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/oom.h>
34 #include <linux/notifier.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/mempolicy.h>
43 #include <linux/stop_machine.h>
44 #include <linux/sort.h>
45 #include <linux/pfn.h>
46 #include <linux/backing-dev.h>
47 #include <linux/fault-inject.h>
48 #include <linux/page-isolation.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/debugobjects.h>
51 #include <linux/kmemleak.h>
52 #include <linux/memory.h>
53 #include <linux/compaction.h>
54 #include <trace/events/kmem.h>
55 #include <linux/ftrace_event.h>
57 #include <asm/tlbflush.h>
58 #include <asm/div64.h>
61 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
62 DEFINE_PER_CPU(int, numa_node);
63 EXPORT_PER_CPU_SYMBOL(numa_node);
66 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
68 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
69 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
70 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
71 * defined in <linux/topology.h>.
73 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
74 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
78 * Array of node states.
80 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
81 [N_POSSIBLE] = NODE_MASK_ALL,
82 [N_ONLINE] = { { [0] = 1UL } },
84 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
86 [N_HIGH_MEMORY] = { { [0] = 1UL } },
88 [N_CPU] = { { [0] = 1UL } },
91 EXPORT_SYMBOL(node_states);
93 unsigned long totalram_pages __read_mostly;
94 unsigned long totalreserve_pages __read_mostly;
95 int percpu_pagelist_fraction;
96 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
98 #ifdef CONFIG_PM_SLEEP
100 * The following functions are used by the suspend/hibernate code to temporarily
101 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
102 * while devices are suspended. To avoid races with the suspend/hibernate code,
103 * they should always be called with pm_mutex held (gfp_allowed_mask also should
104 * only be modified with pm_mutex held, unless the suspend/hibernate code is
105 * guaranteed not to run in parallel with that modification).
107 void set_gfp_allowed_mask(gfp_t mask)
109 WARN_ON(!mutex_is_locked(&pm_mutex));
110 gfp_allowed_mask = mask;
113 gfp_t clear_gfp_allowed_mask(gfp_t mask)
115 gfp_t ret = gfp_allowed_mask;
117 WARN_ON(!mutex_is_locked(&pm_mutex));
118 gfp_allowed_mask &= ~mask;
121 #endif /* CONFIG_PM_SLEEP */
123 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
124 int pageblock_order __read_mostly;
127 static void __free_pages_ok(struct page *page, unsigned int order);
130 * results with 256, 32 in the lowmem_reserve sysctl:
131 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
132 * 1G machine -> (16M dma, 784M normal, 224M high)
133 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
134 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
135 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
137 * TBD: should special case ZONE_DMA32 machines here - in those we normally
138 * don't need any ZONE_NORMAL reservation
140 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
141 #ifdef CONFIG_ZONE_DMA
144 #ifdef CONFIG_ZONE_DMA32
147 #ifdef CONFIG_HIGHMEM
153 EXPORT_SYMBOL(totalram_pages);
155 static char * const zone_names[MAX_NR_ZONES] = {
156 #ifdef CONFIG_ZONE_DMA
159 #ifdef CONFIG_ZONE_DMA32
163 #ifdef CONFIG_HIGHMEM
169 int min_free_kbytes = 1024;
171 static unsigned long __meminitdata nr_kernel_pages;
172 static unsigned long __meminitdata nr_all_pages;
173 static unsigned long __meminitdata dma_reserve;
175 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
177 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
178 * ranges of memory (RAM) that may be registered with add_active_range().
179 * Ranges passed to add_active_range() will be merged if possible
180 * so the number of times add_active_range() can be called is
181 * related to the number of nodes and the number of holes
183 #ifdef CONFIG_MAX_ACTIVE_REGIONS
184 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
185 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
187 #if MAX_NUMNODES >= 32
188 /* If there can be many nodes, allow up to 50 holes per node */
189 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
191 /* By default, allow up to 256 distinct regions */
192 #define MAX_ACTIVE_REGIONS 256
196 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
197 static int __meminitdata nr_nodemap_entries;
198 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
199 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
200 static unsigned long __initdata required_kernelcore;
201 static unsigned long __initdata required_movablecore;
202 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
204 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
206 EXPORT_SYMBOL(movable_zone);
207 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
210 int nr_node_ids __read_mostly = MAX_NUMNODES;
211 int nr_online_nodes __read_mostly = 1;
212 EXPORT_SYMBOL(nr_node_ids);
213 EXPORT_SYMBOL(nr_online_nodes);
216 int page_group_by_mobility_disabled __read_mostly;
218 static void set_pageblock_migratetype(struct page *page, int migratetype)
221 if (unlikely(page_group_by_mobility_disabled))
222 migratetype = MIGRATE_UNMOVABLE;
224 set_pageblock_flags_group(page, (unsigned long)migratetype,
225 PB_migrate, PB_migrate_end);
228 bool oom_killer_disabled __read_mostly;
230 #ifdef CONFIG_DEBUG_VM
231 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
235 unsigned long pfn = page_to_pfn(page);
238 seq = zone_span_seqbegin(zone);
239 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
241 else if (pfn < zone->zone_start_pfn)
243 } while (zone_span_seqretry(zone, seq));
248 static int page_is_consistent(struct zone *zone, struct page *page)
250 if (!pfn_valid_within(page_to_pfn(page)))
252 if (zone != page_zone(page))
258 * Temporary debugging check for pages not lying within a given zone.
260 static int bad_range(struct zone *zone, struct page *page)
262 if (page_outside_zone_boundaries(zone, page))
264 if (!page_is_consistent(zone, page))
270 static inline int bad_range(struct zone *zone, struct page *page)
276 static void bad_page(struct page *page)
278 static unsigned long resume;
279 static unsigned long nr_shown;
280 static unsigned long nr_unshown;
282 /* Don't complain about poisoned pages */
283 if (PageHWPoison(page)) {
284 __ClearPageBuddy(page);
289 * Allow a burst of 60 reports, then keep quiet for that minute;
290 * or allow a steady drip of one report per second.
292 if (nr_shown == 60) {
293 if (time_before(jiffies, resume)) {
299 "BUG: Bad page state: %lu messages suppressed\n",
306 resume = jiffies + 60 * HZ;
308 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
309 current->comm, page_to_pfn(page));
314 /* Leave bad fields for debug, except PageBuddy could make trouble */
315 __ClearPageBuddy(page);
316 add_taint(TAINT_BAD_PAGE);
320 * Higher-order pages are called "compound pages". They are structured thusly:
322 * The first PAGE_SIZE page is called the "head page".
324 * The remaining PAGE_SIZE pages are called "tail pages".
326 * All pages have PG_compound set. All pages have their ->private pointing at
327 * the head page (even the head page has this).
329 * The first tail page's ->lru.next holds the address of the compound page's
330 * put_page() function. Its ->lru.prev holds the order of allocation.
331 * This usage means that zero-order pages may not be compound.
334 static void free_compound_page(struct page *page)
336 __free_pages_ok(page, compound_order(page));
339 void prep_compound_page(struct page *page, unsigned long order)
342 int nr_pages = 1 << order;
344 set_compound_page_dtor(page, free_compound_page);
345 set_compound_order(page, order);
347 for (i = 1; i < nr_pages; i++) {
348 struct page *p = page + i;
351 p->first_page = page;
355 static int destroy_compound_page(struct page *page, unsigned long order)
358 int nr_pages = 1 << order;
361 if (unlikely(compound_order(page) != order) ||
362 unlikely(!PageHead(page))) {
367 __ClearPageHead(page);
369 for (i = 1; i < nr_pages; i++) {
370 struct page *p = page + i;
372 if (unlikely(!PageTail(p) || (p->first_page != page))) {
382 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
387 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
388 * and __GFP_HIGHMEM from hard or soft interrupt context.
390 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
391 for (i = 0; i < (1 << order); i++)
392 clear_highpage(page + i);
395 static inline void set_page_order(struct page *page, int order)
397 set_page_private(page, order);
398 __SetPageBuddy(page);
401 static inline void rmv_page_order(struct page *page)
403 __ClearPageBuddy(page);
404 set_page_private(page, 0);
408 * Locate the struct page for both the matching buddy in our
409 * pair (buddy1) and the combined O(n+1) page they form (page).
411 * 1) Any buddy B1 will have an order O twin B2 which satisfies
412 * the following equation:
414 * For example, if the starting buddy (buddy2) is #8 its order
416 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
418 * 2) Any buddy B will have an order O+1 parent P which
419 * satisfies the following equation:
422 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
424 static inline struct page *
425 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
427 unsigned long buddy_idx = page_idx ^ (1 << order);
429 return page + (buddy_idx - page_idx);
432 static inline unsigned long
433 __find_combined_index(unsigned long page_idx, unsigned int order)
435 return (page_idx & ~(1 << order));
439 * This function checks whether a page is free && is the buddy
440 * we can do coalesce a page and its buddy if
441 * (a) the buddy is not in a hole &&
442 * (b) the buddy is in the buddy system &&
443 * (c) a page and its buddy have the same order &&
444 * (d) a page and its buddy are in the same zone.
446 * For recording whether a page is in the buddy system, we use PG_buddy.
447 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
449 * For recording page's order, we use page_private(page).
451 static inline int page_is_buddy(struct page *page, struct page *buddy,
454 if (!pfn_valid_within(page_to_pfn(buddy)))
457 if (page_zone_id(page) != page_zone_id(buddy))
460 if (PageBuddy(buddy) && page_order(buddy) == order) {
461 VM_BUG_ON(page_count(buddy) != 0);
468 * Freeing function for a buddy system allocator.
470 * The concept of a buddy system is to maintain direct-mapped table
471 * (containing bit values) for memory blocks of various "orders".
472 * The bottom level table contains the map for the smallest allocatable
473 * units of memory (here, pages), and each level above it describes
474 * pairs of units from the levels below, hence, "buddies".
475 * At a high level, all that happens here is marking the table entry
476 * at the bottom level available, and propagating the changes upward
477 * as necessary, plus some accounting needed to play nicely with other
478 * parts of the VM system.
479 * At each level, we keep a list of pages, which are heads of continuous
480 * free pages of length of (1 << order) and marked with PG_buddy. Page's
481 * order is recorded in page_private(page) field.
482 * So when we are allocating or freeing one, we can derive the state of the
483 * other. That is, if we allocate a small block, and both were
484 * free, the remainder of the region must be split into blocks.
485 * If a block is freed, and its buddy is also free, then this
486 * triggers coalescing into a block of larger size.
491 static inline void __free_one_page(struct page *page,
492 struct zone *zone, unsigned int order,
495 unsigned long page_idx;
496 unsigned long combined_idx;
499 if (unlikely(PageCompound(page)))
500 if (unlikely(destroy_compound_page(page, order)))
503 VM_BUG_ON(migratetype == -1);
505 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
507 VM_BUG_ON(page_idx & ((1 << order) - 1));
508 VM_BUG_ON(bad_range(zone, page));
510 while (order < MAX_ORDER-1) {
511 buddy = __page_find_buddy(page, page_idx, order);
512 if (!page_is_buddy(page, buddy, order))
515 /* Our buddy is free, merge with it and move up one order. */
516 list_del(&buddy->lru);
517 zone->free_area[order].nr_free--;
518 rmv_page_order(buddy);
519 combined_idx = __find_combined_index(page_idx, order);
520 page = page + (combined_idx - page_idx);
521 page_idx = combined_idx;
524 set_page_order(page, order);
527 * If this is not the largest possible page, check if the buddy
528 * of the next-highest order is free. If it is, it's possible
529 * that pages are being freed that will coalesce soon. In case,
530 * that is happening, add the free page to the tail of the list
531 * so it's less likely to be used soon and more likely to be merged
532 * as a higher order page
534 if ((order < MAX_ORDER-1) && pfn_valid_within(page_to_pfn(buddy))) {
535 struct page *higher_page, *higher_buddy;
536 combined_idx = __find_combined_index(page_idx, order);
537 higher_page = page + combined_idx - page_idx;
538 higher_buddy = __page_find_buddy(higher_page, combined_idx, order + 1);
539 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
540 list_add_tail(&page->lru,
541 &zone->free_area[order].free_list[migratetype]);
546 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
548 zone->free_area[order].nr_free++;
552 * free_page_mlock() -- clean up attempts to free and mlocked() page.
553 * Page should not be on lru, so no need to fix that up.
554 * free_pages_check() will verify...
556 static inline void free_page_mlock(struct page *page)
558 __dec_zone_page_state(page, NR_MLOCK);
559 __count_vm_event(UNEVICTABLE_MLOCKFREED);
562 static inline int free_pages_check(struct page *page)
564 if (unlikely(page_mapcount(page) |
565 (page->mapping != NULL) |
566 (atomic_read(&page->_count) != 0) |
567 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
571 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
572 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
577 * Frees a number of pages from the PCP lists
578 * Assumes all pages on list are in same zone, and of same order.
579 * count is the number of pages to free.
581 * If the zone was previously in an "all pages pinned" state then look to
582 * see if this freeing clears that state.
584 * And clear the zone's pages_scanned counter, to hold off the "all pages are
585 * pinned" detection logic.
587 static void free_pcppages_bulk(struct zone *zone, int count,
588 struct per_cpu_pages *pcp)
594 spin_lock(&zone->lock);
595 zone->all_unreclaimable = 0;
596 zone->pages_scanned = 0;
600 struct list_head *list;
603 * Remove pages from lists in a round-robin fashion. A
604 * batch_free count is maintained that is incremented when an
605 * empty list is encountered. This is so more pages are freed
606 * off fuller lists instead of spinning excessively around empty
611 if (++migratetype == MIGRATE_PCPTYPES)
613 list = &pcp->lists[migratetype];
614 } while (list_empty(list));
617 page = list_entry(list->prev, struct page, lru);
618 /* must delete as __free_one_page list manipulates */
619 list_del(&page->lru);
620 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
621 __free_one_page(page, zone, 0, page_private(page));
622 trace_mm_page_pcpu_drain(page, 0, page_private(page));
623 } while (--to_free && --batch_free && !list_empty(list));
625 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
626 spin_unlock(&zone->lock);
629 static void free_one_page(struct zone *zone, struct page *page, int order,
632 spin_lock(&zone->lock);
633 zone->all_unreclaimable = 0;
634 zone->pages_scanned = 0;
636 __free_one_page(page, zone, order, migratetype);
637 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
638 spin_unlock(&zone->lock);
641 static bool free_pages_prepare(struct page *page, unsigned int order)
646 trace_mm_page_free_direct(page, order);
647 kmemcheck_free_shadow(page, order);
649 for (i = 0; i < (1 << order); i++) {
650 struct page *pg = page + i;
654 bad += free_pages_check(pg);
659 if (!PageHighMem(page)) {
660 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
661 debug_check_no_obj_freed(page_address(page),
664 arch_free_page(page, order);
665 kernel_map_pages(page, 1 << order, 0);
670 static void __free_pages_ok(struct page *page, unsigned int order)
673 int wasMlocked = __TestClearPageMlocked(page);
675 if (!free_pages_prepare(page, order))
678 local_irq_save(flags);
679 if (unlikely(wasMlocked))
680 free_page_mlock(page);
681 __count_vm_events(PGFREE, 1 << order);
682 free_one_page(page_zone(page), page, order,
683 get_pageblock_migratetype(page));
684 local_irq_restore(flags);
688 * permit the bootmem allocator to evade page validation on high-order frees
690 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
693 __ClearPageReserved(page);
694 set_page_count(page, 0);
695 set_page_refcounted(page);
701 for (loop = 0; loop < BITS_PER_LONG; loop++) {
702 struct page *p = &page[loop];
704 if (loop + 1 < BITS_PER_LONG)
706 __ClearPageReserved(p);
707 set_page_count(p, 0);
710 set_page_refcounted(page);
711 __free_pages(page, order);
717 * The order of subdivision here is critical for the IO subsystem.
718 * Please do not alter this order without good reasons and regression
719 * testing. Specifically, as large blocks of memory are subdivided,
720 * the order in which smaller blocks are delivered depends on the order
721 * they're subdivided in this function. This is the primary factor
722 * influencing the order in which pages are delivered to the IO
723 * subsystem according to empirical testing, and this is also justified
724 * by considering the behavior of a buddy system containing a single
725 * large block of memory acted on by a series of small allocations.
726 * This behavior is a critical factor in sglist merging's success.
