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/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74 DEFINE_PER_CPU(int, numa_node);
75 EXPORT_PER_CPU_SYMBOL(numa_node);
78 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
85 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
86 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
90 * Array of node states.
92 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
93 [N_POSSIBLE] = NODE_MASK_ALL,
94 [N_ONLINE] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY] = { { [0] = 1UL } },
103 [N_CPU] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock);
111 unsigned long totalram_pages __read_mostly;
112 unsigned long totalreserve_pages __read_mostly;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly;
121 int percpu_pagelist_fraction;
122 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex));
139 if (saved_gfp_mask) {
140 gfp_allowed_mask = saved_gfp_mask;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex));
148 WARN_ON(saved_gfp_mask);
149 saved_gfp_mask = gfp_allowed_mask;
150 gfp_allowed_mask &= ~GFP_IOFS;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly;
165 static void __free_pages_ok(struct page *page, unsigned int order);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages);
193 static char * const zone_names[MAX_NR_ZONES] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes = 1024;
208 int user_min_free_kbytes = -1;
210 static unsigned long __meminitdata nr_kernel_pages;
211 static unsigned long __meminitdata nr_all_pages;
212 static unsigned long __meminitdata dma_reserve;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
217 static unsigned long __initdata required_kernelcore;
218 static unsigned long __initdata required_movablecore;
219 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly = MAX_NUMNODES;
228 int nr_online_nodes __read_mostly = 1;
229 EXPORT_SYMBOL(nr_node_ids);
230 EXPORT_SYMBOL(nr_online_nodes);
233 int page_group_by_mobility_disabled __read_mostly;
235 void set_pageblock_migratetype(struct page *page, int migratetype)
237 if (unlikely(page_group_by_mobility_disabled &&
238 migratetype < MIGRATE_PCPTYPES))
239 migratetype = MIGRATE_UNMOVABLE;
241 set_pageblock_flags_group(page, (unsigned long)migratetype,
242 PB_migrate, PB_migrate_end);
245 bool oom_killer_disabled __read_mostly;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
252 unsigned long pfn = page_to_pfn(page);
253 unsigned long sp, start_pfn;
256 seq = zone_span_seqbegin(zone);
257 start_pfn = zone->zone_start_pfn;
258 sp = zone->spanned_pages;
259 if (!zone_spans_pfn(zone, pfn))
261 } while (zone_span_seqretry(zone, seq));
264 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
265 pfn, zone_to_nid(zone), zone->name,
266 start_pfn, start_pfn + sp);
271 static int page_is_consistent(struct zone *zone, struct page *page)
273 if (!pfn_valid_within(page_to_pfn(page)))
275 if (zone != page_zone(page))
281 * Temporary debugging check for pages not lying within a given zone.
283 static int bad_range(struct zone *zone, struct page *page)
285 if (page_outside_zone_boundaries(zone, page))
287 if (!page_is_consistent(zone, page))
293 static inline int bad_range(struct zone *zone, struct page *page)
299 static void bad_page(struct page *page, const char *reason,
300 unsigned long bad_flags)
302 static unsigned long resume;
303 static unsigned long nr_shown;
304 static unsigned long nr_unshown;
306 /* Don't complain about poisoned pages */
307 if (PageHWPoison(page)) {
308 page_mapcount_reset(page); /* remove PageBuddy */
313 * Allow a burst of 60 reports, then keep quiet for that minute;
314 * or allow a steady drip of one report per second.
316 if (nr_shown == 60) {
317 if (time_before(jiffies, resume)) {
323 "BUG: Bad page state: %lu messages suppressed\n",
330 resume = jiffies + 60 * HZ;
332 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
333 current->comm, page_to_pfn(page));
334 dump_page_badflags(page, reason, bad_flags);
339 /* Leave bad fields for debug, except PageBuddy could make trouble */
340 page_mapcount_reset(page); /* remove PageBuddy */
341 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
345 * Higher-order pages are called "compound pages". They are structured thusly:
347 * The first PAGE_SIZE page is called the "head page".
349 * The remaining PAGE_SIZE pages are called "tail pages".
351 * All pages have PG_compound set. All tail pages have their ->first_page
352 * pointing at the head page.
354 * The first tail page's ->lru.next holds the address of the compound page's
355 * put_page() function. Its ->lru.prev holds the order of allocation.
356 * This usage means that zero-order pages may not be compound.
359 static void free_compound_page(struct page *page)
361 __free_pages_ok(page, compound_order(page));
364 void prep_compound_page(struct page *page, unsigned long order)
367 int nr_pages = 1 << order;
369 set_compound_page_dtor(page, free_compound_page);
370 set_compound_order(page, order);
372 for (i = 1; i < nr_pages; i++) {
373 struct page *p = page + i;
374 set_page_count(p, 0);
375 p->first_page = page;
376 /* Make sure p->first_page is always valid for PageTail() */
382 /* update __split_huge_page_refcount if you change this function */
383 static int destroy_compound_page(struct page *page, unsigned long order)
386 int nr_pages = 1 << order;
389 if (unlikely(compound_order(page) != order)) {
390 bad_page(page, "wrong compound order", 0);
394 __ClearPageHead(page);
396 for (i = 1; i < nr_pages; i++) {
397 struct page *p = page + i;
399 if (unlikely(!PageTail(p))) {
400 bad_page(page, "PageTail not set", 0);
402 } else if (unlikely(p->first_page != page)) {
403 bad_page(page, "first_page not consistent", 0);
412 static inline void prep_zero_page(struct page *page, unsigned int order,
418 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
419 * and __GFP_HIGHMEM from hard or soft interrupt context.
421 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
422 for (i = 0; i < (1 << order); i++)
423 clear_highpage(page + i);
426 #ifdef CONFIG_DEBUG_PAGEALLOC
427 unsigned int _debug_guardpage_minorder;
429 static int __init debug_guardpage_minorder_setup(char *buf)
433 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
434 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
437 _debug_guardpage_minorder = res;
438 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
441 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
443 static inline void set_page_guard_flag(struct page *page)
445 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
448 static inline void clear_page_guard_flag(struct page *page)
450 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
453 static inline void set_page_guard_flag(struct page *page) { }
454 static inline void clear_page_guard_flag(struct page *page) { }
457 static inline void set_page_order(struct page *page, unsigned int order)
459 set_page_private(page, order);
460 __SetPageBuddy(page);
463 static inline void rmv_page_order(struct page *page)
465 __ClearPageBuddy(page);
466 set_page_private(page, 0);
470 * Locate the struct page for both the matching buddy in our
471 * pair (buddy1) and the combined O(n+1) page they form (page).
473 * 1) Any buddy B1 will have an order O twin B2 which satisfies
474 * the following equation:
476 * For example, if the starting buddy (buddy2) is #8 its order
478 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
480 * 2) Any buddy B will have an order O+1 parent P which
481 * satisfies the following equation:
484 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
486 static inline unsigned long
487 __find_buddy_index(unsigned long page_idx, unsigned int order)
489 return page_idx ^ (1 << order);
493 * This function checks whether a page is free && is the buddy
494 * we can do coalesce a page and its buddy if
495 * (a) the buddy is not in a hole &&
496 * (b) the buddy is in the buddy system &&
497 * (c) a page and its buddy have the same order &&
498 * (d) a page and its buddy are in the same zone.
500 * For recording whether a page is in the buddy system, we set ->_mapcount
501 * PAGE_BUDDY_MAPCOUNT_VALUE.
502 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
503 * serialized by zone->lock.
505 * For recording page's order, we use page_private(page).
507 static inline int page_is_buddy(struct page *page, struct page *buddy,
510 if (!pfn_valid_within(page_to_pfn(buddy)))
513 if (page_is_guard(buddy) && page_order(buddy) == order) {
514 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
516 if (page_zone_id(page) != page_zone_id(buddy))
522 if (PageBuddy(buddy) && page_order(buddy) == order) {
523 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
526 * zone check is done late to avoid uselessly
527 * calculating zone/node ids for pages that could
530 if (page_zone_id(page) != page_zone_id(buddy))
539 * Freeing function for a buddy system allocator.
541 * The concept of a buddy system is to maintain direct-mapped table
542 * (containing bit values) for memory blocks of various "orders".
543 * The bottom level table contains the map for the smallest allocatable
544 * units of memory (here, pages), and each level above it describes
545 * pairs of units from the levels below, hence, "buddies".
546 * At a high level, all that happens here is marking the table entry
547 * at the bottom level available, and propagating the changes upward
548 * as necessary, plus some accounting needed to play nicely with other
549 * parts of the VM system.
550 * At each level, we keep a list of pages, which are heads of continuous
551 * free pages of length of (1 << order) and marked with _mapcount
552 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
554 * So when we are allocating or freeing one, we can derive the state of the
555 * other. That is, if we allocate a small block, and both were
556 * free, the remainder of the region must be split into blocks.
557 * If a block is freed, and its buddy is also free, then this
558 * triggers coalescing into a block of larger size.
563 static inline void __free_one_page(struct page *page,
565 struct zone *zone, unsigned int order,
568 unsigned long page_idx;
569 unsigned long combined_idx;
570 unsigned long uninitialized_var(buddy_idx);
573 VM_BUG_ON(!zone_is_initialized(zone));
575 if (unlikely(PageCompound(page)))
576 if (unlikely(destroy_compound_page(page, order)))
579 VM_BUG_ON(migratetype == -1);
581 page_idx = pfn & ((1 << MAX_ORDER) - 1);
583 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
584 VM_BUG_ON_PAGE(bad_range(zone, page), page);
586 while (order < MAX_ORDER-1) {
587 buddy_idx = __find_buddy_index(page_idx, order);
588 buddy = page + (buddy_idx - page_idx);
589 if (!page_is_buddy(page, buddy, order))
592 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
593 * merge with it and move up one order.
595 if (page_is_guard(buddy)) {
596 clear_page_guard_flag(buddy);
597 set_page_private(page, 0);
598 __mod_zone_freepage_state(zone, 1 << order,
601 list_del(&buddy->lru);
602 zone->free_area[order].nr_free--;
603 rmv_page_order(buddy);
605 combined_idx = buddy_idx & page_idx;
606 page = page + (combined_idx - page_idx);
607 page_idx = combined_idx;
610 set_page_order(page, order);
613 * If this is not the largest possible page, check if the buddy
614 * of the next-highest order is free. If it is, it's possible
615 * that pages are being freed that will coalesce soon. In case,
616 * that is happening, add the free page to the tail of the list
617 * so it's less likely to be used soon and more likely to be merged
618 * as a higher order page
620 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
621 struct page *higher_page, *higher_buddy;
622 combined_idx = buddy_idx & page_idx;
623 higher_page = page + (combined_idx - page_idx);
624 buddy_idx = __find_buddy_index(combined_idx, order + 1);
625 higher_buddy = higher_page + (buddy_idx - combined_idx);
626 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
627 list_add_tail(&page->lru,
628 &zone->free_area[order].free_list[migratetype]);
633 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
635 zone->free_area[order].nr_free++;
638 static inline int free_pages_check(struct page *page)
640 const char *bad_reason = NULL;
641 unsigned long bad_flags = 0;
643 if (unlikely(page_mapcount(page)))
644 bad_reason = "nonzero mapcount";
645 if (unlikely(page->mapping != NULL))
646 bad_reason = "non-NULL mapping";
647 if (unlikely(atomic_read(&page->_count) != 0))
648 bad_reason = "nonzero _count";
649 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
650 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
651 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
653 if (unlikely(mem_cgroup_bad_page_check(page)))
654 bad_reason = "cgroup check failed";
655 if (unlikely(bad_reason)) {
656 bad_page(page, bad_reason, bad_flags);
659 page_cpupid_reset_last(page);
660 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
661 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
666 * Frees a number of pages from the PCP lists
667 * Assumes all pages on list are in same zone, and of same order.
668 * count is the number of pages to free.
670 * If the zone was previously in an "all pages pinned" state then look to
671 * see if this freeing clears that state.
673 * And clear the zone's pages_scanned counter, to hold off the "all pages are
674 * pinned" detection logic.
676 static void free_pcppages_bulk(struct zone *zone, int count,
677 struct per_cpu_pages *pcp)
683 spin_lock(&zone->lock);
684 zone->pages_scanned = 0;
688 struct list_head *list;
691 * Remove pages from lists in a round-robin fashion. A
692 * batch_free count is maintained that is incremented when an
693 * empty list is encountered. This is so more pages are freed
694 * off fuller lists instead of spinning excessively around empty
699 if (++migratetype == MIGRATE_PCPTYPES)
701 list = &pcp->lists[migratetype];
702 } while (list_empty(list));
704 /* This is the only non-empty list. Free them all. */
705 if (batch_free == MIGRATE_PCPTYPES)
706 batch_free = to_free;
709 int mt; /* migratetype of the to-be-freed page */
711 page = list_entry(list->prev, struct page, lru);
712 /* must delete as __free_one_page list manipulates */
713 list_del(&page->lru);
714 mt = get_freepage_migratetype(page);
715 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
716 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
717 trace_mm_page_pcpu_drain(page, 0, mt);
718 if (likely(!is_migrate_isolate_page(page))) {
719 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
720 if (is_migrate_cma(mt))
721 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
723 } while (--to_free && --batch_free && !list_empty(list));
725 spin_unlock(&zone->lock);
728 static void free_one_page(struct zone *zone,
729 struct page *page, unsigned long pfn,
733 spin_lock(&zone->lock);
734 zone->pages_scanned = 0;
736 __free_one_page(page, pfn, zone, order, migratetype);
737 if (unlikely(!is_migrate_isolate(migratetype)))
738 __mod_zone_freepage_state(zone, 1 << order, migratetype);
739 spin_unlock(&zone->lock);
742 static bool free_pages_prepare(struct page *page, unsigned int order)
747 trace_mm_page_free(page, order);
748 kmemcheck_free_shadow(page, order);
751 page->mapping = NULL;
752 for (i = 0; i < (1 << order); i++)
753 bad += free_pages_check(page + i);
757 if (!PageHighMem(page)) {
758 debug_check_no_locks_freed(page_address(page),
760 debug_check_no_obj_freed(page_address(page),
763 arch_free_page(page, order);
764 kernel_map_pages(page, 1 << order, 0);
769 static void __free_pages_ok(struct page *page, unsigned int order)
773 unsigned long pfn = page_to_pfn(page);
775 if (!free_pages_prepare(page, order))
778 migratetype = get_pfnblock_migratetype(page, pfn);
779 local_irq_save(flags);
780 __count_vm_events(PGFREE, 1 << order);
781 set_freepage_migratetype(page, migratetype);
782 free_one_page(page_zone(page), page, pfn, order, migratetype);
783 local_irq_restore(flags);
786 void __init __free_pages_bootmem(struct page *page, unsigned int order)
788 unsigned int nr_pages = 1 << order;
789 struct page *p = page;
793 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
795 __ClearPageReserved(p);
796 set_page_count(p, 0);
798 __ClearPageReserved(p);
799 set_page_count(p, 0);
801 page_zone(page)->managed_pages += nr_pages;
802 set_page_refcounted(page);
803 __free_pages(page, order);
807 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
808 void __init init_cma_reserved_pageblock(struct page *page)
810 unsigned i = pageblock_nr_pages;
811 struct page *p = page;
814 __ClearPageReserved(p);
815 set_page_count(p, 0);
818 set_page_refcounted(page);
819 set_pageblock_migratetype(page, MIGRATE_CMA);
820 __free_pages(page, pageblock_order);
821 adjust_managed_page_count(page, pageblock_nr_pages);
826 * The order of subdivision here is critical for the IO subsystem.
827 * Please do not alter this order without good reasons and regression
828 * testing. Specifically, as large blocks of memory are subdivided,
829 * the order in which smaller blocks are delivered depends on the order
830 * they're subdivided in this function. This is the primary factor
831 * influencing the order in which pages are delivered to the IO
832 * subsystem according to empirical testing, and this is also justified
833 * by considering the behavior of a buddy system containing a single
834 * large block of memory acted on by a series of small allocations.
835 * This behavior is a critical factor in sglist merging's success.
839 static inline void expand(struct zone *zone, struct page *page,
840 int low, int high, struct free_area *area,
843 unsigned long size = 1 << high;
849 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
851 #ifdef CONFIG_DEBUG_PAGEALLOC
852 if (high < debug_guardpage_minorder()) {
854 * Mark as guard pages (or page), that will allow to
855 * merge back to allocator when buddy will be freed.
856 * Corresponding page table entries will not be touched,
857 * pages will stay not present in virtual address space
859 INIT_LIST_HEAD(&page[size].lru);
860 set_page_guard_flag(&page[size]);
861 set_page_private(&page[size], high);
862 /* Guard pages are not available for any usage */
863 __mod_zone_freepage_state(zone, -(1 << high),
868 list_add(&page[size].lru, &area->free_list[migratetype]);
870 set_page_order(&page[size], high);
875 * This page is about to be returned from the page allocator
877 static inline int check_new_page(struct page *page)
879 const char *bad_reason = NULL;
880 unsigned long bad_flags = 0;
882 if (unlikely(page_mapcount(page)))
883 bad_reason = "nonzero mapcount";
884 if (unlikely(page->mapping != NULL))
885 bad_reason = "non-NULL mapping";
886 if (unlikely(atomic_read(&page->_count) != 0))
887 bad_reason = "nonzero _count";
888 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
889 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
890 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
892 if (unlikely(mem_cgroup_bad_page_check(page)))
893 bad_reason = "cgroup check failed";
894 if (unlikely(bad_reason)) {
895 bad_page(page, bad_reason, bad_flags);
901 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags)
905 for (i = 0; i < (1 << order); i++) {
906 struct page *p = page + i;
907 if (unlikely(check_new_page(p)))
911 set_page_private(page, 0);
912 set_page_refcounted(page);
914 arch_alloc_page(page, order);
915 kernel_map_pages(page, 1 << order, 1);
917 if (gfp_flags & __GFP_ZERO)
918 prep_zero_page(page, order, gfp_flags);
920 if (order && (gfp_flags & __GFP_COMP))
921 prep_compound_page(page, order);
927 * Go through the free lists for the given migratetype and remove
928 * the smallest available page from the freelists
931 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
934 unsigned int current_order;
935 struct free_area *area;
938 /* Find a page of the appropriate size in the preferred list */
939 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
940 area = &(zone->free_area[current_order]);
941 if (list_empty(&area->free_list[migratetype]))
944 page = list_entry(area->free_list[migratetype].next,
946 list_del(&page->lru);
947 rmv_page_order(page);
949 expand(zone, page, order, current_order, area, migratetype);
950 set_freepage_migratetype(page, migratetype);
959 * This array describes the order lists are fallen back to when
960 * the free lists for the desirable migrate type are depleted
962 static int fallbacks[MIGRATE_TYPES][4] = {
963 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
964 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
966 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
967 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
969 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
971 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
972 #ifdef CONFIG_MEMORY_ISOLATION
973 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
978 * Move the free pages in a range to the free lists of the requested type.
979 * Note that start_page and end_pages are not aligned on a pageblock
980 * boundary. If alignment is required, use move_freepages_block()
982 int move_freepages(struct zone *zone,
983 struct page *start_page, struct page *end_page,
990 #ifndef CONFIG_HOLES_IN_ZONE
992 * page_zone is not safe to call in this context when
993 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
994 * anyway as we check zone boundaries in move_freepages_block().
