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/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.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 #ifdef CONFIG_MOVABLE_NODE
94 [N_MEMORY] = { { [0] = 1UL } },
96 [N_CPU] = { { [0] = 1UL } },
99 EXPORT_SYMBOL(node_states);
101 unsigned long totalram_pages __read_mostly;
102 unsigned long totalreserve_pages __read_mostly;
104 * When calculating the number of globally allowed dirty pages, there
105 * is a certain number of per-zone reserves that should not be
106 * considered dirtyable memory. This is the sum of those reserves
107 * over all existing zones that contribute dirtyable memory.
109 unsigned long dirty_balance_reserve __read_mostly;
111 int percpu_pagelist_fraction;
112 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
114 #ifdef CONFIG_PM_SLEEP
116 * The following functions are used by the suspend/hibernate code to temporarily
117 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
118 * while devices are suspended. To avoid races with the suspend/hibernate code,
119 * they should always be called with pm_mutex held (gfp_allowed_mask also should
120 * only be modified with pm_mutex held, unless the suspend/hibernate code is
121 * guaranteed not to run in parallel with that modification).
124 static gfp_t saved_gfp_mask;
126 void pm_restore_gfp_mask(void)
128 WARN_ON(!mutex_is_locked(&pm_mutex));
129 if (saved_gfp_mask) {
130 gfp_allowed_mask = saved_gfp_mask;
135 void pm_restrict_gfp_mask(void)
137 WARN_ON(!mutex_is_locked(&pm_mutex));
138 WARN_ON(saved_gfp_mask);
139 saved_gfp_mask = gfp_allowed_mask;
140 gfp_allowed_mask &= ~GFP_IOFS;
143 bool pm_suspended_storage(void)
145 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
149 #endif /* CONFIG_PM_SLEEP */
151 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
152 int pageblock_order __read_mostly;
155 static void __free_pages_ok(struct page *page, unsigned int order);
158 * results with 256, 32 in the lowmem_reserve sysctl:
159 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
160 * 1G machine -> (16M dma, 784M normal, 224M high)
161 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
162 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
163 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
165 * TBD: should special case ZONE_DMA32 machines here - in those we normally
166 * don't need any ZONE_NORMAL reservation
168 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
169 #ifdef CONFIG_ZONE_DMA
172 #ifdef CONFIG_ZONE_DMA32
175 #ifdef CONFIG_HIGHMEM
181 EXPORT_SYMBOL(totalram_pages);
183 static char * const zone_names[MAX_NR_ZONES] = {
184 #ifdef CONFIG_ZONE_DMA
187 #ifdef CONFIG_ZONE_DMA32
191 #ifdef CONFIG_HIGHMEM
197 int min_free_kbytes = 1024;
199 static unsigned long __meminitdata nr_kernel_pages;
200 static unsigned long __meminitdata nr_all_pages;
201 static unsigned long __meminitdata dma_reserve;
203 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
204 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
205 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
206 static unsigned long __initdata required_kernelcore;
207 static unsigned long __initdata required_movablecore;
208 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
210 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
212 EXPORT_SYMBOL(movable_zone);
213 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
216 int nr_node_ids __read_mostly = MAX_NUMNODES;
217 int nr_online_nodes __read_mostly = 1;
218 EXPORT_SYMBOL(nr_node_ids);
219 EXPORT_SYMBOL(nr_online_nodes);
222 int page_group_by_mobility_disabled __read_mostly;
226 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
227 * Instead, use {un}set_pageblock_isolate.
229 void set_pageblock_migratetype(struct page *page, int migratetype)
232 if (unlikely(page_group_by_mobility_disabled))
233 migratetype = MIGRATE_UNMOVABLE;
235 set_pageblock_flags_group(page, (unsigned long)migratetype,
236 PB_migrate, PB_migrate_end);
239 bool oom_killer_disabled __read_mostly;
241 #ifdef CONFIG_DEBUG_VM
242 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
246 unsigned long pfn = page_to_pfn(page);
249 seq = zone_span_seqbegin(zone);
250 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
252 else if (pfn < zone->zone_start_pfn)
254 } while (zone_span_seqretry(zone, seq));
259 static int page_is_consistent(struct zone *zone, struct page *page)
261 if (!pfn_valid_within(page_to_pfn(page)))
263 if (zone != page_zone(page))
269 * Temporary debugging check for pages not lying within a given zone.
271 static int bad_range(struct zone *zone, struct page *page)
273 if (page_outside_zone_boundaries(zone, page))
275 if (!page_is_consistent(zone, page))
281 static inline int bad_range(struct zone *zone, struct page *page)
287 static void bad_page(struct page *page)
289 static unsigned long resume;
290 static unsigned long nr_shown;
291 static unsigned long nr_unshown;
293 /* Don't complain about poisoned pages */
294 if (PageHWPoison(page)) {
295 reset_page_mapcount(page); /* remove PageBuddy */
300 * Allow a burst of 60 reports, then keep quiet for that minute;
301 * or allow a steady drip of one report per second.
303 if (nr_shown == 60) {
304 if (time_before(jiffies, resume)) {
310 "BUG: Bad page state: %lu messages suppressed\n",
317 resume = jiffies + 60 * HZ;
319 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
320 current->comm, page_to_pfn(page));
326 /* Leave bad fields for debug, except PageBuddy could make trouble */
327 reset_page_mapcount(page); /* remove PageBuddy */
328 add_taint(TAINT_BAD_PAGE);
332 * Higher-order pages are called "compound pages". They are structured thusly:
334 * The first PAGE_SIZE page is called the "head page".
336 * The remaining PAGE_SIZE pages are called "tail pages".
338 * All pages have PG_compound set. All tail pages have their ->first_page
339 * pointing at the head page.
341 * The first tail page's ->lru.next holds the address of the compound page's
342 * put_page() function. Its ->lru.prev holds the order of allocation.
343 * This usage means that zero-order pages may not be compound.
346 static void free_compound_page(struct page *page)
348 __free_pages_ok(page, compound_order(page));
351 void prep_compound_page(struct page *page, unsigned long order)
354 int nr_pages = 1 << order;
356 set_compound_page_dtor(page, free_compound_page);
357 set_compound_order(page, order);
359 for (i = 1; i < nr_pages; i++) {
360 struct page *p = page + i;
362 set_page_count(p, 0);
363 p->first_page = page;
367 /* update __split_huge_page_refcount if you change this function */
368 static int destroy_compound_page(struct page *page, unsigned long order)
371 int nr_pages = 1 << order;
374 if (unlikely(compound_order(page) != order) ||
375 unlikely(!PageHead(page))) {
380 __ClearPageHead(page);
382 for (i = 1; i < nr_pages; i++) {
383 struct page *p = page + i;
385 if (unlikely(!PageTail(p) || (p->first_page != page))) {
395 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
400 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
401 * and __GFP_HIGHMEM from hard or soft interrupt context.
403 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
404 for (i = 0; i < (1 << order); i++)
405 clear_highpage(page + i);
408 #ifdef CONFIG_DEBUG_PAGEALLOC
409 unsigned int _debug_guardpage_minorder;
411 static int __init debug_guardpage_minorder_setup(char *buf)
415 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
416 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
419 _debug_guardpage_minorder = res;
420 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
423 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
425 static inline void set_page_guard_flag(struct page *page)
427 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
430 static inline void clear_page_guard_flag(struct page *page)
432 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
435 static inline void set_page_guard_flag(struct page *page) { }
436 static inline void clear_page_guard_flag(struct page *page) { }
439 static inline void set_page_order(struct page *page, int order)
441 set_page_private(page, order);
442 __SetPageBuddy(page);
445 static inline void rmv_page_order(struct page *page)
447 __ClearPageBuddy(page);
448 set_page_private(page, 0);
452 * Locate the struct page for both the matching buddy in our
453 * pair (buddy1) and the combined O(n+1) page they form (page).
455 * 1) Any buddy B1 will have an order O twin B2 which satisfies
456 * the following equation:
458 * For example, if the starting buddy (buddy2) is #8 its order
460 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
462 * 2) Any buddy B will have an order O+1 parent P which
463 * satisfies the following equation:
466 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
468 static inline unsigned long
469 __find_buddy_index(unsigned long page_idx, unsigned int order)
471 return page_idx ^ (1 << order);
475 * This function checks whether a page is free && is the buddy
476 * we can do coalesce a page and its buddy if
477 * (a) the buddy is not in a hole &&
478 * (b) the buddy is in the buddy system &&
479 * (c) a page and its buddy have the same order &&
480 * (d) a page and its buddy are in the same zone.
482 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
483 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
485 * For recording page's order, we use page_private(page).
487 static inline int page_is_buddy(struct page *page, struct page *buddy,
490 if (!pfn_valid_within(page_to_pfn(buddy)))
493 if (page_zone_id(page) != page_zone_id(buddy))
496 if (page_is_guard(buddy) && page_order(buddy) == order) {
497 VM_BUG_ON(page_count(buddy) != 0);
501 if (PageBuddy(buddy) && page_order(buddy) == order) {
502 VM_BUG_ON(page_count(buddy) != 0);
509 * Freeing function for a buddy system allocator.
511 * The concept of a buddy system is to maintain direct-mapped table
512 * (containing bit values) for memory blocks of various "orders".
513 * The bottom level table contains the map for the smallest allocatable
514 * units of memory (here, pages), and each level above it describes
515 * pairs of units from the levels below, hence, "buddies".
516 * At a high level, all that happens here is marking the table entry
517 * at the bottom level available, and propagating the changes upward
518 * as necessary, plus some accounting needed to play nicely with other
519 * parts of the VM system.
520 * At each level, we keep a list of pages, which are heads of continuous
521 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
522 * order is recorded in page_private(page) field.
523 * So when we are allocating or freeing one, we can derive the state of the
524 * other. That is, if we allocate a small block, and both were
525 * free, the remainder of the region must be split into blocks.
526 * If a block is freed, and its buddy is also free, then this
527 * triggers coalescing into a block of larger size.
532 static inline void __free_one_page(struct page *page,
533 struct zone *zone, unsigned int order,
536 unsigned long page_idx;
537 unsigned long combined_idx;
538 unsigned long uninitialized_var(buddy_idx);
541 if (unlikely(PageCompound(page)))
542 if (unlikely(destroy_compound_page(page, order)))
545 VM_BUG_ON(migratetype == -1);
547 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
549 VM_BUG_ON(page_idx & ((1 << order) - 1));
550 VM_BUG_ON(bad_range(zone, page));
552 while (order < MAX_ORDER-1) {
553 buddy_idx = __find_buddy_index(page_idx, order);
554 buddy = page + (buddy_idx - page_idx);
555 if (!page_is_buddy(page, buddy, order))
558 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
559 * merge with it and move up one order.
561 if (page_is_guard(buddy)) {
562 clear_page_guard_flag(buddy);
563 set_page_private(page, 0);
564 __mod_zone_freepage_state(zone, 1 << order,
567 list_del(&buddy->lru);
568 zone->free_area[order].nr_free--;
569 rmv_page_order(buddy);
571 combined_idx = buddy_idx & page_idx;
572 page = page + (combined_idx - page_idx);
573 page_idx = combined_idx;
576 set_page_order(page, order);
579 * If this is not the largest possible page, check if the buddy
580 * of the next-highest order is free. If it is, it's possible
581 * that pages are being freed that will coalesce soon. In case,
582 * that is happening, add the free page to the tail of the list
583 * so it's less likely to be used soon and more likely to be merged
584 * as a higher order page
586 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
587 struct page *higher_page, *higher_buddy;
588 combined_idx = buddy_idx & page_idx;
589 higher_page = page + (combined_idx - page_idx);
590 buddy_idx = __find_buddy_index(combined_idx, order + 1);
591 higher_buddy = higher_page + (buddy_idx - combined_idx);
592 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
593 list_add_tail(&page->lru,
594 &zone->free_area[order].free_list[migratetype]);
599 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
601 zone->free_area[order].nr_free++;
604 static inline int free_pages_check(struct page *page)
606 if (unlikely(page_mapcount(page) |
607 (page->mapping != NULL) |
608 (atomic_read(&page->_count) != 0) |
609 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
610 (mem_cgroup_bad_page_check(page)))) {
614 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
615 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
620 * Frees a number of pages from the PCP lists
621 * Assumes all pages on list are in same zone, and of same order.
622 * count is the number of pages to free.
624 * If the zone was previously in an "all pages pinned" state then look to
625 * see if this freeing clears that state.
627 * And clear the zone's pages_scanned counter, to hold off the "all pages are
628 * pinned" detection logic.
630 static void free_pcppages_bulk(struct zone *zone, int count,
631 struct per_cpu_pages *pcp)
637 spin_lock(&zone->lock);
638 zone->all_unreclaimable = 0;
639 zone->pages_scanned = 0;
643 struct list_head *list;
646 * Remove pages from lists in a round-robin fashion. A
647 * batch_free count is maintained that is incremented when an
648 * empty list is encountered. This is so more pages are freed
649 * off fuller lists instead of spinning excessively around empty
654 if (++migratetype == MIGRATE_PCPTYPES)
656 list = &pcp->lists[migratetype];
657 } while (list_empty(list));
659 /* This is the only non-empty list. Free them all. */
660 if (batch_free == MIGRATE_PCPTYPES)
661 batch_free = to_free;
664 int mt; /* migratetype of the to-be-freed page */
666 page = list_entry(list->prev, struct page, lru);
667 /* must delete as __free_one_page list manipulates */
668 list_del(&page->lru);
669 mt = get_freepage_migratetype(page);
670 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
671 __free_one_page(page, zone, 0, mt);
672 trace_mm_page_pcpu_drain(page, 0, mt);
673 if (likely(get_pageblock_migratetype(page) != MIGRATE_ISOLATE)) {
674 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
675 if (is_migrate_cma(mt))
676 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
678 } while (--to_free && --batch_free && !list_empty(list));
680 spin_unlock(&zone->lock);
683 static void free_one_page(struct zone *zone, struct page *page, int order,
686 spin_lock(&zone->lock);
687 zone->all_unreclaimable = 0;
688 zone->pages_scanned = 0;
690 __free_one_page(page, zone, order, migratetype);
691 if (unlikely(migratetype != MIGRATE_ISOLATE))
692 __mod_zone_freepage_state(zone, 1 << order, migratetype);
693 spin_unlock(&zone->lock);
696 static bool free_pages_prepare(struct page *page, unsigned int order)
701 trace_mm_page_free(page, order);
702 kmemcheck_free_shadow(page, order);
705 page->mapping = NULL;
706 for (i = 0; i < (1 << order); i++)
707 bad += free_pages_check(page + i);
711 if (!PageHighMem(page)) {
712 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
713 debug_check_no_obj_freed(page_address(page),
716 arch_free_page(page, order);
717 kernel_map_pages(page, 1 << order, 0);
722 static void __free_pages_ok(struct page *page, unsigned int order)
727 if (!free_pages_prepare(page, order))
730 local_irq_save(flags);
731 __count_vm_events(PGFREE, 1 << order);
732 migratetype = get_pageblock_migratetype(page);
733 set_freepage_migratetype(page, migratetype);
734 free_one_page(page_zone(page), page, order, migratetype);
735 local_irq_restore(flags);
738 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
740 unsigned int nr_pages = 1 << order;
744 for (loop = 0; loop < nr_pages; loop++) {
745 struct page *p = &page[loop];
747 if (loop + 1 < nr_pages)
749 __ClearPageReserved(p);
750 set_page_count(p, 0);
753 set_page_refcounted(page);
754 __free_pages(page, order);
758 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
759 void __init init_cma_reserved_pageblock(struct page *page)
761 unsigned i = pageblock_nr_pages;
762 struct page *p = page;
765 __ClearPageReserved(p);
766 set_page_count(p, 0);
769 set_page_refcounted(page);
770 set_pageblock_migratetype(page, MIGRATE_CMA);
771 __free_pages(page, pageblock_order);
772 totalram_pages += pageblock_nr_pages;
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_freepage_state(zone, -(1 << high),
819 list_add(&page[size].lru, &area->free_list[migratetype]);
821 set_page_order(&page[size], high);
826 * This page is about to be returned from the page allocator
828 static inline int check_new_page(struct page *page)
830 if (unlikely(page_mapcount(page) |
831 (page->mapping != NULL) |
832 (atomic_read(&page->_count) != 0) |
833 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
834 (mem_cgroup_bad_page_check(page)))) {
841 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
845 for (i = 0; i < (1 << order); i++) {
846 struct page *p = page + i;
847 if (unlikely(check_new_page(p)))
851 set_page_private(page, 0);
852 set_page_refcounted(page);
854 arch_alloc_page(page, order);
855 kernel_map_pages(page, 1 << order, 1);
857 if (gfp_flags & __GFP_ZERO)
858 prep_zero_page(page, order, gfp_flags);
860 if (order && (gfp_flags & __GFP_COMP))
861 prep_compound_page(page, order);
867 * Go through the free lists for the given migratetype and remove
868 * the smallest available page from the freelists
871 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
874 unsigned int current_order;
875 struct free_area * area;
878 /* Find a page of the appropriate size in the preferred list */
879 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
880 area = &(zone->free_area[current_order]);
881 if (list_empty(&area->free_list[migratetype]))
884 page = list_entry(area->free_list[migratetype].next,
886 list_del(&page->lru);
887 rmv_page_order(page);
889 expand(zone, page, order, current_order, area, migratetype);
898 * This array describes the order lists are fallen back to when
899 * the free lists for the desirable migrate type are depleted
901 static int fallbacks[MIGRATE_TYPES][4] = {
902 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
903 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
905 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
906 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
908 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
910 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
911 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
915 * Move the free pages in a range to the free lists of the requested type.