730 static inline void expand(struct zone *zone, struct page *page,
731 int low, int high, struct free_area *area,
734 unsigned long size = 1 << high;
740 VM_BUG_ON(bad_range(zone, &page[size]));
741 list_add(&page[size].lru, &area->free_list[migratetype]);
743 set_page_order(&page[size], high);
748 * This page is about to be returned from the page allocator
750 static inline int check_new_page(struct page *page)
752 if (unlikely(page_mapcount(page) |
753 (page->mapping != NULL) |
754 (atomic_read(&page->_count) != 0) |
755 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
762 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
766 for (i = 0; i < (1 << order); i++) {
767 struct page *p = page + i;
768 if (unlikely(check_new_page(p)))
772 set_page_private(page, 0);
773 set_page_refcounted(page);
775 arch_alloc_page(page, order);
776 kernel_map_pages(page, 1 << order, 1);
778 if (gfp_flags & __GFP_ZERO)
779 prep_zero_page(page, order, gfp_flags);
781 if (order && (gfp_flags & __GFP_COMP))
782 prep_compound_page(page, order);
788 * Go through the free lists for the given migratetype and remove
789 * the smallest available page from the freelists
792 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
795 unsigned int current_order;
796 struct free_area * area;
799 /* Find a page of the appropriate size in the preferred list */
800 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
801 area = &(zone->free_area[current_order]);
802 if (list_empty(&area->free_list[migratetype]))
805 page = list_entry(area->free_list[migratetype].next,
807 list_del(&page->lru);
808 rmv_page_order(page);
810 expand(zone, page, order, current_order, area, migratetype);
819 * This array describes the order lists are fallen back to when
820 * the free lists for the desirable migrate type are depleted
822 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
823 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
824 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
825 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
826 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
830 * Move the free pages in a range to the free lists of the requested type.
831 * Note that start_page and end_pages are not aligned on a pageblock
832 * boundary. If alignment is required, use move_freepages_block()
834 static int move_freepages(struct zone *zone,
835 struct page *start_page, struct page *end_page,
842 #ifndef CONFIG_HOLES_IN_ZONE
844 * page_zone is not safe to call in this context when
845 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
846 * anyway as we check zone boundaries in move_freepages_block().
847 * Remove at a later date when no bug reports exist related to
848 * grouping pages by mobility
850 BUG_ON(page_zone(start_page) != page_zone(end_page));
853 for (page = start_page; page <= end_page;) {
854 /* Make sure we are not inadvertently changing nodes */
855 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
857 if (!pfn_valid_within(page_to_pfn(page))) {
862 if (!PageBuddy(page)) {
867 order = page_order(page);
868 list_del(&page->lru);
870 &zone->free_area[order].free_list[migratetype]);
872 pages_moved += 1 << order;
878 static int move_freepages_block(struct zone *zone, struct page *page,
881 unsigned long start_pfn, end_pfn;
882 struct page *start_page, *end_page;
884 start_pfn = page_to_pfn(page);
885 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
886 start_page = pfn_to_page(start_pfn);
887 end_page = start_page + pageblock_nr_pages - 1;
888 end_pfn = start_pfn + pageblock_nr_pages - 1;
890 /* Do not cross zone boundaries */
891 if (start_pfn < zone->zone_start_pfn)
893 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
896 return move_freepages(zone, start_page, end_page, migratetype);
899 static void change_pageblock_range(struct page *pageblock_page,
900 int start_order, int migratetype)
902 int nr_pageblocks = 1 << (start_order - pageblock_order);
904 while (nr_pageblocks--) {
905 set_pageblock_migratetype(pageblock_page, migratetype);
906 pageblock_page += pageblock_nr_pages;
910 /* Remove an element from the buddy allocator from the fallback list */
911 static inline struct page *
912 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
914 struct free_area * area;
919 /* Find the largest possible block of pages in the other list */
920 for (current_order = MAX_ORDER-1; current_order >= order;
922 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
923 migratetype = fallbacks[start_migratetype][i];
925 /* MIGRATE_RESERVE handled later if necessary */
926 if (migratetype == MIGRATE_RESERVE)
929 area = &(zone->free_area[current_order]);
930 if (list_empty(&area->free_list[migratetype]))
933 page = list_entry(area->free_list[migratetype].next,
938 * If breaking a large block of pages, move all free
939 * pages to the preferred allocation list. If falling
940 * back for a reclaimable kernel allocation, be more
941 * agressive about taking ownership of free pages
943 if (unlikely(current_order >= (pageblock_order >> 1)) ||
944 start_migratetype == MIGRATE_RECLAIMABLE ||
945 page_group_by_mobility_disabled) {
947 pages = move_freepages_block(zone, page,
950 /* Claim the whole block if over half of it is free */
951 if (pages >= (1 << (pageblock_order-1)) ||
952 page_group_by_mobility_disabled)
953 set_pageblock_migratetype(page,
956 migratetype = start_migratetype;
959 /* Remove the page from the freelists */
960 list_del(&page->lru);
961 rmv_page_order(page);
963 /* Take ownership for orders >= pageblock_order */
964 if (current_order >= pageblock_order)
965 change_pageblock_range(page, current_order,
968 expand(zone, page, order, current_order, area, migratetype);
970 trace_mm_page_alloc_extfrag(page, order, current_order,
971 start_migratetype, migratetype);
981 * Do the hard work of removing an element from the buddy allocator.
982 * Call me with the zone->lock already held.
984 static struct page *__rmqueue(struct zone *zone, unsigned int order,
990 page = __rmqueue_smallest(zone, order, migratetype);
992 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
993 page = __rmqueue_fallback(zone, order, migratetype);
996 * Use MIGRATE_RESERVE rather than fail an allocation. goto
997 * is used because __rmqueue_smallest is an inline function
998 * and we want just one call site
1001 migratetype = MIGRATE_RESERVE;
1006 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1011 * Obtain a specified number of elements from the buddy allocator, all under
1012 * a single hold of the lock, for efficiency. Add them to the supplied list.
1013 * Returns the number of new pages which were placed at *list.
1015 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1016 unsigned long count, struct list_head *list,
1017 int migratetype, int cold)
1021 spin_lock(&zone->lock);
1022 for (i = 0; i < count; ++i) {
1023 struct page *page = __rmqueue(zone, order, migratetype);
1024 if (unlikely(page == NULL))
1028 * Split buddy pages returned by expand() are received here
1029 * in physical page order. The page is added to the callers and
1030 * list and the list head then moves forward. From the callers
1031 * perspective, the linked list is ordered by page number in
1032 * some conditions. This is useful for IO devices that can
1033 * merge IO requests if the physical pages are ordered
1036 if (likely(cold == 0))
1037 list_add(&page->lru, list);
1039 list_add_tail(&page->lru, list);
1040 set_page_private(page, migratetype);
1043 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1044 spin_unlock(&zone->lock);
1050 * Called from the vmstat counter updater to drain pagesets of this
1051 * currently executing processor on remote nodes after they have
1054 * Note that this function must be called with the thread pinned to
1055 * a single processor.
1057 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1059 unsigned long flags;
1062 local_irq_save(flags);
1063 if (pcp->count >= pcp->batch)
1064 to_drain = pcp->batch;
1066 to_drain = pcp->count;
1067 free_pcppages_bulk(zone, to_drain, pcp);
1068 pcp->count -= to_drain;
1069 local_irq_restore(flags);
1074 * Drain pages of the indicated processor.
1076 * The processor must either be the current processor and the
1077 * thread pinned to the current processor or a processor that
1080 static void drain_pages(unsigned int cpu)
1082 unsigned long flags;
1085 for_each_populated_zone(zone) {
1086 struct per_cpu_pageset *pset;
1087 struct per_cpu_pages *pcp;
1089 local_irq_save(flags);
1090 pset = per_cpu_ptr(zone->pageset, cpu);
1093 free_pcppages_bulk(zone, pcp->count, pcp);
1095 local_irq_restore(flags);
1100 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1102 void drain_local_pages(void *arg)
1104 drain_pages(smp_processor_id());
1108 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1110 void drain_all_pages(void)
1112 on_each_cpu(drain_local_pages, NULL, 1);
1115 #ifdef CONFIG_HIBERNATION
1117 void mark_free_pages(struct zone *zone)
1119 unsigned long pfn, max_zone_pfn;
1120 unsigned long flags;
1122 struct list_head *curr;
1124 if (!zone->spanned_pages)
1127 spin_lock_irqsave(&zone->lock, flags);
1129 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1130 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1131 if (pfn_valid(pfn)) {
1132 struct page *page = pfn_to_page(pfn);
1134 if (!swsusp_page_is_forbidden(page))
1135 swsusp_unset_page_free(page);
1138 for_each_migratetype_order(order, t) {
1139 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1142 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1143 for (i = 0; i < (1UL << order); i++)
1144 swsusp_set_page_free(pfn_to_page(pfn + i));
1147 spin_unlock_irqrestore(&zone->lock, flags);
1149 #endif /* CONFIG_PM */
1152 * Free a 0-order page
1153 * cold == 1 ? free a cold page : free a hot page
1155 void free_hot_cold_page(struct page *page, int cold)
1157 struct zone *zone = page_zone(page);
1158 struct per_cpu_pages *pcp;
1159 unsigned long flags;
1161 int wasMlocked = __TestClearPageMlocked(page);
1163 if (!free_pages_prepare(page, 0))
1166 migratetype = get_pageblock_migratetype(page);
1167 set_page_private(page, migratetype);
1168 local_irq_save(flags);
1169 if (unlikely(wasMlocked))
1170 free_page_mlock(page);
1171 __count_vm_event(PGFREE);
1174 * We only track unmovable, reclaimable and movable on pcp lists.
1175 * Free ISOLATE pages back to the allocator because they are being
1176 * offlined but treat RESERVE as movable pages so we can get those
1177 * areas back if necessary. Otherwise, we may have to free
1178 * excessively into the page allocator
1180 if (migratetype >= MIGRATE_PCPTYPES) {
1181 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1182 free_one_page(zone, page, 0, migratetype);
1185 migratetype = MIGRATE_MOVABLE;
1188 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1190 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1192 list_add(&page->lru, &pcp->lists[migratetype]);
1194 if (pcp->count >= pcp->high) {
1195 free_pcppages_bulk(zone, pcp->batch, pcp);
1196 pcp->count -= pcp->batch;
1200 local_irq_restore(flags);
1204 * split_page takes a non-compound higher-order page, and splits it into
1205 * n (1<<order) sub-pages: page[0..n]
1206 * Each sub-page must be freed individually.
1208 * Note: this is probably too low level an operation for use in drivers.
1209 * Please consult with lkml before using this in your driver.
1211 void split_page(struct page *page, unsigned int order)
1215 VM_BUG_ON(PageCompound(page));
1216 VM_BUG_ON(!page_count(page));
1218 #ifdef CONFIG_KMEMCHECK
1220 * Split shadow pages too, because free(page[0]) would
1221 * otherwise free the whole shadow.
1223 if (kmemcheck_page_is_tracked(page))
1224 split_page(virt_to_page(page[0].shadow), order);
1227 for (i = 1; i < (1 << order); i++)
1228 set_page_refcounted(page + i);
1232 * Similar to split_page except the page is already free. As this is only
1233 * being used for migration, the migratetype of the block also changes.
1234 * As this is called with interrupts disabled, the caller is responsible
1235 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1238 * Note: this is probably too low level an operation for use in drivers.
1239 * Please consult with lkml before using this in your driver.
1241 int split_free_page(struct page *page)
1244 unsigned long watermark;
1247 BUG_ON(!PageBuddy(page));
1249 zone = page_zone(page);
1250 order = page_order(page);
1252 /* Obey watermarks as if the page was being allocated */
1253 watermark = low_wmark_pages(zone) + (1 << order);
1254 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1257 /* Remove page from free list */
1258 list_del(&page->lru);
1259 zone->free_area[order].nr_free--;
1260 rmv_page_order(page);
1261 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1263 /* Split into individual pages */
1264 set_page_refcounted(page);
1265 split_page(page, order);
1267 if (order >= pageblock_order - 1) {
1268 struct page *endpage = page + (1 << order) - 1;
1269 for (; page < endpage; page += pageblock_nr_pages)
1270 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1277 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1278 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1282 struct page *buffered_rmqueue(struct zone *preferred_zone,
1283 struct zone *zone, int order, gfp_t gfp_flags,
1286 unsigned long flags;
1288 int cold = !!(gfp_flags & __GFP_COLD);
1291 if (likely(order == 0)) {
1292 struct per_cpu_pages *pcp;
1293 struct list_head *list;
1295 local_irq_save(flags);
1296 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1297 list = &pcp->lists[migratetype];
1298 if (list_empty(list)) {
1299 pcp->count += rmqueue_bulk(zone, 0,
1302 if (unlikely(list_empty(list)))
1307 page = list_entry(list->prev, struct page, lru);
1309 page = list_entry(list->next, struct page, lru);
1311 list_del(&page->lru);
1314 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1316 * __GFP_NOFAIL is not to be used in new code.
1318 * All __GFP_NOFAIL callers should be fixed so that they
1319 * properly detect and handle allocation failures.
1321 * We most definitely don't want callers attempting to
1322 * allocate greater than order-1 page units with
1325 WARN_ON_ONCE(order > 1);
1327 spin_lock_irqsave(&zone->lock, flags);
1328 page = __rmqueue(zone, order, migratetype);
1329 spin_unlock(&zone->lock);
1332 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1335 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1336 zone_statistics(preferred_zone, zone);
1337 local_irq_restore(flags);
1339 VM_BUG_ON(bad_range(zone, page));
1340 if (prep_new_page(page, order, gfp_flags))
1345 local_irq_restore(flags);
1349 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1350 #define ALLOC_WMARK_MIN WMARK_MIN
1351 #define ALLOC_WMARK_LOW WMARK_LOW
1352 #define ALLOC_WMARK_HIGH WMARK_HIGH
1353 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1355 /* Mask to get the watermark bits */
1356 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1358 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1359 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1360 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1362 #ifdef CONFIG_FAIL_PAGE_ALLOC
1364 static struct fail_page_alloc_attr {
1365 struct fault_attr attr;
1367 u32 ignore_gfp_highmem;
1368 u32 ignore_gfp_wait;
1371 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1373 struct dentry *ignore_gfp_highmem_file;
1374 struct dentry *ignore_gfp_wait_file;
1375 struct dentry *min_order_file;
1377 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1379 } fail_page_alloc = {
1380 .attr = FAULT_ATTR_INITIALIZER,
1381 .ignore_gfp_wait = 1,
1382 .ignore_gfp_highmem = 1,
1386 static int __init setup_fail_page_alloc(char *str)
1388 return setup_fault_attr(&fail_page_alloc.attr, str);
1390 __setup("fail_page_alloc=", setup_fail_page_alloc);
1392 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1394 if (order < fail_page_alloc.min_order)
1396 if (gfp_mask & __GFP_NOFAIL)
1398 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1400 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1403 return should_fail(&fail_page_alloc.attr, 1 << order);
1406 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1408 static int __init fail_page_alloc_debugfs(void)
1410 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1414 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1418 dir = fail_page_alloc.attr.dentries.dir;
1420 fail_page_alloc.ignore_gfp_wait_file =
1421 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1422 &fail_page_alloc.ignore_gfp_wait);
1424 fail_page_alloc.ignore_gfp_highmem_file =
1425 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1426 &fail_page_alloc.ignore_gfp_highmem);
1427 fail_page_alloc.min_order_file =
1428 debugfs_create_u32("min-order", mode, dir,
1429 &fail_page_alloc.min_order);
1431 if (!fail_page_alloc.ignore_gfp_wait_file ||
1432 !fail_page_alloc.ignore_gfp_highmem_file ||
1433 !fail_page_alloc.min_order_file) {
1435 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1436 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1437 debugfs_remove(fail_page_alloc.min_order_file);
1438 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1444 late_initcall(fail_page_alloc_debugfs);
1446 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1448 #else /* CONFIG_FAIL_PAGE_ALLOC */
1450 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1455 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1458 * Return 1 if free pages are above 'mark'. This takes into account the order
1459 * of the allocation.
1461 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1462 int classzone_idx, int alloc_flags)
1464 /* free_pages my go negative - that's OK */
1466 long free_pages = zone_nr_free_pages(z) - (1 << order) + 1;
1469 if (alloc_flags & ALLOC_HIGH)
1471 if (alloc_flags & ALLOC_HARDER)
1474 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1476 for (o = 0; o < order; o++) {
1477 /* At the next order, this order's pages become unavailable */
1478 free_pages -= z->free_area[o].nr_free << o;
1480 /* Require fewer higher order pages to be free */
1483 if (free_pages <= min)
1491 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1492 * skip over zones that are not allowed by the cpuset, or that have
1493 * been recently (in last second) found to be nearly full. See further
1494 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1495 * that have to skip over a lot of full or unallowed zones.
