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>
61 #include <asm/tlbflush.h>
62 #include <asm/div64.h>
65 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
66 DEFINE_PER_CPU(int, numa_node);
67 EXPORT_PER_CPU_SYMBOL(numa_node);
70 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
72 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
73 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
74 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
75 * defined in <linux/topology.h>.
77 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
78 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
82 * Array of node states.
84 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
85 [N_POSSIBLE] = NODE_MASK_ALL,
86 [N_ONLINE] = { { [0] = 1UL } },
88 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
90 [N_HIGH_MEMORY] = { { [0] = 1UL } },
92 [N_CPU] = { { [0] = 1UL } },
95 EXPORT_SYMBOL(node_states);
97 unsigned long totalram_pages __read_mostly;
98 unsigned long totalreserve_pages __read_mostly;
99 int percpu_pagelist_fraction;
100 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
102 #ifdef CONFIG_PM_SLEEP
104 * The following functions are used by the suspend/hibernate code to temporarily
105 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
106 * while devices are suspended. To avoid races with the suspend/hibernate code,
107 * they should always be called with pm_mutex held (gfp_allowed_mask also should
108 * only be modified with pm_mutex held, unless the suspend/hibernate code is
109 * guaranteed not to run in parallel with that modification).
112 static gfp_t saved_gfp_mask;
114 void pm_restore_gfp_mask(void)
116 WARN_ON(!mutex_is_locked(&pm_mutex));
117 if (saved_gfp_mask) {
118 gfp_allowed_mask = saved_gfp_mask;
123 void pm_restrict_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 WARN_ON(saved_gfp_mask);
127 saved_gfp_mask = gfp_allowed_mask;
128 gfp_allowed_mask &= ~GFP_IOFS;
131 static bool pm_suspending(void)
133 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
140 static bool pm_suspending(void)
144 #endif /* CONFIG_PM_SLEEP */
146 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
147 int pageblock_order __read_mostly;
150 static void __free_pages_ok(struct page *page, unsigned int order);
153 * results with 256, 32 in the lowmem_reserve sysctl:
154 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
155 * 1G machine -> (16M dma, 784M normal, 224M high)
156 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
157 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
158 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
160 * TBD: should special case ZONE_DMA32 machines here - in those we normally
161 * don't need any ZONE_NORMAL reservation
163 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
164 #ifdef CONFIG_ZONE_DMA
167 #ifdef CONFIG_ZONE_DMA32
170 #ifdef CONFIG_HIGHMEM
176 EXPORT_SYMBOL(totalram_pages);
178 static char * const zone_names[MAX_NR_ZONES] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
186 #ifdef CONFIG_HIGHMEM
192 int min_free_kbytes = 1024;
193 int min_free_order_shift = 1;
195 static unsigned long __meminitdata nr_kernel_pages;
196 static unsigned long __meminitdata nr_all_pages;
197 static unsigned long __meminitdata dma_reserve;
199 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
201 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
202 * ranges of memory (RAM) that may be registered with add_active_range().
203 * Ranges passed to add_active_range() will be merged if possible
204 * so the number of times add_active_range() can be called is
205 * related to the number of nodes and the number of holes
207 #ifdef CONFIG_MAX_ACTIVE_REGIONS
208 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
209 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
211 #if MAX_NUMNODES >= 32
212 /* If there can be many nodes, allow up to 50 holes per node */
213 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
215 /* By default, allow up to 256 distinct regions */
216 #define MAX_ACTIVE_REGIONS 256
220 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
221 static int __meminitdata nr_nodemap_entries;
222 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
223 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
224 static unsigned long __initdata required_kernelcore;
225 static unsigned long __initdata required_movablecore;
226 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
228 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
230 EXPORT_SYMBOL(movable_zone);
231 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
234 int nr_node_ids __read_mostly = MAX_NUMNODES;
235 int nr_online_nodes __read_mostly = 1;
236 EXPORT_SYMBOL(nr_node_ids);
237 EXPORT_SYMBOL(nr_online_nodes);
240 int page_group_by_mobility_disabled __read_mostly;
242 static void set_pageblock_migratetype(struct page *page, int migratetype)
245 if (unlikely(page_group_by_mobility_disabled))
246 migratetype = MIGRATE_UNMOVABLE;
248 set_pageblock_flags_group(page, (unsigned long)migratetype,
249 PB_migrate, PB_migrate_end);
252 bool oom_killer_disabled __read_mostly;
254 #ifdef CONFIG_DEBUG_VM
255 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
259 unsigned long pfn = page_to_pfn(page);
262 seq = zone_span_seqbegin(zone);
263 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
265 else if (pfn < zone->zone_start_pfn)
267 } while (zone_span_seqretry(zone, seq));
272 static int page_is_consistent(struct zone *zone, struct page *page)
274 if (!pfn_valid_within(page_to_pfn(page)))
276 if (zone != page_zone(page))
282 * Temporary debugging check for pages not lying within a given zone.
284 static int bad_range(struct zone *zone, struct page *page)
286 if (page_outside_zone_boundaries(zone, page))
288 if (!page_is_consistent(zone, page))
294 static inline int bad_range(struct zone *zone, struct page *page)
300 static void bad_page(struct page *page)
302 static unsigned long resume;
303 static unsigned long nr_shown;
304 static unsigned long nr_unshown;
306 /* Don't complain about poisoned pages */
307 if (PageHWPoison(page)) {
308 reset_page_mapcount(page); /* remove PageBuddy */
313 * Allow a burst of 60 reports, then keep quiet for that minute;
314 * or allow a steady drip of one report per second.
316 if (nr_shown == 60) {
317 if (time_before(jiffies, resume)) {
323 "BUG: Bad page state: %lu messages suppressed\n",
330 resume = jiffies + 60 * HZ;
332 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
333 current->comm, page_to_pfn(page));
338 /* Leave bad fields for debug, except PageBuddy could make trouble */
339 reset_page_mapcount(page); /* remove PageBuddy */
340 add_taint(TAINT_BAD_PAGE);
344 * Higher-order pages are called "compound pages". They are structured thusly:
346 * The first PAGE_SIZE page is called the "head page".
348 * The remaining PAGE_SIZE pages are called "tail pages".
350 * All pages have PG_compound set. All pages have their ->private pointing at
351 * the head page (even the head page has this).
353 * The first tail page's ->lru.next holds the address of the compound page's
354 * put_page() function. Its ->lru.prev holds the order of allocation.
355 * This usage means that zero-order pages may not be compound.
358 static void free_compound_page(struct page *page)
360 __free_pages_ok(page, compound_order(page));
363 void prep_compound_page(struct page *page, unsigned long order)
366 int nr_pages = 1 << order;
368 set_compound_page_dtor(page, free_compound_page);
369 set_compound_order(page, order);
371 for (i = 1; i < nr_pages; i++) {
372 struct page *p = page + i;
374 set_page_count(p, 0);
375 p->first_page = page;
379 /* update __split_huge_page_refcount if you change this function */
380 static int destroy_compound_page(struct page *page, unsigned long order)
383 int nr_pages = 1 << order;
386 if (unlikely(compound_order(page) != order) ||
387 unlikely(!PageHead(page))) {
392 __ClearPageHead(page);
394 for (i = 1; i < nr_pages; i++) {
395 struct page *p = page + i;
397 if (unlikely(!PageTail(p) || (p->first_page != page))) {
407 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
412 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
413 * and __GFP_HIGHMEM from hard or soft interrupt context.
415 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
416 for (i = 0; i < (1 << order); i++)
417 clear_highpage(page + i);
420 static inline void set_page_order(struct page *page, int order)
422 set_page_private(page, order);
423 __SetPageBuddy(page);
426 static inline void rmv_page_order(struct page *page)
428 __ClearPageBuddy(page);
429 set_page_private(page, 0);
433 * Locate the struct page for both the matching buddy in our
434 * pair (buddy1) and the combined O(n+1) page they form (page).
436 * 1) Any buddy B1 will have an order O twin B2 which satisfies
437 * the following equation:
439 * For example, if the starting buddy (buddy2) is #8 its order
441 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
443 * 2) Any buddy B will have an order O+1 parent P which
444 * satisfies the following equation:
447 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
449 static inline unsigned long
450 __find_buddy_index(unsigned long page_idx, unsigned int order)
452 return page_idx ^ (1 << order);
456 * This function checks whether a page is free && is the buddy
457 * we can do coalesce a page and its buddy if
458 * (a) the buddy is not in a hole &&
459 * (b) the buddy is in the buddy system &&
460 * (c) a page and its buddy have the same order &&
461 * (d) a page and its buddy are in the same zone.
463 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
464 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
466 * For recording page's order, we use page_private(page).
468 static inline int page_is_buddy(struct page *page, struct page *buddy,
471 if (!pfn_valid_within(page_to_pfn(buddy)))
474 if (page_zone_id(page) != page_zone_id(buddy))
477 if (PageBuddy(buddy) && page_order(buddy) == order) {
478 VM_BUG_ON(page_count(buddy) != 0);
485 * Freeing function for a buddy system allocator.
487 * The concept of a buddy system is to maintain direct-mapped table
488 * (containing bit values) for memory blocks of various "orders".
489 * The bottom level table contains the map for the smallest allocatable
490 * units of memory (here, pages), and each level above it describes
491 * pairs of units from the levels below, hence, "buddies".
492 * At a high level, all that happens here is marking the table entry
493 * at the bottom level available, and propagating the changes upward
494 * as necessary, plus some accounting needed to play nicely with other
495 * parts of the VM system.
496 * At each level, we keep a list of pages, which are heads of continuous
497 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
498 * order is recorded in page_private(page) field.
499 * So when we are allocating or freeing one, we can derive the state of the
500 * other. That is, if we allocate a small block, and both were
501 * free, the remainder of the region must be split into blocks.
502 * If a block is freed, and its buddy is also free, then this
503 * triggers coalescing into a block of larger size.
508 static inline void __free_one_page(struct page *page,
509 struct zone *zone, unsigned int order,
512 unsigned long page_idx;
513 unsigned long combined_idx;
514 unsigned long uninitialized_var(buddy_idx);
517 if (unlikely(PageCompound(page)))
518 if (unlikely(destroy_compound_page(page, order)))
521 VM_BUG_ON(migratetype == -1);
523 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
525 VM_BUG_ON(page_idx & ((1 << order) - 1));
526 VM_BUG_ON(bad_range(zone, page));
528 while (order < MAX_ORDER-1) {
529 buddy_idx = __find_buddy_index(page_idx, order);
530 buddy = page + (buddy_idx - page_idx);
531 if (!page_is_buddy(page, buddy, order))
534 /* Our buddy is free, merge with it and move up one order. */
535 list_del(&buddy->lru);
536 zone->free_area[order].nr_free--;
537 rmv_page_order(buddy);
538 combined_idx = buddy_idx & page_idx;
539 page = page + (combined_idx - page_idx);
540 page_idx = combined_idx;
543 set_page_order(page, order);
546 * If this is not the largest possible page, check if the buddy
547 * of the next-highest order is free. If it is, it's possible
548 * that pages are being freed that will coalesce soon. In case,
549 * that is happening, add the free page to the tail of the list
550 * so it's less likely to be used soon and more likely to be merged
551 * as a higher order page
553 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
554 struct page *higher_page, *higher_buddy;
555 combined_idx = buddy_idx & page_idx;
556 higher_page = page + (combined_idx - page_idx);
557 buddy_idx = __find_buddy_index(combined_idx, order + 1);
558 higher_buddy = page + (buddy_idx - combined_idx);
559 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
560 list_add_tail(&page->lru,
561 &zone->free_area[order].free_list[migratetype]);
566 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
568 zone->free_area[order].nr_free++;
572 * free_page_mlock() -- clean up attempts to free and mlocked() page.
573 * Page should not be on lru, so no need to fix that up.
574 * free_pages_check() will verify...
576 static inline void free_page_mlock(struct page *page)
578 __dec_zone_page_state(page, NR_MLOCK);
579 __count_vm_event(UNEVICTABLE_MLOCKFREED);
582 static inline int free_pages_check(struct page *page)
584 if (unlikely(page_mapcount(page) |
585 (page->mapping != NULL) |
586 (atomic_read(&page->_count) != 0) |
587 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
588 (mem_cgroup_bad_page_check(page)))) {
592 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
593 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
598 * Frees a number of pages from the PCP lists
599 * Assumes all pages on list are in same zone, and of same order.
600 * count is the number of pages to free.
602 * If the zone was previously in an "all pages pinned" state then look to
603 * see if this freeing clears that state.
605 * And clear the zone's pages_scanned counter, to hold off the "all pages are
606 * pinned" detection logic.
608 static void free_pcppages_bulk(struct zone *zone, int count,
609 struct per_cpu_pages *pcp)
615 spin_lock(&zone->lock);
616 zone->all_unreclaimable = 0;
617 zone->pages_scanned = 0;
621 struct list_head *list;
624 * Remove pages from lists in a round-robin fashion. A
625 * batch_free count is maintained that is incremented when an
626 * empty list is encountered. This is so more pages are freed
627 * off fuller lists instead of spinning excessively around empty
632 if (++migratetype == MIGRATE_PCPTYPES)
634 list = &pcp->lists[migratetype];
635 } while (list_empty(list));
637 /* This is the only non-empty list. Free them all. */
638 if (batch_free == MIGRATE_PCPTYPES)
639 batch_free = to_free;
642 page = list_entry(list->prev, struct page, lru);
643 /* must delete as __free_one_page list manipulates */
644 list_del(&page->lru);
645 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
646 __free_one_page(page, zone, 0, page_private(page));
647 trace_mm_page_pcpu_drain(page, 0, page_private(page));
648 } while (--to_free && --batch_free && !list_empty(list));
650 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
651 spin_unlock(&zone->lock);
654 static void free_one_page(struct zone *zone, struct page *page, int order,
657 spin_lock(&zone->lock);
658 zone->all_unreclaimable = 0;
659 zone->pages_scanned = 0;
661 __free_one_page(page, zone, order, migratetype);
662 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
663 spin_unlock(&zone->lock);
666 static bool free_pages_prepare(struct page *page, unsigned int order)
671 trace_mm_page_free_direct(page, order);
672 kmemcheck_free_shadow(page, order);
675 page->mapping = NULL;
676 for (i = 0; i < (1 << order); i++)
677 bad += free_pages_check(page + i);
681 if (!PageHighMem(page)) {
682 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
683 debug_check_no_obj_freed(page_address(page),
686 arch_free_page(page, order);
687 kernel_map_pages(page, 1 << order, 0);
692 static void __free_pages_ok(struct page *page, unsigned int order)
695 int wasMlocked = __TestClearPageMlocked(page);
697 if (!free_pages_prepare(page, order))
700 local_irq_save(flags);
701 if (unlikely(wasMlocked))
702 free_page_mlock(page);
703 __count_vm_events(PGFREE, 1 << order);
704 free_one_page(page_zone(page), page, order,
705 get_pageblock_migratetype(page));
706 local_irq_restore(flags);
710 * permit the bootmem allocator to evade page validation on high-order frees
712 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
715 __ClearPageReserved(page);
716 set_page_count(page, 0);
717 set_page_refcounted(page);
723 for (loop = 0; loop < BITS_PER_LONG; loop++) {
724 struct page *p = &page[loop];
726 if (loop + 1 < BITS_PER_LONG)
728 __ClearPageReserved(p);
729 set_page_count(p, 0);
732 set_page_refcounted(page);
733 __free_pages(page, order);
739 * The order of subdivision here is critical for the IO subsystem.
740 * Please do not alter this order without good reasons and regression
741 * testing. Specifically, as large blocks of memory are subdivided,
742 * the order in which smaller blocks are delivered depends on the order
743 * they're subdivided in this function. This is the primary factor
744 * influencing the order in which pages are delivered to the IO
745 * subsystem according to empirical testing, and this is also justified
746 * by considering the behavior of a buddy system containing a single
747 * large block of memory acted on by a series of small allocations.
748 * This behavior is a critical factor in sglist merging's success.
752 static inline void expand(struct zone *zone, struct page *page,
753 int low, int high, struct free_area *area,
756 unsigned long size = 1 << high;
762 VM_BUG_ON(bad_range(zone, &page[size]));
763 list_add(&page[size].lru, &area->free_list[migratetype]);
765 set_page_order(&page[size], high);
770 * This page is about to be returned from the page allocator
772 static inline int check_new_page(struct page *page)
774 if (unlikely(page_mapcount(page) |
775 (page->mapping != NULL) |
776 (atomic_read(&page->_count) != 0) |
777 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
778 (mem_cgroup_bad_page_check(page)))) {
785 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
789 for (i = 0; i < (1 << order); i++) {
790 struct page *p = page + i;
791 if (unlikely(check_new_page(p)))
795 set_page_private(page, 0);
796 set_page_refcounted(page);
798 arch_alloc_page(page, order);
799 kernel_map_pages(page, 1 << order, 1);
801 if (gfp_flags & __GFP_ZERO)
802 prep_zero_page(page, order, gfp_flags);
804 if (order && (gfp_flags & __GFP_COMP))
805 prep_compound_page(page, order);
811 * Go through the free lists for the given migratetype and remove
812 * the smallest available page from the freelists
815 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
818 unsigned int current_order;
819 struct free_area * area;
822 /* Find a page of the appropriate size in the preferred list */
823 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
824 area = &(zone->free_area[current_order]);
825 if (list_empty(&area->free_list[migratetype]))
828 page = list_entry(area->free_list[migratetype].next,
830 list_del(&page->lru);
831 rmv_page_order(page);
833 expand(zone, page, order, current_order, area, migratetype);
842 * This array describes the order lists are fallen back to when
843 * the free lists for the desirable migrate type are depleted
845 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
846 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
847 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
848 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
849 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
853 * Move the free pages in a range to the free lists of the requested type.