995 * Remove at a later date when no bug reports exist related to
996 * grouping pages by mobility
998 BUG_ON(page_zone(start_page) != page_zone(end_page));
1001 for (page = start_page; page <= end_page;) {
1002 /* Make sure we are not inadvertently changing nodes */
1003 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1005 if (!pfn_valid_within(page_to_pfn(page))) {
1010 if (!PageBuddy(page)) {
1015 order = page_order(page);
1016 list_move(&page->lru,
1017 &zone->free_area[order].free_list[migratetype]);
1018 set_freepage_migratetype(page, migratetype);
1020 pages_moved += 1 << order;
1026 int move_freepages_block(struct zone *zone, struct page *page,
1029 unsigned long start_pfn, end_pfn;
1030 struct page *start_page, *end_page;
1032 start_pfn = page_to_pfn(page);
1033 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1034 start_page = pfn_to_page(start_pfn);
1035 end_page = start_page + pageblock_nr_pages - 1;
1036 end_pfn = start_pfn + pageblock_nr_pages - 1;
1038 /* Do not cross zone boundaries */
1039 if (!zone_spans_pfn(zone, start_pfn))
1041 if (!zone_spans_pfn(zone, end_pfn))
1044 return move_freepages(zone, start_page, end_page, migratetype);
1047 static void change_pageblock_range(struct page *pageblock_page,
1048 int start_order, int migratetype)
1050 int nr_pageblocks = 1 << (start_order - pageblock_order);
1052 while (nr_pageblocks--) {
1053 set_pageblock_migratetype(pageblock_page, migratetype);
1054 pageblock_page += pageblock_nr_pages;
1059 * If breaking a large block of pages, move all free pages to the preferred
1060 * allocation list. If falling back for a reclaimable kernel allocation, be
1061 * more aggressive about taking ownership of free pages.
1063 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1064 * nor move CMA pages to different free lists. We don't want unmovable pages
1065 * to be allocated from MIGRATE_CMA areas.
1067 * Returns the new migratetype of the pageblock (or the same old migratetype
1068 * if it was unchanged).
1070 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1071 int start_type, int fallback_type)
1073 int current_order = page_order(page);
1076 * When borrowing from MIGRATE_CMA, we need to release the excess
1077 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1078 * is set to CMA so it is returned to the correct freelist in case
1079 * the page ends up being not actually allocated from the pcp lists.
1081 if (is_migrate_cma(fallback_type))
1082 return fallback_type;
1084 /* Take ownership for orders >= pageblock_order */
1085 if (current_order >= pageblock_order) {
1086 change_pageblock_range(page, current_order, start_type);
1090 if (current_order >= pageblock_order / 2 ||
1091 start_type == MIGRATE_RECLAIMABLE ||
1092 page_group_by_mobility_disabled) {
1095 pages = move_freepages_block(zone, page, start_type);
1097 /* Claim the whole block if over half of it is free */
1098 if (pages >= (1 << (pageblock_order-1)) ||
1099 page_group_by_mobility_disabled) {
1101 set_pageblock_migratetype(page, start_type);
1107 return fallback_type;
1110 /* Remove an element from the buddy allocator from the fallback list */
1111 static inline struct page *
1112 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1114 struct free_area *area;
1115 unsigned int current_order;
1117 int migratetype, new_type, i;
1119 /* Find the largest possible block of pages in the other list */
1120 for (current_order = MAX_ORDER-1;
1121 current_order >= order && current_order <= MAX_ORDER-1;
1124 migratetype = fallbacks[start_migratetype][i];
1126 /* MIGRATE_RESERVE handled later if necessary */
1127 if (migratetype == MIGRATE_RESERVE)
1130 area = &(zone->free_area[current_order]);
1131 if (list_empty(&area->free_list[migratetype]))
1134 page = list_entry(area->free_list[migratetype].next,
1138 new_type = try_to_steal_freepages(zone, page,
1142 /* Remove the page from the freelists */
1143 list_del(&page->lru);
1144 rmv_page_order(page);
1146 expand(zone, page, order, current_order, area,
1148 /* The freepage_migratetype may differ from pageblock's
1149 * migratetype depending on the decisions in
1150 * try_to_steal_freepages. This is OK as long as it does
1151 * not differ for MIGRATE_CMA type.
1153 set_freepage_migratetype(page, new_type);
1155 trace_mm_page_alloc_extfrag(page, order, current_order,
1156 start_migratetype, migratetype, new_type);
1166 * Do the hard work of removing an element from the buddy allocator.
1167 * Call me with the zone->lock already held.
1169 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1175 page = __rmqueue_smallest(zone, order, migratetype);
1177 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1178 page = __rmqueue_fallback(zone, order, migratetype);
1181 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1182 * is used because __rmqueue_smallest is an inline function
1183 * and we want just one call site
1186 migratetype = MIGRATE_RESERVE;
1191 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1196 * Obtain a specified number of elements from the buddy allocator, all under
1197 * a single hold of the lock, for efficiency. Add them to the supplied list.
1198 * Returns the number of new pages which were placed at *list.
1200 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1201 unsigned long count, struct list_head *list,
1202 int migratetype, bool cold)
1206 spin_lock(&zone->lock);
1207 for (i = 0; i < count; ++i) {
1208 struct page *page = __rmqueue(zone, order, migratetype);
1209 if (unlikely(page == NULL))
1213 * Split buddy pages returned by expand() are received here
1214 * in physical page order. The page is added to the callers and
1215 * list and the list head then moves forward. From the callers
1216 * perspective, the linked list is ordered by page number in
1217 * some conditions. This is useful for IO devices that can
1218 * merge IO requests if the physical pages are ordered
1222 list_add(&page->lru, list);
1224 list_add_tail(&page->lru, list);
1226 if (is_migrate_cma(get_freepage_migratetype(page)))
1227 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1230 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1231 spin_unlock(&zone->lock);
1237 * Called from the vmstat counter updater to drain pagesets of this
1238 * currently executing processor on remote nodes after they have
1241 * Note that this function must be called with the thread pinned to
1242 * a single processor.
1244 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1246 unsigned long flags;
1248 unsigned long batch;
1250 local_irq_save(flags);
1251 batch = ACCESS_ONCE(pcp->batch);
1252 if (pcp->count >= batch)
1255 to_drain = pcp->count;
1257 free_pcppages_bulk(zone, to_drain, pcp);
1258 pcp->count -= to_drain;
1260 local_irq_restore(flags);
1265 * Drain pages of the indicated processor.
1267 * The processor must either be the current processor and the
1268 * thread pinned to the current processor or a processor that
1271 static void drain_pages(unsigned int cpu)
1273 unsigned long flags;
1276 for_each_populated_zone(zone) {
1277 struct per_cpu_pageset *pset;
1278 struct per_cpu_pages *pcp;
1280 local_irq_save(flags);
1281 pset = per_cpu_ptr(zone->pageset, cpu);
1285 free_pcppages_bulk(zone, pcp->count, pcp);
1288 local_irq_restore(flags);
1293 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1295 void drain_local_pages(void *arg)
1297 drain_pages(smp_processor_id());
1301 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1303 * Note that this code is protected against sending an IPI to an offline
1304 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1305 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1306 * nothing keeps CPUs from showing up after we populated the cpumask and
1307 * before the call to on_each_cpu_mask().
1309 void drain_all_pages(void)
1312 struct per_cpu_pageset *pcp;
1316 * Allocate in the BSS so we wont require allocation in
1317 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1319 static cpumask_t cpus_with_pcps;
1322 * We don't care about racing with CPU hotplug event
1323 * as offline notification will cause the notified
1324 * cpu to drain that CPU pcps and on_each_cpu_mask
1325 * disables preemption as part of its processing
1327 for_each_online_cpu(cpu) {
1328 bool has_pcps = false;
1329 for_each_populated_zone(zone) {
1330 pcp = per_cpu_ptr(zone->pageset, cpu);
1331 if (pcp->pcp.count) {
1337 cpumask_set_cpu(cpu, &cpus_with_pcps);
1339 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1341 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1344 #ifdef CONFIG_HIBERNATION
1346 void mark_free_pages(struct zone *zone)
1348 unsigned long pfn, max_zone_pfn;
1349 unsigned long flags;
1350 unsigned int order, t;
1351 struct list_head *curr;
1353 if (zone_is_empty(zone))
1356 spin_lock_irqsave(&zone->lock, flags);
1358 max_zone_pfn = zone_end_pfn(zone);
1359 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1360 if (pfn_valid(pfn)) {
1361 struct page *page = pfn_to_page(pfn);
1363 if (!swsusp_page_is_forbidden(page))
1364 swsusp_unset_page_free(page);
1367 for_each_migratetype_order(order, t) {
1368 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1371 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1372 for (i = 0; i < (1UL << order); i++)
1373 swsusp_set_page_free(pfn_to_page(pfn + i));
1376 spin_unlock_irqrestore(&zone->lock, flags);
1378 #endif /* CONFIG_PM */
1381 * Free a 0-order page
1382 * cold == true ? free a cold page : free a hot page
1384 void free_hot_cold_page(struct page *page, bool cold)
1386 struct zone *zone = page_zone(page);
1387 struct per_cpu_pages *pcp;
1388 unsigned long flags;
1389 unsigned long pfn = page_to_pfn(page);
1392 if (!free_pages_prepare(page, 0))
1395 migratetype = get_pfnblock_migratetype(page, pfn);
1396 set_freepage_migratetype(page, migratetype);
1397 local_irq_save(flags);
1398 __count_vm_event(PGFREE);
1401 * We only track unmovable, reclaimable and movable on pcp lists.
1402 * Free ISOLATE pages back to the allocator because they are being
1403 * offlined but treat RESERVE as movable pages so we can get those
1404 * areas back if necessary. Otherwise, we may have to free
1405 * excessively into the page allocator
1407 if (migratetype >= MIGRATE_PCPTYPES) {
1408 if (unlikely(is_migrate_isolate(migratetype))) {
1409 free_one_page(zone, page, pfn, 0, migratetype);
1412 migratetype = MIGRATE_MOVABLE;
1415 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1417 list_add(&page->lru, &pcp->lists[migratetype]);
1419 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1421 if (pcp->count >= pcp->high) {
1422 unsigned long batch = ACCESS_ONCE(pcp->batch);
1423 free_pcppages_bulk(zone, batch, pcp);
1424 pcp->count -= batch;
1428 local_irq_restore(flags);
1432 * Free a list of 0-order pages
1434 void free_hot_cold_page_list(struct list_head *list, bool cold)
1436 struct page *page, *next;
1438 list_for_each_entry_safe(page, next, list, lru) {
1439 trace_mm_page_free_batched(page, cold);
1440 free_hot_cold_page(page, cold);
1445 * split_page takes a non-compound higher-order page, and splits it into
1446 * n (1<<order) sub-pages: page[0..n]
1447 * Each sub-page must be freed individually.
1449 * Note: this is probably too low level an operation for use in drivers.
1450 * Please consult with lkml before using this in your driver.
1452 void split_page(struct page *page, unsigned int order)
1456 VM_BUG_ON_PAGE(PageCompound(page), page);
1457 VM_BUG_ON_PAGE(!page_count(page), page);
1459 #ifdef CONFIG_KMEMCHECK
1461 * Split shadow pages too, because free(page[0]) would
1462 * otherwise free the whole shadow.
1464 if (kmemcheck_page_is_tracked(page))
1465 split_page(virt_to_page(page[0].shadow), order);
1468 for (i = 1; i < (1 << order); i++)
1469 set_page_refcounted(page + i);
1471 EXPORT_SYMBOL_GPL(split_page);
1473 static int __isolate_free_page(struct page *page, unsigned int order)
1475 unsigned long watermark;
1479 BUG_ON(!PageBuddy(page));
1481 zone = page_zone(page);
1482 mt = get_pageblock_migratetype(page);
1484 if (!is_migrate_isolate(mt)) {
1485 /* Obey watermarks as if the page was being allocated */
1486 watermark = low_wmark_pages(zone) + (1 << order);
1487 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1490 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1493 /* Remove page from free list */
1494 list_del(&page->lru);
1495 zone->free_area[order].nr_free--;
1496 rmv_page_order(page);
1498 /* Set the pageblock if the isolated page is at least a pageblock */
1499 if (order >= pageblock_order - 1) {
1500 struct page *endpage = page + (1 << order) - 1;
1501 for (; page < endpage; page += pageblock_nr_pages) {
1502 int mt = get_pageblock_migratetype(page);
1503 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1504 set_pageblock_migratetype(page,
1509 return 1UL << order;
1513 * Similar to split_page except the page is already free. As this is only
1514 * being used for migration, the migratetype of the block also changes.
1515 * As this is called with interrupts disabled, the caller is responsible
1516 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1519 * Note: this is probably too low level an operation for use in drivers.
1520 * Please consult with lkml before using this in your driver.
1522 int split_free_page(struct page *page)
1527 order = page_order(page);
1529 nr_pages = __isolate_free_page(page, order);
1533 /* Split into individual pages */
1534 set_page_refcounted(page);
1535 split_page(page, order);
1540 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1541 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1545 struct page *buffered_rmqueue(struct zone *preferred_zone,
1546 struct zone *zone, unsigned int order,
1547 gfp_t gfp_flags, int migratetype)
1549 unsigned long flags;
1551 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1554 if (likely(order == 0)) {
1555 struct per_cpu_pages *pcp;
1556 struct list_head *list;
1558 local_irq_save(flags);
1559 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1560 list = &pcp->lists[migratetype];
1561 if (list_empty(list)) {
1562 pcp->count += rmqueue_bulk(zone, 0,
1565 if (unlikely(list_empty(list)))
1570 page = list_entry(list->prev, struct page, lru);
1572 page = list_entry(list->next, struct page, lru);
1574 list_del(&page->lru);
1577 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1579 * __GFP_NOFAIL is not to be used in new code.
1581 * All __GFP_NOFAIL callers should be fixed so that they
1582 * properly detect and handle allocation failures.
1584 * We most definitely don't want callers attempting to
1585 * allocate greater than order-1 page units with
1588 WARN_ON_ONCE(order > 1);
1590 spin_lock_irqsave(&zone->lock, flags);
1591 page = __rmqueue(zone, order, migratetype);
1592 spin_unlock(&zone->lock);
1595 __mod_zone_freepage_state(zone, -(1 << order),
1596 get_freepage_migratetype(page));
1599 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1601 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1602 zone_statistics(preferred_zone, zone, gfp_flags);
1603 local_irq_restore(flags);
1605 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1606 if (prep_new_page(page, order, gfp_flags))
1611 local_irq_restore(flags);
1615 #ifdef CONFIG_FAIL_PAGE_ALLOC
1618 struct fault_attr attr;
1620 u32 ignore_gfp_highmem;
1621 u32 ignore_gfp_wait;
1623 } fail_page_alloc = {
1624 .attr = FAULT_ATTR_INITIALIZER,
1625 .ignore_gfp_wait = 1,
1626 .ignore_gfp_highmem = 1,
1630 static int __init setup_fail_page_alloc(char *str)
1632 return setup_fault_attr(&fail_page_alloc.attr, str);
1634 __setup("fail_page_alloc=", setup_fail_page_alloc);
1636 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1638 if (order < fail_page_alloc.min_order)
1640 if (gfp_mask & __GFP_NOFAIL)
1642 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1644 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1647 return should_fail(&fail_page_alloc.attr, 1 << order);
1650 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1652 static int __init fail_page_alloc_debugfs(void)
1654 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1657 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1658 &fail_page_alloc.attr);
1660 return PTR_ERR(dir);
1662 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1663 &fail_page_alloc.ignore_gfp_wait))
1665 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1666 &fail_page_alloc.ignore_gfp_highmem))
1668 if (!debugfs_create_u32("min-order", mode, dir,
1669 &fail_page_alloc.min_order))
1674 debugfs_remove_recursive(dir);
1679 late_initcall(fail_page_alloc_debugfs);
1681 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1683 #else /* CONFIG_FAIL_PAGE_ALLOC */
1685 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1690 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1693 * Return true if free pages are above 'mark'. This takes into account the order
1694 * of the allocation.
1696 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1697 unsigned long mark, int classzone_idx, int alloc_flags,
1700 /* free_pages my go negative - that's OK */
1702 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1706 free_pages -= (1 << order) - 1;
1707 if (alloc_flags & ALLOC_HIGH)
1709 if (alloc_flags & ALLOC_HARDER)
1712 /* If allocation can't use CMA areas don't use free CMA pages */
1713 if (!(alloc_flags & ALLOC_CMA))
1714 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1717 if (free_pages - free_cma <= min + lowmem_reserve)
1719 for (o = 0; o < order; o++) {
1720 /* At the next order, this order's pages become unavailable */
1721 free_pages -= z->free_area[o].nr_free << o;
1723 /* Require fewer higher order pages to be free */
1726 if (free_pages <= min)
1732 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1733 int classzone_idx, int alloc_flags)
1735 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1736 zone_page_state(z, NR_FREE_PAGES));
1739 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1740 unsigned long mark, int classzone_idx, int alloc_flags)
1742 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1744 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1745 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1747 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1753 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1754 * skip over zones that are not allowed by the cpuset, or that have
1755 * been recently (in last second) found to be nearly full. See further
1756 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1757 * that have to skip over a lot of full or unallowed zones.
1759 * If the zonelist cache is present in the passed zonelist, then
1760 * returns a pointer to the allowed node mask (either the current
1761 * tasks mems_allowed, or node_states[N_MEMORY].)
1763 * If the zonelist cache is not available for this zonelist, does
1764 * nothing and returns NULL.
1766 * If the fullzones BITMAP in the zonelist cache is stale (more than
1767 * a second since last zap'd) then we zap it out (clear its bits.)
1769 * We hold off even calling zlc_setup, until after we've checked the
1770 * first zone in the zonelist, on the theory that most allocations will
1771 * be satisfied from that first zone, so best to examine that zone as
1772 * quickly as we can.
1774 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1776 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1777 nodemask_t *allowednodes; /* zonelist_cache approximation */
1779 zlc = zonelist->zlcache_ptr;
1783 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1784 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1785 zlc->last_full_zap = jiffies;
1788 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1789 &cpuset_current_mems_allowed :
1790 &node_states[N_MEMORY];
1791 return allowednodes;
1795 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1796 * if it is worth looking at further for free memory:
1797 * 1) Check that the zone isn't thought to be full (doesn't have its
1798 * bit set in the zonelist_cache fullzones BITMAP).
1799 * 2) Check that the zones node (obtained from the zonelist_cache
1800 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1801 * Return true (non-zero) if zone is worth looking at further, or
1802 * else return false (zero) if it is not.
1804 * This check -ignores- the distinction between various watermarks,
1805 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1806 * found to be full for any variation of these watermarks, it will
1807 * be considered full for up to one second by all requests, unless
1808 * we are so low on memory on all allowed nodes that we are forced
1809 * into the second scan of the zonelist.
1811 * In the second scan we ignore this zonelist cache and exactly
1812 * apply the watermarks to all zones, even it is slower to do so.
1813 * We are low on memory in the second scan, and should leave no stone
1814 * unturned looking for a free page.
