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/memory.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/memcontrol.h>
59 #include <linux/prefetch.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
63 #include <asm/tlbflush.h>
64 #include <asm/div64.h>
67 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
68 DEFINE_PER_CPU(int, numa_node);
69 EXPORT_PER_CPU_SYMBOL(numa_node);
72 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
74 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
75 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
76 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
77 * defined in <linux/topology.h>.
79 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
80 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
84 * Array of node states.
86 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
87 [N_POSSIBLE] = NODE_MASK_ALL,
88 [N_ONLINE] = { { [0] = 1UL } },
90 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
92 [N_HIGH_MEMORY] = { { [0] = 1UL } },
94 [N_CPU] = { { [0] = 1UL } },
97 EXPORT_SYMBOL(node_states);
99 unsigned long totalram_pages __read_mostly;
100 unsigned long totalreserve_pages __read_mostly;
102 * When calculating the number of globally allowed dirty pages, there
103 * is a certain number of per-zone reserves that should not be
104 * considered dirtyable memory. This is the sum of those reserves
105 * over all existing zones that contribute dirtyable memory.
107 unsigned long dirty_balance_reserve __read_mostly;
109 int percpu_pagelist_fraction;
110 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
112 #ifdef CONFIG_PM_SLEEP
114 * The following functions are used by the suspend/hibernate code to temporarily
115 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
116 * while devices are suspended. To avoid races with the suspend/hibernate code,
117 * they should always be called with pm_mutex held (gfp_allowed_mask also should
118 * only be modified with pm_mutex held, unless the suspend/hibernate code is
119 * guaranteed not to run in parallel with that modification).
122 static gfp_t saved_gfp_mask;
124 void pm_restore_gfp_mask(void)
126 WARN_ON(!mutex_is_locked(&pm_mutex));
127 if (saved_gfp_mask) {
128 gfp_allowed_mask = saved_gfp_mask;
133 void pm_restrict_gfp_mask(void)
135 WARN_ON(!mutex_is_locked(&pm_mutex));
136 WARN_ON(saved_gfp_mask);
137 saved_gfp_mask = gfp_allowed_mask;
138 gfp_allowed_mask &= ~GFP_IOFS;
141 bool pm_suspended_storage(void)
143 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
147 #endif /* CONFIG_PM_SLEEP */
149 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
150 int pageblock_order __read_mostly;
153 static void __free_pages_ok(struct page *page, unsigned int order);
156 * results with 256, 32 in the lowmem_reserve sysctl:
157 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
158 * 1G machine -> (16M dma, 784M normal, 224M high)
159 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
160 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
161 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
163 * TBD: should special case ZONE_DMA32 machines here - in those we normally
164 * don't need any ZONE_NORMAL reservation
166 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
167 #ifdef CONFIG_ZONE_DMA
170 #ifdef CONFIG_ZONE_DMA32
173 #ifdef CONFIG_HIGHMEM
179 EXPORT_SYMBOL(totalram_pages);
181 static char * const zone_names[MAX_NR_ZONES] = {
182 #ifdef CONFIG_ZONE_DMA
185 #ifdef CONFIG_ZONE_DMA32
189 #ifdef CONFIG_HIGHMEM
195 int min_free_kbytes = 1024;
197 static unsigned long __meminitdata nr_kernel_pages;
198 static unsigned long __meminitdata nr_all_pages;
199 static unsigned long __meminitdata dma_reserve;
201 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
203 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
204 * ranges of memory (RAM) that may be registered with add_active_range().
205 * Ranges passed to add_active_range() will be merged if possible
206 * so the number of times add_active_range() can be called is
207 * related to the number of nodes and the number of holes
209 #ifdef CONFIG_MAX_ACTIVE_REGIONS
210 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
211 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
213 #if MAX_NUMNODES >= 32
214 /* If there can be many nodes, allow up to 50 holes per node */
215 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
217 /* By default, allow up to 256 distinct regions */
218 #define MAX_ACTIVE_REGIONS 256
222 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
223 static int __meminitdata nr_nodemap_entries;
224 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
225 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
226 static unsigned long __initdata required_kernelcore;
227 static unsigned long __initdata required_movablecore;
228 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
230 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
232 EXPORT_SYMBOL(movable_zone);
233 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
236 int nr_node_ids __read_mostly = MAX_NUMNODES;
237 int nr_online_nodes __read_mostly = 1;
238 EXPORT_SYMBOL(nr_node_ids);
239 EXPORT_SYMBOL(nr_online_nodes);
242 int page_group_by_mobility_disabled __read_mostly;
244 static void set_pageblock_migratetype(struct page *page, int migratetype)
247 if (unlikely(page_group_by_mobility_disabled))
248 migratetype = MIGRATE_UNMOVABLE;
250 set_pageblock_flags_group(page, (unsigned long)migratetype,
251 PB_migrate, PB_migrate_end);
254 bool oom_killer_disabled __read_mostly;
256 #ifdef CONFIG_DEBUG_VM
257 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
261 unsigned long pfn = page_to_pfn(page);
264 seq = zone_span_seqbegin(zone);
265 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
267 else if (pfn < zone->zone_start_pfn)
269 } while (zone_span_seqretry(zone, seq));
274 static int page_is_consistent(struct zone *zone, struct page *page)
276 if (!pfn_valid_within(page_to_pfn(page)))
278 if (zone != page_zone(page))
284 * Temporary debugging check for pages not lying within a given zone.
286 static int bad_range(struct zone *zone, struct page *page)
288 if (page_outside_zone_boundaries(zone, page))
290 if (!page_is_consistent(zone, page))
296 static inline int bad_range(struct zone *zone, struct page *page)
302 static void bad_page(struct page *page)
304 static unsigned long resume;
305 static unsigned long nr_shown;
306 static unsigned long nr_unshown;
308 /* Don't complain about poisoned pages */
309 if (PageHWPoison(page)) {
310 reset_page_mapcount(page); /* remove PageBuddy */
315 * Allow a burst of 60 reports, then keep quiet for that minute;
316 * or allow a steady drip of one report per second.
318 if (nr_shown == 60) {
319 if (time_before(jiffies, resume)) {
325 "BUG: Bad page state: %lu messages suppressed\n",
332 resume = jiffies + 60 * HZ;
334 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
335 current->comm, page_to_pfn(page));
341 /* Leave bad fields for debug, except PageBuddy could make trouble */
342 reset_page_mapcount(page); /* remove PageBuddy */
343 add_taint(TAINT_BAD_PAGE);
347 * Higher-order pages are called "compound pages". They are structured thusly:
349 * The first PAGE_SIZE page is called the "head page".
351 * The remaining PAGE_SIZE pages are called "tail pages".
353 * All pages have PG_compound set. All pages have their ->private pointing at
354 * the head page (even the head page has this).
356 * The first tail page's ->lru.next holds the address of the compound page's
357 * put_page() function. Its ->lru.prev holds the order of allocation.
358 * This usage means that zero-order pages may not be compound.
361 static void free_compound_page(struct page *page)
363 __free_pages_ok(page, compound_order(page));
366 void prep_compound_page(struct page *page, unsigned long order)
369 int nr_pages = 1 << order;
371 set_compound_page_dtor(page, free_compound_page);
372 set_compound_order(page, order);
374 for (i = 1; i < nr_pages; i++) {
375 struct page *p = page + i;
377 set_page_count(p, 0);
378 p->first_page = page;
382 /* update __split_huge_page_refcount if you change this function */
383 static int destroy_compound_page(struct page *page, unsigned long order)
386 int nr_pages = 1 << order;
389 if (unlikely(compound_order(page) != order) ||
390 unlikely(!PageHead(page))) {
395 __ClearPageHead(page);
397 for (i = 1; i < nr_pages; i++) {
398 struct page *p = page + i;
400 if (unlikely(!PageTail(p) || (p->first_page != page))) {
410 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
415 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
416 * and __GFP_HIGHMEM from hard or soft interrupt context.
418 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
419 for (i = 0; i < (1 << order); i++)
420 clear_highpage(page + i);
423 #ifdef CONFIG_DEBUG_PAGEALLOC
424 unsigned int _debug_guardpage_minorder;
426 static int __init debug_guardpage_minorder_setup(char *buf)
430 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
431 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
434 _debug_guardpage_minorder = res;
435 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
438 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
440 static inline void set_page_guard_flag(struct page *page)
442 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
445 static inline void clear_page_guard_flag(struct page *page)
447 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
450 static inline void set_page_guard_flag(struct page *page) { }
451 static inline void clear_page_guard_flag(struct page *page) { }
454 static inline void set_page_order(struct page *page, int order)
456 set_page_private(page, order);
457 __SetPageBuddy(page);
460 static inline void rmv_page_order(struct page *page)
462 __ClearPageBuddy(page);
463 set_page_private(page, 0);
467 * Locate the struct page for both the matching buddy in our
468 * pair (buddy1) and the combined O(n+1) page they form (page).
470 * 1) Any buddy B1 will have an order O twin B2 which satisfies
471 * the following equation:
473 * For example, if the starting buddy (buddy2) is #8 its order
475 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
477 * 2) Any buddy B will have an order O+1 parent P which
478 * satisfies the following equation:
481 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
483 static inline unsigned long
484 __find_buddy_index(unsigned long page_idx, unsigned int order)
486 return page_idx ^ (1 << order);
490 * This function checks whether a page is free && is the buddy
491 * we can do coalesce a page and its buddy if
492 * (a) the buddy is not in a hole &&
493 * (b) the buddy is in the buddy system &&
494 * (c) a page and its buddy have the same order &&
495 * (d) a page and its buddy are in the same zone.
497 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
498 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
500 * For recording page's order, we use page_private(page).
502 static inline int page_is_buddy(struct page *page, struct page *buddy,
505 if (!pfn_valid_within(page_to_pfn(buddy)))
508 if (page_zone_id(page) != page_zone_id(buddy))
511 if (page_is_guard(buddy) && page_order(buddy) == order) {
512 VM_BUG_ON(page_count(buddy) != 0);
516 if (PageBuddy(buddy) && page_order(buddy) == order) {
517 VM_BUG_ON(page_count(buddy) != 0);
524 * Freeing function for a buddy system allocator.
526 * The concept of a buddy system is to maintain direct-mapped table
527 * (containing bit values) for memory blocks of various "orders".
528 * The bottom level table contains the map for the smallest allocatable
529 * units of memory (here, pages), and each level above it describes
530 * pairs of units from the levels below, hence, "buddies".
531 * At a high level, all that happens here is marking the table entry
532 * at the bottom level available, and propagating the changes upward
533 * as necessary, plus some accounting needed to play nicely with other
534 * parts of the VM system.
535 * At each level, we keep a list of pages, which are heads of continuous
536 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
537 * order is recorded in page_private(page) field.
538 * So when we are allocating or freeing one, we can derive the state of the
539 * other. That is, if we allocate a small block, and both were
540 * free, the remainder of the region must be split into blocks.
541 * If a block is freed, and its buddy is also free, then this
542 * triggers coalescing into a block of larger size.
547 static inline void __free_one_page(struct page *page,
548 struct zone *zone, unsigned int order,
551 unsigned long page_idx;
552 unsigned long combined_idx;
553 unsigned long uninitialized_var(buddy_idx);
556 if (unlikely(PageCompound(page)))
557 if (unlikely(destroy_compound_page(page, order)))
560 VM_BUG_ON(migratetype == -1);
562 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
564 VM_BUG_ON(page_idx & ((1 << order) - 1));
565 VM_BUG_ON(bad_range(zone, page));
567 while (order < MAX_ORDER-1) {
568 buddy_idx = __find_buddy_index(page_idx, order);
569 buddy = page + (buddy_idx - page_idx);
570 if (!page_is_buddy(page, buddy, order))
573 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
574 * merge with it and move up one order.
576 if (page_is_guard(buddy)) {
577 clear_page_guard_flag(buddy);
578 set_page_private(page, 0);
579 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
581 list_del(&buddy->lru);
582 zone->free_area[order].nr_free--;
583 rmv_page_order(buddy);
585 combined_idx = buddy_idx & page_idx;
586 page = page + (combined_idx - page_idx);
587 page_idx = combined_idx;
590 set_page_order(page, order);
593 * If this is not the largest possible page, check if the buddy
594 * of the next-highest order is free. If it is, it's possible
595 * that pages are being freed that will coalesce soon. In case,
596 * that is happening, add the free page to the tail of the list
597 * so it's less likely to be used soon and more likely to be merged
598 * as a higher order page
600 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
601 struct page *higher_page, *higher_buddy;
602 combined_idx = buddy_idx & page_idx;
603 higher_page = page + (combined_idx - page_idx);
604 buddy_idx = __find_buddy_index(combined_idx, order + 1);
605 higher_buddy = higher_page + (buddy_idx - combined_idx);
606 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
607 list_add_tail(&page->lru,
608 &zone->free_area[order].free_list[migratetype]);
613 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
615 zone->free_area[order].nr_free++;
619 * free_page_mlock() -- clean up attempts to free and mlocked() page.
620 * Page should not be on lru, so no need to fix that up.
621 * free_pages_check() will verify...
623 static inline void free_page_mlock(struct page *page)
625 __dec_zone_page_state(page, NR_MLOCK);
626 __count_vm_event(UNEVICTABLE_MLOCKFREED);
629 static inline int free_pages_check(struct page *page)
631 if (unlikely(page_mapcount(page) |
632 (page->mapping != NULL) |
633 (atomic_read(&page->_count) != 0) |
634 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
635 (mem_cgroup_bad_page_check(page)))) {
639 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
640 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
645 * Frees a number of pages from the PCP lists
646 * Assumes all pages on list are in same zone, and of same order.
647 * count is the number of pages to free.
649 * If the zone was previously in an "all pages pinned" state then look to
650 * see if this freeing clears that state.
652 * And clear the zone's pages_scanned counter, to hold off the "all pages are
653 * pinned" detection logic.
655 static void free_pcppages_bulk(struct zone *zone, int count,
656 struct per_cpu_pages *pcp)
662 spin_lock(&zone->lock);
663 zone->all_unreclaimable = 0;
664 zone->pages_scanned = 0;
668 struct list_head *list;
671 * Remove pages from lists in a round-robin fashion. A
672 * batch_free count is maintained that is incremented when an
673 * empty list is encountered. This is so more pages are freed
674 * off fuller lists instead of spinning excessively around empty
679 if (++migratetype == MIGRATE_PCPTYPES)
681 list = &pcp->lists[migratetype];
682 } while (list_empty(list));
684 /* This is the only non-empty list. Free them all. */
685 if (batch_free == MIGRATE_PCPTYPES)
686 batch_free = to_free;
689 page = list_entry(list->prev, struct page, lru);
690 /* must delete as __free_one_page list manipulates */
691 list_del(&page->lru);
692 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
693 __free_one_page(page, zone, 0, page_private(page));
694 trace_mm_page_pcpu_drain(page, 0, page_private(page));
695 } while (--to_free && --batch_free && !list_empty(list));
697 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
698 spin_unlock(&zone->lock);
701 static void free_one_page(struct zone *zone, struct page *page, int order,
704 spin_lock(&zone->lock);
705 zone->all_unreclaimable = 0;
706 zone->pages_scanned = 0;
708 __free_one_page(page, zone, order, migratetype);
709 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
710 spin_unlock(&zone->lock);
713 static bool free_pages_prepare(struct page *page, unsigned int order)
718 trace_mm_page_free(page, order);
719 kmemcheck_free_shadow(page, order);
722 page->mapping = NULL;
723 for (i = 0; i < (1 << order); i++)
724 bad += free_pages_check(page + i);
728 if (!PageHighMem(page)) {
729 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
730 debug_check_no_obj_freed(page_address(page),
733 arch_free_page(page, order);
734 kernel_map_pages(page, 1 << order, 0);
739 static void __free_pages_ok(struct page *page, unsigned int order)
742 int wasMlocked = __TestClearPageMlocked(page);
744 if (!free_pages_prepare(page, order))
747 local_irq_save(flags);
748 if (unlikely(wasMlocked))
749 free_page_mlock(page);
750 __count_vm_events(PGFREE, 1 << order);
751 free_one_page(page_zone(page), page, order,
752 get_pageblock_migratetype(page));
753 local_irq_restore(flags);
757 * permit the bootmem allocator to evade page validation on high-order frees
759 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
762 __ClearPageReserved(page);
763 set_page_count(page, 0);
764 set_page_refcounted(page);
770 for (loop = 0; loop < BITS_PER_LONG; loop++) {
771 struct page *p = &page[loop];
773 if (loop + 1 < BITS_PER_LONG)
775 __ClearPageReserved(p);
776 set_page_count(p, 0);
779 set_page_refcounted(page);
780 __free_pages(page, order);
785 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
786 void __init init_cma_reserved_pageblock(struct page *page)
788 unsigned i = pageblock_nr_pages;
789 struct page *p = page;
792 __ClearPageReserved(p);
793 set_page_count(p, 0);
796 set_page_refcounted(page);
797 set_pageblock_migratetype(page, MIGRATE_CMA);
798 __free_pages(page, pageblock_order);
799 totalram_pages += pageblock_nr_pages;
804 * The order of subdivision here is critical for the IO subsystem.
805 * Please do not alter this order without good reasons and regression
806 * testing. Specifically, as large blocks of memory are subdivided,
807 * the order in which smaller blocks are delivered depends on the order
808 * they're subdivided in this function. This is the primary factor
809 * influencing the order in which pages are delivered to the IO
810 * subsystem according to empirical testing, and this is also justified
811 * by considering the behavior of a buddy system containing a single
812 * large block of memory acted on by a series of small allocations.
813 * This behavior is a critical factor in sglist merging's success.
817 static inline void expand(struct zone *zone, struct page *page,
818 int low, int high, struct free_area *area,
821 unsigned long size = 1 << high;
827 VM_BUG_ON(bad_range(zone, &page[size]));
829 #ifdef CONFIG_DEBUG_PAGEALLOC
830 if (high < debug_guardpage_minorder()) {
832 * Mark as guard pages (or page), that will allow to
833 * merge back to allocator when buddy will be freed.