854 * Note that start_page and end_pages are not aligned on a pageblock
855 * boundary. If alignment is required, use move_freepages_block()
857 static int move_freepages(struct zone *zone,
858 struct page *start_page, struct page *end_page,
865 #ifndef CONFIG_HOLES_IN_ZONE
867 * page_zone is not safe to call in this context when
868 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
869 * anyway as we check zone boundaries in move_freepages_block().
870 * Remove at a later date when no bug reports exist related to
871 * grouping pages by mobility
873 BUG_ON(page_zone(start_page) != page_zone(end_page));
876 for (page = start_page; page <= end_page;) {
877 /* Make sure we are not inadvertently changing nodes */
878 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
880 if (!pfn_valid_within(page_to_pfn(page))) {
885 if (!PageBuddy(page)) {
890 order = page_order(page);
891 list_move(&page->lru,
892 &zone->free_area[order].free_list[migratetype]);
894 pages_moved += 1 << order;
900 static int move_freepages_block(struct zone *zone, struct page *page,
903 unsigned long start_pfn, end_pfn;
904 struct page *start_page, *end_page;
906 start_pfn = page_to_pfn(page);
907 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
908 start_page = pfn_to_page(start_pfn);
909 end_page = start_page + pageblock_nr_pages - 1;
910 end_pfn = start_pfn + pageblock_nr_pages - 1;
912 /* Do not cross zone boundaries */
913 if (start_pfn < zone->zone_start_pfn)
915 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
918 return move_freepages(zone, start_page, end_page, migratetype);
921 static void change_pageblock_range(struct page *pageblock_page,
922 int start_order, int migratetype)
924 int nr_pageblocks = 1 << (start_order - pageblock_order);
926 while (nr_pageblocks--) {
927 set_pageblock_migratetype(pageblock_page, migratetype);
928 pageblock_page += pageblock_nr_pages;
932 /* Remove an element from the buddy allocator from the fallback list */
933 static inline struct page *
934 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
936 struct free_area * area;
941 /* Find the largest possible block of pages in the other list */
942 for (current_order = MAX_ORDER-1; current_order >= order;
944 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
945 migratetype = fallbacks[start_migratetype][i];
947 /* MIGRATE_RESERVE handled later if necessary */
948 if (migratetype == MIGRATE_RESERVE)
951 area = &(zone->free_area[current_order]);
952 if (list_empty(&area->free_list[migratetype]))
955 page = list_entry(area->free_list[migratetype].next,
960 * If breaking a large block of pages, move all free
961 * pages to the preferred allocation list. If falling
962 * back for a reclaimable kernel allocation, be more
963 * aggressive about taking ownership of free pages
965 if (unlikely(current_order >= (pageblock_order >> 1)) ||
966 start_migratetype == MIGRATE_RECLAIMABLE ||
967 page_group_by_mobility_disabled) {
969 pages = move_freepages_block(zone, page,
972 /* Claim the whole block if over half of it is free */
973 if (pages >= (1 << (pageblock_order-1)) ||
974 page_group_by_mobility_disabled)
975 set_pageblock_migratetype(page,
978 migratetype = start_migratetype;
981 /* Remove the page from the freelists */
982 list_del(&page->lru);
983 rmv_page_order(page);
985 /* Take ownership for orders >= pageblock_order */
986 if (current_order >= pageblock_order)
987 change_pageblock_range(page, current_order,
990 expand(zone, page, order, current_order, area, migratetype);
992 trace_mm_page_alloc_extfrag(page, order, current_order,
993 start_migratetype, migratetype);
1003 * Do the hard work of removing an element from the buddy allocator.
1004 * Call me with the zone->lock already held.
1006 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1012 page = __rmqueue_smallest(zone, order, migratetype);
1014 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1015 page = __rmqueue_fallback(zone, order, migratetype);
1018 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1019 * is used because __rmqueue_smallest is an inline function
1020 * and we want just one call site
1023 migratetype = MIGRATE_RESERVE;
1028 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1033 * Obtain a specified number of elements from the buddy allocator, all under
1034 * a single hold of the lock, for efficiency. Add them to the supplied list.
1035 * Returns the number of new pages which were placed at *list.
1037 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1038 unsigned long count, struct list_head *list,
1039 int migratetype, int cold)
1043 spin_lock(&zone->lock);
1044 for (i = 0; i < count; ++i) {
1045 struct page *page = __rmqueue(zone, order, migratetype);
1046 if (unlikely(page == NULL))
1050 * Split buddy pages returned by expand() are received here
1051 * in physical page order. The page is added to the callers and
1052 * list and the list head then moves forward. From the callers
1053 * perspective, the linked list is ordered by page number in
1054 * some conditions. This is useful for IO devices that can
1055 * merge IO requests if the physical pages are ordered
1058 if (likely(cold == 0))
1059 list_add(&page->lru, list);
1061 list_add_tail(&page->lru, list);
1062 set_page_private(page, migratetype);
1065 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1066 spin_unlock(&zone->lock);
1072 * Called from the vmstat counter updater to drain pagesets of this
1073 * currently executing processor on remote nodes after they have
1076 * Note that this function must be called with the thread pinned to
1077 * a single processor.
1079 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1081 unsigned long flags;
1084 local_irq_save(flags);
1085 if (pcp->count >= pcp->batch)
1086 to_drain = pcp->batch;
1088 to_drain = pcp->count;
1089 free_pcppages_bulk(zone, to_drain, pcp);
1090 pcp->count -= to_drain;
1091 local_irq_restore(flags);
1096 * Drain pages of the indicated processor.
1098 * The processor must either be the current processor and the
1099 * thread pinned to the current processor or a processor that
1102 static void drain_pages(unsigned int cpu)
1104 unsigned long flags;
1107 for_each_populated_zone(zone) {
1108 struct per_cpu_pageset *pset;
1109 struct per_cpu_pages *pcp;
1111 local_irq_save(flags);
1112 pset = per_cpu_ptr(zone->pageset, cpu);
1116 free_pcppages_bulk(zone, pcp->count, pcp);
1119 local_irq_restore(flags);
1124 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1126 void drain_local_pages(void *arg)
1128 drain_pages(smp_processor_id());
1132 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1134 void drain_all_pages(void)
1136 on_each_cpu(drain_local_pages, NULL, 1);
1139 #ifdef CONFIG_HIBERNATION
1141 void mark_free_pages(struct zone *zone)
1143 unsigned long pfn, max_zone_pfn;
1144 unsigned long flags;
1146 struct list_head *curr;
1148 if (!zone->spanned_pages)
1151 spin_lock_irqsave(&zone->lock, flags);
1153 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1154 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1155 if (pfn_valid(pfn)) {
1156 struct page *page = pfn_to_page(pfn);
1158 if (!swsusp_page_is_forbidden(page))
1159 swsusp_unset_page_free(page);
1162 for_each_migratetype_order(order, t) {
1163 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1166 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1167 for (i = 0; i < (1UL << order); i++)
1168 swsusp_set_page_free(pfn_to_page(pfn + i));
1171 spin_unlock_irqrestore(&zone->lock, flags);
1173 #endif /* CONFIG_PM */
1176 * Free a 0-order page
1177 * cold == 1 ? free a cold page : free a hot page
1179 void free_hot_cold_page(struct page *page, int cold)
1181 struct zone *zone = page_zone(page);
1182 struct per_cpu_pages *pcp;
1183 unsigned long flags;
1185 int wasMlocked = __TestClearPageMlocked(page);
1187 if (!free_pages_prepare(page, 0))
1190 migratetype = get_pageblock_migratetype(page);
1191 set_page_private(page, migratetype);
1192 local_irq_save(flags);
1193 if (unlikely(wasMlocked))
1194 free_page_mlock(page);
1195 __count_vm_event(PGFREE);
1198 * We only track unmovable, reclaimable and movable on pcp lists.
1199 * Free ISOLATE pages back to the allocator because they are being
1200 * offlined but treat RESERVE as movable pages so we can get those
1201 * areas back if necessary. Otherwise, we may have to free
1202 * excessively into the page allocator
1204 if (migratetype >= MIGRATE_PCPTYPES) {
1205 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1206 free_one_page(zone, page, 0, migratetype);
1209 migratetype = MIGRATE_MOVABLE;
1212 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1214 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1216 list_add(&page->lru, &pcp->lists[migratetype]);
1218 if (pcp->count >= pcp->high) {
1219 free_pcppages_bulk(zone, pcp->batch, pcp);
1220 pcp->count -= pcp->batch;
1224 local_irq_restore(flags);
1228 * split_page takes a non-compound higher-order page, and splits it into
1229 * n (1<<order) sub-pages: page[0..n]
1230 * Each sub-page must be freed individually.
1232 * Note: this is probably too low level an operation for use in drivers.
1233 * Please consult with lkml before using this in your driver.
1235 void split_page(struct page *page, unsigned int order)
1239 VM_BUG_ON(PageCompound(page));
1240 VM_BUG_ON(!page_count(page));
1242 #ifdef CONFIG_KMEMCHECK
1244 * Split shadow pages too, because free(page[0]) would
1245 * otherwise free the whole shadow.
1247 if (kmemcheck_page_is_tracked(page))
1248 split_page(virt_to_page(page[0].shadow), order);
1251 for (i = 1; i < (1 << order); i++)
1252 set_page_refcounted(page + i);
1256 * Similar to split_page except the page is already free. As this is only
1257 * being used for migration, the migratetype of the block also changes.
1258 * As this is called with interrupts disabled, the caller is responsible
1259 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1262 * Note: this is probably too low level an operation for use in drivers.
1263 * Please consult with lkml before using this in your driver.
1265 int split_free_page(struct page *page)
1268 unsigned long watermark;
1271 BUG_ON(!PageBuddy(page));
1273 zone = page_zone(page);
1274 order = page_order(page);
1276 /* Obey watermarks as if the page was being allocated */
1277 watermark = low_wmark_pages(zone) + (1 << order);
1278 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1281 /* Remove page from free list */
1282 list_del(&page->lru);
1283 zone->free_area[order].nr_free--;
1284 rmv_page_order(page);
1285 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1287 /* Split into individual pages */
1288 set_page_refcounted(page);
1289 split_page(page, order);
1291 if (order >= pageblock_order - 1) {
1292 struct page *endpage = page + (1 << order) - 1;
1293 for (; page < endpage; page += pageblock_nr_pages)
1294 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1301 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1302 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1306 struct page *buffered_rmqueue(struct zone *preferred_zone,
1307 struct zone *zone, int order, gfp_t gfp_flags,
1310 unsigned long flags;
1312 int cold = !!(gfp_flags & __GFP_COLD);
1315 if (likely(order == 0)) {
1316 struct per_cpu_pages *pcp;
1317 struct list_head *list;
1319 local_irq_save(flags);
1320 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1321 list = &pcp->lists[migratetype];
1322 if (list_empty(list)) {
1323 pcp->count += rmqueue_bulk(zone, 0,
1326 if (unlikely(list_empty(list)))
1331 page = list_entry(list->prev, struct page, lru);
1333 page = list_entry(list->next, struct page, lru);
1335 list_del(&page->lru);
1338 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1340 * __GFP_NOFAIL is not to be used in new code.
1342 * All __GFP_NOFAIL callers should be fixed so that they
1343 * properly detect and handle allocation failures.
1345 * We most definitely don't want callers attempting to
1346 * allocate greater than order-1 page units with
1349 WARN_ON_ONCE(order > 1);
1351 spin_lock_irqsave(&zone->lock, flags);
1352 page = __rmqueue(zone, order, migratetype);
1353 spin_unlock(&zone->lock);
1356 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1359 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1360 zone_statistics(preferred_zone, zone, gfp_flags);
1361 local_irq_restore(flags);
1363 VM_BUG_ON(bad_range(zone, page));
1364 if (prep_new_page(page, order, gfp_flags))
1369 local_irq_restore(flags);
1373 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1374 #define ALLOC_WMARK_MIN WMARK_MIN
1375 #define ALLOC_WMARK_LOW WMARK_LOW
1376 #define ALLOC_WMARK_HIGH WMARK_HIGH
1377 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1379 /* Mask to get the watermark bits */
1380 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1382 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1383 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1384 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1386 #ifdef CONFIG_FAIL_PAGE_ALLOC
1388 static struct fail_page_alloc_attr {
1389 struct fault_attr attr;
1391 u32 ignore_gfp_highmem;
1392 u32 ignore_gfp_wait;
1395 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1397 struct dentry *ignore_gfp_highmem_file;
1398 struct dentry *ignore_gfp_wait_file;
1399 struct dentry *min_order_file;
1401 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1403 } fail_page_alloc = {
1404 .attr = FAULT_ATTR_INITIALIZER,
1405 .ignore_gfp_wait = 1,
1406 .ignore_gfp_highmem = 1,
1410 static int __init setup_fail_page_alloc(char *str)
1412 return setup_fault_attr(&fail_page_alloc.attr, str);
1414 __setup("fail_page_alloc=", setup_fail_page_alloc);
1416 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1418 if (order < fail_page_alloc.min_order)
1420 if (gfp_mask & __GFP_NOFAIL)
1422 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1424 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1427 return should_fail(&fail_page_alloc.attr, 1 << order);
1430 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1432 static int __init fail_page_alloc_debugfs(void)
1434 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1438 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1442 dir = fail_page_alloc.attr.dentries.dir;
1444 fail_page_alloc.ignore_gfp_wait_file =
1445 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1446 &fail_page_alloc.ignore_gfp_wait);
1448 fail_page_alloc.ignore_gfp_highmem_file =
1449 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1450 &fail_page_alloc.ignore_gfp_highmem);
1451 fail_page_alloc.min_order_file =
1452 debugfs_create_u32("min-order", mode, dir,
1453 &fail_page_alloc.min_order);
1455 if (!fail_page_alloc.ignore_gfp_wait_file ||
1456 !fail_page_alloc.ignore_gfp_highmem_file ||
1457 !fail_page_alloc.min_order_file) {
1459 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1460 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1461 debugfs_remove(fail_page_alloc.min_order_file);
1462 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1468 late_initcall(fail_page_alloc_debugfs);
1470 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1472 #else /* CONFIG_FAIL_PAGE_ALLOC */
1474 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1479 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1482 * Return true if free pages are above 'mark'. This takes into account the order
1483 * of the allocation.
1485 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1486 int classzone_idx, int alloc_flags, long free_pages)
1488 /* free_pages my go negative - that's OK */
1492 free_pages -= (1 << order) + 1;
1493 if (alloc_flags & ALLOC_HIGH)
1495 if (alloc_flags & ALLOC_HARDER)
1498 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1500 for (o = 0; o < order; o++) {
1501 /* At the next order, this order's pages become unavailable */
1502 free_pages -= z->free_area[o].nr_free << o;
1504 /* Require fewer higher order pages to be free */
1505 min >>= min_free_order_shift;
1507 if (free_pages <= min)
1513 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1514 int classzone_idx, int alloc_flags)
1516 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1517 zone_page_state(z, NR_FREE_PAGES));
1520 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1521 int classzone_idx, int alloc_flags)
1523 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1525 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1526 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1528 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1534 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1535 * skip over zones that are not allowed by the cpuset, or that have
1536 * been recently (in last second) found to be nearly full. See further
1537 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1538 * that have to skip over a lot of full or unallowed zones.
1540 * If the zonelist cache is present in the passed in zonelist, then
1541 * returns a pointer to the allowed node mask (either the current
1542 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1544 * If the zonelist cache is not available for this zonelist, does
1545 * nothing and returns NULL.
1547 * If the fullzones BITMAP in the zonelist cache is stale (more than
1548 * a second since last zap'd) then we zap it out (clear its bits.)
1550 * We hold off even calling zlc_setup, until after we've checked the
1551 * first zone in the zonelist, on the theory that most allocations will
1552 * be satisfied from that first zone, so best to examine that zone as
1553 * quickly as we can.
1555 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1557 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1558 nodemask_t *allowednodes; /* zonelist_cache approximation */
1560 zlc = zonelist->zlcache_ptr;
1564 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1565 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1566 zlc->last_full_zap = jiffies;
1569 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1570 &cpuset_current_mems_allowed :
1571 &node_states[N_HIGH_MEMORY];
1572 return allowednodes;
1576 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1577 * if it is worth looking at further for free memory:
1578 * 1) Check that the zone isn't thought to be full (doesn't have its
1579 * bit set in the zonelist_cache fullzones BITMAP).
1580 * 2) Check that the zones node (obtained from the zonelist_cache
1581 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1582 * Return true (non-zero) if zone is worth looking at further, or
1583 * else return false (zero) if it is not.
1585 * This check -ignores- the distinction between various watermarks,
1586 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1587 * found to be full for any variation of these watermarks, it will
1588 * be considered full for up to one second by all requests, unless
1589 * we are so low on memory on all allowed nodes that we are forced
1590 * into the second scan of the zonelist.