916 * Note that start_page and end_pages are not aligned on a pageblock
917 * boundary. If alignment is required, use move_freepages_block()
919 int move_freepages(struct zone *zone,
920 struct page *start_page, struct page *end_page,
927 #ifndef CONFIG_HOLES_IN_ZONE
929 * page_zone is not safe to call in this context when
930 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
931 * anyway as we check zone boundaries in move_freepages_block().
932 * Remove at a later date when no bug reports exist related to
933 * grouping pages by mobility
935 BUG_ON(page_zone(start_page) != page_zone(end_page));
938 for (page = start_page; page <= end_page;) {
939 /* Make sure we are not inadvertently changing nodes */
940 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
942 if (!pfn_valid_within(page_to_pfn(page))) {
947 if (!PageBuddy(page)) {
952 order = page_order(page);
953 list_move(&page->lru,
954 &zone->free_area[order].free_list[migratetype]);
955 set_freepage_migratetype(page, migratetype);
957 pages_moved += 1 << order;
963 int move_freepages_block(struct zone *zone, struct page *page,
966 unsigned long start_pfn, end_pfn;
967 struct page *start_page, *end_page;
969 start_pfn = page_to_pfn(page);
970 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
971 start_page = pfn_to_page(start_pfn);
972 end_page = start_page + pageblock_nr_pages - 1;
973 end_pfn = start_pfn + pageblock_nr_pages - 1;
975 /* Do not cross zone boundaries */
976 if (start_pfn < zone->zone_start_pfn)
978 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
981 return move_freepages(zone, start_page, end_page, migratetype);
984 static void change_pageblock_range(struct page *pageblock_page,
985 int start_order, int migratetype)
987 int nr_pageblocks = 1 << (start_order - pageblock_order);
989 while (nr_pageblocks--) {
990 set_pageblock_migratetype(pageblock_page, migratetype);
991 pageblock_page += pageblock_nr_pages;
995 /* Remove an element from the buddy allocator from the fallback list */
996 static inline struct page *
997 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
999 struct free_area * area;
1004 /* Find the largest possible block of pages in the other list */
1005 for (current_order = MAX_ORDER-1; current_order >= order;
1008 migratetype = fallbacks[start_migratetype][i];
1010 /* MIGRATE_RESERVE handled later if necessary */
1011 if (migratetype == MIGRATE_RESERVE)
1014 area = &(zone->free_area[current_order]);
1015 if (list_empty(&area->free_list[migratetype]))
1018 page = list_entry(area->free_list[migratetype].next,
1023 * If breaking a large block of pages, move all free
1024 * pages to the preferred allocation list. If falling
1025 * back for a reclaimable kernel allocation, be more
1026 * aggressive about taking ownership of free pages
1028 * On the other hand, never change migration
1029 * type of MIGRATE_CMA pageblocks nor move CMA
1030 * pages on different free lists. We don't
1031 * want unmovable pages to be allocated from
1032 * MIGRATE_CMA areas.
1034 if (!is_migrate_cma(migratetype) &&
1035 (unlikely(current_order >= pageblock_order / 2) ||
1036 start_migratetype == MIGRATE_RECLAIMABLE ||
1037 page_group_by_mobility_disabled)) {
1039 pages = move_freepages_block(zone, page,
1042 /* Claim the whole block if over half of it is free */
1043 if (pages >= (1 << (pageblock_order-1)) ||
1044 page_group_by_mobility_disabled)
1045 set_pageblock_migratetype(page,
1048 migratetype = start_migratetype;
1051 /* Remove the page from the freelists */
1052 list_del(&page->lru);
1053 rmv_page_order(page);
1055 /* Take ownership for orders >= pageblock_order */
1056 if (current_order >= pageblock_order &&
1057 !is_migrate_cma(migratetype))
1058 change_pageblock_range(page, current_order,
1061 expand(zone, page, order, current_order, area,
1062 is_migrate_cma(migratetype)
1063 ? migratetype : start_migratetype);
1065 trace_mm_page_alloc_extfrag(page, order, current_order,
1066 start_migratetype, migratetype);
1076 * Do the hard work of removing an element from the buddy allocator.
1077 * Call me with the zone->lock already held.
1079 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1085 page = __rmqueue_smallest(zone, order, migratetype);
1087 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1088 page = __rmqueue_fallback(zone, order, migratetype);
1091 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1092 * is used because __rmqueue_smallest is an inline function
1093 * and we want just one call site
1096 migratetype = MIGRATE_RESERVE;
1101 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1106 * Obtain a specified number of elements from the buddy allocator, all under
1107 * a single hold of the lock, for efficiency. Add them to the supplied list.
1108 * Returns the number of new pages which were placed at *list.
1110 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1111 unsigned long count, struct list_head *list,
1112 int migratetype, int cold)
1114 int mt = migratetype, i;
1116 spin_lock(&zone->lock);
1117 for (i = 0; i < count; ++i) {
1118 struct page *page = __rmqueue(zone, order, migratetype);
1119 if (unlikely(page == NULL))
1123 * Split buddy pages returned by expand() are received here
1124 * in physical page order. The page is added to the callers and
1125 * list and the list head then moves forward. From the callers
1126 * perspective, the linked list is ordered by page number in
1127 * some conditions. This is useful for IO devices that can
1128 * merge IO requests if the physical pages are ordered
1131 if (likely(cold == 0))
1132 list_add(&page->lru, list);
1134 list_add_tail(&page->lru, list);
1135 if (IS_ENABLED(CONFIG_CMA)) {
1136 mt = get_pageblock_migratetype(page);
1137 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1140 set_freepage_migratetype(page, mt);
1142 if (is_migrate_cma(mt))
1143 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1146 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1147 spin_unlock(&zone->lock);
1153 * Called from the vmstat counter updater to drain pagesets of this
1154 * currently executing processor on remote nodes after they have
1157 * Note that this function must be called with the thread pinned to
1158 * a single processor.
1160 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1162 unsigned long flags;
1165 local_irq_save(flags);
1166 if (pcp->count >= pcp->batch)
1167 to_drain = pcp->batch;
1169 to_drain = pcp->count;
1171 free_pcppages_bulk(zone, to_drain, pcp);
1172 pcp->count -= to_drain;
1174 local_irq_restore(flags);
1179 * Drain pages of the indicated processor.
1181 * The processor must either be the current processor and the
1182 * thread pinned to the current processor or a processor that
1185 static void drain_pages(unsigned int cpu)
1187 unsigned long flags;
1190 for_each_populated_zone(zone) {
1191 struct per_cpu_pageset *pset;
1192 struct per_cpu_pages *pcp;
1194 local_irq_save(flags);
1195 pset = per_cpu_ptr(zone->pageset, cpu);
1199 free_pcppages_bulk(zone, pcp->count, pcp);
1202 local_irq_restore(flags);
1207 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1209 void drain_local_pages(void *arg)
1211 drain_pages(smp_processor_id());
1215 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1217 * Note that this code is protected against sending an IPI to an offline
1218 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1219 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1220 * nothing keeps CPUs from showing up after we populated the cpumask and
1221 * before the call to on_each_cpu_mask().
1223 void drain_all_pages(void)
1226 struct per_cpu_pageset *pcp;
1230 * Allocate in the BSS so we wont require allocation in
1231 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1233 static cpumask_t cpus_with_pcps;
1236 * We don't care about racing with CPU hotplug event
1237 * as offline notification will cause the notified
1238 * cpu to drain that CPU pcps and on_each_cpu_mask
1239 * disables preemption as part of its processing
1241 for_each_online_cpu(cpu) {
1242 bool has_pcps = false;
1243 for_each_populated_zone(zone) {
1244 pcp = per_cpu_ptr(zone->pageset, cpu);
1245 if (pcp->pcp.count) {
1251 cpumask_set_cpu(cpu, &cpus_with_pcps);
1253 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1255 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1258 #ifdef CONFIG_HIBERNATION
1260 void mark_free_pages(struct zone *zone)
1262 unsigned long pfn, max_zone_pfn;
1263 unsigned long flags;
1265 struct list_head *curr;
1267 if (!zone->spanned_pages)
1270 spin_lock_irqsave(&zone->lock, flags);
1272 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1273 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1274 if (pfn_valid(pfn)) {
1275 struct page *page = pfn_to_page(pfn);
1277 if (!swsusp_page_is_forbidden(page))
1278 swsusp_unset_page_free(page);
1281 for_each_migratetype_order(order, t) {
1282 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1285 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1286 for (i = 0; i < (1UL << order); i++)
1287 swsusp_set_page_free(pfn_to_page(pfn + i));
1290 spin_unlock_irqrestore(&zone->lock, flags);
1292 #endif /* CONFIG_PM */
1295 * Free a 0-order page
1296 * cold == 1 ? free a cold page : free a hot page
1298 void free_hot_cold_page(struct page *page, int cold)
1300 struct zone *zone = page_zone(page);
1301 struct per_cpu_pages *pcp;
1302 unsigned long flags;
1305 if (!free_pages_prepare(page, 0))
1308 migratetype = get_pageblock_migratetype(page);
1309 set_freepage_migratetype(page, migratetype);
1310 local_irq_save(flags);
1311 __count_vm_event(PGFREE);
1314 * We only track unmovable, reclaimable and movable on pcp lists.
1315 * Free ISOLATE pages back to the allocator because they are being
1316 * offlined but treat RESERVE as movable pages so we can get those
1317 * areas back if necessary. Otherwise, we may have to free
1318 * excessively into the page allocator
1320 if (migratetype >= MIGRATE_PCPTYPES) {
1321 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1322 free_one_page(zone, page, 0, migratetype);
1325 migratetype = MIGRATE_MOVABLE;
1328 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1330 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1332 list_add(&page->lru, &pcp->lists[migratetype]);
1334 if (pcp->count >= pcp->high) {
1335 free_pcppages_bulk(zone, pcp->batch, pcp);
1336 pcp->count -= pcp->batch;
1340 local_irq_restore(flags);
1344 * Free a list of 0-order pages
1346 void free_hot_cold_page_list(struct list_head *list, int cold)
1348 struct page *page, *next;
1350 list_for_each_entry_safe(page, next, list, lru) {
1351 trace_mm_page_free_batched(page, cold);
1352 free_hot_cold_page(page, cold);
1357 * split_page takes a non-compound higher-order page, and splits it into
1358 * n (1<<order) sub-pages: page[0..n]
1359 * Each sub-page must be freed individually.
1361 * Note: this is probably too low level an operation for use in drivers.
1362 * Please consult with lkml before using this in your driver.
1364 void split_page(struct page *page, unsigned int order)
1368 VM_BUG_ON(PageCompound(page));
1369 VM_BUG_ON(!page_count(page));
1371 #ifdef CONFIG_KMEMCHECK
1373 * Split shadow pages too, because free(page[0]) would
1374 * otherwise free the whole shadow.
1376 if (kmemcheck_page_is_tracked(page))
1377 split_page(virt_to_page(page[0].shadow), order);
1380 for (i = 1; i < (1 << order); i++)
1381 set_page_refcounted(page + i);
1385 * Similar to the split_page family of functions except that the page
1386 * required at the given order and being isolated now to prevent races
1387 * with parallel allocators
1389 int capture_free_page(struct page *page, int alloc_order, int migratetype)
1392 unsigned long watermark;
1396 BUG_ON(!PageBuddy(page));
1398 zone = page_zone(page);
1399 order = page_order(page);
1400 mt = get_pageblock_migratetype(page);
1402 if (mt != MIGRATE_ISOLATE) {
1403 /* Obey watermarks as if the page was being allocated */
1404 watermark = low_wmark_pages(zone) + (1 << order);
1405 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1408 __mod_zone_freepage_state(zone, -(1UL << alloc_order), mt);
1411 /* Remove page from free list */
1412 list_del(&page->lru);
1413 zone->free_area[order].nr_free--;
1414 rmv_page_order(page);
1416 if (alloc_order != order)
1417 expand(zone, page, alloc_order, order,
1418 &zone->free_area[order], migratetype);
1420 /* Set the pageblock if the captured page is at least a pageblock */
1421 if (order >= pageblock_order - 1) {
1422 struct page *endpage = page + (1 << order) - 1;
1423 for (; page < endpage; page += pageblock_nr_pages) {
1424 int mt = get_pageblock_migratetype(page);
1425 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1426 set_pageblock_migratetype(page,
1431 return 1UL << alloc_order;
1435 * Similar to split_page except the page is already free. As this is only
1436 * being used for migration, the migratetype of the block also changes.
1437 * As this is called with interrupts disabled, the caller is responsible
1438 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1441 * Note: this is probably too low level an operation for use in drivers.
1442 * Please consult with lkml before using this in your driver.
1444 int split_free_page(struct page *page)
1449 BUG_ON(!PageBuddy(page));
1450 order = page_order(page);
1452 nr_pages = capture_free_page(page, order, 0);
1456 /* Split into individual pages */
1457 set_page_refcounted(page);
1458 split_page(page, order);
1463 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1464 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1468 struct page *buffered_rmqueue(struct zone *preferred_zone,
1469 struct zone *zone, int order, gfp_t gfp_flags,
1472 unsigned long flags;
1474 int cold = !!(gfp_flags & __GFP_COLD);
1477 if (likely(order == 0)) {
1478 struct per_cpu_pages *pcp;
1479 struct list_head *list;
1481 local_irq_save(flags);
1482 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1483 list = &pcp->lists[migratetype];
1484 if (list_empty(list)) {
1485 pcp->count += rmqueue_bulk(zone, 0,
1488 if (unlikely(list_empty(list)))
1493 page = list_entry(list->prev, struct page, lru);
1495 page = list_entry(list->next, struct page, lru);
1497 list_del(&page->lru);
1500 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1502 * __GFP_NOFAIL is not to be used in new code.
1504 * All __GFP_NOFAIL callers should be fixed so that they
1505 * properly detect and handle allocation failures.
1507 * We most definitely don't want callers attempting to
1508 * allocate greater than order-1 page units with
1511 WARN_ON_ONCE(order > 1);
1513 spin_lock_irqsave(&zone->lock, flags);
1514 page = __rmqueue(zone, order, migratetype);
1515 spin_unlock(&zone->lock);
1518 __mod_zone_freepage_state(zone, -(1 << order),
1519 get_pageblock_migratetype(page));
1522 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1523 zone_statistics(preferred_zone, zone, gfp_flags);
1524 local_irq_restore(flags);
1526 VM_BUG_ON(bad_range(zone, page));
1527 if (prep_new_page(page, order, gfp_flags))
1532 local_irq_restore(flags);
1536 #ifdef CONFIG_FAIL_PAGE_ALLOC
1539 struct fault_attr attr;
1541 u32 ignore_gfp_highmem;
1542 u32 ignore_gfp_wait;
1544 } fail_page_alloc = {
1545 .attr = FAULT_ATTR_INITIALIZER,
1546 .ignore_gfp_wait = 1,
1547 .ignore_gfp_highmem = 1,
1551 static int __init setup_fail_page_alloc(char *str)
1553 return setup_fault_attr(&fail_page_alloc.attr, str);
1555 __setup("fail_page_alloc=", setup_fail_page_alloc);
1557 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1559 if (order < fail_page_alloc.min_order)
1561 if (gfp_mask & __GFP_NOFAIL)
1563 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1565 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1568 return should_fail(&fail_page_alloc.attr, 1 << order);
1571 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1573 static int __init fail_page_alloc_debugfs(void)
1575 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1578 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1579 &fail_page_alloc.attr);
1581 return PTR_ERR(dir);
1583 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1584 &fail_page_alloc.ignore_gfp_wait))
1586 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1587 &fail_page_alloc.ignore_gfp_highmem))
1589 if (!debugfs_create_u32("min-order", mode, dir,
1590 &fail_page_alloc.min_order))
1595 debugfs_remove_recursive(dir);
1600 late_initcall(fail_page_alloc_debugfs);
1602 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1604 #else /* CONFIG_FAIL_PAGE_ALLOC */
1606 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1611 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1614 * Return true if free pages are above 'mark'. This takes into account the order
1615 * of the allocation.
1617 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1618 int classzone_idx, int alloc_flags, long free_pages)
1620 /* free_pages my go negative - that's OK */
1622 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1625 free_pages -= (1 << order) - 1;
1626 if (alloc_flags & ALLOC_HIGH)
1628 if (alloc_flags & ALLOC_HARDER)
1631 /* If allocation can't use CMA areas don't use free CMA pages */
1632 if (!(alloc_flags & ALLOC_CMA))
1633 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1635 if (free_pages <= min + lowmem_reserve)
1637 for (o = 0; o < order; o++) {
1638 /* At the next order, this order's pages become unavailable */
1639 free_pages -= z->free_area[o].nr_free << o;
1641 /* Require fewer higher order pages to be free */
1644 if (free_pages <= min)
1650 #ifdef CONFIG_MEMORY_ISOLATION
1651 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1653 if (unlikely(zone->nr_pageblock_isolate))
1654 return zone->nr_pageblock_isolate * pageblock_nr_pages;
1658 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1664 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1665 int classzone_idx, int alloc_flags)
1667 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1668 zone_page_state(z, NR_FREE_PAGES));
1671 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1672 int classzone_idx, int alloc_flags)
1674 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1676 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1677 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1680 * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1681 * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1682 * sleep although it could do so. But this is more desirable for memory
1683 * hotplug than sleeping which can cause a livelock in the direct
1686 free_pages -= nr_zone_isolate_freepages(z);
1687 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1693 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1694 * skip over zones that are not allowed by the cpuset, or that have
1695 * been recently (in last second) found to be nearly full. See further
1696 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1697 * that have to skip over a lot of full or unallowed zones.
1699 * If the zonelist cache is present in the passed in zonelist, then
1700 * returns a pointer to the allowed node mask (either the current
1701 * tasks mems_allowed, or node_states[N_MEMORY].)
1703 * If the zonelist cache is not available for this zonelist, does
1704 * nothing and returns NULL.
