2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/memory.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/memcontrol.h>
59 #include <linux/prefetch.h>
61 #include <asm/tlbflush.h>
62 #include <asm/div64.h>
65 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
66 DEFINE_PER_CPU(int, numa_node);
67 EXPORT_PER_CPU_SYMBOL(numa_node);
70 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
72 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
73 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
74 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
75 * defined in <linux/topology.h>.
77 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
78 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
82 * Array of node states.
84 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
85 [N_POSSIBLE] = NODE_MASK_ALL,
86 [N_ONLINE] = { { [0] = 1UL } },
88 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
90 [N_HIGH_MEMORY] = { { [0] = 1UL } },
92 [N_CPU] = { { [0] = 1UL } },
95 EXPORT_SYMBOL(node_states);
97 unsigned long totalram_pages __read_mostly;
98 unsigned long totalreserve_pages __read_mostly;
99 int percpu_pagelist_fraction;
100 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
102 #ifdef CONFIG_PM_SLEEP
104 * The following functions are used by the suspend/hibernate code to temporarily
105 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
106 * while devices are suspended. To avoid races with the suspend/hibernate code,
107 * they should always be called with pm_mutex held (gfp_allowed_mask also should
108 * only be modified with pm_mutex held, unless the suspend/hibernate code is
109 * guaranteed not to run in parallel with that modification).
112 static gfp_t saved_gfp_mask;
114 void pm_restore_gfp_mask(void)
116 WARN_ON(!mutex_is_locked(&pm_mutex));
117 if (saved_gfp_mask) {
118 gfp_allowed_mask = saved_gfp_mask;
123 void pm_restrict_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 WARN_ON(saved_gfp_mask);
127 saved_gfp_mask = gfp_allowed_mask;
128 gfp_allowed_mask &= ~GFP_IOFS;
130 #endif /* CONFIG_PM_SLEEP */
132 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
133 int pageblock_order __read_mostly;
136 static void __free_pages_ok(struct page *page, unsigned int order);
139 * results with 256, 32 in the lowmem_reserve sysctl:
140 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
141 * 1G machine -> (16M dma, 784M normal, 224M high)
142 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
143 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
144 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
146 * TBD: should special case ZONE_DMA32 machines here - in those we normally
147 * don't need any ZONE_NORMAL reservation
149 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
150 #ifdef CONFIG_ZONE_DMA
153 #ifdef CONFIG_ZONE_DMA32
156 #ifdef CONFIG_HIGHMEM
162 EXPORT_SYMBOL(totalram_pages);
164 static char * const zone_names[MAX_NR_ZONES] = {
165 #ifdef CONFIG_ZONE_DMA
168 #ifdef CONFIG_ZONE_DMA32
172 #ifdef CONFIG_HIGHMEM
178 int min_free_kbytes = 1024;
180 static unsigned long __meminitdata nr_kernel_pages;
181 static unsigned long __meminitdata nr_all_pages;
182 static unsigned long __meminitdata dma_reserve;
184 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
186 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
187 * ranges of memory (RAM) that may be registered with add_active_range().
188 * Ranges passed to add_active_range() will be merged if possible
189 * so the number of times add_active_range() can be called is
190 * related to the number of nodes and the number of holes
192 #ifdef CONFIG_MAX_ACTIVE_REGIONS
193 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
194 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
196 #if MAX_NUMNODES >= 32
197 /* If there can be many nodes, allow up to 50 holes per node */
198 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
200 /* By default, allow up to 256 distinct regions */
201 #define MAX_ACTIVE_REGIONS 256
205 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
206 static int __meminitdata nr_nodemap_entries;
207 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
208 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
209 static unsigned long __initdata required_kernelcore;
210 static unsigned long __initdata required_movablecore;
211 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
213 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
215 EXPORT_SYMBOL(movable_zone);
216 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
219 int nr_node_ids __read_mostly = MAX_NUMNODES;
220 int nr_online_nodes __read_mostly = 1;
221 EXPORT_SYMBOL(nr_node_ids);
222 EXPORT_SYMBOL(nr_online_nodes);
225 int page_group_by_mobility_disabled __read_mostly;
227 static void set_pageblock_migratetype(struct page *page, int migratetype)
230 if (unlikely(page_group_by_mobility_disabled))
231 migratetype = MIGRATE_UNMOVABLE;
233 set_pageblock_flags_group(page, (unsigned long)migratetype,
234 PB_migrate, PB_migrate_end);
237 bool oom_killer_disabled __read_mostly;
239 #ifdef CONFIG_DEBUG_VM
240 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
244 unsigned long pfn = page_to_pfn(page);
247 seq = zone_span_seqbegin(zone);
248 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
250 else if (pfn < zone->zone_start_pfn)
252 } while (zone_span_seqretry(zone, seq));
257 static int page_is_consistent(struct zone *zone, struct page *page)
259 if (!pfn_valid_within(page_to_pfn(page)))
261 if (zone != page_zone(page))
267 * Temporary debugging check for pages not lying within a given zone.
269 static int bad_range(struct zone *zone, struct page *page)
271 if (page_outside_zone_boundaries(zone, page))
273 if (!page_is_consistent(zone, page))
279 static inline int bad_range(struct zone *zone, struct page *page)
285 static void bad_page(struct page *page)
287 static unsigned long resume;
288 static unsigned long nr_shown;
289 static unsigned long nr_unshown;
291 /* Don't complain about poisoned pages */
292 if (PageHWPoison(page)) {
293 reset_page_mapcount(page); /* remove PageBuddy */
298 * Allow a burst of 60 reports, then keep quiet for that minute;
299 * or allow a steady drip of one report per second.
301 if (nr_shown == 60) {
302 if (time_before(jiffies, resume)) {
308 "BUG: Bad page state: %lu messages suppressed\n",
315 resume = jiffies + 60 * HZ;
317 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
318 current->comm, page_to_pfn(page));
323 /* Leave bad fields for debug, except PageBuddy could make trouble */
324 reset_page_mapcount(page); /* remove PageBuddy */
325 add_taint(TAINT_BAD_PAGE);
329 * Higher-order pages are called "compound pages". They are structured thusly:
331 * The first PAGE_SIZE page is called the "head page".
333 * The remaining PAGE_SIZE pages are called "tail pages".
335 * All pages have PG_compound set. All pages have their ->private pointing at
336 * the head page (even the head page has this).
338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
343 static void free_compound_page(struct page *page)
345 __free_pages_ok(page, compound_order(page));
348 void prep_compound_page(struct page *page, unsigned long order)
351 int nr_pages = 1 << order;
353 set_compound_page_dtor(page, free_compound_page);
354 set_compound_order(page, order);
356 for (i = 1; i < nr_pages; i++) {
357 struct page *p = page + i;
360 p->first_page = page;
364 /* update __split_huge_page_refcount if you change this function */
365 static int destroy_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
371 if (unlikely(compound_order(page) != order) ||
372 unlikely(!PageHead(page))) {
377 __ClearPageHead(page);
379 for (i = 1; i < nr_pages; i++) {
380 struct page *p = page + i;
382 if (unlikely(!PageTail(p) || (p->first_page != page))) {
392 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
397 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
398 * and __GFP_HIGHMEM from hard or soft interrupt context.
400 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
401 for (i = 0; i < (1 << order); i++)
402 clear_highpage(page + i);
405 static inline void set_page_order(struct page *page, int order)
407 set_page_private(page, order);
408 __SetPageBuddy(page);
411 static inline void rmv_page_order(struct page *page)
413 __ClearPageBuddy(page);
414 set_page_private(page, 0);
418 * Locate the struct page for both the matching buddy in our
419 * pair (buddy1) and the combined O(n+1) page they form (page).
421 * 1) Any buddy B1 will have an order O twin B2 which satisfies
422 * the following equation:
424 * For example, if the starting buddy (buddy2) is #8 its order
426 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
428 * 2) Any buddy B will have an order O+1 parent P which
429 * satisfies the following equation:
432 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
434 static inline unsigned long
435 __find_buddy_index(unsigned long page_idx, unsigned int order)
437 return page_idx ^ (1 << order);
441 * This function checks whether a page is free && is the buddy
442 * we can do coalesce a page and its buddy if
443 * (a) the buddy is not in a hole &&
444 * (b) the buddy is in the buddy system &&
445 * (c) a page and its buddy have the same order &&
446 * (d) a page and its buddy are in the same zone.
448 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
449 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
451 * For recording page's order, we use page_private(page).
453 static inline int page_is_buddy(struct page *page, struct page *buddy,
456 if (!pfn_valid_within(page_to_pfn(buddy)))
459 if (page_zone_id(page) != page_zone_id(buddy))
462 if (PageBuddy(buddy) && page_order(buddy) == order) {
463 VM_BUG_ON(page_count(buddy) != 0);
470 * Freeing function for a buddy system allocator.
472 * The concept of a buddy system is to maintain direct-mapped table
473 * (containing bit values) for memory blocks of various "orders".
474 * The bottom level table contains the map for the smallest allocatable
475 * units of memory (here, pages), and each level above it describes
476 * pairs of units from the levels below, hence, "buddies".
477 * At a high level, all that happens here is marking the table entry
478 * at the bottom level available, and propagating the changes upward
479 * as necessary, plus some accounting needed to play nicely with other
480 * parts of the VM system.
481 * At each level, we keep a list of pages, which are heads of continuous
482 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
483 * order is recorded in page_private(page) field.
484 * So when we are allocating or freeing one, we can derive the state of the
485 * other. That is, if we allocate a small block, and both were
486 * free, the remainder of the region must be split into blocks.
487 * If a block is freed, and its buddy is also free, then this
488 * triggers coalescing into a block of larger size.
493 static inline void __free_one_page(struct page *page,
494 struct zone *zone, unsigned int order,
497 unsigned long page_idx;
498 unsigned long combined_idx;
499 unsigned long uninitialized_var(buddy_idx);
502 if (unlikely(PageCompound(page)))
503 if (unlikely(destroy_compound_page(page, order)))
506 VM_BUG_ON(migratetype == -1);
508 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
510 VM_BUG_ON(page_idx & ((1 << order) - 1));
511 VM_BUG_ON(bad_range(zone, page));
513 while (order < MAX_ORDER-1) {
514 buddy_idx = __find_buddy_index(page_idx, order);
515 buddy = page + (buddy_idx - page_idx);
516 if (!page_is_buddy(page, buddy, order))
519 /* Our buddy is free, merge with it and move up one order. */
520 list_del(&buddy->lru);
521 zone->free_area[order].nr_free--;
522 rmv_page_order(buddy);
523 combined_idx = buddy_idx & page_idx;
524 page = page + (combined_idx - page_idx);
525 page_idx = combined_idx;
528 set_page_order(page, order);
531 * If this is not the largest possible page, check if the buddy
532 * of the next-highest order is free. If it is, it's possible
533 * that pages are being freed that will coalesce soon. In case,
534 * that is happening, add the free page to the tail of the list
535 * so it's less likely to be used soon and more likely to be merged
536 * as a higher order page
538 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
539 struct page *higher_page, *higher_buddy;
540 combined_idx = buddy_idx & page_idx;
541 higher_page = page + (combined_idx - page_idx);
542 buddy_idx = __find_buddy_index(combined_idx, order + 1);
543 higher_buddy = page + (buddy_idx - combined_idx);
544 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
545 list_add_tail(&page->lru,
546 &zone->free_area[order].free_list[migratetype]);
551 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
553 zone->free_area[order].nr_free++;
557 * free_page_mlock() -- clean up attempts to free and mlocked() page.
558 * Page should not be on lru, so no need to fix that up.
559 * free_pages_check() will verify...
561 static inline void free_page_mlock(struct page *page)
563 __dec_zone_page_state(page, NR_MLOCK);
564 __count_vm_event(UNEVICTABLE_MLOCKFREED);
567 static inline int free_pages_check(struct page *page)
569 if (unlikely(page_mapcount(page) |
570 (page->mapping != NULL) |
571 (atomic_read(&page->_count) != 0) |
572 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
573 (mem_cgroup_bad_page_check(page)))) {
577 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
578 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
583 * Frees a number of pages from the PCP lists
584 * Assumes all pages on list are in same zone, and of same order.
585 * count is the number of pages to free.
587 * If the zone was previously in an "all pages pinned" state then look to
588 * see if this freeing clears that state.
590 * And clear the zone's pages_scanned counter, to hold off the "all pages are
591 * pinned" detection logic.
593 static void free_pcppages_bulk(struct zone *zone, int count,
594 struct per_cpu_pages *pcp)
600 spin_lock(&zone->lock);
601 zone->all_unreclaimable = 0;
602 zone->pages_scanned = 0;
606 struct list_head *list;
609 * Remove pages from lists in a round-robin fashion. A
610 * batch_free count is maintained that is incremented when an
611 * empty list is encountered. This is so more pages are freed
612 * off fuller lists instead of spinning excessively around empty
617 if (++migratetype == MIGRATE_PCPTYPES)
619 list = &pcp->lists[migratetype];
620 } while (list_empty(list));
622 /* This is the only non-empty list. Free them all. */
623 if (batch_free == MIGRATE_PCPTYPES)
624 batch_free = to_free;
627 page = list_entry(list->prev, struct page, lru);
628 /* must delete as __free_one_page list manipulates */
629 list_del(&page->lru);
630 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
631 __free_one_page(page, zone, 0, page_private(page));
632 trace_mm_page_pcpu_drain(page, 0, page_private(page));
633 } while (--to_free && --batch_free && !list_empty(list));
635 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
636 spin_unlock(&zone->lock);
639 static void free_one_page(struct zone *zone, struct page *page, int order,
642 spin_lock(&zone->lock);
643 zone->all_unreclaimable = 0;
644 zone->pages_scanned = 0;
646 __free_one_page(page, zone, order, migratetype);
647 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
648 spin_unlock(&zone->lock);
651 static bool free_pages_prepare(struct page *page, unsigned int order)
657 if (PageForeign(page)) {
658 PageForeignDestructor(page, order);
663 trace_mm_page_free_direct(page, order);
664 kmemcheck_free_shadow(page, order);
667 page->mapping = NULL;
668 for (i = 0; i < (1 << order); i++)
669 bad += free_pages_check(page + i);
673 if (!PageHighMem(page)) {
674 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
675 debug_check_no_obj_freed(page_address(page),
678 arch_free_page(page, order);
679 kernel_map_pages(page, 1 << order, 0);
684 static void __free_pages_ok(struct page *page, unsigned int order)
687 int wasMlocked = __TestClearPageMlocked(page);
690 WARN_ON(PageForeign(page) && wasMlocked);
692 if (!free_pages_prepare(page, order))
695 local_irq_save(flags);
696 if (unlikely(wasMlocked))
697 free_page_mlock(page);
698 __count_vm_events(PGFREE, 1 << order);
699 free_one_page(page_zone(page), page, order,
700 get_pageblock_migratetype(page));
701 local_irq_restore(flags);
705 * permit the bootmem allocator to evade page validation on high-order frees
707 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
710 __ClearPageReserved(page);
711 set_page_count(page, 0);
712 set_page_refcounted(page);
718 for (loop = 0; loop < BITS_PER_LONG; loop++) {
719 struct page *p = &page[loop];
721 if (loop + 1 < BITS_PER_LONG)
723 __ClearPageReserved(p);
724 set_page_count(p, 0);
727 set_page_refcounted(page);
728 __free_pages(page, order);
734 * The order of subdivision here is critical for the IO subsystem.
735 * Please do not alter this order without good reasons and regression
736 * testing. Specifically, as large blocks of memory are subdivided,
737 * the order in which smaller blocks are delivered depends on the order
738 * they're subdivided in this function. This is the primary factor
739 * influencing the order in which pages are delivered to the IO
740 * subsystem according to empirical testing, and this is also justified
741 * by considering the behavior of a buddy system containing a single
742 * large block of memory acted on by a series of small allocations.
743 * This behavior is a critical factor in sglist merging's success.
747 static inline void expand(struct zone *zone, struct page *page,
748 int low, int high, struct free_area *area,
751 unsigned long size = 1 << high;
757 VM_BUG_ON(bad_range(zone, &page[size]));
758 list_add(&page[size].lru, &area->free_list[migratetype]);
760 set_page_order(&page[size], high);
765 * This page is about to be returned from the page allocator
767 static inline int check_new_page(struct page *page)
769 if (unlikely(page_mapcount(page) |
770 (page->mapping != NULL) |
771 (atomic_read(&page->_count) != 0) |
772 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
773 (mem_cgroup_bad_page_check(page)))) {
780 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
784 for (i = 0; i < (1 << order); i++) {
785 struct page *p = page + i;
786 if (unlikely(check_new_page(p)))
790 set_page_private(page, 0);
791 set_page_refcounted(page);
793 arch_alloc_page(page, order);
794 kernel_map_pages(page, 1 << order, 1);
796 if (gfp_flags & __GFP_ZERO)
797 prep_zero_page(page, order, gfp_flags);
799 if (order && (gfp_flags & __GFP_COMP))
800 prep_compound_page(page, order);
806 * Go through the free lists for the given migratetype and remove
807 * the smallest available page from the freelists
810 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
813 unsigned int current_order;
814 struct free_area * area;
817 /* Find a page of the appropriate size in the preferred list */
818 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
819 area = &(zone->free_area[current_order]);
820 if (list_empty(&area->free_list[migratetype]))
823 page = list_entry(area->free_list[migratetype].next,
825 list_del(&page->lru);
826 rmv_page_order(page);
828 expand(zone, page, order, current_order, area, migratetype);
837 * This array describes the order lists are fallen back to when
838 * the free lists for the desirable migrate type are depleted
840 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
841 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
842 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
843 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
844 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
848 * Move the free pages in a range to the free lists of the requested type.