1497 * If the zonelist cache is present in the passed in zonelist, then
1498 * returns a pointer to the allowed node mask (either the current
1499 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1501 * If the zonelist cache is not available for this zonelist, does
1502 * nothing and returns NULL.
1504 * If the fullzones BITMAP in the zonelist cache is stale (more than
1505 * a second since last zap'd) then we zap it out (clear its bits.)
1507 * We hold off even calling zlc_setup, until after we've checked the
1508 * first zone in the zonelist, on the theory that most allocations will
1509 * be satisfied from that first zone, so best to examine that zone as
1510 * quickly as we can.
1512 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1514 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1515 nodemask_t *allowednodes; /* zonelist_cache approximation */
1517 zlc = zonelist->zlcache_ptr;
1521 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1522 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1523 zlc->last_full_zap = jiffies;
1526 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1527 &cpuset_current_mems_allowed :
1528 &node_states[N_HIGH_MEMORY];
1529 return allowednodes;
1533 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1534 * if it is worth looking at further for free memory:
1535 * 1) Check that the zone isn't thought to be full (doesn't have its
1536 * bit set in the zonelist_cache fullzones BITMAP).
1537 * 2) Check that the zones node (obtained from the zonelist_cache
1538 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1539 * Return true (non-zero) if zone is worth looking at further, or
1540 * else return false (zero) if it is not.
1542 * This check -ignores- the distinction between various watermarks,
1543 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1544 * found to be full for any variation of these watermarks, it will
1545 * be considered full for up to one second by all requests, unless
1546 * we are so low on memory on all allowed nodes that we are forced
1547 * into the second scan of the zonelist.
1549 * In the second scan we ignore this zonelist cache and exactly
1550 * apply the watermarks to all zones, even it is slower to do so.
1551 * We are low on memory in the second scan, and should leave no stone
1552 * unturned looking for a free page.
1554 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1555 nodemask_t *allowednodes)
1557 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1558 int i; /* index of *z in zonelist zones */
1559 int n; /* node that zone *z is on */
1561 zlc = zonelist->zlcache_ptr;
1565 i = z - zonelist->_zonerefs;
1568 /* This zone is worth trying if it is allowed but not full */
1569 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1573 * Given 'z' scanning a zonelist, set the corresponding bit in
1574 * zlc->fullzones, so that subsequent attempts to allocate a page
1575 * from that zone don't waste time re-examining it.
1577 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1579 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1580 int i; /* index of *z in zonelist zones */
1582 zlc = zonelist->zlcache_ptr;
1586 i = z - zonelist->_zonerefs;
1588 set_bit(i, zlc->fullzones);
1591 #else /* CONFIG_NUMA */
1593 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1598 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1599 nodemask_t *allowednodes)
1604 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1607 #endif /* CONFIG_NUMA */
1610 * get_page_from_freelist goes through the zonelist trying to allocate
1613 static struct page *
1614 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1615 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1616 struct zone *preferred_zone, int migratetype)
1619 struct page *page = NULL;
1622 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1623 int zlc_active = 0; /* set if using zonelist_cache */
1624 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1626 classzone_idx = zone_idx(preferred_zone);
1629 * Scan zonelist, looking for a zone with enough free.
1630 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1632 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1633 high_zoneidx, nodemask) {
1634 if (NUMA_BUILD && zlc_active &&
1635 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1637 if ((alloc_flags & ALLOC_CPUSET) &&
1638 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1641 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1642 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1646 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1647 if (zone_watermark_ok(zone, order, mark,
1648 classzone_idx, alloc_flags))
1651 if (zone_reclaim_mode == 0)
1652 goto this_zone_full;
1654 ret = zone_reclaim(zone, gfp_mask, order);
1656 case ZONE_RECLAIM_NOSCAN:
1659 case ZONE_RECLAIM_FULL:
1660 /* scanned but unreclaimable */
1661 goto this_zone_full;
1663 /* did we reclaim enough */
1664 if (!zone_watermark_ok(zone, order, mark,
1665 classzone_idx, alloc_flags))
1666 goto this_zone_full;
1671 page = buffered_rmqueue(preferred_zone, zone, order,
1672 gfp_mask, migratetype);
1677 zlc_mark_zone_full(zonelist, z);
1679 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1681 * we do zlc_setup after the first zone is tried but only
1682 * if there are multiple nodes make it worthwhile
1684 allowednodes = zlc_setup(zonelist, alloc_flags);
1690 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1691 /* Disable zlc cache for second zonelist scan */
1699 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1700 unsigned long pages_reclaimed)
1702 /* Do not loop if specifically requested */
1703 if (gfp_mask & __GFP_NORETRY)
1707 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1708 * means __GFP_NOFAIL, but that may not be true in other
1711 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1715 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1716 * specified, then we retry until we no longer reclaim any pages
1717 * (above), or we've reclaimed an order of pages at least as
1718 * large as the allocation's order. In both cases, if the
1719 * allocation still fails, we stop retrying.
1721 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1725 * Don't let big-order allocations loop unless the caller
1726 * explicitly requests that.
1728 if (gfp_mask & __GFP_NOFAIL)
1734 static inline struct page *
1735 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1736 struct zonelist *zonelist, enum zone_type high_zoneidx,
1737 nodemask_t *nodemask, struct zone *preferred_zone,
1742 /* Acquire the OOM killer lock for the zones in zonelist */
1743 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1744 schedule_timeout_uninterruptible(1);
1749 * Go through the zonelist yet one more time, keep very high watermark
1750 * here, this is only to catch a parallel oom killing, we must fail if
1751 * we're still under heavy pressure.
1753 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1754 order, zonelist, high_zoneidx,
1755 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1756 preferred_zone, migratetype);
1760 if (!(gfp_mask & __GFP_NOFAIL)) {
1761 /* The OOM killer will not help higher order allocs */
1762 if (order > PAGE_ALLOC_COSTLY_ORDER)
1764 /* The OOM killer does not needlessly kill tasks for lowmem */
1765 if (high_zoneidx < ZONE_NORMAL)
1768 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1769 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1770 * The caller should handle page allocation failure by itself if
1771 * it specifies __GFP_THISNODE.
1772 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1774 if (gfp_mask & __GFP_THISNODE)
1777 /* Exhausted what can be done so it's blamo time */
1778 out_of_memory(zonelist, gfp_mask, order, nodemask);
1781 clear_zonelist_oom(zonelist, gfp_mask);
1785 #ifdef CONFIG_COMPACTION
1786 /* Try memory compaction for high-order allocations before reclaim */
1787 static struct page *
1788 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1789 struct zonelist *zonelist, enum zone_type high_zoneidx,
1790 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1791 int migratetype, unsigned long *did_some_progress)
1795 if (!order || compaction_deferred(preferred_zone))
1798 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1800 if (*did_some_progress != COMPACT_SKIPPED) {
1802 /* Page migration frees to the PCP lists but we want merging */
1803 drain_pages(get_cpu());
1806 page = get_page_from_freelist(gfp_mask, nodemask,
1807 order, zonelist, high_zoneidx,
1808 alloc_flags, preferred_zone,
1811 preferred_zone->compact_considered = 0;
1812 preferred_zone->compact_defer_shift = 0;
1813 count_vm_event(COMPACTSUCCESS);
1818 * It's bad if compaction run occurs and fails.
1819 * The most likely reason is that pages exist,
1820 * but not enough to satisfy watermarks.
1822 count_vm_event(COMPACTFAIL);
1823 defer_compaction(preferred_zone);
1831 static inline struct page *
1832 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1833 struct zonelist *zonelist, enum zone_type high_zoneidx,
1834 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1835 int migratetype, unsigned long *did_some_progress)
1839 #endif /* CONFIG_COMPACTION */
1841 /* The really slow allocator path where we enter direct reclaim */
1842 static inline struct page *
1843 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1844 struct zonelist *zonelist, enum zone_type high_zoneidx,
1845 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1846 int migratetype, unsigned long *did_some_progress)
1848 struct page *page = NULL;
1849 struct reclaim_state reclaim_state;
1850 struct task_struct *p = current;
1851 bool drained = false;
1855 /* We now go into synchronous reclaim */
1856 cpuset_memory_pressure_bump();
1857 p->flags |= PF_MEMALLOC;
1858 lockdep_set_current_reclaim_state(gfp_mask);
1859 reclaim_state.reclaimed_slab = 0;
1860 p->reclaim_state = &reclaim_state;
1862 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1864 p->reclaim_state = NULL;
1865 lockdep_clear_current_reclaim_state();
1866 p->flags &= ~PF_MEMALLOC;
1870 if (unlikely(!(*did_some_progress)))
1874 page = get_page_from_freelist(gfp_mask, nodemask, order,
1875 zonelist, high_zoneidx,
1876 alloc_flags, preferred_zone,
1880 * If an allocation failed after direct reclaim, it could be because
1881 * pages are pinned on the per-cpu lists. Drain them and try again
1883 if (!page && !drained) {
1893 * This is called in the allocator slow-path if the allocation request is of
1894 * sufficient urgency to ignore watermarks and take other desperate measures
1896 static inline struct page *
1897 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1898 struct zonelist *zonelist, enum zone_type high_zoneidx,
1899 nodemask_t *nodemask, struct zone *preferred_zone,
1905 page = get_page_from_freelist(gfp_mask, nodemask, order,
1906 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1907 preferred_zone, migratetype);
1909 if (!page && gfp_mask & __GFP_NOFAIL)
1910 congestion_wait(BLK_RW_ASYNC, HZ/50);
1911 } while (!page && (gfp_mask & __GFP_NOFAIL));
1917 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1918 enum zone_type high_zoneidx)
1923 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1924 wakeup_kswapd(zone, order);
1928 gfp_to_alloc_flags(gfp_t gfp_mask)
1930 struct task_struct *p = current;
1931 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1932 const gfp_t wait = gfp_mask & __GFP_WAIT;
1934 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1935 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1938 * The caller may dip into page reserves a bit more if the caller
1939 * cannot run direct reclaim, or if the caller has realtime scheduling
1940 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1941 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1943 alloc_flags |= (gfp_mask & __GFP_HIGH);
1946 alloc_flags |= ALLOC_HARDER;
1948 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1949 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1951 alloc_flags &= ~ALLOC_CPUSET;
1952 } else if (unlikely(rt_task(p)) && !in_interrupt())
1953 alloc_flags |= ALLOC_HARDER;
1955 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1956 if (!in_interrupt() &&
1957 ((p->flags & PF_MEMALLOC) ||
1958 unlikely(test_thread_flag(TIF_MEMDIE))))
1959 alloc_flags |= ALLOC_NO_WATERMARKS;
1965 static inline struct page *
1966 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1967 struct zonelist *zonelist, enum zone_type high_zoneidx,
1968 nodemask_t *nodemask, struct zone *preferred_zone,
1971 const gfp_t wait = gfp_mask & __GFP_WAIT;
1972 struct page *page = NULL;
1974 unsigned long pages_reclaimed = 0;
1975 unsigned long did_some_progress;
1976 struct task_struct *p = current;
1979 * In the slowpath, we sanity check order to avoid ever trying to
1980 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1981 * be using allocators in order of preference for an area that is
1984 if (order >= MAX_ORDER) {
1985 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1990 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1991 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1992 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1993 * using a larger set of nodes after it has established that the
1994 * allowed per node queues are empty and that nodes are
1997 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2001 wake_all_kswapd(order, zonelist, high_zoneidx);
2004 * OK, we're below the kswapd watermark and have kicked background
2005 * reclaim. Now things get more complex, so set up alloc_flags according
2006 * to how we want to proceed.
2008 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2010 /* This is the last chance, in general, before the goto nopage. */
2011 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2012 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2013 preferred_zone, migratetype);
2018 /* Allocate without watermarks if the context allows */
2019 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2020 page = __alloc_pages_high_priority(gfp_mask, order,
2021 zonelist, high_zoneidx, nodemask,
2022 preferred_zone, migratetype);
2027 /* Atomic allocations - we can't balance anything */
2031 /* Avoid recursion of direct reclaim */
2032 if (p->flags & PF_MEMALLOC)
2035 /* Avoid allocations with no watermarks from looping endlessly */
2036 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2039 /* Try direct compaction */
2040 page = __alloc_pages_direct_compact(gfp_mask, order,
2041 zonelist, high_zoneidx,
2043 alloc_flags, preferred_zone,
2044 migratetype, &did_some_progress);
2048 /* Try direct reclaim and then allocating */
2049 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2050 zonelist, high_zoneidx,
2052 alloc_flags, preferred_zone,
2053 migratetype, &did_some_progress);
2058 * If we failed to make any progress reclaiming, then we are
2059 * running out of options and have to consider going OOM
2061 if (!did_some_progress) {
2062 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2063 if (oom_killer_disabled)
2065 page = __alloc_pages_may_oom(gfp_mask, order,
2066 zonelist, high_zoneidx,
2067 nodemask, preferred_zone,
2072 if (!(gfp_mask & __GFP_NOFAIL)) {
2074 * The oom killer is not called for high-order
2075 * allocations that may fail, so if no progress
2076 * is being made, there are no other options and
2077 * retrying is unlikely to help.
2079 if (order > PAGE_ALLOC_COSTLY_ORDER)
2082 * The oom killer is not called for lowmem
2083 * allocations to prevent needlessly killing
2086 if (high_zoneidx < ZONE_NORMAL)
2094 /* Check if we should retry the allocation */
2095 pages_reclaimed += did_some_progress;
2096 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2097 /* Wait for some write requests to complete then retry */
2098 congestion_wait(BLK_RW_ASYNC, HZ/50);
2103 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
2104 printk(KERN_WARNING "%s: page allocation failure."
2105 " order:%d, mode:0x%x\n",
2106 p->comm, order, gfp_mask);
2112 if (kmemcheck_enabled)
2113 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2119 * This is the 'heart' of the zoned buddy allocator.
2122 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2123 struct zonelist *zonelist, nodemask_t *nodemask)
2125 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2126 struct zone *preferred_zone;
2128 int migratetype = allocflags_to_migratetype(gfp_mask);
2130 gfp_mask &= gfp_allowed_mask;
2132 lockdep_trace_alloc(gfp_mask);
2134 might_sleep_if(gfp_mask & __GFP_WAIT);
2136 if (should_fail_alloc_page(gfp_mask, order))
2140 * Check the zones suitable for the gfp_mask contain at least one
2141 * valid zone. It's possible to have an empty zonelist as a result
2142 * of GFP_THISNODE and a memoryless node
2144 if (unlikely(!zonelist->_zonerefs->zone))
2148 /* The preferred zone is used for statistics later */
2149 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
2150 if (!preferred_zone) {
2155 /* First allocation attempt */
2156 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2157 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2158 preferred_zone, migratetype);
2159 if (unlikely(!page))
2160 page = __alloc_pages_slowpath(gfp_mask, order,
2161 zonelist, high_zoneidx, nodemask,
2162 preferred_zone, migratetype);
2165 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2168 EXPORT_SYMBOL(__alloc_pages_nodemask);
2171 * Common helper functions.
2173 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2178 * __get_free_pages() returns a 32-bit address, which cannot represent
2181 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2183 page = alloc_pages(gfp_mask, order);
2186 return (unsigned long) page_address(page);
2188 EXPORT_SYMBOL(__get_free_pages);
2190 unsigned long get_zeroed_page(gfp_t gfp_mask)
2192 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2194 EXPORT_SYMBOL(get_zeroed_page);
2196 void __pagevec_free(struct pagevec *pvec)
2198 int i = pagevec_count(pvec);
2201 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2202 free_hot_cold_page(pvec->pages[i], pvec->cold);
2206 void __free_pages(struct page *page, unsigned int order)
2208 if (put_page_testzero(page)) {
2210 free_hot_cold_page(page, 0);
2212 __free_pages_ok(page, order);
2216 EXPORT_SYMBOL(__free_pages);
2218 void free_pages(unsigned long addr, unsigned int order)
2221 VM_BUG_ON(!virt_addr_valid((void *)addr));
2222 __free_pages(virt_to_page((void *)addr), order);
2226 EXPORT_SYMBOL(free_pages);
2229 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2230 * @size: the number of bytes to allocate
2231 * @gfp_mask: GFP flags for the allocation
2233 * This function is similar to alloc_pages(), except that it allocates the
2234 * minimum number of pages to satisfy the request. alloc_pages() can only
2235 * allocate memory in power-of-two pages.
2237 * This function is also limited by MAX_ORDER.
2239 * Memory allocated by this function must be released by free_pages_exact().
