pm_mutex));
138 WARN_ON(saved_gfp_mask);
139 saved_gfp_mask = gfp_allowed_mask;
140 gfp_allowed_mask &= ~GFP_IOFS;
143 bool pm_suspended_storage(void)
145 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
149 #endif /* CONFIG_PM_SLEEP */
151 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
152 int pageblock_order __read_mostly;
155 static void __free_pages_ok(struct page *page, unsigned int order);
158 * results with 256, 32 in the lowmem_reserve sysctl:
159 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
160 * 1G machine -> (16M dma, 784M normal, 224M high)
161 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
162 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
163 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
165 * TBD: should special case ZONE_DMA32 machines here - in those we normally
166 * don't need any ZONE_NORMAL reservation
168 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
169 #ifdef CONFIG_ZONE_DMA
172 #ifdef CONFIG_ZONE_DMA32
175 #ifdef CONFIG_HIGHMEM
181 EXPORT_SYMBOL(totalram_pages);
183 static char * const zone_names[MAX_NR_ZONES] = {
184 #ifdef CONFIG_ZONE_DMA
187 #ifdef CONFIG_ZONE_DMA32
191 #ifdef CONFIG_HIGHMEM
197 int min_free_kbytes = 1024;
199 static unsigned long __meminitdata nr_kernel_pages;
200 static unsigned long __meminitdata nr_all_pages;
201 static unsigned long __meminitdata dma_reserve;
203 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
204 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
205 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
206 static unsigned long __initdata required_kernelcore;
207 static unsigned long __initdata required_movablecore;
208 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
210 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
212 EXPORT_SYMBOL(movable_zone);
213 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
216 int nr_node_ids __read_mostly = MAX_NUMNODES;
217 int nr_online_nodes __read_mostly = 1;
218 EXPORT_SYMBOL(nr_node_ids);
219 EXPORT_SYMBOL(nr_online_nodes);
222 int page_group_by_mobility_disabled __read_mostly;
226 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
227 * Instead, use {un}set_pageblock_isolate.
229 void set_pageblock_migratetype(struct page *page, int migratetype)
232 if (unlikely(page_group_by_mobility_disabled))
233 migratetype = MIGRATE_UNMOVABLE;
235 set_pageblock_flags_group(page, (unsigned long)migratetype,
236 PB_migrate, PB_migrate_end);
239 bool oom_killer_disabled __read_mostly;
241 #ifdef CONFIG_DEBUG_VM
242 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
246 unsigned long pfn = page_to_pfn(page);
249 seq = zone_span_seqbegin(zone);
250 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
252 else if (pfn < zone->zone_start_pfn)
254 } while (zone_span_seqretry(zone, seq));
259 static int page_is_consistent(struct zone *zone, struct page *page)
261 if (!pfn_valid_within(page_to_pfn(page)))
263 if (zone != page_zone(page))
269 * Temporary debugging check for pages not lying within a given zone.
271 static int bad_range(struct zone *zone, struct page *page)
273 if (page_outside_zone_boundaries(zone, page))
275 if (!page_is_consistent(zone, page))
281 static inline int bad_range(struct zone *zone, struct page *page)
287 static void bad_page(struct page *page)
289 static unsigned long resume;
290 static unsigned long nr_shown;
291 static unsigned long nr_unshown;
293 /* Don't complain about poisoned pages */
294 if (PageHWPoison(page)) {
295 reset_page_mapcount(page); /* remove PageBuddy */
300 * Allow a burst of 60 reports, then keep quiet for that minute;
301 * or allow a steady drip of one report per second.
303 if (nr_shown == 60) {
304 if (time_before(jiffies, resume)) {
310 "BUG: Bad page state: %lu messages suppressed\n",
317 resume = jiffies + 60 * HZ;
319 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
320 current->comm, page_to_pfn(page));
326 /* Leave bad fields for debug, except PageBuddy could make trouble */
327 reset_page_mapcount(page); /* remove PageBuddy */
328 add_taint(TAINT_BAD_PAGE);
332 * Higher-order pages are called "compound pages". They are structured thusly:
334 * The first PAGE_SIZE page is called the "head page".
336 * The remaining PAGE_SIZE pages are called "tail pages".
338 * All pages have PG_compound set. All tail pages have their ->first_page
339 * pointing at the head page.
341 * The first tail page's ->lru.next holds the address of the compound page's
342 * put_page() function. Its ->lru.prev holds the order of allocation.
343 * This usage means that zero-order pages may not be compound.
346 static void free_compound_page(struct page *page)
348 __free_pages_ok(page, compound_order(page));
351 void prep_compound_page(struct page *page, unsigned long order)
354 int nr_pages = 1 << order;
356 set_compound_page_dtor(page, free_compound_page);
357 set_compound_order(page, order);
359 for (i = 1; i < nr_pages; i++) {
360 struct page *p = page + i;
362 set_page_count(p, 0);
363 p->first_page = page;
367 /* update __split_huge_page_refcount if you change this function */
368 static int destroy_compound_page(struct page *page, unsigned long order)
371 int nr_pages = 1 << order;
374 if (unlikely(compound_order(page) != order) ||
375 unlikely(!PageHead(page))) {
380 __ClearPageHead(page);
382 for (i = 1; i < nr_pages; i++) {
383 struct page *p = page + i;
385 if (unlikely(!PageTail(p) || (p->first_page != page))) {
395 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
400 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
401 * and __GFP_HIGHMEM from hard or soft interrupt context.
403 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
404 for (i = 0; i < (1 << order); i++)
405 clear_highpage(page + i);
408 #ifdef CONFIG_DEBUG_PAGEALLOC
409 unsigned int _debug_guardpage_minorder;
411 static int __init debug_guardpage_minorder_setup(char *buf)
415 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
416 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
419 _debug_guardpage_minorder = res;
420 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
423 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
425 static inline void set_page_guard_flag(struct page *page)
427 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
430 static inline void clear_page_guard_flag(struct page *page)
432 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
435 static inline void set_page_guard_flag(struct page *page) { }
436 static inline void clear_page_guard_flag(struct page *page) { }
439 static inline void set_page_order(struct page *page, int order)
441 set_page_private(page, order);
442 __SetPageBuddy(page);
445 static inline void rmv_page_order(struct page *page)
447 __ClearPageBuddy(page);
448 set_page_private(page, 0);
452 * Locate the struct page for both the matching buddy in our
453 * pair (buddy1) and the combined O(n+1) page they form (page).
455 * 1) Any buddy B1 will have an order O twin B2 which satisfies
456 * the following equation:
458 * For example, if the starting buddy (buddy2) is #8 its order
460 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
462 * 2) Any buddy B will have an order O+1 parent P which
463 * satisfies the following equation:
466 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
468 static inline unsigned long
469 __find_buddy_index(unsigned long page_idx, unsigned int order)
471 return page_idx ^ (1 << order);
475 * This function checks whether a page is free && is the buddy
476 * we can do coalesce a page and its buddy if
477 * (a) the buddy is not in a hole &&
478 * (b) the buddy is in the buddy system &&
479 * (c) a page and its buddy have the same order &&
480 * (d) a page and its buddy are in the same zone.
482 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
483 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
485 * For recording page's order, we use page_private(page).
487 static inline int page_is_buddy(struct page *page, struct page *buddy,
490 if (!pfn_valid_within(page_to_pfn(buddy)))
493 if (page_zone_id(page) != page_zone_id(buddy))
496 if (page_is_guard(buddy) && page_order(buddy) == order) {
497 VM_BUG_ON(page_count(buddy) != 0);
501 if (PageBuddy(buddy) && page_order(buddy) == order) {
502 VM_BUG_ON(page_count(buddy) != 0);
509 * Freeing function for a buddy system allocator.
511 * The concept of a buddy system is to maintain direct-mapped table
512 * (containing bit values) for memory blocks of various "orders".
513 * The bottom level table contains the map for the smallest allocatable
514 * units of memory (here, pages), and each level above it describes
515 * pairs of units from the levels below, hence, "buddies".
516 * At a high level, all that happens here is marking the table entry
517 * at the bottom level available, and propagating the changes upward
518 * as necessary, plus some accounting needed to play nicely with other
519 * parts of the VM system.
520 * At each level, we keep a list of pages, which are heads of continuous
521 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
522 * order is recorded in page_private(page) field.
523 * So when we are allocating or freeing one, we can derive the state of the
524 * other. That is, if we allocate a small block, and both were
525 * free, the remainder of the region must be split into blocks.
526 * If a block is freed, and its buddy is also free, then this
527 * triggers coalescing into a block of larger size.
532 static inline void __free_one_page(struct page *page,
533 struct zone *zone, unsigned int order,
536 unsigned long page_idx;
537 unsigned long combined_idx;
538 unsigned long uninitialized_var(buddy_idx);
541 if (unlikely(PageCompound(page)))
542 if (unlikely(destroy_compound_page(page, order)))
545 VM_BUG_ON(migratetype == -1);
547 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
549 VM_BUG_ON(page_idx & ((1 << order) - 1));
550 VM_BUG_ON(bad_range(zone, page));
552 while (order < MAX_ORDER-1) {
553 buddy_idx = __find_buddy_index(page_idx, order);
554 buddy = page + (buddy_idx - page_idx);
555 if (!page_is_buddy(page, buddy, order))
558 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
559 * merge with it and move up one order.
561 if (page_is_guard(buddy)) {
562 clear_page_guard_flag(buddy);
563 set_page_private(page, 0);
564 __mod_zone_freepage_state(zone, 1 << order,
567 list_del(&buddy->lru);
568 zone->free_area[order].nr_free--;
569 rmv_page_order(buddy);
571 combined_idx = buddy_idx & page_idx;
572 page = page + (combined_idx - page_idx);
573 page_idx = combined_idx;
576 set_page_order(page, order);
579 * If this is not the largest possible page, check if the buddy
580 * of the next-highest order is free. If it is, it's possible
581 * that pages are being freed that will coalesce soon. In case,
582 * that is happening, add the free page to the tail of the list
583 * so it's less likely to be used soon and more likely to be merged
584 * as a higher order page
586 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
587 struct page *higher_page, *higher_buddy;
588 combined_idx = buddy_idx & page_idx;
589 higher_page = page + (combined_idx - page_idx);
590 buddy_idx = __find_buddy_index(combined_idx, order + 1);
591 higher_buddy = higher_page + (buddy_idx - combined_idx);
592 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
593 list_add_tail(&page->lru,
594 &zone->free_area[order].free_list[migratetype]);
599 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
601 zone->free_area[order].nr_free++;
604 static inline int free_pages_check(struct page *page)
606 if (unlikely(page_mapcount(page) |
607 (page->mapping != NULL) |
608 (atomic_read(&page->_count) != 0) |
609 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
610 (mem_cgroup_bad_page_check(page)))) {
614 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
615 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
620 * Frees a number of pages from the PCP lists
621 * Assumes all pages on list are in same zone, and of same order.
622 * count is the number of pages to free.
624 * If the zone was previously in an "all pages pinned" state then look to
625 * see if this freeing clears that state.
627 * And clear the zone's pages_scanned counter, to hold off the "all pages are
628 * pinned" detection logic.
630 static void free_pcppages_bulk(struct zone *zone, int count,
631 struct per_cpu_pages *pcp)
637 spin_lock(&zone->lock);
638 zone->all_unreclaimable = 0;
639 zone->pages_scanned = 0;
643 struct list_head *list;
646 * Remove pages from lists in a round-robin fashion. A
647 * batch_free count is maintained that is incremented when an
648 * empty list is encountered. This is so more pages are freed
649 * off fuller lists instead of spinning excessively around empty
654 if (++migratetype == MIGRATE_PCPTYPES)
656 list = &pcp->lists[migratetype];
657 } while (list_empty(list));
659 /* This is the only non-empty list. Free them all. */
660 if (batch_free == MIGRATE_PCPTYPES)
661 batch_free = to_free;
664 int mt; /* migratetype of the to-be-freed page */
666 page = list_entry(list->prev, struct page, lru);
667 /* must delete as __free_one_page list manipulates */
668 list_del(&page->lru);
669 mt = get_freepage_migratetype(page);
670 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
671 __free_one_page(page, zone, 0, mt);
672 trace_mm_page_pcpu_drain(page, 0, mt);
673 if (likely(get_pageblock_migratetype(page) != MIGRATE_ISOLATE)) {
674 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
675 if (is_migrate_cma(mt))
676 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
678 } while (--to_free && --batch_free && !list_empty(list));
680 spin_unlock(&zone->lock);
683 static void free_one_page(struct zone *zone, struct page *page, int order,
686 spin_lock(&zone->lock);
687 zone->all_unreclaimable = 0;
688 zone->pages_scanned = 0;
690 __free_one_page(page, zone, order, migratetype);
691 if (unlikely(migratetype != MIGRATE_ISOLATE))
692 __mod_zone_freepage_state(zone, 1 << order, migratetype);
693 spin_unlock(&zone->lock);
696 static bool free_pages_prepare(struct page *page, unsigned int order)
701 trace_mm_page_free(page, order);
702 kmemcheck_free_shadow(page, order);
705 page->mapping = NULL;
706 for (i = 0; i < (1 << order); i++)
707 bad += free_pages_check(page + i);
711 if (!PageHighMem(page)) {
712 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
713 debug_check_no_obj_freed(page_address(page),
716 arch_free_page(page, order);
717 kernel_map_pages(page, 1 << order, 0);
722 static void __free_pages_ok(struct page *page, unsigned int order)
727 if (!free_pages_prepare(page, order))
730 local_irq_save(flags);
731 __count_vm_events(PGFREE, 1 << order);
732 migratetype = get_pageblock_migratetype(page);
733 set_freepage_migratetype(page, migratetype);
734 free_one_page(page_zone(page), page, order, migratetype);
735 local_irq_restore(flags);
739 * Read access to zone->managed_pages is safe because it's unsigned long,
740 * but we still need to serialize writers. Currently all callers of
741 * __free_pages_bootmem() except put_page_bootmem() should only be used
742 * at boot time. So for shorter boot time, we shift the burden to
743 * put_page_bootmem() to serialize writers.
745 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
747 unsigned int nr_pages = 1 << order;
751 for (loop = 0; loop < nr_pages; loop++) {
752 struct page *p = &page[loop];
754 if (loop + 1 < nr_pages)
756 __ClearPageReserved(p);
757 set_page_count(p, 0);
760 page_zone(page)->managed_pages += 1 << order;
761 set_page_refcounted(page);
762 __free_pages(page, order);
766 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
767 void __init init_cma_reserved_pageblock(struct page *page)
769 unsigned i = pageblock_nr_pages;
770 struct page *p = page;
773 __ClearPageReserved(p);
774 set_page_count(p, 0);
777 set_page_refcounted(page);
778 set_pageblock_migratetype(page, MIGRATE_CMA);
779 __free_pages(page, pageblock_order);
780 totalram_pages += pageblock_nr_pages;
785 * The order of subdivision here is critical for the IO subsystem.
786 * Please do not alter this order without good reasons and regression
787 * testing. Specifically, as large blocks of memory are subdivided,
788 * the order in which smaller blocks are delivered depends on the order
789 * they're subdivided in this function. This is the primary factor
790 * influencing the order in which pages are delivered to the IO
791 * subsystem according to empirical testing, and this is also justified
792 * by considering the behavior of a buddy system containing a single
793 * large block of memory acted on by a series of small allocations.
794 * This behavior is a critical factor in sglist merging's success.
798 static inline void expand(struct zone *zone, struct page *page,
799 int low, int high, struct free_area *area,
802 unsigned long size = 1 << high;
808 VM_BUG_ON(bad_range(zone, &page[size]));
810 #ifdef CONFIG_DEBUG_PAGEALLOC
811 if (high < debug_guardpage_minorder()) {
813 * Mark as guard pages (or page), that will allow to
814 * merge back to allocator when buddy will be freed.
815 * Corresponding page table entries will not be touched,
816 * pages will stay not present in virtual address space
818 INIT_LIST_HEAD(&page[size].lru);
819 set_page_guard_flag(&page[size]);
820 set_page_private(&page[size], high);
821 /* Guard pages are not available for any usage */
822 __mod_zone_freepage_state(zone, -(1 << high),
827 list_add(&page[size].lru, &area->free_list[migratetype]);
829 set_page_order(&page[size], high);
834 * This page is about to be returned from the page allocator
836 static inline int check_new_page(struct page *page)
838 if (unlikely(page_mapcount(page) |
839 (page->mapping != NULL) |
840 (atomic_read(&page->_count) != 0) |
841 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
842 (mem_cgroup_bad_page_check(page)))) {
849 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
853 for (i = 0; i < (1 << order); i++) {
854 struct page *p = page + i;
855 if (unlikely(check_new_page(p)))
859 set_page_private(page, 0);
860 set_page_refcounted(page);
862 arch_alloc_page(page, order);
863 kernel_map_pages(page, 1 << order, 1);
865 if (gfp_flags & __GFP_ZERO)
866 prep_zero_page(page, order, gfp_flags);
868 if (order && (gfp_flags & __GFP_COMP))
869 prep_compound_page(page, order);
875 * Go through the free lists for the given migratetype and remove
876 * the smallest available page from the freelists
879 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
882 unsigned int current_order;
883 struct free_area * area;
886 /* Find a page of the appropriate size in the preferred list */
887 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
888 area = &(zone->free_area[current_order]);
889 if (list_empty(&area->free_list[migratetype]))
892 page = list_entry(area->free_list[migratetype].next,
894 list_del(&page->lru);
895 rmv_page_order(page);
897 expand(zone, page, order, current_order, area, migratetype);
906 * This array describes the order lists are fallen back to when
907 * the free lists for the desirable migrate type are depleted
909 static int fallbacks[MIGRATE_TYPES][4] = {
910 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
911 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
913 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
914 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
916 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
918 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
919 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
923 * Move the free pages in a range to the free lists of the requested type.
