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;
359 set_page_count(p, 0);
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 = higher_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)
656 trace_mm_page_free_direct(page, order);
657 kmemcheck_free_shadow(page, order);
660 page->mapping = NULL;
661 for (i = 0; i < (1 << order); i++)
662 bad += free_pages_check(page + i);
666 if (!PageHighMem(page)) {
667 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
668 debug_check_no_obj_freed(page_address(page),
671 arch_free_page(page, order);
672 kernel_map_pages(page, 1 << order, 0);
677 static void __free_pages_ok(struct page *page, unsigned int order)
680 int wasMlocked = __TestClearPageMlocked(page);
682 if (!free_pages_prepare(page, order))
685 local_irq_save(flags);
686 if (unlikely(wasMlocked))
687 free_page_mlock(page);
688 __count_vm_events(PGFREE, 1 << order);
689 free_one_page(page_zone(page), page, order,
690 get_pageblock_migratetype(page));
691 local_irq_restore(flags);
695 * permit the bootmem allocator to evade page validation on high-order frees
697 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
700 __ClearPageReserved(page);
701 set_page_count(page, 0);
702 set_page_refcounted(page);
708 for (loop = 0; loop < BITS_PER_LONG; loop++) {
709 struct page *p = &page[loop];
711 if (loop + 1 < BITS_PER_LONG)
713 __ClearPageReserved(p);
714 set_page_count(p, 0);
717 set_page_refcounted(page);
718 __free_pages(page, order);
724 * The order of subdivision here is critical for the IO subsystem.
725 * Please do not alter this order without good reasons and regression
726 * testing. Specifically, as large blocks of memory are subdivided,
727 * the order in which smaller blocks are delivered depends on the order
728 * they're subdivided in this function. This is the primary factor
729 * influencing the order in which pages are delivered to the IO
730 * subsystem according to empirical testing, and this is also justified
731 * by considering the behavior of a buddy system containing a single
732 * large block of memory acted on by a series of small allocations.
733 * This behavior is a critical factor in sglist merging's success.
737 static inline void expand(struct zone *zone, struct page *page,
738 int low, int high, struct free_area *area,
741 unsigned long size = 1 << high;
747 VM_BUG_ON(bad_range(zone, &page[size]));
748 list_add(&page[size].lru, &area->free_list[migratetype]);
750 set_page_order(&page[size], high);
755 * This page is about to be returned from the page allocator
757 static inline int check_new_page(struct page *page)
759 if (unlikely(page_mapcount(page) |
760 (page->mapping != NULL) |
761 (atomic_read(&page->_count) != 0) |
762 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
763 (mem_cgroup_bad_page_check(page)))) {
770 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
774 for (i = 0; i < (1 << order); i++) {
775 struct page *p = page + i;
776 if (unlikely(check_new_page(p)))
780 set_page_private(page, 0);
781 set_page_refcounted(page);
783 arch_alloc_page(page, order);
784 kernel_map_pages(page, 1 << order, 1);
786 if (gfp_flags & __GFP_ZERO)
787 prep_zero_page(page, order, gfp_flags);
789 if (order && (gfp_flags & __GFP_COMP))
790 prep_compound_page(page, order);
796 * Go through the free lists for the given migratetype and remove
797 * the smallest available page from the freelists
800 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
803 unsigned int current_order;
804 struct free_area * area;
807 /* Find a page of the appropriate size in the preferred list */
808 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
809 area = &(zone->free_area[current_order]);
810 if (list_empty(&area->free_list[migratetype]))
813 page = list_entry(area->free_list[migratetype].next,
815 list_del(&page->lru);
816 rmv_page_order(page);
818 expand(zone, page, order, current_order, area, migratetype);
827 * This array describes the order lists are fallen back to when
828 * the free lists for the desirable migrate type are depleted
830 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
831 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
832 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
833 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
834 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
838 * Move the free pages in a range to the free lists of the requested type.
839 * Note that start_page and end_pages are not aligned on a pageblock
840 * boundary. If alignment is required, use move_freepages_block()
842 static int move_freepages(struct zone *zone,
843 struct page *start_page, struct page *end_page,
850 #ifndef CONFIG_HOLES_IN_ZONE
852 * page_zone is not safe to call in this context when
853 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
854 * anyway as we check zone boundaries in move_freepages_block().
855 * Remove at a later date when no bug reports exist related to
856 * grouping pages by mobility
858 BUG_ON(page_zone(start_page) != page_zone(end_page));
861 for (page = start_page; page <= end_page;) {
862 /* Make sure we are not inadvertently changing nodes */
863 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
865 if (!pfn_valid_within(page_to_pfn(page))) {
870 if (!PageBuddy(page)) {
875 order = page_order(page);
876 list_move(&page->lru,
877 &zone->free_area[order].free_list[migratetype]);
879 pages_moved += 1 << order;
885 static int move_freepages_block(struct zone *zone, struct page *page,
888 unsigned long start_pfn, end_pfn;
889 struct page *start_page, *end_page;
891 start_pfn = page_to_pfn(page);
892 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
893 start_page = pfn_to_page(start_pfn);
894 end_page = start_page + pageblock_nr_pages - 1;
895 end_pfn = start_pfn + pageblock_nr_pages - 1;
897 /* Do not cross zone boundaries */
898 if (start_pfn < zone->zone_start_pfn)
900 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
903 return move_freepages(zone, start_page, end_page, migratetype);
906 static void change_pageblock_range(struct page *pageblock_page,
907 int start_order, int migratetype)
909 int nr_pageblocks = 1 << (start_order - pageblock_order);
911 while (nr_pageblocks--) {
912 set_pageblock_migratetype(pageblock_page, migratetype);
913 pageblock_page += pageblock_nr_pages;
917 /* Remove an element from the buddy allocator from the fallback list */
918 static inline struct page *
919 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
921 struct free_area * area;
926 /* Find the largest possible block of pages in the other list */
927 for (current_order = MAX_ORDER-1; current_order >= order;
929 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
930 migratetype = fallbacks[start_migratetype][i];
932 /* MIGRATE_RESERVE handled later if necessary */
933 if (migratetype == MIGRATE_RESERVE)
936 area = &(zone->free_area[current_order]);
937 if (list_empty(&area->free_list[migratetype]))
940 page = list_entry(area->free_list[migratetype].next,
945 * If breaking a large block of pages, move all free
946 * pages to the preferred allocation list. If falling
947 * back for a reclaimable kernel allocation, be more
948 * aggressive about taking ownership of free pages
950 if (unlikely(current_order >= (pageblock_order >> 1)) ||
951 start_migratetype == MIGRATE_RECLAIMABLE ||
952 page_group_by_mobility_disabled) {
954 pages = move_freepages_block(zone, page,
957 /* Claim the whole block if over half of it is free */
958 if (pages >= (1 << (pageblock_order-1)) ||
959 page_group_by_mobility_disabled)
960 set_pageblock_migratetype(page,
963 migratetype = start_migratetype;
966 /* Remove the page from the freelists */
967 list_del(&page->lru);
968 rmv_page_order(page);
970 /* Take ownership for orders >= pageblock_order */
971 if (current_order >= pageblock_order)
972 change_pageblock_range(page, current_order,
975 expand(zone, page, order, current_order, area, migratetype);
977 trace_mm_page_alloc_extfrag(page, order, current_order,
978 start_migratetype, migratetype);
988 * Do the hard work of removing an element from the buddy allocator.
989 * Call me with the zone->lock already held.
991 static struct page *__rmqueue(struct zone *zone, unsigned int order,
997 page = __rmqueue_smallest(zone, order, migratetype);
999 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1000 page = __rmqueue_fallback(zone, order, migratetype);
1003 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1004 * is used because __rmqueue_smallest is an inline function
1005 * and we want just one call site
1008 migratetype = MIGRATE_RESERVE;
1013 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1018 * Obtain a specified number of elements from the buddy allocator, all under
1019 * a single hold of the lock, for efficiency. Add them to the supplied list.
1020 * Returns the number of new pages which were placed at *list.
1022 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1023 unsigned long count, struct list_head *list,
1024 int migratetype, int cold)
1028 spin_lock(&zone->lock);
1029 for (i = 0; i < count; ++i) {
1030 struct page *page = __rmqueue(zone, order, migratetype);
1031 if (unlikely(page == NULL))
1035 * Split buddy pages returned by expand() are received here
1036 * in physical page order. The page is added to the callers and
1037 * list and the list head then moves forward. From the callers
1038 * perspective, the linked list is ordered by page number in
1039 * some conditions. This is useful for IO devices that can
1040 * merge IO requests if the physical pages are ordered
1043 if (likely(cold == 0))
1044 list_add(&page->lru, list);
1046 list_add_tail(&page->lru, list);
1047 set_page_private(page, migratetype);
1050 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1051 spin_unlock(&zone->lock);
1057 * Called from the vmstat counter updater to drain pagesets of this
1058 * currently executing processor on remote nodes after they have
1061 * Note that this function must be called with the thread pinned to
1062 * a single processor.
1064 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1066 unsigned long flags;
1069 local_irq_save(flags);
1070 if (pcp->count >= pcp->batch)
1071 to_drain = pcp->batch;
1073 to_drain = pcp->count;
1074 free_pcppages_bulk(zone, to_drain, pcp);
1075 pcp->count -= to_drain;
1076 local_irq_restore(flags);
1081 * Drain pages of the indicated processor.
1083 * The processor must either be the current processor and the
1084 * thread pinned to the current processor or a processor that
1087 static void drain_pages(unsigned int cpu)
1089 unsigned long flags;
1092 for_each_populated_zone(zone) {
1093 struct per_cpu_pageset *pset;
1094 struct per_cpu_pages *pcp;
1096 local_irq_save(flags);
1097 pset = per_cpu_ptr(zone->pageset, cpu);
1101 free_pcppages_bulk(zone, pcp->count, pcp);
1104 local_irq_restore(flags);
1109 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1111 void drain_local_pages(void *arg)
1113 drain_pages(smp_processor_id());
1117 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1119 void drain_all_pages(void)
1121 on_each_cpu(drain_local_pages, NULL, 1);
1124 #ifdef CONFIG_HIBERNATION
1126 void mark_free_pages(struct zone *zone)
1128 unsigned long pfn, max_zone_pfn;
1129 unsigned long flags;
1131 struct list_head *curr;
1133 if (!zone->spanned_pages)
1136 spin_lock_irqsave(&zone->lock, flags);
1138 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1139 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1140 if (pfn_valid(pfn)) {
1141 struct page *page = pfn_to_page(pfn);
1143 if (!swsusp_page_is_forbidden(page))
1144 swsusp_unset_page_free(page);
1147 for_each_migratetype_order(order, t) {
1148 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1151 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1152 for (i = 0; i < (1UL << order); i++)
1153 swsusp_set_page_free(pfn_to_page(pfn + i));
1156 spin_unlock_irqrestore(&zone->lock, flags);
1158 #endif /* CONFIG_PM */
1161 * Free a 0-order page
1162 * cold == 1 ? free a cold page : free a hot page
1164 void free_hot_cold_page(struct page *page, int cold)
1166 struct zone *zone = page_zone(page);
1167 struct per_cpu_pages *pcp;
1168 unsigned long flags;
1170 int wasMlocked = __TestClearPageMlocked(page);
1172 if (!free_pages_prepare(page, 0))
1175 migratetype = get_pageblock_migratetype(page);
1176 set_page_private(page, migratetype);
1177 local_irq_save(flags);
1178 if (unlikely(wasMlocked))
1179 free_page_mlock(page);
1180 __count_vm_event(PGFREE);
1183 * We only track unmovable, reclaimable and movable on pcp lists.
1184 * Free ISOLATE pages back to the allocator because they are being
1185 * offlined but treat RESERVE as movable pages so we can get those
1186 * areas back if necessary. Otherwise, we may have to free
1187 * excessively into the page allocator
1189 if (migratetype >= MIGRATE_PCPTYPES) {
1190 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1191 free_one_page(zone, page, 0, migratetype);
1194 migratetype = MIGRATE_MOVABLE;
1197 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1199 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1201 list_add(&page->lru, &pcp->lists[migratetype]);
1203 if (pcp->count >= pcp->high) {
1204 free_pcppages_bulk(zone, pcp->batch, pcp);
1205 pcp->count -= pcp->batch;
1209 local_irq_restore(flags);
1213 * split_page takes a non-compound higher-order page, and splits it into
1214 * n (1<<order) sub-pages: page[0..n]
1215 * Each sub-page must be freed individually.
1217 * Note: this is probably too low level an operation for use in drivers.
1218 * Please consult with lkml before using this in your driver.
1220 void split_page(struct page *page, unsigned int order)
1224 VM_BUG_ON(PageCompound(page));
1225 VM_BUG_ON(!page_count(page));
1227 #ifdef CONFIG_KMEMCHECK
1229 * Split shadow pages too, because free(page[0]) would
1230 * otherwise free the whole shadow.
1232 if (kmemcheck_page_is_tracked(page))
1233 split_page(virt_to_page(page[0].shadow), order);
1236 for (i = 1; i < (1 << order); i++)
1237 set_page_refcounted(page + i);
1241 * Similar to split_page except the page is already free. As this is only
1242 * being used for migration, the migratetype of the block also changes.
1243 * As this is called with interrupts disabled, the caller is responsible
1244 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1247 * Note: this is probably too low level an operation for use in drivers.
1248 * Please consult with lkml before using this in your driver.
1250 int split_free_page(struct page *page)
1253 unsigned long watermark;
1256 BUG_ON(!PageBuddy(page));
1258 zone = page_zone(page);
1259 order = page_order(page);
1261 /* Obey watermarks as if the page was being allocated */
1262 watermark = low_wmark_pages(zone) + (1 << order);
1263 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1266 /* Remove page from free list */
1267 list_del(&page->lru);
1268 zone->free_area[order].nr_free--;
1269 rmv_page_order(page);
1270 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1272 /* Split into individual pages */
1273 set_page_refcounted(page);
1274 split_page(page, order);
1276 if (order >= pageblock_order - 1) {
1277 struct page *endpage = page + (1 << order) - 1;
1278 for (; page < endpage; page += pageblock_nr_pages)
1279 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1286 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1287 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1291 struct page *buffered_rmqueue(struct zone *preferred_zone,
1292 struct zone *zone, int order, gfp_t gfp_flags,
1295 unsigned long flags;
1297 int cold = !!(gfp_flags & __GFP_COLD);
1300 if (likely(order == 0)) {
1301 struct per_cpu_pages *pcp;
1302 struct list_head *list;
1304 local_irq_save(flags);
1305 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1306 list = &pcp->lists[migratetype];
1307 if (list_empty(list)) {
1308 pcp->count += rmqueue_bulk(zone, 0,
1311 if (unlikely(list_empty(list)))
1316 page = list_entry(list->prev, struct page, lru);
1318 page = list_entry(list->next, struct page, lru);
1320 list_del(&page->lru);
1323 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1325 * __GFP_NOFAIL is not to be used in new code.
1327 * All __GFP_NOFAIL callers should be fixed so that they
1328 * properly detect and handle allocation failures.
