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>
60 #include <linux/page-debug-flags.h>
62 #include <asm/tlbflush.h>
63 #include <asm/div64.h>
66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67 DEFINE_PER_CPU(int, numa_node);
68 EXPORT_PER_CPU_SYMBOL(numa_node);
71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
78 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
83 * Array of node states.
85 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86 [N_POSSIBLE] = NODE_MASK_ALL,
87 [N_ONLINE] = { { [0] = 1UL } },
89 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
91 [N_HIGH_MEMORY] = { { [0] = 1UL } },
93 [N_CPU] = { { [0] = 1UL } },
96 EXPORT_SYMBOL(node_states);
98 unsigned long totalram_pages __read_mostly;
99 unsigned long totalreserve_pages __read_mostly;
100 int percpu_pagelist_fraction;
101 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
103 #ifdef CONFIG_PM_SLEEP
105 * The following functions are used by the suspend/hibernate code to temporarily
106 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
107 * while devices are suspended. To avoid races with the suspend/hibernate code,
108 * they should always be called with pm_mutex held (gfp_allowed_mask also should
109 * only be modified with pm_mutex held, unless the suspend/hibernate code is
110 * guaranteed not to run in parallel with that modification).
113 static gfp_t saved_gfp_mask;
115 void pm_restore_gfp_mask(void)
117 WARN_ON(!mutex_is_locked(&pm_mutex));
118 if (saved_gfp_mask) {
119 gfp_allowed_mask = saved_gfp_mask;
124 void pm_restrict_gfp_mask(void)
126 WARN_ON(!mutex_is_locked(&pm_mutex));
127 WARN_ON(saved_gfp_mask);
128 saved_gfp_mask = gfp_allowed_mask;
129 gfp_allowed_mask &= ~GFP_IOFS;
132 bool pm_suspended_storage(void)
134 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
138 #endif /* CONFIG_PM_SLEEP */
140 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
141 int pageblock_order __read_mostly;
144 static void __free_pages_ok(struct page *page, unsigned int order);
147 * results with 256, 32 in the lowmem_reserve sysctl:
148 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
149 * 1G machine -> (16M dma, 784M normal, 224M high)
150 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
151 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
152 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
154 * TBD: should special case ZONE_DMA32 machines here - in those we normally
155 * don't need any ZONE_NORMAL reservation
157 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
158 #ifdef CONFIG_ZONE_DMA
161 #ifdef CONFIG_ZONE_DMA32
164 #ifdef CONFIG_HIGHMEM
170 EXPORT_SYMBOL(totalram_pages);
172 static char * const zone_names[MAX_NR_ZONES] = {
173 #ifdef CONFIG_ZONE_DMA
176 #ifdef CONFIG_ZONE_DMA32
180 #ifdef CONFIG_HIGHMEM
186 int min_free_kbytes = 1024;
188 static unsigned long __meminitdata nr_kernel_pages;
189 static unsigned long __meminitdata nr_all_pages;
190 static unsigned long __meminitdata dma_reserve;
192 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
194 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
195 * ranges of memory (RAM) that may be registered with add_active_range().
196 * Ranges passed to add_active_range() will be merged if possible
197 * so the number of times add_active_range() can be called is
198 * related to the number of nodes and the number of holes
200 #ifdef CONFIG_MAX_ACTIVE_REGIONS
201 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
202 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
204 #if MAX_NUMNODES >= 32
205 /* If there can be many nodes, allow up to 50 holes per node */
206 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
208 /* By default, allow up to 256 distinct regions */
209 #define MAX_ACTIVE_REGIONS 256
213 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
214 static int __meminitdata nr_nodemap_entries;
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
217 static unsigned long __initdata required_kernelcore;
218 static unsigned long __initdata required_movablecore;
219 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone);
224 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
227 int nr_node_ids __read_mostly = MAX_NUMNODES;
228 int nr_online_nodes __read_mostly = 1;
229 EXPORT_SYMBOL(nr_node_ids);
230 EXPORT_SYMBOL(nr_online_nodes);
233 int page_group_by_mobility_disabled __read_mostly;
235 static void set_pageblock_migratetype(struct page *page, int migratetype)
238 if (unlikely(page_group_by_mobility_disabled))
239 migratetype = MIGRATE_UNMOVABLE;
241 set_pageblock_flags_group(page, (unsigned long)migratetype,
242 PB_migrate, PB_migrate_end);
245 bool oom_killer_disabled __read_mostly;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
252 unsigned long pfn = page_to_pfn(page);
255 seq = zone_span_seqbegin(zone);
256 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
258 else if (pfn < zone->zone_start_pfn)
260 } while (zone_span_seqretry(zone, seq));
265 static int page_is_consistent(struct zone *zone, struct page *page)
267 if (!pfn_valid_within(page_to_pfn(page)))
269 if (zone != page_zone(page))
275 * Temporary debugging check for pages not lying within a given zone.
277 static int bad_range(struct zone *zone, struct page *page)
279 if (page_outside_zone_boundaries(zone, page))
281 if (!page_is_consistent(zone, page))
287 static inline int bad_range(struct zone *zone, struct page *page)
293 static void bad_page(struct page *page)
295 static unsigned long resume;
296 static unsigned long nr_shown;
297 static unsigned long nr_unshown;
299 /* Don't complain about poisoned pages */
300 if (PageHWPoison(page)) {
301 reset_page_mapcount(page); /* remove PageBuddy */
306 * Allow a burst of 60 reports, then keep quiet for that minute;
307 * or allow a steady drip of one report per second.
309 if (nr_shown == 60) {
310 if (time_before(jiffies, resume)) {
316 "BUG: Bad page state: %lu messages suppressed\n",
323 resume = jiffies + 60 * HZ;
325 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
326 current->comm, page_to_pfn(page));
332 /* Leave bad fields for debug, except PageBuddy could make trouble */
333 reset_page_mapcount(page); /* remove PageBuddy */
334 add_taint(TAINT_BAD_PAGE);
338 * Higher-order pages are called "compound pages". They are structured thusly:
340 * The first PAGE_SIZE page is called the "head page".
342 * The remaining PAGE_SIZE pages are called "tail pages".
344 * All pages have PG_compound set. All pages have their ->private pointing at
345 * the head page (even the head page has this).
347 * The first tail page's ->lru.next holds the address of the compound page's
348 * put_page() function. Its ->lru.prev holds the order of allocation.
349 * This usage means that zero-order pages may not be compound.
352 static void free_compound_page(struct page *page)
354 __free_pages_ok(page, compound_order(page));
357 void prep_compound_page(struct page *page, unsigned long order)
360 int nr_pages = 1 << order;
362 set_compound_page_dtor(page, free_compound_page);
363 set_compound_order(page, order);
365 for (i = 1; i < nr_pages; i++) {
366 struct page *p = page + i;
368 set_page_count(p, 0);
369 p->first_page = page;
373 /* update __split_huge_page_refcount if you change this function */
374 static int destroy_compound_page(struct page *page, unsigned long order)
377 int nr_pages = 1 << order;
380 if (unlikely(compound_order(page) != order) ||
381 unlikely(!PageHead(page))) {
386 __ClearPageHead(page);
388 for (i = 1; i < nr_pages; i++) {
389 struct page *p = page + i;
391 if (unlikely(!PageTail(p) || (p->first_page != page))) {
401 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
406 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
407 * and __GFP_HIGHMEM from hard or soft interrupt context.
409 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
410 for (i = 0; i < (1 << order); i++)
411 clear_highpage(page + i);
414 #ifdef CONFIG_DEBUG_PAGEALLOC
415 unsigned int _debug_guardpage_minorder;
417 static int __init debug_guardpage_minorder_setup(char *buf)
421 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
422 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
425 _debug_guardpage_minorder = res;
426 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
429 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
431 static inline void set_page_guard_flag(struct page *page)
433 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
436 static inline void clear_page_guard_flag(struct page *page)
438 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
441 static inline void set_page_guard_flag(struct page *page) { }
442 static inline void clear_page_guard_flag(struct page *page) { }
445 static inline void set_page_order(struct page *page, int order)
447 set_page_private(page, order);
448 __SetPageBuddy(page);
451 static inline void rmv_page_order(struct page *page)
453 __ClearPageBuddy(page);
454 set_page_private(page, 0);
458 * Locate the struct page for both the matching buddy in our
459 * pair (buddy1) and the combined O(n+1) page they form (page).
461 * 1) Any buddy B1 will have an order O twin B2 which satisfies
462 * the following equation:
464 * For example, if the starting buddy (buddy2) is #8 its order
466 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
468 * 2) Any buddy B will have an order O+1 parent P which
469 * satisfies the following equation:
472 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
474 static inline unsigned long
475 __find_buddy_index(unsigned long page_idx, unsigned int order)
477 return page_idx ^ (1 << order);
481 * This function checks whether a page is free && is the buddy
482 * we can do coalesce a page and its buddy if
483 * (a) the buddy is not in a hole &&
484 * (b) the buddy is in the buddy system &&
485 * (c) a page and its buddy have the same order &&
486 * (d) a page and its buddy are in the same zone.
488 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
489 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
491 * For recording page's order, we use page_private(page).
493 static inline int page_is_buddy(struct page *page, struct page *buddy,
496 if (!pfn_valid_within(page_to_pfn(buddy)))
499 if (page_zone_id(page) != page_zone_id(buddy))
502 if (page_is_guard(buddy) && page_order(buddy) == order) {
503 VM_BUG_ON(page_count(buddy) != 0);
507 if (PageBuddy(buddy) && page_order(buddy) == order) {
508 VM_BUG_ON(page_count(buddy) != 0);
515 * Freeing function for a buddy system allocator.
517 * The concept of a buddy system is to maintain direct-mapped table
518 * (containing bit values) for memory blocks of various "orders".
519 * The bottom level table contains the map for the smallest allocatable
520 * units of memory (here, pages), and each level above it describes
521 * pairs of units from the levels below, hence, "buddies".
522 * At a high level, all that happens here is marking the table entry
523 * at the bottom level available, and propagating the changes upward
524 * as necessary, plus some accounting needed to play nicely with other
525 * parts of the VM system.
526 * At each level, we keep a list of pages, which are heads of continuous
527 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
528 * order is recorded in page_private(page) field.
529 * So when we are allocating or freeing one, we can derive the state of the
530 * other. That is, if we allocate a small block, and both were
531 * free, the remainder of the region must be split into blocks.
532 * If a block is freed, and its buddy is also free, then this
533 * triggers coalescing into a block of larger size.
538 static inline void __free_one_page(struct page *page,
539 struct zone *zone, unsigned int order,
542 unsigned long page_idx;
543 unsigned long combined_idx;
544 unsigned long uninitialized_var(buddy_idx);
547 if (unlikely(PageCompound(page)))
548 if (unlikely(destroy_compound_page(page, order)))
551 VM_BUG_ON(migratetype == -1);
553 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
555 VM_BUG_ON(page_idx & ((1 << order) - 1));
556 VM_BUG_ON(bad_range(zone, page));
558 while (order < MAX_ORDER-1) {
559 buddy_idx = __find_buddy_index(page_idx, order);
560 buddy = page + (buddy_idx - page_idx);
561 if (!page_is_buddy(page, buddy, order))
564 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
565 * merge with it and move up one order.
567 if (page_is_guard(buddy)) {
568 clear_page_guard_flag(buddy);
569 set_page_private(page, 0);
570 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
572 list_del(&buddy->lru);
573 zone->free_area[order].nr_free--;
574 rmv_page_order(buddy);
576 combined_idx = buddy_idx & page_idx;
577 page = page + (combined_idx - page_idx);
578 page_idx = combined_idx;
581 set_page_order(page, order);
584 * If this is not the largest possible page, check if the buddy
585 * of the next-highest order is free. If it is, it's possible
586 * that pages are being freed that will coalesce soon. In case,
587 * that is happening, add the free page to the tail of the list
588 * so it's less likely to be used soon and more likely to be merged
589 * as a higher order page
591 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
592 struct page *higher_page, *higher_buddy;
593 combined_idx = buddy_idx & page_idx;
594 higher_page = page + (combined_idx - page_idx);
595 buddy_idx = __find_buddy_index(combined_idx, order + 1);
596 higher_buddy = higher_page + (buddy_idx - combined_idx);
597 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
598 list_add_tail(&page->lru,
599 &zone->free_area[order].free_list[migratetype]);
604 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
606 zone->free_area[order].nr_free++;
610 * free_page_mlock() -- clean up attempts to free and mlocked() page.
611 * Page should not be on lru, so no need to fix that up.
612 * free_pages_check() will verify...
614 static inline void free_page_mlock(struct page *page)
616 __dec_zone_page_state(page, NR_MLOCK);
617 __count_vm_event(UNEVICTABLE_MLOCKFREED);
620 static inline int free_pages_check(struct page *page)
622 if (unlikely(page_mapcount(page) |
623 (page->mapping != NULL) |
624 (atomic_read(&page->_count) != 0) |
625 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
626 (mem_cgroup_bad_page_check(page)))) {
630 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
631 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
636 * Frees a number of pages from the PCP lists
637 * Assumes all pages on list are in same zone, and of same order.
638 * count is the number of pages to free.
640 * If the zone was previously in an "all pages pinned" state then look to
641 * see if this freeing clears that state.
643 * And clear the zone's pages_scanned counter, to hold off the "all pages are
644 * pinned" detection logic.
646 static void free_pcppages_bulk(struct zone *zone, int count,
647 struct per_cpu_pages *pcp)
653 spin_lock(&zone->lock);
654 zone->all_unreclaimable = 0;
655 zone->pages_scanned = 0;
659 struct list_head *list;
662 * Remove pages from lists in a round-robin fashion. A
663 * batch_free count is maintained that is incremented when an
664 * empty list is encountered. This is so more pages are freed
665 * off fuller lists instead of spinning excessively around empty
670 if (++migratetype == MIGRATE_PCPTYPES)
672 list = &pcp->lists[migratetype];
673 } while (list_empty(list));
675 /* This is the only non-empty list. Free them all. */
676 if (batch_free == MIGRATE_PCPTYPES)
677 batch_free = to_free;
680 page = list_entry(list->prev, struct page, lru);
681 /* must delete as __free_one_page list manipulates */
682 list_del(&page->lru);
683 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
684 __free_one_page(page, zone, 0, page_private(page));
685 trace_mm_page_pcpu_drain(page, 0, page_private(page));
686 } while (--to_free && --batch_free && !list_empty(list));
688 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
689 spin_unlock(&zone->lock);
692 static void free_one_page(struct zone *zone, struct page *page, int order,
695 spin_lock(&zone->lock);
696 zone->all_unreclaimable = 0;
697 zone->pages_scanned = 0;
699 __free_one_page(page, zone, order, migratetype);
700 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
701 spin_unlock(&zone->lock);
704 static bool free_pages_prepare(struct page *page, unsigned int order)
709 trace_mm_page_free(page, order);
710 kmemcheck_free_shadow(page, order);
713 page->mapping = NULL;
714 for (i = 0; i < (1 << order); i++)
715 bad += free_pages_check(page + i);
719 if (!PageHighMem(page)) {
720 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
721 debug_check_no_obj_freed(page_address(page),
724 arch_free_page(page, order);
725 kernel_map_pages(page, 1 << order, 0);
730 static void __free_pages_ok(struct page *page, unsigned int order)
733 int wasMlocked = __TestClearPageMlocked(page);
735 if (!free_pages_prepare(page, order))
738 local_irq_save(flags);
739 if (unlikely(wasMlocked))
740 free_page_mlock(page);
741 __count_vm_events(PGFREE, 1 << order);
742 free_one_page(page_zone(page), page, order,
743 get_pageblock_migratetype(page));
744 local_irq_restore(flags);
748 * permit the bootmem allocator to evade page validation on high-order frees
750 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
753 __ClearPageReserved(page);
754 set_page_count(page, 0);
755 set_page_refcounted(page);
761 for (loop = 0; loop < BITS_PER_LONG; loop++) {
762 struct page *p = &page[loop];
764 if (loop + 1 < BITS_PER_LONG)
766 __ClearPageReserved(p);
767 set_page_count(p, 0);
770 set_page_refcounted(page);
771 __free_pages(page, order);
777 * The order of subdivision here is critical for the IO subsystem.
778 * Please do not alter this order without good reasons and regression
779 * testing. Specifically, as large blocks of memory are subdivided,
780 * the order in which smaller blocks are delivered depends on the order
781 * they're subdivided in this function. This is the primary factor
782 * influencing the order in which pages are delivered to the IO
783 * subsystem according to empirical testing, and this is also justified
784 * by considering the behavior of a buddy system containing a single
785 * large block of memory acted on by a series of small allocations.
786 * This behavior is a critical factor in sglist merging's success.
790 static inline void expand(struct zone *zone, struct page *page,
791 int low, int high, struct free_area *area,
794 unsigned long size = 1 << high;
800 VM_BUG_ON(bad_range(zone, &page[size]));
802 #ifdef CONFIG_DEBUG_PAGEALLOC
803 if (high < debug_guardpage_minorder()) {
805 * Mark as guard pages (or page), that will allow to
806 * merge back to allocator when buddy will be freed.