1592 * In the second scan we ignore this zonelist cache and exactly
1593 * apply the watermarks to all zones, even it is slower to do so.
1594 * We are low on memory in the second scan, and should leave no stone
1595 * unturned looking for a free page.
1597 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1598 nodemask_t *allowednodes)
1600 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1601 int i; /* index of *z in zonelist zones */
1602 int n; /* node that zone *z is on */
1604 zlc = zonelist->zlcache_ptr;
1608 i = z - zonelist->_zonerefs;
1611 /* This zone is worth trying if it is allowed but not full */
1612 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1616 * Given 'z' scanning a zonelist, set the corresponding bit in
1617 * zlc->fullzones, so that subsequent attempts to allocate a page
1618 * from that zone don't waste time re-examining it.
1620 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1622 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1623 int i; /* index of *z in zonelist zones */
1625 zlc = zonelist->zlcache_ptr;
1629 i = z - zonelist->_zonerefs;
1631 set_bit(i, zlc->fullzones);
1635 * clear all zones full, called after direct reclaim makes progress so that
1636 * a zone that was recently full is not skipped over for up to a second
1638 static void zlc_clear_zones_full(struct zonelist *zonelist)
1640 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1642 zlc = zonelist->zlcache_ptr;
1646 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1649 #else /* CONFIG_NUMA */
1651 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1656 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1657 nodemask_t *allowednodes)
1662 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1666 static void zlc_clear_zones_full(struct zonelist *zonelist)
1669 #endif /* CONFIG_NUMA */
1672 * get_page_from_freelist goes through the zonelist trying to allocate
1675 static struct page *
1676 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1677 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1678 struct zone *preferred_zone, int migratetype)
1681 struct page *page = NULL;
1684 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1685 int zlc_active = 0; /* set if using zonelist_cache */
1686 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1688 classzone_idx = zone_idx(preferred_zone);
1691 * Scan zonelist, looking for a zone with enough free.
1692 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1694 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1695 high_zoneidx, nodemask) {
1696 if (NUMA_BUILD && zlc_active &&
1697 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1699 if ((alloc_flags & ALLOC_CPUSET) &&
1700 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1703 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1704 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1708 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1709 if (zone_watermark_ok(zone, order, mark,
1710 classzone_idx, alloc_flags))
1713 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1715 * we do zlc_setup if there are multiple nodes
1716 * and before considering the first zone allowed
1719 allowednodes = zlc_setup(zonelist, alloc_flags);
1724 if (zone_reclaim_mode == 0)
1725 goto this_zone_full;
1728 * As we may have just activated ZLC, check if the first
1729 * eligible zone has failed zone_reclaim recently.
1731 if (NUMA_BUILD && zlc_active &&
1732 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1735 ret = zone_reclaim(zone, gfp_mask, order);
1737 case ZONE_RECLAIM_NOSCAN:
1740 case ZONE_RECLAIM_FULL:
1741 /* scanned but unreclaimable */
1744 /* did we reclaim enough */
1745 if (!zone_watermark_ok(zone, order, mark,
1746 classzone_idx, alloc_flags))
1747 goto this_zone_full;
1752 page = buffered_rmqueue(preferred_zone, zone, order,
1753 gfp_mask, migratetype);
1758 zlc_mark_zone_full(zonelist, z);
1761 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1762 /* Disable zlc cache for second zonelist scan */
1770 * Large machines with many possible nodes should not always dump per-node
1771 * meminfo in irq context.
1773 static inline bool should_suppress_show_mem(void)
1778 ret = in_interrupt();
1783 static DEFINE_RATELIMIT_STATE(nopage_rs,
1784 DEFAULT_RATELIMIT_INTERVAL,
1785 DEFAULT_RATELIMIT_BURST);
1787 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1790 unsigned int filter = SHOW_MEM_FILTER_NODES;
1792 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1796 * This documents exceptions given to allocations in certain
1797 * contexts that are allowed to allocate outside current's set
1800 if (!(gfp_mask & __GFP_NOMEMALLOC))
1801 if (test_thread_flag(TIF_MEMDIE) ||
1802 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1803 filter &= ~SHOW_MEM_FILTER_NODES;
1804 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1805 filter &= ~SHOW_MEM_FILTER_NODES;
1808 printk(KERN_WARNING);
1809 va_start(args, fmt);
1814 pr_warning("%s: page allocation failure: order:%d, mode:0x%x\n",
1815 current->comm, order, gfp_mask);
1818 if (!should_suppress_show_mem())
1823 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1824 unsigned long pages_reclaimed)
1826 /* Do not loop if specifically requested */
1827 if (gfp_mask & __GFP_NORETRY)
1831 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1832 * means __GFP_NOFAIL, but that may not be true in other
1835 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1839 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1840 * specified, then we retry until we no longer reclaim any pages
1841 * (above), or we've reclaimed an order of pages at least as
1842 * large as the allocation's order. In both cases, if the
1843 * allocation still fails, we stop retrying.
1845 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1849 * Don't let big-order allocations loop unless the caller
1850 * explicitly requests that.
1852 if (gfp_mask & __GFP_NOFAIL)
1858 static inline struct page *
1859 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1860 struct zonelist *zonelist, enum zone_type high_zoneidx,
1861 nodemask_t *nodemask, struct zone *preferred_zone,
1866 /* Acquire the OOM killer lock for the zones in zonelist */
1867 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1868 schedule_timeout_uninterruptible(1);
1873 * Go through the zonelist yet one more time, keep very high watermark
1874 * here, this is only to catch a parallel oom killing, we must fail if
1875 * we're still under heavy pressure.
1877 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1878 order, zonelist, high_zoneidx,
1879 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1880 preferred_zone, migratetype);
1884 if (!(gfp_mask & __GFP_NOFAIL)) {
1885 /* The OOM killer will not help higher order allocs */
1886 if (order > PAGE_ALLOC_COSTLY_ORDER)
1888 /* The OOM killer does not needlessly kill tasks for lowmem */
1889 if (high_zoneidx < ZONE_NORMAL)
1892 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1893 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1894 * The caller should handle page allocation failure by itself if
1895 * it specifies __GFP_THISNODE.
1896 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1898 if (gfp_mask & __GFP_THISNODE)
1901 /* Exhausted what can be done so it's blamo time */
1902 out_of_memory(zonelist, gfp_mask, order, nodemask);
1905 clear_zonelist_oom(zonelist, gfp_mask);
1909 #ifdef CONFIG_COMPACTION
1910 /* Try memory compaction for high-order allocations before reclaim */
1911 static struct page *
1912 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1913 struct zonelist *zonelist, enum zone_type high_zoneidx,
1914 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1915 int migratetype, bool sync_migration,
1916 bool *deferred_compaction,
1917 unsigned long *did_some_progress)
1924 if (compaction_deferred(preferred_zone)) {
1925 *deferred_compaction = true;
1929 current->flags |= PF_MEMALLOC;
1930 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1931 nodemask, sync_migration);
1932 current->flags &= ~PF_MEMALLOC;
1933 if (*did_some_progress != COMPACT_SKIPPED) {
1935 /* Page migration frees to the PCP lists but we want merging */
1936 drain_pages(get_cpu());
1939 page = get_page_from_freelist(gfp_mask, nodemask,
1940 order, zonelist, high_zoneidx,
1941 alloc_flags, preferred_zone,
1944 preferred_zone->compact_considered = 0;
1945 preferred_zone->compact_defer_shift = 0;
1946 count_vm_event(COMPACTSUCCESS);
1951 * It's bad if compaction run occurs and fails.
1952 * The most likely reason is that pages exist,
1953 * but not enough to satisfy watermarks.
1955 count_vm_event(COMPACTFAIL);
1958 * As async compaction considers a subset of pageblocks, only
1959 * defer if the failure was a sync compaction failure.
1962 defer_compaction(preferred_zone);
1970 static inline struct page *
1971 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1972 struct zonelist *zonelist, enum zone_type high_zoneidx,
1973 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1974 int migratetype, bool sync_migration,
1975 bool *deferred_compaction,
1976 unsigned long *did_some_progress)
1980 #endif /* CONFIG_COMPACTION */
1982 /* The really slow allocator path where we enter direct reclaim */
1983 static inline struct page *
1984 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1985 struct zonelist *zonelist, enum zone_type high_zoneidx,
1986 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1987 int migratetype, unsigned long *did_some_progress)
1989 struct page *page = NULL;
1990 struct reclaim_state reclaim_state;
1991 bool drained = false;
1995 /* We now go into synchronous reclaim */
1996 cpuset_memory_pressure_bump();
1997 current->flags |= PF_MEMALLOC;
1998 lockdep_set_current_reclaim_state(gfp_mask);
1999 reclaim_state.reclaimed_slab = 0;
2000 current->reclaim_state = &reclaim_state;
2002 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2004 current->reclaim_state = NULL;
2005 lockdep_clear_current_reclaim_state();
2006 current->flags &= ~PF_MEMALLOC;
2010 if (unlikely(!(*did_some_progress)))
2013 /* After successful reclaim, reconsider all zones for allocation */
2015 zlc_clear_zones_full(zonelist);
2018 page = get_page_from_freelist(gfp_mask, nodemask, order,
2019 zonelist, high_zoneidx,
2020 alloc_flags, preferred_zone,
2024 * If an allocation failed after direct reclaim, it could be because
2025 * pages are pinned on the per-cpu lists. Drain them and try again
2027 if (!page && !drained) {
2037 * This is called in the allocator slow-path if the allocation request is of
2038 * sufficient urgency to ignore watermarks and take other desperate measures
2040 static inline struct page *
2041 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2042 struct zonelist *zonelist, enum zone_type high_zoneidx,
2043 nodemask_t *nodemask, struct zone *preferred_zone,
2049 page = get_page_from_freelist(gfp_mask, nodemask, order,
2050 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2051 preferred_zone, migratetype);
2053 if (!page && gfp_mask & __GFP_NOFAIL)
2054 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2055 } while (!page && (gfp_mask & __GFP_NOFAIL));
2061 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2062 enum zone_type high_zoneidx,
2063 enum zone_type classzone_idx)
2068 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2069 wakeup_kswapd(zone, order, classzone_idx);
2073 gfp_to_alloc_flags(gfp_t gfp_mask)
2075 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2076 const gfp_t wait = gfp_mask & __GFP_WAIT;
2078 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2079 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2082 * The caller may dip into page reserves a bit more if the caller
2083 * cannot run direct reclaim, or if the caller has realtime scheduling
2084 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2085 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2087 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2091 * Not worth trying to allocate harder for
2092 * __GFP_NOMEMALLOC even if it can't schedule.
2094 if (!(gfp_mask & __GFP_NOMEMALLOC))
2095 alloc_flags |= ALLOC_HARDER;
2097 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2098 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2100 alloc_flags &= ~ALLOC_CPUSET;
2101 } else if (unlikely(rt_task(current)) && !in_interrupt())
2102 alloc_flags |= ALLOC_HARDER;
2104 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2105 if (!in_interrupt() &&
2106 ((current->flags & PF_MEMALLOC) ||
2107 unlikely(test_thread_flag(TIF_MEMDIE))))
2108 alloc_flags |= ALLOC_NO_WATERMARKS;
2114 static inline struct page *
2115 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2116 struct zonelist *zonelist, enum zone_type high_zoneidx,
2117 nodemask_t *nodemask, struct zone *preferred_zone,
2120 const gfp_t wait = gfp_mask & __GFP_WAIT;
2121 struct page *page = NULL;
2123 unsigned long pages_reclaimed = 0;
2124 unsigned long did_some_progress;
2125 bool sync_migration = false;
2126 bool deferred_compaction = false;
2129 * In the slowpath, we sanity check order to avoid ever trying to
2130 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2131 * be using allocators in order of preference for an area that is
2134 if (order >= MAX_ORDER) {
2135 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2140 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2141 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2142 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2143 * using a larger set of nodes after it has established that the
2144 * allowed per node queues are empty and that nodes are
2147 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2151 if (!(gfp_mask & __GFP_NO_KSWAPD))
2152 wake_all_kswapd(order, zonelist, high_zoneidx,
2153 zone_idx(preferred_zone));
2156 * OK, we're below the kswapd watermark and have kicked background
2157 * reclaim. Now things get more complex, so set up alloc_flags according
2158 * to how we want to proceed.
2160 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2163 * Find the true preferred zone if the allocation is unconstrained by
2166 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2167 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2171 /* This is the last chance, in general, before the goto nopage. */
2172 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2173 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2174 preferred_zone, migratetype);
2178 /* Allocate without watermarks if the context allows */
2179 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2180 page = __alloc_pages_high_priority(gfp_mask, order,
2181 zonelist, high_zoneidx, nodemask,
2182 preferred_zone, migratetype);
2187 /* Atomic allocations - we can't balance anything */
2191 /* Avoid recursion of direct reclaim */
2192 if (current->flags & PF_MEMALLOC)
2195 /* Avoid allocations with no watermarks from looping endlessly */
2196 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2200 * Try direct compaction. The first pass is asynchronous. Subsequent
2201 * attempts after direct reclaim are synchronous
2203 page = __alloc_pages_direct_compact(gfp_mask, order,
2204 zonelist, high_zoneidx,
2206 alloc_flags, preferred_zone,
2207 migratetype, sync_migration,
2208 &deferred_compaction,
2209 &did_some_progress);
2212 sync_migration = true;
2215 * If compaction is deferred for high-order allocations, it is because
2216 * sync compaction recently failed. In this is the case and the caller
2217 * has requested the system not be heavily disrupted, fail the
2218 * allocation now instead of entering direct reclaim
2220 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2223 /* Try direct reclaim and then allocating */
2224 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2225 zonelist, high_zoneidx,
2227 alloc_flags, preferred_zone,
2228 migratetype, &did_some_progress);
2233 * If we failed to make any progress reclaiming, then we are
2234 * running out of options and have to consider going OOM
2236 if (!did_some_progress) {
2237 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2238 if (oom_killer_disabled)
2240 page = __alloc_pages_may_oom(gfp_mask, order,
2241 zonelist, high_zoneidx,
2242 nodemask, preferred_zone,
2247 if (!(gfp_mask & __GFP_NOFAIL)) {
2249 * The oom killer is not called for high-order
2250 * allocations that may fail, so if no progress
2251 * is being made, there are no other options and
2252 * retrying is unlikely to help.
2254 if (order > PAGE_ALLOC_COSTLY_ORDER)
2257 * The oom killer is not called for lowmem
2258 * allocations to prevent needlessly killing
2261 if (high_zoneidx < ZONE_NORMAL)
2269 * Suspend converts GFP_KERNEL to __GFP_WAIT which can
2270 * prevent reclaim making forward progress without
2271 * invoking OOM. Bail if we are suspending
2273 if (pm_suspending())
2277 /* Check if we should retry the allocation */
2278 pages_reclaimed += did_some_progress;
2279 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2280 /* Wait for some write requests to complete then retry */
2281 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2285 * High-order allocations do not necessarily loop after
2286 * direct reclaim and reclaim/compaction depends on compaction
2287 * being called after reclaim so call directly if necessary
2289 page = __alloc_pages_direct_compact(gfp_mask, order,
2290 zonelist, high_zoneidx,
2292 alloc_flags, preferred_zone,
2293 migratetype, sync_migration,
2294 &deferred_compaction,
2295 &did_some_progress);
2301 warn_alloc_failed(gfp_mask, order, NULL);
2304 if (kmemcheck_enabled)
2305 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2311 * This is the 'heart' of the zoned buddy allocator.
2314 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2315 struct zonelist *zonelist, nodemask_t *nodemask)
2317 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2318 struct zone *preferred_zone;
2319 struct page *page = NULL;
2320 int migratetype = allocflags_to_migratetype(gfp_mask);
2321 unsigned int cpuset_mems_cookie;
2323 gfp_mask &= gfp_allowed_mask;
2325 lockdep_trace_alloc(gfp_mask);
2327 might_sleep_if(gfp_mask & __GFP_WAIT);
2329 if (should_fail_alloc_page(gfp_mask, order))
2333 * Check the zones suitable for the gfp_mask contain at least one
2334 * valid zone. It's possible to have an empty zonelist as a result
2335 * of GFP_THISNODE and a memoryless node
2337 if (unlikely(!zonelist->_zonerefs->zone))
2341 cpuset_mems_cookie = get_mems_allowed();
2343 /* The preferred zone is used for statistics later */
2344 first_zones_zonelist(zonelist, high_zoneidx,
2345 nodemask ? : &cpuset_current_mems_allowed,
2347 if (!preferred_zone)
2350 /* First allocation attempt */
2351 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2352 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2353 preferred_zone, migratetype);
2354 if (unlikely(!page))
2355 page = __alloc_pages_slowpath(gfp_mask, order,
2356 zonelist, high_zoneidx, nodemask,
2357 preferred_zone, migratetype);
2359 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2363 * When updating a task's mems_allowed, it is possible to race with
2364 * parallel threads in such a way that an allocation can fail while
2365 * the mask is being updated. If a page allocation is about to fail,
2366 * check if the cpuset changed during allocation and if so, retry.
2368 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2373 EXPORT_SYMBOL(__alloc_pages_nodemask);
2376 * Common helper functions.