849 * Note that start_page and end_pages are not aligned on a pageblock
850 * boundary. If alignment is required, use move_freepages_block()
852 static int move_freepages(struct zone *zone,
853 struct page *start_page, struct page *end_page,
860 #ifndef CONFIG_HOLES_IN_ZONE
862 * page_zone is not safe to call in this context when
863 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
864 * anyway as we check zone boundaries in move_freepages_block().
865 * Remove at a later date when no bug reports exist related to
866 * grouping pages by mobility
868 BUG_ON(page_zone(start_page) != page_zone(end_page));
871 for (page = start_page; page <= end_page;) {
872 /* Make sure we are not inadvertently changing nodes */
873 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
875 if (!pfn_valid_within(page_to_pfn(page))) {
880 if (!PageBuddy(page)) {
885 order = page_order(page);
886 list_move(&page->lru,
887 &zone->free_area[order].free_list[migratetype]);
889 pages_moved += 1 << order;
895 static int move_freepages_block(struct zone *zone, struct page *page,
898 unsigned long start_pfn, end_pfn;
899 struct page *start_page, *end_page;
901 start_pfn = page_to_pfn(page);
902 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
903 start_page = pfn_to_page(start_pfn);
904 end_page = start_page + pageblock_nr_pages - 1;
905 end_pfn = start_pfn + pageblock_nr_pages - 1;
907 /* Do not cross zone boundaries */
908 if (start_pfn < zone->zone_start_pfn)
910 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
913 return move_freepages(zone, start_page, end_page, migratetype);
916 static void change_pageblock_range(struct page *pageblock_page,
917 int start_order, int migratetype)
919 int nr_pageblocks = 1 << (start_order - pageblock_order);
921 while (nr_pageblocks--) {
922 set_pageblock_migratetype(pageblock_page, migratetype);
923 pageblock_page += pageblock_nr_pages;
927 /* Remove an element from the buddy allocator from the fallback list */
928 static inline struct page *
929 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
931 struct free_area * area;
936 /* Find the largest possible block of pages in the other list */
937 for (current_order = MAX_ORDER-1; current_order >= order;
939 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
940 migratetype = fallbacks[start_migratetype][i];
942 /* MIGRATE_RESERVE handled later if necessary */
943 if (migratetype == MIGRATE_RESERVE)
946 area = &(zone->free_area[current_order]);
947 if (list_empty(&area->free_list[migratetype]))
950 page = list_entry(area->free_list[migratetype].next,
955 * If breaking a large block of pages, move all free
956 * pages to the preferred allocation list. If falling
957 * back for a reclaimable kernel allocation, be more
958 * aggressive about taking ownership of free pages
960 if (unlikely(current_order >= (pageblock_order >> 1)) ||
961 start_migratetype == MIGRATE_RECLAIMABLE ||
962 page_group_by_mobility_disabled) {
964 pages = move_freepages_block(zone, page,
967 /* Claim the whole block if over half of it is free */
968 if (pages >= (1 << (pageblock_order-1)) ||
969 page_group_by_mobility_disabled)
970 set_pageblock_migratetype(page,
973 migratetype = start_migratetype;
976 /* Remove the page from the freelists */
977 list_del(&page->lru);
978 rmv_page_order(page);
980 /* Take ownership for orders >= pageblock_order */
981 if (current_order >= pageblock_order)
982 change_pageblock_range(page, current_order,
985 expand(zone, page, order, current_order, area, migratetype);
987 trace_mm_page_alloc_extfrag(page, order, current_order,
988 start_migratetype, migratetype);
998 * Do the hard work of removing an element from the buddy allocator.
999 * Call me with the zone->lock already held.
1001 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1007 page = __rmqueue_smallest(zone, order, migratetype);
1009 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1010 page = __rmqueue_fallback(zone, order, migratetype);
1013 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1014 * is used because __rmqueue_smallest is an inline function
1015 * and we want just one call site
1018 migratetype = MIGRATE_RESERVE;
1023 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1028 * Obtain a specified number of elements from the buddy allocator, all under
1029 * a single hold of the lock, for efficiency. Add them to the supplied list.
1030 * Returns the number of new pages which were placed at *list.
1032 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1033 unsigned long count, struct list_head *list,
1034 int migratetype, int cold)
1038 spin_lock(&zone->lock);
1039 for (i = 0; i < count; ++i) {
1040 struct page *page = __rmqueue(zone, order, migratetype);
1041 if (unlikely(page == NULL))
1045 * Split buddy pages returned by expand() are received here
1046 * in physical page order. The page is added to the callers and
1047 * list and the list head then moves forward. From the callers
1048 * perspective, the linked list is ordered by page number in
1049 * some conditions. This is useful for IO devices that can
1050 * merge IO requests if the physical pages are ordered
1053 if (likely(cold == 0))
1054 list_add(&page->lru, list);
1056 list_add_tail(&page->lru, list);
1057 set_page_private(page, migratetype);
1060 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1061 spin_unlock(&zone->lock);
1067 * Called from the vmstat counter updater to drain pagesets of this
1068 * currently executing processor on remote nodes after they have
1071 * Note that this function must be called with the thread pinned to
1072 * a single processor.
1074 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1076 unsigned long flags;
1079 local_irq_save(flags);
1080 if (pcp->count >= pcp->batch)
1081 to_drain = pcp->batch;
1083 to_drain = pcp->count;
1084 free_pcppages_bulk(zone, to_drain, pcp);
1085 pcp->count -= to_drain;
1086 local_irq_restore(flags);
1091 * Drain pages of the indicated processor.
1093 * The processor must either be the current processor and the
1094 * thread pinned to the current processor or a processor that
1097 static void drain_pages(unsigned int cpu)
1099 unsigned long flags;
1102 for_each_populated_zone(zone) {
1103 struct per_cpu_pageset *pset;
1104 struct per_cpu_pages *pcp;
1106 local_irq_save(flags);
1107 pset = per_cpu_ptr(zone->pageset, cpu);
1111 free_pcppages_bulk(zone, pcp->count, pcp);
1114 local_irq_restore(flags);
1119 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1121 void drain_local_pages(void *arg)
1123 drain_pages(smp_processor_id());
1127 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1129 void drain_all_pages(void)
1131 on_each_cpu(drain_local_pages, NULL, 1);
1134 #ifdef CONFIG_HIBERNATION
1136 void mark_free_pages(struct zone *zone)
1138 unsigned long pfn, max_zone_pfn;
1139 unsigned long flags;
1141 struct list_head *curr;
1143 if (!zone->spanned_pages)
1146 spin_lock_irqsave(&zone->lock, flags);
1148 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1149 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1150 if (pfn_valid(pfn)) {
1151 struct page *page = pfn_to_page(pfn);
1153 if (!swsusp_page_is_forbidden(page))
1154 swsusp_unset_page_free(page);
1157 for_each_migratetype_order(order, t) {
1158 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1161 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1162 for (i = 0; i < (1UL << order); i++)
1163 swsusp_set_page_free(pfn_to_page(pfn + i));
1166 spin_unlock_irqrestore(&zone->lock, flags);
1168 #endif /* CONFIG_PM */
1171 * Free a 0-order page
1172 * cold == 1 ? free a cold page : free a hot page
1174 void free_hot_cold_page(struct page *page, int cold)
1176 struct zone *zone = page_zone(page);
1177 struct per_cpu_pages *pcp;
1178 unsigned long flags;
1180 int wasMlocked = __TestClearPageMlocked(page);
1183 WARN_ON(PageForeign(page) && wasMlocked);
1185 if (!free_pages_prepare(page, 0))
1188 migratetype = get_pageblock_migratetype(page);
1189 set_page_private(page, migratetype);
1190 local_irq_save(flags);
1191 if (unlikely(wasMlocked))
1192 free_page_mlock(page);
1193 __count_vm_event(PGFREE);
1196 * We only track unmovable, reclaimable and movable on pcp lists.
1197 * Free ISOLATE pages back to the allocator because they are being
1198 * offlined but treat RESERVE as movable pages so we can get those
1199 * areas back if necessary. Otherwise, we may have to free
1200 * excessively into the page allocator
1202 if (migratetype >= MIGRATE_PCPTYPES) {
1203 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1204 free_one_page(zone, page, 0, migratetype);
1207 migratetype = MIGRATE_MOVABLE;
1210 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1212 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1214 list_add(&page->lru, &pcp->lists[migratetype]);
1216 if (pcp->count >= pcp->high) {
1217 free_pcppages_bulk(zone, pcp->batch, pcp);
1218 pcp->count -= pcp->batch;
1222 local_irq_restore(flags);
1226 * split_page takes a non-compound higher-order page, and splits it into
1227 * n (1<<order) sub-pages: page[0..n]
1228 * Each sub-page must be freed individually.
1230 * Note: this is probably too low level an operation for use in drivers.
1231 * Please consult with lkml before using this in your driver.
1233 void split_page(struct page *page, unsigned int order)
1237 VM_BUG_ON(PageCompound(page));
1238 VM_BUG_ON(!page_count(page));
1240 #ifdef CONFIG_KMEMCHECK
1242 * Split shadow pages too, because free(page[0]) would
1243 * otherwise free the whole shadow.
1245 if (kmemcheck_page_is_tracked(page))
1246 split_page(virt_to_page(page[0].shadow), order);
1249 for (i = 1; i < (1 << order); i++)
1250 set_page_refcounted(page + i);
1254 * Similar to split_page except the page is already free. As this is only
1255 * being used for migration, the migratetype of the block also changes.
1256 * As this is called with interrupts disabled, the caller is responsible
1257 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1260 * Note: this is probably too low level an operation for use in drivers.
1261 * Please consult with lkml before using this in your driver.
1263 int split_free_page(struct page *page)
1266 unsigned long watermark;
1269 BUG_ON(!PageBuddy(page));
1271 zone = page_zone(page);
1272 order = page_order(page);
1274 /* Obey watermarks as if the page was being allocated */
1275 watermark = low_wmark_pages(zone) + (1 << order);
1276 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1279 /* Remove page from free list */
1280 list_del(&page->lru);
1281 zone->free_area[order].nr_free--;
1282 rmv_page_order(page);
1283 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1285 /* Split into individual pages */
1286 set_page_refcounted(page);
1287 split_page(page, order);
1289 if (order >= pageblock_order - 1) {
1290 struct page *endpage = page + (1 << order) - 1;
1291 for (; page < endpage; page += pageblock_nr_pages)
1292 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1299 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1300 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1304 struct page *buffered_rmqueue(struct zone *preferred_zone,
1305 struct zone *zone, int order, gfp_t gfp_flags,
1308 unsigned long flags;
1310 int cold = !!(gfp_flags & __GFP_COLD);
1313 if (likely(order == 0)) {
1314 struct per_cpu_pages *pcp;
1315 struct list_head *list;
1317 local_irq_save(flags);
1318 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1319 list = &pcp->lists[migratetype];
1320 if (list_empty(list)) {
1321 pcp->count += rmqueue_bulk(zone, 0,
1324 if (unlikely(list_empty(list)))
1329 page = list_entry(list->prev, struct page, lru);
1331 page = list_entry(list->next, struct page, lru);
1333 list_del(&page->lru);
1336 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1338 * __GFP_NOFAIL is not to be used in new code.
1340 * All __GFP_NOFAIL callers should be fixed so that they
1341 * properly detect and handle allocation failures.
1343 * We most definitely don't want callers attempting to
1344 * allocate greater than order-1 page units with
1347 WARN_ON_ONCE(order > 1);
1349 spin_lock_irqsave(&zone->lock, flags);
1350 page = __rmqueue(zone, order, migratetype);
1351 spin_unlock(&zone->lock);
1354 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1357 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1358 zone_statistics(preferred_zone, zone, gfp_flags);
1359 local_irq_restore(flags);
1361 VM_BUG_ON(bad_range(zone, page));
1362 if (prep_new_page(page, order, gfp_flags))
1367 local_irq_restore(flags);
1371 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1372 #define ALLOC_WMARK_MIN WMARK_MIN
1373 #define ALLOC_WMARK_LOW WMARK_LOW
1374 #define ALLOC_WMARK_HIGH WMARK_HIGH
1375 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1377 /* Mask to get the watermark bits */
1378 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1380 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1381 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1382 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1384 #ifdef CONFIG_FAIL_PAGE_ALLOC
1387 struct fault_attr attr;
1389 u32 ignore_gfp_highmem;
1390 u32 ignore_gfp_wait;
1392 } fail_page_alloc = {
1393 .attr = FAULT_ATTR_INITIALIZER,
1394 .ignore_gfp_wait = 1,
1395 .ignore_gfp_highmem = 1,
1399 static int __init setup_fail_page_alloc(char *str)
1401 return setup_fault_attr(&fail_page_alloc.attr, str);
1403 __setup("fail_page_alloc=", setup_fail_page_alloc);
1405 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1407 if (order < fail_page_alloc.min_order)
1409 if (gfp_mask & __GFP_NOFAIL)
1411 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1413 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1416 return should_fail(&fail_page_alloc.attr, 1 << order);
1419 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1421 static int __init fail_page_alloc_debugfs(void)
1423 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1426 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1427 &fail_page_alloc.attr);
1429 return PTR_ERR(dir);
1431 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1432 &fail_page_alloc.ignore_gfp_wait))
1434 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1435 &fail_page_alloc.ignore_gfp_highmem))
1437 if (!debugfs_create_u32("min-order", mode, dir,
1438 &fail_page_alloc.min_order))
1443 debugfs_remove_recursive(dir);
1448 late_initcall(fail_page_alloc_debugfs);
1450 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1452 #else /* CONFIG_FAIL_PAGE_ALLOC */
1454 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1459 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1462 * Return true if free pages are above 'mark'. This takes into account the order
1463 * of the allocation.
1465 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1466 int classzone_idx, int alloc_flags, long free_pages)
1468 /* free_pages my go negative - that's OK */
1472 free_pages -= (1 << order) + 1;
1473 if (alloc_flags & ALLOC_HIGH)
1475 if (alloc_flags & ALLOC_HARDER)
1478 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1480 for (o = 0; o < order; o++) {
1481 /* At the next order, this order's pages become unavailable */
1482 free_pages -= z->free_area[o].nr_free << o;
1484 /* Require fewer higher order pages to be free */
1487 if (free_pages <= min)
1493 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1494 int classzone_idx, int alloc_flags)
1496 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1497 zone_page_state(z, NR_FREE_PAGES));
1500 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1501 int classzone_idx, int alloc_flags)
1503 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1505 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1506 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1508 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1514 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1515 * skip over zones that are not allowed by the cpuset, or that have
1516 * been recently (in last second) found to be nearly full. See further
1517 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1518 * that have to skip over a lot of full or unallowed zones.
1520 * If the zonelist cache is present in the passed in zonelist, then
1521 * returns a pointer to the allowed node mask (either the current
1522 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1524 * If the zonelist cache is not available for this zonelist, does
1525 * nothing and returns NULL.
1527 * If the fullzones BITMAP in the zonelist cache is stale (more than
1528 * a second since last zap'd) then we zap it out (clear its bits.)
1530 * We hold off even calling zlc_setup, until after we've checked the
1531 * first zone in the zonelist, on the theory that most allocations will
1532 * be satisfied from that first zone, so best to examine that zone as
1533 * quickly as we can.
1535 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1537 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1538 nodemask_t *allowednodes; /* zonelist_cache approximation */
1540 zlc = zonelist->zlcache_ptr;
1544 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1545 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1546 zlc->last_full_zap = jiffies;
1549 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1550 &cpuset_current_mems_allowed :
1551 &node_states[N_HIGH_MEMORY];
1552 return allowednodes;
1556 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1557 * if it is worth looking at further for free memory:
1558 * 1) Check that the zone isn't thought to be full (doesn't have its
1559 * bit set in the zonelist_cache fullzones BITMAP).
1560 * 2) Check that the zones node (obtained from the zonelist_cache
1561 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1562 * Return true (non-zero) if zone is worth looking at further, or
1563 * else return false (zero) if it is not.
1565 * This check -ignores- the distinction between various watermarks,
1566 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1567 * found to be full for any variation of these watermarks, it will
1568 * be considered full for up to one second by all requests, unless
1569 * we are so low on memory on all allowed nodes that we are forced
1570 * into the second scan of the zonelist.
1572 * In the second scan we ignore this zonelist cache and exactly
1573 * apply the watermarks to all zones, even it is slower to do so.
1574 * We are low on memory in the second scan, and should leave no stone
1575 * unturned looking for a free page.
1577 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1578 nodemask_t *allowednodes)
1580 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1581 int i; /* index of *z in zonelist zones */
1582 int n; /* node that zone *z is on */
1584 zlc = zonelist->zlcache_ptr;
1588 i = z - zonelist->_zonerefs;
1591 /* This zone is worth trying if it is allowed but not full */
1592 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1596 * Given 'z' scanning a zonelist, set the corresponding bit in
1597 * zlc->fullzones, so that subsequent attempts to allocate a page
1598 * from that zone don't waste time re-examining it.