807 * Corresponding page table entries will not be touched,
808 * pages will stay not present in virtual address space
810 INIT_LIST_HEAD(&page[size].lru);
811 set_page_guard_flag(&page[size]);
812 set_page_private(&page[size], high);
813 /* Guard pages are not available for any usage */
814 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high));
818 list_add(&page[size].lru, &area->free_list[migratetype]);
820 set_page_order(&page[size], high);
825 * This page is about to be returned from the page allocator
827 static inline int check_new_page(struct page *page)
829 if (unlikely(page_mapcount(page) |
830 (page->mapping != NULL) |
831 (atomic_read(&page->_count) != 0) |
832 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
833 (mem_cgroup_bad_page_check(page)))) {
840 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
844 for (i = 0; i < (1 << order); i++) {
845 struct page *p = page + i;
846 if (unlikely(check_new_page(p)))
850 set_page_private(page, 0);
851 set_page_refcounted(page);
853 arch_alloc_page(page, order);
854 kernel_map_pages(page, 1 << order, 1);
856 if (gfp_flags & __GFP_ZERO)
857 prep_zero_page(page, order, gfp_flags);
859 if (order && (gfp_flags & __GFP_COMP))
860 prep_compound_page(page, order);
866 * Go through the free lists for the given migratetype and remove
867 * the smallest available page from the freelists
870 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
873 unsigned int current_order;
874 struct free_area * area;
877 /* Find a page of the appropriate size in the preferred list */
878 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
879 area = &(zone->free_area[current_order]);
880 if (list_empty(&area->free_list[migratetype]))
883 page = list_entry(area->free_list[migratetype].next,
885 list_del(&page->lru);
886 rmv_page_order(page);
888 expand(zone, page, order, current_order, area, migratetype);
897 * This array describes the order lists are fallen back to when
898 * the free lists for the desirable migrate type are depleted
900 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
901 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
902 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
903 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
904 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
908 * Move the free pages in a range to the free lists of the requested type.
909 * Note that start_page and end_pages are not aligned on a pageblock
910 * boundary. If alignment is required, use move_freepages_block()
912 static int move_freepages(struct zone *zone,
913 struct page *start_page, struct page *end_page,
920 #ifndef CONFIG_HOLES_IN_ZONE
922 * page_zone is not safe to call in this context when
923 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
924 * anyway as we check zone boundaries in move_freepages_block().
925 * Remove at a later date when no bug reports exist related to
926 * grouping pages by mobility
928 BUG_ON(page_zone(start_page) != page_zone(end_page));
931 for (page = start_page; page <= end_page;) {
932 /* Make sure we are not inadvertently changing nodes */
933 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
935 if (!pfn_valid_within(page_to_pfn(page))) {
940 if (!PageBuddy(page)) {
945 order = page_order(page);
946 list_move(&page->lru,
947 &zone->free_area[order].free_list[migratetype]);
949 pages_moved += 1 << order;
955 static int move_freepages_block(struct zone *zone, struct page *page,
958 unsigned long start_pfn, end_pfn;
959 struct page *start_page, *end_page;
961 start_pfn = page_to_pfn(page);
962 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
963 start_page = pfn_to_page(start_pfn);
964 end_page = start_page + pageblock_nr_pages - 1;
965 end_pfn = start_pfn + pageblock_nr_pages - 1;
967 /* Do not cross zone boundaries */
968 if (start_pfn < zone->zone_start_pfn)
970 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
973 return move_freepages(zone, start_page, end_page, migratetype);
976 static void change_pageblock_range(struct page *pageblock_page,
977 int start_order, int migratetype)
979 int nr_pageblocks = 1 << (start_order - pageblock_order);
981 while (nr_pageblocks--) {
982 set_pageblock_migratetype(pageblock_page, migratetype);
983 pageblock_page += pageblock_nr_pages;
987 /* Remove an element from the buddy allocator from the fallback list */
988 static inline struct page *
989 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
991 struct free_area * area;
996 /* Find the largest possible block of pages in the other list */
997 for (current_order = MAX_ORDER-1; current_order >= order;
999 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
1000 migratetype = fallbacks[start_migratetype][i];
1002 /* MIGRATE_RESERVE handled later if necessary */
1003 if (migratetype == MIGRATE_RESERVE)
1006 area = &(zone->free_area[current_order]);
1007 if (list_empty(&area->free_list[migratetype]))
1010 page = list_entry(area->free_list[migratetype].next,
1015 * If breaking a large block of pages, move all free
1016 * pages to the preferred allocation list. If falling
1017 * back for a reclaimable kernel allocation, be more
1018 * aggressive about taking ownership of free pages
1020 if (unlikely(current_order >= (pageblock_order >> 1)) ||
1021 start_migratetype == MIGRATE_RECLAIMABLE ||
1022 page_group_by_mobility_disabled) {
1023 unsigned long pages;
1024 pages = move_freepages_block(zone, page,
1027 /* Claim the whole block if over half of it is free */
1028 if (pages >= (1 << (pageblock_order-1)) ||
1029 page_group_by_mobility_disabled)
1030 set_pageblock_migratetype(page,
1033 migratetype = start_migratetype;
1036 /* Remove the page from the freelists */
1037 list_del(&page->lru);
1038 rmv_page_order(page);
1040 /* Take ownership for orders >= pageblock_order */
1041 if (current_order >= pageblock_order)
1042 change_pageblock_range(page, current_order,
1045 expand(zone, page, order, current_order, area, migratetype);
1047 trace_mm_page_alloc_extfrag(page, order, current_order,
1048 start_migratetype, migratetype);
1058 * Do the hard work of removing an element from the buddy allocator.
1059 * Call me with the zone->lock already held.
1061 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1067 page = __rmqueue_smallest(zone, order, migratetype);
1069 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1070 page = __rmqueue_fallback(zone, order, migratetype);
1073 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1074 * is used because __rmqueue_smallest is an inline function
1075 * and we want just one call site
1078 migratetype = MIGRATE_RESERVE;
1083 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1088 * Obtain a specified number of elements from the buddy allocator, all under
1089 * a single hold of the lock, for efficiency. Add them to the supplied list.
1090 * Returns the number of new pages which were placed at *list.
1092 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1093 unsigned long count, struct list_head *list,
1094 int migratetype, int cold)
1098 spin_lock(&zone->lock);
1099 for (i = 0; i < count; ++i) {
1100 struct page *page = __rmqueue(zone, order, migratetype);
1101 if (unlikely(page == NULL))
1105 * Split buddy pages returned by expand() are received here
1106 * in physical page order. The page is added to the callers and
1107 * list and the list head then moves forward. From the callers
1108 * perspective, the linked list is ordered by page number in
1109 * some conditions. This is useful for IO devices that can
1110 * merge IO requests if the physical pages are ordered
1113 if (likely(cold == 0))
1114 list_add(&page->lru, list);
1116 list_add_tail(&page->lru, list);
1117 set_page_private(page, migratetype);
1120 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1121 spin_unlock(&zone->lock);
1127 * Called from the vmstat counter updater to drain pagesets of this
1128 * currently executing processor on remote nodes after they have
1131 * Note that this function must be called with the thread pinned to
1132 * a single processor.
1134 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1136 unsigned long flags;
1139 local_irq_save(flags);
1140 if (pcp->count >= pcp->batch)
1141 to_drain = pcp->batch;
1143 to_drain = pcp->count;
1144 free_pcppages_bulk(zone, to_drain, pcp);
1145 pcp->count -= to_drain;
1146 local_irq_restore(flags);
1151 * Drain pages of the indicated processor.
1153 * The processor must either be the current processor and the
1154 * thread pinned to the current processor or a processor that
1157 static void drain_pages(unsigned int cpu)
1159 unsigned long flags;
1162 for_each_populated_zone(zone) {
1163 struct per_cpu_pageset *pset;
1164 struct per_cpu_pages *pcp;
1166 local_irq_save(flags);
1167 pset = per_cpu_ptr(zone->pageset, cpu);
1171 free_pcppages_bulk(zone, pcp->count, pcp);
1174 local_irq_restore(flags);
1179 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1181 void drain_local_pages(void *arg)
1183 drain_pages(smp_processor_id());
1187 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1189 void drain_all_pages(void)
1191 on_each_cpu(drain_local_pages, NULL, 1);
1194 #ifdef CONFIG_HIBERNATION
1196 void mark_free_pages(struct zone *zone)
1198 unsigned long pfn, max_zone_pfn;
1199 unsigned long flags;
1201 struct list_head *curr;
1203 if (!zone->spanned_pages)
1206 spin_lock_irqsave(&zone->lock, flags);
1208 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1209 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1210 if (pfn_valid(pfn)) {
1211 struct page *page = pfn_to_page(pfn);
1213 if (!swsusp_page_is_forbidden(page))
1214 swsusp_unset_page_free(page);
1217 for_each_migratetype_order(order, t) {
1218 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1221 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1222 for (i = 0; i < (1UL << order); i++)
1223 swsusp_set_page_free(pfn_to_page(pfn + i));
1226 spin_unlock_irqrestore(&zone->lock, flags);
1228 #endif /* CONFIG_PM */
1231 * Free a 0-order page
1232 * cold == 1 ? free a cold page : free a hot page
1234 void free_hot_cold_page(struct page *page, int cold)
1236 struct zone *zone = page_zone(page);
1237 struct per_cpu_pages *pcp;
1238 unsigned long flags;
1240 int wasMlocked = __TestClearPageMlocked(page);
1242 if (!free_pages_prepare(page, 0))
1245 migratetype = get_pageblock_migratetype(page);
1246 set_page_private(page, migratetype);
1247 local_irq_save(flags);
1248 if (unlikely(wasMlocked))
1249 free_page_mlock(page);
1250 __count_vm_event(PGFREE);
1253 * We only track unmovable, reclaimable and movable on pcp lists.
1254 * Free ISOLATE pages back to the allocator because they are being
1255 * offlined but treat RESERVE as movable pages so we can get those
1256 * areas back if necessary. Otherwise, we may have to free
1257 * excessively into the page allocator
1259 if (migratetype >= MIGRATE_PCPTYPES) {
1260 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1261 free_one_page(zone, page, 0, migratetype);
1264 migratetype = MIGRATE_MOVABLE;
1267 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1269 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1271 list_add(&page->lru, &pcp->lists[migratetype]);
1273 if (pcp->count >= pcp->high) {
1274 free_pcppages_bulk(zone, pcp->batch, pcp);
1275 pcp->count -= pcp->batch;
1279 local_irq_restore(flags);
1283 * Free a list of 0-order pages
1285 void free_hot_cold_page_list(struct list_head *list, int cold)
1287 struct page *page, *next;
1289 list_for_each_entry_safe(page, next, list, lru) {
1290 trace_mm_page_free_batched(page, cold);
1291 free_hot_cold_page(page, cold);
1296 * split_page takes a non-compound higher-order page, and splits it into
1297 * n (1<<order) sub-pages: page[0..n]
1298 * Each sub-page must be freed individually.
1300 * Note: this is probably too low level an operation for use in drivers.
1301 * Please consult with lkml before using this in your driver.
1303 void split_page(struct page *page, unsigned int order)
1307 VM_BUG_ON(PageCompound(page));
1308 VM_BUG_ON(!page_count(page));
1310 #ifdef CONFIG_KMEMCHECK
1312 * Split shadow pages too, because free(page[0]) would
1313 * otherwise free the whole shadow.
1315 if (kmemcheck_page_is_tracked(page))
1316 split_page(virt_to_page(page[0].shadow), order);
1319 for (i = 1; i < (1 << order); i++)
1320 set_page_refcounted(page + i);
1324 * Similar to split_page except the page is already free. As this is only
1325 * being used for migration, the migratetype of the block also changes.
1326 * As this is called with interrupts disabled, the caller is responsible
1327 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1330 * Note: this is probably too low level an operation for use in drivers.
1331 * Please consult with lkml before using this in your driver.
1333 int split_free_page(struct page *page)
1336 unsigned long watermark;
1339 BUG_ON(!PageBuddy(page));
1341 zone = page_zone(page);
1342 order = page_order(page);
1344 /* Obey watermarks as if the page was being allocated */
1345 watermark = low_wmark_pages(zone) + (1 << order);
1346 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1349 /* Remove page from free list */
1350 list_del(&page->lru);
1351 zone->free_area[order].nr_free--;
1352 rmv_page_order(page);
1353 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1355 /* Split into individual pages */
1356 set_page_refcounted(page);
1357 split_page(page, order);
1359 if (order >= pageblock_order - 1) {
1360 struct page *endpage = page + (1 << order) - 1;
1361 for (; page < endpage; page += pageblock_nr_pages)
1362 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1369 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1370 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1374 struct page *buffered_rmqueue(struct zone *preferred_zone,
1375 struct zone *zone, int order, gfp_t gfp_flags,
1378 unsigned long flags;
1380 int cold = !!(gfp_flags & __GFP_COLD);
1383 if (likely(order == 0)) {
1384 struct per_cpu_pages *pcp;
1385 struct list_head *list;
1387 local_irq_save(flags);
1388 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1389 list = &pcp->lists[migratetype];
1390 if (list_empty(list)) {
1391 pcp->count += rmqueue_bulk(zone, 0,
1394 if (unlikely(list_empty(list)))
1399 page = list_entry(list->prev, struct page, lru);
1401 page = list_entry(list->next, struct page, lru);
1403 list_del(&page->lru);
1406 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1408 * __GFP_NOFAIL is not to be used in new code.
1410 * All __GFP_NOFAIL callers should be fixed so that they
1411 * properly detect and handle allocation failures.
1413 * We most definitely don't want callers attempting to
1414 * allocate greater than order-1 page units with
1417 WARN_ON_ONCE(order > 1);
1419 spin_lock_irqsave(&zone->lock, flags);
1420 page = __rmqueue(zone, order, migratetype);
1421 spin_unlock(&zone->lock);
1424 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1427 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1428 zone_statistics(preferred_zone, zone, gfp_flags);
1429 local_irq_restore(flags);
1431 VM_BUG_ON(bad_range(zone, page));
1432 if (prep_new_page(page, order, gfp_flags))
1437 local_irq_restore(flags);
1441 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1442 #define ALLOC_WMARK_MIN WMARK_MIN
1443 #define ALLOC_WMARK_LOW WMARK_LOW
1444 #define ALLOC_WMARK_HIGH WMARK_HIGH
1445 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1447 /* Mask to get the watermark bits */
1448 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1450 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1451 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1452 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1454 #ifdef CONFIG_FAIL_PAGE_ALLOC
1457 struct fault_attr attr;
1459 u32 ignore_gfp_highmem;
1460 u32 ignore_gfp_wait;
1462 } fail_page_alloc = {
1463 .attr = FAULT_ATTR_INITIALIZER,
1464 .ignore_gfp_wait = 1,
1465 .ignore_gfp_highmem = 1,
1469 static int __init setup_fail_page_alloc(char *str)
1471 return setup_fault_attr(&fail_page_alloc.attr, str);
1473 __setup("fail_page_alloc=", setup_fail_page_alloc);
1475 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1477 if (order < fail_page_alloc.min_order)
1479 if (gfp_mask & __GFP_NOFAIL)
1481 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1483 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1486 return should_fail(&fail_page_alloc.attr, 1 << order);
1489 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1491 static int __init fail_page_alloc_debugfs(void)
1493 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1496 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1497 &fail_page_alloc.attr);
1499 return PTR_ERR(dir);
1501 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1502 &fail_page_alloc.ignore_gfp_wait))
1504 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1505 &fail_page_alloc.ignore_gfp_highmem))
1507 if (!debugfs_create_u32("min-order", mode, dir,
1508 &fail_page_alloc.min_order))
1513 debugfs_remove_recursive(dir);
1518 late_initcall(fail_page_alloc_debugfs);
1520 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1522 #else /* CONFIG_FAIL_PAGE_ALLOC */
1524 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1529 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1532 * Return true if free pages are above 'mark'. This takes into account the order
1533 * of the allocation.
1535 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1536 int classzone_idx, int alloc_flags, long free_pages)
1538 /* free_pages my go negative - that's OK */
1542 free_pages -= (1 << order) + 1;
1543 if (alloc_flags & ALLOC_HIGH)
1545 if (alloc_flags & ALLOC_HARDER)
1548 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1550 for (o = 0; o < order; o++) {
1551 /* At the next order, this order's pages become unavailable */
1552 free_pages -= z->free_area[o].nr_free << o;
1554 /* Require fewer higher order pages to be free */
1557 if (free_pages <= min)
1563 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1564 int classzone_idx, int alloc_flags)
1566 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1567 zone_page_state(z, NR_FREE_PAGES));
1570 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1571 int classzone_idx, int alloc_flags)
1573 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1575 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1576 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1578 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1584 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1585 * skip over zones that are not allowed by the cpuset, or that have
1586 * been recently (in last second) found to be nearly full. See further
1587 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1588 * that have to skip over a lot of full or unallowed zones.
1590 * If the zonelist cache is present in the passed in zonelist, then
1591 * returns a pointer to the allowed node mask (either the current
1592 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1594 * If the zonelist cache is not available for this zonelist, does
1595 * nothing and returns NULL.
1597 * If the fullzones BITMAP in the zonelist cache is stale (more than
1598 * a second since last zap'd) then we zap it out (clear its bits.)
1600 * We hold off even calling zlc_setup, until after we've checked the
1601 * first zone in the zonelist, on the theory that most allocations will
1602 * be satisfied from that first zone, so best to examine that zone as
1603 * quickly as we can.
1605 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1607 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1608 nodemask_t *allowednodes; /* zonelist_cache approximation */
1610 zlc = zonelist->zlcache_ptr;
1614 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1615 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1616 zlc->last_full_zap = jiffies;
1619 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1620 &cpuset_current_mems_allowed :
1621 &node_states[N_HIGH_MEMORY];
1622 return allowednodes;
1626 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1627 * if it is worth looking at further for free memory:
1628 * 1) Check that the zone isn't thought to be full (doesn't have its
1629 * bit set in the zonelist_cache fullzones BITMAP).
