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));
324 /* Leave bad fields for debug, except PageBuddy could make trouble */
325 reset_page_mapcount(page); /* remove PageBuddy */
326 add_taint(TAINT_BAD_PAGE);
330 * Higher-order pages are called "compound pages". They are structured thusly:
332 * The first PAGE_SIZE page is called the "head page".
334 * The remaining PAGE_SIZE pages are called "tail pages".
336 * All pages have PG_compound set. All pages have their ->private pointing at
337 * the head page (even the head page has this).
339 * The first tail page's ->lru.next holds the address of the compound page's
340 * put_page() function. Its ->lru.prev holds the order of allocation.
341 * This usage means that zero-order pages may not be compound.
344 static void free_compound_page(struct page *page)
346 __free_pages_ok(page, compound_order(page));
349 void prep_compound_page(struct page *page, unsigned long order)
352 int nr_pages = 1 << order;
354 set_compound_page_dtor(page, free_compound_page);
355 set_compound_order(page, order);
357 for (i = 1; i < nr_pages; i++) {
358 struct page *p = page + i;
360 set_page_count(p, 0);
361 p->first_page = page;
365 /* update __split_huge_page_refcount if you change this function */
366 static int destroy_compound_page(struct page *page, unsigned long order)
369 int nr_pages = 1 << order;
372 if (unlikely(compound_order(page) != order) ||
373 unlikely(!PageHead(page))) {
378 __ClearPageHead(page);
380 for (i = 1; i < nr_pages; i++) {
381 struct page *p = page + i;
383 if (unlikely(!PageTail(p) || (p->first_page != page))) {
393 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
398 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
399 * and __GFP_HIGHMEM from hard or soft interrupt context.
401 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
402 for (i = 0; i < (1 << order); i++)
403 clear_highpage(page + i);
406 static inline void set_page_order(struct page *page, int order)
408 set_page_private(page, order);
409 __SetPageBuddy(page);
412 static inline void rmv_page_order(struct page *page)
414 __ClearPageBuddy(page);
415 set_page_private(page, 0);
419 * Locate the struct page for both the matching buddy in our
420 * pair (buddy1) and the combined O(n+1) page they form (page).
422 * 1) Any buddy B1 will have an order O twin B2 which satisfies
423 * the following equation:
425 * For example, if the starting buddy (buddy2) is #8 its order
427 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
429 * 2) Any buddy B will have an order O+1 parent P which
430 * satisfies the following equation:
433 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
435 static inline unsigned long
436 __find_buddy_index(unsigned long page_idx, unsigned int order)
438 return page_idx ^ (1 << order);
442 * This function checks whether a page is free && is the buddy
443 * we can do coalesce a page and its buddy if
444 * (a) the buddy is not in a hole &&
445 * (b) the buddy is in the buddy system &&
446 * (c) a page and its buddy have the same order &&
447 * (d) a page and its buddy are in the same zone.
449 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
450 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
452 * For recording page's order, we use page_private(page).
454 static inline int page_is_buddy(struct page *page, struct page *buddy,
457 if (!pfn_valid_within(page_to_pfn(buddy)))
460 if (page_zone_id(page) != page_zone_id(buddy))
463 if (PageBuddy(buddy) && page_order(buddy) == order) {
464 VM_BUG_ON(page_count(buddy) != 0);
471 * Freeing function for a buddy system allocator.
473 * The concept of a buddy system is to maintain direct-mapped table
474 * (containing bit values) for memory blocks of various "orders".
475 * The bottom level table contains the map for the smallest allocatable
476 * units of memory (here, pages), and each level above it describes
477 * pairs of units from the levels below, hence, "buddies".
478 * At a high level, all that happens here is marking the table entry
479 * at the bottom level available, and propagating the changes upward
480 * as necessary, plus some accounting needed to play nicely with other
481 * parts of the VM system.
482 * At each level, we keep a list of pages, which are heads of continuous
483 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
484 * order is recorded in page_private(page) field.
485 * So when we are allocating or freeing one, we can derive the state of the
486 * other. That is, if we allocate a small block, and both were
487 * free, the remainder of the region must be split into blocks.
488 * If a block is freed, and its buddy is also free, then this
489 * triggers coalescing into a block of larger size.
494 static inline void __free_one_page(struct page *page,
495 struct zone *zone, unsigned int order,
498 unsigned long page_idx;
499 unsigned long combined_idx;
500 unsigned long uninitialized_var(buddy_idx);
503 if (unlikely(PageCompound(page)))
504 if (unlikely(destroy_compound_page(page, order)))
507 VM_BUG_ON(migratetype == -1);
509 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
511 VM_BUG_ON(page_idx & ((1 << order) - 1));
512 VM_BUG_ON(bad_range(zone, page));
514 while (order < MAX_ORDER-1) {
515 buddy_idx = __find_buddy_index(page_idx, order);
516 buddy = page + (buddy_idx - page_idx);
517 if (!page_is_buddy(page, buddy, order))
520 /* Our buddy is free, merge with it and move up one order. */
521 list_del(&buddy->lru);
522 zone->free_area[order].nr_free--;
523 rmv_page_order(buddy);
524 combined_idx = buddy_idx & page_idx;
525 page = page + (combined_idx - page_idx);
526 page_idx = combined_idx;
529 set_page_order(page, order);
532 * If this is not the largest possible page, check if the buddy
533 * of the next-highest order is free. If it is, it's possible
534 * that pages are being freed that will coalesce soon. In case,
535 * that is happening, add the free page to the tail of the list
536 * so it's less likely to be used soon and more likely to be merged
537 * as a higher order page
539 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
540 struct page *higher_page, *higher_buddy;
541 combined_idx = buddy_idx & page_idx;
542 higher_page = page + (combined_idx - page_idx);
543 buddy_idx = __find_buddy_index(combined_idx, order + 1);
544 higher_buddy = page + (buddy_idx - combined_idx);
545 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
546 list_add_tail(&page->lru,
547 &zone->free_area[order].free_list[migratetype]);
552 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
554 zone->free_area[order].nr_free++;
558 * free_page_mlock() -- clean up attempts to free and mlocked() page.
559 * Page should not be on lru, so no need to fix that up.
560 * free_pages_check() will verify...
562 static inline void free_page_mlock(struct page *page)
564 __dec_zone_page_state(page, NR_MLOCK);
565 __count_vm_event(UNEVICTABLE_MLOCKFREED);
568 static inline int free_pages_check(struct page *page)
570 if (unlikely(page_mapcount(page) |
571 (page->mapping != NULL) |
572 (atomic_read(&page->_count) != 0) |
573 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
574 (mem_cgroup_bad_page_check(page)))) {
578 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
579 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
584 * Frees a number of pages from the PCP lists
585 * Assumes all pages on list are in same zone, and of same order.
586 * count is the number of pages to free.
588 * If the zone was previously in an "all pages pinned" state then look to
589 * see if this freeing clears that state.
591 * And clear the zone's pages_scanned counter, to hold off the "all pages are
592 * pinned" detection logic.
594 static void free_pcppages_bulk(struct zone *zone, int count,
595 struct per_cpu_pages *pcp)
601 spin_lock(&zone->lock);
602 zone->all_unreclaimable = 0;
603 zone->pages_scanned = 0;
607 struct list_head *list;
610 * Remove pages from lists in a round-robin fashion. A
611 * batch_free count is maintained that is incremented when an
612 * empty list is encountered. This is so more pages are freed
613 * off fuller lists instead of spinning excessively around empty
618 if (++migratetype == MIGRATE_PCPTYPES)
620 list = &pcp->lists[migratetype];
621 } while (list_empty(list));
623 /* This is the only non-empty list. Free them all. */
624 if (batch_free == MIGRATE_PCPTYPES)
625 batch_free = to_free;
628 page = list_entry(list->prev, struct page, lru);
629 /* must delete as __free_one_page list manipulates */
630 list_del(&page->lru);
631 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
632 __free_one_page(page, zone, 0, page_private(page));
633 trace_mm_page_pcpu_drain(page, 0, page_private(page));
634 } while (--to_free && --batch_free && !list_empty(list));
636 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
637 spin_unlock(&zone->lock);
640 static void free_one_page(struct zone *zone, struct page *page, int order,
643 spin_lock(&zone->lock);
644 zone->all_unreclaimable = 0;
645 zone->pages_scanned = 0;
647 __free_one_page(page, zone, order, migratetype);
648 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
649 spin_unlock(&zone->lock);
652 static bool free_pages_prepare(struct page *page, unsigned int order)
658 if (PageForeign(page)) {
659 PageForeignDestructor(page, order);
664 trace_mm_page_free_direct(page, order);
665 kmemcheck_free_shadow(page, order);
668 page->mapping = NULL;
669 for (i = 0; i < (1 << order); i++)
670 bad += free_pages_check(page + i);
674 if (!PageHighMem(page)) {
675 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
676 debug_check_no_obj_freed(page_address(page),
679 arch_free_page(page, order);
680 kernel_map_pages(page, 1 << order, 0);
685 static void __free_pages_ok(struct page *page, unsigned int order)
688 int wasMlocked = __TestClearPageMlocked(page);
691 WARN_ON(PageForeign(page) && wasMlocked);
693 if (!free_pages_prepare(page, order))
696 local_irq_save(flags);
697 if (unlikely(wasMlocked))
698 free_page_mlock(page);
699 __count_vm_events(PGFREE, 1 << order);
700 free_one_page(page_zone(page), page, order,
701 get_pageblock_migratetype(page));
702 local_irq_restore(flags);
706 * permit the bootmem allocator to evade page validation on high-order frees
708 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
711 __ClearPageReserved(page);
712 set_page_count(page, 0);
713 set_page_refcounted(page);
719 for (loop = 0; loop < BITS_PER_LONG; loop++) {
720 struct page *p = &page[loop];
722 if (loop + 1 < BITS_PER_LONG)
724 __ClearPageReserved(p);
725 set_page_count(p, 0);
728 set_page_refcounted(page);
729 __free_pages(page, order);
735 * The order of subdivision here is critical for the IO subsystem.
736 * Please do not alter this order without good reasons and regression
737 * testing. Specifically, as large blocks of memory are subdivided,
738 * the order in which smaller blocks are delivered depends on the order
739 * they're subdivided in this function. This is the primary factor
740 * influencing the order in which pages are delivered to the IO
741 * subsystem according to empirical testing, and this is also justified
742 * by considering the behavior of a buddy system containing a single
743 * large block of memory acted on by a series of small allocations.
744 * This behavior is a critical factor in sglist merging's success.
748 static inline void expand(struct zone *zone, struct page *page,
749 int low, int high, struct free_area *area,
752 unsigned long size = 1 << high;
758 VM_BUG_ON(bad_range(zone, &page[size]));
759 list_add(&page[size].lru, &area->free_list[migratetype]);
761 set_page_order(&page[size], high);
766 * This page is about to be returned from the page allocator
768 static inline int check_new_page(struct page *page)
770 if (unlikely(page_mapcount(page) |
771 (page->mapping != NULL) |
772 (atomic_read(&page->_count) != 0) |
773 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
774 (mem_cgroup_bad_page_check(page)))) {
781 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
785 for (i = 0; i < (1 << order); i++) {
786 struct page *p = page + i;
787 if (unlikely(check_new_page(p)))
791 set_page_private(page, 0);
792 set_page_refcounted(page);
794 arch_alloc_page(page, order);
795 kernel_map_pages(page, 1 << order, 1);
797 if (gfp_flags & __GFP_ZERO)
798 prep_zero_page(page, order, gfp_flags);
800 if (order && (gfp_flags & __GFP_COMP))
801 prep_compound_page(page, order);
807 * Go through the free lists for the given migratetype and remove
808 * the smallest available page from the freelists
811 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
814 unsigned int current_order;
815 struct free_area * area;
818 /* Find a page of the appropriate size in the preferred list */
819 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
820 area = &(zone->free_area[current_order]);
821 if (list_empty(&area->free_list[migratetype]))
824 page = list_entry(area->free_list[migratetype].next,
826 list_del(&page->lru);
827 rmv_page_order(page);
829 expand(zone, page, order, current_order, area, migratetype);
838 * This array describes the order lists are fallen back to when
839 * the free lists for the desirable migrate type are depleted
841 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
842 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
843 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
844 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
845 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
849 * Move the free pages in a range to the free lists of the requested type.
850 * Note that start_page and end_pages are not aligned on a pageblock
851 * boundary. If alignment is required, use move_freepages_block()
853 static int move_freepages(struct zone *zone,
854 struct page *start_page, struct page *end_page,
861 #ifndef CONFIG_HOLES_IN_ZONE
863 * page_zone is not safe to call in this context when
864 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
865 * anyway as we check zone boundaries in move_freepages_block().
866 * Remove at a later date when no bug reports exist related to
867 * grouping pages by mobility
869 BUG_ON(page_zone(start_page) != page_zone(end_page));
872 for (page = start_page; page <= end_page;) {
873 /* Make sure we are not inadvertently changing nodes */
874 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
876 if (!pfn_valid_within(page_to_pfn(page))) {
881 if (!PageBuddy(page)) {
886 order = page_order(page);
887 list_move(&page->lru,
888 &zone->free_area[order].free_list[migratetype]);
890 pages_moved += 1 << order;
896 static int move_freepages_block(struct zone *zone, struct page *page,
899 unsigned long start_pfn, end_pfn;
900 struct page *start_page, *end_page;
902 start_pfn = page_to_pfn(page);
903 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
904 start_page = pfn_to_page(start_pfn);
905 end_page = start_page + pageblock_nr_pages - 1;
906 end_pfn = start_pfn + pageblock_nr_pages - 1;
908 /* Do not cross zone boundaries */
909 if (start_pfn < zone->zone_start_pfn)
911 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
914 return move_freepages(zone, start_page, end_page, migratetype);
917 static void change_pageblock_range(struct page *pageblock_page,
918 int start_order, int migratetype)
920 int nr_pageblocks = 1 << (start_order - pageblock_order);
922 while (nr_pageblocks--) {
923 set_pageblock_migratetype(pageblock_page, migratetype);
924 pageblock_page += pageblock_nr_pages;
928 /* Remove an element from the buddy allocator from the fallback list */
929 static inline struct page *
930 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
932 struct free_area * area;
937 /* Find the largest possible block of pages in the other list */
938 for (current_order = MAX_ORDER-1; current_order >= order;
940 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
941 migratetype = fallbacks[start_migratetype][i];
943 /* MIGRATE_RESERVE handled later if necessary */
944 if (migratetype == MIGRATE_RESERVE)
947 area = &(zone->free_area[current_order]);
948 if (list_empty(&area->free_list[migratetype]))
951 page = list_entry(area->free_list[migratetype].next,
956 * If breaking a large block of pages, move all free
957 * pages to the preferred allocation list. If falling
958 * back for a reclaimable kernel allocation, be more
959 * aggressive about taking ownership of free pages
961 if (unlikely(current_order >= (pageblock_order >> 1)) ||
962 start_migratetype == MIGRATE_RECLAIMABLE ||
963 page_group_by_mobility_disabled) {
965 pages = move_freepages_block(zone, page,
968 /* Claim the whole block if over half of it is free */
969 if (pages >= (1 << (pageblock_order-1)) ||
970 page_group_by_mobility_disabled)
971 set_pageblock_migratetype(page,
974 migratetype = start_migratetype;
977 /* Remove the page from the freelists */
978 list_del(&page->lru);
979 rmv_page_order(page);
981 /* Take ownership for orders >= pageblock_order */
982 if (current_order >= pageblock_order)
983 change_pageblock_range(page, current_order,
986 expand(zone, page, order, current_order, area, migratetype);
988 trace_mm_page_alloc_extfrag(page, order, current_order,
989 start_migratetype, migratetype);
999 * Do the hard work of removing an element from the buddy allocator.
1000 * Call me with the zone->lock already held.
1002 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1008 page = __rmqueue_smallest(zone, order, migratetype);
1010 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1011 page = __rmqueue_fallback(zone, order, migratetype);
1014 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1015 * is used because __rmqueue_smallest is an inline function
1016 * and we want just one call site
1019 migratetype = MIGRATE_RESERVE;
1024 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1029 * Obtain a specified number of elements from the buddy allocator, all under
1030 * a single hold of the lock, for efficiency. Add them to the supplied list.
1031 * Returns the number of new pages which were placed at *list.