2241 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2243 unsigned int order = get_order(size);
2246 addr = __get_free_pages(gfp_mask, order);
2248 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2249 unsigned long used = addr + PAGE_ALIGN(size);
2251 split_page(virt_to_page((void *)addr), order);
2252 while (used < alloc_end) {
2258 return (void *)addr;
2260 EXPORT_SYMBOL(alloc_pages_exact);
2263 * free_pages_exact - release memory allocated via alloc_pages_exact()
2264 * @virt: the value returned by alloc_pages_exact.
2265 * @size: size of allocation, same value as passed to alloc_pages_exact().
2267 * Release the memory allocated by a previous call to alloc_pages_exact.
2269 void free_pages_exact(void *virt, size_t size)
2271 unsigned long addr = (unsigned long)virt;
2272 unsigned long end = addr + PAGE_ALIGN(size);
2274 while (addr < end) {
2279 EXPORT_SYMBOL(free_pages_exact);
2281 static unsigned int nr_free_zone_pages(int offset)
2286 /* Just pick one node, since fallback list is circular */
2287 unsigned int sum = 0;
2289 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2291 for_each_zone_zonelist(zone, z, zonelist, offset) {
2292 unsigned long size = zone->present_pages;
2293 unsigned long high = high_wmark_pages(zone);
2302 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2304 unsigned int nr_free_buffer_pages(void)
2306 return nr_free_zone_pages(gfp_zone(GFP_USER));
2308 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2311 * Amount of free RAM allocatable within all zones
2313 unsigned int nr_free_pagecache_pages(void)
2315 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2318 static inline void show_node(struct zone *zone)
2321 printk("Node %d ", zone_to_nid(zone));
2324 void si_meminfo(struct sysinfo *val)
2326 val->totalram = totalram_pages;
2328 val->freeram = global_page_state(NR_FREE_PAGES);
2329 val->bufferram = nr_blockdev_pages();
2330 val->totalhigh = totalhigh_pages;
2331 val->freehigh = nr_free_highpages();
2332 val->mem_unit = PAGE_SIZE;
2335 EXPORT_SYMBOL(si_meminfo);
2338 void si_meminfo_node(struct sysinfo *val, int nid)
2340 pg_data_t *pgdat = NODE_DATA(nid);
2342 val->totalram = pgdat->node_present_pages;
2343 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2344 #ifdef CONFIG_HIGHMEM
2345 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2346 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2352 val->mem_unit = PAGE_SIZE;
2356 #define K(x) ((x) << (PAGE_SHIFT-10))
2359 * Show free area list (used inside shift_scroll-lock stuff)
2360 * We also calculate the percentage fragmentation. We do this by counting the
2361 * memory on each free list with the exception of the first item on the list.
2363 void show_free_areas(void)
2368 for_each_populated_zone(zone) {
2370 printk("%s per-cpu:\n", zone->name);
2372 for_each_online_cpu(cpu) {
2373 struct per_cpu_pageset *pageset;
2375 pageset = per_cpu_ptr(zone->pageset, cpu);
2377 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2378 cpu, pageset->pcp.high,
2379 pageset->pcp.batch, pageset->pcp.count);
2383 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2384 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2386 " dirty:%lu writeback:%lu unstable:%lu\n"
2387 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2388 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2389 global_page_state(NR_ACTIVE_ANON),
2390 global_page_state(NR_INACTIVE_ANON),
2391 global_page_state(NR_ISOLATED_ANON),
2392 global_page_state(NR_ACTIVE_FILE),
2393 global_page_state(NR_INACTIVE_FILE),
2394 global_page_state(NR_ISOLATED_FILE),
2395 global_page_state(NR_UNEVICTABLE),
2396 global_page_state(NR_FILE_DIRTY),
2397 global_page_state(NR_WRITEBACK),
2398 global_page_state(NR_UNSTABLE_NFS),
2399 global_page_state(NR_FREE_PAGES),
2400 global_page_state(NR_SLAB_RECLAIMABLE),
2401 global_page_state(NR_SLAB_UNRECLAIMABLE),
2402 global_page_state(NR_FILE_MAPPED),
2403 global_page_state(NR_SHMEM),
2404 global_page_state(NR_PAGETABLE),
2405 global_page_state(NR_BOUNCE));
2407 for_each_populated_zone(zone) {
2416 " active_anon:%lukB"
2417 " inactive_anon:%lukB"
2418 " active_file:%lukB"
2419 " inactive_file:%lukB"
2420 " unevictable:%lukB"
2421 " isolated(anon):%lukB"
2422 " isolated(file):%lukB"
2429 " slab_reclaimable:%lukB"
2430 " slab_unreclaimable:%lukB"
2431 " kernel_stack:%lukB"
2435 " writeback_tmp:%lukB"
2436 " pages_scanned:%lu"
2437 " all_unreclaimable? %s"
2440 K(zone_nr_free_pages(zone)),
2441 K(min_wmark_pages(zone)),
2442 K(low_wmark_pages(zone)),
2443 K(high_wmark_pages(zone)),
2444 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2445 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2446 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2447 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2448 K(zone_page_state(zone, NR_UNEVICTABLE)),
2449 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2450 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2451 K(zone->present_pages),
2452 K(zone_page_state(zone, NR_MLOCK)),
2453 K(zone_page_state(zone, NR_FILE_DIRTY)),
2454 K(zone_page_state(zone, NR_WRITEBACK)),
2455 K(zone_page_state(zone, NR_FILE_MAPPED)),
2456 K(zone_page_state(zone, NR_SHMEM)),
2457 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2458 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2459 zone_page_state(zone, NR_KERNEL_STACK) *
2461 K(zone_page_state(zone, NR_PAGETABLE)),
2462 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2463 K(zone_page_state(zone, NR_BOUNCE)),
2464 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2465 zone->pages_scanned,
2466 (zone->all_unreclaimable ? "yes" : "no")
2468 printk("lowmem_reserve[]:");
2469 for (i = 0; i < MAX_NR_ZONES; i++)
2470 printk(" %lu", zone->lowmem_reserve[i]);
2474 for_each_populated_zone(zone) {
2475 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2478 printk("%s: ", zone->name);
2480 spin_lock_irqsave(&zone->lock, flags);
2481 for (order = 0; order < MAX_ORDER; order++) {
2482 nr[order] = zone->free_area[order].nr_free;
2483 total += nr[order] << order;
2485 spin_unlock_irqrestore(&zone->lock, flags);
2486 for (order = 0; order < MAX_ORDER; order++)
2487 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2488 printk("= %lukB\n", K(total));
2491 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2493 show_swap_cache_info();
2496 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2498 zoneref->zone = zone;
2499 zoneref->zone_idx = zone_idx(zone);
2503 * Builds allocation fallback zone lists.
2505 * Add all populated zones of a node to the zonelist.
2507 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2508 int nr_zones, enum zone_type zone_type)
2512 BUG_ON(zone_type >= MAX_NR_ZONES);
2517 zone = pgdat->node_zones + zone_type;
2518 if (populated_zone(zone)) {
2519 zoneref_set_zone(zone,
2520 &zonelist->_zonerefs[nr_zones++]);
2521 check_highest_zone(zone_type);
2524 } while (zone_type);
2531 * 0 = automatic detection of better ordering.
2532 * 1 = order by ([node] distance, -zonetype)
2533 * 2 = order by (-zonetype, [node] distance)
2535 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2536 * the same zonelist. So only NUMA can configure this param.
2538 #define ZONELIST_ORDER_DEFAULT 0
2539 #define ZONELIST_ORDER_NODE 1
2540 #define ZONELIST_ORDER_ZONE 2
2542 /* zonelist order in the kernel.
2543 * set_zonelist_order() will set this to NODE or ZONE.
2545 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2546 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2550 /* The value user specified ....changed by config */
2551 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2552 /* string for sysctl */
2553 #define NUMA_ZONELIST_ORDER_LEN 16
2554 char numa_zonelist_order[16] = "default";
2557 * interface for configure zonelist ordering.
2558 * command line option "numa_zonelist_order"
2559 * = "[dD]efault - default, automatic configuration.
2560 * = "[nN]ode - order by node locality, then by zone within node
2561 * = "[zZ]one - order by zone, then by locality within zone
2564 static int __parse_numa_zonelist_order(char *s)
2566 if (*s == 'd' || *s == 'D') {
2567 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2568 } else if (*s == 'n' || *s == 'N') {
2569 user_zonelist_order = ZONELIST_ORDER_NODE;
2570 } else if (*s == 'z' || *s == 'Z') {
2571 user_zonelist_order = ZONELIST_ORDER_ZONE;
2574 "Ignoring invalid numa_zonelist_order value: "
2581 static __init int setup_numa_zonelist_order(char *s)
2584 return __parse_numa_zonelist_order(s);
2587 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2590 * sysctl handler for numa_zonelist_order
2592 int numa_zonelist_order_handler(ctl_table *table, int write,
2593 void __user *buffer, size_t *length,
2596 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2598 static DEFINE_MUTEX(zl_order_mutex);
2600 mutex_lock(&zl_order_mutex);
2602 strcpy(saved_string, (char*)table->data);
2603 ret = proc_dostring(table, write, buffer, length, ppos);
2607 int oldval = user_zonelist_order;
2608 if (__parse_numa_zonelist_order((char*)table->data)) {
2610 * bogus value. restore saved string
2612 strncpy((char*)table->data, saved_string,
2613 NUMA_ZONELIST_ORDER_LEN);
2614 user_zonelist_order = oldval;
2615 } else if (oldval != user_zonelist_order) {
2616 mutex_lock(&zonelists_mutex);
2617 build_all_zonelists(NULL);
2618 mutex_unlock(&zonelists_mutex);
2622 mutex_unlock(&zl_order_mutex);
2627 #define MAX_NODE_LOAD (nr_online_nodes)
2628 static int node_load[MAX_NUMNODES];
2631 * find_next_best_node - find the next node that should appear in a given node's fallback list
2632 * @node: node whose fallback list we're appending
2633 * @used_node_mask: nodemask_t of already used nodes
2635 * We use a number of factors to determine which is the next node that should
2636 * appear on a given node's fallback list. The node should not have appeared
2637 * already in @node's fallback list, and it should be the next closest node
2638 * according to the distance array (which contains arbitrary distance values
2639 * from each node to each node in the system), and should also prefer nodes
2640 * with no CPUs, since presumably they'll have very little allocation pressure
2641 * on them otherwise.
2642 * It returns -1 if no node is found.
2644 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2647 int min_val = INT_MAX;
2649 const struct cpumask *tmp = cpumask_of_node(0);
2651 /* Use the local node if we haven't already */
2652 if (!node_isset(node, *used_node_mask)) {
2653 node_set(node, *used_node_mask);
2657 for_each_node_state(n, N_HIGH_MEMORY) {
2659 /* Don't want a node to appear more than once */
2660 if (node_isset(n, *used_node_mask))
2663 /* Use the distance array to find the distance */
2664 val = node_distance(node, n);
2666 /* Penalize nodes under us ("prefer the next node") */
2669 /* Give preference to headless and unused nodes */
2670 tmp = cpumask_of_node(n);
2671 if (!cpumask_empty(tmp))
2672 val += PENALTY_FOR_NODE_WITH_CPUS;
2674 /* Slight preference for less loaded node */
2675 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2676 val += node_load[n];
2678 if (val < min_val) {
2685 node_set(best_node, *used_node_mask);
2692 * Build zonelists ordered by node and zones within node.
2693 * This results in maximum locality--normal zone overflows into local
2694 * DMA zone, if any--but risks exhausting DMA zone.
2696 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2699 struct zonelist *zonelist;
2701 zonelist = &pgdat->node_zonelists[0];
2702 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2704 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2706 zonelist->_zonerefs[j].zone = NULL;
2707 zonelist->_zonerefs[j].zone_idx = 0;
2711 * Build gfp_thisnode zonelists
2713 static void build_thisnode_zonelists(pg_data_t *pgdat)
2716 struct zonelist *zonelist;
2718 zonelist = &pgdat->node_zonelists[1];
2719 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2720 zonelist->_zonerefs[j].zone = NULL;
2721 zonelist->_zonerefs[j].zone_idx = 0;
2725 * Build zonelists ordered by zone and nodes within zones.
2726 * This results in conserving DMA zone[s] until all Normal memory is
2727 * exhausted, but results in overflowing to remote node while memory
2728 * may still exist in local DMA zone.
2730 static int node_order[MAX_NUMNODES];
2732 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2735 int zone_type; /* needs to be signed */
2737 struct zonelist *zonelist;
2739 zonelist = &pgdat->node_zonelists[0];
2741 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2742 for (j = 0; j < nr_nodes; j++) {
2743 node = node_order[j];
2744 z = &NODE_DATA(node)->node_zones[zone_type];
2745 if (populated_zone(z)) {
2747 &zonelist->_zonerefs[pos++]);
2748 check_highest_zone(zone_type);
2752 zonelist->_zonerefs[pos].zone = NULL;
2753 zonelist->_zonerefs[pos].zone_idx = 0;
2756 static int default_zonelist_order(void)
2759 unsigned long low_kmem_size,total_size;
2763 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2764 * If they are really small and used heavily, the system can fall
2765 * into OOM very easily.
2766 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2768 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2771 for_each_online_node(nid) {
2772 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2773 z = &NODE_DATA(nid)->node_zones[zone_type];
2774 if (populated_zone(z)) {
2775 if (zone_type < ZONE_NORMAL)
2776 low_kmem_size += z->present_pages;
2777 total_size += z->present_pages;
2778 } else if (zone_type == ZONE_NORMAL) {
2780 * If any node has only lowmem, then node order
2781 * is preferred to allow kernel allocations
2782 * locally; otherwise, they can easily infringe
2783 * on other nodes when there is an abundance of
2784 * lowmem available to allocate from.
2786 return ZONELIST_ORDER_NODE;
2790 if (!low_kmem_size || /* there are no DMA area. */
2791 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2792 return ZONELIST_ORDER_NODE;
2794 * look into each node's config.
2795 * If there is a node whose DMA/DMA32 memory is very big area on
2796 * local memory, NODE_ORDER may be suitable.
2798 average_size = total_size /
2799 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2800 for_each_online_node(nid) {
2803 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2804 z = &NODE_DATA(nid)->node_zones[zone_type];
2805 if (populated_zone(z)) {
2806 if (zone_type < ZONE_NORMAL)
2807 low_kmem_size += z->present_pages;
2808 total_size += z->present_pages;
2811 if (low_kmem_size &&
2812 total_size > average_size && /* ignore small node */
2813 low_kmem_size > total_size * 70/100)
2814 return ZONELIST_ORDER_NODE;
2816 return ZONELIST_ORDER_ZONE;
2819 static void set_zonelist_order(void)
2821 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2822 current_zonelist_order = default_zonelist_order();
2824 current_zonelist_order = user_zonelist_order;
2827 static void build_zonelists(pg_data_t *pgdat)
2831 nodemask_t used_mask;
2832 int local_node, prev_node;
2833 struct zonelist *zonelist;
2834 int order = current_zonelist_order;
2836 /* initialize zonelists */
2837 for (i = 0; i < MAX_ZONELISTS; i++) {
2838 zonelist = pgdat->node_zonelists + i;
2839 zonelist->_zonerefs[0].zone = NULL;
2840 zonelist->_zonerefs[0].zone_idx = 0;
2843 /* NUMA-aware ordering of nodes */
2844 local_node = pgdat->node_id;
2845 load = nr_online_nodes;
2846 prev_node = local_node;
2847 nodes_clear(used_mask);
2849 memset(node_order, 0, sizeof(node_order));
2852 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2853 int distance = node_distance(local_node, node);
2856 * If another node is sufficiently far away then it is better
2857 * to reclaim pages in a zone before going off node.
2859 if (distance > RECLAIM_DISTANCE)
2860 zone_reclaim_mode = 1;
2863 * We don't want to pressure a particular node.
2864 * So adding penalty to the first node in same
2865 * distance group to make it round-robin.
2867 if (distance != node_distance(local_node, prev_node))
2868 node_load[node] = load;
2872 if (order == ZONELIST_ORDER_NODE)
2873 build_zonelists_in_node_order(pgdat, node);
2875 node_order[j++] = node; /* remember order */
2878 if (order == ZONELIST_ORDER_ZONE) {
2879 /* calculate node order -- i.e., DMA last! */
2880 build_zonelists_in_zone_order(pgdat, j);
2883 build_thisnode_zonelists(pgdat);
2886 /* Construct the zonelist performance cache - see further mmzone.h */
2887 static void build_zonelist_cache(pg_data_t *pgdat)
2889 struct zonelist *zonelist;
2890 struct zonelist_cache *zlc;
2893 zonelist = &pgdat->node_zonelists[0];
2894 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2895 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2896 for (z = zonelist->_zonerefs; z->zone; z++)
2897 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2900 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
2902 * Return node id of node used for "local" allocations.