1630 * 2) Check that the zones node (obtained from the zonelist_cache
1631 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1632 * Return true (non-zero) if zone is worth looking at further, or
1633 * else return false (zero) if it is not.
1635 * This check -ignores- the distinction between various watermarks,
1636 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1637 * found to be full for any variation of these watermarks, it will
1638 * be considered full for up to one second by all requests, unless
1639 * we are so low on memory on all allowed nodes that we are forced
1640 * into the second scan of the zonelist.
1642 * In the second scan we ignore this zonelist cache and exactly
1643 * apply the watermarks to all zones, even it is slower to do so.
1644 * We are low on memory in the second scan, and should leave no stone
1645 * unturned looking for a free page.
1647 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1648 nodemask_t *allowednodes)
1650 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1651 int i; /* index of *z in zonelist zones */
1652 int n; /* node that zone *z is on */
1654 zlc = zonelist->zlcache_ptr;
1658 i = z - zonelist->_zonerefs;
1661 /* This zone is worth trying if it is allowed but not full */
1662 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1666 * Given 'z' scanning a zonelist, set the corresponding bit in
1667 * zlc->fullzones, so that subsequent attempts to allocate a page
1668 * from that zone don't waste time re-examining it.
1670 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1672 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1673 int i; /* index of *z in zonelist zones */
1675 zlc = zonelist->zlcache_ptr;
1679 i = z - zonelist->_zonerefs;
1681 set_bit(i, zlc->fullzones);
1685 * clear all zones full, called after direct reclaim makes progress so that
1686 * a zone that was recently full is not skipped over for up to a second
1688 static void zlc_clear_zones_full(struct zonelist *zonelist)
1690 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1692 zlc = zonelist->zlcache_ptr;
1696 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1699 #else /* CONFIG_NUMA */
1701 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1706 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1707 nodemask_t *allowednodes)
1712 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1716 static void zlc_clear_zones_full(struct zonelist *zonelist)
1719 #endif /* CONFIG_NUMA */
1722 * get_page_from_freelist goes through the zonelist trying to allocate
1725 static struct page *
1726 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1727 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1728 struct zone *preferred_zone, int migratetype)
1731 struct page *page = NULL;
1734 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1735 int zlc_active = 0; /* set if using zonelist_cache */
1736 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1738 classzone_idx = zone_idx(preferred_zone);
1741 * Scan zonelist, looking for a zone with enough free.
1742 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1744 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1745 high_zoneidx, nodemask) {
1746 if (NUMA_BUILD && zlc_active &&
1747 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1749 if ((alloc_flags & ALLOC_CPUSET) &&
1750 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1753 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1754 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1758 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1759 if (zone_watermark_ok(zone, order, mark,
1760 classzone_idx, alloc_flags))
1763 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1765 * we do zlc_setup if there are multiple nodes
1766 * and before considering the first zone allowed
1769 allowednodes = zlc_setup(zonelist, alloc_flags);
1774 if (zone_reclaim_mode == 0)
1775 goto this_zone_full;
1778 * As we may have just activated ZLC, check if the first
1779 * eligible zone has failed zone_reclaim recently.
1781 if (NUMA_BUILD && zlc_active &&
1782 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1785 ret = zone_reclaim(zone, gfp_mask, order);
1787 case ZONE_RECLAIM_NOSCAN:
1790 case ZONE_RECLAIM_FULL:
1791 /* scanned but unreclaimable */
1794 /* did we reclaim enough */
1795 if (!zone_watermark_ok(zone, order, mark,
1796 classzone_idx, alloc_flags))
1797 goto this_zone_full;
1802 page = buffered_rmqueue(preferred_zone, zone, order,
1803 gfp_mask, migratetype);
1808 zlc_mark_zone_full(zonelist, z);
1811 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1812 /* Disable zlc cache for second zonelist scan */
1820 * Large machines with many possible nodes should not always dump per-node
1821 * meminfo in irq context.
1823 static inline bool should_suppress_show_mem(void)
1828 ret = in_interrupt();
1833 static DEFINE_RATELIMIT_STATE(nopage_rs,
1834 DEFAULT_RATELIMIT_INTERVAL,
1835 DEFAULT_RATELIMIT_BURST);
1837 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1839 unsigned int filter = SHOW_MEM_FILTER_NODES;
1841 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1842 debug_guardpage_minorder() > 0)
1846 * This documents exceptions given to allocations in certain
1847 * contexts that are allowed to allocate outside current's set
1850 if (!(gfp_mask & __GFP_NOMEMALLOC))
1851 if (test_thread_flag(TIF_MEMDIE) ||
1852 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1853 filter &= ~SHOW_MEM_FILTER_NODES;
1854 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1855 filter &= ~SHOW_MEM_FILTER_NODES;
1858 struct va_format vaf;
1861 va_start(args, fmt);
1866 pr_warn("%pV", &vaf);
1871 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1872 current->comm, order, gfp_mask);
1875 if (!should_suppress_show_mem())
1880 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1881 unsigned long did_some_progress,
1882 unsigned long pages_reclaimed)
1884 /* Do not loop if specifically requested */
1885 if (gfp_mask & __GFP_NORETRY)
1888 /* Always retry if specifically requested */
1889 if (gfp_mask & __GFP_NOFAIL)
1893 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
1894 * making forward progress without invoking OOM. Suspend also disables
1895 * storage devices so kswapd will not help. Bail if we are suspending.
1897 if (!did_some_progress && pm_suspended_storage())
1901 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1902 * means __GFP_NOFAIL, but that may not be true in other
1905 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1909 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1910 * specified, then we retry until we no longer reclaim any pages
1911 * (above), or we've reclaimed an order of pages at least as
1912 * large as the allocation's order. In both cases, if the
1913 * allocation still fails, we stop retrying.
1915 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1921 static inline struct page *
1922 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1923 struct zonelist *zonelist, enum zone_type high_zoneidx,
1924 nodemask_t *nodemask, struct zone *preferred_zone,
1929 /* Acquire the OOM killer lock for the zones in zonelist */
1930 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1931 schedule_timeout_uninterruptible(1);
1936 * Go through the zonelist yet one more time, keep very high watermark
1937 * here, this is only to catch a parallel oom killing, we must fail if
1938 * we're still under heavy pressure.
1940 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1941 order, zonelist, high_zoneidx,
1942 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1943 preferred_zone, migratetype);
1947 if (!(gfp_mask & __GFP_NOFAIL)) {
1948 /* The OOM killer will not help higher order allocs */
1949 if (order > PAGE_ALLOC_COSTLY_ORDER)
1951 /* The OOM killer does not needlessly kill tasks for lowmem */
1952 if (high_zoneidx < ZONE_NORMAL)
1955 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1956 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1957 * The caller should handle page allocation failure by itself if
1958 * it specifies __GFP_THISNODE.
1959 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1961 if (gfp_mask & __GFP_THISNODE)
1964 /* Exhausted what can be done so it's blamo time */
1965 out_of_memory(zonelist, gfp_mask, order, nodemask);
1968 clear_zonelist_oom(zonelist, gfp_mask);
1972 #ifdef CONFIG_COMPACTION
1973 /* Try memory compaction for high-order allocations before reclaim */
1974 static struct page *
1975 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1976 struct zonelist *zonelist, enum zone_type high_zoneidx,
1977 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1978 int migratetype, bool sync_migration,
1979 bool *deferred_compaction,
1980 unsigned long *did_some_progress)
1987 if (compaction_deferred(preferred_zone)) {
1988 *deferred_compaction = true;
1992 current->flags |= PF_MEMALLOC;
1993 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1994 nodemask, sync_migration);
1995 current->flags &= ~PF_MEMALLOC;
1996 if (*did_some_progress != COMPACT_SKIPPED) {
1998 /* Page migration frees to the PCP lists but we want merging */
1999 drain_pages(get_cpu());
2002 page = get_page_from_freelist(gfp_mask, nodemask,
2003 order, zonelist, high_zoneidx,
2004 alloc_flags, preferred_zone,
2007 preferred_zone->compact_considered = 0;
2008 preferred_zone->compact_defer_shift = 0;
2009 count_vm_event(COMPACTSUCCESS);
2014 * It's bad if compaction run occurs and fails.
2015 * The most likely reason is that pages exist,
2016 * but not enough to satisfy watermarks.
2018 count_vm_event(COMPACTFAIL);
2021 * As async compaction considers a subset of pageblocks, only
2022 * defer if the failure was a sync compaction failure.
2025 defer_compaction(preferred_zone);
2033 static inline struct page *
2034 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2035 struct zonelist *zonelist, enum zone_type high_zoneidx,
2036 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2037 int migratetype, bool sync_migration,
2038 bool *deferred_compaction,
2039 unsigned long *did_some_progress)
2043 #endif /* CONFIG_COMPACTION */
2045 /* The really slow allocator path where we enter direct reclaim */
2046 static inline struct page *
2047 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2048 struct zonelist *zonelist, enum zone_type high_zoneidx,
2049 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2050 int migratetype, unsigned long *did_some_progress)
2052 struct page *page = NULL;
2053 struct reclaim_state reclaim_state;
2054 bool drained = false;
2058 /* We now go into synchronous reclaim */
2059 cpuset_memory_pressure_bump();
2060 current->flags |= PF_MEMALLOC;
2061 lockdep_set_current_reclaim_state(gfp_mask);
2062 reclaim_state.reclaimed_slab = 0;
2063 current->reclaim_state = &reclaim_state;
2065 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2067 current->reclaim_state = NULL;
2068 lockdep_clear_current_reclaim_state();
2069 current->flags &= ~PF_MEMALLOC;
2073 if (unlikely(!(*did_some_progress)))
2076 /* After successful reclaim, reconsider all zones for allocation */
2078 zlc_clear_zones_full(zonelist);
2081 page = get_page_from_freelist(gfp_mask, nodemask, order,
2082 zonelist, high_zoneidx,
2083 alloc_flags, preferred_zone,
2087 * If an allocation failed after direct reclaim, it could be because
2088 * pages are pinned on the per-cpu lists. Drain them and try again
2090 if (!page && !drained) {
2100 * This is called in the allocator slow-path if the allocation request is of
2101 * sufficient urgency to ignore watermarks and take other desperate measures
2103 static inline struct page *
2104 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2105 struct zonelist *zonelist, enum zone_type high_zoneidx,
2106 nodemask_t *nodemask, struct zone *preferred_zone,
2112 page = get_page_from_freelist(gfp_mask, nodemask, order,
2113 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2114 preferred_zone, migratetype);
2116 if (!page && gfp_mask & __GFP_NOFAIL)
2117 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2118 } while (!page && (gfp_mask & __GFP_NOFAIL));
2124 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2125 enum zone_type high_zoneidx,
2126 enum zone_type classzone_idx)
2131 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2132 wakeup_kswapd(zone, order, classzone_idx);
2136 gfp_to_alloc_flags(gfp_t gfp_mask)
2138 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2139 const gfp_t wait = gfp_mask & __GFP_WAIT;
2141 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2142 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2145 * The caller may dip into page reserves a bit more if the caller
2146 * cannot run direct reclaim, or if the caller has realtime scheduling
2147 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2148 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2150 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2154 * Not worth trying to allocate harder for
2155 * __GFP_NOMEMALLOC even if it can't schedule.
2157 if (!(gfp_mask & __GFP_NOMEMALLOC))
2158 alloc_flags |= ALLOC_HARDER;
2160 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2161 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2163 alloc_flags &= ~ALLOC_CPUSET;
2164 } else if (unlikely(rt_task(current)) && !in_interrupt())
2165 alloc_flags |= ALLOC_HARDER;
2167 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2168 if (!in_interrupt() &&
2169 ((current->flags & PF_MEMALLOC) ||
2170 unlikely(test_thread_flag(TIF_MEMDIE))))
2171 alloc_flags |= ALLOC_NO_WATERMARKS;
2177 static inline struct page *
2178 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2179 struct zonelist *zonelist, enum zone_type high_zoneidx,
2180 nodemask_t *nodemask, struct zone *preferred_zone,
2183 const gfp_t wait = gfp_mask & __GFP_WAIT;
2184 struct page *page = NULL;
2186 unsigned long pages_reclaimed = 0;
2187 unsigned long did_some_progress;
2188 bool sync_migration = false;
2189 bool deferred_compaction = false;
2192 * In the slowpath, we sanity check order to avoid ever trying to
2193 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2194 * be using allocators in order of preference for an area that is
2197 if (order >= MAX_ORDER) {
2198 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2203 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2204 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2205 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2206 * using a larger set of nodes after it has established that the
2207 * allowed per node queues are empty and that nodes are
2210 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2214 if (!(gfp_mask & __GFP_NO_KSWAPD))
2215 wake_all_kswapd(order, zonelist, high_zoneidx,
2216 zone_idx(preferred_zone));
2219 * OK, we're below the kswapd watermark and have kicked background
2220 * reclaim. Now things get more complex, so set up alloc_flags according
2221 * to how we want to proceed.
2223 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2226 * Find the true preferred zone if the allocation is unconstrained by
2229 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2230 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2234 /* This is the last chance, in general, before the goto nopage. */
2235 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2236 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2237 preferred_zone, migratetype);
2241 /* Allocate without watermarks if the context allows */
2242 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2243 page = __alloc_pages_high_priority(gfp_mask, order,
2244 zonelist, high_zoneidx, nodemask,
2245 preferred_zone, migratetype);
2250 /* Atomic allocations - we can't balance anything */
2254 /* Avoid recursion of direct reclaim */
2255 if (current->flags & PF_MEMALLOC)
2258 /* Avoid allocations with no watermarks from looping endlessly */
2259 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2263 * Try direct compaction. The first pass is asynchronous. Subsequent
2264 * attempts after direct reclaim are synchronous
2266 page = __alloc_pages_direct_compact(gfp_mask, order,
2267 zonelist, high_zoneidx,
2269 alloc_flags, preferred_zone,
2270 migratetype, sync_migration,
2271 &deferred_compaction,
2272 &did_some_progress);
2275 sync_migration = true;
2278 * If compaction is deferred for high-order allocations, it is because
2279 * sync compaction recently failed. In this is the case and the caller
2280 * has requested the system not be heavily disrupted, fail the
2281 * allocation now instead of entering direct reclaim
2283 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2286 /* Try direct reclaim and then allocating */
2287 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2288 zonelist, high_zoneidx,
2290 alloc_flags, preferred_zone,
2291 migratetype, &did_some_progress);
2296 * If we failed to make any progress reclaiming, then we are
2297 * running out of options and have to consider going OOM
2299 if (!did_some_progress) {
2300 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2301 if (oom_killer_disabled)
2303 page = __alloc_pages_may_oom(gfp_mask, order,
2304 zonelist, high_zoneidx,
2305 nodemask, preferred_zone,
2310 if (!(gfp_mask & __GFP_NOFAIL)) {
2312 * The oom killer is not called for high-order
2313 * allocations that may fail, so if no progress
2314 * is being made, there are no other options and
2315 * retrying is unlikely to help.
2317 if (order > PAGE_ALLOC_COSTLY_ORDER)
2320 * The oom killer is not called for lowmem
2321 * allocations to prevent needlessly killing
2324 if (high_zoneidx < ZONE_NORMAL)
2332 /* Check if we should retry the allocation */
2333 pages_reclaimed += did_some_progress;
2334 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2336 /* Wait for some write requests to complete then retry */
2337 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2341 * High-order allocations do not necessarily loop after
2342 * direct reclaim and reclaim/compaction depends on compaction
2343 * being called after reclaim so call directly if necessary
2345 page = __alloc_pages_direct_compact(gfp_mask, order,
2346 zonelist, high_zoneidx,
2348 alloc_flags, preferred_zone,
2349 migratetype, sync_migration,
2350 &deferred_compaction,
2351 &did_some_progress);
2357 warn_alloc_failed(gfp_mask, order, NULL);
2360 if (kmemcheck_enabled)
2361 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2367 * This is the 'heart' of the zoned buddy allocator.
2370 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2371 struct zonelist *zonelist, nodemask_t *nodemask)
2373 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2374 struct zone *preferred_zone;
2375 struct page *page = NULL;
2376 int migratetype = allocflags_to_migratetype(gfp_mask);
2377 unsigned int cpuset_mems_cookie;
2379 gfp_mask &= gfp_allowed_mask;
2381 lockdep_trace_alloc(gfp_mask);
2383 might_sleep_if(gfp_mask & __GFP_WAIT);
2385 if (should_fail_alloc_page(gfp_mask, order))
2389 * Check the zones suitable for the gfp_mask contain at least one
2390 * valid zone. It's possible to have an empty zonelist as a result
2391 * of GFP_THISNODE and a memoryless node
2393 if (unlikely(!zonelist->_zonerefs->zone))
2397 cpuset_mems_cookie = get_mems_allowed();
2399 /* The preferred zone is used for statistics later */
2400 first_zones_zonelist(zonelist, high_zoneidx,
2401 nodemask ? : &cpuset_current_mems_allowed,
2403 if (!preferred_zone)
2406 /* First allocation attempt */
2407 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2408 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2409 preferred_zone, migratetype);
2410 if (unlikely(!page))
2411 page = __alloc_pages_slowpath(gfp_mask, order,
2412 zonelist, high_zoneidx, nodemask,
2413 preferred_zone, migratetype);
2415 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2419 * When updating a task's mems_allowed, it is possible to race with
2420 * parallel threads in such a way that an allocation can fail while
2421 * the mask is being updated. If a page allocation is about to fail,
2422 * check if the cpuset changed during allocation and if so, retry.