1816 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1817 nodemask_t *allowednodes)
1819 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1820 int i; /* index of *z in zonelist zones */
1821 int n; /* node that zone *z is on */
1823 zlc = zonelist->zlcache_ptr;
1827 i = z - zonelist->_zonerefs;
1830 /* This zone is worth trying if it is allowed but not full */
1831 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1835 * Given 'z' scanning a zonelist, set the corresponding bit in
1836 * zlc->fullzones, so that subsequent attempts to allocate a page
1837 * from that zone don't waste time re-examining it.
1839 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1841 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1842 int i; /* index of *z in zonelist zones */
1844 zlc = zonelist->zlcache_ptr;
1848 i = z - zonelist->_zonerefs;
1850 set_bit(i, zlc->fullzones);
1854 * clear all zones full, called after direct reclaim makes progress so that
1855 * a zone that was recently full is not skipped over for up to a second
1857 static void zlc_clear_zones_full(struct zonelist *zonelist)
1859 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1861 zlc = zonelist->zlcache_ptr;
1865 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1868 static bool zone_local(struct zone *local_zone, struct zone *zone)
1870 return local_zone->node == zone->node;
1873 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1875 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1879 #else /* CONFIG_NUMA */
1881 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1886 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1887 nodemask_t *allowednodes)
1892 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1896 static void zlc_clear_zones_full(struct zonelist *zonelist)
1900 static bool zone_local(struct zone *local_zone, struct zone *zone)
1905 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1910 #endif /* CONFIG_NUMA */
1913 * get_page_from_freelist goes through the zonelist trying to allocate
1916 static struct page *
1917 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1918 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1919 struct zone *preferred_zone, int migratetype)
1922 struct page *page = NULL;
1925 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1926 int zlc_active = 0; /* set if using zonelist_cache */
1927 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1928 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
1929 (gfp_mask & __GFP_WRITE);
1931 classzone_idx = zone_idx(preferred_zone);
1934 * Scan zonelist, looking for a zone with enough free.
1935 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1937 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1938 high_zoneidx, nodemask) {
1941 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1942 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1944 if (cpusets_enabled() &&
1945 (alloc_flags & ALLOC_CPUSET) &&
1946 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1949 * Distribute pages in proportion to the individual
1950 * zone size to ensure fair page aging. The zone a
1951 * page was allocated in should have no effect on the
1952 * time the page has in memory before being reclaimed.
1954 if (alloc_flags & ALLOC_FAIR) {
1955 if (!zone_local(preferred_zone, zone))
1957 if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1961 * When allocating a page cache page for writing, we
1962 * want to get it from a zone that is within its dirty
1963 * limit, such that no single zone holds more than its
1964 * proportional share of globally allowed dirty pages.
1965 * The dirty limits take into account the zone's
1966 * lowmem reserves and high watermark so that kswapd
1967 * should be able to balance it without having to
1968 * write pages from its LRU list.
1970 * This may look like it could increase pressure on
1971 * lower zones by failing allocations in higher zones
1972 * before they are full. But the pages that do spill
1973 * over are limited as the lower zones are protected
1974 * by this very same mechanism. It should not become
1975 * a practical burden to them.
1977 * XXX: For now, allow allocations to potentially
1978 * exceed the per-zone dirty limit in the slowpath
1979 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1980 * which is important when on a NUMA setup the allowed
1981 * zones are together not big enough to reach the
1982 * global limit. The proper fix for these situations
1983 * will require awareness of zones in the
1984 * dirty-throttling and the flusher threads.
1986 if (consider_zone_dirty && !zone_dirty_ok(zone))
1989 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1990 if (!zone_watermark_ok(zone, order, mark,
1991 classzone_idx, alloc_flags)) {
1994 /* Checked here to keep the fast path fast */
1995 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1996 if (alloc_flags & ALLOC_NO_WATERMARKS)
1999 if (IS_ENABLED(CONFIG_NUMA) &&
2000 !did_zlc_setup && nr_online_nodes > 1) {
2002 * we do zlc_setup if there are multiple nodes
2003 * and before considering the first zone allowed
2006 allowednodes = zlc_setup(zonelist, alloc_flags);
2011 if (zone_reclaim_mode == 0 ||
2012 !zone_allows_reclaim(preferred_zone, zone))
2013 goto this_zone_full;
2016 * As we may have just activated ZLC, check if the first
2017 * eligible zone has failed zone_reclaim recently.
2019 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2020 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2023 ret = zone_reclaim(zone, gfp_mask, order);
2025 case ZONE_RECLAIM_NOSCAN:
2028 case ZONE_RECLAIM_FULL:
2029 /* scanned but unreclaimable */
2032 /* did we reclaim enough */
2033 if (zone_watermark_ok(zone, order, mark,
2034 classzone_idx, alloc_flags))
2038 * Failed to reclaim enough to meet watermark.
2039 * Only mark the zone full if checking the min
2040 * watermark or if we failed to reclaim just
2041 * 1<<order pages or else the page allocator
2042 * fastpath will prematurely mark zones full
2043 * when the watermark is between the low and
2046 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2047 ret == ZONE_RECLAIM_SOME)
2048 goto this_zone_full;
2055 page = buffered_rmqueue(preferred_zone, zone, order,
2056 gfp_mask, migratetype);
2060 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2061 zlc_mark_zone_full(zonelist, z);
2064 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2065 /* Disable zlc cache for second zonelist scan */
2072 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2073 * necessary to allocate the page. The expectation is
2074 * that the caller is taking steps that will free more
2075 * memory. The caller should avoid the page being used
2076 * for !PFMEMALLOC purposes.
2078 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2084 * Large machines with many possible nodes should not always dump per-node
2085 * meminfo in irq context.
2087 static inline bool should_suppress_show_mem(void)
2092 ret = in_interrupt();
2097 static DEFINE_RATELIMIT_STATE(nopage_rs,
2098 DEFAULT_RATELIMIT_INTERVAL,
2099 DEFAULT_RATELIMIT_BURST);
2101 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2103 unsigned int filter = SHOW_MEM_FILTER_NODES;
2105 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2106 debug_guardpage_minorder() > 0)
2110 * This documents exceptions given to allocations in certain
2111 * contexts that are allowed to allocate outside current's set
2114 if (!(gfp_mask & __GFP_NOMEMALLOC))
2115 if (test_thread_flag(TIF_MEMDIE) ||
2116 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2117 filter &= ~SHOW_MEM_FILTER_NODES;
2118 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2119 filter &= ~SHOW_MEM_FILTER_NODES;
2122 struct va_format vaf;
2125 va_start(args, fmt);
2130 pr_warn("%pV", &vaf);
2135 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2136 current->comm, order, gfp_mask);
2139 if (!should_suppress_show_mem())
2144 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2145 unsigned long did_some_progress,
2146 unsigned long pages_reclaimed)
2148 /* Do not loop if specifically requested */
2149 if (gfp_mask & __GFP_NORETRY)
2152 /* Always retry if specifically requested */
2153 if (gfp_mask & __GFP_NOFAIL)
2157 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2158 * making forward progress without invoking OOM. Suspend also disables
2159 * storage devices so kswapd will not help. Bail if we are suspending.
2161 if (!did_some_progress && pm_suspended_storage())
2165 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2166 * means __GFP_NOFAIL, but that may not be true in other
2169 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2173 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2174 * specified, then we retry until we no longer reclaim any pages
2175 * (above), or we've reclaimed an order of pages at least as
2176 * large as the allocation's order. In both cases, if the
2177 * allocation still fails, we stop retrying.
2179 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2185 static inline struct page *
2186 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2187 struct zonelist *zonelist, enum zone_type high_zoneidx,
2188 nodemask_t *nodemask, struct zone *preferred_zone,
2193 /* Acquire the OOM killer lock for the zones in zonelist */
2194 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2195 schedule_timeout_uninterruptible(1);
2200 * Go through the zonelist yet one more time, keep very high watermark
2201 * here, this is only to catch a parallel oom killing, we must fail if
2202 * we're still under heavy pressure.
2204 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2205 order, zonelist, high_zoneidx,
2206 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2207 preferred_zone, migratetype);
2211 if (!(gfp_mask & __GFP_NOFAIL)) {
2212 /* The OOM killer will not help higher order allocs */
2213 if (order > PAGE_ALLOC_COSTLY_ORDER)
2215 /* The OOM killer does not needlessly kill tasks for lowmem */
2216 if (high_zoneidx < ZONE_NORMAL)
2219 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2220 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2221 * The caller should handle page allocation failure by itself if
2222 * it specifies __GFP_THISNODE.
2223 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2225 if (gfp_mask & __GFP_THISNODE)
2228 /* Exhausted what can be done so it's blamo time */
2229 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2232 clear_zonelist_oom(zonelist, gfp_mask);
2236 #ifdef CONFIG_COMPACTION
2237 /* Try memory compaction for high-order allocations before reclaim */
2238 static struct page *
2239 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2240 struct zonelist *zonelist, enum zone_type high_zoneidx,
2241 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2242 int migratetype, enum migrate_mode mode,
2243 bool *contended_compaction, bool *deferred_compaction,
2244 unsigned long *did_some_progress)
2249 if (compaction_deferred(preferred_zone, order)) {
2250 *deferred_compaction = true;
2254 current->flags |= PF_MEMALLOC;
2255 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2257 contended_compaction);
2258 current->flags &= ~PF_MEMALLOC;
2260 if (*did_some_progress != COMPACT_SKIPPED) {
2263 /* Page migration frees to the PCP lists but we want merging */
2264 drain_pages(get_cpu());
2267 page = get_page_from_freelist(gfp_mask, nodemask,
2268 order, zonelist, high_zoneidx,
2269 alloc_flags & ~ALLOC_NO_WATERMARKS,
2270 preferred_zone, migratetype);
2272 preferred_zone->compact_blockskip_flush = false;
2273 compaction_defer_reset(preferred_zone, order, true);
2274 count_vm_event(COMPACTSUCCESS);
2279 * It's bad if compaction run occurs and fails.
2280 * The most likely reason is that pages exist,
2281 * but not enough to satisfy watermarks.
2283 count_vm_event(COMPACTFAIL);
2286 * As async compaction considers a subset of pageblocks, only
2287 * defer if the failure was a sync compaction failure.
2289 if (mode != MIGRATE_ASYNC)
2290 defer_compaction(preferred_zone, order);
2298 static inline struct page *
2299 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2300 struct zonelist *zonelist, enum zone_type high_zoneidx,
2301 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2302 int migratetype, enum migrate_mode mode, bool *contended_compaction,
2303 bool *deferred_compaction, unsigned long *did_some_progress)
2307 #endif /* CONFIG_COMPACTION */
2309 /* Perform direct synchronous page reclaim */
2311 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2312 nodemask_t *nodemask)
2314 struct reclaim_state reclaim_state;
2319 /* We now go into synchronous reclaim */
2320 cpuset_memory_pressure_bump();
2321 current->flags |= PF_MEMALLOC;
2322 lockdep_set_current_reclaim_state(gfp_mask);
2323 reclaim_state.reclaimed_slab = 0;
2324 current->reclaim_state = &reclaim_state;
2326 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2328 current->reclaim_state = NULL;
2329 lockdep_clear_current_reclaim_state();
2330 current->flags &= ~PF_MEMALLOC;
2337 /* The really slow allocator path where we enter direct reclaim */
2338 static inline struct page *
2339 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2340 struct zonelist *zonelist, enum zone_type high_zoneidx,
2341 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2342 int migratetype, unsigned long *did_some_progress)
2344 struct page *page = NULL;
2345 bool drained = false;
2347 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2349 if (unlikely(!(*did_some_progress)))
2352 /* After successful reclaim, reconsider all zones for allocation */
2353 if (IS_ENABLED(CONFIG_NUMA))
2354 zlc_clear_zones_full(zonelist);
2357 page = get_page_from_freelist(gfp_mask, nodemask, order,
2358 zonelist, high_zoneidx,
2359 alloc_flags & ~ALLOC_NO_WATERMARKS,
2360 preferred_zone, migratetype);
2363 * If an allocation failed after direct reclaim, it could be because
2364 * pages are pinned on the per-cpu lists. Drain them and try again
2366 if (!page && !drained) {
2376 * This is called in the allocator slow-path if the allocation request is of
2377 * sufficient urgency to ignore watermarks and take other desperate measures
2379 static inline struct page *
2380 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2381 struct zonelist *zonelist, enum zone_type high_zoneidx,
2382 nodemask_t *nodemask, struct zone *preferred_zone,
2388 page = get_page_from_freelist(gfp_mask, nodemask, order,
2389 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2390 preferred_zone, migratetype);
2392 if (!page && gfp_mask & __GFP_NOFAIL)
2393 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2394 } while (!page && (gfp_mask & __GFP_NOFAIL));
2399 static void reset_alloc_batches(struct zonelist *zonelist,
2400 enum zone_type high_zoneidx,
2401 struct zone *preferred_zone)
2406 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2408 * Only reset the batches of zones that were actually
2409 * considered in the fairness pass, we don't want to
2410 * trash fairness information for zones that are not
2411 * actually part of this zonelist's round-robin cycle.
2413 if (!zone_local(preferred_zone, zone))
2415 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2416 high_wmark_pages(zone) - low_wmark_pages(zone) -
2417 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2421 static void wake_all_kswapds(unsigned int order,
2422 struct zonelist *zonelist,
2423 enum zone_type high_zoneidx,
2424 struct zone *preferred_zone)
2429 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2430 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2434 gfp_to_alloc_flags(gfp_t gfp_mask)
2436 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2437 const gfp_t wait = gfp_mask & __GFP_WAIT;
2439 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2440 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2443 * The caller may dip into page reserves a bit more if the caller
2444 * cannot run direct reclaim, or if the caller has realtime scheduling
2445 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2446 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2448 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2452 * Not worth trying to allocate harder for
2453 * __GFP_NOMEMALLOC even if it can't schedule.
2455 if (!(gfp_mask & __GFP_NOMEMALLOC))
2456 alloc_flags |= ALLOC_HARDER;
2458 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2459 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2461 alloc_flags &= ~ALLOC_CPUSET;
2462 } else if (unlikely(rt_task(current)) && !in_interrupt())
2463 alloc_flags |= ALLOC_HARDER;
2465 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2466 if (gfp_mask & __GFP_MEMALLOC)
2467 alloc_flags |= ALLOC_NO_WATERMARKS;
2468 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2469 alloc_flags |= ALLOC_NO_WATERMARKS;
2470 else if (!in_interrupt() &&
2471 ((current->flags & PF_MEMALLOC) ||
2472 unlikely(test_thread_flag(TIF_MEMDIE))))
2473 alloc_flags |= ALLOC_NO_WATERMARKS;
2476 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2477 alloc_flags |= ALLOC_CMA;
2482 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2484 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2487 static inline struct page *
2488 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2489 struct zonelist *zonelist, enum zone_type high_zoneidx,
2490 nodemask_t *nodemask, struct zone *preferred_zone,
2493 const gfp_t wait = gfp_mask & __GFP_WAIT;
2494 struct page *page = NULL;
2496 unsigned long pages_reclaimed = 0;
2497 unsigned long did_some_progress;
2498 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2499 bool deferred_compaction = false;
2500 bool contended_compaction = false;
2503 * In the slowpath, we sanity check order to avoid ever trying to
2504 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2505 * be using allocators in order of preference for an area that is
2508 if (order >= MAX_ORDER) {
2509 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2514 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2515 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2516 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2517 * using a larger set of nodes after it has established that the
2518 * allowed per node queues are empty and that nodes are
2521 if (IS_ENABLED(CONFIG_NUMA) &&
2522 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2526 if (!(gfp_mask & __GFP_NO_KSWAPD))
2527 wake_all_kswapds(order, zonelist, high_zoneidx, preferred_zone);
2530 * OK, we're below the kswapd watermark and have kicked background
2531 * reclaim. Now things get more complex, so set up alloc_flags according
2532 * to how we want to proceed.
2534 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2537 * Find the true preferred zone if the allocation is unconstrained by
2540 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2541 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2545 /* This is the last chance, in general, before the goto nopage. */
2546 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2547 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2548 preferred_zone, migratetype);
2552 /* Allocate without watermarks if the context allows */
2553 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2555 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2556 * the allocation is high priority and these type of
2557 * allocations are system rather than user orientated
2559 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2561 page = __alloc_pages_high_priority(gfp_mask, order,
2562 zonelist, high_zoneidx, nodemask,
2563 preferred_zone, migratetype);
2569 /* Atomic allocations - we can't balance anything */
2572 * All existing users of the deprecated __GFP_NOFAIL are
2573 * blockable, so warn of any new users that actually allow this
2574 * type of allocation to fail.
2576 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2580 /* Avoid recursion of direct reclaim */
2581 if (current->flags & PF_MEMALLOC)
2584 /* Avoid allocations with no watermarks from looping endlessly */
2585 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2589 * Try direct compaction. The first pass is asynchronous. Subsequent
2590 * attempts after direct reclaim are synchronous
2592 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2593 high_zoneidx, nodemask, alloc_flags,
2594 preferred_zone, migratetype,
2595 migration_mode, &contended_compaction,
2596 &deferred_compaction,
2597 &did_some_progress);
2602 * It can become very expensive to allocate transparent hugepages at
2603 * fault, so use asynchronous memory compaction for THP unless it is
2604 * khugepaged trying to collapse.
2606 if (!(gfp_mask & __GFP_NO_KSWAPD) || (current->flags & PF_KTHREAD))
2607 migration_mode = MIGRATE_SYNC_LIGHT;
2610 * If compaction is deferred for high-order allocations, it is because
2611 * sync compaction recently failed. In this is the case and the caller
2612 * requested a movable allocation that does not heavily disrupt the
2613 * system then fail the allocation instead of entering direct reclaim.
2615 if ((deferred_compaction || contended_compaction) &&
2616 (gfp_mask & __GFP_NO_KSWAPD))
2619 /* Try direct reclaim and then allocating */
2620 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2621 zonelist, high_zoneidx,
2623 alloc_flags, preferred_zone,
2624 migratetype, &did_some_progress);
2629 * If we failed to make any progress reclaiming, then we are
2630 * running out of options and have to consider going OOM
2632 if (!did_some_progress) {
2633 if (oom_gfp_allowed(gfp_mask)) {
2634 if (oom_killer_disabled)
2636 /* Coredumps can quickly deplete all memory reserves */
2637 if ((current->flags & PF_DUMPCORE) &&
2638 !(gfp_mask & __GFP_NOFAIL))
2640 page = __alloc_pages_may_oom(gfp_mask, order,
2641 zonelist, high_zoneidx,
2642 nodemask, preferred_zone,
2647 if (!(gfp_mask & __GFP_NOFAIL)) {
2649 * The oom killer is not called for high-order
2650 * allocations that may fail, so if no progress
2651 * is being made, there are no other options and
2652 * retrying is unlikely to help.
2654 if (order > PAGE_ALLOC_COSTLY_ORDER)
2657 * The oom killer is not called for lowmem
2658 * allocations to prevent needlessly killing
2661 if (high_zoneidx < ZONE_NORMAL)
2669 /* Check if we should retry the allocation */
2670 pages_reclaimed += did_some_progress;
2671 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2673 /* Wait for some write requests to complete then retry */
2674 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2678 * High-order allocations do not necessarily loop after
2679 * direct reclaim and reclaim/compaction depends on compaction
2680 * being called after reclaim so call directly if necessary
2682 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2683 high_zoneidx, nodemask, alloc_flags,
2684 preferred_zone, migratetype,
2685 migration_mode, &contended_compaction,
2686 &deferred_compaction,
2687 &did_some_progress);
2693 warn_alloc_failed(gfp_mask, order, NULL);
2696 if (kmemcheck_enabled)
2697 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2703 * This is the 'heart' of the zoned buddy allocator.