1706 * If the fullzones BITMAP in the zonelist cache is stale (more than
1707 * a second since last zap'd) then we zap it out (clear its bits.)
1709 * We hold off even calling zlc_setup, until after we've checked the
1710 * first zone in the zonelist, on the theory that most allocations will
1711 * be satisfied from that first zone, so best to examine that zone as
1712 * quickly as we can.
1714 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1716 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1717 nodemask_t *allowednodes; /* zonelist_cache approximation */
1719 zlc = zonelist->zlcache_ptr;
1723 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1724 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1725 zlc->last_full_zap = jiffies;
1728 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1729 &cpuset_current_mems_allowed :
1730 &node_states[N_MEMORY];
1731 return allowednodes;
1735 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1736 * if it is worth looking at further for free memory:
1737 * 1) Check that the zone isn't thought to be full (doesn't have its
1738 * bit set in the zonelist_cache fullzones BITMAP).
1739 * 2) Check that the zones node (obtained from the zonelist_cache
1740 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1741 * Return true (non-zero) if zone is worth looking at further, or
1742 * else return false (zero) if it is not.
1744 * This check -ignores- the distinction between various watermarks,
1745 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1746 * found to be full for any variation of these watermarks, it will
1747 * be considered full for up to one second by all requests, unless
1748 * we are so low on memory on all allowed nodes that we are forced
1749 * into the second scan of the zonelist.
1751 * In the second scan we ignore this zonelist cache and exactly
1752 * apply the watermarks to all zones, even it is slower to do so.
1753 * We are low on memory in the second scan, and should leave no stone
1754 * unturned looking for a free page.
1756 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1757 nodemask_t *allowednodes)
1759 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1760 int i; /* index of *z in zonelist zones */
1761 int n; /* node that zone *z is on */
1763 zlc = zonelist->zlcache_ptr;
1767 i = z - zonelist->_zonerefs;
1770 /* This zone is worth trying if it is allowed but not full */
1771 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1775 * Given 'z' scanning a zonelist, set the corresponding bit in
1776 * zlc->fullzones, so that subsequent attempts to allocate a page
1777 * from that zone don't waste time re-examining it.
1779 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1781 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1782 int i; /* index of *z in zonelist zones */
1784 zlc = zonelist->zlcache_ptr;
1788 i = z - zonelist->_zonerefs;
1790 set_bit(i, zlc->fullzones);
1794 * clear all zones full, called after direct reclaim makes progress so that
1795 * a zone that was recently full is not skipped over for up to a second
1797 static void zlc_clear_zones_full(struct zonelist *zonelist)
1799 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1801 zlc = zonelist->zlcache_ptr;
1805 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1808 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1810 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1813 static void __paginginit init_zone_allows_reclaim(int nid)
1817 for_each_online_node(i)
1818 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1819 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1821 zone_reclaim_mode = 1;
1824 #else /* CONFIG_NUMA */
1826 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1831 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1832 nodemask_t *allowednodes)
1837 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1841 static void zlc_clear_zones_full(struct zonelist *zonelist)
1845 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1850 static inline void init_zone_allows_reclaim(int nid)
1853 #endif /* CONFIG_NUMA */
1856 * get_page_from_freelist goes through the zonelist trying to allocate
1859 static struct page *
1860 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1861 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1862 struct zone *preferred_zone, int migratetype)
1865 struct page *page = NULL;
1868 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1869 int zlc_active = 0; /* set if using zonelist_cache */
1870 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1872 classzone_idx = zone_idx(preferred_zone);
1875 * Scan zonelist, looking for a zone with enough free.
1876 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1878 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1879 high_zoneidx, nodemask) {
1880 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1881 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1883 if ((alloc_flags & ALLOC_CPUSET) &&
1884 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1887 * When allocating a page cache page for writing, we
1888 * want to get it from a zone that is within its dirty
1889 * limit, such that no single zone holds more than its
1890 * proportional share of globally allowed dirty pages.
1891 * The dirty limits take into account the zone's
1892 * lowmem reserves and high watermark so that kswapd
1893 * should be able to balance it without having to
1894 * write pages from its LRU list.
1896 * This may look like it could increase pressure on
1897 * lower zones by failing allocations in higher zones
1898 * before they are full. But the pages that do spill
1899 * over are limited as the lower zones are protected
1900 * by this very same mechanism. It should not become
1901 * a practical burden to them.
1903 * XXX: For now, allow allocations to potentially
1904 * exceed the per-zone dirty limit in the slowpath
1905 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1906 * which is important when on a NUMA setup the allowed
1907 * zones are together not big enough to reach the
1908 * global limit. The proper fix for these situations
1909 * will require awareness of zones in the
1910 * dirty-throttling and the flusher threads.
1912 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1913 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1914 goto this_zone_full;
1916 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1917 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1921 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1922 if (zone_watermark_ok(zone, order, mark,
1923 classzone_idx, alloc_flags))
1926 if (IS_ENABLED(CONFIG_NUMA) &&
1927 !did_zlc_setup && nr_online_nodes > 1) {
1929 * we do zlc_setup if there are multiple nodes
1930 * and before considering the first zone allowed
1933 allowednodes = zlc_setup(zonelist, alloc_flags);
1938 if (zone_reclaim_mode == 0 ||
1939 !zone_allows_reclaim(preferred_zone, zone))
1940 goto this_zone_full;
1943 * As we may have just activated ZLC, check if the first
1944 * eligible zone has failed zone_reclaim recently.
1946 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1947 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1950 ret = zone_reclaim(zone, gfp_mask, order);
1952 case ZONE_RECLAIM_NOSCAN:
1955 case ZONE_RECLAIM_FULL:
1956 /* scanned but unreclaimable */
1959 /* did we reclaim enough */
1960 if (!zone_watermark_ok(zone, order, mark,
1961 classzone_idx, alloc_flags))
1962 goto this_zone_full;
1967 page = buffered_rmqueue(preferred_zone, zone, order,
1968 gfp_mask, migratetype);
1972 if (IS_ENABLED(CONFIG_NUMA))
1973 zlc_mark_zone_full(zonelist, z);
1976 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1977 /* Disable zlc cache for second zonelist scan */
1984 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1985 * necessary to allocate the page. The expectation is
1986 * that the caller is taking steps that will free more
1987 * memory. The caller should avoid the page being used
1988 * for !PFMEMALLOC purposes.
1990 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1996 * Large machines with many possible nodes should not always dump per-node
1997 * meminfo in irq context.
1999 static inline bool should_suppress_show_mem(void)
2004 ret = in_interrupt();
2009 static DEFINE_RATELIMIT_STATE(nopage_rs,
2010 DEFAULT_RATELIMIT_INTERVAL,
2011 DEFAULT_RATELIMIT_BURST);
2013 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2015 unsigned int filter = SHOW_MEM_FILTER_NODES;
2017 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2018 debug_guardpage_minorder() > 0)
2022 * This documents exceptions given to allocations in certain
2023 * contexts that are allowed to allocate outside current's set
2026 if (!(gfp_mask & __GFP_NOMEMALLOC))
2027 if (test_thread_flag(TIF_MEMDIE) ||
2028 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2029 filter &= ~SHOW_MEM_FILTER_NODES;
2030 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2031 filter &= ~SHOW_MEM_FILTER_NODES;
2034 struct va_format vaf;
2037 va_start(args, fmt);
2042 pr_warn("%pV", &vaf);
2047 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2048 current->comm, order, gfp_mask);
2051 if (!should_suppress_show_mem())
2056 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2057 unsigned long did_some_progress,
2058 unsigned long pages_reclaimed)
2060 /* Do not loop if specifically requested */
2061 if (gfp_mask & __GFP_NORETRY)
2064 /* Always retry if specifically requested */
2065 if (gfp_mask & __GFP_NOFAIL)
2069 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2070 * making forward progress without invoking OOM. Suspend also disables
2071 * storage devices so kswapd will not help. Bail if we are suspending.
2073 if (!did_some_progress && pm_suspended_storage())
2077 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2078 * means __GFP_NOFAIL, but that may not be true in other
2081 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2085 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2086 * specified, then we retry until we no longer reclaim any pages
2087 * (above), or we've reclaimed an order of pages at least as
2088 * large as the allocation's order. In both cases, if the
2089 * allocation still fails, we stop retrying.
2091 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2097 static inline struct page *
2098 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2099 struct zonelist *zonelist, enum zone_type high_zoneidx,
2100 nodemask_t *nodemask, struct zone *preferred_zone,
2105 /* Acquire the OOM killer lock for the zones in zonelist */
2106 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2107 schedule_timeout_uninterruptible(1);
2112 * Go through the zonelist yet one more time, keep very high watermark
2113 * here, this is only to catch a parallel oom killing, we must fail if
2114 * we're still under heavy pressure.
2116 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2117 order, zonelist, high_zoneidx,
2118 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2119 preferred_zone, migratetype);
2123 if (!(gfp_mask & __GFP_NOFAIL)) {
2124 /* The OOM killer will not help higher order allocs */
2125 if (order > PAGE_ALLOC_COSTLY_ORDER)
2127 /* The OOM killer does not needlessly kill tasks for lowmem */
2128 if (high_zoneidx < ZONE_NORMAL)
2131 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2132 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2133 * The caller should handle page allocation failure by itself if
2134 * it specifies __GFP_THISNODE.
2135 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2137 if (gfp_mask & __GFP_THISNODE)
2140 /* Exhausted what can be done so it's blamo time */
2141 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2144 clear_zonelist_oom(zonelist, gfp_mask);
2148 #ifdef CONFIG_COMPACTION
2149 /* Try memory compaction for high-order allocations before reclaim */
2150 static struct page *
2151 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2152 struct zonelist *zonelist, enum zone_type high_zoneidx,
2153 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2154 int migratetype, bool sync_migration,
2155 bool *contended_compaction, bool *deferred_compaction,
2156 unsigned long *did_some_progress)
2158 struct page *page = NULL;
2163 if (compaction_deferred(preferred_zone, order)) {
2164 *deferred_compaction = true;
2168 current->flags |= PF_MEMALLOC;
2169 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2170 nodemask, sync_migration,
2171 contended_compaction, &page);
2172 current->flags &= ~PF_MEMALLOC;
2174 /* If compaction captured a page, prep and use it */
2176 prep_new_page(page, order, gfp_mask);
2180 if (*did_some_progress != COMPACT_SKIPPED) {
2181 /* Page migration frees to the PCP lists but we want merging */
2182 drain_pages(get_cpu());
2185 page = get_page_from_freelist(gfp_mask, nodemask,
2186 order, zonelist, high_zoneidx,
2187 alloc_flags & ~ALLOC_NO_WATERMARKS,
2188 preferred_zone, migratetype);
2191 preferred_zone->compact_blockskip_flush = false;
2192 preferred_zone->compact_considered = 0;
2193 preferred_zone->compact_defer_shift = 0;
2194 if (order >= preferred_zone->compact_order_failed)
2195 preferred_zone->compact_order_failed = order + 1;
2196 count_vm_event(COMPACTSUCCESS);
2201 * It's bad if compaction run occurs and fails.
2202 * The most likely reason is that pages exist,
2203 * but not enough to satisfy watermarks.
2205 count_vm_event(COMPACTFAIL);
2208 * As async compaction considers a subset of pageblocks, only
2209 * defer if the failure was a sync compaction failure.
2212 defer_compaction(preferred_zone, order);
2220 static inline struct page *
2221 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2222 struct zonelist *zonelist, enum zone_type high_zoneidx,
2223 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2224 int migratetype, bool sync_migration,
2225 bool *contended_compaction, bool *deferred_compaction,
2226 unsigned long *did_some_progress)
2230 #endif /* CONFIG_COMPACTION */
2232 /* Perform direct synchronous page reclaim */
2234 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2235 nodemask_t *nodemask)
2237 struct reclaim_state reclaim_state;
2242 /* We now go into synchronous reclaim */
2243 cpuset_memory_pressure_bump();
2244 current->flags |= PF_MEMALLOC;
2245 lockdep_set_current_reclaim_state(gfp_mask);
2246 reclaim_state.reclaimed_slab = 0;
2247 current->reclaim_state = &reclaim_state;
2249 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2251 current->reclaim_state = NULL;
2252 lockdep_clear_current_reclaim_state();
2253 current->flags &= ~PF_MEMALLOC;
2260 /* The really slow allocator path where we enter direct reclaim */
2261 static inline struct page *
2262 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2263 struct zonelist *zonelist, enum zone_type high_zoneidx,
2264 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2265 int migratetype, unsigned long *did_some_progress)
2267 struct page *page = NULL;
2268 bool drained = false;
2270 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2272 if (unlikely(!(*did_some_progress)))
2275 /* After successful reclaim, reconsider all zones for allocation */
2276 if (IS_ENABLED(CONFIG_NUMA))
2277 zlc_clear_zones_full(zonelist);
2280 page = get_page_from_freelist(gfp_mask, nodemask, order,
2281 zonelist, high_zoneidx,
2282 alloc_flags & ~ALLOC_NO_WATERMARKS,
2283 preferred_zone, migratetype);
2286 * If an allocation failed after direct reclaim, it could be because
2287 * pages are pinned on the per-cpu lists. Drain them and try again
2289 if (!page && !drained) {
2299 * This is called in the allocator slow-path if the allocation request is of
2300 * sufficient urgency to ignore watermarks and take other desperate measures
2302 static inline struct page *
2303 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2304 struct zonelist *zonelist, enum zone_type high_zoneidx,
2305 nodemask_t *nodemask, struct zone *preferred_zone,
2311 page = get_page_from_freelist(gfp_mask, nodemask, order,
2312 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2313 preferred_zone, migratetype);
2315 if (!page && gfp_mask & __GFP_NOFAIL)
2316 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2317 } while (!page && (gfp_mask & __GFP_NOFAIL));
2323 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2324 enum zone_type high_zoneidx,
2325 enum zone_type classzone_idx)
2330 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2331 wakeup_kswapd(zone, order, classzone_idx);
2335 gfp_to_alloc_flags(gfp_t gfp_mask)
2337 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2338 const gfp_t wait = gfp_mask & __GFP_WAIT;
2340 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2341 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2344 * The caller may dip into page reserves a bit more if the caller
2345 * cannot run direct reclaim, or if the caller has realtime scheduling
2346 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2347 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2349 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2353 * Not worth trying to allocate harder for
2354 * __GFP_NOMEMALLOC even if it can't schedule.
2356 if (!(gfp_mask & __GFP_NOMEMALLOC))
2357 alloc_flags |= ALLOC_HARDER;
2359 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2360 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2362 alloc_flags &= ~ALLOC_CPUSET;
2363 } else if (unlikely(rt_task(current)) && !in_interrupt())
2364 alloc_flags |= ALLOC_HARDER;
2366 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2367 if (gfp_mask & __GFP_MEMALLOC)
2368 alloc_flags |= ALLOC_NO_WATERMARKS;
2369 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2370 alloc_flags |= ALLOC_NO_WATERMARKS;
2371 else if (!in_interrupt() &&
2372 ((current->flags & PF_MEMALLOC) ||
2373 unlikely(test_thread_flag(TIF_MEMDIE))))
2374 alloc_flags |= ALLOC_NO_WATERMARKS;
2377 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2378 alloc_flags |= ALLOC_CMA;
2383 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2385 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2388 static inline struct page *
2389 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2390 struct zonelist *zonelist, enum zone_type high_zoneidx,
2391 nodemask_t *nodemask, struct zone *preferred_zone,
2394 const gfp_t wait = gfp_mask & __GFP_WAIT;
2395 struct page *page = NULL;
2397 unsigned long pages_reclaimed = 0;
2398 unsigned long did_some_progress;
2399 bool sync_migration = false;
2400 bool deferred_compaction = false;
2401 bool contended_compaction = false;
2404 * In the slowpath, we sanity check order to avoid ever trying to
2405 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2406 * be using allocators in order of preference for an area that is
2409 if (order >= MAX_ORDER) {
2410 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2415 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2416 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2417 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2418 * using a larger set of nodes after it has established that the
2419 * allowed per node queues are empty and that nodes are
2422 if (IS_ENABLED(CONFIG_NUMA) &&
2423 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2427 if (!(gfp_mask & __GFP_NO_KSWAPD))
2428 wake_all_kswapd(order, zonelist, high_zoneidx,
2429 zone_idx(preferred_zone));
2432 * OK, we're below the kswapd watermark and have kicked background
2433 * reclaim. Now things get more complex, so set up alloc_flags according
2434 * to how we want to proceed.
2436 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2439 * Find the true preferred zone if the allocation is unconstrained by
2442 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2443 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2447 /* This is the last chance, in general, before the goto nopage. */
2448 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2449 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2450 preferred_zone, migratetype);
2454 /* Allocate without watermarks if the context allows */
2455 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2457 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2458 * the allocation is high priority and these type of
2459 * allocations are system rather than user orientated
2461 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2463 page = __alloc_pages_high_priority(gfp_mask, order,
2464 zonelist, high_zoneidx, nodemask,
2465 preferred_zone, migratetype);
2471 /* Atomic allocations - we can't balance anything */
2475 /* Avoid recursion of direct reclaim */
2476 if (current->flags & PF_MEMALLOC)
2479 /* Avoid allocations with no watermarks from looping endlessly */
2480 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2484 * Try direct compaction. The first pass is asynchronous. Subsequent
2485 * attempts after direct reclaim are synchronous
2487 page = __alloc_pages_direct_compact(gfp_mask, order,
2488 zonelist, high_zoneidx,
2490 alloc_flags, preferred_zone,
2491 migratetype, sync_migration,
2492 &contended_compaction,
2493 &deferred_compaction,
2494 &did_some_progress);
2497 sync_migration = true;
2500 * If compaction is deferred for high-order allocations, it is because
2501 * sync compaction recently failed. In this is the case and the caller
2502 * requested a movable allocation that does not heavily disrupt the
2503 * system then fail the allocation instead of entering direct reclaim.