2424 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2429 EXPORT_SYMBOL(__alloc_pages_nodemask);
2432 * Common helper functions.
2434 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2439 * __get_free_pages() returns a 32-bit address, which cannot represent
2442 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2444 page = alloc_pages(gfp_mask, order);
2447 return (unsigned long) page_address(page);
2449 EXPORT_SYMBOL(__get_free_pages);
2451 unsigned long get_zeroed_page(gfp_t gfp_mask)
2453 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2455 EXPORT_SYMBOL(get_zeroed_page);
2457 void __free_pages(struct page *page, unsigned int order)
2459 if (put_page_testzero(page)) {
2461 free_hot_cold_page(page, 0);
2463 __free_pages_ok(page, order);
2467 EXPORT_SYMBOL(__free_pages);
2469 void free_pages(unsigned long addr, unsigned int order)
2472 VM_BUG_ON(!virt_addr_valid((void *)addr));
2473 __free_pages(virt_to_page((void *)addr), order);
2477 EXPORT_SYMBOL(free_pages);
2479 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2482 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2483 unsigned long used = addr + PAGE_ALIGN(size);
2485 split_page(virt_to_page((void *)addr), order);
2486 while (used < alloc_end) {
2491 return (void *)addr;
2495 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2496 * @size: the number of bytes to allocate
2497 * @gfp_mask: GFP flags for the allocation
2499 * This function is similar to alloc_pages(), except that it allocates the
2500 * minimum number of pages to satisfy the request. alloc_pages() can only
2501 * allocate memory in power-of-two pages.
2503 * This function is also limited by MAX_ORDER.
2505 * Memory allocated by this function must be released by free_pages_exact().
2507 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2509 unsigned int order = get_order(size);
2512 addr = __get_free_pages(gfp_mask, order);
2513 return make_alloc_exact(addr, order, size);
2515 EXPORT_SYMBOL(alloc_pages_exact);
2518 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2520 * @nid: the preferred node ID where memory should be allocated
2521 * @size: the number of bytes to allocate
2522 * @gfp_mask: GFP flags for the allocation
2524 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2526 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2529 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2531 unsigned order = get_order(size);
2532 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2535 return make_alloc_exact((unsigned long)page_address(p), order, size);
2537 EXPORT_SYMBOL(alloc_pages_exact_nid);
2540 * free_pages_exact - release memory allocated via alloc_pages_exact()
2541 * @virt: the value returned by alloc_pages_exact.
2542 * @size: size of allocation, same value as passed to alloc_pages_exact().
2544 * Release the memory allocated by a previous call to alloc_pages_exact.
2546 void free_pages_exact(void *virt, size_t size)
2548 unsigned long addr = (unsigned long)virt;
2549 unsigned long end = addr + PAGE_ALIGN(size);
2551 while (addr < end) {
2556 EXPORT_SYMBOL(free_pages_exact);
2558 static unsigned int nr_free_zone_pages(int offset)
2563 /* Just pick one node, since fallback list is circular */
2564 unsigned int sum = 0;
2566 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2568 for_each_zone_zonelist(zone, z, zonelist, offset) {
2569 unsigned long size = zone->present_pages;
2570 unsigned long high = high_wmark_pages(zone);
2579 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2581 unsigned int nr_free_buffer_pages(void)
2583 return nr_free_zone_pages(gfp_zone(GFP_USER));
2585 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2588 * Amount of free RAM allocatable within all zones
2590 unsigned int nr_free_pagecache_pages(void)
2592 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2595 static inline void show_node(struct zone *zone)
2598 printk("Node %d ", zone_to_nid(zone));
2601 void si_meminfo(struct sysinfo *val)
2603 val->totalram = totalram_pages;
2605 val->freeram = global_page_state(NR_FREE_PAGES);
2606 val->bufferram = nr_blockdev_pages();
2607 val->totalhigh = totalhigh_pages;
2608 val->freehigh = nr_free_highpages();
2609 val->mem_unit = PAGE_SIZE;
2612 EXPORT_SYMBOL(si_meminfo);
2615 void si_meminfo_node(struct sysinfo *val, int nid)
2617 pg_data_t *pgdat = NODE_DATA(nid);
2619 val->totalram = pgdat->node_present_pages;
2620 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2621 #ifdef CONFIG_HIGHMEM
2622 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2623 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2629 val->mem_unit = PAGE_SIZE;
2634 * Determine whether the node should be displayed or not, depending on whether
2635 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2637 bool skip_free_areas_node(unsigned int flags, int nid)
2640 unsigned int cpuset_mems_cookie;
2642 if (!(flags & SHOW_MEM_FILTER_NODES))
2646 cpuset_mems_cookie = get_mems_allowed();
2647 ret = !node_isset(nid, cpuset_current_mems_allowed);
2648 } while (!put_mems_allowed(cpuset_mems_cookie));
2653 #define K(x) ((x) << (PAGE_SHIFT-10))
2656 * Show free area list (used inside shift_scroll-lock stuff)
2657 * We also calculate the percentage fragmentation. We do this by counting the
2658 * memory on each free list with the exception of the first item on the list.
2659 * Suppresses nodes that are not allowed by current's cpuset if
2660 * SHOW_MEM_FILTER_NODES is passed.
2662 void show_free_areas(unsigned int filter)
2667 for_each_populated_zone(zone) {
2668 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2671 printk("%s per-cpu:\n", zone->name);
2673 for_each_online_cpu(cpu) {
2674 struct per_cpu_pageset *pageset;
2676 pageset = per_cpu_ptr(zone->pageset, cpu);
2678 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2679 cpu, pageset->pcp.high,
2680 pageset->pcp.batch, pageset->pcp.count);
2684 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2685 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2687 " dirty:%lu writeback:%lu unstable:%lu\n"
2688 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2689 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2690 global_page_state(NR_ACTIVE_ANON),
2691 global_page_state(NR_INACTIVE_ANON),
2692 global_page_state(NR_ISOLATED_ANON),
2693 global_page_state(NR_ACTIVE_FILE),
2694 global_page_state(NR_INACTIVE_FILE),
2695 global_page_state(NR_ISOLATED_FILE),
2696 global_page_state(NR_UNEVICTABLE),
2697 global_page_state(NR_FILE_DIRTY),
2698 global_page_state(NR_WRITEBACK),
2699 global_page_state(NR_UNSTABLE_NFS),
2700 global_page_state(NR_FREE_PAGES),
2701 global_page_state(NR_SLAB_RECLAIMABLE),
2702 global_page_state(NR_SLAB_UNRECLAIMABLE),
2703 global_page_state(NR_FILE_MAPPED),
2704 global_page_state(NR_SHMEM),
2705 global_page_state(NR_PAGETABLE),
2706 global_page_state(NR_BOUNCE));
2708 for_each_populated_zone(zone) {
2711 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2719 " active_anon:%lukB"
2720 " inactive_anon:%lukB"
2721 " active_file:%lukB"
2722 " inactive_file:%lukB"
2723 " unevictable:%lukB"
2724 " isolated(anon):%lukB"
2725 " isolated(file):%lukB"
2732 " slab_reclaimable:%lukB"
2733 " slab_unreclaimable:%lukB"
2734 " kernel_stack:%lukB"
2738 " writeback_tmp:%lukB"
2739 " pages_scanned:%lu"
2740 " all_unreclaimable? %s"
2743 K(zone_page_state(zone, NR_FREE_PAGES)),
2744 K(min_wmark_pages(zone)),
2745 K(low_wmark_pages(zone)),
2746 K(high_wmark_pages(zone)),
2747 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2748 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2749 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2750 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2751 K(zone_page_state(zone, NR_UNEVICTABLE)),
2752 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2753 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2754 K(zone->present_pages),
2755 K(zone_page_state(zone, NR_MLOCK)),
2756 K(zone_page_state(zone, NR_FILE_DIRTY)),
2757 K(zone_page_state(zone, NR_WRITEBACK)),
2758 K(zone_page_state(zone, NR_FILE_MAPPED)),
2759 K(zone_page_state(zone, NR_SHMEM)),
2760 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2761 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2762 zone_page_state(zone, NR_KERNEL_STACK) *
2764 K(zone_page_state(zone, NR_PAGETABLE)),
2765 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2766 K(zone_page_state(zone, NR_BOUNCE)),
2767 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2768 zone->pages_scanned,
2769 (zone->all_unreclaimable ? "yes" : "no")
2771 printk("lowmem_reserve[]:");
2772 for (i = 0; i < MAX_NR_ZONES; i++)
2773 printk(" %lu", zone->lowmem_reserve[i]);
2777 for_each_populated_zone(zone) {
2778 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2780 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2783 printk("%s: ", zone->name);
2785 spin_lock_irqsave(&zone->lock, flags);
2786 for (order = 0; order < MAX_ORDER; order++) {
2787 nr[order] = zone->free_area[order].nr_free;
2788 total += nr[order] << order;
2790 spin_unlock_irqrestore(&zone->lock, flags);
2791 for (order = 0; order < MAX_ORDER; order++)
2792 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2793 printk("= %lukB\n", K(total));
2796 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2798 show_swap_cache_info();
2801 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2803 zoneref->zone = zone;
2804 zoneref->zone_idx = zone_idx(zone);
2808 * Builds allocation fallback zone lists.
2810 * Add all populated zones of a node to the zonelist.
2812 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2813 int nr_zones, enum zone_type zone_type)
2817 BUG_ON(zone_type >= MAX_NR_ZONES);
2822 zone = pgdat->node_zones + zone_type;
2823 if (populated_zone(zone)) {
2824 zoneref_set_zone(zone,
2825 &zonelist->_zonerefs[nr_zones++]);
2826 check_highest_zone(zone_type);
2829 } while (zone_type);
2836 * 0 = automatic detection of better ordering.
2837 * 1 = order by ([node] distance, -zonetype)
2838 * 2 = order by (-zonetype, [node] distance)
2840 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2841 * the same zonelist. So only NUMA can configure this param.
2843 #define ZONELIST_ORDER_DEFAULT 0
2844 #define ZONELIST_ORDER_NODE 1
2845 #define ZONELIST_ORDER_ZONE 2
2847 /* zonelist order in the kernel.
2848 * set_zonelist_order() will set this to NODE or ZONE.
2850 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2851 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2855 /* The value user specified ....changed by config */
2856 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2857 /* string for sysctl */
2858 #define NUMA_ZONELIST_ORDER_LEN 16
2859 char numa_zonelist_order[16] = "default";
2862 * interface for configure zonelist ordering.
2863 * command line option "numa_zonelist_order"
2864 * = "[dD]efault - default, automatic configuration.
2865 * = "[nN]ode - order by node locality, then by zone within node
2866 * = "[zZ]one - order by zone, then by locality within zone
2869 static int __parse_numa_zonelist_order(char *s)
2871 if (*s == 'd' || *s == 'D') {
2872 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2873 } else if (*s == 'n' || *s == 'N') {
2874 user_zonelist_order = ZONELIST_ORDER_NODE;
2875 } else if (*s == 'z' || *s == 'Z') {
2876 user_zonelist_order = ZONELIST_ORDER_ZONE;
2879 "Ignoring invalid numa_zonelist_order value: "
2886 static __init int setup_numa_zonelist_order(char *s)
2893 ret = __parse_numa_zonelist_order(s);
2895 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2899 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2902 * sysctl handler for numa_zonelist_order
2904 int numa_zonelist_order_handler(ctl_table *table, int write,
2905 void __user *buffer, size_t *length,
2908 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2910 static DEFINE_MUTEX(zl_order_mutex);
2912 mutex_lock(&zl_order_mutex);
2914 strcpy(saved_string, (char*)table->data);
2915 ret = proc_dostring(table, write, buffer, length, ppos);
2919 int oldval = user_zonelist_order;
2920 if (__parse_numa_zonelist_order((char*)table->data)) {
2922 * bogus value. restore saved string
2924 strncpy((char*)table->data, saved_string,
2925 NUMA_ZONELIST_ORDER_LEN);
2926 user_zonelist_order = oldval;
2927 } else if (oldval != user_zonelist_order) {
2928 mutex_lock(&zonelists_mutex);
2929 build_all_zonelists(NULL);
2930 mutex_unlock(&zonelists_mutex);
2934 mutex_unlock(&zl_order_mutex);
2939 #define MAX_NODE_LOAD (nr_online_nodes)
2940 static int node_load[MAX_NUMNODES];
2943 * find_next_best_node - find the next node that should appear in a given node's fallback list
2944 * @node: node whose fallback list we're appending
2945 * @used_node_mask: nodemask_t of already used nodes
2947 * We use a number of factors to determine which is the next node that should
2948 * appear on a given node's fallback list. The node should not have appeared
2949 * already in @node's fallback list, and it should be the next closest node
2950 * according to the distance array (which contains arbitrary distance values
2951 * from each node to each node in the system), and should also prefer nodes
2952 * with no CPUs, since presumably they'll have very little allocation pressure
2953 * on them otherwise.
2954 * It returns -1 if no node is found.
2956 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2959 int min_val = INT_MAX;
2961 const struct cpumask *tmp = cpumask_of_node(0);
2963 /* Use the local node if we haven't already */
2964 if (!node_isset(node, *used_node_mask)) {
2965 node_set(node, *used_node_mask);
2969 for_each_node_state(n, N_HIGH_MEMORY) {
2971 /* Don't want a node to appear more than once */
2972 if (node_isset(n, *used_node_mask))
2975 /* Use the distance array to find the distance */
2976 val = node_distance(node, n);
2978 /* Penalize nodes under us ("prefer the next node") */
2981 /* Give preference to headless and unused nodes */
2982 tmp = cpumask_of_node(n);
2983 if (!cpumask_empty(tmp))
2984 val += PENALTY_FOR_NODE_WITH_CPUS;
2986 /* Slight preference for less loaded node */
2987 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2988 val += node_load[n];
2990 if (val < min_val) {
2997 node_set(best_node, *used_node_mask);
3004 * Build zonelists ordered by node and zones within node.
3005 * This results in maximum locality--normal zone overflows into local
3006 * DMA zone, if any--but risks exhausting DMA zone.
3008 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3011 struct zonelist *zonelist;
3013 zonelist = &pgdat->node_zonelists[0];
3014 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3016 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3018 zonelist->_zonerefs[j].zone = NULL;
3019 zonelist->_zonerefs[j].zone_idx = 0;
3023 * Build gfp_thisnode zonelists
3025 static void build_thisnode_zonelists(pg_data_t *pgdat)
3028 struct zonelist *zonelist;
3030 zonelist = &pgdat->node_zonelists[1];
3031 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3032 zonelist->_zonerefs[j].zone = NULL;
3033 zonelist->_zonerefs[j].zone_idx = 0;
3037 * Build zonelists ordered by zone and nodes within zones.
3038 * This results in conserving DMA zone[s] until all Normal memory is
3039 * exhausted, but results in overflowing to remote node while memory
3040 * may still exist in local DMA zone.
3042 static int node_order[MAX_NUMNODES];
3044 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3047 int zone_type; /* needs to be signed */
3049 struct zonelist *zonelist;
3051 zonelist = &pgdat->node_zonelists[0];
3053 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3054 for (j = 0; j < nr_nodes; j++) {
3055 node = node_order[j];
3056 z = &NODE_DATA(node)->node_zones[zone_type];
3057 if (populated_zone(z)) {
3059 &zonelist->_zonerefs[pos++]);
3060 check_highest_zone(zone_type);
3064 zonelist->_zonerefs[pos].zone = NULL;
3065 zonelist->_zonerefs[pos].zone_idx = 0;
3068 static int default_zonelist_order(void)
3071 unsigned long low_kmem_size,total_size;
3075 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3076 * If they are really small and used heavily, the system can fall
3077 * into OOM very easily.
3078 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3080 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3083 for_each_online_node(nid) {
3084 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3085 z = &NODE_DATA(nid)->node_zones[zone_type];
3086 if (populated_zone(z)) {
3087 if (zone_type < ZONE_NORMAL)
3088 low_kmem_size += z->present_pages;
3089 total_size += z->present_pages;
3090 } else if (zone_type == ZONE_NORMAL) {
3092 * If any node has only lowmem, then node order
3093 * is preferred to allow kernel allocations
3094 * locally; otherwise, they can easily infringe
3095 * on other nodes when there is an abundance of
3096 * lowmem available to allocate from.
3098 return ZONELIST_ORDER_NODE;
3102 if (!low_kmem_size || /* there are no DMA area. */
3103 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3104 return ZONELIST_ORDER_NODE;
3106 * look into each node's config.
3107 * If there is a node whose DMA/DMA32 memory is very big area on
3108 * local memory, NODE_ORDER may be suitable.
3110 average_size = total_size /
3111 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
3112 for_each_online_node(nid) {
3115 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3116 z = &NODE_DATA(nid)->node_zones[zone_type];
3117 if (populated_zone(z)) {
3118 if (zone_type < ZONE_NORMAL)
3119 low_kmem_size += z->present_pages;
3120 total_size += z->present_pages;
3123 if (low_kmem_size &&
3124 total_size > average_size && /* ignore small node */
3125 low_kmem_size > total_size * 70/100)
3126 return ZONELIST_ORDER_NODE;
3128 return ZONELIST_ORDER_ZONE;
3131 static void set_zonelist_order(void)
3133 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3134 current_zonelist_order = default_zonelist_order();
3136 current_zonelist_order = user_zonelist_order;
3139 static void build_zonelists(pg_data_t *pgdat)
3143 nodemask_t used_mask;
3144 int local_node, prev_node;
3145 struct zonelist *zonelist;
3146 int order = current_zonelist_order;
3148 /* initialize zonelists */
3149 for (i = 0; i < MAX_ZONELISTS; i++) {
3150 zonelist = pgdat->node_zonelists + i;
3151 zonelist->_zonerefs[0].zone = NULL;
3152 zonelist->_zonerefs[0].zone_idx = 0;
3155 /* NUMA-aware ordering of nodes */
3156 local_node = pgdat->node_id;
3157 load = nr_online_nodes;
3158 prev_node = local_node;
3159 nodes_clear(used_mask);
3161 memset(node_order, 0, sizeof(node_order));
3164 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3165 int distance = node_distance(local_node, node);
3168 * If another node is sufficiently far away then it is better
3169 * to reclaim pages in a zone before going off node.