2706 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2707 struct zonelist *zonelist, nodemask_t *nodemask)
2709 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2710 struct zone *preferred_zone;
2711 struct page *page = NULL;
2712 int migratetype = allocflags_to_migratetype(gfp_mask);
2713 unsigned int cpuset_mems_cookie;
2714 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2716 gfp_mask &= gfp_allowed_mask;
2718 lockdep_trace_alloc(gfp_mask);
2720 might_sleep_if(gfp_mask & __GFP_WAIT);
2722 if (should_fail_alloc_page(gfp_mask, order))
2726 * Check the zones suitable for the gfp_mask contain at least one
2727 * valid zone. It's possible to have an empty zonelist as a result
2728 * of GFP_THISNODE and a memoryless node
2730 if (unlikely(!zonelist->_zonerefs->zone))
2734 cpuset_mems_cookie = read_mems_allowed_begin();
2736 /* The preferred zone is used for statistics later */
2737 first_zones_zonelist(zonelist, high_zoneidx,
2738 nodemask ? : &cpuset_current_mems_allowed,
2740 if (!preferred_zone)
2744 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2745 alloc_flags |= ALLOC_CMA;
2748 /* First allocation attempt */
2749 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2750 zonelist, high_zoneidx, alloc_flags,
2751 preferred_zone, migratetype);
2752 if (unlikely(!page)) {
2754 * The first pass makes sure allocations are spread
2755 * fairly within the local node. However, the local
2756 * node might have free pages left after the fairness
2757 * batches are exhausted, and remote zones haven't
2758 * even been considered yet. Try once more without
2759 * fairness, and include remote zones now, before
2760 * entering the slowpath and waking kswapd: prefer
2761 * spilling to a remote zone over swapping locally.
2763 if (alloc_flags & ALLOC_FAIR) {
2764 reset_alloc_batches(zonelist, high_zoneidx,
2766 alloc_flags &= ~ALLOC_FAIR;
2770 * Runtime PM, block IO and its error handling path
2771 * can deadlock because I/O on the device might not
2774 gfp_mask = memalloc_noio_flags(gfp_mask);
2775 page = __alloc_pages_slowpath(gfp_mask, order,
2776 zonelist, high_zoneidx, nodemask,
2777 preferred_zone, migratetype);
2780 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2784 * When updating a task's mems_allowed, it is possible to race with
2785 * parallel threads in such a way that an allocation can fail while
2786 * the mask is being updated. If a page allocation is about to fail,
2787 * check if the cpuset changed during allocation and if so, retry.
2789 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2794 EXPORT_SYMBOL(__alloc_pages_nodemask);
2797 * Common helper functions.
2799 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2804 * __get_free_pages() returns a 32-bit address, which cannot represent
2807 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2809 page = alloc_pages(gfp_mask, order);
2812 return (unsigned long) page_address(page);
2814 EXPORT_SYMBOL(__get_free_pages);
2816 unsigned long get_zeroed_page(gfp_t gfp_mask)
2818 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2820 EXPORT_SYMBOL(get_zeroed_page);
2822 void __free_pages(struct page *page, unsigned int order)
2824 if (put_page_testzero(page)) {
2826 free_hot_cold_page(page, false);
2828 __free_pages_ok(page, order);
2832 EXPORT_SYMBOL(__free_pages);
2834 void free_pages(unsigned long addr, unsigned int order)
2837 VM_BUG_ON(!virt_addr_valid((void *)addr));
2838 __free_pages(virt_to_page((void *)addr), order);
2842 EXPORT_SYMBOL(free_pages);
2845 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2846 * of the current memory cgroup.
2848 * It should be used when the caller would like to use kmalloc, but since the
2849 * allocation is large, it has to fall back to the page allocator.
2851 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2854 struct mem_cgroup *memcg = NULL;
2856 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2858 page = alloc_pages(gfp_mask, order);
2859 memcg_kmem_commit_charge(page, memcg, order);
2863 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2866 struct mem_cgroup *memcg = NULL;
2868 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2870 page = alloc_pages_node(nid, gfp_mask, order);
2871 memcg_kmem_commit_charge(page, memcg, order);
2876 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2879 void __free_kmem_pages(struct page *page, unsigned int order)
2881 memcg_kmem_uncharge_pages(page, order);
2882 __free_pages(page, order);
2885 void free_kmem_pages(unsigned long addr, unsigned int order)
2888 VM_BUG_ON(!virt_addr_valid((void *)addr));
2889 __free_kmem_pages(virt_to_page((void *)addr), order);
2893 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2896 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2897 unsigned long used = addr + PAGE_ALIGN(size);
2899 split_page(virt_to_page((void *)addr), order);
2900 while (used < alloc_end) {
2905 return (void *)addr;
2909 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2910 * @size: the number of bytes to allocate
2911 * @gfp_mask: GFP flags for the allocation
2913 * This function is similar to alloc_pages(), except that it allocates the
2914 * minimum number of pages to satisfy the request. alloc_pages() can only
2915 * allocate memory in power-of-two pages.
2917 * This function is also limited by MAX_ORDER.
2919 * Memory allocated by this function must be released by free_pages_exact().
2921 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2923 unsigned int order = get_order(size);
2926 addr = __get_free_pages(gfp_mask, order);
2927 return make_alloc_exact(addr, order, size);
2929 EXPORT_SYMBOL(alloc_pages_exact);
2932 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2934 * @nid: the preferred node ID where memory should be allocated
2935 * @size: the number of bytes to allocate
2936 * @gfp_mask: GFP flags for the allocation
2938 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2940 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2943 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2945 unsigned order = get_order(size);
2946 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2949 return make_alloc_exact((unsigned long)page_address(p), order, size);
2951 EXPORT_SYMBOL(alloc_pages_exact_nid);
2954 * free_pages_exact - release memory allocated via alloc_pages_exact()
2955 * @virt: the value returned by alloc_pages_exact.
2956 * @size: size of allocation, same value as passed to alloc_pages_exact().
2958 * Release the memory allocated by a previous call to alloc_pages_exact.
2960 void free_pages_exact(void *virt, size_t size)
2962 unsigned long addr = (unsigned long)virt;
2963 unsigned long end = addr + PAGE_ALIGN(size);
2965 while (addr < end) {
2970 EXPORT_SYMBOL(free_pages_exact);
2973 * nr_free_zone_pages - count number of pages beyond high watermark
2974 * @offset: The zone index of the highest zone
2976 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2977 * high watermark within all zones at or below a given zone index. For each
2978 * zone, the number of pages is calculated as:
2979 * managed_pages - high_pages
2981 static unsigned long nr_free_zone_pages(int offset)
2986 /* Just pick one node, since fallback list is circular */
2987 unsigned long sum = 0;
2989 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2991 for_each_zone_zonelist(zone, z, zonelist, offset) {
2992 unsigned long size = zone->managed_pages;
2993 unsigned long high = high_wmark_pages(zone);
3002 * nr_free_buffer_pages - count number of pages beyond high watermark
3004 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3005 * watermark within ZONE_DMA and ZONE_NORMAL.
3007 unsigned long nr_free_buffer_pages(void)
3009 return nr_free_zone_pages(gfp_zone(GFP_USER));
3011 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3014 * nr_free_pagecache_pages - count number of pages beyond high watermark
3016 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3017 * high watermark within all zones.
3019 unsigned long nr_free_pagecache_pages(void)
3021 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3024 static inline void show_node(struct zone *zone)
3026 if (IS_ENABLED(CONFIG_NUMA))
3027 printk("Node %d ", zone_to_nid(zone));
3030 void si_meminfo(struct sysinfo *val)
3032 val->totalram = totalram_pages;
3034 val->freeram = global_page_state(NR_FREE_PAGES);
3035 val->bufferram = nr_blockdev_pages();
3036 val->totalhigh = totalhigh_pages;
3037 val->freehigh = nr_free_highpages();
3038 val->mem_unit = PAGE_SIZE;
3041 EXPORT_SYMBOL(si_meminfo);
3044 void si_meminfo_node(struct sysinfo *val, int nid)
3046 int zone_type; /* needs to be signed */
3047 unsigned long managed_pages = 0;
3048 pg_data_t *pgdat = NODE_DATA(nid);
3050 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3051 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3052 val->totalram = managed_pages;
3053 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3054 #ifdef CONFIG_HIGHMEM
3055 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3056 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3062 val->mem_unit = PAGE_SIZE;
3067 * Determine whether the node should be displayed or not, depending on whether
3068 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3070 bool skip_free_areas_node(unsigned int flags, int nid)
3073 unsigned int cpuset_mems_cookie;
3075 if (!(flags & SHOW_MEM_FILTER_NODES))
3079 cpuset_mems_cookie = read_mems_allowed_begin();
3080 ret = !node_isset(nid, cpuset_current_mems_allowed);
3081 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3086 #define K(x) ((x) << (PAGE_SHIFT-10))
3088 static void show_migration_types(unsigned char type)
3090 static const char types[MIGRATE_TYPES] = {
3091 [MIGRATE_UNMOVABLE] = 'U',
3092 [MIGRATE_RECLAIMABLE] = 'E',
3093 [MIGRATE_MOVABLE] = 'M',
3094 [MIGRATE_RESERVE] = 'R',
3096 [MIGRATE_CMA] = 'C',
3098 #ifdef CONFIG_MEMORY_ISOLATION
3099 [MIGRATE_ISOLATE] = 'I',
3102 char tmp[MIGRATE_TYPES + 1];
3106 for (i = 0; i < MIGRATE_TYPES; i++) {
3107 if (type & (1 << i))
3112 printk("(%s) ", tmp);
3116 * Show free area list (used inside shift_scroll-lock stuff)
3117 * We also calculate the percentage fragmentation. We do this by counting the
3118 * memory on each free list with the exception of the first item on the list.
3119 * Suppresses nodes that are not allowed by current's cpuset if
3120 * SHOW_MEM_FILTER_NODES is passed.
3122 void show_free_areas(unsigned int filter)
3127 for_each_populated_zone(zone) {
3128 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3131 printk("%s per-cpu:\n", zone->name);
3133 for_each_online_cpu(cpu) {
3134 struct per_cpu_pageset *pageset;
3136 pageset = per_cpu_ptr(zone->pageset, cpu);
3138 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3139 cpu, pageset->pcp.high,
3140 pageset->pcp.batch, pageset->pcp.count);
3144 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3145 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3147 " dirty:%lu writeback:%lu unstable:%lu\n"
3148 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3149 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3151 global_page_state(NR_ACTIVE_ANON),
3152 global_page_state(NR_INACTIVE_ANON),
3153 global_page_state(NR_ISOLATED_ANON),
3154 global_page_state(NR_ACTIVE_FILE),
3155 global_page_state(NR_INACTIVE_FILE),
3156 global_page_state(NR_ISOLATED_FILE),
3157 global_page_state(NR_UNEVICTABLE),
3158 global_page_state(NR_FILE_DIRTY),
3159 global_page_state(NR_WRITEBACK),
3160 global_page_state(NR_UNSTABLE_NFS),
3161 global_page_state(NR_FREE_PAGES),
3162 global_page_state(NR_SLAB_RECLAIMABLE),
3163 global_page_state(NR_SLAB_UNRECLAIMABLE),
3164 global_page_state(NR_FILE_MAPPED),
3165 global_page_state(NR_SHMEM),
3166 global_page_state(NR_PAGETABLE),
3167 global_page_state(NR_BOUNCE),
3168 global_page_state(NR_FREE_CMA_PAGES));
3170 for_each_populated_zone(zone) {
3173 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3181 " active_anon:%lukB"
3182 " inactive_anon:%lukB"
3183 " active_file:%lukB"
3184 " inactive_file:%lukB"
3185 " unevictable:%lukB"
3186 " isolated(anon):%lukB"
3187 " isolated(file):%lukB"
3195 " slab_reclaimable:%lukB"
3196 " slab_unreclaimable:%lukB"
3197 " kernel_stack:%lukB"
3202 " writeback_tmp:%lukB"
3203 " pages_scanned:%lu"
3204 " all_unreclaimable? %s"
3207 K(zone_page_state(zone, NR_FREE_PAGES)),
3208 K(min_wmark_pages(zone)),
3209 K(low_wmark_pages(zone)),
3210 K(high_wmark_pages(zone)),
3211 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3212 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3213 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3214 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3215 K(zone_page_state(zone, NR_UNEVICTABLE)),
3216 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3217 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3218 K(zone->present_pages),
3219 K(zone->managed_pages),
3220 K(zone_page_state(zone, NR_MLOCK)),
3221 K(zone_page_state(zone, NR_FILE_DIRTY)),
3222 K(zone_page_state(zone, NR_WRITEBACK)),
3223 K(zone_page_state(zone, NR_FILE_MAPPED)),
3224 K(zone_page_state(zone, NR_SHMEM)),
3225 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3226 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3227 zone_page_state(zone, NR_KERNEL_STACK) *
3229 K(zone_page_state(zone, NR_PAGETABLE)),
3230 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3231 K(zone_page_state(zone, NR_BOUNCE)),
3232 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3233 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3234 zone->pages_scanned,
3235 (!zone_reclaimable(zone) ? "yes" : "no")
3237 printk("lowmem_reserve[]:");
3238 for (i = 0; i < MAX_NR_ZONES; i++)
3239 printk(" %lu", zone->lowmem_reserve[i]);
3243 for_each_populated_zone(zone) {
3244 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3245 unsigned char types[MAX_ORDER];
3247 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3250 printk("%s: ", zone->name);
3252 spin_lock_irqsave(&zone->lock, flags);
3253 for (order = 0; order < MAX_ORDER; order++) {
3254 struct free_area *area = &zone->free_area[order];
3257 nr[order] = area->nr_free;
3258 total += nr[order] << order;
3261 for (type = 0; type < MIGRATE_TYPES; type++) {
3262 if (!list_empty(&area->free_list[type]))
3263 types[order] |= 1 << type;
3266 spin_unlock_irqrestore(&zone->lock, flags);
3267 for (order = 0; order < MAX_ORDER; order++) {
3268 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3270 show_migration_types(types[order]);
3272 printk("= %lukB\n", K(total));
3275 hugetlb_show_meminfo();
3277 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3279 show_swap_cache_info();
3282 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3284 zoneref->zone = zone;
3285 zoneref->zone_idx = zone_idx(zone);
3289 * Builds allocation fallback zone lists.
3291 * Add all populated zones of a node to the zonelist.
3293 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3297 enum zone_type zone_type = MAX_NR_ZONES;
3301 zone = pgdat->node_zones + zone_type;
3302 if (populated_zone(zone)) {
3303 zoneref_set_zone(zone,
3304 &zonelist->_zonerefs[nr_zones++]);
3305 check_highest_zone(zone_type);
3307 } while (zone_type);
3315 * 0 = automatic detection of better ordering.
3316 * 1 = order by ([node] distance, -zonetype)
3317 * 2 = order by (-zonetype, [node] distance)
3319 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3320 * the same zonelist. So only NUMA can configure this param.
3322 #define ZONELIST_ORDER_DEFAULT 0
3323 #define ZONELIST_ORDER_NODE 1
3324 #define ZONELIST_ORDER_ZONE 2
3326 /* zonelist order in the kernel.
3327 * set_zonelist_order() will set this to NODE or ZONE.
3329 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3330 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3334 /* The value user specified ....changed by config */
3335 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3336 /* string for sysctl */
3337 #define NUMA_ZONELIST_ORDER_LEN 16
3338 char numa_zonelist_order[16] = "default";
3341 * interface for configure zonelist ordering.
3342 * command line option "numa_zonelist_order"
3343 * = "[dD]efault - default, automatic configuration.
3344 * = "[nN]ode - order by node locality, then by zone within node
3345 * = "[zZ]one - order by zone, then by locality within zone
3348 static int __parse_numa_zonelist_order(char *s)
3350 if (*s == 'd' || *s == 'D') {
3351 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3352 } else if (*s == 'n' || *s == 'N') {
3353 user_zonelist_order = ZONELIST_ORDER_NODE;
3354 } else if (*s == 'z' || *s == 'Z') {
3355 user_zonelist_order = ZONELIST_ORDER_ZONE;
3358 "Ignoring invalid numa_zonelist_order value: "
3365 static __init int setup_numa_zonelist_order(char *s)
3372 ret = __parse_numa_zonelist_order(s);
3374 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3378 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3381 * sysctl handler for numa_zonelist_order
3383 int numa_zonelist_order_handler(ctl_table *table, int write,
3384 void __user *buffer, size_t *length,
3387 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3389 static DEFINE_MUTEX(zl_order_mutex);
3391 mutex_lock(&zl_order_mutex);
3393 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3397 strcpy(saved_string, (char *)table->data);
3399 ret = proc_dostring(table, write, buffer, length, ppos);
3403 int oldval = user_zonelist_order;
3405 ret = __parse_numa_zonelist_order((char *)table->data);
3408 * bogus value. restore saved string
3410 strncpy((char *)table->data, saved_string,
3411 NUMA_ZONELIST_ORDER_LEN);
3412 user_zonelist_order = oldval;
3413 } else if (oldval != user_zonelist_order) {
3414 mutex_lock(&zonelists_mutex);
3415 build_all_zonelists(NULL, NULL);
3416 mutex_unlock(&zonelists_mutex);
3420 mutex_unlock(&zl_order_mutex);
3425 #define MAX_NODE_LOAD (nr_online_nodes)
3426 static int node_load[MAX_NUMNODES];
3429 * find_next_best_node - find the next node that should appear in a given node's fallback list
3430 * @node: node whose fallback list we're appending
3431 * @used_node_mask: nodemask_t of already used nodes
3433 * We use a number of factors to determine which is the next node that should
3434 * appear on a given node's fallback list. The node should not have appeared
3435 * already in @node's fallback list, and it should be the next closest node
3436 * according to the distance array (which contains arbitrary distance values
3437 * from each node to each node in the system), and should also prefer nodes
3438 * with no CPUs, since presumably they'll have very little allocation pressure
3439 * on them otherwise.
3440 * It returns -1 if no node is found.
3442 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3445 int min_val = INT_MAX;
3446 int best_node = NUMA_NO_NODE;
3447 const struct cpumask *tmp = cpumask_of_node(0);
3449 /* Use the local node if we haven't already */
3450 if (!node_isset(node, *used_node_mask)) {
3451 node_set(node, *used_node_mask);
3455 for_each_node_state(n, N_MEMORY) {
3457 /* Don't want a node to appear more than once */
3458 if (node_isset(n, *used_node_mask))
3461 /* Use the distance array to find the distance */
3462 val = node_distance(node, n);
3464 /* Penalize nodes under us ("prefer the next node") */
3467 /* Give preference to headless and unused nodes */
3468 tmp = cpumask_of_node(n);
3469 if (!cpumask_empty(tmp))
3470 val += PENALTY_FOR_NODE_WITH_CPUS;
3472 /* Slight preference for less loaded node */
3473 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3474 val += node_load[n];
3476 if (val < min_val) {
3483 node_set(best_node, *used_node_mask);
3490 * Build zonelists ordered by node and zones within node.