834 * Corresponding page table entries will not be touched,
835 * pages will stay not present in virtual address space
837 INIT_LIST_HEAD(&page[size].lru);
838 set_page_guard_flag(&page[size]);
839 set_page_private(&page[size], high);
840 /* Guard pages are not available for any usage */
841 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high));
845 list_add(&page[size].lru, &area->free_list[migratetype]);
847 set_page_order(&page[size], high);
852 * This page is about to be returned from the page allocator
854 static inline int check_new_page(struct page *page)
856 if (unlikely(page_mapcount(page) |
857 (page->mapping != NULL) |
858 (atomic_read(&page->_count) != 0) |
859 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
860 (mem_cgroup_bad_page_check(page)))) {
867 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
871 for (i = 0; i < (1 << order); i++) {
872 struct page *p = page + i;
873 if (unlikely(check_new_page(p)))
877 set_page_private(page, 0);
878 set_page_refcounted(page);
880 arch_alloc_page(page, order);
881 kernel_map_pages(page, 1 << order, 1);
883 if (gfp_flags & __GFP_ZERO)
884 prep_zero_page(page, order, gfp_flags);
886 if (order && (gfp_flags & __GFP_COMP))
887 prep_compound_page(page, order);
893 * Go through the free lists for the given migratetype and remove
894 * the smallest available page from the freelists
897 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
900 unsigned int current_order;
901 struct free_area * area;
904 /* Find a page of the appropriate size in the preferred list */
905 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
906 area = &(zone->free_area[current_order]);
907 if (list_empty(&area->free_list[migratetype]))
910 page = list_entry(area->free_list[migratetype].next,
912 list_del(&page->lru);
913 rmv_page_order(page);
915 expand(zone, page, order, current_order, area, migratetype);
924 * This array describes the order lists are fallen back to when
925 * the free lists for the desirable migrate type are depleted
927 static int fallbacks[MIGRATE_TYPES][4] = {
928 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
929 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
931 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
932 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
934 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
936 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
937 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
941 * Move the free pages in a range to the free lists of the requested type.
942 * Note that start_page and end_pages are not aligned on a pageblock
943 * boundary. If alignment is required, use move_freepages_block()
945 static int move_freepages(struct zone *zone,
946 struct page *start_page, struct page *end_page,
953 #ifndef CONFIG_HOLES_IN_ZONE
955 * page_zone is not safe to call in this context when
956 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
957 * anyway as we check zone boundaries in move_freepages_block().
958 * Remove at a later date when no bug reports exist related to
959 * grouping pages by mobility
961 BUG_ON(page_zone(start_page) != page_zone(end_page));
964 for (page = start_page; page <= end_page;) {
965 /* Make sure we are not inadvertently changing nodes */
966 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
968 if (!pfn_valid_within(page_to_pfn(page))) {
973 if (!PageBuddy(page)) {
978 order = page_order(page);
979 list_move(&page->lru,
980 &zone->free_area[order].free_list[migratetype]);
982 pages_moved += 1 << order;
988 static int move_freepages_block(struct zone *zone, struct page *page,
991 unsigned long start_pfn, end_pfn;
992 struct page *start_page, *end_page;
994 start_pfn = page_to_pfn(page);
995 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
996 start_page = pfn_to_page(start_pfn);
997 end_page = start_page + pageblock_nr_pages - 1;
998 end_pfn = start_pfn + pageblock_nr_pages - 1;
1000 /* Do not cross zone boundaries */
1001 if (start_pfn < zone->zone_start_pfn)
1003 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
1006 return move_freepages(zone, start_page, end_page, migratetype);
1009 static void change_pageblock_range(struct page *pageblock_page,
1010 int start_order, int migratetype)
1012 int nr_pageblocks = 1 << (start_order - pageblock_order);
1014 while (nr_pageblocks--) {
1015 set_pageblock_migratetype(pageblock_page, migratetype);
1016 pageblock_page += pageblock_nr_pages;
1020 /* Remove an element from the buddy allocator from the fallback list */
1021 static inline struct page *
1022 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1024 struct free_area * area;
1029 /* Find the largest possible block of pages in the other list */
1030 for (current_order = MAX_ORDER-1; current_order >= order;
1033 migratetype = fallbacks[start_migratetype][i];
1035 /* MIGRATE_RESERVE handled later if necessary */
1036 if (migratetype == MIGRATE_RESERVE)
1039 area = &(zone->free_area[current_order]);
1040 if (list_empty(&area->free_list[migratetype]))
1043 page = list_entry(area->free_list[migratetype].next,
1048 * If breaking a large block of pages, move all free
1049 * pages to the preferred allocation list. If falling
1050 * back for a reclaimable kernel allocation, be more
1051 * aggressive about taking ownership of free pages
1053 * On the other hand, never change migration
1054 * type of MIGRATE_CMA pageblocks nor move CMA
1055 * pages on different free lists. We don't
1056 * want unmovable pages to be allocated from
1057 * MIGRATE_CMA areas.
1059 if (!is_migrate_cma(migratetype) &&
1060 (unlikely(current_order >= pageblock_order / 2) ||
1061 start_migratetype == MIGRATE_RECLAIMABLE ||
1062 page_group_by_mobility_disabled)) {
1064 pages = move_freepages_block(zone, page,
1067 /* Claim the whole block if over half of it is free */
1068 if (pages >= (1 << (pageblock_order-1)) ||
1069 page_group_by_mobility_disabled)
1070 set_pageblock_migratetype(page,
1073 migratetype = start_migratetype;
1076 /* Remove the page from the freelists */
1077 list_del(&page->lru);
1078 rmv_page_order(page);
1080 /* Take ownership for orders >= pageblock_order */
1081 if (current_order >= pageblock_order &&
1082 !is_migrate_cma(migratetype))
1083 change_pageblock_range(page, current_order,
1086 expand(zone, page, order, current_order, area,
1087 is_migrate_cma(migratetype)
1088 ? migratetype : start_migratetype);
1090 trace_mm_page_alloc_extfrag(page, order, current_order,
1091 start_migratetype, migratetype);
1101 * Do the hard work of removing an element from the buddy allocator.
1102 * Call me with the zone->lock already held.
1104 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1110 page = __rmqueue_smallest(zone, order, migratetype);
1112 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1113 page = __rmqueue_fallback(zone, order, migratetype);
1116 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1117 * is used because __rmqueue_smallest is an inline function
1118 * and we want just one call site
1121 migratetype = MIGRATE_RESERVE;
1126 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1131 * Obtain a specified number of elements from the buddy allocator, all under
1132 * a single hold of the lock, for efficiency. Add them to the supplied list.
1133 * Returns the number of new pages which were placed at *list.
1135 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1136 unsigned long count, struct list_head *list,
1137 int migratetype, int cold)
1139 int mt = migratetype, i;
1141 spin_lock(&zone->lock);
1142 for (i = 0; i < count; ++i) {
1143 struct page *page = __rmqueue(zone, order, migratetype);
1144 if (unlikely(page == NULL))
1148 * Split buddy pages returned by expand() are received here
1149 * in physical page order. The page is added to the callers and
1150 * list and the list head then moves forward. From the callers
1151 * perspective, the linked list is ordered by page number in
1152 * some conditions. This is useful for IO devices that can
1153 * merge IO requests if the physical pages are ordered
1156 if (likely(cold == 0))
1157 list_add(&page->lru, list);
1159 list_add_tail(&page->lru, list);
1160 if (IS_ENABLED(CONFIG_CMA)) {
1161 mt = get_pageblock_migratetype(page);
1162 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1165 set_page_private(page, mt);
1168 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1169 spin_unlock(&zone->lock);
1175 * Called from the vmstat counter updater to drain pagesets of this
1176 * currently executing processor on remote nodes after they have
1179 * Note that this function must be called with the thread pinned to
1180 * a single processor.
1182 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1184 unsigned long flags;
1187 local_irq_save(flags);
1188 if (pcp->count >= pcp->batch)
1189 to_drain = pcp->batch;
1191 to_drain = pcp->count;
1192 free_pcppages_bulk(zone, to_drain, pcp);
1193 pcp->count -= to_drain;
1194 local_irq_restore(flags);
1199 * Drain pages of the indicated processor.
1201 * The processor must either be the current processor and the
1202 * thread pinned to the current processor or a processor that
1205 static void drain_pages(unsigned int cpu)
1207 unsigned long flags;
1210 for_each_populated_zone(zone) {
1211 struct per_cpu_pageset *pset;
1212 struct per_cpu_pages *pcp;
1214 local_irq_save(flags);
1215 pset = per_cpu_ptr(zone->pageset, cpu);
1219 free_pcppages_bulk(zone, pcp->count, pcp);
1222 local_irq_restore(flags);
1227 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1229 void drain_local_pages(void *arg)
1231 drain_pages(smp_processor_id());
1235 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1237 void drain_all_pages(void)
1239 on_each_cpu(drain_local_pages, NULL, 1);
1242 #ifdef CONFIG_HIBERNATION
1244 void mark_free_pages(struct zone *zone)
1246 unsigned long pfn, max_zone_pfn;
1247 unsigned long flags;
1249 struct list_head *curr;
1251 if (!zone->spanned_pages)
1254 spin_lock_irqsave(&zone->lock, flags);
1256 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1257 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1258 if (pfn_valid(pfn)) {
1259 struct page *page = pfn_to_page(pfn);
1261 if (!swsusp_page_is_forbidden(page))
1262 swsusp_unset_page_free(page);
1265 for_each_migratetype_order(order, t) {
1266 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1269 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1270 for (i = 0; i < (1UL << order); i++)
1271 swsusp_set_page_free(pfn_to_page(pfn + i));
1274 spin_unlock_irqrestore(&zone->lock, flags);
1276 #endif /* CONFIG_PM */
1279 * Free a 0-order page
1280 * cold == 1 ? free a cold page : free a hot page
1282 void free_hot_cold_page(struct page *page, int cold)
1284 struct zone *zone = page_zone(page);
1285 struct per_cpu_pages *pcp;
1286 unsigned long flags;
1288 int wasMlocked = __TestClearPageMlocked(page);
1290 if (!free_pages_prepare(page, 0))
1293 migratetype = get_pageblock_migratetype(page);
1294 set_page_private(page, migratetype);
1295 local_irq_save(flags);
1296 if (unlikely(wasMlocked))
1297 free_page_mlock(page);
1298 __count_vm_event(PGFREE);
1301 * We only track unmovable, reclaimable and movable on pcp lists.
1302 * Free ISOLATE pages back to the allocator because they are being
1303 * offlined but treat RESERVE as movable pages so we can get those
1304 * areas back if necessary. Otherwise, we may have to free
1305 * excessively into the page allocator
1307 if (migratetype >= MIGRATE_PCPTYPES) {
1308 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1309 free_one_page(zone, page, 0, migratetype);
1312 migratetype = MIGRATE_MOVABLE;
1315 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1317 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1319 list_add(&page->lru, &pcp->lists[migratetype]);
1321 if (pcp->count >= pcp->high) {
1322 free_pcppages_bulk(zone, pcp->batch, pcp);
1323 pcp->count -= pcp->batch;
1327 local_irq_restore(flags);
1331 * Free a list of 0-order pages
1333 void free_hot_cold_page_list(struct list_head *list, int cold)
1335 struct page *page, *next;
1337 list_for_each_entry_safe(page, next, list, lru) {
1338 trace_mm_page_free_batched(page, cold);
1339 free_hot_cold_page(page, cold);
1344 * split_page takes a non-compound higher-order page, and splits it into
1345 * n (1<<order) sub-pages: page[0..n]
1346 * Each sub-page must be freed individually.
1348 * Note: this is probably too low level an operation for use in drivers.
1349 * Please consult with lkml before using this in your driver.
1351 void split_page(struct page *page, unsigned int order)
1355 VM_BUG_ON(PageCompound(page));
1356 VM_BUG_ON(!page_count(page));
1358 #ifdef CONFIG_KMEMCHECK
1360 * Split shadow pages too, because free(page[0]) would
1361 * otherwise free the whole shadow.
1363 if (kmemcheck_page_is_tracked(page))
1364 split_page(virt_to_page(page[0].shadow), order);
1367 for (i = 1; i < (1 << order); i++)
1368 set_page_refcounted(page + i);
1372 * Similar to split_page except the page is already free. As this is only
1373 * being used for migration, the migratetype of the block also changes.
1374 * As this is called with interrupts disabled, the caller is responsible
1375 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1378 * Note: this is probably too low level an operation for use in drivers.
1379 * Please consult with lkml before using this in your driver.
1381 int split_free_page(struct page *page)
1384 unsigned long watermark;
1387 BUG_ON(!PageBuddy(page));
1389 zone = page_zone(page);
1390 order = page_order(page);
1392 /* Obey watermarks as if the page was being allocated */
1393 watermark = low_wmark_pages(zone) + (1 << order);
1394 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1397 /* Remove page from free list */
1398 list_del(&page->lru);
1399 zone->free_area[order].nr_free--;
1400 rmv_page_order(page);
1401 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1403 /* Split into individual pages */
1404 set_page_refcounted(page);
1405 split_page(page, order);
1407 if (order >= pageblock_order - 1) {
1408 struct page *endpage = page + (1 << order) - 1;
1409 for (; page < endpage; page += pageblock_nr_pages) {
1410 int mt = get_pageblock_migratetype(page);
1411 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1412 set_pageblock_migratetype(page,
1421 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1422 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1426 struct page *buffered_rmqueue(struct zone *preferred_zone,
1427 struct zone *zone, int order, gfp_t gfp_flags,
1430 unsigned long flags;
1432 int cold = !!(gfp_flags & __GFP_COLD);
1435 if (likely(order == 0)) {
1436 struct per_cpu_pages *pcp;
1437 struct list_head *list;
1439 local_irq_save(flags);
1440 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1441 list = &pcp->lists[migratetype];
1442 if (list_empty(list)) {
1443 pcp->count += rmqueue_bulk(zone, 0,
1446 if (unlikely(list_empty(list)))
1451 page = list_entry(list->prev, struct page, lru);
1453 page = list_entry(list->next, struct page, lru);
1455 list_del(&page->lru);
1458 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1460 * __GFP_NOFAIL is not to be used in new code.
1462 * All __GFP_NOFAIL callers should be fixed so that they
1463 * properly detect and handle allocation failures.
1465 * We most definitely don't want callers attempting to
1466 * allocate greater than order-1 page units with
1469 WARN_ON_ONCE(order > 1);
1471 spin_lock_irqsave(&zone->lock, flags);
1472 page = __rmqueue(zone, order, migratetype);
1473 spin_unlock(&zone->lock);
1476 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1479 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1480 zone_statistics(preferred_zone, zone, gfp_flags);
1481 local_irq_restore(flags);
1483 VM_BUG_ON(bad_range(zone, page));
1484 if (prep_new_page(page, order, gfp_flags))
1489 local_irq_restore(flags);
1493 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1494 #define ALLOC_WMARK_MIN WMARK_MIN
1495 #define ALLOC_WMARK_LOW WMARK_LOW
1496 #define ALLOC_WMARK_HIGH WMARK_HIGH
1497 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1499 /* Mask to get the watermark bits */
1500 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1502 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1503 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1504 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1506 #ifdef CONFIG_FAIL_PAGE_ALLOC
1509 struct fault_attr attr;
1511 u32 ignore_gfp_highmem;
1512 u32 ignore_gfp_wait;
1514 } fail_page_alloc = {
1515 .attr = FAULT_ATTR_INITIALIZER,
1516 .ignore_gfp_wait = 1,
1517 .ignore_gfp_highmem = 1,
1521 static int __init setup_fail_page_alloc(char *str)
1523 return setup_fault_attr(&fail_page_alloc.attr, str);
1525 __setup("fail_page_alloc=", setup_fail_page_alloc);
1527 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1529 if (order < fail_page_alloc.min_order)
1531 if (gfp_mask & __GFP_NOFAIL)
1533 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1535 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1538 return should_fail(&fail_page_alloc.attr, 1 << order);
1541 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1543 static int __init fail_page_alloc_debugfs(void)
1545 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1548 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1549 &fail_page_alloc.attr);
1551 return PTR_ERR(dir);
1553 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1554 &fail_page_alloc.ignore_gfp_wait))
1556 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1557 &fail_page_alloc.ignore_gfp_highmem))
1559 if (!debugfs_create_u32("min-order", mode, dir,
1560 &fail_page_alloc.min_order))
1565 debugfs_remove_recursive(dir);
1570 late_initcall(fail_page_alloc_debugfs);
1572 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1574 #else /* CONFIG_FAIL_PAGE_ALLOC */
1576 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1581 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1584 * Return true if free pages are above 'mark'. This takes into account the order
1585 * of the allocation.
1587 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1588 int classzone_idx, int alloc_flags, long free_pages)
1590 /* free_pages my go negative - that's OK */
1594 free_pages -= (1 << order) - 1;
1595 if (alloc_flags & ALLOC_HIGH)
1597 if (alloc_flags & ALLOC_HARDER)
1600 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1602 for (o = 0; o < order; o++) {
1603 /* At the next order, this order's pages become unavailable */
1604 free_pages -= z->free_area[o].nr_free << o;
1606 /* Require fewer higher order pages to be free */
1609 if (free_pages <= min)
1615 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1616 int classzone_idx, int alloc_flags)
1618 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1619 zone_page_state(z, NR_FREE_PAGES));
1622 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1623 int classzone_idx, int alloc_flags)
1625 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1627 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1628 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1630 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1636 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1637 * skip over zones that are not allowed by the cpuset, or that have
1638 * been recently (in last second) found to be nearly full. See further
1639 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1640 * that have to skip over a lot of full or unallowed zones.
1642 * If the zonelist cache is present in the passed in zonelist, then
1643 * returns a pointer to the allowed node mask (either the current
1644 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1646 * If the zonelist cache is not available for this zonelist, does
1647 * nothing and returns NULL.
1649 * If the fullzones BITMAP in the zonelist cache is stale (more than
1650 * a second since last zap'd) then we zap it out (clear its bits.)
1652 * We hold off even calling zlc_setup, until after we've checked the
1653 * first zone in the zonelist, on the theory that most allocations will
1654 * be satisfied from that first zone, so best to examine that zone as
1655 * quickly as we can.
1657 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1659 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1660 nodemask_t *allowednodes; /* zonelist_cache approximation */
1662 zlc = zonelist->zlcache_ptr;
1666 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1667 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1668 zlc->last_full_zap = jiffies;
1671 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1672 &cpuset_current_mems_allowed :
1673 &node_states[N_HIGH_MEMORY];
1674 return allowednodes;
1678 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1679 * if it is worth looking at further for free memory:
1680 * 1) Check that the zone isn't thought to be full (doesn't have its
1681 * bit set in the zonelist_cache fullzones BITMAP).
