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)
657 trace_mm_page_free_direct(page, order);
658 kmemcheck_free_shadow(page, order);
661 page->mapping = NULL;
662 for (i = 0; i < (1 << order); i++)
663 bad += free_pages_check(page + i);
667 if (!PageHighMem(page)) {
668 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
669 debug_check_no_obj_freed(page_address(page),
672 arch_free_page(page, order);
673 kernel_map_pages(page, 1 << order, 0);
678 static void __free_pages_ok(struct page *page, unsigned int order)
681 int wasMlocked = __TestClearPageMlocked(page);
683 if (!free_pages_prepare(page, order))
686 local_irq_save(flags);
687 if (unlikely(wasMlocked))
688 free_page_mlock(page);
689 __count_vm_events(PGFREE, 1 << order);
690 free_one_page(page_zone(page), page, order,
691 get_pageblock_migratetype(page));
692 local_irq_restore(flags);
696 * permit the bootmem allocator to evade page validation on high-order frees
698 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
701 __ClearPageReserved(page);
702 set_page_count(page, 0);
703 set_page_refcounted(page);
709 for (loop = 0; loop < BITS_PER_LONG; loop++) {
710 struct page *p = &page[loop];
712 if (loop + 1 < BITS_PER_LONG)
714 __ClearPageReserved(p);
715 set_page_count(p, 0);
718 set_page_refcounted(page);
719 __free_pages(page, order);
725 * The order of subdivision here is critical for the IO subsystem.
726 * Please do not alter this order without good reasons and regression
727 * testing. Specifically, as large blocks of memory are subdivided,
728 * the order in which smaller blocks are delivered depends on the order
729 * they're subdivided in this function. This is the primary factor
730 * influencing the order in which pages are delivered to the IO
731 * subsystem according to empirical testing, and this is also justified
732 * by considering the behavior of a buddy system containing a single
733 * large block of memory acted on by a series of small allocations.
734 * This behavior is a critical factor in sglist merging's success.
738 static inline void expand(struct zone *zone, struct page *page,
739 int low, int high, struct free_area *area,
742 unsigned long size = 1 << high;
748 VM_BUG_ON(bad_range(zone, &page[size]));
749 list_add(&page[size].lru, &area->free_list[migratetype]);
751 set_page_order(&page[size], high);
756 * This page is about to be returned from the page allocator
758 static inline int check_new_page(struct page *page)
760 if (unlikely(page_mapcount(page) |
761 (page->mapping != NULL) |
762 (atomic_read(&page->_count) != 0) |
763 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
764 (mem_cgroup_bad_page_check(page)))) {
771 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
775 for (i = 0; i < (1 << order); i++) {
776 struct page *p = page + i;
777 if (unlikely(check_new_page(p)))
781 set_page_private(page, 0);
782 set_page_refcounted(page);
784 arch_alloc_page(page, order);
785 kernel_map_pages(page, 1 << order, 1);
787 if (gfp_flags & __GFP_ZERO)
788 prep_zero_page(page, order, gfp_flags);
790 if (order && (gfp_flags & __GFP_COMP))
791 prep_compound_page(page, order);
797 * Go through the free lists for the given migratetype and remove
798 * the smallest available page from the freelists
801 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
804 unsigned int current_order;
805 struct free_area * area;
808 /* Find a page of the appropriate size in the preferred list */
809 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
810 area = &(zone->free_area[current_order]);
811 if (list_empty(&area->free_list[migratetype]))
814 page = list_entry(area->free_list[migratetype].next,
816 list_del(&page->lru);
817 rmv_page_order(page);
819 expand(zone, page, order, current_order, area, migratetype);
828 * This array describes the order lists are fallen back to when
829 * the free lists for the desirable migrate type are depleted
831 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
832 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
833 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
834 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
835 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
839 * Move the free pages in a range to the free lists of the requested type.
840 * Note that start_page and end_pages are not aligned on a pageblock
841 * boundary. If alignment is required, use move_freepages_block()
843 static int move_freepages(struct zone *zone,
844 struct page *start_page, struct page *end_page,
851 #ifndef CONFIG_HOLES_IN_ZONE
853 * page_zone is not safe to call in this context when
854 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
855 * anyway as we check zone boundaries in move_freepages_block().
856 * Remove at a later date when no bug reports exist related to
857 * grouping pages by mobility
859 BUG_ON(page_zone(start_page) != page_zone(end_page));
862 for (page = start_page; page <= end_page;) {
863 /* Make sure we are not inadvertently changing nodes */
864 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
866 if (!pfn_valid_within(page_to_pfn(page))) {
871 if (!PageBuddy(page)) {
876 order = page_order(page);
877 list_move(&page->lru,
878 &zone->free_area[order].free_list[migratetype]);
880 pages_moved += 1 << order;
886 static int move_freepages_block(struct zone *zone, struct page *page,
889 unsigned long start_pfn, end_pfn;
890 struct page *start_page, *end_page;
892 start_pfn = page_to_pfn(page);
893 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
894 start_page = pfn_to_page(start_pfn);
895 end_page = start_page + pageblock_nr_pages - 1;
896 end_pfn = start_pfn + pageblock_nr_pages - 1;
898 /* Do not cross zone boundaries */
899 if (start_pfn < zone->zone_start_pfn)
901 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
904 return move_freepages(zone, start_page, end_page, migratetype);
907 static void change_pageblock_range(struct page *pageblock_page,
908 int start_order, int migratetype)
910 int nr_pageblocks = 1 << (start_order - pageblock_order);
912 while (nr_pageblocks--) {
913 set_pageblock_migratetype(pageblock_page, migratetype);
914 pageblock_page += pageblock_nr_pages;
918 /* Remove an element from the buddy allocator from the fallback list */
919 static inline struct page *
920 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
922 struct free_area * area;
927 /* Find the largest possible block of pages in the other list */
928 for (current_order = MAX_ORDER-1; current_order >= order;
930 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
931 migratetype = fallbacks[start_migratetype][i];
933 /* MIGRATE_RESERVE handled later if necessary */
934 if (migratetype == MIGRATE_RESERVE)
937 area = &(zone->free_area[current_order]);
938 if (list_empty(&area->free_list[migratetype]))
941 page = list_entry(area->free_list[migratetype].next,
946 * If breaking a large block of pages, move all free
947 * pages to the preferred allocation list. If falling
948 * back for a reclaimable kernel allocation, be more
949 * aggressive about taking ownership of free pages
951 if (unlikely(current_order >= (pageblock_order >> 1)) ||
952 start_migratetype == MIGRATE_RECLAIMABLE ||
953 page_group_by_mobility_disabled) {
955 pages = move_freepages_block(zone, page,
958 /* Claim the whole block if over half of it is free */
959 if (pages >= (1 << (pageblock_order-1)) ||
960 page_group_by_mobility_disabled)
961 set_pageblock_migratetype(page,
964 migratetype = start_migratetype;
967 /* Remove the page from the freelists */
968 list_del(&page->lru);
969 rmv_page_order(page);
971 /* Take ownership for orders >= pageblock_order */
972 if (current_order >= pageblock_order)
973 change_pageblock_range(page, current_order,
976 expand(zone, page, order, current_order, area, migratetype);
978 trace_mm_page_alloc_extfrag(page, order, current_order,
979 start_migratetype, migratetype);
989 * Do the hard work of removing an element from the buddy allocator.
990 * Call me with the zone->lock already held.
992 static struct page *__rmqueue(struct zone *zone, unsigned int order,
998 page = __rmqueue_smallest(zone, order, migratetype);
1000 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1001 page = __rmqueue_fallback(zone, order, migratetype);
1004 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1005 * is used because __rmqueue_smallest is an inline function
1006 * and we want just one call site
1009 migratetype = MIGRATE_RESERVE;
1014 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1019 * Obtain a specified number of elements from the buddy allocator, all under
1020 * a single hold of the lock, for efficiency. Add them to the supplied list.
1021 * Returns the number of new pages which were placed at *list.
1023 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1024 unsigned long count, struct list_head *list,
1025 int migratetype, int cold)
1029 spin_lock(&zone->lock);
1030 for (i = 0; i < count; ++i) {
1031 struct page *page = __rmqueue(zone, order, migratetype);
1032 if (unlikely(page == NULL))
1036 * Split buddy pages returned by expand() are received here
1037 * in physical page order. The page is added to the callers and
1038 * list and the list head then moves forward. From the callers
1039 * perspective, the linked list is ordered by page number in
1040 * some conditions. This is useful for IO devices that can
1041 * merge IO requests if the physical pages are ordered
1044 if (likely(cold == 0))
1045 list_add(&page->lru, list);
1047 list_add_tail(&page->lru, list);
1048 set_page_private(page, migratetype);
1051 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1052 spin_unlock(&zone->lock);
1058 * Called from the vmstat counter updater to drain pagesets of this
1059 * currently executing processor on remote nodes after they have
1062 * Note that this function must be called with the thread pinned to
1063 * a single processor.
1065 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1067 unsigned long flags;
1070 local_irq_save(flags);
1071 if (pcp->count >= pcp->batch)
1072 to_drain = pcp->batch;
1074 to_drain = pcp->count;
1075 free_pcppages_bulk(zone, to_drain, pcp);
1076 pcp->count -= to_drain;
1077 local_irq_restore(flags);
1082 * Drain pages of the indicated processor.
1084 * The processor must either be the current processor and the
1085 * thread pinned to the current processor or a processor that
1088 static void drain_pages(unsigned int cpu)
1090 unsigned long flags;
1093 for_each_populated_zone(zone) {
1094 struct per_cpu_pageset *pset;
1095 struct per_cpu_pages *pcp;
1097 local_irq_save(flags);
1098 pset = per_cpu_ptr(zone->pageset, cpu);
1102 free_pcppages_bulk(zone, pcp->count, pcp);
1105 local_irq_restore(flags);
1110 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1112 void drain_local_pages(void *arg)
1114 drain_pages(smp_processor_id());
1118 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1120 void drain_all_pages(void)
1122 on_each_cpu(drain_local_pages, NULL, 1);
1125 #ifdef CONFIG_HIBERNATION
1127 void mark_free_pages(struct zone *zone)
1129 unsigned long pfn, max_zone_pfn;
1130 unsigned long flags;
1132 struct list_head *curr;
1134 if (!zone->spanned_pages)
1137 spin_lock_irqsave(&zone->lock, flags);
1139 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1140 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1141 if (pfn_valid(pfn)) {
1142 struct page *page = pfn_to_page(pfn);
1144 if (!swsusp_page_is_forbidden(page))
1145 swsusp_unset_page_free(page);
1148 for_each_migratetype_order(order, t) {
1149 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1152 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1153 for (i = 0; i < (1UL << order); i++)
1154 swsusp_set_page_free(pfn_to_page(pfn + i));
1157 spin_unlock_irqrestore(&zone->lock, flags);
1159 #endif /* CONFIG_PM */
1162 * Free a 0-order page
1163 * cold == 1 ? free a cold page : free a hot page
1165 void free_hot_cold_page(struct page *page, int cold)
1167 struct zone *zone = page_zone(page);
1168 struct per_cpu_pages *pcp;
1169 unsigned long flags;
1171 int wasMlocked = __TestClearPageMlocked(page);
1173 if (!free_pages_prepare(page, 0))
1176 migratetype = get_pageblock_migratetype(page);
1177 set_page_private(page, migratetype);
1178 local_irq_save(flags);
1179 if (unlikely(wasMlocked))
1180 free_page_mlock(page);
1181 __count_vm_event(PGFREE);
1184 * We only track unmovable, reclaimable and movable on pcp lists.
1185 * Free ISOLATE pages back to the allocator because they are being
1186 * offlined but treat RESERVE as movable pages so we can get those
1187 * areas back if necessary. Otherwise, we may have to free
1188 * excessively into the page allocator
1190 if (migratetype >= MIGRATE_PCPTYPES) {
1191 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1192 free_one_page(zone, page, 0, migratetype);
1195 migratetype = MIGRATE_MOVABLE;
1198 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1200 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1202 list_add(&page->lru, &pcp->lists[migratetype]);
1204 if (pcp->count >= pcp->high) {
1205 free_pcppages_bulk(zone, pcp->batch, pcp);
1206 pcp->count -= pcp->batch;
1210 local_irq_restore(flags);
1214 * split_page takes a non-compound higher-order page, and splits it into
1215 * n (1<<order) sub-pages: page[0..n]
1216 * Each sub-page must be freed individually.
1218 * Note: this is probably too low level an operation for use in drivers.
1219 * Please consult with lkml before using this in your driver.
1221 void split_page(struct page *page, unsigned int order)
1225 VM_BUG_ON(PageCompound(page));
1226 VM_BUG_ON(!page_count(page));
1228 #ifdef CONFIG_KMEMCHECK
1230 * Split shadow pages too, because free(page[0]) would
1231 * otherwise free the whole shadow.
1233 if (kmemcheck_page_is_tracked(page))
1234 split_page(virt_to_page(page[0].shadow), order);
1237 for (i = 1; i < (1 << order); i++)
1238 set_page_refcounted(page + i);
1242 * Similar to split_page except the page is already free. As this is only
1243 * being used for migration, the migratetype of the block also changes.
1244 * As this is called with interrupts disabled, the caller is responsible
1245 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1248 * Note: this is probably too low level an operation for use in drivers.
1249 * Please consult with lkml before using this in your driver.
1251 int split_free_page(struct page *page)
1254 unsigned long watermark;
1257 BUG_ON(!PageBuddy(page));
1259 zone = page_zone(page);
1260 order = page_order(page);
1262 /* Obey watermarks as if the page was being allocated */
1263 watermark = low_wmark_pages(zone) + (1 << order);
1264 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1267 /* Remove page from free list */
1268 list_del(&page->lru);
1269 zone->free_area[order].nr_free--;
1270 rmv_page_order(page);
1271 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1273 /* Split into individual pages */
1274 set_page_refcounted(page);
1275 split_page(page, order);
1277 if (order >= pageblock_order - 1) {
1278 struct page *endpage = page + (1 << order) - 1;
1279 for (; page < endpage; page += pageblock_nr_pages)
1280 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1287 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1288 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1292 struct page *buffered_rmqueue(struct zone *preferred_zone,
1293 struct zone *zone, int order, gfp_t gfp_flags,
1296 unsigned long flags;
1298 int cold = !!(gfp_flags & __GFP_COLD);
1301 if (likely(order == 0)) {
1302 struct per_cpu_pages *pcp;
1303 struct list_head *list;
1305 local_irq_save(flags);
1306 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1307 list = &pcp->lists[migratetype];
1308 if (list_empty(list)) {
1309 pcp->count += rmqueue_bulk(zone, 0,
1312 if (unlikely(list_empty(list)))
1317 page = list_entry(list->prev, struct page, lru);
1319 page = list_entry(list->next, struct page, lru);
1321 list_del(&page->lru);
1324 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1326 * __GFP_NOFAIL is not to be used in new code.
1328 * All __GFP_NOFAIL callers should be fixed so that they
1329 * properly detect and handle allocation failures.