1330 * We most definitely don't want callers attempting to
1331 * allocate greater than order-1 page units with
1334 WARN_ON_ONCE(order > 1);
1336 spin_lock_irqsave(&zone->lock, flags);
1337 page = __rmqueue(zone, order, migratetype);
1338 spin_unlock(&zone->lock);
1341 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1344 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1345 zone_statistics(preferred_zone, zone, gfp_flags);
1346 local_irq_restore(flags);
1348 VM_BUG_ON(bad_range(zone, page));
1349 if (prep_new_page(page, order, gfp_flags))
1354 local_irq_restore(flags);
1358 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1359 #define ALLOC_WMARK_MIN WMARK_MIN
1360 #define ALLOC_WMARK_LOW WMARK_LOW
1361 #define ALLOC_WMARK_HIGH WMARK_HIGH
1362 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1364 /* Mask to get the watermark bits */
1365 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1367 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1368 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1369 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1371 #ifdef CONFIG_FAIL_PAGE_ALLOC
1373 static struct fail_page_alloc_attr {
1374 struct fault_attr attr;
1376 u32 ignore_gfp_highmem;
1377 u32 ignore_gfp_wait;
1380 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1382 struct dentry *ignore_gfp_highmem_file;
1383 struct dentry *ignore_gfp_wait_file;
1384 struct dentry *min_order_file;
1386 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1388 } fail_page_alloc = {
1389 .attr = FAULT_ATTR_INITIALIZER,
1390 .ignore_gfp_wait = 1,
1391 .ignore_gfp_highmem = 1,
1395 static int __init setup_fail_page_alloc(char *str)
1397 return setup_fault_attr(&fail_page_alloc.attr, str);
1399 __setup("fail_page_alloc=", setup_fail_page_alloc);
1401 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1403 if (order < fail_page_alloc.min_order)
1405 if (gfp_mask & __GFP_NOFAIL)
1407 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1409 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1412 return should_fail(&fail_page_alloc.attr, 1 << order);
1415 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1417 static int __init fail_page_alloc_debugfs(void)
1419 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1423 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1427 dir = fail_page_alloc.attr.dentries.dir;
1429 fail_page_alloc.ignore_gfp_wait_file =
1430 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1431 &fail_page_alloc.ignore_gfp_wait);
1433 fail_page_alloc.ignore_gfp_highmem_file =
1434 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1435 &fail_page_alloc.ignore_gfp_highmem);
1436 fail_page_alloc.min_order_file =
1437 debugfs_create_u32("min-order", mode, dir,
1438 &fail_page_alloc.min_order);
1440 if (!fail_page_alloc.ignore_gfp_wait_file ||
1441 !fail_page_alloc.ignore_gfp_highmem_file ||
1442 !fail_page_alloc.min_order_file) {
1444 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1445 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1446 debugfs_remove(fail_page_alloc.min_order_file);
1447 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1453 late_initcall(fail_page_alloc_debugfs);
1455 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1457 #else /* CONFIG_FAIL_PAGE_ALLOC */
1459 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1464 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1467 * Return true if free pages are above 'mark'. This takes into account the order
1468 * of the allocation.
1470 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1471 int classzone_idx, int alloc_flags, long free_pages)
1473 /* free_pages my go negative - that's OK */
1477 free_pages -= (1 << order) + 1;
1478 if (alloc_flags & ALLOC_HIGH)
1480 if (alloc_flags & ALLOC_HARDER)
1483 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1485 for (o = 0; o < order; o++) {
1486 /* At the next order, this order's pages become unavailable */
1487 free_pages -= z->free_area[o].nr_free << o;
1489 /* Require fewer higher order pages to be free */
1492 if (free_pages <= min)
1498 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1499 int classzone_idx, int alloc_flags)
1501 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1502 zone_page_state(z, NR_FREE_PAGES));
1505 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1506 int classzone_idx, int alloc_flags)
1508 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1510 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1511 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1513 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1519 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1520 * skip over zones that are not allowed by the cpuset, or that have
1521 * been recently (in last second) found to be nearly full. See further
1522 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1523 * that have to skip over a lot of full or unallowed zones.
1525 * If the zonelist cache is present in the passed in zonelist, then
1526 * returns a pointer to the allowed node mask (either the current
1527 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1529 * If the zonelist cache is not available for this zonelist, does
1530 * nothing and returns NULL.
1532 * If the fullzones BITMAP in the zonelist cache is stale (more than
1533 * a second since last zap'd) then we zap it out (clear its bits.)
1535 * We hold off even calling zlc_setup, until after we've checked the
1536 * first zone in the zonelist, on the theory that most allocations will
1537 * be satisfied from that first zone, so best to examine that zone as
1538 * quickly as we can.
1540 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1542 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1543 nodemask_t *allowednodes; /* zonelist_cache approximation */
1545 zlc = zonelist->zlcache_ptr;
1549 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1550 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1551 zlc->last_full_zap = jiffies;
1554 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1555 &cpuset_current_mems_allowed :
1556 &node_states[N_HIGH_MEMORY];
1557 return allowednodes;
1561 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1562 * if it is worth looking at further for free memory:
1563 * 1) Check that the zone isn't thought to be full (doesn't have its
1564 * bit set in the zonelist_cache fullzones BITMAP).
1565 * 2) Check that the zones node (obtained from the zonelist_cache
1566 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1567 * Return true (non-zero) if zone is worth looking at further, or
1568 * else return false (zero) if it is not.
1570 * This check -ignores- the distinction between various watermarks,
1571 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1572 * found to be full for any variation of these watermarks, it will
1573 * be considered full for up to one second by all requests, unless
1574 * we are so low on memory on all allowed nodes that we are forced
1575 * into the second scan of the zonelist.
1577 * In the second scan we ignore this zonelist cache and exactly
1578 * apply the watermarks to all zones, even it is slower to do so.
1579 * We are low on memory in the second scan, and should leave no stone
1580 * unturned looking for a free page.
1582 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1583 nodemask_t *allowednodes)
1585 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1586 int i; /* index of *z in zonelist zones */
1587 int n; /* node that zone *z is on */
1589 zlc = zonelist->zlcache_ptr;
1593 i = z - zonelist->_zonerefs;
1596 /* This zone is worth trying if it is allowed but not full */
1597 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1601 * Given 'z' scanning a zonelist, set the corresponding bit in
1602 * zlc->fullzones, so that subsequent attempts to allocate a page
1603 * from that zone don't waste time re-examining it.
1605 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1607 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1608 int i; /* index of *z in zonelist zones */
1610 zlc = zonelist->zlcache_ptr;
1614 i = z - zonelist->_zonerefs;
1616 set_bit(i, zlc->fullzones);
1620 * clear all zones full, called after direct reclaim makes progress so that
1621 * a zone that was recently full is not skipped over for up to a second
1623 static void zlc_clear_zones_full(struct zonelist *zonelist)
1625 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1627 zlc = zonelist->zlcache_ptr;
1631 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1634 #else /* CONFIG_NUMA */
1636 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1641 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1642 nodemask_t *allowednodes)
1647 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1651 static void zlc_clear_zones_full(struct zonelist *zonelist)
1654 #endif /* CONFIG_NUMA */
1657 * get_page_from_freelist goes through the zonelist trying to allocate
1660 static struct page *
1661 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1662 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1663 struct zone *preferred_zone, int migratetype)
1666 struct page *page = NULL;
1669 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1670 int zlc_active = 0; /* set if using zonelist_cache */
1671 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1673 classzone_idx = zone_idx(preferred_zone);
1676 * Scan zonelist, looking for a zone with enough free.
1677 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1679 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1680 high_zoneidx, nodemask) {
1681 if (NUMA_BUILD && zlc_active &&
1682 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1684 if ((alloc_flags & ALLOC_CPUSET) &&
1685 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1688 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1689 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1693 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1694 if (zone_watermark_ok(zone, order, mark,
1695 classzone_idx, alloc_flags))
1698 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1700 * we do zlc_setup if there are multiple nodes
1701 * and before considering the first zone allowed
1704 allowednodes = zlc_setup(zonelist, alloc_flags);
1709 if (zone_reclaim_mode == 0)
1710 goto this_zone_full;
1713 * As we may have just activated ZLC, check if the first
1714 * eligible zone has failed zone_reclaim recently.
1716 if (NUMA_BUILD && zlc_active &&
1717 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1720 ret = zone_reclaim(zone, gfp_mask, order);
1722 case ZONE_RECLAIM_NOSCAN:
1725 case ZONE_RECLAIM_FULL:
1726 /* scanned but unreclaimable */
1729 /* did we reclaim enough */
1730 if (!zone_watermark_ok(zone, order, mark,
1731 classzone_idx, alloc_flags))
1732 goto this_zone_full;
1737 page = buffered_rmqueue(preferred_zone, zone, order,
1738 gfp_mask, migratetype);
1743 zlc_mark_zone_full(zonelist, z);
1746 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1747 /* Disable zlc cache for second zonelist scan */
1755 * Large machines with many possible nodes should not always dump per-node
1756 * meminfo in irq context.
1758 static inline bool should_suppress_show_mem(void)
1763 ret = in_interrupt();
1768 static DEFINE_RATELIMIT_STATE(nopage_rs,
1769 DEFAULT_RATELIMIT_INTERVAL,
1770 DEFAULT_RATELIMIT_BURST);
1772 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1775 unsigned int filter = SHOW_MEM_FILTER_NODES;
1777 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1781 * This documents exceptions given to allocations in certain
1782 * contexts that are allowed to allocate outside current's set
1785 if (!(gfp_mask & __GFP_NOMEMALLOC))
1786 if (test_thread_flag(TIF_MEMDIE) ||
1787 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1788 filter &= ~SHOW_MEM_FILTER_NODES;
1789 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1790 filter &= ~SHOW_MEM_FILTER_NODES;
1793 printk(KERN_WARNING);
1794 va_start(args, fmt);
1799 pr_warning("%s: page allocation failure: order:%d, mode:0x%x\n",
1800 current->comm, order, gfp_mask);
1803 if (!should_suppress_show_mem())
1808 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1809 unsigned long pages_reclaimed)
1811 /* Do not loop if specifically requested */
1812 if (gfp_mask & __GFP_NORETRY)
1816 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1817 * means __GFP_NOFAIL, but that may not be true in other
1820 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1824 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1825 * specified, then we retry until we no longer reclaim any pages
1826 * (above), or we've reclaimed an order of pages at least as
1827 * large as the allocation's order. In both cases, if the
1828 * allocation still fails, we stop retrying.
1830 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1834 * Don't let big-order allocations loop unless the caller
1835 * explicitly requests that.
1837 if (gfp_mask & __GFP_NOFAIL)
1843 static inline struct page *
1844 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1845 struct zonelist *zonelist, enum zone_type high_zoneidx,
1846 nodemask_t *nodemask, struct zone *preferred_zone,
1851 /* Acquire the OOM killer lock for the zones in zonelist */
1852 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1853 schedule_timeout_uninterruptible(1);
1858 * Go through the zonelist yet one more time, keep very high watermark
1859 * here, this is only to catch a parallel oom killing, we must fail if
1860 * we're still under heavy pressure.
1862 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1863 order, zonelist, high_zoneidx,
1864 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1865 preferred_zone, migratetype);
1869 if (!(gfp_mask & __GFP_NOFAIL)) {
1870 /* The OOM killer will not help higher order allocs */
1871 if (order > PAGE_ALLOC_COSTLY_ORDER)
1873 /* The OOM killer does not needlessly kill tasks for lowmem */
1874 if (high_zoneidx < ZONE_NORMAL)
1877 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1878 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1879 * The caller should handle page allocation failure by itself if
1880 * it specifies __GFP_THISNODE.
1881 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1883 if (gfp_mask & __GFP_THISNODE)
1886 /* Exhausted what can be done so it's blamo time */
1887 out_of_memory(zonelist, gfp_mask, order, nodemask);
1890 clear_zonelist_oom(zonelist, gfp_mask);
1894 #ifdef CONFIG_COMPACTION
1895 /* Try memory compaction for high-order allocations before reclaim */
1896 static struct page *
1897 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1898 struct zonelist *zonelist, enum zone_type high_zoneidx,
1899 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1900 int migratetype, bool sync_migration,
1901 bool *deferred_compaction,
1902 unsigned long *did_some_progress)
1909 if (compaction_deferred(preferred_zone)) {
1910 *deferred_compaction = true;
1914 current->flags |= PF_MEMALLOC;
1915 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1916 nodemask, sync_migration);
1917 current->flags &= ~PF_MEMALLOC;
1918 if (*did_some_progress != COMPACT_SKIPPED) {
1920 /* Page migration frees to the PCP lists but we want merging */
1921 drain_pages(get_cpu());
1924 page = get_page_from_freelist(gfp_mask, nodemask,
1925 order, zonelist, high_zoneidx,
1926 alloc_flags, preferred_zone,
1929 preferred_zone->compact_considered = 0;
1930 preferred_zone->compact_defer_shift = 0;
1931 count_vm_event(COMPACTSUCCESS);
1936 * It's bad if compaction run occurs and fails.
1937 * The most likely reason is that pages exist,
1938 * but not enough to satisfy watermarks.
1940 count_vm_event(COMPACTFAIL);
1943 * As async compaction considers a subset of pageblocks, only
1944 * defer if the failure was a sync compaction failure.
1947 defer_compaction(preferred_zone);
1955 static inline struct page *
1956 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1957 struct zonelist *zonelist, enum zone_type high_zoneidx,
1958 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1959 int migratetype, bool sync_migration,
1960 bool *deferred_compaction,
1961 unsigned long *did_some_progress)
1965 #endif /* CONFIG_COMPACTION */
1967 /* The really slow allocator path where we enter direct reclaim */
1968 static inline struct page *
1969 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1970 struct zonelist *zonelist, enum zone_type high_zoneidx,
1971 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1972 int migratetype, unsigned long *did_some_progress)
1974 struct page *page = NULL;
1975 struct reclaim_state reclaim_state;
1976 bool drained = false;
1980 /* We now go into synchronous reclaim */
1981 cpuset_memory_pressure_bump();
1982 current->flags |= PF_MEMALLOC;
1983 lockdep_set_current_reclaim_state(gfp_mask);
1984 reclaim_state.reclaimed_slab = 0;
1985 current->reclaim_state = &reclaim_state;
1987 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1989 current->reclaim_state = NULL;
1990 lockdep_clear_current_reclaim_state();
1991 current->flags &= ~PF_MEMALLOC;
1995 if (unlikely(!(*did_some_progress)))
1998 /* After successful reclaim, reconsider all zones for allocation */
2000 zlc_clear_zones_full(zonelist);
2003 page = get_page_from_freelist(gfp_mask, nodemask, order,
2004 zonelist, high_zoneidx,
2005 alloc_flags, preferred_zone,
2009 * If an allocation failed after direct reclaim, it could be because
2010 * pages are pinned on the per-cpu lists. Drain them and try again
2012 if (!page && !drained) {
2022 * This is called in the allocator slow-path if the allocation request is of
2023 * sufficient urgency to ignore watermarks and take other desperate measures
2025 static inline struct page *
2026 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2027 struct zonelist *zonelist, enum zone_type high_zoneidx,
2028 nodemask_t *nodemask, struct zone *preferred_zone,
2034 page = get_page_from_freelist(gfp_mask, nodemask, order,
2035 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2036 preferred_zone, migratetype);
2038 if (!page && gfp_mask & __GFP_NOFAIL)
2039 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2040 } while (!page && (gfp_mask & __GFP_NOFAIL));
2046 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2047 enum zone_type high_zoneidx,
2048 enum zone_type classzone_idx)
2053 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2054 wakeup_kswapd(zone, order, classzone_idx);
2058 gfp_to_alloc_flags(gfp_t gfp_mask)
2060 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2061 const gfp_t wait = gfp_mask & __GFP_WAIT;
2063 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2064 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2067 * The caller may dip into page reserves a bit more if the caller
2068 * cannot run direct reclaim, or if the caller has realtime scheduling
2069 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2070 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2072 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2076 * Not worth trying to allocate harder for
2077 * __GFP_NOMEMALLOC even if it can't schedule.