3171 if (distance > RECLAIM_DISTANCE)
3172 zone_reclaim_mode = 1;
3175 * We don't want to pressure a particular node.
3176 * So adding penalty to the first node in same
3177 * distance group to make it round-robin.
3179 if (distance != node_distance(local_node, prev_node))
3180 node_load[node] = load;
3184 if (order == ZONELIST_ORDER_NODE)
3185 build_zonelists_in_node_order(pgdat, node);
3187 node_order[j++] = node; /* remember order */
3190 if (order == ZONELIST_ORDER_ZONE) {
3191 /* calculate node order -- i.e., DMA last! */
3192 build_zonelists_in_zone_order(pgdat, j);
3195 build_thisnode_zonelists(pgdat);
3198 /* Construct the zonelist performance cache - see further mmzone.h */
3199 static void build_zonelist_cache(pg_data_t *pgdat)
3201 struct zonelist *zonelist;
3202 struct zonelist_cache *zlc;
3205 zonelist = &pgdat->node_zonelists[0];
3206 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3207 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3208 for (z = zonelist->_zonerefs; z->zone; z++)
3209 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3212 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3214 * Return node id of node used for "local" allocations.
3215 * I.e., first node id of first zone in arg node's generic zonelist.
3216 * Used for initializing percpu 'numa_mem', which is used primarily
3217 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3219 int local_memory_node(int node)
3223 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3224 gfp_zone(GFP_KERNEL),
3231 #else /* CONFIG_NUMA */
3233 static void set_zonelist_order(void)
3235 current_zonelist_order = ZONELIST_ORDER_ZONE;
3238 static void build_zonelists(pg_data_t *pgdat)
3240 int node, local_node;
3242 struct zonelist *zonelist;
3244 local_node = pgdat->node_id;
3246 zonelist = &pgdat->node_zonelists[0];
3247 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3250 * Now we build the zonelist so that it contains the zones
3251 * of all the other nodes.
3252 * We don't want to pressure a particular node, so when
3253 * building the zones for node N, we make sure that the
3254 * zones coming right after the local ones are those from
3255 * node N+1 (modulo N)
3257 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3258 if (!node_online(node))
3260 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3263 for (node = 0; node < local_node; node++) {
3264 if (!node_online(node))
3266 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3270 zonelist->_zonerefs[j].zone = NULL;
3271 zonelist->_zonerefs[j].zone_idx = 0;
3274 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3275 static void build_zonelist_cache(pg_data_t *pgdat)
3277 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3280 #endif /* CONFIG_NUMA */
3283 * Boot pageset table. One per cpu which is going to be used for all
3284 * zones and all nodes. The parameters will be set in such a way
3285 * that an item put on a list will immediately be handed over to
3286 * the buddy list. This is safe since pageset manipulation is done
3287 * with interrupts disabled.
3289 * The boot_pagesets must be kept even after bootup is complete for
3290 * unused processors and/or zones. They do play a role for bootstrapping
3291 * hotplugged processors.
3293 * zoneinfo_show() and maybe other functions do
3294 * not check if the processor is online before following the pageset pointer.
3295 * Other parts of the kernel may not check if the zone is available.
3297 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3298 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3299 static void setup_zone_pageset(struct zone *zone);
3302 * Global mutex to protect against size modification of zonelists
3303 * as well as to serialize pageset setup for the new populated zone.
3305 DEFINE_MUTEX(zonelists_mutex);
3307 /* return values int ....just for stop_machine() */
3308 static __init_refok int __build_all_zonelists(void *data)
3314 memset(node_load, 0, sizeof(node_load));
3316 for_each_online_node(nid) {
3317 pg_data_t *pgdat = NODE_DATA(nid);
3319 build_zonelists(pgdat);
3320 build_zonelist_cache(pgdat);
3324 * Initialize the boot_pagesets that are going to be used
3325 * for bootstrapping processors. The real pagesets for
3326 * each zone will be allocated later when the per cpu
3327 * allocator is available.
3329 * boot_pagesets are used also for bootstrapping offline
3330 * cpus if the system is already booted because the pagesets
3331 * are needed to initialize allocators on a specific cpu too.
3332 * F.e. the percpu allocator needs the page allocator which
3333 * needs the percpu allocator in order to allocate its pagesets
3334 * (a chicken-egg dilemma).
3336 for_each_possible_cpu(cpu) {
3337 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3339 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3341 * We now know the "local memory node" for each node--
3342 * i.e., the node of the first zone in the generic zonelist.
3343 * Set up numa_mem percpu variable for on-line cpus. During
3344 * boot, only the boot cpu should be on-line; we'll init the
3345 * secondary cpus' numa_mem as they come on-line. During
3346 * node/memory hotplug, we'll fixup all on-line cpus.
3348 if (cpu_online(cpu))
3349 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3357 * Called with zonelists_mutex held always
3358 * unless system_state == SYSTEM_BOOTING.
3360 void __ref build_all_zonelists(void *data)
3362 set_zonelist_order();
3364 if (system_state == SYSTEM_BOOTING) {
3365 __build_all_zonelists(NULL);
3366 mminit_verify_zonelist();
3367 cpuset_init_current_mems_allowed();
3369 /* we have to stop all cpus to guarantee there is no user
3371 #ifdef CONFIG_MEMORY_HOTPLUG
3373 setup_zone_pageset((struct zone *)data);
3375 stop_machine(__build_all_zonelists, NULL, NULL);
3376 /* cpuset refresh routine should be here */
3378 vm_total_pages = nr_free_pagecache_pages();
3380 * Disable grouping by mobility if the number of pages in the
3381 * system is too low to allow the mechanism to work. It would be
3382 * more accurate, but expensive to check per-zone. This check is
3383 * made on memory-hotadd so a system can start with mobility
3384 * disabled and enable it later
3386 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3387 page_group_by_mobility_disabled = 1;
3389 page_group_by_mobility_disabled = 0;
3391 printk("Built %i zonelists in %s order, mobility grouping %s. "
3392 "Total pages: %ld\n",
3394 zonelist_order_name[current_zonelist_order],
3395 page_group_by_mobility_disabled ? "off" : "on",
3398 printk("Policy zone: %s\n", zone_names[policy_zone]);
3403 * Helper functions to size the waitqueue hash table.
3404 * Essentially these want to choose hash table sizes sufficiently
3405 * large so that collisions trying to wait on pages are rare.
3406 * But in fact, the number of active page waitqueues on typical
3407 * systems is ridiculously low, less than 200. So this is even
3408 * conservative, even though it seems large.
3410 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3411 * waitqueues, i.e. the size of the waitq table given the number of pages.
3413 #define PAGES_PER_WAITQUEUE 256
3415 #ifndef CONFIG_MEMORY_HOTPLUG
3416 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3418 unsigned long size = 1;
3420 pages /= PAGES_PER_WAITQUEUE;
3422 while (size < pages)
3426 * Once we have dozens or even hundreds of threads sleeping
3427 * on IO we've got bigger problems than wait queue collision.
3428 * Limit the size of the wait table to a reasonable size.
3430 size = min(size, 4096UL);
3432 return max(size, 4UL);
3436 * A zone's size might be changed by hot-add, so it is not possible to determine
3437 * a suitable size for its wait_table. So we use the maximum size now.
3439 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3441 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3442 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3443 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3445 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3446 * or more by the traditional way. (See above). It equals:
3448 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3449 * ia64(16K page size) : = ( 8G + 4M)byte.
3450 * powerpc (64K page size) : = (32G +16M)byte.
3452 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3459 * This is an integer logarithm so that shifts can be used later
3460 * to extract the more random high bits from the multiplicative
3461 * hash function before the remainder is taken.
3463 static inline unsigned long wait_table_bits(unsigned long size)
3468 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3471 * Check if a pageblock contains reserved pages
3473 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3477 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3478 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3485 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3486 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3487 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3488 * higher will lead to a bigger reserve which will get freed as contiguous
3489 * blocks as reclaim kicks in
3491 static void setup_zone_migrate_reserve(struct zone *zone)
3493 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3495 unsigned long block_migratetype;
3499 * Get the start pfn, end pfn and the number of blocks to reserve
3500 * We have to be careful to be aligned to pageblock_nr_pages to
3501 * make sure that we always check pfn_valid for the first page in
3504 start_pfn = zone->zone_start_pfn;
3505 end_pfn = start_pfn + zone->spanned_pages;
3506 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3507 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3511 * Reserve blocks are generally in place to help high-order atomic
3512 * allocations that are short-lived. A min_free_kbytes value that
3513 * would result in more than 2 reserve blocks for atomic allocations
3514 * is assumed to be in place to help anti-fragmentation for the
3515 * future allocation of hugepages at runtime.
3517 reserve = min(2, reserve);
3519 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3520 if (!pfn_valid(pfn))
3522 page = pfn_to_page(pfn);
3524 /* Watch out for overlapping nodes */
3525 if (page_to_nid(page) != zone_to_nid(zone))
3528 block_migratetype = get_pageblock_migratetype(page);
3530 /* Only test what is necessary when the reserves are not met */
3533 * Blocks with reserved pages will never free, skip
3536 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3537 if (pageblock_is_reserved(pfn, block_end_pfn))
3540 /* If this block is reserved, account for it */
3541 if (block_migratetype == MIGRATE_RESERVE) {
3546 /* Suitable for reserving if this block is movable */
3547 if (block_migratetype == MIGRATE_MOVABLE) {
3548 set_pageblock_migratetype(page,
3550 move_freepages_block(zone, page,
3558 * If the reserve is met and this is a previous reserved block,
3561 if (block_migratetype == MIGRATE_RESERVE) {
3562 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3563 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3569 * Initially all pages are reserved - free ones are freed
3570 * up by free_all_bootmem() once the early boot process is
3571 * done. Non-atomic initialization, single-pass.
3573 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3574 unsigned long start_pfn, enum memmap_context context)
3577 unsigned long end_pfn = start_pfn + size;
3581 if (highest_memmap_pfn < end_pfn - 1)
3582 highest_memmap_pfn = end_pfn - 1;
3584 z = &NODE_DATA(nid)->node_zones[zone];
3585 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3587 * There can be holes in boot-time mem_map[]s
3588 * handed to this function. They do not
3589 * exist on hotplugged memory.
3591 if (context == MEMMAP_EARLY) {
3592 if (!early_pfn_valid(pfn))
3594 if (!early_pfn_in_nid(pfn, nid))
3597 page = pfn_to_page(pfn);
3598 set_page_links(page, zone, nid, pfn);
3599 mminit_verify_page_links(page, zone, nid, pfn);
3600 init_page_count(page);
3601 reset_page_mapcount(page);
3602 SetPageReserved(page);
3604 * Mark the block movable so that blocks are reserved for
3605 * movable at startup. This will force kernel allocations
3606 * to reserve their blocks rather than leaking throughout
3607 * the address space during boot when many long-lived
3608 * kernel allocations are made. Later some blocks near
3609 * the start are marked MIGRATE_RESERVE by
3610 * setup_zone_migrate_reserve()
3612 * bitmap is created for zone's valid pfn range. but memmap
3613 * can be created for invalid pages (for alignment)
3614 * check here not to call set_pageblock_migratetype() against
3617 if ((z->zone_start_pfn <= pfn)
3618 && (pfn < z->zone_start_pfn + z->spanned_pages)
3619 && !(pfn & (pageblock_nr_pages - 1)))
3620 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3622 INIT_LIST_HEAD(&page->lru);
3623 #ifdef WANT_PAGE_VIRTUAL
3624 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3625 if (!is_highmem_idx(zone))
3626 set_page_address(page, __va(pfn << PAGE_SHIFT));
3631 static void __meminit zone_init_free_lists(struct zone *zone)
3634 for_each_migratetype_order(order, t) {
3635 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3636 zone->free_area[order].nr_free = 0;
3640 #ifndef __HAVE_ARCH_MEMMAP_INIT
3641 #define memmap_init(size, nid, zone, start_pfn) \
3642 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3645 static int zone_batchsize(struct zone *zone)
3651 * The per-cpu-pages pools are set to around 1000th of the
3652 * size of the zone. But no more than 1/2 of a meg.
3654 * OK, so we don't know how big the cache is. So guess.
3656 batch = zone->present_pages / 1024;
3657 if (batch * PAGE_SIZE > 512 * 1024)
3658 batch = (512 * 1024) / PAGE_SIZE;
3659 batch /= 4; /* We effectively *= 4 below */
3664 * Clamp the batch to a 2^n - 1 value. Having a power
3665 * of 2 value was found to be more likely to have
3666 * suboptimal cache aliasing properties in some cases.
3668 * For example if 2 tasks are alternately allocating
3669 * batches of pages, one task can end up with a lot
3670 * of pages of one half of the possible page colors
3671 * and the other with pages of the other colors.
3673 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3678 /* The deferral and batching of frees should be suppressed under NOMMU
3681 * The problem is that NOMMU needs to be able to allocate large chunks
3682 * of contiguous memory as there's no hardware page translation to
3683 * assemble apparent contiguous memory from discontiguous pages.
3685 * Queueing large contiguous runs of pages for batching, however,
3686 * causes the pages to actually be freed in smaller chunks. As there
3687 * can be a significant delay between the individual batches being
3688 * recycled, this leads to the once large chunks of space being
3689 * fragmented and becoming unavailable for high-order allocations.
3695 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3697 struct per_cpu_pages *pcp;
3700 memset(p, 0, sizeof(*p));
3704 pcp->high = 6 * batch;
3705 pcp->batch = max(1UL, 1 * batch);
3706 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3707 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3711 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3712 * to the value high for the pageset p.
3715 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3718 struct per_cpu_pages *pcp;
3722 pcp->batch = max(1UL, high/4);
3723 if ((high/4) > (PAGE_SHIFT * 8))
3724 pcp->batch = PAGE_SHIFT * 8;
3727 static void setup_zone_pageset(struct zone *zone)
3731 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3733 for_each_possible_cpu(cpu) {
3734 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3736 setup_pageset(pcp, zone_batchsize(zone));
3738 if (percpu_pagelist_fraction)
3739 setup_pagelist_highmark(pcp,
3740 (zone->present_pages /
3741 percpu_pagelist_fraction));
3746 * Allocate per cpu pagesets and initialize them.
3747 * Before this call only boot pagesets were available.
3749 void __init setup_per_cpu_pageset(void)
3753 for_each_populated_zone(zone)
3754 setup_zone_pageset(zone);
3757 static noinline __init_refok
3758 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3761 struct pglist_data *pgdat = zone->zone_pgdat;
3765 * The per-page waitqueue mechanism uses hashed waitqueues
3768 zone->wait_table_hash_nr_entries =
3769 wait_table_hash_nr_entries(zone_size_pages);
3770 zone->wait_table_bits =
3771 wait_table_bits(zone->wait_table_hash_nr_entries);
3772 alloc_size = zone->wait_table_hash_nr_entries
3773 * sizeof(wait_queue_head_t);
3775 if (!slab_is_available()) {
3776 zone->wait_table = (wait_queue_head_t *)
3777 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3780 * This case means that a zone whose size was 0 gets new memory
3781 * via memory hot-add.
3782 * But it may be the case that a new node was hot-added. In
3783 * this case vmalloc() will not be able to use this new node's
3784 * memory - this wait_table must be initialized to use this new
3785 * node itself as well.
3786 * To use this new node's memory, further consideration will be
3789 zone->wait_table = vmalloc(alloc_size);
3791 if (!zone->wait_table)
3794 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3795 init_waitqueue_head(zone->wait_table + i);
3800 static int __zone_pcp_update(void *data)
3802 struct zone *zone = data;
3804 unsigned long batch = zone_batchsize(zone), flags;
3806 for_each_possible_cpu(cpu) {
3807 struct per_cpu_pageset *pset;
3808 struct per_cpu_pages *pcp;
3810 pset = per_cpu_ptr(zone->pageset, cpu);
3813 local_irq_save(flags);
3814 free_pcppages_bulk(zone, pcp->count, pcp);
3815 setup_pageset(pset, batch);
3816 local_irq_restore(flags);
3821 void zone_pcp_update(struct zone *zone)
3823 stop_machine(__zone_pcp_update, zone, NULL);
3826 static __meminit void zone_pcp_init(struct zone *zone)
3829 * per cpu subsystem is not up at this point. The following code
3830 * relies on the ability of the linker to provide the
3831 * offset of a (static) per cpu variable into the per cpu area.