1682 * 2) Check that the zones node (obtained from the zonelist_cache
1683 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1684 * Return true (non-zero) if zone is worth looking at further, or
1685 * else return false (zero) if it is not.
1687 * This check -ignores- the distinction between various watermarks,
1688 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1689 * found to be full for any variation of these watermarks, it will
1690 * be considered full for up to one second by all requests, unless
1691 * we are so low on memory on all allowed nodes that we are forced
1692 * into the second scan of the zonelist.
1694 * In the second scan we ignore this zonelist cache and exactly
1695 * apply the watermarks to all zones, even it is slower to do so.
1696 * We are low on memory in the second scan, and should leave no stone
1697 * unturned looking for a free page.
1699 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1700 nodemask_t *allowednodes)
1702 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1703 int i; /* index of *z in zonelist zones */
1704 int n; /* node that zone *z is on */
1706 zlc = zonelist->zlcache_ptr;
1710 i = z - zonelist->_zonerefs;
1713 /* This zone is worth trying if it is allowed but not full */
1714 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1718 * Given 'z' scanning a zonelist, set the corresponding bit in
1719 * zlc->fullzones, so that subsequent attempts to allocate a page
1720 * from that zone don't waste time re-examining it.
1722 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1724 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1725 int i; /* index of *z in zonelist zones */
1727 zlc = zonelist->zlcache_ptr;
1731 i = z - zonelist->_zonerefs;
1733 set_bit(i, zlc->fullzones);
1737 * clear all zones full, called after direct reclaim makes progress so that
1738 * a zone that was recently full is not skipped over for up to a second
1740 static void zlc_clear_zones_full(struct zonelist *zonelist)
1742 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1744 zlc = zonelist->zlcache_ptr;
1748 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1751 #else /* CONFIG_NUMA */
1753 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1758 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1759 nodemask_t *allowednodes)
1764 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1768 static void zlc_clear_zones_full(struct zonelist *zonelist)
1771 #endif /* CONFIG_NUMA */
1774 * get_page_from_freelist goes through the zonelist trying to allocate
1777 static struct page *
1778 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1779 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1780 struct zone *preferred_zone, int migratetype)
1783 struct page *page = NULL;
1786 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1787 int zlc_active = 0; /* set if using zonelist_cache */
1788 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1790 classzone_idx = zone_idx(preferred_zone);
1793 * Scan zonelist, looking for a zone with enough free.
1794 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1796 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1797 high_zoneidx, nodemask) {
1798 if (NUMA_BUILD && zlc_active &&
1799 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1801 if ((alloc_flags & ALLOC_CPUSET) &&
1802 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1805 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1806 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1810 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1811 if (zone_watermark_ok(zone, order, mark,
1812 classzone_idx, alloc_flags))
1815 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1817 * we do zlc_setup if there are multiple nodes
1818 * and before considering the first zone allowed
1821 allowednodes = zlc_setup(zonelist, alloc_flags);
1826 if (zone_reclaim_mode == 0)
1827 goto this_zone_full;
1830 * As we may have just activated ZLC, check if the first
1831 * eligible zone has failed zone_reclaim recently.
1833 if (NUMA_BUILD && zlc_active &&
1834 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1837 ret = zone_reclaim(zone, gfp_mask, order);
1839 case ZONE_RECLAIM_NOSCAN:
1842 case ZONE_RECLAIM_FULL:
1843 /* scanned but unreclaimable */
1846 /* did we reclaim enough */
1847 if (!zone_watermark_ok(zone, order, mark,
1848 classzone_idx, alloc_flags))
1849 goto this_zone_full;
1854 page = buffered_rmqueue(preferred_zone, zone, order,
1855 gfp_mask, migratetype);
1860 zlc_mark_zone_full(zonelist, z);
1863 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1864 /* Disable zlc cache for second zonelist scan */
1872 * Large machines with many possible nodes should not always dump per-node
1873 * meminfo in irq context.
1875 static inline bool should_suppress_show_mem(void)
1880 ret = in_interrupt();
1885 static DEFINE_RATELIMIT_STATE(nopage_rs,
1886 DEFAULT_RATELIMIT_INTERVAL,
1887 DEFAULT_RATELIMIT_BURST);
1889 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1891 unsigned int filter = SHOW_MEM_FILTER_NODES;
1893 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1894 debug_guardpage_minorder() > 0)
1898 * This documents exceptions given to allocations in certain
1899 * contexts that are allowed to allocate outside current's set
1902 if (!(gfp_mask & __GFP_NOMEMALLOC))
1903 if (test_thread_flag(TIF_MEMDIE) ||
1904 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1905 filter &= ~SHOW_MEM_FILTER_NODES;
1906 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1907 filter &= ~SHOW_MEM_FILTER_NODES;
1910 struct va_format vaf;
1913 va_start(args, fmt);
1918 pr_warn("%pV", &vaf);
1923 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1924 current->comm, order, gfp_mask);
1927 if (!should_suppress_show_mem())
1932 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1933 unsigned long did_some_progress,
1934 unsigned long pages_reclaimed)
1936 /* Do not loop if specifically requested */
1937 if (gfp_mask & __GFP_NORETRY)
1940 /* Always retry if specifically requested */
1941 if (gfp_mask & __GFP_NOFAIL)
1945 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
1946 * making forward progress without invoking OOM. Suspend also disables
1947 * storage devices so kswapd will not help. Bail if we are suspending.
1949 if (!did_some_progress && pm_suspended_storage())
1953 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1954 * means __GFP_NOFAIL, but that may not be true in other
1957 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1961 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1962 * specified, then we retry until we no longer reclaim any pages
1963 * (above), or we've reclaimed an order of pages at least as
1964 * large as the allocation's order. In both cases, if the
1965 * allocation still fails, we stop retrying.
1967 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1973 static inline struct page *
1974 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1975 struct zonelist *zonelist, enum zone_type high_zoneidx,
1976 nodemask_t *nodemask, struct zone *preferred_zone,
1981 /* Acquire the OOM killer lock for the zones in zonelist */
1982 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1983 schedule_timeout_uninterruptible(1);
1988 * Go through the zonelist yet one more time, keep very high watermark
1989 * here, this is only to catch a parallel oom killing, we must fail if
1990 * we're still under heavy pressure.
1992 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1993 order, zonelist, high_zoneidx,
1994 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1995 preferred_zone, migratetype);
1999 if (!(gfp_mask & __GFP_NOFAIL)) {
2000 /* The OOM killer will not help higher order allocs */
2001 if (order > PAGE_ALLOC_COSTLY_ORDER)
2003 /* The OOM killer does not needlessly kill tasks for lowmem */
2004 if (high_zoneidx < ZONE_NORMAL)
2007 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2008 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2009 * The caller should handle page allocation failure by itself if
2010 * it specifies __GFP_THISNODE.
2011 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2013 if (gfp_mask & __GFP_THISNODE)
2016 /* Exhausted what can be done so it's blamo time */
2017 out_of_memory(zonelist, gfp_mask, order, nodemask);
2020 clear_zonelist_oom(zonelist, gfp_mask);
2024 #ifdef CONFIG_COMPACTION
2025 /* Try memory compaction for high-order allocations before reclaim */
2026 static struct page *
2027 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2028 struct zonelist *zonelist, enum zone_type high_zoneidx,
2029 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2030 int migratetype, bool sync_migration,
2031 bool *deferred_compaction,
2032 unsigned long *did_some_progress)
2039 if (compaction_deferred(preferred_zone)) {
2040 *deferred_compaction = true;
2044 current->flags |= PF_MEMALLOC;
2045 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2046 nodemask, sync_migration);
2047 current->flags &= ~PF_MEMALLOC;
2048 if (*did_some_progress != COMPACT_SKIPPED) {
2050 /* Page migration frees to the PCP lists but we want merging */
2051 drain_pages(get_cpu());
2054 page = get_page_from_freelist(gfp_mask, nodemask,
2055 order, zonelist, high_zoneidx,
2056 alloc_flags, preferred_zone,
2059 preferred_zone->compact_considered = 0;
2060 preferred_zone->compact_defer_shift = 0;
2061 count_vm_event(COMPACTSUCCESS);
2066 * It's bad if compaction run occurs and fails.
2067 * The most likely reason is that pages exist,
2068 * but not enough to satisfy watermarks.
2070 count_vm_event(COMPACTFAIL);
2073 * As async compaction considers a subset of pageblocks, only
2074 * defer if the failure was a sync compaction failure.
2077 defer_compaction(preferred_zone);
2085 static inline struct page *
2086 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2087 struct zonelist *zonelist, enum zone_type high_zoneidx,
2088 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2089 int migratetype, bool sync_migration,
2090 bool *deferred_compaction,
2091 unsigned long *did_some_progress)
2095 #endif /* CONFIG_COMPACTION */
2097 /* Perform direct synchronous page reclaim */
2099 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2100 nodemask_t *nodemask)
2102 struct reclaim_state reclaim_state;
2107 /* We now go into synchronous reclaim */
2108 cpuset_memory_pressure_bump();
2109 current->flags |= PF_MEMALLOC;
2110 lockdep_set_current_reclaim_state(gfp_mask);
2111 reclaim_state.reclaimed_slab = 0;
2112 current->reclaim_state = &reclaim_state;
2114 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2116 current->reclaim_state = NULL;
2117 lockdep_clear_current_reclaim_state();
2118 current->flags &= ~PF_MEMALLOC;
2125 /* The really slow allocator path where we enter direct reclaim */
2126 static inline struct page *
2127 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2128 struct zonelist *zonelist, enum zone_type high_zoneidx,
2129 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2130 int migratetype, unsigned long *did_some_progress)
2132 struct page *page = NULL;
2133 bool drained = false;
2135 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2137 if (unlikely(!(*did_some_progress)))
2140 /* After successful reclaim, reconsider all zones for allocation */
2142 zlc_clear_zones_full(zonelist);
2145 page = get_page_from_freelist(gfp_mask, nodemask, order,
2146 zonelist, high_zoneidx,
2147 alloc_flags, preferred_zone,
2151 * If an allocation failed after direct reclaim, it could be because
2152 * pages are pinned on the per-cpu lists. Drain them and try again
2154 if (!page && !drained) {
2164 * This is called in the allocator slow-path if the allocation request is of
2165 * sufficient urgency to ignore watermarks and take other desperate measures
2167 static inline struct page *
2168 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2169 struct zonelist *zonelist, enum zone_type high_zoneidx,
2170 nodemask_t *nodemask, struct zone *preferred_zone,
2176 page = get_page_from_freelist(gfp_mask, nodemask, order,
2177 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2178 preferred_zone, migratetype);
2180 if (!page && gfp_mask & __GFP_NOFAIL)
2181 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2182 } while (!page && (gfp_mask & __GFP_NOFAIL));
2188 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2189 enum zone_type high_zoneidx,
2190 enum zone_type classzone_idx)
2195 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2196 wakeup_kswapd(zone, order, classzone_idx);
2200 gfp_to_alloc_flags(gfp_t gfp_mask)
2202 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2203 const gfp_t wait = gfp_mask & __GFP_WAIT;
2205 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2206 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2209 * The caller may dip into page reserves a bit more if the caller
2210 * cannot run direct reclaim, or if the caller has realtime scheduling
2211 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2212 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2214 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2218 * Not worth trying to allocate harder for
2219 * __GFP_NOMEMALLOC even if it can't schedule.
2221 if (!(gfp_mask & __GFP_NOMEMALLOC))
2222 alloc_flags |= ALLOC_HARDER;
2224 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2225 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2227 alloc_flags &= ~ALLOC_CPUSET;
2228 } else if (unlikely(rt_task(current)) && !in_interrupt())
2229 alloc_flags |= ALLOC_HARDER;
2231 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2232 if (!in_interrupt() &&
2233 ((current->flags & PF_MEMALLOC) ||
2234 unlikely(test_thread_flag(TIF_MEMDIE))))
2235 alloc_flags |= ALLOC_NO_WATERMARKS;
2241 static inline struct page *
2242 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2243 struct zonelist *zonelist, enum zone_type high_zoneidx,
2244 nodemask_t *nodemask, struct zone *preferred_zone,
2247 const gfp_t wait = gfp_mask & __GFP_WAIT;
2248 struct page *page = NULL;
2250 unsigned long pages_reclaimed = 0;
2251 unsigned long did_some_progress;
2252 bool sync_migration = false;
2253 bool deferred_compaction = false;
2256 * In the slowpath, we sanity check order to avoid ever trying to
2257 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2258 * be using allocators in order of preference for an area that is
2261 if (order >= MAX_ORDER) {
2262 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2267 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2268 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2269 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2270 * using a larger set of nodes after it has established that the
2271 * allowed per node queues are empty and that nodes are
2274 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2278 if (!(gfp_mask & __GFP_NO_KSWAPD))
2279 wake_all_kswapd(order, zonelist, high_zoneidx,
2280 zone_idx(preferred_zone));
2283 * OK, we're below the kswapd watermark and have kicked background
2284 * reclaim. Now things get more complex, so set up alloc_flags according
2285 * to how we want to proceed.
2287 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2290 * Find the true preferred zone if the allocation is unconstrained by
2293 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2294 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2298 /* This is the last chance, in general, before the goto nopage. */
2299 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2300 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2301 preferred_zone, migratetype);
2305 /* Allocate without watermarks if the context allows */
2306 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2307 page = __alloc_pages_high_priority(gfp_mask, order,
2308 zonelist, high_zoneidx, nodemask,
2309 preferred_zone, migratetype);
2314 /* Atomic allocations - we can't balance anything */
2318 /* Avoid recursion of direct reclaim */
2319 if (current->flags & PF_MEMALLOC)
2322 /* Avoid allocations with no watermarks from looping endlessly */
2323 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2327 * Try direct compaction. The first pass is asynchronous. Subsequent
2328 * attempts after direct reclaim are synchronous
2330 page = __alloc_pages_direct_compact(gfp_mask, order,
2331 zonelist, high_zoneidx,
2333 alloc_flags, preferred_zone,
2334 migratetype, sync_migration,
2335 &deferred_compaction,
2336 &did_some_progress);
2339 sync_migration = true;
2342 * If compaction is deferred for high-order allocations, it is because
2343 * sync compaction recently failed. In this is the case and the caller
2344 * has requested the system not be heavily disrupted, fail the
2345 * allocation now instead of entering direct reclaim
2347 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2350 /* Try direct reclaim and then allocating */
2351 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2352 zonelist, high_zoneidx,
2354 alloc_flags, preferred_zone,
2355 migratetype, &did_some_progress);
2360 * If we failed to make any progress reclaiming, then we are
2361 * running out of options and have to consider going OOM
2363 if (!did_some_progress) {
2364 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2365 if (oom_killer_disabled)
2367 page = __alloc_pages_may_oom(gfp_mask, order,
2368 zonelist, high_zoneidx,
2369 nodemask, preferred_zone,
2374 if (!(gfp_mask & __GFP_NOFAIL)) {
2376 * The oom killer is not called for high-order
2377 * allocations that may fail, so if no progress
2378 * is being made, there are no other options and
2379 * retrying is unlikely to help.
2381 if (order > PAGE_ALLOC_COSTLY_ORDER)
2384 * The oom killer is not called for lowmem
2385 * allocations to prevent needlessly killing
2388 if (high_zoneidx < ZONE_NORMAL)
2396 /* Check if we should retry the allocation */
2397 pages_reclaimed += did_some_progress;
2398 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2400 /* Wait for some write requests to complete then retry */
2401 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2405 * High-order allocations do not necessarily loop after
2406 * direct reclaim and reclaim/compaction depends on compaction
2407 * being called after reclaim so call directly if necessary
2409 page = __alloc_pages_direct_compact(gfp_mask, order,
2410 zonelist, high_zoneidx,
2412 alloc_flags, preferred_zone,
2413 migratetype, sync_migration,
2414 &deferred_compaction,
2415 &did_some_progress);
2421 warn_alloc_failed(gfp_mask, order, NULL);
2424 if (kmemcheck_enabled)
2425 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2431 * This is the 'heart' of the zoned buddy allocator.
2434 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2435 struct zonelist *zonelist, nodemask_t *nodemask)
2437 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2438 struct zone *preferred_zone;
2439 struct page *page = NULL;
2440 int migratetype = allocflags_to_migratetype(gfp_mask);
2441 unsigned int cpuset_mems_cookie;
2443 gfp_mask &= gfp_allowed_mask;
2445 lockdep_trace_alloc(gfp_mask);
2447 might_sleep_if(gfp_mask & __GFP_WAIT);
2449 if (should_fail_alloc_page(gfp_mask, order))
2453 * Check the zones suitable for the gfp_mask contain at least one
2454 * valid zone. It's possible to have an empty zonelist as a result
2455 * of GFP_THISNODE and a memoryless node
2457 if (unlikely(!zonelist->_zonerefs->zone))
2461 cpuset_mems_cookie = get_mems_allowed();
2463 /* The preferred zone is used for statistics later */
2464 first_zones_zonelist(zonelist, high_zoneidx,
2465 nodemask ? : &cpuset_current_mems_allowed,
2467 if (!preferred_zone)
2470 /* First allocation attempt */
2471 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2472 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2473 preferred_zone, migratetype);
2474 if (unlikely(!page))
2475 page = __alloc_pages_slowpath(gfp_mask, order,
2476 zonelist, high_zoneidx, nodemask,
2477 preferred_zone, migratetype);
2479 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2483 * When updating a task's mems_allowed, it is possible to race with
2484 * parallel threads in such a way that an allocation can fail while
2485 * the mask is being updated. If a page allocation is about to fail,
2486 * check if the cpuset changed during allocation and if so, retry.
2488 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2493 EXPORT_SYMBOL(__alloc_pages_nodemask);
2496 * Common helper functions.
2498 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2503 * __get_free_pages() returns a 32-bit address, which cannot represent
2506 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2508 page = alloc_pages(gfp_mask, order);
2511 return (unsigned long) page_address(page);
2513 EXPORT_SYMBOL(__get_free_pages);
2515 unsigned long get_zeroed_page(gfp_t gfp_mask)
2517 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2519 EXPORT_SYMBOL(get_zeroed_page);
2521 void __free_pages(struct page *page, unsigned int order)
2523 if (put_page_testzero(page)) {
2525 free_hot_cold_page(page, 0);
2527 __free_pages_ok(page, order);
2531 EXPORT_SYMBOL(__free_pages);
2533 void free_pages(unsigned long addr, unsigned int order)
2536 VM_BUG_ON(!virt_addr_valid((void *)addr));
2537 __free_pages(virt_to_page((void *)addr), order);
2541 EXPORT_SYMBOL(free_pages);
2543 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2546 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2547 unsigned long used = addr + PAGE_ALIGN(size);
2549 split_page(virt_to_page((void *)addr), order);
2550 while (used < alloc_end) {
2555 return (void *)addr;
2559 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2560 * @size: the number of bytes to allocate
2561 * @gfp_mask: GFP flags for the allocation
2563 * This function is similar to alloc_pages(), except that it allocates the
2564 * minimum number of pages to satisfy the request. alloc_pages() can only
2565 * allocate memory in power-of-two pages.