924 * Note that start_page and end_pages are not aligned on a pageblock
925 * boundary. If alignment is required, use move_freepages_block()
927 int move_freepages(struct zone *zone,
928 struct page *start_page, struct page *end_page,
935 #ifndef CONFIG_HOLES_IN_ZONE
937 * page_zone is not safe to call in this context when
938 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
939 * anyway as we check zone boundaries in move_freepages_block().
940 * Remove at a later date when no bug reports exist related to
941 * grouping pages by mobility
943 BUG_ON(page_zone(start_page) != page_zone(end_page));
946 for (page = start_page; page <= end_page;) {
947 /* Make sure we are not inadvertently changing nodes */
948 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
950 if (!pfn_valid_within(page_to_pfn(page))) {
955 if (!PageBuddy(page)) {
960 order = page_order(page);
961 list_move(&page->lru,
962 &zone->free_area[order].free_list[migratetype]);
963 set_freepage_migratetype(page, migratetype);
965 pages_moved += 1 << order;
971 int move_freepages_block(struct zone *zone, struct page *page,
974 unsigned long start_pfn, end_pfn;
975 struct page *start_page, *end_page;
977 start_pfn = page_to_pfn(page);
978 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
979 start_page = pfn_to_page(start_pfn);
980 end_page = start_page + pageblock_nr_pages - 1;
981 end_pfn = start_pfn + pageblock_nr_pages - 1;
983 /* Do not cross zone boundaries */
984 if (start_pfn < zone->zone_start_pfn)
986 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
989 return move_freepages(zone, start_page, end_page, migratetype);
992 static void change_pageblock_range(struct page *pageblock_page,
993 int start_order, int migratetype)
995 int nr_pageblocks = 1 << (start_order - pageblock_order);
997 while (nr_pageblocks--) {
998 set_pageblock_migratetype(pageblock_page, migratetype);
999 pageblock_page += pageblock_nr_pages;
1003 /* Remove an element from the buddy allocator from the fallback list */
1004 static inline struct page *
1005 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1007 struct free_area * area;
1012 /* Find the largest possible block of pages in the other list */
1013 for (current_order = MAX_ORDER-1; current_order >= order;
1016 migratetype = fallbacks[start_migratetype][i];
1018 /* MIGRATE_RESERVE handled later if necessary */
1019 if (migratetype == MIGRATE_RESERVE)
1022 area = &(zone->free_area[current_order]);
1023 if (list_empty(&area->free_list[migratetype]))
1026 page = list_entry(area->free_list[migratetype].next,
1031 * If breaking a large block of pages, move all free
1032 * pages to the preferred allocation list. If falling
1033 * back for a reclaimable kernel allocation, be more
1034 * aggressive about taking ownership of free pages
1036 * On the other hand, never change migration
1037 * type of MIGRATE_CMA pageblocks nor move CMA
1038 * pages on different free lists. We don't
1039 * want unmovable pages to be allocated from
1040 * MIGRATE_CMA areas.
1042 if (!is_migrate_cma(migratetype) &&
1043 (unlikely(current_order >= pageblock_order / 2) ||
1044 start_migratetype == MIGRATE_RECLAIMABLE ||
1045 page_group_by_mobility_disabled)) {
1047 pages = move_freepages_block(zone, page,
1050 /* Claim the whole block if over half of it is free */
1051 if (pages >= (1 << (pageblock_order-1)) ||
1052 page_group_by_mobility_disabled)
1053 set_pageblock_migratetype(page,
1056 migratetype = start_migratetype;
1059 /* Remove the page from the freelists */
1060 list_del(&page->lru);
1061 rmv_page_order(page);
1063 /* Take ownership for orders >= pageblock_order */
1064 if (current_order >= pageblock_order &&
1065 !is_migrate_cma(migratetype))
1066 change_pageblock_range(page, current_order,
1069 expand(zone, page, order, current_order, area,
1070 is_migrate_cma(migratetype)
1071 ? migratetype : start_migratetype);
1073 trace_mm_page_alloc_extfrag(page, order, current_order,
1074 start_migratetype, migratetype);
1084 * Do the hard work of removing an element from the buddy allocator.
1085 * Call me with the zone->lock already held.
1087 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1093 page = __rmqueue_smallest(zone, order, migratetype);
1095 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1096 page = __rmqueue_fallback(zone, order, migratetype);
1099 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1100 * is used because __rmqueue_smallest is an inline function
1101 * and we want just one call site
1104 migratetype = MIGRATE_RESERVE;
1109 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1114 * Obtain a specified number of elements from the buddy allocator, all under
1115 * a single hold of the lock, for efficiency. Add them to the supplied list.
1116 * Returns the number of new pages which were placed at *list.
1118 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1119 unsigned long count, struct list_head *list,
1120 int migratetype, int cold)
1122 int mt = migratetype, i;
1124 spin_lock(&zone->lock);
1125 for (i = 0; i < count; ++i) {
1126 struct page *page = __rmqueue(zone, order, migratetype);
1127 if (unlikely(page == NULL))
1131 * Split buddy pages returned by expand() are received here
1132 * in physical page order. The page is added to the callers and
1133 * list and the list head then moves forward. From the callers
1134 * perspective, the linked list is ordered by page number in
1135 * some conditions. This is useful for IO devices that can
1136 * merge IO requests if the physical pages are ordered
1139 if (likely(cold == 0))
1140 list_add(&page->lru, list);
1142 list_add_tail(&page->lru, list);
1143 if (IS_ENABLED(CONFIG_CMA)) {
1144 mt = get_pageblock_migratetype(page);
1145 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1148 set_freepage_migratetype(page, mt);
1150 if (is_migrate_cma(mt))
1151 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1154 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1155 spin_unlock(&zone->lock);
1161 * Called from the vmstat counter updater to drain pagesets of this
1162 * currently executing processor on remote nodes after they have
1165 * Note that this function must be called with the thread pinned to
1166 * a single processor.
1168 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1170 unsigned long flags;
1173 local_irq_save(flags);
1174 if (pcp->count >= pcp->batch)
1175 to_drain = pcp->batch;
1177 to_drain = pcp->count;
1179 free_pcppages_bulk(zone, to_drain, pcp);
1180 pcp->count -= to_drain;
1182 local_irq_restore(flags);
1187 * Drain pages of the indicated processor.
1189 * The processor must either be the current processor and the
1190 * thread pinned to the current processor or a processor that
1193 static void drain_pages(unsigned int cpu)
1195 unsigned long flags;
1198 for_each_populated_zone(zone) {
1199 struct per_cpu_pageset *pset;
1200 struct per_cpu_pages *pcp;
1202 local_irq_save(flags);
1203 pset = per_cpu_ptr(zone->pageset, cpu);
1207 free_pcppages_bulk(zone, pcp->count, pcp);
1210 local_irq_restore(flags);
1215 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1217 void drain_local_pages(void *arg)
1219 drain_pages(smp_processor_id());
1223 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1225 * Note that this code is protected against sending an IPI to an offline
1226 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1227 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1228 * nothing keeps CPUs from showing up after we populated the cpumask and
1229 * before the call to on_each_cpu_mask().
1231 void drain_all_pages(void)
1234 struct per_cpu_pageset *pcp;
1238 * Allocate in the BSS so we wont require allocation in
1239 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1241 static cpumask_t cpus_with_pcps;
1244 * We don't care about racing with CPU hotplug event
1245 * as offline notification will cause the notified
1246 * cpu to drain that CPU pcps and on_each_cpu_mask
1247 * disables preemption as part of its processing
1249 for_each_online_cpu(cpu) {
1250 bool has_pcps = false;
1251 for_each_populated_zone(zone) {
1252 pcp = per_cpu_ptr(zone->pageset, cpu);
1253 if (pcp->pcp.count) {
1259 cpumask_set_cpu(cpu, &cpus_with_pcps);
1261 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1263 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1266 #ifdef CONFIG_HIBERNATION
1268 void mark_free_pages(struct zone *zone)
1270 unsigned long pfn, max_zone_pfn;
1271 unsigned long flags;
1273 struct list_head *curr;
1275 if (!zone->spanned_pages)
1278 spin_lock_irqsave(&zone->lock, flags);
1280 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1281 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1282 if (pfn_valid(pfn)) {
1283 struct page *page = pfn_to_page(pfn);
1285 if (!swsusp_page_is_forbidden(page))
1286 swsusp_unset_page_free(page);
1289 for_each_migratetype_order(order, t) {
1290 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1293 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1294 for (i = 0; i < (1UL << order); i++)
1295 swsusp_set_page_free(pfn_to_page(pfn + i));
1298 spin_unlock_irqrestore(&zone->lock, flags);
1300 #endif /* CONFIG_PM */
1303 * Free a 0-order page
1304 * cold == 1 ? free a cold page : free a hot page
1306 void free_hot_cold_page(struct page *page, int cold)
1308 struct zone *zone = page_zone(page);
1309 struct per_cpu_pages *pcp;
1310 unsigned long flags;
1313 if (!free_pages_prepare(page, 0))
1316 migratetype = get_pageblock_migratetype(page);
1317 set_freepage_migratetype(page, migratetype);
1318 local_irq_save(flags);
1319 __count_vm_event(PGFREE);
1322 * We only track unmovable, reclaimable and movable on pcp lists.
1323 * Free ISOLATE pages back to the allocator because they are being
1324 * offlined but treat RESERVE as movable pages so we can get those
1325 * areas back if necessary. Otherwise, we may have to free
1326 * excessively into the page allocator
1328 if (migratetype >= MIGRATE_PCPTYPES) {
1329 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1330 free_one_page(zone, page, 0, migratetype);
1333 migratetype = MIGRATE_MOVABLE;
1336 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1338 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1340 list_add(&page->lru, &pcp->lists[migratetype]);
1342 if (pcp->count >= pcp->high) {
1343 free_pcppages_bulk(zone, pcp->batch, pcp);
1344 pcp->count -= pcp->batch;
1348 local_irq_restore(flags);
1352 * Free a list of 0-order pages
1354 void free_hot_cold_page_list(struct list_head *list, int cold)
1356 struct page *page, *next;
1358 list_for_each_entry_safe(page, next, list, lru) {
1359 trace_mm_page_free_batched(page, cold);
1360 free_hot_cold_page(page, cold);
1365 * split_page takes a non-compound higher-order page, and splits it into
1366 * n (1<<order) sub-pages: page[0..n]
1367 * Each sub-page must be freed individually.
1369 * Note: this is probably too low level an operation for use in drivers.
1370 * Please consult with lkml before using this in your driver.
1372 void split_page(struct page *page, unsigned int order)
1376 VM_BUG_ON(PageCompound(page));
1377 VM_BUG_ON(!page_count(page));
1379 #ifdef CONFIG_KMEMCHECK
1381 * Split shadow pages too, because free(page[0]) would
1382 * otherwise free the whole shadow.
1384 if (kmemcheck_page_is_tracked(page))
1385 split_page(virt_to_page(page[0].shadow), order);
1388 for (i = 1; i < (1 << order); i++)
1389 set_page_refcounted(page + i);
1393 * Similar to the split_page family of functions except that the page
1394 * required at the given order and being isolated now to prevent races
1395 * with parallel allocators
1397 int capture_free_page(struct page *page, int alloc_order, int migratetype)
1400 unsigned long watermark;
1404 BUG_ON(!PageBuddy(page));
1406 zone = page_zone(page);
1407 order = page_order(page);
1408 mt = get_pageblock_migratetype(page);
1410 if (mt != MIGRATE_ISOLATE) {
1411 /* Obey watermarks as if the page was being allocated */
1412 watermark = low_wmark_pages(zone) + (1 << order);
1413 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1416 __mod_zone_freepage_state(zone, -(1UL << alloc_order), mt);
1419 /* Remove page from free list */
1420 list_del(&page->lru);
1421 zone->free_area[order].nr_free--;
1422 rmv_page_order(page);
1424 if (alloc_order != order)
1425 expand(zone, page, alloc_order, order,
1426 &zone->free_area[order], migratetype);
1428 /* Set the pageblock if the captured page is at least a pageblock */
1429 if (order >= pageblock_order - 1) {
1430 struct page *endpage = page + (1 << order) - 1;
1431 for (; page < endpage; page += pageblock_nr_pages) {
1432 int mt = get_pageblock_migratetype(page);
1433 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1434 set_pageblock_migratetype(page,
1439 return 1UL << alloc_order;
1443 * Similar to split_page except the page is already free. As this is only
1444 * being used for migration, the migratetype of the block also changes.
1445 * As this is called with interrupts disabled, the caller is responsible
1446 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1449 * Note: this is probably too low level an operation for use in drivers.
1450 * Please consult with lkml before using this in your driver.
1452 int split_free_page(struct page *page)
1457 BUG_ON(!PageBuddy(page));
1458 order = page_order(page);
1460 nr_pages = capture_free_page(page, order, 0);
1464 /* Split into individual pages */
1465 set_page_refcounted(page);
1466 split_page(page, order);
1471 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1472 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1476 struct page *buffered_rmqueue(struct zone *preferred_zone,
1477 struct zone *zone, int order, gfp_t gfp_flags,
1480 unsigned long flags;
1482 int cold = !!(gfp_flags & __GFP_COLD);
1485 if (likely(order == 0)) {
1486 struct per_cpu_pages *pcp;
1487 struct list_head *list;
1489 local_irq_save(flags);
1490 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1491 list = &pcp->lists[migratetype];
1492 if (list_empty(list)) {
1493 pcp->count += rmqueue_bulk(zone, 0,
1496 if (unlikely(list_empty(list)))
1501 page = list_entry(list->prev, struct page, lru);
1503 page = list_entry(list->next, struct page, lru);
1505 list_del(&page->lru);
1508 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1510 * __GFP_NOFAIL is not to be used in new code.
1512 * All __GFP_NOFAIL callers should be fixed so that they
1513 * properly detect and handle allocation failures.
1515 * We most definitely don't want callers attempting to
1516 * allocate greater than order-1 page units with
1519 WARN_ON_ONCE(order > 1);
1521 spin_lock_irqsave(&zone->lock, flags);
1522 page = __rmqueue(zone, order, migratetype);
1523 spin_unlock(&zone->lock);
1526 __mod_zone_freepage_state(zone, -(1 << order),
1527 get_pageblock_migratetype(page));
1530 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1531 zone_statistics(preferred_zone, zone, gfp_flags);
1532 local_irq_restore(flags);
1534 VM_BUG_ON(bad_range(zone, page));
1535 if (prep_new_page(page, order, gfp_flags))
1540 local_irq_restore(flags);
1544 #ifdef CONFIG_FAIL_PAGE_ALLOC
1547 struct fault_attr attr;
1549 u32 ignore_gfp_highmem;
1550 u32 ignore_gfp_wait;
1552 } fail_page_alloc = {
1553 .attr = FAULT_ATTR_INITIALIZER,
1554 .ignore_gfp_wait = 1,
1555 .ignore_gfp_highmem = 1,
1559 static int __init setup_fail_page_alloc(char *str)
1561 return setup_fault_attr(&fail_page_alloc.attr, str);
1563 __setup("fail_page_alloc=", setup_fail_page_alloc);
1565 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1567 if (order < fail_page_alloc.min_order)
1569 if (gfp_mask & __GFP_NOFAIL)
1571 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1573 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1576 return should_fail(&fail_page_alloc.attr, 1 << order);
1579 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1581 static int __init fail_page_alloc_debugfs(void)
1583 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1586 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1587 &fail_page_alloc.attr);
1589 return PTR_ERR(dir);
1591 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1592 &fail_page_alloc.ignore_gfp_wait))
1594 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1595 &fail_page_alloc.ignore_gfp_highmem))
1597 if (!debugfs_create_u32("min-order", mode, dir,
1598 &fail_page_alloc.min_order))
1603 debugfs_remove_recursive(dir);
1608 late_initcall(fail_page_alloc_debugfs);
1610 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1612 #else /* CONFIG_FAIL_PAGE_ALLOC */
1614 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1619 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1622 * Return true if free pages are above 'mark'. This takes into account the order
1623 * of the allocation.
1625 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1626 int classzone_idx, int alloc_flags, long free_pages)
1628 /* free_pages my go negative - that's OK */
1630 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1633 free_pages -= (1 << order) - 1;
1634 if (alloc_flags & ALLOC_HIGH)
1636 if (alloc_flags & ALLOC_HARDER)
1639 /* If allocation can't use CMA areas don't use free CMA pages */
1640 if (!(alloc_flags & ALLOC_CMA))
1641 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1643 if (free_pages <= min + lowmem_reserve)
1645 for (o = 0; o < order; o++) {
1646 /* At the next order, this order's pages become unavailable */
1647 free_pages -= z->free_area[o].nr_free << o;
1649 /* Require fewer higher order pages to be free */
1652 if (free_pages <= min)
1658 #ifdef CONFIG_MEMORY_ISOLATION
1659 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1661 if (unlikely(zone->nr_pageblock_isolate))
1662 return zone->nr_pageblock_isolate * pageblock_nr_pages;
1666 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1672 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1673 int classzone_idx, int alloc_flags)
1675 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1676 zone_page_state(z, NR_FREE_PAGES));
1679 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1680 int classzone_idx, int alloc_flags)
1682 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1684 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1685 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1688 * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1689 * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1690 * sleep although it could do so. But this is more desirable for memory
1691 * hotplug than sleeping which can cause a livelock in the direct
1694 free_pages -= nr_zone_isolate_freepages(z);
1695 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1701 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1702 * skip over zones that are not allowed by the cpuset, or that have
1703 * been recently (in last second) found to be nearly full. See further
1704 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1705 * that have to skip over a lot of full or unallowed zones.
1707 * If the zonelist cache is present in the passed in zonelist, then
1708 * returns a pointer to the allowed node mask (either the current
1709 * tasks mems_allowed, or node_states[N_MEMORY].)