2378 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2383 * __get_free_pages() returns a 32-bit address, which cannot represent
2386 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2388 page = alloc_pages(gfp_mask, order);
2391 return (unsigned long) page_address(page);
2393 EXPORT_SYMBOL(__get_free_pages);
2395 unsigned long get_zeroed_page(gfp_t gfp_mask)
2397 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2399 EXPORT_SYMBOL(get_zeroed_page);
2401 void __pagevec_free(struct pagevec *pvec)
2403 int i = pagevec_count(pvec);
2406 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2407 free_hot_cold_page(pvec->pages[i], pvec->cold);
2411 void __free_pages(struct page *page, unsigned int order)
2413 if (put_page_testzero(page)) {
2415 free_hot_cold_page(page, 0);
2417 __free_pages_ok(page, order);
2421 EXPORT_SYMBOL(__free_pages);
2423 void free_pages(unsigned long addr, unsigned int order)
2426 VM_BUG_ON(!virt_addr_valid((void *)addr));
2427 __free_pages(virt_to_page((void *)addr), order);
2431 EXPORT_SYMBOL(free_pages);
2433 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2436 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2437 unsigned long used = addr + PAGE_ALIGN(size);
2439 split_page(virt_to_page((void *)addr), order);
2440 while (used < alloc_end) {
2445 return (void *)addr;
2449 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2450 * @size: the number of bytes to allocate
2451 * @gfp_mask: GFP flags for the allocation
2453 * This function is similar to alloc_pages(), except that it allocates the
2454 * minimum number of pages to satisfy the request. alloc_pages() can only
2455 * allocate memory in power-of-two pages.
2457 * This function is also limited by MAX_ORDER.
2459 * Memory allocated by this function must be released by free_pages_exact().
2461 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2463 unsigned int order = get_order(size);
2466 addr = __get_free_pages(gfp_mask, order);
2467 return make_alloc_exact(addr, order, size);
2469 EXPORT_SYMBOL(alloc_pages_exact);
2472 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2474 * @nid: the preferred node ID where memory should be allocated
2475 * @size: the number of bytes to allocate
2476 * @gfp_mask: GFP flags for the allocation
2478 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2480 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2483 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2485 unsigned order = get_order(size);
2486 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2489 return make_alloc_exact((unsigned long)page_address(p), order, size);
2491 EXPORT_SYMBOL(alloc_pages_exact_nid);
2494 * free_pages_exact - release memory allocated via alloc_pages_exact()
2495 * @virt: the value returned by alloc_pages_exact.
2496 * @size: size of allocation, same value as passed to alloc_pages_exact().
2498 * Release the memory allocated by a previous call to alloc_pages_exact.
2500 void free_pages_exact(void *virt, size_t size)
2502 unsigned long addr = (unsigned long)virt;
2503 unsigned long end = addr + PAGE_ALIGN(size);
2505 while (addr < end) {
2510 EXPORT_SYMBOL(free_pages_exact);
2512 static unsigned int nr_free_zone_pages(int offset)
2517 /* Just pick one node, since fallback list is circular */
2518 unsigned int sum = 0;
2520 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2522 for_each_zone_zonelist(zone, z, zonelist, offset) {
2523 unsigned long size = zone->present_pages;
2524 unsigned long high = high_wmark_pages(zone);
2533 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2535 unsigned int nr_free_buffer_pages(void)
2537 return nr_free_zone_pages(gfp_zone(GFP_USER));
2539 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2542 * Amount of free RAM allocatable within all zones
2544 unsigned int nr_free_pagecache_pages(void)
2546 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2549 static inline void show_node(struct zone *zone)
2552 printk("Node %d ", zone_to_nid(zone));
2555 void si_meminfo(struct sysinfo *val)
2557 val->totalram = totalram_pages;
2559 val->freeram = global_page_state(NR_FREE_PAGES);
2560 val->bufferram = nr_blockdev_pages();
2561 val->totalhigh = totalhigh_pages;
2562 val->freehigh = nr_free_highpages();
2563 val->mem_unit = PAGE_SIZE;
2566 EXPORT_SYMBOL(si_meminfo);
2569 void si_meminfo_node(struct sysinfo *val, int nid)
2571 pg_data_t *pgdat = NODE_DATA(nid);
2573 val->totalram = pgdat->node_present_pages;
2574 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2575 #ifdef CONFIG_HIGHMEM
2576 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2577 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2583 val->mem_unit = PAGE_SIZE;
2588 * Determine whether the node should be displayed or not, depending on whether
2589 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2591 bool skip_free_areas_node(unsigned int flags, int nid)
2594 unsigned int cpuset_mems_cookie;
2596 if (!(flags & SHOW_MEM_FILTER_NODES))
2600 cpuset_mems_cookie = get_mems_allowed();
2601 ret = !node_isset(nid, cpuset_current_mems_allowed);
2602 } while (!put_mems_allowed(cpuset_mems_cookie));
2607 #define K(x) ((x) << (PAGE_SHIFT-10))
2610 * Show free area list (used inside shift_scroll-lock stuff)
2611 * We also calculate the percentage fragmentation. We do this by counting the
2612 * memory on each free list with the exception of the first item on the list.
2613 * Suppresses nodes that are not allowed by current's cpuset if
2614 * SHOW_MEM_FILTER_NODES is passed.
2616 void show_free_areas(unsigned int filter)
2621 for_each_populated_zone(zone) {
2622 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2625 printk("%s per-cpu:\n", zone->name);
2627 for_each_online_cpu(cpu) {
2628 struct per_cpu_pageset *pageset;
2630 pageset = per_cpu_ptr(zone->pageset, cpu);
2632 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2633 cpu, pageset->pcp.high,
2634 pageset->pcp.batch, pageset->pcp.count);
2638 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2639 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2641 " dirty:%lu writeback:%lu unstable:%lu\n"
2642 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2643 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2644 global_page_state(NR_ACTIVE_ANON),
2645 global_page_state(NR_INACTIVE_ANON),
2646 global_page_state(NR_ISOLATED_ANON),
2647 global_page_state(NR_ACTIVE_FILE),
2648 global_page_state(NR_INACTIVE_FILE),
2649 global_page_state(NR_ISOLATED_FILE),
2650 global_page_state(NR_UNEVICTABLE),
2651 global_page_state(NR_FILE_DIRTY),
2652 global_page_state(NR_WRITEBACK),
2653 global_page_state(NR_UNSTABLE_NFS),
2654 global_page_state(NR_FREE_PAGES),
2655 global_page_state(NR_SLAB_RECLAIMABLE),
2656 global_page_state(NR_SLAB_UNRECLAIMABLE),
2657 global_page_state(NR_FILE_MAPPED),
2658 global_page_state(NR_SHMEM),
2659 global_page_state(NR_PAGETABLE),
2660 global_page_state(NR_BOUNCE));
2662 for_each_populated_zone(zone) {
2665 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2673 " active_anon:%lukB"
2674 " inactive_anon:%lukB"
2675 " active_file:%lukB"
2676 " inactive_file:%lukB"
2677 " unevictable:%lukB"
2678 " isolated(anon):%lukB"
2679 " isolated(file):%lukB"
2686 " slab_reclaimable:%lukB"
2687 " slab_unreclaimable:%lukB"
2688 " kernel_stack:%lukB"
2692 " writeback_tmp:%lukB"
2693 " pages_scanned:%lu"
2694 " all_unreclaimable? %s"
2697 K(zone_page_state(zone, NR_FREE_PAGES)),
2698 K(min_wmark_pages(zone)),
2699 K(low_wmark_pages(zone)),
2700 K(high_wmark_pages(zone)),
2701 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2702 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2703 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2704 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2705 K(zone_page_state(zone, NR_UNEVICTABLE)),
2706 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2707 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2708 K(zone->present_pages),
2709 K(zone_page_state(zone, NR_MLOCK)),
2710 K(zone_page_state(zone, NR_FILE_DIRTY)),
2711 K(zone_page_state(zone, NR_WRITEBACK)),
2712 K(zone_page_state(zone, NR_FILE_MAPPED)),
2713 K(zone_page_state(zone, NR_SHMEM)),
2714 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2715 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2716 zone_page_state(zone, NR_KERNEL_STACK) *
2718 K(zone_page_state(zone, NR_PAGETABLE)),
2719 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2720 K(zone_page_state(zone, NR_BOUNCE)),
2721 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2722 zone->pages_scanned,
2723 (zone->all_unreclaimable ? "yes" : "no")
2725 printk("lowmem_reserve[]:");
2726 for (i = 0; i < MAX_NR_ZONES; i++)
2727 printk(" %lu", zone->lowmem_reserve[i]);
2731 for_each_populated_zone(zone) {
2732 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2734 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2737 printk("%s: ", zone->name);
2739 spin_lock_irqsave(&zone->lock, flags);
2740 for (order = 0; order < MAX_ORDER; order++) {
2741 nr[order] = zone->free_area[order].nr_free;
2742 total += nr[order] << order;
2744 spin_unlock_irqrestore(&zone->lock, flags);
2745 for (order = 0; order < MAX_ORDER; order++)
2746 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2747 printk("= %lukB\n", K(total));
2750 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2752 show_swap_cache_info();
2755 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2757 zoneref->zone = zone;
2758 zoneref->zone_idx = zone_idx(zone);
2762 * Builds allocation fallback zone lists.
2764 * Add all populated zones of a node to the zonelist.
2766 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2767 int nr_zones, enum zone_type zone_type)
2771 BUG_ON(zone_type >= MAX_NR_ZONES);
2776 zone = pgdat->node_zones + zone_type;
2777 if (populated_zone(zone)) {
2778 zoneref_set_zone(zone,
2779 &zonelist->_zonerefs[nr_zones++]);
2780 check_highest_zone(zone_type);
2783 } while (zone_type);
2790 * 0 = automatic detection of better ordering.
2791 * 1 = order by ([node] distance, -zonetype)
2792 * 2 = order by (-zonetype, [node] distance)
2794 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2795 * the same zonelist. So only NUMA can configure this param.
2797 #define ZONELIST_ORDER_DEFAULT 0
2798 #define ZONELIST_ORDER_NODE 1
2799 #define ZONELIST_ORDER_ZONE 2
2801 /* zonelist order in the kernel.
2802 * set_zonelist_order() will set this to NODE or ZONE.
2804 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2805 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2809 /* The value user specified ....changed by config */
2810 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2811 /* string for sysctl */
2812 #define NUMA_ZONELIST_ORDER_LEN 16
2813 char numa_zonelist_order[16] = "default";
2816 * interface for configure zonelist ordering.
2817 * command line option "numa_zonelist_order"
2818 * = "[dD]efault - default, automatic configuration.
2819 * = "[nN]ode - order by node locality, then by zone within node
2820 * = "[zZ]one - order by zone, then by locality within zone
2823 static int __parse_numa_zonelist_order(char *s)
2825 if (*s == 'd' || *s == 'D') {
2826 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2827 } else if (*s == 'n' || *s == 'N') {
2828 user_zonelist_order = ZONELIST_ORDER_NODE;
2829 } else if (*s == 'z' || *s == 'Z') {
2830 user_zonelist_order = ZONELIST_ORDER_ZONE;
2833 "Ignoring invalid numa_zonelist_order value: "
2840 static __init int setup_numa_zonelist_order(char *s)
2847 ret = __parse_numa_zonelist_order(s);
2849 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2853 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2856 * sysctl handler for numa_zonelist_order
2858 int numa_zonelist_order_handler(ctl_table *table, int write,
2859 void __user *buffer, size_t *length,
2862 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2864 static DEFINE_MUTEX(zl_order_mutex);
2866 mutex_lock(&zl_order_mutex);
2868 strcpy(saved_string, (char*)table->data);
2869 ret = proc_dostring(table, write, buffer, length, ppos);
2873 int oldval = user_zonelist_order;
2874 if (__parse_numa_zonelist_order((char*)table->data)) {
2876 * bogus value. restore saved string
2878 strncpy((char*)table->data, saved_string,
2879 NUMA_ZONELIST_ORDER_LEN);
2880 user_zonelist_order = oldval;
2881 } else if (oldval != user_zonelist_order) {
2882 mutex_lock(&zonelists_mutex);
2883 build_all_zonelists(NULL);
2884 mutex_unlock(&zonelists_mutex);
2888 mutex_unlock(&zl_order_mutex);
2893 #define MAX_NODE_LOAD (nr_online_nodes)
2894 static int node_load[MAX_NUMNODES];
2897 * find_next_best_node - find the next node that should appear in a given node's fallback list
2898 * @node: node whose fallback list we're appending
2899 * @used_node_mask: nodemask_t of already used nodes
2901 * We use a number of factors to determine which is the next node that should
2902 * appear on a given node's fallback list. The node should not have appeared
2903 * already in @node's fallback list, and it should be the next closest node
2904 * according to the distance array (which contains arbitrary distance values
2905 * from each node to each node in the system), and should also prefer nodes
2906 * with no CPUs, since presumably they'll have very little allocation pressure
2907 * on them otherwise.
2908 * It returns -1 if no node is found.
2910 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2913 int min_val = INT_MAX;
2915 const struct cpumask *tmp = cpumask_of_node(0);
2917 /* Use the local node if we haven't already */
2918 if (!node_isset(node, *used_node_mask)) {
2919 node_set(node, *used_node_mask);
2923 for_each_node_state(n, N_HIGH_MEMORY) {
2925 /* Don't want a node to appear more than once */
2926 if (node_isset(n, *used_node_mask))
2929 /* Use the distance array to find the distance */
2930 val = node_distance(node, n);
2932 /* Penalize nodes under us ("prefer the next node") */
2935 /* Give preference to headless and unused nodes */
2936 tmp = cpumask_of_node(n);
2937 if (!cpumask_empty(tmp))
2938 val += PENALTY_FOR_NODE_WITH_CPUS;
2940 /* Slight preference for less loaded node */
2941 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2942 val += node_load[n];
2944 if (val < min_val) {
2951 node_set(best_node, *used_node_mask);
2958 * Build zonelists ordered by node and zones within node.
2959 * This results in maximum locality--normal zone overflows into local
2960 * DMA zone, if any--but risks exhausting DMA zone.
2962 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2965 struct zonelist *zonelist;
2967 zonelist = &pgdat->node_zonelists[0];
2968 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2970 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2972 zonelist->_zonerefs[j].zone = NULL;
2973 zonelist->_zonerefs[j].zone_idx = 0;
2977 * Build gfp_thisnode zonelists
2979 static void build_thisnode_zonelists(pg_data_t *pgdat)
2982 struct zonelist *zonelist;
2984 zonelist = &pgdat->node_zonelists[1];
2985 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2986 zonelist->_zonerefs[j].zone = NULL;
2987 zonelist->_zonerefs[j].zone_idx = 0;
2991 * Build zonelists ordered by zone and nodes within zones.
2992 * This results in conserving DMA zone[s] until all Normal memory is
2993 * exhausted, but results in overflowing to remote node while memory
2994 * may still exist in local DMA zone.
2996 static int node_order[MAX_NUMNODES];
2998 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3001 int zone_type; /* needs to be signed */
3003 struct zonelist *zonelist;
3005 zonelist = &pgdat->node_zonelists[0];
3007 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3008 for (j = 0; j < nr_nodes; j++) {
3009 node = node_order[j];
3010 z = &NODE_DATA(node)->node_zones[zone_type];
3011 if (populated_zone(z)) {
3013 &zonelist->_zonerefs[pos++]);
3014 check_highest_zone(zone_type);
3018 zonelist->_zonerefs[pos].zone = NULL;
3019 zonelist->_zonerefs[pos].zone_idx = 0;
3022 static int default_zonelist_order(void)
3025 unsigned long low_kmem_size,total_size;
3029 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3030 * If they are really small and used heavily, the system can fall
3031 * into OOM very easily.
3032 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3034 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3037 for_each_online_node(nid) {
3038 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3039 z = &NODE_DATA(nid)->node_zones[zone_type];
3040 if (populated_zone(z)) {
3041 if (zone_type < ZONE_NORMAL)
3042 low_kmem_size += z->present_pages;
3043 total_size += z->present_pages;
3044 } else if (zone_type == ZONE_NORMAL) {
3046 * If any node has only lowmem, then node order
3047 * is preferred to allow kernel allocations
3048 * locally; otherwise, they can easily infringe
3049 * on other nodes when there is an abundance of
3050 * lowmem available to allocate from.
3052 return ZONELIST_ORDER_NODE;
3056 if (!low_kmem_size || /* there are no DMA area. */
3057 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3058 return ZONELIST_ORDER_NODE;
3060 * look into each node's config.
3061 * If there is a node whose DMA/DMA32 memory is very big area on
3062 * local memory, NODE_ORDER may be suitable.
3064 average_size = total_size /
3065 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3066 for_each_online_node(nid) {
3069 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3070 z = &NODE_DATA(nid)->node_zones[zone_type];
3071 if (populated_zone(z)) {
3072 if (zone_type < ZONE_NORMAL)
3073 low_kmem_size += z->present_pages;
3074 total_size += z->present_pages;
3077 if (low_kmem_size &&
3078 total_size > average_size && /* ignore small node */
3079 low_kmem_size > total_size * 70/100)
3080 return ZONELIST_ORDER_NODE;
3082 return ZONELIST_ORDER_ZONE;
3085 static void set_zonelist_order(void)
3087 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3088 current_zonelist_order = default_zonelist_order();
3090 current_zonelist_order = user_zonelist_order;
3093 static void build_zonelists(pg_data_t *pgdat)
3097 nodemask_t used_mask;
3098 int local_node, prev_node;
3099 struct zonelist *zonelist;
3100 int order = current_zonelist_order;
3102 /* initialize zonelists */
3103 for (i = 0; i < MAX_ZONELISTS; i++) {
3104 zonelist = pgdat->node_zonelists + i;
3105 zonelist->_zonerefs[0].zone = NULL;
3106 zonelist->_zonerefs[0].zone_idx = 0;
3109 /* NUMA-aware ordering of nodes */
3110 local_node = pgdat->node_id;
3111 load = nr_online_nodes;
3112 prev_node = local_node;
3113 nodes_clear(used_mask);
3115 memset(node_order, 0, sizeof(node_order));
3118 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3119 int distance = node_distance(local_node, node);
3122 * If another node is sufficiently far away then it is better
3123 * to reclaim pages in a zone before going off node.