2567 * This function is also limited by MAX_ORDER.
2569 * Memory allocated by this function must be released by free_pages_exact().
2571 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2573 unsigned int order = get_order(size);
2576 addr = __get_free_pages(gfp_mask, order);
2577 return make_alloc_exact(addr, order, size);
2579 EXPORT_SYMBOL(alloc_pages_exact);
2582 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2584 * @nid: the preferred node ID where memory should be allocated
2585 * @size: the number of bytes to allocate
2586 * @gfp_mask: GFP flags for the allocation
2588 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2590 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2593 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2595 unsigned order = get_order(size);
2596 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2599 return make_alloc_exact((unsigned long)page_address(p), order, size);
2601 EXPORT_SYMBOL(alloc_pages_exact_nid);
2604 * free_pages_exact - release memory allocated via alloc_pages_exact()
2605 * @virt: the value returned by alloc_pages_exact.
2606 * @size: size of allocation, same value as passed to alloc_pages_exact().
2608 * Release the memory allocated by a previous call to alloc_pages_exact.
2610 void free_pages_exact(void *virt, size_t size)
2612 unsigned long addr = (unsigned long)virt;
2613 unsigned long end = addr + PAGE_ALIGN(size);
2615 while (addr < end) {
2620 EXPORT_SYMBOL(free_pages_exact);
2622 static unsigned int nr_free_zone_pages(int offset)
2627 /* Just pick one node, since fallback list is circular */
2628 unsigned int sum = 0;
2630 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2632 for_each_zone_zonelist(zone, z, zonelist, offset) {
2633 unsigned long size = zone->present_pages;
2634 unsigned long high = high_wmark_pages(zone);
2643 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2645 unsigned int nr_free_buffer_pages(void)
2647 return nr_free_zone_pages(gfp_zone(GFP_USER));
2649 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2652 * Amount of free RAM allocatable within all zones
2654 unsigned int nr_free_pagecache_pages(void)
2656 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2659 static inline void show_node(struct zone *zone)
2662 printk("Node %d ", zone_to_nid(zone));
2665 void si_meminfo(struct sysinfo *val)
2667 val->totalram = totalram_pages;
2669 val->freeram = global_page_state(NR_FREE_PAGES);
2670 val->bufferram = nr_blockdev_pages();
2671 val->totalhigh = totalhigh_pages;
2672 val->freehigh = nr_free_highpages();
2673 val->mem_unit = PAGE_SIZE;
2676 EXPORT_SYMBOL(si_meminfo);
2679 void si_meminfo_node(struct sysinfo *val, int nid)
2681 pg_data_t *pgdat = NODE_DATA(nid);
2683 val->totalram = pgdat->node_present_pages;
2684 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2685 #ifdef CONFIG_HIGHMEM
2686 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2687 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2693 val->mem_unit = PAGE_SIZE;
2698 * Determine whether the node should be displayed or not, depending on whether
2699 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2701 bool skip_free_areas_node(unsigned int flags, int nid)
2704 unsigned int cpuset_mems_cookie;
2706 if (!(flags & SHOW_MEM_FILTER_NODES))
2710 cpuset_mems_cookie = get_mems_allowed();
2711 ret = !node_isset(nid, cpuset_current_mems_allowed);
2712 } while (!put_mems_allowed(cpuset_mems_cookie));
2717 #define K(x) ((x) << (PAGE_SHIFT-10))
2720 * Show free area list (used inside shift_scroll-lock stuff)
2721 * We also calculate the percentage fragmentation. We do this by counting the
2722 * memory on each free list with the exception of the first item on the list.
2723 * Suppresses nodes that are not allowed by current's cpuset if
2724 * SHOW_MEM_FILTER_NODES is passed.
2726 void show_free_areas(unsigned int filter)
2731 for_each_populated_zone(zone) {
2732 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2735 printk("%s per-cpu:\n", zone->name);
2737 for_each_online_cpu(cpu) {
2738 struct per_cpu_pageset *pageset;
2740 pageset = per_cpu_ptr(zone->pageset, cpu);
2742 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2743 cpu, pageset->pcp.high,
2744 pageset->pcp.batch, pageset->pcp.count);
2748 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2749 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2751 " dirty:%lu writeback:%lu unstable:%lu\n"
2752 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2753 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2754 global_page_state(NR_ACTIVE_ANON),
2755 global_page_state(NR_INACTIVE_ANON),
2756 global_page_state(NR_ISOLATED_ANON),
2757 global_page_state(NR_ACTIVE_FILE),
2758 global_page_state(NR_INACTIVE_FILE),
2759 global_page_state(NR_ISOLATED_FILE),
2760 global_page_state(NR_UNEVICTABLE),
2761 global_page_state(NR_FILE_DIRTY),
2762 global_page_state(NR_WRITEBACK),
2763 global_page_state(NR_UNSTABLE_NFS),
2764 global_page_state(NR_FREE_PAGES),
2765 global_page_state(NR_SLAB_RECLAIMABLE),
2766 global_page_state(NR_SLAB_UNRECLAIMABLE),
2767 global_page_state(NR_FILE_MAPPED),
2768 global_page_state(NR_SHMEM),
2769 global_page_state(NR_PAGETABLE),
2770 global_page_state(NR_BOUNCE));
2772 for_each_populated_zone(zone) {
2775 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2783 " active_anon:%lukB"
2784 " inactive_anon:%lukB"
2785 " active_file:%lukB"
2786 " inactive_file:%lukB"
2787 " unevictable:%lukB"
2788 " isolated(anon):%lukB"
2789 " isolated(file):%lukB"
2796 " slab_reclaimable:%lukB"
2797 " slab_unreclaimable:%lukB"
2798 " kernel_stack:%lukB"
2802 " writeback_tmp:%lukB"
2803 " pages_scanned:%lu"
2804 " all_unreclaimable? %s"
2807 K(zone_page_state(zone, NR_FREE_PAGES)),
2808 K(min_wmark_pages(zone)),
2809 K(low_wmark_pages(zone)),
2810 K(high_wmark_pages(zone)),
2811 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2812 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2813 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2814 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2815 K(zone_page_state(zone, NR_UNEVICTABLE)),
2816 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2817 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2818 K(zone->present_pages),
2819 K(zone_page_state(zone, NR_MLOCK)),
2820 K(zone_page_state(zone, NR_FILE_DIRTY)),
2821 K(zone_page_state(zone, NR_WRITEBACK)),
2822 K(zone_page_state(zone, NR_FILE_MAPPED)),
2823 K(zone_page_state(zone, NR_SHMEM)),
2824 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2825 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2826 zone_page_state(zone, NR_KERNEL_STACK) *
2828 K(zone_page_state(zone, NR_PAGETABLE)),
2829 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2830 K(zone_page_state(zone, NR_BOUNCE)),
2831 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2832 zone->pages_scanned,
2833 (zone->all_unreclaimable ? "yes" : "no")
2835 printk("lowmem_reserve[]:");
2836 for (i = 0; i < MAX_NR_ZONES; i++)
2837 printk(" %lu", zone->lowmem_reserve[i]);
2841 for_each_populated_zone(zone) {
2842 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2844 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2847 printk("%s: ", zone->name);
2849 spin_lock_irqsave(&zone->lock, flags);
2850 for (order = 0; order < MAX_ORDER; order++) {
2851 nr[order] = zone->free_area[order].nr_free;
2852 total += nr[order] << order;
2854 spin_unlock_irqrestore(&zone->lock, flags);
2855 for (order = 0; order < MAX_ORDER; order++)
2856 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2857 printk("= %lukB\n", K(total));
2860 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2862 show_swap_cache_info();
2865 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2867 zoneref->zone = zone;
2868 zoneref->zone_idx = zone_idx(zone);
2872 * Builds allocation fallback zone lists.
2874 * Add all populated zones of a node to the zonelist.
2876 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2877 int nr_zones, enum zone_type zone_type)
2881 BUG_ON(zone_type >= MAX_NR_ZONES);
2886 zone = pgdat->node_zones + zone_type;
2887 if (populated_zone(zone)) {
2888 zoneref_set_zone(zone,
2889 &zonelist->_zonerefs[nr_zones++]);
2890 check_highest_zone(zone_type);
2893 } while (zone_type);
2900 * 0 = automatic detection of better ordering.
2901 * 1 = order by ([node] distance, -zonetype)
2902 * 2 = order by (-zonetype, [node] distance)
2904 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2905 * the same zonelist. So only NUMA can configure this param.
2907 #define ZONELIST_ORDER_DEFAULT 0
2908 #define ZONELIST_ORDER_NODE 1
2909 #define ZONELIST_ORDER_ZONE 2
2911 /* zonelist order in the kernel.
2912 * set_zonelist_order() will set this to NODE or ZONE.
2914 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2915 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2919 /* The value user specified ....changed by config */
2920 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2921 /* string for sysctl */
2922 #define NUMA_ZONELIST_ORDER_LEN 16
2923 char numa_zonelist_order[16] = "default";
2926 * interface for configure zonelist ordering.
2927 * command line option "numa_zonelist_order"
2928 * = "[dD]efault - default, automatic configuration.
2929 * = "[nN]ode - order by node locality, then by zone within node
2930 * = "[zZ]one - order by zone, then by locality within zone
2933 static int __parse_numa_zonelist_order(char *s)
2935 if (*s == 'd' || *s == 'D') {
2936 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2937 } else if (*s == 'n' || *s == 'N') {
2938 user_zonelist_order = ZONELIST_ORDER_NODE;
2939 } else if (*s == 'z' || *s == 'Z') {
2940 user_zonelist_order = ZONELIST_ORDER_ZONE;
2943 "Ignoring invalid numa_zonelist_order value: "
2950 static __init int setup_numa_zonelist_order(char *s)
2957 ret = __parse_numa_zonelist_order(s);
2959 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2963 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2966 * sysctl handler for numa_zonelist_order
2968 int numa_zonelist_order_handler(ctl_table *table, int write,
2969 void __user *buffer, size_t *length,
2972 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2974 static DEFINE_MUTEX(zl_order_mutex);
2976 mutex_lock(&zl_order_mutex);
2978 strcpy(saved_string, (char*)table->data);
2979 ret = proc_dostring(table, write, buffer, length, ppos);
2983 int oldval = user_zonelist_order;
2984 if (__parse_numa_zonelist_order((char*)table->data)) {
2986 * bogus value. restore saved string
2988 strncpy((char*)table->data, saved_string,
2989 NUMA_ZONELIST_ORDER_LEN);
2990 user_zonelist_order = oldval;
2991 } else if (oldval != user_zonelist_order) {
2992 mutex_lock(&zonelists_mutex);
2993 build_all_zonelists(NULL);
2994 mutex_unlock(&zonelists_mutex);
2998 mutex_unlock(&zl_order_mutex);
3003 #define MAX_NODE_LOAD (nr_online_nodes)
3004 static int node_load[MAX_NUMNODES];
3007 * find_next_best_node - find the next node that should appear in a given node's fallback list
3008 * @node: node whose fallback list we're appending
3009 * @used_node_mask: nodemask_t of already used nodes
3011 * We use a number of factors to determine which is the next node that should
3012 * appear on a given node's fallback list. The node should not have appeared
3013 * already in @node's fallback list, and it should be the next closest node
3014 * according to the distance array (which contains arbitrary distance values
3015 * from each node to each node in the system), and should also prefer nodes
3016 * with no CPUs, since presumably they'll have very little allocation pressure
3017 * on them otherwise.
3018 * It returns -1 if no node is found.
3020 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3023 int min_val = INT_MAX;
3025 const struct cpumask *tmp = cpumask_of_node(0);
3027 /* Use the local node if we haven't already */
3028 if (!node_isset(node, *used_node_mask)) {
3029 node_set(node, *used_node_mask);
3033 for_each_node_state(n, N_HIGH_MEMORY) {
3035 /* Don't want a node to appear more than once */
3036 if (node_isset(n, *used_node_mask))
3039 /* Use the distance array to find the distance */
3040 val = node_distance(node, n);
3042 /* Penalize nodes under us ("prefer the next node") */
3045 /* Give preference to headless and unused nodes */
3046 tmp = cpumask_of_node(n);
3047 if (!cpumask_empty(tmp))
3048 val += PENALTY_FOR_NODE_WITH_CPUS;
3050 /* Slight preference for less loaded node */
3051 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3052 val += node_load[n];
3054 if (val < min_val) {
3061 node_set(best_node, *used_node_mask);
3068 * Build zonelists ordered by node and zones within node.
3069 * This results in maximum locality--normal zone overflows into local
3070 * DMA zone, if any--but risks exhausting DMA zone.
3072 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3075 struct zonelist *zonelist;
3077 zonelist = &pgdat->node_zonelists[0];
3078 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3080 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3082 zonelist->_zonerefs[j].zone = NULL;
3083 zonelist->_zonerefs[j].zone_idx = 0;
3087 * Build gfp_thisnode zonelists
3089 static void build_thisnode_zonelists(pg_data_t *pgdat)
3092 struct zonelist *zonelist;
3094 zonelist = &pgdat->node_zonelists[1];
3095 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3096 zonelist->_zonerefs[j].zone = NULL;
3097 zonelist->_zonerefs[j].zone_idx = 0;
3101 * Build zonelists ordered by zone and nodes within zones.
3102 * This results in conserving DMA zone[s] until all Normal memory is
3103 * exhausted, but results in overflowing to remote node while memory
3104 * may still exist in local DMA zone.
3106 static int node_order[MAX_NUMNODES];
3108 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3111 int zone_type; /* needs to be signed */
3113 struct zonelist *zonelist;
3115 zonelist = &pgdat->node_zonelists[0];
3117 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3118 for (j = 0; j < nr_nodes; j++) {
3119 node = node_order[j];
3120 z = &NODE_DATA(node)->node_zones[zone_type];
3121 if (populated_zone(z)) {
3123 &zonelist->_zonerefs[pos++]);
3124 check_highest_zone(zone_type);
3128 zonelist->_zonerefs[pos].zone = NULL;
3129 zonelist->_zonerefs[pos].zone_idx = 0;
3132 static int default_zonelist_order(void)
3135 unsigned long low_kmem_size,total_size;
3139 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3140 * If they are really small and used heavily, the system can fall
3141 * into OOM very easily.
3142 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3144 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3147 for_each_online_node(nid) {
3148 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3149 z = &NODE_DATA(nid)->node_zones[zone_type];
3150 if (populated_zone(z)) {
3151 if (zone_type < ZONE_NORMAL)
3152 low_kmem_size += z->present_pages;
3153 total_size += z->present_pages;
3154 } else if (zone_type == ZONE_NORMAL) {
3156 * If any node has only lowmem, then node order
3157 * is preferred to allow kernel allocations
3158 * locally; otherwise, they can easily infringe
3159 * on other nodes when there is an abundance of
3160 * lowmem available to allocate from.
3162 return ZONELIST_ORDER_NODE;
3166 if (!low_kmem_size || /* there are no DMA area. */
3167 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3168 return ZONELIST_ORDER_NODE;
3170 * look into each node's config.
3171 * If there is a node whose DMA/DMA32 memory is very big area on
3172 * local memory, NODE_ORDER may be suitable.
3174 average_size = total_size /
3175 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3176 for_each_online_node(nid) {
3179 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3180 z = &NODE_DATA(nid)->node_zones[zone_type];
3181 if (populated_zone(z)) {
3182 if (zone_type < ZONE_NORMAL)
3183 low_kmem_size += z->present_pages;
3184 total_size += z->present_pages;
3187 if (low_kmem_size &&
3188 total_size > average_size && /* ignore small node */
3189 low_kmem_size > total_size * 70/100)
3190 return ZONELIST_ORDER_NODE;
3192 return ZONELIST_ORDER_ZONE;
3195 static void set_zonelist_order(void)
3197 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3198 current_zonelist_order = default_zonelist_order();
3200 current_zonelist_order = user_zonelist_order;
3203 static void build_zonelists(pg_data_t *pgdat)
3207 nodemask_t used_mask;
3208 int local_node, prev_node;
3209 struct zonelist *zonelist;
3210 int order = current_zonelist_order;
3212 /* initialize zonelists */
3213 for (i = 0; i < MAX_ZONELISTS; i++) {
3214 zonelist = pgdat->node_zonelists + i;
3215 zonelist->_zonerefs[0].zone = NULL;
3216 zonelist->_zonerefs[0].zone_idx = 0;
3219 /* NUMA-aware ordering of nodes */
3220 local_node = pgdat->node_id;
3221 load = nr_online_nodes;
3222 prev_node = local_node;
3223 nodes_clear(used_mask);
3225 memset(node_order, 0, sizeof(node_order));
3228 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3229 int distance = node_distance(local_node, node);
3232 * If another node is sufficiently far away then it is better
3233 * to reclaim pages in a zone before going off node.
3235 if (distance > RECLAIM_DISTANCE)
3236 zone_reclaim_mode = 1;
3239 * We don't want to pressure a particular node.
3240 * So adding penalty to the first node in same
3241 * distance group to make it round-robin.
3243 if (distance != node_distance(local_node, prev_node))
3244 node_load[node] = load;
3248 if (order == ZONELIST_ORDER_NODE)
3249 build_zonelists_in_node_order(pgdat, node);
3251 node_order[j++] = node; /* remember order */
3254 if (order == ZONELIST_ORDER_ZONE) {
3255 /* calculate node order -- i.e., DMA last! */
3256 build_zonelists_in_zone_order(pgdat, j);
3259 build_thisnode_zonelists(pgdat);
3262 /* Construct the zonelist performance cache - see further mmzone.h */
3263 static void build_zonelist_cache(pg_data_t *pgdat)
3265 struct zonelist *zonelist;
3266 struct zonelist_cache *zlc;
3269 zonelist = &pgdat->node_zonelists[0];
3270 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3271 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3272 for (z = zonelist->_zonerefs; z->zone; z++)
3273 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3276 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3278 * Return node id of node used for "local" allocations.