2903 * I.e., first node id of first zone in arg node's generic zonelist.
2904 * Used for initializing percpu 'numa_mem', which is used primarily
2905 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
2907 int local_memory_node(int node)
2911 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
2912 gfp_zone(GFP_KERNEL),
2919 #else /* CONFIG_NUMA */
2921 static void set_zonelist_order(void)
2923 current_zonelist_order = ZONELIST_ORDER_ZONE;
2926 static void build_zonelists(pg_data_t *pgdat)
2928 int node, local_node;
2930 struct zonelist *zonelist;
2932 local_node = pgdat->node_id;
2934 zonelist = &pgdat->node_zonelists[0];
2935 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2938 * Now we build the zonelist so that it contains the zones
2939 * of all the other nodes.
2940 * We don't want to pressure a particular node, so when
2941 * building the zones for node N, we make sure that the
2942 * zones coming right after the local ones are those from
2943 * node N+1 (modulo N)
2945 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2946 if (!node_online(node))
2948 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2951 for (node = 0; node < local_node; node++) {
2952 if (!node_online(node))
2954 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2958 zonelist->_zonerefs[j].zone = NULL;
2959 zonelist->_zonerefs[j].zone_idx = 0;
2962 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2963 static void build_zonelist_cache(pg_data_t *pgdat)
2965 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2968 #endif /* CONFIG_NUMA */
2971 * Boot pageset table. One per cpu which is going to be used for all
2972 * zones and all nodes. The parameters will be set in such a way
2973 * that an item put on a list will immediately be handed over to
2974 * the buddy list. This is safe since pageset manipulation is done
2975 * with interrupts disabled.
2977 * The boot_pagesets must be kept even after bootup is complete for
2978 * unused processors and/or zones. They do play a role for bootstrapping
2979 * hotplugged processors.
2981 * zoneinfo_show() and maybe other functions do
2982 * not check if the processor is online before following the pageset pointer.
2983 * Other parts of the kernel may not check if the zone is available.
2985 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
2986 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
2987 static void setup_zone_pageset(struct zone *zone);
2990 * Global mutex to protect against size modification of zonelists
2991 * as well as to serialize pageset setup for the new populated zone.
2993 DEFINE_MUTEX(zonelists_mutex);
2995 /* return values int ....just for stop_machine() */
2996 static __init_refok int __build_all_zonelists(void *data)
3002 memset(node_load, 0, sizeof(node_load));
3004 for_each_online_node(nid) {
3005 pg_data_t *pgdat = NODE_DATA(nid);
3007 build_zonelists(pgdat);
3008 build_zonelist_cache(pgdat);
3011 #ifdef CONFIG_MEMORY_HOTPLUG
3012 /* Setup real pagesets for the new zone */
3014 struct zone *zone = data;
3015 setup_zone_pageset(zone);
3020 * Initialize the boot_pagesets that are going to be used
3021 * for bootstrapping processors. The real pagesets for
3022 * each zone will be allocated later when the per cpu
3023 * allocator is available.
3025 * boot_pagesets are used also for bootstrapping offline
3026 * cpus if the system is already booted because the pagesets
3027 * are needed to initialize allocators on a specific cpu too.
3028 * F.e. the percpu allocator needs the page allocator which
3029 * needs the percpu allocator in order to allocate its pagesets
3030 * (a chicken-egg dilemma).
3032 for_each_possible_cpu(cpu) {
3033 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3035 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3037 * We now know the "local memory node" for each node--
3038 * i.e., the node of the first zone in the generic zonelist.
3039 * Set up numa_mem percpu variable for on-line cpus. During
3040 * boot, only the boot cpu should be on-line; we'll init the
3041 * secondary cpus' numa_mem as they come on-line. During
3042 * node/memory hotplug, we'll fixup all on-line cpus.
3044 if (cpu_online(cpu))
3045 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3053 * Called with zonelists_mutex held always
3054 * unless system_state == SYSTEM_BOOTING.
3056 void build_all_zonelists(void *data)
3058 set_zonelist_order();
3060 if (system_state == SYSTEM_BOOTING) {
3061 __build_all_zonelists(NULL);
3062 mminit_verify_zonelist();
3063 cpuset_init_current_mems_allowed();
3065 /* we have to stop all cpus to guarantee there is no user
3067 stop_machine(__build_all_zonelists, data, NULL);
3068 /* cpuset refresh routine should be here */
3070 vm_total_pages = nr_free_pagecache_pages();
3072 * Disable grouping by mobility if the number of pages in the
3073 * system is too low to allow the mechanism to work. It would be
3074 * more accurate, but expensive to check per-zone. This check is
3075 * made on memory-hotadd so a system can start with mobility
3076 * disabled and enable it later
3078 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3079 page_group_by_mobility_disabled = 1;
3081 page_group_by_mobility_disabled = 0;
3083 printk("Built %i zonelists in %s order, mobility grouping %s. "
3084 "Total pages: %ld\n",
3086 zonelist_order_name[current_zonelist_order],
3087 page_group_by_mobility_disabled ? "off" : "on",
3090 printk("Policy zone: %s\n", zone_names[policy_zone]);
3095 * Helper functions to size the waitqueue hash table.
3096 * Essentially these want to choose hash table sizes sufficiently
3097 * large so that collisions trying to wait on pages are rare.
3098 * But in fact, the number of active page waitqueues on typical
3099 * systems is ridiculously low, less than 200. So this is even
3100 * conservative, even though it seems large.
3102 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3103 * waitqueues, i.e. the size of the waitq table given the number of pages.
3105 #define PAGES_PER_WAITQUEUE 256
3107 #ifndef CONFIG_MEMORY_HOTPLUG
3108 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3110 unsigned long size = 1;
3112 pages /= PAGES_PER_WAITQUEUE;
3114 while (size < pages)
3118 * Once we have dozens or even hundreds of threads sleeping
3119 * on IO we've got bigger problems than wait queue collision.
3120 * Limit the size of the wait table to a reasonable size.
3122 size = min(size, 4096UL);
3124 return max(size, 4UL);
3128 * A zone's size might be changed by hot-add, so it is not possible to determine
3129 * a suitable size for its wait_table. So we use the maximum size now.
3131 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3133 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3134 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3135 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3137 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3138 * or more by the traditional way. (See above). It equals:
3140 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3141 * ia64(16K page size) : = ( 8G + 4M)byte.
3142 * powerpc (64K page size) : = (32G +16M)byte.
3144 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3151 * This is an integer logarithm so that shifts can be used later
3152 * to extract the more random high bits from the multiplicative
3153 * hash function before the remainder is taken.
3155 static inline unsigned long wait_table_bits(unsigned long size)
3160 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3163 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3164 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3165 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3166 * higher will lead to a bigger reserve which will get freed as contiguous
3167 * blocks as reclaim kicks in
3169 static void setup_zone_migrate_reserve(struct zone *zone)
3171 unsigned long start_pfn, pfn, end_pfn;
3173 unsigned long block_migratetype;
3176 /* Get the start pfn, end pfn and the number of blocks to reserve */
3177 start_pfn = zone->zone_start_pfn;
3178 end_pfn = start_pfn + zone->spanned_pages;
3179 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3183 * Reserve blocks are generally in place to help high-order atomic
3184 * allocations that are short-lived. A min_free_kbytes value that
3185 * would result in more than 2 reserve blocks for atomic allocations
3186 * is assumed to be in place to help anti-fragmentation for the
3187 * future allocation of hugepages at runtime.
3189 reserve = min(2, reserve);
3191 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3192 if (!pfn_valid(pfn))
3194 page = pfn_to_page(pfn);
3196 /* Watch out for overlapping nodes */
3197 if (page_to_nid(page) != zone_to_nid(zone))
3200 /* Blocks with reserved pages will never free, skip them. */
3201 if (PageReserved(page))
3204 block_migratetype = get_pageblock_migratetype(page);
3206 /* If this block is reserved, account for it */
3207 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3212 /* Suitable for reserving if this block is movable */
3213 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3214 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3215 move_freepages_block(zone, page, MIGRATE_RESERVE);
3221 * If the reserve is met and this is a previous reserved block,
3224 if (block_migratetype == MIGRATE_RESERVE) {
3225 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3226 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3232 * Initially all pages are reserved - free ones are freed
3233 * up by free_all_bootmem() once the early boot process is
3234 * done. Non-atomic initialization, single-pass.
3236 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3237 unsigned long start_pfn, enum memmap_context context)
3240 unsigned long end_pfn = start_pfn + size;
3244 if (highest_memmap_pfn < end_pfn - 1)
3245 highest_memmap_pfn = end_pfn - 1;
3247 z = &NODE_DATA(nid)->node_zones[zone];
3248 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3250 * There can be holes in boot-time mem_map[]s
3251 * handed to this function. They do not
3252 * exist on hotplugged memory.
3254 if (context == MEMMAP_EARLY) {
3255 if (!early_pfn_valid(pfn))
3257 if (!early_pfn_in_nid(pfn, nid))
3260 page = pfn_to_page(pfn);
3261 set_page_links(page, zone, nid, pfn);
3262 mminit_verify_page_links(page, zone, nid, pfn);
3263 init_page_count(page);
3264 reset_page_mapcount(page);
3265 SetPageReserved(page);
3267 * Mark the block movable so that blocks are reserved for
3268 * movable at startup. This will force kernel allocations
3269 * to reserve their blocks rather than leaking throughout
3270 * the address space during boot when many long-lived
3271 * kernel allocations are made. Later some blocks near
3272 * the start are marked MIGRATE_RESERVE by
3273 * setup_zone_migrate_reserve()
3275 * bitmap is created for zone's valid pfn range. but memmap
3276 * can be created for invalid pages (for alignment)
3277 * check here not to call set_pageblock_migratetype() against
3280 if ((z->zone_start_pfn <= pfn)
3281 && (pfn < z->zone_start_pfn + z->spanned_pages)
3282 && !(pfn & (pageblock_nr_pages - 1)))
3283 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3285 INIT_LIST_HEAD(&page->lru);
3286 #ifdef WANT_PAGE_VIRTUAL
3287 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3288 if (!is_highmem_idx(zone))
3289 set_page_address(page, __va(pfn << PAGE_SHIFT));
3294 static void __meminit zone_init_free_lists(struct zone *zone)
3297 for_each_migratetype_order(order, t) {
3298 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3299 zone->free_area[order].nr_free = 0;
3303 #ifndef __HAVE_ARCH_MEMMAP_INIT
3304 #define memmap_init(size, nid, zone, start_pfn) \
3305 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3308 static int zone_batchsize(struct zone *zone)
3314 * The per-cpu-pages pools are set to around 1000th of the
3315 * size of the zone. But no more than 1/2 of a meg.
3317 * OK, so we don't know how big the cache is. So guess.
3319 batch = zone->present_pages / 1024;
3320 if (batch * PAGE_SIZE > 512 * 1024)
3321 batch = (512 * 1024) / PAGE_SIZE;
3322 batch /= 4; /* We effectively *= 4 below */
3327 * Clamp the batch to a 2^n - 1 value. Having a power
3328 * of 2 value was found to be more likely to have
3329 * suboptimal cache aliasing properties in some cases.
3331 * For example if 2 tasks are alternately allocating
3332 * batches of pages, one task can end up with a lot
3333 * of pages of one half of the possible page colors
3334 * and the other with pages of the other colors.
3336 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3341 /* The deferral and batching of frees should be suppressed under NOMMU
3344 * The problem is that NOMMU needs to be able to allocate large chunks
3345 * of contiguous memory as there's no hardware page translation to
3346 * assemble apparent contiguous memory from discontiguous pages.
3348 * Queueing large contiguous runs of pages for batching, however,
3349 * causes the pages to actually be freed in smaller chunks. As there
3350 * can be a significant delay between the individual batches being
3351 * recycled, this leads to the once large chunks of space being
3352 * fragmented and becoming unavailable for high-order allocations.
3358 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3360 struct per_cpu_pages *pcp;
3363 memset(p, 0, sizeof(*p));
3367 pcp->high = 6 * batch;
3368 pcp->batch = max(1UL, 1 * batch);
3369 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3370 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3374 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3375 * to the value high for the pageset p.
3378 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3381 struct per_cpu_pages *pcp;
3385 pcp->batch = max(1UL, high/4);
3386 if ((high/4) > (PAGE_SHIFT * 8))
3387 pcp->batch = PAGE_SHIFT * 8;
3390 static __meminit void setup_zone_pageset(struct zone *zone)
3394 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3396 for_each_possible_cpu(cpu) {
3397 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3399 setup_pageset(pcp, zone_batchsize(zone));
3401 if (percpu_pagelist_fraction)
3402 setup_pagelist_highmark(pcp,
3403 (zone->present_pages /
3404 percpu_pagelist_fraction));
3409 * Allocate per cpu pagesets and initialize them.
3410 * Before this call only boot pagesets were available.
3412 void __init setup_per_cpu_pageset(void)
3416 for_each_populated_zone(zone)
3417 setup_zone_pageset(zone);
3420 static noinline __init_refok
3421 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3424 struct pglist_data *pgdat = zone->zone_pgdat;
3428 * The per-page waitqueue mechanism uses hashed waitqueues
3431 zone->wait_table_hash_nr_entries =
3432 wait_table_hash_nr_entries(zone_size_pages);
3433 zone->wait_table_bits =
3434 wait_table_bits(zone->wait_table_hash_nr_entries);
3435 alloc_size = zone->wait_table_hash_nr_entries
3436 * sizeof(wait_queue_head_t);
3438 if (!slab_is_available()) {
3439 zone->wait_table = (wait_queue_head_t *)
3440 alloc_bootmem_node(pgdat, alloc_size);
3443 * This case means that a zone whose size was 0 gets new memory
3444 * via memory hot-add.
3445 * But it may be the case that a new node was hot-added. In
3446 * this case vmalloc() will not be able to use this new node's
3447 * memory - this wait_table must be initialized to use this new
3448 * node itself as well.
3449 * To use this new node's memory, further consideration will be
3452 zone->wait_table = vmalloc(alloc_size);
3454 if (!zone->wait_table)
3457 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3458 init_waitqueue_head(zone->wait_table + i);
3463 static int __zone_pcp_update(void *data)
3465 struct zone *zone = data;
3467 unsigned long batch = zone_batchsize(zone), flags;
3469 for_each_possible_cpu(cpu) {
3470 struct per_cpu_pageset *pset;
3471 struct per_cpu_pages *pcp;
3473 pset = per_cpu_ptr(zone->pageset, cpu);
3476 local_irq_save(flags);
3477 free_pcppages_bulk(zone, pcp->count, pcp);
3478 setup_pageset(pset, batch);
3479 local_irq_restore(flags);
3484 void zone_pcp_update(struct zone *zone)
3486 stop_machine(__zone_pcp_update, zone, NULL);
3489 static __meminit void zone_pcp_init(struct zone *zone)
3492 * per cpu subsystem is not up at this point. The following code
3493 * relies on the ability of the linker to provide the
3494 * offset of a (static) per cpu variable into the per cpu area.
3496 zone->pageset = &boot_pageset;
3498 if (zone->present_pages)
3499 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3500 zone->name, zone->present_pages,
3501 zone_batchsize(zone));
3504 __meminit int init_currently_empty_zone(struct zone *zone,
3505 unsigned long zone_start_pfn,
3507 enum memmap_context context)
3509 struct pglist_data *pgdat = zone->zone_pgdat;
3511 ret = zone_wait_table_init(zone, size);
3514 pgdat->nr_zones = zone_idx(zone) + 1;
3516 zone->zone_start_pfn = zone_start_pfn;
3518 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3519 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3521 (unsigned long)zone_idx(zone),
3522 zone_start_pfn, (zone_start_pfn + size));
3524 zone_init_free_lists(zone);
3529 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3531 * Basic iterator support. Return the first range of PFNs for a node
3532 * Note: nid == MAX_NUMNODES returns first region regardless of node
3534 static int __meminit first_active_region_index_in_nid(int nid)
3538 for (i = 0; i < nr_nodemap_entries; i++)
3539 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3546 * Basic iterator support. Return the next active range of PFNs for a node
3547 * Note: nid == MAX_NUMNODES returns next region regardless of node
3549 static int __meminit next_active_region_index_in_nid(int index, int nid)
3551 for (index = index + 1; index < nr_nodemap_entries; index++)
3552 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3558 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3560 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3561 * Architectures may implement their own version but if add_active_range()
3562 * was used and there are no special requirements, this is a convenient
3565 int __meminit __early_pfn_to_nid(unsigned long pfn)
3569 for (i = 0; i < nr_nodemap_entries; i++) {
3570 unsigned long start_pfn = early_node_map[i].start_pfn;
3571 unsigned long end_pfn = early_node_map[i].end_pfn;
3573 if (start_pfn <= pfn && pfn < end_pfn)
3574 return early_node_map[i].nid;
3576 /* This is a memory hole */
3579 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3581 int __meminit early_pfn_to_nid(unsigned long pfn)
3585 nid = __early_pfn_to_nid(pfn);
3588 /* just returns 0 */
3592 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3593 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3597 nid = __early_pfn_to_nid(pfn);
3598 if (nid >= 0 && nid != node)
3604 /* Basic iterator support to walk early_node_map[] */
3605 #define for_each_active_range_index_in_nid(i, nid) \
3606 for (i = first_active_region_index_in_nid(nid); i != -1; \
3607 i = next_active_region_index_in_nid(i, nid))
3610 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3611 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3612 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3614 * If an architecture guarantees that all ranges registered with
3615 * add_active_ranges() contain no holes and may be freed, this
3616 * this function may be used instead of calling free_bootmem() manually.