3125 if (distance > RECLAIM_DISTANCE)
3126 zone_reclaim_mode = 1;
3129 * We don't want to pressure a particular node.
3130 * So adding penalty to the first node in same
3131 * distance group to make it round-robin.
3133 if (distance != node_distance(local_node, prev_node))
3134 node_load[node] = load;
3138 if (order == ZONELIST_ORDER_NODE)
3139 build_zonelists_in_node_order(pgdat, node);
3141 node_order[j++] = node; /* remember order */
3144 if (order == ZONELIST_ORDER_ZONE) {
3145 /* calculate node order -- i.e., DMA last! */
3146 build_zonelists_in_zone_order(pgdat, j);
3149 build_thisnode_zonelists(pgdat);
3152 /* Construct the zonelist performance cache - see further mmzone.h */
3153 static void build_zonelist_cache(pg_data_t *pgdat)
3155 struct zonelist *zonelist;
3156 struct zonelist_cache *zlc;
3159 zonelist = &pgdat->node_zonelists[0];
3160 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3161 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3162 for (z = zonelist->_zonerefs; z->zone; z++)
3163 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3166 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3168 * Return node id of node used for "local" allocations.
3169 * I.e., first node id of first zone in arg node's generic zonelist.
3170 * Used for initializing percpu 'numa_mem', which is used primarily
3171 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3173 int local_memory_node(int node)
3177 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3178 gfp_zone(GFP_KERNEL),
3185 #else /* CONFIG_NUMA */
3187 static void set_zonelist_order(void)
3189 current_zonelist_order = ZONELIST_ORDER_ZONE;
3192 static void build_zonelists(pg_data_t *pgdat)
3194 int node, local_node;
3196 struct zonelist *zonelist;
3198 local_node = pgdat->node_id;
3200 zonelist = &pgdat->node_zonelists[0];
3201 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3204 * Now we build the zonelist so that it contains the zones
3205 * of all the other nodes.
3206 * We don't want to pressure a particular node, so when
3207 * building the zones for node N, we make sure that the
3208 * zones coming right after the local ones are those from
3209 * node N+1 (modulo N)
3211 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3212 if (!node_online(node))
3214 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3217 for (node = 0; node < local_node; node++) {
3218 if (!node_online(node))
3220 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3224 zonelist->_zonerefs[j].zone = NULL;
3225 zonelist->_zonerefs[j].zone_idx = 0;
3228 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3229 static void build_zonelist_cache(pg_data_t *pgdat)
3231 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3234 #endif /* CONFIG_NUMA */
3237 * Boot pageset table. One per cpu which is going to be used for all
3238 * zones and all nodes. The parameters will be set in such a way
3239 * that an item put on a list will immediately be handed over to
3240 * the buddy list. This is safe since pageset manipulation is done
3241 * with interrupts disabled.
3243 * The boot_pagesets must be kept even after bootup is complete for
3244 * unused processors and/or zones. They do play a role for bootstrapping
3245 * hotplugged processors.
3247 * zoneinfo_show() and maybe other functions do
3248 * not check if the processor is online before following the pageset pointer.
3249 * Other parts of the kernel may not check if the zone is available.
3251 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3252 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3253 static void setup_zone_pageset(struct zone *zone);
3256 * Global mutex to protect against size modification of zonelists
3257 * as well as to serialize pageset setup for the new populated zone.
3259 DEFINE_MUTEX(zonelists_mutex);
3261 /* return values int ....just for stop_machine() */
3262 static __init_refok int __build_all_zonelists(void *data)
3268 memset(node_load, 0, sizeof(node_load));
3270 for_each_online_node(nid) {
3271 pg_data_t *pgdat = NODE_DATA(nid);
3273 build_zonelists(pgdat);
3274 build_zonelist_cache(pgdat);
3278 * Initialize the boot_pagesets that are going to be used
3279 * for bootstrapping processors. The real pagesets for
3280 * each zone will be allocated later when the per cpu
3281 * allocator is available.
3283 * boot_pagesets are used also for bootstrapping offline
3284 * cpus if the system is already booted because the pagesets
3285 * are needed to initialize allocators on a specific cpu too.
3286 * F.e. the percpu allocator needs the page allocator which
3287 * needs the percpu allocator in order to allocate its pagesets
3288 * (a chicken-egg dilemma).
3290 for_each_possible_cpu(cpu) {
3291 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3293 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3295 * We now know the "local memory node" for each node--
3296 * i.e., the node of the first zone in the generic zonelist.
3297 * Set up numa_mem percpu variable for on-line cpus. During
3298 * boot, only the boot cpu should be on-line; we'll init the
3299 * secondary cpus' numa_mem as they come on-line. During
3300 * node/memory hotplug, we'll fixup all on-line cpus.
3302 if (cpu_online(cpu))
3303 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3311 * Called with zonelists_mutex held always
3312 * unless system_state == SYSTEM_BOOTING.
3314 void __ref build_all_zonelists(void *data)
3316 set_zonelist_order();
3318 if (system_state == SYSTEM_BOOTING) {
3319 __build_all_zonelists(NULL);
3320 mminit_verify_zonelist();
3321 cpuset_init_current_mems_allowed();
3323 /* we have to stop all cpus to guarantee there is no user
3325 #ifdef CONFIG_MEMORY_HOTPLUG
3327 setup_zone_pageset((struct zone *)data);
3329 stop_machine(__build_all_zonelists, NULL, NULL);
3330 /* cpuset refresh routine should be here */
3332 vm_total_pages = nr_free_pagecache_pages();
3334 * Disable grouping by mobility if the number of pages in the
3335 * system is too low to allow the mechanism to work. It would be
3336 * more accurate, but expensive to check per-zone. This check is
3337 * made on memory-hotadd so a system can start with mobility
3338 * disabled and enable it later
3340 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3341 page_group_by_mobility_disabled = 1;
3343 page_group_by_mobility_disabled = 0;
3345 printk("Built %i zonelists in %s order, mobility grouping %s. "
3346 "Total pages: %ld\n",
3348 zonelist_order_name[current_zonelist_order],
3349 page_group_by_mobility_disabled ? "off" : "on",
3352 printk("Policy zone: %s\n", zone_names[policy_zone]);
3357 * Helper functions to size the waitqueue hash table.
3358 * Essentially these want to choose hash table sizes sufficiently
3359 * large so that collisions trying to wait on pages are rare.
3360 * But in fact, the number of active page waitqueues on typical
3361 * systems is ridiculously low, less than 200. So this is even
3362 * conservative, even though it seems large.
3364 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3365 * waitqueues, i.e. the size of the waitq table given the number of pages.
3367 #define PAGES_PER_WAITQUEUE 256
3369 #ifndef CONFIG_MEMORY_HOTPLUG
3370 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3372 unsigned long size = 1;
3374 pages /= PAGES_PER_WAITQUEUE;
3376 while (size < pages)
3380 * Once we have dozens or even hundreds of threads sleeping
3381 * on IO we've got bigger problems than wait queue collision.
3382 * Limit the size of the wait table to a reasonable size.
3384 size = min(size, 4096UL);
3386 return max(size, 4UL);
3390 * A zone's size might be changed by hot-add, so it is not possible to determine
3391 * a suitable size for its wait_table. So we use the maximum size now.
3393 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3395 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3396 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3397 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3399 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3400 * or more by the traditional way. (See above). It equals:
3402 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3403 * ia64(16K page size) : = ( 8G + 4M)byte.
3404 * powerpc (64K page size) : = (32G +16M)byte.
3406 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3413 * This is an integer logarithm so that shifts can be used later
3414 * to extract the more random high bits from the multiplicative
3415 * hash function before the remainder is taken.
3417 static inline unsigned long wait_table_bits(unsigned long size)
3422 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3425 * Check if a pageblock contains reserved pages
3427 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3431 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3432 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3439 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3440 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3441 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3442 * higher will lead to a bigger reserve which will get freed as contiguous
3443 * blocks as reclaim kicks in
3445 static void setup_zone_migrate_reserve(struct zone *zone)
3447 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3449 unsigned long block_migratetype;
3453 * Get the start pfn, end pfn and the number of blocks to reserve
3454 * We have to be careful to be aligned to pageblock_nr_pages to
3455 * make sure that we always check pfn_valid for the first page in
3458 start_pfn = zone->zone_start_pfn;
3459 end_pfn = start_pfn + zone->spanned_pages;
3460 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3461 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3465 * Reserve blocks are generally in place to help high-order atomic
3466 * allocations that are short-lived. A min_free_kbytes value that
3467 * would result in more than 2 reserve blocks for atomic allocations
3468 * is assumed to be in place to help anti-fragmentation for the
3469 * future allocation of hugepages at runtime.
3471 reserve = min(2, reserve);
3473 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3474 if (!pfn_valid(pfn))
3476 page = pfn_to_page(pfn);
3478 /* Watch out for overlapping nodes */
3479 if (page_to_nid(page) != zone_to_nid(zone))
3482 block_migratetype = get_pageblock_migratetype(page);
3484 /* Only test what is necessary when the reserves are not met */
3487 * Blocks with reserved pages will never free, skip
3490 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3491 if (pageblock_is_reserved(pfn, block_end_pfn))
3494 /* If this block is reserved, account for it */
3495 if (block_migratetype == MIGRATE_RESERVE) {
3500 /* Suitable for reserving if this block is movable */
3501 if (block_migratetype == MIGRATE_MOVABLE) {
3502 set_pageblock_migratetype(page,
3504 move_freepages_block(zone, page,
3512 * If the reserve is met and this is a previous reserved block,
3515 if (block_migratetype == MIGRATE_RESERVE) {
3516 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3517 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3523 * Initially all pages are reserved - free ones are freed
3524 * up by free_all_bootmem() once the early boot process is
3525 * done. Non-atomic initialization, single-pass.
3527 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3528 unsigned long start_pfn, enum memmap_context context)
3531 unsigned long end_pfn = start_pfn + size;
3535 if (highest_memmap_pfn < end_pfn - 1)
3536 highest_memmap_pfn = end_pfn - 1;
3538 z = &NODE_DATA(nid)->node_zones[zone];
3539 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3541 * There can be holes in boot-time mem_map[]s
3542 * handed to this function. They do not
3543 * exist on hotplugged memory.
3545 if (context == MEMMAP_EARLY) {
3546 if (!early_pfn_valid(pfn))
3548 if (!early_pfn_in_nid(pfn, nid))
3551 page = pfn_to_page(pfn);
3552 set_page_links(page, zone, nid, pfn);
3553 mminit_verify_page_links(page, zone, nid, pfn);
3554 init_page_count(page);
3555 reset_page_mapcount(page);
3556 SetPageReserved(page);
3558 * Mark the block movable so that blocks are reserved for
3559 * movable at startup. This will force kernel allocations
3560 * to reserve their blocks rather than leaking throughout
3561 * the address space during boot when many long-lived
3562 * kernel allocations are made. Later some blocks near
3563 * the start are marked MIGRATE_RESERVE by
3564 * setup_zone_migrate_reserve()
3566 * bitmap is created for zone's valid pfn range. but memmap
3567 * can be created for invalid pages (for alignment)
3568 * check here not to call set_pageblock_migratetype() against
3571 if ((z->zone_start_pfn <= pfn)
3572 && (pfn < z->zone_start_pfn + z->spanned_pages)
3573 && !(pfn & (pageblock_nr_pages - 1)))
3574 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3576 INIT_LIST_HEAD(&page->lru);
3577 #ifdef WANT_PAGE_VIRTUAL
3578 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3579 if (!is_highmem_idx(zone))
3580 set_page_address(page, __va(pfn << PAGE_SHIFT));
3585 static void __meminit zone_init_free_lists(struct zone *zone)
3588 for_each_migratetype_order(order, t) {
3589 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3590 zone->free_area[order].nr_free = 0;
3594 #ifndef __HAVE_ARCH_MEMMAP_INIT
3595 #define memmap_init(size, nid, zone, start_pfn) \
3596 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3599 static int zone_batchsize(struct zone *zone)
3605 * The per-cpu-pages pools are set to around 1000th of the
3606 * size of the zone. But no more than 1/2 of a meg.
3608 * OK, so we don't know how big the cache is. So guess.
3610 batch = zone->present_pages / 1024;
3611 if (batch * PAGE_SIZE > 512 * 1024)
3612 batch = (512 * 1024) / PAGE_SIZE;
3613 batch /= 4; /* We effectively *= 4 below */
3618 * Clamp the batch to a 2^n - 1 value. Having a power
3619 * of 2 value was found to be more likely to have
3620 * suboptimal cache aliasing properties in some cases.
3622 * For example if 2 tasks are alternately allocating
3623 * batches of pages, one task can end up with a lot
3624 * of pages of one half of the possible page colors
3625 * and the other with pages of the other colors.
3627 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3632 /* The deferral and batching of frees should be suppressed under NOMMU
3635 * The problem is that NOMMU needs to be able to allocate large chunks
3636 * of contiguous memory as there's no hardware page translation to
3637 * assemble apparent contiguous memory from discontiguous pages.
3639 * Queueing large contiguous runs of pages for batching, however,
3640 * causes the pages to actually be freed in smaller chunks. As there
3641 * can be a significant delay between the individual batches being
3642 * recycled, this leads to the once large chunks of space being
3643 * fragmented and becoming unavailable for high-order allocations.
3649 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3651 struct per_cpu_pages *pcp;
3654 memset(p, 0, sizeof(*p));
3658 pcp->high = 6 * batch;
3659 pcp->batch = max(1UL, 1 * batch);
3660 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3661 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3665 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3666 * to the value high for the pageset p.
3669 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3672 struct per_cpu_pages *pcp;
3676 pcp->batch = max(1UL, high/4);
3677 if ((high/4) > (PAGE_SHIFT * 8))
3678 pcp->batch = PAGE_SHIFT * 8;
3681 static void setup_zone_pageset(struct zone *zone)
3685 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3687 for_each_possible_cpu(cpu) {
3688 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3690 setup_pageset(pcp, zone_batchsize(zone));
3692 if (percpu_pagelist_fraction)
3693 setup_pagelist_highmark(pcp,
3694 (zone->present_pages /
3695 percpu_pagelist_fraction));
3700 * Allocate per cpu pagesets and initialize them.
3701 * Before this call only boot pagesets were available.
3703 void __init setup_per_cpu_pageset(void)
3707 for_each_populated_zone(zone)
3708 setup_zone_pageset(zone);
3711 static noinline __init_refok
3712 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3715 struct pglist_data *pgdat = zone->zone_pgdat;
3719 * The per-page waitqueue mechanism uses hashed waitqueues
3722 zone->wait_table_hash_nr_entries =
3723 wait_table_hash_nr_entries(zone_size_pages);
3724 zone->wait_table_bits =
3725 wait_table_bits(zone->wait_table_hash_nr_entries);
3726 alloc_size = zone->wait_table_hash_nr_entries
3727 * sizeof(wait_queue_head_t);
3729 if (!slab_is_available()) {
3730 zone->wait_table = (wait_queue_head_t *)
3731 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3734 * This case means that a zone whose size was 0 gets new memory
3735 * via memory hot-add.
3736 * But it may be the case that a new node was hot-added. In
3737 * this case vmalloc() will not be able to use this new node's
3738 * memory - this wait_table must be initialized to use this new
3739 * node itself as well.
3740 * To use this new node's memory, further consideration will be
3743 zone->wait_table = vmalloc(alloc_size);
3745 if (!zone->wait_table)
3748 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3749 init_waitqueue_head(zone->wait_table + i);
3754 static int __zone_pcp_update(void *data)
3756 struct zone *zone = data;
3758 unsigned long batch = zone_batchsize(zone), flags;
3760 for_each_possible_cpu(cpu) {
3761 struct per_cpu_pageset *pset;
3762 struct per_cpu_pages *pcp;
3764 pset = per_cpu_ptr(zone->pageset, cpu);
3767 local_irq_save(flags);
3768 free_pcppages_bulk(zone, pcp->count, pcp);
3769 setup_pageset(pset, batch);
3770 local_irq_restore(flags);
3775 void zone_pcp_update(struct zone *zone)
3777 stop_machine(__zone_pcp_update, zone, NULL);
3780 static __meminit void zone_pcp_init(struct zone *zone)
3783 * per cpu subsystem is not up at this point. The following code
3784 * relies on the ability of the linker to provide the
3785 * offset of a (static) per cpu variable into the per cpu area.