1331 * We most definitely don't want callers attempting to
1332 * allocate greater than order-1 page units with
1335 WARN_ON_ONCE(order > 1);
1337 spin_lock_irqsave(&zone->lock, flags);
1338 page = __rmqueue(zone, order, migratetype);
1339 spin_unlock(&zone->lock);
1342 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1345 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1346 zone_statistics(preferred_zone, zone, gfp_flags);
1347 local_irq_restore(flags);
1349 VM_BUG_ON(bad_range(zone, page));
1350 if (prep_new_page(page, order, gfp_flags))
1355 local_irq_restore(flags);
1359 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1360 #define ALLOC_WMARK_MIN WMARK_MIN
1361 #define ALLOC_WMARK_LOW WMARK_LOW
1362 #define ALLOC_WMARK_HIGH WMARK_HIGH
1363 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1365 /* Mask to get the watermark bits */
1366 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1368 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1369 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1370 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1372 #ifdef CONFIG_FAIL_PAGE_ALLOC
1375 struct fault_attr attr;
1377 u32 ignore_gfp_highmem;
1378 u32 ignore_gfp_wait;
1380 } fail_page_alloc = {
1381 .attr = FAULT_ATTR_INITIALIZER,
1382 .ignore_gfp_wait = 1,
1383 .ignore_gfp_highmem = 1,
1387 static int __init setup_fail_page_alloc(char *str)
1389 return setup_fault_attr(&fail_page_alloc.attr, str);
1391 __setup("fail_page_alloc=", setup_fail_page_alloc);
1393 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1395 if (order < fail_page_alloc.min_order)
1397 if (gfp_mask & __GFP_NOFAIL)
1399 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1401 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1404 return should_fail(&fail_page_alloc.attr, 1 << order);
1407 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1409 static int __init fail_page_alloc_debugfs(void)
1411 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1414 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1415 &fail_page_alloc.attr);
1417 return PTR_ERR(dir);
1419 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1420 &fail_page_alloc.ignore_gfp_wait))
1422 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1423 &fail_page_alloc.ignore_gfp_highmem))
1425 if (!debugfs_create_u32("min-order", mode, dir,
1426 &fail_page_alloc.min_order))
1431 debugfs_remove_recursive(dir);
1436 late_initcall(fail_page_alloc_debugfs);
1438 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1440 #else /* CONFIG_FAIL_PAGE_ALLOC */
1442 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1447 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1450 * Return true if free pages are above 'mark'. This takes into account the order
1451 * of the allocation.
1453 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1454 int classzone_idx, int alloc_flags, long free_pages)
1456 /* free_pages my go negative - that's OK */
1460 free_pages -= (1 << order) + 1;
1461 if (alloc_flags & ALLOC_HIGH)
1463 if (alloc_flags & ALLOC_HARDER)
1466 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1468 for (o = 0; o < order; o++) {
1469 /* At the next order, this order's pages become unavailable */
1470 free_pages -= z->free_area[o].nr_free << o;
1472 /* Require fewer higher order pages to be free */
1475 if (free_pages <= min)
1481 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1482 int classzone_idx, int alloc_flags)
1484 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1485 zone_page_state(z, NR_FREE_PAGES));
1488 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1489 int classzone_idx, int alloc_flags)
1491 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1493 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1494 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1496 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1502 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1503 * skip over zones that are not allowed by the cpuset, or that have
1504 * been recently (in last second) found to be nearly full. See further
1505 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1506 * that have to skip over a lot of full or unallowed zones.
1508 * If the zonelist cache is present in the passed in zonelist, then
1509 * returns a pointer to the allowed node mask (either the current
1510 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1512 * If the zonelist cache is not available for this zonelist, does
1513 * nothing and returns NULL.
1515 * If the fullzones BITMAP in the zonelist cache is stale (more than
1516 * a second since last zap'd) then we zap it out (clear its bits.)
1518 * We hold off even calling zlc_setup, until after we've checked the
1519 * first zone in the zonelist, on the theory that most allocations will
1520 * be satisfied from that first zone, so best to examine that zone as
1521 * quickly as we can.
1523 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1525 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1526 nodemask_t *allowednodes; /* zonelist_cache approximation */
1528 zlc = zonelist->zlcache_ptr;
1532 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1533 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1534 zlc->last_full_zap = jiffies;
1537 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1538 &cpuset_current_mems_allowed :
1539 &node_states[N_HIGH_MEMORY];
1540 return allowednodes;
1544 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1545 * if it is worth looking at further for free memory:
1546 * 1) Check that the zone isn't thought to be full (doesn't have its
1547 * bit set in the zonelist_cache fullzones BITMAP).
1548 * 2) Check that the zones node (obtained from the zonelist_cache
1549 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1550 * Return true (non-zero) if zone is worth looking at further, or
1551 * else return false (zero) if it is not.
1553 * This check -ignores- the distinction between various watermarks,
1554 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1555 * found to be full for any variation of these watermarks, it will
1556 * be considered full for up to one second by all requests, unless
1557 * we are so low on memory on all allowed nodes that we are forced
1558 * into the second scan of the zonelist.
1560 * In the second scan we ignore this zonelist cache and exactly
1561 * apply the watermarks to all zones, even it is slower to do so.
1562 * We are low on memory in the second scan, and should leave no stone
1563 * unturned looking for a free page.
1565 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1566 nodemask_t *allowednodes)
1568 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1569 int i; /* index of *z in zonelist zones */
1570 int n; /* node that zone *z is on */
1572 zlc = zonelist->zlcache_ptr;
1576 i = z - zonelist->_zonerefs;
1579 /* This zone is worth trying if it is allowed but not full */
1580 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1584 * Given 'z' scanning a zonelist, set the corresponding bit in
1585 * zlc->fullzones, so that subsequent attempts to allocate a page
1586 * from that zone don't waste time re-examining it.
1588 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1590 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1591 int i; /* index of *z in zonelist zones */
1593 zlc = zonelist->zlcache_ptr;
1597 i = z - zonelist->_zonerefs;
1599 set_bit(i, zlc->fullzones);
1603 * clear all zones full, called after direct reclaim makes progress so that
1604 * a zone that was recently full is not skipped over for up to a second
1606 static void zlc_clear_zones_full(struct zonelist *zonelist)
1608 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1610 zlc = zonelist->zlcache_ptr;
1614 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1617 #else /* CONFIG_NUMA */
1619 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1624 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1625 nodemask_t *allowednodes)
1630 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1634 static void zlc_clear_zones_full(struct zonelist *zonelist)
1637 #endif /* CONFIG_NUMA */
1640 * get_page_from_freelist goes through the zonelist trying to allocate
1643 static struct page *
1644 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1645 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1646 struct zone *preferred_zone, int migratetype)
1649 struct page *page = NULL;
1652 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1653 int zlc_active = 0; /* set if using zonelist_cache */
1654 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1656 classzone_idx = zone_idx(preferred_zone);
1659 * Scan zonelist, looking for a zone with enough free.
1660 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1662 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1663 high_zoneidx, nodemask) {
1664 if (NUMA_BUILD && zlc_active &&
1665 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1667 if ((alloc_flags & ALLOC_CPUSET) &&
1668 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1671 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1672 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1676 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1677 if (zone_watermark_ok(zone, order, mark,
1678 classzone_idx, alloc_flags))
1681 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1683 * we do zlc_setup if there are multiple nodes
1684 * and before considering the first zone allowed
1687 allowednodes = zlc_setup(zonelist, alloc_flags);
1692 if (zone_reclaim_mode == 0)
1693 goto this_zone_full;
1696 * As we may have just activated ZLC, check if the first
1697 * eligible zone has failed zone_reclaim recently.
1699 if (NUMA_BUILD && zlc_active &&
1700 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1703 ret = zone_reclaim(zone, gfp_mask, order);
1705 case ZONE_RECLAIM_NOSCAN:
1708 case ZONE_RECLAIM_FULL:
1709 /* scanned but unreclaimable */
1712 /* did we reclaim enough */
1713 if (!zone_watermark_ok(zone, order, mark,
1714 classzone_idx, alloc_flags))
1715 goto this_zone_full;
1720 page = buffered_rmqueue(preferred_zone, zone, order,
1721 gfp_mask, migratetype);
1726 zlc_mark_zone_full(zonelist, z);
1729 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1730 /* Disable zlc cache for second zonelist scan */
1738 * Large machines with many possible nodes should not always dump per-node
1739 * meminfo in irq context.
1741 static inline bool should_suppress_show_mem(void)
1746 ret = in_interrupt();
1751 static DEFINE_RATELIMIT_STATE(nopage_rs,
1752 DEFAULT_RATELIMIT_INTERVAL,
1753 DEFAULT_RATELIMIT_BURST);
1755 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1757 unsigned int filter = SHOW_MEM_FILTER_NODES;
1759 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1763 * This documents exceptions given to allocations in certain
1764 * contexts that are allowed to allocate outside current's set
1767 if (!(gfp_mask & __GFP_NOMEMALLOC))
1768 if (test_thread_flag(TIF_MEMDIE) ||
1769 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1770 filter &= ~SHOW_MEM_FILTER_NODES;
1771 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1772 filter &= ~SHOW_MEM_FILTER_NODES;
1775 struct va_format vaf;
1778 va_start(args, fmt);
1783 pr_warn("%pV", &vaf);
1788 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1789 current->comm, order, gfp_mask);
1792 if (!should_suppress_show_mem())
1797 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1798 unsigned long pages_reclaimed)
1800 /* Do not loop if specifically requested */
1801 if (gfp_mask & __GFP_NORETRY)
1805 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1806 * means __GFP_NOFAIL, but that may not be true in other
1809 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1813 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1814 * specified, then we retry until we no longer reclaim any pages
1815 * (above), or we've reclaimed an order of pages at least as
1816 * large as the allocation's order. In both cases, if the
1817 * allocation still fails, we stop retrying.
1819 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1823 * Don't let big-order allocations loop unless the caller
1824 * explicitly requests that.
1826 if (gfp_mask & __GFP_NOFAIL)
1832 static inline struct page *
1833 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1834 struct zonelist *zonelist, enum zone_type high_zoneidx,
1835 nodemask_t *nodemask, struct zone *preferred_zone,
1840 /* Acquire the OOM killer lock for the zones in zonelist */
1841 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1842 schedule_timeout_uninterruptible(1);
1847 * Go through the zonelist yet one more time, keep very high watermark
1848 * here, this is only to catch a parallel oom killing, we must fail if
1849 * we're still under heavy pressure.
1851 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1852 order, zonelist, high_zoneidx,
1853 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1854 preferred_zone, migratetype);
1858 if (!(gfp_mask & __GFP_NOFAIL)) {
1859 /* The OOM killer will not help higher order allocs */
1860 if (order > PAGE_ALLOC_COSTLY_ORDER)
1862 /* The OOM killer does not needlessly kill tasks for lowmem */
1863 if (high_zoneidx < ZONE_NORMAL)
1866 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1867 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1868 * The caller should handle page allocation failure by itself if
1869 * it specifies __GFP_THISNODE.
1870 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1872 if (gfp_mask & __GFP_THISNODE)
1875 /* Exhausted what can be done so it's blamo time */
1876 out_of_memory(zonelist, gfp_mask, order, nodemask);
1879 clear_zonelist_oom(zonelist, gfp_mask);
1883 #ifdef CONFIG_COMPACTION
1884 /* Try memory compaction for high-order allocations before reclaim */
1885 static struct page *
1886 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1887 struct zonelist *zonelist, enum zone_type high_zoneidx,
1888 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1889 int migratetype, unsigned long *did_some_progress,
1890 bool sync_migration)
1894 if (!order || compaction_deferred(preferred_zone))
1897 current->flags |= PF_MEMALLOC;
1898 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1899 nodemask, sync_migration);
1900 current->flags &= ~PF_MEMALLOC;
1901 if (*did_some_progress != COMPACT_SKIPPED) {
1903 /* Page migration frees to the PCP lists but we want merging */
1904 drain_pages(get_cpu());
1907 page = get_page_from_freelist(gfp_mask, nodemask,
1908 order, zonelist, high_zoneidx,
1909 alloc_flags, preferred_zone,
1912 preferred_zone->compact_considered = 0;
1913 preferred_zone->compact_defer_shift = 0;
1914 count_vm_event(COMPACTSUCCESS);
1919 * It's bad if compaction run occurs and fails.
1920 * The most likely reason is that pages exist,
1921 * but not enough to satisfy watermarks.
1923 count_vm_event(COMPACTFAIL);
1924 defer_compaction(preferred_zone);
1932 static inline struct page *
1933 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1934 struct zonelist *zonelist, enum zone_type high_zoneidx,
1935 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1936 int migratetype, unsigned long *did_some_progress,
1937 bool sync_migration)
1941 #endif /* CONFIG_COMPACTION */
1943 /* The really slow allocator path where we enter direct reclaim */
1944 static inline struct page *
1945 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1946 struct zonelist *zonelist, enum zone_type high_zoneidx,
1947 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1948 int migratetype, unsigned long *did_some_progress)
1950 struct page *page = NULL;
1951 struct reclaim_state reclaim_state;
1952 bool drained = false;
1956 /* We now go into synchronous reclaim */
1957 cpuset_memory_pressure_bump();
1958 current->flags |= PF_MEMALLOC;
1959 lockdep_set_current_reclaim_state(gfp_mask);
1960 reclaim_state.reclaimed_slab = 0;
1961 current->reclaim_state = &reclaim_state;
1963 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1965 current->reclaim_state = NULL;
1966 lockdep_clear_current_reclaim_state();
1967 current->flags &= ~PF_MEMALLOC;
1971 if (unlikely(!(*did_some_progress)))
1974 /* After successful reclaim, reconsider all zones for allocation */
1976 zlc_clear_zones_full(zonelist);
1979 page = get_page_from_freelist(gfp_mask, nodemask, order,
1980 zonelist, high_zoneidx,
1981 alloc_flags, preferred_zone,
1985 * If an allocation failed after direct reclaim, it could be because
1986 * pages are pinned on the per-cpu lists. Drain them and try again
1988 if (!page && !drained) {
1998 * This is called in the allocator slow-path if the allocation request is of
1999 * sufficient urgency to ignore watermarks and take other desperate measures
2001 static inline struct page *
2002 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2003 struct zonelist *zonelist, enum zone_type high_zoneidx,
2004 nodemask_t *nodemask, struct zone *preferred_zone,
2010 page = get_page_from_freelist(gfp_mask, nodemask, order,
2011 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2012 preferred_zone, migratetype);
2014 if (!page && gfp_mask & __GFP_NOFAIL)
2015 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2016 } while (!page && (gfp_mask & __GFP_NOFAIL));
2022 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2023 enum zone_type high_zoneidx,
2024 enum zone_type classzone_idx)
2029 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2030 wakeup_kswapd(zone, order, classzone_idx);
2034 gfp_to_alloc_flags(gfp_t gfp_mask)
2036 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2037 const gfp_t wait = gfp_mask & __GFP_WAIT;
2039 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2040 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2043 * The caller may dip into page reserves a bit more if the caller
2044 * cannot run direct reclaim, or if the caller has realtime scheduling
2045 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2046 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2048 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2052 * Not worth trying to allocate harder for
2053 * __GFP_NOMEMALLOC even if it can't schedule.