1033 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1034 unsigned long count, struct list_head *list,
1035 int migratetype, int cold)
1039 spin_lock(&zone->lock);
1040 for (i = 0; i < count; ++i) {
1041 struct page *page = __rmqueue(zone, order, migratetype);
1042 if (unlikely(page == NULL))
1046 * Split buddy pages returned by expand() are received here
1047 * in physical page order. The page is added to the callers and
1048 * list and the list head then moves forward. From the callers
1049 * perspective, the linked list is ordered by page number in
1050 * some conditions. This is useful for IO devices that can
1051 * merge IO requests if the physical pages are ordered
1054 if (likely(cold == 0))
1055 list_add(&page->lru, list);
1057 list_add_tail(&page->lru, list);
1058 set_page_private(page, migratetype);
1061 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1062 spin_unlock(&zone->lock);
1068 * Called from the vmstat counter updater to drain pagesets of this
1069 * currently executing processor on remote nodes after they have
1072 * Note that this function must be called with the thread pinned to
1073 * a single processor.
1075 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1077 unsigned long flags;
1080 local_irq_save(flags);
1081 if (pcp->count >= pcp->batch)
1082 to_drain = pcp->batch;
1084 to_drain = pcp->count;
1085 free_pcppages_bulk(zone, to_drain, pcp);
1086 pcp->count -= to_drain;
1087 local_irq_restore(flags);
1092 * Drain pages of the indicated processor.
1094 * The processor must either be the current processor and the
1095 * thread pinned to the current processor or a processor that
1098 static void drain_pages(unsigned int cpu)
1100 unsigned long flags;
1103 for_each_populated_zone(zone) {
1104 struct per_cpu_pageset *pset;
1105 struct per_cpu_pages *pcp;
1107 local_irq_save(flags);
1108 pset = per_cpu_ptr(zone->pageset, cpu);
1112 free_pcppages_bulk(zone, pcp->count, pcp);
1115 local_irq_restore(flags);
1120 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1122 void drain_local_pages(void *arg)
1124 drain_pages(smp_processor_id());
1128 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1130 void drain_all_pages(void)
1132 on_each_cpu(drain_local_pages, NULL, 1);
1135 #ifdef CONFIG_HIBERNATION
1137 void mark_free_pages(struct zone *zone)
1139 unsigned long pfn, max_zone_pfn;
1140 unsigned long flags;
1142 struct list_head *curr;
1144 if (!zone->spanned_pages)
1147 spin_lock_irqsave(&zone->lock, flags);
1149 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1150 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1151 if (pfn_valid(pfn)) {
1152 struct page *page = pfn_to_page(pfn);
1154 if (!swsusp_page_is_forbidden(page))
1155 swsusp_unset_page_free(page);
1158 for_each_migratetype_order(order, t) {
1159 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1162 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1163 for (i = 0; i < (1UL << order); i++)
1164 swsusp_set_page_free(pfn_to_page(pfn + i));
1167 spin_unlock_irqrestore(&zone->lock, flags);
1169 #endif /* CONFIG_PM */
1172 * Free a 0-order page
1173 * cold == 1 ? free a cold page : free a hot page
1175 void free_hot_cold_page(struct page *page, int cold)
1177 struct zone *zone = page_zone(page);
1178 struct per_cpu_pages *pcp;
1179 unsigned long flags;
1181 int wasMlocked = __TestClearPageMlocked(page);
1184 WARN_ON(PageForeign(page) && wasMlocked);
1186 if (!free_pages_prepare(page, 0))
1189 migratetype = get_pageblock_migratetype(page);
1190 set_page_private(page, migratetype);
1191 local_irq_save(flags);
1192 if (unlikely(wasMlocked))
1193 free_page_mlock(page);
1194 __count_vm_event(PGFREE);
1197 * We only track unmovable, reclaimable and movable on pcp lists.
1198 * Free ISOLATE pages back to the allocator because they are being
1199 * offlined but treat RESERVE as movable pages so we can get those
1200 * areas back if necessary. Otherwise, we may have to free
1201 * excessively into the page allocator
1203 if (migratetype >= MIGRATE_PCPTYPES) {
1204 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1205 free_one_page(zone, page, 0, migratetype);
1208 migratetype = MIGRATE_MOVABLE;
1211 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1213 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1215 list_add(&page->lru, &pcp->lists[migratetype]);
1217 if (pcp->count >= pcp->high) {
1218 free_pcppages_bulk(zone, pcp->batch, pcp);
1219 pcp->count -= pcp->batch;
1223 local_irq_restore(flags);
1227 * split_page takes a non-compound higher-order page, and splits it into
1228 * n (1<<order) sub-pages: page[0..n]
1229 * Each sub-page must be freed individually.
1231 * Note: this is probably too low level an operation for use in drivers.
1232 * Please consult with lkml before using this in your driver.
1234 void split_page(struct page *page, unsigned int order)
1238 VM_BUG_ON(PageCompound(page));
1239 VM_BUG_ON(!page_count(page));
1241 #ifdef CONFIG_KMEMCHECK
1243 * Split shadow pages too, because free(page[0]) would
1244 * otherwise free the whole shadow.
1246 if (kmemcheck_page_is_tracked(page))
1247 split_page(virt_to_page(page[0].shadow), order);
1250 for (i = 1; i < (1 << order); i++)
1251 set_page_refcounted(page + i);
1255 * Similar to split_page except the page is already free. As this is only
1256 * being used for migration, the migratetype of the block also changes.
1257 * As this is called with interrupts disabled, the caller is responsible
1258 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1261 * Note: this is probably too low level an operation for use in drivers.
1262 * Please consult with lkml before using this in your driver.
1264 int split_free_page(struct page *page)
1267 unsigned long watermark;
1270 BUG_ON(!PageBuddy(page));
1272 zone = page_zone(page);
1273 order = page_order(page);
1275 /* Obey watermarks as if the page was being allocated */
1276 watermark = low_wmark_pages(zone) + (1 << order);
1277 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1280 /* Remove page from free list */
1281 list_del(&page->lru);
1282 zone->free_area[order].nr_free--;
1283 rmv_page_order(page);
1284 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1286 /* Split into individual pages */
1287 set_page_refcounted(page);
1288 split_page(page, order);
1290 if (order >= pageblock_order - 1) {
1291 struct page *endpage = page + (1 << order) - 1;
1292 for (; page < endpage; page += pageblock_nr_pages)
1293 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1300 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1301 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1305 struct page *buffered_rmqueue(struct zone *preferred_zone,
1306 struct zone *zone, int order, gfp_t gfp_flags,
1309 unsigned long flags;
1311 int cold = !!(gfp_flags & __GFP_COLD);
1314 if (likely(order == 0)) {
1315 struct per_cpu_pages *pcp;
1316 struct list_head *list;
1318 local_irq_save(flags);
1319 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1320 list = &pcp->lists[migratetype];
1321 if (list_empty(list)) {
1322 pcp->count += rmqueue_bulk(zone, 0,
1325 if (unlikely(list_empty(list)))
1330 page = list_entry(list->prev, struct page, lru);
1332 page = list_entry(list->next, struct page, lru);
1334 list_del(&page->lru);
1337 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1339 * __GFP_NOFAIL is not to be used in new code.
1341 * All __GFP_NOFAIL callers should be fixed so that they
1342 * properly detect and handle allocation failures.
1344 * We most definitely don't want callers attempting to
1345 * allocate greater than order-1 page units with
1348 WARN_ON_ONCE(order > 1);
1350 spin_lock_irqsave(&zone->lock, flags);
1351 page = __rmqueue(zone, order, migratetype);
1352 spin_unlock(&zone->lock);
1355 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1358 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1359 zone_statistics(preferred_zone, zone, gfp_flags);
1360 local_irq_restore(flags);
1362 VM_BUG_ON(bad_range(zone, page));
1363 if (prep_new_page(page, order, gfp_flags))
1368 local_irq_restore(flags);
1372 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1373 #define ALLOC_WMARK_MIN WMARK_MIN
1374 #define ALLOC_WMARK_LOW WMARK_LOW
1375 #define ALLOC_WMARK_HIGH WMARK_HIGH
1376 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1378 /* Mask to get the watermark bits */
1379 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1381 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1382 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1383 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1385 #ifdef CONFIG_FAIL_PAGE_ALLOC
1388 struct fault_attr attr;
1390 u32 ignore_gfp_highmem;
1391 u32 ignore_gfp_wait;
1393 } fail_page_alloc = {
1394 .attr = FAULT_ATTR_INITIALIZER,
1395 .ignore_gfp_wait = 1,
1396 .ignore_gfp_highmem = 1,
1400 static int __init setup_fail_page_alloc(char *str)
1402 return setup_fault_attr(&fail_page_alloc.attr, str);
1404 __setup("fail_page_alloc=", setup_fail_page_alloc);
1406 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1408 if (order < fail_page_alloc.min_order)
1410 if (gfp_mask & __GFP_NOFAIL)
1412 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1414 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1417 return should_fail(&fail_page_alloc.attr, 1 << order);
1420 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1422 static int __init fail_page_alloc_debugfs(void)
1424 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1427 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1428 &fail_page_alloc.attr);
1430 return PTR_ERR(dir);
1432 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1433 &fail_page_alloc.ignore_gfp_wait))
1435 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1436 &fail_page_alloc.ignore_gfp_highmem))
1438 if (!debugfs_create_u32("min-order", mode, dir,
1439 &fail_page_alloc.min_order))
1444 debugfs_remove_recursive(dir);
1449 late_initcall(fail_page_alloc_debugfs);
1451 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1453 #else /* CONFIG_FAIL_PAGE_ALLOC */
1455 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1460 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1463 * Return true if free pages are above 'mark'. This takes into account the order
1464 * of the allocation.
1466 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1467 int classzone_idx, int alloc_flags, long free_pages)
1469 /* free_pages my go negative - that's OK */
1473 free_pages -= (1 << order) + 1;
1474 if (alloc_flags & ALLOC_HIGH)
1476 if (alloc_flags & ALLOC_HARDER)
1479 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1481 for (o = 0; o < order; o++) {
1482 /* At the next order, this order's pages become unavailable */
1483 free_pages -= z->free_area[o].nr_free << o;
1485 /* Require fewer higher order pages to be free */
1488 if (free_pages <= min)
1494 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1495 int classzone_idx, int alloc_flags)
1497 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1498 zone_page_state(z, NR_FREE_PAGES));
1501 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1502 int classzone_idx, int alloc_flags)
1504 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1506 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1507 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1509 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1515 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1516 * skip over zones that are not allowed by the cpuset, or that have
1517 * been recently (in last second) found to be nearly full. See further
1518 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1519 * that have to skip over a lot of full or unallowed zones.
1521 * If the zonelist cache is present in the passed in zonelist, then
1522 * returns a pointer to the allowed node mask (either the current
1523 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1525 * If the zonelist cache is not available for this zonelist, does
1526 * nothing and returns NULL.
1528 * If the fullzones BITMAP in the zonelist cache is stale (more than
1529 * a second since last zap'd) then we zap it out (clear its bits.)
1531 * We hold off even calling zlc_setup, until after we've checked the
1532 * first zone in the zonelist, on the theory that most allocations will
1533 * be satisfied from that first zone, so best to examine that zone as
1534 * quickly as we can.
1536 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1538 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1539 nodemask_t *allowednodes; /* zonelist_cache approximation */
1541 zlc = zonelist->zlcache_ptr;
1545 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1546 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1547 zlc->last_full_zap = jiffies;
1550 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1551 &cpuset_current_mems_allowed :
1552 &node_states[N_HIGH_MEMORY];
1553 return allowednodes;
1557 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1558 * if it is worth looking at further for free memory:
1559 * 1) Check that the zone isn't thought to be full (doesn't have its
1560 * bit set in the zonelist_cache fullzones BITMAP).
1561 * 2) Check that the zones node (obtained from the zonelist_cache
1562 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1563 * Return true (non-zero) if zone is worth looking at further, or
1564 * else return false (zero) if it is not.
1566 * This check -ignores- the distinction between various watermarks,
1567 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1568 * found to be full for any variation of these watermarks, it will
1569 * be considered full for up to one second by all requests, unless
1570 * we are so low on memory on all allowed nodes that we are forced
1571 * into the second scan of the zonelist.
1573 * In the second scan we ignore this zonelist cache and exactly
1574 * apply the watermarks to all zones, even it is slower to do so.
1575 * We are low on memory in the second scan, and should leave no stone
1576 * unturned looking for a free page.
1578 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1579 nodemask_t *allowednodes)
1581 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1582 int i; /* index of *z in zonelist zones */
1583 int n; /* node that zone *z is on */
1585 zlc = zonelist->zlcache_ptr;
1589 i = z - zonelist->_zonerefs;
1592 /* This zone is worth trying if it is allowed but not full */
1593 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1597 * Given 'z' scanning a zonelist, set the corresponding bit in
1598 * zlc->fullzones, so that subsequent attempts to allocate a page
1599 * from that zone don't waste time re-examining it.
1601 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1603 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1604 int i; /* index of *z in zonelist zones */
1606 zlc = zonelist->zlcache_ptr;
1610 i = z - zonelist->_zonerefs;
1612 set_bit(i, zlc->fullzones);
1616 * clear all zones full, called after direct reclaim makes progress so that
1617 * a zone that was recently full is not skipped over for up to a second
1619 static void zlc_clear_zones_full(struct zonelist *zonelist)
1621 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1623 zlc = zonelist->zlcache_ptr;
1627 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1630 #else /* CONFIG_NUMA */
1632 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1637 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1638 nodemask_t *allowednodes)
1643 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1647 static void zlc_clear_zones_full(struct zonelist *zonelist)
1650 #endif /* CONFIG_NUMA */
1653 * get_page_from_freelist goes through the zonelist trying to allocate
1656 static struct page *
1657 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1658 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1659 struct zone *preferred_zone, int migratetype)
1662 struct page *page = NULL;
1665 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1666 int zlc_active = 0; /* set if using zonelist_cache */
1667 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1669 classzone_idx = zone_idx(preferred_zone);
1672 * Scan zonelist, looking for a zone with enough free.
1673 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1675 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1676 high_zoneidx, nodemask) {
1677 if (NUMA_BUILD && zlc_active &&
1678 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1680 if ((alloc_flags & ALLOC_CPUSET) &&
1681 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1684 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1685 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1689 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1690 if (zone_watermark_ok(zone, order, mark,
1691 classzone_idx, alloc_flags))
1694 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1696 * we do zlc_setup if there are multiple nodes
1697 * and before considering the first zone allowed
1700 allowednodes = zlc_setup(zonelist, alloc_flags);
1705 if (zone_reclaim_mode == 0)
1706 goto this_zone_full;
1709 * As we may have just activated ZLC, check if the first
1710 * eligible zone has failed zone_reclaim recently.
1712 if (NUMA_BUILD && zlc_active &&
1713 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1716 ret = zone_reclaim(zone, gfp_mask, order);
1718 case ZONE_RECLAIM_NOSCAN:
1721 case ZONE_RECLAIM_FULL:
1722 /* scanned but unreclaimable */
1725 /* did we reclaim enough */
1726 if (!zone_watermark_ok(zone, order, mark,
1727 classzone_idx, alloc_flags))
1728 goto this_zone_full;
1733 page = buffered_rmqueue(preferred_zone, zone, order,
1734 gfp_mask, migratetype);
1739 zlc_mark_zone_full(zonelist, z);
1742 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1743 /* Disable zlc cache for second zonelist scan */
1751 * Large machines with many possible nodes should not always dump per-node
1752 * meminfo in irq context.
1754 static inline bool should_suppress_show_mem(void)
1759 ret = in_interrupt();
1764 static DEFINE_RATELIMIT_STATE(nopage_rs,
1765 DEFAULT_RATELIMIT_INTERVAL,
1766 DEFAULT_RATELIMIT_BURST);
1768 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1770 unsigned int filter = SHOW_MEM_FILTER_NODES;
1772 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1776 * This documents exceptions given to allocations in certain
1777 * contexts that are allowed to allocate outside current's set
1780 if (!(gfp_mask & __GFP_NOMEMALLOC))
1781 if (test_thread_flag(TIF_MEMDIE) ||
1782 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1783 filter &= ~SHOW_MEM_FILTER_NODES;
1784 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1785 filter &= ~SHOW_MEM_FILTER_NODES;
1788 struct va_format vaf;
1791 va_start(args, fmt);
1796 pr_warn("%pV", &vaf);
1801 if (!(gfp_mask & __GFP_WAIT)) {
1802 pr_info("The following is only an harmless informational message.\n");
1803 pr_info("Unless you get a _continuous_flood_ of these messages it means\n");
1804 pr_info("everything is working fine. Allocations from irqs cannot be\n");
1805 pr_info("perfectly reliable and the kernel is designed to handle that.\n");
1807 pr_info("%s: page allocation failure. order:%d, mode:0x%x\n",
1808 current->comm, order, gfp_mask);
1811 if (!should_suppress_show_mem())
1816 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1817 unsigned long pages_reclaimed)
1819 /* Do not loop if specifically requested */
1820 if (gfp_mask & __GFP_NORETRY)
1824 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1825 * means __GFP_NOFAIL, but that may not be true in other
1828 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1832 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1833 * specified, then we retry until we no longer reclaim any pages
1834 * (above), or we've reclaimed an order of pages at least as
1835 * large as the allocation's order. In both cases, if the
1836 * allocation still fails, we stop retrying.