3787 zone->pageset = &boot_pageset;
3789 if (zone->present_pages)
3790 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3791 zone->name, zone->present_pages,
3792 zone_batchsize(zone));
3795 __meminit int init_currently_empty_zone(struct zone *zone,
3796 unsigned long zone_start_pfn,
3798 enum memmap_context context)
3800 struct pglist_data *pgdat = zone->zone_pgdat;
3802 ret = zone_wait_table_init(zone, size);
3805 pgdat->nr_zones = zone_idx(zone) + 1;
3807 zone->zone_start_pfn = zone_start_pfn;
3809 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3810 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3812 (unsigned long)zone_idx(zone),
3813 zone_start_pfn, (zone_start_pfn + size));
3815 zone_init_free_lists(zone);
3820 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3822 * Basic iterator support. Return the first range of PFNs for a node
3823 * Note: nid == MAX_NUMNODES returns first region regardless of node
3825 static int __meminit first_active_region_index_in_nid(int nid)
3829 for (i = 0; i < nr_nodemap_entries; i++)
3830 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3837 * Basic iterator support. Return the next active range of PFNs for a node
3838 * Note: nid == MAX_NUMNODES returns next region regardless of node
3840 static int __meminit next_active_region_index_in_nid(int index, int nid)
3842 for (index = index + 1; index < nr_nodemap_entries; index++)
3843 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3849 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3851 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3852 * Architectures may implement their own version but if add_active_range()
3853 * was used and there are no special requirements, this is a convenient
3856 int __meminit __early_pfn_to_nid(unsigned long pfn)
3860 for (i = 0; i < nr_nodemap_entries; i++) {
3861 unsigned long start_pfn = early_node_map[i].start_pfn;
3862 unsigned long end_pfn = early_node_map[i].end_pfn;
3864 if (start_pfn <= pfn && pfn < end_pfn)
3865 return early_node_map[i].nid;
3867 /* This is a memory hole */
3870 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3872 int __meminit early_pfn_to_nid(unsigned long pfn)
3876 nid = __early_pfn_to_nid(pfn);
3879 /* just returns 0 */
3883 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3884 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3888 nid = __early_pfn_to_nid(pfn);
3889 if (nid >= 0 && nid != node)
3895 /* Basic iterator support to walk early_node_map[] */
3896 #define for_each_active_range_index_in_nid(i, nid) \
3897 for (i = first_active_region_index_in_nid(nid); i != -1; \
3898 i = next_active_region_index_in_nid(i, nid))
3901 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3902 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3903 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3905 * If an architecture guarantees that all ranges registered with
3906 * add_active_ranges() contain no holes and may be freed, this
3907 * this function may be used instead of calling free_bootmem() manually.
3909 void __init free_bootmem_with_active_regions(int nid,
3910 unsigned long max_low_pfn)
3914 for_each_active_range_index_in_nid(i, nid) {
3915 unsigned long size_pages = 0;
3916 unsigned long end_pfn = early_node_map[i].end_pfn;
3918 if (early_node_map[i].start_pfn >= max_low_pfn)
3921 if (end_pfn > max_low_pfn)
3922 end_pfn = max_low_pfn;
3924 size_pages = end_pfn - early_node_map[i].start_pfn;
3925 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3926 PFN_PHYS(early_node_map[i].start_pfn),
3927 size_pages << PAGE_SHIFT);
3931 #ifdef CONFIG_HAVE_MEMBLOCK
3933 * Basic iterator support. Return the last range of PFNs for a node
3934 * Note: nid == MAX_NUMNODES returns last region regardless of node
3936 static int __meminit last_active_region_index_in_nid(int nid)
3940 for (i = nr_nodemap_entries - 1; i >= 0; i--)
3941 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3948 * Basic iterator support. Return the previous active range of PFNs for a node
3949 * Note: nid == MAX_NUMNODES returns next region regardless of node
3951 static int __meminit previous_active_region_index_in_nid(int index, int nid)
3953 for (index = index - 1; index >= 0; index--)
3954 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3960 #define for_each_active_range_index_in_nid_reverse(i, nid) \
3961 for (i = last_active_region_index_in_nid(nid); i != -1; \
3962 i = previous_active_region_index_in_nid(i, nid))
3964 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3965 u64 goal, u64 limit)
3969 /* Need to go over early_node_map to find out good range for node */
3970 for_each_active_range_index_in_nid_reverse(i, nid) {
3972 u64 ei_start, ei_last;
3973 u64 final_start, final_end;
3975 ei_last = early_node_map[i].end_pfn;
3976 ei_last <<= PAGE_SHIFT;
3977 ei_start = early_node_map[i].start_pfn;
3978 ei_start <<= PAGE_SHIFT;
3980 final_start = max(ei_start, goal);
3981 final_end = min(ei_last, limit);
3983 if (final_start >= final_end)
3986 addr = memblock_find_in_range(final_start, final_end, size, align);
3988 if (addr == MEMBLOCK_ERROR)
3994 return MEMBLOCK_ERROR;
3998 int __init add_from_early_node_map(struct range *range, int az,
3999 int nr_range, int nid)
4004 /* need to go over early_node_map to find out good range for node */
4005 for_each_active_range_index_in_nid(i, nid) {
4006 start = early_node_map[i].start_pfn;
4007 end = early_node_map[i].end_pfn;
4008 nr_range = add_range(range, az, nr_range, start, end);
4013 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
4018 for_each_active_range_index_in_nid(i, nid) {
4019 ret = work_fn(early_node_map[i].start_pfn,
4020 early_node_map[i].end_pfn, data);
4026 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4027 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4029 * If an architecture guarantees that all ranges registered with
4030 * add_active_ranges() contain no holes and may be freed, this
4031 * function may be used instead of calling memory_present() manually.
4033 void __init sparse_memory_present_with_active_regions(int nid)
4037 for_each_active_range_index_in_nid(i, nid)
4038 memory_present(early_node_map[i].nid,
4039 early_node_map[i].start_pfn,
4040 early_node_map[i].end_pfn);
4044 * get_pfn_range_for_nid - Return the start and end page frames for a node
4045 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4046 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4047 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4049 * It returns the start and end page frame of a node based on information
4050 * provided by an arch calling add_active_range(). If called for a node
4051 * with no available memory, a warning is printed and the start and end
4054 void __meminit get_pfn_range_for_nid(unsigned int nid,
4055 unsigned long *start_pfn, unsigned long *end_pfn)
4061 for_each_active_range_index_in_nid(i, nid) {
4062 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
4063 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
4066 if (*start_pfn == -1UL)
4071 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4072 * assumption is made that zones within a node are ordered in monotonic
4073 * increasing memory addresses so that the "highest" populated zone is used
4075 static void __init find_usable_zone_for_movable(void)
4078 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4079 if (zone_index == ZONE_MOVABLE)
4082 if (arch_zone_highest_possible_pfn[zone_index] >
4083 arch_zone_lowest_possible_pfn[zone_index])
4087 VM_BUG_ON(zone_index == -1);
4088 movable_zone = zone_index;
4092 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4093 * because it is sized independent of architecture. Unlike the other zones,
4094 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4095 * in each node depending on the size of each node and how evenly kernelcore
4096 * is distributed. This helper function adjusts the zone ranges
4097 * provided by the architecture for a given node by using the end of the
4098 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4099 * zones within a node are in order of monotonic increases memory addresses
4101 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4102 unsigned long zone_type,
4103 unsigned long node_start_pfn,
4104 unsigned long node_end_pfn,
4105 unsigned long *zone_start_pfn,
4106 unsigned long *zone_end_pfn)
4108 /* Only adjust if ZONE_MOVABLE is on this node */
4109 if (zone_movable_pfn[nid]) {
4110 /* Size ZONE_MOVABLE */
4111 if (zone_type == ZONE_MOVABLE) {
4112 *zone_start_pfn = zone_movable_pfn[nid];
4113 *zone_end_pfn = min(node_end_pfn,
4114 arch_zone_highest_possible_pfn[movable_zone]);
4116 /* Adjust for ZONE_MOVABLE starting within this range */
4117 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4118 *zone_end_pfn > zone_movable_pfn[nid]) {
4119 *zone_end_pfn = zone_movable_pfn[nid];
4121 /* Check if this whole range is within ZONE_MOVABLE */
4122 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4123 *zone_start_pfn = *zone_end_pfn;
4128 * Return the number of pages a zone spans in a node, including holes
4129 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4131 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4132 unsigned long zone_type,
4133 unsigned long *ignored)
4135 unsigned long node_start_pfn, node_end_pfn;
4136 unsigned long zone_start_pfn, zone_end_pfn;
4138 /* Get the start and end of the node and zone */
4139 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4140 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4141 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4142 adjust_zone_range_for_zone_movable(nid, zone_type,
4143 node_start_pfn, node_end_pfn,
4144 &zone_start_pfn, &zone_end_pfn);
4146 /* Check that this node has pages within the zone's required range */
4147 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4150 /* Move the zone boundaries inside the node if necessary */
4151 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4152 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4154 /* Return the spanned pages */
4155 return zone_end_pfn - zone_start_pfn;
4159 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4160 * then all holes in the requested range will be accounted for.
4162 unsigned long __meminit __absent_pages_in_range(int nid,
4163 unsigned long range_start_pfn,
4164 unsigned long range_end_pfn)
4167 unsigned long prev_end_pfn = 0, hole_pages = 0;
4168 unsigned long start_pfn;
4170 /* Find the end_pfn of the first active range of pfns in the node */
4171 i = first_active_region_index_in_nid(nid);
4175 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4177 /* Account for ranges before physical memory on this node */
4178 if (early_node_map[i].start_pfn > range_start_pfn)
4179 hole_pages = prev_end_pfn - range_start_pfn;
4181 /* Find all holes for the zone within the node */
4182 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4184 /* No need to continue if prev_end_pfn is outside the zone */
4185 if (prev_end_pfn >= range_end_pfn)
4188 /* Make sure the end of the zone is not within the hole */
4189 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4190 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4192 /* Update the hole size cound and move on */
4193 if (start_pfn > range_start_pfn) {
4194 BUG_ON(prev_end_pfn > start_pfn);
4195 hole_pages += start_pfn - prev_end_pfn;
4197 prev_end_pfn = early_node_map[i].end_pfn;
4200 /* Account for ranges past physical memory on this node */
4201 if (range_end_pfn > prev_end_pfn)
4202 hole_pages += range_end_pfn -
4203 max(range_start_pfn, prev_end_pfn);
4209 * absent_pages_in_range - Return number of page frames in holes within a range
4210 * @start_pfn: The start PFN to start searching for holes
4211 * @end_pfn: The end PFN to stop searching for holes
4213 * It returns the number of pages frames in memory holes within a range.
4215 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4216 unsigned long end_pfn)
4218 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4221 /* Return the number of page frames in holes in a zone on a node */
4222 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4223 unsigned long zone_type,
4224 unsigned long *ignored)
4226 unsigned long node_start_pfn, node_end_pfn;
4227 unsigned long zone_start_pfn, zone_end_pfn;
4229 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4230 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4232 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4235 adjust_zone_range_for_zone_movable(nid, zone_type,
4236 node_start_pfn, node_end_pfn,
4237 &zone_start_pfn, &zone_end_pfn);
4238 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4242 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4243 unsigned long zone_type,
4244 unsigned long *zones_size)
4246 return zones_size[zone_type];
4249 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4250 unsigned long zone_type,
4251 unsigned long *zholes_size)
4256 return zholes_size[zone_type];
4261 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4262 unsigned long *zones_size, unsigned long *zholes_size)
4264 unsigned long realtotalpages, totalpages = 0;
4267 for (i = 0; i < MAX_NR_ZONES; i++)
4268 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4270 pgdat->node_spanned_pages = totalpages;
4272 realtotalpages = totalpages;
4273 for (i = 0; i < MAX_NR_ZONES; i++)
4275 zone_absent_pages_in_node(pgdat->node_id, i,
4277 pgdat->node_present_pages = realtotalpages;
4278 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4282 #ifndef CONFIG_SPARSEMEM
4284 * Calculate the size of the zone->blockflags rounded to an unsigned long
4285 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4286 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4287 * round what is now in bits to nearest long in bits, then return it in
4290 static unsigned long __init usemap_size(unsigned long zonesize)
4292 unsigned long usemapsize;
4294 usemapsize = roundup(zonesize, pageblock_nr_pages);
4295 usemapsize = usemapsize >> pageblock_order;
4296 usemapsize *= NR_PAGEBLOCK_BITS;
4297 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4299 return usemapsize / 8;
4302 static void __init setup_usemap(struct pglist_data *pgdat,
4303 struct zone *zone, unsigned long zonesize)
4305 unsigned long usemapsize = usemap_size(zonesize);
4306 zone->pageblock_flags = NULL;
4308 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4312 static inline void setup_usemap(struct pglist_data *pgdat,
4313 struct zone *zone, unsigned long zonesize) {}
4314 #endif /* CONFIG_SPARSEMEM */
4316 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4318 /* Return a sensible default order for the pageblock size. */
4319 static inline int pageblock_default_order(void)
4321 if (HPAGE_SHIFT > PAGE_SHIFT)
4322 return HUGETLB_PAGE_ORDER;
4327 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4328 static inline void __init set_pageblock_order(unsigned int order)
4330 /* Check that pageblock_nr_pages has not already been setup */
4331 if (pageblock_order)
4335 * Assume the largest contiguous order of interest is a huge page.
4336 * This value may be variable depending on boot parameters on IA64
4338 pageblock_order = order;
4340 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4343 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4344 * and pageblock_default_order() are unused as pageblock_order is set
4345 * at compile-time. See include/linux/pageblock-flags.h for the values of
4346 * pageblock_order based on the kernel config
4348 static inline int pageblock_default_order(unsigned int order)
4352 #define set_pageblock_order(x) do {} while (0)
4354 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4357 * Set up the zone data structures:
4358 * - mark all pages reserved
4359 * - mark all memory queues empty
4360 * - clear the memory bitmaps
4362 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4363 unsigned long *zones_size, unsigned long *zholes_size)
4366 int nid = pgdat->node_id;
4367 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4370 pgdat_resize_init(pgdat);
4371 pgdat->nr_zones = 0;
4372 init_waitqueue_head(&pgdat->kswapd_wait);
4373 pgdat->kswapd_max_order = 0;
4374 pgdat_page_cgroup_init(pgdat);
4376 for (j = 0; j < MAX_NR_ZONES; j++) {
4377 struct zone *zone = pgdat->node_zones + j;
4378 unsigned long size, realsize, memmap_pages;
4381 size = zone_spanned_pages_in_node(nid, j, zones_size);
4382 realsize = size - zone_absent_pages_in_node(nid, j,
4386 * Adjust realsize so that it accounts for how much memory
4387 * is used by this zone for memmap. This affects the watermark
4388 * and per-cpu initialisations
4391 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4392 if (realsize >= memmap_pages) {
4393 realsize -= memmap_pages;
4396 " %s zone: %lu pages used for memmap\n",
4397 zone_names[j], memmap_pages);
4400 " %s zone: %lu pages exceeds realsize %lu\n",
4401 zone_names[j], memmap_pages, realsize);
4403 /* Account for reserved pages */
4404 if (j == 0 && realsize > dma_reserve) {
4405 realsize -= dma_reserve;
4406 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4407 zone_names[0], dma_reserve);
4410 if (!is_highmem_idx(j))
4411 nr_kernel_pages += realsize;
4412 nr_all_pages += realsize;
4414 zone->spanned_pages = size;
4415 zone->present_pages = realsize;
4418 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4420 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4422 zone->name = zone_names[j];
4423 spin_lock_init(&zone->lock);
4424 spin_lock_init(&zone->lru_lock);
4425 zone_seqlock_init(zone);
4426 zone->zone_pgdat = pgdat;
4428 zone_pcp_init(zone);
4430 INIT_LIST_HEAD(&zone->lru[l].list);
4431 zone->reclaim_stat.recent_rotated[0] = 0;
4432 zone->reclaim_stat.recent_rotated[1] = 0;
4433 zone->reclaim_stat.recent_scanned[0] = 0;
4434 zone->reclaim_stat.recent_scanned[1] = 0;
4435 zap_zone_vm_stats(zone);
4440 set_pageblock_order(pageblock_default_order());
4441 setup_usemap(pgdat, zone, size);
4442 ret = init_currently_empty_zone(zone, zone_start_pfn,
4443 size, MEMMAP_EARLY);
4445 memmap_init(size, nid, j, zone_start_pfn);
4446 zone_start_pfn += size;
4450 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4452 /* Skip empty nodes */
4453 if (!pgdat->node_spanned_pages)
4456 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4457 /* ia64 gets its own node_mem_map, before this, without bootmem */
4458 if (!pgdat->node_mem_map) {
4459 unsigned long size, start, end;
4463 * The zone's endpoints aren't required to be MAX_ORDER
4464 * aligned but the node_mem_map endpoints must be in order
4465 * for the buddy allocator to function correctly.