2079 if (!(gfp_mask & __GFP_NOMEMALLOC))
2080 alloc_flags |= ALLOC_HARDER;
2082 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2083 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2085 alloc_flags &= ~ALLOC_CPUSET;
2086 } else if (unlikely(rt_task(current)) && !in_interrupt())
2087 alloc_flags |= ALLOC_HARDER;
2089 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2090 if (!in_interrupt() &&
2091 ((current->flags & PF_MEMALLOC) ||
2092 unlikely(test_thread_flag(TIF_MEMDIE))))
2093 alloc_flags |= ALLOC_NO_WATERMARKS;
2099 static inline struct page *
2100 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2101 struct zonelist *zonelist, enum zone_type high_zoneidx,
2102 nodemask_t *nodemask, struct zone *preferred_zone,
2105 const gfp_t wait = gfp_mask & __GFP_WAIT;
2106 struct page *page = NULL;
2108 unsigned long pages_reclaimed = 0;
2109 unsigned long did_some_progress;
2110 bool sync_migration = false;
2111 bool deferred_compaction = false;
2114 * In the slowpath, we sanity check order to avoid ever trying to
2115 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2116 * be using allocators in order of preference for an area that is
2119 if (order >= MAX_ORDER) {
2120 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2125 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2126 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2127 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2128 * using a larger set of nodes after it has established that the
2129 * allowed per node queues are empty and that nodes are
2132 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2136 if (!(gfp_mask & __GFP_NO_KSWAPD))
2137 wake_all_kswapd(order, zonelist, high_zoneidx,
2138 zone_idx(preferred_zone));
2141 * OK, we're below the kswapd watermark and have kicked background
2142 * reclaim. Now things get more complex, so set up alloc_flags according
2143 * to how we want to proceed.
2145 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2148 * Find the true preferred zone if the allocation is unconstrained by
2151 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2152 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2156 /* This is the last chance, in general, before the goto nopage. */
2157 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2158 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2159 preferred_zone, migratetype);
2163 /* Allocate without watermarks if the context allows */
2164 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2165 page = __alloc_pages_high_priority(gfp_mask, order,
2166 zonelist, high_zoneidx, nodemask,
2167 preferred_zone, migratetype);
2172 /* Atomic allocations - we can't balance anything */
2176 /* Avoid recursion of direct reclaim */
2177 if (current->flags & PF_MEMALLOC)
2180 /* Avoid allocations with no watermarks from looping endlessly */
2181 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2185 * Try direct compaction. The first pass is asynchronous. Subsequent
2186 * attempts after direct reclaim are synchronous
2188 page = __alloc_pages_direct_compact(gfp_mask, order,
2189 zonelist, high_zoneidx,
2191 alloc_flags, preferred_zone,
2192 migratetype, sync_migration,
2193 &deferred_compaction,
2194 &did_some_progress);
2197 sync_migration = true;
2200 * If compaction is deferred for high-order allocations, it is because
2201 * sync compaction recently failed. In this is the case and the caller
2202 * has requested the system not be heavily disrupted, fail the
2203 * allocation now instead of entering direct reclaim
2205 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2208 /* Try direct reclaim and then allocating */
2209 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2210 zonelist, high_zoneidx,
2212 alloc_flags, preferred_zone,
2213 migratetype, &did_some_progress);
2218 * If we failed to make any progress reclaiming, then we are
2219 * running out of options and have to consider going OOM
2221 if (!did_some_progress) {
2222 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2223 if (oom_killer_disabled)
2225 page = __alloc_pages_may_oom(gfp_mask, order,
2226 zonelist, high_zoneidx,
2227 nodemask, preferred_zone,
2232 if (!(gfp_mask & __GFP_NOFAIL)) {
2234 * The oom killer is not called for high-order
2235 * allocations that may fail, so if no progress
2236 * is being made, there are no other options and
2237 * retrying is unlikely to help.
2239 if (order > PAGE_ALLOC_COSTLY_ORDER)
2242 * The oom killer is not called for lowmem
2243 * allocations to prevent needlessly killing
2246 if (high_zoneidx < ZONE_NORMAL)
2254 /* Check if we should retry the allocation */
2255 pages_reclaimed += did_some_progress;
2256 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2257 /* Wait for some write requests to complete then retry */
2258 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2262 * High-order allocations do not necessarily loop after
2263 * direct reclaim and reclaim/compaction depends on compaction
2264 * being called after reclaim so call directly if necessary
2266 page = __alloc_pages_direct_compact(gfp_mask, order,
2267 zonelist, high_zoneidx,
2269 alloc_flags, preferred_zone,
2270 migratetype, sync_migration,
2271 &deferred_compaction,
2272 &did_some_progress);
2278 warn_alloc_failed(gfp_mask, order, NULL);
2281 if (kmemcheck_enabled)
2282 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2288 * This is the 'heart' of the zoned buddy allocator.
2291 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2292 struct zonelist *zonelist, nodemask_t *nodemask)
2294 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2295 struct zone *preferred_zone;
2296 struct page *page = NULL;
2297 int migratetype = allocflags_to_migratetype(gfp_mask);
2298 unsigned int cpuset_mems_cookie;
2300 gfp_mask &= gfp_allowed_mask;
2302 lockdep_trace_alloc(gfp_mask);
2304 might_sleep_if(gfp_mask & __GFP_WAIT);
2306 if (should_fail_alloc_page(gfp_mask, order))
2310 * Check the zones suitable for the gfp_mask contain at least one
2311 * valid zone. It's possible to have an empty zonelist as a result
2312 * of GFP_THISNODE and a memoryless node
2314 if (unlikely(!zonelist->_zonerefs->zone))
2318 cpuset_mems_cookie = get_mems_allowed();
2320 /* The preferred zone is used for statistics later */
2321 first_zones_zonelist(zonelist, high_zoneidx,
2322 nodemask ? : &cpuset_current_mems_allowed,
2324 if (!preferred_zone)
2327 /* First allocation attempt */
2328 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2329 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2330 preferred_zone, migratetype);
2331 if (unlikely(!page))
2332 page = __alloc_pages_slowpath(gfp_mask, order,
2333 zonelist, high_zoneidx, nodemask,
2334 preferred_zone, migratetype);
2336 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2340 * When updating a task's mems_allowed, it is possible to race with
2341 * parallel threads in such a way that an allocation can fail while
2342 * the mask is being updated. If a page allocation is about to fail,
2343 * check if the cpuset changed during allocation and if so, retry.
2345 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2350 EXPORT_SYMBOL(__alloc_pages_nodemask);
2353 * Common helper functions.
2355 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2360 * __get_free_pages() returns a 32-bit address, which cannot represent
2363 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2365 page = alloc_pages(gfp_mask, order);
2368 return (unsigned long) page_address(page);
2370 EXPORT_SYMBOL(__get_free_pages);
2372 unsigned long get_zeroed_page(gfp_t gfp_mask)
2374 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2376 EXPORT_SYMBOL(get_zeroed_page);
2378 void __pagevec_free(struct pagevec *pvec)
2380 int i = pagevec_count(pvec);
2383 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2384 free_hot_cold_page(pvec->pages[i], pvec->cold);
2388 void __free_pages(struct page *page, unsigned int order)
2390 if (put_page_testzero(page)) {
2392 free_hot_cold_page(page, 0);
2394 __free_pages_ok(page, order);
2398 EXPORT_SYMBOL(__free_pages);
2400 void free_pages(unsigned long addr, unsigned int order)
2403 VM_BUG_ON(!virt_addr_valid((void *)addr));
2404 __free_pages(virt_to_page((void *)addr), order);
2408 EXPORT_SYMBOL(free_pages);
2410 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2413 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2414 unsigned long used = addr + PAGE_ALIGN(size);
2416 split_page(virt_to_page((void *)addr), order);
2417 while (used < alloc_end) {
2422 return (void *)addr;
2426 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2427 * @size: the number of bytes to allocate
2428 * @gfp_mask: GFP flags for the allocation
2430 * This function is similar to alloc_pages(), except that it allocates the
2431 * minimum number of pages to satisfy the request. alloc_pages() can only
2432 * allocate memory in power-of-two pages.
2434 * This function is also limited by MAX_ORDER.
2436 * Memory allocated by this function must be released by free_pages_exact().
2438 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2440 unsigned int order = get_order(size);
2443 addr = __get_free_pages(gfp_mask, order);
2444 return make_alloc_exact(addr, order, size);
2446 EXPORT_SYMBOL(alloc_pages_exact);
2449 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2451 * @nid: the preferred node ID where memory should be allocated
2452 * @size: the number of bytes to allocate
2453 * @gfp_mask: GFP flags for the allocation
2455 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2457 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2460 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2462 unsigned order = get_order(size);
2463 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2466 return make_alloc_exact((unsigned long)page_address(p), order, size);
2468 EXPORT_SYMBOL(alloc_pages_exact_nid);
2471 * free_pages_exact - release memory allocated via alloc_pages_exact()
2472 * @virt: the value returned by alloc_pages_exact.
2473 * @size: size of allocation, same value as passed to alloc_pages_exact().
2475 * Release the memory allocated by a previous call to alloc_pages_exact.
2477 void free_pages_exact(void *virt, size_t size)
2479 unsigned long addr = (unsigned long)virt;
2480 unsigned long end = addr + PAGE_ALIGN(size);
2482 while (addr < end) {
2487 EXPORT_SYMBOL(free_pages_exact);
2489 static unsigned int nr_free_zone_pages(int offset)
2494 /* Just pick one node, since fallback list is circular */
2495 unsigned int sum = 0;
2497 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2499 for_each_zone_zonelist(zone, z, zonelist, offset) {
2500 unsigned long size = zone->present_pages;
2501 unsigned long high = high_wmark_pages(zone);
2510 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2512 unsigned int nr_free_buffer_pages(void)
2514 return nr_free_zone_pages(gfp_zone(GFP_USER));
2516 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2519 * Amount of free RAM allocatable within all zones
2521 unsigned int nr_free_pagecache_pages(void)
2523 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2526 static inline void show_node(struct zone *zone)
2529 printk("Node %d ", zone_to_nid(zone));
2532 void si_meminfo(struct sysinfo *val)
2534 val->totalram = totalram_pages;
2536 val->freeram = global_page_state(NR_FREE_PAGES);
2537 val->bufferram = nr_blockdev_pages();
2538 val->totalhigh = totalhigh_pages;
2539 val->freehigh = nr_free_highpages();
2540 val->mem_unit = PAGE_SIZE;
2543 EXPORT_SYMBOL(si_meminfo);
2546 void si_meminfo_node(struct sysinfo *val, int nid)
2548 pg_data_t *pgdat = NODE_DATA(nid);
2550 val->totalram = pgdat->node_present_pages;
2551 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2552 #ifdef CONFIG_HIGHMEM
2553 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2554 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2560 val->mem_unit = PAGE_SIZE;
2565 * Determine whether the node should be displayed or not, depending on whether
2566 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2568 bool skip_free_areas_node(unsigned int flags, int nid)
2571 unsigned int cpuset_mems_cookie;
2573 if (!(flags & SHOW_MEM_FILTER_NODES))
2577 cpuset_mems_cookie = get_mems_allowed();
2578 ret = !node_isset(nid, cpuset_current_mems_allowed);
2579 } while (!put_mems_allowed(cpuset_mems_cookie));
2584 #define K(x) ((x) << (PAGE_SHIFT-10))
2587 * Show free area list (used inside shift_scroll-lock stuff)
2588 * We also calculate the percentage fragmentation. We do this by counting the
2589 * memory on each free list with the exception of the first item on the list.
2590 * Suppresses nodes that are not allowed by current's cpuset if
2591 * SHOW_MEM_FILTER_NODES is passed.
2593 void show_free_areas(unsigned int filter)
2598 for_each_populated_zone(zone) {
2599 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2602 printk("%s per-cpu:\n", zone->name);
2604 for_each_online_cpu(cpu) {
2605 struct per_cpu_pageset *pageset;
2607 pageset = per_cpu_ptr(zone->pageset, cpu);
2609 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2610 cpu, pageset->pcp.high,
2611 pageset->pcp.batch, pageset->pcp.count);
2615 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2616 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2618 " dirty:%lu writeback:%lu unstable:%lu\n"
2619 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2620 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2621 global_page_state(NR_ACTIVE_ANON),
2622 global_page_state(NR_INACTIVE_ANON),
2623 global_page_state(NR_ISOLATED_ANON),
2624 global_page_state(NR_ACTIVE_FILE),
2625 global_page_state(NR_INACTIVE_FILE),
2626 global_page_state(NR_ISOLATED_FILE),
2627 global_page_state(NR_UNEVICTABLE),
2628 global_page_state(NR_FILE_DIRTY),
2629 global_page_state(NR_WRITEBACK),
2630 global_page_state(NR_UNSTABLE_NFS),
2631 global_page_state(NR_FREE_PAGES),
2632 global_page_state(NR_SLAB_RECLAIMABLE),
2633 global_page_state(NR_SLAB_UNRECLAIMABLE),
2634 global_page_state(NR_FILE_MAPPED),
2635 global_page_state(NR_SHMEM),
2636 global_page_state(NR_PAGETABLE),
2637 global_page_state(NR_BOUNCE));
2639 for_each_populated_zone(zone) {
2642 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2650 " active_anon:%lukB"
2651 " inactive_anon:%lukB"
2652 " active_file:%lukB"
2653 " inactive_file:%lukB"
2654 " unevictable:%lukB"
2655 " isolated(anon):%lukB"
2656 " isolated(file):%lukB"
2663 " slab_reclaimable:%lukB"
2664 " slab_unreclaimable:%lukB"
2665 " kernel_stack:%lukB"
2669 " writeback_tmp:%lukB"
2670 " pages_scanned:%lu"
2671 " all_unreclaimable? %s"
2674 K(zone_page_state(zone, NR_FREE_PAGES)),
2675 K(min_wmark_pages(zone)),
2676 K(low_wmark_pages(zone)),
2677 K(high_wmark_pages(zone)),
2678 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2679 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2680 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2681 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2682 K(zone_page_state(zone, NR_UNEVICTABLE)),
2683 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2684 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2685 K(zone->present_pages),
2686 K(zone_page_state(zone, NR_MLOCK)),
2687 K(zone_page_state(zone, NR_FILE_DIRTY)),
2688 K(zone_page_state(zone, NR_WRITEBACK)),
2689 K(zone_page_state(zone, NR_FILE_MAPPED)),
2690 K(zone_page_state(zone, NR_SHMEM)),
2691 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2692 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2693 zone_page_state(zone, NR_KERNEL_STACK) *
2695 K(zone_page_state(zone, NR_PAGETABLE)),
2696 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2697 K(zone_page_state(zone, NR_BOUNCE)),
2698 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2699 zone->pages_scanned,
2700 (zone->all_unreclaimable ? "yes" : "no")
2702 printk("lowmem_reserve[]:");
2703 for (i = 0; i < MAX_NR_ZONES; i++)
2704 printk(" %lu", zone->lowmem_reserve[i]);
2708 for_each_populated_zone(zone) {
2709 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2711 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2714 printk("%s: ", zone->name);
2716 spin_lock_irqsave(&zone->lock, flags);
2717 for (order = 0; order < MAX_ORDER; order++) {
2718 nr[order] = zone->free_area[order].nr_free;
2719 total += nr[order] << order;
2721 spin_unlock_irqrestore(&zone->lock, flags);
2722 for (order = 0; order < MAX_ORDER; order++)
2723 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2724 printk("= %lukB\n", K(total));
2727 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2729 show_swap_cache_info();
2732 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2734 zoneref->zone = zone;
2735 zoneref->zone_idx = zone_idx(zone);
2739 * Builds allocation fallback zone lists.