1711 * If the zonelist cache is not available for this zonelist, does
1712 * nothing and returns NULL.
1714 * If the fullzones BITMAP in the zonelist cache is stale (more than
1715 * a second since last zap'd) then we zap it out (clear its bits.)
1717 * We hold off even calling zlc_setup, until after we've checked the
1718 * first zone in the zonelist, on the theory that most allocations will
1719 * be satisfied from that first zone, so best to examine that zone as
1720 * quickly as we can.
1722 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1724 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1725 nodemask_t *allowednodes; /* zonelist_cache approximation */
1727 zlc = zonelist->zlcache_ptr;
1731 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1732 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1733 zlc->last_full_zap = jiffies;
1736 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1737 &cpuset_current_mems_allowed :
1738 &node_states[N_MEMORY];
1739 return allowednodes;
1743 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1744 * if it is worth looking at further for free memory:
1745 * 1) Check that the zone isn't thought to be full (doesn't have its
1746 * bit set in the zonelist_cache fullzones BITMAP).
1747 * 2) Check that the zones node (obtained from the zonelist_cache
1748 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1749 * Return true (non-zero) if zone is worth looking at further, or
1750 * else return false (zero) if it is not.
1752 * This check -ignores- the distinction between various watermarks,
1753 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1754 * found to be full for any variation of these watermarks, it will
1755 * be considered full for up to one second by all requests, unless
1756 * we are so low on memory on all allowed nodes that we are forced
1757 * into the second scan of the zonelist.
1759 * In the second scan we ignore this zonelist cache and exactly
1760 * apply the watermarks to all zones, even it is slower to do so.
1761 * We are low on memory in the second scan, and should leave no stone
1762 * unturned looking for a free page.
1764 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1765 nodemask_t *allowednodes)
1767 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1768 int i; /* index of *z in zonelist zones */
1769 int n; /* node that zone *z is on */
1771 zlc = zonelist->zlcache_ptr;
1775 i = z - zonelist->_zonerefs;
1778 /* This zone is worth trying if it is allowed but not full */
1779 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1783 * Given 'z' scanning a zonelist, set the corresponding bit in
1784 * zlc->fullzones, so that subsequent attempts to allocate a page
1785 * from that zone don't waste time re-examining it.
1787 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1789 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1790 int i; /* index of *z in zonelist zones */
1792 zlc = zonelist->zlcache_ptr;
1796 i = z - zonelist->_zonerefs;
1798 set_bit(i, zlc->fullzones);
1802 * clear all zones full, called after direct reclaim makes progress so that
1803 * a zone that was recently full is not skipped over for up to a second
1805 static void zlc_clear_zones_full(struct zonelist *zonelist)
1807 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1809 zlc = zonelist->zlcache_ptr;
1813 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1816 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1818 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1821 static void __paginginit init_zone_allows_reclaim(int nid)
1825 for_each_online_node(i)
1826 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1827 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1829 zone_reclaim_mode = 1;
1832 #else /* CONFIG_NUMA */
1834 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1839 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1840 nodemask_t *allowednodes)
1845 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1849 static void zlc_clear_zones_full(struct zonelist *zonelist)
1853 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1858 static inline void init_zone_allows_reclaim(int nid)
1861 #endif /* CONFIG_NUMA */
1864 * get_page_from_freelist goes through the zonelist trying to allocate
1867 static struct page *
1868 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1869 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1870 struct zone *preferred_zone, int migratetype)
1873 struct page *page = NULL;
1876 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1877 int zlc_active = 0; /* set if using zonelist_cache */
1878 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1880 classzone_idx = zone_idx(preferred_zone);
1883 * Scan zonelist, looking for a zone with enough free.
1884 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1886 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1887 high_zoneidx, nodemask) {
1888 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1889 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1891 if ((alloc_flags & ALLOC_CPUSET) &&
1892 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1895 * When allocating a page cache page for writing, we
1896 * want to get it from a zone that is within its dirty
1897 * limit, such that no single zone holds more than its
1898 * proportional share of globally allowed dirty pages.
1899 * The dirty limits take into account the zone's
1900 * lowmem reserves and high watermark so that kswapd
1901 * should be able to balance it without having to
1902 * write pages from its LRU list.
1904 * This may look like it could increase pressure on
1905 * lower zones by failing allocations in higher zones
1906 * before they are full. But the pages that do spill
1907 * over are limited as the lower zones are protected
1908 * by this very same mechanism. It should not become
1909 * a practical burden to them.
1911 * XXX: For now, allow allocations to potentially
1912 * exceed the per-zone dirty limit in the slowpath
1913 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1914 * which is important when on a NUMA setup the allowed
1915 * zones are together not big enough to reach the
1916 * global limit. The proper fix for these situations
1917 * will require awareness of zones in the
1918 * dirty-throttling and the flusher threads.
1920 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1921 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1922 goto this_zone_full;
1924 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1925 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1929 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1930 if (zone_watermark_ok(zone, order, mark,
1931 classzone_idx, alloc_flags))
1934 if (IS_ENABLED(CONFIG_NUMA) &&
1935 !did_zlc_setup && nr_online_nodes > 1) {
1937 * we do zlc_setup if there are multiple nodes
1938 * and before considering the first zone allowed
1941 allowednodes = zlc_setup(zonelist, alloc_flags);
1946 if (zone_reclaim_mode == 0 ||
1947 !zone_allows_reclaim(preferred_zone, zone))
1948 goto this_zone_full;
1951 * As we may have just activated ZLC, check if the first
1952 * eligible zone has failed zone_reclaim recently.
1954 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1955 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1958 ret = zone_reclaim(zone, gfp_mask, order);
1960 case ZONE_RECLAIM_NOSCAN:
1963 case ZONE_RECLAIM_FULL:
1964 /* scanned but unreclaimable */
1967 /* did we reclaim enough */
1968 if (!zone_watermark_ok(zone, order, mark,
1969 classzone_idx, alloc_flags))
1970 goto this_zone_full;
1975 page = buffered_rmqueue(preferred_zone, zone, order,
1976 gfp_mask, migratetype);
1980 if (IS_ENABLED(CONFIG_NUMA))
1981 zlc_mark_zone_full(zonelist, z);
1984 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1985 /* Disable zlc cache for second zonelist scan */
1992 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1993 * necessary to allocate the page. The expectation is
1994 * that the caller is taking steps that will free more
1995 * memory. The caller should avoid the page being used
1996 * for !PFMEMALLOC purposes.
1998 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2004 * Large machines with many possible nodes should not always dump per-node
2005 * meminfo in irq context.
2007 static inline bool should_suppress_show_mem(void)
2012 ret = in_interrupt();
2017 static DEFINE_RATELIMIT_STATE(nopage_rs,
2018 DEFAULT_RATELIMIT_INTERVAL,
2019 DEFAULT_RATELIMIT_BURST);
2021 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2023 unsigned int filter = SHOW_MEM_FILTER_NODES;
2025 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2026 debug_guardpage_minorder() > 0)
2030 * This documents exceptions given to allocations in certain
2031 * contexts that are allowed to allocate outside current's set
2034 if (!(gfp_mask & __GFP_NOMEMALLOC))
2035 if (test_thread_flag(TIF_MEMDIE) ||
2036 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2037 filter &= ~SHOW_MEM_FILTER_NODES;
2038 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2039 filter &= ~SHOW_MEM_FILTER_NODES;
2042 struct va_format vaf;
2045 va_start(args, fmt);
2050 pr_warn("%pV", &vaf);
2055 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2056 current->comm, order, gfp_mask);
2059 if (!should_suppress_show_mem())
2064 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2065 unsigned long did_some_progress,
2066 unsigned long pages_reclaimed)
2068 /* Do not loop if specifically requested */
2069 if (gfp_mask & __GFP_NORETRY)
2072 /* Always retry if specifically requested */
2073 if (gfp_mask & __GFP_NOFAIL)
2077 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2078 * making forward progress without invoking OOM. Suspend also disables
2079 * storage devices so kswapd will not help. Bail if we are suspending.
2081 if (!did_some_progress && pm_suspended_storage())
2085 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2086 * means __GFP_NOFAIL, but that may not be true in other
2089 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2093 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2094 * specified, then we retry until we no longer reclaim any pages
2095 * (above), or we've reclaimed an order of pages at least as
2096 * large as the allocation's order. In both cases, if the
2097 * allocation still fails, we stop retrying.
2099 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2105 static inline struct page *
2106 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2107 struct zonelist *zonelist, enum zone_type high_zoneidx,
2108 nodemask_t *nodemask, struct zone *preferred_zone,
2113 /* Acquire the OOM killer lock for the zones in zonelist */
2114 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2115 schedule_timeout_uninterruptible(1);
2120 * Go through the zonelist yet one more time, keep very high watermark
2121 * here, this is only to catch a parallel oom killing, we must fail if
2122 * we're still under heavy pressure.
2124 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2125 order, zonelist, high_zoneidx,
2126 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2127 preferred_zone, migratetype);
2131 if (!(gfp_mask & __GFP_NOFAIL)) {
2132 /* The OOM killer will not help higher order allocs */
2133 if (order > PAGE_ALLOC_COSTLY_ORDER)
2135 /* The OOM killer does not needlessly kill tasks for lowmem */
2136 if (high_zoneidx < ZONE_NORMAL)
2139 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2140 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2141 * The caller should handle page allocation failure by itself if
2142 * it specifies __GFP_THISNODE.
2143 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2145 if (gfp_mask & __GFP_THISNODE)
2148 /* Exhausted what can be done so it's blamo time */
2149 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2152 clear_zonelist_oom(zonelist, gfp_mask);
2156 #ifdef CONFIG_COMPACTION
2157 /* Try memory compaction for high-order allocations before reclaim */
2158 static struct page *
2159 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2160 struct zonelist *zonelist, enum zone_type high_zoneidx,
2161 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2162 int migratetype, bool sync_migration,
2163 bool *contended_compaction, bool *deferred_compaction,
2164 unsigned long *did_some_progress)
2166 struct page *page = NULL;
2171 if (compaction_deferred(preferred_zone, order)) {
2172 *deferred_compaction = true;
2176 current->flags |= PF_MEMALLOC;
2177 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2178 nodemask, sync_migration,
2179 contended_compaction, &page);
2180 current->flags &= ~PF_MEMALLOC;
2182 /* If compaction captured a page, prep and use it */
2184 prep_new_page(page, order, gfp_mask);
2188 if (*did_some_progress != COMPACT_SKIPPED) {
2189 /* Page migration frees to the PCP lists but we want merging */
2190 drain_pages(get_cpu());
2193 page = get_page_from_freelist(gfp_mask, nodemask,
2194 order, zonelist, high_zoneidx,
2195 alloc_flags & ~ALLOC_NO_WATERMARKS,
2196 preferred_zone, migratetype);
2199 preferred_zone->compact_blockskip_flush = false;
2200 preferred_zone->compact_considered = 0;
2201 preferred_zone->compact_defer_shift = 0;
2202 if (order >= preferred_zone->compact_order_failed)
2203 preferred_zone->compact_order_failed = order + 1;
2204 count_vm_event(COMPACTSUCCESS);
2209 * It's bad if compaction run occurs and fails.
2210 * The most likely reason is that pages exist,
2211 * but not enough to satisfy watermarks.
2213 count_vm_event(COMPACTFAIL);
2216 * As async compaction considers a subset of pageblocks, only
2217 * defer if the failure was a sync compaction failure.
2220 defer_compaction(preferred_zone, order);
2228 static inline struct page *
2229 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2230 struct zonelist *zonelist, enum zone_type high_zoneidx,
2231 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2232 int migratetype, bool sync_migration,
2233 bool *contended_compaction, bool *deferred_compaction,
2234 unsigned long *did_some_progress)
2238 #endif /* CONFIG_COMPACTION */
2240 /* Perform direct synchronous page reclaim */
2242 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2243 nodemask_t *nodemask)
2245 struct reclaim_state reclaim_state;
2250 /* We now go into synchronous reclaim */
2251 cpuset_memory_pressure_bump();
2252 current->flags |= PF_MEMALLOC;
2253 lockdep_set_current_reclaim_state(gfp_mask);
2254 reclaim_state.reclaimed_slab = 0;
2255 current->reclaim_state = &reclaim_state;
2257 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2259 current->reclaim_state = NULL;
2260 lockdep_clear_current_reclaim_state();
2261 current->flags &= ~PF_MEMALLOC;
2268 /* The really slow allocator path where we enter direct reclaim */
2269 static inline struct page *
2270 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2271 struct zonelist *zonelist, enum zone_type high_zoneidx,
2272 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2273 int migratetype, unsigned long *did_some_progress)
2275 struct page *page = NULL;
2276 bool drained = false;
2278 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2280 if (unlikely(!(*did_some_progress)))
2283 /* After successful reclaim, reconsider all zones for allocation */
2284 if (IS_ENABLED(CONFIG_NUMA))
2285 zlc_clear_zones_full(zonelist);
2288 page = get_page_from_freelist(gfp_mask, nodemask, order,
2289 zonelist, high_zoneidx,
2290 alloc_flags & ~ALLOC_NO_WATERMARKS,
2291 preferred_zone, migratetype);
2294 * If an allocation failed after direct reclaim, it could be because
2295 * pages are pinned on the per-cpu lists. Drain them and try again
2297 if (!page && !drained) {
2307 * This is called in the allocator slow-path if the allocation request is of
2308 * sufficient urgency to ignore watermarks and take other desperate measures
2310 static inline struct page *
2311 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2312 struct zonelist *zonelist, enum zone_type high_zoneidx,
2313 nodemask_t *nodemask, struct zone *preferred_zone,
2319 page = get_page_from_freelist(gfp_mask, nodemask, order,
2320 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2321 preferred_zone, migratetype);
2323 if (!page && gfp_mask & __GFP_NOFAIL)
2324 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2325 } while (!page && (gfp_mask & __GFP_NOFAIL));
2331 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2332 enum zone_type high_zoneidx,
2333 enum zone_type classzone_idx)
2338 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2339 wakeup_kswapd(zone, order, classzone_idx);
2343 gfp_to_alloc_flags(gfp_t gfp_mask)
2345 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2346 const gfp_t wait = gfp_mask & __GFP_WAIT;
2348 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2349 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2352 * The caller may dip into page reserves a bit more if the caller
2353 * cannot run direct reclaim, or if the caller has realtime scheduling
2354 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2355 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2357 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2361 * Not worth trying to allocate harder for
2362 * __GFP_NOMEMALLOC even if it can't schedule.
2364 if (!(gfp_mask & __GFP_NOMEMALLOC))
2365 alloc_flags |= ALLOC_HARDER;
2367 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2368 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2370 alloc_flags &= ~ALLOC_CPUSET;
2371 } else if (unlikely(rt_task(current)) && !in_interrupt())
2372 alloc_flags |= ALLOC_HARDER;
2374 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2375 if (gfp_mask & __GFP_MEMALLOC)
2376 alloc_flags |= ALLOC_NO_WATERMARKS;
2377 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2378 alloc_flags |= ALLOC_NO_WATERMARKS;
2379 else if (!in_interrupt() &&
2380 ((current->flags & PF_MEMALLOC) ||
2381 unlikely(test_thread_flag(TIF_MEMDIE))))
2382 alloc_flags |= ALLOC_NO_WATERMARKS;
2385 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2386 alloc_flags |= ALLOC_CMA;
2391 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2393 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2396 static inline struct page *
2397 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2398 struct zonelist *zonelist, enum zone_type high_zoneidx,
2399 nodemask_t *nodemask, struct zone *preferred_zone,
2402 const gfp_t wait = gfp_mask & __GFP_WAIT;
2403 struct page *page = NULL;
2405 unsigned long pages_reclaimed = 0;
2406 unsigned long did_some_progress;
2407 bool sync_migration = false;
2408 bool deferred_compaction = false;
2409 bool contended_compaction = false;
2412 * In the slowpath, we sanity check order to avoid ever trying to
2413 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2414 * be using allocators in order of preference for an area that is
2417 if (order >= MAX_ORDER) {
2418 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2423 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2424 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2425 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2426 * using a larger set of nodes after it has established that the
2427 * allowed per node queues are empty and that nodes are
2430 if (IS_ENABLED(CONFIG_NUMA) &&
2431 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2435 if (!(gfp_mask & __GFP_NO_KSWAPD))
2436 wake_all_kswapd(order, zonelist, high_zoneidx,
2437 zone_idx(preferred_zone));
2440 * OK, we're below the kswapd watermark and have kicked background
2441 * reclaim. Now things get more complex, so set up alloc_flags according
2442 * to how we want to proceed.
2444 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2447 * Find the true preferred zone if the allocation is unconstrained by
2450 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2451 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2455 /* This is the last chance, in general, before the goto nopage. */
2456 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2457 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2458 preferred_zone, migratetype);
2462 /* Allocate without watermarks if the context allows */
2463 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2465 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2466 * the allocation is high priority and these type of
2467 * allocations are system rather than user orientated
2469 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2471 page = __alloc_pages_high_priority(gfp_mask, order,
2472 zonelist, high_zoneidx, nodemask,
2473 preferred_zone, migratetype);
2479 /* Atomic allocations - we can't balance anything */
2483 /* Avoid recursion of direct reclaim */
2484 if (current->flags & PF_MEMALLOC)
2487 /* Avoid allocations with no watermarks from looping endlessly */
2488 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2492 * Try direct compaction. The first pass is asynchronous. Subsequent
2493 * attempts after direct reclaim are synchronous
2495 page = __alloc_pages_direct_compact(gfp_mask, order,
2496 zonelist, high_zoneidx,
2498 alloc_flags, preferred_zone,
2499 migratetype, sync_migration,
2500 &contended_compaction,
2501 &deferred_compaction,
2502 &did_some_progress);
2505 sync_migration = true;
2508 * If compaction is deferred for high-order allocations, it is because
2509 * sync compaction recently failed. In this is the case and the caller
2510 * requested a movable allocation that does not heavily disrupt the
2511 * system then fail the allocation instead of entering direct reclaim.