1600 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1602 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1603 int i; /* index of *z in zonelist zones */
1605 zlc = zonelist->zlcache_ptr;
1609 i = z - zonelist->_zonerefs;
1611 set_bit(i, zlc->fullzones);
1615 * clear all zones full, called after direct reclaim makes progress so that
1616 * a zone that was recently full is not skipped over for up to a second
1618 static void zlc_clear_zones_full(struct zonelist *zonelist)
1620 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1622 zlc = zonelist->zlcache_ptr;
1626 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1629 #else /* CONFIG_NUMA */
1631 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1636 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1637 nodemask_t *allowednodes)
1642 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1646 static void zlc_clear_zones_full(struct zonelist *zonelist)
1649 #endif /* CONFIG_NUMA */
1652 * get_page_from_freelist goes through the zonelist trying to allocate
1655 static struct page *
1656 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1657 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1658 struct zone *preferred_zone, int migratetype)
1661 struct page *page = NULL;
1664 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1665 int zlc_active = 0; /* set if using zonelist_cache */
1666 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1668 classzone_idx = zone_idx(preferred_zone);
1671 * Scan zonelist, looking for a zone with enough free.
1672 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1674 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1675 high_zoneidx, nodemask) {
1676 if (NUMA_BUILD && zlc_active &&
1677 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1679 if ((alloc_flags & ALLOC_CPUSET) &&
1680 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1683 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1684 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1688 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1689 if (zone_watermark_ok(zone, order, mark,
1690 classzone_idx, alloc_flags))
1693 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1695 * we do zlc_setup if there are multiple nodes
1696 * and before considering the first zone allowed
1699 allowednodes = zlc_setup(zonelist, alloc_flags);
1704 if (zone_reclaim_mode == 0)
1705 goto this_zone_full;
1708 * As we may have just activated ZLC, check if the first
1709 * eligible zone has failed zone_reclaim recently.
1711 if (NUMA_BUILD && zlc_active &&
1712 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1715 ret = zone_reclaim(zone, gfp_mask, order);
1717 case ZONE_RECLAIM_NOSCAN:
1720 case ZONE_RECLAIM_FULL:
1721 /* scanned but unreclaimable */
1724 /* did we reclaim enough */
1725 if (!zone_watermark_ok(zone, order, mark,
1726 classzone_idx, alloc_flags))
1727 goto this_zone_full;
1732 page = buffered_rmqueue(preferred_zone, zone, order,
1733 gfp_mask, migratetype);
1738 zlc_mark_zone_full(zonelist, z);
1741 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1742 /* Disable zlc cache for second zonelist scan */
1750 * Large machines with many possible nodes should not always dump per-node
1751 * meminfo in irq context.
1753 static inline bool should_suppress_show_mem(void)
1758 ret = in_interrupt();
1763 static DEFINE_RATELIMIT_STATE(nopage_rs,
1764 DEFAULT_RATELIMIT_INTERVAL,
1765 DEFAULT_RATELIMIT_BURST);
1767 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1770 unsigned int filter = SHOW_MEM_FILTER_NODES;
1772 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1776 * This documents exceptions given to allocations in certain
1777 * contexts that are allowed to allocate outside current's set
1780 if (!(gfp_mask & __GFP_NOMEMALLOC))
1781 if (test_thread_flag(TIF_MEMDIE) ||
1782 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1783 filter &= ~SHOW_MEM_FILTER_NODES;
1784 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1785 filter &= ~SHOW_MEM_FILTER_NODES;
1788 printk(KERN_WARNING);
1789 va_start(args, fmt);
1794 if (!(gfp_mask & __GFP_WAIT)) {
1795 pr_info("The following is only an harmless informational message.\n");
1796 pr_info("Unless you get a _continuous_flood_ of these messages it means\n");
1797 pr_info("everything is working fine. Allocations from irqs cannot be\n");
1798 pr_info("perfectly reliable and the kernel is designed to handle that.\n");
1800 pr_info("%s: page allocation failure. order:%d, mode:0x%x\n",
1801 current->comm, order, gfp_mask);
1804 if (!should_suppress_show_mem())
1809 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1810 unsigned long pages_reclaimed)
1812 /* Do not loop if specifically requested */
1813 if (gfp_mask & __GFP_NORETRY)
1817 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1818 * means __GFP_NOFAIL, but that may not be true in other
1821 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1825 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1826 * specified, then we retry until we no longer reclaim any pages
1827 * (above), or we've reclaimed an order of pages at least as
1828 * large as the allocation's order. In both cases, if the
1829 * allocation still fails, we stop retrying.
1831 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1835 * Don't let big-order allocations loop unless the caller
1836 * explicitly requests that.
1838 if (gfp_mask & __GFP_NOFAIL)
1844 static inline struct page *
1845 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1846 struct zonelist *zonelist, enum zone_type high_zoneidx,
1847 nodemask_t *nodemask, struct zone *preferred_zone,
1852 /* Acquire the OOM killer lock for the zones in zonelist */
1853 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1854 schedule_timeout_uninterruptible(1);
1859 * Go through the zonelist yet one more time, keep very high watermark
1860 * here, this is only to catch a parallel oom killing, we must fail if
1861 * we're still under heavy pressure.
1863 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1864 order, zonelist, high_zoneidx,
1865 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1866 preferred_zone, migratetype);
1870 if (!(gfp_mask & __GFP_NOFAIL)) {
1871 /* The OOM killer will not help higher order allocs */
1872 if (order > PAGE_ALLOC_COSTLY_ORDER)
1874 /* The OOM killer does not needlessly kill tasks for lowmem */
1875 if (high_zoneidx < ZONE_NORMAL)
1878 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1879 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1880 * The caller should handle page allocation failure by itself if
1881 * it specifies __GFP_THISNODE.
1882 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1884 if (gfp_mask & __GFP_THISNODE)
1887 /* Exhausted what can be done so it's blamo time */
1888 out_of_memory(zonelist, gfp_mask, order, nodemask);
1891 clear_zonelist_oom(zonelist, gfp_mask);
1895 #ifdef CONFIG_COMPACTION
1896 /* Try memory compaction for high-order allocations before reclaim */
1897 static struct page *
1898 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1899 struct zonelist *zonelist, enum zone_type high_zoneidx,
1900 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1901 int migratetype, unsigned long *did_some_progress,
1902 bool sync_migration)
1906 if (!order || compaction_deferred(preferred_zone))
1909 current->flags |= PF_MEMALLOC;
1910 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1911 nodemask, sync_migration);
1912 current->flags &= ~PF_MEMALLOC;
1913 if (*did_some_progress != COMPACT_SKIPPED) {
1915 /* Page migration frees to the PCP lists but we want merging */
1916 drain_pages(get_cpu());
1919 page = get_page_from_freelist(gfp_mask, nodemask,
1920 order, zonelist, high_zoneidx,
1921 alloc_flags, preferred_zone,
1924 preferred_zone->compact_considered = 0;
1925 preferred_zone->compact_defer_shift = 0;
1926 count_vm_event(COMPACTSUCCESS);
1931 * It's bad if compaction run occurs and fails.
1932 * The most likely reason is that pages exist,
1933 * but not enough to satisfy watermarks.
1935 count_vm_event(COMPACTFAIL);
1936 defer_compaction(preferred_zone);
1944 static inline struct page *
1945 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1946 struct zonelist *zonelist, enum zone_type high_zoneidx,
1947 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1948 int migratetype, unsigned long *did_some_progress,
1949 bool sync_migration)
1953 #endif /* CONFIG_COMPACTION */
1955 /* The really slow allocator path where we enter direct reclaim */
1956 static inline struct page *
1957 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1958 struct zonelist *zonelist, enum zone_type high_zoneidx,
1959 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1960 int migratetype, unsigned long *did_some_progress)
1962 struct page *page = NULL;
1963 struct reclaim_state reclaim_state;
1964 bool drained = false;
1968 /* We now go into synchronous reclaim */
1969 cpuset_memory_pressure_bump();
1970 current->flags |= PF_MEMALLOC;
1971 lockdep_set_current_reclaim_state(gfp_mask);
1972 reclaim_state.reclaimed_slab = 0;
1973 current->reclaim_state = &reclaim_state;
1975 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1977 current->reclaim_state = NULL;
1978 lockdep_clear_current_reclaim_state();
1979 current->flags &= ~PF_MEMALLOC;
1983 if (unlikely(!(*did_some_progress)))
1986 /* After successful reclaim, reconsider all zones for allocation */
1988 zlc_clear_zones_full(zonelist);
1991 page = get_page_from_freelist(gfp_mask, nodemask, order,
1992 zonelist, high_zoneidx,
1993 alloc_flags, preferred_zone,
1997 * If an allocation failed after direct reclaim, it could be because
1998 * pages are pinned on the per-cpu lists. Drain them and try again
2000 if (!page && !drained) {
2010 * This is called in the allocator slow-path if the allocation request is of
2011 * sufficient urgency to ignore watermarks and take other desperate measures
2013 static inline struct page *
2014 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2015 struct zonelist *zonelist, enum zone_type high_zoneidx,
2016 nodemask_t *nodemask, struct zone *preferred_zone,
2022 page = get_page_from_freelist(gfp_mask, nodemask, order,
2023 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2024 preferred_zone, migratetype);
2026 if (!page && gfp_mask & __GFP_NOFAIL)
2027 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2028 } while (!page && (gfp_mask & __GFP_NOFAIL));
2034 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2035 enum zone_type high_zoneidx,
2036 enum zone_type classzone_idx)
2041 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2042 wakeup_kswapd(zone, order, classzone_idx);
2046 gfp_to_alloc_flags(gfp_t gfp_mask)
2048 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2049 const gfp_t wait = gfp_mask & __GFP_WAIT;
2051 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2052 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2055 * The caller may dip into page reserves a bit more if the caller
2056 * cannot run direct reclaim, or if the caller has realtime scheduling
2057 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2058 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2060 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2064 * Not worth trying to allocate harder for
2065 * __GFP_NOMEMALLOC even if it can't schedule.
2067 if (!(gfp_mask & __GFP_NOMEMALLOC))
2068 alloc_flags |= ALLOC_HARDER;
2070 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2071 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2073 alloc_flags &= ~ALLOC_CPUSET;
2074 } else if (unlikely(rt_task(current)) && !in_interrupt())
2075 alloc_flags |= ALLOC_HARDER;
2077 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2078 if (!in_interrupt() &&
2079 ((current->flags & PF_MEMALLOC) ||
2080 unlikely(test_thread_flag(TIF_MEMDIE))))
2081 alloc_flags |= ALLOC_NO_WATERMARKS;
2087 static inline struct page *
2088 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2089 struct zonelist *zonelist, enum zone_type high_zoneidx,
2090 nodemask_t *nodemask, struct zone *preferred_zone,
2093 const gfp_t wait = gfp_mask & __GFP_WAIT;
2094 struct page *page = NULL;
2096 unsigned long pages_reclaimed = 0;
2097 unsigned long did_some_progress;
2098 bool sync_migration = false;
2101 * In the slowpath, we sanity check order to avoid ever trying to
2102 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2103 * be using allocators in order of preference for an area that is
2106 if (order >= MAX_ORDER) {
2107 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2112 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2113 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2114 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2115 * using a larger set of nodes after it has established that the
2116 * allowed per node queues are empty and that nodes are
2119 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2123 if (!(gfp_mask & __GFP_NO_KSWAPD))
2124 wake_all_kswapd(order, zonelist, high_zoneidx,
2125 zone_idx(preferred_zone));
2128 * OK, we're below the kswapd watermark and have kicked background
2129 * reclaim. Now things get more complex, so set up alloc_flags according
2130 * to how we want to proceed.
2132 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2135 * Find the true preferred zone if the allocation is unconstrained by
2138 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2139 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2143 /* This is the last chance, in general, before the goto nopage. */
2144 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2145 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2146 preferred_zone, migratetype);
2150 /* Allocate without watermarks if the context allows */
2151 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2152 page = __alloc_pages_high_priority(gfp_mask, order,
2153 zonelist, high_zoneidx, nodemask,
2154 preferred_zone, migratetype);
2159 /* Atomic allocations - we can't balance anything */
2163 /* Avoid recursion of direct reclaim */
2164 if (current->flags & PF_MEMALLOC)
2167 /* Avoid allocations with no watermarks from looping endlessly */
2168 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2172 * Try direct compaction. The first pass is asynchronous. Subsequent
2173 * attempts after direct reclaim are synchronous
2175 page = __alloc_pages_direct_compact(gfp_mask, order,
2176 zonelist, high_zoneidx,
2178 alloc_flags, preferred_zone,
2179 migratetype, &did_some_progress,
2183 sync_migration = true;
2185 /* Try direct reclaim and then allocating */
2186 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2187 zonelist, high_zoneidx,
2189 alloc_flags, preferred_zone,
2190 migratetype, &did_some_progress);
2195 * If we failed to make any progress reclaiming, then we are
2196 * running out of options and have to consider going OOM
2198 if (!did_some_progress) {
2199 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2200 if (oom_killer_disabled)
2202 page = __alloc_pages_may_oom(gfp_mask, order,
2203 zonelist, high_zoneidx,
2204 nodemask, preferred_zone,
2209 if (!(gfp_mask & __GFP_NOFAIL)) {
2211 * The oom killer is not called for high-order
2212 * allocations that may fail, so if no progress
2213 * is being made, there are no other options and
2214 * retrying is unlikely to help.
2216 if (order > PAGE_ALLOC_COSTLY_ORDER)
2219 * The oom killer is not called for lowmem
2220 * allocations to prevent needlessly killing
2223 if (high_zoneidx < ZONE_NORMAL)
2231 /* Check if we should retry the allocation */
2232 pages_reclaimed += did_some_progress;
2233 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2234 /* Wait for some write requests to complete then retry */
2235 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2239 * High-order allocations do not necessarily loop after
2240 * direct reclaim and reclaim/compaction depends on compaction
2241 * being called after reclaim so call directly if necessary
2243 page = __alloc_pages_direct_compact(gfp_mask, order,
2244 zonelist, high_zoneidx,
2246 alloc_flags, preferred_zone,
2247 migratetype, &did_some_progress,
2254 warn_alloc_failed(gfp_mask, order, NULL);
2257 if (kmemcheck_enabled)
2258 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2264 * This is the 'heart' of the zoned buddy allocator.
2267 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2268 struct zonelist *zonelist, nodemask_t *nodemask)
2270 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2271 struct zone *preferred_zone;
2273 int migratetype = allocflags_to_migratetype(gfp_mask);
2275 gfp_mask &= gfp_allowed_mask;
2277 lockdep_trace_alloc(gfp_mask);
2279 might_sleep_if(gfp_mask & __GFP_WAIT);
2281 if (should_fail_alloc_page(gfp_mask, order))
2285 * Check the zones suitable for the gfp_mask contain at least one
2286 * valid zone. It's possible to have an empty zonelist as a result
2287 * of GFP_THISNODE and a memoryless node
2289 if (unlikely(!zonelist->_zonerefs->zone))
2293 /* The preferred zone is used for statistics later */
2294 first_zones_zonelist(zonelist, high_zoneidx,
2295 nodemask ? : &cpuset_current_mems_allowed,
2297 if (!preferred_zone) {
2302 /* First allocation attempt */
2303 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2304 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2305 preferred_zone, migratetype);
2306 if (unlikely(!page))
2307 page = __alloc_pages_slowpath(gfp_mask, order,
2308 zonelist, high_zoneidx, nodemask,
2309 preferred_zone, migratetype);
2312 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2315 EXPORT_SYMBOL(__alloc_pages_nodemask);
2318 * Common helper functions.
2320 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2325 * __get_free_pages() returns a 32-bit address, which cannot represent
2328 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2330 page = alloc_pages(gfp_mask, order);
2333 return (unsigned long) page_address(page);
2335 EXPORT_SYMBOL(__get_free_pages);
2337 unsigned long get_zeroed_page(gfp_t gfp_mask)
2339 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2341 EXPORT_SYMBOL(get_zeroed_page);
2343 void __pagevec_free(struct pagevec *pvec)
2345 int i = pagevec_count(pvec);
2348 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2349 free_hot_cold_page(pvec->pages[i], pvec->cold);
2353 void __free_pages(struct page *page, unsigned int order)
2355 if (put_page_testzero(page)) {
2357 free_hot_cold_page(page, 0);
2359 __free_pages_ok(page, order);
2363 EXPORT_SYMBOL(__free_pages);
2365 void free_pages(unsigned long addr, unsigned int order)
2368 VM_BUG_ON(!virt_addr_valid((void *)addr));
2369 __free_pages(virt_to_page((void *)addr), order);
2373 EXPORT_SYMBOL(free_pages);
2375 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2378 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2379 unsigned long used = addr + PAGE_ALIGN(size);
2381 split_page(virt_to_page((void *)addr), order);
2382 while (used < alloc_end) {
2387 return (void *)addr;
2391 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2392 * @size: the number of bytes to allocate
2393 * @gfp_mask: GFP flags for the allocation
2395 * This function is similar to alloc_pages(), except that it allocates the
2396 * minimum number of pages to satisfy the request. alloc_pages() can only
2397 * allocate memory in power-of-two pages.