2505 if ((deferred_compaction || contended_compaction) &&
2506 (gfp_mask & __GFP_NO_KSWAPD))
2509 /* Try direct reclaim and then allocating */
2510 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2511 zonelist, high_zoneidx,
2513 alloc_flags, preferred_zone,
2514 migratetype, &did_some_progress);
2519 * If we failed to make any progress reclaiming, then we are
2520 * running out of options and have to consider going OOM
2522 if (!did_some_progress) {
2523 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2524 if (oom_killer_disabled)
2526 /* Coredumps can quickly deplete all memory reserves */
2527 if ((current->flags & PF_DUMPCORE) &&
2528 !(gfp_mask & __GFP_NOFAIL))
2530 page = __alloc_pages_may_oom(gfp_mask, order,
2531 zonelist, high_zoneidx,
2532 nodemask, preferred_zone,
2537 if (!(gfp_mask & __GFP_NOFAIL)) {
2539 * The oom killer is not called for high-order
2540 * allocations that may fail, so if no progress
2541 * is being made, there are no other options and
2542 * retrying is unlikely to help.
2544 if (order > PAGE_ALLOC_COSTLY_ORDER)
2547 * The oom killer is not called for lowmem
2548 * allocations to prevent needlessly killing
2551 if (high_zoneidx < ZONE_NORMAL)
2559 /* Check if we should retry the allocation */
2560 pages_reclaimed += did_some_progress;
2561 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2563 /* Wait for some write requests to complete then retry */
2564 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2568 * High-order allocations do not necessarily loop after
2569 * direct reclaim and reclaim/compaction depends on compaction
2570 * being called after reclaim so call directly if necessary
2572 page = __alloc_pages_direct_compact(gfp_mask, order,
2573 zonelist, high_zoneidx,
2575 alloc_flags, preferred_zone,
2576 migratetype, sync_migration,
2577 &contended_compaction,
2578 &deferred_compaction,
2579 &did_some_progress);
2585 warn_alloc_failed(gfp_mask, order, NULL);
2588 if (kmemcheck_enabled)
2589 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2595 * This is the 'heart' of the zoned buddy allocator.
2598 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2599 struct zonelist *zonelist, nodemask_t *nodemask)
2601 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2602 struct zone *preferred_zone;
2603 struct page *page = NULL;
2604 int migratetype = allocflags_to_migratetype(gfp_mask);
2605 unsigned int cpuset_mems_cookie;
2606 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2608 gfp_mask &= gfp_allowed_mask;
2610 lockdep_trace_alloc(gfp_mask);
2612 might_sleep_if(gfp_mask & __GFP_WAIT);
2614 if (should_fail_alloc_page(gfp_mask, order))
2618 * Check the zones suitable for the gfp_mask contain at least one
2619 * valid zone. It's possible to have an empty zonelist as a result
2620 * of GFP_THISNODE and a memoryless node
2622 if (unlikely(!zonelist->_zonerefs->zone))
2626 cpuset_mems_cookie = get_mems_allowed();
2628 /* The preferred zone is used for statistics later */
2629 first_zones_zonelist(zonelist, high_zoneidx,
2630 nodemask ? : &cpuset_current_mems_allowed,
2632 if (!preferred_zone)
2636 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2637 alloc_flags |= ALLOC_CMA;
2639 /* First allocation attempt */
2640 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2641 zonelist, high_zoneidx, alloc_flags,
2642 preferred_zone, migratetype);
2643 if (unlikely(!page))
2644 page = __alloc_pages_slowpath(gfp_mask, order,
2645 zonelist, high_zoneidx, nodemask,
2646 preferred_zone, migratetype);
2648 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2652 * When updating a task's mems_allowed, it is possible to race with
2653 * parallel threads in such a way that an allocation can fail while
2654 * the mask is being updated. If a page allocation is about to fail,
2655 * check if the cpuset changed during allocation and if so, retry.
2657 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2662 EXPORT_SYMBOL(__alloc_pages_nodemask);
2665 * Common helper functions.
2667 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2672 * __get_free_pages() returns a 32-bit address, which cannot represent
2675 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2677 page = alloc_pages(gfp_mask, order);
2680 return (unsigned long) page_address(page);
2682 EXPORT_SYMBOL(__get_free_pages);
2684 unsigned long get_zeroed_page(gfp_t gfp_mask)
2686 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2688 EXPORT_SYMBOL(get_zeroed_page);
2690 void __free_pages(struct page *page, unsigned int order)
2692 if (put_page_testzero(page)) {
2694 free_hot_cold_page(page, 0);
2696 __free_pages_ok(page, order);
2700 EXPORT_SYMBOL(__free_pages);
2702 void free_pages(unsigned long addr, unsigned int order)
2705 VM_BUG_ON(!virt_addr_valid((void *)addr));
2706 __free_pages(virt_to_page((void *)addr), order);
2710 EXPORT_SYMBOL(free_pages);
2712 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2715 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2716 unsigned long used = addr + PAGE_ALIGN(size);
2718 split_page(virt_to_page((void *)addr), order);
2719 while (used < alloc_end) {
2724 return (void *)addr;
2728 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2729 * @size: the number of bytes to allocate
2730 * @gfp_mask: GFP flags for the allocation
2732 * This function is similar to alloc_pages(), except that it allocates the
2733 * minimum number of pages to satisfy the request. alloc_pages() can only
2734 * allocate memory in power-of-two pages.
2736 * This function is also limited by MAX_ORDER.
2738 * Memory allocated by this function must be released by free_pages_exact().
2740 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2742 unsigned int order = get_order(size);
2745 addr = __get_free_pages(gfp_mask, order);
2746 return make_alloc_exact(addr, order, size);
2748 EXPORT_SYMBOL(alloc_pages_exact);
2751 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2753 * @nid: the preferred node ID where memory should be allocated
2754 * @size: the number of bytes to allocate
2755 * @gfp_mask: GFP flags for the allocation
2757 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2759 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2762 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2764 unsigned order = get_order(size);
2765 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2768 return make_alloc_exact((unsigned long)page_address(p), order, size);
2770 EXPORT_SYMBOL(alloc_pages_exact_nid);
2773 * free_pages_exact - release memory allocated via alloc_pages_exact()
2774 * @virt: the value returned by alloc_pages_exact.
2775 * @size: size of allocation, same value as passed to alloc_pages_exact().
2777 * Release the memory allocated by a previous call to alloc_pages_exact.
2779 void free_pages_exact(void *virt, size_t size)
2781 unsigned long addr = (unsigned long)virt;
2782 unsigned long end = addr + PAGE_ALIGN(size);
2784 while (addr < end) {
2789 EXPORT_SYMBOL(free_pages_exact);
2791 static unsigned int nr_free_zone_pages(int offset)
2796 /* Just pick one node, since fallback list is circular */
2797 unsigned int sum = 0;
2799 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2801 for_each_zone_zonelist(zone, z, zonelist, offset) {
2802 unsigned long size = zone->present_pages;
2803 unsigned long high = high_wmark_pages(zone);
2812 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2814 unsigned int nr_free_buffer_pages(void)
2816 return nr_free_zone_pages(gfp_zone(GFP_USER));
2818 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2821 * Amount of free RAM allocatable within all zones
2823 unsigned int nr_free_pagecache_pages(void)
2825 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2828 static inline void show_node(struct zone *zone)
2830 if (IS_ENABLED(CONFIG_NUMA))
2831 printk("Node %d ", zone_to_nid(zone));
2834 void si_meminfo(struct sysinfo *val)
2836 val->totalram = totalram_pages;
2838 val->freeram = global_page_state(NR_FREE_PAGES);
2839 val->bufferram = nr_blockdev_pages();
2840 val->totalhigh = totalhigh_pages;
2841 val->freehigh = nr_free_highpages();
2842 val->mem_unit = PAGE_SIZE;
2845 EXPORT_SYMBOL(si_meminfo);
2848 void si_meminfo_node(struct sysinfo *val, int nid)
2850 pg_data_t *pgdat = NODE_DATA(nid);
2852 val->totalram = pgdat->node_present_pages;
2853 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2854 #ifdef CONFIG_HIGHMEM
2855 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2856 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2862 val->mem_unit = PAGE_SIZE;
2867 * Determine whether the node should be displayed or not, depending on whether
2868 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2870 bool skip_free_areas_node(unsigned int flags, int nid)
2873 unsigned int cpuset_mems_cookie;
2875 if (!(flags & SHOW_MEM_FILTER_NODES))
2879 cpuset_mems_cookie = get_mems_allowed();
2880 ret = !node_isset(nid, cpuset_current_mems_allowed);
2881 } while (!put_mems_allowed(cpuset_mems_cookie));
2886 #define K(x) ((x) << (PAGE_SHIFT-10))
2888 static void show_migration_types(unsigned char type)
2890 static const char types[MIGRATE_TYPES] = {
2891 [MIGRATE_UNMOVABLE] = 'U',
2892 [MIGRATE_RECLAIMABLE] = 'E',
2893 [MIGRATE_MOVABLE] = 'M',
2894 [MIGRATE_RESERVE] = 'R',
2896 [MIGRATE_CMA] = 'C',
2898 [MIGRATE_ISOLATE] = 'I',
2900 char tmp[MIGRATE_TYPES + 1];
2904 for (i = 0; i < MIGRATE_TYPES; i++) {
2905 if (type & (1 << i))
2910 printk("(%s) ", tmp);
2914 * Show free area list (used inside shift_scroll-lock stuff)
2915 * We also calculate the percentage fragmentation. We do this by counting the
2916 * memory on each free list with the exception of the first item on the list.
2917 * Suppresses nodes that are not allowed by current's cpuset if
2918 * SHOW_MEM_FILTER_NODES is passed.
2920 void show_free_areas(unsigned int filter)
2925 for_each_populated_zone(zone) {
2926 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2929 printk("%s per-cpu:\n", zone->name);
2931 for_each_online_cpu(cpu) {
2932 struct per_cpu_pageset *pageset;
2934 pageset = per_cpu_ptr(zone->pageset, cpu);
2936 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2937 cpu, pageset->pcp.high,
2938 pageset->pcp.batch, pageset->pcp.count);
2942 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2943 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2945 " dirty:%lu writeback:%lu unstable:%lu\n"
2946 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2947 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2949 global_page_state(NR_ACTIVE_ANON),
2950 global_page_state(NR_INACTIVE_ANON),
2951 global_page_state(NR_ISOLATED_ANON),
2952 global_page_state(NR_ACTIVE_FILE),
2953 global_page_state(NR_INACTIVE_FILE),
2954 global_page_state(NR_ISOLATED_FILE),
2955 global_page_state(NR_UNEVICTABLE),
2956 global_page_state(NR_FILE_DIRTY),
2957 global_page_state(NR_WRITEBACK),
2958 global_page_state(NR_UNSTABLE_NFS),
2959 global_page_state(NR_FREE_PAGES),
2960 global_page_state(NR_SLAB_RECLAIMABLE),
2961 global_page_state(NR_SLAB_UNRECLAIMABLE),
2962 global_page_state(NR_FILE_MAPPED),
2963 global_page_state(NR_SHMEM),
2964 global_page_state(NR_PAGETABLE),
2965 global_page_state(NR_BOUNCE),
2966 global_page_state(NR_FREE_CMA_PAGES));
2968 for_each_populated_zone(zone) {
2971 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2979 " active_anon:%lukB"
2980 " inactive_anon:%lukB"
2981 " active_file:%lukB"
2982 " inactive_file:%lukB"
2983 " unevictable:%lukB"
2984 " isolated(anon):%lukB"
2985 " isolated(file):%lukB"
2992 " slab_reclaimable:%lukB"
2993 " slab_unreclaimable:%lukB"
2994 " kernel_stack:%lukB"
2999 " writeback_tmp:%lukB"
3000 " pages_scanned:%lu"
3001 " all_unreclaimable? %s"
3004 K(zone_page_state(zone, NR_FREE_PAGES)),
3005 K(min_wmark_pages(zone)),
3006 K(low_wmark_pages(zone)),
3007 K(high_wmark_pages(zone)),
3008 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3009 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3010 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3011 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3012 K(zone_page_state(zone, NR_UNEVICTABLE)),
3013 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3014 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3015 K(zone->present_pages),
3016 K(zone_page_state(zone, NR_MLOCK)),
3017 K(zone_page_state(zone, NR_FILE_DIRTY)),
3018 K(zone_page_state(zone, NR_WRITEBACK)),
3019 K(zone_page_state(zone, NR_FILE_MAPPED)),
3020 K(zone_page_state(zone, NR_SHMEM)),
3021 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3022 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3023 zone_page_state(zone, NR_KERNEL_STACK) *
3025 K(zone_page_state(zone, NR_PAGETABLE)),
3026 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3027 K(zone_page_state(zone, NR_BOUNCE)),
3028 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3029 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3030 zone->pages_scanned,
3031 (zone->all_unreclaimable ? "yes" : "no")
3033 printk("lowmem_reserve[]:");
3034 for (i = 0; i < MAX_NR_ZONES; i++)
3035 printk(" %lu", zone->lowmem_reserve[i]);
3039 for_each_populated_zone(zone) {
3040 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3041 unsigned char types[MAX_ORDER];
3043 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3046 printk("%s: ", zone->name);
3048 spin_lock_irqsave(&zone->lock, flags);
3049 for (order = 0; order < MAX_ORDER; order++) {
3050 struct free_area *area = &zone->free_area[order];
3053 nr[order] = area->nr_free;
3054 total += nr[order] << order;
3057 for (type = 0; type < MIGRATE_TYPES; type++) {
3058 if (!list_empty(&area->free_list[type]))
3059 types[order] |= 1 << type;
3062 spin_unlock_irqrestore(&zone->lock, flags);
3063 for (order = 0; order < MAX_ORDER; order++) {
3064 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3066 show_migration_types(types[order]);
3068 printk("= %lukB\n", K(total));
3071 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3073 show_swap_cache_info();
3076 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3078 zoneref->zone = zone;
3079 zoneref->zone_idx = zone_idx(zone);
3083 * Builds allocation fallback zone lists.
3085 * Add all populated zones of a node to the zonelist.
3087 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3088 int nr_zones, enum zone_type zone_type)
3092 BUG_ON(zone_type >= MAX_NR_ZONES);
3097 zone = pgdat->node_zones + zone_type;
3098 if (populated_zone(zone)) {
3099 zoneref_set_zone(zone,
3100 &zonelist->_zonerefs[nr_zones++]);
3101 check_highest_zone(zone_type);
3104 } while (zone_type);
3111 * 0 = automatic detection of better ordering.
3112 * 1 = order by ([node] distance, -zonetype)
3113 * 2 = order by (-zonetype, [node] distance)
3115 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3116 * the same zonelist. So only NUMA can configure this param.
3118 #define ZONELIST_ORDER_DEFAULT 0
3119 #define ZONELIST_ORDER_NODE 1
3120 #define ZONELIST_ORDER_ZONE 2
3122 /* zonelist order in the kernel.
3123 * set_zonelist_order() will set this to NODE or ZONE.
3125 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3126 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3130 /* The value user specified ....changed by config */
3131 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3132 /* string for sysctl */
3133 #define NUMA_ZONELIST_ORDER_LEN 16
3134 char numa_zonelist_order[16] = "default";
3137 * interface for configure zonelist ordering.
3138 * command line option "numa_zonelist_order"
3139 * = "[dD]efault - default, automatic configuration.
3140 * = "[nN]ode - order by node locality, then by zone within node
3141 * = "[zZ]one - order by zone, then by locality within zone
3144 static int __parse_numa_zonelist_order(char *s)
3146 if (*s == 'd' || *s == 'D') {
3147 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3148 } else if (*s == 'n' || *s == 'N') {
3149 user_zonelist_order = ZONELIST_ORDER_NODE;
3150 } else if (*s == 'z' || *s == 'Z') {
3151 user_zonelist_order = ZONELIST_ORDER_ZONE;
3154 "Ignoring invalid numa_zonelist_order value: "
3161 static __init int setup_numa_zonelist_order(char *s)
3168 ret = __parse_numa_zonelist_order(s);
3170 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3174 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3177 * sysctl handler for numa_zonelist_order
3179 int numa_zonelist_order_handler(ctl_table *table, int write,
3180 void __user *buffer, size_t *length,
3183 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3185 static DEFINE_MUTEX(zl_order_mutex);
3187 mutex_lock(&zl_order_mutex);
3189 strcpy(saved_string, (char*)table->data);
3190 ret = proc_dostring(table, write, buffer, length, ppos);
3194 int oldval = user_zonelist_order;
3195 if (__parse_numa_zonelist_order((char*)table->data)) {
3197 * bogus value. restore saved string
3199 strncpy((char*)table->data, saved_string,
3200 NUMA_ZONELIST_ORDER_LEN);
3201 user_zonelist_order = oldval;
3202 } else if (oldval != user_zonelist_order) {
3203 mutex_lock(&zonelists_mutex);
3204 build_all_zonelists(NULL, NULL);
3205 mutex_unlock(&zonelists_mutex);
3209 mutex_unlock(&zl_order_mutex);
3214 #define MAX_NODE_LOAD (nr_online_nodes)
3215 static int node_load[MAX_NUMNODES];
3218 * find_next_best_node - find the next node that should appear in a given node's fallback list
3219 * @node: node whose fallback list we're appending
3220 * @used_node_mask: nodemask_t of already used nodes
3222 * We use a number of factors to determine which is the next node that should
3223 * appear on a given node's fallback list. The node should not have appeared
3224 * already in @node's fallback list, and it should be the next closest node
3225 * according to the distance array (which contains arbitrary distance values
3226 * from each node to each node in the system), and should also prefer nodes
3227 * with no CPUs, since presumably they'll have very little allocation pressure
3228 * on them otherwise.