3833 zone->pageset = &boot_pageset;
3835 if (zone->present_pages)
3836 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3837 zone->name, zone->present_pages,
3838 zone_batchsize(zone));
3841 __meminit int init_currently_empty_zone(struct zone *zone,
3842 unsigned long zone_start_pfn,
3844 enum memmap_context context)
3846 struct pglist_data *pgdat = zone->zone_pgdat;
3848 ret = zone_wait_table_init(zone, size);
3851 pgdat->nr_zones = zone_idx(zone) + 1;
3853 zone->zone_start_pfn = zone_start_pfn;
3855 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3856 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3858 (unsigned long)zone_idx(zone),
3859 zone_start_pfn, (zone_start_pfn + size));
3861 zone_init_free_lists(zone);
3866 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3868 * Basic iterator support. Return the first range of PFNs for a node
3869 * Note: nid == MAX_NUMNODES returns first region regardless of node
3871 static int __meminit first_active_region_index_in_nid(int nid)
3875 for (i = 0; i < nr_nodemap_entries; i++)
3876 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3883 * Basic iterator support. Return the next active range of PFNs for a node
3884 * Note: nid == MAX_NUMNODES returns next region regardless of node
3886 static int __meminit next_active_region_index_in_nid(int index, int nid)
3888 for (index = index + 1; index < nr_nodemap_entries; index++)
3889 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3895 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3897 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3898 * Architectures may implement their own version but if add_active_range()
3899 * was used and there are no special requirements, this is a convenient
3902 int __meminit __early_pfn_to_nid(unsigned long pfn)
3906 for (i = 0; i < nr_nodemap_entries; i++) {
3907 unsigned long start_pfn = early_node_map[i].start_pfn;
3908 unsigned long end_pfn = early_node_map[i].end_pfn;
3910 if (start_pfn <= pfn && pfn < end_pfn)
3911 return early_node_map[i].nid;
3913 /* This is a memory hole */
3916 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3918 int __meminit early_pfn_to_nid(unsigned long pfn)
3922 nid = __early_pfn_to_nid(pfn);
3925 /* just returns 0 */
3929 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3930 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3934 nid = __early_pfn_to_nid(pfn);
3935 if (nid >= 0 && nid != node)
3941 /* Basic iterator support to walk early_node_map[] */
3942 #define for_each_active_range_index_in_nid(i, nid) \
3943 for (i = first_active_region_index_in_nid(nid); i != -1; \
3944 i = next_active_region_index_in_nid(i, nid))
3947 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3948 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3949 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3951 * If an architecture guarantees that all ranges registered with
3952 * add_active_ranges() contain no holes and may be freed, this
3953 * this function may be used instead of calling free_bootmem() manually.
3955 void __init free_bootmem_with_active_regions(int nid,
3956 unsigned long max_low_pfn)
3960 for_each_active_range_index_in_nid(i, nid) {
3961 unsigned long size_pages = 0;
3962 unsigned long end_pfn = early_node_map[i].end_pfn;
3964 if (early_node_map[i].start_pfn >= max_low_pfn)
3967 if (end_pfn > max_low_pfn)
3968 end_pfn = max_low_pfn;
3970 size_pages = end_pfn - early_node_map[i].start_pfn;
3971 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3972 PFN_PHYS(early_node_map[i].start_pfn),
3973 size_pages << PAGE_SHIFT);
3977 #ifdef CONFIG_HAVE_MEMBLOCK
3979 * Basic iterator support. Return the last range of PFNs for a node
3980 * Note: nid == MAX_NUMNODES returns last region regardless of node
3982 static int __meminit last_active_region_index_in_nid(int nid)
3986 for (i = nr_nodemap_entries - 1; i >= 0; i--)
3987 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3994 * Basic iterator support. Return the previous active range of PFNs for a node
3995 * Note: nid == MAX_NUMNODES returns next region regardless of node
3997 static int __meminit previous_active_region_index_in_nid(int index, int nid)
3999 for (index = index - 1; index >= 0; index--)
4000 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
4006 #define for_each_active_range_index_in_nid_reverse(i, nid) \
4007 for (i = last_active_region_index_in_nid(nid); i != -1; \
4008 i = previous_active_region_index_in_nid(i, nid))
4010 u64 __init find_memory_core_early(int nid, u64 size, u64 align,
4011 u64 goal, u64 limit)
4015 /* Need to go over early_node_map to find out good range for node */
4016 for_each_active_range_index_in_nid_reverse(i, nid) {
4018 u64 ei_start, ei_last;
4019 u64 final_start, final_end;
4021 ei_last = early_node_map[i].end_pfn;
4022 ei_last <<= PAGE_SHIFT;
4023 ei_start = early_node_map[i].start_pfn;
4024 ei_start <<= PAGE_SHIFT;
4026 final_start = max(ei_start, goal);
4027 final_end = min(ei_last, limit);
4029 if (final_start >= final_end)
4032 addr = memblock_find_in_range(final_start, final_end, size, align);
4034 if (addr == MEMBLOCK_ERROR)
4040 return MEMBLOCK_ERROR;
4044 int __init add_from_early_node_map(struct range *range, int az,
4045 int nr_range, int nid)
4050 /* need to go over early_node_map to find out good range for node */
4051 for_each_active_range_index_in_nid(i, nid) {
4052 start = early_node_map[i].start_pfn;
4053 end = early_node_map[i].end_pfn;
4054 nr_range = add_range(range, az, nr_range, start, end);
4059 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
4064 for_each_active_range_index_in_nid(i, nid) {
4065 ret = work_fn(early_node_map[i].start_pfn,
4066 early_node_map[i].end_pfn, data);
4072 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4073 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4075 * If an architecture guarantees that all ranges registered with
4076 * add_active_ranges() contain no holes and may be freed, this
4077 * function may be used instead of calling memory_present() manually.
4079 void __init sparse_memory_present_with_active_regions(int nid)
4083 for_each_active_range_index_in_nid(i, nid)
4084 memory_present(early_node_map[i].nid,
4085 early_node_map[i].start_pfn,
4086 early_node_map[i].end_pfn);
4090 * get_pfn_range_for_nid - Return the start and end page frames for a node
4091 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4092 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4093 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4095 * It returns the start and end page frame of a node based on information
4096 * provided by an arch calling add_active_range(). If called for a node
4097 * with no available memory, a warning is printed and the start and end
4100 void __meminit get_pfn_range_for_nid(unsigned int nid,
4101 unsigned long *start_pfn, unsigned long *end_pfn)
4107 for_each_active_range_index_in_nid(i, nid) {
4108 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
4109 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
4112 if (*start_pfn == -1UL)
4117 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4118 * assumption is made that zones within a node are ordered in monotonic
4119 * increasing memory addresses so that the "highest" populated zone is used
4121 static void __init find_usable_zone_for_movable(void)
4124 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4125 if (zone_index == ZONE_MOVABLE)
4128 if (arch_zone_highest_possible_pfn[zone_index] >
4129 arch_zone_lowest_possible_pfn[zone_index])
4133 VM_BUG_ON(zone_index == -1);
4134 movable_zone = zone_index;
4138 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4139 * because it is sized independent of architecture. Unlike the other zones,
4140 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4141 * in each node depending on the size of each node and how evenly kernelcore
4142 * is distributed. This helper function adjusts the zone ranges
4143 * provided by the architecture for a given node by using the end of the
4144 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4145 * zones within a node are in order of monotonic increases memory addresses
4147 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4148 unsigned long zone_type,
4149 unsigned long node_start_pfn,
4150 unsigned long node_end_pfn,
4151 unsigned long *zone_start_pfn,
4152 unsigned long *zone_end_pfn)
4154 /* Only adjust if ZONE_MOVABLE is on this node */
4155 if (zone_movable_pfn[nid]) {
4156 /* Size ZONE_MOVABLE */
4157 if (zone_type == ZONE_MOVABLE) {
4158 *zone_start_pfn = zone_movable_pfn[nid];
4159 *zone_end_pfn = min(node_end_pfn,
4160 arch_zone_highest_possible_pfn[movable_zone]);
4162 /* Adjust for ZONE_MOVABLE starting within this range */
4163 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4164 *zone_end_pfn > zone_movable_pfn[nid]) {
4165 *zone_end_pfn = zone_movable_pfn[nid];
4167 /* Check if this whole range is within ZONE_MOVABLE */
4168 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4169 *zone_start_pfn = *zone_end_pfn;
4174 * Return the number of pages a zone spans in a node, including holes
4175 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4177 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4178 unsigned long zone_type,
4179 unsigned long *ignored)
4181 unsigned long node_start_pfn, node_end_pfn;
4182 unsigned long zone_start_pfn, zone_end_pfn;
4184 /* Get the start and end of the node and zone */
4185 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4186 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4187 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4188 adjust_zone_range_for_zone_movable(nid, zone_type,
4189 node_start_pfn, node_end_pfn,
4190 &zone_start_pfn, &zone_end_pfn);
4192 /* Check that this node has pages within the zone's required range */
4193 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4196 /* Move the zone boundaries inside the node if necessary */
4197 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4198 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4200 /* Return the spanned pages */
4201 return zone_end_pfn - zone_start_pfn;
4205 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4206 * then all holes in the requested range will be accounted for.
4208 unsigned long __meminit __absent_pages_in_range(int nid,
4209 unsigned long range_start_pfn,
4210 unsigned long range_end_pfn)
4213 unsigned long prev_end_pfn = 0, hole_pages = 0;
4214 unsigned long start_pfn;
4216 /* Find the end_pfn of the first active range of pfns in the node */
4217 i = first_active_region_index_in_nid(nid);
4221 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4223 /* Account for ranges before physical memory on this node */
4224 if (early_node_map[i].start_pfn > range_start_pfn)
4225 hole_pages = prev_end_pfn - range_start_pfn;
4227 /* Find all holes for the zone within the node */
4228 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4230 /* No need to continue if prev_end_pfn is outside the zone */
4231 if (prev_end_pfn >= range_end_pfn)
4234 /* Make sure the end of the zone is not within the hole */
4235 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4236 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4238 /* Update the hole size cound and move on */
4239 if (start_pfn > range_start_pfn) {
4240 BUG_ON(prev_end_pfn > start_pfn);
4241 hole_pages += start_pfn - prev_end_pfn;
4243 prev_end_pfn = early_node_map[i].end_pfn;
4246 /* Account for ranges past physical memory on this node */
4247 if (range_end_pfn > prev_end_pfn)
4248 hole_pages += range_end_pfn -
4249 max(range_start_pfn, prev_end_pfn);
4255 * absent_pages_in_range - Return number of page frames in holes within a range
4256 * @start_pfn: The start PFN to start searching for holes
4257 * @end_pfn: The end PFN to stop searching for holes
4259 * It returns the number of pages frames in memory holes within a range.
4261 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4262 unsigned long end_pfn)
4264 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4267 /* Return the number of page frames in holes in a zone on a node */
4268 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4269 unsigned long zone_type,
4270 unsigned long *ignored)
4272 unsigned long node_start_pfn, node_end_pfn;
4273 unsigned long zone_start_pfn, zone_end_pfn;
4275 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4276 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4278 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4281 adjust_zone_range_for_zone_movable(nid, zone_type,
4282 node_start_pfn, node_end_pfn,
4283 &zone_start_pfn, &zone_end_pfn);
4284 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4288 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4289 unsigned long zone_type,
4290 unsigned long *zones_size)
4292 return zones_size[zone_type];
4295 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4296 unsigned long zone_type,
4297 unsigned long *zholes_size)
4302 return zholes_size[zone_type];
4307 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4308 unsigned long *zones_size, unsigned long *zholes_size)
4310 unsigned long realtotalpages, totalpages = 0;
4313 for (i = 0; i < MAX_NR_ZONES; i++)
4314 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4316 pgdat->node_spanned_pages = totalpages;
4318 realtotalpages = totalpages;
4319 for (i = 0; i < MAX_NR_ZONES; i++)
4321 zone_absent_pages_in_node(pgdat->node_id, i,
4323 pgdat->node_present_pages = realtotalpages;
4324 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4328 #ifndef CONFIG_SPARSEMEM
4330 * Calculate the size of the zone->blockflags rounded to an unsigned long
4331 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4332 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4333 * round what is now in bits to nearest long in bits, then return it in
4336 static unsigned long __init usemap_size(unsigned long zonesize)
4338 unsigned long usemapsize;
4340 usemapsize = roundup(zonesize, pageblock_nr_pages);
4341 usemapsize = usemapsize >> pageblock_order;
4342 usemapsize *= NR_PAGEBLOCK_BITS;
4343 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4345 return usemapsize / 8;
4348 static void __init setup_usemap(struct pglist_data *pgdat,
4349 struct zone *zone, unsigned long zonesize)
4351 unsigned long usemapsize = usemap_size(zonesize);
4352 zone->pageblock_flags = NULL;
4354 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4358 static inline void setup_usemap(struct pglist_data *pgdat,
4359 struct zone *zone, unsigned long zonesize) {}
4360 #endif /* CONFIG_SPARSEMEM */
4362 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4364 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4365 void __init set_pageblock_order(void)
4369 /* Check that pageblock_nr_pages has not already been setup */
4370 if (pageblock_order)
4373 if (HPAGE_SHIFT > PAGE_SHIFT)
4374 order = HUGETLB_PAGE_ORDER;
4376 order = MAX_ORDER - 1;
4379 * Assume the largest contiguous order of interest is a huge page.
4380 * This value may be variable depending on boot parameters on IA64 and
4383 pageblock_order = order;
4385 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4388 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4389 * is unused as pageblock_order is set at compile-time. See
4390 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4393 void __init set_pageblock_order(void)
4397 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4400 * Set up the zone data structures:
4401 * - mark all pages reserved
4402 * - mark all memory queues empty
4403 * - clear the memory bitmaps
4405 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4406 unsigned long *zones_size, unsigned long *zholes_size)
4409 int nid = pgdat->node_id;
4410 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4413 pgdat_resize_init(pgdat);
4414 pgdat->nr_zones = 0;
4415 init_waitqueue_head(&pgdat->kswapd_wait);
4416 pgdat->kswapd_max_order = 0;
4417 pgdat_page_cgroup_init(pgdat);
4419 for (j = 0; j < MAX_NR_ZONES; j++) {
4420 struct zone *zone = pgdat->node_zones + j;
4421 unsigned long size, realsize, memmap_pages;
4424 size = zone_spanned_pages_in_node(nid, j, zones_size);
4425 realsize = size - zone_absent_pages_in_node(nid, j,
4429 * Adjust realsize so that it accounts for how much memory
4430 * is used by this zone for memmap. This affects the watermark
4431 * and per-cpu initialisations
4434 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4435 if (realsize >= memmap_pages) {
4436 realsize -= memmap_pages;
4439 " %s zone: %lu pages used for memmap\n",
4440 zone_names[j], memmap_pages);
4443 " %s zone: %lu pages exceeds realsize %lu\n",
4444 zone_names[j], memmap_pages, realsize);
4446 /* Account for reserved pages */
4447 if (j == 0 && realsize > dma_reserve) {
4448 realsize -= dma_reserve;
4449 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4450 zone_names[0], dma_reserve);
4453 if (!is_highmem_idx(j))
4454 nr_kernel_pages += realsize;
4455 nr_all_pages += realsize;
4457 zone->spanned_pages = size;
4458 zone->present_pages = realsize;
4461 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4463 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4465 zone->name = zone_names[j];
4466 spin_lock_init(&zone->lock);
4467 spin_lock_init(&zone->lru_lock);
4468 zone_seqlock_init(zone);
4469 zone->zone_pgdat = pgdat;
4471 zone_pcp_init(zone);
4473 INIT_LIST_HEAD(&zone->lru[l].list);
4474 zone->reclaim_stat.recent_rotated[0] = 0;
4475 zone->reclaim_stat.recent_rotated[1] = 0;
4476 zone->reclaim_stat.recent_scanned[0] = 0;
4477 zone->reclaim_stat.recent_scanned[1] = 0;
4478 zap_zone_vm_stats(zone);
4483 set_pageblock_order();
4484 setup_usemap(pgdat, zone, size);
4485 ret = init_currently_empty_zone(zone, zone_start_pfn,
4486 size, MEMMAP_EARLY);
4488 memmap_init(size, nid, j, zone_start_pfn);
4489 zone_start_pfn += size;
4493 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4495 /* Skip empty nodes */
4496 if (!pgdat->node_spanned_pages)
4499 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4500 /* ia64 gets its own node_mem_map, before this, without bootmem */
4501 if (!pgdat->node_mem_map) {
4502 unsigned long size, start, end;
4506 * The zone's endpoints aren't required to be MAX_ORDER
4507 * aligned but the node_mem_map endpoints must be in order
4508 * for the buddy allocator to function correctly.
4510 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4511 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4512 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4513 size = (end - start) * sizeof(struct page);
4514 map = alloc_remap(pgdat->node_id, size);
4516 map = alloc_bootmem_node_nopanic(pgdat, size);
4517 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4519 #ifndef CONFIG_NEED_MULTIPLE_NODES
4521 * With no DISCONTIG, the global mem_map is just set as node 0's
4523 if (pgdat == NODE_DATA(0)) {
4524 mem_map = NODE_DATA(0)->node_mem_map;
4525 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4526 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4527 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4528 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4531 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4534 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4535 unsigned long node_start_pfn, unsigned long *zholes_size)
4537 pg_data_t *pgdat = NODE_DATA(nid);
4539 pgdat->node_id = nid;
4540 pgdat->node_start_pfn = node_start_pfn;
4541 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4543 alloc_node_mem_map(pgdat);
4544 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4545 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4546 nid, (unsigned long)pgdat,
4547 (unsigned long)pgdat->node_mem_map);
4550 free_area_init_core(pgdat, zones_size, zholes_size);
4553 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4555 #if MAX_NUMNODES > 1
4557 * Figure out the number of possible node ids.