4467 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4468 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4469 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4470 size = (end - start) * sizeof(struct page);
4471 map = alloc_remap(pgdat->node_id, size);
4473 map = alloc_bootmem_node_nopanic(pgdat, size);
4474 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4476 #ifndef CONFIG_NEED_MULTIPLE_NODES
4478 * With no DISCONTIG, the global mem_map is just set as node 0's
4480 if (pgdat == NODE_DATA(0)) {
4481 mem_map = NODE_DATA(0)->node_mem_map;
4482 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4483 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4484 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4485 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4488 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4491 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4492 unsigned long node_start_pfn, unsigned long *zholes_size)
4494 pg_data_t *pgdat = NODE_DATA(nid);
4496 pgdat->node_id = nid;
4497 pgdat->node_start_pfn = node_start_pfn;
4498 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4500 alloc_node_mem_map(pgdat);
4501 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4502 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4503 nid, (unsigned long)pgdat,
4504 (unsigned long)pgdat->node_mem_map);
4507 free_area_init_core(pgdat, zones_size, zholes_size);
4510 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4512 #if MAX_NUMNODES > 1
4514 * Figure out the number of possible node ids.
4516 static void __init setup_nr_node_ids(void)
4519 unsigned int highest = 0;
4521 for_each_node_mask(node, node_possible_map)
4523 nr_node_ids = highest + 1;
4526 static inline void setup_nr_node_ids(void)
4532 * add_active_range - Register a range of PFNs backed by physical memory
4533 * @nid: The node ID the range resides on
4534 * @start_pfn: The start PFN of the available physical memory
4535 * @end_pfn: The end PFN of the available physical memory
4537 * These ranges are stored in an early_node_map[] and later used by
4538 * free_area_init_nodes() to calculate zone sizes and holes. If the
4539 * range spans a memory hole, it is up to the architecture to ensure
4540 * the memory is not freed by the bootmem allocator. If possible
4541 * the range being registered will be merged with existing ranges.
4543 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4544 unsigned long end_pfn)
4548 mminit_dprintk(MMINIT_TRACE, "memory_register",
4549 "Entering add_active_range(%d, %#lx, %#lx) "
4550 "%d entries of %d used\n",
4551 nid, start_pfn, end_pfn,
4552 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4554 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4556 /* Merge with existing active regions if possible */
4557 for (i = 0; i < nr_nodemap_entries; i++) {
4558 if (early_node_map[i].nid != nid)
4561 /* Skip if an existing region covers this new one */
4562 if (start_pfn >= early_node_map[i].start_pfn &&
4563 end_pfn <= early_node_map[i].end_pfn)
4566 /* Merge forward if suitable */
4567 if (start_pfn <= early_node_map[i].end_pfn &&
4568 end_pfn > early_node_map[i].end_pfn) {
4569 early_node_map[i].end_pfn = end_pfn;
4573 /* Merge backward if suitable */
4574 if (start_pfn < early_node_map[i].start_pfn &&
4575 end_pfn >= early_node_map[i].start_pfn) {
4576 early_node_map[i].start_pfn = start_pfn;
4581 /* Check that early_node_map is large enough */
4582 if (i >= MAX_ACTIVE_REGIONS) {
4583 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4584 MAX_ACTIVE_REGIONS);
4588 early_node_map[i].nid = nid;
4589 early_node_map[i].start_pfn = start_pfn;
4590 early_node_map[i].end_pfn = end_pfn;
4591 nr_nodemap_entries = i + 1;
4595 * remove_active_range - Shrink an existing registered range of PFNs
4596 * @nid: The node id the range is on that should be shrunk
4597 * @start_pfn: The new PFN of the range
4598 * @end_pfn: The new PFN of the range
4600 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4601 * The map is kept near the end physical page range that has already been
4602 * registered. This function allows an arch to shrink an existing registered
4605 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4606 unsigned long end_pfn)
4611 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4612 nid, start_pfn, end_pfn);
4614 /* Find the old active region end and shrink */
4615 for_each_active_range_index_in_nid(i, nid) {
4616 if (early_node_map[i].start_pfn >= start_pfn &&
4617 early_node_map[i].end_pfn <= end_pfn) {
4619 early_node_map[i].start_pfn = 0;
4620 early_node_map[i].end_pfn = 0;
4624 if (early_node_map[i].start_pfn < start_pfn &&
4625 early_node_map[i].end_pfn > start_pfn) {
4626 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4627 early_node_map[i].end_pfn = start_pfn;
4628 if (temp_end_pfn > end_pfn)
4629 add_active_range(nid, end_pfn, temp_end_pfn);
4632 if (early_node_map[i].start_pfn >= start_pfn &&
4633 early_node_map[i].end_pfn > end_pfn &&
4634 early_node_map[i].start_pfn < end_pfn) {
4635 early_node_map[i].start_pfn = end_pfn;
4643 /* remove the blank ones */
4644 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4645 if (early_node_map[i].nid != nid)
4647 if (early_node_map[i].end_pfn)
4649 /* we found it, get rid of it */
4650 for (j = i; j < nr_nodemap_entries - 1; j++)
4651 memcpy(&early_node_map[j], &early_node_map[j+1],
4652 sizeof(early_node_map[j]));
4653 j = nr_nodemap_entries - 1;
4654 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4655 nr_nodemap_entries--;
4660 * remove_all_active_ranges - Remove all currently registered regions
4662 * During discovery, it may be found that a table like SRAT is invalid
4663 * and an alternative discovery method must be used. This function removes
4664 * all currently registered regions.
4666 void __init remove_all_active_ranges(void)
4668 memset(early_node_map, 0, sizeof(early_node_map));
4669 nr_nodemap_entries = 0;
4672 /* Compare two active node_active_regions */
4673 static int __init cmp_node_active_region(const void *a, const void *b)
4675 struct node_active_region *arange = (struct node_active_region *)a;
4676 struct node_active_region *brange = (struct node_active_region *)b;
4678 /* Done this way to avoid overflows */
4679 if (arange->start_pfn > brange->start_pfn)
4681 if (arange->start_pfn < brange->start_pfn)
4687 /* sort the node_map by start_pfn */
4688 void __init sort_node_map(void)
4690 sort(early_node_map, (size_t)nr_nodemap_entries,
4691 sizeof(struct node_active_region),
4692 cmp_node_active_region, NULL);
4695 /* Find the lowest pfn for a node */
4696 static unsigned long __init find_min_pfn_for_node(int nid)
4699 unsigned long min_pfn = ULONG_MAX;
4701 /* Assuming a sorted map, the first range found has the starting pfn */
4702 for_each_active_range_index_in_nid(i, nid)
4703 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4705 if (min_pfn == ULONG_MAX) {
4707 "Could not find start_pfn for node %d\n", nid);
4715 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4717 * It returns the minimum PFN based on information provided via
4718 * add_active_range().
4720 unsigned long __init find_min_pfn_with_active_regions(void)
4722 return find_min_pfn_for_node(MAX_NUMNODES);
4726 * early_calculate_totalpages()
4727 * Sum pages in active regions for movable zone.
4728 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4730 static unsigned long __init early_calculate_totalpages(void)
4733 unsigned long totalpages = 0;
4735 for (i = 0; i < nr_nodemap_entries; i++) {
4736 unsigned long pages = early_node_map[i].end_pfn -
4737 early_node_map[i].start_pfn;
4738 totalpages += pages;
4740 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4746 * Find the PFN the Movable zone begins in each node. Kernel memory
4747 * is spread evenly between nodes as long as the nodes have enough
4748 * memory. When they don't, some nodes will have more kernelcore than
4751 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4754 unsigned long usable_startpfn;
4755 unsigned long kernelcore_node, kernelcore_remaining;
4756 /* save the state before borrow the nodemask */
4757 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4758 unsigned long totalpages = early_calculate_totalpages();
4759 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4762 * If movablecore was specified, calculate what size of
4763 * kernelcore that corresponds so that memory usable for
4764 * any allocation type is evenly spread. If both kernelcore
4765 * and movablecore are specified, then the value of kernelcore
4766 * will be used for required_kernelcore if it's greater than
4767 * what movablecore would have allowed.
4769 if (required_movablecore) {
4770 unsigned long corepages;
4773 * Round-up so that ZONE_MOVABLE is at least as large as what
4774 * was requested by the user
4776 required_movablecore =
4777 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4778 corepages = totalpages - required_movablecore;
4780 required_kernelcore = max(required_kernelcore, corepages);
4783 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4784 if (!required_kernelcore)
4787 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4788 find_usable_zone_for_movable();
4789 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4792 /* Spread kernelcore memory as evenly as possible throughout nodes */
4793 kernelcore_node = required_kernelcore / usable_nodes;
4794 for_each_node_state(nid, N_HIGH_MEMORY) {
4796 * Recalculate kernelcore_node if the division per node
4797 * now exceeds what is necessary to satisfy the requested
4798 * amount of memory for the kernel
4800 if (required_kernelcore < kernelcore_node)
4801 kernelcore_node = required_kernelcore / usable_nodes;
4804 * As the map is walked, we track how much memory is usable
4805 * by the kernel using kernelcore_remaining. When it is
4806 * 0, the rest of the node is usable by ZONE_MOVABLE
4808 kernelcore_remaining = kernelcore_node;
4810 /* Go through each range of PFNs within this node */
4811 for_each_active_range_index_in_nid(i, nid) {
4812 unsigned long start_pfn, end_pfn;
4813 unsigned long size_pages;
4815 start_pfn = max(early_node_map[i].start_pfn,
4816 zone_movable_pfn[nid]);
4817 end_pfn = early_node_map[i].end_pfn;
4818 if (start_pfn >= end_pfn)
4821 /* Account for what is only usable for kernelcore */
4822 if (start_pfn < usable_startpfn) {
4823 unsigned long kernel_pages;
4824 kernel_pages = min(end_pfn, usable_startpfn)
4827 kernelcore_remaining -= min(kernel_pages,
4828 kernelcore_remaining);
4829 required_kernelcore -= min(kernel_pages,
4830 required_kernelcore);
4832 /* Continue if range is now fully accounted */
4833 if (end_pfn <= usable_startpfn) {
4836 * Push zone_movable_pfn to the end so
4837 * that if we have to rebalance
4838 * kernelcore across nodes, we will
4839 * not double account here
4841 zone_movable_pfn[nid] = end_pfn;
4844 start_pfn = usable_startpfn;
4848 * The usable PFN range for ZONE_MOVABLE is from
4849 * start_pfn->end_pfn. Calculate size_pages as the
4850 * number of pages used as kernelcore
4852 size_pages = end_pfn - start_pfn;
4853 if (size_pages > kernelcore_remaining)
4854 size_pages = kernelcore_remaining;
4855 zone_movable_pfn[nid] = start_pfn + size_pages;
4858 * Some kernelcore has been met, update counts and
4859 * break if the kernelcore for this node has been
4862 required_kernelcore -= min(required_kernelcore,
4864 kernelcore_remaining -= size_pages;
4865 if (!kernelcore_remaining)
4871 * If there is still required_kernelcore, we do another pass with one
4872 * less node in the count. This will push zone_movable_pfn[nid] further
4873 * along on the nodes that still have memory until kernelcore is
4877 if (usable_nodes && required_kernelcore > usable_nodes)
4880 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4881 for (nid = 0; nid < MAX_NUMNODES; nid++)
4882 zone_movable_pfn[nid] =
4883 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4886 /* restore the node_state */
4887 node_states[N_HIGH_MEMORY] = saved_node_state;
4890 /* Any regular memory on that node ? */
4891 static void check_for_regular_memory(pg_data_t *pgdat)
4893 #ifdef CONFIG_HIGHMEM
4894 enum zone_type zone_type;
4896 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4897 struct zone *zone = &pgdat->node_zones[zone_type];
4898 if (zone->present_pages)
4899 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4905 * free_area_init_nodes - Initialise all pg_data_t and zone data
4906 * @max_zone_pfn: an array of max PFNs for each zone
4908 * This will call free_area_init_node() for each active node in the system.
4909 * Using the page ranges provided by add_active_range(), the size of each
4910 * zone in each node and their holes is calculated. If the maximum PFN
4911 * between two adjacent zones match, it is assumed that the zone is empty.
4912 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4913 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4914 * starts where the previous one ended. For example, ZONE_DMA32 starts
4915 * at arch_max_dma_pfn.
4917 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4922 /* Sort early_node_map as initialisation assumes it is sorted */
4925 /* Record where the zone boundaries are */
4926 memset(arch_zone_lowest_possible_pfn, 0,
4927 sizeof(arch_zone_lowest_possible_pfn));
4928 memset(arch_zone_highest_possible_pfn, 0,
4929 sizeof(arch_zone_highest_possible_pfn));
4930 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4931 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4932 for (i = 1; i < MAX_NR_ZONES; i++) {
4933 if (i == ZONE_MOVABLE)
4935 arch_zone_lowest_possible_pfn[i] =
4936 arch_zone_highest_possible_pfn[i-1];
4937 arch_zone_highest_possible_pfn[i] =
4938 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4940 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4941 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4943 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4944 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4945 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4947 /* Print out the zone ranges */
4948 printk("Zone PFN ranges:\n");
4949 for (i = 0; i < MAX_NR_ZONES; i++) {
4950 if (i == ZONE_MOVABLE)
4952 printk(" %-8s ", zone_names[i]);
4953 if (arch_zone_lowest_possible_pfn[i] ==
4954 arch_zone_highest_possible_pfn[i])
4957 printk("%0#10lx -> %0#10lx\n",
4958 arch_zone_lowest_possible_pfn[i],
4959 arch_zone_highest_possible_pfn[i]);
4962 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4963 printk("Movable zone start PFN for each node\n");
4964 for (i = 0; i < MAX_NUMNODES; i++) {
4965 if (zone_movable_pfn[i])
4966 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4969 /* Print out the early_node_map[] */
4970 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4971 for (i = 0; i < nr_nodemap_entries; i++)
4972 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4973 early_node_map[i].start_pfn,
4974 early_node_map[i].end_pfn);
4976 /* Initialise every node */
4977 mminit_verify_pageflags_layout();
4978 setup_nr_node_ids();
4979 for_each_online_node(nid) {
4980 pg_data_t *pgdat = NODE_DATA(nid);
4981 free_area_init_node(nid, NULL,
4982 find_min_pfn_for_node(nid), NULL);
4984 /* Any memory on that node */
4985 if (pgdat->node_present_pages)
4986 node_set_state(nid, N_HIGH_MEMORY);
4987 check_for_regular_memory(pgdat);
4991 static int __init cmdline_parse_core(char *p, unsigned long *core)
4993 unsigned long long coremem;
4997 coremem = memparse(p, &p);
4998 *core = coremem >> PAGE_SHIFT;
5000 /* Paranoid check that UL is enough for the coremem value */
5001 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5007 * kernelcore=size sets the amount of memory for use for allocations that
5008 * cannot be reclaimed or migrated.
5010 static int __init cmdline_parse_kernelcore(char *p)
5012 return cmdline_parse_core(p, &required_kernelcore);
5016 * movablecore=size sets the amount of memory for use for allocations that
5017 * can be reclaimed or migrated.
5019 static int __init cmdline_parse_movablecore(char *p)
5021 return cmdline_parse_core(p, &required_movablecore);
5024 early_param("kernelcore", cmdline_parse_kernelcore);
5025 early_param("movablecore", cmdline_parse_movablecore);
5027 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
5030 * set_dma_reserve - set the specified number of pages reserved in the first zone
5031 * @new_dma_reserve: The number of pages to mark reserved
5033 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5034 * In the DMA zone, a significant percentage may be consumed by kernel image
5035 * and other unfreeable allocations which can skew the watermarks badly. This
5036 * function may optionally be used to account for unfreeable pages in the
5037 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5038 * smaller per-cpu batchsize.
5040 void __init set_dma_reserve(unsigned long new_dma_reserve)
5042 dma_reserve = new_dma_reserve;
5045 void __init free_area_init(unsigned long *zones_size)
5047 free_area_init_node(0, zones_size,
5048 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5051 static int page_alloc_cpu_notify(struct notifier_block *self,
5052 unsigned long action, void *hcpu)
5054 int cpu = (unsigned long)hcpu;
5056 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5060 * Spill the event counters of the dead processor
5061 * into the current processors event counters.
5062 * This artificially elevates the count of the current
5065 vm_events_fold_cpu(cpu);
5068 * Zero the differential counters of the dead processor
5069 * so that the vm statistics are consistent.
5071 * This is only okay since the processor is dead and cannot
5072 * race with what we are doing.
5074 refresh_cpu_vm_stats(cpu);
5079 void __init page_alloc_init(void)
5081 hotcpu_notifier(page_alloc_cpu_notify, 0);
5085 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5086 * or min_free_kbytes changes.
5088 static void calculate_totalreserve_pages(void)
5090 struct pglist_data *pgdat;
5091 unsigned long reserve_pages = 0;
5092 enum zone_type i, j;
5094 for_each_online_pgdat(pgdat) {
5095 for (i = 0; i < MAX_NR_ZONES; i++) {
5096 struct zone *zone = pgdat->node_zones + i;
5097 unsigned long max = 0;
5099 /* Find valid and maximum lowmem_reserve in the zone */
5100 for (j = i; j < MAX_NR_ZONES; j++) {
5101 if (zone->lowmem_reserve[j] > max)
5102 max = zone->lowmem_reserve[j];
5105 /* we treat the high watermark as reserved pages. */
5106 max += high_wmark_pages(zone);
5108 if (max > zone->present_pages)
5109 max = zone->present_pages;
5110 reserve_pages += max;
5113 totalreserve_pages = reserve_pages;
5117 * setup_per_zone_lowmem_reserve - called whenever
5118 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5119 * has a correct pages reserved value, so an adequate number of
5120 * pages are left in the zone after a successful __alloc_pages().