1838 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1842 * Don't let big-order allocations loop unless the caller
1843 * explicitly requests that.
1845 if (gfp_mask & __GFP_NOFAIL)
1851 static inline struct page *
1852 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1853 struct zonelist *zonelist, enum zone_type high_zoneidx,
1854 nodemask_t *nodemask, struct zone *preferred_zone,
1859 /* Acquire the OOM killer lock for the zones in zonelist */
1860 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1861 schedule_timeout_uninterruptible(1);
1866 * Go through the zonelist yet one more time, keep very high watermark
1867 * here, this is only to catch a parallel oom killing, we must fail if
1868 * we're still under heavy pressure.
1870 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1871 order, zonelist, high_zoneidx,
1872 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1873 preferred_zone, migratetype);
1877 if (!(gfp_mask & __GFP_NOFAIL)) {
1878 /* The OOM killer will not help higher order allocs */
1879 if (order > PAGE_ALLOC_COSTLY_ORDER)
1881 /* The OOM killer does not needlessly kill tasks for lowmem */
1882 if (high_zoneidx < ZONE_NORMAL)
1885 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1886 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1887 * The caller should handle page allocation failure by itself if
1888 * it specifies __GFP_THISNODE.
1889 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1891 if (gfp_mask & __GFP_THISNODE)
1894 /* Exhausted what can be done so it's blamo time */
1895 out_of_memory(zonelist, gfp_mask, order, nodemask);
1898 clear_zonelist_oom(zonelist, gfp_mask);
1902 #ifdef CONFIG_COMPACTION
1903 /* Try memory compaction for high-order allocations before reclaim */
1904 static struct page *
1905 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1906 struct zonelist *zonelist, enum zone_type high_zoneidx,
1907 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1908 int migratetype, unsigned long *did_some_progress,
1909 bool sync_migration)
1913 if (!order || compaction_deferred(preferred_zone))
1916 current->flags |= PF_MEMALLOC;
1917 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1918 nodemask, sync_migration);
1919 current->flags &= ~PF_MEMALLOC;
1920 if (*did_some_progress != COMPACT_SKIPPED) {
1922 /* Page migration frees to the PCP lists but we want merging */
1923 drain_pages(get_cpu());
1926 page = get_page_from_freelist(gfp_mask, nodemask,
1927 order, zonelist, high_zoneidx,
1928 alloc_flags, preferred_zone,
1931 preferred_zone->compact_considered = 0;
1932 preferred_zone->compact_defer_shift = 0;
1933 count_vm_event(COMPACTSUCCESS);
1938 * It's bad if compaction run occurs and fails.
1939 * The most likely reason is that pages exist,
1940 * but not enough to satisfy watermarks.
1942 count_vm_event(COMPACTFAIL);
1943 defer_compaction(preferred_zone);
1951 static inline struct page *
1952 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1953 struct zonelist *zonelist, enum zone_type high_zoneidx,
1954 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1955 int migratetype, unsigned long *did_some_progress,
1956 bool sync_migration)
1960 #endif /* CONFIG_COMPACTION */
1962 /* The really slow allocator path where we enter direct reclaim */
1963 static inline struct page *
1964 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1965 struct zonelist *zonelist, enum zone_type high_zoneidx,
1966 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1967 int migratetype, unsigned long *did_some_progress)
1969 struct page *page = NULL;
1970 struct reclaim_state reclaim_state;
1971 bool drained = false;
1975 /* We now go into synchronous reclaim */
1976 cpuset_memory_pressure_bump();
1977 current->flags |= PF_MEMALLOC;
1978 lockdep_set_current_reclaim_state(gfp_mask);
1979 reclaim_state.reclaimed_slab = 0;
1980 current->reclaim_state = &reclaim_state;
1982 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1984 current->reclaim_state = NULL;
1985 lockdep_clear_current_reclaim_state();
1986 current->flags &= ~PF_MEMALLOC;
1990 if (unlikely(!(*did_some_progress)))
1993 /* After successful reclaim, reconsider all zones for allocation */
1995 zlc_clear_zones_full(zonelist);
1998 page = get_page_from_freelist(gfp_mask, nodemask, order,
1999 zonelist, high_zoneidx,
2000 alloc_flags, preferred_zone,
2004 * If an allocation failed after direct reclaim, it could be because
2005 * pages are pinned on the per-cpu lists. Drain them and try again
2007 if (!page && !drained) {
2017 * This is called in the allocator slow-path if the allocation request is of
2018 * sufficient urgency to ignore watermarks and take other desperate measures
2020 static inline struct page *
2021 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2022 struct zonelist *zonelist, enum zone_type high_zoneidx,
2023 nodemask_t *nodemask, struct zone *preferred_zone,
2029 page = get_page_from_freelist(gfp_mask, nodemask, order,
2030 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2031 preferred_zone, migratetype);
2033 if (!page && gfp_mask & __GFP_NOFAIL)
2034 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2035 } while (!page && (gfp_mask & __GFP_NOFAIL));
2041 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2042 enum zone_type high_zoneidx,
2043 enum zone_type classzone_idx)
2048 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2049 wakeup_kswapd(zone, order, classzone_idx);
2053 gfp_to_alloc_flags(gfp_t gfp_mask)
2055 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2056 const gfp_t wait = gfp_mask & __GFP_WAIT;
2058 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2059 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2062 * The caller may dip into page reserves a bit more if the caller
2063 * cannot run direct reclaim, or if the caller has realtime scheduling
2064 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2065 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2067 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2071 * Not worth trying to allocate harder for
2072 * __GFP_NOMEMALLOC even if it can't schedule.
2074 if (!(gfp_mask & __GFP_NOMEMALLOC))
2075 alloc_flags |= ALLOC_HARDER;
2077 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2078 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2080 alloc_flags &= ~ALLOC_CPUSET;
2081 } else if (unlikely(rt_task(current)) && !in_interrupt())
2082 alloc_flags |= ALLOC_HARDER;
2084 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2085 if (!in_interrupt() &&
2086 ((current->flags & PF_MEMALLOC) ||
2087 unlikely(test_thread_flag(TIF_MEMDIE))))
2088 alloc_flags |= ALLOC_NO_WATERMARKS;
2094 static inline struct page *
2095 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2096 struct zonelist *zonelist, enum zone_type high_zoneidx,
2097 nodemask_t *nodemask, struct zone *preferred_zone,
2100 const gfp_t wait = gfp_mask & __GFP_WAIT;
2101 struct page *page = NULL;
2103 unsigned long pages_reclaimed = 0;
2104 unsigned long did_some_progress;
2105 bool sync_migration = false;
2108 * In the slowpath, we sanity check order to avoid ever trying to
2109 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2110 * be using allocators in order of preference for an area that is
2113 if (order >= MAX_ORDER) {
2114 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2119 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2120 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2121 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2122 * using a larger set of nodes after it has established that the
2123 * allowed per node queues are empty and that nodes are
2126 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2130 if (!(gfp_mask & __GFP_NO_KSWAPD))
2131 wake_all_kswapd(order, zonelist, high_zoneidx,
2132 zone_idx(preferred_zone));
2135 * OK, we're below the kswapd watermark and have kicked background
2136 * reclaim. Now things get more complex, so set up alloc_flags according
2137 * to how we want to proceed.
2139 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2142 * Find the true preferred zone if the allocation is unconstrained by
2145 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2146 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2150 /* This is the last chance, in general, before the goto nopage. */
2151 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2152 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2153 preferred_zone, migratetype);
2157 /* Allocate without watermarks if the context allows */
2158 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2159 page = __alloc_pages_high_priority(gfp_mask, order,
2160 zonelist, high_zoneidx, nodemask,
2161 preferred_zone, migratetype);
2166 /* Atomic allocations - we can't balance anything */
2170 /* Avoid recursion of direct reclaim */
2171 if (current->flags & PF_MEMALLOC)
2174 /* Avoid allocations with no watermarks from looping endlessly */
2175 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2179 * Try direct compaction. The first pass is asynchronous. Subsequent
2180 * attempts after direct reclaim are synchronous
2182 page = __alloc_pages_direct_compact(gfp_mask, order,
2183 zonelist, high_zoneidx,
2185 alloc_flags, preferred_zone,
2186 migratetype, &did_some_progress,
2190 sync_migration = true;
2192 /* Try direct reclaim and then allocating */
2193 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2194 zonelist, high_zoneidx,
2196 alloc_flags, preferred_zone,
2197 migratetype, &did_some_progress);
2202 * If we failed to make any progress reclaiming, then we are
2203 * running out of options and have to consider going OOM
2205 if (!did_some_progress) {
2206 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2207 if (oom_killer_disabled)
2209 page = __alloc_pages_may_oom(gfp_mask, order,
2210 zonelist, high_zoneidx,
2211 nodemask, preferred_zone,
2216 if (!(gfp_mask & __GFP_NOFAIL)) {
2218 * The oom killer is not called for high-order
2219 * allocations that may fail, so if no progress
2220 * is being made, there are no other options and
2221 * retrying is unlikely to help.
2223 if (order > PAGE_ALLOC_COSTLY_ORDER)
2226 * The oom killer is not called for lowmem
2227 * allocations to prevent needlessly killing
2230 if (high_zoneidx < ZONE_NORMAL)
2238 /* Check if we should retry the allocation */
2239 pages_reclaimed += did_some_progress;
2240 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2241 /* Wait for some write requests to complete then retry */
2242 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2246 * High-order allocations do not necessarily loop after
2247 * direct reclaim and reclaim/compaction depends on compaction
2248 * being called after reclaim so call directly if necessary
2250 page = __alloc_pages_direct_compact(gfp_mask, order,
2251 zonelist, high_zoneidx,
2253 alloc_flags, preferred_zone,
2254 migratetype, &did_some_progress,
2261 warn_alloc_failed(gfp_mask, order, NULL);
2264 if (kmemcheck_enabled)
2265 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2271 * This is the 'heart' of the zoned buddy allocator.
2274 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2275 struct zonelist *zonelist, nodemask_t *nodemask)
2277 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2278 struct zone *preferred_zone;
2280 int migratetype = allocflags_to_migratetype(gfp_mask);
2282 gfp_mask &= gfp_allowed_mask;
2284 lockdep_trace_alloc(gfp_mask);
2286 might_sleep_if(gfp_mask & __GFP_WAIT);
2288 if (should_fail_alloc_page(gfp_mask, order))
2292 * Check the zones suitable for the gfp_mask contain at least one
2293 * valid zone. It's possible to have an empty zonelist as a result
2294 * of GFP_THISNODE and a memoryless node
2296 if (unlikely(!zonelist->_zonerefs->zone))
2300 /* The preferred zone is used for statistics later */
2301 first_zones_zonelist(zonelist, high_zoneidx,
2302 nodemask ? : &cpuset_current_mems_allowed,
2304 if (!preferred_zone) {
2309 /* First allocation attempt */
2310 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2311 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2312 preferred_zone, migratetype);
2313 if (unlikely(!page))
2314 page = __alloc_pages_slowpath(gfp_mask, order,
2315 zonelist, high_zoneidx, nodemask,
2316 preferred_zone, migratetype);
2319 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2322 EXPORT_SYMBOL(__alloc_pages_nodemask);
2325 * Common helper functions.
2327 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2332 * __get_free_pages() returns a 32-bit address, which cannot represent
2335 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2337 page = alloc_pages(gfp_mask, order);
2340 return (unsigned long) page_address(page);
2342 EXPORT_SYMBOL(__get_free_pages);
2344 unsigned long get_zeroed_page(gfp_t gfp_mask)
2346 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2348 EXPORT_SYMBOL(get_zeroed_page);
2350 void __pagevec_free(struct pagevec *pvec)
2352 int i = pagevec_count(pvec);
2355 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2356 free_hot_cold_page(pvec->pages[i], pvec->cold);
2360 void __free_pages(struct page *page, unsigned int order)
2362 if (put_page_testzero(page)) {
2364 free_hot_cold_page(page, 0);
2366 __free_pages_ok(page, order);
2370 EXPORT_SYMBOL(__free_pages);
2372 void free_pages(unsigned long addr, unsigned int order)
2375 VM_BUG_ON(!virt_addr_valid((void *)addr));
2376 __free_pages(virt_to_page((void *)addr), order);
2380 EXPORT_SYMBOL(free_pages);
2382 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2385 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2386 unsigned long used = addr + PAGE_ALIGN(size);
2388 split_page(virt_to_page((void *)addr), order);
2389 while (used < alloc_end) {
2394 return (void *)addr;
2398 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2399 * @size: the number of bytes to allocate
2400 * @gfp_mask: GFP flags for the allocation
2402 * This function is similar to alloc_pages(), except that it allocates the
2403 * minimum number of pages to satisfy the request. alloc_pages() can only
2404 * allocate memory in power-of-two pages.
2406 * This function is also limited by MAX_ORDER.
2408 * Memory allocated by this function must be released by free_pages_exact().
2410 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2412 unsigned int order = get_order(size);
2415 addr = __get_free_pages(gfp_mask, order);
2416 return make_alloc_exact(addr, order, size);
2418 EXPORT_SYMBOL(alloc_pages_exact);
2421 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2423 * @nid: the preferred node ID where memory should be allocated
2424 * @size: the number of bytes to allocate
2425 * @gfp_mask: GFP flags for the allocation
2427 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2429 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2432 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2434 unsigned order = get_order(size);
2435 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2438 return make_alloc_exact((unsigned long)page_address(p), order, size);
2440 EXPORT_SYMBOL(alloc_pages_exact_nid);
2443 * free_pages_exact - release memory allocated via alloc_pages_exact()
2444 * @virt: the value returned by alloc_pages_exact.
2445 * @size: size of allocation, same value as passed to alloc_pages_exact().
2447 * Release the memory allocated by a previous call to alloc_pages_exact.
2449 void free_pages_exact(void *virt, size_t size)
2451 unsigned long addr = (unsigned long)virt;
2452 unsigned long end = addr + PAGE_ALIGN(size);
2454 while (addr < end) {
2459 EXPORT_SYMBOL(free_pages_exact);
2461 static unsigned int nr_free_zone_pages(int offset)
2466 /* Just pick one node, since fallback list is circular */
2467 unsigned int sum = 0;
2469 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2471 for_each_zone_zonelist(zone, z, zonelist, offset) {
2472 unsigned long size = zone->present_pages;
2473 unsigned long high = high_wmark_pages(zone);
2482 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2484 unsigned int nr_free_buffer_pages(void)
2486 return nr_free_zone_pages(gfp_zone(GFP_USER));
2488 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2491 * Amount of free RAM allocatable within all zones
2493 unsigned int nr_free_pagecache_pages(void)
2495 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2498 static inline void show_node(struct zone *zone)
2501 printk("Node %d ", zone_to_nid(zone));
2504 void si_meminfo(struct sysinfo *val)
2506 val->totalram = totalram_pages;
2508 val->freeram = global_page_state(NR_FREE_PAGES);
2509 val->bufferram = nr_blockdev_pages();
2510 val->totalhigh = totalhigh_pages;
2511 val->freehigh = nr_free_highpages();
2512 val->mem_unit = PAGE_SIZE;
2515 EXPORT_SYMBOL(si_meminfo);
2518 void si_meminfo_node(struct sysinfo *val, int nid)
2520 pg_data_t *pgdat = NODE_DATA(nid);
2522 val->totalram = pgdat->node_present_pages;
2523 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2524 #ifdef CONFIG_HIGHMEM
2525 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2526 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2532 val->mem_unit = PAGE_SIZE;
2537 * Determine whether the node should be displayed or not, depending on whether
2538 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2540 bool skip_free_areas_node(unsigned int flags, int nid)
2544 if (!(flags & SHOW_MEM_FILTER_NODES))
2548 ret = !node_isset(nid, cpuset_current_mems_allowed);
2554 #define K(x) ((x) << (PAGE_SHIFT-10))
2557 * Show free area list (used inside shift_scroll-lock stuff)
2558 * We also calculate the percentage fragmentation. We do this by counting the
2559 * memory on each free list with the exception of the first item on the list.