3618 void __init free_bootmem_with_active_regions(int nid,
3619 unsigned long max_low_pfn)
3623 for_each_active_range_index_in_nid(i, nid) {
3624 unsigned long size_pages = 0;
3625 unsigned long end_pfn = early_node_map[i].end_pfn;
3627 if (early_node_map[i].start_pfn >= max_low_pfn)
3630 if (end_pfn > max_low_pfn)
3631 end_pfn = max_low_pfn;
3633 size_pages = end_pfn - early_node_map[i].start_pfn;
3634 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3635 PFN_PHYS(early_node_map[i].start_pfn),
3636 size_pages << PAGE_SHIFT);
3640 #ifdef CONFIG_HAVE_MEMBLOCK
3641 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3642 u64 goal, u64 limit)
3646 /* Need to go over early_node_map to find out good range for node */
3647 for_each_active_range_index_in_nid(i, nid) {
3649 u64 ei_start, ei_last;
3650 u64 final_start, final_end;
3652 ei_last = early_node_map[i].end_pfn;
3653 ei_last <<= PAGE_SHIFT;
3654 ei_start = early_node_map[i].start_pfn;
3655 ei_start <<= PAGE_SHIFT;
3657 final_start = max(ei_start, goal);
3658 final_end = min(ei_last, limit);
3660 if (final_start >= final_end)
3663 addr = memblock_find_in_range(final_start, final_end, size, align);
3665 if (addr == MEMBLOCK_ERROR)
3671 return MEMBLOCK_ERROR;
3675 int __init add_from_early_node_map(struct range *range, int az,
3676 int nr_range, int nid)
3681 /* need to go over early_node_map to find out good range for node */
3682 for_each_active_range_index_in_nid(i, nid) {
3683 start = early_node_map[i].start_pfn;
3684 end = early_node_map[i].end_pfn;
3685 nr_range = add_range(range, az, nr_range, start, end);
3690 #ifdef CONFIG_NO_BOOTMEM
3691 void * __init __alloc_memory_core_early(int nid, u64 size, u64 align,
3692 u64 goal, u64 limit)
3697 if (limit > memblock.current_limit)
3698 limit = memblock.current_limit;
3700 addr = find_memory_core_early(nid, size, align, goal, limit);
3702 if (addr == MEMBLOCK_ERROR)
3705 ptr = phys_to_virt(addr);
3706 memset(ptr, 0, size);
3707 memblock_x86_reserve_range(addr, addr + size, "BOOTMEM");
3709 * The min_count is set to 0 so that bootmem allocated blocks
3710 * are never reported as leaks.
3712 kmemleak_alloc(ptr, size, 0, 0);
3718 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3723 for_each_active_range_index_in_nid(i, nid) {
3724 ret = work_fn(early_node_map[i].start_pfn,
3725 early_node_map[i].end_pfn, data);
3731 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3732 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3734 * If an architecture guarantees that all ranges registered with
3735 * add_active_ranges() contain no holes and may be freed, this
3736 * function may be used instead of calling memory_present() manually.
3738 void __init sparse_memory_present_with_active_regions(int nid)
3742 for_each_active_range_index_in_nid(i, nid)
3743 memory_present(early_node_map[i].nid,
3744 early_node_map[i].start_pfn,
3745 early_node_map[i].end_pfn);
3749 * get_pfn_range_for_nid - Return the start and end page frames for a node
3750 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3751 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3752 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3754 * It returns the start and end page frame of a node based on information
3755 * provided by an arch calling add_active_range(). If called for a node
3756 * with no available memory, a warning is printed and the start and end
3759 void __meminit get_pfn_range_for_nid(unsigned int nid,
3760 unsigned long *start_pfn, unsigned long *end_pfn)
3766 for_each_active_range_index_in_nid(i, nid) {
3767 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3768 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3771 if (*start_pfn == -1UL)
3776 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3777 * assumption is made that zones within a node are ordered in monotonic
3778 * increasing memory addresses so that the "highest" populated zone is used
3780 static void __init find_usable_zone_for_movable(void)
3783 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3784 if (zone_index == ZONE_MOVABLE)
3787 if (arch_zone_highest_possible_pfn[zone_index] >
3788 arch_zone_lowest_possible_pfn[zone_index])
3792 VM_BUG_ON(zone_index == -1);
3793 movable_zone = zone_index;
3797 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3798 * because it is sized independant of architecture. Unlike the other zones,
3799 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3800 * in each node depending on the size of each node and how evenly kernelcore
3801 * is distributed. This helper function adjusts the zone ranges
3802 * provided by the architecture for a given node by using the end of the
3803 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3804 * zones within a node are in order of monotonic increases memory addresses
3806 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3807 unsigned long zone_type,
3808 unsigned long node_start_pfn,
3809 unsigned long node_end_pfn,
3810 unsigned long *zone_start_pfn,
3811 unsigned long *zone_end_pfn)
3813 /* Only adjust if ZONE_MOVABLE is on this node */
3814 if (zone_movable_pfn[nid]) {
3815 /* Size ZONE_MOVABLE */
3816 if (zone_type == ZONE_MOVABLE) {
3817 *zone_start_pfn = zone_movable_pfn[nid];
3818 *zone_end_pfn = min(node_end_pfn,
3819 arch_zone_highest_possible_pfn[movable_zone]);
3821 /* Adjust for ZONE_MOVABLE starting within this range */
3822 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3823 *zone_end_pfn > zone_movable_pfn[nid]) {
3824 *zone_end_pfn = zone_movable_pfn[nid];
3826 /* Check if this whole range is within ZONE_MOVABLE */
3827 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3828 *zone_start_pfn = *zone_end_pfn;
3833 * Return the number of pages a zone spans in a node, including holes
3834 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3836 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3837 unsigned long zone_type,
3838 unsigned long *ignored)
3840 unsigned long node_start_pfn, node_end_pfn;
3841 unsigned long zone_start_pfn, zone_end_pfn;
3843 /* Get the start and end of the node and zone */
3844 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3845 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3846 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3847 adjust_zone_range_for_zone_movable(nid, zone_type,
3848 node_start_pfn, node_end_pfn,
3849 &zone_start_pfn, &zone_end_pfn);
3851 /* Check that this node has pages within the zone's required range */
3852 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3855 /* Move the zone boundaries inside the node if necessary */
3856 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3857 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3859 /* Return the spanned pages */
3860 return zone_end_pfn - zone_start_pfn;
3864 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3865 * then all holes in the requested range will be accounted for.
3867 unsigned long __meminit __absent_pages_in_range(int nid,
3868 unsigned long range_start_pfn,
3869 unsigned long range_end_pfn)
3872 unsigned long prev_end_pfn = 0, hole_pages = 0;
3873 unsigned long start_pfn;
3875 /* Find the end_pfn of the first active range of pfns in the node */
3876 i = first_active_region_index_in_nid(nid);
3880 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3882 /* Account for ranges before physical memory on this node */
3883 if (early_node_map[i].start_pfn > range_start_pfn)
3884 hole_pages = prev_end_pfn - range_start_pfn;
3886 /* Find all holes for the zone within the node */
3887 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3889 /* No need to continue if prev_end_pfn is outside the zone */
3890 if (prev_end_pfn >= range_end_pfn)
3893 /* Make sure the end of the zone is not within the hole */
3894 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3895 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3897 /* Update the hole size cound and move on */
3898 if (start_pfn > range_start_pfn) {
3899 BUG_ON(prev_end_pfn > start_pfn);
3900 hole_pages += start_pfn - prev_end_pfn;
3902 prev_end_pfn = early_node_map[i].end_pfn;
3905 /* Account for ranges past physical memory on this node */
3906 if (range_end_pfn > prev_end_pfn)
3907 hole_pages += range_end_pfn -
3908 max(range_start_pfn, prev_end_pfn);
3914 * absent_pages_in_range - Return number of page frames in holes within a range
3915 * @start_pfn: The start PFN to start searching for holes
3916 * @end_pfn: The end PFN to stop searching for holes
3918 * It returns the number of pages frames in memory holes within a range.
3920 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3921 unsigned long end_pfn)
3923 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3926 /* Return the number of page frames in holes in a zone on a node */
3927 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3928 unsigned long zone_type,
3929 unsigned long *ignored)
3931 unsigned long node_start_pfn, node_end_pfn;
3932 unsigned long zone_start_pfn, zone_end_pfn;
3934 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3935 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3937 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3940 adjust_zone_range_for_zone_movable(nid, zone_type,
3941 node_start_pfn, node_end_pfn,
3942 &zone_start_pfn, &zone_end_pfn);
3943 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3947 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3948 unsigned long zone_type,
3949 unsigned long *zones_size)
3951 return zones_size[zone_type];
3954 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3955 unsigned long zone_type,
3956 unsigned long *zholes_size)
3961 return zholes_size[zone_type];
3966 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3967 unsigned long *zones_size, unsigned long *zholes_size)
3969 unsigned long realtotalpages, totalpages = 0;
3972 for (i = 0; i < MAX_NR_ZONES; i++)
3973 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3975 pgdat->node_spanned_pages = totalpages;
3977 realtotalpages = totalpages;
3978 for (i = 0; i < MAX_NR_ZONES; i++)
3980 zone_absent_pages_in_node(pgdat->node_id, i,
3982 pgdat->node_present_pages = realtotalpages;
3983 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3987 #ifndef CONFIG_SPARSEMEM
3989 * Calculate the size of the zone->blockflags rounded to an unsigned long
3990 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3991 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3992 * round what is now in bits to nearest long in bits, then return it in
3995 static unsigned long __init usemap_size(unsigned long zonesize)
3997 unsigned long usemapsize;
3999 usemapsize = roundup(zonesize, pageblock_nr_pages);
4000 usemapsize = usemapsize >> pageblock_order;
4001 usemapsize *= NR_PAGEBLOCK_BITS;
4002 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4004 return usemapsize / 8;
4007 static void __init setup_usemap(struct pglist_data *pgdat,
4008 struct zone *zone, unsigned long zonesize)
4010 unsigned long usemapsize = usemap_size(zonesize);
4011 zone->pageblock_flags = NULL;
4013 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
4016 static void inline setup_usemap(struct pglist_data *pgdat,
4017 struct zone *zone, unsigned long zonesize) {}
4018 #endif /* CONFIG_SPARSEMEM */
4020 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4022 /* Return a sensible default order for the pageblock size. */
4023 static inline int pageblock_default_order(void)
4025 if (HPAGE_SHIFT > PAGE_SHIFT)
4026 return HUGETLB_PAGE_ORDER;
4031 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4032 static inline void __init set_pageblock_order(unsigned int order)
4034 /* Check that pageblock_nr_pages has not already been setup */
4035 if (pageblock_order)
4039 * Assume the largest contiguous order of interest is a huge page.
4040 * This value may be variable depending on boot parameters on IA64
4042 pageblock_order = order;
4044 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4047 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4048 * and pageblock_default_order() are unused as pageblock_order is set
4049 * at compile-time. See include/linux/pageblock-flags.h for the values of
4050 * pageblock_order based on the kernel config
4052 static inline int pageblock_default_order(unsigned int order)
4056 #define set_pageblock_order(x) do {} while (0)
4058 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4061 * Set up the zone data structures:
4062 * - mark all pages reserved
4063 * - mark all memory queues empty
4064 * - clear the memory bitmaps
4066 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4067 unsigned long *zones_size, unsigned long *zholes_size)
4070 int nid = pgdat->node_id;
4071 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4074 pgdat_resize_init(pgdat);
4075 pgdat->nr_zones = 0;
4076 init_waitqueue_head(&pgdat->kswapd_wait);
4077 pgdat->kswapd_max_order = 0;
4078 pgdat_page_cgroup_init(pgdat);
4080 for (j = 0; j < MAX_NR_ZONES; j++) {
4081 struct zone *zone = pgdat->node_zones + j;
4082 unsigned long size, realsize, memmap_pages;
4085 size = zone_spanned_pages_in_node(nid, j, zones_size);
4086 realsize = size - zone_absent_pages_in_node(nid, j,
4090 * Adjust realsize so that it accounts for how much memory
4091 * is used by this zone for memmap. This affects the watermark
4092 * and per-cpu initialisations
4095 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4096 if (realsize >= memmap_pages) {
4097 realsize -= memmap_pages;
4100 " %s zone: %lu pages used for memmap\n",
4101 zone_names[j], memmap_pages);
4104 " %s zone: %lu pages exceeds realsize %lu\n",
4105 zone_names[j], memmap_pages, realsize);
4107 /* Account for reserved pages */
4108 if (j == 0 && realsize > dma_reserve) {
4109 realsize -= dma_reserve;
4110 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4111 zone_names[0], dma_reserve);
4114 if (!is_highmem_idx(j))
4115 nr_kernel_pages += realsize;
4116 nr_all_pages += realsize;
4118 zone->spanned_pages = size;
4119 zone->present_pages = realsize;
4122 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4124 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4126 zone->name = zone_names[j];
4127 spin_lock_init(&zone->lock);
4128 spin_lock_init(&zone->lru_lock);
4129 zone_seqlock_init(zone);
4130 zone->zone_pgdat = pgdat;
4132 zone_pcp_init(zone);
4134 INIT_LIST_HEAD(&zone->lru[l].list);
4135 zone->reclaim_stat.nr_saved_scan[l] = 0;
4137 zone->reclaim_stat.recent_rotated[0] = 0;
4138 zone->reclaim_stat.recent_rotated[1] = 0;
4139 zone->reclaim_stat.recent_scanned[0] = 0;
4140 zone->reclaim_stat.recent_scanned[1] = 0;
4141 zap_zone_vm_stats(zone);
4146 set_pageblock_order(pageblock_default_order());
4147 setup_usemap(pgdat, zone, size);
4148 ret = init_currently_empty_zone(zone, zone_start_pfn,
4149 size, MEMMAP_EARLY);
4151 memmap_init(size, nid, j, zone_start_pfn);
4152 zone_start_pfn += size;
4156 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4158 /* Skip empty nodes */
4159 if (!pgdat->node_spanned_pages)
4162 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4163 /* ia64 gets its own node_mem_map, before this, without bootmem */
4164 if (!pgdat->node_mem_map) {
4165 unsigned long size, start, end;
4169 * The zone's endpoints aren't required to be MAX_ORDER
4170 * aligned but the node_mem_map endpoints must be in order
4171 * for the buddy allocator to function correctly.
4173 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4174 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4175 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4176 size = (end - start) * sizeof(struct page);
4177 map = alloc_remap(pgdat->node_id, size);
4179 map = alloc_bootmem_node(pgdat, size);
4180 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4182 #ifndef CONFIG_NEED_MULTIPLE_NODES
4184 * With no DISCONTIG, the global mem_map is just set as node 0's
4186 if (pgdat == NODE_DATA(0)) {
4187 mem_map = NODE_DATA(0)->node_mem_map;
4188 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4189 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4190 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4191 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4194 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4197 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4198 unsigned long node_start_pfn, unsigned long *zholes_size)
4200 pg_data_t *pgdat = NODE_DATA(nid);
4202 pgdat->node_id = nid;
4203 pgdat->node_start_pfn = node_start_pfn;
4204 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4206 alloc_node_mem_map(pgdat);
4207 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4208 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4209 nid, (unsigned long)pgdat,
4210 (unsigned long)pgdat->node_mem_map);
4213 free_area_init_core(pgdat, zones_size, zholes_size);
4216 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4218 #if MAX_NUMNODES > 1
4220 * Figure out the number of possible node ids.