2399 * This function is also limited by MAX_ORDER.
2401 * Memory allocated by this function must be released by free_pages_exact().
2403 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2405 unsigned int order = get_order(size);
2408 addr = __get_free_pages(gfp_mask, order);
2409 return make_alloc_exact(addr, order, size);
2411 EXPORT_SYMBOL(alloc_pages_exact);
2414 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2416 * @nid: the preferred node ID where memory should be allocated
2417 * @size: the number of bytes to allocate
2418 * @gfp_mask: GFP flags for the allocation
2420 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2422 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2425 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2427 unsigned order = get_order(size);
2428 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2431 return make_alloc_exact((unsigned long)page_address(p), order, size);
2433 EXPORT_SYMBOL(alloc_pages_exact_nid);
2436 * free_pages_exact - release memory allocated via alloc_pages_exact()
2437 * @virt: the value returned by alloc_pages_exact.
2438 * @size: size of allocation, same value as passed to alloc_pages_exact().
2440 * Release the memory allocated by a previous call to alloc_pages_exact.
2442 void free_pages_exact(void *virt, size_t size)
2444 unsigned long addr = (unsigned long)virt;
2445 unsigned long end = addr + PAGE_ALIGN(size);
2447 while (addr < end) {
2452 EXPORT_SYMBOL(free_pages_exact);
2454 static unsigned int nr_free_zone_pages(int offset)
2459 /* Just pick one node, since fallback list is circular */
2460 unsigned int sum = 0;
2462 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2464 for_each_zone_zonelist(zone, z, zonelist, offset) {
2465 unsigned long size = zone->present_pages;
2466 unsigned long high = high_wmark_pages(zone);
2475 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2477 unsigned int nr_free_buffer_pages(void)
2479 return nr_free_zone_pages(gfp_zone(GFP_USER));
2481 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2484 * Amount of free RAM allocatable within all zones
2486 unsigned int nr_free_pagecache_pages(void)
2488 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2491 static inline void show_node(struct zone *zone)
2494 printk("Node %d ", zone_to_nid(zone));
2497 void si_meminfo(struct sysinfo *val)
2499 val->totalram = totalram_pages;
2501 val->freeram = global_page_state(NR_FREE_PAGES);
2502 val->bufferram = nr_blockdev_pages();
2503 val->totalhigh = totalhigh_pages;
2504 val->freehigh = nr_free_highpages();
2505 val->mem_unit = PAGE_SIZE;
2508 EXPORT_SYMBOL(si_meminfo);
2511 void si_meminfo_node(struct sysinfo *val, int nid)
2513 pg_data_t *pgdat = NODE_DATA(nid);
2515 val->totalram = pgdat->node_present_pages;
2516 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2517 #ifdef CONFIG_HIGHMEM
2518 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2519 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2525 val->mem_unit = PAGE_SIZE;
2530 * Determine whether the node should be displayed or not, depending on whether
2531 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2533 bool skip_free_areas_node(unsigned int flags, int nid)
2537 if (!(flags & SHOW_MEM_FILTER_NODES))
2541 ret = !node_isset(nid, cpuset_current_mems_allowed);
2547 #define K(x) ((x) << (PAGE_SHIFT-10))
2550 * Show free area list (used inside shift_scroll-lock stuff)
2551 * We also calculate the percentage fragmentation. We do this by counting the
2552 * memory on each free list with the exception of the first item on the list.
2553 * Suppresses nodes that are not allowed by current's cpuset if
2554 * SHOW_MEM_FILTER_NODES is passed.
2556 void show_free_areas(unsigned int filter)
2561 for_each_populated_zone(zone) {
2562 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2565 printk("%s per-cpu:\n", zone->name);
2567 for_each_online_cpu(cpu) {
2568 struct per_cpu_pageset *pageset;
2570 pageset = per_cpu_ptr(zone->pageset, cpu);
2572 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2573 cpu, pageset->pcp.high,
2574 pageset->pcp.batch, pageset->pcp.count);
2578 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2579 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2581 " dirty:%lu writeback:%lu unstable:%lu\n"
2582 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2583 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2584 global_page_state(NR_ACTIVE_ANON),
2585 global_page_state(NR_INACTIVE_ANON),
2586 global_page_state(NR_ISOLATED_ANON),
2587 global_page_state(NR_ACTIVE_FILE),
2588 global_page_state(NR_INACTIVE_FILE),
2589 global_page_state(NR_ISOLATED_FILE),
2590 global_page_state(NR_UNEVICTABLE),
2591 global_page_state(NR_FILE_DIRTY),
2592 global_page_state(NR_WRITEBACK),
2593 global_page_state(NR_UNSTABLE_NFS),
2594 global_page_state(NR_FREE_PAGES),
2595 global_page_state(NR_SLAB_RECLAIMABLE),
2596 global_page_state(NR_SLAB_UNRECLAIMABLE),
2597 global_page_state(NR_FILE_MAPPED),
2598 global_page_state(NR_SHMEM),
2599 global_page_state(NR_PAGETABLE),
2600 global_page_state(NR_BOUNCE));
2602 for_each_populated_zone(zone) {
2605 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2613 " active_anon:%lukB"
2614 " inactive_anon:%lukB"
2615 " active_file:%lukB"
2616 " inactive_file:%lukB"
2617 " unevictable:%lukB"
2618 " isolated(anon):%lukB"
2619 " isolated(file):%lukB"
2626 " slab_reclaimable:%lukB"
2627 " slab_unreclaimable:%lukB"
2628 " kernel_stack:%lukB"
2632 " writeback_tmp:%lukB"
2633 " pages_scanned:%lu"
2634 " all_unreclaimable? %s"
2637 K(zone_page_state(zone, NR_FREE_PAGES)),
2638 K(min_wmark_pages(zone)),
2639 K(low_wmark_pages(zone)),
2640 K(high_wmark_pages(zone)),
2641 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2642 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2643 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2644 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2645 K(zone_page_state(zone, NR_UNEVICTABLE)),
2646 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2647 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2648 K(zone->present_pages),
2649 K(zone_page_state(zone, NR_MLOCK)),
2650 K(zone_page_state(zone, NR_FILE_DIRTY)),
2651 K(zone_page_state(zone, NR_WRITEBACK)),
2652 K(zone_page_state(zone, NR_FILE_MAPPED)),
2653 K(zone_page_state(zone, NR_SHMEM)),
2654 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2655 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2656 zone_page_state(zone, NR_KERNEL_STACK) *
2658 K(zone_page_state(zone, NR_PAGETABLE)),
2659 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2660 K(zone_page_state(zone, NR_BOUNCE)),
2661 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2662 zone->pages_scanned,
2663 (zone->all_unreclaimable ? "yes" : "no")
2665 printk("lowmem_reserve[]:");
2666 for (i = 0; i < MAX_NR_ZONES; i++)
2667 printk(" %lu", zone->lowmem_reserve[i]);
2671 for_each_populated_zone(zone) {
2672 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2674 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2677 printk("%s: ", zone->name);
2679 spin_lock_irqsave(&zone->lock, flags);
2680 for (order = 0; order < MAX_ORDER; order++) {
2681 nr[order] = zone->free_area[order].nr_free;
2682 total += nr[order] << order;
2684 spin_unlock_irqrestore(&zone->lock, flags);
2685 for (order = 0; order < MAX_ORDER; order++)
2686 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2687 printk("= %lukB\n", K(total));
2690 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2692 show_swap_cache_info();
2695 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2697 zoneref->zone = zone;
2698 zoneref->zone_idx = zone_idx(zone);
2702 * Builds allocation fallback zone lists.
2704 * Add all populated zones of a node to the zonelist.
2706 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2707 int nr_zones, enum zone_type zone_type)
2711 BUG_ON(zone_type >= MAX_NR_ZONES);
2716 zone = pgdat->node_zones + zone_type;
2717 if (populated_zone(zone)) {
2718 zoneref_set_zone(zone,
2719 &zonelist->_zonerefs[nr_zones++]);
2720 check_highest_zone(zone_type);
2723 } while (zone_type);
2730 * 0 = automatic detection of better ordering.
2731 * 1 = order by ([node] distance, -zonetype)
2732 * 2 = order by (-zonetype, [node] distance)
2734 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2735 * the same zonelist. So only NUMA can configure this param.
2737 #define ZONELIST_ORDER_DEFAULT 0
2738 #define ZONELIST_ORDER_NODE 1
2739 #define ZONELIST_ORDER_ZONE 2
2741 /* zonelist order in the kernel.
2742 * set_zonelist_order() will set this to NODE or ZONE.
2744 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2745 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2749 /* The value user specified ....changed by config */
2750 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2751 /* string for sysctl */
2752 #define NUMA_ZONELIST_ORDER_LEN 16
2753 char numa_zonelist_order[16] = "default";
2756 * interface for configure zonelist ordering.
2757 * command line option "numa_zonelist_order"
2758 * = "[dD]efault - default, automatic configuration.
2759 * = "[nN]ode - order by node locality, then by zone within node
2760 * = "[zZ]one - order by zone, then by locality within zone
2763 static int __parse_numa_zonelist_order(char *s)
2765 if (*s == 'd' || *s == 'D') {
2766 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2767 } else if (*s == 'n' || *s == 'N') {
2768 user_zonelist_order = ZONELIST_ORDER_NODE;
2769 } else if (*s == 'z' || *s == 'Z') {
2770 user_zonelist_order = ZONELIST_ORDER_ZONE;
2773 "Ignoring invalid numa_zonelist_order value: "
2780 static __init int setup_numa_zonelist_order(char *s)
2787 ret = __parse_numa_zonelist_order(s);
2789 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2793 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2796 * sysctl handler for numa_zonelist_order
2798 int numa_zonelist_order_handler(ctl_table *table, int write,
2799 void __user *buffer, size_t *length,
2802 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2804 static DEFINE_MUTEX(zl_order_mutex);
2806 mutex_lock(&zl_order_mutex);
2808 strcpy(saved_string, (char*)table->data);
2809 ret = proc_dostring(table, write, buffer, length, ppos);
2813 int oldval = user_zonelist_order;
2814 if (__parse_numa_zonelist_order((char*)table->data)) {
2816 * bogus value. restore saved string
2818 strncpy((char*)table->data, saved_string,
2819 NUMA_ZONELIST_ORDER_LEN);
2820 user_zonelist_order = oldval;
2821 } else if (oldval != user_zonelist_order) {
2822 mutex_lock(&zonelists_mutex);
2823 build_all_zonelists(NULL);
2824 mutex_unlock(&zonelists_mutex);
2828 mutex_unlock(&zl_order_mutex);
2833 #define MAX_NODE_LOAD (nr_online_nodes)
2834 static int node_load[MAX_NUMNODES];
2837 * find_next_best_node - find the next node that should appear in a given node's fallback list
2838 * @node: node whose fallback list we're appending
2839 * @used_node_mask: nodemask_t of already used nodes
2841 * We use a number of factors to determine which is the next node that should
2842 * appear on a given node's fallback list. The node should not have appeared
2843 * already in @node's fallback list, and it should be the next closest node
2844 * according to the distance array (which contains arbitrary distance values
2845 * from each node to each node in the system), and should also prefer nodes
2846 * with no CPUs, since presumably they'll have very little allocation pressure
2847 * on them otherwise.
2848 * It returns -1 if no node is found.
2850 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2853 int min_val = INT_MAX;
2855 const struct cpumask *tmp = cpumask_of_node(0);
2857 /* Use the local node if we haven't already */
2858 if (!node_isset(node, *used_node_mask)) {
2859 node_set(node, *used_node_mask);
2863 for_each_node_state(n, N_HIGH_MEMORY) {
2865 /* Don't want a node to appear more than once */
2866 if (node_isset(n, *used_node_mask))
2869 /* Use the distance array to find the distance */
2870 val = node_distance(node, n);
2872 /* Penalize nodes under us ("prefer the next node") */
2875 /* Give preference to headless and unused nodes */
2876 tmp = cpumask_of_node(n);
2877 if (!cpumask_empty(tmp))
2878 val += PENALTY_FOR_NODE_WITH_CPUS;
2880 /* Slight preference for less loaded node */
2881 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2882 val += node_load[n];
2884 if (val < min_val) {
2891 node_set(best_node, *used_node_mask);
2898 * Build zonelists ordered by node and zones within node.
2899 * This results in maximum locality--normal zone overflows into local
2900 * DMA zone, if any--but risks exhausting DMA zone.
2902 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2905 struct zonelist *zonelist;
2907 zonelist = &pgdat->node_zonelists[0];
2908 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2910 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2912 zonelist->_zonerefs[j].zone = NULL;
2913 zonelist->_zonerefs[j].zone_idx = 0;
2917 * Build gfp_thisnode zonelists
2919 static void build_thisnode_zonelists(pg_data_t *pgdat)
2922 struct zonelist *zonelist;
2924 zonelist = &pgdat->node_zonelists[1];
2925 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2926 zonelist->_zonerefs[j].zone = NULL;
2927 zonelist->_zonerefs[j].zone_idx = 0;
2931 * Build zonelists ordered by zone and nodes within zones.
2932 * This results in conserving DMA zone[s] until all Normal memory is
2933 * exhausted, but results in overflowing to remote node while memory
2934 * may still exist in local DMA zone.
2936 static int node_order[MAX_NUMNODES];
2938 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2941 int zone_type; /* needs to be signed */
2943 struct zonelist *zonelist;
2945 zonelist = &pgdat->node_zonelists[0];
2947 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2948 for (j = 0; j < nr_nodes; j++) {
2949 node = node_order[j];
2950 z = &NODE_DATA(node)->node_zones[zone_type];
2951 if (populated_zone(z)) {
2953 &zonelist->_zonerefs[pos++]);
2954 check_highest_zone(zone_type);
2958 zonelist->_zonerefs[pos].zone = NULL;
2959 zonelist->_zonerefs[pos].zone_idx = 0;
2962 static int default_zonelist_order(void)
2965 unsigned long low_kmem_size,total_size;
2969 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2970 * If they are really small and used heavily, the system can fall
2971 * into OOM very easily.
2972 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2974 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2977 for_each_online_node(nid) {
2978 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2979 z = &NODE_DATA(nid)->node_zones[zone_type];
2980 if (populated_zone(z)) {
2981 if (zone_type < ZONE_NORMAL)
2982 low_kmem_size += z->present_pages;
2983 total_size += z->present_pages;
2984 } else if (zone_type == ZONE_NORMAL) {
2986 * If any node has only lowmem, then node order
2987 * is preferred to allow kernel allocations
2988 * locally; otherwise, they can easily infringe
2989 * on other nodes when there is an abundance of
2990 * lowmem available to allocate from.
2992 return ZONELIST_ORDER_NODE;
2996 if (!low_kmem_size || /* there are no DMA area. */
2997 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2998 return ZONELIST_ORDER_NODE;
3000 * look into each node's config.
3001 * If there is a node whose DMA/DMA32 memory is very big area on
3002 * local memory, NODE_ORDER may be suitable.
3004 average_size = total_size /
3005 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3006 for_each_online_node(nid) {
3009 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3010 z = &NODE_DATA(nid)->node_zones[zone_type];
3011 if (populated_zone(z)) {
3012 if (zone_type < ZONE_NORMAL)
3013 low_kmem_size += z->present_pages;
3014 total_size += z->present_pages;
3017 if (low_kmem_size &&
3018 total_size > average_size && /* ignore small node */
3019 low_kmem_size > total_size * 70/100)
3020 return ZONELIST_ORDER_NODE;
3022 return ZONELIST_ORDER_ZONE;
3025 static void set_zonelist_order(void)
3027 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3028 current_zonelist_order = default_zonelist_order();
3030 current_zonelist_order = user_zonelist_order;
3033 static void build_zonelists(pg_data_t *pgdat)
3037 nodemask_t used_mask;
3038 int local_node, prev_node;
3039 struct zonelist *zonelist;
3040 int order = current_zonelist_order;
3042 /* initialize zonelists */
3043 for (i = 0; i < MAX_ZONELISTS; i++) {
3044 zonelist = pgdat->node_zonelists + i;
3045 zonelist->_zonerefs[0].zone = NULL;
3046 zonelist->_zonerefs[0].zone_idx = 0;
3049 /* NUMA-aware ordering of nodes */
3050 local_node = pgdat->node_id;
3051 load = nr_online_nodes;
3052 prev_node = local_node;
3053 nodes_clear(used_mask);
3055 memset(node_order, 0, sizeof(node_order));
3058 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3059 int distance = node_distance(local_node, node);
3062 * If another node is sufficiently far away then it is better
3063 * to reclaim pages in a zone before going off node.
3065 if (distance > RECLAIM_DISTANCE)
3066 zone_reclaim_mode = 1;
3069 * We don't want to pressure a particular node.
3070 * So adding penalty to the first node in same
3071 * distance group to make it round-robin.