3279 * I.e., first node id of first zone in arg node's generic zonelist.
3280 * Used for initializing percpu 'numa_mem', which is used primarily
3281 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3283 int local_memory_node(int node)
3287 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3288 gfp_zone(GFP_KERNEL),
3295 #else /* CONFIG_NUMA */
3297 static void set_zonelist_order(void)
3299 current_zonelist_order = ZONELIST_ORDER_ZONE;
3302 static void build_zonelists(pg_data_t *pgdat)
3304 int node, local_node;
3306 struct zonelist *zonelist;
3308 local_node = pgdat->node_id;
3310 zonelist = &pgdat->node_zonelists[0];
3311 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3314 * Now we build the zonelist so that it contains the zones
3315 * of all the other nodes.
3316 * We don't want to pressure a particular node, so when
3317 * building the zones for node N, we make sure that the
3318 * zones coming right after the local ones are those from
3319 * node N+1 (modulo N)
3321 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3322 if (!node_online(node))
3324 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3327 for (node = 0; node < local_node; node++) {
3328 if (!node_online(node))
3330 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3334 zonelist->_zonerefs[j].zone = NULL;
3335 zonelist->_zonerefs[j].zone_idx = 0;
3338 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3339 static void build_zonelist_cache(pg_data_t *pgdat)
3341 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3344 #endif /* CONFIG_NUMA */
3347 * Boot pageset table. One per cpu which is going to be used for all
3348 * zones and all nodes. The parameters will be set in such a way
3349 * that an item put on a list will immediately be handed over to
3350 * the buddy list. This is safe since pageset manipulation is done
3351 * with interrupts disabled.
3353 * The boot_pagesets must be kept even after bootup is complete for
3354 * unused processors and/or zones. They do play a role for bootstrapping
3355 * hotplugged processors.
3357 * zoneinfo_show() and maybe other functions do
3358 * not check if the processor is online before following the pageset pointer.
3359 * Other parts of the kernel may not check if the zone is available.
3361 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3362 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3363 static void setup_zone_pageset(struct zone *zone);
3366 * Global mutex to protect against size modification of zonelists
3367 * as well as to serialize pageset setup for the new populated zone.
3369 DEFINE_MUTEX(zonelists_mutex);
3371 /* return values int ....just for stop_machine() */
3372 static __init_refok int __build_all_zonelists(void *data)
3378 memset(node_load, 0, sizeof(node_load));
3380 for_each_online_node(nid) {
3381 pg_data_t *pgdat = NODE_DATA(nid);
3383 build_zonelists(pgdat);
3384 build_zonelist_cache(pgdat);
3388 * Initialize the boot_pagesets that are going to be used
3389 * for bootstrapping processors. The real pagesets for
3390 * each zone will be allocated later when the per cpu
3391 * allocator is available.
3393 * boot_pagesets are used also for bootstrapping offline
3394 * cpus if the system is already booted because the pagesets
3395 * are needed to initialize allocators on a specific cpu too.
3396 * F.e. the percpu allocator needs the page allocator which
3397 * needs the percpu allocator in order to allocate its pagesets
3398 * (a chicken-egg dilemma).
3400 for_each_possible_cpu(cpu) {
3401 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3403 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3405 * We now know the "local memory node" for each node--
3406 * i.e., the node of the first zone in the generic zonelist.
3407 * Set up numa_mem percpu variable for on-line cpus. During
3408 * boot, only the boot cpu should be on-line; we'll init the
3409 * secondary cpus' numa_mem as they come on-line. During
3410 * node/memory hotplug, we'll fixup all on-line cpus.
3412 if (cpu_online(cpu))
3413 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3421 * Called with zonelists_mutex held always
3422 * unless system_state == SYSTEM_BOOTING.
3424 void __ref build_all_zonelists(void *data)
3426 set_zonelist_order();
3428 if (system_state == SYSTEM_BOOTING) {
3429 __build_all_zonelists(NULL);
3430 mminit_verify_zonelist();
3431 cpuset_init_current_mems_allowed();
3433 /* we have to stop all cpus to guarantee there is no user
3435 #ifdef CONFIG_MEMORY_HOTPLUG
3437 setup_zone_pageset((struct zone *)data);
3439 stop_machine(__build_all_zonelists, NULL, NULL);
3440 /* cpuset refresh routine should be here */
3442 vm_total_pages = nr_free_pagecache_pages();
3444 * Disable grouping by mobility if the number of pages in the
3445 * system is too low to allow the mechanism to work. It would be
3446 * more accurate, but expensive to check per-zone. This check is
3447 * made on memory-hotadd so a system can start with mobility
3448 * disabled and enable it later
3450 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3451 page_group_by_mobility_disabled = 1;
3453 page_group_by_mobility_disabled = 0;
3455 printk("Built %i zonelists in %s order, mobility grouping %s. "
3456 "Total pages: %ld\n",
3458 zonelist_order_name[current_zonelist_order],
3459 page_group_by_mobility_disabled ? "off" : "on",
3462 printk("Policy zone: %s\n", zone_names[policy_zone]);
3467 * Helper functions to size the waitqueue hash table.
3468 * Essentially these want to choose hash table sizes sufficiently
3469 * large so that collisions trying to wait on pages are rare.
3470 * But in fact, the number of active page waitqueues on typical
3471 * systems is ridiculously low, less than 200. So this is even
3472 * conservative, even though it seems large.
3474 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3475 * waitqueues, i.e. the size of the waitq table given the number of pages.
3477 #define PAGES_PER_WAITQUEUE 256
3479 #ifndef CONFIG_MEMORY_HOTPLUG
3480 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3482 unsigned long size = 1;
3484 pages /= PAGES_PER_WAITQUEUE;
3486 while (size < pages)
3490 * Once we have dozens or even hundreds of threads sleeping
3491 * on IO we've got bigger problems than wait queue collision.
3492 * Limit the size of the wait table to a reasonable size.
3494 size = min(size, 4096UL);
3496 return max(size, 4UL);
3500 * A zone's size might be changed by hot-add, so it is not possible to determine
3501 * a suitable size for its wait_table. So we use the maximum size now.
3503 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3505 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3506 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3507 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3509 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3510 * or more by the traditional way. (See above). It equals:
3512 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3513 * ia64(16K page size) : = ( 8G + 4M)byte.
3514 * powerpc (64K page size) : = (32G +16M)byte.
3516 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3523 * This is an integer logarithm so that shifts can be used later
3524 * to extract the more random high bits from the multiplicative
3525 * hash function before the remainder is taken.
3527 static inline unsigned long wait_table_bits(unsigned long size)
3532 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3535 * Check if a pageblock contains reserved pages
3537 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3541 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3542 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3549 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3550 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3551 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3552 * higher will lead to a bigger reserve which will get freed as contiguous
3553 * blocks as reclaim kicks in
3555 static void setup_zone_migrate_reserve(struct zone *zone)
3557 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3559 unsigned long block_migratetype;
3563 * Get the start pfn, end pfn and the number of blocks to reserve
3564 * We have to be careful to be aligned to pageblock_nr_pages to
3565 * make sure that we always check pfn_valid for the first page in
3568 start_pfn = zone->zone_start_pfn;
3569 end_pfn = start_pfn + zone->spanned_pages;
3570 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3571 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3575 * Reserve blocks are generally in place to help high-order atomic
3576 * allocations that are short-lived. A min_free_kbytes value that
3577 * would result in more than 2 reserve blocks for atomic allocations
3578 * is assumed to be in place to help anti-fragmentation for the
3579 * future allocation of hugepages at runtime.
3581 reserve = min(2, reserve);
3583 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3584 if (!pfn_valid(pfn))
3586 page = pfn_to_page(pfn);
3588 /* Watch out for overlapping nodes */
3589 if (page_to_nid(page) != zone_to_nid(zone))
3592 block_migratetype = get_pageblock_migratetype(page);
3594 /* Only test what is necessary when the reserves are not met */
3597 * Blocks with reserved pages will never free, skip
3600 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3601 if (pageblock_is_reserved(pfn, block_end_pfn))
3604 /* If this block is reserved, account for it */
3605 if (block_migratetype == MIGRATE_RESERVE) {
3610 /* Suitable for reserving if this block is movable */
3611 if (block_migratetype == MIGRATE_MOVABLE) {
3612 set_pageblock_migratetype(page,
3614 move_freepages_block(zone, page,
3622 * If the reserve is met and this is a previous reserved block,
3625 if (block_migratetype == MIGRATE_RESERVE) {
3626 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3627 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3633 * Initially all pages are reserved - free ones are freed
3634 * up by free_all_bootmem() once the early boot process is
3635 * done. Non-atomic initialization, single-pass.
3637 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3638 unsigned long start_pfn, enum memmap_context context)
3641 unsigned long end_pfn = start_pfn + size;
3645 if (highest_memmap_pfn < end_pfn - 1)
3646 highest_memmap_pfn = end_pfn - 1;
3648 z = &NODE_DATA(nid)->node_zones[zone];
3649 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3651 * There can be holes in boot-time mem_map[]s
3652 * handed to this function. They do not
3653 * exist on hotplugged memory.
3655 if (context == MEMMAP_EARLY) {
3656 if (!early_pfn_valid(pfn))
3658 if (!early_pfn_in_nid(pfn, nid))
3661 page = pfn_to_page(pfn);
3662 set_page_links(page, zone, nid, pfn);
3663 mminit_verify_page_links(page, zone, nid, pfn);
3664 init_page_count(page);
3665 reset_page_mapcount(page);
3666 SetPageReserved(page);
3668 * Mark the block movable so that blocks are reserved for
3669 * movable at startup. This will force kernel allocations
3670 * to reserve their blocks rather than leaking throughout
3671 * the address space during boot when many long-lived
3672 * kernel allocations are made. Later some blocks near
3673 * the start are marked MIGRATE_RESERVE by
3674 * setup_zone_migrate_reserve()
3676 * bitmap is created for zone's valid pfn range. but memmap
3677 * can be created for invalid pages (for alignment)
3678 * check here not to call set_pageblock_migratetype() against
3681 if ((z->zone_start_pfn <= pfn)
3682 && (pfn < z->zone_start_pfn + z->spanned_pages)
3683 && !(pfn & (pageblock_nr_pages - 1)))
3684 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3686 INIT_LIST_HEAD(&page->lru);
3687 #ifdef WANT_PAGE_VIRTUAL
3688 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3689 if (!is_highmem_idx(zone))
3690 set_page_address(page, __va(pfn << PAGE_SHIFT));
3695 static void __meminit zone_init_free_lists(struct zone *zone)
3698 for_each_migratetype_order(order, t) {
3699 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3700 zone->free_area[order].nr_free = 0;
3704 #ifndef __HAVE_ARCH_MEMMAP_INIT
3705 #define memmap_init(size, nid, zone, start_pfn) \
3706 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3709 static int zone_batchsize(struct zone *zone)
3715 * The per-cpu-pages pools are set to around 1000th of the
3716 * size of the zone. But no more than 1/2 of a meg.
3718 * OK, so we don't know how big the cache is. So guess.
3720 batch = zone->present_pages / 1024;
3721 if (batch * PAGE_SIZE > 512 * 1024)
3722 batch = (512 * 1024) / PAGE_SIZE;
3723 batch /= 4; /* We effectively *= 4 below */
3728 * Clamp the batch to a 2^n - 1 value. Having a power
3729 * of 2 value was found to be more likely to have
3730 * suboptimal cache aliasing properties in some cases.
3732 * For example if 2 tasks are alternately allocating
3733 * batches of pages, one task can end up with a lot
3734 * of pages of one half of the possible page colors
3735 * and the other with pages of the other colors.
3737 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3742 /* The deferral and batching of frees should be suppressed under NOMMU
3745 * The problem is that NOMMU needs to be able to allocate large chunks
3746 * of contiguous memory as there's no hardware page translation to
3747 * assemble apparent contiguous memory from discontiguous pages.
3749 * Queueing large contiguous runs of pages for batching, however,
3750 * causes the pages to actually be freed in smaller chunks. As there
3751 * can be a significant delay between the individual batches being
3752 * recycled, this leads to the once large chunks of space being
3753 * fragmented and becoming unavailable for high-order allocations.
3759 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3761 struct per_cpu_pages *pcp;
3764 memset(p, 0, sizeof(*p));
3768 pcp->high = 6 * batch;
3769 pcp->batch = max(1UL, 1 * batch);
3770 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3771 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3775 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3776 * to the value high for the pageset p.
3779 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3782 struct per_cpu_pages *pcp;
3786 pcp->batch = max(1UL, high/4);
3787 if ((high/4) > (PAGE_SHIFT * 8))
3788 pcp->batch = PAGE_SHIFT * 8;
3791 static void setup_zone_pageset(struct zone *zone)
3795 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3797 for_each_possible_cpu(cpu) {
3798 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3800 setup_pageset(pcp, zone_batchsize(zone));
3802 if (percpu_pagelist_fraction)
3803 setup_pagelist_highmark(pcp,
3804 (zone->present_pages /
3805 percpu_pagelist_fraction));
3810 * Allocate per cpu pagesets and initialize them.
3811 * Before this call only boot pagesets were available.
3813 void __init setup_per_cpu_pageset(void)
3817 for_each_populated_zone(zone)
3818 setup_zone_pageset(zone);
3821 static noinline __init_refok
3822 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3825 struct pglist_data *pgdat = zone->zone_pgdat;
3829 * The per-page waitqueue mechanism uses hashed waitqueues
3832 zone->wait_table_hash_nr_entries =
3833 wait_table_hash_nr_entries(zone_size_pages);
3834 zone->wait_table_bits =
3835 wait_table_bits(zone->wait_table_hash_nr_entries);
3836 alloc_size = zone->wait_table_hash_nr_entries
3837 * sizeof(wait_queue_head_t);
3839 if (!slab_is_available()) {
3840 zone->wait_table = (wait_queue_head_t *)
3841 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3844 * This case means that a zone whose size was 0 gets new memory
3845 * via memory hot-add.
3846 * But it may be the case that a new node was hot-added. In
3847 * this case vmalloc() will not be able to use this new node's
3848 * memory - this wait_table must be initialized to use this new
3849 * node itself as well.
3850 * To use this new node's memory, further consideration will be
3853 zone->wait_table = vmalloc(alloc_size);
3855 if (!zone->wait_table)
3858 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3859 init_waitqueue_head(zone->wait_table + i);
3864 static int __zone_pcp_update(void *data)
3866 struct zone *zone = data;
3868 unsigned long batch = zone_batchsize(zone), flags;
3870 for_each_possible_cpu(cpu) {
3871 struct per_cpu_pageset *pset;
3872 struct per_cpu_pages *pcp;
3874 pset = per_cpu_ptr(zone->pageset, cpu);
3877 local_irq_save(flags);
3878 free_pcppages_bulk(zone, pcp->count, pcp);
3879 setup_pageset(pset, batch);
3880 local_irq_restore(flags);
3885 void zone_pcp_update(struct zone *zone)
3887 stop_machine(__zone_pcp_update, zone, NULL);
3890 static __meminit void zone_pcp_init(struct zone *zone)
3893 * per cpu subsystem is not up at this point. The following code
3894 * relies on the ability of the linker to provide the
3895 * offset of a (static) per cpu variable into the per cpu area.
3897 zone->pageset = &boot_pageset;
3899 if (zone->present_pages)
3900 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3901 zone->name, zone->present_pages,
3902 zone_batchsize(zone));
3905 __meminit int init_currently_empty_zone(struct zone *zone,
3906 unsigned long zone_start_pfn,
3908 enum memmap_context context)
3910 struct pglist_data *pgdat = zone->zone_pgdat;
3912 ret = zone_wait_table_init(zone, size);
3915 pgdat->nr_zones = zone_idx(zone) + 1;
3917 zone->zone_start_pfn = zone_start_pfn;
3919 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3920 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3922 (unsigned long)zone_idx(zone),
3923 zone_start_pfn, (zone_start_pfn + size));
3925 zone_init_free_lists(zone);
3930 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3932 * Basic iterator support. Return the first range of PFNs for a node
3933 * Note: nid == MAX_NUMNODES returns first region regardless of node
3935 static int __meminit first_active_region_index_in_nid(int nid)
3939 for (i = 0; i < nr_nodemap_entries; i++)
3940 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3947 * Basic iterator support. Return the next active range of PFNs for a node
3948 * Note: nid == MAX_NUMNODES returns next region regardless of node
3950 static int __meminit next_active_region_index_in_nid(int index, int nid)
3952 for (index = index + 1; index < nr_nodemap_entries; index++)
3953 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3959 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3961 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3962 * Architectures may implement their own version but if add_active_range()
3963 * was used and there are no special requirements, this is a convenient
3966 int __meminit __early_pfn_to_nid(unsigned long pfn)
3970 for (i = 0; i < nr_nodemap_entries; i++) {
3971 unsigned long start_pfn = early_node_map[i].start_pfn;
3972 unsigned long end_pfn = early_node_map[i].end_pfn;
3974 if (start_pfn <= pfn && pfn < end_pfn)
3975 return early_node_map[i].nid;
3977 /* This is a memory hole */
3980 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3982 int __meminit early_pfn_to_nid(unsigned long pfn)
3986 nid = __early_pfn_to_nid(pfn);
3989 /* just returns 0 */
3993 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3994 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3998 nid = __early_pfn_to_nid(pfn);
3999 if (nid >= 0 && nid != node)
4005 /* Basic iterator support to walk early_node_map[] */
4006 #define for_each_active_range_index_in_nid(i, nid) \
4007 for (i = first_active_region_index_in_nid(nid); i != -1; \
4008 i = next_active_region_index_in_nid(i, nid))
4011 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4012 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4013 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4015 * If an architecture guarantees that all ranges registered with
4016 * add_active_ranges() contain no holes and may be freed, this
4017 * this function may be used instead of calling free_bootmem() manually.