4222 static void __init setup_nr_node_ids(void)
4225 unsigned int highest = 0;
4227 for_each_node_mask(node, node_possible_map)
4229 nr_node_ids = highest + 1;
4232 static inline void setup_nr_node_ids(void)
4238 * add_active_range - Register a range of PFNs backed by physical memory
4239 * @nid: The node ID the range resides on
4240 * @start_pfn: The start PFN of the available physical memory
4241 * @end_pfn: The end PFN of the available physical memory
4243 * These ranges are stored in an early_node_map[] and later used by
4244 * free_area_init_nodes() to calculate zone sizes and holes. If the
4245 * range spans a memory hole, it is up to the architecture to ensure
4246 * the memory is not freed by the bootmem allocator. If possible
4247 * the range being registered will be merged with existing ranges.
4249 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4250 unsigned long end_pfn)
4254 mminit_dprintk(MMINIT_TRACE, "memory_register",
4255 "Entering add_active_range(%d, %#lx, %#lx) "
4256 "%d entries of %d used\n",
4257 nid, start_pfn, end_pfn,
4258 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4260 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4262 /* Merge with existing active regions if possible */
4263 for (i = 0; i < nr_nodemap_entries; i++) {
4264 if (early_node_map[i].nid != nid)
4267 /* Skip if an existing region covers this new one */
4268 if (start_pfn >= early_node_map[i].start_pfn &&
4269 end_pfn <= early_node_map[i].end_pfn)
4272 /* Merge forward if suitable */
4273 if (start_pfn <= early_node_map[i].end_pfn &&
4274 end_pfn > early_node_map[i].end_pfn) {
4275 early_node_map[i].end_pfn = end_pfn;
4279 /* Merge backward if suitable */
4280 if (start_pfn < early_node_map[i].start_pfn &&
4281 end_pfn >= early_node_map[i].start_pfn) {
4282 early_node_map[i].start_pfn = start_pfn;
4287 /* Check that early_node_map is large enough */
4288 if (i >= MAX_ACTIVE_REGIONS) {
4289 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4290 MAX_ACTIVE_REGIONS);
4294 early_node_map[i].nid = nid;
4295 early_node_map[i].start_pfn = start_pfn;
4296 early_node_map[i].end_pfn = end_pfn;
4297 nr_nodemap_entries = i + 1;
4301 * remove_active_range - Shrink an existing registered range of PFNs
4302 * @nid: The node id the range is on that should be shrunk
4303 * @start_pfn: The new PFN of the range
4304 * @end_pfn: The new PFN of the range
4306 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4307 * The map is kept near the end physical page range that has already been
4308 * registered. This function allows an arch to shrink an existing registered
4311 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4312 unsigned long end_pfn)
4317 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4318 nid, start_pfn, end_pfn);
4320 /* Find the old active region end and shrink */
4321 for_each_active_range_index_in_nid(i, nid) {
4322 if (early_node_map[i].start_pfn >= start_pfn &&
4323 early_node_map[i].end_pfn <= end_pfn) {
4325 early_node_map[i].start_pfn = 0;
4326 early_node_map[i].end_pfn = 0;
4330 if (early_node_map[i].start_pfn < start_pfn &&
4331 early_node_map[i].end_pfn > start_pfn) {
4332 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4333 early_node_map[i].end_pfn = start_pfn;
4334 if (temp_end_pfn > end_pfn)
4335 add_active_range(nid, end_pfn, temp_end_pfn);
4338 if (early_node_map[i].start_pfn >= start_pfn &&
4339 early_node_map[i].end_pfn > end_pfn &&
4340 early_node_map[i].start_pfn < end_pfn) {
4341 early_node_map[i].start_pfn = end_pfn;
4349 /* remove the blank ones */
4350 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4351 if (early_node_map[i].nid != nid)
4353 if (early_node_map[i].end_pfn)
4355 /* we found it, get rid of it */
4356 for (j = i; j < nr_nodemap_entries - 1; j++)
4357 memcpy(&early_node_map[j], &early_node_map[j+1],
4358 sizeof(early_node_map[j]));
4359 j = nr_nodemap_entries - 1;
4360 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4361 nr_nodemap_entries--;
4366 * remove_all_active_ranges - Remove all currently registered regions
4368 * During discovery, it may be found that a table like SRAT is invalid
4369 * and an alternative discovery method must be used. This function removes
4370 * all currently registered regions.
4372 void __init remove_all_active_ranges(void)
4374 memset(early_node_map, 0, sizeof(early_node_map));
4375 nr_nodemap_entries = 0;
4378 /* Compare two active node_active_regions */
4379 static int __init cmp_node_active_region(const void *a, const void *b)
4381 struct node_active_region *arange = (struct node_active_region *)a;
4382 struct node_active_region *brange = (struct node_active_region *)b;
4384 /* Done this way to avoid overflows */
4385 if (arange->start_pfn > brange->start_pfn)
4387 if (arange->start_pfn < brange->start_pfn)
4393 /* sort the node_map by start_pfn */
4394 void __init sort_node_map(void)
4396 sort(early_node_map, (size_t)nr_nodemap_entries,
4397 sizeof(struct node_active_region),
4398 cmp_node_active_region, NULL);
4401 /* Find the lowest pfn for a node */
4402 static unsigned long __init find_min_pfn_for_node(int nid)
4405 unsigned long min_pfn = ULONG_MAX;
4407 /* Assuming a sorted map, the first range found has the starting pfn */
4408 for_each_active_range_index_in_nid(i, nid)
4409 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4411 if (min_pfn == ULONG_MAX) {
4413 "Could not find start_pfn for node %d\n", nid);
4421 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4423 * It returns the minimum PFN based on information provided via
4424 * add_active_range().
4426 unsigned long __init find_min_pfn_with_active_regions(void)
4428 return find_min_pfn_for_node(MAX_NUMNODES);
4432 * early_calculate_totalpages()
4433 * Sum pages in active regions for movable zone.
4434 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4436 static unsigned long __init early_calculate_totalpages(void)
4439 unsigned long totalpages = 0;
4441 for (i = 0; i < nr_nodemap_entries; i++) {
4442 unsigned long pages = early_node_map[i].end_pfn -
4443 early_node_map[i].start_pfn;
4444 totalpages += pages;
4446 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4452 * Find the PFN the Movable zone begins in each node. Kernel memory
4453 * is spread evenly between nodes as long as the nodes have enough
4454 * memory. When they don't, some nodes will have more kernelcore than
4457 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4460 unsigned long usable_startpfn;
4461 unsigned long kernelcore_node, kernelcore_remaining;
4462 /* save the state before borrow the nodemask */
4463 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4464 unsigned long totalpages = early_calculate_totalpages();
4465 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4468 * If movablecore was specified, calculate what size of
4469 * kernelcore that corresponds so that memory usable for
4470 * any allocation type is evenly spread. If both kernelcore
4471 * and movablecore are specified, then the value of kernelcore
4472 * will be used for required_kernelcore if it's greater than
4473 * what movablecore would have allowed.
4475 if (required_movablecore) {
4476 unsigned long corepages;
4479 * Round-up so that ZONE_MOVABLE is at least as large as what
4480 * was requested by the user
4482 required_movablecore =
4483 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4484 corepages = totalpages - required_movablecore;
4486 required_kernelcore = max(required_kernelcore, corepages);
4489 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4490 if (!required_kernelcore)
4493 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4494 find_usable_zone_for_movable();
4495 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4498 /* Spread kernelcore memory as evenly as possible throughout nodes */
4499 kernelcore_node = required_kernelcore / usable_nodes;
4500 for_each_node_state(nid, N_HIGH_MEMORY) {
4502 * Recalculate kernelcore_node if the division per node
4503 * now exceeds what is necessary to satisfy the requested
4504 * amount of memory for the kernel
4506 if (required_kernelcore < kernelcore_node)
4507 kernelcore_node = required_kernelcore / usable_nodes;
4510 * As the map is walked, we track how much memory is usable
4511 * by the kernel using kernelcore_remaining. When it is
4512 * 0, the rest of the node is usable by ZONE_MOVABLE
4514 kernelcore_remaining = kernelcore_node;
4516 /* Go through each range of PFNs within this node */
4517 for_each_active_range_index_in_nid(i, nid) {
4518 unsigned long start_pfn, end_pfn;
4519 unsigned long size_pages;
4521 start_pfn = max(early_node_map[i].start_pfn,
4522 zone_movable_pfn[nid]);
4523 end_pfn = early_node_map[i].end_pfn;
4524 if (start_pfn >= end_pfn)
4527 /* Account for what is only usable for kernelcore */
4528 if (start_pfn < usable_startpfn) {
4529 unsigned long kernel_pages;
4530 kernel_pages = min(end_pfn, usable_startpfn)
4533 kernelcore_remaining -= min(kernel_pages,
4534 kernelcore_remaining);
4535 required_kernelcore -= min(kernel_pages,
4536 required_kernelcore);
4538 /* Continue if range is now fully accounted */
4539 if (end_pfn <= usable_startpfn) {
4542 * Push zone_movable_pfn to the end so
4543 * that if we have to rebalance
4544 * kernelcore across nodes, we will
4545 * not double account here
4547 zone_movable_pfn[nid] = end_pfn;
4550 start_pfn = usable_startpfn;
4554 * The usable PFN range for ZONE_MOVABLE is from
4555 * start_pfn->end_pfn. Calculate size_pages as the
4556 * number of pages used as kernelcore
4558 size_pages = end_pfn - start_pfn;
4559 if (size_pages > kernelcore_remaining)
4560 size_pages = kernelcore_remaining;
4561 zone_movable_pfn[nid] = start_pfn + size_pages;
4564 * Some kernelcore has been met, update counts and
4565 * break if the kernelcore for this node has been
4568 required_kernelcore -= min(required_kernelcore,
4570 kernelcore_remaining -= size_pages;
4571 if (!kernelcore_remaining)
4577 * If there is still required_kernelcore, we do another pass with one
4578 * less node in the count. This will push zone_movable_pfn[nid] further
4579 * along on the nodes that still have memory until kernelcore is
4583 if (usable_nodes && required_kernelcore > usable_nodes)
4586 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4587 for (nid = 0; nid < MAX_NUMNODES; nid++)
4588 zone_movable_pfn[nid] =
4589 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4592 /* restore the node_state */
4593 node_states[N_HIGH_MEMORY] = saved_node_state;
4596 /* Any regular memory on that node ? */
4597 static void check_for_regular_memory(pg_data_t *pgdat)
4599 #ifdef CONFIG_HIGHMEM
4600 enum zone_type zone_type;
4602 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4603 struct zone *zone = &pgdat->node_zones[zone_type];
4604 if (zone->present_pages)
4605 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4611 * free_area_init_nodes - Initialise all pg_data_t and zone data
4612 * @max_zone_pfn: an array of max PFNs for each zone
4614 * This will call free_area_init_node() for each active node in the system.
4615 * Using the page ranges provided by add_active_range(), the size of each
4616 * zone in each node and their holes is calculated. If the maximum PFN
4617 * between two adjacent zones match, it is assumed that the zone is empty.
4618 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4619 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4620 * starts where the previous one ended. For example, ZONE_DMA32 starts
4621 * at arch_max_dma_pfn.
4623 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4628 /* Sort early_node_map as initialisation assumes it is sorted */
4631 /* Record where the zone boundaries are */
4632 memset(arch_zone_lowest_possible_pfn, 0,
4633 sizeof(arch_zone_lowest_possible_pfn));
4634 memset(arch_zone_highest_possible_pfn, 0,
4635 sizeof(arch_zone_highest_possible_pfn));
4636 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4637 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4638 for (i = 1; i < MAX_NR_ZONES; i++) {
4639 if (i == ZONE_MOVABLE)
4641 arch_zone_lowest_possible_pfn[i] =
4642 arch_zone_highest_possible_pfn[i-1];
4643 arch_zone_highest_possible_pfn[i] =
4644 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4646 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4647 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4649 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4650 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4651 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4653 /* Print out the zone ranges */
4654 printk("Zone PFN ranges:\n");
4655 for (i = 0; i < MAX_NR_ZONES; i++) {
4656 if (i == ZONE_MOVABLE)
4658 printk(" %-8s ", zone_names[i]);
4659 if (arch_zone_lowest_possible_pfn[i] ==
4660 arch_zone_highest_possible_pfn[i])
4663 printk("%0#10lx -> %0#10lx\n",
4664 arch_zone_lowest_possible_pfn[i],
4665 arch_zone_highest_possible_pfn[i]);
4668 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4669 printk("Movable zone start PFN for each node\n");
4670 for (i = 0; i < MAX_NUMNODES; i++) {
4671 if (zone_movable_pfn[i])
4672 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4675 /* Print out the early_node_map[] */
4676 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4677 for (i = 0; i < nr_nodemap_entries; i++)
4678 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4679 early_node_map[i].start_pfn,
4680 early_node_map[i].end_pfn);
4682 /* Initialise every node */
4683 mminit_verify_pageflags_layout();
4684 setup_nr_node_ids();
4685 for_each_online_node(nid) {
4686 pg_data_t *pgdat = NODE_DATA(nid);
4687 free_area_init_node(nid, NULL,
4688 find_min_pfn_for_node(nid), NULL);
4690 /* Any memory on that node */
4691 if (pgdat->node_present_pages)
4692 node_set_state(nid, N_HIGH_MEMORY);
4693 check_for_regular_memory(pgdat);
4697 static int __init cmdline_parse_core(char *p, unsigned long *core)
4699 unsigned long long coremem;
4703 coremem = memparse(p, &p);
4704 *core = coremem >> PAGE_SHIFT;
4706 /* Paranoid check that UL is enough for the coremem value */
4707 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4713 * kernelcore=size sets the amount of memory for use for allocations that
4714 * cannot be reclaimed or migrated.
4716 static int __init cmdline_parse_kernelcore(char *p)
4718 return cmdline_parse_core(p, &required_kernelcore);
4722 * movablecore=size sets the amount of memory for use for allocations that
4723 * can be reclaimed or migrated.
4725 static int __init cmdline_parse_movablecore(char *p)
4727 return cmdline_parse_core(p, &required_movablecore);
4730 early_param("kernelcore", cmdline_parse_kernelcore);
4731 early_param("movablecore", cmdline_parse_movablecore);
4733 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4736 * set_dma_reserve - set the specified number of pages reserved in the first zone
4737 * @new_dma_reserve: The number of pages to mark reserved
4739 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4740 * In the DMA zone, a significant percentage may be consumed by kernel image
4741 * and other unfreeable allocations which can skew the watermarks badly. This
4742 * function may optionally be used to account for unfreeable pages in the
4743 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4744 * smaller per-cpu batchsize.
4746 void __init set_dma_reserve(unsigned long new_dma_reserve)
4748 dma_reserve = new_dma_reserve;
4751 #ifndef CONFIG_NEED_MULTIPLE_NODES
4752 struct pglist_data __refdata contig_page_data = {
4753 #ifndef CONFIG_NO_BOOTMEM
4754 .bdata = &bootmem_node_data[0]
4757 EXPORT_SYMBOL(contig_page_data);
4760 void __init free_area_init(unsigned long *zones_size)
4762 free_area_init_node(0, zones_size,
4763 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4766 static int page_alloc_cpu_notify(struct notifier_block *self,
4767 unsigned long action, void *hcpu)
4769 int cpu = (unsigned long)hcpu;
4771 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4775 * Spill the event counters of the dead processor
4776 * into the current processors event counters.
4777 * This artificially elevates the count of the current
4780 vm_events_fold_cpu(cpu);
4783 * Zero the differential counters of the dead processor
4784 * so that the vm statistics are consistent.
4786 * This is only okay since the processor is dead and cannot
4787 * race with what we are doing.
4789 refresh_cpu_vm_stats(cpu);
4794 void __init page_alloc_init(void)
4796 hotcpu_notifier(page_alloc_cpu_notify, 0);
4800 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4801 * or min_free_kbytes changes.
4803 static void calculate_totalreserve_pages(void)
4805 struct pglist_data *pgdat;
4806 unsigned long reserve_pages = 0;
4807 enum zone_type i, j;
4809 for_each_online_pgdat(pgdat) {
4810 for (i = 0; i < MAX_NR_ZONES; i++) {
4811 struct zone *zone = pgdat->node_zones + i;
4812 unsigned long max = 0;
4814 /* Find valid and maximum lowmem_reserve in the zone */
4815 for (j = i; j < MAX_NR_ZONES; j++) {
4816 if (zone->lowmem_reserve[j] > max)
4817 max = zone->lowmem_reserve[j];
4820 /* we treat the high watermark as reserved pages. */
4821 max += high_wmark_pages(zone);
4823 if (max > zone->present_pages)
4824 max = zone->present_pages;
4825 reserve_pages += max;
4828 totalreserve_pages = reserve_pages;
4832 * setup_per_zone_lowmem_reserve - called whenever
4833 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4834 * has a correct pages reserved value, so an adequate number of
4835 * pages are left in the zone after a successful __alloc_pages().