4559 static void __init setup_nr_node_ids(void)
4562 unsigned int highest = 0;
4564 for_each_node_mask(node, node_possible_map)
4566 nr_node_ids = highest + 1;
4569 static inline void setup_nr_node_ids(void)
4575 * add_active_range - Register a range of PFNs backed by physical memory
4576 * @nid: The node ID the range resides on
4577 * @start_pfn: The start PFN of the available physical memory
4578 * @end_pfn: The end PFN of the available physical memory
4580 * These ranges are stored in an early_node_map[] and later used by
4581 * free_area_init_nodes() to calculate zone sizes and holes. If the
4582 * range spans a memory hole, it is up to the architecture to ensure
4583 * the memory is not freed by the bootmem allocator. If possible
4584 * the range being registered will be merged with existing ranges.
4586 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4587 unsigned long end_pfn)
4591 mminit_dprintk(MMINIT_TRACE, "memory_register",
4592 "Entering add_active_range(%d, %#lx, %#lx) "
4593 "%d entries of %d used\n",
4594 nid, start_pfn, end_pfn,
4595 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4597 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4599 /* Merge with existing active regions if possible */
4600 for (i = 0; i < nr_nodemap_entries; i++) {
4601 if (early_node_map[i].nid != nid)
4604 /* Skip if an existing region covers this new one */
4605 if (start_pfn >= early_node_map[i].start_pfn &&
4606 end_pfn <= early_node_map[i].end_pfn)
4609 /* Merge forward if suitable */
4610 if (start_pfn <= early_node_map[i].end_pfn &&
4611 end_pfn > early_node_map[i].end_pfn) {
4612 early_node_map[i].end_pfn = end_pfn;
4616 /* Merge backward if suitable */
4617 if (start_pfn < early_node_map[i].start_pfn &&
4618 end_pfn >= early_node_map[i].start_pfn) {
4619 early_node_map[i].start_pfn = start_pfn;
4624 /* Check that early_node_map is large enough */
4625 if (i >= MAX_ACTIVE_REGIONS) {
4626 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4627 MAX_ACTIVE_REGIONS);
4631 early_node_map[i].nid = nid;
4632 early_node_map[i].start_pfn = start_pfn;
4633 early_node_map[i].end_pfn = end_pfn;
4634 nr_nodemap_entries = i + 1;
4638 * remove_active_range - Shrink an existing registered range of PFNs
4639 * @nid: The node id the range is on that should be shrunk
4640 * @start_pfn: The new PFN of the range
4641 * @end_pfn: The new PFN of the range
4643 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4644 * The map is kept near the end physical page range that has already been
4645 * registered. This function allows an arch to shrink an existing registered
4648 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4649 unsigned long end_pfn)
4654 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4655 nid, start_pfn, end_pfn);
4657 /* Find the old active region end and shrink */
4658 for_each_active_range_index_in_nid(i, nid) {
4659 if (early_node_map[i].start_pfn >= start_pfn &&
4660 early_node_map[i].end_pfn <= end_pfn) {
4662 early_node_map[i].start_pfn = 0;
4663 early_node_map[i].end_pfn = 0;
4667 if (early_node_map[i].start_pfn < start_pfn &&
4668 early_node_map[i].end_pfn > start_pfn) {
4669 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4670 early_node_map[i].end_pfn = start_pfn;
4671 if (temp_end_pfn > end_pfn)
4672 add_active_range(nid, end_pfn, temp_end_pfn);
4675 if (early_node_map[i].start_pfn >= start_pfn &&
4676 early_node_map[i].end_pfn > end_pfn &&
4677 early_node_map[i].start_pfn < end_pfn) {
4678 early_node_map[i].start_pfn = end_pfn;
4686 /* remove the blank ones */
4687 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4688 if (early_node_map[i].nid != nid)
4690 if (early_node_map[i].end_pfn)
4692 /* we found it, get rid of it */
4693 for (j = i; j < nr_nodemap_entries - 1; j++)
4694 memcpy(&early_node_map[j], &early_node_map[j+1],
4695 sizeof(early_node_map[j]));
4696 j = nr_nodemap_entries - 1;
4697 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4698 nr_nodemap_entries--;
4703 * remove_all_active_ranges - Remove all currently registered regions
4705 * During discovery, it may be found that a table like SRAT is invalid
4706 * and an alternative discovery method must be used. This function removes
4707 * all currently registered regions.
4709 void __init remove_all_active_ranges(void)
4711 memset(early_node_map, 0, sizeof(early_node_map));
4712 nr_nodemap_entries = 0;
4715 /* Compare two active node_active_regions */
4716 static int __init cmp_node_active_region(const void *a, const void *b)
4718 struct node_active_region *arange = (struct node_active_region *)a;
4719 struct node_active_region *brange = (struct node_active_region *)b;
4721 /* Done this way to avoid overflows */
4722 if (arange->start_pfn > brange->start_pfn)
4724 if (arange->start_pfn < brange->start_pfn)
4730 /* sort the node_map by start_pfn */
4731 void __init sort_node_map(void)
4733 sort(early_node_map, (size_t)nr_nodemap_entries,
4734 sizeof(struct node_active_region),
4735 cmp_node_active_region, NULL);
4739 * node_map_pfn_alignment - determine the maximum internode alignment
4741 * This function should be called after node map is populated and sorted.
4742 * It calculates the maximum power of two alignment which can distinguish
4745 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4746 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4747 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4748 * shifted, 1GiB is enough and this function will indicate so.
4750 * This is used to test whether pfn -> nid mapping of the chosen memory
4751 * model has fine enough granularity to avoid incorrect mapping for the
4752 * populated node map.
4754 * Returns the determined alignment in pfn's. 0 if there is no alignment
4755 * requirement (single node).
4757 unsigned long __init node_map_pfn_alignment(void)
4759 unsigned long accl_mask = 0, last_end = 0;
4763 for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4764 int nid = early_node_map[i].nid;
4765 unsigned long start = early_node_map[i].start_pfn;
4766 unsigned long end = early_node_map[i].end_pfn;
4769 if (!start || last_nid < 0 || last_nid == nid) {
4776 * Start with a mask granular enough to pin-point to the
4777 * start pfn and tick off bits one-by-one until it becomes
4778 * too coarse to separate the current node from the last.
4780 mask = ~((1 << __ffs(start)) - 1);
4781 while (mask && last_end <= (start & (mask << 1)))
4784 /* accumulate all internode masks */
4788 /* convert mask to number of pages */
4789 return ~accl_mask + 1;
4792 /* Find the lowest pfn for a node */
4793 static unsigned long __init find_min_pfn_for_node(int nid)
4796 unsigned long min_pfn = ULONG_MAX;
4798 /* Assuming a sorted map, the first range found has the starting pfn */
4799 for_each_active_range_index_in_nid(i, nid)
4800 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4802 if (min_pfn == ULONG_MAX) {
4804 "Could not find start_pfn for node %d\n", nid);
4812 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4814 * It returns the minimum PFN based on information provided via
4815 * add_active_range().
4817 unsigned long __init find_min_pfn_with_active_regions(void)
4819 return find_min_pfn_for_node(MAX_NUMNODES);
4823 * early_calculate_totalpages()
4824 * Sum pages in active regions for movable zone.
4825 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4827 static unsigned long __init early_calculate_totalpages(void)
4830 unsigned long totalpages = 0;
4832 for (i = 0; i < nr_nodemap_entries; i++) {
4833 unsigned long pages = early_node_map[i].end_pfn -
4834 early_node_map[i].start_pfn;
4835 totalpages += pages;
4837 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4843 * Find the PFN the Movable zone begins in each node. Kernel memory
4844 * is spread evenly between nodes as long as the nodes have enough
4845 * memory. When they don't, some nodes will have more kernelcore than
4848 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4851 unsigned long usable_startpfn;
4852 unsigned long kernelcore_node, kernelcore_remaining;
4853 /* save the state before borrow the nodemask */
4854 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4855 unsigned long totalpages = early_calculate_totalpages();
4856 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4859 * If movablecore was specified, calculate what size of
4860 * kernelcore that corresponds so that memory usable for
4861 * any allocation type is evenly spread. If both kernelcore
4862 * and movablecore are specified, then the value of kernelcore
4863 * will be used for required_kernelcore if it's greater than
4864 * what movablecore would have allowed.
4866 if (required_movablecore) {
4867 unsigned long corepages;
4870 * Round-up so that ZONE_MOVABLE is at least as large as what
4871 * was requested by the user
4873 required_movablecore =
4874 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4875 corepages = totalpages - required_movablecore;
4877 required_kernelcore = max(required_kernelcore, corepages);
4880 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4881 if (!required_kernelcore)
4884 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4885 find_usable_zone_for_movable();
4886 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4889 /* Spread kernelcore memory as evenly as possible throughout nodes */
4890 kernelcore_node = required_kernelcore / usable_nodes;
4891 for_each_node_state(nid, N_HIGH_MEMORY) {
4893 * Recalculate kernelcore_node if the division per node
4894 * now exceeds what is necessary to satisfy the requested
4895 * amount of memory for the kernel
4897 if (required_kernelcore < kernelcore_node)
4898 kernelcore_node = required_kernelcore / usable_nodes;
4901 * As the map is walked, we track how much memory is usable
4902 * by the kernel using kernelcore_remaining. When it is
4903 * 0, the rest of the node is usable by ZONE_MOVABLE
4905 kernelcore_remaining = kernelcore_node;
4907 /* Go through each range of PFNs within this node */
4908 for_each_active_range_index_in_nid(i, nid) {
4909 unsigned long start_pfn, end_pfn;
4910 unsigned long size_pages;
4912 start_pfn = max(early_node_map[i].start_pfn,
4913 zone_movable_pfn[nid]);
4914 end_pfn = early_node_map[i].end_pfn;
4915 if (start_pfn >= end_pfn)
4918 /* Account for what is only usable for kernelcore */
4919 if (start_pfn < usable_startpfn) {
4920 unsigned long kernel_pages;
4921 kernel_pages = min(end_pfn, usable_startpfn)
4924 kernelcore_remaining -= min(kernel_pages,
4925 kernelcore_remaining);
4926 required_kernelcore -= min(kernel_pages,
4927 required_kernelcore);
4929 /* Continue if range is now fully accounted */
4930 if (end_pfn <= usable_startpfn) {
4933 * Push zone_movable_pfn to the end so
4934 * that if we have to rebalance
4935 * kernelcore across nodes, we will
4936 * not double account here
4938 zone_movable_pfn[nid] = end_pfn;
4941 start_pfn = usable_startpfn;
4945 * The usable PFN range for ZONE_MOVABLE is from
4946 * start_pfn->end_pfn. Calculate size_pages as the
4947 * number of pages used as kernelcore
4949 size_pages = end_pfn - start_pfn;
4950 if (size_pages > kernelcore_remaining)
4951 size_pages = kernelcore_remaining;
4952 zone_movable_pfn[nid] = start_pfn + size_pages;
4955 * Some kernelcore has been met, update counts and
4956 * break if the kernelcore for this node has been
4959 required_kernelcore -= min(required_kernelcore,
4961 kernelcore_remaining -= size_pages;
4962 if (!kernelcore_remaining)
4968 * If there is still required_kernelcore, we do another pass with one
4969 * less node in the count. This will push zone_movable_pfn[nid] further
4970 * along on the nodes that still have memory until kernelcore is
4974 if (usable_nodes && required_kernelcore > usable_nodes)
4977 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4978 for (nid = 0; nid < MAX_NUMNODES; nid++)
4979 zone_movable_pfn[nid] =
4980 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4983 /* restore the node_state */
4984 node_states[N_HIGH_MEMORY] = saved_node_state;
4987 /* Any regular memory on that node ? */
4988 static void check_for_regular_memory(pg_data_t *pgdat)
4990 #ifdef CONFIG_HIGHMEM
4991 enum zone_type zone_type;
4993 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4994 struct zone *zone = &pgdat->node_zones[zone_type];
4995 if (zone->present_pages)
4996 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
5002 * free_area_init_nodes - Initialise all pg_data_t and zone data
5003 * @max_zone_pfn: an array of max PFNs for each zone
5005 * This will call free_area_init_node() for each active node in the system.
5006 * Using the page ranges provided by add_active_range(), the size of each
5007 * zone in each node and their holes is calculated. If the maximum PFN
5008 * between two adjacent zones match, it is assumed that the zone is empty.
5009 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5010 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5011 * starts where the previous one ended. For example, ZONE_DMA32 starts
5012 * at arch_max_dma_pfn.
5014 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5019 /* Sort early_node_map as initialisation assumes it is sorted */
5022 /* Record where the zone boundaries are */
5023 memset(arch_zone_lowest_possible_pfn, 0,
5024 sizeof(arch_zone_lowest_possible_pfn));
5025 memset(arch_zone_highest_possible_pfn, 0,
5026 sizeof(arch_zone_highest_possible_pfn));
5027 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5028 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5029 for (i = 1; i < MAX_NR_ZONES; i++) {
5030 if (i == ZONE_MOVABLE)
5032 arch_zone_lowest_possible_pfn[i] =
5033 arch_zone_highest_possible_pfn[i-1];
5034 arch_zone_highest_possible_pfn[i] =
5035 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5037 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5038 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5040 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5041 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5042 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
5044 /* Print out the zone ranges */
5045 printk("Zone PFN ranges:\n");
5046 for (i = 0; i < MAX_NR_ZONES; i++) {
5047 if (i == ZONE_MOVABLE)
5049 printk(" %-8s ", zone_names[i]);
5050 if (arch_zone_lowest_possible_pfn[i] ==
5051 arch_zone_highest_possible_pfn[i])
5054 printk("%0#10lx -> %0#10lx\n",
5055 arch_zone_lowest_possible_pfn[i],
5056 arch_zone_highest_possible_pfn[i]);
5059 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5060 printk("Movable zone start PFN for each node\n");
5061 for (i = 0; i < MAX_NUMNODES; i++) {
5062 if (zone_movable_pfn[i])
5063 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
5066 /* Print out the early_node_map[] */
5067 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
5068 for (i = 0; i < nr_nodemap_entries; i++)
5069 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
5070 early_node_map[i].start_pfn,
5071 early_node_map[i].end_pfn);
5073 /* Initialise every node */
5074 mminit_verify_pageflags_layout();
5075 setup_nr_node_ids();
5076 for_each_online_node(nid) {
5077 pg_data_t *pgdat = NODE_DATA(nid);
5078 free_area_init_node(nid, NULL,
5079 find_min_pfn_for_node(nid), NULL);
5081 /* Any memory on that node */
5082 if (pgdat->node_present_pages)
5083 node_set_state(nid, N_HIGH_MEMORY);
5084 check_for_regular_memory(pgdat);
5088 static int __init cmdline_parse_core(char *p, unsigned long *core)
5090 unsigned long long coremem;
5094 coremem = memparse(p, &p);
5095 *core = coremem >> PAGE_SHIFT;
5097 /* Paranoid check that UL is enough for the coremem value */
5098 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5104 * kernelcore=size sets the amount of memory for use for allocations that
5105 * cannot be reclaimed or migrated.
5107 static int __init cmdline_parse_kernelcore(char *p)
5109 return cmdline_parse_core(p, &required_kernelcore);
5113 * movablecore=size sets the amount of memory for use for allocations that
5114 * can be reclaimed or migrated.
5116 static int __init cmdline_parse_movablecore(char *p)
5118 return cmdline_parse_core(p, &required_movablecore);
5121 early_param("kernelcore", cmdline_parse_kernelcore);
5122 early_param("movablecore", cmdline_parse_movablecore);
5124 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
5127 * set_dma_reserve - set the specified number of pages reserved in the first zone
5128 * @new_dma_reserve: The number of pages to mark reserved
5130 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5131 * In the DMA zone, a significant percentage may be consumed by kernel image
5132 * and other unfreeable allocations which can skew the watermarks badly. This
5133 * function may optionally be used to account for unfreeable pages in the
5134 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5135 * smaller per-cpu batchsize.
5137 void __init set_dma_reserve(unsigned long new_dma_reserve)
5139 dma_reserve = new_dma_reserve;
5142 void __init free_area_init(unsigned long *zones_size)
5144 free_area_init_node(0, zones_size,
5145 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5148 static int page_alloc_cpu_notify(struct notifier_block *self,
5149 unsigned long action, void *hcpu)
5151 int cpu = (unsigned long)hcpu;
5153 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5157 * Spill the event counters of the dead processor
5158 * into the current processors event counters.
5159 * This artificially elevates the count of the current
5162 vm_events_fold_cpu(cpu);
5165 * Zero the differential counters of the dead processor
5166 * so that the vm statistics are consistent.
5168 * This is only okay since the processor is dead and cannot
5169 * race with what we are doing.
5171 refresh_cpu_vm_stats(cpu);
5176 void __init page_alloc_init(void)
5178 hotcpu_notifier(page_alloc_cpu_notify, 0);
5182 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5183 * or min_free_kbytes changes.
5185 static void calculate_totalreserve_pages(void)
5187 struct pglist_data *pgdat;
5188 unsigned long reserve_pages = 0;
5189 enum zone_type i, j;
5191 for_each_online_pgdat(pgdat) {
5192 for (i = 0; i < MAX_NR_ZONES; i++) {
5193 struct zone *zone = pgdat->node_zones + i;
5194 unsigned long max = 0;
5196 /* Find valid and maximum lowmem_reserve in the zone */
5197 for (j = i; j < MAX_NR_ZONES; j++) {
5198 if (zone->lowmem_reserve[j] > max)
5199 max = zone->lowmem_reserve[j];
5202 /* we treat the high watermark as reserved pages. */
5203 max += high_wmark_pages(zone);
5205 if (max > zone->present_pages)
5206 max = zone->present_pages;
5207 reserve_pages += max;
5210 totalreserve_pages = reserve_pages;
5214 * setup_per_zone_lowmem_reserve - called whenever
5215 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5216 * has a correct pages reserved value, so an adequate number of
5217 * pages are left in the zone after a successful __alloc_pages().