2055 if (!(gfp_mask & __GFP_NOMEMALLOC))
2056 alloc_flags |= ALLOC_HARDER;
2058 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2059 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2061 alloc_flags &= ~ALLOC_CPUSET;
2062 } else if (unlikely(rt_task(current)) && !in_interrupt())
2063 alloc_flags |= ALLOC_HARDER;
2065 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2066 if (!in_interrupt() &&
2067 ((current->flags & PF_MEMALLOC) ||
2068 unlikely(test_thread_flag(TIF_MEMDIE))))
2069 alloc_flags |= ALLOC_NO_WATERMARKS;
2075 static inline struct page *
2076 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2077 struct zonelist *zonelist, enum zone_type high_zoneidx,
2078 nodemask_t *nodemask, struct zone *preferred_zone,
2081 const gfp_t wait = gfp_mask & __GFP_WAIT;
2082 struct page *page = NULL;
2084 unsigned long pages_reclaimed = 0;
2085 unsigned long did_some_progress;
2086 bool sync_migration = false;
2089 * In the slowpath, we sanity check order to avoid ever trying to
2090 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2091 * be using allocators in order of preference for an area that is
2094 if (order >= MAX_ORDER) {
2095 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2100 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2101 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2102 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2103 * using a larger set of nodes after it has established that the
2104 * allowed per node queues are empty and that nodes are
2107 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2111 if (!(gfp_mask & __GFP_NO_KSWAPD))
2112 wake_all_kswapd(order, zonelist, high_zoneidx,
2113 zone_idx(preferred_zone));
2116 * OK, we're below the kswapd watermark and have kicked background
2117 * reclaim. Now things get more complex, so set up alloc_flags according
2118 * to how we want to proceed.
2120 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2123 * Find the true preferred zone if the allocation is unconstrained by
2126 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2127 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2131 /* This is the last chance, in general, before the goto nopage. */
2132 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2133 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2134 preferred_zone, migratetype);
2138 /* Allocate without watermarks if the context allows */
2139 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2140 page = __alloc_pages_high_priority(gfp_mask, order,
2141 zonelist, high_zoneidx, nodemask,
2142 preferred_zone, migratetype);
2147 /* Atomic allocations - we can't balance anything */
2151 /* Avoid recursion of direct reclaim */
2152 if (current->flags & PF_MEMALLOC)
2155 /* Avoid allocations with no watermarks from looping endlessly */
2156 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2160 * Try direct compaction. The first pass is asynchronous. Subsequent
2161 * attempts after direct reclaim are synchronous
2163 page = __alloc_pages_direct_compact(gfp_mask, order,
2164 zonelist, high_zoneidx,
2166 alloc_flags, preferred_zone,
2167 migratetype, &did_some_progress,
2171 sync_migration = true;
2173 /* Try direct reclaim and then allocating */
2174 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2175 zonelist, high_zoneidx,
2177 alloc_flags, preferred_zone,
2178 migratetype, &did_some_progress);
2183 * If we failed to make any progress reclaiming, then we are
2184 * running out of options and have to consider going OOM
2186 if (!did_some_progress) {
2187 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2188 if (oom_killer_disabled)
2190 page = __alloc_pages_may_oom(gfp_mask, order,
2191 zonelist, high_zoneidx,
2192 nodemask, preferred_zone,
2197 if (!(gfp_mask & __GFP_NOFAIL)) {
2199 * The oom killer is not called for high-order
2200 * allocations that may fail, so if no progress
2201 * is being made, there are no other options and
2202 * retrying is unlikely to help.
2204 if (order > PAGE_ALLOC_COSTLY_ORDER)
2207 * The oom killer is not called for lowmem
2208 * allocations to prevent needlessly killing
2211 if (high_zoneidx < ZONE_NORMAL)
2219 /* Check if we should retry the allocation */
2220 pages_reclaimed += did_some_progress;
2221 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2222 /* Wait for some write requests to complete then retry */
2223 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2227 * High-order allocations do not necessarily loop after
2228 * direct reclaim and reclaim/compaction depends on compaction
2229 * being called after reclaim so call directly if necessary
2231 page = __alloc_pages_direct_compact(gfp_mask, order,
2232 zonelist, high_zoneidx,
2234 alloc_flags, preferred_zone,
2235 migratetype, &did_some_progress,
2242 warn_alloc_failed(gfp_mask, order, NULL);
2245 if (kmemcheck_enabled)
2246 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2252 * This is the 'heart' of the zoned buddy allocator.
2255 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2256 struct zonelist *zonelist, nodemask_t *nodemask)
2258 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2259 struct zone *preferred_zone;
2261 int migratetype = allocflags_to_migratetype(gfp_mask);
2263 gfp_mask &= gfp_allowed_mask;
2265 lockdep_trace_alloc(gfp_mask);
2267 might_sleep_if(gfp_mask & __GFP_WAIT);
2269 if (should_fail_alloc_page(gfp_mask, order))
2273 * Check the zones suitable for the gfp_mask contain at least one
2274 * valid zone. It's possible to have an empty zonelist as a result
2275 * of GFP_THISNODE and a memoryless node
2277 if (unlikely(!zonelist->_zonerefs->zone))
2281 /* The preferred zone is used for statistics later */
2282 first_zones_zonelist(zonelist, high_zoneidx,
2283 nodemask ? : &cpuset_current_mems_allowed,
2285 if (!preferred_zone) {
2290 /* First allocation attempt */
2291 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2292 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2293 preferred_zone, migratetype);
2294 if (unlikely(!page))
2295 page = __alloc_pages_slowpath(gfp_mask, order,
2296 zonelist, high_zoneidx, nodemask,
2297 preferred_zone, migratetype);
2300 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2303 EXPORT_SYMBOL(__alloc_pages_nodemask);
2306 * Common helper functions.
2308 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2313 * __get_free_pages() returns a 32-bit address, which cannot represent
2316 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2318 page = alloc_pages(gfp_mask, order);
2321 return (unsigned long) page_address(page);
2323 EXPORT_SYMBOL(__get_free_pages);
2325 unsigned long get_zeroed_page(gfp_t gfp_mask)
2327 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2329 EXPORT_SYMBOL(get_zeroed_page);
2331 void __pagevec_free(struct pagevec *pvec)
2333 int i = pagevec_count(pvec);
2336 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2337 free_hot_cold_page(pvec->pages[i], pvec->cold);
2341 void __free_pages(struct page *page, unsigned int order)
2343 if (put_page_testzero(page)) {
2345 free_hot_cold_page(page, 0);
2347 __free_pages_ok(page, order);
2351 EXPORT_SYMBOL(__free_pages);
2353 void free_pages(unsigned long addr, unsigned int order)
2356 VM_BUG_ON(!virt_addr_valid((void *)addr));
2357 __free_pages(virt_to_page((void *)addr), order);
2361 EXPORT_SYMBOL(free_pages);
2363 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2366 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2367 unsigned long used = addr + PAGE_ALIGN(size);
2369 split_page(virt_to_page((void *)addr), order);
2370 while (used < alloc_end) {
2375 return (void *)addr;
2379 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2380 * @size: the number of bytes to allocate
2381 * @gfp_mask: GFP flags for the allocation
2383 * This function is similar to alloc_pages(), except that it allocates the
2384 * minimum number of pages to satisfy the request. alloc_pages() can only
2385 * allocate memory in power-of-two pages.
2387 * This function is also limited by MAX_ORDER.
2389 * Memory allocated by this function must be released by free_pages_exact().
2391 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2393 unsigned int order = get_order(size);
2396 addr = __get_free_pages(gfp_mask, order);
2397 return make_alloc_exact(addr, order, size);
2399 EXPORT_SYMBOL(alloc_pages_exact);
2402 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2404 * @nid: the preferred node ID where memory should be allocated
2405 * @size: the number of bytes to allocate
2406 * @gfp_mask: GFP flags for the allocation
2408 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2410 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2413 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2415 unsigned order = get_order(size);
2416 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2419 return make_alloc_exact((unsigned long)page_address(p), order, size);
2421 EXPORT_SYMBOL(alloc_pages_exact_nid);
2424 * free_pages_exact - release memory allocated via alloc_pages_exact()
2425 * @virt: the value returned by alloc_pages_exact.
2426 * @size: size of allocation, same value as passed to alloc_pages_exact().
2428 * Release the memory allocated by a previous call to alloc_pages_exact.
2430 void free_pages_exact(void *virt, size_t size)
2432 unsigned long addr = (unsigned long)virt;
2433 unsigned long end = addr + PAGE_ALIGN(size);
2435 while (addr < end) {
2440 EXPORT_SYMBOL(free_pages_exact);
2442 static unsigned int nr_free_zone_pages(int offset)
2447 /* Just pick one node, since fallback list is circular */
2448 unsigned int sum = 0;
2450 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2452 for_each_zone_zonelist(zone, z, zonelist, offset) {
2453 unsigned long size = zone->present_pages;
2454 unsigned long high = high_wmark_pages(zone);
2463 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2465 unsigned int nr_free_buffer_pages(void)
2467 return nr_free_zone_pages(gfp_zone(GFP_USER));
2469 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2472 * Amount of free RAM allocatable within all zones
2474 unsigned int nr_free_pagecache_pages(void)
2476 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2479 static inline void show_node(struct zone *zone)
2482 printk("Node %d ", zone_to_nid(zone));
2485 void si_meminfo(struct sysinfo *val)
2487 val->totalram = totalram_pages;
2489 val->freeram = global_page_state(NR_FREE_PAGES);
2490 val->bufferram = nr_blockdev_pages();
2491 val->totalhigh = totalhigh_pages;
2492 val->freehigh = nr_free_highpages();
2493 val->mem_unit = PAGE_SIZE;
2496 EXPORT_SYMBOL(si_meminfo);
2499 void si_meminfo_node(struct sysinfo *val, int nid)
2501 pg_data_t *pgdat = NODE_DATA(nid);
2503 val->totalram = pgdat->node_present_pages;
2504 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2505 #ifdef CONFIG_HIGHMEM
2506 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2507 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2513 val->mem_unit = PAGE_SIZE;
2518 * Determine whether the node should be displayed or not, depending on whether
2519 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2521 bool skip_free_areas_node(unsigned int flags, int nid)
2525 if (!(flags & SHOW_MEM_FILTER_NODES))
2529 ret = !node_isset(nid, cpuset_current_mems_allowed);
2535 #define K(x) ((x) << (PAGE_SHIFT-10))
2538 * Show free area list (used inside shift_scroll-lock stuff)
2539 * We also calculate the percentage fragmentation. We do this by counting the
2540 * memory on each free list with the exception of the first item on the list.
2541 * Suppresses nodes that are not allowed by current's cpuset if
2542 * SHOW_MEM_FILTER_NODES is passed.
2544 void show_free_areas(unsigned int filter)
2549 for_each_populated_zone(zone) {
2550 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2553 printk("%s per-cpu:\n", zone->name);
2555 for_each_online_cpu(cpu) {
2556 struct per_cpu_pageset *pageset;
2558 pageset = per_cpu_ptr(zone->pageset, cpu);
2560 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2561 cpu, pageset->pcp.high,
2562 pageset->pcp.batch, pageset->pcp.count);
2566 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2567 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2569 " dirty:%lu writeback:%lu unstable:%lu\n"
2570 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2571 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2572 global_page_state(NR_ACTIVE_ANON),
2573 global_page_state(NR_INACTIVE_ANON),
2574 global_page_state(NR_ISOLATED_ANON),
2575 global_page_state(NR_ACTIVE_FILE),
2576 global_page_state(NR_INACTIVE_FILE),
2577 global_page_state(NR_ISOLATED_FILE),
2578 global_page_state(NR_UNEVICTABLE),
2579 global_page_state(NR_FILE_DIRTY),
2580 global_page_state(NR_WRITEBACK),
2581 global_page_state(NR_UNSTABLE_NFS),
2582 global_page_state(NR_FREE_PAGES),
2583 global_page_state(NR_SLAB_RECLAIMABLE),
2584 global_page_state(NR_SLAB_UNRECLAIMABLE),
2585 global_page_state(NR_FILE_MAPPED),
2586 global_page_state(NR_SHMEM),
2587 global_page_state(NR_PAGETABLE),
2588 global_page_state(NR_BOUNCE));
2590 for_each_populated_zone(zone) {
2593 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2601 " active_anon:%lukB"
2602 " inactive_anon:%lukB"
2603 " active_file:%lukB"
2604 " inactive_file:%lukB"
2605 " unevictable:%lukB"
2606 " isolated(anon):%lukB"
2607 " isolated(file):%lukB"
2614 " slab_reclaimable:%lukB"
2615 " slab_unreclaimable:%lukB"
2616 " kernel_stack:%lukB"
2620 " writeback_tmp:%lukB"
2621 " pages_scanned:%lu"
2622 " all_unreclaimable? %s"
2625 K(zone_page_state(zone, NR_FREE_PAGES)),
2626 K(min_wmark_pages(zone)),
2627 K(low_wmark_pages(zone)),
2628 K(high_wmark_pages(zone)),
2629 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2630 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2631 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2632 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2633 K(zone_page_state(zone, NR_UNEVICTABLE)),
2634 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2635 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2636 K(zone->present_pages),
2637 K(zone_page_state(zone, NR_MLOCK)),
2638 K(zone_page_state(zone, NR_FILE_DIRTY)),
2639 K(zone_page_state(zone, NR_WRITEBACK)),
2640 K(zone_page_state(zone, NR_FILE_MAPPED)),
2641 K(zone_page_state(zone, NR_SHMEM)),
2642 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2643 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2644 zone_page_state(zone, NR_KERNEL_STACK) *
2646 K(zone_page_state(zone, NR_PAGETABLE)),
2647 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2648 K(zone_page_state(zone, NR_BOUNCE)),
2649 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2650 zone->pages_scanned,
2651 (zone->all_unreclaimable ? "yes" : "no")
2653 printk("lowmem_reserve[]:");
2654 for (i = 0; i < MAX_NR_ZONES; i++)
2655 printk(" %lu", zone->lowmem_reserve[i]);
2659 for_each_populated_zone(zone) {
2660 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2662 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2665 printk("%s: ", zone->name);
2667 spin_lock_irqsave(&zone->lock, flags);
2668 for (order = 0; order < MAX_ORDER; order++) {
2669 nr[order] = zone->free_area[order].nr_free;
2670 total += nr[order] << order;
2672 spin_unlock_irqrestore(&zone->lock, flags);
2673 for (order = 0; order < MAX_ORDER; order++)
2674 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2675 printk("= %lukB\n", K(total));
2678 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2680 show_swap_cache_info();
2683 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2685 zoneref->zone = zone;
2686 zoneref->zone_idx = zone_idx(zone);
2690 * Builds allocation fallback zone lists.
2692 * Add all populated zones of a node to the zonelist.
2694 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2695 int nr_zones, enum zone_type zone_type)
2699 BUG_ON(zone_type >= MAX_NR_ZONES);
2704 zone = pgdat->node_zones + zone_type;
2705 if (populated_zone(zone)) {
2706 zoneref_set_zone(zone,
2707 &zonelist->_zonerefs[nr_zones++]);
2708 check_highest_zone(zone_type);
2711 } while (zone_type);
2718 * 0 = automatic detection of better ordering.
2719 * 1 = order by ([node] distance, -zonetype)
2720 * 2 = order by (-zonetype, [node] distance)
2722 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2723 * the same zonelist. So only NUMA can configure this param.
2725 #define ZONELIST_ORDER_DEFAULT 0
2726 #define ZONELIST_ORDER_NODE 1
2727 #define ZONELIST_ORDER_ZONE 2
2729 /* zonelist order in the kernel.
2730 * set_zonelist_order() will set this to NODE or ZONE.
2732 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2733 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2737 /* The value user specified ....changed by config */
2738 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2739 /* string for sysctl */
2740 #define NUMA_ZONELIST_ORDER_LEN 16
2741 char numa_zonelist_order[16] = "default";
2744 * interface for configure zonelist ordering.
2745 * command line option "numa_zonelist_order"
2746 * = "[dD]efault - default, automatic configuration.