4837 static void setup_per_zone_lowmem_reserve(void)
4839 struct pglist_data *pgdat;
4840 enum zone_type j, idx;
4842 for_each_online_pgdat(pgdat) {
4843 for (j = 0; j < MAX_NR_ZONES; j++) {
4844 struct zone *zone = pgdat->node_zones + j;
4845 unsigned long present_pages = zone->present_pages;
4847 zone->lowmem_reserve[j] = 0;
4851 struct zone *lower_zone;
4855 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4856 sysctl_lowmem_reserve_ratio[idx] = 1;
4858 lower_zone = pgdat->node_zones + idx;
4859 lower_zone->lowmem_reserve[j] = present_pages /
4860 sysctl_lowmem_reserve_ratio[idx];
4861 present_pages += lower_zone->present_pages;
4866 /* update totalreserve_pages */
4867 calculate_totalreserve_pages();
4871 * setup_per_zone_wmarks - called when min_free_kbytes changes
4872 * or when memory is hot-{added|removed}
4874 * Ensures that the watermark[min,low,high] values for each zone are set
4875 * correctly with respect to min_free_kbytes.
4877 void setup_per_zone_wmarks(void)
4879 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4880 unsigned long lowmem_pages = 0;
4882 unsigned long flags;
4884 /* Calculate total number of !ZONE_HIGHMEM pages */
4885 for_each_zone(zone) {
4886 if (!is_highmem(zone))
4887 lowmem_pages += zone->present_pages;
4890 for_each_zone(zone) {
4893 spin_lock_irqsave(&zone->lock, flags);
4894 tmp = (u64)pages_min * zone->present_pages;
4895 do_div(tmp, lowmem_pages);
4896 if (is_highmem(zone)) {
4898 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4899 * need highmem pages, so cap pages_min to a small
4902 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4903 * deltas controls asynch page reclaim, and so should
4904 * not be capped for highmem.
4908 min_pages = zone->present_pages / 1024;
4909 if (min_pages < SWAP_CLUSTER_MAX)
4910 min_pages = SWAP_CLUSTER_MAX;
4911 if (min_pages > 128)
4913 zone->watermark[WMARK_MIN] = min_pages;
4916 * If it's a lowmem zone, reserve a number of pages
4917 * proportionate to the zone's size.
4919 zone->watermark[WMARK_MIN] = tmp;
4922 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4923 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4924 setup_zone_migrate_reserve(zone);
4925 spin_unlock_irqrestore(&zone->lock, flags);
4928 /* update totalreserve_pages */
4929 calculate_totalreserve_pages();
4933 * The inactive anon list should be small enough that the VM never has to
4934 * do too much work, but large enough that each inactive page has a chance
4935 * to be referenced again before it is swapped out.
4937 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4938 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4939 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4940 * the anonymous pages are kept on the inactive list.
4943 * memory ratio inactive anon
4944 * -------------------------------------
4953 void calculate_zone_inactive_ratio(struct zone *zone)
4955 unsigned int gb, ratio;
4957 /* Zone size in gigabytes */
4958 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4960 ratio = int_sqrt(10 * gb);
4964 zone->inactive_ratio = ratio;
4967 static void __init setup_per_zone_inactive_ratio(void)
4972 calculate_zone_inactive_ratio(zone);
4976 * Initialise min_free_kbytes.
4978 * For small machines we want it small (128k min). For large machines
4979 * we want it large (64MB max). But it is not linear, because network
4980 * bandwidth does not increase linearly with machine size. We use
4982 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4983 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4999 static int __init init_per_zone_wmark_min(void)
5001 unsigned long lowmem_kbytes;
5003 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5005 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5006 if (min_free_kbytes < 128)
5007 min_free_kbytes = 128;
5008 if (min_free_kbytes > 65536)
5009 min_free_kbytes = 65536;
5010 setup_per_zone_wmarks();
5011 setup_per_zone_lowmem_reserve();
5012 setup_per_zone_inactive_ratio();
5015 module_init(init_per_zone_wmark_min)
5018 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5019 * that we can call two helper functions whenever min_free_kbytes
5022 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5023 void __user *buffer, size_t *length, loff_t *ppos)
5025 proc_dointvec(table, write, buffer, length, ppos);
5027 setup_per_zone_wmarks();
5032 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5033 void __user *buffer, size_t *length, loff_t *ppos)
5038 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5043 zone->min_unmapped_pages = (zone->present_pages *
5044 sysctl_min_unmapped_ratio) / 100;
5048 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5049 void __user *buffer, size_t *length, loff_t *ppos)
5054 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5059 zone->min_slab_pages = (zone->present_pages *
5060 sysctl_min_slab_ratio) / 100;
5066 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5067 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5068 * whenever sysctl_lowmem_reserve_ratio changes.
5070 * The reserve ratio obviously has absolutely no relation with the
5071 * minimum watermarks. The lowmem reserve ratio can only make sense
5072 * if in function of the boot time zone sizes.
5074 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5075 void __user *buffer, size_t *length, loff_t *ppos)
5077 proc_dointvec_minmax(table, write, buffer, length, ppos);
5078 setup_per_zone_lowmem_reserve();
5083 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5084 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5085 * can have before it gets flushed back to buddy allocator.
5088 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5089 void __user *buffer, size_t *length, loff_t *ppos)
5095 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5096 if (!write || (ret == -EINVAL))
5098 for_each_populated_zone(zone) {
5099 for_each_possible_cpu(cpu) {
5101 high = zone->present_pages / percpu_pagelist_fraction;
5102 setup_pagelist_highmark(
5103 per_cpu_ptr(zone->pageset, cpu), high);
5109 int hashdist = HASHDIST_DEFAULT;
5112 static int __init set_hashdist(char *str)
5116 hashdist = simple_strtoul(str, &str, 0);
5119 __setup("hashdist=", set_hashdist);
5123 * allocate a large system hash table from bootmem
5124 * - it is assumed that the hash table must contain an exact power-of-2
5125 * quantity of entries
5126 * - limit is the number of hash buckets, not the total allocation size
5128 void *__init alloc_large_system_hash(const char *tablename,
5129 unsigned long bucketsize,
5130 unsigned long numentries,
5133 unsigned int *_hash_shift,
5134 unsigned int *_hash_mask,
5135 unsigned long limit)
5137 unsigned long long max = limit;
5138 unsigned long log2qty, size;
5141 /* allow the kernel cmdline to have a say */
5143 /* round applicable memory size up to nearest megabyte */
5144 numentries = nr_kernel_pages;
5145 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5146 numentries >>= 20 - PAGE_SHIFT;
5147 numentries <<= 20 - PAGE_SHIFT;
5149 /* limit to 1 bucket per 2^scale bytes of low memory */
5150 if (scale > PAGE_SHIFT)
5151 numentries >>= (scale - PAGE_SHIFT);
5153 numentries <<= (PAGE_SHIFT - scale);
5155 /* Make sure we've got at least a 0-order allocation.. */
5156 if (unlikely(flags & HASH_SMALL)) {
5157 /* Makes no sense without HASH_EARLY */
5158 WARN_ON(!(flags & HASH_EARLY));
5159 if (!(numentries >> *_hash_shift)) {
5160 numentries = 1UL << *_hash_shift;
5161 BUG_ON(!numentries);
5163 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5164 numentries = PAGE_SIZE / bucketsize;
5166 numentries = roundup_pow_of_two(numentries);
5168 /* limit allocation size to 1/16 total memory by default */
5170 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5171 do_div(max, bucketsize);
5174 if (numentries > max)
5177 log2qty = ilog2(numentries);
5180 size = bucketsize << log2qty;
5181 if (flags & HASH_EARLY)
5182 table = alloc_bootmem_nopanic(size);
5184 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5187 * If bucketsize is not a power-of-two, we may free
5188 * some pages at the end of hash table which
5189 * alloc_pages_exact() automatically does
5191 if (get_order(size) < MAX_ORDER) {
5192 table = alloc_pages_exact(size, GFP_ATOMIC);
5193 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5196 } while (!table && size > PAGE_SIZE && --log2qty);
5199 panic("Failed to allocate %s hash table\n", tablename);
5201 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5204 ilog2(size) - PAGE_SHIFT,
5208 *_hash_shift = log2qty;
5210 *_hash_mask = (1 << log2qty) - 1;
5215 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5216 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5219 #ifdef CONFIG_SPARSEMEM
5220 return __pfn_to_section(pfn)->pageblock_flags;
5222 return zone->pageblock_flags;
5223 #endif /* CONFIG_SPARSEMEM */
5226 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5228 #ifdef CONFIG_SPARSEMEM
5229 pfn &= (PAGES_PER_SECTION-1);
5230 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5232 pfn = pfn - zone->zone_start_pfn;
5233 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5234 #endif /* CONFIG_SPARSEMEM */
5238 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5239 * @page: The page within the block of interest
5240 * @start_bitidx: The first bit of interest to retrieve
5241 * @end_bitidx: The last bit of interest
5242 * returns pageblock_bits flags
5244 unsigned long get_pageblock_flags_group(struct page *page,
5245 int start_bitidx, int end_bitidx)
5248 unsigned long *bitmap;
5249 unsigned long pfn, bitidx;
5250 unsigned long flags = 0;
5251 unsigned long value = 1;
5253 zone = page_zone(page);
5254 pfn = page_to_pfn(page);
5255 bitmap = get_pageblock_bitmap(zone, pfn);
5256 bitidx = pfn_to_bitidx(zone, pfn);
5258 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5259 if (test_bit(bitidx + start_bitidx, bitmap))
5266 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5267 * @page: The page within the block of interest
5268 * @start_bitidx: The first bit of interest
5269 * @end_bitidx: The last bit of interest
5270 * @flags: The flags to set
5272 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5273 int start_bitidx, int end_bitidx)
5276 unsigned long *bitmap;
5277 unsigned long pfn, bitidx;
5278 unsigned long value = 1;
5280 zone = page_zone(page);
5281 pfn = page_to_pfn(page);
5282 bitmap = get_pageblock_bitmap(zone, pfn);
5283 bitidx = pfn_to_bitidx(zone, pfn);
5284 VM_BUG_ON(pfn < zone->zone_start_pfn);
5285 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5287 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5289 __set_bit(bitidx + start_bitidx, bitmap);
5291 __clear_bit(bitidx + start_bitidx, bitmap);
5295 * This is designed as sub function...plz see page_isolation.c also.
5296 * set/clear page block's type to be ISOLATE.
5297 * page allocater never alloc memory from ISOLATE block.
5300 int set_migratetype_isolate(struct page *page)
5303 struct page *curr_page;
5304 unsigned long flags, pfn, iter;
5305 unsigned long immobile = 0;
5306 struct memory_isolate_notify arg;
5311 zone = page_zone(page);
5312 zone_idx = zone_idx(zone);
5314 spin_lock_irqsave(&zone->lock, flags);
5315 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE ||
5316 zone_idx == ZONE_MOVABLE) {
5321 pfn = page_to_pfn(page);
5322 arg.start_pfn = pfn;
5323 arg.nr_pages = pageblock_nr_pages;
5324 arg.pages_found = 0;
5327 * It may be possible to isolate a pageblock even if the
5328 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5329 * notifier chain is used by balloon drivers to return the
5330 * number of pages in a range that are held by the balloon
5331 * driver to shrink memory. If all the pages are accounted for
5332 * by balloons, are free, or on the LRU, isolation can continue.
5333 * Later, for example, when memory hotplug notifier runs, these
5334 * pages reported as "can be isolated" should be isolated(freed)
5335 * by the balloon driver through the memory notifier chain.
5337 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5338 notifier_ret = notifier_to_errno(notifier_ret);
5339 if (notifier_ret || !arg.pages_found)
5342 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) {
5343 if (!pfn_valid_within(pfn))
5346 curr_page = pfn_to_page(iter);
5347 if (!page_count(curr_page) || PageLRU(curr_page))
5353 if (arg.pages_found == immobile)
5358 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5359 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5362 spin_unlock_irqrestore(&zone->lock, flags);
5368 void unset_migratetype_isolate(struct page *page)
5371 unsigned long flags;
5372 zone = page_zone(page);
5373 spin_lock_irqsave(&zone->lock, flags);
5374 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5376 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5377 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5379 spin_unlock_irqrestore(&zone->lock, flags);
5382 #ifdef CONFIG_MEMORY_HOTREMOVE
5384 * All pages in the range must be isolated before calling this.
5387 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5393 unsigned long flags;
5394 /* find the first valid pfn */
5395 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5400 zone = page_zone(pfn_to_page(pfn));
5401 spin_lock_irqsave(&zone->lock, flags);
5403 while (pfn < end_pfn) {
5404 if (!pfn_valid(pfn)) {
5408 page = pfn_to_page(pfn);
5409 BUG_ON(page_count(page));
5410 BUG_ON(!PageBuddy(page));
5411 order = page_order(page);
5412 #ifdef CONFIG_DEBUG_VM
5413 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5414 pfn, 1 << order, end_pfn);
5416 list_del(&page->lru);
5417 rmv_page_order(page);
5418 zone->free_area[order].nr_free--;
5419 __mod_zone_page_state(zone, NR_FREE_PAGES,
5421 for (i = 0; i < (1 << order); i++)
5422 SetPageReserved((page+i));
5423 pfn += (1 << order);
5425 spin_unlock_irqrestore(&zone->lock, flags);
5429 #ifdef CONFIG_MEMORY_FAILURE
5430 bool is_free_buddy_page(struct page *page)
5432 struct zone *zone = page_zone(page);
5433 unsigned long pfn = page_to_pfn(page);
5434 unsigned long flags;
5437 spin_lock_irqsave(&zone->lock, flags);
5438 for (order = 0; order < MAX_ORDER; order++) {
5439 struct page *page_head = page - (pfn & ((1 << order) - 1));
5441 if (PageBuddy(page_head) && page_order(page_head) >= order)
5444 spin_unlock_irqrestore(&zone->lock, flags);
5446 return order < MAX_ORDER;
5450 static struct trace_print_flags pageflag_names[] = {
5451 {1UL << PG_locked, "locked" },
5452 {1UL << PG_error, "error" },
5453 {1UL << PG_referenced, "referenced" },
5454 {1UL << PG_uptodate, "uptodate" },
5455 {1UL << PG_dirty, "dirty" },
5456 {1UL << PG_lru, "lru" },
5457 {1UL << PG_active, "active" },
5458 {1UL << PG_slab, "slab" },
5459 {1UL << PG_owner_priv_1, "owner_priv_1" },
5460 {1UL << PG_arch_1, "arch_1" },
5461 {1UL << PG_reserved, "reserved" },
5462 {1UL << PG_private, "private" },
5463 {1UL << PG_private_2, "private_2" },
5464 {1UL << PG_writeback, "writeback" },
5465 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5466 {1UL << PG_head, "head" },
5467 {1UL << PG_tail, "tail" },
5469 {1UL << PG_compound, "compound" },
5471 {1UL << PG_swapcache, "swapcache" },
5472 {1UL << PG_mappedtodisk, "mappedtodisk" },
5473 {1UL << PG_reclaim, "reclaim" },
5474 {1UL << PG_buddy, "buddy" },
5475 {1UL << PG_swapbacked, "swapbacked" },
5476 {1UL << PG_unevictable, "unevictable" },
5478 {1UL << PG_mlocked, "mlocked" },
5480 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5481 {1UL << PG_uncached, "uncached" },
5483 #ifdef CONFIG_MEMORY_FAILURE
5484 {1UL << PG_hwpoison, "hwpoison" },
5489 static void dump_page_flags(unsigned long flags)
5491 const char *delim = "";
5495 printk(KERN_ALERT "page flags: %#lx(", flags);
5497 /* remove zone id */
5498 flags &= (1UL << NR_PAGEFLAGS) - 1;
5500 for (i = 0; pageflag_names[i].name && flags; i++) {
5502 mask = pageflag_names[i].mask;
5503 if ((flags & mask) != mask)
5507 printk("%s%s", delim, pageflag_names[i].name);
5511 /* check for left over flags */
5513 printk("%s%#lx", delim, flags);
5518 void dump_page(struct page *page)
5521 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5522 page, page_count(page), page_mapcount(page),
5523 page->mapping, page->index);
5524 dump_page_flags(page->flags);