3073 if (distance != node_distance(local_node, prev_node))
3074 node_load[node] = load;
3078 if (order == ZONELIST_ORDER_NODE)
3079 build_zonelists_in_node_order(pgdat, node);
3081 node_order[j++] = node; /* remember order */
3084 if (order == ZONELIST_ORDER_ZONE) {
3085 /* calculate node order -- i.e., DMA last! */
3086 build_zonelists_in_zone_order(pgdat, j);
3089 build_thisnode_zonelists(pgdat);
3092 /* Construct the zonelist performance cache - see further mmzone.h */
3093 static void build_zonelist_cache(pg_data_t *pgdat)
3095 struct zonelist *zonelist;
3096 struct zonelist_cache *zlc;
3099 zonelist = &pgdat->node_zonelists[0];
3100 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3101 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3102 for (z = zonelist->_zonerefs; z->zone; z++)
3103 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3106 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3108 * Return node id of node used for "local" allocations.
3109 * I.e., first node id of first zone in arg node's generic zonelist.
3110 * Used for initializing percpu 'numa_mem', which is used primarily
3111 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3113 int local_memory_node(int node)
3117 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3118 gfp_zone(GFP_KERNEL),
3125 #else /* CONFIG_NUMA */
3127 static void set_zonelist_order(void)
3129 current_zonelist_order = ZONELIST_ORDER_ZONE;
3132 static void build_zonelists(pg_data_t *pgdat)
3134 int node, local_node;
3136 struct zonelist *zonelist;
3138 local_node = pgdat->node_id;
3140 zonelist = &pgdat->node_zonelists[0];
3141 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3144 * Now we build the zonelist so that it contains the zones
3145 * of all the other nodes.
3146 * We don't want to pressure a particular node, so when
3147 * building the zones for node N, we make sure that the
3148 * zones coming right after the local ones are those from
3149 * node N+1 (modulo N)
3151 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3152 if (!node_online(node))
3154 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3157 for (node = 0; node < local_node; node++) {
3158 if (!node_online(node))
3160 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3164 zonelist->_zonerefs[j].zone = NULL;
3165 zonelist->_zonerefs[j].zone_idx = 0;
3168 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3169 static void build_zonelist_cache(pg_data_t *pgdat)
3171 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3174 #endif /* CONFIG_NUMA */
3177 * Boot pageset table. One per cpu which is going to be used for all
3178 * zones and all nodes. The parameters will be set in such a way
3179 * that an item put on a list will immediately be handed over to
3180 * the buddy list. This is safe since pageset manipulation is done
3181 * with interrupts disabled.
3183 * The boot_pagesets must be kept even after bootup is complete for
3184 * unused processors and/or zones. They do play a role for bootstrapping
3185 * hotplugged processors.
3187 * zoneinfo_show() and maybe other functions do
3188 * not check if the processor is online before following the pageset pointer.
3189 * Other parts of the kernel may not check if the zone is available.
3191 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3192 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3193 static void setup_zone_pageset(struct zone *zone);
3196 * Global mutex to protect against size modification of zonelists
3197 * as well as to serialize pageset setup for the new populated zone.
3199 DEFINE_MUTEX(zonelists_mutex);
3201 /* return values int ....just for stop_machine() */
3202 static __init_refok int __build_all_zonelists(void *data)
3208 memset(node_load, 0, sizeof(node_load));
3210 for_each_online_node(nid) {
3211 pg_data_t *pgdat = NODE_DATA(nid);
3213 build_zonelists(pgdat);
3214 build_zonelist_cache(pgdat);
3218 * Initialize the boot_pagesets that are going to be used
3219 * for bootstrapping processors. The real pagesets for
3220 * each zone will be allocated later when the per cpu
3221 * allocator is available.
3223 * boot_pagesets are used also for bootstrapping offline
3224 * cpus if the system is already booted because the pagesets
3225 * are needed to initialize allocators on a specific cpu too.
3226 * F.e. the percpu allocator needs the page allocator which
3227 * needs the percpu allocator in order to allocate its pagesets
3228 * (a chicken-egg dilemma).
3230 for_each_possible_cpu(cpu) {
3231 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3233 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3235 * We now know the "local memory node" for each node--
3236 * i.e., the node of the first zone in the generic zonelist.
3237 * Set up numa_mem percpu variable for on-line cpus. During
3238 * boot, only the boot cpu should be on-line; we'll init the
3239 * secondary cpus' numa_mem as they come on-line. During
3240 * node/memory hotplug, we'll fixup all on-line cpus.
3242 if (cpu_online(cpu))
3243 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3251 * Called with zonelists_mutex held always
3252 * unless system_state == SYSTEM_BOOTING.
3254 void __ref build_all_zonelists(void *data)
3256 set_zonelist_order();
3258 if (system_state == SYSTEM_BOOTING) {
3259 __build_all_zonelists(NULL);
3260 mminit_verify_zonelist();
3261 cpuset_init_current_mems_allowed();
3263 /* we have to stop all cpus to guarantee there is no user
3265 #ifdef CONFIG_MEMORY_HOTPLUG
3267 setup_zone_pageset((struct zone *)data);
3269 stop_machine(__build_all_zonelists, NULL, NULL);
3270 /* cpuset refresh routine should be here */
3272 vm_total_pages = nr_free_pagecache_pages();
3274 * Disable grouping by mobility if the number of pages in the
3275 * system is too low to allow the mechanism to work. It would be
3276 * more accurate, but expensive to check per-zone. This check is
3277 * made on memory-hotadd so a system can start with mobility
3278 * disabled and enable it later
3280 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3281 page_group_by_mobility_disabled = 1;
3283 page_group_by_mobility_disabled = 0;
3285 printk("Built %i zonelists in %s order, mobility grouping %s. "
3286 "Total pages: %ld\n",
3288 zonelist_order_name[current_zonelist_order],
3289 page_group_by_mobility_disabled ? "off" : "on",
3292 printk("Policy zone: %s\n", zone_names[policy_zone]);
3297 * Helper functions to size the waitqueue hash table.
3298 * Essentially these want to choose hash table sizes sufficiently
3299 * large so that collisions trying to wait on pages are rare.
3300 * But in fact, the number of active page waitqueues on typical
3301 * systems is ridiculously low, less than 200. So this is even
3302 * conservative, even though it seems large.
3304 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3305 * waitqueues, i.e. the size of the waitq table given the number of pages.
3307 #define PAGES_PER_WAITQUEUE 256
3309 #ifndef CONFIG_MEMORY_HOTPLUG
3310 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3312 unsigned long size = 1;
3314 pages /= PAGES_PER_WAITQUEUE;
3316 while (size < pages)
3320 * Once we have dozens or even hundreds of threads sleeping
3321 * on IO we've got bigger problems than wait queue collision.
3322 * Limit the size of the wait table to a reasonable size.
3324 size = min(size, 4096UL);
3326 return max(size, 4UL);
3330 * A zone's size might be changed by hot-add, so it is not possible to determine
3331 * a suitable size for its wait_table. So we use the maximum size now.
3333 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3335 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3336 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3337 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3339 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3340 * or more by the traditional way. (See above). It equals:
3342 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3343 * ia64(16K page size) : = ( 8G + 4M)byte.
3344 * powerpc (64K page size) : = (32G +16M)byte.
3346 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3353 * This is an integer logarithm so that shifts can be used later
3354 * to extract the more random high bits from the multiplicative
3355 * hash function before the remainder is taken.
3357 static inline unsigned long wait_table_bits(unsigned long size)
3362 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3365 * Check if a pageblock contains reserved pages
3367 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3371 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3372 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3379 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3380 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3381 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3382 * higher will lead to a bigger reserve which will get freed as contiguous
3383 * blocks as reclaim kicks in
3385 static void setup_zone_migrate_reserve(struct zone *zone)
3387 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3389 unsigned long block_migratetype;
3392 /* Get the start pfn, end pfn and the number of blocks to reserve */
3393 start_pfn = zone->zone_start_pfn;
3394 end_pfn = start_pfn + zone->spanned_pages;
3395 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3399 * Reserve blocks are generally in place to help high-order atomic
3400 * allocations that are short-lived. A min_free_kbytes value that
3401 * would result in more than 2 reserve blocks for atomic allocations
3402 * is assumed to be in place to help anti-fragmentation for the
3403 * future allocation of hugepages at runtime.
3405 reserve = min(2, reserve);
3407 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3408 if (!pfn_valid(pfn))
3410 page = pfn_to_page(pfn);
3412 /* Watch out for overlapping nodes */
3413 if (page_to_nid(page) != zone_to_nid(zone))
3416 /* Blocks with reserved pages will never free, skip them. */
3417 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3418 if (pageblock_is_reserved(pfn, block_end_pfn))
3421 block_migratetype = get_pageblock_migratetype(page);
3423 /* If this block is reserved, account for it */
3424 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3429 /* Suitable for reserving if this block is movable */
3430 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3431 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3432 move_freepages_block(zone, page, MIGRATE_RESERVE);
3438 * If the reserve is met and this is a previous reserved block,
3441 if (block_migratetype == MIGRATE_RESERVE) {
3442 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3443 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3449 * Initially all pages are reserved - free ones are freed
3450 * up by free_all_bootmem() once the early boot process is
3451 * done. Non-atomic initialization, single-pass.
3453 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3454 unsigned long start_pfn, enum memmap_context context)
3457 unsigned long end_pfn = start_pfn + size;
3461 if (highest_memmap_pfn < end_pfn - 1)
3462 highest_memmap_pfn = end_pfn - 1;
3464 z = &NODE_DATA(nid)->node_zones[zone];
3465 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3467 * There can be holes in boot-time mem_map[]s
3468 * handed to this function. They do not
3469 * exist on hotplugged memory.
3471 if (context == MEMMAP_EARLY) {
3472 if (!early_pfn_valid(pfn))
3474 if (!early_pfn_in_nid(pfn, nid))
3477 page = pfn_to_page(pfn);
3478 set_page_links(page, zone, nid, pfn);
3479 mminit_verify_page_links(page, zone, nid, pfn);
3480 init_page_count(page);
3481 reset_page_mapcount(page);
3482 SetPageReserved(page);
3484 * Mark the block movable so that blocks are reserved for
3485 * movable at startup. This will force kernel allocations
3486 * to reserve their blocks rather than leaking throughout
3487 * the address space during boot when many long-lived
3488 * kernel allocations are made. Later some blocks near
3489 * the start are marked MIGRATE_RESERVE by
3490 * setup_zone_migrate_reserve()
3492 * bitmap is created for zone's valid pfn range. but memmap
3493 * can be created for invalid pages (for alignment)
3494 * check here not to call set_pageblock_migratetype() against
3497 if ((z->zone_start_pfn <= pfn)
3498 && (pfn < z->zone_start_pfn + z->spanned_pages)
3499 && !(pfn & (pageblock_nr_pages - 1)))
3500 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3502 INIT_LIST_HEAD(&page->lru);
3503 #ifdef WANT_PAGE_VIRTUAL
3504 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3505 if (!is_highmem_idx(zone))
3506 set_page_address(page, __va(pfn << PAGE_SHIFT));
3511 static void __meminit zone_init_free_lists(struct zone *zone)
3514 for_each_migratetype_order(order, t) {
3515 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3516 zone->free_area[order].nr_free = 0;
3520 #ifndef __HAVE_ARCH_MEMMAP_INIT
3521 #define memmap_init(size, nid, zone, start_pfn) \
3522 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3525 static int zone_batchsize(struct zone *zone)
3531 * The per-cpu-pages pools are set to around 1000th of the
3532 * size of the zone. But no more than 1/2 of a meg.
3534 * OK, so we don't know how big the cache is. So guess.
3536 batch = zone->present_pages / 1024;
3537 if (batch * PAGE_SIZE > 512 * 1024)
3538 batch = (512 * 1024) / PAGE_SIZE;
3539 batch /= 4; /* We effectively *= 4 below */
3544 * Clamp the batch to a 2^n - 1 value. Having a power
3545 * of 2 value was found to be more likely to have
3546 * suboptimal cache aliasing properties in some cases.
3548 * For example if 2 tasks are alternately allocating
3549 * batches of pages, one task can end up with a lot
3550 * of pages of one half of the possible page colors
3551 * and the other with pages of the other colors.
3553 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3558 /* The deferral and batching of frees should be suppressed under NOMMU
3561 * The problem is that NOMMU needs to be able to allocate large chunks
3562 * of contiguous memory as there's no hardware page translation to
3563 * assemble apparent contiguous memory from discontiguous pages.
3565 * Queueing large contiguous runs of pages for batching, however,
3566 * causes the pages to actually be freed in smaller chunks. As there
3567 * can be a significant delay between the individual batches being
3568 * recycled, this leads to the once large chunks of space being
3569 * fragmented and becoming unavailable for high-order allocations.
3575 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3577 struct per_cpu_pages *pcp;
3580 memset(p, 0, sizeof(*p));
3584 pcp->high = 6 * batch;
3585 pcp->batch = max(1UL, 1 * batch);
3586 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3587 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3591 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3592 * to the value high for the pageset p.
3595 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3598 struct per_cpu_pages *pcp;
3602 pcp->batch = max(1UL, high/4);
3603 if ((high/4) > (PAGE_SHIFT * 8))
3604 pcp->batch = PAGE_SHIFT * 8;
3607 static void setup_zone_pageset(struct zone *zone)
3611 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3613 for_each_possible_cpu(cpu) {
3614 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3616 setup_pageset(pcp, zone_batchsize(zone));
3618 if (percpu_pagelist_fraction)
3619 setup_pagelist_highmark(pcp,
3620 (zone->present_pages /
3621 percpu_pagelist_fraction));
3626 * Allocate per cpu pagesets and initialize them.
3627 * Before this call only boot pagesets were available.
3629 void __init setup_per_cpu_pageset(void)
3633 for_each_populated_zone(zone)
3634 setup_zone_pageset(zone);
3637 static noinline __init_refok
3638 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3641 struct pglist_data *pgdat = zone->zone_pgdat;
3645 * The per-page waitqueue mechanism uses hashed waitqueues
3648 zone->wait_table_hash_nr_entries =
3649 wait_table_hash_nr_entries(zone_size_pages);
3650 zone->wait_table_bits =
3651 wait_table_bits(zone->wait_table_hash_nr_entries);
3652 alloc_size = zone->wait_table_hash_nr_entries
3653 * sizeof(wait_queue_head_t);
3655 if (!slab_is_available()) {
3656 zone->wait_table = (wait_queue_head_t *)
3657 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3660 * This case means that a zone whose size was 0 gets new memory
3661 * via memory hot-add.
3662 * But it may be the case that a new node was hot-added. In
3663 * this case vmalloc() will not be able to use this new node's
3664 * memory - this wait_table must be initialized to use this new
3665 * node itself as well.
3666 * To use this new node's memory, further consideration will be
3669 zone->wait_table = vmalloc(alloc_size);
3671 if (!zone->wait_table)
3674 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3675 init_waitqueue_head(zone->wait_table + i);
3680 static int __zone_pcp_update(void *data)
3682 struct zone *zone = data;
3684 unsigned long batch = zone_batchsize(zone), flags;
3686 for_each_possible_cpu(cpu) {
3687 struct per_cpu_pageset *pset;
3688 struct per_cpu_pages *pcp;
3690 pset = per_cpu_ptr(zone->pageset, cpu);
3693 local_irq_save(flags);
3694 free_pcppages_bulk(zone, pcp->count, pcp);
3695 setup_pageset(pset, batch);
3696 local_irq_restore(flags);
3701 void zone_pcp_update(struct zone *zone)
3703 stop_machine(__zone_pcp_update, zone, NULL);
3706 static __meminit void zone_pcp_init(struct zone *zone)
3709 * per cpu subsystem is not up at this point. The following code
3710 * relies on the ability of the linker to provide the
3711 * offset of a (static) per cpu variable into the per cpu area.
3713 zone->pageset = &boot_pageset;
3715 if (zone->present_pages)
3716 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3717 zone->name, zone->present_pages,
3718 zone_batchsize(zone));
3721 __meminit int init_currently_empty_zone(struct zone *zone,
3722 unsigned long zone_start_pfn,
3724 enum memmap_context context)
3726 struct pglist_data *pgdat = zone->zone_pgdat;
3728 ret = zone_wait_table_init(zone, size);
3731 pgdat->nr_zones = zone_idx(zone) + 1;
3733 zone->zone_start_pfn = zone_start_pfn;
3735 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3736 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3738 (unsigned long)zone_idx(zone),
3739 zone_start_pfn, (zone_start_pfn + size));
3741 zone_init_free_lists(zone);
3746 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3748 * Basic iterator support. Return the first range of PFNs for a node
3749 * Note: nid == MAX_NUMNODES returns first region regardless of node
3751 static int __meminit first_active_region_index_in_nid(int nid)
3755 for (i = 0; i < nr_nodemap_entries; i++)
3756 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3763 * Basic iterator support. Return the next active range of PFNs for a node
3764 * Note: nid == MAX_NUMNODES returns next region regardless of node
3766 static int __meminit next_active_region_index_in_nid(int index, int nid)
3768 for (index = index + 1; index < nr_nodemap_entries; index++)
3769 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3775 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3777 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3778 * Architectures may implement their own version but if add_active_range()
3779 * was used and there are no special requirements, this is a convenient
3782 int __meminit __early_pfn_to_nid(unsigned long pfn)
3786 for (i = 0; i < nr_nodemap_entries; i++) {
3787 unsigned long start_pfn = early_node_map[i].start_pfn;
3788 unsigned long end_pfn = early_node_map[i].end_pfn;
3790 if (start_pfn <= pfn && pfn < end_pfn)
3791 return early_node_map[i].nid;
3793 /* This is a memory hole */
3796 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3798 int __meminit early_pfn_to_nid(unsigned long pfn)
3802 nid = __early_pfn_to_nid(pfn);
3805 /* just returns 0 */
3809 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3810 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3814 nid = __early_pfn_to_nid(pfn);
3815 if (nid >= 0 && nid != node)
3821 /* Basic iterator support to walk early_node_map[] */
3822 #define for_each_active_range_index_in_nid(i, nid) \
3823 for (i = first_active_region_index_in_nid(nid); i != -1; \
3824 i = next_active_region_index_in_nid(i, nid))
3827 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3828 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3829 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3831 * If an architecture guarantees that all ranges registered with
3832 * add_active_ranges() contain no holes and may be freed, this
3833 * this function may be used instead of calling free_bootmem() manually.