2747 * = "[nN]ode - order by node locality, then by zone within node
2748 * = "[zZ]one - order by zone, then by locality within zone
2751 static int __parse_numa_zonelist_order(char *s)
2753 if (*s == 'd' || *s == 'D') {
2754 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2755 } else if (*s == 'n' || *s == 'N') {
2756 user_zonelist_order = ZONELIST_ORDER_NODE;
2757 } else if (*s == 'z' || *s == 'Z') {
2758 user_zonelist_order = ZONELIST_ORDER_ZONE;
2761 "Ignoring invalid numa_zonelist_order value: "
2768 static __init int setup_numa_zonelist_order(char *s)
2775 ret = __parse_numa_zonelist_order(s);
2777 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2781 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2784 * sysctl handler for numa_zonelist_order
2786 int numa_zonelist_order_handler(ctl_table *table, int write,
2787 void __user *buffer, size_t *length,
2790 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2792 static DEFINE_MUTEX(zl_order_mutex);
2794 mutex_lock(&zl_order_mutex);
2796 strcpy(saved_string, (char*)table->data);
2797 ret = proc_dostring(table, write, buffer, length, ppos);
2801 int oldval = user_zonelist_order;
2802 if (__parse_numa_zonelist_order((char*)table->data)) {
2804 * bogus value. restore saved string
2806 strncpy((char*)table->data, saved_string,
2807 NUMA_ZONELIST_ORDER_LEN);
2808 user_zonelist_order = oldval;
2809 } else if (oldval != user_zonelist_order) {
2810 mutex_lock(&zonelists_mutex);
2811 build_all_zonelists(NULL);
2812 mutex_unlock(&zonelists_mutex);
2816 mutex_unlock(&zl_order_mutex);
2821 #define MAX_NODE_LOAD (nr_online_nodes)
2822 static int node_load[MAX_NUMNODES];
2825 * find_next_best_node - find the next node that should appear in a given node's fallback list
2826 * @node: node whose fallback list we're appending
2827 * @used_node_mask: nodemask_t of already used nodes
2829 * We use a number of factors to determine which is the next node that should
2830 * appear on a given node's fallback list. The node should not have appeared
2831 * already in @node's fallback list, and it should be the next closest node
2832 * according to the distance array (which contains arbitrary distance values
2833 * from each node to each node in the system), and should also prefer nodes
2834 * with no CPUs, since presumably they'll have very little allocation pressure
2835 * on them otherwise.
2836 * It returns -1 if no node is found.
2838 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2841 int min_val = INT_MAX;
2843 const struct cpumask *tmp = cpumask_of_node(0);
2845 /* Use the local node if we haven't already */
2846 if (!node_isset(node, *used_node_mask)) {
2847 node_set(node, *used_node_mask);
2851 for_each_node_state(n, N_HIGH_MEMORY) {
2853 /* Don't want a node to appear more than once */
2854 if (node_isset(n, *used_node_mask))
2857 /* Use the distance array to find the distance */
2858 val = node_distance(node, n);
2860 /* Penalize nodes under us ("prefer the next node") */
2863 /* Give preference to headless and unused nodes */
2864 tmp = cpumask_of_node(n);
2865 if (!cpumask_empty(tmp))
2866 val += PENALTY_FOR_NODE_WITH_CPUS;
2868 /* Slight preference for less loaded node */
2869 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2870 val += node_load[n];
2872 if (val < min_val) {
2879 node_set(best_node, *used_node_mask);
2886 * Build zonelists ordered by node and zones within node.
2887 * This results in maximum locality--normal zone overflows into local
2888 * DMA zone, if any--but risks exhausting DMA zone.
2890 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2893 struct zonelist *zonelist;
2895 zonelist = &pgdat->node_zonelists[0];
2896 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2898 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2900 zonelist->_zonerefs[j].zone = NULL;
2901 zonelist->_zonerefs[j].zone_idx = 0;
2905 * Build gfp_thisnode zonelists
2907 static void build_thisnode_zonelists(pg_data_t *pgdat)
2910 struct zonelist *zonelist;
2912 zonelist = &pgdat->node_zonelists[1];
2913 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2914 zonelist->_zonerefs[j].zone = NULL;
2915 zonelist->_zonerefs[j].zone_idx = 0;
2919 * Build zonelists ordered by zone and nodes within zones.
2920 * This results in conserving DMA zone[s] until all Normal memory is
2921 * exhausted, but results in overflowing to remote node while memory
2922 * may still exist in local DMA zone.
2924 static int node_order[MAX_NUMNODES];
2926 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2929 int zone_type; /* needs to be signed */
2931 struct zonelist *zonelist;
2933 zonelist = &pgdat->node_zonelists[0];
2935 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2936 for (j = 0; j < nr_nodes; j++) {
2937 node = node_order[j];
2938 z = &NODE_DATA(node)->node_zones[zone_type];
2939 if (populated_zone(z)) {
2941 &zonelist->_zonerefs[pos++]);
2942 check_highest_zone(zone_type);
2946 zonelist->_zonerefs[pos].zone = NULL;
2947 zonelist->_zonerefs[pos].zone_idx = 0;
2950 static int default_zonelist_order(void)
2953 unsigned long low_kmem_size,total_size;
2957 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2958 * If they are really small and used heavily, the system can fall
2959 * into OOM very easily.
2960 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2962 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2965 for_each_online_node(nid) {
2966 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2967 z = &NODE_DATA(nid)->node_zones[zone_type];
2968 if (populated_zone(z)) {
2969 if (zone_type < ZONE_NORMAL)
2970 low_kmem_size += z->present_pages;
2971 total_size += z->present_pages;
2972 } else if (zone_type == ZONE_NORMAL) {
2974 * If any node has only lowmem, then node order
2975 * is preferred to allow kernel allocations
2976 * locally; otherwise, they can easily infringe
2977 * on other nodes when there is an abundance of
2978 * lowmem available to allocate from.
2980 return ZONELIST_ORDER_NODE;
2984 if (!low_kmem_size || /* there are no DMA area. */
2985 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2986 return ZONELIST_ORDER_NODE;
2988 * look into each node's config.
2989 * If there is a node whose DMA/DMA32 memory is very big area on
2990 * local memory, NODE_ORDER may be suitable.
2992 average_size = total_size /
2993 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2994 for_each_online_node(nid) {
2997 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2998 z = &NODE_DATA(nid)->node_zones[zone_type];
2999 if (populated_zone(z)) {
3000 if (zone_type < ZONE_NORMAL)
3001 low_kmem_size += z->present_pages;
3002 total_size += z->present_pages;
3005 if (low_kmem_size &&
3006 total_size > average_size && /* ignore small node */
3007 low_kmem_size > total_size * 70/100)
3008 return ZONELIST_ORDER_NODE;
3010 return ZONELIST_ORDER_ZONE;
3013 static void set_zonelist_order(void)
3015 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3016 current_zonelist_order = default_zonelist_order();
3018 current_zonelist_order = user_zonelist_order;
3021 static void build_zonelists(pg_data_t *pgdat)
3025 nodemask_t used_mask;
3026 int local_node, prev_node;
3027 struct zonelist *zonelist;
3028 int order = current_zonelist_order;
3030 /* initialize zonelists */
3031 for (i = 0; i < MAX_ZONELISTS; i++) {
3032 zonelist = pgdat->node_zonelists + i;
3033 zonelist->_zonerefs[0].zone = NULL;
3034 zonelist->_zonerefs[0].zone_idx = 0;
3037 /* NUMA-aware ordering of nodes */
3038 local_node = pgdat->node_id;
3039 load = nr_online_nodes;
3040 prev_node = local_node;
3041 nodes_clear(used_mask);
3043 memset(node_order, 0, sizeof(node_order));
3046 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3047 int distance = node_distance(local_node, node);
3050 * If another node is sufficiently far away then it is better
3051 * to reclaim pages in a zone before going off node.
3053 if (distance > RECLAIM_DISTANCE)
3054 zone_reclaim_mode = 1;
3057 * We don't want to pressure a particular node.
3058 * So adding penalty to the first node in same
3059 * distance group to make it round-robin.
3061 if (distance != node_distance(local_node, prev_node))
3062 node_load[node] = load;
3066 if (order == ZONELIST_ORDER_NODE)
3067 build_zonelists_in_node_order(pgdat, node);
3069 node_order[j++] = node; /* remember order */
3072 if (order == ZONELIST_ORDER_ZONE) {
3073 /* calculate node order -- i.e., DMA last! */
3074 build_zonelists_in_zone_order(pgdat, j);
3077 build_thisnode_zonelists(pgdat);
3080 /* Construct the zonelist performance cache - see further mmzone.h */
3081 static void build_zonelist_cache(pg_data_t *pgdat)
3083 struct zonelist *zonelist;
3084 struct zonelist_cache *zlc;
3087 zonelist = &pgdat->node_zonelists[0];
3088 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3089 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3090 for (z = zonelist->_zonerefs; z->zone; z++)
3091 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3094 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3096 * Return node id of node used for "local" allocations.
3097 * I.e., first node id of first zone in arg node's generic zonelist.
3098 * Used for initializing percpu 'numa_mem', which is used primarily
3099 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3101 int local_memory_node(int node)
3105 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3106 gfp_zone(GFP_KERNEL),
3113 #else /* CONFIG_NUMA */
3115 static void set_zonelist_order(void)
3117 current_zonelist_order = ZONELIST_ORDER_ZONE;
3120 static void build_zonelists(pg_data_t *pgdat)
3122 int node, local_node;
3124 struct zonelist *zonelist;
3126 local_node = pgdat->node_id;
3128 zonelist = &pgdat->node_zonelists[0];
3129 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3132 * Now we build the zonelist so that it contains the zones
3133 * of all the other nodes.
3134 * We don't want to pressure a particular node, so when
3135 * building the zones for node N, we make sure that the
3136 * zones coming right after the local ones are those from
3137 * node N+1 (modulo N)
3139 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3140 if (!node_online(node))
3142 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3145 for (node = 0; node < local_node; node++) {
3146 if (!node_online(node))
3148 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3152 zonelist->_zonerefs[j].zone = NULL;
3153 zonelist->_zonerefs[j].zone_idx = 0;
3156 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3157 static void build_zonelist_cache(pg_data_t *pgdat)
3159 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3162 #endif /* CONFIG_NUMA */
3165 * Boot pageset table. One per cpu which is going to be used for all
3166 * zones and all nodes. The parameters will be set in such a way
3167 * that an item put on a list will immediately be handed over to
3168 * the buddy list. This is safe since pageset manipulation is done
3169 * with interrupts disabled.
3171 * The boot_pagesets must be kept even after bootup is complete for
3172 * unused processors and/or zones. They do play a role for bootstrapping
3173 * hotplugged processors.
3175 * zoneinfo_show() and maybe other functions do
3176 * not check if the processor is online before following the pageset pointer.
3177 * Other parts of the kernel may not check if the zone is available.
3179 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3180 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3181 static void setup_zone_pageset(struct zone *zone);
3184 * Global mutex to protect against size modification of zonelists
3185 * as well as to serialize pageset setup for the new populated zone.
3187 DEFINE_MUTEX(zonelists_mutex);
3189 /* return values int ....just for stop_machine() */
3190 static __init_refok int __build_all_zonelists(void *data)
3196 memset(node_load, 0, sizeof(node_load));
3198 for_each_online_node(nid) {
3199 pg_data_t *pgdat = NODE_DATA(nid);
3201 build_zonelists(pgdat);
3202 build_zonelist_cache(pgdat);
3206 * Initialize the boot_pagesets that are going to be used
3207 * for bootstrapping processors. The real pagesets for
3208 * each zone will be allocated later when the per cpu
3209 * allocator is available.
3211 * boot_pagesets are used also for bootstrapping offline
3212 * cpus if the system is already booted because the pagesets
3213 * are needed to initialize allocators on a specific cpu too.
3214 * F.e. the percpu allocator needs the page allocator which
3215 * needs the percpu allocator in order to allocate its pagesets
3216 * (a chicken-egg dilemma).
3218 for_each_possible_cpu(cpu) {
3219 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3221 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3223 * We now know the "local memory node" for each node--
3224 * i.e., the node of the first zone in the generic zonelist.
3225 * Set up numa_mem percpu variable for on-line cpus. During
3226 * boot, only the boot cpu should be on-line; we'll init the
3227 * secondary cpus' numa_mem as they come on-line. During
3228 * node/memory hotplug, we'll fixup all on-line cpus.
3230 if (cpu_online(cpu))
3231 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3239 * Called with zonelists_mutex held always
3240 * unless system_state == SYSTEM_BOOTING.
3242 void __ref build_all_zonelists(void *data)
3244 set_zonelist_order();
3246 if (system_state == SYSTEM_BOOTING) {
3247 __build_all_zonelists(NULL);
3248 mminit_verify_zonelist();
3249 cpuset_init_current_mems_allowed();
3251 /* we have to stop all cpus to guarantee there is no user
3253 #ifdef CONFIG_MEMORY_HOTPLUG
3255 setup_zone_pageset((struct zone *)data);
3257 stop_machine(__build_all_zonelists, NULL, NULL);
3258 /* cpuset refresh routine should be here */
3260 vm_total_pages = nr_free_pagecache_pages();
3262 * Disable grouping by mobility if the number of pages in the
3263 * system is too low to allow the mechanism to work. It would be
3264 * more accurate, but expensive to check per-zone. This check is
3265 * made on memory-hotadd so a system can start with mobility
3266 * disabled and enable it later
3268 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3269 page_group_by_mobility_disabled = 1;
3271 page_group_by_mobility_disabled = 0;
3273 printk("Built %i zonelists in %s order, mobility grouping %s. "
3274 "Total pages: %ld\n",
3276 zonelist_order_name[current_zonelist_order],
3277 page_group_by_mobility_disabled ? "off" : "on",
3280 printk("Policy zone: %s\n", zone_names[policy_zone]);
3285 * Helper functions to size the waitqueue hash table.
3286 * Essentially these want to choose hash table sizes sufficiently
3287 * large so that collisions trying to wait on pages are rare.
3288 * But in fact, the number of active page waitqueues on typical
3289 * systems is ridiculously low, less than 200. So this is even
3290 * conservative, even though it seems large.
3292 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3293 * waitqueues, i.e. the size of the waitq table given the number of pages.
3295 #define PAGES_PER_WAITQUEUE 256
3297 #ifndef CONFIG_MEMORY_HOTPLUG
3298 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3300 unsigned long size = 1;
3302 pages /= PAGES_PER_WAITQUEUE;
3304 while (size < pages)
3308 * Once we have dozens or even hundreds of threads sleeping
3309 * on IO we've got bigger problems than wait queue collision.
3310 * Limit the size of the wait table to a reasonable size.
3312 size = min(size, 4096UL);
3314 return max(size, 4UL);
3318 * A zone's size might be changed by hot-add, so it is not possible to determine
3319 * a suitable size for its wait_table. So we use the maximum size now.
3321 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3323 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3324 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3325 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3327 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3328 * or more by the traditional way. (See above). It equals:
3330 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3331 * ia64(16K page size) : = ( 8G + 4M)byte.
3332 * powerpc (64K page size) : = (32G +16M)byte.
3334 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3341 * This is an integer logarithm so that shifts can be used later
3342 * to extract the more random high bits from the multiplicative
3343 * hash function before the remainder is taken.