2513 if ((deferred_compaction || contended_compaction) &&
2514 (gfp_mask & __GFP_NO_KSWAPD))
2517 /* Try direct reclaim and then allocating */
2518 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2519 zonelist, high_zoneidx,
2521 alloc_flags, preferred_zone,
2522 migratetype, &did_some_progress);
2527 * If we failed to make any progress reclaiming, then we are
2528 * running out of options and have to consider going OOM
2530 if (!did_some_progress) {
2531 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2532 if (oom_killer_disabled)
2534 /* Coredumps can quickly deplete all memory reserves */
2535 if ((current->flags & PF_DUMPCORE) &&
2536 !(gfp_mask & __GFP_NOFAIL))
2538 page = __alloc_pages_may_oom(gfp_mask, order,
2539 zonelist, high_zoneidx,
2540 nodemask, preferred_zone,
2545 if (!(gfp_mask & __GFP_NOFAIL)) {
2547 * The oom killer is not called for high-order
2548 * allocations that may fail, so if no progress
2549 * is being made, there are no other options and
2550 * retrying is unlikely to help.
2552 if (order > PAGE_ALLOC_COSTLY_ORDER)
2555 * The oom killer is not called for lowmem
2556 * allocations to prevent needlessly killing
2559 if (high_zoneidx < ZONE_NORMAL)
2567 /* Check if we should retry the allocation */
2568 pages_reclaimed += did_some_progress;
2569 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2571 /* Wait for some write requests to complete then retry */
2572 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2576 * High-order allocations do not necessarily loop after
2577 * direct reclaim and reclaim/compaction depends on compaction
2578 * being called after reclaim so call directly if necessary
2580 page = __alloc_pages_direct_compact(gfp_mask, order,
2581 zonelist, high_zoneidx,
2583 alloc_flags, preferred_zone,
2584 migratetype, sync_migration,
2585 &contended_compaction,
2586 &deferred_compaction,
2587 &did_some_progress);
2593 warn_alloc_failed(gfp_mask, order, NULL);
2596 if (kmemcheck_enabled)
2597 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2603 * This is the 'heart' of the zoned buddy allocator.
2606 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2607 struct zonelist *zonelist, nodemask_t *nodemask)
2609 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2610 struct zone *preferred_zone;
2611 struct page *page = NULL;
2612 int migratetype = allocflags_to_migratetype(gfp_mask);
2613 unsigned int cpuset_mems_cookie;
2614 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2616 gfp_mask &= gfp_allowed_mask;
2618 lockdep_trace_alloc(gfp_mask);
2620 might_sleep_if(gfp_mask & __GFP_WAIT);
2622 if (should_fail_alloc_page(gfp_mask, order))
2626 * Check the zones suitable for the gfp_mask contain at least one
2627 * valid zone. It's possible to have an empty zonelist as a result
2628 * of GFP_THISNODE and a memoryless node
2630 if (unlikely(!zonelist->_zonerefs->zone))
2634 cpuset_mems_cookie = get_mems_allowed();
2636 /* The preferred zone is used for statistics later */
2637 first_zones_zonelist(zonelist, high_zoneidx,
2638 nodemask ? : &cpuset_current_mems_allowed,
2640 if (!preferred_zone)
2644 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2645 alloc_flags |= ALLOC_CMA;
2647 /* First allocation attempt */
2648 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2649 zonelist, high_zoneidx, alloc_flags,
2650 preferred_zone, migratetype);
2651 if (unlikely(!page))
2652 page = __alloc_pages_slowpath(gfp_mask, order,
2653 zonelist, high_zoneidx, nodemask,
2654 preferred_zone, migratetype);
2656 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2660 * When updating a task's mems_allowed, it is possible to race with
2661 * parallel threads in such a way that an allocation can fail while
2662 * the mask is being updated. If a page allocation is about to fail,
2663 * check if the cpuset changed during allocation and if so, retry.
2665 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2670 EXPORT_SYMBOL(__alloc_pages_nodemask);
2673 * Common helper functions.
2675 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2680 * __get_free_pages() returns a 32-bit address, which cannot represent
2683 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2685 page = alloc_pages(gfp_mask, order);
2688 return (unsigned long) page_address(page);
2690 EXPORT_SYMBOL(__get_free_pages);
2692 unsigned long get_zeroed_page(gfp_t gfp_mask)
2694 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2696 EXPORT_SYMBOL(get_zeroed_page);
2698 void __free_pages(struct page *page, unsigned int order)
2700 if (put_page_testzero(page)) {
2702 free_hot_cold_page(page, 0);
2704 __free_pages_ok(page, order);
2708 EXPORT_SYMBOL(__free_pages);
2710 void free_pages(unsigned long addr, unsigned int order)
2713 VM_BUG_ON(!virt_addr_valid((void *)addr));
2714 __free_pages(virt_to_page((void *)addr), order);
2718 EXPORT_SYMBOL(free_pages);
2720 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2723 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2724 unsigned long used = addr + PAGE_ALIGN(size);
2726 split_page(virt_to_page((void *)addr), order);
2727 while (used < alloc_end) {
2732 return (void *)addr;
2736 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2737 * @size: the number of bytes to allocate
2738 * @gfp_mask: GFP flags for the allocation
2740 * This function is similar to alloc_pages(), except that it allocates the
2741 * minimum number of pages to satisfy the request. alloc_pages() can only
2742 * allocate memory in power-of-two pages.
2744 * This function is also limited by MAX_ORDER.
2746 * Memory allocated by this function must be released by free_pages_exact().
2748 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2750 unsigned int order = get_order(size);
2753 addr = __get_free_pages(gfp_mask, order);
2754 return make_alloc_exact(addr, order, size);
2756 EXPORT_SYMBOL(alloc_pages_exact);
2759 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2761 * @nid: the preferred node ID where memory should be allocated
2762 * @size: the number of bytes to allocate
2763 * @gfp_mask: GFP flags for the allocation
2765 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2767 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2770 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2772 unsigned order = get_order(size);
2773 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2776 return make_alloc_exact((unsigned long)page_address(p), order, size);
2778 EXPORT_SYMBOL(alloc_pages_exact_nid);
2781 * free_pages_exact - release memory allocated via alloc_pages_exact()
2782 * @virt: the value returned by alloc_pages_exact.
2783 * @size: size of allocation, same value as passed to alloc_pages_exact().
2785 * Release the memory allocated by a previous call to alloc_pages_exact.
2787 void free_pages_exact(void *virt, size_t size)
2789 unsigned long addr = (unsigned long)virt;
2790 unsigned long end = addr + PAGE_ALIGN(size);
2792 while (addr < end) {
2797 EXPORT_SYMBOL(free_pages_exact);
2799 static unsigned int nr_free_zone_pages(int offset)
2804 /* Just pick one node, since fallback list is circular */
2805 unsigned int sum = 0;
2807 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2809 for_each_zone_zonelist(zone, z, zonelist, offset) {
2810 unsigned long size = zone->present_pages;
2811 unsigned long high = high_wmark_pages(zone);
2820 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2822 unsigned int nr_free_buffer_pages(void)
2824 return nr_free_zone_pages(gfp_zone(GFP_USER));
2826 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2829 * Amount of free RAM allocatable within all zones
2831 unsigned int nr_free_pagecache_pages(void)
2833 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2836 static inline void show_node(struct zone *zone)
2838 if (IS_ENABLED(CONFIG_NUMA))
2839 printk("Node %d ", zone_to_nid(zone));
2842 void si_meminfo(struct sysinfo *val)
2844 val->totalram = totalram_pages;
2846 val->freeram = global_page_state(NR_FREE_PAGES);
2847 val->bufferram = nr_blockdev_pages();
2848 val->totalhigh = totalhigh_pages;
2849 val->freehigh = nr_free_highpages();
2850 val->mem_unit = PAGE_SIZE;
2853 EXPORT_SYMBOL(si_meminfo);
2856 void si_meminfo_node(struct sysinfo *val, int nid)
2858 pg_data_t *pgdat = NODE_DATA(nid);
2860 val->totalram = pgdat->node_present_pages;
2861 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2862 #ifdef CONFIG_HIGHMEM
2863 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2864 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2870 val->mem_unit = PAGE_SIZE;
2875 * Determine whether the node should be displayed or not, depending on whether
2876 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2878 bool skip_free_areas_node(unsigned int flags, int nid)
2881 unsigned int cpuset_mems_cookie;
2883 if (!(flags & SHOW_MEM_FILTER_NODES))
2887 cpuset_mems_cookie = get_mems_allowed();
2888 ret = !node_isset(nid, cpuset_current_mems_allowed);
2889 } while (!put_mems_allowed(cpuset_mems_cookie));
2894 #define K(x) ((x) << (PAGE_SHIFT-10))
2896 static void show_migration_types(unsigned char type)
2898 static const char types[MIGRATE_TYPES] = {
2899 [MIGRATE_UNMOVABLE] = 'U',
2900 [MIGRATE_RECLAIMABLE] = 'E',
2901 [MIGRATE_MOVABLE] = 'M',
2902 [MIGRATE_RESERVE] = 'R',
2904 [MIGRATE_CMA] = 'C',
2906 [MIGRATE_ISOLATE] = 'I',
2908 char tmp[MIGRATE_TYPES + 1];
2912 for (i = 0; i < MIGRATE_TYPES; i++) {
2913 if (type & (1 << i))
2918 printk("(%s) ", tmp);
2922 * Show free area list (used inside shift_scroll-lock stuff)
2923 * We also calculate the percentage fragmentation. We do this by counting the
2924 * memory on each free list with the exception of the first item on the list.
2925 * Suppresses nodes that are not allowed by current's cpuset if
2926 * SHOW_MEM_FILTER_NODES is passed.
2928 void show_free_areas(unsigned int filter)
2933 for_each_populated_zone(zone) {
2934 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2937 printk("%s per-cpu:\n", zone->name);
2939 for_each_online_cpu(cpu) {
2940 struct per_cpu_pageset *pageset;
2942 pageset = per_cpu_ptr(zone->pageset, cpu);
2944 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2945 cpu, pageset->pcp.high,
2946 pageset->pcp.batch, pageset->pcp.count);
2950 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2951 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2953 " dirty:%lu writeback:%lu unstable:%lu\n"
2954 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2955 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2957 global_page_state(NR_ACTIVE_ANON),
2958 global_page_state(NR_INACTIVE_ANON),
2959 global_page_state(NR_ISOLATED_ANON),
2960 global_page_state(NR_ACTIVE_FILE),
2961 global_page_state(NR_INACTIVE_FILE),
2962 global_page_state(NR_ISOLATED_FILE),
2963 global_page_state(NR_UNEVICTABLE),
2964 global_page_state(NR_FILE_DIRTY),
2965 global_page_state(NR_WRITEBACK),
2966 global_page_state(NR_UNSTABLE_NFS),
2967 global_page_state(NR_FREE_PAGES),
2968 global_page_state(NR_SLAB_RECLAIMABLE),
2969 global_page_state(NR_SLAB_UNRECLAIMABLE),
2970 global_page_state(NR_FILE_MAPPED),
2971 global_page_state(NR_SHMEM),
2972 global_page_state(NR_PAGETABLE),
2973 global_page_state(NR_BOUNCE),
2974 global_page_state(NR_FREE_CMA_PAGES));
2976 for_each_populated_zone(zone) {
2979 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2987 " active_anon:%lukB"
2988 " inactive_anon:%lukB"
2989 " active_file:%lukB"
2990 " inactive_file:%lukB"
2991 " unevictable:%lukB"
2992 " isolated(anon):%lukB"
2993 " isolated(file):%lukB"
3001 " slab_reclaimable:%lukB"
3002 " slab_unreclaimable:%lukB"
3003 " kernel_stack:%lukB"
3008 " writeback_tmp:%lukB"
3009 " pages_scanned:%lu"
3010 " all_unreclaimable? %s"
3013 K(zone_page_state(zone, NR_FREE_PAGES)),
3014 K(min_wmark_pages(zone)),
3015 K(low_wmark_pages(zone)),
3016 K(high_wmark_pages(zone)),
3017 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3018 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3019 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3020 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3021 K(zone_page_state(zone, NR_UNEVICTABLE)),
3022 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3023 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3024 K(zone->present_pages),
3025 K(zone->managed_pages),
3026 K(zone_page_state(zone, NR_MLOCK)),
3027 K(zone_page_state(zone, NR_FILE_DIRTY)),
3028 K(zone_page_state(zone, NR_WRITEBACK)),
3029 K(zone_page_state(zone, NR_FILE_MAPPED)),
3030 K(zone_page_state(zone, NR_SHMEM)),
3031 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3032 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3033 zone_page_state(zone, NR_KERNEL_STACK) *
3035 K(zone_page_state(zone, NR_PAGETABLE)),
3036 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3037 K(zone_page_state(zone, NR_BOUNCE)),
3038 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3039 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3040 zone->pages_scanned,
3041 (zone->all_unreclaimable ? "yes" : "no")
3043 printk("lowmem_reserve[]:");
3044 for (i = 0; i < MAX_NR_ZONES; i++)
3045 printk(" %lu", zone->lowmem_reserve[i]);
3049 for_each_populated_zone(zone) {
3050 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3051 unsigned char types[MAX_ORDER];
3053 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3056 printk("%s: ", zone->name);
3058 spin_lock_irqsave(&zone->lock, flags);
3059 for (order = 0; order < MAX_ORDER; order++) {
3060 struct free_area *area = &zone->free_area[order];
3063 nr[order] = area->nr_free;
3064 total += nr[order] << order;
3067 for (type = 0; type < MIGRATE_TYPES; type++) {
3068 if (!list_empty(&area->free_list[type]))
3069 types[order] |= 1 << type;
3072 spin_unlock_irqrestore(&zone->lock, flags);
3073 for (order = 0; order < MAX_ORDER; order++) {
3074 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3076 show_migration_types(types[order]);
3078 printk("= %lukB\n", K(total));
3081 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3083 show_swap_cache_info();
3086 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3088 zoneref->zone = zone;
3089 zoneref->zone_idx = zone_idx(zone);
3093 * Builds allocation fallback zone lists.
3095 * Add all populated zones of a node to the zonelist.
3097 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3098 int nr_zones, enum zone_type zone_type)
3102 BUG_ON(zone_type >= MAX_NR_ZONES);
3107 zone = pgdat->node_zones + zone_type;
3108 if (populated_zone(zone)) {
3109 zoneref_set_zone(zone,
3110 &zonelist->_zonerefs[nr_zones++]);
3111 check_highest_zone(zone_type);
3114 } while (zone_type);
3121 * 0 = automatic detection of better ordering.
3122 * 1 = order by ([node] distance, -zonetype)
3123 * 2 = order by (-zonetype, [node] distance)
3125 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3126 * the same zonelist. So only NUMA can configure this param.
3128 #define ZONELIST_ORDER_DEFAULT 0
3129 #define ZONELIST_ORDER_NODE 1
3130 #define ZONELIST_ORDER_ZONE 2
3132 /* zonelist order in the kernel.
3133 * set_zonelist_order() will set this to NODE or ZONE.
3135 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3136 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3140 /* The value user specified ....changed by config */
3141 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3142 /* string for sysctl */
3143 #define NUMA_ZONELIST_ORDER_LEN 16
3144 char numa_zonelist_order[16] = "default";
3147 * interface for configure zonelist ordering.
3148 * command line option "numa_zonelist_order"
3149 * = "[dD]efault - default, automatic configuration.
3150 * = "[nN]ode - order by node locality, then by zone within node
3151 * = "[zZ]one - order by zone, then by locality within zone
3154 static int __parse_numa_zonelist_order(char *s)
3156 if (*s == 'd' || *s == 'D') {
3157 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3158 } else if (*s == 'n' || *s == 'N') {
3159 user_zonelist_order = ZONELIST_ORDER_NODE;
3160 } else if (*s == 'z' || *s == 'Z') {
3161 user_zonelist_order = ZONELIST_ORDER_ZONE;
3164 "Ignoring invalid numa_zonelist_order value: "
3171 static __init int setup_numa_zonelist_order(char *s)
3178 ret = __parse_numa_zonelist_order(s);
3180 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3184 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3187 * sysctl handler for numa_zonelist_order
3189 int numa_zonelist_order_handler(ctl_table *table, int write,
3190 void __user *buffer, size_t *length,
3193 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3195 static DEFINE_MUTEX(zl_order_mutex);
3197 mutex_lock(&zl_order_mutex);
3199 strcpy(saved_string, (char*)table->data);
3200 ret = proc_dostring(table, write, buffer, length, ppos);
3204 int oldval = user_zonelist_order;
3205 if (__parse_numa_zonelist_order((char*)table->data)) {
3207 * bogus value. restore saved string
3209 strncpy((char*)table->data, saved_string,
3210 NUMA_ZONELIST_ORDER_LEN);
3211 user_zonelist_order = oldval;
3212 } else if (oldval != user_zonelist_order) {
3213 mutex_lock(&zonelists_mutex);
3214 build_all_zonelists(NULL, NULL);
3215 mutex_unlock(&zonelists_mutex);
3219 mutex_unlock(&zl_order_mutex);
3224 #define MAX_NODE_LOAD (nr_online_nodes)
3225 static int node_load[MAX_NUMNODES];
3228 * find_next_best_node - find the next node that should appear in a given node's fallback list
3229 * @node: node whose fallback list we're appending
3230 * @used_node_mask: nodemask_t of already used nodes
3232 * We use a number of factors to determine which is the next node that should
3233 * appear on a given node's fallback list. The node should not have appeared
3234 * already in @node's fallback list, and it should be the next closest node
3235 * according to the distance array (which contains arbitrary distance values
3236 * from each node to each node in the system), and should also prefer nodes
3237 * with no CPUs, since presumably they'll have very little allocation pressure
3238 * on them otherwise.