2741 * Add all populated zones of a node to the zonelist.
2743 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2744 int nr_zones, enum zone_type zone_type)
2748 BUG_ON(zone_type >= MAX_NR_ZONES);
2753 zone = pgdat->node_zones + zone_type;
2754 if (populated_zone(zone)) {
2755 zoneref_set_zone(zone,
2756 &zonelist->_zonerefs[nr_zones++]);
2757 check_highest_zone(zone_type);
2760 } while (zone_type);
2767 * 0 = automatic detection of better ordering.
2768 * 1 = order by ([node] distance, -zonetype)
2769 * 2 = order by (-zonetype, [node] distance)
2771 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2772 * the same zonelist. So only NUMA can configure this param.
2774 #define ZONELIST_ORDER_DEFAULT 0
2775 #define ZONELIST_ORDER_NODE 1
2776 #define ZONELIST_ORDER_ZONE 2
2778 /* zonelist order in the kernel.
2779 * set_zonelist_order() will set this to NODE or ZONE.
2781 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2782 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2786 /* The value user specified ....changed by config */
2787 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2788 /* string for sysctl */
2789 #define NUMA_ZONELIST_ORDER_LEN 16
2790 char numa_zonelist_order[16] = "default";
2793 * interface for configure zonelist ordering.
2794 * command line option "numa_zonelist_order"
2795 * = "[dD]efault - default, automatic configuration.
2796 * = "[nN]ode - order by node locality, then by zone within node
2797 * = "[zZ]one - order by zone, then by locality within zone
2800 static int __parse_numa_zonelist_order(char *s)
2802 if (*s == 'd' || *s == 'D') {
2803 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2804 } else if (*s == 'n' || *s == 'N') {
2805 user_zonelist_order = ZONELIST_ORDER_NODE;
2806 } else if (*s == 'z' || *s == 'Z') {
2807 user_zonelist_order = ZONELIST_ORDER_ZONE;
2810 "Ignoring invalid numa_zonelist_order value: "
2817 static __init int setup_numa_zonelist_order(char *s)
2824 ret = __parse_numa_zonelist_order(s);
2826 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2830 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2833 * sysctl handler for numa_zonelist_order
2835 int numa_zonelist_order_handler(ctl_table *table, int write,
2836 void __user *buffer, size_t *length,
2839 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2841 static DEFINE_MUTEX(zl_order_mutex);
2843 mutex_lock(&zl_order_mutex);
2845 strcpy(saved_string, (char*)table->data);
2846 ret = proc_dostring(table, write, buffer, length, ppos);
2850 int oldval = user_zonelist_order;
2851 if (__parse_numa_zonelist_order((char*)table->data)) {
2853 * bogus value. restore saved string
2855 strncpy((char*)table->data, saved_string,
2856 NUMA_ZONELIST_ORDER_LEN);
2857 user_zonelist_order = oldval;
2858 } else if (oldval != user_zonelist_order) {
2859 mutex_lock(&zonelists_mutex);
2860 build_all_zonelists(NULL);
2861 mutex_unlock(&zonelists_mutex);
2865 mutex_unlock(&zl_order_mutex);
2870 #define MAX_NODE_LOAD (nr_online_nodes)
2871 static int node_load[MAX_NUMNODES];
2874 * find_next_best_node - find the next node that should appear in a given node's fallback list
2875 * @node: node whose fallback list we're appending
2876 * @used_node_mask: nodemask_t of already used nodes
2878 * We use a number of factors to determine which is the next node that should
2879 * appear on a given node's fallback list. The node should not have appeared
2880 * already in @node's fallback list, and it should be the next closest node
2881 * according to the distance array (which contains arbitrary distance values
2882 * from each node to each node in the system), and should also prefer nodes
2883 * with no CPUs, since presumably they'll have very little allocation pressure
2884 * on them otherwise.
2885 * It returns -1 if no node is found.
2887 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2890 int min_val = INT_MAX;
2892 const struct cpumask *tmp = cpumask_of_node(0);
2894 /* Use the local node if we haven't already */
2895 if (!node_isset(node, *used_node_mask)) {
2896 node_set(node, *used_node_mask);
2900 for_each_node_state(n, N_HIGH_MEMORY) {
2902 /* Don't want a node to appear more than once */
2903 if (node_isset(n, *used_node_mask))
2906 /* Use the distance array to find the distance */
2907 val = node_distance(node, n);
2909 /* Penalize nodes under us ("prefer the next node") */
2912 /* Give preference to headless and unused nodes */
2913 tmp = cpumask_of_node(n);
2914 if (!cpumask_empty(tmp))
2915 val += PENALTY_FOR_NODE_WITH_CPUS;
2917 /* Slight preference for less loaded node */
2918 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2919 val += node_load[n];
2921 if (val < min_val) {
2928 node_set(best_node, *used_node_mask);
2935 * Build zonelists ordered by node and zones within node.
2936 * This results in maximum locality--normal zone overflows into local
2937 * DMA zone, if any--but risks exhausting DMA zone.
2939 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2942 struct zonelist *zonelist;
2944 zonelist = &pgdat->node_zonelists[0];
2945 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2947 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2949 zonelist->_zonerefs[j].zone = NULL;
2950 zonelist->_zonerefs[j].zone_idx = 0;
2954 * Build gfp_thisnode zonelists
2956 static void build_thisnode_zonelists(pg_data_t *pgdat)
2959 struct zonelist *zonelist;
2961 zonelist = &pgdat->node_zonelists[1];
2962 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2963 zonelist->_zonerefs[j].zone = NULL;
2964 zonelist->_zonerefs[j].zone_idx = 0;
2968 * Build zonelists ordered by zone and nodes within zones.
2969 * This results in conserving DMA zone[s] until all Normal memory is
2970 * exhausted, but results in overflowing to remote node while memory
2971 * may still exist in local DMA zone.
2973 static int node_order[MAX_NUMNODES];
2975 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2978 int zone_type; /* needs to be signed */
2980 struct zonelist *zonelist;
2982 zonelist = &pgdat->node_zonelists[0];
2984 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2985 for (j = 0; j < nr_nodes; j++) {
2986 node = node_order[j];
2987 z = &NODE_DATA(node)->node_zones[zone_type];
2988 if (populated_zone(z)) {
2990 &zonelist->_zonerefs[pos++]);
2991 check_highest_zone(zone_type);
2995 zonelist->_zonerefs[pos].zone = NULL;
2996 zonelist->_zonerefs[pos].zone_idx = 0;
2999 static int default_zonelist_order(void)
3002 unsigned long low_kmem_size,total_size;
3006 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3007 * If they are really small and used heavily, the system can fall
3008 * into OOM very easily.
3009 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3011 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3014 for_each_online_node(nid) {
3015 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3016 z = &NODE_DATA(nid)->node_zones[zone_type];
3017 if (populated_zone(z)) {
3018 if (zone_type < ZONE_NORMAL)
3019 low_kmem_size += z->present_pages;
3020 total_size += z->present_pages;
3021 } else if (zone_type == ZONE_NORMAL) {
3023 * If any node has only lowmem, then node order
3024 * is preferred to allow kernel allocations
3025 * locally; otherwise, they can easily infringe
3026 * on other nodes when there is an abundance of
3027 * lowmem available to allocate from.
3029 return ZONELIST_ORDER_NODE;
3033 if (!low_kmem_size || /* there are no DMA area. */
3034 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3035 return ZONELIST_ORDER_NODE;
3037 * look into each node's config.
3038 * If there is a node whose DMA/DMA32 memory is very big area on
3039 * local memory, NODE_ORDER may be suitable.
3041 average_size = total_size /
3042 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3043 for_each_online_node(nid) {
3046 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3047 z = &NODE_DATA(nid)->node_zones[zone_type];
3048 if (populated_zone(z)) {
3049 if (zone_type < ZONE_NORMAL)
3050 low_kmem_size += z->present_pages;
3051 total_size += z->present_pages;
3054 if (low_kmem_size &&
3055 total_size > average_size && /* ignore small node */
3056 low_kmem_size > total_size * 70/100)
3057 return ZONELIST_ORDER_NODE;
3059 return ZONELIST_ORDER_ZONE;
3062 static void set_zonelist_order(void)
3064 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3065 current_zonelist_order = default_zonelist_order();
3067 current_zonelist_order = user_zonelist_order;
3070 static void build_zonelists(pg_data_t *pgdat)
3074 nodemask_t used_mask;
3075 int local_node, prev_node;
3076 struct zonelist *zonelist;
3077 int order = current_zonelist_order;
3079 /* initialize zonelists */
3080 for (i = 0; i < MAX_ZONELISTS; i++) {
3081 zonelist = pgdat->node_zonelists + i;
3082 zonelist->_zonerefs[0].zone = NULL;
3083 zonelist->_zonerefs[0].zone_idx = 0;
3086 /* NUMA-aware ordering of nodes */
3087 local_node = pgdat->node_id;
3088 load = nr_online_nodes;
3089 prev_node = local_node;
3090 nodes_clear(used_mask);
3092 memset(node_order, 0, sizeof(node_order));
3095 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3096 int distance = node_distance(local_node, node);
3099 * If another node is sufficiently far away then it is better
3100 * to reclaim pages in a zone before going off node.
3102 if (distance > RECLAIM_DISTANCE)
3103 zone_reclaim_mode = 1;
3106 * We don't want to pressure a particular node.
3107 * So adding penalty to the first node in same
3108 * distance group to make it round-robin.
3110 if (distance != node_distance(local_node, prev_node))
3111 node_load[node] = load;
3115 if (order == ZONELIST_ORDER_NODE)
3116 build_zonelists_in_node_order(pgdat, node);
3118 node_order[j++] = node; /* remember order */
3121 if (order == ZONELIST_ORDER_ZONE) {
3122 /* calculate node order -- i.e., DMA last! */
3123 build_zonelists_in_zone_order(pgdat, j);
3126 build_thisnode_zonelists(pgdat);
3129 /* Construct the zonelist performance cache - see further mmzone.h */
3130 static void build_zonelist_cache(pg_data_t *pgdat)
3132 struct zonelist *zonelist;
3133 struct zonelist_cache *zlc;
3136 zonelist = &pgdat->node_zonelists[0];
3137 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3138 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3139 for (z = zonelist->_zonerefs; z->zone; z++)
3140 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3143 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3145 * Return node id of node used for "local" allocations.
3146 * I.e., first node id of first zone in arg node's generic zonelist.
3147 * Used for initializing percpu 'numa_mem', which is used primarily
3148 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3150 int local_memory_node(int node)
3154 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3155 gfp_zone(GFP_KERNEL),
3162 #else /* CONFIG_NUMA */
3164 static void set_zonelist_order(void)
3166 current_zonelist_order = ZONELIST_ORDER_ZONE;
3169 static void build_zonelists(pg_data_t *pgdat)
3171 int node, local_node;
3173 struct zonelist *zonelist;
3175 local_node = pgdat->node_id;
3177 zonelist = &pgdat->node_zonelists[0];
3178 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3181 * Now we build the zonelist so that it contains the zones
3182 * of all the other nodes.
3183 * We don't want to pressure a particular node, so when
3184 * building the zones for node N, we make sure that the
3185 * zones coming right after the local ones are those from
3186 * node N+1 (modulo N)
3188 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3189 if (!node_online(node))
3191 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3194 for (node = 0; node < local_node; node++) {
3195 if (!node_online(node))
3197 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3201 zonelist->_zonerefs[j].zone = NULL;
3202 zonelist->_zonerefs[j].zone_idx = 0;
3205 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3206 static void build_zonelist_cache(pg_data_t *pgdat)
3208 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3211 #endif /* CONFIG_NUMA */
3214 * Boot pageset table. One per cpu which is going to be used for all
3215 * zones and all nodes. The parameters will be set in such a way
3216 * that an item put on a list will immediately be handed over to
3217 * the buddy list. This is safe since pageset manipulation is done
3218 * with interrupts disabled.
3220 * The boot_pagesets must be kept even after bootup is complete for
3221 * unused processors and/or zones. They do play a role for bootstrapping
3222 * hotplugged processors.
3224 * zoneinfo_show() and maybe other functions do
3225 * not check if the processor is online before following the pageset pointer.
3226 * Other parts of the kernel may not check if the zone is available.
3228 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3229 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3230 static void setup_zone_pageset(struct zone *zone);
3233 * Global mutex to protect against size modification of zonelists
3234 * as well as to serialize pageset setup for the new populated zone.
3236 DEFINE_MUTEX(zonelists_mutex);
3238 /* return values int ....just for stop_machine() */
3239 static __init_refok int __build_all_zonelists(void *data)
3245 memset(node_load, 0, sizeof(node_load));
3247 for_each_online_node(nid) {
3248 pg_data_t *pgdat = NODE_DATA(nid);
3250 build_zonelists(pgdat);
3251 build_zonelist_cache(pgdat);
3255 * Initialize the boot_pagesets that are going to be used
3256 * for bootstrapping processors. The real pagesets for
3257 * each zone will be allocated later when the per cpu
3258 * allocator is available.
3260 * boot_pagesets are used also for bootstrapping offline
3261 * cpus if the system is already booted because the pagesets
3262 * are needed to initialize allocators on a specific cpu too.
3263 * F.e. the percpu allocator needs the page allocator which
3264 * needs the percpu allocator in order to allocate its pagesets
3265 * (a chicken-egg dilemma).
3267 for_each_possible_cpu(cpu) {
3268 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3270 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3272 * We now know the "local memory node" for each node--
3273 * i.e., the node of the first zone in the generic zonelist.
3274 * Set up numa_mem percpu variable for on-line cpus. During
3275 * boot, only the boot cpu should be on-line; we'll init the
3276 * secondary cpus' numa_mem as they come on-line. During
3277 * node/memory hotplug, we'll fixup all on-line cpus.
3279 if (cpu_online(cpu))
3280 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3288 * Called with zonelists_mutex held always
3289 * unless system_state == SYSTEM_BOOTING.
3291 void __ref build_all_zonelists(void *data)
3293 set_zonelist_order();
3295 if (system_state == SYSTEM_BOOTING) {
3296 __build_all_zonelists(NULL);
3297 mminit_verify_zonelist();
3298 cpuset_init_current_mems_allowed();
3300 /* we have to stop all cpus to guarantee there is no user
3302 #ifdef CONFIG_MEMORY_HOTPLUG
3304 setup_zone_pageset((struct zone *)data);
3306 stop_machine(__build_all_zonelists, NULL, NULL);
3307 /* cpuset refresh routine should be here */
3309 vm_total_pages = nr_free_pagecache_pages();
3311 * Disable grouping by mobility if the number of pages in the
3312 * system is too low to allow the mechanism to work. It would be
3313 * more accurate, but expensive to check per-zone. This check is
3314 * made on memory-hotadd so a system can start with mobility
3315 * disabled and enable it later
3317 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3318 page_group_by_mobility_disabled = 1;
3320 page_group_by_mobility_disabled = 0;
3322 printk("Built %i zonelists in %s order, mobility grouping %s. "
3323 "Total pages: %ld\n",
3325 zonelist_order_name[current_zonelist_order],
3326 page_group_by_mobility_disabled ? "off" : "on",
3329 printk("Policy zone: %s\n", zone_names[policy_zone]);
3334 * Helper functions to size the waitqueue hash table.
3335 * Essentially these want to choose hash table sizes sufficiently
3336 * large so that collisions trying to wait on pages are rare.