2560 * Suppresses nodes that are not allowed by current's cpuset if
2561 * SHOW_MEM_FILTER_NODES is passed.
2563 void show_free_areas(unsigned int filter)
2568 for_each_populated_zone(zone) {
2569 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2572 printk("%s per-cpu:\n", zone->name);
2574 for_each_online_cpu(cpu) {
2575 struct per_cpu_pageset *pageset;
2577 pageset = per_cpu_ptr(zone->pageset, cpu);
2579 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2580 cpu, pageset->pcp.high,
2581 pageset->pcp.batch, pageset->pcp.count);
2585 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2586 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2588 " dirty:%lu writeback:%lu unstable:%lu\n"
2589 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2590 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2591 global_page_state(NR_ACTIVE_ANON),
2592 global_page_state(NR_INACTIVE_ANON),
2593 global_page_state(NR_ISOLATED_ANON),
2594 global_page_state(NR_ACTIVE_FILE),
2595 global_page_state(NR_INACTIVE_FILE),
2596 global_page_state(NR_ISOLATED_FILE),
2597 global_page_state(NR_UNEVICTABLE),
2598 global_page_state(NR_FILE_DIRTY),
2599 global_page_state(NR_WRITEBACK),
2600 global_page_state(NR_UNSTABLE_NFS),
2601 global_page_state(NR_FREE_PAGES),
2602 global_page_state(NR_SLAB_RECLAIMABLE),
2603 global_page_state(NR_SLAB_UNRECLAIMABLE),
2604 global_page_state(NR_FILE_MAPPED),
2605 global_page_state(NR_SHMEM),
2606 global_page_state(NR_PAGETABLE),
2607 global_page_state(NR_BOUNCE));
2609 for_each_populated_zone(zone) {
2612 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2620 " active_anon:%lukB"
2621 " inactive_anon:%lukB"
2622 " active_file:%lukB"
2623 " inactive_file:%lukB"
2624 " unevictable:%lukB"
2625 " isolated(anon):%lukB"
2626 " isolated(file):%lukB"
2633 " slab_reclaimable:%lukB"
2634 " slab_unreclaimable:%lukB"
2635 " kernel_stack:%lukB"
2639 " writeback_tmp:%lukB"
2640 " pages_scanned:%lu"
2641 " all_unreclaimable? %s"
2644 K(zone_page_state(zone, NR_FREE_PAGES)),
2645 K(min_wmark_pages(zone)),
2646 K(low_wmark_pages(zone)),
2647 K(high_wmark_pages(zone)),
2648 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2649 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2650 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2651 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2652 K(zone_page_state(zone, NR_UNEVICTABLE)),
2653 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2654 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2655 K(zone->present_pages),
2656 K(zone_page_state(zone, NR_MLOCK)),
2657 K(zone_page_state(zone, NR_FILE_DIRTY)),
2658 K(zone_page_state(zone, NR_WRITEBACK)),
2659 K(zone_page_state(zone, NR_FILE_MAPPED)),
2660 K(zone_page_state(zone, NR_SHMEM)),
2661 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2662 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2663 zone_page_state(zone, NR_KERNEL_STACK) *
2665 K(zone_page_state(zone, NR_PAGETABLE)),
2666 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2667 K(zone_page_state(zone, NR_BOUNCE)),
2668 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2669 zone->pages_scanned,
2670 (zone->all_unreclaimable ? "yes" : "no")
2672 printk("lowmem_reserve[]:");
2673 for (i = 0; i < MAX_NR_ZONES; i++)
2674 printk(" %lu", zone->lowmem_reserve[i]);
2678 for_each_populated_zone(zone) {
2679 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2681 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2684 printk("%s: ", zone->name);
2686 spin_lock_irqsave(&zone->lock, flags);
2687 for (order = 0; order < MAX_ORDER; order++) {
2688 nr[order] = zone->free_area[order].nr_free;
2689 total += nr[order] << order;
2691 spin_unlock_irqrestore(&zone->lock, flags);
2692 for (order = 0; order < MAX_ORDER; order++)
2693 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2694 printk("= %lukB\n", K(total));
2697 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2699 show_swap_cache_info();
2702 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2704 zoneref->zone = zone;
2705 zoneref->zone_idx = zone_idx(zone);
2709 * Builds allocation fallback zone lists.
2711 * Add all populated zones of a node to the zonelist.
2713 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2714 int nr_zones, enum zone_type zone_type)
2718 BUG_ON(zone_type >= MAX_NR_ZONES);
2723 zone = pgdat->node_zones + zone_type;
2724 if (populated_zone(zone)) {
2725 zoneref_set_zone(zone,
2726 &zonelist->_zonerefs[nr_zones++]);
2727 check_highest_zone(zone_type);
2730 } while (zone_type);
2737 * 0 = automatic detection of better ordering.
2738 * 1 = order by ([node] distance, -zonetype)
2739 * 2 = order by (-zonetype, [node] distance)
2741 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2742 * the same zonelist. So only NUMA can configure this param.
2744 #define ZONELIST_ORDER_DEFAULT 0
2745 #define ZONELIST_ORDER_NODE 1
2746 #define ZONELIST_ORDER_ZONE 2
2748 /* zonelist order in the kernel.
2749 * set_zonelist_order() will set this to NODE or ZONE.
2751 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2752 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2756 /* The value user specified ....changed by config */
2757 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2758 /* string for sysctl */
2759 #define NUMA_ZONELIST_ORDER_LEN 16
2760 char numa_zonelist_order[16] = "default";
2763 * interface for configure zonelist ordering.
2764 * command line option "numa_zonelist_order"
2765 * = "[dD]efault - default, automatic configuration.
2766 * = "[nN]ode - order by node locality, then by zone within node
2767 * = "[zZ]one - order by zone, then by locality within zone
2770 static int __parse_numa_zonelist_order(char *s)
2772 if (*s == 'd' || *s == 'D') {
2773 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2774 } else if (*s == 'n' || *s == 'N') {
2775 user_zonelist_order = ZONELIST_ORDER_NODE;
2776 } else if (*s == 'z' || *s == 'Z') {
2777 user_zonelist_order = ZONELIST_ORDER_ZONE;
2780 "Ignoring invalid numa_zonelist_order value: "
2787 static __init int setup_numa_zonelist_order(char *s)
2794 ret = __parse_numa_zonelist_order(s);
2796 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2800 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2803 * sysctl handler for numa_zonelist_order
2805 int numa_zonelist_order_handler(ctl_table *table, int write,
2806 void __user *buffer, size_t *length,
2809 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2811 static DEFINE_MUTEX(zl_order_mutex);
2813 mutex_lock(&zl_order_mutex);
2815 strcpy(saved_string, (char*)table->data);
2816 ret = proc_dostring(table, write, buffer, length, ppos);
2820 int oldval = user_zonelist_order;
2821 if (__parse_numa_zonelist_order((char*)table->data)) {
2823 * bogus value. restore saved string
2825 strncpy((char*)table->data, saved_string,
2826 NUMA_ZONELIST_ORDER_LEN);
2827 user_zonelist_order = oldval;
2828 } else if (oldval != user_zonelist_order) {
2829 mutex_lock(&zonelists_mutex);
2830 build_all_zonelists(NULL);
2831 mutex_unlock(&zonelists_mutex);
2835 mutex_unlock(&zl_order_mutex);
2840 #define MAX_NODE_LOAD (nr_online_nodes)
2841 static int node_load[MAX_NUMNODES];
2844 * find_next_best_node - find the next node that should appear in a given node's fallback list
2845 * @node: node whose fallback list we're appending
2846 * @used_node_mask: nodemask_t of already used nodes
2848 * We use a number of factors to determine which is the next node that should
2849 * appear on a given node's fallback list. The node should not have appeared
2850 * already in @node's fallback list, and it should be the next closest node
2851 * according to the distance array (which contains arbitrary distance values
2852 * from each node to each node in the system), and should also prefer nodes
2853 * with no CPUs, since presumably they'll have very little allocation pressure
2854 * on them otherwise.
2855 * It returns -1 if no node is found.
2857 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2860 int min_val = INT_MAX;
2862 const struct cpumask *tmp = cpumask_of_node(0);
2864 /* Use the local node if we haven't already */
2865 if (!node_isset(node, *used_node_mask)) {
2866 node_set(node, *used_node_mask);
2870 for_each_node_state(n, N_HIGH_MEMORY) {
2872 /* Don't want a node to appear more than once */
2873 if (node_isset(n, *used_node_mask))
2876 /* Use the distance array to find the distance */
2877 val = node_distance(node, n);
2879 /* Penalize nodes under us ("prefer the next node") */
2882 /* Give preference to headless and unused nodes */
2883 tmp = cpumask_of_node(n);
2884 if (!cpumask_empty(tmp))
2885 val += PENALTY_FOR_NODE_WITH_CPUS;
2887 /* Slight preference for less loaded node */
2888 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2889 val += node_load[n];
2891 if (val < min_val) {
2898 node_set(best_node, *used_node_mask);
2905 * Build zonelists ordered by node and zones within node.
2906 * This results in maximum locality--normal zone overflows into local
2907 * DMA zone, if any--but risks exhausting DMA zone.
2909 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2912 struct zonelist *zonelist;
2914 zonelist = &pgdat->node_zonelists[0];
2915 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2917 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2919 zonelist->_zonerefs[j].zone = NULL;
2920 zonelist->_zonerefs[j].zone_idx = 0;
2924 * Build gfp_thisnode zonelists
2926 static void build_thisnode_zonelists(pg_data_t *pgdat)
2929 struct zonelist *zonelist;
2931 zonelist = &pgdat->node_zonelists[1];
2932 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2933 zonelist->_zonerefs[j].zone = NULL;
2934 zonelist->_zonerefs[j].zone_idx = 0;
2938 * Build zonelists ordered by zone and nodes within zones.
2939 * This results in conserving DMA zone[s] until all Normal memory is
2940 * exhausted, but results in overflowing to remote node while memory
2941 * may still exist in local DMA zone.
2943 static int node_order[MAX_NUMNODES];
2945 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2948 int zone_type; /* needs to be signed */
2950 struct zonelist *zonelist;
2952 zonelist = &pgdat->node_zonelists[0];
2954 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2955 for (j = 0; j < nr_nodes; j++) {
2956 node = node_order[j];
2957 z = &NODE_DATA(node)->node_zones[zone_type];
2958 if (populated_zone(z)) {
2960 &zonelist->_zonerefs[pos++]);
2961 check_highest_zone(zone_type);
2965 zonelist->_zonerefs[pos].zone = NULL;
2966 zonelist->_zonerefs[pos].zone_idx = 0;
2969 static int default_zonelist_order(void)
2972 unsigned long low_kmem_size,total_size;
2976 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2977 * If they are really small and used heavily, the system can fall
2978 * into OOM very easily.
2979 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2981 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2984 for_each_online_node(nid) {
2985 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2986 z = &NODE_DATA(nid)->node_zones[zone_type];
2987 if (populated_zone(z)) {
2988 if (zone_type < ZONE_NORMAL)
2989 low_kmem_size += z->present_pages;
2990 total_size += z->present_pages;
2991 } else if (zone_type == ZONE_NORMAL) {
2993 * If any node has only lowmem, then node order
2994 * is preferred to allow kernel allocations
2995 * locally; otherwise, they can easily infringe
2996 * on other nodes when there is an abundance of
2997 * lowmem available to allocate from.
2999 return ZONELIST_ORDER_NODE;
3003 if (!low_kmem_size || /* there are no DMA area. */
3004 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3005 return ZONELIST_ORDER_NODE;
3007 * look into each node's config.
3008 * If there is a node whose DMA/DMA32 memory is very big area on
3009 * local memory, NODE_ORDER may be suitable.
3011 average_size = total_size /
3012 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3013 for_each_online_node(nid) {
3016 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3017 z = &NODE_DATA(nid)->node_zones[zone_type];
3018 if (populated_zone(z)) {
3019 if (zone_type < ZONE_NORMAL)
3020 low_kmem_size += z->present_pages;
3021 total_size += z->present_pages;
3024 if (low_kmem_size &&
3025 total_size > average_size && /* ignore small node */
3026 low_kmem_size > total_size * 70/100)
3027 return ZONELIST_ORDER_NODE;
3029 return ZONELIST_ORDER_ZONE;
3032 static void set_zonelist_order(void)
3034 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3035 current_zonelist_order = default_zonelist_order();
3037 current_zonelist_order = user_zonelist_order;
3040 static void build_zonelists(pg_data_t *pgdat)
3044 nodemask_t used_mask;
3045 int local_node, prev_node;
3046 struct zonelist *zonelist;
3047 int order = current_zonelist_order;
3049 /* initialize zonelists */
3050 for (i = 0; i < MAX_ZONELISTS; i++) {
3051 zonelist = pgdat->node_zonelists + i;
3052 zonelist->_zonerefs[0].zone = NULL;
3053 zonelist->_zonerefs[0].zone_idx = 0;
3056 /* NUMA-aware ordering of nodes */
3057 local_node = pgdat->node_id;
3058 load = nr_online_nodes;
3059 prev_node = local_node;
3060 nodes_clear(used_mask);
3062 memset(node_order, 0, sizeof(node_order));
3065 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3066 int distance = node_distance(local_node, node);
3069 * If another node is sufficiently far away then it is better
3070 * to reclaim pages in a zone before going off node.
3072 if (distance > RECLAIM_DISTANCE)
3073 zone_reclaim_mode = 1;
3076 * We don't want to pressure a particular node.
3077 * So adding penalty to the first node in same
3078 * distance group to make it round-robin.
3080 if (distance != node_distance(local_node, prev_node))
3081 node_load[node] = load;
3085 if (order == ZONELIST_ORDER_NODE)
3086 build_zonelists_in_node_order(pgdat, node);
3088 node_order[j++] = node; /* remember order */
3091 if (order == ZONELIST_ORDER_ZONE) {
3092 /* calculate node order -- i.e., DMA last! */
3093 build_zonelists_in_zone_order(pgdat, j);
3096 build_thisnode_zonelists(pgdat);
3099 /* Construct the zonelist performance cache - see further mmzone.h */
3100 static void build_zonelist_cache(pg_data_t *pgdat)
3102 struct zonelist *zonelist;
3103 struct zonelist_cache *zlc;
3106 zonelist = &pgdat->node_zonelists[0];
3107 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3108 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3109 for (z = zonelist->_zonerefs; z->zone; z++)
3110 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3113 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3115 * Return node id of node used for "local" allocations.
3116 * I.e., first node id of first zone in arg node's generic zonelist.
3117 * Used for initializing percpu 'numa_mem', which is used primarily
3118 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3120 int local_memory_node(int node)
3124 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3125 gfp_zone(GFP_KERNEL),
3132 #else /* CONFIG_NUMA */
3134 static void set_zonelist_order(void)
3136 current_zonelist_order = ZONELIST_ORDER_ZONE;
3139 static void build_zonelists(pg_data_t *pgdat)
3141 int node, local_node;
3143 struct zonelist *zonelist;
3145 local_node = pgdat->node_id;
3147 zonelist = &pgdat->node_zonelists[0];
3148 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3151 * Now we build the zonelist so that it contains the zones
3152 * of all the other nodes.
3153 * We don't want to pressure a particular node, so when
3154 * building the zones for node N, we make sure that the
3155 * zones coming right after the local ones are those from
3156 * node N+1 (modulo N)
3158 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3159 if (!node_online(node))
3161 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3164 for (node = 0; node < local_node; node++) {
3165 if (!node_online(node))
3167 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3171 zonelist->_zonerefs[j].zone = NULL;
3172 zonelist->_zonerefs[j].zone_idx = 0;
3175 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3176 static void build_zonelist_cache(pg_data_t *pgdat)
3178 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3181 #endif /* CONFIG_NUMA */
3184 * Boot pageset table. One per cpu which is going to be used for all
3185 * zones and all nodes. The parameters will be set in such a way
3186 * that an item put on a list will immediately be handed over to
3187 * the buddy list. This is safe since pageset manipulation is done
3188 * with interrupts disabled.
3190 * The boot_pagesets must be kept even after bootup is complete for
3191 * unused processors and/or zones. They do play a role for bootstrapping
3192 * hotplugged processors.
3194 * zoneinfo_show() and maybe other functions do
3195 * not check if the processor is online before following the pageset pointer.
3196 * Other parts of the kernel may not check if the zone is available.
3198 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3199 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3200 static void setup_zone_pageset(struct zone *zone);
3203 * Global mutex to protect against size modification of zonelists
3204 * as well as to serialize pageset setup for the new populated zone.