3229 * It returns -1 if no node is found.
3231 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3234 int min_val = INT_MAX;
3236 const struct cpumask *tmp = cpumask_of_node(0);
3238 /* Use the local node if we haven't already */
3239 if (!node_isset(node, *used_node_mask)) {
3240 node_set(node, *used_node_mask);
3244 for_each_node_state(n, N_MEMORY) {
3246 /* Don't want a node to appear more than once */
3247 if (node_isset(n, *used_node_mask))
3250 /* Use the distance array to find the distance */
3251 val = node_distance(node, n);
3253 /* Penalize nodes under us ("prefer the next node") */
3256 /* Give preference to headless and unused nodes */
3257 tmp = cpumask_of_node(n);
3258 if (!cpumask_empty(tmp))
3259 val += PENALTY_FOR_NODE_WITH_CPUS;
3261 /* Slight preference for less loaded node */
3262 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3263 val += node_load[n];
3265 if (val < min_val) {
3272 node_set(best_node, *used_node_mask);
3279 * Build zonelists ordered by node and zones within node.
3280 * This results in maximum locality--normal zone overflows into local
3281 * DMA zone, if any--but risks exhausting DMA zone.
3283 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3286 struct zonelist *zonelist;
3288 zonelist = &pgdat->node_zonelists[0];
3289 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3291 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3293 zonelist->_zonerefs[j].zone = NULL;
3294 zonelist->_zonerefs[j].zone_idx = 0;
3298 * Build gfp_thisnode zonelists
3300 static void build_thisnode_zonelists(pg_data_t *pgdat)
3303 struct zonelist *zonelist;
3305 zonelist = &pgdat->node_zonelists[1];
3306 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3307 zonelist->_zonerefs[j].zone = NULL;
3308 zonelist->_zonerefs[j].zone_idx = 0;
3312 * Build zonelists ordered by zone and nodes within zones.
3313 * This results in conserving DMA zone[s] until all Normal memory is
3314 * exhausted, but results in overflowing to remote node while memory
3315 * may still exist in local DMA zone.
3317 static int node_order[MAX_NUMNODES];
3319 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3322 int zone_type; /* needs to be signed */
3324 struct zonelist *zonelist;
3326 zonelist = &pgdat->node_zonelists[0];
3328 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3329 for (j = 0; j < nr_nodes; j++) {
3330 node = node_order[j];
3331 z = &NODE_DATA(node)->node_zones[zone_type];
3332 if (populated_zone(z)) {
3334 &zonelist->_zonerefs[pos++]);
3335 check_highest_zone(zone_type);
3339 zonelist->_zonerefs[pos].zone = NULL;
3340 zonelist->_zonerefs[pos].zone_idx = 0;
3343 static int default_zonelist_order(void)
3346 unsigned long low_kmem_size,total_size;
3350 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3351 * If they are really small and used heavily, the system can fall
3352 * into OOM very easily.
3353 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3355 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3358 for_each_online_node(nid) {
3359 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3360 z = &NODE_DATA(nid)->node_zones[zone_type];
3361 if (populated_zone(z)) {
3362 if (zone_type < ZONE_NORMAL)
3363 low_kmem_size += z->present_pages;
3364 total_size += z->present_pages;
3365 } else if (zone_type == ZONE_NORMAL) {
3367 * If any node has only lowmem, then node order
3368 * is preferred to allow kernel allocations
3369 * locally; otherwise, they can easily infringe
3370 * on other nodes when there is an abundance of
3371 * lowmem available to allocate from.
3373 return ZONELIST_ORDER_NODE;
3377 if (!low_kmem_size || /* there are no DMA area. */
3378 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3379 return ZONELIST_ORDER_NODE;
3381 * look into each node's config.
3382 * If there is a node whose DMA/DMA32 memory is very big area on
3383 * local memory, NODE_ORDER may be suitable.
3385 average_size = total_size /
3386 (nodes_weight(node_states[N_MEMORY]) + 1);
3387 for_each_online_node(nid) {
3390 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3391 z = &NODE_DATA(nid)->node_zones[zone_type];
3392 if (populated_zone(z)) {
3393 if (zone_type < ZONE_NORMAL)
3394 low_kmem_size += z->present_pages;
3395 total_size += z->present_pages;
3398 if (low_kmem_size &&
3399 total_size > average_size && /* ignore small node */
3400 low_kmem_size > total_size * 70/100)
3401 return ZONELIST_ORDER_NODE;
3403 return ZONELIST_ORDER_ZONE;
3406 static void set_zonelist_order(void)
3408 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3409 current_zonelist_order = default_zonelist_order();
3411 current_zonelist_order = user_zonelist_order;
3414 static void build_zonelists(pg_data_t *pgdat)
3418 nodemask_t used_mask;
3419 int local_node, prev_node;
3420 struct zonelist *zonelist;
3421 int order = current_zonelist_order;
3423 /* initialize zonelists */
3424 for (i = 0; i < MAX_ZONELISTS; i++) {
3425 zonelist = pgdat->node_zonelists + i;
3426 zonelist->_zonerefs[0].zone = NULL;
3427 zonelist->_zonerefs[0].zone_idx = 0;
3430 /* NUMA-aware ordering of nodes */
3431 local_node = pgdat->node_id;
3432 load = nr_online_nodes;
3433 prev_node = local_node;
3434 nodes_clear(used_mask);
3436 memset(node_order, 0, sizeof(node_order));
3439 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3441 * We don't want to pressure a particular node.
3442 * So adding penalty to the first node in same
3443 * distance group to make it round-robin.
3445 if (node_distance(local_node, node) !=
3446 node_distance(local_node, prev_node))
3447 node_load[node] = load;
3451 if (order == ZONELIST_ORDER_NODE)
3452 build_zonelists_in_node_order(pgdat, node);
3454 node_order[j++] = node; /* remember order */
3457 if (order == ZONELIST_ORDER_ZONE) {
3458 /* calculate node order -- i.e., DMA last! */
3459 build_zonelists_in_zone_order(pgdat, j);
3462 build_thisnode_zonelists(pgdat);
3465 /* Construct the zonelist performance cache - see further mmzone.h */
3466 static void build_zonelist_cache(pg_data_t *pgdat)
3468 struct zonelist *zonelist;
3469 struct zonelist_cache *zlc;
3472 zonelist = &pgdat->node_zonelists[0];
3473 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3474 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3475 for (z = zonelist->_zonerefs; z->zone; z++)
3476 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3479 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3481 * Return node id of node used for "local" allocations.
3482 * I.e., first node id of first zone in arg node's generic zonelist.
3483 * Used for initializing percpu 'numa_mem', which is used primarily
3484 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3486 int local_memory_node(int node)
3490 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3491 gfp_zone(GFP_KERNEL),
3498 #else /* CONFIG_NUMA */
3500 static void set_zonelist_order(void)
3502 current_zonelist_order = ZONELIST_ORDER_ZONE;
3505 static void build_zonelists(pg_data_t *pgdat)
3507 int node, local_node;
3509 struct zonelist *zonelist;
3511 local_node = pgdat->node_id;
3513 zonelist = &pgdat->node_zonelists[0];
3514 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3517 * Now we build the zonelist so that it contains the zones
3518 * of all the other nodes.
3519 * We don't want to pressure a particular node, so when
3520 * building the zones for node N, we make sure that the
3521 * zones coming right after the local ones are those from
3522 * node N+1 (modulo N)
3524 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3525 if (!node_online(node))
3527 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3530 for (node = 0; node < local_node; node++) {
3531 if (!node_online(node))
3533 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3537 zonelist->_zonerefs[j].zone = NULL;
3538 zonelist->_zonerefs[j].zone_idx = 0;
3541 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3542 static void build_zonelist_cache(pg_data_t *pgdat)
3544 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3547 #endif /* CONFIG_NUMA */
3550 * Boot pageset table. One per cpu which is going to be used for all
3551 * zones and all nodes. The parameters will be set in such a way
3552 * that an item put on a list will immediately be handed over to
3553 * the buddy list. This is safe since pageset manipulation is done
3554 * with interrupts disabled.
3556 * The boot_pagesets must be kept even after bootup is complete for
3557 * unused processors and/or zones. They do play a role for bootstrapping
3558 * hotplugged processors.
3560 * zoneinfo_show() and maybe other functions do
3561 * not check if the processor is online before following the pageset pointer.
3562 * Other parts of the kernel may not check if the zone is available.
3564 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3565 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3566 static void setup_zone_pageset(struct zone *zone);
3569 * Global mutex to protect against size modification of zonelists
3570 * as well as to serialize pageset setup for the new populated zone.
3572 DEFINE_MUTEX(zonelists_mutex);
3574 /* return values int ....just for stop_machine() */
3575 static int __build_all_zonelists(void *data)
3579 pg_data_t *self = data;
3582 memset(node_load, 0, sizeof(node_load));
3585 if (self && !node_online(self->node_id)) {
3586 build_zonelists(self);
3587 build_zonelist_cache(self);
3590 for_each_online_node(nid) {
3591 pg_data_t *pgdat = NODE_DATA(nid);
3593 build_zonelists(pgdat);
3594 build_zonelist_cache(pgdat);
3598 * Initialize the boot_pagesets that are going to be used
3599 * for bootstrapping processors. The real pagesets for
3600 * each zone will be allocated later when the per cpu
3601 * allocator is available.
3603 * boot_pagesets are used also for bootstrapping offline
3604 * cpus if the system is already booted because the pagesets
3605 * are needed to initialize allocators on a specific cpu too.
3606 * F.e. the percpu allocator needs the page allocator which
3607 * needs the percpu allocator in order to allocate its pagesets
3608 * (a chicken-egg dilemma).
3610 for_each_possible_cpu(cpu) {
3611 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3613 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3615 * We now know the "local memory node" for each node--
3616 * i.e., the node of the first zone in the generic zonelist.
3617 * Set up numa_mem percpu variable for on-line cpus. During
3618 * boot, only the boot cpu should be on-line; we'll init the
3619 * secondary cpus' numa_mem as they come on-line. During
3620 * node/memory hotplug, we'll fixup all on-line cpus.
3622 if (cpu_online(cpu))
3623 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3631 * Called with zonelists_mutex held always
3632 * unless system_state == SYSTEM_BOOTING.
3634 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3636 set_zonelist_order();
3638 if (system_state == SYSTEM_BOOTING) {
3639 __build_all_zonelists(NULL);
3640 mminit_verify_zonelist();
3641 cpuset_init_current_mems_allowed();
3643 /* we have to stop all cpus to guarantee there is no user
3645 #ifdef CONFIG_MEMORY_HOTPLUG
3647 setup_zone_pageset(zone);
3649 stop_machine(__build_all_zonelists, pgdat, NULL);
3650 /* cpuset refresh routine should be here */
3652 vm_total_pages = nr_free_pagecache_pages();
3654 * Disable grouping by mobility if the number of pages in the
3655 * system is too low to allow the mechanism to work. It would be
3656 * more accurate, but expensive to check per-zone. This check is
3657 * made on memory-hotadd so a system can start with mobility
3658 * disabled and enable it later
3660 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3661 page_group_by_mobility_disabled = 1;
3663 page_group_by_mobility_disabled = 0;
3665 printk("Built %i zonelists in %s order, mobility grouping %s. "
3666 "Total pages: %ld\n",
3668 zonelist_order_name[current_zonelist_order],
3669 page_group_by_mobility_disabled ? "off" : "on",
3672 printk("Policy zone: %s\n", zone_names[policy_zone]);
3677 * Helper functions to size the waitqueue hash table.
3678 * Essentially these want to choose hash table sizes sufficiently
3679 * large so that collisions trying to wait on pages are rare.
3680 * But in fact, the number of active page waitqueues on typical
3681 * systems is ridiculously low, less than 200. So this is even
3682 * conservative, even though it seems large.
3684 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3685 * waitqueues, i.e. the size of the waitq table given the number of pages.
3687 #define PAGES_PER_WAITQUEUE 256
3689 #ifndef CONFIG_MEMORY_HOTPLUG
3690 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3692 unsigned long size = 1;
3694 pages /= PAGES_PER_WAITQUEUE;
3696 while (size < pages)
3700 * Once we have dozens or even hundreds of threads sleeping
3701 * on IO we've got bigger problems than wait queue collision.
3702 * Limit the size of the wait table to a reasonable size.
3704 size = min(size, 4096UL);
3706 return max(size, 4UL);
3710 * A zone's size might be changed by hot-add, so it is not possible to determine
3711 * a suitable size for its wait_table. So we use the maximum size now.
3713 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3715 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3716 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3717 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3719 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3720 * or more by the traditional way. (See above). It equals:
3722 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3723 * ia64(16K page size) : = ( 8G + 4M)byte.
3724 * powerpc (64K page size) : = (32G +16M)byte.
3726 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3733 * This is an integer logarithm so that shifts can be used later
3734 * to extract the more random high bits from the multiplicative
3735 * hash function before the remainder is taken.
3737 static inline unsigned long wait_table_bits(unsigned long size)
3742 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3745 * Check if a pageblock contains reserved pages
3747 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3751 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3752 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3759 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3760 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3761 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3762 * higher will lead to a bigger reserve which will get freed as contiguous
3763 * blocks as reclaim kicks in
3765 static void setup_zone_migrate_reserve(struct zone *zone)
3767 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3769 unsigned long block_migratetype;
3773 * Get the start pfn, end pfn and the number of blocks to reserve
3774 * We have to be careful to be aligned to pageblock_nr_pages to
3775 * make sure that we always check pfn_valid for the first page in
3778 start_pfn = zone->zone_start_pfn;
3779 end_pfn = start_pfn + zone->spanned_pages;
3780 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3781 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3785 * Reserve blocks are generally in place to help high-order atomic
3786 * allocations that are short-lived. A min_free_kbytes value that
3787 * would result in more than 2 reserve blocks for atomic allocations
3788 * is assumed to be in place to help anti-fragmentation for the
3789 * future allocation of hugepages at runtime.
3791 reserve = min(2, reserve);
3793 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3794 if (!pfn_valid(pfn))
3796 page = pfn_to_page(pfn);
3798 /* Watch out for overlapping nodes */
3799 if (page_to_nid(page) != zone_to_nid(zone))
3802 block_migratetype = get_pageblock_migratetype(page);
3804 /* Only test what is necessary when the reserves are not met */
3807 * Blocks with reserved pages will never free, skip
3810 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3811 if (pageblock_is_reserved(pfn, block_end_pfn))
3814 /* If this block is reserved, account for it */
3815 if (block_migratetype == MIGRATE_RESERVE) {
3820 /* Suitable for reserving if this block is movable */
3821 if (block_migratetype == MIGRATE_MOVABLE) {
3822 set_pageblock_migratetype(page,
3824 move_freepages_block(zone, page,
3832 * If the reserve is met and this is a previous reserved block,
3835 if (block_migratetype == MIGRATE_RESERVE) {
3836 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3837 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3843 * Initially all pages are reserved - free ones are freed
3844 * up by free_all_bootmem() once the early boot process is
3845 * done. Non-atomic initialization, single-pass.
3847 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3848 unsigned long start_pfn, enum memmap_context context)
3851 unsigned long end_pfn = start_pfn + size;
3855 if (highest_memmap_pfn < end_pfn - 1)
3856 highest_memmap_pfn = end_pfn - 1;
3858 z = &NODE_DATA(nid)->node_zones[zone];
3859 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3861 * There can be holes in boot-time mem_map[]s
3862 * handed to this function. They do not
3863 * exist on hotplugged memory.
3865 if (context == MEMMAP_EARLY) {
3866 if (!early_pfn_valid(pfn))
3868 if (!early_pfn_in_nid(pfn, nid))
3871 page = pfn_to_page(pfn);
3872 set_page_links(page, zone, nid, pfn);
3873 mminit_verify_page_links(page, zone, nid, pfn);
3874 init_page_count(page);
3875 reset_page_mapcount(page);
3876 SetPageReserved(page);
3878 * Mark the block movable so that blocks are reserved for
3879 * movable at startup. This will force kernel allocations
3880 * to reserve their blocks rather than leaking throughout
3881 * the address space during boot when many long-lived
3882 * kernel allocations are made. Later some blocks near
3883 * the start are marked MIGRATE_RESERVE by
3884 * setup_zone_migrate_reserve()
3886 * bitmap is created for zone's valid pfn range. but memmap
3887 * can be created for invalid pages (for alignment)
3888 * check here not to call set_pageblock_migratetype() against
3891 if ((z->zone_start_pfn <= pfn)
3892 && (pfn < z->zone_start_pfn + z->spanned_pages)
3893 && !(pfn & (pageblock_nr_pages - 1)))
3894 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3896 INIT_LIST_HEAD(&page->lru);
3897 #ifdef WANT_PAGE_VIRTUAL
3898 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3899 if (!is_highmem_idx(zone))
3900 set_page_address(page, __va(pfn << PAGE_SHIFT));
3905 static void __meminit zone_init_free_lists(struct zone *zone)
3908 for_each_migratetype_order(order, t) {
3909 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3910 zone->free_area[order].nr_free = 0;
3914 #ifndef __HAVE_ARCH_MEMMAP_INIT
3915 #define memmap_init(size, nid, zone, start_pfn) \
3916 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3919 static int __meminit zone_batchsize(struct zone *zone)
3925 * The per-cpu-pages pools are set to around 1000th of the
3926 * size of the zone. But no more than 1/2 of a meg.
3928 * OK, so we don't know how big the cache is. So guess.