5219 static void setup_per_zone_lowmem_reserve(void)
5221 struct pglist_data *pgdat;
5222 enum zone_type j, idx;
5224 for_each_online_pgdat(pgdat) {
5225 for (j = 0; j < MAX_NR_ZONES; j++) {
5226 struct zone *zone = pgdat->node_zones + j;
5227 unsigned long present_pages = zone->present_pages;
5229 zone->lowmem_reserve[j] = 0;
5233 struct zone *lower_zone;
5237 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5238 sysctl_lowmem_reserve_ratio[idx] = 1;
5240 lower_zone = pgdat->node_zones + idx;
5241 lower_zone->lowmem_reserve[j] = present_pages /
5242 sysctl_lowmem_reserve_ratio[idx];
5243 present_pages += lower_zone->present_pages;
5248 /* update totalreserve_pages */
5249 calculate_totalreserve_pages();
5253 * setup_per_zone_wmarks - called when min_free_kbytes changes
5254 * or when memory is hot-{added|removed}
5256 * Ensures that the watermark[min,low,high] values for each zone are set
5257 * correctly with respect to min_free_kbytes.
5259 void setup_per_zone_wmarks(void)
5261 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5262 unsigned long lowmem_pages = 0;
5264 unsigned long flags;
5266 /* Calculate total number of !ZONE_HIGHMEM pages */
5267 for_each_zone(zone) {
5268 if (!is_highmem(zone))
5269 lowmem_pages += zone->present_pages;
5272 for_each_zone(zone) {
5275 spin_lock_irqsave(&zone->lock, flags);
5276 tmp = (u64)pages_min * zone->present_pages;
5277 do_div(tmp, lowmem_pages);
5278 if (is_highmem(zone)) {
5280 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5281 * need highmem pages, so cap pages_min to a small
5284 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5285 * deltas controls asynch page reclaim, and so should
5286 * not be capped for highmem.
5290 min_pages = zone->present_pages / 1024;
5291 if (min_pages < SWAP_CLUSTER_MAX)
5292 min_pages = SWAP_CLUSTER_MAX;
5293 if (min_pages > 128)
5295 zone->watermark[WMARK_MIN] = min_pages;
5298 * If it's a lowmem zone, reserve a number of pages
5299 * proportionate to the zone's size.
5301 zone->watermark[WMARK_MIN] = tmp;
5304 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5305 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5306 setup_zone_migrate_reserve(zone);
5307 spin_unlock_irqrestore(&zone->lock, flags);
5310 /* update totalreserve_pages */
5311 calculate_totalreserve_pages();
5315 * The inactive anon list should be small enough that the VM never has to
5316 * do too much work, but large enough that each inactive page has a chance
5317 * to be referenced again before it is swapped out.
5319 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5320 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5321 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5322 * the anonymous pages are kept on the inactive list.
5325 * memory ratio inactive anon
5326 * -------------------------------------
5335 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5337 unsigned int gb, ratio;
5339 /* Zone size in gigabytes */
5340 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5342 ratio = int_sqrt(10 * gb);
5346 zone->inactive_ratio = ratio;
5349 static void __meminit setup_per_zone_inactive_ratio(void)
5354 calculate_zone_inactive_ratio(zone);
5358 * Initialise min_free_kbytes.
5360 * For small machines we want it small (128k min). For large machines
5361 * we want it large (64MB max). But it is not linear, because network
5362 * bandwidth does not increase linearly with machine size. We use
5364 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5365 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5381 int __meminit init_per_zone_wmark_min(void)
5383 unsigned long lowmem_kbytes;
5385 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5387 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5388 if (min_free_kbytes < 128)
5389 min_free_kbytes = 128;
5390 if (min_free_kbytes > 65536)
5391 min_free_kbytes = 65536;
5392 setup_per_zone_wmarks();
5393 refresh_zone_stat_thresholds();
5394 setup_per_zone_lowmem_reserve();
5395 setup_per_zone_inactive_ratio();
5398 module_init(init_per_zone_wmark_min)
5401 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5402 * that we can call two helper functions whenever min_free_kbytes
5405 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5406 void __user *buffer, size_t *length, loff_t *ppos)
5408 proc_dointvec(table, write, buffer, length, ppos);
5410 setup_per_zone_wmarks();
5415 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5416 void __user *buffer, size_t *length, loff_t *ppos)
5421 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5426 zone->min_unmapped_pages = (zone->present_pages *
5427 sysctl_min_unmapped_ratio) / 100;
5431 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5432 void __user *buffer, size_t *length, loff_t *ppos)
5437 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5442 zone->min_slab_pages = (zone->present_pages *
5443 sysctl_min_slab_ratio) / 100;
5449 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5450 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5451 * whenever sysctl_lowmem_reserve_ratio changes.
5453 * The reserve ratio obviously has absolutely no relation with the
5454 * minimum watermarks. The lowmem reserve ratio can only make sense
5455 * if in function of the boot time zone sizes.
5457 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5458 void __user *buffer, size_t *length, loff_t *ppos)
5460 proc_dointvec_minmax(table, write, buffer, length, ppos);
5461 setup_per_zone_lowmem_reserve();
5466 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5467 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5468 * can have before it gets flushed back to buddy allocator.
5471 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5472 void __user *buffer, size_t *length, loff_t *ppos)
5478 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5479 if (!write || (ret == -EINVAL))
5481 for_each_populated_zone(zone) {
5482 for_each_possible_cpu(cpu) {
5484 high = zone->present_pages / percpu_pagelist_fraction;
5485 setup_pagelist_highmark(
5486 per_cpu_ptr(zone->pageset, cpu), high);
5492 int hashdist = HASHDIST_DEFAULT;
5495 static int __init set_hashdist(char *str)
5499 hashdist = simple_strtoul(str, &str, 0);
5502 __setup("hashdist=", set_hashdist);
5506 * allocate a large system hash table from bootmem
5507 * - it is assumed that the hash table must contain an exact power-of-2
5508 * quantity of entries
5509 * - limit is the number of hash buckets, not the total allocation size
5511 void *__init alloc_large_system_hash(const char *tablename,
5512 unsigned long bucketsize,
5513 unsigned long numentries,
5516 unsigned int *_hash_shift,
5517 unsigned int *_hash_mask,
5518 unsigned long limit)
5520 unsigned long long max = limit;
5521 unsigned long log2qty, size;
5524 /* allow the kernel cmdline to have a say */
5526 /* round applicable memory size up to nearest megabyte */
5527 numentries = nr_kernel_pages;
5528 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5529 numentries >>= 20 - PAGE_SHIFT;
5530 numentries <<= 20 - PAGE_SHIFT;
5532 /* limit to 1 bucket per 2^scale bytes of low memory */
5533 if (scale > PAGE_SHIFT)
5534 numentries >>= (scale - PAGE_SHIFT);
5536 numentries <<= (PAGE_SHIFT - scale);
5538 /* Make sure we've got at least a 0-order allocation.. */
5539 if (unlikely(flags & HASH_SMALL)) {
5540 /* Makes no sense without HASH_EARLY */
5541 WARN_ON(!(flags & HASH_EARLY));
5542 if (!(numentries >> *_hash_shift)) {
5543 numentries = 1UL << *_hash_shift;
5544 BUG_ON(!numentries);
5546 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5547 numentries = PAGE_SIZE / bucketsize;
5549 numentries = roundup_pow_of_two(numentries);
5551 /* limit allocation size to 1/16 total memory by default */
5553 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5554 do_div(max, bucketsize);
5557 if (numentries > max)
5560 log2qty = ilog2(numentries);
5563 size = bucketsize << log2qty;
5564 if (flags & HASH_EARLY)
5565 table = alloc_bootmem_nopanic(size);
5567 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5570 * If bucketsize is not a power-of-two, we may free
5571 * some pages at the end of hash table which
5572 * alloc_pages_exact() automatically does
5574 if (get_order(size) < MAX_ORDER) {
5575 table = alloc_pages_exact(size, GFP_ATOMIC);
5576 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5579 } while (!table && size > PAGE_SIZE && --log2qty);
5582 panic("Failed to allocate %s hash table\n", tablename);
5584 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5587 ilog2(size) - PAGE_SHIFT,
5591 *_hash_shift = log2qty;
5593 *_hash_mask = (1 << log2qty) - 1;
5598 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5599 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5602 #ifdef CONFIG_SPARSEMEM
5603 return __pfn_to_section(pfn)->pageblock_flags;
5605 return zone->pageblock_flags;
5606 #endif /* CONFIG_SPARSEMEM */
5609 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5611 #ifdef CONFIG_SPARSEMEM
5612 pfn &= (PAGES_PER_SECTION-1);
5613 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5615 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5616 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5617 #endif /* CONFIG_SPARSEMEM */
5621 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5622 * @page: The page within the block of interest
5623 * @start_bitidx: The first bit of interest to retrieve
5624 * @end_bitidx: The last bit of interest
5625 * returns pageblock_bits flags
5627 unsigned long get_pageblock_flags_group(struct page *page,
5628 int start_bitidx, int end_bitidx)
5631 unsigned long *bitmap;
5632 unsigned long pfn, bitidx;
5633 unsigned long flags = 0;
5634 unsigned long value = 1;
5636 zone = page_zone(page);
5637 pfn = page_to_pfn(page);
5638 bitmap = get_pageblock_bitmap(zone, pfn);
5639 bitidx = pfn_to_bitidx(zone, pfn);
5641 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5642 if (test_bit(bitidx + start_bitidx, bitmap))
5649 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5650 * @page: The page within the block of interest
5651 * @start_bitidx: The first bit of interest
5652 * @end_bitidx: The last bit of interest
5653 * @flags: The flags to set
5655 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5656 int start_bitidx, int end_bitidx)
5659 unsigned long *bitmap;
5660 unsigned long pfn, bitidx;
5661 unsigned long value = 1;
5663 zone = page_zone(page);
5664 pfn = page_to_pfn(page);
5665 bitmap = get_pageblock_bitmap(zone, pfn);
5666 bitidx = pfn_to_bitidx(zone, pfn);
5667 VM_BUG_ON(pfn < zone->zone_start_pfn);
5668 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5670 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5672 __set_bit(bitidx + start_bitidx, bitmap);
5674 __clear_bit(bitidx + start_bitidx, bitmap);
5678 * This is designed as sub function...plz see page_isolation.c also.
5679 * set/clear page block's type to be ISOLATE.
5680 * page allocater never alloc memory from ISOLATE block.
5684 __count_immobile_pages(struct zone *zone, struct page *page, int count)
5686 unsigned long pfn, iter, found;
5688 * For avoiding noise data, lru_add_drain_all() should be called
5689 * If ZONE_MOVABLE, the zone never contains immobile pages
5691 if (zone_idx(zone) == ZONE_MOVABLE)
5694 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5697 pfn = page_to_pfn(page);
5698 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5699 unsigned long check = pfn + iter;
5701 if (!pfn_valid_within(check))
5704 page = pfn_to_page(check);
5705 if (!page_count(page)) {
5706 if (PageBuddy(page))
5707 iter += (1 << page_order(page)) - 1;
5713 * If there are RECLAIMABLE pages, we need to check it.
5714 * But now, memory offline itself doesn't call shrink_slab()
5715 * and it still to be fixed.
5718 * If the page is not RAM, page_count()should be 0.
5719 * we don't need more check. This is an _used_ not-movable page.
5721 * The problematic thing here is PG_reserved pages. PG_reserved
5722 * is set to both of a memory hole page and a _used_ kernel
5731 bool is_pageblock_removable_nolock(struct page *page)
5733 struct zone *zone = page_zone(page);
5734 unsigned long pfn = page_to_pfn(page);
5737 * We have to be careful here because we are iterating over memory
5738 * sections which are not zone aware so we might end up outside of
5739 * the zone but still within the section.
5741 if (!zone || zone->zone_start_pfn > pfn ||
5742 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5745 return __count_immobile_pages(zone, page, 0);
5748 int set_migratetype_isolate(struct page *page)
5751 unsigned long flags, pfn;
5752 struct memory_isolate_notify arg;
5756 zone = page_zone(page);
5758 spin_lock_irqsave(&zone->lock, flags);
5760 pfn = page_to_pfn(page);
5761 arg.start_pfn = pfn;
5762 arg.nr_pages = pageblock_nr_pages;
5763 arg.pages_found = 0;
5766 * It may be possible to isolate a pageblock even if the
5767 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5768 * notifier chain is used by balloon drivers to return the
5769 * number of pages in a range that are held by the balloon
5770 * driver to shrink memory. If all the pages are accounted for
5771 * by balloons, are free, or on the LRU, isolation can continue.
5772 * Later, for example, when memory hotplug notifier runs, these
5773 * pages reported as "can be isolated" should be isolated(freed)
5774 * by the balloon driver through the memory notifier chain.
5776 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5777 notifier_ret = notifier_to_errno(notifier_ret);
5781 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5782 * We just check MOVABLE pages.
5784 if (__count_immobile_pages(zone, page, arg.pages_found))
5788 * immobile means "not-on-lru" paes. If immobile is larger than
5789 * removable-by-driver pages reported by notifier, we'll fail.
5794 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5795 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5798 spin_unlock_irqrestore(&zone->lock, flags);
5804 void unset_migratetype_isolate(struct page *page)
5807 unsigned long flags;
5808 zone = page_zone(page);
5809 spin_lock_irqsave(&zone->lock, flags);
5810 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5812 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5813 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5815 spin_unlock_irqrestore(&zone->lock, flags);
5818 #ifdef CONFIG_MEMORY_HOTREMOVE
5820 * All pages in the range must be isolated before calling this.
5823 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5829 unsigned long flags;
5830 /* find the first valid pfn */
5831 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5836 zone = page_zone(pfn_to_page(pfn));
5837 spin_lock_irqsave(&zone->lock, flags);
5839 while (pfn < end_pfn) {
5840 if (!pfn_valid(pfn)) {
5844 page = pfn_to_page(pfn);
5845 BUG_ON(page_count(page));
5846 BUG_ON(!PageBuddy(page));
5847 order = page_order(page);
5848 #ifdef CONFIG_DEBUG_VM
5849 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5850 pfn, 1 << order, end_pfn);
5852 list_del(&page->lru);
5853 rmv_page_order(page);
5854 zone->free_area[order].nr_free--;
5855 __mod_zone_page_state(zone, NR_FREE_PAGES,
5857 for (i = 0; i < (1 << order); i++)
5858 SetPageReserved((page+i));
5859 pfn += (1 << order);
5861 spin_unlock_irqrestore(&zone->lock, flags);
5865 #ifdef CONFIG_MEMORY_FAILURE
5866 bool is_free_buddy_page(struct page *page)
5868 struct zone *zone = page_zone(page);
5869 unsigned long pfn = page_to_pfn(page);
5870 unsigned long flags;
5873 spin_lock_irqsave(&zone->lock, flags);
5874 for (order = 0; order < MAX_ORDER; order++) {
5875 struct page *page_head = page - (pfn & ((1 << order) - 1));
5877 if (PageBuddy(page_head) && page_order(page_head) >= order)
5880 spin_unlock_irqrestore(&zone->lock, flags);
5882 return order < MAX_ORDER;
5886 static struct trace_print_flags pageflag_names[] = {
5887 {1UL << PG_locked, "locked" },
5888 {1UL << PG_error, "error" },
5889 {1UL << PG_referenced, "referenced" },
5890 {1UL << PG_uptodate, "uptodate" },
5891 {1UL << PG_dirty, "dirty" },
5892 {1UL << PG_lru, "lru" },
5893 {1UL << PG_active, "active" },
5894 {1UL << PG_slab, "slab" },
5895 {1UL << PG_owner_priv_1, "owner_priv_1" },
5896 {1UL << PG_arch_1, "arch_1" },
5897 {1UL << PG_reserved, "reserved" },
5898 {1UL << PG_private, "private" },
5899 {1UL << PG_private_2, "private_2" },
5900 {1UL << PG_writeback, "writeback" },
5901 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5902 {1UL << PG_head, "head" },
5903 {1UL << PG_tail, "tail" },
5905 {1UL << PG_compound, "compound" },
5907 {1UL << PG_swapcache, "swapcache" },
5908 {1UL << PG_mappedtodisk, "mappedtodisk" },
5909 {1UL << PG_reclaim, "reclaim" },
5910 {1UL << PG_swapbacked, "swapbacked" },
5911 {1UL << PG_unevictable, "unevictable" },
5913 {1UL << PG_mlocked, "mlocked" },
5915 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5916 {1UL << PG_uncached, "uncached" },
5918 #ifdef CONFIG_MEMORY_FAILURE
5919 {1UL << PG_hwpoison, "hwpoison" },
5924 static void dump_page_flags(unsigned long flags)
5926 const char *delim = "";
5930 printk(KERN_ALERT "page flags: %#lx(", flags);
5932 /* remove zone id */
5933 flags &= (1UL << NR_PAGEFLAGS) - 1;
5935 for (i = 0; pageflag_names[i].name && flags; i++) {
5937 mask = pageflag_names[i].mask;
5938 if ((flags & mask) != mask)
5942 printk("%s%s", delim, pageflag_names[i].name);
5946 /* check for left over flags */
5948 printk("%s%#lx", delim, flags);
5953 void dump_page(struct page *page)
5956 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5957 page, atomic_read(&page->_count), page_mapcount(page),
5958 page->mapping, page->index);
5959 dump_page_flags(page->flags);
5960 mem_cgroup_print_bad_page(page);