3345 static inline unsigned long wait_table_bits(unsigned long size)
3350 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3353 * Check if a pageblock contains reserved pages
3355 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3359 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3360 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3367 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3368 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3369 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3370 * higher will lead to a bigger reserve which will get freed as contiguous
3371 * blocks as reclaim kicks in
3373 static void setup_zone_migrate_reserve(struct zone *zone)
3375 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3377 unsigned long block_migratetype;
3381 * Get the start pfn, end pfn and the number of blocks to reserve
3382 * We have to be careful to be aligned to pageblock_nr_pages to
3383 * make sure that we always check pfn_valid for the first page in
3386 start_pfn = zone->zone_start_pfn;
3387 end_pfn = start_pfn + zone->spanned_pages;
3388 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3389 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3393 * Reserve blocks are generally in place to help high-order atomic
3394 * allocations that are short-lived. A min_free_kbytes value that
3395 * would result in more than 2 reserve blocks for atomic allocations
3396 * is assumed to be in place to help anti-fragmentation for the
3397 * future allocation of hugepages at runtime.
3399 reserve = min(2, reserve);
3401 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3402 if (!pfn_valid(pfn))
3404 page = pfn_to_page(pfn);
3406 /* Watch out for overlapping nodes */
3407 if (page_to_nid(page) != zone_to_nid(zone))
3410 /* Blocks with reserved pages will never free, skip them. */
3411 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3412 if (pageblock_is_reserved(pfn, block_end_pfn))
3415 block_migratetype = get_pageblock_migratetype(page);
3417 /* If this block is reserved, account for it */
3418 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3423 /* Suitable for reserving if this block is movable */
3424 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3425 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3426 move_freepages_block(zone, page, MIGRATE_RESERVE);
3432 * If the reserve is met and this is a previous reserved block,
3435 if (block_migratetype == MIGRATE_RESERVE) {
3436 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3437 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3443 * Initially all pages are reserved - free ones are freed
3444 * up by free_all_bootmem() once the early boot process is
3445 * done. Non-atomic initialization, single-pass.
3447 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3448 unsigned long start_pfn, enum memmap_context context)
3451 unsigned long end_pfn = start_pfn + size;
3455 if (highest_memmap_pfn < end_pfn - 1)
3456 highest_memmap_pfn = end_pfn - 1;
3458 z = &NODE_DATA(nid)->node_zones[zone];
3459 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3461 * There can be holes in boot-time mem_map[]s
3462 * handed to this function. They do not
3463 * exist on hotplugged memory.
3465 if (context == MEMMAP_EARLY) {
3466 if (!early_pfn_valid(pfn))
3468 if (!early_pfn_in_nid(pfn, nid))
3471 page = pfn_to_page(pfn);
3472 set_page_links(page, zone, nid, pfn);
3473 mminit_verify_page_links(page, zone, nid, pfn);
3474 init_page_count(page);
3475 reset_page_mapcount(page);
3476 SetPageReserved(page);
3478 * Mark the block movable so that blocks are reserved for
3479 * movable at startup. This will force kernel allocations
3480 * to reserve their blocks rather than leaking throughout
3481 * the address space during boot when many long-lived
3482 * kernel allocations are made. Later some blocks near
3483 * the start are marked MIGRATE_RESERVE by
3484 * setup_zone_migrate_reserve()
3486 * bitmap is created for zone's valid pfn range. but memmap
3487 * can be created for invalid pages (for alignment)
3488 * check here not to call set_pageblock_migratetype() against
3491 if ((z->zone_start_pfn <= pfn)
3492 && (pfn < z->zone_start_pfn + z->spanned_pages)
3493 && !(pfn & (pageblock_nr_pages - 1)))
3494 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3496 INIT_LIST_HEAD(&page->lru);
3497 #ifdef WANT_PAGE_VIRTUAL
3498 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3499 if (!is_highmem_idx(zone))
3500 set_page_address(page, __va(pfn << PAGE_SHIFT));
3505 static void __meminit zone_init_free_lists(struct zone *zone)
3508 for_each_migratetype_order(order, t) {
3509 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3510 zone->free_area[order].nr_free = 0;
3514 #ifndef __HAVE_ARCH_MEMMAP_INIT
3515 #define memmap_init(size, nid, zone, start_pfn) \
3516 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3519 static int zone_batchsize(struct zone *zone)
3525 * The per-cpu-pages pools are set to around 1000th of the
3526 * size of the zone. But no more than 1/2 of a meg.
3528 * OK, so we don't know how big the cache is. So guess.
3530 batch = zone->present_pages / 1024;
3531 if (batch * PAGE_SIZE > 512 * 1024)
3532 batch = (512 * 1024) / PAGE_SIZE;
3533 batch /= 4; /* We effectively *= 4 below */
3538 * Clamp the batch to a 2^n - 1 value. Having a power
3539 * of 2 value was found to be more likely to have
3540 * suboptimal cache aliasing properties in some cases.
3542 * For example if 2 tasks are alternately allocating
3543 * batches of pages, one task can end up with a lot
3544 * of pages of one half of the possible page colors
3545 * and the other with pages of the other colors.
3547 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3552 /* The deferral and batching of frees should be suppressed under NOMMU
3555 * The problem is that NOMMU needs to be able to allocate large chunks
3556 * of contiguous memory as there's no hardware page translation to
3557 * assemble apparent contiguous memory from discontiguous pages.
3559 * Queueing large contiguous runs of pages for batching, however,
3560 * causes the pages to actually be freed in smaller chunks. As there
3561 * can be a significant delay between the individual batches being
3562 * recycled, this leads to the once large chunks of space being
3563 * fragmented and becoming unavailable for high-order allocations.
3569 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3571 struct per_cpu_pages *pcp;
3574 memset(p, 0, sizeof(*p));
3578 pcp->high = 6 * batch;
3579 pcp->batch = max(1UL, 1 * batch);
3580 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3581 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3585 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3586 * to the value high for the pageset p.
3589 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3592 struct per_cpu_pages *pcp;
3596 pcp->batch = max(1UL, high/4);
3597 if ((high/4) > (PAGE_SHIFT * 8))
3598 pcp->batch = PAGE_SHIFT * 8;
3601 static void setup_zone_pageset(struct zone *zone)
3605 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3607 for_each_possible_cpu(cpu) {
3608 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3610 setup_pageset(pcp, zone_batchsize(zone));
3612 if (percpu_pagelist_fraction)
3613 setup_pagelist_highmark(pcp,
3614 (zone->present_pages /
3615 percpu_pagelist_fraction));
3620 * Allocate per cpu pagesets and initialize them.
3621 * Before this call only boot pagesets were available.
3623 void __init setup_per_cpu_pageset(void)
3627 for_each_populated_zone(zone)
3628 setup_zone_pageset(zone);
3631 static noinline __init_refok
3632 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3635 struct pglist_data *pgdat = zone->zone_pgdat;
3639 * The per-page waitqueue mechanism uses hashed waitqueues
3642 zone->wait_table_hash_nr_entries =
3643 wait_table_hash_nr_entries(zone_size_pages);
3644 zone->wait_table_bits =
3645 wait_table_bits(zone->wait_table_hash_nr_entries);
3646 alloc_size = zone->wait_table_hash_nr_entries
3647 * sizeof(wait_queue_head_t);
3649 if (!slab_is_available()) {
3650 zone->wait_table = (wait_queue_head_t *)
3651 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3654 * This case means that a zone whose size was 0 gets new memory
3655 * via memory hot-add.
3656 * But it may be the case that a new node was hot-added. In
3657 * this case vmalloc() will not be able to use this new node's
3658 * memory - this wait_table must be initialized to use this new
3659 * node itself as well.
3660 * To use this new node's memory, further consideration will be
3663 zone->wait_table = vmalloc(alloc_size);
3665 if (!zone->wait_table)
3668 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3669 init_waitqueue_head(zone->wait_table + i);
3674 static int __zone_pcp_update(void *data)
3676 struct zone *zone = data;
3678 unsigned long batch = zone_batchsize(zone), flags;
3680 for_each_possible_cpu(cpu) {
3681 struct per_cpu_pageset *pset;
3682 struct per_cpu_pages *pcp;
3684 pset = per_cpu_ptr(zone->pageset, cpu);
3687 local_irq_save(flags);
3688 free_pcppages_bulk(zone, pcp->count, pcp);
3689 setup_pageset(pset, batch);
3690 local_irq_restore(flags);
3695 void zone_pcp_update(struct zone *zone)
3697 stop_machine(__zone_pcp_update, zone, NULL);
3700 static __meminit void zone_pcp_init(struct zone *zone)
3703 * per cpu subsystem is not up at this point. The following code
3704 * relies on the ability of the linker to provide the
3705 * offset of a (static) per cpu variable into the per cpu area.
3707 zone->pageset = &boot_pageset;
3709 if (zone->present_pages)
3710 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3711 zone->name, zone->present_pages,
3712 zone_batchsize(zone));
3715 __meminit int init_currently_empty_zone(struct zone *zone,
3716 unsigned long zone_start_pfn,
3718 enum memmap_context context)
3720 struct pglist_data *pgdat = zone->zone_pgdat;
3722 ret = zone_wait_table_init(zone, size);
3725 pgdat->nr_zones = zone_idx(zone) + 1;
3727 zone->zone_start_pfn = zone_start_pfn;
3729 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3730 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3732 (unsigned long)zone_idx(zone),
3733 zone_start_pfn, (zone_start_pfn + size));
3735 zone_init_free_lists(zone);
3740 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3742 * Basic iterator support. Return the first range of PFNs for a node
3743 * Note: nid == MAX_NUMNODES returns first region regardless of node
3745 static int __meminit first_active_region_index_in_nid(int nid)
3749 for (i = 0; i < nr_nodemap_entries; i++)
3750 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3757 * Basic iterator support. Return the next active range of PFNs for a node
3758 * Note: nid == MAX_NUMNODES returns next region regardless of node
3760 static int __meminit next_active_region_index_in_nid(int index, int nid)
3762 for (index = index + 1; index < nr_nodemap_entries; index++)
3763 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3769 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3771 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3772 * Architectures may implement their own version but if add_active_range()
3773 * was used and there are no special requirements, this is a convenient
3776 int __meminit __early_pfn_to_nid(unsigned long pfn)
3780 for (i = 0; i < nr_nodemap_entries; i++) {
3781 unsigned long start_pfn = early_node_map[i].start_pfn;
3782 unsigned long end_pfn = early_node_map[i].end_pfn;
3784 if (start_pfn <= pfn && pfn < end_pfn)
3785 return early_node_map[i].nid;
3787 /* This is a memory hole */
3790 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3792 int __meminit early_pfn_to_nid(unsigned long pfn)
3796 nid = __early_pfn_to_nid(pfn);
3799 /* just returns 0 */
3803 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3804 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3808 nid = __early_pfn_to_nid(pfn);
3809 if (nid >= 0 && nid != node)
3815 /* Basic iterator support to walk early_node_map[] */
3816 #define for_each_active_range_index_in_nid(i, nid) \
3817 for (i = first_active_region_index_in_nid(nid); i != -1; \
3818 i = next_active_region_index_in_nid(i, nid))
3821 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3822 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3823 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3825 * If an architecture guarantees that all ranges registered with
3826 * add_active_ranges() contain no holes and may be freed, this
3827 * this function may be used instead of calling free_bootmem() manually.
3829 void __init free_bootmem_with_active_regions(int nid,
3830 unsigned long max_low_pfn)
3834 for_each_active_range_index_in_nid(i, nid) {
3835 unsigned long size_pages = 0;
3836 unsigned long end_pfn = early_node_map[i].end_pfn;
3838 if (early_node_map[i].start_pfn >= max_low_pfn)
3841 if (end_pfn > max_low_pfn)
3842 end_pfn = max_low_pfn;
3844 size_pages = end_pfn - early_node_map[i].start_pfn;
3845 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3846 PFN_PHYS(early_node_map[i].start_pfn),
3847 size_pages << PAGE_SHIFT);
3851 #ifdef CONFIG_HAVE_MEMBLOCK
3853 * Basic iterator support. Return the last range of PFNs for a node
3854 * Note: nid == MAX_NUMNODES returns last region regardless of node
3856 static int __meminit last_active_region_index_in_nid(int nid)
3860 for (i = nr_nodemap_entries - 1; i >= 0; i--)
3861 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3868 * Basic iterator support. Return the previous active range of PFNs for a node
3869 * Note: nid == MAX_NUMNODES returns next region regardless of node
3871 static int __meminit previous_active_region_index_in_nid(int index, int nid)
3873 for (index = index - 1; index >= 0; index--)
3874 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3880 #define for_each_active_range_index_in_nid_reverse(i, nid) \
3881 for (i = last_active_region_index_in_nid(nid); i != -1; \
3882 i = previous_active_region_index_in_nid(i, nid))
3884 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3885 u64 goal, u64 limit)
3889 /* Need to go over early_node_map to find out good range for node */
3890 for_each_active_range_index_in_nid_reverse(i, nid) {
3892 u64 ei_start, ei_last;
3893 u64 final_start, final_end;
3895 ei_last = early_node_map[i].end_pfn;
3896 ei_last <<= PAGE_SHIFT;
3897 ei_start = early_node_map[i].start_pfn;
3898 ei_start <<= PAGE_SHIFT;
3900 final_start = max(ei_start, goal);
3901 final_end = min(ei_last, limit);
3903 if (final_start >= final_end)
3906 addr = memblock_find_in_range(final_start, final_end, size, align);
3908 if (addr == MEMBLOCK_ERROR)
3914 return MEMBLOCK_ERROR;
3918 int __init add_from_early_node_map(struct range *range, int az,
3919 int nr_range, int nid)
3924 /* need to go over early_node_map to find out good range for node */
3925 for_each_active_range_index_in_nid(i, nid) {
3926 start = early_node_map[i].start_pfn;
3927 end = early_node_map[i].end_pfn;
3928 nr_range = add_range(range, az, nr_range, start, end);
3933 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3938 for_each_active_range_index_in_nid(i, nid) {
3939 ret = work_fn(early_node_map[i].start_pfn,
3940 early_node_map[i].end_pfn, data);
3946 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3947 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3949 * If an architecture guarantees that all ranges registered with
3950 * add_active_ranges() contain no holes and may be freed, this
3951 * function may be used instead of calling memory_present() manually.