3337 * But in fact, the number of active page waitqueues on typical
3338 * systems is ridiculously low, less than 200. So this is even
3339 * conservative, even though it seems large.
3341 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3342 * waitqueues, i.e. the size of the waitq table given the number of pages.
3344 #define PAGES_PER_WAITQUEUE 256
3346 #ifndef CONFIG_MEMORY_HOTPLUG
3347 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3349 unsigned long size = 1;
3351 pages /= PAGES_PER_WAITQUEUE;
3353 while (size < pages)
3357 * Once we have dozens or even hundreds of threads sleeping
3358 * on IO we've got bigger problems than wait queue collision.
3359 * Limit the size of the wait table to a reasonable size.
3361 size = min(size, 4096UL);
3363 return max(size, 4UL);
3367 * A zone's size might be changed by hot-add, so it is not possible to determine
3368 * a suitable size for its wait_table. So we use the maximum size now.
3370 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3372 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3373 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3374 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3376 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3377 * or more by the traditional way. (See above). It equals:
3379 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3380 * ia64(16K page size) : = ( 8G + 4M)byte.
3381 * powerpc (64K page size) : = (32G +16M)byte.
3383 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3390 * This is an integer logarithm so that shifts can be used later
3391 * to extract the more random high bits from the multiplicative
3392 * hash function before the remainder is taken.
3394 static inline unsigned long wait_table_bits(unsigned long size)
3399 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3402 * Check if a pageblock contains reserved pages
3404 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3408 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3409 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3416 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3417 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3418 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3419 * higher will lead to a bigger reserve which will get freed as contiguous
3420 * blocks as reclaim kicks in
3422 static void setup_zone_migrate_reserve(struct zone *zone)
3424 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3426 unsigned long block_migratetype;
3430 * Get the start pfn, end pfn and the number of blocks to reserve
3431 * We have to be careful to be aligned to pageblock_nr_pages to
3432 * make sure that we always check pfn_valid for the first page in
3435 start_pfn = zone->zone_start_pfn;
3436 end_pfn = start_pfn + zone->spanned_pages;
3437 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3438 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3442 * Reserve blocks are generally in place to help high-order atomic
3443 * allocations that are short-lived. A min_free_kbytes value that
3444 * would result in more than 2 reserve blocks for atomic allocations
3445 * is assumed to be in place to help anti-fragmentation for the
3446 * future allocation of hugepages at runtime.
3448 reserve = min(2, reserve);
3450 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3451 if (!pfn_valid(pfn))
3453 page = pfn_to_page(pfn);
3455 /* Watch out for overlapping nodes */
3456 if (page_to_nid(page) != zone_to_nid(zone))
3459 block_migratetype = get_pageblock_migratetype(page);
3461 /* Only test what is necessary when the reserves are not met */
3464 * Blocks with reserved pages will never free, skip
3467 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3468 if (pageblock_is_reserved(pfn, block_end_pfn))
3471 /* If this block is reserved, account for it */
3472 if (block_migratetype == MIGRATE_RESERVE) {
3477 /* Suitable for reserving if this block is movable */
3478 if (block_migratetype == MIGRATE_MOVABLE) {
3479 set_pageblock_migratetype(page,
3481 move_freepages_block(zone, page,
3489 * If the reserve is met and this is a previous reserved block,
3492 if (block_migratetype == MIGRATE_RESERVE) {
3493 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3494 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3500 * Initially all pages are reserved - free ones are freed
3501 * up by free_all_bootmem() once the early boot process is
3502 * done. Non-atomic initialization, single-pass.
3504 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3505 unsigned long start_pfn, enum memmap_context context)
3508 unsigned long end_pfn = start_pfn + size;
3512 if (highest_memmap_pfn < end_pfn - 1)
3513 highest_memmap_pfn = end_pfn - 1;
3515 z = &NODE_DATA(nid)->node_zones[zone];
3516 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3518 * There can be holes in boot-time mem_map[]s
3519 * handed to this function. They do not
3520 * exist on hotplugged memory.
3522 if (context == MEMMAP_EARLY) {
3523 if (!early_pfn_valid(pfn))
3525 if (!early_pfn_in_nid(pfn, nid))
3528 page = pfn_to_page(pfn);
3529 set_page_links(page, zone, nid, pfn);
3530 mminit_verify_page_links(page, zone, nid, pfn);
3531 init_page_count(page);
3532 reset_page_mapcount(page);
3533 SetPageReserved(page);
3535 * Mark the block movable so that blocks are reserved for
3536 * movable at startup. This will force kernel allocations
3537 * to reserve their blocks rather than leaking throughout
3538 * the address space during boot when many long-lived
3539 * kernel allocations are made. Later some blocks near
3540 * the start are marked MIGRATE_RESERVE by
3541 * setup_zone_migrate_reserve()
3543 * bitmap is created for zone's valid pfn range. but memmap
3544 * can be created for invalid pages (for alignment)
3545 * check here not to call set_pageblock_migratetype() against
3548 if ((z->zone_start_pfn <= pfn)
3549 && (pfn < z->zone_start_pfn + z->spanned_pages)
3550 && !(pfn & (pageblock_nr_pages - 1)))
3551 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3553 INIT_LIST_HEAD(&page->lru);
3554 #ifdef WANT_PAGE_VIRTUAL
3555 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3556 if (!is_highmem_idx(zone))
3557 set_page_address(page, __va(pfn << PAGE_SHIFT));
3562 static void __meminit zone_init_free_lists(struct zone *zone)
3565 for_each_migratetype_order(order, t) {
3566 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3567 zone->free_area[order].nr_free = 0;
3571 #ifndef __HAVE_ARCH_MEMMAP_INIT
3572 #define memmap_init(size, nid, zone, start_pfn) \
3573 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3576 static int zone_batchsize(struct zone *zone)
3582 * The per-cpu-pages pools are set to around 1000th of the
3583 * size of the zone. But no more than 1/2 of a meg.
3585 * OK, so we don't know how big the cache is. So guess.
3587 batch = zone->present_pages / 1024;
3588 if (batch * PAGE_SIZE > 512 * 1024)
3589 batch = (512 * 1024) / PAGE_SIZE;
3590 batch /= 4; /* We effectively *= 4 below */
3595 * Clamp the batch to a 2^n - 1 value. Having a power
3596 * of 2 value was found to be more likely to have
3597 * suboptimal cache aliasing properties in some cases.
3599 * For example if 2 tasks are alternately allocating
3600 * batches of pages, one task can end up with a lot
3601 * of pages of one half of the possible page colors
3602 * and the other with pages of the other colors.
3604 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3609 /* The deferral and batching of frees should be suppressed under NOMMU
3612 * The problem is that NOMMU needs to be able to allocate large chunks
3613 * of contiguous memory as there's no hardware page translation to
3614 * assemble apparent contiguous memory from discontiguous pages.
3616 * Queueing large contiguous runs of pages for batching, however,
3617 * causes the pages to actually be freed in smaller chunks. As there
3618 * can be a significant delay between the individual batches being
3619 * recycled, this leads to the once large chunks of space being
3620 * fragmented and becoming unavailable for high-order allocations.
3626 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3628 struct per_cpu_pages *pcp;
3631 memset(p, 0, sizeof(*p));
3635 pcp->high = 6 * batch;
3636 pcp->batch = max(1UL, 1 * batch);
3637 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3638 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3642 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3643 * to the value high for the pageset p.
3646 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3649 struct per_cpu_pages *pcp;
3653 pcp->batch = max(1UL, high/4);
3654 if ((high/4) > (PAGE_SHIFT * 8))
3655 pcp->batch = PAGE_SHIFT * 8;
3658 static void setup_zone_pageset(struct zone *zone)
3662 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3664 for_each_possible_cpu(cpu) {
3665 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3667 setup_pageset(pcp, zone_batchsize(zone));
3669 if (percpu_pagelist_fraction)
3670 setup_pagelist_highmark(pcp,
3671 (zone->present_pages /
3672 percpu_pagelist_fraction));
3677 * Allocate per cpu pagesets and initialize them.
3678 * Before this call only boot pagesets were available.
3680 void __init setup_per_cpu_pageset(void)
3684 for_each_populated_zone(zone)
3685 setup_zone_pageset(zone);
3688 static noinline __init_refok
3689 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3692 struct pglist_data *pgdat = zone->zone_pgdat;
3696 * The per-page waitqueue mechanism uses hashed waitqueues
3699 zone->wait_table_hash_nr_entries =
3700 wait_table_hash_nr_entries(zone_size_pages);
3701 zone->wait_table_bits =
3702 wait_table_bits(zone->wait_table_hash_nr_entries);
3703 alloc_size = zone->wait_table_hash_nr_entries
3704 * sizeof(wait_queue_head_t);
3706 if (!slab_is_available()) {
3707 zone->wait_table = (wait_queue_head_t *)
3708 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3711 * This case means that a zone whose size was 0 gets new memory
3712 * via memory hot-add.
3713 * But it may be the case that a new node was hot-added. In
3714 * this case vmalloc() will not be able to use this new node's
3715 * memory - this wait_table must be initialized to use this new
3716 * node itself as well.
3717 * To use this new node's memory, further consideration will be
3720 zone->wait_table = vmalloc(alloc_size);
3722 if (!zone->wait_table)
3725 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3726 init_waitqueue_head(zone->wait_table + i);
3731 static int __zone_pcp_update(void *data)
3733 struct zone *zone = data;
3735 unsigned long batch = zone_batchsize(zone), flags;
3737 for_each_possible_cpu(cpu) {
3738 struct per_cpu_pageset *pset;
3739 struct per_cpu_pages *pcp;
3741 pset = per_cpu_ptr(zone->pageset, cpu);
3744 local_irq_save(flags);
3745 free_pcppages_bulk(zone, pcp->count, pcp);
3746 setup_pageset(pset, batch);
3747 local_irq_restore(flags);
3752 void zone_pcp_update(struct zone *zone)
3754 stop_machine(__zone_pcp_update, zone, NULL);
3757 static __meminit void zone_pcp_init(struct zone *zone)
3760 * per cpu subsystem is not up at this point. The following code
3761 * relies on the ability of the linker to provide the
3762 * offset of a (static) per cpu variable into the per cpu area.
3764 zone->pageset = &boot_pageset;
3766 if (zone->present_pages)
3767 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3768 zone->name, zone->present_pages,
3769 zone_batchsize(zone));
3772 __meminit int init_currently_empty_zone(struct zone *zone,
3773 unsigned long zone_start_pfn,
3775 enum memmap_context context)
3777 struct pglist_data *pgdat = zone->zone_pgdat;
3779 ret = zone_wait_table_init(zone, size);
3782 pgdat->nr_zones = zone_idx(zone) + 1;
3784 zone->zone_start_pfn = zone_start_pfn;
3786 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3787 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3789 (unsigned long)zone_idx(zone),
3790 zone_start_pfn, (zone_start_pfn + size));
3792 zone_init_free_lists(zone);
3797 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3799 * Basic iterator support. Return the first range of PFNs for a node
3800 * Note: nid == MAX_NUMNODES returns first region regardless of node
3802 static int __meminit first_active_region_index_in_nid(int nid)
3806 for (i = 0; i < nr_nodemap_entries; i++)
3807 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3814 * Basic iterator support. Return the next active range of PFNs for a node
3815 * Note: nid == MAX_NUMNODES returns next region regardless of node
3817 static int __meminit next_active_region_index_in_nid(int index, int nid)
3819 for (index = index + 1; index < nr_nodemap_entries; index++)
3820 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3826 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3828 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3829 * Architectures may implement their own version but if add_active_range()
3830 * was used and there are no special requirements, this is a convenient
3833 int __meminit __early_pfn_to_nid(unsigned long pfn)
3837 for (i = 0; i < nr_nodemap_entries; i++) {
3838 unsigned long start_pfn = early_node_map[i].start_pfn;
3839 unsigned long end_pfn = early_node_map[i].end_pfn;
3841 if (start_pfn <= pfn && pfn < end_pfn)
3842 return early_node_map[i].nid;
3844 /* This is a memory hole */
3847 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3849 int __meminit early_pfn_to_nid(unsigned long pfn)
3853 nid = __early_pfn_to_nid(pfn);
3856 /* just returns 0 */
3860 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3861 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3865 nid = __early_pfn_to_nid(pfn);
3866 if (nid >= 0 && nid != node)
3872 /* Basic iterator support to walk early_node_map[] */
3873 #define for_each_active_range_index_in_nid(i, nid) \
3874 for (i = first_active_region_index_in_nid(nid); i != -1; \
3875 i = next_active_region_index_in_nid(i, nid))
3878 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3879 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3880 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3882 * If an architecture guarantees that all ranges registered with
3883 * add_active_ranges() contain no holes and may be freed, this
3884 * this function may be used instead of calling free_bootmem() manually.
3886 void __init free_bootmem_with_active_regions(int nid,
3887 unsigned long max_low_pfn)
3891 for_each_active_range_index_in_nid(i, nid) {
3892 unsigned long size_pages = 0;
3893 unsigned long end_pfn = early_node_map[i].end_pfn;
3895 if (early_node_map[i].start_pfn >= max_low_pfn)
3898 if (end_pfn > max_low_pfn)
3899 end_pfn = max_low_pfn;
3901 size_pages = end_pfn - early_node_map[i].start_pfn;
3902 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3903 PFN_PHYS(early_node_map[i].start_pfn),
3904 size_pages << PAGE_SHIFT);
3908 #ifdef CONFIG_HAVE_MEMBLOCK
3910 * Basic iterator support. Return the last range of PFNs for a node
3911 * Note: nid == MAX_NUMNODES returns last region regardless of node
3913 static int __meminit last_active_region_index_in_nid(int nid)
3917 for (i = nr_nodemap_entries - 1; i >= 0; i--)
3918 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3925 * Basic iterator support. Return the previous active range of PFNs for a node
3926 * Note: nid == MAX_NUMNODES returns next region regardless of node
3928 static int __meminit previous_active_region_index_in_nid(int index, int nid)
3930 for (index = index - 1; index >= 0; index--)
3931 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3937 #define for_each_active_range_index_in_nid_reverse(i, nid) \
3938 for (i = last_active_region_index_in_nid(nid); i != -1; \
3939 i = previous_active_region_index_in_nid(i, nid))
3941 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3942 u64 goal, u64 limit)
3946 /* Need to go over early_node_map to find out good range for node */
3947 for_each_active_range_index_in_nid_reverse(i, nid) {
3949 u64 ei_start, ei_last;
3950 u64 final_start, final_end;
3952 ei_last = early_node_map[i].end_pfn;
3953 ei_last <<= PAGE_SHIFT;
3954 ei_start = early_node_map[i].start_pfn;
3955 ei_start <<= PAGE_SHIFT;
3957 final_start = max(ei_start, goal);
3958 final_end = min(ei_last, limit);
3960 if (final_start >= final_end)
3963 addr = memblock_find_in_range(final_start, final_end, size, align);
3965 if (addr == MEMBLOCK_ERROR)
3971 return MEMBLOCK_ERROR;
3975 int __init add_from_early_node_map(struct range *range, int az,
3976 int nr_range, int nid)
3981 /* need to go over early_node_map to find out good range for node */
3982 for_each_active_range_index_in_nid(i, nid) {
3983 start = early_node_map[i].start_pfn;
3984 end = early_node_map[i].end_pfn;
3985 nr_range = add_range(range, az, nr_range, start, end);
3990 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3995 for_each_active_range_index_in_nid(i, nid) {
3996 ret = work_fn(early_node_map[i].start_pfn,
3997 early_node_map[i].end_pfn, data);
4003 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4004 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4006 * If an architecture guarantees that all ranges registered with
4007 * add_active_ranges() contain no holes and may be freed, this
4008 * function may be used instead of calling memory_present() manually.