4019 void __init free_bootmem_with_active_regions(int nid,
4020 unsigned long max_low_pfn)
4024 for_each_active_range_index_in_nid(i, nid) {
4025 unsigned long size_pages = 0;
4026 unsigned long end_pfn = early_node_map[i].end_pfn;
4028 if (early_node_map[i].start_pfn >= max_low_pfn)
4031 if (end_pfn > max_low_pfn)
4032 end_pfn = max_low_pfn;
4034 size_pages = end_pfn - early_node_map[i].start_pfn;
4035 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
4036 PFN_PHYS(early_node_map[i].start_pfn),
4037 size_pages << PAGE_SHIFT);
4041 #ifdef CONFIG_HAVE_MEMBLOCK
4043 * Basic iterator support. Return the last range of PFNs for a node
4044 * Note: nid == MAX_NUMNODES returns last region regardless of node
4046 static int __meminit last_active_region_index_in_nid(int nid)
4050 for (i = nr_nodemap_entries - 1; i >= 0; i--)
4051 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
4058 * Basic iterator support. Return the previous active range of PFNs for a node
4059 * Note: nid == MAX_NUMNODES returns next region regardless of node
4061 static int __meminit previous_active_region_index_in_nid(int index, int nid)
4063 for (index = index - 1; index >= 0; index--)
4064 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
4070 #define for_each_active_range_index_in_nid_reverse(i, nid) \
4071 for (i = last_active_region_index_in_nid(nid); i != -1; \
4072 i = previous_active_region_index_in_nid(i, nid))
4074 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
4075 u64 goal, u64 limit)
4079 /* Need to go over early_node_map to find out good range for node */
4080 for_each_active_range_index_in_nid_reverse(i, nid) {
4082 u64 ei_start, ei_last;
4083 u64 final_start, final_end;
4085 ei_last = early_node_map[i].end_pfn;
4086 ei_last <<= PAGE_SHIFT;
4087 ei_start = early_node_map[i].start_pfn;
4088 ei_start <<= PAGE_SHIFT;
4090 final_start = max(ei_start, goal);
4091 final_end = min(ei_last, limit);
4093 if (final_start >= final_end)
4096 addr = memblock_find_in_range(final_start, final_end, size, align);
4098 if (addr == MEMBLOCK_ERROR)
4104 return MEMBLOCK_ERROR;
4108 int __init add_from_early_node_map(struct range *range, int az,
4109 int nr_range, int nid)
4114 /* need to go over early_node_map to find out good range for node */
4115 for_each_active_range_index_in_nid(i, nid) {
4116 start = early_node_map[i].start_pfn;
4117 end = early_node_map[i].end_pfn;
4118 nr_range = add_range(range, az, nr_range, start, end);
4123 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
4128 for_each_active_range_index_in_nid(i, nid) {
4129 ret = work_fn(early_node_map[i].start_pfn,
4130 early_node_map[i].end_pfn, data);
4136 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4137 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4139 * If an architecture guarantees that all ranges registered with
4140 * add_active_ranges() contain no holes and may be freed, this
4141 * function may be used instead of calling memory_present() manually.
4143 void __init sparse_memory_present_with_active_regions(int nid)
4147 for_each_active_range_index_in_nid(i, nid)
4148 memory_present(early_node_map[i].nid,
4149 early_node_map[i].start_pfn,
4150 early_node_map[i].end_pfn);
4154 * get_pfn_range_for_nid - Return the start and end page frames for a node
4155 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4156 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4157 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4159 * It returns the start and end page frame of a node based on information
4160 * provided by an arch calling add_active_range(). If called for a node
4161 * with no available memory, a warning is printed and the start and end
4164 void __meminit get_pfn_range_for_nid(unsigned int nid,
4165 unsigned long *start_pfn, unsigned long *end_pfn)
4171 for_each_active_range_index_in_nid(i, nid) {
4172 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
4173 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
4176 if (*start_pfn == -1UL)
4181 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4182 * assumption is made that zones within a node are ordered in monotonic
4183 * increasing memory addresses so that the "highest" populated zone is used
4185 static void __init find_usable_zone_for_movable(void)
4188 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4189 if (zone_index == ZONE_MOVABLE)
4192 if (arch_zone_highest_possible_pfn[zone_index] >
4193 arch_zone_lowest_possible_pfn[zone_index])
4197 VM_BUG_ON(zone_index == -1);
4198 movable_zone = zone_index;
4202 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4203 * because it is sized independent of architecture. Unlike the other zones,
4204 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4205 * in each node depending on the size of each node and how evenly kernelcore
4206 * is distributed. This helper function adjusts the zone ranges
4207 * provided by the architecture for a given node by using the end of the
4208 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4209 * zones within a node are in order of monotonic increases memory addresses
4211 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4212 unsigned long zone_type,
4213 unsigned long node_start_pfn,
4214 unsigned long node_end_pfn,
4215 unsigned long *zone_start_pfn,
4216 unsigned long *zone_end_pfn)
4218 /* Only adjust if ZONE_MOVABLE is on this node */
4219 if (zone_movable_pfn[nid]) {
4220 /* Size ZONE_MOVABLE */
4221 if (zone_type == ZONE_MOVABLE) {
4222 *zone_start_pfn = zone_movable_pfn[nid];
4223 *zone_end_pfn = min(node_end_pfn,
4224 arch_zone_highest_possible_pfn[movable_zone]);
4226 /* Adjust for ZONE_MOVABLE starting within this range */
4227 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4228 *zone_end_pfn > zone_movable_pfn[nid]) {
4229 *zone_end_pfn = zone_movable_pfn[nid];
4231 /* Check if this whole range is within ZONE_MOVABLE */
4232 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4233 *zone_start_pfn = *zone_end_pfn;
4238 * Return the number of pages a zone spans in a node, including holes
4239 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4241 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4242 unsigned long zone_type,
4243 unsigned long *ignored)
4245 unsigned long node_start_pfn, node_end_pfn;
4246 unsigned long zone_start_pfn, zone_end_pfn;
4248 /* Get the start and end of the node and zone */
4249 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4250 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4251 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4252 adjust_zone_range_for_zone_movable(nid, zone_type,
4253 node_start_pfn, node_end_pfn,
4254 &zone_start_pfn, &zone_end_pfn);
4256 /* Check that this node has pages within the zone's required range */
4257 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4260 /* Move the zone boundaries inside the node if necessary */
4261 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4262 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4264 /* Return the spanned pages */
4265 return zone_end_pfn - zone_start_pfn;
4269 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4270 * then all holes in the requested range will be accounted for.
4272 unsigned long __meminit __absent_pages_in_range(int nid,
4273 unsigned long range_start_pfn,
4274 unsigned long range_end_pfn)
4277 unsigned long prev_end_pfn = 0, hole_pages = 0;
4278 unsigned long start_pfn;
4280 /* Find the end_pfn of the first active range of pfns in the node */
4281 i = first_active_region_index_in_nid(nid);
4285 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4287 /* Account for ranges before physical memory on this node */
4288 if (early_node_map[i].start_pfn > range_start_pfn)
4289 hole_pages = prev_end_pfn - range_start_pfn;
4291 /* Find all holes for the zone within the node */
4292 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4294 /* No need to continue if prev_end_pfn is outside the zone */
4295 if (prev_end_pfn >= range_end_pfn)
4298 /* Make sure the end of the zone is not within the hole */
4299 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4300 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4302 /* Update the hole size cound and move on */
4303 if (start_pfn > range_start_pfn) {
4304 BUG_ON(prev_end_pfn > start_pfn);
4305 hole_pages += start_pfn - prev_end_pfn;
4307 prev_end_pfn = early_node_map[i].end_pfn;
4310 /* Account for ranges past physical memory on this node */
4311 if (range_end_pfn > prev_end_pfn)
4312 hole_pages += range_end_pfn -
4313 max(range_start_pfn, prev_end_pfn);
4319 * absent_pages_in_range - Return number of page frames in holes within a range
4320 * @start_pfn: The start PFN to start searching for holes
4321 * @end_pfn: The end PFN to stop searching for holes
4323 * It returns the number of pages frames in memory holes within a range.
4325 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4326 unsigned long end_pfn)
4328 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4331 /* Return the number of page frames in holes in a zone on a node */
4332 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4333 unsigned long zone_type,
4334 unsigned long *ignored)
4336 unsigned long node_start_pfn, node_end_pfn;
4337 unsigned long zone_start_pfn, zone_end_pfn;
4339 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4340 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4342 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4345 adjust_zone_range_for_zone_movable(nid, zone_type,
4346 node_start_pfn, node_end_pfn,
4347 &zone_start_pfn, &zone_end_pfn);
4348 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4352 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4353 unsigned long zone_type,
4354 unsigned long *zones_size)
4356 return zones_size[zone_type];
4359 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4360 unsigned long zone_type,
4361 unsigned long *zholes_size)
4366 return zholes_size[zone_type];
4371 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4372 unsigned long *zones_size, unsigned long *zholes_size)
4374 unsigned long realtotalpages, totalpages = 0;
4377 for (i = 0; i < MAX_NR_ZONES; i++)
4378 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4380 pgdat->node_spanned_pages = totalpages;
4382 realtotalpages = totalpages;
4383 for (i = 0; i < MAX_NR_ZONES; i++)
4385 zone_absent_pages_in_node(pgdat->node_id, i,
4387 pgdat->node_present_pages = realtotalpages;
4388 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4392 #ifndef CONFIG_SPARSEMEM
4394 * Calculate the size of the zone->blockflags rounded to an unsigned long
4395 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4396 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4397 * round what is now in bits to nearest long in bits, then return it in
4400 static unsigned long __init usemap_size(unsigned long zonesize)
4402 unsigned long usemapsize;
4404 usemapsize = roundup(zonesize, pageblock_nr_pages);
4405 usemapsize = usemapsize >> pageblock_order;
4406 usemapsize *= NR_PAGEBLOCK_BITS;
4407 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4409 return usemapsize / 8;
4412 static void __init setup_usemap(struct pglist_data *pgdat,
4413 struct zone *zone, unsigned long zonesize)
4415 unsigned long usemapsize = usemap_size(zonesize);
4416 zone->pageblock_flags = NULL;
4418 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4422 static inline void setup_usemap(struct pglist_data *pgdat,
4423 struct zone *zone, unsigned long zonesize) {}
4424 #endif /* CONFIG_SPARSEMEM */
4426 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4428 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4429 void __init set_pageblock_order(void)
4433 /* Check that pageblock_nr_pages has not already been setup */
4434 if (pageblock_order)
4437 if (HPAGE_SHIFT > PAGE_SHIFT)
4438 order = HUGETLB_PAGE_ORDER;
4440 order = MAX_ORDER - 1;
4443 * Assume the largest contiguous order of interest is a huge page.
4444 * This value may be variable depending on boot parameters on IA64 and
4447 pageblock_order = order;
4449 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4452 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4453 * is unused as pageblock_order is set at compile-time. See
4454 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4457 void __init set_pageblock_order(void)
4461 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4464 * Set up the zone data structures:
4465 * - mark all pages reserved
4466 * - mark all memory queues empty
4467 * - clear the memory bitmaps
4469 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4470 unsigned long *zones_size, unsigned long *zholes_size)
4473 int nid = pgdat->node_id;
4474 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4477 pgdat_resize_init(pgdat);
4478 pgdat->nr_zones = 0;
4479 init_waitqueue_head(&pgdat->kswapd_wait);
4480 pgdat->kswapd_max_order = 0;
4481 pgdat_page_cgroup_init(pgdat);
4483 for (j = 0; j < MAX_NR_ZONES; j++) {
4484 struct zone *zone = pgdat->node_zones + j;
4485 unsigned long size, realsize, memmap_pages;
4488 size = zone_spanned_pages_in_node(nid, j, zones_size);
4489 realsize = size - zone_absent_pages_in_node(nid, j,
4493 * Adjust realsize so that it accounts for how much memory
4494 * is used by this zone for memmap. This affects the watermark
4495 * and per-cpu initialisations
4498 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4499 if (realsize >= memmap_pages) {
4500 realsize -= memmap_pages;
4503 " %s zone: %lu pages used for memmap\n",
4504 zone_names[j], memmap_pages);
4507 " %s zone: %lu pages exceeds realsize %lu\n",
4508 zone_names[j], memmap_pages, realsize);
4510 /* Account for reserved pages */
4511 if (j == 0 && realsize > dma_reserve) {
4512 realsize -= dma_reserve;
4513 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4514 zone_names[0], dma_reserve);
4517 if (!is_highmem_idx(j))
4518 nr_kernel_pages += realsize;
4519 nr_all_pages += realsize;
4521 zone->spanned_pages = size;
4522 zone->present_pages = realsize;
4525 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4527 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4529 zone->name = zone_names[j];
4530 spin_lock_init(&zone->lock);
4531 spin_lock_init(&zone->lru_lock);
4532 zone_seqlock_init(zone);
4533 zone->zone_pgdat = pgdat;
4535 zone_pcp_init(zone);
4537 INIT_LIST_HEAD(&zone->lru[l].list);
4538 zone->reclaim_stat.recent_rotated[0] = 0;
4539 zone->reclaim_stat.recent_rotated[1] = 0;
4540 zone->reclaim_stat.recent_scanned[0] = 0;
4541 zone->reclaim_stat.recent_scanned[1] = 0;
4542 zap_zone_vm_stats(zone);
4547 set_pageblock_order();
4548 setup_usemap(pgdat, zone, size);
4549 ret = init_currently_empty_zone(zone, zone_start_pfn,
4550 size, MEMMAP_EARLY);
4552 memmap_init(size, nid, j, zone_start_pfn);
4553 zone_start_pfn += size;
4557 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4559 /* Skip empty nodes */
4560 if (!pgdat->node_spanned_pages)
4563 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4564 /* ia64 gets its own node_mem_map, before this, without bootmem */
4565 if (!pgdat->node_mem_map) {
4566 unsigned long size, start, end;
4570 * The zone's endpoints aren't required to be MAX_ORDER
4571 * aligned but the node_mem_map endpoints must be in order
4572 * for the buddy allocator to function correctly.
4574 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4575 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4576 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4577 size = (end - start) * sizeof(struct page);
4578 map = alloc_remap(pgdat->node_id, size);
4580 map = alloc_bootmem_node_nopanic(pgdat, size);
4581 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4583 #ifndef CONFIG_NEED_MULTIPLE_NODES
4585 * With no DISCONTIG, the global mem_map is just set as node 0's
4587 if (pgdat == NODE_DATA(0)) {
4588 mem_map = NODE_DATA(0)->node_mem_map;
4589 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4590 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4591 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4592 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4595 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4598 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4599 unsigned long node_start_pfn, unsigned long *zholes_size)
4601 pg_data_t *pgdat = NODE_DATA(nid);
4603 pgdat->node_id = nid;
4604 pgdat->node_start_pfn = node_start_pfn;
4605 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4607 alloc_node_mem_map(pgdat);
4608 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4609 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4610 nid, (unsigned long)pgdat,
4611 (unsigned long)pgdat->node_mem_map);
4614 free_area_init_core(pgdat, zones_size, zholes_size);
4617 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4619 #if MAX_NUMNODES > 1
4621 * Figure out the number of possible node ids.
4623 static void __init setup_nr_node_ids(void)
4626 unsigned int highest = 0;
4628 for_each_node_mask(node, node_possible_map)
4630 nr_node_ids = highest + 1;
4633 static inline void setup_nr_node_ids(void)
4639 * add_active_range - Register a range of PFNs backed by physical memory
4640 * @nid: The node ID the range resides on
4641 * @start_pfn: The start PFN of the available physical memory
4642 * @end_pfn: The end PFN of the available physical memory
4644 * These ranges are stored in an early_node_map[] and later used by
4645 * free_area_init_nodes() to calculate zone sizes and holes. If the
4646 * range spans a memory hole, it is up to the architecture to ensure
4647 * the memory is not freed by the bootmem allocator. If possible
4648 * the range being registered will be merged with existing ranges.
4650 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4651 unsigned long end_pfn)
4655 mminit_dprintk(MMINIT_TRACE, "memory_register",
4656 "Entering add_active_range(%d, %#lx, %#lx) "
4657 "%d entries of %d used\n",
4658 nid, start_pfn, end_pfn,
4659 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4661 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4663 /* Merge with existing active regions if possible */
4664 for (i = 0; i < nr_nodemap_entries; i++) {
4665 if (early_node_map[i].nid != nid)
4668 /* Skip if an existing region covers this new one */
4669 if (start_pfn >= early_node_map[i].start_pfn &&
4670 end_pfn <= early_node_map[i].end_pfn)
4673 /* Merge forward if suitable */
4674 if (start_pfn <= early_node_map[i].end_pfn &&
4675 end_pfn > early_node_map[i].end_pfn) {
4676 early_node_map[i].end_pfn = end_pfn;
4680 /* Merge backward if suitable */
4681 if (start_pfn < early_node_map[i].start_pfn &&
4682 end_pfn >= early_node_map[i].start_pfn) {
4683 early_node_map[i].start_pfn = start_pfn;
4688 /* Check that early_node_map is large enough */
4689 if (i >= MAX_ACTIVE_REGIONS) {
4690 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4691 MAX_ACTIVE_REGIONS);
4695 early_node_map[i].nid = nid;
4696 early_node_map[i].start_pfn = start_pfn;
4697 early_node_map[i].end_pfn = end_pfn;
4698 nr_nodemap_entries = i + 1;
4702 * remove_active_range - Shrink an existing registered range of PFNs
4703 * @nid: The node id the range is on that should be shrunk
4704 * @start_pfn: The new PFN of the range
4705 * @end_pfn: The new PFN of the range
4707 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4708 * The map is kept near the end physical page range that has already been
4709 * registered. This function allows an arch to shrink an existing registered
4712 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4713 unsigned long end_pfn)
4718 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4719 nid, start_pfn, end_pfn);
4721 /* Find the old active region end and shrink */
4722 for_each_active_range_index_in_nid(i, nid) {
4723 if (early_node_map[i].start_pfn >= start_pfn &&
4724 early_node_map[i].end_pfn <= end_pfn) {
4726 early_node_map[i].start_pfn = 0;
4727 early_node_map[i].end_pfn = 0;
4731 if (early_node_map[i].start_pfn < start_pfn &&
4732 early_node_map[i].end_pfn > start_pfn) {
4733 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4734 early_node_map[i].end_pfn = start_pfn;
4735 if (temp_end_pfn > end_pfn)
4736 add_active_range(nid, end_pfn, temp_end_pfn);
4739 if (early_node_map[i].start_pfn >= start_pfn &&
4740 early_node_map[i].end_pfn > end_pfn &&
4741 early_node_map[i].start_pfn < end_pfn) {
4742 early_node_map[i].start_pfn = end_pfn;
4750 /* remove the blank ones */
4751 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4752 if (early_node_map[i].nid != nid)
4754 if (early_node_map[i].end_pfn)
4756 /* we found it, get rid of it */
4757 for (j = i; j < nr_nodemap_entries - 1; j++)
4758 memcpy(&early_node_map[j], &early_node_map[j+1],
4759 sizeof(early_node_map[j]));
4760 j = nr_nodemap_entries - 1;
4761 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4762 nr_nodemap_entries--;
4767 * remove_all_active_ranges - Remove all currently registered regions
4769 * During discovery, it may be found that a table like SRAT is invalid
4770 * and an alternative discovery method must be used. This function removes
4771 * all currently registered regions.