3835 void __init free_bootmem_with_active_regions(int nid,
3836 unsigned long max_low_pfn)
3840 for_each_active_range_index_in_nid(i, nid) {
3841 unsigned long size_pages = 0;
3842 unsigned long end_pfn = early_node_map[i].end_pfn;
3844 if (early_node_map[i].start_pfn >= max_low_pfn)
3847 if (end_pfn > max_low_pfn)
3848 end_pfn = max_low_pfn;
3850 size_pages = end_pfn - early_node_map[i].start_pfn;
3851 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3852 PFN_PHYS(early_node_map[i].start_pfn),
3853 size_pages << PAGE_SHIFT);
3857 #ifdef CONFIG_HAVE_MEMBLOCK
3859 * Basic iterator support. Return the last range of PFNs for a node
3860 * Note: nid == MAX_NUMNODES returns last region regardless of node
3862 static int __meminit last_active_region_index_in_nid(int nid)
3866 for (i = nr_nodemap_entries - 1; i >= 0; i--)
3867 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3874 * Basic iterator support. Return the previous active range of PFNs for a node
3875 * Note: nid == MAX_NUMNODES returns next region regardless of node
3877 static int __meminit previous_active_region_index_in_nid(int index, int nid)
3879 for (index = index - 1; index >= 0; index--)
3880 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3886 #define for_each_active_range_index_in_nid_reverse(i, nid) \
3887 for (i = last_active_region_index_in_nid(nid); i != -1; \
3888 i = previous_active_region_index_in_nid(i, nid))
3890 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3891 u64 goal, u64 limit)
3895 /* Need to go over early_node_map to find out good range for node */
3896 for_each_active_range_index_in_nid_reverse(i, nid) {
3898 u64 ei_start, ei_last;
3899 u64 final_start, final_end;
3901 ei_last = early_node_map[i].end_pfn;
3902 ei_last <<= PAGE_SHIFT;
3903 ei_start = early_node_map[i].start_pfn;
3904 ei_start <<= PAGE_SHIFT;
3906 final_start = max(ei_start, goal);
3907 final_end = min(ei_last, limit);
3909 if (final_start >= final_end)
3912 addr = memblock_find_in_range(final_start, final_end, size, align);
3914 if (addr == MEMBLOCK_ERROR)
3920 return MEMBLOCK_ERROR;
3924 int __init add_from_early_node_map(struct range *range, int az,
3925 int nr_range, int nid)
3930 /* need to go over early_node_map to find out good range for node */
3931 for_each_active_range_index_in_nid(i, nid) {
3932 start = early_node_map[i].start_pfn;
3933 end = early_node_map[i].end_pfn;
3934 nr_range = add_range(range, az, nr_range, start, end);
3939 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3944 for_each_active_range_index_in_nid(i, nid) {
3945 ret = work_fn(early_node_map[i].start_pfn,
3946 early_node_map[i].end_pfn, data);
3952 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3953 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3955 * If an architecture guarantees that all ranges registered with
3956 * add_active_ranges() contain no holes and may be freed, this
3957 * function may be used instead of calling memory_present() manually.
3959 void __init sparse_memory_present_with_active_regions(int nid)
3963 for_each_active_range_index_in_nid(i, nid)
3964 memory_present(early_node_map[i].nid,
3965 early_node_map[i].start_pfn,
3966 early_node_map[i].end_pfn);
3970 * get_pfn_range_for_nid - Return the start and end page frames for a node
3971 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3972 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3973 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3975 * It returns the start and end page frame of a node based on information
3976 * provided by an arch calling add_active_range(). If called for a node
3977 * with no available memory, a warning is printed and the start and end
3980 void __meminit get_pfn_range_for_nid(unsigned int nid,
3981 unsigned long *start_pfn, unsigned long *end_pfn)
3987 for_each_active_range_index_in_nid(i, nid) {
3988 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3989 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3992 if (*start_pfn == -1UL)
3997 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3998 * assumption is made that zones within a node are ordered in monotonic
3999 * increasing memory addresses so that the "highest" populated zone is used
4001 static void __init find_usable_zone_for_movable(void)
4004 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4005 if (zone_index == ZONE_MOVABLE)
4008 if (arch_zone_highest_possible_pfn[zone_index] >
4009 arch_zone_lowest_possible_pfn[zone_index])
4013 VM_BUG_ON(zone_index == -1);
4014 movable_zone = zone_index;
4018 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4019 * because it is sized independent of architecture. Unlike the other zones,
4020 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4021 * in each node depending on the size of each node and how evenly kernelcore
4022 * is distributed. This helper function adjusts the zone ranges
4023 * provided by the architecture for a given node by using the end of the
4024 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4025 * zones within a node are in order of monotonic increases memory addresses
4027 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4028 unsigned long zone_type,
4029 unsigned long node_start_pfn,
4030 unsigned long node_end_pfn,
4031 unsigned long *zone_start_pfn,
4032 unsigned long *zone_end_pfn)
4034 /* Only adjust if ZONE_MOVABLE is on this node */
4035 if (zone_movable_pfn[nid]) {
4036 /* Size ZONE_MOVABLE */
4037 if (zone_type == ZONE_MOVABLE) {
4038 *zone_start_pfn = zone_movable_pfn[nid];
4039 *zone_end_pfn = min(node_end_pfn,
4040 arch_zone_highest_possible_pfn[movable_zone]);
4042 /* Adjust for ZONE_MOVABLE starting within this range */
4043 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4044 *zone_end_pfn > zone_movable_pfn[nid]) {
4045 *zone_end_pfn = zone_movable_pfn[nid];
4047 /* Check if this whole range is within ZONE_MOVABLE */
4048 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4049 *zone_start_pfn = *zone_end_pfn;
4054 * Return the number of pages a zone spans in a node, including holes
4055 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4057 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4058 unsigned long zone_type,
4059 unsigned long *ignored)
4061 unsigned long node_start_pfn, node_end_pfn;
4062 unsigned long zone_start_pfn, zone_end_pfn;
4064 /* Get the start and end of the node and zone */
4065 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4066 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4067 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4068 adjust_zone_range_for_zone_movable(nid, zone_type,
4069 node_start_pfn, node_end_pfn,
4070 &zone_start_pfn, &zone_end_pfn);
4072 /* Check that this node has pages within the zone's required range */
4073 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4076 /* Move the zone boundaries inside the node if necessary */
4077 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4078 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4080 /* Return the spanned pages */
4081 return zone_end_pfn - zone_start_pfn;
4085 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4086 * then all holes in the requested range will be accounted for.
4088 unsigned long __meminit __absent_pages_in_range(int nid,
4089 unsigned long range_start_pfn,
4090 unsigned long range_end_pfn)
4093 unsigned long prev_end_pfn = 0, hole_pages = 0;
4094 unsigned long start_pfn;
4096 /* Find the end_pfn of the first active range of pfns in the node */
4097 i = first_active_region_index_in_nid(nid);
4101 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4103 /* Account for ranges before physical memory on this node */
4104 if (early_node_map[i].start_pfn > range_start_pfn)
4105 hole_pages = prev_end_pfn - range_start_pfn;
4107 /* Find all holes for the zone within the node */
4108 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4110 /* No need to continue if prev_end_pfn is outside the zone */
4111 if (prev_end_pfn >= range_end_pfn)
4114 /* Make sure the end of the zone is not within the hole */
4115 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4116 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4118 /* Update the hole size cound and move on */
4119 if (start_pfn > range_start_pfn) {
4120 BUG_ON(prev_end_pfn > start_pfn);
4121 hole_pages += start_pfn - prev_end_pfn;
4123 prev_end_pfn = early_node_map[i].end_pfn;
4126 /* Account for ranges past physical memory on this node */
4127 if (range_end_pfn > prev_end_pfn)
4128 hole_pages += range_end_pfn -
4129 max(range_start_pfn, prev_end_pfn);
4135 * absent_pages_in_range - Return number of page frames in holes within a range
4136 * @start_pfn: The start PFN to start searching for holes
4137 * @end_pfn: The end PFN to stop searching for holes
4139 * It returns the number of pages frames in memory holes within a range.
4141 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4142 unsigned long end_pfn)
4144 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4147 /* Return the number of page frames in holes in a zone on a node */
4148 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4149 unsigned long zone_type,
4150 unsigned long *ignored)
4152 unsigned long node_start_pfn, node_end_pfn;
4153 unsigned long zone_start_pfn, zone_end_pfn;
4155 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4156 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4158 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4161 adjust_zone_range_for_zone_movable(nid, zone_type,
4162 node_start_pfn, node_end_pfn,
4163 &zone_start_pfn, &zone_end_pfn);
4164 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4168 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4169 unsigned long zone_type,
4170 unsigned long *zones_size)
4172 return zones_size[zone_type];
4175 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4176 unsigned long zone_type,
4177 unsigned long *zholes_size)
4182 return zholes_size[zone_type];
4187 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4188 unsigned long *zones_size, unsigned long *zholes_size)
4190 unsigned long realtotalpages, totalpages = 0;
4193 for (i = 0; i < MAX_NR_ZONES; i++)
4194 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4196 pgdat->node_spanned_pages = totalpages;
4198 realtotalpages = totalpages;
4199 for (i = 0; i < MAX_NR_ZONES; i++)
4201 zone_absent_pages_in_node(pgdat->node_id, i,
4203 pgdat->node_present_pages = realtotalpages;
4204 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4208 #ifndef CONFIG_SPARSEMEM
4210 * Calculate the size of the zone->blockflags rounded to an unsigned long
4211 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4212 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4213 * round what is now in bits to nearest long in bits, then return it in
4216 static unsigned long __init usemap_size(unsigned long zonesize)
4218 unsigned long usemapsize;
4220 usemapsize = roundup(zonesize, pageblock_nr_pages);
4221 usemapsize = usemapsize >> pageblock_order;
4222 usemapsize *= NR_PAGEBLOCK_BITS;
4223 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4225 return usemapsize / 8;
4228 static void __init setup_usemap(struct pglist_data *pgdat,
4229 struct zone *zone, unsigned long zonesize)
4231 unsigned long usemapsize = usemap_size(zonesize);
4232 zone->pageblock_flags = NULL;
4234 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4238 static inline void setup_usemap(struct pglist_data *pgdat,
4239 struct zone *zone, unsigned long zonesize) {}
4240 #endif /* CONFIG_SPARSEMEM */
4242 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4244 /* Return a sensible default order for the pageblock size. */
4245 static inline int pageblock_default_order(void)
4247 if (HPAGE_SHIFT > PAGE_SHIFT)
4248 return HUGETLB_PAGE_ORDER;
4253 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4254 static inline void __init set_pageblock_order(unsigned int order)
4256 /* Check that pageblock_nr_pages has not already been setup */
4257 if (pageblock_order)
4261 * Assume the largest contiguous order of interest is a huge page.
4262 * This value may be variable depending on boot parameters on IA64
4264 pageblock_order = order;
4266 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4269 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4270 * and pageblock_default_order() are unused as pageblock_order is set
4271 * at compile-time. See include/linux/pageblock-flags.h for the values of
4272 * pageblock_order based on the kernel config
4274 static inline int pageblock_default_order(unsigned int order)
4278 #define set_pageblock_order(x) do {} while (0)
4280 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4283 * Set up the zone data structures:
4284 * - mark all pages reserved
4285 * - mark all memory queues empty
4286 * - clear the memory bitmaps
4288 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4289 unsigned long *zones_size, unsigned long *zholes_size)
4292 int nid = pgdat->node_id;
4293 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4296 pgdat_resize_init(pgdat);
4297 pgdat->nr_zones = 0;
4298 init_waitqueue_head(&pgdat->kswapd_wait);
4299 pgdat->kswapd_max_order = 0;
4300 pgdat_page_cgroup_init(pgdat);
4302 for (j = 0; j < MAX_NR_ZONES; j++) {
4303 struct zone *zone = pgdat->node_zones + j;
4304 unsigned long size, realsize, memmap_pages;
4307 size = zone_spanned_pages_in_node(nid, j, zones_size);
4308 realsize = size - zone_absent_pages_in_node(nid, j,
4312 * Adjust realsize so that it accounts for how much memory
4313 * is used by this zone for memmap. This affects the watermark
4314 * and per-cpu initialisations
4317 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4318 if (realsize >= memmap_pages) {
4319 realsize -= memmap_pages;
4322 " %s zone: %lu pages used for memmap\n",
4323 zone_names[j], memmap_pages);
4326 " %s zone: %lu pages exceeds realsize %lu\n",
4327 zone_names[j], memmap_pages, realsize);
4329 /* Account for reserved pages */
4330 if (j == 0 && realsize > dma_reserve) {
4331 realsize -= dma_reserve;
4332 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4333 zone_names[0], dma_reserve);
4336 if (!is_highmem_idx(j))
4337 nr_kernel_pages += realsize;
4338 nr_all_pages += realsize;
4340 zone->spanned_pages = size;
4341 zone->present_pages = realsize;
4344 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4346 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4348 zone->name = zone_names[j];
4349 spin_lock_init(&zone->lock);
4350 spin_lock_init(&zone->lru_lock);
4351 zone_seqlock_init(zone);
4352 zone->zone_pgdat = pgdat;
4354 zone_pcp_init(zone);
4356 INIT_LIST_HEAD(&zone->lru[l].list);
4357 zone->reclaim_stat.recent_rotated[0] = 0;
4358 zone->reclaim_stat.recent_rotated[1] = 0;
4359 zone->reclaim_stat.recent_scanned[0] = 0;
4360 zone->reclaim_stat.recent_scanned[1] = 0;
4361 zap_zone_vm_stats(zone);
4366 set_pageblock_order(pageblock_default_order());
4367 setup_usemap(pgdat, zone, size);
4368 ret = init_currently_empty_zone(zone, zone_start_pfn,
4369 size, MEMMAP_EARLY);
4371 memmap_init(size, nid, j, zone_start_pfn);
4372 zone_start_pfn += size;
4376 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4378 /* Skip empty nodes */
4379 if (!pgdat->node_spanned_pages)
4382 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4383 /* ia64 gets its own node_mem_map, before this, without bootmem */
4384 if (!pgdat->node_mem_map) {
4385 unsigned long size, start, end;
4389 * The zone's endpoints aren't required to be MAX_ORDER
4390 * aligned but the node_mem_map endpoints must be in order
4391 * for the buddy allocator to function correctly.
4393 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4394 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4395 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4396 size = (end - start) * sizeof(struct page);
4397 map = alloc_remap(pgdat->node_id, size);
4399 map = alloc_bootmem_node_nopanic(pgdat, size);
4400 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4402 #ifndef CONFIG_NEED_MULTIPLE_NODES
4404 * With no DISCONTIG, the global mem_map is just set as node 0's
4406 if (pgdat == NODE_DATA(0)) {
4407 mem_map = NODE_DATA(0)->node_mem_map;
4408 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4409 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4410 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4411 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4414 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4417 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4418 unsigned long node_start_pfn, unsigned long *zholes_size)
4420 pg_data_t *pgdat = NODE_DATA(nid);
4422 pgdat->node_id = nid;
4423 pgdat->node_start_pfn = node_start_pfn;
4424 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4426 alloc_node_mem_map(pgdat);
4427 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4428 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4429 nid, (unsigned long)pgdat,
4430 (unsigned long)pgdat->node_mem_map);
4433 free_area_init_core(pgdat, zones_size, zholes_size);
4436 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4438 #if MAX_NUMNODES > 1
4440 * Figure out the number of possible node ids.
4442 static void __init setup_nr_node_ids(void)
4445 unsigned int highest = 0;
4447 for_each_node_mask(node, node_possible_map)
4449 nr_node_ids = highest + 1;
4452 static inline void setup_nr_node_ids(void)
4458 * add_active_range - Register a range of PFNs backed by physical memory
4459 * @nid: The node ID the range resides on
4460 * @start_pfn: The start PFN of the available physical memory
4461 * @end_pfn: The end PFN of the available physical memory
4463 * These ranges are stored in an early_node_map[] and later used by
4464 * free_area_init_nodes() to calculate zone sizes and holes. If the
4465 * range spans a memory hole, it is up to the architecture to ensure
4466 * the memory is not freed by the bootmem allocator. If possible
4467 * the range being registered will be merged with existing ranges.