4010 void __init sparse_memory_present_with_active_regions(int nid)
4014 for_each_active_range_index_in_nid(i, nid)
4015 memory_present(early_node_map[i].nid,
4016 early_node_map[i].start_pfn,
4017 early_node_map[i].end_pfn);
4021 * get_pfn_range_for_nid - Return the start and end page frames for a node
4022 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4023 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4024 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4026 * It returns the start and end page frame of a node based on information
4027 * provided by an arch calling add_active_range(). If called for a node
4028 * with no available memory, a warning is printed and the start and end
4031 void __meminit get_pfn_range_for_nid(unsigned int nid,
4032 unsigned long *start_pfn, unsigned long *end_pfn)
4038 for_each_active_range_index_in_nid(i, nid) {
4039 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
4040 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
4043 if (*start_pfn == -1UL)
4048 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4049 * assumption is made that zones within a node are ordered in monotonic
4050 * increasing memory addresses so that the "highest" populated zone is used
4052 static void __init find_usable_zone_for_movable(void)
4055 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4056 if (zone_index == ZONE_MOVABLE)
4059 if (arch_zone_highest_possible_pfn[zone_index] >
4060 arch_zone_lowest_possible_pfn[zone_index])
4064 VM_BUG_ON(zone_index == -1);
4065 movable_zone = zone_index;
4069 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4070 * because it is sized independent of architecture. Unlike the other zones,
4071 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4072 * in each node depending on the size of each node and how evenly kernelcore
4073 * is distributed. This helper function adjusts the zone ranges
4074 * provided by the architecture for a given node by using the end of the
4075 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4076 * zones within a node are in order of monotonic increases memory addresses
4078 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4079 unsigned long zone_type,
4080 unsigned long node_start_pfn,
4081 unsigned long node_end_pfn,
4082 unsigned long *zone_start_pfn,
4083 unsigned long *zone_end_pfn)
4085 /* Only adjust if ZONE_MOVABLE is on this node */
4086 if (zone_movable_pfn[nid]) {
4087 /* Size ZONE_MOVABLE */
4088 if (zone_type == ZONE_MOVABLE) {
4089 *zone_start_pfn = zone_movable_pfn[nid];
4090 *zone_end_pfn = min(node_end_pfn,
4091 arch_zone_highest_possible_pfn[movable_zone]);
4093 /* Adjust for ZONE_MOVABLE starting within this range */
4094 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4095 *zone_end_pfn > zone_movable_pfn[nid]) {
4096 *zone_end_pfn = zone_movable_pfn[nid];
4098 /* Check if this whole range is within ZONE_MOVABLE */
4099 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4100 *zone_start_pfn = *zone_end_pfn;
4105 * Return the number of pages a zone spans in a node, including holes
4106 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4108 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4109 unsigned long zone_type,
4110 unsigned long *ignored)
4112 unsigned long node_start_pfn, node_end_pfn;
4113 unsigned long zone_start_pfn, zone_end_pfn;
4115 /* Get the start and end of the node and zone */
4116 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4117 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4118 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4119 adjust_zone_range_for_zone_movable(nid, zone_type,
4120 node_start_pfn, node_end_pfn,
4121 &zone_start_pfn, &zone_end_pfn);
4123 /* Check that this node has pages within the zone's required range */
4124 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4127 /* Move the zone boundaries inside the node if necessary */
4128 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4129 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4131 /* Return the spanned pages */
4132 return zone_end_pfn - zone_start_pfn;
4136 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4137 * then all holes in the requested range will be accounted for.
4139 unsigned long __meminit __absent_pages_in_range(int nid,
4140 unsigned long range_start_pfn,
4141 unsigned long range_end_pfn)
4144 unsigned long prev_end_pfn = 0, hole_pages = 0;
4145 unsigned long start_pfn;
4147 /* Find the end_pfn of the first active range of pfns in the node */
4148 i = first_active_region_index_in_nid(nid);
4152 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4154 /* Account for ranges before physical memory on this node */
4155 if (early_node_map[i].start_pfn > range_start_pfn)
4156 hole_pages = prev_end_pfn - range_start_pfn;
4158 /* Find all holes for the zone within the node */
4159 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4161 /* No need to continue if prev_end_pfn is outside the zone */
4162 if (prev_end_pfn >= range_end_pfn)
4165 /* Make sure the end of the zone is not within the hole */
4166 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4167 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4169 /* Update the hole size cound and move on */
4170 if (start_pfn > range_start_pfn) {
4171 BUG_ON(prev_end_pfn > start_pfn);
4172 hole_pages += start_pfn - prev_end_pfn;
4174 prev_end_pfn = early_node_map[i].end_pfn;
4177 /* Account for ranges past physical memory on this node */
4178 if (range_end_pfn > prev_end_pfn)
4179 hole_pages += range_end_pfn -
4180 max(range_start_pfn, prev_end_pfn);
4186 * absent_pages_in_range - Return number of page frames in holes within a range
4187 * @start_pfn: The start PFN to start searching for holes
4188 * @end_pfn: The end PFN to stop searching for holes
4190 * It returns the number of pages frames in memory holes within a range.
4192 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4193 unsigned long end_pfn)
4195 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4198 /* Return the number of page frames in holes in a zone on a node */
4199 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4200 unsigned long zone_type,
4201 unsigned long *ignored)
4203 unsigned long node_start_pfn, node_end_pfn;
4204 unsigned long zone_start_pfn, zone_end_pfn;
4206 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4207 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4209 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4212 adjust_zone_range_for_zone_movable(nid, zone_type,
4213 node_start_pfn, node_end_pfn,
4214 &zone_start_pfn, &zone_end_pfn);
4215 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4219 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4220 unsigned long zone_type,
4221 unsigned long *zones_size)
4223 return zones_size[zone_type];
4226 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4227 unsigned long zone_type,
4228 unsigned long *zholes_size)
4233 return zholes_size[zone_type];
4238 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4239 unsigned long *zones_size, unsigned long *zholes_size)
4241 unsigned long realtotalpages, totalpages = 0;
4244 for (i = 0; i < MAX_NR_ZONES; i++)
4245 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4247 pgdat->node_spanned_pages = totalpages;
4249 realtotalpages = totalpages;
4250 for (i = 0; i < MAX_NR_ZONES; i++)
4252 zone_absent_pages_in_node(pgdat->node_id, i,
4254 pgdat->node_present_pages = realtotalpages;
4255 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4259 #ifndef CONFIG_SPARSEMEM
4261 * Calculate the size of the zone->blockflags rounded to an unsigned long
4262 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4263 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4264 * round what is now in bits to nearest long in bits, then return it in
4267 static unsigned long __init usemap_size(unsigned long zonesize)
4269 unsigned long usemapsize;
4271 usemapsize = roundup(zonesize, pageblock_nr_pages);
4272 usemapsize = usemapsize >> pageblock_order;
4273 usemapsize *= NR_PAGEBLOCK_BITS;
4274 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4276 return usemapsize / 8;
4279 static void __init setup_usemap(struct pglist_data *pgdat,
4280 struct zone *zone, unsigned long zonesize)
4282 unsigned long usemapsize = usemap_size(zonesize);
4283 zone->pageblock_flags = NULL;
4285 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4289 static inline void setup_usemap(struct pglist_data *pgdat,
4290 struct zone *zone, unsigned long zonesize) {}
4291 #endif /* CONFIG_SPARSEMEM */
4293 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4295 /* Return a sensible default order for the pageblock size. */
4296 static inline int pageblock_default_order(void)
4298 if (HPAGE_SHIFT > PAGE_SHIFT)
4299 return HUGETLB_PAGE_ORDER;
4304 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4305 static inline void __init set_pageblock_order(unsigned int order)
4307 /* Check that pageblock_nr_pages has not already been setup */
4308 if (pageblock_order)
4312 * Assume the largest contiguous order of interest is a huge page.
4313 * This value may be variable depending on boot parameters on IA64
4315 pageblock_order = order;
4317 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4320 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4321 * and pageblock_default_order() are unused as pageblock_order is set
4322 * at compile-time. See include/linux/pageblock-flags.h for the values of
4323 * pageblock_order based on the kernel config
4325 static inline int pageblock_default_order(unsigned int order)
4329 #define set_pageblock_order(x) do {} while (0)
4331 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4334 * Set up the zone data structures:
4335 * - mark all pages reserved
4336 * - mark all memory queues empty
4337 * - clear the memory bitmaps
4339 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4340 unsigned long *zones_size, unsigned long *zholes_size)
4343 int nid = pgdat->node_id;
4344 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4347 pgdat_resize_init(pgdat);
4348 pgdat->nr_zones = 0;
4349 init_waitqueue_head(&pgdat->kswapd_wait);
4350 pgdat->kswapd_max_order = 0;
4351 pgdat_page_cgroup_init(pgdat);
4353 for (j = 0; j < MAX_NR_ZONES; j++) {
4354 struct zone *zone = pgdat->node_zones + j;
4355 unsigned long size, realsize, memmap_pages;
4358 size = zone_spanned_pages_in_node(nid, j, zones_size);
4359 realsize = size - zone_absent_pages_in_node(nid, j,
4363 * Adjust realsize so that it accounts for how much memory
4364 * is used by this zone for memmap. This affects the watermark
4365 * and per-cpu initialisations
4368 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4369 if (realsize >= memmap_pages) {
4370 realsize -= memmap_pages;
4373 " %s zone: %lu pages used for memmap\n",
4374 zone_names[j], memmap_pages);
4377 " %s zone: %lu pages exceeds realsize %lu\n",
4378 zone_names[j], memmap_pages, realsize);
4380 /* Account for reserved pages */
4381 if (j == 0 && realsize > dma_reserve) {
4382 realsize -= dma_reserve;
4383 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4384 zone_names[0], dma_reserve);
4387 if (!is_highmem_idx(j))
4388 nr_kernel_pages += realsize;
4389 nr_all_pages += realsize;
4391 zone->spanned_pages = size;
4392 zone->present_pages = realsize;
4395 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4397 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4399 zone->name = zone_names[j];
4400 spin_lock_init(&zone->lock);
4401 spin_lock_init(&zone->lru_lock);
4402 zone_seqlock_init(zone);
4403 zone->zone_pgdat = pgdat;
4405 zone_pcp_init(zone);
4407 INIT_LIST_HEAD(&zone->lru[l].list);
4408 zone->reclaim_stat.recent_rotated[0] = 0;
4409 zone->reclaim_stat.recent_rotated[1] = 0;
4410 zone->reclaim_stat.recent_scanned[0] = 0;
4411 zone->reclaim_stat.recent_scanned[1] = 0;
4412 zap_zone_vm_stats(zone);
4417 set_pageblock_order(pageblock_default_order());
4418 setup_usemap(pgdat, zone, size);
4419 ret = init_currently_empty_zone(zone, zone_start_pfn,
4420 size, MEMMAP_EARLY);
4422 memmap_init(size, nid, j, zone_start_pfn);
4423 zone_start_pfn += size;
4427 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4429 /* Skip empty nodes */
4430 if (!pgdat->node_spanned_pages)
4433 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4434 /* ia64 gets its own node_mem_map, before this, without bootmem */
4435 if (!pgdat->node_mem_map) {
4436 unsigned long size, start, end;
4440 * The zone's endpoints aren't required to be MAX_ORDER
4441 * aligned but the node_mem_map endpoints must be in order
4442 * for the buddy allocator to function correctly.
4444 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4445 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4446 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4447 size = (end - start) * sizeof(struct page);
4448 map = alloc_remap(pgdat->node_id, size);
4450 map = alloc_bootmem_node_nopanic(pgdat, size);
4451 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4453 #ifndef CONFIG_NEED_MULTIPLE_NODES
4455 * With no DISCONTIG, the global mem_map is just set as node 0's
4457 if (pgdat == NODE_DATA(0)) {
4458 mem_map = NODE_DATA(0)->node_mem_map;
4459 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4460 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4461 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4462 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4465 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4468 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4469 unsigned long node_start_pfn, unsigned long *zholes_size)
4471 pg_data_t *pgdat = NODE_DATA(nid);
4473 pgdat->node_id = nid;
4474 pgdat->node_start_pfn = node_start_pfn;
4475 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4477 alloc_node_mem_map(pgdat);
4478 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4479 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4480 nid, (unsigned long)pgdat,
4481 (unsigned long)pgdat->node_mem_map);
4484 free_area_init_core(pgdat, zones_size, zholes_size);
4487 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4489 #if MAX_NUMNODES > 1
4491 * Figure out the number of possible node ids.
4493 static void __init setup_nr_node_ids(void)
4496 unsigned int highest = 0;
4498 for_each_node_mask(node, node_possible_map)
4500 nr_node_ids = highest + 1;
4503 static inline void setup_nr_node_ids(void)
4509 * add_active_range - Register a range of PFNs backed by physical memory
4510 * @nid: The node ID the range resides on
4511 * @start_pfn: The start PFN of the available physical memory
4512 * @end_pfn: The end PFN of the available physical memory
4514 * These ranges are stored in an early_node_map[] and later used by
4515 * free_area_init_nodes() to calculate zone sizes and holes. If the
4516 * range spans a memory hole, it is up to the architecture to ensure
4517 * the memory is not freed by the bootmem allocator. If possible
4518 * the range being registered will be merged with existing ranges.
4520 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4521 unsigned long end_pfn)
4525 mminit_dprintk(MMINIT_TRACE, "memory_register",
4526 "Entering add_active_range(%d, %#lx, %#lx) "
4527 "%d entries of %d used\n",
4528 nid, start_pfn, end_pfn,
4529 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4531 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4533 /* Merge with existing active regions if possible */
4534 for (i = 0; i < nr_nodemap_entries; i++) {
4535 if (early_node_map[i].nid != nid)
4538 /* Skip if an existing region covers this new one */
4539 if (start_pfn >= early_node_map[i].start_pfn &&
4540 end_pfn <= early_node_map[i].end_pfn)
4543 /* Merge forward if suitable */
4544 if (start_pfn <= early_node_map[i].end_pfn &&
4545 end_pfn > early_node_map[i].end_pfn) {
4546 early_node_map[i].end_pfn = end_pfn;
4550 /* Merge backward if suitable */
4551 if (start_pfn < early_node_map[i].start_pfn &&
4552 end_pfn >= early_node_map[i].start_pfn) {
4553 early_node_map[i].start_pfn = start_pfn;
4558 /* Check that early_node_map is large enough */
4559 if (i >= MAX_ACTIVE_REGIONS) {
4560 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4561 MAX_ACTIVE_REGIONS);
4565 early_node_map[i].nid = nid;
4566 early_node_map[i].start_pfn = start_pfn;
4567 early_node_map[i].end_pfn = end_pfn;
4568 nr_nodemap_entries = i + 1;
4572 * remove_active_range - Shrink an existing registered range of PFNs
4573 * @nid: The node id the range is on that should be shrunk
4574 * @start_pfn: The new PFN of the range
4575 * @end_pfn: The new PFN of the range
4577 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4578 * The map is kept near the end physical page range that has already been
4579 * registered. This function allows an arch to shrink an existing registered
4582 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4583 unsigned long end_pfn)
4588 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4589 nid, start_pfn, end_pfn);
4591 /* Find the old active region end and shrink */
4592 for_each_active_range_index_in_nid(i, nid) {
4593 if (early_node_map[i].start_pfn >= start_pfn &&
4594 early_node_map[i].end_pfn <= end_pfn) {
4596 early_node_map[i].start_pfn = 0;
4597 early_node_map[i].end_pfn = 0;
4601 if (early_node_map[i].start_pfn < start_pfn &&
4602 early_node_map[i].end_pfn > start_pfn) {
4603 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4604 early_node_map[i].end_pfn = start_pfn;
4605 if (temp_end_pfn > end_pfn)
4606 add_active_range(nid, end_pfn, temp_end_pfn);
4609 if (early_node_map[i].start_pfn >= start_pfn &&
4610 early_node_map[i].end_pfn > end_pfn &&
4611 early_node_map[i].start_pfn < end_pfn) {
4612 early_node_map[i].start_pfn = end_pfn;
4620 /* remove the blank ones */
4621 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4622 if (early_node_map[i].nid != nid)
4624 if (early_node_map[i].end_pfn)
4626 /* we found it, get rid of it */
4627 for (j = i; j < nr_nodemap_entries - 1; j++)
4628 memcpy(&early_node_map[j], &early_node_map[j+1],
4629 sizeof(early_node_map[j]));
4630 j = nr_nodemap_entries - 1;
4631 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4632 nr_nodemap_entries--;
4637 * remove_all_active_ranges - Remove all currently registered regions
4639 * During discovery, it may be found that a table like SRAT is invalid
4640 * and an alternative discovery method must be used. This function removes
4641 * all currently registered regions.