3491 * This results in maximum locality--normal zone overflows into local
3492 * DMA zone, if any--but risks exhausting DMA zone.
3494 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3497 struct zonelist *zonelist;
3499 zonelist = &pgdat->node_zonelists[0];
3500 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3502 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3503 zonelist->_zonerefs[j].zone = NULL;
3504 zonelist->_zonerefs[j].zone_idx = 0;
3508 * Build gfp_thisnode zonelists
3510 static void build_thisnode_zonelists(pg_data_t *pgdat)
3513 struct zonelist *zonelist;
3515 zonelist = &pgdat->node_zonelists[1];
3516 j = build_zonelists_node(pgdat, zonelist, 0);
3517 zonelist->_zonerefs[j].zone = NULL;
3518 zonelist->_zonerefs[j].zone_idx = 0;
3522 * Build zonelists ordered by zone and nodes within zones.
3523 * This results in conserving DMA zone[s] until all Normal memory is
3524 * exhausted, but results in overflowing to remote node while memory
3525 * may still exist in local DMA zone.
3527 static int node_order[MAX_NUMNODES];
3529 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3532 int zone_type; /* needs to be signed */
3534 struct zonelist *zonelist;
3536 zonelist = &pgdat->node_zonelists[0];
3538 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3539 for (j = 0; j < nr_nodes; j++) {
3540 node = node_order[j];
3541 z = &NODE_DATA(node)->node_zones[zone_type];
3542 if (populated_zone(z)) {
3544 &zonelist->_zonerefs[pos++]);
3545 check_highest_zone(zone_type);
3549 zonelist->_zonerefs[pos].zone = NULL;
3550 zonelist->_zonerefs[pos].zone_idx = 0;
3553 static int default_zonelist_order(void)
3556 unsigned long low_kmem_size, total_size;
3560 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3561 * If they are really small and used heavily, the system can fall
3562 * into OOM very easily.
3563 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3565 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3568 for_each_online_node(nid) {
3569 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3570 z = &NODE_DATA(nid)->node_zones[zone_type];
3571 if (populated_zone(z)) {
3572 if (zone_type < ZONE_NORMAL)
3573 low_kmem_size += z->managed_pages;
3574 total_size += z->managed_pages;
3575 } else if (zone_type == ZONE_NORMAL) {
3577 * If any node has only lowmem, then node order
3578 * is preferred to allow kernel allocations
3579 * locally; otherwise, they can easily infringe
3580 * on other nodes when there is an abundance of
3581 * lowmem available to allocate from.
3583 return ZONELIST_ORDER_NODE;
3587 if (!low_kmem_size || /* there are no DMA area. */
3588 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3589 return ZONELIST_ORDER_NODE;
3591 * look into each node's config.
3592 * If there is a node whose DMA/DMA32 memory is very big area on
3593 * local memory, NODE_ORDER may be suitable.
3595 average_size = total_size /
3596 (nodes_weight(node_states[N_MEMORY]) + 1);
3597 for_each_online_node(nid) {
3600 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3601 z = &NODE_DATA(nid)->node_zones[zone_type];
3602 if (populated_zone(z)) {
3603 if (zone_type < ZONE_NORMAL)
3604 low_kmem_size += z->present_pages;
3605 total_size += z->present_pages;
3608 if (low_kmem_size &&
3609 total_size > average_size && /* ignore small node */
3610 low_kmem_size > total_size * 70/100)
3611 return ZONELIST_ORDER_NODE;
3613 return ZONELIST_ORDER_ZONE;
3616 static void set_zonelist_order(void)
3618 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3619 current_zonelist_order = default_zonelist_order();
3621 current_zonelist_order = user_zonelist_order;
3624 static void build_zonelists(pg_data_t *pgdat)
3628 nodemask_t used_mask;
3629 int local_node, prev_node;
3630 struct zonelist *zonelist;
3631 int order = current_zonelist_order;
3633 /* initialize zonelists */
3634 for (i = 0; i < MAX_ZONELISTS; i++) {
3635 zonelist = pgdat->node_zonelists + i;
3636 zonelist->_zonerefs[0].zone = NULL;
3637 zonelist->_zonerefs[0].zone_idx = 0;
3640 /* NUMA-aware ordering of nodes */
3641 local_node = pgdat->node_id;
3642 load = nr_online_nodes;
3643 prev_node = local_node;
3644 nodes_clear(used_mask);
3646 memset(node_order, 0, sizeof(node_order));
3649 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3651 * We don't want to pressure a particular node.
3652 * So adding penalty to the first node in same
3653 * distance group to make it round-robin.
3655 if (node_distance(local_node, node) !=
3656 node_distance(local_node, prev_node))
3657 node_load[node] = load;
3661 if (order == ZONELIST_ORDER_NODE)
3662 build_zonelists_in_node_order(pgdat, node);
3664 node_order[j++] = node; /* remember order */
3667 if (order == ZONELIST_ORDER_ZONE) {
3668 /* calculate node order -- i.e., DMA last! */
3669 build_zonelists_in_zone_order(pgdat, j);
3672 build_thisnode_zonelists(pgdat);
3675 /* Construct the zonelist performance cache - see further mmzone.h */
3676 static void build_zonelist_cache(pg_data_t *pgdat)
3678 struct zonelist *zonelist;
3679 struct zonelist_cache *zlc;
3682 zonelist = &pgdat->node_zonelists[0];
3683 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3684 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3685 for (z = zonelist->_zonerefs; z->zone; z++)
3686 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3689 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3691 * Return node id of node used for "local" allocations.
3692 * I.e., first node id of first zone in arg node's generic zonelist.
3693 * Used for initializing percpu 'numa_mem', which is used primarily
3694 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3696 int local_memory_node(int node)
3700 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3701 gfp_zone(GFP_KERNEL),
3708 #else /* CONFIG_NUMA */
3710 static void set_zonelist_order(void)
3712 current_zonelist_order = ZONELIST_ORDER_ZONE;
3715 static void build_zonelists(pg_data_t *pgdat)
3717 int node, local_node;
3719 struct zonelist *zonelist;
3721 local_node = pgdat->node_id;
3723 zonelist = &pgdat->node_zonelists[0];
3724 j = build_zonelists_node(pgdat, zonelist, 0);
3727 * Now we build the zonelist so that it contains the zones
3728 * of all the other nodes.
3729 * We don't want to pressure a particular node, so when
3730 * building the zones for node N, we make sure that the
3731 * zones coming right after the local ones are those from
3732 * node N+1 (modulo N)
3734 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3735 if (!node_online(node))
3737 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3739 for (node = 0; node < local_node; node++) {
3740 if (!node_online(node))
3742 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3745 zonelist->_zonerefs[j].zone = NULL;
3746 zonelist->_zonerefs[j].zone_idx = 0;
3749 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3750 static void build_zonelist_cache(pg_data_t *pgdat)
3752 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3755 #endif /* CONFIG_NUMA */
3758 * Boot pageset table. One per cpu which is going to be used for all
3759 * zones and all nodes. The parameters will be set in such a way
3760 * that an item put on a list will immediately be handed over to
3761 * the buddy list. This is safe since pageset manipulation is done
3762 * with interrupts disabled.
3764 * The boot_pagesets must be kept even after bootup is complete for
3765 * unused processors and/or zones. They do play a role for bootstrapping
3766 * hotplugged processors.
3768 * zoneinfo_show() and maybe other functions do
3769 * not check if the processor is online before following the pageset pointer.
3770 * Other parts of the kernel may not check if the zone is available.
3772 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3773 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3774 static void setup_zone_pageset(struct zone *zone);
3777 * Global mutex to protect against size modification of zonelists
3778 * as well as to serialize pageset setup for the new populated zone.
3780 DEFINE_MUTEX(zonelists_mutex);
3782 /* return values int ....just for stop_machine() */
3783 static int __build_all_zonelists(void *data)
3787 pg_data_t *self = data;
3790 memset(node_load, 0, sizeof(node_load));
3793 if (self && !node_online(self->node_id)) {
3794 build_zonelists(self);
3795 build_zonelist_cache(self);
3798 for_each_online_node(nid) {
3799 pg_data_t *pgdat = NODE_DATA(nid);
3801 build_zonelists(pgdat);
3802 build_zonelist_cache(pgdat);
3806 * Initialize the boot_pagesets that are going to be used
3807 * for bootstrapping processors. The real pagesets for
3808 * each zone will be allocated later when the per cpu
3809 * allocator is available.
3811 * boot_pagesets are used also for bootstrapping offline
3812 * cpus if the system is already booted because the pagesets
3813 * are needed to initialize allocators on a specific cpu too.
3814 * F.e. the percpu allocator needs the page allocator which
3815 * needs the percpu allocator in order to allocate its pagesets
3816 * (a chicken-egg dilemma).
3818 for_each_possible_cpu(cpu) {
3819 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3821 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3823 * We now know the "local memory node" for each node--
3824 * i.e., the node of the first zone in the generic zonelist.
3825 * Set up numa_mem percpu variable for on-line cpus. During
3826 * boot, only the boot cpu should be on-line; we'll init the
3827 * secondary cpus' numa_mem as they come on-line. During
3828 * node/memory hotplug, we'll fixup all on-line cpus.
3830 if (cpu_online(cpu))
3831 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3839 * Called with zonelists_mutex held always
3840 * unless system_state == SYSTEM_BOOTING.
3842 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3844 set_zonelist_order();
3846 if (system_state == SYSTEM_BOOTING) {
3847 __build_all_zonelists(NULL);
3848 mminit_verify_zonelist();
3849 cpuset_init_current_mems_allowed();
3851 #ifdef CONFIG_MEMORY_HOTPLUG
3853 setup_zone_pageset(zone);
3855 /* we have to stop all cpus to guarantee there is no user
3857 stop_machine(__build_all_zonelists, pgdat, NULL);
3858 /* cpuset refresh routine should be here */
3860 vm_total_pages = nr_free_pagecache_pages();
3862 * Disable grouping by mobility if the number of pages in the
3863 * system is too low to allow the mechanism to work. It would be
3864 * more accurate, but expensive to check per-zone. This check is
3865 * made on memory-hotadd so a system can start with mobility
3866 * disabled and enable it later
3868 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3869 page_group_by_mobility_disabled = 1;
3871 page_group_by_mobility_disabled = 0;
3873 printk("Built %i zonelists in %s order, mobility grouping %s. "
3874 "Total pages: %ld\n",
3876 zonelist_order_name[current_zonelist_order],
3877 page_group_by_mobility_disabled ? "off" : "on",
3880 printk("Policy zone: %s\n", zone_names[policy_zone]);
3885 * Helper functions to size the waitqueue hash table.
3886 * Essentially these want to choose hash table sizes sufficiently
3887 * large so that collisions trying to wait on pages are rare.
3888 * But in fact, the number of active page waitqueues on typical
3889 * systems is ridiculously low, less than 200. So this is even
3890 * conservative, even though it seems large.
3892 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3893 * waitqueues, i.e. the size of the waitq table given the number of pages.
3895 #define PAGES_PER_WAITQUEUE 256
3897 #ifndef CONFIG_MEMORY_HOTPLUG
3898 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3900 unsigned long size = 1;
3902 pages /= PAGES_PER_WAITQUEUE;
3904 while (size < pages)
3908 * Once we have dozens or even hundreds of threads sleeping
3909 * on IO we've got bigger problems than wait queue collision.
3910 * Limit the size of the wait table to a reasonable size.
3912 size = min(size, 4096UL);
3914 return max(size, 4UL);
3918 * A zone's size might be changed by hot-add, so it is not possible to determine
3919 * a suitable size for its wait_table. So we use the maximum size now.
3921 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3923 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3924 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3925 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3927 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3928 * or more by the traditional way. (See above). It equals:
3930 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3931 * ia64(16K page size) : = ( 8G + 4M)byte.
3932 * powerpc (64K page size) : = (32G +16M)byte.
3934 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3941 * This is an integer logarithm so that shifts can be used later
3942 * to extract the more random high bits from the multiplicative
3943 * hash function before the remainder is taken.
3945 static inline unsigned long wait_table_bits(unsigned long size)
3951 * Check if a pageblock contains reserved pages
3953 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3957 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3958 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3965 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3966 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3967 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3968 * higher will lead to a bigger reserve which will get freed as contiguous
3969 * blocks as reclaim kicks in
3971 static void setup_zone_migrate_reserve(struct zone *zone)
3973 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3975 unsigned long block_migratetype;
3980 * Get the start pfn, end pfn and the number of blocks to reserve
3981 * We have to be careful to be aligned to pageblock_nr_pages to
3982 * make sure that we always check pfn_valid for the first page in
3985 start_pfn = zone->zone_start_pfn;
3986 end_pfn = zone_end_pfn(zone);
3987 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3988 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3992 * Reserve blocks are generally in place to help high-order atomic
3993 * allocations that are short-lived. A min_free_kbytes value that
3994 * would result in more than 2 reserve blocks for atomic allocations
3995 * is assumed to be in place to help anti-fragmentation for the
3996 * future allocation of hugepages at runtime.
3998 reserve = min(2, reserve);
3999 old_reserve = zone->nr_migrate_reserve_block;
4001 /* When memory hot-add, we almost always need to do nothing */
4002 if (reserve == old_reserve)
4004 zone->nr_migrate_reserve_block = reserve;
4006 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4007 if (!pfn_valid(pfn))
4009 page = pfn_to_page(pfn);
4011 /* Watch out for overlapping nodes */
4012 if (page_to_nid(page) != zone_to_nid(zone))
4015 block_migratetype = get_pageblock_migratetype(page);
4017 /* Only test what is necessary when the reserves are not met */
4020 * Blocks with reserved pages will never free, skip
4023 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4024 if (pageblock_is_reserved(pfn, block_end_pfn))
4027 /* If this block is reserved, account for it */
4028 if (block_migratetype == MIGRATE_RESERVE) {
4033 /* Suitable for reserving if this block is movable */
4034 if (block_migratetype == MIGRATE_MOVABLE) {
4035 set_pageblock_migratetype(page,
4037 move_freepages_block(zone, page,
4042 } else if (!old_reserve) {
4044 * At boot time we don't need to scan the whole zone
4045 * for turning off MIGRATE_RESERVE.
4051 * If the reserve is met and this is a previous reserved block,
4054 if (block_migratetype == MIGRATE_RESERVE) {
4055 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4056 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4062 * Initially all pages are reserved - free ones are freed
4063 * up by free_all_bootmem() once the early boot process is
4064 * done. Non-atomic initialization, single-pass.
4066 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4067 unsigned long start_pfn, enum memmap_context context)
4070 unsigned long end_pfn = start_pfn + size;
4074 if (highest_memmap_pfn < end_pfn - 1)
4075 highest_memmap_pfn = end_pfn - 1;
4077 z = &NODE_DATA(nid)->node_zones[zone];
4078 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4080 * There can be holes in boot-time mem_map[]s
4081 * handed to this function. They do not
4082 * exist on hotplugged memory.
4084 if (context == MEMMAP_EARLY) {
4085 if (!early_pfn_valid(pfn))
4087 if (!early_pfn_in_nid(pfn, nid))
4090 page = pfn_to_page(pfn);
4091 set_page_links(page, zone, nid, pfn);
4092 mminit_verify_page_links(page, zone, nid, pfn);
4093 init_page_count(page);
4094 page_mapcount_reset(page);
4095 page_cpupid_reset_last(page);
4096 SetPageReserved(page);
4098 * Mark the block movable so that blocks are reserved for
4099 * movable at startup. This will force kernel allocations
4100 * to reserve their blocks rather than leaking throughout
4101 * the address space during boot when many long-lived
4102 * kernel allocations are made. Later some blocks near
4103 * the start are marked MIGRATE_RESERVE by
4104 * setup_zone_migrate_reserve()
4106 * bitmap is created for zone's valid pfn range. but memmap
4107 * can be created for invalid pages (for alignment)
4108 * check here not to call set_pageblock_migratetype() against
4111 if ((z->zone_start_pfn <= pfn)
4112 && (pfn < zone_end_pfn(z))
4113 && !(pfn & (pageblock_nr_pages - 1)))
4114 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4116 INIT_LIST_HEAD(&page->lru);
4117 #ifdef WANT_PAGE_VIRTUAL
4118 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4119 if (!is_highmem_idx(zone))
4120 set_page_address(page, __va(pfn << PAGE_SHIFT));
4125 static void __meminit zone_init_free_lists(struct zone *zone)
4127 unsigned int order, t;
4128 for_each_migratetype_order(order, t) {
4129 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4130 zone->free_area[order].nr_free = 0;
4134 #ifndef __HAVE_ARCH_MEMMAP_INIT
4135 #define memmap_init(size, nid, zone, start_pfn) \
4136 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4139 static int __meminit zone_batchsize(struct zone *zone)
4145 * The per-cpu-pages pools are set to around 1000th of the
4146 * size of the zone. But no more than 1/2 of a meg.
4148 * OK, so we don't know how big the cache is. So guess.
4150 batch = zone->managed_pages / 1024;
4151 if (batch * PAGE_SIZE > 512 * 1024)
4152 batch = (512 * 1024) / PAGE_SIZE;
4153 batch /= 4; /* We effectively *= 4 below */
4158 * Clamp the batch to a 2^n - 1 value. Having a power
4159 * of 2 value was found to be more likely to have
4160 * suboptimal cache aliasing properties in some cases.
4162 * For example if 2 tasks are alternately allocating
4163 * batches of pages, one task can end up with a lot
4164 * of pages of one half of the possible page colors
4165 * and the other with pages of the other colors.
4167 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4172 /* The deferral and batching of frees should be suppressed under NOMMU
4175 * The problem is that NOMMU needs to be able to allocate large chunks
4176 * of contiguous memory as there's no hardware page translation to
4177 * assemble apparent contiguous memory from discontiguous pages.
4179 * Queueing large contiguous runs of pages for batching, however,
4180 * causes the pages to actually be freed in smaller chunks. As there
4181 * can be a significant delay between the individual batches being
4182 * recycled, this leads to the once large chunks of space being
4183 * fragmented and becoming unavailable for high-order allocations.
4190 * pcp->high and pcp->batch values are related and dependent on one another:
4191 * ->batch must never be higher then ->high.
4192 * The following function updates them in a safe manner without read side
4195 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4196 * those fields changing asynchronously (acording the the above rule).
4198 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4199 * outside of boot time (or some other assurance that no concurrent updaters
4202 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4203 unsigned long batch)
4205 /* start with a fail safe value for batch */
4209 /* Update high, then batch, in order */
4216 /* a companion to pageset_set_high() */
4217 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4219 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4222 static void pageset_init(struct per_cpu_pageset *p)
4224 struct per_cpu_pages *pcp;
4227 memset(p, 0, sizeof(*p));
4231 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4232 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4235 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4238 pageset_set_batch(p, batch);
4242 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4243 * to the value high for the pageset p.