4773 void __init remove_all_active_ranges(void)
4775 memset(early_node_map, 0, sizeof(early_node_map));
4776 nr_nodemap_entries = 0;
4779 /* Compare two active node_active_regions */
4780 static int __init cmp_node_active_region(const void *a, const void *b)
4782 struct node_active_region *arange = (struct node_active_region *)a;
4783 struct node_active_region *brange = (struct node_active_region *)b;
4785 /* Done this way to avoid overflows */
4786 if (arange->start_pfn > brange->start_pfn)
4788 if (arange->start_pfn < brange->start_pfn)
4794 /* sort the node_map by start_pfn */
4795 void __init sort_node_map(void)
4797 sort(early_node_map, (size_t)nr_nodemap_entries,
4798 sizeof(struct node_active_region),
4799 cmp_node_active_region, NULL);
4803 * node_map_pfn_alignment - determine the maximum internode alignment
4805 * This function should be called after node map is populated and sorted.
4806 * It calculates the maximum power of two alignment which can distinguish
4809 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4810 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4811 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4812 * shifted, 1GiB is enough and this function will indicate so.
4814 * This is used to test whether pfn -> nid mapping of the chosen memory
4815 * model has fine enough granularity to avoid incorrect mapping for the
4816 * populated node map.
4818 * Returns the determined alignment in pfn's. 0 if there is no alignment
4819 * requirement (single node).
4821 unsigned long __init node_map_pfn_alignment(void)
4823 unsigned long accl_mask = 0, last_end = 0;
4827 for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4828 int nid = early_node_map[i].nid;
4829 unsigned long start = early_node_map[i].start_pfn;
4830 unsigned long end = early_node_map[i].end_pfn;
4833 if (!start || last_nid < 0 || last_nid == nid) {
4840 * Start with a mask granular enough to pin-point to the
4841 * start pfn and tick off bits one-by-one until it becomes
4842 * too coarse to separate the current node from the last.
4844 mask = ~((1 << __ffs(start)) - 1);
4845 while (mask && last_end <= (start & (mask << 1)))
4848 /* accumulate all internode masks */
4852 /* convert mask to number of pages */
4853 return ~accl_mask + 1;
4856 /* Find the lowest pfn for a node */
4857 static unsigned long __init find_min_pfn_for_node(int nid)
4860 unsigned long min_pfn = ULONG_MAX;
4862 /* Assuming a sorted map, the first range found has the starting pfn */
4863 for_each_active_range_index_in_nid(i, nid)
4864 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4866 if (min_pfn == ULONG_MAX) {
4868 "Could not find start_pfn for node %d\n", nid);
4876 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4878 * It returns the minimum PFN based on information provided via
4879 * add_active_range().
4881 unsigned long __init find_min_pfn_with_active_regions(void)
4883 return find_min_pfn_for_node(MAX_NUMNODES);
4887 * early_calculate_totalpages()
4888 * Sum pages in active regions for movable zone.
4889 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4891 static unsigned long __init early_calculate_totalpages(void)
4894 unsigned long totalpages = 0;
4896 for (i = 0; i < nr_nodemap_entries; i++) {
4897 unsigned long pages = early_node_map[i].end_pfn -
4898 early_node_map[i].start_pfn;
4899 totalpages += pages;
4901 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4907 * Find the PFN the Movable zone begins in each node. Kernel memory
4908 * is spread evenly between nodes as long as the nodes have enough
4909 * memory. When they don't, some nodes will have more kernelcore than
4912 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4915 unsigned long usable_startpfn;
4916 unsigned long kernelcore_node, kernelcore_remaining;
4917 /* save the state before borrow the nodemask */
4918 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4919 unsigned long totalpages = early_calculate_totalpages();
4920 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4923 * If movablecore was specified, calculate what size of
4924 * kernelcore that corresponds so that memory usable for
4925 * any allocation type is evenly spread. If both kernelcore
4926 * and movablecore are specified, then the value of kernelcore
4927 * will be used for required_kernelcore if it's greater than
4928 * what movablecore would have allowed.
4930 if (required_movablecore) {
4931 unsigned long corepages;
4934 * Round-up so that ZONE_MOVABLE is at least as large as what
4935 * was requested by the user
4937 required_movablecore =
4938 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4939 corepages = totalpages - required_movablecore;
4941 required_kernelcore = max(required_kernelcore, corepages);
4944 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4945 if (!required_kernelcore)
4948 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4949 find_usable_zone_for_movable();
4950 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4953 /* Spread kernelcore memory as evenly as possible throughout nodes */
4954 kernelcore_node = required_kernelcore / usable_nodes;
4955 for_each_node_state(nid, N_HIGH_MEMORY) {
4957 * Recalculate kernelcore_node if the division per node
4958 * now exceeds what is necessary to satisfy the requested
4959 * amount of memory for the kernel
4961 if (required_kernelcore < kernelcore_node)
4962 kernelcore_node = required_kernelcore / usable_nodes;
4965 * As the map is walked, we track how much memory is usable
4966 * by the kernel using kernelcore_remaining. When it is
4967 * 0, the rest of the node is usable by ZONE_MOVABLE
4969 kernelcore_remaining = kernelcore_node;
4971 /* Go through each range of PFNs within this node */
4972 for_each_active_range_index_in_nid(i, nid) {
4973 unsigned long start_pfn, end_pfn;
4974 unsigned long size_pages;
4976 start_pfn = max(early_node_map[i].start_pfn,
4977 zone_movable_pfn[nid]);
4978 end_pfn = early_node_map[i].end_pfn;
4979 if (start_pfn >= end_pfn)
4982 /* Account for what is only usable for kernelcore */
4983 if (start_pfn < usable_startpfn) {
4984 unsigned long kernel_pages;
4985 kernel_pages = min(end_pfn, usable_startpfn)
4988 kernelcore_remaining -= min(kernel_pages,
4989 kernelcore_remaining);
4990 required_kernelcore -= min(kernel_pages,
4991 required_kernelcore);
4993 /* Continue if range is now fully accounted */
4994 if (end_pfn <= usable_startpfn) {
4997 * Push zone_movable_pfn to the end so
4998 * that if we have to rebalance
4999 * kernelcore across nodes, we will
5000 * not double account here
5002 zone_movable_pfn[nid] = end_pfn;
5005 start_pfn = usable_startpfn;
5009 * The usable PFN range for ZONE_MOVABLE is from
5010 * start_pfn->end_pfn. Calculate size_pages as the
5011 * number of pages used as kernelcore
5013 size_pages = end_pfn - start_pfn;
5014 if (size_pages > kernelcore_remaining)
5015 size_pages = kernelcore_remaining;
5016 zone_movable_pfn[nid] = start_pfn + size_pages;
5019 * Some kernelcore has been met, update counts and
5020 * break if the kernelcore for this node has been
5023 required_kernelcore -= min(required_kernelcore,
5025 kernelcore_remaining -= size_pages;
5026 if (!kernelcore_remaining)
5032 * If there is still required_kernelcore, we do another pass with one
5033 * less node in the count. This will push zone_movable_pfn[nid] further
5034 * along on the nodes that still have memory until kernelcore is
5038 if (usable_nodes && required_kernelcore > usable_nodes)
5041 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5042 for (nid = 0; nid < MAX_NUMNODES; nid++)
5043 zone_movable_pfn[nid] =
5044 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5047 /* restore the node_state */
5048 node_states[N_HIGH_MEMORY] = saved_node_state;
5051 /* Any regular memory on that node ? */
5052 static void check_for_regular_memory(pg_data_t *pgdat)
5054 #ifdef CONFIG_HIGHMEM
5055 enum zone_type zone_type;
5057 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
5058 struct zone *zone = &pgdat->node_zones[zone_type];
5059 if (zone->present_pages)
5060 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
5066 * free_area_init_nodes - Initialise all pg_data_t and zone data
5067 * @max_zone_pfn: an array of max PFNs for each zone
5069 * This will call free_area_init_node() for each active node in the system.
5070 * Using the page ranges provided by add_active_range(), the size of each
5071 * zone in each node and their holes is calculated. If the maximum PFN
5072 * between two adjacent zones match, it is assumed that the zone is empty.
5073 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5074 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5075 * starts where the previous one ended. For example, ZONE_DMA32 starts
5076 * at arch_max_dma_pfn.
5078 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5083 /* Sort early_node_map as initialisation assumes it is sorted */
5086 /* Record where the zone boundaries are */
5087 memset(arch_zone_lowest_possible_pfn, 0,
5088 sizeof(arch_zone_lowest_possible_pfn));
5089 memset(arch_zone_highest_possible_pfn, 0,
5090 sizeof(arch_zone_highest_possible_pfn));
5091 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5092 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5093 for (i = 1; i < MAX_NR_ZONES; i++) {
5094 if (i == ZONE_MOVABLE)
5096 arch_zone_lowest_possible_pfn[i] =
5097 arch_zone_highest_possible_pfn[i-1];
5098 arch_zone_highest_possible_pfn[i] =
5099 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5101 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5102 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5104 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5105 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5106 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
5108 /* Print out the zone ranges */
5109 printk("Zone PFN ranges:\n");
5110 for (i = 0; i < MAX_NR_ZONES; i++) {
5111 if (i == ZONE_MOVABLE)
5113 printk(" %-8s ", zone_names[i]);
5114 if (arch_zone_lowest_possible_pfn[i] ==
5115 arch_zone_highest_possible_pfn[i])
5118 printk("%0#10lx -> %0#10lx\n",
5119 arch_zone_lowest_possible_pfn[i],
5120 arch_zone_highest_possible_pfn[i]);
5123 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5124 printk("Movable zone start PFN for each node\n");
5125 for (i = 0; i < MAX_NUMNODES; i++) {
5126 if (zone_movable_pfn[i])
5127 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
5130 /* Print out the early_node_map[] */
5131 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
5132 for (i = 0; i < nr_nodemap_entries; i++)
5133 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
5134 early_node_map[i].start_pfn,
5135 early_node_map[i].end_pfn);
5137 /* Initialise every node */
5138 mminit_verify_pageflags_layout();
5139 setup_nr_node_ids();
5140 for_each_online_node(nid) {
5141 pg_data_t *pgdat = NODE_DATA(nid);
5142 free_area_init_node(nid, NULL,
5143 find_min_pfn_for_node(nid), NULL);
5145 /* Any memory on that node */
5146 if (pgdat->node_present_pages)
5147 node_set_state(nid, N_HIGH_MEMORY);
5148 check_for_regular_memory(pgdat);
5152 static int __init cmdline_parse_core(char *p, unsigned long *core)
5154 unsigned long long coremem;
5158 coremem = memparse(p, &p);
5159 *core = coremem >> PAGE_SHIFT;
5161 /* Paranoid check that UL is enough for the coremem value */
5162 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5168 * kernelcore=size sets the amount of memory for use for allocations that
5169 * cannot be reclaimed or migrated.
5171 static int __init cmdline_parse_kernelcore(char *p)
5173 return cmdline_parse_core(p, &required_kernelcore);
5177 * movablecore=size sets the amount of memory for use for allocations that
5178 * can be reclaimed or migrated.
5180 static int __init cmdline_parse_movablecore(char *p)
5182 return cmdline_parse_core(p, &required_movablecore);
5185 early_param("kernelcore", cmdline_parse_kernelcore);
5186 early_param("movablecore", cmdline_parse_movablecore);
5188 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
5191 * set_dma_reserve - set the specified number of pages reserved in the first zone
5192 * @new_dma_reserve: The number of pages to mark reserved
5194 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5195 * In the DMA zone, a significant percentage may be consumed by kernel image
5196 * and other unfreeable allocations which can skew the watermarks badly. This
5197 * function may optionally be used to account for unfreeable pages in the
5198 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5199 * smaller per-cpu batchsize.
5201 void __init set_dma_reserve(unsigned long new_dma_reserve)
5203 dma_reserve = new_dma_reserve;
5206 void __init free_area_init(unsigned long *zones_size)
5208 free_area_init_node(0, zones_size,
5209 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5212 static int page_alloc_cpu_notify(struct notifier_block *self,
5213 unsigned long action, void *hcpu)
5215 int cpu = (unsigned long)hcpu;
5217 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5221 * Spill the event counters of the dead processor
5222 * into the current processors event counters.
5223 * This artificially elevates the count of the current
5226 vm_events_fold_cpu(cpu);
5229 * Zero the differential counters of the dead processor
5230 * so that the vm statistics are consistent.
5232 * This is only okay since the processor is dead and cannot
5233 * race with what we are doing.
5235 refresh_cpu_vm_stats(cpu);
5240 void __init page_alloc_init(void)
5242 hotcpu_notifier(page_alloc_cpu_notify, 0);
5246 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5247 * or min_free_kbytes changes.
5249 static void calculate_totalreserve_pages(void)
5251 struct pglist_data *pgdat;
5252 unsigned long reserve_pages = 0;
5253 enum zone_type i, j;
5255 for_each_online_pgdat(pgdat) {
5256 for (i = 0; i < MAX_NR_ZONES; i++) {
5257 struct zone *zone = pgdat->node_zones + i;
5258 unsigned long max = 0;
5260 /* Find valid and maximum lowmem_reserve in the zone */
5261 for (j = i; j < MAX_NR_ZONES; j++) {
5262 if (zone->lowmem_reserve[j] > max)
5263 max = zone->lowmem_reserve[j];
5266 /* we treat the high watermark as reserved pages. */
5267 max += high_wmark_pages(zone);
5269 if (max > zone->present_pages)
5270 max = zone->present_pages;
5271 reserve_pages += max;
5273 * Lowmem reserves are not available to
5274 * GFP_HIGHUSER page cache allocations and
5275 * kswapd tries to balance zones to their high
5276 * watermark. As a result, neither should be
5277 * regarded as dirtyable memory, to prevent a
5278 * situation where reclaim has to clean pages
5279 * in order to balance the zones.
5281 zone->dirty_balance_reserve = max;
5284 dirty_balance_reserve = reserve_pages;
5285 totalreserve_pages = reserve_pages;
5289 * setup_per_zone_lowmem_reserve - called whenever
5290 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5291 * has a correct pages reserved value, so an adequate number of
5292 * pages are left in the zone after a successful __alloc_pages().
5294 static void setup_per_zone_lowmem_reserve(void)
5296 struct pglist_data *pgdat;
5297 enum zone_type j, idx;
5299 for_each_online_pgdat(pgdat) {
5300 for (j = 0; j < MAX_NR_ZONES; j++) {
5301 struct zone *zone = pgdat->node_zones + j;
5302 unsigned long present_pages = zone->present_pages;
5304 zone->lowmem_reserve[j] = 0;
5308 struct zone *lower_zone;
5312 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5313 sysctl_lowmem_reserve_ratio[idx] = 1;
5315 lower_zone = pgdat->node_zones + idx;
5316 lower_zone->lowmem_reserve[j] = present_pages /
5317 sysctl_lowmem_reserve_ratio[idx];
5318 present_pages += lower_zone->present_pages;
5323 /* update totalreserve_pages */
5324 calculate_totalreserve_pages();
5327 static void __setup_per_zone_wmarks(void)
5329 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5330 unsigned long lowmem_pages = 0;
5332 unsigned long flags;
5334 /* Calculate total number of !ZONE_HIGHMEM pages */
5335 for_each_zone(zone) {
5336 if (!is_highmem(zone))
5337 lowmem_pages += zone->present_pages;
5340 for_each_zone(zone) {
5343 spin_lock_irqsave(&zone->lock, flags);
5344 tmp = (u64)pages_min * zone->present_pages;
5345 do_div(tmp, lowmem_pages);
5346 if (is_highmem(zone)) {
5348 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5349 * need highmem pages, so cap pages_min to a small
5352 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5353 * deltas controls asynch page reclaim, and so should
5354 * not be capped for highmem.
5358 min_pages = zone->present_pages / 1024;
5359 if (min_pages < SWAP_CLUSTER_MAX)
5360 min_pages = SWAP_CLUSTER_MAX;
5361 if (min_pages > 128)
5363 zone->watermark[WMARK_MIN] = min_pages;
5366 * If it's a lowmem zone, reserve a number of pages
5367 * proportionate to the zone's size.
5369 zone->watermark[WMARK_MIN] = tmp;
5372 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5373 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5374 setup_zone_migrate_reserve(zone);
5375 spin_unlock_irqrestore(&zone->lock, flags);
5378 /* update totalreserve_pages */
5379 calculate_totalreserve_pages();
5383 * setup_per_zone_wmarks - called when min_free_kbytes changes
5384 * or when memory is hot-{added|removed}
5386 * Ensures that the watermark[min,low,high] values for each zone are set
5387 * correctly with respect to min_free_kbytes.
5389 void setup_per_zone_wmarks(void)
5391 mutex_lock(&zonelists_mutex);
5392 __setup_per_zone_wmarks();
5393 mutex_unlock(&zonelists_mutex);
5397 * The inactive anon list should be small enough that the VM never has to
5398 * do too much work, but large enough that each inactive page has a chance
5399 * to be referenced again before it is swapped out.
5401 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5402 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5403 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5404 * the anonymous pages are kept on the inactive list.
5407 * memory ratio inactive anon
5408 * -------------------------------------
5417 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5419 unsigned int gb, ratio;
5421 /* Zone size in gigabytes */
5422 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5424 ratio = int_sqrt(10 * gb);
5428 zone->inactive_ratio = ratio;
5431 static void __meminit setup_per_zone_inactive_ratio(void)
5436 calculate_zone_inactive_ratio(zone);
5440 * Initialise min_free_kbytes.