3239 * It returns -1 if no node is found.
3241 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3244 int min_val = INT_MAX;
3246 const struct cpumask *tmp = cpumask_of_node(0);
3248 /* Use the local node if we haven't already */
3249 if (!node_isset(node, *used_node_mask)) {
3250 node_set(node, *used_node_mask);
3254 for_each_node_state(n, N_MEMORY) {
3256 /* Don't want a node to appear more than once */
3257 if (node_isset(n, *used_node_mask))
3260 /* Use the distance array to find the distance */
3261 val = node_distance(node, n);
3263 /* Penalize nodes under us ("prefer the next node") */
3266 /* Give preference to headless and unused nodes */
3267 tmp = cpumask_of_node(n);
3268 if (!cpumask_empty(tmp))
3269 val += PENALTY_FOR_NODE_WITH_CPUS;
3271 /* Slight preference for less loaded node */
3272 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3273 val += node_load[n];
3275 if (val < min_val) {
3282 node_set(best_node, *used_node_mask);
3289 * Build zonelists ordered by node and zones within node.
3290 * This results in maximum locality--normal zone overflows into local
3291 * DMA zone, if any--but risks exhausting DMA zone.
3293 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3296 struct zonelist *zonelist;
3298 zonelist = &pgdat->node_zonelists[0];
3299 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3301 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3303 zonelist->_zonerefs[j].zone = NULL;
3304 zonelist->_zonerefs[j].zone_idx = 0;
3308 * Build gfp_thisnode zonelists
3310 static void build_thisnode_zonelists(pg_data_t *pgdat)
3313 struct zonelist *zonelist;
3315 zonelist = &pgdat->node_zonelists[1];
3316 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3317 zonelist->_zonerefs[j].zone = NULL;
3318 zonelist->_zonerefs[j].zone_idx = 0;
3322 * Build zonelists ordered by zone and nodes within zones.
3323 * This results in conserving DMA zone[s] until all Normal memory is
3324 * exhausted, but results in overflowing to remote node while memory
3325 * may still exist in local DMA zone.
3327 static int node_order[MAX_NUMNODES];
3329 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3332 int zone_type; /* needs to be signed */
3334 struct zonelist *zonelist;
3336 zonelist = &pgdat->node_zonelists[0];
3338 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3339 for (j = 0; j < nr_nodes; j++) {
3340 node = node_order[j];
3341 z = &NODE_DATA(node)->node_zones[zone_type];
3342 if (populated_zone(z)) {
3344 &zonelist->_zonerefs[pos++]);
3345 check_highest_zone(zone_type);
3349 zonelist->_zonerefs[pos].zone = NULL;
3350 zonelist->_zonerefs[pos].zone_idx = 0;
3353 static int default_zonelist_order(void)
3356 unsigned long low_kmem_size,total_size;
3360 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3361 * If they are really small and used heavily, the system can fall
3362 * into OOM very easily.
3363 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3365 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3368 for_each_online_node(nid) {
3369 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3370 z = &NODE_DATA(nid)->node_zones[zone_type];
3371 if (populated_zone(z)) {
3372 if (zone_type < ZONE_NORMAL)
3373 low_kmem_size += z->present_pages;
3374 total_size += z->present_pages;
3375 } else if (zone_type == ZONE_NORMAL) {
3377 * If any node has only lowmem, then node order
3378 * is preferred to allow kernel allocations
3379 * locally; otherwise, they can easily infringe
3380 * on other nodes when there is an abundance of
3381 * lowmem available to allocate from.
3383 return ZONELIST_ORDER_NODE;
3387 if (!low_kmem_size || /* there are no DMA area. */
3388 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3389 return ZONELIST_ORDER_NODE;
3391 * look into each node's config.
3392 * If there is a node whose DMA/DMA32 memory is very big area on
3393 * local memory, NODE_ORDER may be suitable.
3395 average_size = total_size /
3396 (nodes_weight(node_states[N_MEMORY]) + 1);
3397 for_each_online_node(nid) {
3400 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3401 z = &NODE_DATA(nid)->node_zones[zone_type];
3402 if (populated_zone(z)) {
3403 if (zone_type < ZONE_NORMAL)
3404 low_kmem_size += z->present_pages;
3405 total_size += z->present_pages;
3408 if (low_kmem_size &&
3409 total_size > average_size && /* ignore small node */
3410 low_kmem_size > total_size * 70/100)
3411 return ZONELIST_ORDER_NODE;
3413 return ZONELIST_ORDER_ZONE;
3416 static void set_zonelist_order(void)
3418 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3419 current_zonelist_order = default_zonelist_order();
3421 current_zonelist_order = user_zonelist_order;
3424 static void build_zonelists(pg_data_t *pgdat)
3428 nodemask_t used_mask;
3429 int local_node, prev_node;
3430 struct zonelist *zonelist;
3431 int order = current_zonelist_order;
3433 /* initialize zonelists */
3434 for (i = 0; i < MAX_ZONELISTS; i++) {
3435 zonelist = pgdat->node_zonelists + i;
3436 zonelist->_zonerefs[0].zone = NULL;
3437 zonelist->_zonerefs[0].zone_idx = 0;
3440 /* NUMA-aware ordering of nodes */
3441 local_node = pgdat->node_id;
3442 load = nr_online_nodes;
3443 prev_node = local_node;
3444 nodes_clear(used_mask);
3446 memset(node_order, 0, sizeof(node_order));
3449 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3451 * We don't want to pressure a particular node.
3452 * So adding penalty to the first node in same
3453 * distance group to make it round-robin.
3455 if (node_distance(local_node, node) !=
3456 node_distance(local_node, prev_node))
3457 node_load[node] = load;
3461 if (order == ZONELIST_ORDER_NODE)
3462 build_zonelists_in_node_order(pgdat, node);
3464 node_order[j++] = node; /* remember order */
3467 if (order == ZONELIST_ORDER_ZONE) {
3468 /* calculate node order -- i.e., DMA last! */
3469 build_zonelists_in_zone_order(pgdat, j);
3472 build_thisnode_zonelists(pgdat);
3475 /* Construct the zonelist performance cache - see further mmzone.h */
3476 static void build_zonelist_cache(pg_data_t *pgdat)
3478 struct zonelist *zonelist;
3479 struct zonelist_cache *zlc;
3482 zonelist = &pgdat->node_zonelists[0];
3483 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3484 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3485 for (z = zonelist->_zonerefs; z->zone; z++)
3486 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3489 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3491 * Return node id of node used for "local" allocations.
3492 * I.e., first node id of first zone in arg node's generic zonelist.
3493 * Used for initializing percpu 'numa_mem', which is used primarily
3494 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3496 int local_memory_node(int node)
3500 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3501 gfp_zone(GFP_KERNEL),
3508 #else /* CONFIG_NUMA */
3510 static void set_zonelist_order(void)
3512 current_zonelist_order = ZONELIST_ORDER_ZONE;
3515 static void build_zonelists(pg_data_t *pgdat)
3517 int node, local_node;
3519 struct zonelist *zonelist;
3521 local_node = pgdat->node_id;
3523 zonelist = &pgdat->node_zonelists[0];
3524 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3527 * Now we build the zonelist so that it contains the zones
3528 * of all the other nodes.
3529 * We don't want to pressure a particular node, so when
3530 * building the zones for node N, we make sure that the
3531 * zones coming right after the local ones are those from
3532 * node N+1 (modulo N)
3534 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3535 if (!node_online(node))
3537 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3540 for (node = 0; node < local_node; node++) {
3541 if (!node_online(node))
3543 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3547 zonelist->_zonerefs[j].zone = NULL;
3548 zonelist->_zonerefs[j].zone_idx = 0;
3551 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3552 static void build_zonelist_cache(pg_data_t *pgdat)
3554 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3557 #endif /* CONFIG_NUMA */
3560 * Boot pageset table. One per cpu which is going to be used for all
3561 * zones and all nodes. The parameters will be set in such a way
3562 * that an item put on a list will immediately be handed over to
3563 * the buddy list. This is safe since pageset manipulation is done
3564 * with interrupts disabled.
3566 * The boot_pagesets must be kept even after bootup is complete for
3567 * unused processors and/or zones. They do play a role for bootstrapping
3568 * hotplugged processors.
3570 * zoneinfo_show() and maybe other functions do
3571 * not check if the processor is online before following the pageset pointer.
3572 * Other parts of the kernel may not check if the zone is available.
3574 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3575 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3576 static void setup_zone_pageset(struct zone *zone);
3579 * Global mutex to protect against size modification of zonelists
3580 * as well as to serialize pageset setup for the new populated zone.
3582 DEFINE_MUTEX(zonelists_mutex);
3584 /* return values int ....just for stop_machine() */
3585 static int __build_all_zonelists(void *data)
3589 pg_data_t *self = data;
3592 memset(node_load, 0, sizeof(node_load));
3595 if (self && !node_online(self->node_id)) {
3596 build_zonelists(self);
3597 build_zonelist_cache(self);
3600 for_each_online_node(nid) {
3601 pg_data_t *pgdat = NODE_DATA(nid);
3603 build_zonelists(pgdat);
3604 build_zonelist_cache(pgdat);
3608 * Initialize the boot_pagesets that are going to be used
3609 * for bootstrapping processors. The real pagesets for
3610 * each zone will be allocated later when the per cpu
3611 * allocator is available.
3613 * boot_pagesets are used also for bootstrapping offline
3614 * cpus if the system is already booted because the pagesets
3615 * are needed to initialize allocators on a specific cpu too.
3616 * F.e. the percpu allocator needs the page allocator which
3617 * needs the percpu allocator in order to allocate its pagesets
3618 * (a chicken-egg dilemma).
3620 for_each_possible_cpu(cpu) {
3621 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3623 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3625 * We now know the "local memory node" for each node--
3626 * i.e., the node of the first zone in the generic zonelist.
3627 * Set up numa_mem percpu variable for on-line cpus. During
3628 * boot, only the boot cpu should be on-line; we'll init the
3629 * secondary cpus' numa_mem as they come on-line. During
3630 * node/memory hotplug, we'll fixup all on-line cpus.
3632 if (cpu_online(cpu))
3633 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3641 * Called with zonelists_mutex held always
3642 * unless system_state == SYSTEM_BOOTING.
3644 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3646 set_zonelist_order();
3648 if (system_state == SYSTEM_BOOTING) {
3649 __build_all_zonelists(NULL);
3650 mminit_verify_zonelist();
3651 cpuset_init_current_mems_allowed();
3653 /* we have to stop all cpus to guarantee there is no user
3655 #ifdef CONFIG_MEMORY_HOTPLUG
3657 setup_zone_pageset(zone);
3659 stop_machine(__build_all_zonelists, pgdat, NULL);
3660 /* cpuset refresh routine should be here */
3662 vm_total_pages = nr_free_pagecache_pages();
3664 * Disable grouping by mobility if the number of pages in the
3665 * system is too low to allow the mechanism to work. It would be
3666 * more accurate, but expensive to check per-zone. This check is
3667 * made on memory-hotadd so a system can start with mobility
3668 * disabled and enable it later
3670 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3671 page_group_by_mobility_disabled = 1;
3673 page_group_by_mobility_disabled = 0;
3675 printk("Built %i zonelists in %s order, mobility grouping %s. "
3676 "Total pages: %ld\n",
3678 zonelist_order_name[current_zonelist_order],
3679 page_group_by_mobility_disabled ? "off" : "on",
3682 printk("Policy zone: %s\n", zone_names[policy_zone]);
3687 * Helper functions to size the waitqueue hash table.
3688 * Essentially these want to choose hash table sizes sufficiently
3689 * large so that collisions trying to wait on pages are rare.
3690 * But in fact, the number of active page waitqueues on typical
3691 * systems is ridiculously low, less than 200. So this is even
3692 * conservative, even though it seems large.
3694 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3695 * waitqueues, i.e. the size of the waitq table given the number of pages.
3697 #define PAGES_PER_WAITQUEUE 256
3699 #ifndef CONFIG_MEMORY_HOTPLUG
3700 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3702 unsigned long size = 1;
3704 pages /= PAGES_PER_WAITQUEUE;
3706 while (size < pages)
3710 * Once we have dozens or even hundreds of threads sleeping
3711 * on IO we've got bigger problems than wait queue collision.
3712 * Limit the size of the wait table to a reasonable size.
3714 size = min(size, 4096UL);
3716 return max(size, 4UL);
3720 * A zone's size might be changed by hot-add, so it is not possible to determine
3721 * a suitable size for its wait_table. So we use the maximum size now.
3723 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3725 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3726 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3727 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3729 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3730 * or more by the traditional way. (See above). It equals:
3732 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3733 * ia64(16K page size) : = ( 8G + 4M)byte.
3734 * powerpc (64K page size) : = (32G +16M)byte.
3736 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3743 * This is an integer logarithm so that shifts can be used later
3744 * to extract the more random high bits from the multiplicative
3745 * hash function before the remainder is taken.
3747 static inline unsigned long wait_table_bits(unsigned long size)
3752 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3755 * Check if a pageblock contains reserved pages
3757 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3761 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3762 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3769 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3770 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3771 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3772 * higher will lead to a bigger reserve which will get freed as contiguous
3773 * blocks as reclaim kicks in
3775 static void setup_zone_migrate_reserve(struct zone *zone)
3777 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3779 unsigned long block_migratetype;
3783 * Get the start pfn, end pfn and the number of blocks to reserve
3784 * We have to be careful to be aligned to pageblock_nr_pages to
3785 * make sure that we always check pfn_valid for the first page in
3788 start_pfn = zone->zone_start_pfn;
3789 end_pfn = start_pfn + zone->spanned_pages;
3790 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3791 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3795 * Reserve blocks are generally in place to help high-order atomic
3796 * allocations that are short-lived. A min_free_kbytes value that
3797 * would result in more than 2 reserve blocks for atomic allocations
3798 * is assumed to be in place to help anti-fragmentation for the
3799 * future allocation of hugepages at runtime.
3801 reserve = min(2, reserve);
3803 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3804 if (!pfn_valid(pfn))
3806 page = pfn_to_page(pfn);
3808 /* Watch out for overlapping nodes */
3809 if (page_to_nid(page) != zone_to_nid(zone))
3812 block_migratetype = get_pageblock_migratetype(page);
3814 /* Only test what is necessary when the reserves are not met */
3817 * Blocks with reserved pages will never free, skip
3820 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3821 if (pageblock_is_reserved(pfn, block_end_pfn))
3824 /* If this block is reserved, account for it */
3825 if (block_migratetype == MIGRATE_RESERVE) {
3830 /* Suitable for reserving if this block is movable */
3831 if (block_migratetype == MIGRATE_MOVABLE) {
3832 set_pageblock_migratetype(page,
3834 move_freepages_block(zone, page,
3842 * If the reserve is met and this is a previous reserved block,
3845 if (block_migratetype == MIGRATE_RESERVE) {
3846 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3847 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3853 * Initially all pages are reserved - free ones are freed
3854 * up by free_all_bootmem() once the early boot process is
3855 * done. Non-atomic initialization, single-pass.
3857 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3858 unsigned long start_pfn, enum memmap_context context)
3861 unsigned long end_pfn = start_pfn + size;
3865 if (highest_memmap_pfn < end_pfn - 1)
3866 highest_memmap_pfn = end_pfn - 1;
3868 z = &NODE_DATA(nid)->node_zones[zone];
3869 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3871 * There can be holes in boot-time mem_map[]s
3872 * handed to this function. They do not
3873 * exist on hotplugged memory.
3875 if (context == MEMMAP_EARLY) {
3876 if (!early_pfn_valid(pfn))
3878 if (!early_pfn_in_nid(pfn, nid))
3881 page = pfn_to_page(pfn);
3882 set_page_links(page, zone, nid, pfn);
3883 mminit_verify_page_links(page, zone, nid, pfn);
3884 init_page_count(page);
3885 reset_page_mapcount(page);
3886 SetPageReserved(page);
3888 * Mark the block movable so that blocks are reserved for
3889 * movable at startup. This will force kernel allocations
3890 * to reserve their blocks rather than leaking throughout
3891 * the address space during boot when many long-lived
3892 * kernel allocations are made. Later some blocks near
3893 * the start are marked MIGRATE_RESERVE by
3894 * setup_zone_migrate_reserve()
3896 * bitmap is created for zone's valid pfn range. but memmap
3897 * can be created for invalid pages (for alignment)
3898 * check here not to call set_pageblock_migratetype() against
3901 if ((z->zone_start_pfn <= pfn)
3902 && (pfn < z->zone_start_pfn + z->spanned_pages)
3903 && !(pfn & (pageblock_nr_pages - 1)))
3904 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3906 INIT_LIST_HEAD(&page->lru);
3907 #ifdef WANT_PAGE_VIRTUAL
3908 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3909 if (!is_highmem_idx(zone))
3910 set_page_address(page, __va(pfn << PAGE_SHIFT));
3915 static void __meminit zone_init_free_lists(struct zone *zone)
3918 for_each_migratetype_order(order, t) {
3919 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3920 zone->free_area[order].nr_free = 0;
3924 #ifndef __HAVE_ARCH_MEMMAP_INIT
3925 #define memmap_init(size, nid, zone, start_pfn) \
3926 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3929 static int __meminit zone_batchsize(struct zone *zone)
3935 * The per-cpu-pages pools are set to around 1000th of the
3936 * size of the zone. But no more than 1/2 of a meg.
3938 * OK, so we don't know how big the cache is. So guess.