4245 static void pageset_set_high(struct per_cpu_pageset *p,
4248 unsigned long batch = max(1UL, high / 4);
4249 if ((high / 4) > (PAGE_SHIFT * 8))
4250 batch = PAGE_SHIFT * 8;
4252 pageset_update(&p->pcp, high, batch);
4255 static void __meminit pageset_set_high_and_batch(struct zone *zone,
4256 struct per_cpu_pageset *pcp)
4258 if (percpu_pagelist_fraction)
4259 pageset_set_high(pcp,
4260 (zone->managed_pages /
4261 percpu_pagelist_fraction));
4263 pageset_set_batch(pcp, zone_batchsize(zone));
4266 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4268 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4271 pageset_set_high_and_batch(zone, pcp);
4274 static void __meminit setup_zone_pageset(struct zone *zone)
4277 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4278 for_each_possible_cpu(cpu)
4279 zone_pageset_init(zone, cpu);
4283 * Allocate per cpu pagesets and initialize them.
4284 * Before this call only boot pagesets were available.
4286 void __init setup_per_cpu_pageset(void)
4290 for_each_populated_zone(zone)
4291 setup_zone_pageset(zone);
4294 static noinline __init_refok
4295 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4301 * The per-page waitqueue mechanism uses hashed waitqueues
4304 zone->wait_table_hash_nr_entries =
4305 wait_table_hash_nr_entries(zone_size_pages);
4306 zone->wait_table_bits =
4307 wait_table_bits(zone->wait_table_hash_nr_entries);
4308 alloc_size = zone->wait_table_hash_nr_entries
4309 * sizeof(wait_queue_head_t);
4311 if (!slab_is_available()) {
4312 zone->wait_table = (wait_queue_head_t *)
4313 memblock_virt_alloc_node_nopanic(
4314 alloc_size, zone->zone_pgdat->node_id);
4317 * This case means that a zone whose size was 0 gets new memory
4318 * via memory hot-add.
4319 * But it may be the case that a new node was hot-added. In
4320 * this case vmalloc() will not be able to use this new node's
4321 * memory - this wait_table must be initialized to use this new
4322 * node itself as well.
4323 * To use this new node's memory, further consideration will be
4326 zone->wait_table = vmalloc(alloc_size);
4328 if (!zone->wait_table)
4331 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4332 init_waitqueue_head(zone->wait_table + i);
4337 static __meminit void zone_pcp_init(struct zone *zone)
4340 * per cpu subsystem is not up at this point. The following code
4341 * relies on the ability of the linker to provide the
4342 * offset of a (static) per cpu variable into the per cpu area.
4344 zone->pageset = &boot_pageset;
4346 if (populated_zone(zone))
4347 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4348 zone->name, zone->present_pages,
4349 zone_batchsize(zone));
4352 int __meminit init_currently_empty_zone(struct zone *zone,
4353 unsigned long zone_start_pfn,
4355 enum memmap_context context)
4357 struct pglist_data *pgdat = zone->zone_pgdat;
4359 ret = zone_wait_table_init(zone, size);
4362 pgdat->nr_zones = zone_idx(zone) + 1;
4364 zone->zone_start_pfn = zone_start_pfn;
4366 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4367 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4369 (unsigned long)zone_idx(zone),
4370 zone_start_pfn, (zone_start_pfn + size));
4372 zone_init_free_lists(zone);
4377 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4378 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4380 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4381 * Architectures may implement their own version but if add_active_range()
4382 * was used and there are no special requirements, this is a convenient
4385 int __meminit __early_pfn_to_nid(unsigned long pfn)
4387 unsigned long start_pfn, end_pfn;
4390 * NOTE: The following SMP-unsafe globals are only used early in boot
4391 * when the kernel is running single-threaded.
4393 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4394 static int __meminitdata last_nid;
4396 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4399 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4401 last_start_pfn = start_pfn;
4402 last_end_pfn = end_pfn;
4408 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4410 int __meminit early_pfn_to_nid(unsigned long pfn)
4414 nid = __early_pfn_to_nid(pfn);
4417 /* just returns 0 */
4421 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4422 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4426 nid = __early_pfn_to_nid(pfn);
4427 if (nid >= 0 && nid != node)
4434 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4435 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4436 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4438 * If an architecture guarantees that all ranges registered with
4439 * add_active_ranges() contain no holes and may be freed, this
4440 * this function may be used instead of calling memblock_free_early_nid()
4443 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4445 unsigned long start_pfn, end_pfn;
4448 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4449 start_pfn = min(start_pfn, max_low_pfn);
4450 end_pfn = min(end_pfn, max_low_pfn);
4452 if (start_pfn < end_pfn)
4453 memblock_free_early_nid(PFN_PHYS(start_pfn),
4454 (end_pfn - start_pfn) << PAGE_SHIFT,
4460 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4461 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4463 * If an architecture guarantees that all ranges registered with
4464 * add_active_ranges() contain no holes and may be freed, this
4465 * function may be used instead of calling memory_present() manually.
4467 void __init sparse_memory_present_with_active_regions(int nid)
4469 unsigned long start_pfn, end_pfn;
4472 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4473 memory_present(this_nid, start_pfn, end_pfn);
4477 * get_pfn_range_for_nid - Return the start and end page frames for a node
4478 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4479 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4480 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4482 * It returns the start and end page frame of a node based on information
4483 * provided by an arch calling add_active_range(). If called for a node
4484 * with no available memory, a warning is printed and the start and end
4487 void __meminit get_pfn_range_for_nid(unsigned int nid,
4488 unsigned long *start_pfn, unsigned long *end_pfn)
4490 unsigned long this_start_pfn, this_end_pfn;
4496 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4497 *start_pfn = min(*start_pfn, this_start_pfn);
4498 *end_pfn = max(*end_pfn, this_end_pfn);
4501 if (*start_pfn == -1UL)
4506 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4507 * assumption is made that zones within a node are ordered in monotonic
4508 * increasing memory addresses so that the "highest" populated zone is used
4510 static void __init find_usable_zone_for_movable(void)
4513 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4514 if (zone_index == ZONE_MOVABLE)
4517 if (arch_zone_highest_possible_pfn[zone_index] >
4518 arch_zone_lowest_possible_pfn[zone_index])
4522 VM_BUG_ON(zone_index == -1);
4523 movable_zone = zone_index;
4527 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4528 * because it is sized independent of architecture. Unlike the other zones,
4529 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4530 * in each node depending on the size of each node and how evenly kernelcore
4531 * is distributed. This helper function adjusts the zone ranges
4532 * provided by the architecture for a given node by using the end of the
4533 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4534 * zones within a node are in order of monotonic increases memory addresses
4536 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4537 unsigned long zone_type,
4538 unsigned long node_start_pfn,
4539 unsigned long node_end_pfn,
4540 unsigned long *zone_start_pfn,
4541 unsigned long *zone_end_pfn)
4543 /* Only adjust if ZONE_MOVABLE is on this node */
4544 if (zone_movable_pfn[nid]) {
4545 /* Size ZONE_MOVABLE */
4546 if (zone_type == ZONE_MOVABLE) {
4547 *zone_start_pfn = zone_movable_pfn[nid];
4548 *zone_end_pfn = min(node_end_pfn,
4549 arch_zone_highest_possible_pfn[movable_zone]);
4551 /* Adjust for ZONE_MOVABLE starting within this range */
4552 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4553 *zone_end_pfn > zone_movable_pfn[nid]) {
4554 *zone_end_pfn = zone_movable_pfn[nid];
4556 /* Check if this whole range is within ZONE_MOVABLE */
4557 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4558 *zone_start_pfn = *zone_end_pfn;
4563 * Return the number of pages a zone spans in a node, including holes
4564 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4566 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4567 unsigned long zone_type,
4568 unsigned long node_start_pfn,
4569 unsigned long node_end_pfn,
4570 unsigned long *ignored)
4572 unsigned long zone_start_pfn, zone_end_pfn;
4574 /* Get the start and end of the zone */
4575 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4576 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4577 adjust_zone_range_for_zone_movable(nid, zone_type,
4578 node_start_pfn, node_end_pfn,
4579 &zone_start_pfn, &zone_end_pfn);
4581 /* Check that this node has pages within the zone's required range */
4582 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4585 /* Move the zone boundaries inside the node if necessary */
4586 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4587 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4589 /* Return the spanned pages */
4590 return zone_end_pfn - zone_start_pfn;
4594 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4595 * then all holes in the requested range will be accounted for.
4597 unsigned long __meminit __absent_pages_in_range(int nid,
4598 unsigned long range_start_pfn,
4599 unsigned long range_end_pfn)
4601 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4602 unsigned long start_pfn, end_pfn;
4605 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4606 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4607 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4608 nr_absent -= end_pfn - start_pfn;
4614 * absent_pages_in_range - Return number of page frames in holes within a range
4615 * @start_pfn: The start PFN to start searching for holes
4616 * @end_pfn: The end PFN to stop searching for holes
4618 * It returns the number of pages frames in memory holes within a range.
4620 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4621 unsigned long end_pfn)
4623 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4626 /* Return the number of page frames in holes in a zone on a node */
4627 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4628 unsigned long zone_type,
4629 unsigned long node_start_pfn,
4630 unsigned long node_end_pfn,
4631 unsigned long *ignored)
4633 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4634 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4635 unsigned long zone_start_pfn, zone_end_pfn;
4637 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4638 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4640 adjust_zone_range_for_zone_movable(nid, zone_type,
4641 node_start_pfn, node_end_pfn,
4642 &zone_start_pfn, &zone_end_pfn);
4643 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4646 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4647 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4648 unsigned long zone_type,
4649 unsigned long node_start_pfn,
4650 unsigned long node_end_pfn,
4651 unsigned long *zones_size)
4653 return zones_size[zone_type];
4656 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4657 unsigned long zone_type,
4658 unsigned long node_start_pfn,
4659 unsigned long node_end_pfn,
4660 unsigned long *zholes_size)
4665 return zholes_size[zone_type];
4668 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4670 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4671 unsigned long node_start_pfn,
4672 unsigned long node_end_pfn,
4673 unsigned long *zones_size,
4674 unsigned long *zholes_size)
4676 unsigned long realtotalpages, totalpages = 0;
4679 for (i = 0; i < MAX_NR_ZONES; i++)
4680 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4684 pgdat->node_spanned_pages = totalpages;
4686 realtotalpages = totalpages;
4687 for (i = 0; i < MAX_NR_ZONES; i++)
4689 zone_absent_pages_in_node(pgdat->node_id, i,
4690 node_start_pfn, node_end_pfn,
4692 pgdat->node_present_pages = realtotalpages;
4693 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4697 #ifndef CONFIG_SPARSEMEM
4699 * Calculate the size of the zone->blockflags rounded to an unsigned long
4700 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4701 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4702 * round what is now in bits to nearest long in bits, then return it in
4705 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4707 unsigned long usemapsize;
4709 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4710 usemapsize = roundup(zonesize, pageblock_nr_pages);
4711 usemapsize = usemapsize >> pageblock_order;
4712 usemapsize *= NR_PAGEBLOCK_BITS;
4713 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4715 return usemapsize / 8;
4718 static void __init setup_usemap(struct pglist_data *pgdat,
4720 unsigned long zone_start_pfn,
4721 unsigned long zonesize)
4723 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4724 zone->pageblock_flags = NULL;
4726 zone->pageblock_flags =
4727 memblock_virt_alloc_node_nopanic(usemapsize,
4731 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4732 unsigned long zone_start_pfn, unsigned long zonesize) {}
4733 #endif /* CONFIG_SPARSEMEM */
4735 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4737 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4738 void __paginginit set_pageblock_order(void)
4742 /* Check that pageblock_nr_pages has not already been setup */
4743 if (pageblock_order)
4746 if (HPAGE_SHIFT > PAGE_SHIFT)
4747 order = HUGETLB_PAGE_ORDER;
4749 order = MAX_ORDER - 1;
4752 * Assume the largest contiguous order of interest is a huge page.
4753 * This value may be variable depending on boot parameters on IA64 and
4756 pageblock_order = order;
4758 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4761 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4762 * is unused as pageblock_order is set at compile-time. See
4763 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4766 void __paginginit set_pageblock_order(void)
4770 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4772 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4773 unsigned long present_pages)
4775 unsigned long pages = spanned_pages;
4778 * Provide a more accurate estimation if there are holes within
4779 * the zone and SPARSEMEM is in use. If there are holes within the
4780 * zone, each populated memory region may cost us one or two extra
4781 * memmap pages due to alignment because memmap pages for each
4782 * populated regions may not naturally algined on page boundary.
4783 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4785 if (spanned_pages > present_pages + (present_pages >> 4) &&
4786 IS_ENABLED(CONFIG_SPARSEMEM))
4787 pages = present_pages;
4789 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4793 * Set up the zone data structures:
4794 * - mark all pages reserved
4795 * - mark all memory queues empty
4796 * - clear the memory bitmaps
4798 * NOTE: pgdat should get zeroed by caller.
4800 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4801 unsigned long node_start_pfn, unsigned long node_end_pfn,
4802 unsigned long *zones_size, unsigned long *zholes_size)
4805 int nid = pgdat->node_id;
4806 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4809 pgdat_resize_init(pgdat);
4810 #ifdef CONFIG_NUMA_BALANCING
4811 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4812 pgdat->numabalancing_migrate_nr_pages = 0;
4813 pgdat->numabalancing_migrate_next_window = jiffies;
4815 init_waitqueue_head(&pgdat->kswapd_wait);
4816 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4817 pgdat_page_cgroup_init(pgdat);
4819 for (j = 0; j < MAX_NR_ZONES; j++) {
4820 struct zone *zone = pgdat->node_zones + j;
4821 unsigned long size, realsize, freesize, memmap_pages;
4823 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4824 node_end_pfn, zones_size);
4825 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4831 * Adjust freesize so that it accounts for how much memory
4832 * is used by this zone for memmap. This affects the watermark
4833 * and per-cpu initialisations
4835 memmap_pages = calc_memmap_size(size, realsize);
4836 if (freesize >= memmap_pages) {
4837 freesize -= memmap_pages;
4840 " %s zone: %lu pages used for memmap\n",
4841 zone_names[j], memmap_pages);
4844 " %s zone: %lu pages exceeds freesize %lu\n",
4845 zone_names[j], memmap_pages, freesize);
4847 /* Account for reserved pages */
4848 if (j == 0 && freesize > dma_reserve) {
4849 freesize -= dma_reserve;
4850 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4851 zone_names[0], dma_reserve);
4854 if (!is_highmem_idx(j))
4855 nr_kernel_pages += freesize;
4856 /* Charge for highmem memmap if there are enough kernel pages */
4857 else if (nr_kernel_pages > memmap_pages * 2)
4858 nr_kernel_pages -= memmap_pages;
4859 nr_all_pages += freesize;
4861 zone->spanned_pages = size;
4862 zone->present_pages = realsize;
4864 * Set an approximate value for lowmem here, it will be adjusted
4865 * when the bootmem allocator frees pages into the buddy system.
4866 * And all highmem pages will be managed by the buddy system.
4868 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4871 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4873 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4875 zone->name = zone_names[j];
4876 spin_lock_init(&zone->lock);
4877 spin_lock_init(&zone->lru_lock);
4878 zone_seqlock_init(zone);
4879 zone->zone_pgdat = pgdat;
4880 zone_pcp_init(zone);
4882 /* For bootup, initialized properly in watermark setup */
4883 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4885 lruvec_init(&zone->lruvec);
4889 set_pageblock_order();
4890 setup_usemap(pgdat, zone, zone_start_pfn, size);
4891 ret = init_currently_empty_zone(zone, zone_start_pfn,
4892 size, MEMMAP_EARLY);
4894 memmap_init(size, nid, j, zone_start_pfn);
4895 zone_start_pfn += size;
4899 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4901 /* Skip empty nodes */
4902 if (!pgdat->node_spanned_pages)
4905 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4906 /* ia64 gets its own node_mem_map, before this, without bootmem */
4907 if (!pgdat->node_mem_map) {
4908 unsigned long size, start, end;
4912 * The zone's endpoints aren't required to be MAX_ORDER
4913 * aligned but the node_mem_map endpoints must be in order
4914 * for the buddy allocator to function correctly.
4916 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4917 end = pgdat_end_pfn(pgdat);
4918 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4919 size = (end - start) * sizeof(struct page);
4920 map = alloc_remap(pgdat->node_id, size);
4922 map = memblock_virt_alloc_node_nopanic(size,
4924 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4926 #ifndef CONFIG_NEED_MULTIPLE_NODES
4928 * With no DISCONTIG, the global mem_map is just set as node 0's
4930 if (pgdat == NODE_DATA(0)) {
4931 mem_map = NODE_DATA(0)->node_mem_map;
4932 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4933 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4934 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4935 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4938 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4941 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4942 unsigned long node_start_pfn, unsigned long *zholes_size)
4944 pg_data_t *pgdat = NODE_DATA(nid);
4945 unsigned long start_pfn = 0;
4946 unsigned long end_pfn = 0;
4948 /* pg_data_t should be reset to zero when it's allocated */
4949 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4951 pgdat->node_id = nid;
4952 pgdat->node_start_pfn = node_start_pfn;
4953 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4954 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4956 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4957 zones_size, zholes_size);
4959 alloc_node_mem_map(pgdat);
4960 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4961 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4962 nid, (unsigned long)pgdat,
4963 (unsigned long)pgdat->node_mem_map);
4966 free_area_init_core(pgdat, start_pfn, end_pfn,
4967 zones_size, zholes_size);
4970 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4972 #if MAX_NUMNODES > 1
4974 * Figure out the number of possible node ids.
4976 void __init setup_nr_node_ids(void)
4979 unsigned int highest = 0;
4981 for_each_node_mask(node, node_possible_map)
4983 nr_node_ids = highest + 1;
4988 * node_map_pfn_alignment - determine the maximum internode alignment
4990 * This function should be called after node map is populated and sorted.
4991 * It calculates the maximum power of two alignment which can distinguish
4994 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4995 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4996 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4997 * shifted, 1GiB is enough and this function will indicate so.
4999 * This is used to test whether pfn -> nid mapping of the chosen memory
5000 * model has fine enough granularity to avoid incorrect mapping for the
5001 * populated node map.
5003 * Returns the determined alignment in pfn's. 0 if there is no alignment
5004 * requirement (single node).
5006 unsigned long __init node_map_pfn_alignment(void)
5008 unsigned long accl_mask = 0, last_end = 0;
5009 unsigned long start, end, mask;
5013 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5014 if (!start || last_nid < 0 || last_nid == nid) {
5021 * Start with a mask granular enough to pin-point to the
5022 * start pfn and tick off bits one-by-one until it becomes
5023 * too coarse to separate the current node from the last.
5025 mask = ~((1 << __ffs(start)) - 1);
5026 while (mask && last_end <= (start & (mask << 1)))
5029 /* accumulate all internode masks */
5033 /* convert mask to number of pages */
5034 return ~accl_mask + 1;
5037 /* Find the lowest pfn for a node */
5038 static unsigned long __init find_min_pfn_for_node(int nid)
5040 unsigned long min_pfn = ULONG_MAX;
5041 unsigned long start_pfn;
5044 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5045 min_pfn = min(min_pfn, start_pfn);
5047 if (min_pfn == ULONG_MAX) {
5049 "Could not find start_pfn for node %d\n", nid);
5057 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5059 * It returns the minimum PFN based on information provided via
5060 * add_active_range().
5062 unsigned long __init find_min_pfn_with_active_regions(void)
5064 return find_min_pfn_for_node(MAX_NUMNODES);
5068 * early_calculate_totalpages()
5069 * Sum pages in active regions for movable zone.