3930 batch = zone->present_pages / 1024;
3931 if (batch * PAGE_SIZE > 512 * 1024)
3932 batch = (512 * 1024) / PAGE_SIZE;
3933 batch /= 4; /* We effectively *= 4 below */
3938 * Clamp the batch to a 2^n - 1 value. Having a power
3939 * of 2 value was found to be more likely to have
3940 * suboptimal cache aliasing properties in some cases.
3942 * For example if 2 tasks are alternately allocating
3943 * batches of pages, one task can end up with a lot
3944 * of pages of one half of the possible page colors
3945 * and the other with pages of the other colors.
3947 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3952 /* The deferral and batching of frees should be suppressed under NOMMU
3955 * The problem is that NOMMU needs to be able to allocate large chunks
3956 * of contiguous memory as there's no hardware page translation to
3957 * assemble apparent contiguous memory from discontiguous pages.
3959 * Queueing large contiguous runs of pages for batching, however,
3960 * causes the pages to actually be freed in smaller chunks. As there
3961 * can be a significant delay between the individual batches being
3962 * recycled, this leads to the once large chunks of space being
3963 * fragmented and becoming unavailable for high-order allocations.
3969 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3971 struct per_cpu_pages *pcp;
3974 memset(p, 0, sizeof(*p));
3978 pcp->high = 6 * batch;
3979 pcp->batch = max(1UL, 1 * batch);
3980 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3981 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3985 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3986 * to the value high for the pageset p.
3989 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3992 struct per_cpu_pages *pcp;
3996 pcp->batch = max(1UL, high/4);
3997 if ((high/4) > (PAGE_SHIFT * 8))
3998 pcp->batch = PAGE_SHIFT * 8;
4001 static void __meminit setup_zone_pageset(struct zone *zone)
4005 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4007 for_each_possible_cpu(cpu) {
4008 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4010 setup_pageset(pcp, zone_batchsize(zone));
4012 if (percpu_pagelist_fraction)
4013 setup_pagelist_highmark(pcp,
4014 (zone->present_pages /
4015 percpu_pagelist_fraction));
4020 * Allocate per cpu pagesets and initialize them.
4021 * Before this call only boot pagesets were available.
4023 void __init setup_per_cpu_pageset(void)
4027 for_each_populated_zone(zone)
4028 setup_zone_pageset(zone);
4031 static noinline __init_refok
4032 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4035 struct pglist_data *pgdat = zone->zone_pgdat;
4039 * The per-page waitqueue mechanism uses hashed waitqueues
4042 zone->wait_table_hash_nr_entries =
4043 wait_table_hash_nr_entries(zone_size_pages);
4044 zone->wait_table_bits =
4045 wait_table_bits(zone->wait_table_hash_nr_entries);
4046 alloc_size = zone->wait_table_hash_nr_entries
4047 * sizeof(wait_queue_head_t);
4049 if (!slab_is_available()) {
4050 zone->wait_table = (wait_queue_head_t *)
4051 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4054 * This case means that a zone whose size was 0 gets new memory
4055 * via memory hot-add.
4056 * But it may be the case that a new node was hot-added. In
4057 * this case vmalloc() will not be able to use this new node's
4058 * memory - this wait_table must be initialized to use this new
4059 * node itself as well.
4060 * To use this new node's memory, further consideration will be
4063 zone->wait_table = vmalloc(alloc_size);
4065 if (!zone->wait_table)
4068 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4069 init_waitqueue_head(zone->wait_table + i);
4074 static __meminit void zone_pcp_init(struct zone *zone)
4077 * per cpu subsystem is not up at this point. The following code
4078 * relies on the ability of the linker to provide the
4079 * offset of a (static) per cpu variable into the per cpu area.
4081 zone->pageset = &boot_pageset;
4083 if (zone->present_pages)
4084 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4085 zone->name, zone->present_pages,
4086 zone_batchsize(zone));
4089 int __meminit init_currently_empty_zone(struct zone *zone,
4090 unsigned long zone_start_pfn,
4092 enum memmap_context context)
4094 struct pglist_data *pgdat = zone->zone_pgdat;
4096 ret = zone_wait_table_init(zone, size);
4099 pgdat->nr_zones = zone_idx(zone) + 1;
4101 zone->zone_start_pfn = zone_start_pfn;
4103 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4104 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4106 (unsigned long)zone_idx(zone),
4107 zone_start_pfn, (zone_start_pfn + size));
4109 zone_init_free_lists(zone);
4114 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4115 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4117 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4118 * Architectures may implement their own version but if add_active_range()
4119 * was used and there are no special requirements, this is a convenient
4122 int __meminit __early_pfn_to_nid(unsigned long pfn)
4124 unsigned long start_pfn, end_pfn;
4127 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4128 if (start_pfn <= pfn && pfn < end_pfn)
4130 /* This is a memory hole */
4133 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4135 int __meminit early_pfn_to_nid(unsigned long pfn)
4139 nid = __early_pfn_to_nid(pfn);
4142 /* just returns 0 */
4146 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4147 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4151 nid = __early_pfn_to_nid(pfn);
4152 if (nid >= 0 && nid != node)
4159 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4160 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4161 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4163 * If an architecture guarantees that all ranges registered with
4164 * add_active_ranges() contain no holes and may be freed, this
4165 * this function may be used instead of calling free_bootmem() manually.
4167 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4169 unsigned long start_pfn, end_pfn;
4172 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4173 start_pfn = min(start_pfn, max_low_pfn);
4174 end_pfn = min(end_pfn, max_low_pfn);
4176 if (start_pfn < end_pfn)
4177 free_bootmem_node(NODE_DATA(this_nid),
4178 PFN_PHYS(start_pfn),
4179 (end_pfn - start_pfn) << PAGE_SHIFT);
4184 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4185 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4187 * If an architecture guarantees that all ranges registered with
4188 * add_active_ranges() contain no holes and may be freed, this
4189 * function may be used instead of calling memory_present() manually.
4191 void __init sparse_memory_present_with_active_regions(int nid)
4193 unsigned long start_pfn, end_pfn;
4196 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4197 memory_present(this_nid, start_pfn, end_pfn);
4201 * get_pfn_range_for_nid - Return the start and end page frames for a node
4202 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4203 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4204 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4206 * It returns the start and end page frame of a node based on information
4207 * provided by an arch calling add_active_range(). If called for a node
4208 * with no available memory, a warning is printed and the start and end
4211 void __meminit get_pfn_range_for_nid(unsigned int nid,
4212 unsigned long *start_pfn, unsigned long *end_pfn)
4214 unsigned long this_start_pfn, this_end_pfn;
4220 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4221 *start_pfn = min(*start_pfn, this_start_pfn);
4222 *end_pfn = max(*end_pfn, this_end_pfn);
4225 if (*start_pfn == -1UL)
4230 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4231 * assumption is made that zones within a node are ordered in monotonic
4232 * increasing memory addresses so that the "highest" populated zone is used
4234 static void __init find_usable_zone_for_movable(void)
4237 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4238 if (zone_index == ZONE_MOVABLE)
4241 if (arch_zone_highest_possible_pfn[zone_index] >
4242 arch_zone_lowest_possible_pfn[zone_index])
4246 VM_BUG_ON(zone_index == -1);
4247 movable_zone = zone_index;
4251 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4252 * because it is sized independent of architecture. Unlike the other zones,
4253 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4254 * in each node depending on the size of each node and how evenly kernelcore
4255 * is distributed. This helper function adjusts the zone ranges
4256 * provided by the architecture for a given node by using the end of the
4257 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4258 * zones within a node are in order of monotonic increases memory addresses
4260 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4261 unsigned long zone_type,
4262 unsigned long node_start_pfn,
4263 unsigned long node_end_pfn,
4264 unsigned long *zone_start_pfn,
4265 unsigned long *zone_end_pfn)
4267 /* Only adjust if ZONE_MOVABLE is on this node */
4268 if (zone_movable_pfn[nid]) {
4269 /* Size ZONE_MOVABLE */
4270 if (zone_type == ZONE_MOVABLE) {
4271 *zone_start_pfn = zone_movable_pfn[nid];
4272 *zone_end_pfn = min(node_end_pfn,
4273 arch_zone_highest_possible_pfn[movable_zone]);
4275 /* Adjust for ZONE_MOVABLE starting within this range */
4276 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4277 *zone_end_pfn > zone_movable_pfn[nid]) {
4278 *zone_end_pfn = zone_movable_pfn[nid];
4280 /* Check if this whole range is within ZONE_MOVABLE */
4281 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4282 *zone_start_pfn = *zone_end_pfn;
4287 * Return the number of pages a zone spans in a node, including holes
4288 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4290 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4291 unsigned long zone_type,
4292 unsigned long *ignored)
4294 unsigned long node_start_pfn, node_end_pfn;
4295 unsigned long zone_start_pfn, zone_end_pfn;
4297 /* Get the start and end of the node and zone */
4298 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4299 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4300 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4301 adjust_zone_range_for_zone_movable(nid, zone_type,
4302 node_start_pfn, node_end_pfn,
4303 &zone_start_pfn, &zone_end_pfn);
4305 /* Check that this node has pages within the zone's required range */
4306 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4309 /* Move the zone boundaries inside the node if necessary */
4310 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4311 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4313 /* Return the spanned pages */
4314 return zone_end_pfn - zone_start_pfn;
4318 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4319 * then all holes in the requested range will be accounted for.
4321 unsigned long __meminit __absent_pages_in_range(int nid,
4322 unsigned long range_start_pfn,
4323 unsigned long range_end_pfn)
4325 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4326 unsigned long start_pfn, end_pfn;
4329 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4330 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4331 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4332 nr_absent -= end_pfn - start_pfn;
4338 * absent_pages_in_range - Return number of page frames in holes within a range
4339 * @start_pfn: The start PFN to start searching for holes
4340 * @end_pfn: The end PFN to stop searching for holes
4342 * It returns the number of pages frames in memory holes within a range.
4344 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4345 unsigned long end_pfn)
4347 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4350 /* Return the number of page frames in holes in a zone on a node */
4351 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4352 unsigned long zone_type,
4353 unsigned long *ignored)
4355 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4356 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4357 unsigned long node_start_pfn, node_end_pfn;
4358 unsigned long zone_start_pfn, zone_end_pfn;
4360 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4361 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4362 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4364 adjust_zone_range_for_zone_movable(nid, zone_type,
4365 node_start_pfn, node_end_pfn,
4366 &zone_start_pfn, &zone_end_pfn);
4367 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4370 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4371 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4372 unsigned long zone_type,
4373 unsigned long *zones_size)
4375 return zones_size[zone_type];
4378 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4379 unsigned long zone_type,
4380 unsigned long *zholes_size)
4385 return zholes_size[zone_type];
4388 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4390 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4391 unsigned long *zones_size, unsigned long *zholes_size)
4393 unsigned long realtotalpages, totalpages = 0;
4396 for (i = 0; i < MAX_NR_ZONES; i++)
4397 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4399 pgdat->node_spanned_pages = totalpages;
4401 realtotalpages = totalpages;
4402 for (i = 0; i < MAX_NR_ZONES; i++)
4404 zone_absent_pages_in_node(pgdat->node_id, i,
4406 pgdat->node_present_pages = realtotalpages;
4407 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4411 #ifndef CONFIG_SPARSEMEM
4413 * Calculate the size of the zone->blockflags rounded to an unsigned long
4414 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4415 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4416 * round what is now in bits to nearest long in bits, then return it in
4419 static unsigned long __init usemap_size(unsigned long zonesize)
4421 unsigned long usemapsize;
4423 usemapsize = roundup(zonesize, pageblock_nr_pages);
4424 usemapsize = usemapsize >> pageblock_order;
4425 usemapsize *= NR_PAGEBLOCK_BITS;
4426 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4428 return usemapsize / 8;
4431 static void __init setup_usemap(struct pglist_data *pgdat,
4432 struct zone *zone, unsigned long zonesize)
4434 unsigned long usemapsize = usemap_size(zonesize);
4435 zone->pageblock_flags = NULL;
4437 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4441 static inline void setup_usemap(struct pglist_data *pgdat,
4442 struct zone *zone, unsigned long zonesize) {}
4443 #endif /* CONFIG_SPARSEMEM */
4445 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4447 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4448 void __init set_pageblock_order(void)
4452 /* Check that pageblock_nr_pages has not already been setup */
4453 if (pageblock_order)
4456 if (HPAGE_SHIFT > PAGE_SHIFT)
4457 order = HUGETLB_PAGE_ORDER;
4459 order = MAX_ORDER - 1;
4462 * Assume the largest contiguous order of interest is a huge page.
4463 * This value may be variable depending on boot parameters on IA64 and
4466 pageblock_order = order;
4468 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4471 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4472 * is unused as pageblock_order is set at compile-time. See
4473 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4476 void __init set_pageblock_order(void)
4480 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4483 * Set up the zone data structures:
4484 * - mark all pages reserved
4485 * - mark all memory queues empty
4486 * - clear the memory bitmaps
4488 * NOTE: pgdat should get zeroed by caller.
4490 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4491 unsigned long *zones_size, unsigned long *zholes_size)
4494 int nid = pgdat->node_id;
4495 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4498 pgdat_resize_init(pgdat);
4499 init_waitqueue_head(&pgdat->kswapd_wait);
4500 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4501 pgdat_page_cgroup_init(pgdat);
4503 for (j = 0; j < MAX_NR_ZONES; j++) {
4504 struct zone *zone = pgdat->node_zones + j;
4505 unsigned long size, realsize, memmap_pages;
4507 size = zone_spanned_pages_in_node(nid, j, zones_size);
4508 realsize = size - zone_absent_pages_in_node(nid, j,
4512 * Adjust realsize so that it accounts for how much memory
4513 * is used by this zone for memmap. This affects the watermark
4514 * and per-cpu initialisations
4517 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4518 if (realsize >= memmap_pages) {
4519 realsize -= memmap_pages;
4522 " %s zone: %lu pages used for memmap\n",
4523 zone_names[j], memmap_pages);
4526 " %s zone: %lu pages exceeds realsize %lu\n",
4527 zone_names[j], memmap_pages, realsize);
4529 /* Account for reserved pages */
4530 if (j == 0 && realsize > dma_reserve) {
4531 realsize -= dma_reserve;
4532 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4533 zone_names[0], dma_reserve);
4536 if (!is_highmem_idx(j))
4537 nr_kernel_pages += realsize;
4538 nr_all_pages += realsize;
4540 zone->spanned_pages = size;
4541 zone->present_pages = realsize;
4544 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4546 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4548 zone->name = zone_names[j];
4549 spin_lock_init(&zone->lock);
4550 spin_lock_init(&zone->lru_lock);
4551 zone_seqlock_init(zone);
4552 zone->zone_pgdat = pgdat;
4554 zone_pcp_init(zone);
4555 lruvec_init(&zone->lruvec);
4559 set_pageblock_order();
4560 setup_usemap(pgdat, zone, size);
4561 ret = init_currently_empty_zone(zone, zone_start_pfn,
4562 size, MEMMAP_EARLY);
4564 memmap_init(size, nid, j, zone_start_pfn);
4565 zone_start_pfn += size;
4569 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4571 /* Skip empty nodes */
4572 if (!pgdat->node_spanned_pages)
4575 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4576 /* ia64 gets its own node_mem_map, before this, without bootmem */
4577 if (!pgdat->node_mem_map) {
4578 unsigned long size, start, end;
4582 * The zone's endpoints aren't required to be MAX_ORDER
4583 * aligned but the node_mem_map endpoints must be in order
4584 * for the buddy allocator to function correctly.
4586 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4587 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4588 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4589 size = (end - start) * sizeof(struct page);
4590 map = alloc_remap(pgdat->node_id, size);
4592 map = alloc_bootmem_node_nopanic(pgdat, size);
4593 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4595 #ifndef CONFIG_NEED_MULTIPLE_NODES
4597 * With no DISCONTIG, the global mem_map is just set as node 0's
4599 if (pgdat == NODE_DATA(0)) {
4600 mem_map = NODE_DATA(0)->node_mem_map;
4601 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4602 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4603 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4604 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4607 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4610 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4611 unsigned long node_start_pfn, unsigned long *zholes_size)
4613 pg_data_t *pgdat = NODE_DATA(nid);
4615 /* pg_data_t should be reset to zero when it's allocated */
4616 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4618 pgdat->node_id = nid;
4619 pgdat->node_start_pfn = node_start_pfn;
4620 init_zone_allows_reclaim(nid);
4621 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4623 alloc_node_mem_map(pgdat);
4624 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4625 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4626 nid, (unsigned long)pgdat,
4627 (unsigned long)pgdat->node_mem_map);
4630 free_area_init_core(pgdat, zones_size, zholes_size);
4633 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4635 #if MAX_NUMNODES > 1
4637 * Figure out the number of possible node ids.
4639 static void __init setup_nr_node_ids(void)
4642 unsigned int highest = 0;
4644 for_each_node_mask(node, node_possible_map)
4646 nr_node_ids = highest + 1;
4649 static inline void setup_nr_node_ids(void)
4655 * node_map_pfn_alignment - determine the maximum internode alignment
4657 * This function should be called after node map is populated and sorted.
4658 * It calculates the maximum power of two alignment which can distinguish
4661 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4662 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4663 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4664 * shifted, 1GiB is enough and this function will indicate so.
4666 * This is used to test whether pfn -> nid mapping of the chosen memory
4667 * model has fine enough granularity to avoid incorrect mapping for the
4668 * populated node map.
4670 * Returns the determined alignment in pfn's. 0 if there is no alignment
4671 * requirement (single node).
4673 unsigned long __init node_map_pfn_alignment(void)
4675 unsigned long accl_mask = 0, last_end = 0;
4676 unsigned long start, end, mask;
4680 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4681 if (!start || last_nid < 0 || last_nid == nid) {
4688 * Start with a mask granular enough to pin-point to the
4689 * start pfn and tick off bits one-by-one until it becomes
4690 * too coarse to separate the current node from the last.