3940 batch = zone->present_pages / 1024;
3941 if (batch * PAGE_SIZE > 512 * 1024)
3942 batch = (512 * 1024) / PAGE_SIZE;
3943 batch /= 4; /* We effectively *= 4 below */
3948 * Clamp the batch to a 2^n - 1 value. Having a power
3949 * of 2 value was found to be more likely to have
3950 * suboptimal cache aliasing properties in some cases.
3952 * For example if 2 tasks are alternately allocating
3953 * batches of pages, one task can end up with a lot
3954 * of pages of one half of the possible page colors
3955 * and the other with pages of the other colors.
3957 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3962 /* The deferral and batching of frees should be suppressed under NOMMU
3965 * The problem is that NOMMU needs to be able to allocate large chunks
3966 * of contiguous memory as there's no hardware page translation to
3967 * assemble apparent contiguous memory from discontiguous pages.
3969 * Queueing large contiguous runs of pages for batching, however,
3970 * causes the pages to actually be freed in smaller chunks. As there
3971 * can be a significant delay between the individual batches being
3972 * recycled, this leads to the once large chunks of space being
3973 * fragmented and becoming unavailable for high-order allocations.
3979 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3981 struct per_cpu_pages *pcp;
3984 memset(p, 0, sizeof(*p));
3988 pcp->high = 6 * batch;
3989 pcp->batch = max(1UL, 1 * batch);
3990 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3991 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3995 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3996 * to the value high for the pageset p.
3999 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
4002 struct per_cpu_pages *pcp;
4006 pcp->batch = max(1UL, high/4);
4007 if ((high/4) > (PAGE_SHIFT * 8))
4008 pcp->batch = PAGE_SHIFT * 8;
4011 static void __meminit setup_zone_pageset(struct zone *zone)
4015 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4017 for_each_possible_cpu(cpu) {
4018 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4020 setup_pageset(pcp, zone_batchsize(zone));
4022 if (percpu_pagelist_fraction)
4023 setup_pagelist_highmark(pcp,
4024 (zone->present_pages /
4025 percpu_pagelist_fraction));
4030 * Allocate per cpu pagesets and initialize them.
4031 * Before this call only boot pagesets were available.
4033 void __init setup_per_cpu_pageset(void)
4037 for_each_populated_zone(zone)
4038 setup_zone_pageset(zone);
4041 static noinline __init_refok
4042 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4045 struct pglist_data *pgdat = zone->zone_pgdat;
4049 * The per-page waitqueue mechanism uses hashed waitqueues
4052 zone->wait_table_hash_nr_entries =
4053 wait_table_hash_nr_entries(zone_size_pages);
4054 zone->wait_table_bits =
4055 wait_table_bits(zone->wait_table_hash_nr_entries);
4056 alloc_size = zone->wait_table_hash_nr_entries
4057 * sizeof(wait_queue_head_t);
4059 if (!slab_is_available()) {
4060 zone->wait_table = (wait_queue_head_t *)
4061 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4064 * This case means that a zone whose size was 0 gets new memory
4065 * via memory hot-add.
4066 * But it may be the case that a new node was hot-added. In
4067 * this case vmalloc() will not be able to use this new node's
4068 * memory - this wait_table must be initialized to use this new
4069 * node itself as well.
4070 * To use this new node's memory, further consideration will be
4073 zone->wait_table = vmalloc(alloc_size);
4075 if (!zone->wait_table)
4078 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4079 init_waitqueue_head(zone->wait_table + i);
4084 static __meminit void zone_pcp_init(struct zone *zone)
4087 * per cpu subsystem is not up at this point. The following code
4088 * relies on the ability of the linker to provide the
4089 * offset of a (static) per cpu variable into the per cpu area.
4091 zone->pageset = &boot_pageset;
4093 if (zone->present_pages)
4094 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4095 zone->name, zone->present_pages,
4096 zone_batchsize(zone));
4099 int __meminit init_currently_empty_zone(struct zone *zone,
4100 unsigned long zone_start_pfn,
4102 enum memmap_context context)
4104 struct pglist_data *pgdat = zone->zone_pgdat;
4106 ret = zone_wait_table_init(zone, size);
4109 pgdat->nr_zones = zone_idx(zone) + 1;
4111 zone->zone_start_pfn = zone_start_pfn;
4113 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4114 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4116 (unsigned long)zone_idx(zone),
4117 zone_start_pfn, (zone_start_pfn + size));
4119 zone_init_free_lists(zone);
4124 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4125 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4127 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4128 * Architectures may implement their own version but if add_active_range()
4129 * was used and there are no special requirements, this is a convenient
4132 int __meminit __early_pfn_to_nid(unsigned long pfn)
4134 unsigned long start_pfn, end_pfn;
4137 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4138 if (start_pfn <= pfn && pfn < end_pfn)
4140 /* This is a memory hole */
4143 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4145 int __meminit early_pfn_to_nid(unsigned long pfn)
4149 nid = __early_pfn_to_nid(pfn);
4152 /* just returns 0 */
4156 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4157 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4161 nid = __early_pfn_to_nid(pfn);
4162 if (nid >= 0 && nid != node)
4169 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4170 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4171 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4173 * If an architecture guarantees that all ranges registered with
4174 * add_active_ranges() contain no holes and may be freed, this
4175 * this function may be used instead of calling free_bootmem() manually.
4177 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4179 unsigned long start_pfn, end_pfn;
4182 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4183 start_pfn = min(start_pfn, max_low_pfn);
4184 end_pfn = min(end_pfn, max_low_pfn);
4186 if (start_pfn < end_pfn)
4187 free_bootmem_node(NODE_DATA(this_nid),
4188 PFN_PHYS(start_pfn),
4189 (end_pfn - start_pfn) << PAGE_SHIFT);
4194 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4195 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4197 * If an architecture guarantees that all ranges registered with
4198 * add_active_ranges() contain no holes and may be freed, this
4199 * function may be used instead of calling memory_present() manually.
4201 void __init sparse_memory_present_with_active_regions(int nid)
4203 unsigned long start_pfn, end_pfn;
4206 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4207 memory_present(this_nid, start_pfn, end_pfn);
4211 * get_pfn_range_for_nid - Return the start and end page frames for a node
4212 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4213 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4214 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4216 * It returns the start and end page frame of a node based on information
4217 * provided by an arch calling add_active_range(). If called for a node
4218 * with no available memory, a warning is printed and the start and end
4221 void __meminit get_pfn_range_for_nid(unsigned int nid,
4222 unsigned long *start_pfn, unsigned long *end_pfn)
4224 unsigned long this_start_pfn, this_end_pfn;
4230 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4231 *start_pfn = min(*start_pfn, this_start_pfn);
4232 *end_pfn = max(*end_pfn, this_end_pfn);
4235 if (*start_pfn == -1UL)
4240 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4241 * assumption is made that zones within a node are ordered in monotonic
4242 * increasing memory addresses so that the "highest" populated zone is used
4244 static void __init find_usable_zone_for_movable(void)
4247 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4248 if (zone_index == ZONE_MOVABLE)
4251 if (arch_zone_highest_possible_pfn[zone_index] >
4252 arch_zone_lowest_possible_pfn[zone_index])
4256 VM_BUG_ON(zone_index == -1);
4257 movable_zone = zone_index;
4261 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4262 * because it is sized independent of architecture. Unlike the other zones,
4263 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4264 * in each node depending on the size of each node and how evenly kernelcore
4265 * is distributed. This helper function adjusts the zone ranges
4266 * provided by the architecture for a given node by using the end of the
4267 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4268 * zones within a node are in order of monotonic increases memory addresses
4270 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4271 unsigned long zone_type,
4272 unsigned long node_start_pfn,
4273 unsigned long node_end_pfn,
4274 unsigned long *zone_start_pfn,
4275 unsigned long *zone_end_pfn)
4277 /* Only adjust if ZONE_MOVABLE is on this node */
4278 if (zone_movable_pfn[nid]) {
4279 /* Size ZONE_MOVABLE */
4280 if (zone_type == ZONE_MOVABLE) {
4281 *zone_start_pfn = zone_movable_pfn[nid];
4282 *zone_end_pfn = min(node_end_pfn,
4283 arch_zone_highest_possible_pfn[movable_zone]);
4285 /* Adjust for ZONE_MOVABLE starting within this range */
4286 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4287 *zone_end_pfn > zone_movable_pfn[nid]) {
4288 *zone_end_pfn = zone_movable_pfn[nid];
4290 /* Check if this whole range is within ZONE_MOVABLE */
4291 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4292 *zone_start_pfn = *zone_end_pfn;
4297 * Return the number of pages a zone spans in a node, including holes
4298 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4300 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4301 unsigned long zone_type,
4302 unsigned long *ignored)
4304 unsigned long node_start_pfn, node_end_pfn;
4305 unsigned long zone_start_pfn, zone_end_pfn;
4307 /* Get the start and end of the node and zone */
4308 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4309 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4310 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4311 adjust_zone_range_for_zone_movable(nid, zone_type,
4312 node_start_pfn, node_end_pfn,
4313 &zone_start_pfn, &zone_end_pfn);
4315 /* Check that this node has pages within the zone's required range */
4316 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4319 /* Move the zone boundaries inside the node if necessary */
4320 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4321 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4323 /* Return the spanned pages */
4324 return zone_end_pfn - zone_start_pfn;
4328 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4329 * then all holes in the requested range will be accounted for.
4331 unsigned long __meminit __absent_pages_in_range(int nid,
4332 unsigned long range_start_pfn,
4333 unsigned long range_end_pfn)
4335 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4336 unsigned long start_pfn, end_pfn;
4339 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4340 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4341 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4342 nr_absent -= end_pfn - start_pfn;
4348 * absent_pages_in_range - Return number of page frames in holes within a range
4349 * @start_pfn: The start PFN to start searching for holes
4350 * @end_pfn: The end PFN to stop searching for holes
4352 * It returns the number of pages frames in memory holes within a range.
4354 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4355 unsigned long end_pfn)
4357 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4360 /* Return the number of page frames in holes in a zone on a node */
4361 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4362 unsigned long zone_type,
4363 unsigned long *ignored)
4365 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4366 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4367 unsigned long node_start_pfn, node_end_pfn;
4368 unsigned long zone_start_pfn, zone_end_pfn;
4370 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4371 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4372 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4374 adjust_zone_range_for_zone_movable(nid, zone_type,
4375 node_start_pfn, node_end_pfn,
4376 &zone_start_pfn, &zone_end_pfn);
4377 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4380 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4381 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4382 unsigned long zone_type,
4383 unsigned long *zones_size)
4385 return zones_size[zone_type];
4388 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4389 unsigned long zone_type,
4390 unsigned long *zholes_size)
4395 return zholes_size[zone_type];
4398 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4400 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4401 unsigned long *zones_size, unsigned long *zholes_size)
4403 unsigned long realtotalpages, totalpages = 0;
4406 for (i = 0; i < MAX_NR_ZONES; i++)
4407 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4409 pgdat->node_spanned_pages = totalpages;
4411 realtotalpages = totalpages;
4412 for (i = 0; i < MAX_NR_ZONES; i++)
4414 zone_absent_pages_in_node(pgdat->node_id, i,
4416 pgdat->node_present_pages = realtotalpages;
4417 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4421 #ifndef CONFIG_SPARSEMEM
4423 * Calculate the size of the zone->blockflags rounded to an unsigned long
4424 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4425 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4426 * round what is now in bits to nearest long in bits, then return it in
4429 static unsigned long __init usemap_size(unsigned long zonesize)
4431 unsigned long usemapsize;
4433 usemapsize = roundup(zonesize, pageblock_nr_pages);
4434 usemapsize = usemapsize >> pageblock_order;
4435 usemapsize *= NR_PAGEBLOCK_BITS;
4436 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4438 return usemapsize / 8;
4441 static void __init setup_usemap(struct pglist_data *pgdat,
4442 struct zone *zone, unsigned long zonesize)
4444 unsigned long usemapsize = usemap_size(zonesize);
4445 zone->pageblock_flags = NULL;
4447 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4451 static inline void setup_usemap(struct pglist_data *pgdat,
4452 struct zone *zone, unsigned long zonesize) {}
4453 #endif /* CONFIG_SPARSEMEM */
4455 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4457 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4458 void __init set_pageblock_order(void)
4462 /* Check that pageblock_nr_pages has not already been setup */
4463 if (pageblock_order)
4466 if (HPAGE_SHIFT > PAGE_SHIFT)
4467 order = HUGETLB_PAGE_ORDER;
4469 order = MAX_ORDER - 1;
4472 * Assume the largest contiguous order of interest is a huge page.
4473 * This value may be variable depending on boot parameters on IA64 and
4476 pageblock_order = order;
4478 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4481 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4482 * is unused as pageblock_order is set at compile-time. See
4483 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4486 void __init set_pageblock_order(void)
4490 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4493 * Set up the zone data structures:
4494 * - mark all pages reserved
4495 * - mark all memory queues empty
4496 * - clear the memory bitmaps
4498 * NOTE: pgdat should get zeroed by caller.
4500 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4501 unsigned long *zones_size, unsigned long *zholes_size)
4504 int nid = pgdat->node_id;
4505 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4508 pgdat_resize_init(pgdat);
4509 init_waitqueue_head(&pgdat->kswapd_wait);
4510 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4511 pgdat_page_cgroup_init(pgdat);
4513 for (j = 0; j < MAX_NR_ZONES; j++) {
4514 struct zone *zone = pgdat->node_zones + j;
4515 unsigned long size, realsize, freesize, memmap_pages;
4517 size = zone_spanned_pages_in_node(nid, j, zones_size);
4518 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4522 * Adjust freesize so that it accounts for how much memory
4523 * is used by this zone for memmap. This affects the watermark
4524 * and per-cpu initialisations
4527 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4528 if (freesize >= memmap_pages) {
4529 freesize -= memmap_pages;
4532 " %s zone: %lu pages used for memmap\n",
4533 zone_names[j], memmap_pages);
4536 " %s zone: %lu pages exceeds freesize %lu\n",
4537 zone_names[j], memmap_pages, freesize);
4539 /* Account for reserved pages */
4540 if (j == 0 && freesize > dma_reserve) {
4541 freesize -= dma_reserve;
4542 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4543 zone_names[0], dma_reserve);
4546 if (!is_highmem_idx(j))
4547 nr_kernel_pages += freesize;
4548 nr_all_pages += freesize;
4550 zone->spanned_pages = size;
4551 zone->present_pages = freesize;
4553 * Set an approximate value for lowmem here, it will be adjusted
4554 * when the bootmem allocator frees pages into the buddy system.
4555 * And all highmem pages will be managed by the buddy system.
4557 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4560 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4562 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4564 zone->name = zone_names[j];
4565 spin_lock_init(&zone->lock);
4566 spin_lock_init(&zone->lru_lock);
4567 zone_seqlock_init(zone);
4568 zone->zone_pgdat = pgdat;
4570 zone_pcp_init(zone);
4571 lruvec_init(&zone->lruvec);
4575 set_pageblock_order();
4576 setup_usemap(pgdat, zone, size);
4577 ret = init_currently_empty_zone(zone, zone_start_pfn,
4578 size, MEMMAP_EARLY);
4580 memmap_init(size, nid, j, zone_start_pfn);
4581 zone_start_pfn += size;
4585 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4587 /* Skip empty nodes */
4588 if (!pgdat->node_spanned_pages)
4591 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4592 /* ia64 gets its own node_mem_map, before this, without bootmem */
4593 if (!pgdat->node_mem_map) {
4594 unsigned long size, start, end;
4598 * The zone's endpoints aren't required to be MAX_ORDER
4599 * aligned but the node_mem_map endpoints must be in order
4600 * for the buddy allocator to function correctly.
4602 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4603 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4604 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4605 size = (end - start) * sizeof(struct page);
4606 map = alloc_remap(pgdat->node_id, size);
4608 map = alloc_bootmem_node_nopanic(pgdat, size);
4609 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4611 #ifndef CONFIG_NEED_MULTIPLE_NODES
4613 * With no DISCONTIG, the global mem_map is just set as node 0's
4615 if (pgdat == NODE_DATA(0)) {
4616 mem_map = NODE_DATA(0)->node_mem_map;
4617 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4618 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4619 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4620 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4623 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4626 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4627 unsigned long node_start_pfn, unsigned long *zholes_size)
4629 pg_data_t *pgdat = NODE_DATA(nid);
4631 /* pg_data_t should be reset to zero when it's allocated */
4632 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4634 pgdat->node_id = nid;
4635 pgdat->node_start_pfn = node_start_pfn;
4636 init_zone_allows_reclaim(nid);
4637 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4639 alloc_node_mem_map(pgdat);
4640 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4641 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4642 nid, (unsigned long)pgdat,
4643 (unsigned long)pgdat->node_mem_map);
4646 free_area_init_core(pgdat, zones_size, zholes_size);
4649 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4651 #if MAX_NUMNODES > 1
4653 * Figure out the number of possible node ids.
4655 static void __init setup_nr_node_ids(void)
4658 unsigned int highest = 0;
4660 for_each_node_mask(node, node_possible_map)
4662 nr_node_ids = highest + 1;
4665 static inline void setup_nr_node_ids(void)
4671 * node_map_pfn_alignment - determine the maximum internode alignment
4673 * This function should be called after node map is populated and sorted.
4674 * It calculates the maximum power of two alignment which can distinguish
4677 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4678 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4679 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4680 * shifted, 1GiB is enough and this function will indicate so.
4682 * This is used to test whether pfn -> nid mapping of the chosen memory
4683 * model has fine enough granularity to avoid incorrect mapping for the
4684 * populated node map.
4686 * Returns the determined alignment in pfn's. 0 if there is no alignment
4687 * requirement (single node).