5122 static void setup_per_zone_lowmem_reserve(void)
5124 struct pglist_data *pgdat;
5125 enum zone_type j, idx;
5127 for_each_online_pgdat(pgdat) {
5128 for (j = 0; j < MAX_NR_ZONES; j++) {
5129 struct zone *zone = pgdat->node_zones + j;
5130 unsigned long present_pages = zone->present_pages;
5132 zone->lowmem_reserve[j] = 0;
5136 struct zone *lower_zone;
5140 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5141 sysctl_lowmem_reserve_ratio[idx] = 1;
5143 lower_zone = pgdat->node_zones + idx;
5144 lower_zone->lowmem_reserve[j] = present_pages /
5145 sysctl_lowmem_reserve_ratio[idx];
5146 present_pages += lower_zone->present_pages;
5151 /* update totalreserve_pages */
5152 calculate_totalreserve_pages();
5156 * setup_per_zone_wmarks - called when min_free_kbytes changes
5157 * or when memory is hot-{added|removed}
5159 * Ensures that the watermark[min,low,high] values for each zone are set
5160 * correctly with respect to min_free_kbytes.
5162 void setup_per_zone_wmarks(void)
5164 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5165 unsigned long lowmem_pages = 0;
5167 unsigned long flags;
5169 /* Calculate total number of !ZONE_HIGHMEM pages */
5170 for_each_zone(zone) {
5171 if (!is_highmem(zone))
5172 lowmem_pages += zone->present_pages;
5175 for_each_zone(zone) {
5178 spin_lock_irqsave(&zone->lock, flags);
5179 tmp = (u64)pages_min * zone->present_pages;
5180 do_div(tmp, lowmem_pages);
5181 if (is_highmem(zone)) {
5183 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5184 * need highmem pages, so cap pages_min to a small
5187 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5188 * deltas controls asynch page reclaim, and so should
5189 * not be capped for highmem.
5193 min_pages = zone->present_pages / 1024;
5194 if (min_pages < SWAP_CLUSTER_MAX)
5195 min_pages = SWAP_CLUSTER_MAX;
5196 if (min_pages > 128)
5198 zone->watermark[WMARK_MIN] = min_pages;
5201 * If it's a lowmem zone, reserve a number of pages
5202 * proportionate to the zone's size.
5204 zone->watermark[WMARK_MIN] = tmp;
5207 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5208 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5209 setup_zone_migrate_reserve(zone);
5210 spin_unlock_irqrestore(&zone->lock, flags);
5213 /* update totalreserve_pages */
5214 calculate_totalreserve_pages();
5218 * The inactive anon list should be small enough that the VM never has to
5219 * do too much work, but large enough that each inactive page has a chance
5220 * to be referenced again before it is swapped out.
5222 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5223 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5224 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5225 * the anonymous pages are kept on the inactive list.
5228 * memory ratio inactive anon
5229 * -------------------------------------
5238 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5240 unsigned int gb, ratio;
5242 /* Zone size in gigabytes */
5243 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5245 ratio = int_sqrt(10 * gb);
5249 zone->inactive_ratio = ratio;
5252 static void __meminit setup_per_zone_inactive_ratio(void)
5257 calculate_zone_inactive_ratio(zone);
5261 * Initialise min_free_kbytes.
5263 * For small machines we want it small (128k min). For large machines
5264 * we want it large (64MB max). But it is not linear, because network
5265 * bandwidth does not increase linearly with machine size. We use
5267 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5268 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5284 int __meminit init_per_zone_wmark_min(void)
5286 unsigned long lowmem_kbytes;
5288 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5290 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5291 if (min_free_kbytes < 128)
5292 min_free_kbytes = 128;
5293 if (min_free_kbytes > 65536)
5294 min_free_kbytes = 65536;
5295 setup_per_zone_wmarks();
5296 refresh_zone_stat_thresholds();
5297 setup_per_zone_lowmem_reserve();
5298 setup_per_zone_inactive_ratio();
5301 module_init(init_per_zone_wmark_min)
5304 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5305 * that we can call two helper functions whenever min_free_kbytes
5308 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5309 void __user *buffer, size_t *length, loff_t *ppos)
5311 proc_dointvec(table, write, buffer, length, ppos);
5313 setup_per_zone_wmarks();
5318 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5319 void __user *buffer, size_t *length, loff_t *ppos)
5324 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5329 zone->min_unmapped_pages = (zone->present_pages *
5330 sysctl_min_unmapped_ratio) / 100;
5334 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5335 void __user *buffer, size_t *length, loff_t *ppos)
5340 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5345 zone->min_slab_pages = (zone->present_pages *
5346 sysctl_min_slab_ratio) / 100;
5352 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5353 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5354 * whenever sysctl_lowmem_reserve_ratio changes.
5356 * The reserve ratio obviously has absolutely no relation with the
5357 * minimum watermarks. The lowmem reserve ratio can only make sense
5358 * if in function of the boot time zone sizes.
5360 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5361 void __user *buffer, size_t *length, loff_t *ppos)
5363 proc_dointvec_minmax(table, write, buffer, length, ppos);
5364 setup_per_zone_lowmem_reserve();
5369 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5370 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5371 * can have before it gets flushed back to buddy allocator.
5374 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5375 void __user *buffer, size_t *length, loff_t *ppos)
5381 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5382 if (!write || (ret == -EINVAL))
5384 for_each_populated_zone(zone) {
5385 for_each_possible_cpu(cpu) {
5387 high = zone->present_pages / percpu_pagelist_fraction;
5388 setup_pagelist_highmark(
5389 per_cpu_ptr(zone->pageset, cpu), high);
5395 int hashdist = HASHDIST_DEFAULT;
5398 static int __init set_hashdist(char *str)
5402 hashdist = simple_strtoul(str, &str, 0);
5405 __setup("hashdist=", set_hashdist);
5409 * allocate a large system hash table from bootmem
5410 * - it is assumed that the hash table must contain an exact power-of-2
5411 * quantity of entries
5412 * - limit is the number of hash buckets, not the total allocation size
5414 void *__init alloc_large_system_hash(const char *tablename,
5415 unsigned long bucketsize,
5416 unsigned long numentries,
5419 unsigned int *_hash_shift,
5420 unsigned int *_hash_mask,
5421 unsigned long limit)
5423 unsigned long long max = limit;
5424 unsigned long log2qty, size;
5427 /* allow the kernel cmdline to have a say */
5429 /* round applicable memory size up to nearest megabyte */
5430 numentries = nr_kernel_pages;
5431 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5432 numentries >>= 20 - PAGE_SHIFT;
5433 numentries <<= 20 - PAGE_SHIFT;
5435 /* limit to 1 bucket per 2^scale bytes of low memory */
5436 if (scale > PAGE_SHIFT)
5437 numentries >>= (scale - PAGE_SHIFT);
5439 numentries <<= (PAGE_SHIFT - scale);
5441 /* Make sure we've got at least a 0-order allocation.. */
5442 if (unlikely(flags & HASH_SMALL)) {
5443 /* Makes no sense without HASH_EARLY */
5444 WARN_ON(!(flags & HASH_EARLY));
5445 if (!(numentries >> *_hash_shift)) {
5446 numentries = 1UL << *_hash_shift;
5447 BUG_ON(!numentries);
5449 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5450 numentries = PAGE_SIZE / bucketsize;
5452 numentries = roundup_pow_of_two(numentries);
5454 /* limit allocation size to 1/16 total memory by default */
5456 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5457 do_div(max, bucketsize);
5460 if (numentries > max)
5463 log2qty = ilog2(numentries);
5466 size = bucketsize << log2qty;
5467 if (flags & HASH_EARLY)
5468 table = alloc_bootmem_nopanic(size);
5470 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5473 * If bucketsize is not a power-of-two, we may free
5474 * some pages at the end of hash table which
5475 * alloc_pages_exact() automatically does
5477 if (get_order(size) < MAX_ORDER) {
5478 table = alloc_pages_exact(size, GFP_ATOMIC);
5479 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5482 } while (!table && size > PAGE_SIZE && --log2qty);
5485 panic("Failed to allocate %s hash table\n", tablename);
5487 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5490 ilog2(size) - PAGE_SHIFT,
5494 *_hash_shift = log2qty;
5496 *_hash_mask = (1 << log2qty) - 1;
5501 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5502 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5505 #ifdef CONFIG_SPARSEMEM
5506 return __pfn_to_section(pfn)->pageblock_flags;
5508 return zone->pageblock_flags;
5509 #endif /* CONFIG_SPARSEMEM */
5512 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5514 #ifdef CONFIG_SPARSEMEM
5515 pfn &= (PAGES_PER_SECTION-1);
5516 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5518 pfn = pfn - zone->zone_start_pfn;
5519 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5520 #endif /* CONFIG_SPARSEMEM */
5524 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5525 * @page: The page within the block of interest
5526 * @start_bitidx: The first bit of interest to retrieve
5527 * @end_bitidx: The last bit of interest
5528 * returns pageblock_bits flags
5530 unsigned long get_pageblock_flags_group(struct page *page,
5531 int start_bitidx, int end_bitidx)
5534 unsigned long *bitmap;
5535 unsigned long pfn, bitidx;
5536 unsigned long flags = 0;
5537 unsigned long value = 1;
5539 zone = page_zone(page);
5540 pfn = page_to_pfn(page);
5541 bitmap = get_pageblock_bitmap(zone, pfn);
5542 bitidx = pfn_to_bitidx(zone, pfn);
5544 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5545 if (test_bit(bitidx + start_bitidx, bitmap))
5552 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5553 * @page: The page within the block of interest
5554 * @start_bitidx: The first bit of interest
5555 * @end_bitidx: The last bit of interest
5556 * @flags: The flags to set
5558 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5559 int start_bitidx, int end_bitidx)
5562 unsigned long *bitmap;
5563 unsigned long pfn, bitidx;
5564 unsigned long value = 1;
5566 zone = page_zone(page);
5567 pfn = page_to_pfn(page);
5568 bitmap = get_pageblock_bitmap(zone, pfn);
5569 bitidx = pfn_to_bitidx(zone, pfn);
5570 VM_BUG_ON(pfn < zone->zone_start_pfn);
5571 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5573 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5575 __set_bit(bitidx + start_bitidx, bitmap);
5577 __clear_bit(bitidx + start_bitidx, bitmap);
5581 * This is designed as sub function...plz see page_isolation.c also.
5582 * set/clear page block's type to be ISOLATE.
5583 * page allocater never alloc memory from ISOLATE block.
5587 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5589 unsigned long pfn, iter, found;
5591 * For avoiding noise data, lru_add_drain_all() should be called
5592 * If ZONE_MOVABLE, the zone never contains immobile pages
5594 if (zone_idx(zone) == ZONE_MOVABLE)
5597 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5600 pfn = page_to_pfn(page);
5601 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5602 unsigned long check = pfn + iter;
5604 if (!pfn_valid_within(check))
5607 page = pfn_to_page(check);
5608 if (!page_count(page)) {
5609 if (PageBuddy(page))
5610 iter += (1 << page_order(page)) - 1;
5616 * If there are RECLAIMABLE pages, we need to check it.
5617 * But now, memory offline itself doesn't call shrink_slab()
5618 * and it still to be fixed.
5621 * If the page is not RAM, page_count()should be 0.
5622 * we don't need more check. This is an _used_ not-movable page.
5624 * The problematic thing here is PG_reserved pages. PG_reserved
5625 * is set to both of a memory hole page and a _used_ kernel
5634 bool is_pageblock_removable_nolock(struct page *page)
5636 struct zone *zone = page_zone(page);
5637 unsigned long pfn = page_to_pfn(page);
5640 * We have to be careful here because we are iterating over memory
5641 * sections which are not zone aware so we might end up outside of
5642 * the zone but still within the section.
5644 if (!zone || zone->zone_start_pfn > pfn ||
5645 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5648 return __count_immobile_pages(zone, page, 0);
5651 int set_migratetype_isolate(struct page *page)
5654 unsigned long flags, pfn;
5655 struct memory_isolate_notify arg;
5659 zone = page_zone(page);
5661 spin_lock_irqsave(&zone->lock, flags);
5663 pfn = page_to_pfn(page);
5664 arg.start_pfn = pfn;
5665 arg.nr_pages = pageblock_nr_pages;
5666 arg.pages_found = 0;
5669 * It may be possible to isolate a pageblock even if the
5670 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5671 * notifier chain is used by balloon drivers to return the
5672 * number of pages in a range that are held by the balloon
5673 * driver to shrink memory. If all the pages are accounted for
5674 * by balloons, are free, or on the LRU, isolation can continue.
5675 * Later, for example, when memory hotplug notifier runs, these
5676 * pages reported as "can be isolated" should be isolated(freed)
5677 * by the balloon driver through the memory notifier chain.
5679 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5680 notifier_ret = notifier_to_errno(notifier_ret);
5684 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5685 * We just check MOVABLE pages.
5687 if (__count_immobile_pages(zone, page, arg.pages_found))
5691 * immobile means "not-on-lru" paes. If immobile is larger than
5692 * removable-by-driver pages reported by notifier, we'll fail.
5697 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5698 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5701 spin_unlock_irqrestore(&zone->lock, flags);
5707 void unset_migratetype_isolate(struct page *page)
5710 unsigned long flags;
5711 zone = page_zone(page);
5712 spin_lock_irqsave(&zone->lock, flags);
5713 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5715 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5716 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5718 spin_unlock_irqrestore(&zone->lock, flags);
5721 #ifdef CONFIG_MEMORY_HOTREMOVE
5723 * All pages in the range must be isolated before calling this.
5726 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5732 unsigned long flags;
5733 /* find the first valid pfn */
5734 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5739 zone = page_zone(pfn_to_page(pfn));
5740 spin_lock_irqsave(&zone->lock, flags);
5742 while (pfn < end_pfn) {
5743 if (!pfn_valid(pfn)) {
5747 page = pfn_to_page(pfn);
5748 BUG_ON(page_count(page));
5749 BUG_ON(!PageBuddy(page));
5750 order = page_order(page);
5751 #ifdef CONFIG_DEBUG_VM
5752 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5753 pfn, 1 << order, end_pfn);
5755 list_del(&page->lru);
5756 rmv_page_order(page);
5757 zone->free_area[order].nr_free--;
5758 __mod_zone_page_state(zone, NR_FREE_PAGES,
5760 for (i = 0; i < (1 << order); i++)
5761 SetPageReserved((page+i));
5762 pfn += (1 << order);
5764 spin_unlock_irqrestore(&zone->lock, flags);
5768 #ifdef CONFIG_MEMORY_FAILURE
5769 bool is_free_buddy_page(struct page *page)
5771 struct zone *zone = page_zone(page);
5772 unsigned long pfn = page_to_pfn(page);
5773 unsigned long flags;
5776 spin_lock_irqsave(&zone->lock, flags);
5777 for (order = 0; order < MAX_ORDER; order++) {
5778 struct page *page_head = page - (pfn & ((1 << order) - 1));
5780 if (PageBuddy(page_head) && page_order(page_head) >= order)
5783 spin_unlock_irqrestore(&zone->lock, flags);
5785 return order < MAX_ORDER;
5789 static struct trace_print_flags pageflag_names[] = {
5790 {1UL << PG_locked, "locked" },
5791 {1UL << PG_error, "error" },
5792 {1UL << PG_referenced, "referenced" },
5793 {1UL << PG_uptodate, "uptodate" },
5794 {1UL << PG_dirty, "dirty" },
5795 {1UL << PG_lru, "lru" },
5796 {1UL << PG_active, "active" },
5797 {1UL << PG_slab, "slab" },
5798 {1UL << PG_owner_priv_1, "owner_priv_1" },
5799 {1UL << PG_arch_1, "arch_1" },
5800 {1UL << PG_reserved, "reserved" },
5801 {1UL << PG_private, "private" },
5802 {1UL << PG_private_2, "private_2" },
5803 {1UL << PG_writeback, "writeback" },
5804 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5805 {1UL << PG_head, "head" },
5806 {1UL << PG_tail, "tail" },
5808 {1UL << PG_compound, "compound" },
5810 {1UL << PG_swapcache, "swapcache" },
5811 {1UL << PG_mappedtodisk, "mappedtodisk" },
5812 {1UL << PG_reclaim, "reclaim" },
5813 {1UL << PG_swapbacked, "swapbacked" },
5814 {1UL << PG_unevictable, "unevictable" },
5816 {1UL << PG_mlocked, "mlocked" },
5818 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5819 {1UL << PG_uncached, "uncached" },
5821 #ifdef CONFIG_MEMORY_FAILURE
5822 {1UL << PG_hwpoison, "hwpoison" },
5827 static void dump_page_flags(unsigned long flags)
5829 const char *delim = "";
5833 printk(KERN_ALERT "page flags: %#lx(", flags);
5835 /* remove zone id */
5836 flags &= (1UL << NR_PAGEFLAGS) - 1;
5838 for (i = 0; pageflag_names[i].name && flags; i++) {
5840 mask = pageflag_names[i].mask;
5841 if ((flags & mask) != mask)
5845 printk("%s%s", delim, pageflag_names[i].name);
5849 /* check for left over flags */
5851 printk("%s%#lx", delim, flags);
5856 void dump_page(struct page *page)
5859 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5860 page, atomic_read(&page->_count), page_mapcount(page),
5861 page->mapping, page->index);
5862 dump_page_flags(page->flags);
5863 mem_cgroup_print_bad_page(page);