5070 * Populate N_MEMORY for calculating usable_nodes.
5072 static unsigned long __init early_calculate_totalpages(void)
5074 unsigned long totalpages = 0;
5075 unsigned long start_pfn, end_pfn;
5078 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5079 unsigned long pages = end_pfn - start_pfn;
5081 totalpages += pages;
5083 node_set_state(nid, N_MEMORY);
5089 * Find the PFN the Movable zone begins in each node. Kernel memory
5090 * is spread evenly between nodes as long as the nodes have enough
5091 * memory. When they don't, some nodes will have more kernelcore than
5094 static void __init find_zone_movable_pfns_for_nodes(void)
5097 unsigned long usable_startpfn;
5098 unsigned long kernelcore_node, kernelcore_remaining;
5099 /* save the state before borrow the nodemask */
5100 nodemask_t saved_node_state = node_states[N_MEMORY];
5101 unsigned long totalpages = early_calculate_totalpages();
5102 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5103 struct memblock_region *r;
5105 /* Need to find movable_zone earlier when movable_node is specified. */
5106 find_usable_zone_for_movable();
5109 * If movable_node is specified, ignore kernelcore and movablecore
5112 if (movable_node_is_enabled()) {
5113 for_each_memblock(memory, r) {
5114 if (!memblock_is_hotpluggable(r))
5119 usable_startpfn = PFN_DOWN(r->base);
5120 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5121 min(usable_startpfn, zone_movable_pfn[nid]) :
5129 * If movablecore=nn[KMG] was specified, calculate what size of
5130 * kernelcore that corresponds so that memory usable for
5131 * any allocation type is evenly spread. If both kernelcore
5132 * and movablecore are specified, then the value of kernelcore
5133 * will be used for required_kernelcore if it's greater than
5134 * what movablecore would have allowed.
5136 if (required_movablecore) {
5137 unsigned long corepages;
5140 * Round-up so that ZONE_MOVABLE is at least as large as what
5141 * was requested by the user
5143 required_movablecore =
5144 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5145 corepages = totalpages - required_movablecore;
5147 required_kernelcore = max(required_kernelcore, corepages);
5150 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5151 if (!required_kernelcore)
5154 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5155 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5158 /* Spread kernelcore memory as evenly as possible throughout nodes */
5159 kernelcore_node = required_kernelcore / usable_nodes;
5160 for_each_node_state(nid, N_MEMORY) {
5161 unsigned long start_pfn, end_pfn;
5164 * Recalculate kernelcore_node if the division per node
5165 * now exceeds what is necessary to satisfy the requested
5166 * amount of memory for the kernel
5168 if (required_kernelcore < kernelcore_node)
5169 kernelcore_node = required_kernelcore / usable_nodes;
5172 * As the map is walked, we track how much memory is usable
5173 * by the kernel using kernelcore_remaining. When it is
5174 * 0, the rest of the node is usable by ZONE_MOVABLE
5176 kernelcore_remaining = kernelcore_node;
5178 /* Go through each range of PFNs within this node */
5179 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5180 unsigned long size_pages;
5182 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5183 if (start_pfn >= end_pfn)
5186 /* Account for what is only usable for kernelcore */
5187 if (start_pfn < usable_startpfn) {
5188 unsigned long kernel_pages;
5189 kernel_pages = min(end_pfn, usable_startpfn)
5192 kernelcore_remaining -= min(kernel_pages,
5193 kernelcore_remaining);
5194 required_kernelcore -= min(kernel_pages,
5195 required_kernelcore);
5197 /* Continue if range is now fully accounted */
5198 if (end_pfn <= usable_startpfn) {
5201 * Push zone_movable_pfn to the end so
5202 * that if we have to rebalance
5203 * kernelcore across nodes, we will
5204 * not double account here
5206 zone_movable_pfn[nid] = end_pfn;
5209 start_pfn = usable_startpfn;
5213 * The usable PFN range for ZONE_MOVABLE is from
5214 * start_pfn->end_pfn. Calculate size_pages as the
5215 * number of pages used as kernelcore
5217 size_pages = end_pfn - start_pfn;
5218 if (size_pages > kernelcore_remaining)
5219 size_pages = kernelcore_remaining;
5220 zone_movable_pfn[nid] = start_pfn + size_pages;
5223 * Some kernelcore has been met, update counts and
5224 * break if the kernelcore for this node has been
5227 required_kernelcore -= min(required_kernelcore,
5229 kernelcore_remaining -= size_pages;
5230 if (!kernelcore_remaining)
5236 * If there is still required_kernelcore, we do another pass with one
5237 * less node in the count. This will push zone_movable_pfn[nid] further
5238 * along on the nodes that still have memory until kernelcore is
5242 if (usable_nodes && required_kernelcore > usable_nodes)
5246 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5247 for (nid = 0; nid < MAX_NUMNODES; nid++)
5248 zone_movable_pfn[nid] =
5249 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5252 /* restore the node_state */
5253 node_states[N_MEMORY] = saved_node_state;
5256 /* Any regular or high memory on that node ? */
5257 static void check_for_memory(pg_data_t *pgdat, int nid)
5259 enum zone_type zone_type;
5261 if (N_MEMORY == N_NORMAL_MEMORY)
5264 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5265 struct zone *zone = &pgdat->node_zones[zone_type];
5266 if (populated_zone(zone)) {
5267 node_set_state(nid, N_HIGH_MEMORY);
5268 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5269 zone_type <= ZONE_NORMAL)
5270 node_set_state(nid, N_NORMAL_MEMORY);
5277 * free_area_init_nodes - Initialise all pg_data_t and zone data
5278 * @max_zone_pfn: an array of max PFNs for each zone
5280 * This will call free_area_init_node() for each active node in the system.
5281 * Using the page ranges provided by add_active_range(), the size of each
5282 * zone in each node and their holes is calculated. If the maximum PFN
5283 * between two adjacent zones match, it is assumed that the zone is empty.
5284 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5285 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5286 * starts where the previous one ended. For example, ZONE_DMA32 starts
5287 * at arch_max_dma_pfn.
5289 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5291 unsigned long start_pfn, end_pfn;
5294 /* Record where the zone boundaries are */
5295 memset(arch_zone_lowest_possible_pfn, 0,
5296 sizeof(arch_zone_lowest_possible_pfn));
5297 memset(arch_zone_highest_possible_pfn, 0,
5298 sizeof(arch_zone_highest_possible_pfn));
5299 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5300 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5301 for (i = 1; i < MAX_NR_ZONES; i++) {
5302 if (i == ZONE_MOVABLE)
5304 arch_zone_lowest_possible_pfn[i] =
5305 arch_zone_highest_possible_pfn[i-1];
5306 arch_zone_highest_possible_pfn[i] =
5307 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5309 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5310 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5312 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5313 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5314 find_zone_movable_pfns_for_nodes();
5316 /* Print out the zone ranges */
5317 printk("Zone ranges:\n");
5318 for (i = 0; i < MAX_NR_ZONES; i++) {
5319 if (i == ZONE_MOVABLE)
5321 printk(KERN_CONT " %-8s ", zone_names[i]);
5322 if (arch_zone_lowest_possible_pfn[i] ==
5323 arch_zone_highest_possible_pfn[i])
5324 printk(KERN_CONT "empty\n");
5326 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5327 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5328 (arch_zone_highest_possible_pfn[i]
5329 << PAGE_SHIFT) - 1);
5332 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5333 printk("Movable zone start for each node\n");
5334 for (i = 0; i < MAX_NUMNODES; i++) {
5335 if (zone_movable_pfn[i])
5336 printk(" Node %d: %#010lx\n", i,
5337 zone_movable_pfn[i] << PAGE_SHIFT);
5340 /* Print out the early node map */
5341 printk("Early memory node ranges\n");
5342 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5343 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5344 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5346 /* Initialise every node */
5347 mminit_verify_pageflags_layout();
5348 setup_nr_node_ids();
5349 for_each_online_node(nid) {
5350 pg_data_t *pgdat = NODE_DATA(nid);
5351 free_area_init_node(nid, NULL,
5352 find_min_pfn_for_node(nid), NULL);
5354 /* Any memory on that node */
5355 if (pgdat->node_present_pages)
5356 node_set_state(nid, N_MEMORY);
5357 check_for_memory(pgdat, nid);
5361 static int __init cmdline_parse_core(char *p, unsigned long *core)
5363 unsigned long long coremem;
5367 coremem = memparse(p, &p);
5368 *core = coremem >> PAGE_SHIFT;
5370 /* Paranoid check that UL is enough for the coremem value */
5371 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5377 * kernelcore=size sets the amount of memory for use for allocations that
5378 * cannot be reclaimed or migrated.
5380 static int __init cmdline_parse_kernelcore(char *p)
5382 return cmdline_parse_core(p, &required_kernelcore);
5386 * movablecore=size sets the amount of memory for use for allocations that
5387 * can be reclaimed or migrated.
5389 static int __init cmdline_parse_movablecore(char *p)
5391 return cmdline_parse_core(p, &required_movablecore);
5394 early_param("kernelcore", cmdline_parse_kernelcore);
5395 early_param("movablecore", cmdline_parse_movablecore);
5397 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5399 void adjust_managed_page_count(struct page *page, long count)
5401 spin_lock(&managed_page_count_lock);
5402 page_zone(page)->managed_pages += count;
5403 totalram_pages += count;
5404 #ifdef CONFIG_HIGHMEM
5405 if (PageHighMem(page))
5406 totalhigh_pages += count;
5408 spin_unlock(&managed_page_count_lock);
5410 EXPORT_SYMBOL(adjust_managed_page_count);
5412 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5415 unsigned long pages = 0;
5417 start = (void *)PAGE_ALIGN((unsigned long)start);
5418 end = (void *)((unsigned long)end & PAGE_MASK);
5419 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5420 if ((unsigned int)poison <= 0xFF)
5421 memset(pos, poison, PAGE_SIZE);
5422 free_reserved_page(virt_to_page(pos));
5426 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5427 s, pages << (PAGE_SHIFT - 10), start, end);
5431 EXPORT_SYMBOL(free_reserved_area);
5433 #ifdef CONFIG_HIGHMEM
5434 void free_highmem_page(struct page *page)
5436 __free_reserved_page(page);
5438 page_zone(page)->managed_pages++;
5444 void __init mem_init_print_info(const char *str)
5446 unsigned long physpages, codesize, datasize, rosize, bss_size;
5447 unsigned long init_code_size, init_data_size;
5449 physpages = get_num_physpages();
5450 codesize = _etext - _stext;
5451 datasize = _edata - _sdata;
5452 rosize = __end_rodata - __start_rodata;
5453 bss_size = __bss_stop - __bss_start;
5454 init_data_size = __init_end - __init_begin;
5455 init_code_size = _einittext - _sinittext;
5458 * Detect special cases and adjust section sizes accordingly:
5459 * 1) .init.* may be embedded into .data sections
5460 * 2) .init.text.* may be out of [__init_begin, __init_end],
5461 * please refer to arch/tile/kernel/vmlinux.lds.S.
5462 * 3) .rodata.* may be embedded into .text or .data sections.
5464 #define adj_init_size(start, end, size, pos, adj) \
5466 if (start <= pos && pos < end && size > adj) \
5470 adj_init_size(__init_begin, __init_end, init_data_size,
5471 _sinittext, init_code_size);
5472 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5473 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5474 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5475 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5477 #undef adj_init_size
5479 printk("Memory: %luK/%luK available "
5480 "(%luK kernel code, %luK rwdata, %luK rodata, "
5481 "%luK init, %luK bss, %luK reserved"
5482 #ifdef CONFIG_HIGHMEM
5486 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5487 codesize >> 10, datasize >> 10, rosize >> 10,
5488 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5489 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5490 #ifdef CONFIG_HIGHMEM
5491 totalhigh_pages << (PAGE_SHIFT-10),
5493 str ? ", " : "", str ? str : "");
5497 * set_dma_reserve - set the specified number of pages reserved in the first zone
5498 * @new_dma_reserve: The number of pages to mark reserved
5500 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5501 * In the DMA zone, a significant percentage may be consumed by kernel image
5502 * and other unfreeable allocations which can skew the watermarks badly. This
5503 * function may optionally be used to account for unfreeable pages in the
5504 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5505 * smaller per-cpu batchsize.
5507 void __init set_dma_reserve(unsigned long new_dma_reserve)
5509 dma_reserve = new_dma_reserve;
5512 void __init free_area_init(unsigned long *zones_size)
5514 free_area_init_node(0, zones_size,
5515 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5518 static int page_alloc_cpu_notify(struct notifier_block *self,
5519 unsigned long action, void *hcpu)
5521 int cpu = (unsigned long)hcpu;
5523 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5524 lru_add_drain_cpu(cpu);
5528 * Spill the event counters of the dead processor
5529 * into the current processors event counters.
5530 * This artificially elevates the count of the current
5533 vm_events_fold_cpu(cpu);
5536 * Zero the differential counters of the dead processor
5537 * so that the vm statistics are consistent.
5539 * This is only okay since the processor is dead and cannot
5540 * race with what we are doing.
5542 cpu_vm_stats_fold(cpu);
5547 void __init page_alloc_init(void)
5549 hotcpu_notifier(page_alloc_cpu_notify, 0);
5553 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5554 * or min_free_kbytes changes.
5556 static void calculate_totalreserve_pages(void)
5558 struct pglist_data *pgdat;
5559 unsigned long reserve_pages = 0;
5560 enum zone_type i, j;
5562 for_each_online_pgdat(pgdat) {
5563 for (i = 0; i < MAX_NR_ZONES; i++) {
5564 struct zone *zone = pgdat->node_zones + i;
5565 unsigned long max = 0;
5567 /* Find valid and maximum lowmem_reserve in the zone */
5568 for (j = i; j < MAX_NR_ZONES; j++) {
5569 if (zone->lowmem_reserve[j] > max)
5570 max = zone->lowmem_reserve[j];
5573 /* we treat the high watermark as reserved pages. */
5574 max += high_wmark_pages(zone);
5576 if (max > zone->managed_pages)
5577 max = zone->managed_pages;
5578 reserve_pages += max;
5580 * Lowmem reserves are not available to
5581 * GFP_HIGHUSER page cache allocations and
5582 * kswapd tries to balance zones to their high
5583 * watermark. As a result, neither should be
5584 * regarded as dirtyable memory, to prevent a
5585 * situation where reclaim has to clean pages
5586 * in order to balance the zones.
5588 zone->dirty_balance_reserve = max;
5591 dirty_balance_reserve = reserve_pages;
5592 totalreserve_pages = reserve_pages;
5596 * setup_per_zone_lowmem_reserve - called whenever
5597 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5598 * has a correct pages reserved value, so an adequate number of
5599 * pages are left in the zone after a successful __alloc_pages().
5601 static void setup_per_zone_lowmem_reserve(void)
5603 struct pglist_data *pgdat;
5604 enum zone_type j, idx;
5606 for_each_online_pgdat(pgdat) {
5607 for (j = 0; j < MAX_NR_ZONES; j++) {
5608 struct zone *zone = pgdat->node_zones + j;
5609 unsigned long managed_pages = zone->managed_pages;
5611 zone->lowmem_reserve[j] = 0;
5615 struct zone *lower_zone;
5619 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5620 sysctl_lowmem_reserve_ratio[idx] = 1;
5622 lower_zone = pgdat->node_zones + idx;
5623 lower_zone->lowmem_reserve[j] = managed_pages /
5624 sysctl_lowmem_reserve_ratio[idx];
5625 managed_pages += lower_zone->managed_pages;
5630 /* update totalreserve_pages */
5631 calculate_totalreserve_pages();
5634 static void __setup_per_zone_wmarks(void)
5636 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5637 unsigned long lowmem_pages = 0;
5639 unsigned long flags;
5641 /* Calculate total number of !ZONE_HIGHMEM pages */
5642 for_each_zone(zone) {
5643 if (!is_highmem(zone))
5644 lowmem_pages += zone->managed_pages;
5647 for_each_zone(zone) {
5650 spin_lock_irqsave(&zone->lock, flags);
5651 tmp = (u64)pages_min * zone->managed_pages;
5652 do_div(tmp, lowmem_pages);
5653 if (is_highmem(zone)) {
5655 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5656 * need highmem pages, so cap pages_min to a small
5659 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5660 * deltas controls asynch page reclaim, and so should
5661 * not be capped for highmem.
5663 unsigned long min_pages;
5665 min_pages = zone->managed_pages / 1024;
5666 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5667 zone->watermark[WMARK_MIN] = min_pages;
5670 * If it's a lowmem zone, reserve a number of pages
5671 * proportionate to the zone's size.
5673 zone->watermark[WMARK_MIN] = tmp;
5676 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5677 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5679 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5680 high_wmark_pages(zone) -
5681 low_wmark_pages(zone) -
5682 zone_page_state(zone, NR_ALLOC_BATCH));
5684 setup_zone_migrate_reserve(zone);
5685 spin_unlock_irqrestore(&zone->lock, flags);
5688 /* update totalreserve_pages */
5689 calculate_totalreserve_pages();
5693 * setup_per_zone_wmarks - called when min_free_kbytes changes
5694 * or when memory is hot-{added|removed}
5696 * Ensures that the watermark[min,low,high] values for each zone are set
5697 * correctly with respect to min_free_kbytes.
5699 void setup_per_zone_wmarks(void)
5701 mutex_lock(&zonelists_mutex);
5702 __setup_per_zone_wmarks();
5703 mutex_unlock(&zonelists_mutex);
5707 * The inactive anon list should be small enough that the VM never has to
5708 * do too much work, but large enough that each inactive page has a chance
5709 * to be referenced again before it is swapped out.
5711 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5712 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5713 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5714 * the anonymous pages are kept on the inactive list.
5717 * memory ratio inactive anon
5718 * -------------------------------------
5727 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5729 unsigned int gb, ratio;
5731 /* Zone size in gigabytes */
5732 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5734 ratio = int_sqrt(10 * gb);
5738 zone->inactive_ratio = ratio;
5741 static void __meminit setup_per_zone_inactive_ratio(void)
5746 calculate_zone_inactive_ratio(zone);
5750 * Initialise min_free_kbytes.
5752 * For small machines we want it small (128k min). For large machines
5753 * we want it large (64MB max). But it is not linear, because network
5754 * bandwidth does not increase linearly with machine size. We use
5756 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5757 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5773 int __meminit init_per_zone_wmark_min(void)
5775 unsigned long lowmem_kbytes;
5776 int new_min_free_kbytes;
5778 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5779 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5781 if (new_min_free_kbytes > user_min_free_kbytes) {
5782 min_free_kbytes = new_min_free_kbytes;
5783 if (min_free_kbytes < 128)
5784 min_free_kbytes = 128;
5785 if (min_free_kbytes > 65536)
5786 min_free_kbytes = 65536;
5788 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5789 new_min_free_kbytes, user_min_free_kbytes);
5791 setup_per_zone_wmarks();
5792 refresh_zone_stat_thresholds();
5793 setup_per_zone_lowmem_reserve();
5794 setup_per_zone_inactive_ratio();
5797 module_init(init_per_zone_wmark_min)
5800 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5801 * that we can call two helper functions whenever min_free_kbytes
5804 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5805 void __user *buffer, size_t *length, loff_t *ppos)
5809 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5814 user_min_free_kbytes = min_free_kbytes;
5815 setup_per_zone_wmarks();
5821 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5822 void __user *buffer, size_t *length, loff_t *ppos)
5827 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5832 zone->min_unmapped_pages = (zone->managed_pages *
5833 sysctl_min_unmapped_ratio) / 100;
5837 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5838 void __user *buffer, size_t *length, loff_t *ppos)
5843 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5848 zone->min_slab_pages = (zone->managed_pages *
5849 sysctl_min_slab_ratio) / 100;
5855 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5856 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5857 * whenever sysctl_lowmem_reserve_ratio changes.