4689 unsigned long __init node_map_pfn_alignment(void)
4691 unsigned long accl_mask = 0, last_end = 0;
4692 unsigned long start, end, mask;
4696 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4697 if (!start || last_nid < 0 || last_nid == nid) {
4704 * Start with a mask granular enough to pin-point to the
4705 * start pfn and tick off bits one-by-one until it becomes
4706 * too coarse to separate the current node from the last.
4708 mask = ~((1 << __ffs(start)) - 1);
4709 while (mask && last_end <= (start & (mask << 1)))
4712 /* accumulate all internode masks */
4716 /* convert mask to number of pages */
4717 return ~accl_mask + 1;
4720 /* Find the lowest pfn for a node */
4721 static unsigned long __init find_min_pfn_for_node(int nid)
4723 unsigned long min_pfn = ULONG_MAX;
4724 unsigned long start_pfn;
4727 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4728 min_pfn = min(min_pfn, start_pfn);
4730 if (min_pfn == ULONG_MAX) {
4732 "Could not find start_pfn for node %d\n", nid);
4740 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4742 * It returns the minimum PFN based on information provided via
4743 * add_active_range().
4745 unsigned long __init find_min_pfn_with_active_regions(void)
4747 return find_min_pfn_for_node(MAX_NUMNODES);
4751 * early_calculate_totalpages()
4752 * Sum pages in active regions for movable zone.
4753 * Populate N_MEMORY for calculating usable_nodes.
4755 static unsigned long __init early_calculate_totalpages(void)
4757 unsigned long totalpages = 0;
4758 unsigned long start_pfn, end_pfn;
4761 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4762 unsigned long pages = end_pfn - start_pfn;
4764 totalpages += pages;
4766 node_set_state(nid, N_MEMORY);
4772 * Find the PFN the Movable zone begins in each node. Kernel memory
4773 * is spread evenly between nodes as long as the nodes have enough
4774 * memory. When they don't, some nodes will have more kernelcore than
4777 static void __init find_zone_movable_pfns_for_nodes(void)
4780 unsigned long usable_startpfn;
4781 unsigned long kernelcore_node, kernelcore_remaining;
4782 /* save the state before borrow the nodemask */
4783 nodemask_t saved_node_state = node_states[N_MEMORY];
4784 unsigned long totalpages = early_calculate_totalpages();
4785 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4788 * If movablecore was specified, calculate what size of
4789 * kernelcore that corresponds so that memory usable for
4790 * any allocation type is evenly spread. If both kernelcore
4791 * and movablecore are specified, then the value of kernelcore
4792 * will be used for required_kernelcore if it's greater than
4793 * what movablecore would have allowed.
4795 if (required_movablecore) {
4796 unsigned long corepages;
4799 * Round-up so that ZONE_MOVABLE is at least as large as what
4800 * was requested by the user
4802 required_movablecore =
4803 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4804 corepages = totalpages - required_movablecore;
4806 required_kernelcore = max(required_kernelcore, corepages);
4809 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4810 if (!required_kernelcore)
4813 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4814 find_usable_zone_for_movable();
4815 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4818 /* Spread kernelcore memory as evenly as possible throughout nodes */
4819 kernelcore_node = required_kernelcore / usable_nodes;
4820 for_each_node_state(nid, N_MEMORY) {
4821 unsigned long start_pfn, end_pfn;
4824 * Recalculate kernelcore_node if the division per node
4825 * now exceeds what is necessary to satisfy the requested
4826 * amount of memory for the kernel
4828 if (required_kernelcore < kernelcore_node)
4829 kernelcore_node = required_kernelcore / usable_nodes;
4832 * As the map is walked, we track how much memory is usable
4833 * by the kernel using kernelcore_remaining. When it is
4834 * 0, the rest of the node is usable by ZONE_MOVABLE
4836 kernelcore_remaining = kernelcore_node;
4838 /* Go through each range of PFNs within this node */
4839 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4840 unsigned long size_pages;
4842 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4843 if (start_pfn >= end_pfn)
4846 /* Account for what is only usable for kernelcore */
4847 if (start_pfn < usable_startpfn) {
4848 unsigned long kernel_pages;
4849 kernel_pages = min(end_pfn, usable_startpfn)
4852 kernelcore_remaining -= min(kernel_pages,
4853 kernelcore_remaining);
4854 required_kernelcore -= min(kernel_pages,
4855 required_kernelcore);
4857 /* Continue if range is now fully accounted */
4858 if (end_pfn <= usable_startpfn) {
4861 * Push zone_movable_pfn to the end so
4862 * that if we have to rebalance
4863 * kernelcore across nodes, we will
4864 * not double account here
4866 zone_movable_pfn[nid] = end_pfn;
4869 start_pfn = usable_startpfn;
4873 * The usable PFN range for ZONE_MOVABLE is from
4874 * start_pfn->end_pfn. Calculate size_pages as the
4875 * number of pages used as kernelcore
4877 size_pages = end_pfn - start_pfn;
4878 if (size_pages > kernelcore_remaining)
4879 size_pages = kernelcore_remaining;
4880 zone_movable_pfn[nid] = start_pfn + size_pages;
4883 * Some kernelcore has been met, update counts and
4884 * break if the kernelcore for this node has been
4887 required_kernelcore -= min(required_kernelcore,
4889 kernelcore_remaining -= size_pages;
4890 if (!kernelcore_remaining)
4896 * If there is still required_kernelcore, we do another pass with one
4897 * less node in the count. This will push zone_movable_pfn[nid] further
4898 * along on the nodes that still have memory until kernelcore is
4902 if (usable_nodes && required_kernelcore > usable_nodes)
4905 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4906 for (nid = 0; nid < MAX_NUMNODES; nid++)
4907 zone_movable_pfn[nid] =
4908 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4911 /* restore the node_state */
4912 node_states[N_MEMORY] = saved_node_state;
4915 /* Any regular or high memory on that node ? */
4916 static void check_for_memory(pg_data_t *pgdat, int nid)
4918 enum zone_type zone_type;
4920 if (N_MEMORY == N_NORMAL_MEMORY)
4923 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
4924 struct zone *zone = &pgdat->node_zones[zone_type];
4925 if (zone->present_pages) {
4926 node_set_state(nid, N_HIGH_MEMORY);
4927 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
4928 zone_type <= ZONE_NORMAL)
4929 node_set_state(nid, N_NORMAL_MEMORY);
4936 * free_area_init_nodes - Initialise all pg_data_t and zone data
4937 * @max_zone_pfn: an array of max PFNs for each zone
4939 * This will call free_area_init_node() for each active node in the system.
4940 * Using the page ranges provided by add_active_range(), the size of each
4941 * zone in each node and their holes is calculated. If the maximum PFN
4942 * between two adjacent zones match, it is assumed that the zone is empty.
4943 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4944 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4945 * starts where the previous one ended. For example, ZONE_DMA32 starts
4946 * at arch_max_dma_pfn.
4948 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4950 unsigned long start_pfn, end_pfn;
4953 /* Record where the zone boundaries are */
4954 memset(arch_zone_lowest_possible_pfn, 0,
4955 sizeof(arch_zone_lowest_possible_pfn));
4956 memset(arch_zone_highest_possible_pfn, 0,
4957 sizeof(arch_zone_highest_possible_pfn));
4958 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4959 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4960 for (i = 1; i < MAX_NR_ZONES; i++) {
4961 if (i == ZONE_MOVABLE)
4963 arch_zone_lowest_possible_pfn[i] =
4964 arch_zone_highest_possible_pfn[i-1];
4965 arch_zone_highest_possible_pfn[i] =
4966 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4968 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4969 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4971 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4972 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4973 find_zone_movable_pfns_for_nodes();
4975 /* Print out the zone ranges */
4976 printk("Zone ranges:\n");
4977 for (i = 0; i < MAX_NR_ZONES; i++) {
4978 if (i == ZONE_MOVABLE)
4980 printk(KERN_CONT " %-8s ", zone_names[i]);
4981 if (arch_zone_lowest_possible_pfn[i] ==
4982 arch_zone_highest_possible_pfn[i])
4983 printk(KERN_CONT "empty\n");
4985 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4986 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4987 (arch_zone_highest_possible_pfn[i]
4988 << PAGE_SHIFT) - 1);
4991 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4992 printk("Movable zone start for each node\n");
4993 for (i = 0; i < MAX_NUMNODES; i++) {
4994 if (zone_movable_pfn[i])
4995 printk(" Node %d: %#010lx\n", i,
4996 zone_movable_pfn[i] << PAGE_SHIFT);
4999 /* Print out the early node map */
5000 printk("Early memory node ranges\n");
5001 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5002 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5003 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5005 /* Initialise every node */
5006 mminit_verify_pageflags_layout();
5007 setup_nr_node_ids();
5008 for_each_online_node(nid) {
5009 pg_data_t *pgdat = NODE_DATA(nid);
5010 free_area_init_node(nid, NULL,
5011 find_min_pfn_for_node(nid), NULL);
5013 /* Any memory on that node */
5014 if (pgdat->node_present_pages)
5015 node_set_state(nid, N_MEMORY);
5016 check_for_memory(pgdat, nid);
5020 static int __init cmdline_parse_core(char *p, unsigned long *core)
5022 unsigned long long coremem;
5026 coremem = memparse(p, &p);
5027 *core = coremem >> PAGE_SHIFT;
5029 /* Paranoid check that UL is enough for the coremem value */
5030 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5036 * kernelcore=size sets the amount of memory for use for allocations that
5037 * cannot be reclaimed or migrated.
5039 static int __init cmdline_parse_kernelcore(char *p)
5041 return cmdline_parse_core(p, &required_kernelcore);
5045 * movablecore=size sets the amount of memory for use for allocations that
5046 * can be reclaimed or migrated.
5048 static int __init cmdline_parse_movablecore(char *p)
5050 return cmdline_parse_core(p, &required_movablecore);
5053 early_param("kernelcore", cmdline_parse_kernelcore);
5054 early_param("movablecore", cmdline_parse_movablecore);
5056 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5059 * set_dma_reserve - set the specified number of pages reserved in the first zone
5060 * @new_dma_reserve: The number of pages to mark reserved
5062 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5063 * In the DMA zone, a significant percentage may be consumed by kernel image
5064 * and other unfreeable allocations which can skew the watermarks badly. This
5065 * function may optionally be used to account for unfreeable pages in the
5066 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5067 * smaller per-cpu batchsize.
5069 void __init set_dma_reserve(unsigned long new_dma_reserve)
5071 dma_reserve = new_dma_reserve;
5074 void __init free_area_init(unsigned long *zones_size)
5076 free_area_init_node(0, zones_size,
5077 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5080 static int page_alloc_cpu_notify(struct notifier_block *self,
5081 unsigned long action, void *hcpu)
5083 int cpu = (unsigned long)hcpu;
5085 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5086 lru_add_drain_cpu(cpu);
5090 * Spill the event counters of the dead processor
5091 * into the current processors event counters.
5092 * This artificially elevates the count of the current
5095 vm_events_fold_cpu(cpu);
5098 * Zero the differential counters of the dead processor
5099 * so that the vm statistics are consistent.
5101 * This is only okay since the processor is dead and cannot
5102 * race with what we are doing.
5104 refresh_cpu_vm_stats(cpu);
5109 void __init page_alloc_init(void)
5111 hotcpu_notifier(page_alloc_cpu_notify, 0);
5115 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5116 * or min_free_kbytes changes.
5118 static void calculate_totalreserve_pages(void)
5120 struct pglist_data *pgdat;
5121 unsigned long reserve_pages = 0;
5122 enum zone_type i, j;
5124 for_each_online_pgdat(pgdat) {
5125 for (i = 0; i < MAX_NR_ZONES; i++) {
5126 struct zone *zone = pgdat->node_zones + i;
5127 unsigned long max = 0;
5129 /* Find valid and maximum lowmem_reserve in the zone */
5130 for (j = i; j < MAX_NR_ZONES; j++) {
5131 if (zone->lowmem_reserve[j] > max)
5132 max = zone->lowmem_reserve[j];
5135 /* we treat the high watermark as reserved pages. */
5136 max += high_wmark_pages(zone);
5138 if (max > zone->present_pages)
5139 max = zone->present_pages;
5140 reserve_pages += max;
5142 * Lowmem reserves are not available to
5143 * GFP_HIGHUSER page cache allocations and
5144 * kswapd tries to balance zones to their high
5145 * watermark. As a result, neither should be
5146 * regarded as dirtyable memory, to prevent a
5147 * situation where reclaim has to clean pages
5148 * in order to balance the zones.
5150 zone->dirty_balance_reserve = max;
5153 dirty_balance_reserve = reserve_pages;
5154 totalreserve_pages = reserve_pages;
5158 * setup_per_zone_lowmem_reserve - called whenever
5159 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5160 * has a correct pages reserved value, so an adequate number of
5161 * pages are left in the zone after a successful __alloc_pages().
5163 static void setup_per_zone_lowmem_reserve(void)
5165 struct pglist_data *pgdat;
5166 enum zone_type j, idx;
5168 for_each_online_pgdat(pgdat) {
5169 for (j = 0; j < MAX_NR_ZONES; j++) {
5170 struct zone *zone = pgdat->node_zones + j;
5171 unsigned long present_pages = zone->present_pages;
5173 zone->lowmem_reserve[j] = 0;
5177 struct zone *lower_zone;
5181 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5182 sysctl_lowmem_reserve_ratio[idx] = 1;
5184 lower_zone = pgdat->node_zones + idx;
5185 lower_zone->lowmem_reserve[j] = present_pages /
5186 sysctl_lowmem_reserve_ratio[idx];
5187 present_pages += lower_zone->present_pages;
5192 /* update totalreserve_pages */
5193 calculate_totalreserve_pages();
5196 static void __setup_per_zone_wmarks(void)
5198 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5199 unsigned long lowmem_pages = 0;
5201 unsigned long flags;
5203 /* Calculate total number of !ZONE_HIGHMEM pages */
5204 for_each_zone(zone) {
5205 if (!is_highmem(zone))
5206 lowmem_pages += zone->present_pages;
5209 for_each_zone(zone) {
5212 spin_lock_irqsave(&zone->lock, flags);
5213 tmp = (u64)pages_min * zone->present_pages;
5214 do_div(tmp, lowmem_pages);
5215 if (is_highmem(zone)) {
5217 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5218 * need highmem pages, so cap pages_min to a small
5221 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5222 * deltas controls asynch page reclaim, and so should
5223 * not be capped for highmem.
5227 min_pages = zone->present_pages / 1024;
5228 if (min_pages < SWAP_CLUSTER_MAX)
5229 min_pages = SWAP_CLUSTER_MAX;
5230 if (min_pages > 128)
5232 zone->watermark[WMARK_MIN] = min_pages;
5235 * If it's a lowmem zone, reserve a number of pages
5236 * proportionate to the zone's size.
5238 zone->watermark[WMARK_MIN] = tmp;
5241 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5242 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5244 setup_zone_migrate_reserve(zone);
5245 spin_unlock_irqrestore(&zone->lock, flags);
5248 /* update totalreserve_pages */
5249 calculate_totalreserve_pages();
5253 * setup_per_zone_wmarks - called when min_free_kbytes changes
5254 * or when memory is hot-{added|removed}
5256 * Ensures that the watermark[min,low,high] values for each zone are set
5257 * correctly with respect to min_free_kbytes.
5259 void setup_per_zone_wmarks(void)
5261 mutex_lock(&zonelists_mutex);
5262 __setup_per_zone_wmarks();
5263 mutex_unlock(&zonelists_mutex);
5267 * The inactive anon list should be small enough that the VM never has to
5268 * do too much work, but large enough that each inactive page has a chance
5269 * to be referenced again before it is swapped out.
5271 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5272 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5273 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5274 * the anonymous pages are kept on the inactive list.
5277 * memory ratio inactive anon
5278 * -------------------------------------
5287 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5289 unsigned int gb, ratio;
5291 /* Zone size in gigabytes */
5292 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5294 ratio = int_sqrt(10 * gb);
5298 zone->inactive_ratio = ratio;
5301 static void __meminit setup_per_zone_inactive_ratio(void)
5306 calculate_zone_inactive_ratio(zone);
5310 * Initialise min_free_kbytes.
5312 * For small machines we want it small (128k min). For large machines
5313 * we want it large (64MB max). But it is not linear, because network
5314 * bandwidth does not increase linearly with machine size. We use
5316 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5317 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5333 int __meminit init_per_zone_wmark_min(void)
5335 unsigned long lowmem_kbytes;
5337 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5339 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5340 if (min_free_kbytes < 128)
5341 min_free_kbytes = 128;
5342 if (min_free_kbytes > 65536)
5343 min_free_kbytes = 65536;
5344 setup_per_zone_wmarks();
5345 refresh_zone_stat_thresholds();
5346 setup_per_zone_lowmem_reserve();
5347 setup_per_zone_inactive_ratio();
5350 module_init(init_per_zone_wmark_min)
5353 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5354 * that we can call two helper functions whenever min_free_kbytes
5357 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5358 void __user *buffer, size_t *length, loff_t *ppos)
5360 proc_dointvec(table, write, buffer, length, ppos);
5362 setup_per_zone_wmarks();
5367 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5368 void __user *buffer, size_t *length, loff_t *ppos)
5373 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5378 zone->min_unmapped_pages = (zone->present_pages *
5379 sysctl_min_unmapped_ratio) / 100;
5383 int sysctl_min_slab_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_slab_pages = (zone->present_pages *
5395 sysctl_min_slab_ratio) / 100;
5401 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5402 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5403 * whenever sysctl_lowmem_reserve_ratio changes.
5405 * The reserve ratio obviously has absolutely no relation with the
5406 * minimum watermarks. The lowmem reserve ratio can only make sense
5407 * if in function of the boot time zone sizes.
5409 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5410 void __user *buffer, size_t *length, loff_t *ppos)
5412 proc_dointvec_minmax(table, write, buffer, length, ppos);
5413 setup_per_zone_lowmem_reserve();
5418 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5419 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5420 * can have before it gets flushed back to buddy allocator.