3953 void __init sparse_memory_present_with_active_regions(int nid)
3957 for_each_active_range_index_in_nid(i, nid)
3958 memory_present(early_node_map[i].nid,
3959 early_node_map[i].start_pfn,
3960 early_node_map[i].end_pfn);
3964 * get_pfn_range_for_nid - Return the start and end page frames for a node
3965 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3966 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3967 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3969 * It returns the start and end page frame of a node based on information
3970 * provided by an arch calling add_active_range(). If called for a node
3971 * with no available memory, a warning is printed and the start and end
3974 void __meminit get_pfn_range_for_nid(unsigned int nid,
3975 unsigned long *start_pfn, unsigned long *end_pfn)
3981 for_each_active_range_index_in_nid(i, nid) {
3982 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3983 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3986 if (*start_pfn == -1UL)
3991 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3992 * assumption is made that zones within a node are ordered in monotonic
3993 * increasing memory addresses so that the "highest" populated zone is used
3995 static void __init find_usable_zone_for_movable(void)
3998 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3999 if (zone_index == ZONE_MOVABLE)
4002 if (arch_zone_highest_possible_pfn[zone_index] >
4003 arch_zone_lowest_possible_pfn[zone_index])
4007 VM_BUG_ON(zone_index == -1);
4008 movable_zone = zone_index;
4012 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4013 * because it is sized independent of architecture. Unlike the other zones,
4014 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4015 * in each node depending on the size of each node and how evenly kernelcore
4016 * is distributed. This helper function adjusts the zone ranges
4017 * provided by the architecture for a given node by using the end of the
4018 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4019 * zones within a node are in order of monotonic increases memory addresses
4021 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4022 unsigned long zone_type,
4023 unsigned long node_start_pfn,
4024 unsigned long node_end_pfn,
4025 unsigned long *zone_start_pfn,
4026 unsigned long *zone_end_pfn)
4028 /* Only adjust if ZONE_MOVABLE is on this node */
4029 if (zone_movable_pfn[nid]) {
4030 /* Size ZONE_MOVABLE */
4031 if (zone_type == ZONE_MOVABLE) {
4032 *zone_start_pfn = zone_movable_pfn[nid];
4033 *zone_end_pfn = min(node_end_pfn,
4034 arch_zone_highest_possible_pfn[movable_zone]);
4036 /* Adjust for ZONE_MOVABLE starting within this range */
4037 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4038 *zone_end_pfn > zone_movable_pfn[nid]) {
4039 *zone_end_pfn = zone_movable_pfn[nid];
4041 /* Check if this whole range is within ZONE_MOVABLE */
4042 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4043 *zone_start_pfn = *zone_end_pfn;
4048 * Return the number of pages a zone spans in a node, including holes
4049 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4051 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4052 unsigned long zone_type,
4053 unsigned long *ignored)
4055 unsigned long node_start_pfn, node_end_pfn;
4056 unsigned long zone_start_pfn, zone_end_pfn;
4058 /* Get the start and end of the node and zone */
4059 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4060 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4061 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4062 adjust_zone_range_for_zone_movable(nid, zone_type,
4063 node_start_pfn, node_end_pfn,
4064 &zone_start_pfn, &zone_end_pfn);
4066 /* Check that this node has pages within the zone's required range */
4067 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4070 /* Move the zone boundaries inside the node if necessary */
4071 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4072 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4074 /* Return the spanned pages */
4075 return zone_end_pfn - zone_start_pfn;
4079 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4080 * then all holes in the requested range will be accounted for.
4082 unsigned long __meminit __absent_pages_in_range(int nid,
4083 unsigned long range_start_pfn,
4084 unsigned long range_end_pfn)
4087 unsigned long prev_end_pfn = 0, hole_pages = 0;
4088 unsigned long start_pfn;
4090 /* Find the end_pfn of the first active range of pfns in the node */
4091 i = first_active_region_index_in_nid(nid);
4095 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4097 /* Account for ranges before physical memory on this node */
4098 if (early_node_map[i].start_pfn > range_start_pfn)
4099 hole_pages = prev_end_pfn - range_start_pfn;
4101 /* Find all holes for the zone within the node */
4102 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4104 /* No need to continue if prev_end_pfn is outside the zone */
4105 if (prev_end_pfn >= range_end_pfn)
4108 /* Make sure the end of the zone is not within the hole */
4109 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4110 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4112 /* Update the hole size cound and move on */
4113 if (start_pfn > range_start_pfn) {
4114 BUG_ON(prev_end_pfn > start_pfn);
4115 hole_pages += start_pfn - prev_end_pfn;
4117 prev_end_pfn = early_node_map[i].end_pfn;
4120 /* Account for ranges past physical memory on this node */
4121 if (range_end_pfn > prev_end_pfn)
4122 hole_pages += range_end_pfn -
4123 max(range_start_pfn, prev_end_pfn);
4129 * absent_pages_in_range - Return number of page frames in holes within a range
4130 * @start_pfn: The start PFN to start searching for holes
4131 * @end_pfn: The end PFN to stop searching for holes
4133 * It returns the number of pages frames in memory holes within a range.
4135 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4136 unsigned long end_pfn)
4138 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4141 /* Return the number of page frames in holes in a zone on a node */
4142 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4143 unsigned long zone_type,
4144 unsigned long *ignored)
4146 unsigned long node_start_pfn, node_end_pfn;
4147 unsigned long zone_start_pfn, zone_end_pfn;
4149 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4150 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4152 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4155 adjust_zone_range_for_zone_movable(nid, zone_type,
4156 node_start_pfn, node_end_pfn,
4157 &zone_start_pfn, &zone_end_pfn);
4158 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4162 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4163 unsigned long zone_type,
4164 unsigned long *zones_size)
4166 return zones_size[zone_type];
4169 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4170 unsigned long zone_type,
4171 unsigned long *zholes_size)
4176 return zholes_size[zone_type];
4181 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4182 unsigned long *zones_size, unsigned long *zholes_size)
4184 unsigned long realtotalpages, totalpages = 0;
4187 for (i = 0; i < MAX_NR_ZONES; i++)
4188 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4190 pgdat->node_spanned_pages = totalpages;
4192 realtotalpages = totalpages;
4193 for (i = 0; i < MAX_NR_ZONES; i++)
4195 zone_absent_pages_in_node(pgdat->node_id, i,
4197 pgdat->node_present_pages = realtotalpages;
4198 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4202 #ifndef CONFIG_SPARSEMEM
4204 * Calculate the size of the zone->blockflags rounded to an unsigned long
4205 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4206 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4207 * round what is now in bits to nearest long in bits, then return it in
4210 static unsigned long __init usemap_size(unsigned long zonesize)
4212 unsigned long usemapsize;
4214 usemapsize = roundup(zonesize, pageblock_nr_pages);
4215 usemapsize = usemapsize >> pageblock_order;
4216 usemapsize *= NR_PAGEBLOCK_BITS;
4217 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4219 return usemapsize / 8;
4222 static void __init setup_usemap(struct pglist_data *pgdat,
4223 struct zone *zone, unsigned long zonesize)
4225 unsigned long usemapsize = usemap_size(zonesize);
4226 zone->pageblock_flags = NULL;
4228 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4232 static inline void setup_usemap(struct pglist_data *pgdat,
4233 struct zone *zone, unsigned long zonesize) {}
4234 #endif /* CONFIG_SPARSEMEM */
4236 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4238 /* Return a sensible default order for the pageblock size. */
4239 static inline int pageblock_default_order(void)
4241 if (HPAGE_SHIFT > PAGE_SHIFT)
4242 return HUGETLB_PAGE_ORDER;
4247 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4248 static inline void __init set_pageblock_order(unsigned int order)
4250 /* Check that pageblock_nr_pages has not already been setup */
4251 if (pageblock_order)
4255 * Assume the largest contiguous order of interest is a huge page.
4256 * This value may be variable depending on boot parameters on IA64
4258 pageblock_order = order;
4260 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4263 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4264 * and pageblock_default_order() are unused as pageblock_order is set
4265 * at compile-time. See include/linux/pageblock-flags.h for the values of
4266 * pageblock_order based on the kernel config
4268 static inline int pageblock_default_order(unsigned int order)
4272 #define set_pageblock_order(x) do {} while (0)
4274 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4277 * Set up the zone data structures:
4278 * - mark all pages reserved
4279 * - mark all memory queues empty
4280 * - clear the memory bitmaps
4282 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4283 unsigned long *zones_size, unsigned long *zholes_size)
4286 int nid = pgdat->node_id;
4287 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4290 pgdat_resize_init(pgdat);
4291 pgdat->nr_zones = 0;
4292 init_waitqueue_head(&pgdat->kswapd_wait);
4293 pgdat->kswapd_max_order = 0;
4294 pgdat_page_cgroup_init(pgdat);
4296 for (j = 0; j < MAX_NR_ZONES; j++) {
4297 struct zone *zone = pgdat->node_zones + j;
4298 unsigned long size, realsize, memmap_pages;
4301 size = zone_spanned_pages_in_node(nid, j, zones_size);
4302 realsize = size - zone_absent_pages_in_node(nid, j,
4306 * Adjust realsize so that it accounts for how much memory
4307 * is used by this zone for memmap. This affects the watermark
4308 * and per-cpu initialisations
4311 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4312 if (realsize >= memmap_pages) {
4313 realsize -= memmap_pages;
4316 " %s zone: %lu pages used for memmap\n",
4317 zone_names[j], memmap_pages);
4320 " %s zone: %lu pages exceeds realsize %lu\n",
4321 zone_names[j], memmap_pages, realsize);
4323 /* Account for reserved pages */
4324 if (j == 0 && realsize > dma_reserve) {
4325 realsize -= dma_reserve;
4326 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4327 zone_names[0], dma_reserve);
4330 if (!is_highmem_idx(j))
4331 nr_kernel_pages += realsize;
4332 nr_all_pages += realsize;
4334 zone->spanned_pages = size;
4335 zone->present_pages = realsize;
4338 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4340 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4342 zone->name = zone_names[j];
4343 spin_lock_init(&zone->lock);
4344 spin_lock_init(&zone->lru_lock);
4345 zone_seqlock_init(zone);
4346 zone->zone_pgdat = pgdat;
4348 zone_pcp_init(zone);
4350 INIT_LIST_HEAD(&zone->lru[l].list);
4351 zone->reclaim_stat.recent_rotated[0] = 0;
4352 zone->reclaim_stat.recent_rotated[1] = 0;
4353 zone->reclaim_stat.recent_scanned[0] = 0;
4354 zone->reclaim_stat.recent_scanned[1] = 0;
4355 zap_zone_vm_stats(zone);
4360 set_pageblock_order(pageblock_default_order());
4361 setup_usemap(pgdat, zone, size);
4362 ret = init_currently_empty_zone(zone, zone_start_pfn,
4363 size, MEMMAP_EARLY);
4365 memmap_init(size, nid, j, zone_start_pfn);
4366 zone_start_pfn += size;
4370 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4372 /* Skip empty nodes */
4373 if (!pgdat->node_spanned_pages)
4376 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4377 /* ia64 gets its own node_mem_map, before this, without bootmem */
4378 if (!pgdat->node_mem_map) {
4379 unsigned long size, start, end;
4383 * The zone's endpoints aren't required to be MAX_ORDER
4384 * aligned but the node_mem_map endpoints must be in order
4385 * for the buddy allocator to function correctly.
4387 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4388 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4389 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4390 size = (end - start) * sizeof(struct page);
4391 map = alloc_remap(pgdat->node_id, size);
4393 map = alloc_bootmem_node_nopanic(pgdat, size);
4394 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4396 #ifndef CONFIG_NEED_MULTIPLE_NODES
4398 * With no DISCONTIG, the global mem_map is just set as node 0's
4400 if (pgdat == NODE_DATA(0)) {
4401 mem_map = NODE_DATA(0)->node_mem_map;
4402 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4403 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4404 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4405 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4408 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4411 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4412 unsigned long node_start_pfn, unsigned long *zholes_size)
4414 pg_data_t *pgdat = NODE_DATA(nid);
4416 pgdat->node_id = nid;
4417 pgdat->node_start_pfn = node_start_pfn;
4418 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4420 alloc_node_mem_map(pgdat);
4421 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4422 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4423 nid, (unsigned long)pgdat,
4424 (unsigned long)pgdat->node_mem_map);
4427 free_area_init_core(pgdat, zones_size, zholes_size);
4430 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4432 #if MAX_NUMNODES > 1
4434 * Figure out the number of possible node ids.
4436 static void __init setup_nr_node_ids(void)
4439 unsigned int highest = 0;
4441 for_each_node_mask(node, node_possible_map)
4443 nr_node_ids = highest + 1;
4446 static inline void setup_nr_node_ids(void)
4452 * add_active_range - Register a range of PFNs backed by physical memory
4453 * @nid: The node ID the range resides on
4454 * @start_pfn: The start PFN of the available physical memory
4455 * @end_pfn: The end PFN of the available physical memory
4457 * These ranges are stored in an early_node_map[] and later used by
4458 * free_area_init_nodes() to calculate zone sizes and holes. If the
4459 * range spans a memory hole, it is up to the architecture to ensure
4460 * the memory is not freed by the bootmem allocator. If possible
4461 * the range being registered will be merged with existing ranges.
4463 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4464 unsigned long end_pfn)
4468 mminit_dprintk(MMINIT_TRACE, "memory_register",
4469 "Entering add_active_range(%d, %#lx, %#lx) "
4470 "%d entries of %d used\n",
4471 nid, start_pfn, end_pfn,
4472 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4474 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4476 /* Merge with existing active regions if possible */
4477 for (i = 0; i < nr_nodemap_entries; i++) {
4478 if (early_node_map[i].nid != nid)
4481 /* Skip if an existing region covers this new one */
4482 if (start_pfn >= early_node_map[i].start_pfn &&
4483 end_pfn <= early_node_map[i].end_pfn)
4486 /* Merge forward if suitable */
4487 if (start_pfn <= early_node_map[i].end_pfn &&
4488 end_pfn > early_node_map[i].end_pfn) {
4489 early_node_map[i].end_pfn = end_pfn;
4493 /* Merge backward if suitable */
4494 if (start_pfn < early_node_map[i].start_pfn &&
4495 end_pfn >= early_node_map[i].start_pfn) {
4496 early_node_map[i].start_pfn = start_pfn;
4501 /* Check that early_node_map is large enough */
4502 if (i >= MAX_ACTIVE_REGIONS) {
4503 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4504 MAX_ACTIVE_REGIONS);
4508 early_node_map[i].nid = nid;
4509 early_node_map[i].start_pfn = start_pfn;
4510 early_node_map[i].end_pfn = end_pfn;
4511 nr_nodemap_entries = i + 1;
4515 * remove_active_range - Shrink an existing registered range of PFNs
4516 * @nid: The node id the range is on that should be shrunk
4517 * @start_pfn: The new PFN of the range
4518 * @end_pfn: The new PFN of the range
4520 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4521 * The map is kept near the end physical page range that has already been
4522 * registered. This function allows an arch to shrink an existing registered
4525 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4526 unsigned long end_pfn)
4531 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4532 nid, start_pfn, end_pfn);
4534 /* Find the old active region end and shrink */
4535 for_each_active_range_index_in_nid(i, nid) {
4536 if (early_node_map[i].start_pfn >= start_pfn &&
4537 early_node_map[i].end_pfn <= end_pfn) {
4539 early_node_map[i].start_pfn = 0;
4540 early_node_map[i].end_pfn = 0;
4544 if (early_node_map[i].start_pfn < start_pfn &&
4545 early_node_map[i].end_pfn > start_pfn) {
4546 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4547 early_node_map[i].end_pfn = start_pfn;
4548 if (temp_end_pfn > end_pfn)
4549 add_active_range(nid, end_pfn, temp_end_pfn);
4552 if (early_node_map[i].start_pfn >= start_pfn &&
4553 early_node_map[i].end_pfn > end_pfn &&
4554 early_node_map[i].start_pfn < end_pfn) {
4555 early_node_map[i].start_pfn = end_pfn;
4563 /* remove the blank ones */
4564 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4565 if (early_node_map[i].nid != nid)
4567 if (early_node_map[i].end_pfn)
4569 /* we found it, get rid of it */
4570 for (j = i; j < nr_nodemap_entries - 1; j++)
4571 memcpy(&early_node_map[j], &early_node_map[j+1],
4572 sizeof(early_node_map[j]));
4573 j = nr_nodemap_entries - 1;
4574 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4575 nr_nodemap_entries--;
4580 * remove_all_active_ranges - Remove all currently registered regions
4582 * During discovery, it may be found that a table like SRAT is invalid
4583 * and an alternative discovery method must be used. This function removes
4584 * all currently registered regions.