5859 * The reserve ratio obviously has absolutely no relation with the
5860 * minimum watermarks. The lowmem reserve ratio can only make sense
5861 * if in function of the boot time zone sizes.
5863 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5864 void __user *buffer, size_t *length, loff_t *ppos)
5866 proc_dointvec_minmax(table, write, buffer, length, ppos);
5867 setup_per_zone_lowmem_reserve();
5872 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5873 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5874 * pagelist can have before it gets flushed back to buddy allocator.
5876 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5877 void __user *buffer, size_t *length, loff_t *ppos)
5883 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5884 if (!write || (ret < 0))
5887 mutex_lock(&pcp_batch_high_lock);
5888 for_each_populated_zone(zone) {
5890 high = zone->managed_pages / percpu_pagelist_fraction;
5891 for_each_possible_cpu(cpu)
5892 pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5895 mutex_unlock(&pcp_batch_high_lock);
5899 int hashdist = HASHDIST_DEFAULT;
5902 static int __init set_hashdist(char *str)
5906 hashdist = simple_strtoul(str, &str, 0);
5909 __setup("hashdist=", set_hashdist);
5913 * allocate a large system hash table from bootmem
5914 * - it is assumed that the hash table must contain an exact power-of-2
5915 * quantity of entries
5916 * - limit is the number of hash buckets, not the total allocation size
5918 void *__init alloc_large_system_hash(const char *tablename,
5919 unsigned long bucketsize,
5920 unsigned long numentries,
5923 unsigned int *_hash_shift,
5924 unsigned int *_hash_mask,
5925 unsigned long low_limit,
5926 unsigned long high_limit)
5928 unsigned long long max = high_limit;
5929 unsigned long log2qty, size;
5932 /* allow the kernel cmdline to have a say */
5934 /* round applicable memory size up to nearest megabyte */
5935 numentries = nr_kernel_pages;
5937 /* It isn't necessary when PAGE_SIZE >= 1MB */
5938 if (PAGE_SHIFT < 20)
5939 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5941 /* limit to 1 bucket per 2^scale bytes of low memory */
5942 if (scale > PAGE_SHIFT)
5943 numentries >>= (scale - PAGE_SHIFT);
5945 numentries <<= (PAGE_SHIFT - scale);
5947 /* Make sure we've got at least a 0-order allocation.. */
5948 if (unlikely(flags & HASH_SMALL)) {
5949 /* Makes no sense without HASH_EARLY */
5950 WARN_ON(!(flags & HASH_EARLY));
5951 if (!(numentries >> *_hash_shift)) {
5952 numentries = 1UL << *_hash_shift;
5953 BUG_ON(!numentries);
5955 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5956 numentries = PAGE_SIZE / bucketsize;
5958 numentries = roundup_pow_of_two(numentries);
5960 /* limit allocation size to 1/16 total memory by default */
5962 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5963 do_div(max, bucketsize);
5965 max = min(max, 0x80000000ULL);
5967 if (numentries < low_limit)
5968 numentries = low_limit;
5969 if (numentries > max)
5972 log2qty = ilog2(numentries);
5975 size = bucketsize << log2qty;
5976 if (flags & HASH_EARLY)
5977 table = memblock_virt_alloc_nopanic(size, 0);
5979 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5982 * If bucketsize is not a power-of-two, we may free
5983 * some pages at the end of hash table which
5984 * alloc_pages_exact() automatically does
5986 if (get_order(size) < MAX_ORDER) {
5987 table = alloc_pages_exact(size, GFP_ATOMIC);
5988 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5991 } while (!table && size > PAGE_SIZE && --log2qty);
5994 panic("Failed to allocate %s hash table\n", tablename);
5996 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5999 ilog2(size) - PAGE_SHIFT,
6003 *_hash_shift = log2qty;
6005 *_hash_mask = (1 << log2qty) - 1;
6010 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6011 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6014 #ifdef CONFIG_SPARSEMEM
6015 return __pfn_to_section(pfn)->pageblock_flags;
6017 return zone->pageblock_flags;
6018 #endif /* CONFIG_SPARSEMEM */
6021 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6023 #ifdef CONFIG_SPARSEMEM
6024 pfn &= (PAGES_PER_SECTION-1);
6025 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6027 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6028 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6029 #endif /* CONFIG_SPARSEMEM */
6033 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
6034 * @page: The page within the block of interest
6035 * @start_bitidx: The first bit of interest to retrieve
6036 * @end_bitidx: The last bit of interest
6037 * returns pageblock_bits flags
6039 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6040 unsigned long end_bitidx,
6044 unsigned long *bitmap;
6045 unsigned long bitidx, word_bitidx;
6048 zone = page_zone(page);
6049 bitmap = get_pageblock_bitmap(zone, pfn);
6050 bitidx = pfn_to_bitidx(zone, pfn);
6051 word_bitidx = bitidx / BITS_PER_LONG;
6052 bitidx &= (BITS_PER_LONG-1);
6054 word = bitmap[word_bitidx];
6055 bitidx += end_bitidx;
6056 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6060 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6061 * @page: The page within the block of interest
6062 * @start_bitidx: The first bit of interest
6063 * @end_bitidx: The last bit of interest
6064 * @flags: The flags to set
6066 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6068 unsigned long end_bitidx,
6072 unsigned long *bitmap;
6073 unsigned long bitidx, word_bitidx;
6074 unsigned long old_word, word;
6076 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6078 zone = page_zone(page);
6079 bitmap = get_pageblock_bitmap(zone, pfn);
6080 bitidx = pfn_to_bitidx(zone, pfn);
6081 word_bitidx = bitidx / BITS_PER_LONG;
6082 bitidx &= (BITS_PER_LONG-1);
6084 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6086 bitidx += end_bitidx;
6087 mask <<= (BITS_PER_LONG - bitidx - 1);
6088 flags <<= (BITS_PER_LONG - bitidx - 1);
6090 word = ACCESS_ONCE(bitmap[word_bitidx]);
6092 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6093 if (word == old_word)
6100 * This function checks whether pageblock includes unmovable pages or not.
6101 * If @count is not zero, it is okay to include less @count unmovable pages
6103 * PageLRU check without isolation or lru_lock could race so that
6104 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6105 * expect this function should be exact.
6107 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6108 bool skip_hwpoisoned_pages)
6110 unsigned long pfn, iter, found;
6114 * For avoiding noise data, lru_add_drain_all() should be called
6115 * If ZONE_MOVABLE, the zone never contains unmovable pages
6117 if (zone_idx(zone) == ZONE_MOVABLE)
6119 mt = get_pageblock_migratetype(page);
6120 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6123 pfn = page_to_pfn(page);
6124 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6125 unsigned long check = pfn + iter;
6127 if (!pfn_valid_within(check))
6130 page = pfn_to_page(check);
6133 * Hugepages are not in LRU lists, but they're movable.
6134 * We need not scan over tail pages bacause we don't
6135 * handle each tail page individually in migration.
6137 if (PageHuge(page)) {
6138 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6143 * We can't use page_count without pin a page
6144 * because another CPU can free compound page.
6145 * This check already skips compound tails of THP
6146 * because their page->_count is zero at all time.
6148 if (!atomic_read(&page->_count)) {
6149 if (PageBuddy(page))
6150 iter += (1 << page_order(page)) - 1;
6155 * The HWPoisoned page may be not in buddy system, and
6156 * page_count() is not 0.
6158 if (skip_hwpoisoned_pages && PageHWPoison(page))
6164 * If there are RECLAIMABLE pages, we need to check it.
6165 * But now, memory offline itself doesn't call shrink_slab()
6166 * and it still to be fixed.
6169 * If the page is not RAM, page_count()should be 0.
6170 * we don't need more check. This is an _used_ not-movable page.
6172 * The problematic thing here is PG_reserved pages. PG_reserved
6173 * is set to both of a memory hole page and a _used_ kernel
6182 bool is_pageblock_removable_nolock(struct page *page)
6188 * We have to be careful here because we are iterating over memory
6189 * sections which are not zone aware so we might end up outside of
6190 * the zone but still within the section.
6191 * We have to take care about the node as well. If the node is offline
6192 * its NODE_DATA will be NULL - see page_zone.
6194 if (!node_online(page_to_nid(page)))
6197 zone = page_zone(page);
6198 pfn = page_to_pfn(page);
6199 if (!zone_spans_pfn(zone, pfn))
6202 return !has_unmovable_pages(zone, page, 0, true);
6207 static unsigned long pfn_max_align_down(unsigned long pfn)
6209 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6210 pageblock_nr_pages) - 1);
6213 static unsigned long pfn_max_align_up(unsigned long pfn)
6215 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6216 pageblock_nr_pages));
6219 /* [start, end) must belong to a single zone. */
6220 static int __alloc_contig_migrate_range(struct compact_control *cc,
6221 unsigned long start, unsigned long end)
6223 /* This function is based on compact_zone() from compaction.c. */
6224 unsigned long nr_reclaimed;
6225 unsigned long pfn = start;
6226 unsigned int tries = 0;
6231 while (pfn < end || !list_empty(&cc->migratepages)) {
6232 if (fatal_signal_pending(current)) {
6237 if (list_empty(&cc->migratepages)) {
6238 cc->nr_migratepages = 0;
6239 pfn = isolate_migratepages_range(cc->zone, cc,
6246 } else if (++tries == 5) {
6247 ret = ret < 0 ? ret : -EBUSY;
6251 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6253 cc->nr_migratepages -= nr_reclaimed;
6255 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6256 NULL, 0, cc->mode, MR_CMA);
6259 putback_movable_pages(&cc->migratepages);
6266 * alloc_contig_range() -- tries to allocate given range of pages
6267 * @start: start PFN to allocate
6268 * @end: one-past-the-last PFN to allocate
6269 * @migratetype: migratetype of the underlaying pageblocks (either
6270 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6271 * in range must have the same migratetype and it must
6272 * be either of the two.
6274 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6275 * aligned, however it's the caller's responsibility to guarantee that
6276 * we are the only thread that changes migrate type of pageblocks the
6279 * The PFN range must belong to a single zone.
6281 * Returns zero on success or negative error code. On success all
6282 * pages which PFN is in [start, end) are allocated for the caller and
6283 * need to be freed with free_contig_range().
6285 int alloc_contig_range(unsigned long start, unsigned long end,
6286 unsigned migratetype)
6288 unsigned long outer_start, outer_end;
6291 struct compact_control cc = {
6292 .nr_migratepages = 0,
6294 .zone = page_zone(pfn_to_page(start)),
6295 .mode = MIGRATE_SYNC,
6296 .ignore_skip_hint = true,
6298 INIT_LIST_HEAD(&cc.migratepages);
6301 * What we do here is we mark all pageblocks in range as
6302 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6303 * have different sizes, and due to the way page allocator
6304 * work, we align the range to biggest of the two pages so
6305 * that page allocator won't try to merge buddies from
6306 * different pageblocks and change MIGRATE_ISOLATE to some
6307 * other migration type.
6309 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6310 * migrate the pages from an unaligned range (ie. pages that
6311 * we are interested in). This will put all the pages in
6312 * range back to page allocator as MIGRATE_ISOLATE.
6314 * When this is done, we take the pages in range from page
6315 * allocator removing them from the buddy system. This way
6316 * page allocator will never consider using them.
6318 * This lets us mark the pageblocks back as
6319 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6320 * aligned range but not in the unaligned, original range are
6321 * put back to page allocator so that buddy can use them.
6324 ret = start_isolate_page_range(pfn_max_align_down(start),
6325 pfn_max_align_up(end), migratetype,
6330 ret = __alloc_contig_migrate_range(&cc, start, end);
6335 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6336 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6337 * more, all pages in [start, end) are free in page allocator.
6338 * What we are going to do is to allocate all pages from
6339 * [start, end) (that is remove them from page allocator).
6341 * The only problem is that pages at the beginning and at the
6342 * end of interesting range may be not aligned with pages that
6343 * page allocator holds, ie. they can be part of higher order
6344 * pages. Because of this, we reserve the bigger range and
6345 * once this is done free the pages we are not interested in.
6347 * We don't have to hold zone->lock here because the pages are
6348 * isolated thus they won't get removed from buddy.
6351 lru_add_drain_all();
6355 outer_start = start;
6356 while (!PageBuddy(pfn_to_page(outer_start))) {
6357 if (++order >= MAX_ORDER) {
6361 outer_start &= ~0UL << order;
6364 /* Make sure the range is really isolated. */
6365 if (test_pages_isolated(outer_start, end, false)) {
6366 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6373 /* Grab isolated pages from freelists. */
6374 outer_end = isolate_freepages_range(&cc, outer_start, end);
6380 /* Free head and tail (if any) */
6381 if (start != outer_start)
6382 free_contig_range(outer_start, start - outer_start);
6383 if (end != outer_end)
6384 free_contig_range(end, outer_end - end);
6387 undo_isolate_page_range(pfn_max_align_down(start),
6388 pfn_max_align_up(end), migratetype);
6392 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6394 unsigned int count = 0;
6396 for (; nr_pages--; pfn++) {
6397 struct page *page = pfn_to_page(pfn);
6399 count += page_count(page) != 1;
6402 WARN(count != 0, "%d pages are still in use!\n", count);
6406 #ifdef CONFIG_MEMORY_HOTPLUG
6408 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6409 * page high values need to be recalulated.
6411 void __meminit zone_pcp_update(struct zone *zone)
6414 mutex_lock(&pcp_batch_high_lock);
6415 for_each_possible_cpu(cpu)
6416 pageset_set_high_and_batch(zone,
6417 per_cpu_ptr(zone->pageset, cpu));
6418 mutex_unlock(&pcp_batch_high_lock);
6422 void zone_pcp_reset(struct zone *zone)
6424 unsigned long flags;
6426 struct per_cpu_pageset *pset;
6428 /* avoid races with drain_pages() */
6429 local_irq_save(flags);
6430 if (zone->pageset != &boot_pageset) {
6431 for_each_online_cpu(cpu) {
6432 pset = per_cpu_ptr(zone->pageset, cpu);
6433 drain_zonestat(zone, pset);
6435 free_percpu(zone->pageset);
6436 zone->pageset = &boot_pageset;
6438 local_irq_restore(flags);
6441 #ifdef CONFIG_MEMORY_HOTREMOVE
6443 * All pages in the range must be isolated before calling this.
6446 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6450 unsigned int order, i;
6452 unsigned long flags;
6453 /* find the first valid pfn */
6454 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6459 zone = page_zone(pfn_to_page(pfn));
6460 spin_lock_irqsave(&zone->lock, flags);
6462 while (pfn < end_pfn) {
6463 if (!pfn_valid(pfn)) {
6467 page = pfn_to_page(pfn);
6469 * The HWPoisoned page may be not in buddy system, and
6470 * page_count() is not 0.
6472 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6474 SetPageReserved(page);
6478 BUG_ON(page_count(page));
6479 BUG_ON(!PageBuddy(page));
6480 order = page_order(page);
6481 #ifdef CONFIG_DEBUG_VM
6482 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6483 pfn, 1 << order, end_pfn);
6485 list_del(&page->lru);
6486 rmv_page_order(page);
6487 zone->free_area[order].nr_free--;
6488 for (i = 0; i < (1 << order); i++)
6489 SetPageReserved((page+i));
6490 pfn += (1 << order);
6492 spin_unlock_irqrestore(&zone->lock, flags);
6496 #ifdef CONFIG_MEMORY_FAILURE
6497 bool is_free_buddy_page(struct page *page)
6499 struct zone *zone = page_zone(page);
6500 unsigned long pfn = page_to_pfn(page);
6501 unsigned long flags;
6504 spin_lock_irqsave(&zone->lock, flags);
6505 for (order = 0; order < MAX_ORDER; order++) {
6506 struct page *page_head = page - (pfn & ((1 << order) - 1));
6508 if (PageBuddy(page_head) && page_order(page_head) >= order)
6511 spin_unlock_irqrestore(&zone->lock, flags);
6513 return order < MAX_ORDER;
6517 static const struct trace_print_flags pageflag_names[] = {
6518 {1UL << PG_locked, "locked" },
6519 {1UL << PG_error, "error" },
6520 {1UL << PG_referenced, "referenced" },
6521 {1UL << PG_uptodate, "uptodate" },
6522 {1UL << PG_dirty, "dirty" },
6523 {1UL << PG_lru, "lru" },
6524 {1UL << PG_active, "active" },
6525 {1UL << PG_slab, "slab" },
6526 {1UL << PG_owner_priv_1, "owner_priv_1" },
6527 {1UL << PG_arch_1, "arch_1" },
6528 {1UL << PG_reserved, "reserved" },
6529 {1UL << PG_private, "private" },
6530 {1UL << PG_private_2, "private_2" },
6531 {1UL << PG_writeback, "writeback" },
6532 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6533 {1UL << PG_head, "head" },
6534 {1UL << PG_tail, "tail" },
6536 {1UL << PG_compound, "compound" },
6538 {1UL << PG_swapcache, "swapcache" },
6539 {1UL << PG_mappedtodisk, "mappedtodisk" },
6540 {1UL << PG_reclaim, "reclaim" },
6541 {1UL << PG_swapbacked, "swapbacked" },
6542 {1UL << PG_unevictable, "unevictable" },
6544 {1UL << PG_mlocked, "mlocked" },
6546 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6547 {1UL << PG_uncached, "uncached" },
6549 #ifdef CONFIG_MEMORY_FAILURE
6550 {1UL << PG_hwpoison, "hwpoison" },
6552 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6553 {1UL << PG_compound_lock, "compound_lock" },
6557 static void dump_page_flags(unsigned long flags)
6559 const char *delim = "";
6563 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6565 printk(KERN_ALERT "page flags: %#lx(", flags);
6567 /* remove zone id */
6568 flags &= (1UL << NR_PAGEFLAGS) - 1;
6570 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6572 mask = pageflag_names[i].mask;
6573 if ((flags & mask) != mask)
6577 printk("%s%s", delim, pageflag_names[i].name);
6581 /* check for left over flags */
6583 printk("%s%#lx", delim, flags);
6588 void dump_page_badflags(struct page *page, const char *reason,
6589 unsigned long badflags)
6592 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6593 page, atomic_read(&page->_count), page_mapcount(page),
6594 page->mapping, page->index);
6595 dump_page_flags(page->flags);
6597 pr_alert("page dumped because: %s\n", reason);
6598 if (page->flags & badflags) {
6599 pr_alert("bad because of flags:\n");
6600 dump_page_flags(page->flags & badflags);
6602 mem_cgroup_print_bad_page(page);
6605 void dump_page(struct page *page, const char *reason)
6607 dump_page_badflags(page, reason, 0);
6609 EXPORT_SYMBOL(dump_page);