5423 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5424 void __user *buffer, size_t *length, loff_t *ppos)
5430 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5431 if (!write || (ret < 0))
5433 for_each_populated_zone(zone) {
5434 for_each_possible_cpu(cpu) {
5436 high = zone->present_pages / percpu_pagelist_fraction;
5437 setup_pagelist_highmark(
5438 per_cpu_ptr(zone->pageset, cpu), high);
5444 int hashdist = HASHDIST_DEFAULT;
5447 static int __init set_hashdist(char *str)
5451 hashdist = simple_strtoul(str, &str, 0);
5454 __setup("hashdist=", set_hashdist);
5458 * allocate a large system hash table from bootmem
5459 * - it is assumed that the hash table must contain an exact power-of-2
5460 * quantity of entries
5461 * - limit is the number of hash buckets, not the total allocation size
5463 void *__init alloc_large_system_hash(const char *tablename,
5464 unsigned long bucketsize,
5465 unsigned long numentries,
5468 unsigned int *_hash_shift,
5469 unsigned int *_hash_mask,
5470 unsigned long low_limit,
5471 unsigned long high_limit)
5473 unsigned long long max = high_limit;
5474 unsigned long log2qty, size;
5477 /* allow the kernel cmdline to have a say */
5479 /* round applicable memory size up to nearest megabyte */
5480 numentries = nr_kernel_pages;
5481 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5482 numentries >>= 20 - PAGE_SHIFT;
5483 numentries <<= 20 - PAGE_SHIFT;
5485 /* limit to 1 bucket per 2^scale bytes of low memory */
5486 if (scale > PAGE_SHIFT)
5487 numentries >>= (scale - PAGE_SHIFT);
5489 numentries <<= (PAGE_SHIFT - scale);
5491 /* Make sure we've got at least a 0-order allocation.. */
5492 if (unlikely(flags & HASH_SMALL)) {
5493 /* Makes no sense without HASH_EARLY */
5494 WARN_ON(!(flags & HASH_EARLY));
5495 if (!(numentries >> *_hash_shift)) {
5496 numentries = 1UL << *_hash_shift;
5497 BUG_ON(!numentries);
5499 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5500 numentries = PAGE_SIZE / bucketsize;
5502 numentries = roundup_pow_of_two(numentries);
5504 /* limit allocation size to 1/16 total memory by default */
5506 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5507 do_div(max, bucketsize);
5509 max = min(max, 0x80000000ULL);
5511 if (numentries < low_limit)
5512 numentries = low_limit;
5513 if (numentries > max)
5516 log2qty = ilog2(numentries);
5519 size = bucketsize << log2qty;
5520 if (flags & HASH_EARLY)
5521 table = alloc_bootmem_nopanic(size);
5523 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5526 * If bucketsize is not a power-of-two, we may free
5527 * some pages at the end of hash table which
5528 * alloc_pages_exact() automatically does
5530 if (get_order(size) < MAX_ORDER) {
5531 table = alloc_pages_exact(size, GFP_ATOMIC);
5532 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5535 } while (!table && size > PAGE_SIZE && --log2qty);
5538 panic("Failed to allocate %s hash table\n", tablename);
5540 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5543 ilog2(size) - PAGE_SHIFT,
5547 *_hash_shift = log2qty;
5549 *_hash_mask = (1 << log2qty) - 1;
5554 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5555 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5558 #ifdef CONFIG_SPARSEMEM
5559 return __pfn_to_section(pfn)->pageblock_flags;
5561 return zone->pageblock_flags;
5562 #endif /* CONFIG_SPARSEMEM */
5565 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5567 #ifdef CONFIG_SPARSEMEM
5568 pfn &= (PAGES_PER_SECTION-1);
5569 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5571 pfn = pfn - zone->zone_start_pfn;
5572 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5573 #endif /* CONFIG_SPARSEMEM */
5577 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5578 * @page: The page within the block of interest
5579 * @start_bitidx: The first bit of interest to retrieve
5580 * @end_bitidx: The last bit of interest
5581 * returns pageblock_bits flags
5583 unsigned long get_pageblock_flags_group(struct page *page,
5584 int start_bitidx, int end_bitidx)
5587 unsigned long *bitmap;
5588 unsigned long pfn, bitidx;
5589 unsigned long flags = 0;
5590 unsigned long value = 1;
5592 zone = page_zone(page);
5593 pfn = page_to_pfn(page);
5594 bitmap = get_pageblock_bitmap(zone, pfn);
5595 bitidx = pfn_to_bitidx(zone, pfn);
5597 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5598 if (test_bit(bitidx + start_bitidx, bitmap))
5605 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5606 * @page: The page within the block of interest
5607 * @start_bitidx: The first bit of interest
5608 * @end_bitidx: The last bit of interest
5609 * @flags: The flags to set
5611 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5612 int start_bitidx, int end_bitidx)
5615 unsigned long *bitmap;
5616 unsigned long pfn, bitidx;
5617 unsigned long value = 1;
5619 zone = page_zone(page);
5620 pfn = page_to_pfn(page);
5621 bitmap = get_pageblock_bitmap(zone, pfn);
5622 bitidx = pfn_to_bitidx(zone, pfn);
5623 VM_BUG_ON(pfn < zone->zone_start_pfn);
5624 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5626 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5628 __set_bit(bitidx + start_bitidx, bitmap);
5630 __clear_bit(bitidx + start_bitidx, bitmap);
5634 * This function checks whether pageblock includes unmovable pages or not.
5635 * If @count is not zero, it is okay to include less @count unmovable pages
5637 * PageLRU check wihtout isolation or lru_lock could race so that
5638 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5639 * expect this function should be exact.
5641 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5642 bool skip_hwpoisoned_pages)
5644 unsigned long pfn, iter, found;
5648 * For avoiding noise data, lru_add_drain_all() should be called
5649 * If ZONE_MOVABLE, the zone never contains unmovable pages
5651 if (zone_idx(zone) == ZONE_MOVABLE)
5653 mt = get_pageblock_migratetype(page);
5654 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5657 pfn = page_to_pfn(page);
5658 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5659 unsigned long check = pfn + iter;
5661 if (!pfn_valid_within(check))
5664 page = pfn_to_page(check);
5666 * We can't use page_count without pin a page
5667 * because another CPU can free compound page.
5668 * This check already skips compound tails of THP
5669 * because their page->_count is zero at all time.
5671 if (!atomic_read(&page->_count)) {
5672 if (PageBuddy(page))
5673 iter += (1 << page_order(page)) - 1;
5678 * The HWPoisoned page may be not in buddy system, and
5679 * page_count() is not 0.
5681 if (skip_hwpoisoned_pages && PageHWPoison(page))
5687 * If there are RECLAIMABLE pages, we need to check it.
5688 * But now, memory offline itself doesn't call shrink_slab()
5689 * and it still to be fixed.
5692 * If the page is not RAM, page_count()should be 0.
5693 * we don't need more check. This is an _used_ not-movable page.
5695 * The problematic thing here is PG_reserved pages. PG_reserved
5696 * is set to both of a memory hole page and a _used_ kernel
5705 bool is_pageblock_removable_nolock(struct page *page)
5711 * We have to be careful here because we are iterating over memory
5712 * sections which are not zone aware so we might end up outside of
5713 * the zone but still within the section.
5714 * We have to take care about the node as well. If the node is offline
5715 * its NODE_DATA will be NULL - see page_zone.
5717 if (!node_online(page_to_nid(page)))
5720 zone = page_zone(page);
5721 pfn = page_to_pfn(page);
5722 if (zone->zone_start_pfn > pfn ||
5723 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5726 return !has_unmovable_pages(zone, page, 0, true);
5731 static unsigned long pfn_max_align_down(unsigned long pfn)
5733 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5734 pageblock_nr_pages) - 1);
5737 static unsigned long pfn_max_align_up(unsigned long pfn)
5739 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5740 pageblock_nr_pages));
5743 /* [start, end) must belong to a single zone. */
5744 static int __alloc_contig_migrate_range(struct compact_control *cc,
5745 unsigned long start, unsigned long end)
5747 /* This function is based on compact_zone() from compaction.c. */
5748 unsigned long nr_reclaimed;
5749 unsigned long pfn = start;
5750 unsigned int tries = 0;
5755 while (pfn < end || !list_empty(&cc->migratepages)) {
5756 if (fatal_signal_pending(current)) {
5761 if (list_empty(&cc->migratepages)) {
5762 cc->nr_migratepages = 0;
5763 pfn = isolate_migratepages_range(cc->zone, cc,
5770 } else if (++tries == 5) {
5771 ret = ret < 0 ? ret : -EBUSY;
5775 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5777 cc->nr_migratepages -= nr_reclaimed;
5779 ret = migrate_pages(&cc->migratepages,
5780 alloc_migrate_target,
5781 0, false, MIGRATE_SYNC);
5784 putback_movable_pages(&cc->migratepages);
5785 return ret > 0 ? 0 : ret;
5789 * alloc_contig_range() -- tries to allocate given range of pages
5790 * @start: start PFN to allocate
5791 * @end: one-past-the-last PFN to allocate
5792 * @migratetype: migratetype of the underlaying pageblocks (either
5793 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5794 * in range must have the same migratetype and it must
5795 * be either of the two.
5797 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5798 * aligned, however it's the caller's responsibility to guarantee that
5799 * we are the only thread that changes migrate type of pageblocks the
5802 * The PFN range must belong to a single zone.
5804 * Returns zero on success or negative error code. On success all
5805 * pages which PFN is in [start, end) are allocated for the caller and
5806 * need to be freed with free_contig_range().
5808 int alloc_contig_range(unsigned long start, unsigned long end,
5809 unsigned migratetype)
5811 unsigned long outer_start, outer_end;
5814 struct compact_control cc = {
5815 .nr_migratepages = 0,
5817 .zone = page_zone(pfn_to_page(start)),
5819 .ignore_skip_hint = true,
5821 INIT_LIST_HEAD(&cc.migratepages);
5824 * What we do here is we mark all pageblocks in range as
5825 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5826 * have different sizes, and due to the way page allocator
5827 * work, we align the range to biggest of the two pages so
5828 * that page allocator won't try to merge buddies from
5829 * different pageblocks and change MIGRATE_ISOLATE to some
5830 * other migration type.
5832 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5833 * migrate the pages from an unaligned range (ie. pages that
5834 * we are interested in). This will put all the pages in
5835 * range back to page allocator as MIGRATE_ISOLATE.
5837 * When this is done, we take the pages in range from page
5838 * allocator removing them from the buddy system. This way
5839 * page allocator will never consider using them.
5841 * This lets us mark the pageblocks back as
5842 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5843 * aligned range but not in the unaligned, original range are
5844 * put back to page allocator so that buddy can use them.
5847 ret = start_isolate_page_range(pfn_max_align_down(start),
5848 pfn_max_align_up(end), migratetype,
5853 ret = __alloc_contig_migrate_range(&cc, start, end);
5858 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5859 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5860 * more, all pages in [start, end) are free in page allocator.
5861 * What we are going to do is to allocate all pages from
5862 * [start, end) (that is remove them from page allocator).
5864 * The only problem is that pages at the beginning and at the
5865 * end of interesting range may be not aligned with pages that
5866 * page allocator holds, ie. they can be part of higher order
5867 * pages. Because of this, we reserve the bigger range and
5868 * once this is done free the pages we are not interested in.
5870 * We don't have to hold zone->lock here because the pages are
5871 * isolated thus they won't get removed from buddy.
5874 lru_add_drain_all();
5878 outer_start = start;
5879 while (!PageBuddy(pfn_to_page(outer_start))) {
5880 if (++order >= MAX_ORDER) {
5884 outer_start &= ~0UL << order;
5887 /* Make sure the range is really isolated. */
5888 if (test_pages_isolated(outer_start, end, false)) {
5889 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5896 /* Grab isolated pages from freelists. */
5897 outer_end = isolate_freepages_range(&cc, outer_start, end);
5903 /* Free head and tail (if any) */
5904 if (start != outer_start)
5905 free_contig_range(outer_start, start - outer_start);
5906 if (end != outer_end)
5907 free_contig_range(end, outer_end - end);
5910 undo_isolate_page_range(pfn_max_align_down(start),
5911 pfn_max_align_up(end), migratetype);
5915 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5917 for (; nr_pages--; ++pfn)
5918 __free_page(pfn_to_page(pfn));
5922 #ifdef CONFIG_MEMORY_HOTPLUG
5923 static int __meminit __zone_pcp_update(void *data)
5925 struct zone *zone = data;
5927 unsigned long batch = zone_batchsize(zone), flags;
5929 for_each_possible_cpu(cpu) {
5930 struct per_cpu_pageset *pset;
5931 struct per_cpu_pages *pcp;
5933 pset = per_cpu_ptr(zone->pageset, cpu);
5936 local_irq_save(flags);
5938 free_pcppages_bulk(zone, pcp->count, pcp);
5939 drain_zonestat(zone, pset);
5940 setup_pageset(pset, batch);
5941 local_irq_restore(flags);
5946 void __meminit zone_pcp_update(struct zone *zone)
5948 stop_machine(__zone_pcp_update, zone, NULL);
5952 void zone_pcp_reset(struct zone *zone)
5954 unsigned long flags;
5956 struct per_cpu_pageset *pset;
5958 /* avoid races with drain_pages() */
5959 local_irq_save(flags);
5960 if (zone->pageset != &boot_pageset) {
5961 for_each_online_cpu(cpu) {
5962 pset = per_cpu_ptr(zone->pageset, cpu);
5963 drain_zonestat(zone, pset);
5965 free_percpu(zone->pageset);
5966 zone->pageset = &boot_pageset;
5968 local_irq_restore(flags);
5971 #ifdef CONFIG_MEMORY_HOTREMOVE
5973 * All pages in the range must be isolated before calling this.
5976 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5982 unsigned long flags;
5983 /* find the first valid pfn */
5984 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5989 zone = page_zone(pfn_to_page(pfn));
5990 spin_lock_irqsave(&zone->lock, flags);
5992 while (pfn < end_pfn) {
5993 if (!pfn_valid(pfn)) {
5997 page = pfn_to_page(pfn);
5999 * The HWPoisoned page may be not in buddy system, and
6000 * page_count() is not 0.
6002 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6004 SetPageReserved(page);
6008 BUG_ON(page_count(page));
6009 BUG_ON(!PageBuddy(page));
6010 order = page_order(page);
6011 #ifdef CONFIG_DEBUG_VM
6012 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6013 pfn, 1 << order, end_pfn);
6015 list_del(&page->lru);
6016 rmv_page_order(page);
6017 zone->free_area[order].nr_free--;
6018 for (i = 0; i < (1 << order); i++)
6019 SetPageReserved((page+i));
6020 pfn += (1 << order);
6022 spin_unlock_irqrestore(&zone->lock, flags);
6026 #ifdef CONFIG_MEMORY_FAILURE
6027 bool is_free_buddy_page(struct page *page)
6029 struct zone *zone = page_zone(page);
6030 unsigned long pfn = page_to_pfn(page);
6031 unsigned long flags;
6034 spin_lock_irqsave(&zone->lock, flags);
6035 for (order = 0; order < MAX_ORDER; order++) {
6036 struct page *page_head = page - (pfn & ((1 << order) - 1));
6038 if (PageBuddy(page_head) && page_order(page_head) >= order)
6041 spin_unlock_irqrestore(&zone->lock, flags);
6043 return order < MAX_ORDER;
6047 static const struct trace_print_flags pageflag_names[] = {
6048 {1UL << PG_locked, "locked" },
6049 {1UL << PG_error, "error" },
6050 {1UL << PG_referenced, "referenced" },
6051 {1UL << PG_uptodate, "uptodate" },
6052 {1UL << PG_dirty, "dirty" },
6053 {1UL << PG_lru, "lru" },
6054 {1UL << PG_active, "active" },
6055 {1UL << PG_slab, "slab" },
6056 {1UL << PG_owner_priv_1, "owner_priv_1" },
6057 {1UL << PG_arch_1, "arch_1" },
6058 {1UL << PG_reserved, "reserved" },
6059 {1UL << PG_private, "private" },
6060 {1UL << PG_private_2, "private_2" },
6061 {1UL << PG_writeback, "writeback" },
6062 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6063 {1UL << PG_head, "head" },
6064 {1UL << PG_tail, "tail" },
6066 {1UL << PG_compound, "compound" },
6068 {1UL << PG_swapcache, "swapcache" },
6069 {1UL << PG_mappedtodisk, "mappedtodisk" },
6070 {1UL << PG_reclaim, "reclaim" },
6071 {1UL << PG_swapbacked, "swapbacked" },
6072 {1UL << PG_unevictable, "unevictable" },
6074 {1UL << PG_mlocked, "mlocked" },
6076 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6077 {1UL << PG_uncached, "uncached" },
6079 #ifdef CONFIG_MEMORY_FAILURE
6080 {1UL << PG_hwpoison, "hwpoison" },
6082 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6083 {1UL << PG_compound_lock, "compound_lock" },
6087 static void dump_page_flags(unsigned long flags)
6089 const char *delim = "";
6093 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6095 printk(KERN_ALERT "page flags: %#lx(", flags);
6097 /* remove zone id */
6098 flags &= (1UL << NR_PAGEFLAGS) - 1;
6100 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6102 mask = pageflag_names[i].mask;
6103 if ((flags & mask) != mask)
6107 printk("%s%s", delim, pageflag_names[i].name);
6111 /* check for left over flags */
6113 printk("%s%#lx", delim, flags);
6118 void dump_page(struct page *page)
6121 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6122 page, atomic_read(&page->_count), page_mapcount(page),
6123 page->mapping, page->index);
6124 dump_page_flags(page->flags);
6125 mem_cgroup_print_bad_page(page);