4469 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4470 unsigned long end_pfn)
4474 mminit_dprintk(MMINIT_TRACE, "memory_register",
4475 "Entering add_active_range(%d, %#lx, %#lx) "
4476 "%d entries of %d used\n",
4477 nid, start_pfn, end_pfn,
4478 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4480 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4482 /* Merge with existing active regions if possible */
4483 for (i = 0; i < nr_nodemap_entries; i++) {
4484 if (early_node_map[i].nid != nid)
4487 /* Skip if an existing region covers this new one */
4488 if (start_pfn >= early_node_map[i].start_pfn &&
4489 end_pfn <= early_node_map[i].end_pfn)
4492 /* Merge forward if suitable */
4493 if (start_pfn <= early_node_map[i].end_pfn &&
4494 end_pfn > early_node_map[i].end_pfn) {
4495 early_node_map[i].end_pfn = end_pfn;
4499 /* Merge backward if suitable */
4500 if (start_pfn < early_node_map[i].start_pfn &&
4501 end_pfn >= early_node_map[i].start_pfn) {
4502 early_node_map[i].start_pfn = start_pfn;
4507 /* Check that early_node_map is large enough */
4508 if (i >= MAX_ACTIVE_REGIONS) {
4509 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4510 MAX_ACTIVE_REGIONS);
4514 early_node_map[i].nid = nid;
4515 early_node_map[i].start_pfn = start_pfn;
4516 early_node_map[i].end_pfn = end_pfn;
4517 nr_nodemap_entries = i + 1;
4521 * remove_active_range - Shrink an existing registered range of PFNs
4522 * @nid: The node id the range is on that should be shrunk
4523 * @start_pfn: The new PFN of the range
4524 * @end_pfn: The new PFN of the range
4526 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4527 * The map is kept near the end physical page range that has already been
4528 * registered. This function allows an arch to shrink an existing registered
4531 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4532 unsigned long end_pfn)
4537 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4538 nid, start_pfn, end_pfn);
4540 /* Find the old active region end and shrink */
4541 for_each_active_range_index_in_nid(i, nid) {
4542 if (early_node_map[i].start_pfn >= start_pfn &&
4543 early_node_map[i].end_pfn <= end_pfn) {
4545 early_node_map[i].start_pfn = 0;
4546 early_node_map[i].end_pfn = 0;
4550 if (early_node_map[i].start_pfn < start_pfn &&
4551 early_node_map[i].end_pfn > start_pfn) {
4552 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4553 early_node_map[i].end_pfn = start_pfn;
4554 if (temp_end_pfn > end_pfn)
4555 add_active_range(nid, end_pfn, temp_end_pfn);
4558 if (early_node_map[i].start_pfn >= start_pfn &&
4559 early_node_map[i].end_pfn > end_pfn &&
4560 early_node_map[i].start_pfn < end_pfn) {
4561 early_node_map[i].start_pfn = end_pfn;
4569 /* remove the blank ones */
4570 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4571 if (early_node_map[i].nid != nid)
4573 if (early_node_map[i].end_pfn)
4575 /* we found it, get rid of it */
4576 for (j = i; j < nr_nodemap_entries - 1; j++)
4577 memcpy(&early_node_map[j], &early_node_map[j+1],
4578 sizeof(early_node_map[j]));
4579 j = nr_nodemap_entries - 1;
4580 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4581 nr_nodemap_entries--;
4586 * remove_all_active_ranges - Remove all currently registered regions
4588 * During discovery, it may be found that a table like SRAT is invalid
4589 * and an alternative discovery method must be used. This function removes
4590 * all currently registered regions.
4592 void __init remove_all_active_ranges(void)
4594 memset(early_node_map, 0, sizeof(early_node_map));
4595 nr_nodemap_entries = 0;
4598 /* Compare two active node_active_regions */
4599 static int __init cmp_node_active_region(const void *a, const void *b)
4601 struct node_active_region *arange = (struct node_active_region *)a;
4602 struct node_active_region *brange = (struct node_active_region *)b;
4604 /* Done this way to avoid overflows */
4605 if (arange->start_pfn > brange->start_pfn)
4607 if (arange->start_pfn < brange->start_pfn)
4613 /* sort the node_map by start_pfn */
4614 void __init sort_node_map(void)
4616 sort(early_node_map, (size_t)nr_nodemap_entries,
4617 sizeof(struct node_active_region),
4618 cmp_node_active_region, NULL);
4622 * node_map_pfn_alignment - determine the maximum internode alignment
4624 * This function should be called after node map is populated and sorted.
4625 * It calculates the maximum power of two alignment which can distinguish
4628 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4629 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4630 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4631 * shifted, 1GiB is enough and this function will indicate so.
4633 * This is used to test whether pfn -> nid mapping of the chosen memory
4634 * model has fine enough granularity to avoid incorrect mapping for the
4635 * populated node map.
4637 * Returns the determined alignment in pfn's. 0 if there is no alignment
4638 * requirement (single node).
4640 unsigned long __init node_map_pfn_alignment(void)
4642 unsigned long accl_mask = 0, last_end = 0;
4646 for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4647 int nid = early_node_map[i].nid;
4648 unsigned long start = early_node_map[i].start_pfn;
4649 unsigned long end = early_node_map[i].end_pfn;
4652 if (!start || last_nid < 0 || last_nid == nid) {
4659 * Start with a mask granular enough to pin-point to the
4660 * start pfn and tick off bits one-by-one until it becomes
4661 * too coarse to separate the current node from the last.
4663 mask = ~((1 << __ffs(start)) - 1);
4664 while (mask && last_end <= (start & (mask << 1)))
4667 /* accumulate all internode masks */
4671 /* convert mask to number of pages */
4672 return ~accl_mask + 1;
4675 /* Find the lowest pfn for a node */
4676 static unsigned long __init find_min_pfn_for_node(int nid)
4679 unsigned long min_pfn = ULONG_MAX;
4681 /* Assuming a sorted map, the first range found has the starting pfn */
4682 for_each_active_range_index_in_nid(i, nid)
4683 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4685 if (min_pfn == ULONG_MAX) {
4687 "Could not find start_pfn for node %d\n", nid);
4695 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4697 * It returns the minimum PFN based on information provided via
4698 * add_active_range().
4700 unsigned long __init find_min_pfn_with_active_regions(void)
4702 return find_min_pfn_for_node(MAX_NUMNODES);
4706 * early_calculate_totalpages()
4707 * Sum pages in active regions for movable zone.
4708 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4710 static unsigned long __init early_calculate_totalpages(void)
4713 unsigned long totalpages = 0;
4715 for (i = 0; i < nr_nodemap_entries; i++) {
4716 unsigned long pages = early_node_map[i].end_pfn -
4717 early_node_map[i].start_pfn;
4718 totalpages += pages;
4720 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4726 * Find the PFN the Movable zone begins in each node. Kernel memory
4727 * is spread evenly between nodes as long as the nodes have enough
4728 * memory. When they don't, some nodes will have more kernelcore than
4731 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4734 unsigned long usable_startpfn;
4735 unsigned long kernelcore_node, kernelcore_remaining;
4736 /* save the state before borrow the nodemask */
4737 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4738 unsigned long totalpages = early_calculate_totalpages();
4739 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4742 * If movablecore was specified, calculate what size of
4743 * kernelcore that corresponds so that memory usable for
4744 * any allocation type is evenly spread. If both kernelcore
4745 * and movablecore are specified, then the value of kernelcore
4746 * will be used for required_kernelcore if it's greater than
4747 * what movablecore would have allowed.
4749 if (required_movablecore) {
4750 unsigned long corepages;
4753 * Round-up so that ZONE_MOVABLE is at least as large as what
4754 * was requested by the user
4756 required_movablecore =
4757 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4758 corepages = totalpages - required_movablecore;
4760 required_kernelcore = max(required_kernelcore, corepages);
4763 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4764 if (!required_kernelcore)
4767 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4768 find_usable_zone_for_movable();
4769 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4772 /* Spread kernelcore memory as evenly as possible throughout nodes */
4773 kernelcore_node = required_kernelcore / usable_nodes;
4774 for_each_node_state(nid, N_HIGH_MEMORY) {
4776 * Recalculate kernelcore_node if the division per node
4777 * now exceeds what is necessary to satisfy the requested
4778 * amount of memory for the kernel
4780 if (required_kernelcore < kernelcore_node)
4781 kernelcore_node = required_kernelcore / usable_nodes;
4784 * As the map is walked, we track how much memory is usable
4785 * by the kernel using kernelcore_remaining. When it is
4786 * 0, the rest of the node is usable by ZONE_MOVABLE
4788 kernelcore_remaining = kernelcore_node;
4790 /* Go through each range of PFNs within this node */
4791 for_each_active_range_index_in_nid(i, nid) {
4792 unsigned long start_pfn, end_pfn;
4793 unsigned long size_pages;
4795 start_pfn = max(early_node_map[i].start_pfn,
4796 zone_movable_pfn[nid]);
4797 end_pfn = early_node_map[i].end_pfn;
4798 if (start_pfn >= end_pfn)
4801 /* Account for what is only usable for kernelcore */
4802 if (start_pfn < usable_startpfn) {
4803 unsigned long kernel_pages;
4804 kernel_pages = min(end_pfn, usable_startpfn)
4807 kernelcore_remaining -= min(kernel_pages,
4808 kernelcore_remaining);
4809 required_kernelcore -= min(kernel_pages,
4810 required_kernelcore);
4812 /* Continue if range is now fully accounted */
4813 if (end_pfn <= usable_startpfn) {
4816 * Push zone_movable_pfn to the end so
4817 * that if we have to rebalance
4818 * kernelcore across nodes, we will
4819 * not double account here
4821 zone_movable_pfn[nid] = end_pfn;
4824 start_pfn = usable_startpfn;
4828 * The usable PFN range for ZONE_MOVABLE is from
4829 * start_pfn->end_pfn. Calculate size_pages as the
4830 * number of pages used as kernelcore
4832 size_pages = end_pfn - start_pfn;
4833 if (size_pages > kernelcore_remaining)
4834 size_pages = kernelcore_remaining;
4835 zone_movable_pfn[nid] = start_pfn + size_pages;
4838 * Some kernelcore has been met, update counts and
4839 * break if the kernelcore for this node has been
4842 required_kernelcore -= min(required_kernelcore,
4844 kernelcore_remaining -= size_pages;
4845 if (!kernelcore_remaining)
4851 * If there is still required_kernelcore, we do another pass with one
4852 * less node in the count. This will push zone_movable_pfn[nid] further
4853 * along on the nodes that still have memory until kernelcore is
4857 if (usable_nodes && required_kernelcore > usable_nodes)
4860 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4861 for (nid = 0; nid < MAX_NUMNODES; nid++)
4862 zone_movable_pfn[nid] =
4863 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4866 /* restore the node_state */
4867 node_states[N_HIGH_MEMORY] = saved_node_state;
4870 /* Any regular memory on that node ? */
4871 static void check_for_regular_memory(pg_data_t *pgdat)
4873 #ifdef CONFIG_HIGHMEM
4874 enum zone_type zone_type;
4876 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4877 struct zone *zone = &pgdat->node_zones[zone_type];
4878 if (zone->present_pages)
4879 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4885 * free_area_init_nodes - Initialise all pg_data_t and zone data
4886 * @max_zone_pfn: an array of max PFNs for each zone
4888 * This will call free_area_init_node() for each active node in the system.
4889 * Using the page ranges provided by add_active_range(), the size of each
4890 * zone in each node and their holes is calculated. If the maximum PFN
4891 * between two adjacent zones match, it is assumed that the zone is empty.
4892 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4893 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4894 * starts where the previous one ended. For example, ZONE_DMA32 starts
4895 * at arch_max_dma_pfn.
4897 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4902 /* Sort early_node_map as initialisation assumes it is sorted */
4905 /* Record where the zone boundaries are */
4906 memset(arch_zone_lowest_possible_pfn, 0,
4907 sizeof(arch_zone_lowest_possible_pfn));
4908 memset(arch_zone_highest_possible_pfn, 0,
4909 sizeof(arch_zone_highest_possible_pfn));
4910 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4911 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4912 for (i = 1; i < MAX_NR_ZONES; i++) {
4913 if (i == ZONE_MOVABLE)
4915 arch_zone_lowest_possible_pfn[i] =
4916 arch_zone_highest_possible_pfn[i-1];
4917 arch_zone_highest_possible_pfn[i] =
4918 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4920 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4921 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4923 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4924 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4925 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4927 /* Print out the zone ranges */
4928 printk("Zone PFN ranges:\n");
4929 for (i = 0; i < MAX_NR_ZONES; i++) {
4930 if (i == ZONE_MOVABLE)
4932 printk(" %-8s ", zone_names[i]);
4933 if (arch_zone_lowest_possible_pfn[i] ==
4934 arch_zone_highest_possible_pfn[i])
4937 printk("%0#10lx -> %0#10lx\n",
4938 arch_zone_lowest_possible_pfn[i],
4939 arch_zone_highest_possible_pfn[i]);
4942 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4943 printk("Movable zone start PFN for each node\n");
4944 for (i = 0; i < MAX_NUMNODES; i++) {
4945 if (zone_movable_pfn[i])
4946 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4949 /* Print out the early_node_map[] */
4950 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4951 for (i = 0; i < nr_nodemap_entries; i++)
4952 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4953 early_node_map[i].start_pfn,
4954 early_node_map[i].end_pfn);
4956 /* Initialise every node */
4957 mminit_verify_pageflags_layout();
4958 setup_nr_node_ids();
4959 for_each_online_node(nid) {
4960 pg_data_t *pgdat = NODE_DATA(nid);
4961 free_area_init_node(nid, NULL,
4962 find_min_pfn_for_node(nid), NULL);
4964 /* Any memory on that node */
4965 if (pgdat->node_present_pages)
4966 node_set_state(nid, N_HIGH_MEMORY);
4967 check_for_regular_memory(pgdat);
4971 static int __init cmdline_parse_core(char *p, unsigned long *core)
4973 unsigned long long coremem;
4977 coremem = memparse(p, &p);
4978 *core = coremem >> PAGE_SHIFT;
4980 /* Paranoid check that UL is enough for the coremem value */
4981 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4987 * kernelcore=size sets the amount of memory for use for allocations that
4988 * cannot be reclaimed or migrated.
4990 static int __init cmdline_parse_kernelcore(char *p)
4992 return cmdline_parse_core(p, &required_kernelcore);
4996 * movablecore=size sets the amount of memory for use for allocations that
4997 * can be reclaimed or migrated.
4999 static int __init cmdline_parse_movablecore(char *p)
5001 return cmdline_parse_core(p, &required_movablecore);
5004 early_param("kernelcore", cmdline_parse_kernelcore);
5005 early_param("movablecore", cmdline_parse_movablecore);
5007 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
5010 * set_dma_reserve - set the specified number of pages reserved in the first zone
5011 * @new_dma_reserve: The number of pages to mark reserved
5013 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5014 * In the DMA zone, a significant percentage may be consumed by kernel image
5015 * and other unfreeable allocations which can skew the watermarks badly. This
5016 * function may optionally be used to account for unfreeable pages in the
5017 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5018 * smaller per-cpu batchsize.
5020 void __init set_dma_reserve(unsigned long new_dma_reserve)
5022 dma_reserve = new_dma_reserve;
5025 void __init free_area_init(unsigned long *zones_size)
5027 free_area_init_node(0, zones_size,
5028 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5031 static int page_alloc_cpu_notify(struct notifier_block *self,
5032 unsigned long action, void *hcpu)
5034 int cpu = (unsigned long)hcpu;
5036 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5040 * Spill the event counters of the dead processor
5041 * into the current processors event counters.
5042 * This artificially elevates the count of the current
5045 vm_events_fold_cpu(cpu);
5048 * Zero the differential counters of the dead processor
5049 * so that the vm statistics are consistent.
5051 * This is only okay since the processor is dead and cannot
5052 * race with what we are doing.
5054 refresh_cpu_vm_stats(cpu);
5059 void __init page_alloc_init(void)
5061 hotcpu_notifier(page_alloc_cpu_notify, 0);
5065 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5066 * or min_free_kbytes changes.
5068 static void calculate_totalreserve_pages(void)
5070 struct pglist_data *pgdat;
5071 unsigned long reserve_pages = 0;
5072 enum zone_type i, j;
5074 for_each_online_pgdat(pgdat) {
5075 for (i = 0; i < MAX_NR_ZONES; i++) {
5076 struct zone *zone = pgdat->node_zones + i;
5077 unsigned long max = 0;
5079 /* Find valid and maximum lowmem_reserve in the zone */
5080 for (j = i; j < MAX_NR_ZONES; j++) {
5081 if (zone->lowmem_reserve[j] > max)
5082 max = zone->lowmem_reserve[j];
5085 /* we treat the high watermark as reserved pages. */
5086 max += high_wmark_pages(zone);
5088 if (max > zone->present_pages)
5089 max = zone->present_pages;
5090 reserve_pages += max;
5093 totalreserve_pages = reserve_pages;
5097 * setup_per_zone_lowmem_reserve - called whenever
5098 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5099 * has a correct pages reserved value, so an adequate number of
5100 * pages are left in the zone after a successful __alloc_pages().