3206 DEFINE_MUTEX(zonelists_mutex);
3208 /* return values int ....just for stop_machine() */
3209 static __init_refok int __build_all_zonelists(void *data)
3215 memset(node_load, 0, sizeof(node_load));
3217 for_each_online_node(nid) {
3218 pg_data_t *pgdat = NODE_DATA(nid);
3220 build_zonelists(pgdat);
3221 build_zonelist_cache(pgdat);
3225 * Initialize the boot_pagesets that are going to be used
3226 * for bootstrapping processors. The real pagesets for
3227 * each zone will be allocated later when the per cpu
3228 * allocator is available.
3230 * boot_pagesets are used also for bootstrapping offline
3231 * cpus if the system is already booted because the pagesets
3232 * are needed to initialize allocators on a specific cpu too.
3233 * F.e. the percpu allocator needs the page allocator which
3234 * needs the percpu allocator in order to allocate its pagesets
3235 * (a chicken-egg dilemma).
3237 for_each_possible_cpu(cpu) {
3238 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3240 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3242 * We now know the "local memory node" for each node--
3243 * i.e., the node of the first zone in the generic zonelist.
3244 * Set up numa_mem percpu variable for on-line cpus. During
3245 * boot, only the boot cpu should be on-line; we'll init the
3246 * secondary cpus' numa_mem as they come on-line. During
3247 * node/memory hotplug, we'll fixup all on-line cpus.
3249 if (cpu_online(cpu))
3250 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3258 * Called with zonelists_mutex held always
3259 * unless system_state == SYSTEM_BOOTING.
3261 void __ref build_all_zonelists(void *data)
3263 set_zonelist_order();
3265 if (system_state == SYSTEM_BOOTING) {
3266 __build_all_zonelists(NULL);
3267 mminit_verify_zonelist();
3268 cpuset_init_current_mems_allowed();
3270 /* we have to stop all cpus to guarantee there is no user
3272 #ifdef CONFIG_MEMORY_HOTPLUG
3274 setup_zone_pageset((struct zone *)data);
3276 stop_machine(__build_all_zonelists, NULL, NULL);
3277 /* cpuset refresh routine should be here */
3279 vm_total_pages = nr_free_pagecache_pages();
3281 * Disable grouping by mobility if the number of pages in the
3282 * system is too low to allow the mechanism to work. It would be
3283 * more accurate, but expensive to check per-zone. This check is
3284 * made on memory-hotadd so a system can start with mobility
3285 * disabled and enable it later
3287 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3288 page_group_by_mobility_disabled = 1;
3290 page_group_by_mobility_disabled = 0;
3292 printk("Built %i zonelists in %s order, mobility grouping %s. "
3293 "Total pages: %ld\n",
3295 zonelist_order_name[current_zonelist_order],
3296 page_group_by_mobility_disabled ? "off" : "on",
3299 printk("Policy zone: %s\n", zone_names[policy_zone]);
3304 * Helper functions to size the waitqueue hash table.
3305 * Essentially these want to choose hash table sizes sufficiently
3306 * large so that collisions trying to wait on pages are rare.
3307 * But in fact, the number of active page waitqueues on typical
3308 * systems is ridiculously low, less than 200. So this is even
3309 * conservative, even though it seems large.
3311 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3312 * waitqueues, i.e. the size of the waitq table given the number of pages.
3314 #define PAGES_PER_WAITQUEUE 256
3316 #ifndef CONFIG_MEMORY_HOTPLUG
3317 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3319 unsigned long size = 1;
3321 pages /= PAGES_PER_WAITQUEUE;
3323 while (size < pages)
3327 * Once we have dozens or even hundreds of threads sleeping
3328 * on IO we've got bigger problems than wait queue collision.
3329 * Limit the size of the wait table to a reasonable size.
3331 size = min(size, 4096UL);
3333 return max(size, 4UL);
3337 * A zone's size might be changed by hot-add, so it is not possible to determine
3338 * a suitable size for its wait_table. So we use the maximum size now.
3340 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3342 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3343 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3344 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3346 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3347 * or more by the traditional way. (See above). It equals:
3349 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3350 * ia64(16K page size) : = ( 8G + 4M)byte.
3351 * powerpc (64K page size) : = (32G +16M)byte.
3353 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3360 * This is an integer logarithm so that shifts can be used later
3361 * to extract the more random high bits from the multiplicative
3362 * hash function before the remainder is taken.
3364 static inline unsigned long wait_table_bits(unsigned long size)
3369 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3372 * Check if a pageblock contains reserved pages
3374 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3378 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3379 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3386 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3387 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3388 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3389 * higher will lead to a bigger reserve which will get freed as contiguous
3390 * blocks as reclaim kicks in
3392 static void setup_zone_migrate_reserve(struct zone *zone)
3394 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3396 unsigned long block_migratetype;
3400 * Get the start pfn, end pfn and the number of blocks to reserve
3401 * We have to be careful to be aligned to pageblock_nr_pages to
3402 * make sure that we always check pfn_valid for the first page in
3405 start_pfn = zone->zone_start_pfn;
3406 end_pfn = start_pfn + zone->spanned_pages;
3407 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3408 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3412 * Reserve blocks are generally in place to help high-order atomic
3413 * allocations that are short-lived. A min_free_kbytes value that
3414 * would result in more than 2 reserve blocks for atomic allocations
3415 * is assumed to be in place to help anti-fragmentation for the
3416 * future allocation of hugepages at runtime.
3418 reserve = min(2, reserve);
3420 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3421 if (!pfn_valid(pfn))
3423 page = pfn_to_page(pfn);
3425 /* Watch out for overlapping nodes */
3426 if (page_to_nid(page) != zone_to_nid(zone))
3429 /* Blocks with reserved pages will never free, skip them. */
3430 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3431 if (pageblock_is_reserved(pfn, block_end_pfn))
3434 block_migratetype = get_pageblock_migratetype(page);
3436 /* If this block is reserved, account for it */
3437 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3442 /* Suitable for reserving if this block is movable */
3443 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3444 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3445 move_freepages_block(zone, page, MIGRATE_RESERVE);
3451 * If the reserve is met and this is a previous reserved block,
3454 if (block_migratetype == MIGRATE_RESERVE) {
3455 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3456 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3462 * Initially all pages are reserved - free ones are freed
3463 * up by free_all_bootmem() once the early boot process is
3464 * done. Non-atomic initialization, single-pass.
3466 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3467 unsigned long start_pfn, enum memmap_context context)
3470 unsigned long end_pfn = start_pfn + size;
3474 if (highest_memmap_pfn < end_pfn - 1)
3475 highest_memmap_pfn = end_pfn - 1;
3477 z = &NODE_DATA(nid)->node_zones[zone];
3478 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3480 * There can be holes in boot-time mem_map[]s
3481 * handed to this function. They do not
3482 * exist on hotplugged memory.
3484 if (context == MEMMAP_EARLY) {
3485 if (!early_pfn_valid(pfn))
3487 if (!early_pfn_in_nid(pfn, nid))
3490 page = pfn_to_page(pfn);
3491 set_page_links(page, zone, nid, pfn);
3492 mminit_verify_page_links(page, zone, nid, pfn);
3493 init_page_count(page);
3494 reset_page_mapcount(page);
3495 SetPageReserved(page);
3497 * Mark the block movable so that blocks are reserved for
3498 * movable at startup. This will force kernel allocations
3499 * to reserve their blocks rather than leaking throughout
3500 * the address space during boot when many long-lived
3501 * kernel allocations are made. Later some blocks near
3502 * the start are marked MIGRATE_RESERVE by
3503 * setup_zone_migrate_reserve()
3505 * bitmap is created for zone's valid pfn range. but memmap
3506 * can be created for invalid pages (for alignment)
3507 * check here not to call set_pageblock_migratetype() against
3510 if ((z->zone_start_pfn <= pfn)
3511 && (pfn < z->zone_start_pfn + z->spanned_pages)
3512 && !(pfn & (pageblock_nr_pages - 1)))
3513 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3515 INIT_LIST_HEAD(&page->lru);
3516 #ifdef WANT_PAGE_VIRTUAL
3517 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3518 if (!is_highmem_idx(zone))
3519 set_page_address(page, __va(pfn << PAGE_SHIFT));
3524 static void __meminit zone_init_free_lists(struct zone *zone)
3527 for_each_migratetype_order(order, t) {
3528 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3529 zone->free_area[order].nr_free = 0;
3533 #ifndef __HAVE_ARCH_MEMMAP_INIT
3534 #define memmap_init(size, nid, zone, start_pfn) \
3535 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3538 static int zone_batchsize(struct zone *zone)
3544 * The per-cpu-pages pools are set to around 1000th of the
3545 * size of the zone. But no more than 1/2 of a meg.
3547 * OK, so we don't know how big the cache is. So guess.
3549 batch = zone->present_pages / 1024;
3550 if (batch * PAGE_SIZE > 512 * 1024)
3551 batch = (512 * 1024) / PAGE_SIZE;
3552 batch /= 4; /* We effectively *= 4 below */
3557 * Clamp the batch to a 2^n - 1 value. Having a power
3558 * of 2 value was found to be more likely to have
3559 * suboptimal cache aliasing properties in some cases.
3561 * For example if 2 tasks are alternately allocating
3562 * batches of pages, one task can end up with a lot
3563 * of pages of one half of the possible page colors
3564 * and the other with pages of the other colors.
3566 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3571 /* The deferral and batching of frees should be suppressed under NOMMU
3574 * The problem is that NOMMU needs to be able to allocate large chunks
3575 * of contiguous memory as there's no hardware page translation to
3576 * assemble apparent contiguous memory from discontiguous pages.
3578 * Queueing large contiguous runs of pages for batching, however,
3579 * causes the pages to actually be freed in smaller chunks. As there
3580 * can be a significant delay between the individual batches being
3581 * recycled, this leads to the once large chunks of space being
3582 * fragmented and becoming unavailable for high-order allocations.
3588 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3590 struct per_cpu_pages *pcp;
3593 memset(p, 0, sizeof(*p));
3597 pcp->high = 6 * batch;
3598 pcp->batch = max(1UL, 1 * batch);
3599 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3600 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3604 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3605 * to the value high for the pageset p.
3608 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3611 struct per_cpu_pages *pcp;
3615 pcp->batch = max(1UL, high/4);
3616 if ((high/4) > (PAGE_SHIFT * 8))
3617 pcp->batch = PAGE_SHIFT * 8;
3620 static void setup_zone_pageset(struct zone *zone)
3624 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3626 for_each_possible_cpu(cpu) {
3627 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3629 setup_pageset(pcp, zone_batchsize(zone));
3631 if (percpu_pagelist_fraction)
3632 setup_pagelist_highmark(pcp,
3633 (zone->present_pages /
3634 percpu_pagelist_fraction));
3639 * Allocate per cpu pagesets and initialize them.
3640 * Before this call only boot pagesets were available.
3642 void __init setup_per_cpu_pageset(void)
3646 for_each_populated_zone(zone)
3647 setup_zone_pageset(zone);
3650 static noinline __init_refok
3651 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3654 struct pglist_data *pgdat = zone->zone_pgdat;
3658 * The per-page waitqueue mechanism uses hashed waitqueues
3661 zone->wait_table_hash_nr_entries =
3662 wait_table_hash_nr_entries(zone_size_pages);
3663 zone->wait_table_bits =
3664 wait_table_bits(zone->wait_table_hash_nr_entries);
3665 alloc_size = zone->wait_table_hash_nr_entries
3666 * sizeof(wait_queue_head_t);
3668 if (!slab_is_available()) {
3669 zone->wait_table = (wait_queue_head_t *)
3670 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3673 * This case means that a zone whose size was 0 gets new memory
3674 * via memory hot-add.
3675 * But it may be the case that a new node was hot-added. In
3676 * this case vmalloc() will not be able to use this new node's
3677 * memory - this wait_table must be initialized to use this new
3678 * node itself as well.
3679 * To use this new node's memory, further consideration will be
3682 zone->wait_table = vmalloc(alloc_size);
3684 if (!zone->wait_table)
3687 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3688 init_waitqueue_head(zone->wait_table + i);
3693 static int __zone_pcp_update(void *data)
3695 struct zone *zone = data;
3697 unsigned long batch = zone_batchsize(zone), flags;
3699 for_each_possible_cpu(cpu) {
3700 struct per_cpu_pageset *pset;
3701 struct per_cpu_pages *pcp;
3703 pset = per_cpu_ptr(zone->pageset, cpu);
3706 local_irq_save(flags);
3707 free_pcppages_bulk(zone, pcp->count, pcp);
3708 setup_pageset(pset, batch);
3709 local_irq_restore(flags);
3714 void zone_pcp_update(struct zone *zone)
3716 stop_machine(__zone_pcp_update, zone, NULL);
3719 static __meminit void zone_pcp_init(struct zone *zone)
3722 * per cpu subsystem is not up at this point. The following code
3723 * relies on the ability of the linker to provide the
3724 * offset of a (static) per cpu variable into the per cpu area.
3726 zone->pageset = &boot_pageset;
3728 if (zone->present_pages)
3729 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3730 zone->name, zone->present_pages,
3731 zone_batchsize(zone));
3734 __meminit int init_currently_empty_zone(struct zone *zone,
3735 unsigned long zone_start_pfn,
3737 enum memmap_context context)
3739 struct pglist_data *pgdat = zone->zone_pgdat;
3741 ret = zone_wait_table_init(zone, size);
3744 pgdat->nr_zones = zone_idx(zone) + 1;
3746 zone->zone_start_pfn = zone_start_pfn;
3748 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3749 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3751 (unsigned long)zone_idx(zone),
3752 zone_start_pfn, (zone_start_pfn + size));
3754 zone_init_free_lists(zone);
3759 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3761 * Basic iterator support. Return the first range of PFNs for a node
3762 * Note: nid == MAX_NUMNODES returns first region regardless of node
3764 static int __meminit first_active_region_index_in_nid(int nid)
3768 for (i = 0; i < nr_nodemap_entries; i++)
3769 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3776 * Basic iterator support. Return the next active range of PFNs for a node
3777 * Note: nid == MAX_NUMNODES returns next region regardless of node
3779 static int __meminit next_active_region_index_in_nid(int index, int nid)
3781 for (index = index + 1; index < nr_nodemap_entries; index++)
3782 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3788 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3790 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3791 * Architectures may implement their own version but if add_active_range()
3792 * was used and there are no special requirements, this is a convenient
3795 int __meminit __early_pfn_to_nid(unsigned long pfn)
3799 for (i = 0; i < nr_nodemap_entries; i++) {
3800 unsigned long start_pfn = early_node_map[i].start_pfn;
3801 unsigned long end_pfn = early_node_map[i].end_pfn;
3803 if (start_pfn <= pfn && pfn < end_pfn)
3804 return early_node_map[i].nid;
3806 /* This is a memory hole */
3809 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3811 int __meminit early_pfn_to_nid(unsigned long pfn)
3815 nid = __early_pfn_to_nid(pfn);
3818 /* just returns 0 */
3822 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3823 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3827 nid = __early_pfn_to_nid(pfn);
3828 if (nid >= 0 && nid != node)
3834 /* Basic iterator support to walk early_node_map[] */
3835 #define for_each_active_range_index_in_nid(i, nid) \
3836 for (i = first_active_region_index_in_nid(nid); i != -1; \
3837 i = next_active_region_index_in_nid(i, nid))
3840 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3841 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3842 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3844 * If an architecture guarantees that all ranges registered with
3845 * add_active_ranges() contain no holes and may be freed, this
3846 * this function may be used instead of calling free_bootmem() manually.