4643 void __init remove_all_active_ranges(void)
4645 memset(early_node_map, 0, sizeof(early_node_map));
4646 nr_nodemap_entries = 0;
4649 /* Compare two active node_active_regions */
4650 static int __init cmp_node_active_region(const void *a, const void *b)
4652 struct node_active_region *arange = (struct node_active_region *)a;
4653 struct node_active_region *brange = (struct node_active_region *)b;
4655 /* Done this way to avoid overflows */
4656 if (arange->start_pfn > brange->start_pfn)
4658 if (arange->start_pfn < brange->start_pfn)
4664 /* sort the node_map by start_pfn */
4665 void __init sort_node_map(void)
4667 sort(early_node_map, (size_t)nr_nodemap_entries,
4668 sizeof(struct node_active_region),
4669 cmp_node_active_region, NULL);
4672 /* Find the lowest pfn for a node */
4673 static unsigned long __init find_min_pfn_for_node(int nid)
4676 unsigned long min_pfn = ULONG_MAX;
4678 /* Assuming a sorted map, the first range found has the starting pfn */
4679 for_each_active_range_index_in_nid(i, nid)
4680 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4682 if (min_pfn == ULONG_MAX) {
4684 "Could not find start_pfn for node %d\n", nid);
4692 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4694 * It returns the minimum PFN based on information provided via
4695 * add_active_range().
4697 unsigned long __init find_min_pfn_with_active_regions(void)
4699 return find_min_pfn_for_node(MAX_NUMNODES);
4703 * early_calculate_totalpages()
4704 * Sum pages in active regions for movable zone.
4705 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4707 static unsigned long __init early_calculate_totalpages(void)
4710 unsigned long totalpages = 0;
4712 for (i = 0; i < nr_nodemap_entries; i++) {
4713 unsigned long pages = early_node_map[i].end_pfn -
4714 early_node_map[i].start_pfn;
4715 totalpages += pages;
4717 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4723 * Find the PFN the Movable zone begins in each node. Kernel memory
4724 * is spread evenly between nodes as long as the nodes have enough
4725 * memory. When they don't, some nodes will have more kernelcore than
4728 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4731 unsigned long usable_startpfn;
4732 unsigned long kernelcore_node, kernelcore_remaining;
4733 /* save the state before borrow the nodemask */
4734 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4735 unsigned long totalpages = early_calculate_totalpages();
4736 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4739 * If movablecore was specified, calculate what size of
4740 * kernelcore that corresponds so that memory usable for
4741 * any allocation type is evenly spread. If both kernelcore
4742 * and movablecore are specified, then the value of kernelcore
4743 * will be used for required_kernelcore if it's greater than
4744 * what movablecore would have allowed.
4746 if (required_movablecore) {
4747 unsigned long corepages;
4750 * Round-up so that ZONE_MOVABLE is at least as large as what
4751 * was requested by the user
4753 required_movablecore =
4754 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4755 corepages = totalpages - required_movablecore;
4757 required_kernelcore = max(required_kernelcore, corepages);
4760 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4761 if (!required_kernelcore)
4764 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4765 find_usable_zone_for_movable();
4766 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4769 /* Spread kernelcore memory as evenly as possible throughout nodes */
4770 kernelcore_node = required_kernelcore / usable_nodes;
4771 for_each_node_state(nid, N_HIGH_MEMORY) {
4773 * Recalculate kernelcore_node if the division per node
4774 * now exceeds what is necessary to satisfy the requested
4775 * amount of memory for the kernel
4777 if (required_kernelcore < kernelcore_node)
4778 kernelcore_node = required_kernelcore / usable_nodes;
4781 * As the map is walked, we track how much memory is usable
4782 * by the kernel using kernelcore_remaining. When it is
4783 * 0, the rest of the node is usable by ZONE_MOVABLE
4785 kernelcore_remaining = kernelcore_node;
4787 /* Go through each range of PFNs within this node */
4788 for_each_active_range_index_in_nid(i, nid) {
4789 unsigned long start_pfn, end_pfn;
4790 unsigned long size_pages;
4792 start_pfn = max(early_node_map[i].start_pfn,
4793 zone_movable_pfn[nid]);
4794 end_pfn = early_node_map[i].end_pfn;
4795 if (start_pfn >= end_pfn)
4798 /* Account for what is only usable for kernelcore */
4799 if (start_pfn < usable_startpfn) {
4800 unsigned long kernel_pages;
4801 kernel_pages = min(end_pfn, usable_startpfn)
4804 kernelcore_remaining -= min(kernel_pages,
4805 kernelcore_remaining);
4806 required_kernelcore -= min(kernel_pages,
4807 required_kernelcore);
4809 /* Continue if range is now fully accounted */
4810 if (end_pfn <= usable_startpfn) {
4813 * Push zone_movable_pfn to the end so
4814 * that if we have to rebalance
4815 * kernelcore across nodes, we will
4816 * not double account here
4818 zone_movable_pfn[nid] = end_pfn;
4821 start_pfn = usable_startpfn;
4825 * The usable PFN range for ZONE_MOVABLE is from
4826 * start_pfn->end_pfn. Calculate size_pages as the
4827 * number of pages used as kernelcore
4829 size_pages = end_pfn - start_pfn;
4830 if (size_pages > kernelcore_remaining)
4831 size_pages = kernelcore_remaining;
4832 zone_movable_pfn[nid] = start_pfn + size_pages;
4835 * Some kernelcore has been met, update counts and
4836 * break if the kernelcore for this node has been
4839 required_kernelcore -= min(required_kernelcore,
4841 kernelcore_remaining -= size_pages;
4842 if (!kernelcore_remaining)
4848 * If there is still required_kernelcore, we do another pass with one
4849 * less node in the count. This will push zone_movable_pfn[nid] further
4850 * along on the nodes that still have memory until kernelcore is
4854 if (usable_nodes && required_kernelcore > usable_nodes)
4857 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4858 for (nid = 0; nid < MAX_NUMNODES; nid++)
4859 zone_movable_pfn[nid] =
4860 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4863 /* restore the node_state */
4864 node_states[N_HIGH_MEMORY] = saved_node_state;
4867 /* Any regular memory on that node ? */
4868 static void check_for_regular_memory(pg_data_t *pgdat)
4870 #ifdef CONFIG_HIGHMEM
4871 enum zone_type zone_type;
4873 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4874 struct zone *zone = &pgdat->node_zones[zone_type];
4875 if (zone->present_pages)
4876 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4882 * free_area_init_nodes - Initialise all pg_data_t and zone data
4883 * @max_zone_pfn: an array of max PFNs for each zone
4885 * This will call free_area_init_node() for each active node in the system.
4886 * Using the page ranges provided by add_active_range(), the size of each
4887 * zone in each node and their holes is calculated. If the maximum PFN
4888 * between two adjacent zones match, it is assumed that the zone is empty.
4889 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4890 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4891 * starts where the previous one ended. For example, ZONE_DMA32 starts
4892 * at arch_max_dma_pfn.
4894 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4899 /* Sort early_node_map as initialisation assumes it is sorted */
4902 /* Record where the zone boundaries are */
4903 memset(arch_zone_lowest_possible_pfn, 0,
4904 sizeof(arch_zone_lowest_possible_pfn));
4905 memset(arch_zone_highest_possible_pfn, 0,
4906 sizeof(arch_zone_highest_possible_pfn));
4907 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4908 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4909 for (i = 1; i < MAX_NR_ZONES; i++) {
4910 if (i == ZONE_MOVABLE)
4912 arch_zone_lowest_possible_pfn[i] =
4913 arch_zone_highest_possible_pfn[i-1];
4914 arch_zone_highest_possible_pfn[i] =
4915 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4917 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4918 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4920 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4921 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4922 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4924 /* Print out the zone ranges */
4925 printk("Zone PFN ranges:\n");
4926 for (i = 0; i < MAX_NR_ZONES; i++) {
4927 if (i == ZONE_MOVABLE)
4929 printk(" %-8s ", zone_names[i]);
4930 if (arch_zone_lowest_possible_pfn[i] ==
4931 arch_zone_highest_possible_pfn[i])
4934 printk("%0#10lx -> %0#10lx\n",
4935 arch_zone_lowest_possible_pfn[i],
4936 arch_zone_highest_possible_pfn[i]);
4939 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4940 printk("Movable zone start PFN for each node\n");
4941 for (i = 0; i < MAX_NUMNODES; i++) {
4942 if (zone_movable_pfn[i])
4943 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4946 /* Print out the early_node_map[] */
4947 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4948 for (i = 0; i < nr_nodemap_entries; i++)
4949 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4950 early_node_map[i].start_pfn,
4951 early_node_map[i].end_pfn);
4953 /* Initialise every node */
4954 mminit_verify_pageflags_layout();
4955 setup_nr_node_ids();
4956 for_each_online_node(nid) {
4957 pg_data_t *pgdat = NODE_DATA(nid);
4958 free_area_init_node(nid, NULL,
4959 find_min_pfn_for_node(nid), NULL);
4961 /* Any memory on that node */
4962 if (pgdat->node_present_pages)
4963 node_set_state(nid, N_HIGH_MEMORY);
4964 check_for_regular_memory(pgdat);
4968 static int __init cmdline_parse_core(char *p, unsigned long *core)
4970 unsigned long long coremem;
4974 coremem = memparse(p, &p);
4975 *core = coremem >> PAGE_SHIFT;
4977 /* Paranoid check that UL is enough for the coremem value */
4978 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4984 * kernelcore=size sets the amount of memory for use for allocations that
4985 * cannot be reclaimed or migrated.
4987 static int __init cmdline_parse_kernelcore(char *p)
4989 return cmdline_parse_core(p, &required_kernelcore);
4993 * movablecore=size sets the amount of memory for use for allocations that
4994 * can be reclaimed or migrated.
4996 static int __init cmdline_parse_movablecore(char *p)
4998 return cmdline_parse_core(p, &required_movablecore);
5001 early_param("kernelcore", cmdline_parse_kernelcore);
5002 early_param("movablecore", cmdline_parse_movablecore);
5004 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
5007 * set_dma_reserve - set the specified number of pages reserved in the first zone
5008 * @new_dma_reserve: The number of pages to mark reserved
5010 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5011 * In the DMA zone, a significant percentage may be consumed by kernel image
5012 * and other unfreeable allocations which can skew the watermarks badly. This
5013 * function may optionally be used to account for unfreeable pages in the
5014 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5015 * smaller per-cpu batchsize.
5017 void __init set_dma_reserve(unsigned long new_dma_reserve)
5019 dma_reserve = new_dma_reserve;
5022 void __init free_area_init(unsigned long *zones_size)
5024 free_area_init_node(0, zones_size,
5025 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5028 static int page_alloc_cpu_notify(struct notifier_block *self,
5029 unsigned long action, void *hcpu)
5031 int cpu = (unsigned long)hcpu;
5033 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5037 * Spill the event counters of the dead processor
5038 * into the current processors event counters.
5039 * This artificially elevates the count of the current
5042 vm_events_fold_cpu(cpu);
5045 * Zero the differential counters of the dead processor
5046 * so that the vm statistics are consistent.
5048 * This is only okay since the processor is dead and cannot
5049 * race with what we are doing.
5051 refresh_cpu_vm_stats(cpu);
5056 void __init page_alloc_init(void)
5058 hotcpu_notifier(page_alloc_cpu_notify, 0);
5062 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5063 * or min_free_kbytes changes.
5065 static void calculate_totalreserve_pages(void)
5067 struct pglist_data *pgdat;
5068 unsigned long reserve_pages = 0;
5069 enum zone_type i, j;
5071 for_each_online_pgdat(pgdat) {
5072 for (i = 0; i < MAX_NR_ZONES; i++) {
5073 struct zone *zone = pgdat->node_zones + i;
5074 unsigned long max = 0;
5076 /* Find valid and maximum lowmem_reserve in the zone */
5077 for (j = i; j < MAX_NR_ZONES; j++) {
5078 if (zone->lowmem_reserve[j] > max)
5079 max = zone->lowmem_reserve[j];
5082 /* we treat the high watermark as reserved pages. */
5083 max += high_wmark_pages(zone);
5085 if (max > zone->present_pages)
5086 max = zone->present_pages;
5087 reserve_pages += max;
5090 totalreserve_pages = reserve_pages;
5094 * setup_per_zone_lowmem_reserve - called whenever
5095 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5096 * has a correct pages reserved value, so an adequate number of
5097 * pages are left in the zone after a successful __alloc_pages().
5099 static void setup_per_zone_lowmem_reserve(void)
5101 struct pglist_data *pgdat;
5102 enum zone_type j, idx;
5104 for_each_online_pgdat(pgdat) {
5105 for (j = 0; j < MAX_NR_ZONES; j++) {
5106 struct zone *zone = pgdat->node_zones + j;
5107 unsigned long present_pages = zone->present_pages;
5109 zone->lowmem_reserve[j] = 0;
5113 struct zone *lower_zone;
5117 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5118 sysctl_lowmem_reserve_ratio[idx] = 1;
5120 lower_zone = pgdat->node_zones + idx;
5121 lower_zone->lowmem_reserve[j] = present_pages /
5122 sysctl_lowmem_reserve_ratio[idx];
5123 present_pages += lower_zone->present_pages;
5128 /* update totalreserve_pages */
5129 calculate_totalreserve_pages();
5133 * setup_per_zone_wmarks - called when min_free_kbytes changes
5134 * or when memory is hot-{added|removed}
5136 * Ensures that the watermark[min,low,high] values for each zone are set
5137 * correctly with respect to min_free_kbytes.
5139 void setup_per_zone_wmarks(void)
5141 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5142 unsigned long lowmem_pages = 0;
5144 unsigned long flags;
5146 /* Calculate total number of !ZONE_HIGHMEM pages */
5147 for_each_zone(zone) {
5148 if (!is_highmem(zone))
5149 lowmem_pages += zone->present_pages;
5152 for_each_zone(zone) {
5155 spin_lock_irqsave(&zone->lock, flags);
5156 tmp = (u64)pages_min * zone->present_pages;
5157 do_div(tmp, lowmem_pages);
5158 if (is_highmem(zone)) {
5160 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5161 * need highmem pages, so cap pages_min to a small
5164 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5165 * deltas controls asynch page reclaim, and so should
5166 * not be capped for highmem.