4586 void __init remove_all_active_ranges(void)
4588 memset(early_node_map, 0, sizeof(early_node_map));
4589 nr_nodemap_entries = 0;
4592 /* Compare two active node_active_regions */
4593 static int __init cmp_node_active_region(const void *a, const void *b)
4595 struct node_active_region *arange = (struct node_active_region *)a;
4596 struct node_active_region *brange = (struct node_active_region *)b;
4598 /* Done this way to avoid overflows */
4599 if (arange->start_pfn > brange->start_pfn)
4601 if (arange->start_pfn < brange->start_pfn)
4607 /* sort the node_map by start_pfn */
4608 void __init sort_node_map(void)
4610 sort(early_node_map, (size_t)nr_nodemap_entries,
4611 sizeof(struct node_active_region),
4612 cmp_node_active_region, NULL);
4616 * node_map_pfn_alignment - determine the maximum internode alignment
4618 * This function should be called after node map is populated and sorted.
4619 * It calculates the maximum power of two alignment which can distinguish
4622 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4623 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4624 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4625 * shifted, 1GiB is enough and this function will indicate so.
4627 * This is used to test whether pfn -> nid mapping of the chosen memory
4628 * model has fine enough granularity to avoid incorrect mapping for the
4629 * populated node map.
4631 * Returns the determined alignment in pfn's. 0 if there is no alignment
4632 * requirement (single node).
4634 unsigned long __init node_map_pfn_alignment(void)
4636 unsigned long accl_mask = 0, last_end = 0;
4640 for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4641 int nid = early_node_map[i].nid;
4642 unsigned long start = early_node_map[i].start_pfn;
4643 unsigned long end = early_node_map[i].end_pfn;
4646 if (!start || last_nid < 0 || last_nid == nid) {
4653 * Start with a mask granular enough to pin-point to the
4654 * start pfn and tick off bits one-by-one until it becomes
4655 * too coarse to separate the current node from the last.
4657 mask = ~((1 << __ffs(start)) - 1);
4658 while (mask && last_end <= (start & (mask << 1)))
4661 /* accumulate all internode masks */
4665 /* convert mask to number of pages */
4666 return ~accl_mask + 1;
4669 /* Find the lowest pfn for a node */
4670 static unsigned long __init find_min_pfn_for_node(int nid)
4673 unsigned long min_pfn = ULONG_MAX;
4675 /* Assuming a sorted map, the first range found has the starting pfn */
4676 for_each_active_range_index_in_nid(i, nid)
4677 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4679 if (min_pfn == ULONG_MAX) {
4681 "Could not find start_pfn for node %d\n", nid);
4689 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4691 * It returns the minimum PFN based on information provided via
4692 * add_active_range().
4694 unsigned long __init find_min_pfn_with_active_regions(void)
4696 return find_min_pfn_for_node(MAX_NUMNODES);
4700 * early_calculate_totalpages()
4701 * Sum pages in active regions for movable zone.
4702 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4704 static unsigned long __init early_calculate_totalpages(void)
4707 unsigned long totalpages = 0;
4709 for (i = 0; i < nr_nodemap_entries; i++) {
4710 unsigned long pages = early_node_map[i].end_pfn -
4711 early_node_map[i].start_pfn;
4712 totalpages += pages;
4714 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4720 * Find the PFN the Movable zone begins in each node. Kernel memory
4721 * is spread evenly between nodes as long as the nodes have enough
4722 * memory. When they don't, some nodes will have more kernelcore than
4725 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4728 unsigned long usable_startpfn;
4729 unsigned long kernelcore_node, kernelcore_remaining;
4730 /* save the state before borrow the nodemask */
4731 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4732 unsigned long totalpages = early_calculate_totalpages();
4733 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4736 * If movablecore was specified, calculate what size of
4737 * kernelcore that corresponds so that memory usable for
4738 * any allocation type is evenly spread. If both kernelcore
4739 * and movablecore are specified, then the value of kernelcore
4740 * will be used for required_kernelcore if it's greater than
4741 * what movablecore would have allowed.
4743 if (required_movablecore) {
4744 unsigned long corepages;
4747 * Round-up so that ZONE_MOVABLE is at least as large as what
4748 * was requested by the user
4750 required_movablecore =
4751 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4752 corepages = totalpages - required_movablecore;
4754 required_kernelcore = max(required_kernelcore, corepages);
4757 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4758 if (!required_kernelcore)
4761 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4762 find_usable_zone_for_movable();
4763 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4766 /* Spread kernelcore memory as evenly as possible throughout nodes */
4767 kernelcore_node = required_kernelcore / usable_nodes;
4768 for_each_node_state(nid, N_HIGH_MEMORY) {
4770 * Recalculate kernelcore_node if the division per node
4771 * now exceeds what is necessary to satisfy the requested
4772 * amount of memory for the kernel
4774 if (required_kernelcore < kernelcore_node)
4775 kernelcore_node = required_kernelcore / usable_nodes;
4778 * As the map is walked, we track how much memory is usable
4779 * by the kernel using kernelcore_remaining. When it is
4780 * 0, the rest of the node is usable by ZONE_MOVABLE
4782 kernelcore_remaining = kernelcore_node;
4784 /* Go through each range of PFNs within this node */
4785 for_each_active_range_index_in_nid(i, nid) {
4786 unsigned long start_pfn, end_pfn;
4787 unsigned long size_pages;
4789 start_pfn = max(early_node_map[i].start_pfn,
4790 zone_movable_pfn[nid]);
4791 end_pfn = early_node_map[i].end_pfn;
4792 if (start_pfn >= end_pfn)
4795 /* Account for what is only usable for kernelcore */
4796 if (start_pfn < usable_startpfn) {
4797 unsigned long kernel_pages;
4798 kernel_pages = min(end_pfn, usable_startpfn)
4801 kernelcore_remaining -= min(kernel_pages,
4802 kernelcore_remaining);
4803 required_kernelcore -= min(kernel_pages,
4804 required_kernelcore);
4806 /* Continue if range is now fully accounted */
4807 if (end_pfn <= usable_startpfn) {
4810 * Push zone_movable_pfn to the end so
4811 * that if we have to rebalance
4812 * kernelcore across nodes, we will
4813 * not double account here
4815 zone_movable_pfn[nid] = end_pfn;
4818 start_pfn = usable_startpfn;
4822 * The usable PFN range for ZONE_MOVABLE is from
4823 * start_pfn->end_pfn. Calculate size_pages as the
4824 * number of pages used as kernelcore
4826 size_pages = end_pfn - start_pfn;
4827 if (size_pages > kernelcore_remaining)
4828 size_pages = kernelcore_remaining;
4829 zone_movable_pfn[nid] = start_pfn + size_pages;
4832 * Some kernelcore has been met, update counts and
4833 * break if the kernelcore for this node has been
4836 required_kernelcore -= min(required_kernelcore,
4838 kernelcore_remaining -= size_pages;
4839 if (!kernelcore_remaining)
4845 * If there is still required_kernelcore, we do another pass with one
4846 * less node in the count. This will push zone_movable_pfn[nid] further
4847 * along on the nodes that still have memory until kernelcore is
4851 if (usable_nodes && required_kernelcore > usable_nodes)
4854 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4855 for (nid = 0; nid < MAX_NUMNODES; nid++)
4856 zone_movable_pfn[nid] =
4857 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4860 /* restore the node_state */
4861 node_states[N_HIGH_MEMORY] = saved_node_state;
4864 /* Any regular memory on that node ? */
4865 static void check_for_regular_memory(pg_data_t *pgdat)
4867 #ifdef CONFIG_HIGHMEM
4868 enum zone_type zone_type;
4870 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4871 struct zone *zone = &pgdat->node_zones[zone_type];
4872 if (zone->present_pages)
4873 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4879 * free_area_init_nodes - Initialise all pg_data_t and zone data
4880 * @max_zone_pfn: an array of max PFNs for each zone
4882 * This will call free_area_init_node() for each active node in the system.
4883 * Using the page ranges provided by add_active_range(), the size of each
4884 * zone in each node and their holes is calculated. If the maximum PFN
4885 * between two adjacent zones match, it is assumed that the zone is empty.
4886 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4887 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4888 * starts where the previous one ended. For example, ZONE_DMA32 starts
4889 * at arch_max_dma_pfn.
4891 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4896 /* Sort early_node_map as initialisation assumes it is sorted */
4899 /* Record where the zone boundaries are */
4900 memset(arch_zone_lowest_possible_pfn, 0,
4901 sizeof(arch_zone_lowest_possible_pfn));
4902 memset(arch_zone_highest_possible_pfn, 0,
4903 sizeof(arch_zone_highest_possible_pfn));
4904 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4905 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4906 for (i = 1; i < MAX_NR_ZONES; i++) {
4907 if (i == ZONE_MOVABLE)
4909 arch_zone_lowest_possible_pfn[i] =
4910 arch_zone_highest_possible_pfn[i-1];
4911 arch_zone_highest_possible_pfn[i] =
4912 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4914 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4915 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4917 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4918 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4919 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4921 /* Print out the zone ranges */
4922 printk("Zone PFN ranges:\n");
4923 for (i = 0; i < MAX_NR_ZONES; i++) {
4924 if (i == ZONE_MOVABLE)
4926 printk(" %-8s ", zone_names[i]);
4927 if (arch_zone_lowest_possible_pfn[i] ==
4928 arch_zone_highest_possible_pfn[i])
4931 printk("%0#10lx -> %0#10lx\n",
4932 arch_zone_lowest_possible_pfn[i],
4933 arch_zone_highest_possible_pfn[i]);
4936 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4937 printk("Movable zone start PFN for each node\n");
4938 for (i = 0; i < MAX_NUMNODES; i++) {
4939 if (zone_movable_pfn[i])
4940 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4943 /* Print out the early_node_map[] */
4944 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4945 for (i = 0; i < nr_nodemap_entries; i++)
4946 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4947 early_node_map[i].start_pfn,
4948 early_node_map[i].end_pfn);
4950 /* Initialise every node */
4951 mminit_verify_pageflags_layout();
4952 setup_nr_node_ids();
4953 for_each_online_node(nid) {
4954 pg_data_t *pgdat = NODE_DATA(nid);
4955 free_area_init_node(nid, NULL,
4956 find_min_pfn_for_node(nid), NULL);
4958 /* Any memory on that node */
4959 if (pgdat->node_present_pages)
4960 node_set_state(nid, N_HIGH_MEMORY);
4961 check_for_regular_memory(pgdat);
4965 static int __init cmdline_parse_core(char *p, unsigned long *core)
4967 unsigned long long coremem;
4971 coremem = memparse(p, &p);
4972 *core = coremem >> PAGE_SHIFT;
4974 /* Paranoid check that UL is enough for the coremem value */
4975 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4981 * kernelcore=size sets the amount of memory for use for allocations that
4982 * cannot be reclaimed or migrated.
4984 static int __init cmdline_parse_kernelcore(char *p)
4986 return cmdline_parse_core(p, &required_kernelcore);
4990 * movablecore=size sets the amount of memory for use for allocations that
4991 * can be reclaimed or migrated.
4993 static int __init cmdline_parse_movablecore(char *p)
4995 return cmdline_parse_core(p, &required_movablecore);
4998 early_param("kernelcore", cmdline_parse_kernelcore);
4999 early_param("movablecore", cmdline_parse_movablecore);
5001 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
5004 * set_dma_reserve - set the specified number of pages reserved in the first zone
5005 * @new_dma_reserve: The number of pages to mark reserved
5007 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5008 * In the DMA zone, a significant percentage may be consumed by kernel image
5009 * and other unfreeable allocations which can skew the watermarks badly. This
5010 * function may optionally be used to account for unfreeable pages in the
5011 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5012 * smaller per-cpu batchsize.
5014 void __init set_dma_reserve(unsigned long new_dma_reserve)
5016 dma_reserve = new_dma_reserve;
5019 void __init free_area_init(unsigned long *zones_size)
5021 free_area_init_node(0, zones_size,
5022 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5025 static int page_alloc_cpu_notify(struct notifier_block *self,
5026 unsigned long action, void *hcpu)
5028 int cpu = (unsigned long)hcpu;
5030 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5034 * Spill the event counters of the dead processor
5035 * into the current processors event counters.
5036 * This artificially elevates the count of the current
5039 vm_events_fold_cpu(cpu);
5042 * Zero the differential counters of the dead processor
5043 * so that the vm statistics are consistent.
5045 * This is only okay since the processor is dead and cannot
5046 * race with what we are doing.
5048 refresh_cpu_vm_stats(cpu);
5053 void __init page_alloc_init(void)
5055 hotcpu_notifier(page_alloc_cpu_notify, 0);
5059 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5060 * or min_free_kbytes changes.
5062 static void calculate_totalreserve_pages(void)
5064 struct pglist_data *pgdat;
5065 unsigned long reserve_pages = 0;
5066 enum zone_type i, j;
5068 for_each_online_pgdat(pgdat) {
5069 for (i = 0; i < MAX_NR_ZONES; i++) {
5070 struct zone *zone = pgdat->node_zones + i;
5071 unsigned long max = 0;
5073 /* Find valid and maximum lowmem_reserve in the zone */
5074 for (j = i; j < MAX_NR_ZONES; j++) {
5075 if (zone->lowmem_reserve[j] > max)
5076 max = zone->lowmem_reserve[j];
5079 /* we treat the high watermark as reserved pages. */
5080 max += high_wmark_pages(zone);
5082 if (max > zone->present_pages)
5083 max = zone->present_pages;
5084 reserve_pages += max;
5087 totalreserve_pages = reserve_pages;
5091 * setup_per_zone_lowmem_reserve - called whenever
5092 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5093 * has a correct pages reserved value, so an adequate number of
5094 * pages are left in the zone after a successful __alloc_pages().
5096 static void setup_per_zone_lowmem_reserve(void)
5098 struct pglist_data *pgdat;
5099 enum zone_type j, idx;
5101 for_each_online_pgdat(pgdat) {
5102 for (j = 0; j < MAX_NR_ZONES; j++) {
5103 struct zone *zone = pgdat->node_zones + j;
5104 unsigned long present_pages = zone->present_pages;
5106 zone->lowmem_reserve[j] = 0;
5110 struct zone *lower_zone;
5114 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5115 sysctl_lowmem_reserve_ratio[idx] = 1;
5117 lower_zone = pgdat->node_zones + idx;
5118 lower_zone->lowmem_reserve[j] = present_pages /
5119 sysctl_lowmem_reserve_ratio[idx];
5120 present_pages += lower_zone->present_pages;
5125 /* update totalreserve_pages */
5126 calculate_totalreserve_pages();
5130 * setup_per_zone_wmarks - called when min_free_kbytes changes
5131 * or when memory is hot-{added|removed}
5133 * Ensures that the watermark[min,low,high] values for each zone are set
5134 * correctly with respect to min_free_kbytes.
5136 void setup_per_zone_wmarks(void)
5138 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5139 unsigned long lowmem_pages = 0;
5141 unsigned long flags;
5143 /* Calculate total number of !ZONE_HIGHMEM pages */
5144 for_each_zone(zone) {
5145 if (!is_highmem(zone))
5146 lowmem_pages += zone->present_pages;
5149 for_each_zone(zone) {
5152 spin_lock_irqsave(&zone->lock, flags);
5153 tmp = (u64)pages_min * zone->present_pages;
5154 do_div(tmp, lowmem_pages);
5155 if (is_highmem(zone)) {
5157 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5158 * need highmem pages, so cap pages_min to a small
5161 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5162 * deltas controls asynch page reclaim, and so should
5163 * not be capped for highmem.