5102 static void setup_per_zone_lowmem_reserve(void)
5104 struct pglist_data *pgdat;
5105 enum zone_type j, idx;
5107 for_each_online_pgdat(pgdat) {
5108 for (j = 0; j < MAX_NR_ZONES; j++) {
5109 struct zone *zone = pgdat->node_zones + j;
5110 unsigned long present_pages = zone->present_pages;
5112 zone->lowmem_reserve[j] = 0;
5116 struct zone *lower_zone;
5120 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5121 sysctl_lowmem_reserve_ratio[idx] = 1;
5123 lower_zone = pgdat->node_zones + idx;
5124 lower_zone->lowmem_reserve[j] = present_pages /
5125 sysctl_lowmem_reserve_ratio[idx];
5126 present_pages += lower_zone->present_pages;
5131 /* update totalreserve_pages */
5132 calculate_totalreserve_pages();
5136 * setup_per_zone_wmarks - called when min_free_kbytes changes
5137 * or when memory is hot-{added|removed}
5139 * Ensures that the watermark[min,low,high] values for each zone are set
5140 * correctly with respect to min_free_kbytes.
5142 void setup_per_zone_wmarks(void)
5144 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5145 unsigned long lowmem_pages = 0;
5147 unsigned long flags;
5149 /* Calculate total number of !ZONE_HIGHMEM pages */
5150 for_each_zone(zone) {
5151 if (!is_highmem(zone))
5152 lowmem_pages += zone->present_pages;
5155 for_each_zone(zone) {
5158 spin_lock_irqsave(&zone->lock, flags);
5159 tmp = (u64)pages_min * zone->present_pages;
5160 do_div(tmp, lowmem_pages);
5161 if (is_highmem(zone)) {
5163 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5164 * need highmem pages, so cap pages_min to a small
5167 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5168 * deltas controls asynch page reclaim, and so should
5169 * not be capped for highmem.
5173 min_pages = zone->present_pages / 1024;
5174 if (min_pages < SWAP_CLUSTER_MAX)
5175 min_pages = SWAP_CLUSTER_MAX;
5176 if (min_pages > 128)
5178 zone->watermark[WMARK_MIN] = min_pages;
5181 * If it's a lowmem zone, reserve a number of pages
5182 * proportionate to the zone's size.
5184 zone->watermark[WMARK_MIN] = tmp;
5187 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5188 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5189 setup_zone_migrate_reserve(zone);
5190 spin_unlock_irqrestore(&zone->lock, flags);
5194 for_each_populated_zone(zone) {
5197 for_each_online_cpu(cpu) {
5200 high = percpu_pagelist_fraction
5201 ? zone->present_pages / percpu_pagelist_fraction
5202 : 5 * zone_batchsize(zone);
5203 setup_pagelist_highmark(
5204 per_cpu_ptr(zone->pageset, cpu), high);
5209 /* update totalreserve_pages */
5210 calculate_totalreserve_pages();
5214 * The inactive anon list should be small enough that the VM never has to
5215 * do too much work, but large enough that each inactive page has a chance
5216 * to be referenced again before it is swapped out.
5218 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5219 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5220 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5221 * the anonymous pages are kept on the inactive list.
5224 * memory ratio inactive anon
5225 * -------------------------------------
5234 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5236 unsigned int gb, ratio;
5238 /* Zone size in gigabytes */
5239 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5241 ratio = int_sqrt(10 * gb);
5245 zone->inactive_ratio = ratio;
5248 static void __meminit setup_per_zone_inactive_ratio(void)
5253 calculate_zone_inactive_ratio(zone);
5257 * Initialise min_free_kbytes.
5259 * For small machines we want it small (128k min). For large machines
5260 * we want it large (64MB max). But it is not linear, because network
5261 * bandwidth does not increase linearly with machine size. We use
5263 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5264 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5280 int __meminit init_per_zone_wmark_min(void)
5282 unsigned long lowmem_kbytes;
5284 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5286 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5287 if (min_free_kbytes < 128)
5288 min_free_kbytes = 128;
5289 if (min_free_kbytes > 65536)
5290 min_free_kbytes = 65536;
5291 setup_per_zone_wmarks();
5292 refresh_zone_stat_thresholds();
5293 setup_per_zone_lowmem_reserve();
5294 setup_per_zone_inactive_ratio();
5297 module_init(init_per_zone_wmark_min)
5300 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5301 * that we can call two helper functions whenever min_free_kbytes
5304 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5305 void __user *buffer, size_t *length, loff_t *ppos)
5307 proc_dointvec(table, write, buffer, length, ppos);
5309 setup_per_zone_wmarks();
5314 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5315 void __user *buffer, size_t *length, loff_t *ppos)
5320 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5325 zone->min_unmapped_pages = (zone->present_pages *
5326 sysctl_min_unmapped_ratio) / 100;
5330 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5331 void __user *buffer, size_t *length, loff_t *ppos)
5336 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5341 zone->min_slab_pages = (zone->present_pages *
5342 sysctl_min_slab_ratio) / 100;
5348 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5349 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5350 * whenever sysctl_lowmem_reserve_ratio changes.
5352 * The reserve ratio obviously has absolutely no relation with the
5353 * minimum watermarks. The lowmem reserve ratio can only make sense
5354 * if in function of the boot time zone sizes.
5356 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5357 void __user *buffer, size_t *length, loff_t *ppos)
5359 proc_dointvec_minmax(table, write, buffer, length, ppos);
5360 setup_per_zone_lowmem_reserve();
5365 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5366 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5367 * can have before it gets flushed back to buddy allocator.
5370 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5371 void __user *buffer, size_t *length, loff_t *ppos)
5377 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5378 if (!write || (ret == -EINVAL))
5380 for_each_populated_zone(zone) {
5381 for_each_possible_cpu(cpu) {
5383 high = zone->present_pages / percpu_pagelist_fraction;
5384 setup_pagelist_highmark(
5385 per_cpu_ptr(zone->pageset, cpu), high);
5391 int hashdist = HASHDIST_DEFAULT;
5394 static int __init set_hashdist(char *str)
5398 hashdist = simple_strtoul(str, &str, 0);
5401 __setup("hashdist=", set_hashdist);
5405 * allocate a large system hash table from bootmem
5406 * - it is assumed that the hash table must contain an exact power-of-2
5407 * quantity of entries
5408 * - limit is the number of hash buckets, not the total allocation size
5410 void *__init alloc_large_system_hash(const char *tablename,
5411 unsigned long bucketsize,
5412 unsigned long numentries,
5415 unsigned int *_hash_shift,
5416 unsigned int *_hash_mask,
5417 unsigned long limit)
5419 unsigned long long max = limit;
5420 unsigned long log2qty, size;
5423 /* allow the kernel cmdline to have a say */
5425 /* round applicable memory size up to nearest megabyte */
5426 numentries = nr_kernel_pages;
5427 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5428 numentries >>= 20 - PAGE_SHIFT;
5429 numentries <<= 20 - PAGE_SHIFT;
5431 /* limit to 1 bucket per 2^scale bytes of low memory */
5432 if (scale > PAGE_SHIFT)
5433 numentries >>= (scale - PAGE_SHIFT);
5435 numentries <<= (PAGE_SHIFT - scale);
5437 /* Make sure we've got at least a 0-order allocation.. */
5438 if (unlikely(flags & HASH_SMALL)) {
5439 /* Makes no sense without HASH_EARLY */
5440 WARN_ON(!(flags & HASH_EARLY));
5441 if (!(numentries >> *_hash_shift)) {
5442 numentries = 1UL << *_hash_shift;
5443 BUG_ON(!numentries);
5445 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5446 numentries = PAGE_SIZE / bucketsize;
5448 numentries = roundup_pow_of_two(numentries);
5450 /* limit allocation size to 1/16 total memory by default */
5452 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5453 do_div(max, bucketsize);
5456 if (numentries > max)
5459 log2qty = ilog2(numentries);
5462 size = bucketsize << log2qty;
5463 if (flags & HASH_EARLY)
5464 table = alloc_bootmem_nopanic(size);
5466 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5469 * If bucketsize is not a power-of-two, we may free
5470 * some pages at the end of hash table which
5471 * alloc_pages_exact() automatically does
5473 if (get_order(size) < MAX_ORDER) {
5474 table = alloc_pages_exact(size, GFP_ATOMIC);
5475 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5478 } while (!table && size > PAGE_SIZE && --log2qty);
5481 panic("Failed to allocate %s hash table\n", tablename);
5483 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5486 ilog2(size) - PAGE_SHIFT,
5490 *_hash_shift = log2qty;
5492 *_hash_mask = (1 << log2qty) - 1;
5497 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5498 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5501 #ifdef CONFIG_SPARSEMEM
5502 return __pfn_to_section(pfn)->pageblock_flags;
5504 return zone->pageblock_flags;
5505 #endif /* CONFIG_SPARSEMEM */
5508 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5510 #ifdef CONFIG_SPARSEMEM
5511 pfn &= (PAGES_PER_SECTION-1);
5512 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5514 pfn = pfn - zone->zone_start_pfn;
5515 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5516 #endif /* CONFIG_SPARSEMEM */
5520 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5521 * @page: The page within the block of interest
5522 * @start_bitidx: The first bit of interest to retrieve
5523 * @end_bitidx: The last bit of interest
5524 * returns pageblock_bits flags
5526 unsigned long get_pageblock_flags_group(struct page *page,
5527 int start_bitidx, int end_bitidx)
5530 unsigned long *bitmap;
5531 unsigned long pfn, bitidx;
5532 unsigned long flags = 0;
5533 unsigned long value = 1;
5535 zone = page_zone(page);
5536 pfn = page_to_pfn(page);
5537 bitmap = get_pageblock_bitmap(zone, pfn);
5538 bitidx = pfn_to_bitidx(zone, pfn);
5540 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5541 if (test_bit(bitidx + start_bitidx, bitmap))
5548 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5549 * @page: The page within the block of interest
5550 * @start_bitidx: The first bit of interest
5551 * @end_bitidx: The last bit of interest
5552 * @flags: The flags to set
5554 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5555 int start_bitidx, int end_bitidx)
5558 unsigned long *bitmap;
5559 unsigned long pfn, bitidx;
5560 unsigned long value = 1;
5562 zone = page_zone(page);
5563 pfn = page_to_pfn(page);
5564 bitmap = get_pageblock_bitmap(zone, pfn);
5565 bitidx = pfn_to_bitidx(zone, pfn);
5566 VM_BUG_ON(pfn < zone->zone_start_pfn);
5567 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5569 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5571 __set_bit(bitidx + start_bitidx, bitmap);
5573 __clear_bit(bitidx + start_bitidx, bitmap);
5577 * This is designed as sub function...plz see page_isolation.c also.
5578 * set/clear page block's type to be ISOLATE.
5579 * page allocater never alloc memory from ISOLATE block.
5583 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5585 unsigned long pfn, iter, found;
5587 * For avoiding noise data, lru_add_drain_all() should be called
5588 * If ZONE_MOVABLE, the zone never contains immobile pages
5590 if (zone_idx(zone) == ZONE_MOVABLE)
5593 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5596 pfn = page_to_pfn(page);
5597 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5598 unsigned long check = pfn + iter;
5600 if (!pfn_valid_within(check))
5603 page = pfn_to_page(check);
5604 if (!page_count(page)) {
5605 if (PageBuddy(page))
5606 iter += (1 << page_order(page)) - 1;
5612 * If there are RECLAIMABLE pages, we need to check it.
5613 * But now, memory offline itself doesn't call shrink_slab()
5614 * and it still to be fixed.
5617 * If the page is not RAM, page_count()should be 0.
5618 * we don't need more check. This is an _used_ not-movable page.
5620 * The problematic thing here is PG_reserved pages. PG_reserved
5621 * is set to both of a memory hole page and a _used_ kernel
5630 bool is_pageblock_removable_nolock(struct page *page)
5632 struct zone *zone = page_zone(page);
5633 return __count_immobile_pages(zone, page, 0);
5636 int set_migratetype_isolate(struct page *page)
5639 unsigned long flags, pfn;
5640 struct memory_isolate_notify arg;
5644 zone = page_zone(page);
5646 spin_lock_irqsave(&zone->lock, flags);
5648 pfn = page_to_pfn(page);
5649 arg.start_pfn = pfn;
5650 arg.nr_pages = pageblock_nr_pages;
5651 arg.pages_found = 0;
5654 * It may be possible to isolate a pageblock even if the
5655 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5656 * notifier chain is used by balloon drivers to return the
5657 * number of pages in a range that are held by the balloon
5658 * driver to shrink memory. If all the pages are accounted for
5659 * by balloons, are free, or on the LRU, isolation can continue.
5660 * Later, for example, when memory hotplug notifier runs, these
5661 * pages reported as "can be isolated" should be isolated(freed)
5662 * by the balloon driver through the memory notifier chain.
5664 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5665 notifier_ret = notifier_to_errno(notifier_ret);
5669 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5670 * We just check MOVABLE pages.
5672 if (__count_immobile_pages(zone, page, arg.pages_found))
5676 * immobile means "not-on-lru" paes. If immobile is larger than
5677 * removable-by-driver pages reported by notifier, we'll fail.
5682 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5683 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5686 spin_unlock_irqrestore(&zone->lock, flags);
5692 void unset_migratetype_isolate(struct page *page)
5695 unsigned long flags;
5696 zone = page_zone(page);
5697 spin_lock_irqsave(&zone->lock, flags);
5698 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5700 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5701 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5703 spin_unlock_irqrestore(&zone->lock, flags);
5706 #ifdef CONFIG_MEMORY_HOTREMOVE
5708 * All pages in the range must be isolated before calling this.
5711 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5717 unsigned long flags;
5718 /* find the first valid pfn */
5719 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5724 zone = page_zone(pfn_to_page(pfn));
5725 spin_lock_irqsave(&zone->lock, flags);
5727 while (pfn < end_pfn) {
5728 if (!pfn_valid(pfn)) {
5732 page = pfn_to_page(pfn);
5733 BUG_ON(page_count(page));
5734 BUG_ON(!PageBuddy(page));
5735 order = page_order(page);
5736 #ifdef CONFIG_DEBUG_VM
5737 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5738 pfn, 1 << order, end_pfn);
5740 list_del(&page->lru);
5741 rmv_page_order(page);
5742 zone->free_area[order].nr_free--;
5743 __mod_zone_page_state(zone, NR_FREE_PAGES,
5745 for (i = 0; i < (1 << order); i++)
5746 SetPageReserved((page+i));
5747 pfn += (1 << order);
5749 spin_unlock_irqrestore(&zone->lock, flags);
5753 #ifdef CONFIG_MEMORY_FAILURE
5754 bool is_free_buddy_page(struct page *page)
5756 struct zone *zone = page_zone(page);
5757 unsigned long pfn = page_to_pfn(page);
5758 unsigned long flags;
5761 spin_lock_irqsave(&zone->lock, flags);
5762 for (order = 0; order < MAX_ORDER; order++) {
5763 struct page *page_head = page - (pfn & ((1 << order) - 1));
5765 if (PageBuddy(page_head) && page_order(page_head) >= order)
5768 spin_unlock_irqrestore(&zone->lock, flags);
5770 return order < MAX_ORDER;
5774 static struct trace_print_flags pageflag_names[] = {
5775 {1UL << PG_locked, "locked" },
5776 {1UL << PG_error, "error" },
5777 {1UL << PG_referenced, "referenced" },
5778 {1UL << PG_uptodate, "uptodate" },
5779 {1UL << PG_dirty, "dirty" },
5780 {1UL << PG_lru, "lru" },
5781 {1UL << PG_active, "active" },
5782 {1UL << PG_slab, "slab" },
5783 {1UL << PG_owner_priv_1, "owner_priv_1" },
5784 {1UL << PG_arch_1, "arch_1" },
5785 {1UL << PG_reserved, "reserved" },
5786 {1UL << PG_private, "private" },
5787 {1UL << PG_private_2, "private_2" },
5788 {1UL << PG_writeback, "writeback" },
5789 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5790 {1UL << PG_head, "head" },
5791 {1UL << PG_tail, "tail" },
5793 {1UL << PG_compound, "compound" },
5795 {1UL << PG_swapcache, "swapcache" },
5796 {1UL << PG_mappedtodisk, "mappedtodisk" },
5797 {1UL << PG_reclaim, "reclaim" },
5798 {1UL << PG_swapbacked, "swapbacked" },
5799 {1UL << PG_unevictable, "unevictable" },
5801 {1UL << PG_mlocked, "mlocked" },
5803 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5804 {1UL << PG_uncached, "uncached" },
5806 #ifdef CONFIG_MEMORY_FAILURE
5807 {1UL << PG_hwpoison, "hwpoison" },
5812 static void dump_page_flags(unsigned long flags)
5814 const char *delim = "";
5818 printk(KERN_ALERT "page flags: %#lx(", flags);
5820 /* remove zone id */
5821 flags &= (1UL << NR_PAGEFLAGS) - 1;
5823 for (i = 0; pageflag_names[i].name && flags; i++) {
5825 mask = pageflag_names[i].mask;
5826 if ((flags & mask) != mask)
5830 printk("%s%s", delim, pageflag_names[i].name);
5834 /* check for left over flags */
5836 printk("%s%#lx", delim, flags);
5841 void dump_page(struct page *page)
5844 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5845 page, atomic_read(&page->_count), page_mapcount(page),
5846 page->mapping, page->index);
5847 dump_page_flags(page->flags);
5848 mem_cgroup_print_bad_page(page);