4692 mask = ~((1 << __ffs(start)) - 1);
4693 while (mask && last_end <= (start & (mask << 1)))
4696 /* accumulate all internode masks */
4700 /* convert mask to number of pages */
4701 return ~accl_mask + 1;
4704 /* Find the lowest pfn for a node */
4705 static unsigned long __init find_min_pfn_for_node(int nid)
4707 unsigned long min_pfn = ULONG_MAX;
4708 unsigned long start_pfn;
4711 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4712 min_pfn = min(min_pfn, start_pfn);
4714 if (min_pfn == ULONG_MAX) {
4716 "Could not find start_pfn for node %d\n", nid);
4724 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4726 * It returns the minimum PFN based on information provided via
4727 * add_active_range().
4729 unsigned long __init find_min_pfn_with_active_regions(void)
4731 return find_min_pfn_for_node(MAX_NUMNODES);
4735 * early_calculate_totalpages()
4736 * Sum pages in active regions for movable zone.
4737 * Populate N_MEMORY for calculating usable_nodes.
4739 static unsigned long __init early_calculate_totalpages(void)
4741 unsigned long totalpages = 0;
4742 unsigned long start_pfn, end_pfn;
4745 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4746 unsigned long pages = end_pfn - start_pfn;
4748 totalpages += pages;
4750 node_set_state(nid, N_MEMORY);
4756 * Find the PFN the Movable zone begins in each node. Kernel memory
4757 * is spread evenly between nodes as long as the nodes have enough
4758 * memory. When they don't, some nodes will have more kernelcore than
4761 static void __init find_zone_movable_pfns_for_nodes(void)
4764 unsigned long usable_startpfn;
4765 unsigned long kernelcore_node, kernelcore_remaining;
4766 /* save the state before borrow the nodemask */
4767 nodemask_t saved_node_state = node_states[N_MEMORY];
4768 unsigned long totalpages = early_calculate_totalpages();
4769 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4772 * If movablecore was specified, calculate what size of
4773 * kernelcore that corresponds so that memory usable for
4774 * any allocation type is evenly spread. If both kernelcore
4775 * and movablecore are specified, then the value of kernelcore
4776 * will be used for required_kernelcore if it's greater than
4777 * what movablecore would have allowed.
4779 if (required_movablecore) {
4780 unsigned long corepages;
4783 * Round-up so that ZONE_MOVABLE is at least as large as what
4784 * was requested by the user
4786 required_movablecore =
4787 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4788 corepages = totalpages - required_movablecore;
4790 required_kernelcore = max(required_kernelcore, corepages);
4793 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4794 if (!required_kernelcore)
4797 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4798 find_usable_zone_for_movable();
4799 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4802 /* Spread kernelcore memory as evenly as possible throughout nodes */
4803 kernelcore_node = required_kernelcore / usable_nodes;
4804 for_each_node_state(nid, N_MEMORY) {
4805 unsigned long start_pfn, end_pfn;
4808 * Recalculate kernelcore_node if the division per node
4809 * now exceeds what is necessary to satisfy the requested
4810 * amount of memory for the kernel
4812 if (required_kernelcore < kernelcore_node)
4813 kernelcore_node = required_kernelcore / usable_nodes;
4816 * As the map is walked, we track how much memory is usable
4817 * by the kernel using kernelcore_remaining. When it is
4818 * 0, the rest of the node is usable by ZONE_MOVABLE
4820 kernelcore_remaining = kernelcore_node;
4822 /* Go through each range of PFNs within this node */
4823 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4824 unsigned long size_pages;
4826 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4827 if (start_pfn >= end_pfn)
4830 /* Account for what is only usable for kernelcore */
4831 if (start_pfn < usable_startpfn) {
4832 unsigned long kernel_pages;
4833 kernel_pages = min(end_pfn, usable_startpfn)
4836 kernelcore_remaining -= min(kernel_pages,
4837 kernelcore_remaining);
4838 required_kernelcore -= min(kernel_pages,
4839 required_kernelcore);
4841 /* Continue if range is now fully accounted */
4842 if (end_pfn <= usable_startpfn) {
4845 * Push zone_movable_pfn to the end so
4846 * that if we have to rebalance
4847 * kernelcore across nodes, we will
4848 * not double account here
4850 zone_movable_pfn[nid] = end_pfn;
4853 start_pfn = usable_startpfn;
4857 * The usable PFN range for ZONE_MOVABLE is from
4858 * start_pfn->end_pfn. Calculate size_pages as the
4859 * number of pages used as kernelcore
4861 size_pages = end_pfn - start_pfn;
4862 if (size_pages > kernelcore_remaining)
4863 size_pages = kernelcore_remaining;
4864 zone_movable_pfn[nid] = start_pfn + size_pages;
4867 * Some kernelcore has been met, update counts and
4868 * break if the kernelcore for this node has been
4871 required_kernelcore -= min(required_kernelcore,
4873 kernelcore_remaining -= size_pages;
4874 if (!kernelcore_remaining)
4880 * If there is still required_kernelcore, we do another pass with one
4881 * less node in the count. This will push zone_movable_pfn[nid] further
4882 * along on the nodes that still have memory until kernelcore is
4886 if (usable_nodes && required_kernelcore > usable_nodes)
4889 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4890 for (nid = 0; nid < MAX_NUMNODES; nid++)
4891 zone_movable_pfn[nid] =
4892 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4895 /* restore the node_state */
4896 node_states[N_MEMORY] = saved_node_state;
4899 /* Any regular or high memory on that node ? */
4900 static void check_for_memory(pg_data_t *pgdat, int nid)
4902 enum zone_type zone_type;
4904 if (N_MEMORY == N_NORMAL_MEMORY)
4907 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
4908 struct zone *zone = &pgdat->node_zones[zone_type];
4909 if (zone->present_pages) {
4910 node_set_state(nid, N_HIGH_MEMORY);
4911 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
4912 zone_type <= ZONE_NORMAL)
4913 node_set_state(nid, N_NORMAL_MEMORY);
4920 * free_area_init_nodes - Initialise all pg_data_t and zone data
4921 * @max_zone_pfn: an array of max PFNs for each zone
4923 * This will call free_area_init_node() for each active node in the system.
4924 * Using the page ranges provided by add_active_range(), the size of each
4925 * zone in each node and their holes is calculated. If the maximum PFN
4926 * between two adjacent zones match, it is assumed that the zone is empty.
4927 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4928 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4929 * starts where the previous one ended. For example, ZONE_DMA32 starts
4930 * at arch_max_dma_pfn.
4932 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4934 unsigned long start_pfn, end_pfn;
4937 /* Record where the zone boundaries are */
4938 memset(arch_zone_lowest_possible_pfn, 0,
4939 sizeof(arch_zone_lowest_possible_pfn));
4940 memset(arch_zone_highest_possible_pfn, 0,
4941 sizeof(arch_zone_highest_possible_pfn));
4942 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4943 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4944 for (i = 1; i < MAX_NR_ZONES; i++) {
4945 if (i == ZONE_MOVABLE)
4947 arch_zone_lowest_possible_pfn[i] =
4948 arch_zone_highest_possible_pfn[i-1];
4949 arch_zone_highest_possible_pfn[i] =
4950 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4952 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4953 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4955 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4956 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4957 find_zone_movable_pfns_for_nodes();
4959 /* Print out the zone ranges */
4960 printk("Zone ranges:\n");
4961 for (i = 0; i < MAX_NR_ZONES; i++) {
4962 if (i == ZONE_MOVABLE)
4964 printk(KERN_CONT " %-8s ", zone_names[i]);
4965 if (arch_zone_lowest_possible_pfn[i] ==
4966 arch_zone_highest_possible_pfn[i])
4967 printk(KERN_CONT "empty\n");
4969 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4970 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4971 (arch_zone_highest_possible_pfn[i]
4972 << PAGE_SHIFT) - 1);
4975 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4976 printk("Movable zone start for each node\n");
4977 for (i = 0; i < MAX_NUMNODES; i++) {
4978 if (zone_movable_pfn[i])
4979 printk(" Node %d: %#010lx\n", i,
4980 zone_movable_pfn[i] << PAGE_SHIFT);
4983 /* Print out the early node map */
4984 printk("Early memory node ranges\n");
4985 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4986 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
4987 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
4989 /* Initialise every node */
4990 mminit_verify_pageflags_layout();
4991 setup_nr_node_ids();
4992 for_each_online_node(nid) {
4993 pg_data_t *pgdat = NODE_DATA(nid);
4994 free_area_init_node(nid, NULL,
4995 find_min_pfn_for_node(nid), NULL);
4997 /* Any memory on that node */
4998 if (pgdat->node_present_pages)
4999 node_set_state(nid, N_MEMORY);
5000 check_for_memory(pgdat, nid);
5004 static int __init cmdline_parse_core(char *p, unsigned long *core)
5006 unsigned long long coremem;
5010 coremem = memparse(p, &p);
5011 *core = coremem >> PAGE_SHIFT;
5013 /* Paranoid check that UL is enough for the coremem value */
5014 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5020 * kernelcore=size sets the amount of memory for use for allocations that
5021 * cannot be reclaimed or migrated.
5023 static int __init cmdline_parse_kernelcore(char *p)
5025 return cmdline_parse_core(p, &required_kernelcore);
5029 * movablecore=size sets the amount of memory for use for allocations that
5030 * can be reclaimed or migrated.
5032 static int __init cmdline_parse_movablecore(char *p)
5034 return cmdline_parse_core(p, &required_movablecore);
5037 early_param("kernelcore", cmdline_parse_kernelcore);
5038 early_param("movablecore", cmdline_parse_movablecore);
5040 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5043 * set_dma_reserve - set the specified number of pages reserved in the first zone
5044 * @new_dma_reserve: The number of pages to mark reserved
5046 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5047 * In the DMA zone, a significant percentage may be consumed by kernel image
5048 * and other unfreeable allocations which can skew the watermarks badly. This
5049 * function may optionally be used to account for unfreeable pages in the
5050 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5051 * smaller per-cpu batchsize.
5053 void __init set_dma_reserve(unsigned long new_dma_reserve)
5055 dma_reserve = new_dma_reserve;
5058 void __init free_area_init(unsigned long *zones_size)
5060 free_area_init_node(0, zones_size,
5061 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5064 static int page_alloc_cpu_notify(struct notifier_block *self,
5065 unsigned long action, void *hcpu)
5067 int cpu = (unsigned long)hcpu;
5069 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5070 lru_add_drain_cpu(cpu);
5074 * Spill the event counters of the dead processor
5075 * into the current processors event counters.
5076 * This artificially elevates the count of the current
5079 vm_events_fold_cpu(cpu);
5082 * Zero the differential counters of the dead processor
5083 * so that the vm statistics are consistent.
5085 * This is only okay since the processor is dead and cannot
5086 * race with what we are doing.
5088 refresh_cpu_vm_stats(cpu);
5093 void __init page_alloc_init(void)
5095 hotcpu_notifier(page_alloc_cpu_notify, 0);
5099 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5100 * or min_free_kbytes changes.
5102 static void calculate_totalreserve_pages(void)
5104 struct pglist_data *pgdat;
5105 unsigned long reserve_pages = 0;
5106 enum zone_type i, j;
5108 for_each_online_pgdat(pgdat) {
5109 for (i = 0; i < MAX_NR_ZONES; i++) {
5110 struct zone *zone = pgdat->node_zones + i;
5111 unsigned long max = 0;
5113 /* Find valid and maximum lowmem_reserve in the zone */
5114 for (j = i; j < MAX_NR_ZONES; j++) {
5115 if (zone->lowmem_reserve[j] > max)
5116 max = zone->lowmem_reserve[j];
5119 /* we treat the high watermark as reserved pages. */
5120 max += high_wmark_pages(zone);
5122 if (max > zone->present_pages)
5123 max = zone->present_pages;
5124 reserve_pages += max;
5126 * Lowmem reserves are not available to
5127 * GFP_HIGHUSER page cache allocations and
5128 * kswapd tries to balance zones to their high
5129 * watermark. As a result, neither should be
5130 * regarded as dirtyable memory, to prevent a
5131 * situation where reclaim has to clean pages
5132 * in order to balance the zones.
5134 zone->dirty_balance_reserve = max;
5137 dirty_balance_reserve = reserve_pages;
5138 totalreserve_pages = reserve_pages;
5142 * setup_per_zone_lowmem_reserve - called whenever
5143 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5144 * has a correct pages reserved value, so an adequate number of
5145 * pages are left in the zone after a successful __alloc_pages().
5147 static void setup_per_zone_lowmem_reserve(void)
5149 struct pglist_data *pgdat;
5150 enum zone_type j, idx;
5152 for_each_online_pgdat(pgdat) {
5153 for (j = 0; j < MAX_NR_ZONES; j++) {
5154 struct zone *zone = pgdat->node_zones + j;
5155 unsigned long present_pages = zone->present_pages;
5157 zone->lowmem_reserve[j] = 0;
5161 struct zone *lower_zone;
5165 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5166 sysctl_lowmem_reserve_ratio[idx] = 1;
5168 lower_zone = pgdat->node_zones + idx;
5169 lower_zone->lowmem_reserve[j] = present_pages /
5170 sysctl_lowmem_reserve_ratio[idx];
5171 present_pages += lower_zone->present_pages;
5176 /* update totalreserve_pages */
5177 calculate_totalreserve_pages();
5180 static void __setup_per_zone_wmarks(void)
5182 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5183 unsigned long lowmem_pages = 0;
5185 unsigned long flags;
5187 /* Calculate total number of !ZONE_HIGHMEM pages */
5188 for_each_zone(zone) {
5189 if (!is_highmem(zone))
5190 lowmem_pages += zone->present_pages;
5193 for_each_zone(zone) {
5196 spin_lock_irqsave(&zone->lock, flags);
5197 tmp = (u64)pages_min * zone->present_pages;
5198 do_div(tmp, lowmem_pages);
5199 if (is_highmem(zone)) {
5201 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5202 * need highmem pages, so cap pages_min to a small
5205 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5206 * deltas controls asynch page reclaim, and so should
5207 * not be capped for highmem.
5211 min_pages = zone->present_pages / 1024;
5212 if (min_pages < SWAP_CLUSTER_MAX)
5213 min_pages = SWAP_CLUSTER_MAX;
5214 if (min_pages > 128)
5216 zone->watermark[WMARK_MIN] = min_pages;
5219 * If it's a lowmem zone, reserve a number of pages
5220 * proportionate to the zone's size.
5222 zone->watermark[WMARK_MIN] = tmp;
5225 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5226 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5228 setup_zone_migrate_reserve(zone);
5229 spin_unlock_irqrestore(&zone->lock, flags);
5232 /* update totalreserve_pages */
5233 calculate_totalreserve_pages();
5237 * setup_per_zone_wmarks - called when min_free_kbytes changes
5238 * or when memory is hot-{added|removed}
5240 * Ensures that the watermark[min,low,high] values for each zone are set
5241 * correctly with respect to min_free_kbytes.
5243 void setup_per_zone_wmarks(void)
5245 mutex_lock(&zonelists_mutex);
5246 __setup_per_zone_wmarks();
5247 mutex_unlock(&zonelists_mutex);
5251 * The inactive anon list should be small enough that the VM never has to
5252 * do too much work, but large enough that each inactive page has a chance
5253 * to be referenced again before it is swapped out.
5255 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5256 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5257 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5258 * the anonymous pages are kept on the inactive list.
5261 * memory ratio inactive anon
5262 * -------------------------------------
5271 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5273 unsigned int gb, ratio;
5275 /* Zone size in gigabytes */
5276 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5278 ratio = int_sqrt(10 * gb);
5282 zone->inactive_ratio = ratio;
5285 static void __meminit setup_per_zone_inactive_ratio(void)
5290 calculate_zone_inactive_ratio(zone);
5294 * Initialise min_free_kbytes.
5296 * For small machines we want it small (128k min). For large machines
5297 * we want it large (64MB max). But it is not linear, because network
5298 * bandwidth does not increase linearly with machine size. We use
5300 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5301 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5317 int __meminit init_per_zone_wmark_min(void)
5319 unsigned long lowmem_kbytes;
5321 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5323 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5324 if (min_free_kbytes < 128)
5325 min_free_kbytes = 128;
5326 if (min_free_kbytes > 65536)
5327 min_free_kbytes = 65536;
5328 setup_per_zone_wmarks();
5329 refresh_zone_stat_thresholds();
5330 setup_per_zone_lowmem_reserve();
5331 setup_per_zone_inactive_ratio();
5334 module_init(init_per_zone_wmark_min)
5337 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5338 * that we can call two helper functions whenever min_free_kbytes
5341 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5342 void __user *buffer, size_t *length, loff_t *ppos)
5344 proc_dointvec(table, write, buffer, length, ppos);
5346 setup_per_zone_wmarks();
5351 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5352 void __user *buffer, size_t *length, loff_t *ppos)
5357 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5362 zone->min_unmapped_pages = (zone->present_pages *
5363 sysctl_min_unmapped_ratio) / 100;
5367 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5368 void __user *buffer, size_t *length, loff_t *ppos)
5373 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5378 zone->min_slab_pages = (zone->present_pages *
5379 sysctl_min_slab_ratio) / 100;
5385 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5386 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5387 * whenever sysctl_lowmem_reserve_ratio changes.
5389 * The reserve ratio obviously has absolutely no relation with the
5390 * minimum watermarks. The lowmem reserve ratio can only make sense
5391 * if in function of the boot time zone sizes.
5393 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5394 void __user *buffer, size_t *length, loff_t *ppos)
5396 proc_dointvec_minmax(table, write, buffer, length, ppos);
5397 setup_per_zone_lowmem_reserve();
5402 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5403 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5404 * can have before it gets flushed back to buddy allocator.