3848 void __init free_bootmem_with_active_regions(int nid,
3849 unsigned long max_low_pfn)
3853 for_each_active_range_index_in_nid(i, nid) {
3854 unsigned long size_pages = 0;
3855 unsigned long end_pfn = early_node_map[i].end_pfn;
3857 if (early_node_map[i].start_pfn >= max_low_pfn)
3860 if (end_pfn > max_low_pfn)
3861 end_pfn = max_low_pfn;
3863 size_pages = end_pfn - early_node_map[i].start_pfn;
3864 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3865 PFN_PHYS(early_node_map[i].start_pfn),
3866 size_pages << PAGE_SHIFT);
3870 #ifdef CONFIG_HAVE_MEMBLOCK
3872 * Basic iterator support. Return the last range of PFNs for a node
3873 * Note: nid == MAX_NUMNODES returns last region regardless of node
3875 static int __meminit last_active_region_index_in_nid(int nid)
3879 for (i = nr_nodemap_entries - 1; i >= 0; i--)
3880 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3887 * Basic iterator support. Return the previous active range of PFNs for a node
3888 * Note: nid == MAX_NUMNODES returns next region regardless of node
3890 static int __meminit previous_active_region_index_in_nid(int index, int nid)
3892 for (index = index - 1; index >= 0; index--)
3893 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3899 #define for_each_active_range_index_in_nid_reverse(i, nid) \
3900 for (i = last_active_region_index_in_nid(nid); i != -1; \
3901 i = previous_active_region_index_in_nid(i, nid))
3903 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3904 u64 goal, u64 limit)
3908 /* Need to go over early_node_map to find out good range for node */
3909 for_each_active_range_index_in_nid_reverse(i, nid) {
3911 u64 ei_start, ei_last;
3912 u64 final_start, final_end;
3914 ei_last = early_node_map[i].end_pfn;
3915 ei_last <<= PAGE_SHIFT;
3916 ei_start = early_node_map[i].start_pfn;
3917 ei_start <<= PAGE_SHIFT;
3919 final_start = max(ei_start, goal);
3920 final_end = min(ei_last, limit);
3922 if (final_start >= final_end)
3925 addr = memblock_find_in_range(final_start, final_end, size, align);
3927 if (addr == MEMBLOCK_ERROR)
3933 return MEMBLOCK_ERROR;
3937 int __init add_from_early_node_map(struct range *range, int az,
3938 int nr_range, int nid)
3943 /* need to go over early_node_map to find out good range for node */
3944 for_each_active_range_index_in_nid(i, nid) {
3945 start = early_node_map[i].start_pfn;
3946 end = early_node_map[i].end_pfn;
3947 nr_range = add_range(range, az, nr_range, start, end);
3952 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3957 for_each_active_range_index_in_nid(i, nid) {
3958 ret = work_fn(early_node_map[i].start_pfn,
3959 early_node_map[i].end_pfn, data);
3965 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3966 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3968 * If an architecture guarantees that all ranges registered with
3969 * add_active_ranges() contain no holes and may be freed, this
3970 * function may be used instead of calling memory_present() manually.
3972 void __init sparse_memory_present_with_active_regions(int nid)
3976 for_each_active_range_index_in_nid(i, nid)
3977 memory_present(early_node_map[i].nid,
3978 early_node_map[i].start_pfn,
3979 early_node_map[i].end_pfn);
3983 * get_pfn_range_for_nid - Return the start and end page frames for a node
3984 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3985 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3986 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3988 * It returns the start and end page frame of a node based on information
3989 * provided by an arch calling add_active_range(). If called for a node
3990 * with no available memory, a warning is printed and the start and end
3993 void __meminit get_pfn_range_for_nid(unsigned int nid,
3994 unsigned long *start_pfn, unsigned long *end_pfn)
4000 for_each_active_range_index_in_nid(i, nid) {
4001 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
4002 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
4005 if (*start_pfn == -1UL)
4010 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4011 * assumption is made that zones within a node are ordered in monotonic
4012 * increasing memory addresses so that the "highest" populated zone is used
4014 static void __init find_usable_zone_for_movable(void)
4017 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4018 if (zone_index == ZONE_MOVABLE)
4021 if (arch_zone_highest_possible_pfn[zone_index] >
4022 arch_zone_lowest_possible_pfn[zone_index])
4026 VM_BUG_ON(zone_index == -1);
4027 movable_zone = zone_index;
4031 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4032 * because it is sized independent of architecture. Unlike the other zones,
4033 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4034 * in each node depending on the size of each node and how evenly kernelcore
4035 * is distributed. This helper function adjusts the zone ranges
4036 * provided by the architecture for a given node by using the end of the
4037 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4038 * zones within a node are in order of monotonic increases memory addresses
4040 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4041 unsigned long zone_type,
4042 unsigned long node_start_pfn,
4043 unsigned long node_end_pfn,
4044 unsigned long *zone_start_pfn,
4045 unsigned long *zone_end_pfn)
4047 /* Only adjust if ZONE_MOVABLE is on this node */
4048 if (zone_movable_pfn[nid]) {
4049 /* Size ZONE_MOVABLE */
4050 if (zone_type == ZONE_MOVABLE) {
4051 *zone_start_pfn = zone_movable_pfn[nid];
4052 *zone_end_pfn = min(node_end_pfn,
4053 arch_zone_highest_possible_pfn[movable_zone]);
4055 /* Adjust for ZONE_MOVABLE starting within this range */
4056 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4057 *zone_end_pfn > zone_movable_pfn[nid]) {
4058 *zone_end_pfn = zone_movable_pfn[nid];
4060 /* Check if this whole range is within ZONE_MOVABLE */
4061 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4062 *zone_start_pfn = *zone_end_pfn;
4067 * Return the number of pages a zone spans in a node, including holes
4068 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4070 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4071 unsigned long zone_type,
4072 unsigned long *ignored)
4074 unsigned long node_start_pfn, node_end_pfn;
4075 unsigned long zone_start_pfn, zone_end_pfn;
4077 /* Get the start and end of the node and zone */
4078 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4079 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4080 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4081 adjust_zone_range_for_zone_movable(nid, zone_type,
4082 node_start_pfn, node_end_pfn,
4083 &zone_start_pfn, &zone_end_pfn);
4085 /* Check that this node has pages within the zone's required range */
4086 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4089 /* Move the zone boundaries inside the node if necessary */
4090 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4091 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4093 /* Return the spanned pages */
4094 return zone_end_pfn - zone_start_pfn;
4098 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4099 * then all holes in the requested range will be accounted for.
4101 unsigned long __meminit __absent_pages_in_range(int nid,
4102 unsigned long range_start_pfn,
4103 unsigned long range_end_pfn)
4106 unsigned long prev_end_pfn = 0, hole_pages = 0;
4107 unsigned long start_pfn;
4109 /* Find the end_pfn of the first active range of pfns in the node */
4110 i = first_active_region_index_in_nid(nid);
4114 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4116 /* Account for ranges before physical memory on this node */
4117 if (early_node_map[i].start_pfn > range_start_pfn)
4118 hole_pages = prev_end_pfn - range_start_pfn;
4120 /* Find all holes for the zone within the node */
4121 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4123 /* No need to continue if prev_end_pfn is outside the zone */
4124 if (prev_end_pfn >= range_end_pfn)
4127 /* Make sure the end of the zone is not within the hole */
4128 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4129 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4131 /* Update the hole size cound and move on */
4132 if (start_pfn > range_start_pfn) {
4133 BUG_ON(prev_end_pfn > start_pfn);
4134 hole_pages += start_pfn - prev_end_pfn;
4136 prev_end_pfn = early_node_map[i].end_pfn;
4139 /* Account for ranges past physical memory on this node */
4140 if (range_end_pfn > prev_end_pfn)
4141 hole_pages += range_end_pfn -
4142 max(range_start_pfn, prev_end_pfn);
4148 * absent_pages_in_range - Return number of page frames in holes within a range
4149 * @start_pfn: The start PFN to start searching for holes
4150 * @end_pfn: The end PFN to stop searching for holes
4152 * It returns the number of pages frames in memory holes within a range.
4154 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4155 unsigned long end_pfn)
4157 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4160 /* Return the number of page frames in holes in a zone on a node */
4161 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4162 unsigned long zone_type,
4163 unsigned long *ignored)
4165 unsigned long node_start_pfn, node_end_pfn;
4166 unsigned long zone_start_pfn, zone_end_pfn;
4168 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4169 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4171 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4174 adjust_zone_range_for_zone_movable(nid, zone_type,
4175 node_start_pfn, node_end_pfn,
4176 &zone_start_pfn, &zone_end_pfn);
4177 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4181 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4182 unsigned long zone_type,
4183 unsigned long *zones_size)
4185 return zones_size[zone_type];
4188 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4189 unsigned long zone_type,
4190 unsigned long *zholes_size)
4195 return zholes_size[zone_type];
4200 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4201 unsigned long *zones_size, unsigned long *zholes_size)
4203 unsigned long realtotalpages, totalpages = 0;
4206 for (i = 0; i < MAX_NR_ZONES; i++)
4207 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4209 pgdat->node_spanned_pages = totalpages;
4211 realtotalpages = totalpages;
4212 for (i = 0; i < MAX_NR_ZONES; i++)
4214 zone_absent_pages_in_node(pgdat->node_id, i,
4216 pgdat->node_present_pages = realtotalpages;
4217 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4221 #ifndef CONFIG_SPARSEMEM
4223 * Calculate the size of the zone->blockflags rounded to an unsigned long
4224 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4225 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4226 * round what is now in bits to nearest long in bits, then return it in
4229 static unsigned long __init usemap_size(unsigned long zonesize)
4231 unsigned long usemapsize;
4233 usemapsize = roundup(zonesize, pageblock_nr_pages);
4234 usemapsize = usemapsize >> pageblock_order;
4235 usemapsize *= NR_PAGEBLOCK_BITS;
4236 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4238 return usemapsize / 8;
4241 static void __init setup_usemap(struct pglist_data *pgdat,
4242 struct zone *zone, unsigned long zonesize)
4244 unsigned long usemapsize = usemap_size(zonesize);
4245 zone->pageblock_flags = NULL;
4247 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4251 static inline void setup_usemap(struct pglist_data *pgdat,
4252 struct zone *zone, unsigned long zonesize) {}
4253 #endif /* CONFIG_SPARSEMEM */
4255 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4257 /* Return a sensible default order for the pageblock size. */
4258 static inline int pageblock_default_order(void)
4260 if (HPAGE_SHIFT > PAGE_SHIFT)
4261 return HUGETLB_PAGE_ORDER;
4266 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4267 static inline void __init set_pageblock_order(unsigned int order)
4269 /* Check that pageblock_nr_pages has not already been setup */
4270 if (pageblock_order)
4274 * Assume the largest contiguous order of interest is a huge page.
4275 * This value may be variable depending on boot parameters on IA64
4277 pageblock_order = order;
4279 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4282 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4283 * and pageblock_default_order() are unused as pageblock_order is set
4284 * at compile-time. See include/linux/pageblock-flags.h for the values of
4285 * pageblock_order based on the kernel config
4287 static inline int pageblock_default_order(unsigned int order)
4291 #define set_pageblock_order(x) do {} while (0)
4293 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4296 * Set up the zone data structures:
4297 * - mark all pages reserved
4298 * - mark all memory queues empty
4299 * - clear the memory bitmaps
4301 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4302 unsigned long *zones_size, unsigned long *zholes_size)
4305 int nid = pgdat->node_id;
4306 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4309 pgdat_resize_init(pgdat);
4310 pgdat->nr_zones = 0;
4311 init_waitqueue_head(&pgdat->kswapd_wait);
4312 pgdat->kswapd_max_order = 0;
4313 pgdat_page_cgroup_init(pgdat);
4315 for (j = 0; j < MAX_NR_ZONES; j++) {
4316 struct zone *zone = pgdat->node_zones + j;
4317 unsigned long size, realsize, memmap_pages;
4320 size = zone_spanned_pages_in_node(nid, j, zones_size);
4321 realsize = size - zone_absent_pages_in_node(nid, j,
4325 * Adjust realsize so that it accounts for how much memory
4326 * is used by this zone for memmap. This affects the watermark
4327 * and per-cpu initialisations
4330 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4331 if (realsize >= memmap_pages) {
4332 realsize -= memmap_pages;
4335 " %s zone: %lu pages used for memmap\n",
4336 zone_names[j], memmap_pages);
4339 " %s zone: %lu pages exceeds realsize %lu\n",
4340 zone_names[j], memmap_pages, realsize);
4342 /* Account for reserved pages */
4343 if (j == 0 && realsize > dma_reserve) {
4344 realsize -= dma_reserve;
4345 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4346 zone_names[0], dma_reserve);
4349 if (!is_highmem_idx(j))
4350 nr_kernel_pages += realsize;
4351 nr_all_pages += realsize;
4353 zone->spanned_pages = size;
4354 zone->present_pages = realsize;
4357 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4359 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4361 zone->name = zone_names[j];
4362 spin_lock_init(&zone->lock);
4363 spin_lock_init(&zone->lru_lock);
4364 zone_seqlock_init(zone);
4365 zone->zone_pgdat = pgdat;
4367 zone_pcp_init(zone);
4369 INIT_LIST_HEAD(&zone->lru[l].list);
4370 zone->reclaim_stat.recent_rotated[0] = 0;
4371 zone->reclaim_stat.recent_rotated[1] = 0;
4372 zone->reclaim_stat.recent_scanned[0] = 0;
4373 zone->reclaim_stat.recent_scanned[1] = 0;
4374 zap_zone_vm_stats(zone);
4379 set_pageblock_order(pageblock_default_order());
4380 setup_usemap(pgdat, zone, size);
4381 ret = init_currently_empty_zone(zone, zone_start_pfn,
4382 size, MEMMAP_EARLY);
4384 memmap_init(size, nid, j, zone_start_pfn);
4385 zone_start_pfn += size;
4389 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4391 /* Skip empty nodes */
4392 if (!pgdat->node_spanned_pages)
4395 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4396 /* ia64 gets its own node_mem_map, before this, without bootmem */
4397 if (!pgdat->node_mem_map) {
4398 unsigned long size, start, end;
4402 * The zone's endpoints aren't required to be MAX_ORDER
4403 * aligned but the node_mem_map endpoints must be in order
4404 * for the buddy allocator to function correctly.
4406 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4407 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4408 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4409 size = (end - start) * sizeof(struct page);
4410 map = alloc_remap(pgdat->node_id, size);
4412 map = alloc_bootmem_node_nopanic(pgdat, size);
4413 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4415 #ifndef CONFIG_NEED_MULTIPLE_NODES
4417 * With no DISCONTIG, the global mem_map is just set as node 0's
4419 if (pgdat == NODE_DATA(0)) {
4420 mem_map = NODE_DATA(0)->node_mem_map;
4421 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4422 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4423 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4424 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4427 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4430 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4431 unsigned long node_start_pfn, unsigned long *zholes_size)
4433 pg_data_t *pgdat = NODE_DATA(nid);
4435 pgdat->node_id = nid;
4436 pgdat->node_start_pfn = node_start_pfn;
4437 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4439 alloc_node_mem_map(pgdat);
4440 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4441 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4442 nid, (unsigned long)pgdat,
4443 (unsigned long)pgdat->node_mem_map);
4446 free_area_init_core(pgdat, zones_size, zholes_size);
4449 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4451 #if MAX_NUMNODES > 1
4453 * Figure out the number of possible node ids.
4455 static void __init setup_nr_node_ids(void)
4458 unsigned int highest = 0;
4460 for_each_node_mask(node, node_possible_map)
4462 nr_node_ids = highest + 1;
4465 static inline void setup_nr_node_ids(void)
4471 * add_active_range - Register a range of PFNs backed by physical memory
4472 * @nid: The node ID the range resides on
4473 * @start_pfn: The start PFN of the available physical memory
4474 * @end_pfn: The end PFN of the available physical memory
4476 * These ranges are stored in an early_node_map[] and later used by
4477 * free_area_init_nodes() to calculate zone sizes and holes. If the
4478 * range spans a memory hole, it is up to the architecture to ensure
4479 * the memory is not freed by the bootmem allocator. If possible
4480 * the range being registered will be merged with existing ranges.