5170 min_pages = zone->present_pages / 1024;
5171 if (min_pages < SWAP_CLUSTER_MAX)
5172 min_pages = SWAP_CLUSTER_MAX;
5173 if (min_pages > 128)
5175 zone->watermark[WMARK_MIN] = min_pages;
5178 * If it's a lowmem zone, reserve a number of pages
5179 * proportionate to the zone's size.
5181 zone->watermark[WMARK_MIN] = tmp;
5184 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5185 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5186 setup_zone_migrate_reserve(zone);
5187 spin_unlock_irqrestore(&zone->lock, flags);
5190 /* update totalreserve_pages */
5191 calculate_totalreserve_pages();
5195 * The inactive anon list should be small enough that the VM never has to
5196 * do too much work, but large enough that each inactive page has a chance
5197 * to be referenced again before it is swapped out.
5199 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5200 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5201 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5202 * the anonymous pages are kept on the inactive list.
5205 * memory ratio inactive anon
5206 * -------------------------------------
5215 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5217 unsigned int gb, ratio;
5219 /* Zone size in gigabytes */
5220 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5222 ratio = int_sqrt(10 * gb);
5226 zone->inactive_ratio = ratio;
5229 static void __meminit setup_per_zone_inactive_ratio(void)
5234 calculate_zone_inactive_ratio(zone);
5238 * Initialise min_free_kbytes.
5240 * For small machines we want it small (128k min). For large machines
5241 * we want it large (64MB max). But it is not linear, because network
5242 * bandwidth does not increase linearly with machine size. We use
5244 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5245 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5261 int __meminit init_per_zone_wmark_min(void)
5263 unsigned long lowmem_kbytes;
5265 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5267 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5268 if (min_free_kbytes < 128)
5269 min_free_kbytes = 128;
5270 if (min_free_kbytes > 65536)
5271 min_free_kbytes = 65536;
5272 setup_per_zone_wmarks();
5273 refresh_zone_stat_thresholds();
5274 setup_per_zone_lowmem_reserve();
5275 setup_per_zone_inactive_ratio();
5278 module_init(init_per_zone_wmark_min)
5281 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5282 * that we can call two helper functions whenever min_free_kbytes
5285 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5286 void __user *buffer, size_t *length, loff_t *ppos)
5288 proc_dointvec(table, write, buffer, length, ppos);
5290 setup_per_zone_wmarks();
5295 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5296 void __user *buffer, size_t *length, loff_t *ppos)
5301 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5306 zone->min_unmapped_pages = (zone->present_pages *
5307 sysctl_min_unmapped_ratio) / 100;
5311 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5312 void __user *buffer, size_t *length, loff_t *ppos)
5317 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5322 zone->min_slab_pages = (zone->present_pages *
5323 sysctl_min_slab_ratio) / 100;
5329 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5330 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5331 * whenever sysctl_lowmem_reserve_ratio changes.
5333 * The reserve ratio obviously has absolutely no relation with the
5334 * minimum watermarks. The lowmem reserve ratio can only make sense
5335 * if in function of the boot time zone sizes.
5337 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5338 void __user *buffer, size_t *length, loff_t *ppos)
5340 proc_dointvec_minmax(table, write, buffer, length, ppos);
5341 setup_per_zone_lowmem_reserve();
5346 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5347 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5348 * can have before it gets flushed back to buddy allocator.
5351 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5352 void __user *buffer, size_t *length, loff_t *ppos)
5358 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5359 if (!write || (ret == -EINVAL))
5361 for_each_populated_zone(zone) {
5362 for_each_possible_cpu(cpu) {
5364 high = zone->present_pages / percpu_pagelist_fraction;
5365 setup_pagelist_highmark(
5366 per_cpu_ptr(zone->pageset, cpu), high);
5372 int hashdist = HASHDIST_DEFAULT;
5375 static int __init set_hashdist(char *str)
5379 hashdist = simple_strtoul(str, &str, 0);
5382 __setup("hashdist=", set_hashdist);
5386 * allocate a large system hash table from bootmem
5387 * - it is assumed that the hash table must contain an exact power-of-2
5388 * quantity of entries
5389 * - limit is the number of hash buckets, not the total allocation size
5391 void *__init alloc_large_system_hash(const char *tablename,
5392 unsigned long bucketsize,
5393 unsigned long numentries,
5396 unsigned int *_hash_shift,
5397 unsigned int *_hash_mask,
5398 unsigned long limit)
5400 unsigned long long max = limit;
5401 unsigned long log2qty, size;
5404 /* allow the kernel cmdline to have a say */
5406 /* round applicable memory size up to nearest megabyte */
5407 numentries = nr_kernel_pages;
5408 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5409 numentries >>= 20 - PAGE_SHIFT;
5410 numentries <<= 20 - PAGE_SHIFT;
5412 /* limit to 1 bucket per 2^scale bytes of low memory */
5413 if (scale > PAGE_SHIFT)
5414 numentries >>= (scale - PAGE_SHIFT);
5416 numentries <<= (PAGE_SHIFT - scale);
5418 /* Make sure we've got at least a 0-order allocation.. */
5419 if (unlikely(flags & HASH_SMALL)) {
5420 /* Makes no sense without HASH_EARLY */
5421 WARN_ON(!(flags & HASH_EARLY));
5422 if (!(numentries >> *_hash_shift)) {
5423 numentries = 1UL << *_hash_shift;
5424 BUG_ON(!numentries);
5426 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5427 numentries = PAGE_SIZE / bucketsize;
5429 numentries = roundup_pow_of_two(numentries);
5431 /* limit allocation size to 1/16 total memory by default */
5433 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5434 do_div(max, bucketsize);
5437 if (numentries > max)
5440 log2qty = ilog2(numentries);
5443 size = bucketsize << log2qty;
5444 if (flags & HASH_EARLY)
5445 table = alloc_bootmem_nopanic(size);
5447 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5450 * If bucketsize is not a power-of-two, we may free
5451 * some pages at the end of hash table which
5452 * alloc_pages_exact() automatically does
5454 if (get_order(size) < MAX_ORDER) {
5455 table = alloc_pages_exact(size, GFP_ATOMIC);
5456 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5459 } while (!table && size > PAGE_SIZE && --log2qty);
5462 panic("Failed to allocate %s hash table\n", tablename);
5464 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5467 ilog2(size) - PAGE_SHIFT,
5471 *_hash_shift = log2qty;
5473 *_hash_mask = (1 << log2qty) - 1;
5478 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5479 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5482 #ifdef CONFIG_SPARSEMEM
5483 return __pfn_to_section(pfn)->pageblock_flags;
5485 return zone->pageblock_flags;
5486 #endif /* CONFIG_SPARSEMEM */
5489 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5491 #ifdef CONFIG_SPARSEMEM
5492 pfn &= (PAGES_PER_SECTION-1);
5493 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5495 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5496 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5497 #endif /* CONFIG_SPARSEMEM */
5501 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5502 * @page: The page within the block of interest
5503 * @start_bitidx: The first bit of interest to retrieve
5504 * @end_bitidx: The last bit of interest
5505 * returns pageblock_bits flags
5507 unsigned long get_pageblock_flags_group(struct page *page,
5508 int start_bitidx, int end_bitidx)
5511 unsigned long *bitmap;
5512 unsigned long pfn, bitidx;
5513 unsigned long flags = 0;
5514 unsigned long value = 1;
5516 zone = page_zone(page);
5517 pfn = page_to_pfn(page);
5518 bitmap = get_pageblock_bitmap(zone, pfn);
5519 bitidx = pfn_to_bitidx(zone, pfn);
5521 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5522 if (test_bit(bitidx + start_bitidx, bitmap))
5529 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5530 * @page: The page within the block of interest
5531 * @start_bitidx: The first bit of interest
5532 * @end_bitidx: The last bit of interest
5533 * @flags: The flags to set
5535 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5536 int start_bitidx, int end_bitidx)
5539 unsigned long *bitmap;
5540 unsigned long pfn, bitidx;
5541 unsigned long value = 1;
5543 zone = page_zone(page);
5544 pfn = page_to_pfn(page);
5545 bitmap = get_pageblock_bitmap(zone, pfn);
5546 bitidx = pfn_to_bitidx(zone, pfn);
5547 VM_BUG_ON(pfn < zone->zone_start_pfn);
5548 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5550 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5552 __set_bit(bitidx + start_bitidx, bitmap);
5554 __clear_bit(bitidx + start_bitidx, bitmap);
5558 * This is designed as sub function...plz see page_isolation.c also.
5559 * set/clear page block's type to be ISOLATE.
5560 * page allocater never alloc memory from ISOLATE block.
5564 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5566 unsigned long pfn, iter, found;
5568 * For avoiding noise data, lru_add_drain_all() should be called
5569 * If ZONE_MOVABLE, the zone never contains immobile pages
5571 if (zone_idx(zone) == ZONE_MOVABLE)
5574 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5577 pfn = page_to_pfn(page);
5578 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5579 unsigned long check = pfn + iter;
5581 if (!pfn_valid_within(check))
5584 page = pfn_to_page(check);
5585 if (!page_count(page)) {
5586 if (PageBuddy(page))
5587 iter += (1 << page_order(page)) - 1;
5593 * If there are RECLAIMABLE pages, we need to check it.
5594 * But now, memory offline itself doesn't call shrink_slab()
5595 * and it still to be fixed.
5598 * If the page is not RAM, page_count()should be 0.
5599 * we don't need more check. This is an _used_ not-movable page.
5601 * The problematic thing here is PG_reserved pages. PG_reserved
5602 * is set to both of a memory hole page and a _used_ kernel
5611 bool is_pageblock_removable_nolock(struct page *page)
5613 struct zone *zone = page_zone(page);
5614 unsigned long pfn = page_to_pfn(page);
5617 * We have to be careful here because we are iterating over memory
5618 * sections which are not zone aware so we might end up outside of
5619 * the zone but still within the section.
5621 if (!zone || zone->zone_start_pfn > pfn ||
5622 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5625 return __count_immobile_pages(zone, page, 0);
5628 int set_migratetype_isolate(struct page *page)
5631 unsigned long flags, pfn;
5632 struct memory_isolate_notify arg;
5636 zone = page_zone(page);
5638 spin_lock_irqsave(&zone->lock, flags);
5640 pfn = page_to_pfn(page);
5641 arg.start_pfn = pfn;
5642 arg.nr_pages = pageblock_nr_pages;
5643 arg.pages_found = 0;
5646 * It may be possible to isolate a pageblock even if the
5647 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5648 * notifier chain is used by balloon drivers to return the
5649 * number of pages in a range that are held by the balloon
5650 * driver to shrink memory. If all the pages are accounted for
5651 * by balloons, are free, or on the LRU, isolation can continue.
5652 * Later, for example, when memory hotplug notifier runs, these
5653 * pages reported as "can be isolated" should be isolated(freed)
5654 * by the balloon driver through the memory notifier chain.
5656 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5657 notifier_ret = notifier_to_errno(notifier_ret);
5661 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5662 * We just check MOVABLE pages.
5664 if (__count_immobile_pages(zone, page, arg.pages_found))
5668 * immobile means "not-on-lru" paes. If immobile is larger than
5669 * removable-by-driver pages reported by notifier, we'll fail.
5674 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5675 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5678 spin_unlock_irqrestore(&zone->lock, flags);
5684 void unset_migratetype_isolate(struct page *page)
5687 unsigned long flags;
5688 zone = page_zone(page);
5689 spin_lock_irqsave(&zone->lock, flags);
5690 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5692 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5693 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5695 spin_unlock_irqrestore(&zone->lock, flags);
5698 #ifdef CONFIG_MEMORY_HOTREMOVE
5700 * All pages in the range must be isolated before calling this.
5703 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5709 unsigned long flags;
5710 /* find the first valid pfn */
5711 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5716 zone = page_zone(pfn_to_page(pfn));
5717 spin_lock_irqsave(&zone->lock, flags);
5719 while (pfn < end_pfn) {
5720 if (!pfn_valid(pfn)) {
5724 page = pfn_to_page(pfn);
5725 BUG_ON(page_count(page));
5726 BUG_ON(!PageBuddy(page));
5727 order = page_order(page);
5728 #ifdef CONFIG_DEBUG_VM
5729 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5730 pfn, 1 << order, end_pfn);
5732 list_del(&page->lru);
5733 rmv_page_order(page);
5734 zone->free_area[order].nr_free--;
5735 __mod_zone_page_state(zone, NR_FREE_PAGES,
5737 for (i = 0; i < (1 << order); i++)
5738 SetPageReserved((page+i));
5739 pfn += (1 << order);
5741 spin_unlock_irqrestore(&zone->lock, flags);
5745 #ifdef CONFIG_MEMORY_FAILURE
5746 bool is_free_buddy_page(struct page *page)
5748 struct zone *zone = page_zone(page);
5749 unsigned long pfn = page_to_pfn(page);
5750 unsigned long flags;
5753 spin_lock_irqsave(&zone->lock, flags);
5754 for (order = 0; order < MAX_ORDER; order++) {
5755 struct page *page_head = page - (pfn & ((1 << order) - 1));
5757 if (PageBuddy(page_head) && page_order(page_head) >= order)
5760 spin_unlock_irqrestore(&zone->lock, flags);
5762 return order < MAX_ORDER;
5766 static struct trace_print_flags pageflag_names[] = {
5767 {1UL << PG_locked, "locked" },
5768 {1UL << PG_error, "error" },
5769 {1UL << PG_referenced, "referenced" },
5770 {1UL << PG_uptodate, "uptodate" },
5771 {1UL << PG_dirty, "dirty" },
5772 {1UL << PG_lru, "lru" },
5773 {1UL << PG_active, "active" },
5774 {1UL << PG_slab, "slab" },
5775 {1UL << PG_owner_priv_1, "owner_priv_1" },
5776 {1UL << PG_arch_1, "arch_1" },
5777 {1UL << PG_reserved, "reserved" },
5778 {1UL << PG_private, "private" },
5779 {1UL << PG_private_2, "private_2" },
5780 {1UL << PG_writeback, "writeback" },
5781 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5782 {1UL << PG_head, "head" },
5783 {1UL << PG_tail, "tail" },
5785 {1UL << PG_compound, "compound" },
5787 {1UL << PG_swapcache, "swapcache" },
5788 {1UL << PG_mappedtodisk, "mappedtodisk" },
5789 {1UL << PG_reclaim, "reclaim" },
5790 {1UL << PG_swapbacked, "swapbacked" },
5791 {1UL << PG_unevictable, "unevictable" },
5793 {1UL << PG_mlocked, "mlocked" },
5795 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5796 {1UL << PG_uncached, "uncached" },
5798 #ifdef CONFIG_MEMORY_FAILURE
5799 {1UL << PG_hwpoison, "hwpoison" },
5804 static void dump_page_flags(unsigned long flags)
5806 const char *delim = "";
5810 printk(KERN_ALERT "page flags: %#lx(", flags);
5812 /* remove zone id */
5813 flags &= (1UL << NR_PAGEFLAGS) - 1;
5815 for (i = 0; pageflag_names[i].name && flags; i++) {
5817 mask = pageflag_names[i].mask;
5818 if ((flags & mask) != mask)
5822 printk("%s%s", delim, pageflag_names[i].name);
5826 /* check for left over flags */
5828 printk("%s%#lx", delim, flags);
5833 void dump_page(struct page *page)
5836 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5837 page, atomic_read(&page->_count), page_mapcount(page),
5838 page->mapping, page->index);
5839 dump_page_flags(page->flags);
5840 mem_cgroup_print_bad_page(page);