5407 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5408 void __user *buffer, size_t *length, loff_t *ppos)
5414 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5415 if (!write || (ret < 0))
5417 for_each_populated_zone(zone) {
5418 for_each_possible_cpu(cpu) {
5420 high = zone->present_pages / percpu_pagelist_fraction;
5421 setup_pagelist_highmark(
5422 per_cpu_ptr(zone->pageset, cpu), high);
5428 int hashdist = HASHDIST_DEFAULT;
5431 static int __init set_hashdist(char *str)
5435 hashdist = simple_strtoul(str, &str, 0);
5438 __setup("hashdist=", set_hashdist);
5442 * allocate a large system hash table from bootmem
5443 * - it is assumed that the hash table must contain an exact power-of-2
5444 * quantity of entries
5445 * - limit is the number of hash buckets, not the total allocation size
5447 void *__init alloc_large_system_hash(const char *tablename,
5448 unsigned long bucketsize,
5449 unsigned long numentries,
5452 unsigned int *_hash_shift,
5453 unsigned int *_hash_mask,
5454 unsigned long low_limit,
5455 unsigned long high_limit)
5457 unsigned long long max = high_limit;
5458 unsigned long log2qty, size;
5461 /* allow the kernel cmdline to have a say */
5463 /* round applicable memory size up to nearest megabyte */
5464 numentries = nr_kernel_pages;
5465 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5466 numentries >>= 20 - PAGE_SHIFT;
5467 numentries <<= 20 - PAGE_SHIFT;
5469 /* limit to 1 bucket per 2^scale bytes of low memory */
5470 if (scale > PAGE_SHIFT)
5471 numentries >>= (scale - PAGE_SHIFT);
5473 numentries <<= (PAGE_SHIFT - scale);
5475 /* Make sure we've got at least a 0-order allocation.. */
5476 if (unlikely(flags & HASH_SMALL)) {
5477 /* Makes no sense without HASH_EARLY */
5478 WARN_ON(!(flags & HASH_EARLY));
5479 if (!(numentries >> *_hash_shift)) {
5480 numentries = 1UL << *_hash_shift;
5481 BUG_ON(!numentries);
5483 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5484 numentries = PAGE_SIZE / bucketsize;
5486 numentries = roundup_pow_of_two(numentries);
5488 /* limit allocation size to 1/16 total memory by default */
5490 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5491 do_div(max, bucketsize);
5493 max = min(max, 0x80000000ULL);
5495 if (numentries < low_limit)
5496 numentries = low_limit;
5497 if (numentries > max)
5500 log2qty = ilog2(numentries);
5503 size = bucketsize << log2qty;
5504 if (flags & HASH_EARLY)
5505 table = alloc_bootmem_nopanic(size);
5507 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5510 * If bucketsize is not a power-of-two, we may free
5511 * some pages at the end of hash table which
5512 * alloc_pages_exact() automatically does
5514 if (get_order(size) < MAX_ORDER) {
5515 table = alloc_pages_exact(size, GFP_ATOMIC);
5516 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5519 } while (!table && size > PAGE_SIZE && --log2qty);
5522 panic("Failed to allocate %s hash table\n", tablename);
5524 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5527 ilog2(size) - PAGE_SHIFT,
5531 *_hash_shift = log2qty;
5533 *_hash_mask = (1 << log2qty) - 1;
5538 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5539 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5542 #ifdef CONFIG_SPARSEMEM
5543 return __pfn_to_section(pfn)->pageblock_flags;
5545 return zone->pageblock_flags;
5546 #endif /* CONFIG_SPARSEMEM */
5549 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5551 #ifdef CONFIG_SPARSEMEM
5552 pfn &= (PAGES_PER_SECTION-1);
5553 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5555 pfn = pfn - zone->zone_start_pfn;
5556 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5557 #endif /* CONFIG_SPARSEMEM */
5561 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5562 * @page: The page within the block of interest
5563 * @start_bitidx: The first bit of interest to retrieve
5564 * @end_bitidx: The last bit of interest
5565 * returns pageblock_bits flags
5567 unsigned long get_pageblock_flags_group(struct page *page,
5568 int start_bitidx, int end_bitidx)
5571 unsigned long *bitmap;
5572 unsigned long pfn, bitidx;
5573 unsigned long flags = 0;
5574 unsigned long value = 1;
5576 zone = page_zone(page);
5577 pfn = page_to_pfn(page);
5578 bitmap = get_pageblock_bitmap(zone, pfn);
5579 bitidx = pfn_to_bitidx(zone, pfn);
5581 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5582 if (test_bit(bitidx + start_bitidx, bitmap))
5589 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5590 * @page: The page within the block of interest
5591 * @start_bitidx: The first bit of interest
5592 * @end_bitidx: The last bit of interest
5593 * @flags: The flags to set
5595 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5596 int start_bitidx, int end_bitidx)
5599 unsigned long *bitmap;
5600 unsigned long pfn, bitidx;
5601 unsigned long value = 1;
5603 zone = page_zone(page);
5604 pfn = page_to_pfn(page);
5605 bitmap = get_pageblock_bitmap(zone, pfn);
5606 bitidx = pfn_to_bitidx(zone, pfn);
5607 VM_BUG_ON(pfn < zone->zone_start_pfn);
5608 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5610 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5612 __set_bit(bitidx + start_bitidx, bitmap);
5614 __clear_bit(bitidx + start_bitidx, bitmap);
5618 * This function checks whether pageblock includes unmovable pages or not.
5619 * If @count is not zero, it is okay to include less @count unmovable pages
5621 * PageLRU check wihtout isolation or lru_lock could race so that
5622 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5623 * expect this function should be exact.
5625 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5626 bool skip_hwpoisoned_pages)
5628 unsigned long pfn, iter, found;
5632 * For avoiding noise data, lru_add_drain_all() should be called
5633 * If ZONE_MOVABLE, the zone never contains unmovable pages
5635 if (zone_idx(zone) == ZONE_MOVABLE)
5637 mt = get_pageblock_migratetype(page);
5638 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5641 pfn = page_to_pfn(page);
5642 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5643 unsigned long check = pfn + iter;
5645 if (!pfn_valid_within(check))
5648 page = pfn_to_page(check);
5650 * We can't use page_count without pin a page
5651 * because another CPU can free compound page.
5652 * This check already skips compound tails of THP
5653 * because their page->_count is zero at all time.
5655 if (!atomic_read(&page->_count)) {
5656 if (PageBuddy(page))
5657 iter += (1 << page_order(page)) - 1;
5662 * The HWPoisoned page may be not in buddy system, and
5663 * page_count() is not 0.
5665 if (skip_hwpoisoned_pages && PageHWPoison(page))
5671 * If there are RECLAIMABLE pages, we need to check it.
5672 * But now, memory offline itself doesn't call shrink_slab()
5673 * and it still to be fixed.
5676 * If the page is not RAM, page_count()should be 0.
5677 * we don't need more check. This is an _used_ not-movable page.
5679 * The problematic thing here is PG_reserved pages. PG_reserved
5680 * is set to both of a memory hole page and a _used_ kernel
5689 bool is_pageblock_removable_nolock(struct page *page)
5695 * We have to be careful here because we are iterating over memory
5696 * sections which are not zone aware so we might end up outside of
5697 * the zone but still within the section.
5698 * We have to take care about the node as well. If the node is offline
5699 * its NODE_DATA will be NULL - see page_zone.
5701 if (!node_online(page_to_nid(page)))
5704 zone = page_zone(page);
5705 pfn = page_to_pfn(page);
5706 if (zone->zone_start_pfn > pfn ||
5707 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5710 return !has_unmovable_pages(zone, page, 0, true);
5715 static unsigned long pfn_max_align_down(unsigned long pfn)
5717 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5718 pageblock_nr_pages) - 1);
5721 static unsigned long pfn_max_align_up(unsigned long pfn)
5723 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5724 pageblock_nr_pages));
5727 /* [start, end) must belong to a single zone. */
5728 static int __alloc_contig_migrate_range(struct compact_control *cc,
5729 unsigned long start, unsigned long end)
5731 /* This function is based on compact_zone() from compaction.c. */
5732 unsigned long nr_reclaimed;
5733 unsigned long pfn = start;
5734 unsigned int tries = 0;
5739 while (pfn < end || !list_empty(&cc->migratepages)) {
5740 if (fatal_signal_pending(current)) {
5745 if (list_empty(&cc->migratepages)) {
5746 cc->nr_migratepages = 0;
5747 pfn = isolate_migratepages_range(cc->zone, cc,
5754 } else if (++tries == 5) {
5755 ret = ret < 0 ? ret : -EBUSY;
5759 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5761 cc->nr_migratepages -= nr_reclaimed;
5763 ret = migrate_pages(&cc->migratepages,
5764 alloc_migrate_target,
5765 0, false, MIGRATE_SYNC);
5768 putback_movable_pages(&cc->migratepages);
5769 return ret > 0 ? 0 : ret;
5773 * alloc_contig_range() -- tries to allocate given range of pages
5774 * @start: start PFN to allocate
5775 * @end: one-past-the-last PFN to allocate
5776 * @migratetype: migratetype of the underlaying pageblocks (either
5777 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5778 * in range must have the same migratetype and it must
5779 * be either of the two.
5781 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5782 * aligned, however it's the caller's responsibility to guarantee that
5783 * we are the only thread that changes migrate type of pageblocks the
5786 * The PFN range must belong to a single zone.
5788 * Returns zero on success or negative error code. On success all
5789 * pages which PFN is in [start, end) are allocated for the caller and
5790 * need to be freed with free_contig_range().
5792 int alloc_contig_range(unsigned long start, unsigned long end,
5793 unsigned migratetype)
5795 unsigned long outer_start, outer_end;
5798 struct compact_control cc = {
5799 .nr_migratepages = 0,
5801 .zone = page_zone(pfn_to_page(start)),
5803 .ignore_skip_hint = true,
5805 INIT_LIST_HEAD(&cc.migratepages);
5808 * What we do here is we mark all pageblocks in range as
5809 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5810 * have different sizes, and due to the way page allocator
5811 * work, we align the range to biggest of the two pages so
5812 * that page allocator won't try to merge buddies from
5813 * different pageblocks and change MIGRATE_ISOLATE to some
5814 * other migration type.
5816 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5817 * migrate the pages from an unaligned range (ie. pages that
5818 * we are interested in). This will put all the pages in
5819 * range back to page allocator as MIGRATE_ISOLATE.
5821 * When this is done, we take the pages in range from page
5822 * allocator removing them from the buddy system. This way
5823 * page allocator will never consider using them.
5825 * This lets us mark the pageblocks back as
5826 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5827 * aligned range but not in the unaligned, original range are
5828 * put back to page allocator so that buddy can use them.
5831 ret = start_isolate_page_range(pfn_max_align_down(start),
5832 pfn_max_align_up(end), migratetype,
5837 ret = __alloc_contig_migrate_range(&cc, start, end);
5842 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5843 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5844 * more, all pages in [start, end) are free in page allocator.
5845 * What we are going to do is to allocate all pages from
5846 * [start, end) (that is remove them from page allocator).
5848 * The only problem is that pages at the beginning and at the
5849 * end of interesting range may be not aligned with pages that
5850 * page allocator holds, ie. they can be part of higher order
5851 * pages. Because of this, we reserve the bigger range and
5852 * once this is done free the pages we are not interested in.
5854 * We don't have to hold zone->lock here because the pages are
5855 * isolated thus they won't get removed from buddy.
5858 lru_add_drain_all();
5862 outer_start = start;
5863 while (!PageBuddy(pfn_to_page(outer_start))) {
5864 if (++order >= MAX_ORDER) {
5868 outer_start &= ~0UL << order;
5871 /* Make sure the range is really isolated. */
5872 if (test_pages_isolated(outer_start, end, false)) {
5873 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5880 /* Grab isolated pages from freelists. */
5881 outer_end = isolate_freepages_range(&cc, outer_start, end);
5887 /* Free head and tail (if any) */
5888 if (start != outer_start)
5889 free_contig_range(outer_start, start - outer_start);
5890 if (end != outer_end)
5891 free_contig_range(end, outer_end - end);
5894 undo_isolate_page_range(pfn_max_align_down(start),
5895 pfn_max_align_up(end), migratetype);
5899 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5901 for (; nr_pages--; ++pfn)
5902 __free_page(pfn_to_page(pfn));
5906 #ifdef CONFIG_MEMORY_HOTPLUG
5907 static int __meminit __zone_pcp_update(void *data)
5909 struct zone *zone = data;
5911 unsigned long batch = zone_batchsize(zone), flags;
5913 for_each_possible_cpu(cpu) {
5914 struct per_cpu_pageset *pset;
5915 struct per_cpu_pages *pcp;
5917 pset = per_cpu_ptr(zone->pageset, cpu);
5920 local_irq_save(flags);
5922 free_pcppages_bulk(zone, pcp->count, pcp);
5923 drain_zonestat(zone, pset);
5924 setup_pageset(pset, batch);
5925 local_irq_restore(flags);
5930 void __meminit zone_pcp_update(struct zone *zone)
5932 stop_machine(__zone_pcp_update, zone, NULL);
5936 void zone_pcp_reset(struct zone *zone)
5938 unsigned long flags;
5940 struct per_cpu_pageset *pset;
5942 /* avoid races with drain_pages() */
5943 local_irq_save(flags);
5944 if (zone->pageset != &boot_pageset) {
5945 for_each_online_cpu(cpu) {
5946 pset = per_cpu_ptr(zone->pageset, cpu);
5947 drain_zonestat(zone, pset);
5949 free_percpu(zone->pageset);
5950 zone->pageset = &boot_pageset;
5952 local_irq_restore(flags);
5955 #ifdef CONFIG_MEMORY_HOTREMOVE
5957 * All pages in the range must be isolated before calling this.
5960 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5966 unsigned long flags;
5967 /* find the first valid pfn */
5968 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5973 zone = page_zone(pfn_to_page(pfn));
5974 spin_lock_irqsave(&zone->lock, flags);
5976 while (pfn < end_pfn) {
5977 if (!pfn_valid(pfn)) {
5981 page = pfn_to_page(pfn);
5983 * The HWPoisoned page may be not in buddy system, and
5984 * page_count() is not 0.
5986 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
5988 SetPageReserved(page);
5992 BUG_ON(page_count(page));
5993 BUG_ON(!PageBuddy(page));
5994 order = page_order(page);
5995 #ifdef CONFIG_DEBUG_VM
5996 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5997 pfn, 1 << order, end_pfn);
5999 list_del(&page->lru);
6000 rmv_page_order(page);
6001 zone->free_area[order].nr_free--;
6002 for (i = 0; i < (1 << order); i++)
6003 SetPageReserved((page+i));
6004 pfn += (1 << order);
6006 spin_unlock_irqrestore(&zone->lock, flags);
6010 #ifdef CONFIG_MEMORY_FAILURE
6011 bool is_free_buddy_page(struct page *page)
6013 struct zone *zone = page_zone(page);
6014 unsigned long pfn = page_to_pfn(page);
6015 unsigned long flags;
6018 spin_lock_irqsave(&zone->lock, flags);
6019 for (order = 0; order < MAX_ORDER; order++) {
6020 struct page *page_head = page - (pfn & ((1 << order) - 1));
6022 if (PageBuddy(page_head) && page_order(page_head) >= order)
6025 spin_unlock_irqrestore(&zone->lock, flags);
6027 return order < MAX_ORDER;
6031 static const struct trace_print_flags pageflag_names[] = {
6032 {1UL << PG_locked, "locked" },
6033 {1UL << PG_error, "error" },
6034 {1UL << PG_referenced, "referenced" },
6035 {1UL << PG_uptodate, "uptodate" },
6036 {1UL << PG_dirty, "dirty" },
6037 {1UL << PG_lru, "lru" },
6038 {1UL << PG_active, "active" },
6039 {1UL << PG_slab, "slab" },
6040 {1UL << PG_owner_priv_1, "owner_priv_1" },
6041 {1UL << PG_arch_1, "arch_1" },
6042 {1UL << PG_reserved, "reserved" },
6043 {1UL << PG_private, "private" },
6044 {1UL << PG_private_2, "private_2" },
6045 {1UL << PG_writeback, "writeback" },
6046 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6047 {1UL << PG_head, "head" },
6048 {1UL << PG_tail, "tail" },
6050 {1UL << PG_compound, "compound" },
6052 {1UL << PG_swapcache, "swapcache" },
6053 {1UL << PG_mappedtodisk, "mappedtodisk" },
6054 {1UL << PG_reclaim, "reclaim" },
6055 {1UL << PG_swapbacked, "swapbacked" },
6056 {1UL << PG_unevictable, "unevictable" },
6058 {1UL << PG_mlocked, "mlocked" },
6060 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6061 {1UL << PG_uncached, "uncached" },
6063 #ifdef CONFIG_MEMORY_FAILURE
6064 {1UL << PG_hwpoison, "hwpoison" },
6066 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6067 {1UL << PG_compound_lock, "compound_lock" },
6071 static void dump_page_flags(unsigned long flags)
6073 const char *delim = "";
6077 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6079 printk(KERN_ALERT "page flags: %#lx(", flags);
6081 /* remove zone id */
6082 flags &= (1UL << NR_PAGEFLAGS) - 1;
6084 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6086 mask = pageflag_names[i].mask;
6087 if ((flags & mask) != mask)
6091 printk("%s%s", delim, pageflag_names[i].name);
6095 /* check for left over flags */
6097 printk("%s%#lx", delim, flags);
6102 void dump_page(struct page *page)
6105 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6106 page, atomic_read(&page->_count), page_mapcount(page),
6107 page->mapping, page->index);
6108 dump_page_flags(page->flags);
6109 mem_cgroup_print_bad_page(page);