5442 * For small machines we want it small (128k min). For large machines
5443 * we want it large (64MB max). But it is not linear, because network
5444 * bandwidth does not increase linearly with machine size. We use
5446 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5447 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5463 int __meminit init_per_zone_wmark_min(void)
5465 unsigned long lowmem_kbytes;
5467 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5469 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5470 if (min_free_kbytes < 128)
5471 min_free_kbytes = 128;
5472 if (min_free_kbytes > 65536)
5473 min_free_kbytes = 65536;
5474 setup_per_zone_wmarks();
5475 refresh_zone_stat_thresholds();
5476 setup_per_zone_lowmem_reserve();
5477 setup_per_zone_inactive_ratio();
5480 module_init(init_per_zone_wmark_min)
5483 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5484 * that we can call two helper functions whenever min_free_kbytes
5487 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5488 void __user *buffer, size_t *length, loff_t *ppos)
5490 proc_dointvec(table, write, buffer, length, ppos);
5492 setup_per_zone_wmarks();
5497 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5498 void __user *buffer, size_t *length, loff_t *ppos)
5503 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5508 zone->min_unmapped_pages = (zone->present_pages *
5509 sysctl_min_unmapped_ratio) / 100;
5513 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5514 void __user *buffer, size_t *length, loff_t *ppos)
5519 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5524 zone->min_slab_pages = (zone->present_pages *
5525 sysctl_min_slab_ratio) / 100;
5531 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5532 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5533 * whenever sysctl_lowmem_reserve_ratio changes.
5535 * The reserve ratio obviously has absolutely no relation with the
5536 * minimum watermarks. The lowmem reserve ratio can only make sense
5537 * if in function of the boot time zone sizes.
5539 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5540 void __user *buffer, size_t *length, loff_t *ppos)
5542 proc_dointvec_minmax(table, write, buffer, length, ppos);
5543 setup_per_zone_lowmem_reserve();
5548 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5549 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5550 * can have before it gets flushed back to buddy allocator.
5553 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5554 void __user *buffer, size_t *length, loff_t *ppos)
5560 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5561 if (!write || (ret == -EINVAL))
5563 for_each_populated_zone(zone) {
5564 for_each_possible_cpu(cpu) {
5566 high = zone->present_pages / percpu_pagelist_fraction;
5567 setup_pagelist_highmark(
5568 per_cpu_ptr(zone->pageset, cpu), high);
5574 int hashdist = HASHDIST_DEFAULT;
5577 static int __init set_hashdist(char *str)
5581 hashdist = simple_strtoul(str, &str, 0);
5584 __setup("hashdist=", set_hashdist);
5588 * allocate a large system hash table from bootmem
5589 * - it is assumed that the hash table must contain an exact power-of-2
5590 * quantity of entries
5591 * - limit is the number of hash buckets, not the total allocation size
5593 void *__init alloc_large_system_hash(const char *tablename,
5594 unsigned long bucketsize,
5595 unsigned long numentries,
5598 unsigned int *_hash_shift,
5599 unsigned int *_hash_mask,
5600 unsigned long limit)
5602 unsigned long long max = limit;
5603 unsigned long log2qty, size;
5606 /* allow the kernel cmdline to have a say */
5608 /* round applicable memory size up to nearest megabyte */
5609 numentries = nr_kernel_pages;
5610 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5611 numentries >>= 20 - PAGE_SHIFT;
5612 numentries <<= 20 - PAGE_SHIFT;
5614 /* limit to 1 bucket per 2^scale bytes of low memory */
5615 if (scale > PAGE_SHIFT)
5616 numentries >>= (scale - PAGE_SHIFT);
5618 numentries <<= (PAGE_SHIFT - scale);
5620 /* Make sure we've got at least a 0-order allocation.. */
5621 if (unlikely(flags & HASH_SMALL)) {
5622 /* Makes no sense without HASH_EARLY */
5623 WARN_ON(!(flags & HASH_EARLY));
5624 if (!(numentries >> *_hash_shift)) {
5625 numentries = 1UL << *_hash_shift;
5626 BUG_ON(!numentries);
5628 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5629 numentries = PAGE_SIZE / bucketsize;
5631 numentries = roundup_pow_of_two(numentries);
5633 /* limit allocation size to 1/16 total memory by default */
5635 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5636 do_div(max, bucketsize);
5639 if (numentries > max)
5642 log2qty = ilog2(numentries);
5645 size = bucketsize << log2qty;
5646 if (flags & HASH_EARLY)
5647 table = alloc_bootmem_nopanic(size);
5649 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5652 * If bucketsize is not a power-of-two, we may free
5653 * some pages at the end of hash table which
5654 * alloc_pages_exact() automatically does
5656 if (get_order(size) < MAX_ORDER) {
5657 table = alloc_pages_exact(size, GFP_ATOMIC);
5658 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5661 } while (!table && size > PAGE_SIZE && --log2qty);
5664 panic("Failed to allocate %s hash table\n", tablename);
5666 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5669 ilog2(size) - PAGE_SHIFT,
5673 *_hash_shift = log2qty;
5675 *_hash_mask = (1 << log2qty) - 1;
5680 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5681 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5684 #ifdef CONFIG_SPARSEMEM
5685 return __pfn_to_section(pfn)->pageblock_flags;
5687 return zone->pageblock_flags;
5688 #endif /* CONFIG_SPARSEMEM */
5691 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5693 #ifdef CONFIG_SPARSEMEM
5694 pfn &= (PAGES_PER_SECTION-1);
5695 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5697 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5698 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5699 #endif /* CONFIG_SPARSEMEM */
5703 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5704 * @page: The page within the block of interest
5705 * @start_bitidx: The first bit of interest to retrieve
5706 * @end_bitidx: The last bit of interest
5707 * returns pageblock_bits flags
5709 unsigned long get_pageblock_flags_group(struct page *page,
5710 int start_bitidx, int end_bitidx)
5713 unsigned long *bitmap;
5714 unsigned long pfn, bitidx;
5715 unsigned long flags = 0;
5716 unsigned long value = 1;
5718 zone = page_zone(page);
5719 pfn = page_to_pfn(page);
5720 bitmap = get_pageblock_bitmap(zone, pfn);
5721 bitidx = pfn_to_bitidx(zone, pfn);
5723 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5724 if (test_bit(bitidx + start_bitidx, bitmap))
5731 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5732 * @page: The page within the block of interest
5733 * @start_bitidx: The first bit of interest
5734 * @end_bitidx: The last bit of interest
5735 * @flags: The flags to set
5737 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5738 int start_bitidx, int end_bitidx)
5741 unsigned long *bitmap;
5742 unsigned long pfn, bitidx;
5743 unsigned long value = 1;
5745 zone = page_zone(page);
5746 pfn = page_to_pfn(page);
5747 bitmap = get_pageblock_bitmap(zone, pfn);
5748 bitidx = pfn_to_bitidx(zone, pfn);
5749 VM_BUG_ON(pfn < zone->zone_start_pfn);
5750 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5752 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5754 __set_bit(bitidx + start_bitidx, bitmap);
5756 __clear_bit(bitidx + start_bitidx, bitmap);
5760 * This is designed as sub function...plz see page_isolation.c also.
5761 * set/clear page block's type to be ISOLATE.
5762 * page allocater never alloc memory from ISOLATE block.
5766 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5768 unsigned long pfn, iter, found;
5772 * For avoiding noise data, lru_add_drain_all() should be called
5773 * If ZONE_MOVABLE, the zone never contains immobile pages
5775 if (zone_idx(zone) == ZONE_MOVABLE)
5777 mt = get_pageblock_migratetype(page);
5778 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5781 pfn = page_to_pfn(page);
5782 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5783 unsigned long check = pfn + iter;
5785 if (!pfn_valid_within(check))
5788 page = pfn_to_page(check);
5789 if (!page_count(page)) {
5790 if (PageBuddy(page))
5791 iter += (1 << page_order(page)) - 1;
5797 * If there are RECLAIMABLE pages, we need to check it.
5798 * But now, memory offline itself doesn't call shrink_slab()
5799 * and it still to be fixed.
5802 * If the page is not RAM, page_count()should be 0.
5803 * we don't need more check. This is an _used_ not-movable page.
5805 * The problematic thing here is PG_reserved pages. PG_reserved
5806 * is set to both of a memory hole page and a _used_ kernel
5815 bool is_pageblock_removable_nolock(struct page *page)
5817 struct zone *zone = page_zone(page);
5818 unsigned long pfn = page_to_pfn(page);
5821 * We have to be careful here because we are iterating over memory
5822 * sections which are not zone aware so we might end up outside of
5823 * the zone but still within the section.
5825 if (!zone || zone->zone_start_pfn > pfn ||
5826 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5829 return __count_immobile_pages(zone, page, 0);
5832 int set_migratetype_isolate(struct page *page)
5835 unsigned long flags, pfn;
5836 struct memory_isolate_notify arg;
5840 zone = page_zone(page);
5842 spin_lock_irqsave(&zone->lock, flags);
5844 pfn = page_to_pfn(page);
5845 arg.start_pfn = pfn;
5846 arg.nr_pages = pageblock_nr_pages;
5847 arg.pages_found = 0;
5850 * It may be possible to isolate a pageblock even if the
5851 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5852 * notifier chain is used by balloon drivers to return the
5853 * number of pages in a range that are held by the balloon
5854 * driver to shrink memory. If all the pages are accounted for
5855 * by balloons, are free, or on the LRU, isolation can continue.
5856 * Later, for example, when memory hotplug notifier runs, these
5857 * pages reported as "can be isolated" should be isolated(freed)
5858 * by the balloon driver through the memory notifier chain.
5860 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5861 notifier_ret = notifier_to_errno(notifier_ret);
5865 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5866 * We just check MOVABLE pages.
5868 if (__count_immobile_pages(zone, page, arg.pages_found))
5872 * immobile means "not-on-lru" paes. If immobile is larger than
5873 * removable-by-driver pages reported by notifier, we'll fail.
5878 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5879 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5882 spin_unlock_irqrestore(&zone->lock, flags);
5888 void unset_migratetype_isolate(struct page *page, unsigned migratetype)
5891 unsigned long flags;
5892 zone = page_zone(page);
5893 spin_lock_irqsave(&zone->lock, flags);
5894 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5896 set_pageblock_migratetype(page, migratetype);
5897 move_freepages_block(zone, page, migratetype);
5899 spin_unlock_irqrestore(&zone->lock, flags);
5904 static unsigned long pfn_max_align_down(unsigned long pfn)
5906 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5907 pageblock_nr_pages) - 1);
5910 static unsigned long pfn_max_align_up(unsigned long pfn)
5912 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5913 pageblock_nr_pages));
5916 static struct page *
5917 __alloc_contig_migrate_alloc(struct page *page, unsigned long private,
5920 return alloc_page(GFP_HIGHUSER_MOVABLE);
5923 /* [start, end) must belong to a single zone. */
5924 static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5926 /* This function is based on compact_zone() from compaction.c. */
5928 unsigned long pfn = start;
5929 unsigned int tries = 0;
5932 struct compact_control cc = {
5933 .nr_migratepages = 0,
5935 .zone = page_zone(pfn_to_page(start)),
5938 INIT_LIST_HEAD(&cc.migratepages);
5940 migrate_prep_local();
5942 while (pfn < end || !list_empty(&cc.migratepages)) {
5943 if (fatal_signal_pending(current)) {
5948 if (list_empty(&cc.migratepages)) {
5949 cc.nr_migratepages = 0;
5950 pfn = isolate_migratepages_range(cc.zone, &cc,
5957 } else if (++tries == 5) {
5958 ret = ret < 0 ? ret : -EBUSY;
5962 ret = migrate_pages(&cc.migratepages,
5963 __alloc_contig_migrate_alloc,
5967 putback_lru_pages(&cc.migratepages);
5968 return ret > 0 ? 0 : ret;
5972 * alloc_contig_range() -- tries to allocate given range of pages
5973 * @start: start PFN to allocate
5974 * @end: one-past-the-last PFN to allocate
5975 * @migratetype: migratetype of the underlaying pageblocks (either
5976 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5977 * in range must have the same migratetype and it must
5978 * be either of the two.
5980 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5981 * aligned, however it's the caller's responsibility to guarantee that
5982 * we are the only thread that changes migrate type of pageblocks the
5985 * The PFN range must belong to a single zone.
5987 * Returns zero on success or negative error code. On success all
5988 * pages which PFN is in [start, end) are allocated for the caller and
5989 * need to be freed with free_contig_range().
5991 int alloc_contig_range(unsigned long start, unsigned long end,
5992 unsigned migratetype)
5994 struct zone *zone = page_zone(pfn_to_page(start));
5995 unsigned long outer_start, outer_end;
5999 * What we do here is we mark all pageblocks in range as
6000 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6001 * have different sizes, and due to the way page allocator
6002 * work, we align the range to biggest of the two pages so
6003 * that page allocator won't try to merge buddies from
6004 * different pageblocks and change MIGRATE_ISOLATE to some
6005 * other migration type.
6007 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6008 * migrate the pages from an unaligned range (ie. pages that
6009 * we are interested in). This will put all the pages in
6010 * range back to page allocator as MIGRATE_ISOLATE.
6012 * When this is done, we take the pages in range from page
6013 * allocator removing them from the buddy system. This way
6014 * page allocator will never consider using them.
6016 * This lets us mark the pageblocks back as
6017 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6018 * aligned range but not in the unaligned, original range are
6019 * put back to page allocator so that buddy can use them.
6022 ret = start_isolate_page_range(pfn_max_align_down(start),
6023 pfn_max_align_up(end), migratetype);
6027 ret = __alloc_contig_migrate_range(start, end);
6032 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6033 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6034 * more, all pages in [start, end) are free in page allocator.
6035 * What we are going to do is to allocate all pages from
6036 * [start, end) (that is remove them from page allocator).
6038 * The only problem is that pages at the beginning and at the
6039 * end of interesting range may be not aligned with pages that
6040 * page allocator holds, ie. they can be part of higher order
6041 * pages. Because of this, we reserve the bigger range and
6042 * once this is done free the pages we are not interested in.
6044 * We don't have to hold zone->lock here because the pages are
6045 * isolated thus they won't get removed from buddy.
6048 lru_add_drain_all();
6052 outer_start = start;
6053 while (!PageBuddy(pfn_to_page(outer_start))) {
6054 if (++order >= MAX_ORDER) {
6058 outer_start &= ~0UL << order;
6061 /* Make sure the range is really isolated. */
6062 if (test_pages_isolated(outer_start, end)) {
6063 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6069 outer_end = isolate_freepages_range(outer_start, end);
6075 /* Free head and tail (if any) */
6076 if (start != outer_start)
6077 free_contig_range(outer_start, start - outer_start);
6078 if (end != outer_end)
6079 free_contig_range(end, outer_end - end);
6082 undo_isolate_page_range(pfn_max_align_down(start),
6083 pfn_max_align_up(end), migratetype);
6087 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6089 for (; nr_pages--; ++pfn)
6090 __free_page(pfn_to_page(pfn));
6094 #ifdef CONFIG_MEMORY_HOTREMOVE
6096 * All pages in the range must be isolated before calling this.
6099 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6105 unsigned long flags;
6106 /* find the first valid pfn */
6107 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6112 zone = page_zone(pfn_to_page(pfn));
6113 spin_lock_irqsave(&zone->lock, flags);
6115 while (pfn < end_pfn) {
6116 if (!pfn_valid(pfn)) {
6120 page = pfn_to_page(pfn);
6121 BUG_ON(page_count(page));
6122 BUG_ON(!PageBuddy(page));
6123 order = page_order(page);
6124 #ifdef CONFIG_DEBUG_VM
6125 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6126 pfn, 1 << order, end_pfn);
6128 list_del(&page->lru);
6129 rmv_page_order(page);
6130 zone->free_area[order].nr_free--;
6131 __mod_zone_page_state(zone, NR_FREE_PAGES,
6133 for (i = 0; i < (1 << order); i++)
6134 SetPageReserved((page+i));
6135 pfn += (1 << order);
6137 spin_unlock_irqrestore(&zone->lock, flags);
6141 #ifdef CONFIG_MEMORY_FAILURE
6142 bool is_free_buddy_page(struct page *page)
6144 struct zone *zone = page_zone(page);
6145 unsigned long pfn = page_to_pfn(page);
6146 unsigned long flags;
6149 spin_lock_irqsave(&zone->lock, flags);
6150 for (order = 0; order < MAX_ORDER; order++) {
6151 struct page *page_head = page - (pfn & ((1 << order) - 1));
6153 if (PageBuddy(page_head) && page_order(page_head) >= order)
6156 spin_unlock_irqrestore(&zone->lock, flags);
6158 return order < MAX_ORDER;
6162 static struct trace_print_flags pageflag_names[] = {
6163 {1UL << PG_locked, "locked" },
6164 {1UL << PG_error, "error" },
6165 {1UL << PG_referenced, "referenced" },
6166 {1UL << PG_uptodate, "uptodate" },
6167 {1UL << PG_dirty, "dirty" },
6168 {1UL << PG_lru, "lru" },
6169 {1UL << PG_active, "active" },
6170 {1UL << PG_slab, "slab" },
6171 {1UL << PG_owner_priv_1, "owner_priv_1" },
6172 {1UL << PG_arch_1, "arch_1" },
6173 {1UL << PG_reserved, "reserved" },
6174 {1UL << PG_private, "private" },
6175 {1UL << PG_private_2, "private_2" },
6176 {1UL << PG_writeback, "writeback" },
6177 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6178 {1UL << PG_head, "head" },
6179 {1UL << PG_tail, "tail" },
6181 {1UL << PG_compound, "compound" },
6183 {1UL << PG_swapcache, "swapcache" },
6184 {1UL << PG_mappedtodisk, "mappedtodisk" },
6185 {1UL << PG_reclaim, "reclaim" },
6186 {1UL << PG_swapbacked, "swapbacked" },
6187 {1UL << PG_unevictable, "unevictable" },
6189 {1UL << PG_mlocked, "mlocked" },
6191 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6192 {1UL << PG_uncached, "uncached" },
6194 #ifdef CONFIG_MEMORY_FAILURE
6195 {1UL << PG_hwpoison, "hwpoison" },
6200 static void dump_page_flags(unsigned long flags)
6202 const char *delim = "";
6206 printk(KERN_ALERT "page flags: %#lx(", flags);
6208 /* remove zone id */
6209 flags &= (1UL << NR_PAGEFLAGS) - 1;
6211 for (i = 0; pageflag_names[i].name && flags; i++) {
6213 mask = pageflag_names[i].mask;
6214 if ((flags & mask) != mask)
6218 printk("%s%s", delim, pageflag_names[i].name);
6222 /* check for left over flags */
6224 printk("%s%#lx", delim, flags);
6229 void dump_page(struct page *page)
6232 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6233 page, atomic_read(&page->_count), page_mapcount(page),
6234 page->mapping, page->index);
6235 dump_page_flags(page->flags);
6236 mem_cgroup_print_bad_page(page);