4482 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4483 unsigned long end_pfn)
4487 mminit_dprintk(MMINIT_TRACE, "memory_register",
4488 "Entering add_active_range(%d, %#lx, %#lx) "
4489 "%d entries of %d used\n",
4490 nid, start_pfn, end_pfn,
4491 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4493 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4495 /* Merge with existing active regions if possible */
4496 for (i = 0; i < nr_nodemap_entries; i++) {
4497 if (early_node_map[i].nid != nid)
4500 /* Skip if an existing region covers this new one */
4501 if (start_pfn >= early_node_map[i].start_pfn &&
4502 end_pfn <= early_node_map[i].end_pfn)
4505 /* Merge forward if suitable */
4506 if (start_pfn <= early_node_map[i].end_pfn &&
4507 end_pfn > early_node_map[i].end_pfn) {
4508 early_node_map[i].end_pfn = end_pfn;
4512 /* Merge backward if suitable */
4513 if (start_pfn < early_node_map[i].start_pfn &&
4514 end_pfn >= early_node_map[i].start_pfn) {
4515 early_node_map[i].start_pfn = start_pfn;
4520 /* Check that early_node_map is large enough */
4521 if (i >= MAX_ACTIVE_REGIONS) {
4522 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4523 MAX_ACTIVE_REGIONS);
4527 early_node_map[i].nid = nid;
4528 early_node_map[i].start_pfn = start_pfn;
4529 early_node_map[i].end_pfn = end_pfn;
4530 nr_nodemap_entries = i + 1;
4534 * remove_active_range - Shrink an existing registered range of PFNs
4535 * @nid: The node id the range is on that should be shrunk
4536 * @start_pfn: The new PFN of the range
4537 * @end_pfn: The new PFN of the range
4539 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4540 * The map is kept near the end physical page range that has already been
4541 * registered. This function allows an arch to shrink an existing registered
4544 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4545 unsigned long end_pfn)
4550 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4551 nid, start_pfn, end_pfn);
4553 /* Find the old active region end and shrink */
4554 for_each_active_range_index_in_nid(i, nid) {
4555 if (early_node_map[i].start_pfn >= start_pfn &&
4556 early_node_map[i].end_pfn <= end_pfn) {
4558 early_node_map[i].start_pfn = 0;
4559 early_node_map[i].end_pfn = 0;
4563 if (early_node_map[i].start_pfn < start_pfn &&
4564 early_node_map[i].end_pfn > start_pfn) {
4565 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4566 early_node_map[i].end_pfn = start_pfn;
4567 if (temp_end_pfn > end_pfn)
4568 add_active_range(nid, end_pfn, temp_end_pfn);
4571 if (early_node_map[i].start_pfn >= start_pfn &&
4572 early_node_map[i].end_pfn > end_pfn &&
4573 early_node_map[i].start_pfn < end_pfn) {
4574 early_node_map[i].start_pfn = end_pfn;
4582 /* remove the blank ones */
4583 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4584 if (early_node_map[i].nid != nid)
4586 if (early_node_map[i].end_pfn)
4588 /* we found it, get rid of it */
4589 for (j = i; j < nr_nodemap_entries - 1; j++)
4590 memcpy(&early_node_map[j], &early_node_map[j+1],
4591 sizeof(early_node_map[j]));
4592 j = nr_nodemap_entries - 1;
4593 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4594 nr_nodemap_entries--;
4599 * remove_all_active_ranges - Remove all currently registered regions
4601 * During discovery, it may be found that a table like SRAT is invalid
4602 * and an alternative discovery method must be used. This function removes
4603 * all currently registered regions.
4605 void __init remove_all_active_ranges(void)
4607 memset(early_node_map, 0, sizeof(early_node_map));
4608 nr_nodemap_entries = 0;
4611 /* Compare two active node_active_regions */
4612 static int __init cmp_node_active_region(const void *a, const void *b)
4614 struct node_active_region *arange = (struct node_active_region *)a;
4615 struct node_active_region *brange = (struct node_active_region *)b;
4617 /* Done this way to avoid overflows */
4618 if (arange->start_pfn > brange->start_pfn)
4620 if (arange->start_pfn < brange->start_pfn)
4626 /* sort the node_map by start_pfn */
4627 void __init sort_node_map(void)
4629 sort(early_node_map, (size_t)nr_nodemap_entries,
4630 sizeof(struct node_active_region),
4631 cmp_node_active_region, NULL);
4635 * node_map_pfn_alignment - determine the maximum internode alignment
4637 * This function should be called after node map is populated and sorted.
4638 * It calculates the maximum power of two alignment which can distinguish
4641 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4642 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4643 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4644 * shifted, 1GiB is enough and this function will indicate so.
4646 * This is used to test whether pfn -> nid mapping of the chosen memory
4647 * model has fine enough granularity to avoid incorrect mapping for the
4648 * populated node map.
4650 * Returns the determined alignment in pfn's. 0 if there is no alignment
4651 * requirement (single node).
4653 unsigned long __init node_map_pfn_alignment(void)
4655 unsigned long accl_mask = 0, last_end = 0;
4659 for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4660 int nid = early_node_map[i].nid;
4661 unsigned long start = early_node_map[i].start_pfn;
4662 unsigned long end = early_node_map[i].end_pfn;
4665 if (!start || last_nid < 0 || last_nid == nid) {
4672 * Start with a mask granular enough to pin-point to the
4673 * start pfn and tick off bits one-by-one until it becomes
4674 * too coarse to separate the current node from the last.
4676 mask = ~((1 << __ffs(start)) - 1);
4677 while (mask && last_end <= (start & (mask << 1)))
4680 /* accumulate all internode masks */
4684 /* convert mask to number of pages */
4685 return ~accl_mask + 1;
4688 /* Find the lowest pfn for a node */
4689 static unsigned long __init find_min_pfn_for_node(int nid)
4692 unsigned long min_pfn = ULONG_MAX;
4694 /* Assuming a sorted map, the first range found has the starting pfn */
4695 for_each_active_range_index_in_nid(i, nid)
4696 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4698 if (min_pfn == ULONG_MAX) {
4700 "Could not find start_pfn for node %d\n", nid);
4708 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4710 * It returns the minimum PFN based on information provided via
4711 * add_active_range().
4713 unsigned long __init find_min_pfn_with_active_regions(void)
4715 return find_min_pfn_for_node(MAX_NUMNODES);
4719 * early_calculate_totalpages()
4720 * Sum pages in active regions for movable zone.
4721 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4723 static unsigned long __init early_calculate_totalpages(void)
4726 unsigned long totalpages = 0;
4728 for (i = 0; i < nr_nodemap_entries; i++) {
4729 unsigned long pages = early_node_map[i].end_pfn -
4730 early_node_map[i].start_pfn;
4731 totalpages += pages;
4733 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4739 * Find the PFN the Movable zone begins in each node. Kernel memory
4740 * is spread evenly between nodes as long as the nodes have enough
4741 * memory. When they don't, some nodes will have more kernelcore than
4744 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4747 unsigned long usable_startpfn;
4748 unsigned long kernelcore_node, kernelcore_remaining;
4749 /* save the state before borrow the nodemask */
4750 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4751 unsigned long totalpages = early_calculate_totalpages();
4752 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4755 * If movablecore was specified, calculate what size of
4756 * kernelcore that corresponds so that memory usable for
4757 * any allocation type is evenly spread. If both kernelcore
4758 * and movablecore are specified, then the value of kernelcore
4759 * will be used for required_kernelcore if it's greater than
4760 * what movablecore would have allowed.
4762 if (required_movablecore) {
4763 unsigned long corepages;
4766 * Round-up so that ZONE_MOVABLE is at least as large as what
4767 * was requested by the user
4769 required_movablecore =
4770 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4771 corepages = totalpages - required_movablecore;
4773 required_kernelcore = max(required_kernelcore, corepages);
4776 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4777 if (!required_kernelcore)
4780 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4781 find_usable_zone_for_movable();
4782 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4785 /* Spread kernelcore memory as evenly as possible throughout nodes */
4786 kernelcore_node = required_kernelcore / usable_nodes;
4787 for_each_node_state(nid, N_HIGH_MEMORY) {
4789 * Recalculate kernelcore_node if the division per node
4790 * now exceeds what is necessary to satisfy the requested
4791 * amount of memory for the kernel
4793 if (required_kernelcore < kernelcore_node)
4794 kernelcore_node = required_kernelcore / usable_nodes;
4797 * As the map is walked, we track how much memory is usable
4798 * by the kernel using kernelcore_remaining. When it is
4799 * 0, the rest of the node is usable by ZONE_MOVABLE
4801 kernelcore_remaining = kernelcore_node;
4803 /* Go through each range of PFNs within this node */
4804 for_each_active_range_index_in_nid(i, nid) {
4805 unsigned long start_pfn, end_pfn;
4806 unsigned long size_pages;
4808 start_pfn = max(early_node_map[i].start_pfn,
4809 zone_movable_pfn[nid]);
4810 end_pfn = early_node_map[i].end_pfn;
4811 if (start_pfn >= end_pfn)
4814 /* Account for what is only usable for kernelcore */
4815 if (start_pfn < usable_startpfn) {
4816 unsigned long kernel_pages;
4817 kernel_pages = min(end_pfn, usable_startpfn)
4820 kernelcore_remaining -= min(kernel_pages,
4821 kernelcore_remaining);
4822 required_kernelcore -= min(kernel_pages,
4823 required_kernelcore);
4825 /* Continue if range is now fully accounted */
4826 if (end_pfn <= usable_startpfn) {
4829 * Push zone_movable_pfn to the end so
4830 * that if we have to rebalance
4831 * kernelcore across nodes, we will
4832 * not double account here
4834 zone_movable_pfn[nid] = end_pfn;
4837 start_pfn = usable_startpfn;
4841 * The usable PFN range for ZONE_MOVABLE is from
4842 * start_pfn->end_pfn. Calculate size_pages as the
4843 * number of pages used as kernelcore
4845 size_pages = end_pfn - start_pfn;
4846 if (size_pages > kernelcore_remaining)
4847 size_pages = kernelcore_remaining;
4848 zone_movable_pfn[nid] = start_pfn + size_pages;
4851 * Some kernelcore has been met, update counts and
4852 * break if the kernelcore for this node has been
4855 required_kernelcore -= min(required_kernelcore,
4857 kernelcore_remaining -= size_pages;
4858 if (!kernelcore_remaining)
4864 * If there is still required_kernelcore, we do another pass with one
4865 * less node in the count. This will push zone_movable_pfn[nid] further
4866 * along on the nodes that still have memory until kernelcore is
4870 if (usable_nodes && required_kernelcore > usable_nodes)
4873 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4874 for (nid = 0; nid < MAX_NUMNODES; nid++)
4875 zone_movable_pfn[nid] =
4876 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4879 /* restore the node_state */
4880 node_states[N_HIGH_MEMORY] = saved_node_state;
4883 /* Any regular memory on that node ? */
4884 static void check_for_regular_memory(pg_data_t *pgdat)
4886 #ifdef CONFIG_HIGHMEM
4887 enum zone_type zone_type;
4889 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4890 struct zone *zone = &pgdat->node_zones[zone_type];
4891 if (zone->present_pages)
4892 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4898 * free_area_init_nodes - Initialise all pg_data_t and zone data
4899 * @max_zone_pfn: an array of max PFNs for each zone
4901 * This will call free_area_init_node() for each active node in the system.
4902 * Using the page ranges provided by add_active_range(), the size of each
4903 * zone in each node and their holes is calculated. If the maximum PFN
4904 * between two adjacent zones match, it is assumed that the zone is empty.
4905 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4906 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4907 * starts where the previous one ended. For example, ZONE_DMA32 starts
4908 * at arch_max_dma_pfn.
4910 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4915 /* Sort early_node_map as initialisation assumes it is sorted */
4918 /* Record where the zone boundaries are */
4919 memset(arch_zone_lowest_possible_pfn, 0,
4920 sizeof(arch_zone_lowest_possible_pfn));
4921 memset(arch_zone_highest_possible_pfn, 0,
4922 sizeof(arch_zone_highest_possible_pfn));
4923 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4924 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4925 for (i = 1; i < MAX_NR_ZONES; i++) {
4926 if (i == ZONE_MOVABLE)
4928 arch_zone_lowest_possible_pfn[i] =
4929 arch_zone_highest_possible_pfn[i-1];
4930 arch_zone_highest_possible_pfn[i] =
4931 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4933 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4934 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4936 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4937 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4938 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4940 /* Print out the zone ranges */
4941 printk("Zone PFN ranges:\n");
4942 for (i = 0; i < MAX_NR_ZONES; i++) {
4943 if (i == ZONE_MOVABLE)
4945 printk(" %-8s ", zone_names[i]);
4946 if (arch_zone_lowest_possible_pfn[i] ==
4947 arch_zone_highest_possible_pfn[i])
4950 printk("%0#10lx -> %0#10lx\n",
4951 arch_zone_lowest_possible_pfn[i],
4952 arch_zone_highest_possible_pfn[i]);
4955 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4956 printk("Movable zone start PFN for each node\n");
4957 for (i = 0; i < MAX_NUMNODES; i++) {
4958 if (zone_movable_pfn[i])
4959 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4962 /* Print out the early_node_map[] */
4963 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4964 for (i = 0; i < nr_nodemap_entries; i++)
4965 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4966 early_node_map[i].start_pfn,
4967 early_node_map[i].end_pfn);
4969 /* Initialise every node */
4970 mminit_verify_pageflags_layout();
4971 setup_nr_node_ids();
4972 for_each_online_node(nid) {
4973 pg_data_t *pgdat = NODE_DATA(nid);
4974 free_area_init_node(nid, NULL,
4975 find_min_pfn_for_node(nid), NULL);
4977 /* Any memory on that node */
4978 if (pgdat->node_present_pages)
4979 node_set_state(nid, N_HIGH_MEMORY);
4980 check_for_regular_memory(pgdat);
4984 static int __init cmdline_parse_core(char *p, unsigned long *core)
4986 unsigned long long coremem;
4990 coremem = memparse(p, &p);
4991 *core = coremem >> PAGE_SHIFT;
4993 /* Paranoid check that UL is enough for the coremem value */
4994 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5000 * kernelcore=size sets the amount of memory for use for allocations that
5001 * cannot be reclaimed or migrated.
5003 static int __init cmdline_parse_kernelcore(char *p)
5005 return cmdline_parse_core(p, &required_kernelcore);
5009 * movablecore=size sets the amount of memory for use for allocations that
5010 * can be reclaimed or migrated.
5012 static int __init cmdline_parse_movablecore(char *p)
5014 return cmdline_parse_core(p, &required_movablecore);
5017 early_param("kernelcore", cmdline_parse_kernelcore);
5018 early_param("movablecore", cmdline_parse_movablecore);
5020 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
5023 * set_dma_reserve - set the specified number of pages reserved in the first zone
5024 * @new_dma_reserve: The number of pages to mark reserved
5026 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5027 * In the DMA zone, a significant percentage may be consumed by kernel image
5028 * and other unfreeable allocations which can skew the watermarks badly. This
5029 * function may optionally be used to account for unfreeable pages in the
5030 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5031 * smaller per-cpu batchsize.
5033 void __init set_dma_reserve(unsigned long new_dma_reserve)
5035 dma_reserve = new_dma_reserve;
5038 void __init free_area_init(unsigned long *zones_size)
5040 free_area_init_node(0, zones_size,
5041 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5044 static int page_alloc_cpu_notify(struct notifier_block *self,
5045 unsigned long action, void *hcpu)
5047 int cpu = (unsigned long)hcpu;
5049 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5053 * Spill the event counters of the dead processor
5054 * into the current processors event counters.
5055 * This artificially elevates the count of the current
5058 vm_events_fold_cpu(cpu);
5061 * Zero the differential counters of the dead processor
5062 * so that the vm statistics are consistent.
5064 * This is only okay since the processor is dead and cannot
5065 * race with what we are doing.
5067 refresh_cpu_vm_stats(cpu);
5072 void __init page_alloc_init(void)
5074 hotcpu_notifier(page_alloc_cpu_notify, 0);
5078 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5079 * or min_free_kbytes changes.
5081 static void calculate_totalreserve_pages(void)
5083 struct pglist_data *pgdat;
5084 unsigned long reserve_pages = 0;
5085 enum zone_type i, j;
5087 for_each_online_pgdat(pgdat) {
5088 for (i = 0; i < MAX_NR_ZONES; i++) {
5089 struct zone *zone = pgdat->node_zones + i;
5090 unsigned long max = 0;
5092 /* Find valid and maximum lowmem_reserve in the zone */
5093 for (j = i; j < MAX_NR_ZONES; j++) {
5094 if (zone->lowmem_reserve[j] > max)
5095 max = zone->lowmem_reserve[j];
5098 /* we treat the high watermark as reserved pages. */
5099 max += high_wmark_pages(zone);
5101 if (max > zone->present_pages)
5102 max = zone->present_pages;
5103 reserve_pages += max;
5106 totalreserve_pages = reserve_pages;
5110 * setup_per_zone_lowmem_reserve - called whenever
5111 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5112 * has a correct pages reserved value, so an adequate number of
5113 * pages are left in the zone after a successful __alloc_pages().
5115 static void setup_per_zone_lowmem_reserve(void)
5117 struct pglist_data *pgdat;
5118 enum zone_type j, idx;
5120 for_each_online_pgdat(pgdat) {
5121 for (j = 0; j < MAX_NR_ZONES; j++) {
5122 struct zone *zone = pgdat->node_zones + j;
5123 unsigned long present_pages = zone->present_pages;
5125 zone->lowmem_reserve[j] = 0;
5129 struct zone *lower_zone;
5133 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5134 sysctl_lowmem_reserve_ratio[idx] = 1;
5136 lower_zone = pgdat->node_zones + idx;
5137 lower_zone->lowmem_reserve[j] = present_pages /
5138 sysctl_lowmem_reserve_ratio[idx];
5139 present_pages += lower_zone->present_pages;
5144 /* update totalreserve_pages */
5145 calculate_totalreserve_pages();
5149 * setup_per_zone_wmarks - called when min_free_kbytes changes
5150 * or when memory is hot-{added|removed}
5152 * Ensures that the watermark[min,low,high] values for each zone are set
5153 * correctly with respect to min_free_kbytes.