5167 min_pages = zone->present_pages / 1024;
5168 if (min_pages < SWAP_CLUSTER_MAX)
5169 min_pages = SWAP_CLUSTER_MAX;
5170 if (min_pages > 128)
5172 zone->watermark[WMARK_MIN] = min_pages;
5175 * If it's a lowmem zone, reserve a number of pages
5176 * proportionate to the zone's size.
5178 zone->watermark[WMARK_MIN] = tmp;
5181 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5182 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5183 setup_zone_migrate_reserve(zone);
5184 spin_unlock_irqrestore(&zone->lock, flags);
5187 /* update totalreserve_pages */
5188 calculate_totalreserve_pages();
5192 * The inactive anon list should be small enough that the VM never has to
5193 * do too much work, but large enough that each inactive page has a chance
5194 * to be referenced again before it is swapped out.
5196 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5197 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5198 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5199 * the anonymous pages are kept on the inactive list.
5202 * memory ratio inactive anon
5203 * -------------------------------------
5212 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5214 unsigned int gb, ratio;
5216 /* Zone size in gigabytes */
5217 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5219 ratio = int_sqrt(10 * gb);
5223 zone->inactive_ratio = ratio;
5226 static void __meminit setup_per_zone_inactive_ratio(void)
5231 calculate_zone_inactive_ratio(zone);
5235 * Initialise min_free_kbytes.
5237 * For small machines we want it small (128k min). For large machines
5238 * we want it large (64MB max). But it is not linear, because network
5239 * bandwidth does not increase linearly with machine size. We use
5241 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5242 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5258 int __meminit init_per_zone_wmark_min(void)
5260 unsigned long lowmem_kbytes;
5262 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5264 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5265 if (min_free_kbytes < 128)
5266 min_free_kbytes = 128;
5267 if (min_free_kbytes > 65536)
5268 min_free_kbytes = 65536;
5269 setup_per_zone_wmarks();
5270 refresh_zone_stat_thresholds();
5271 setup_per_zone_lowmem_reserve();
5272 setup_per_zone_inactive_ratio();
5275 module_init(init_per_zone_wmark_min)
5278 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5279 * that we can call two helper functions whenever min_free_kbytes
5282 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5283 void __user *buffer, size_t *length, loff_t *ppos)
5285 proc_dointvec(table, write, buffer, length, ppos);
5287 setup_per_zone_wmarks();
5292 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5293 void __user *buffer, size_t *length, loff_t *ppos)
5298 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5303 zone->min_unmapped_pages = (zone->present_pages *
5304 sysctl_min_unmapped_ratio) / 100;
5308 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5309 void __user *buffer, size_t *length, loff_t *ppos)
5314 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5319 zone->min_slab_pages = (zone->present_pages *
5320 sysctl_min_slab_ratio) / 100;
5326 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5327 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5328 * whenever sysctl_lowmem_reserve_ratio changes.
5330 * The reserve ratio obviously has absolutely no relation with the
5331 * minimum watermarks. The lowmem reserve ratio can only make sense
5332 * if in function of the boot time zone sizes.
5334 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5335 void __user *buffer, size_t *length, loff_t *ppos)
5337 proc_dointvec_minmax(table, write, buffer, length, ppos);
5338 setup_per_zone_lowmem_reserve();
5343 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5344 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5345 * can have before it gets flushed back to buddy allocator.
5348 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5349 void __user *buffer, size_t *length, loff_t *ppos)
5355 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5356 if (!write || (ret == -EINVAL))
5358 for_each_populated_zone(zone) {
5359 for_each_possible_cpu(cpu) {
5361 high = zone->present_pages / percpu_pagelist_fraction;
5362 setup_pagelist_highmark(
5363 per_cpu_ptr(zone->pageset, cpu), high);
5369 int hashdist = HASHDIST_DEFAULT;
5372 static int __init set_hashdist(char *str)
5376 hashdist = simple_strtoul(str, &str, 0);
5379 __setup("hashdist=", set_hashdist);
5383 * allocate a large system hash table from bootmem
5384 * - it is assumed that the hash table must contain an exact power-of-2
5385 * quantity of entries
5386 * - limit is the number of hash buckets, not the total allocation size
5388 void *__init alloc_large_system_hash(const char *tablename,
5389 unsigned long bucketsize,
5390 unsigned long numentries,
5393 unsigned int *_hash_shift,
5394 unsigned int *_hash_mask,
5395 unsigned long limit)
5397 unsigned long long max = limit;
5398 unsigned long log2qty, size;
5401 /* allow the kernel cmdline to have a say */
5403 /* round applicable memory size up to nearest megabyte */
5404 numentries = nr_kernel_pages;
5405 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5406 numentries >>= 20 - PAGE_SHIFT;
5407 numentries <<= 20 - PAGE_SHIFT;
5409 /* limit to 1 bucket per 2^scale bytes of low memory */
5410 if (scale > PAGE_SHIFT)
5411 numentries >>= (scale - PAGE_SHIFT);
5413 numentries <<= (PAGE_SHIFT - scale);
5415 /* Make sure we've got at least a 0-order allocation.. */
5416 if (unlikely(flags & HASH_SMALL)) {
5417 /* Makes no sense without HASH_EARLY */
5418 WARN_ON(!(flags & HASH_EARLY));
5419 if (!(numentries >> *_hash_shift)) {
5420 numentries = 1UL << *_hash_shift;
5421 BUG_ON(!numentries);
5423 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5424 numentries = PAGE_SIZE / bucketsize;
5426 numentries = roundup_pow_of_two(numentries);
5428 /* limit allocation size to 1/16 total memory by default */
5430 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5431 do_div(max, bucketsize);
5434 if (numentries > max)
5437 log2qty = ilog2(numentries);
5440 size = bucketsize << log2qty;
5441 if (flags & HASH_EARLY)
5442 table = alloc_bootmem_nopanic(size);
5444 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5447 * If bucketsize is not a power-of-two, we may free
5448 * some pages at the end of hash table which
5449 * alloc_pages_exact() automatically does
5451 if (get_order(size) < MAX_ORDER) {
5452 table = alloc_pages_exact(size, GFP_ATOMIC);
5453 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5456 } while (!table && size > PAGE_SIZE && --log2qty);
5459 panic("Failed to allocate %s hash table\n", tablename);
5461 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5464 ilog2(size) - PAGE_SHIFT,
5468 *_hash_shift = log2qty;
5470 *_hash_mask = (1 << log2qty) - 1;
5475 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5476 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5479 #ifdef CONFIG_SPARSEMEM
5480 return __pfn_to_section(pfn)->pageblock_flags;
5482 return zone->pageblock_flags;
5483 #endif /* CONFIG_SPARSEMEM */
5486 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5488 #ifdef CONFIG_SPARSEMEM
5489 pfn &= (PAGES_PER_SECTION-1);
5490 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5492 pfn = pfn - zone->zone_start_pfn;
5493 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5494 #endif /* CONFIG_SPARSEMEM */
5498 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5499 * @page: The page within the block of interest
5500 * @start_bitidx: The first bit of interest to retrieve
5501 * @end_bitidx: The last bit of interest
5502 * returns pageblock_bits flags
5504 unsigned long get_pageblock_flags_group(struct page *page,
5505 int start_bitidx, int end_bitidx)
5508 unsigned long *bitmap;
5509 unsigned long pfn, bitidx;
5510 unsigned long flags = 0;
5511 unsigned long value = 1;
5513 zone = page_zone(page);
5514 pfn = page_to_pfn(page);
5515 bitmap = get_pageblock_bitmap(zone, pfn);
5516 bitidx = pfn_to_bitidx(zone, pfn);
5518 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5519 if (test_bit(bitidx + start_bitidx, bitmap))
5526 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5527 * @page: The page within the block of interest
5528 * @start_bitidx: The first bit of interest
5529 * @end_bitidx: The last bit of interest
5530 * @flags: The flags to set
5532 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5533 int start_bitidx, int end_bitidx)
5536 unsigned long *bitmap;
5537 unsigned long pfn, bitidx;
5538 unsigned long value = 1;
5540 zone = page_zone(page);
5541 pfn = page_to_pfn(page);
5542 bitmap = get_pageblock_bitmap(zone, pfn);
5543 bitidx = pfn_to_bitidx(zone, pfn);
5544 VM_BUG_ON(pfn < zone->zone_start_pfn);
5545 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5547 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5549 __set_bit(bitidx + start_bitidx, bitmap);
5551 __clear_bit(bitidx + start_bitidx, bitmap);
5555 * This is designed as sub function...plz see page_isolation.c also.
5556 * set/clear page block's type to be ISOLATE.
5557 * page allocater never alloc memory from ISOLATE block.
5561 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5563 unsigned long pfn, iter, found;
5565 * For avoiding noise data, lru_add_drain_all() should be called
5566 * If ZONE_MOVABLE, the zone never contains immobile pages
5568 if (zone_idx(zone) == ZONE_MOVABLE)
5571 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5574 pfn = page_to_pfn(page);
5575 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5576 unsigned long check = pfn + iter;
5578 if (!pfn_valid_within(check))
5581 page = pfn_to_page(check);
5582 if (!page_count(page)) {
5583 if (PageBuddy(page))
5584 iter += (1 << page_order(page)) - 1;
5590 * If there are RECLAIMABLE pages, we need to check it.
5591 * But now, memory offline itself doesn't call shrink_slab()
5592 * and it still to be fixed.
5595 * If the page is not RAM, page_count()should be 0.
5596 * we don't need more check. This is an _used_ not-movable page.
5598 * The problematic thing here is PG_reserved pages. PG_reserved
5599 * is set to both of a memory hole page and a _used_ kernel
5608 bool is_pageblock_removable_nolock(struct page *page)
5610 struct zone *zone = page_zone(page);
5611 return __count_immobile_pages(zone, page, 0);
5614 int set_migratetype_isolate(struct page *page)
5617 unsigned long flags, pfn;
5618 struct memory_isolate_notify arg;
5622 zone = page_zone(page);
5624 spin_lock_irqsave(&zone->lock, flags);
5626 pfn = page_to_pfn(page);
5627 arg.start_pfn = pfn;
5628 arg.nr_pages = pageblock_nr_pages;
5629 arg.pages_found = 0;
5632 * It may be possible to isolate a pageblock even if the
5633 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5634 * notifier chain is used by balloon drivers to return the
5635 * number of pages in a range that are held by the balloon
5636 * driver to shrink memory. If all the pages are accounted for
5637 * by balloons, are free, or on the LRU, isolation can continue.
5638 * Later, for example, when memory hotplug notifier runs, these
5639 * pages reported as "can be isolated" should be isolated(freed)
5640 * by the balloon driver through the memory notifier chain.
5642 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5643 notifier_ret = notifier_to_errno(notifier_ret);
5647 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5648 * We just check MOVABLE pages.
5650 if (__count_immobile_pages(zone, page, arg.pages_found))
5654 * immobile means "not-on-lru" paes. If immobile is larger than
5655 * removable-by-driver pages reported by notifier, we'll fail.
5660 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5661 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5664 spin_unlock_irqrestore(&zone->lock, flags);
5670 void unset_migratetype_isolate(struct page *page)
5673 unsigned long flags;
5674 zone = page_zone(page);
5675 spin_lock_irqsave(&zone->lock, flags);
5676 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5678 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5679 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5681 spin_unlock_irqrestore(&zone->lock, flags);
5684 #ifdef CONFIG_MEMORY_HOTREMOVE
5686 * All pages in the range must be isolated before calling this.
5689 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5695 unsigned long flags;
5696 /* find the first valid pfn */
5697 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5702 zone = page_zone(pfn_to_page(pfn));
5703 spin_lock_irqsave(&zone->lock, flags);
5705 while (pfn < end_pfn) {
5706 if (!pfn_valid(pfn)) {
5710 page = pfn_to_page(pfn);
5711 BUG_ON(page_count(page));
5712 BUG_ON(!PageBuddy(page));
5713 order = page_order(page);
5714 #ifdef CONFIG_DEBUG_VM
5715 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5716 pfn, 1 << order, end_pfn);
5718 list_del(&page->lru);
5719 rmv_page_order(page);
5720 zone->free_area[order].nr_free--;
5721 __mod_zone_page_state(zone, NR_FREE_PAGES,
5723 for (i = 0; i < (1 << order); i++)
5724 SetPageReserved((page+i));
5725 pfn += (1 << order);
5727 spin_unlock_irqrestore(&zone->lock, flags);
5731 #ifdef CONFIG_MEMORY_FAILURE
5732 bool is_free_buddy_page(struct page *page)
5734 struct zone *zone = page_zone(page);
5735 unsigned long pfn = page_to_pfn(page);
5736 unsigned long flags;
5739 spin_lock_irqsave(&zone->lock, flags);
5740 for (order = 0; order < MAX_ORDER; order++) {
5741 struct page *page_head = page - (pfn & ((1 << order) - 1));
5743 if (PageBuddy(page_head) && page_order(page_head) >= order)
5746 spin_unlock_irqrestore(&zone->lock, flags);
5748 return order < MAX_ORDER;
5752 static struct trace_print_flags pageflag_names[] = {
5753 {1UL << PG_locked, "locked" },
5754 {1UL << PG_error, "error" },
5755 {1UL << PG_referenced, "referenced" },
5756 {1UL << PG_uptodate, "uptodate" },
5757 {1UL << PG_dirty, "dirty" },
5758 {1UL << PG_lru, "lru" },
5759 {1UL << PG_active, "active" },
5760 {1UL << PG_slab, "slab" },
5761 {1UL << PG_owner_priv_1, "owner_priv_1" },
5762 {1UL << PG_arch_1, "arch_1" },
5763 {1UL << PG_reserved, "reserved" },
5764 {1UL << PG_private, "private" },
5765 {1UL << PG_private_2, "private_2" },
5766 {1UL << PG_writeback, "writeback" },
5767 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5768 {1UL << PG_head, "head" },
5769 {1UL << PG_tail, "tail" },
5771 {1UL << PG_compound, "compound" },
5773 {1UL << PG_swapcache, "swapcache" },
5774 {1UL << PG_mappedtodisk, "mappedtodisk" },
5775 {1UL << PG_reclaim, "reclaim" },
5776 {1UL << PG_swapbacked, "swapbacked" },
5777 {1UL << PG_unevictable, "unevictable" },
5779 {1UL << PG_mlocked, "mlocked" },
5781 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5782 {1UL << PG_uncached, "uncached" },
5784 #ifdef CONFIG_MEMORY_FAILURE
5785 {1UL << PG_hwpoison, "hwpoison" },
5790 static void dump_page_flags(unsigned long flags)
5792 const char *delim = "";
5796 printk(KERN_ALERT "page flags: %#lx(", flags);
5798 /* remove zone id */
5799 flags &= (1UL << NR_PAGEFLAGS) - 1;
5801 for (i = 0; pageflag_names[i].name && flags; i++) {
5803 mask = pageflag_names[i].mask;
5804 if ((flags & mask) != mask)
5808 printk("%s%s", delim, pageflag_names[i].name);
5812 /* check for left over flags */
5814 printk("%s%#lx", delim, flags);
5819 void dump_page(struct page *page)
5822 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5823 page, atomic_read(&page->_count), page_mapcount(page),
5824 page->mapping, page->index);
5825 dump_page_flags(page->flags);
5826 mem_cgroup_print_bad_page(page);