5155 void setup_per_zone_wmarks(void)
5157 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5158 unsigned long lowmem_pages = 0;
5160 unsigned long flags;
5162 /* Calculate total number of !ZONE_HIGHMEM pages */
5163 for_each_zone(zone) {
5164 if (!is_highmem(zone))
5165 lowmem_pages += zone->present_pages;
5168 for_each_zone(zone) {
5171 spin_lock_irqsave(&zone->lock, flags);
5172 tmp = (u64)pages_min * zone->present_pages;
5173 do_div(tmp, lowmem_pages);
5174 if (is_highmem(zone)) {
5176 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5177 * need highmem pages, so cap pages_min to a small
5180 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5181 * deltas controls asynch page reclaim, and so should
5182 * not be capped for highmem.
5186 min_pages = zone->present_pages / 1024;
5187 if (min_pages < SWAP_CLUSTER_MAX)
5188 min_pages = SWAP_CLUSTER_MAX;
5189 if (min_pages > 128)
5191 zone->watermark[WMARK_MIN] = min_pages;
5194 * If it's a lowmem zone, reserve a number of pages
5195 * proportionate to the zone's size.
5197 zone->watermark[WMARK_MIN] = tmp;
5200 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5201 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5202 setup_zone_migrate_reserve(zone);
5203 spin_unlock_irqrestore(&zone->lock, flags);
5207 for_each_populated_zone(zone) {
5210 for_each_online_cpu(cpu) {
5213 high = percpu_pagelist_fraction
5214 ? zone->present_pages / percpu_pagelist_fraction
5215 : 5 * zone_batchsize(zone);
5216 setup_pagelist_highmark(
5217 per_cpu_ptr(zone->pageset, cpu), high);
5222 /* update totalreserve_pages */
5223 calculate_totalreserve_pages();
5227 * The inactive anon list should be small enough that the VM never has to
5228 * do too much work, but large enough that each inactive page has a chance
5229 * to be referenced again before it is swapped out.
5231 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5232 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5233 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5234 * the anonymous pages are kept on the inactive list.
5237 * memory ratio inactive anon
5238 * -------------------------------------
5247 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5249 unsigned int gb, ratio;
5251 /* Zone size in gigabytes */
5252 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5254 ratio = int_sqrt(10 * gb);
5258 zone->inactive_ratio = ratio;
5261 static void __meminit setup_per_zone_inactive_ratio(void)
5266 calculate_zone_inactive_ratio(zone);
5270 * Initialise min_free_kbytes.
5272 * For small machines we want it small (128k min). For large machines
5273 * we want it large (64MB max). But it is not linear, because network
5274 * bandwidth does not increase linearly with machine size. We use
5276 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5277 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5293 int __meminit init_per_zone_wmark_min(void)
5295 unsigned long lowmem_kbytes;
5297 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5299 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5300 if (min_free_kbytes < 128)
5301 min_free_kbytes = 128;
5302 if (min_free_kbytes > 65536)
5303 min_free_kbytes = 65536;
5304 setup_per_zone_wmarks();
5305 refresh_zone_stat_thresholds();
5306 setup_per_zone_lowmem_reserve();
5307 setup_per_zone_inactive_ratio();
5310 module_init(init_per_zone_wmark_min)
5313 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5314 * that we can call two helper functions whenever min_free_kbytes
5317 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5318 void __user *buffer, size_t *length, loff_t *ppos)
5320 proc_dointvec(table, write, buffer, length, ppos);
5322 setup_per_zone_wmarks();
5327 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5328 void __user *buffer, size_t *length, loff_t *ppos)
5333 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5338 zone->min_unmapped_pages = (zone->present_pages *
5339 sysctl_min_unmapped_ratio) / 100;
5343 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5344 void __user *buffer, size_t *length, loff_t *ppos)
5349 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5354 zone->min_slab_pages = (zone->present_pages *
5355 sysctl_min_slab_ratio) / 100;
5361 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5362 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5363 * whenever sysctl_lowmem_reserve_ratio changes.
5365 * The reserve ratio obviously has absolutely no relation with the
5366 * minimum watermarks. The lowmem reserve ratio can only make sense
5367 * if in function of the boot time zone sizes.
5369 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5370 void __user *buffer, size_t *length, loff_t *ppos)
5372 proc_dointvec_minmax(table, write, buffer, length, ppos);
5373 setup_per_zone_lowmem_reserve();
5378 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5379 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5380 * can have before it gets flushed back to buddy allocator.
5383 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5384 void __user *buffer, size_t *length, loff_t *ppos)
5390 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5391 if (!write || (ret == -EINVAL))
5393 for_each_populated_zone(zone) {
5394 for_each_possible_cpu(cpu) {
5396 high = zone->present_pages / percpu_pagelist_fraction;
5397 setup_pagelist_highmark(
5398 per_cpu_ptr(zone->pageset, cpu), high);
5404 int hashdist = HASHDIST_DEFAULT;
5407 static int __init set_hashdist(char *str)
5411 hashdist = simple_strtoul(str, &str, 0);
5414 __setup("hashdist=", set_hashdist);
5418 * allocate a large system hash table from bootmem
5419 * - it is assumed that the hash table must contain an exact power-of-2
5420 * quantity of entries
5421 * - limit is the number of hash buckets, not the total allocation size
5423 void *__init alloc_large_system_hash(const char *tablename,
5424 unsigned long bucketsize,
5425 unsigned long numentries,
5428 unsigned int *_hash_shift,
5429 unsigned int *_hash_mask,
5430 unsigned long limit)
5432 unsigned long long max = limit;
5433 unsigned long log2qty, size;
5436 /* allow the kernel cmdline to have a say */
5438 /* round applicable memory size up to nearest megabyte */
5439 numentries = nr_kernel_pages;
5440 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5441 numentries >>= 20 - PAGE_SHIFT;
5442 numentries <<= 20 - PAGE_SHIFT;
5444 /* limit to 1 bucket per 2^scale bytes of low memory */
5445 if (scale > PAGE_SHIFT)
5446 numentries >>= (scale - PAGE_SHIFT);
5448 numentries <<= (PAGE_SHIFT - scale);
5450 /* Make sure we've got at least a 0-order allocation.. */
5451 if (unlikely(flags & HASH_SMALL)) {
5452 /* Makes no sense without HASH_EARLY */
5453 WARN_ON(!(flags & HASH_EARLY));
5454 if (!(numentries >> *_hash_shift)) {
5455 numentries = 1UL << *_hash_shift;
5456 BUG_ON(!numentries);
5458 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5459 numentries = PAGE_SIZE / bucketsize;
5461 numentries = roundup_pow_of_two(numentries);
5463 /* limit allocation size to 1/16 total memory by default */
5465 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5466 do_div(max, bucketsize);
5469 if (numentries > max)
5472 log2qty = ilog2(numentries);
5475 size = bucketsize << log2qty;
5476 if (flags & HASH_EARLY)
5477 table = alloc_bootmem_nopanic(size);
5479 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5482 * If bucketsize is not a power-of-two, we may free
5483 * some pages at the end of hash table which
5484 * alloc_pages_exact() automatically does
5486 if (get_order(size) < MAX_ORDER) {
5487 table = alloc_pages_exact(size, GFP_ATOMIC);
5488 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5491 } while (!table && size > PAGE_SIZE && --log2qty);
5494 panic("Failed to allocate %s hash table\n", tablename);
5496 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5499 ilog2(size) - PAGE_SHIFT,
5503 *_hash_shift = log2qty;
5505 *_hash_mask = (1 << log2qty) - 1;
5510 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5511 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5514 #ifdef CONFIG_SPARSEMEM
5515 return __pfn_to_section(pfn)->pageblock_flags;
5517 return zone->pageblock_flags;
5518 #endif /* CONFIG_SPARSEMEM */
5521 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5523 #ifdef CONFIG_SPARSEMEM
5524 pfn &= (PAGES_PER_SECTION-1);
5525 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5527 pfn = pfn - zone->zone_start_pfn;
5528 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5529 #endif /* CONFIG_SPARSEMEM */
5533 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5534 * @page: The page within the block of interest
5535 * @start_bitidx: The first bit of interest to retrieve
5536 * @end_bitidx: The last bit of interest
5537 * returns pageblock_bits flags
5539 unsigned long get_pageblock_flags_group(struct page *page,
5540 int start_bitidx, int end_bitidx)
5543 unsigned long *bitmap;
5544 unsigned long pfn, bitidx;
5545 unsigned long flags = 0;
5546 unsigned long value = 1;
5548 zone = page_zone(page);
5549 pfn = page_to_pfn(page);
5550 bitmap = get_pageblock_bitmap(zone, pfn);
5551 bitidx = pfn_to_bitidx(zone, pfn);
5553 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5554 if (test_bit(bitidx + start_bitidx, bitmap))
5561 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5562 * @page: The page within the block of interest
5563 * @start_bitidx: The first bit of interest
5564 * @end_bitidx: The last bit of interest
5565 * @flags: The flags to set
5567 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5568 int start_bitidx, int end_bitidx)
5571 unsigned long *bitmap;
5572 unsigned long pfn, bitidx;
5573 unsigned long value = 1;
5575 zone = page_zone(page);
5576 pfn = page_to_pfn(page);
5577 bitmap = get_pageblock_bitmap(zone, pfn);
5578 bitidx = pfn_to_bitidx(zone, pfn);
5579 VM_BUG_ON(pfn < zone->zone_start_pfn);
5580 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5582 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5584 __set_bit(bitidx + start_bitidx, bitmap);
5586 __clear_bit(bitidx + start_bitidx, bitmap);
5590 * This is designed as sub function...plz see page_isolation.c also.
5591 * set/clear page block's type to be ISOLATE.
5592 * page allocater never alloc memory from ISOLATE block.
5596 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5598 unsigned long pfn, iter, found;
5600 * For avoiding noise data, lru_add_drain_all() should be called
5601 * If ZONE_MOVABLE, the zone never contains immobile pages
5603 if (zone_idx(zone) == ZONE_MOVABLE)
5606 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5609 pfn = page_to_pfn(page);
5610 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5611 unsigned long check = pfn + iter;
5613 if (!pfn_valid_within(check))
5616 page = pfn_to_page(check);
5617 if (!page_count(page)) {
5618 if (PageBuddy(page))
5619 iter += (1 << page_order(page)) - 1;
5625 * If there are RECLAIMABLE pages, we need to check it.
5626 * But now, memory offline itself doesn't call shrink_slab()
5627 * and it still to be fixed.
5630 * If the page is not RAM, page_count()should be 0.
5631 * we don't need more check. This is an _used_ not-movable page.
5633 * The problematic thing here is PG_reserved pages. PG_reserved
5634 * is set to both of a memory hole page and a _used_ kernel
5643 bool is_pageblock_removable_nolock(struct page *page)
5645 struct zone *zone = page_zone(page);
5646 return __count_immobile_pages(zone, page, 0);
5649 int set_migratetype_isolate(struct page *page)
5652 unsigned long flags, pfn;
5653 struct memory_isolate_notify arg;
5657 zone = page_zone(page);
5659 spin_lock_irqsave(&zone->lock, flags);
5661 pfn = page_to_pfn(page);
5662 arg.start_pfn = pfn;
5663 arg.nr_pages = pageblock_nr_pages;
5664 arg.pages_found = 0;
5667 * It may be possible to isolate a pageblock even if the
5668 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5669 * notifier chain is used by balloon drivers to return the
5670 * number of pages in a range that are held by the balloon
5671 * driver to shrink memory. If all the pages are accounted for
5672 * by balloons, are free, or on the LRU, isolation can continue.
5673 * Later, for example, when memory hotplug notifier runs, these
5674 * pages reported as "can be isolated" should be isolated(freed)
5675 * by the balloon driver through the memory notifier chain.
5677 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5678 notifier_ret = notifier_to_errno(notifier_ret);
5682 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5683 * We just check MOVABLE pages.
5685 if (__count_immobile_pages(zone, page, arg.pages_found))
5689 * immobile means "not-on-lru" paes. If immobile is larger than
5690 * removable-by-driver pages reported by notifier, we'll fail.
5695 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5696 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5699 spin_unlock_irqrestore(&zone->lock, flags);
5705 void unset_migratetype_isolate(struct page *page)
5708 unsigned long flags;
5709 zone = page_zone(page);
5710 spin_lock_irqsave(&zone->lock, flags);
5711 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5713 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5714 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5716 spin_unlock_irqrestore(&zone->lock, flags);
5719 #ifdef CONFIG_MEMORY_HOTREMOVE
5721 * All pages in the range must be isolated before calling this.
5724 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5730 unsigned long flags;
5731 /* find the first valid pfn */
5732 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5737 zone = page_zone(pfn_to_page(pfn));
5738 spin_lock_irqsave(&zone->lock, flags);
5740 while (pfn < end_pfn) {
5741 if (!pfn_valid(pfn)) {
5745 page = pfn_to_page(pfn);
5746 BUG_ON(page_count(page));
5747 BUG_ON(!PageBuddy(page));
5748 order = page_order(page);
5749 #ifdef CONFIG_DEBUG_VM
5750 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5751 pfn, 1 << order, end_pfn);
5753 list_del(&page->lru);
5754 rmv_page_order(page);
5755 zone->free_area[order].nr_free--;
5756 __mod_zone_page_state(zone, NR_FREE_PAGES,
5758 for (i = 0; i < (1 << order); i++)
5759 SetPageReserved((page+i));
5760 pfn += (1 << order);
5762 spin_unlock_irqrestore(&zone->lock, flags);
5766 #ifdef CONFIG_MEMORY_FAILURE
5767 bool is_free_buddy_page(struct page *page)
5769 struct zone *zone = page_zone(page);
5770 unsigned long pfn = page_to_pfn(page);
5771 unsigned long flags;
5774 spin_lock_irqsave(&zone->lock, flags);
5775 for (order = 0; order < MAX_ORDER; order++) {
5776 struct page *page_head = page - (pfn & ((1 << order) - 1));
5778 if (PageBuddy(page_head) && page_order(page_head) >= order)
5781 spin_unlock_irqrestore(&zone->lock, flags);
5783 return order < MAX_ORDER;
5787 static struct trace_print_flags pageflag_names[] = {
5788 {1UL << PG_locked, "locked" },
5789 {1UL << PG_error, "error" },
5790 {1UL << PG_referenced, "referenced" },
5791 {1UL << PG_uptodate, "uptodate" },
5792 {1UL << PG_dirty, "dirty" },
5793 {1UL << PG_lru, "lru" },
5794 {1UL << PG_active, "active" },
5795 {1UL << PG_slab, "slab" },
5796 {1UL << PG_owner_priv_1, "owner_priv_1" },
5797 {1UL << PG_arch_1, "arch_1" },
5798 {1UL << PG_reserved, "reserved" },
5799 {1UL << PG_private, "private" },
5800 {1UL << PG_private_2, "private_2" },
5801 {1UL << PG_writeback, "writeback" },
5802 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5803 {1UL << PG_head, "head" },
5804 {1UL << PG_tail, "tail" },
5806 {1UL << PG_compound, "compound" },
5808 {1UL << PG_swapcache, "swapcache" },
5809 {1UL << PG_mappedtodisk, "mappedtodisk" },
5810 {1UL << PG_reclaim, "reclaim" },
5811 {1UL << PG_swapbacked, "swapbacked" },
5812 {1UL << PG_unevictable, "unevictable" },
5814 {1UL << PG_mlocked, "mlocked" },
5816 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5817 {1UL << PG_uncached, "uncached" },
5819 #ifdef CONFIG_MEMORY_FAILURE
5820 {1UL << PG_hwpoison, "hwpoison" },
5825 static void dump_page_flags(unsigned long flags)
5827 const char *delim = "";
5831 printk(KERN_ALERT "page flags: %#lx(", flags);
5833 /* remove zone id */
5834 flags &= (1UL << NR_PAGEFLAGS) - 1;
5836 for (i = 0; pageflag_names[i].name && flags; i++) {
5838 mask = pageflag_names[i].mask;
5839 if ((flags & mask) != mask)
5843 printk("%s%s", delim, pageflag_names[i].name);
5847 /* check for left over flags */
5849 printk("%s%#lx", delim, flags);
5854 void dump_page(struct page *page)
5857 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5858 page, atomic_read(&page->_count), page_mapcount(page),
5859 page->mapping, page->index);
5860 dump_page_flags(page->flags);
5861 mem_cgroup_print_bad_page(page);