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 %lu outside zone [ %lu - %lu ]\n",
265 pfn, start_pfn, start_pfn + sp);
270 static int page_is_consistent(struct zone *zone, struct page *page)
272 if (!pfn_valid_within(page_to_pfn(page)))
274 if (zone != page_zone(page))
280 * Temporary debugging check for pages not lying within a given zone.
282 static int bad_range(struct zone *zone, struct page *page)
284 if (page_outside_zone_boundaries(zone, page))
286 if (!page_is_consistent(zone, page))
292 static inline int bad_range(struct zone *zone, struct page *page)
298 static void bad_page(struct page *page, const char *reason,
299 unsigned long bad_flags)
301 static unsigned long resume;
302 static unsigned long nr_shown;
303 static unsigned long nr_unshown;
305 /* Don't complain about poisoned pages */
306 if (PageHWPoison(page)) {
307 page_mapcount_reset(page); /* remove PageBuddy */
312 * Allow a burst of 60 reports, then keep quiet for that minute;
313 * or allow a steady drip of one report per second.
315 if (nr_shown == 60) {
316 if (time_before(jiffies, resume)) {
322 "BUG: Bad page state: %lu messages suppressed\n",
329 resume = jiffies + 60 * HZ;
331 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
332 current->comm, page_to_pfn(page));
333 dump_page_badflags(page, reason, bad_flags);
338 /* Leave bad fields for debug, except PageBuddy could make trouble */
339 page_mapcount_reset(page); /* remove PageBuddy */
340 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
344 * Higher-order pages are called "compound pages". They are structured thusly:
346 * The first PAGE_SIZE page is called the "head page".
348 * The remaining PAGE_SIZE pages are called "tail pages".
350 * All pages have PG_compound set. All tail pages have their ->first_page
351 * pointing at the head page.
353 * The first tail page's ->lru.next holds the address of the compound page's
354 * put_page() function. Its ->lru.prev holds the order of allocation.
355 * This usage means that zero-order pages may not be compound.
358 static void free_compound_page(struct page *page)
360 __free_pages_ok(page, compound_order(page));
363 void prep_compound_page(struct page *page, unsigned long order)
366 int nr_pages = 1 << order;
368 set_compound_page_dtor(page, free_compound_page);
369 set_compound_order(page, order);
371 for (i = 1; i < nr_pages; i++) {
372 struct page *p = page + i;
373 set_page_count(p, 0);
374 p->first_page = page;
375 /* Make sure p->first_page is always valid for PageTail() */
381 /* update __split_huge_page_refcount if you change this function */
382 static int destroy_compound_page(struct page *page, unsigned long order)
385 int nr_pages = 1 << order;
388 if (unlikely(compound_order(page) != order)) {
389 bad_page(page, "wrong compound order", 0);
393 __ClearPageHead(page);
395 for (i = 1; i < nr_pages; i++) {
396 struct page *p = page + i;
398 if (unlikely(!PageTail(p))) {
399 bad_page(page, "PageTail not set", 0);
401 } else if (unlikely(p->first_page != page)) {
402 bad_page(page, "first_page not consistent", 0);
411 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
416 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
417 * and __GFP_HIGHMEM from hard or soft interrupt context.
419 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
420 for (i = 0; i < (1 << order); i++)
421 clear_highpage(page + i);
424 #ifdef CONFIG_DEBUG_PAGEALLOC
425 unsigned int _debug_guardpage_minorder;
427 static int __init debug_guardpage_minorder_setup(char *buf)
431 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
432 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
435 _debug_guardpage_minorder = res;
436 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
439 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
441 static inline void set_page_guard_flag(struct page *page)
443 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
446 static inline void clear_page_guard_flag(struct page *page)
448 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
451 static inline void set_page_guard_flag(struct page *page) { }
452 static inline void clear_page_guard_flag(struct page *page) { }
455 static inline void set_page_order(struct page *page, int order)
457 set_page_private(page, order);
458 __SetPageBuddy(page);
461 static inline void rmv_page_order(struct page *page)
463 __ClearPageBuddy(page);
464 set_page_private(page, 0);
468 * Locate the struct page for both the matching buddy in our
469 * pair (buddy1) and the combined O(n+1) page they form (page).
471 * 1) Any buddy B1 will have an order O twin B2 which satisfies
472 * the following equation:
474 * For example, if the starting buddy (buddy2) is #8 its order
476 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
478 * 2) Any buddy B will have an order O+1 parent P which
479 * satisfies the following equation:
482 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
484 static inline unsigned long
485 __find_buddy_index(unsigned long page_idx, unsigned int order)
487 return page_idx ^ (1 << order);
491 * This function checks whether a page is free && is the buddy
492 * we can do coalesce a page and its buddy if
493 * (a) the buddy is not in a hole &&
494 * (b) the buddy is in the buddy system &&
495 * (c) a page and its buddy have the same order &&
496 * (d) a page and its buddy are in the same zone.
498 * For recording whether a page is in the buddy system, we set ->_mapcount
499 * PAGE_BUDDY_MAPCOUNT_VALUE.
500 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
501 * serialized by zone->lock.
503 * For recording page's order, we use page_private(page).
505 static inline int page_is_buddy(struct page *page, struct page *buddy,
508 if (!pfn_valid_within(page_to_pfn(buddy)))
511 if (page_zone_id(page) != page_zone_id(buddy))
514 if (page_is_guard(buddy) && page_order(buddy) == order) {
515 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
519 if (PageBuddy(buddy) && page_order(buddy) == order) {
520 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
527 * Freeing function for a buddy system allocator.
529 * The concept of a buddy system is to maintain direct-mapped table
530 * (containing bit values) for memory blocks of various "orders".
531 * The bottom level table contains the map for the smallest allocatable
532 * units of memory (here, pages), and each level above it describes
533 * pairs of units from the levels below, hence, "buddies".
534 * At a high level, all that happens here is marking the table entry
535 * at the bottom level available, and propagating the changes upward
536 * as necessary, plus some accounting needed to play nicely with other
537 * parts of the VM system.
538 * At each level, we keep a list of pages, which are heads of continuous
539 * free pages of length of (1 << order) and marked with _mapcount
540 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
542 * So when we are allocating or freeing one, we can derive the state of the
543 * other. That is, if we allocate a small block, and both were
544 * free, the remainder of the region must be split into blocks.
545 * If a block is freed, and its buddy is also free, then this
546 * triggers coalescing into a block of larger size.
551 static inline void __free_one_page(struct page *page,
552 struct zone *zone, unsigned int order,
555 unsigned long page_idx;
556 unsigned long combined_idx;
557 unsigned long uninitialized_var(buddy_idx);
560 VM_BUG_ON(!zone_is_initialized(zone));
562 if (unlikely(PageCompound(page)))
563 if (unlikely(destroy_compound_page(page, order)))
566 VM_BUG_ON(migratetype == -1);
568 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
570 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
571 VM_BUG_ON_PAGE(bad_range(zone, page), page);
573 while (order < MAX_ORDER-1) {
574 buddy_idx = __find_buddy_index(page_idx, order);
575 buddy = page + (buddy_idx - page_idx);
576 if (!page_is_buddy(page, buddy, order))
579 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
580 * merge with it and move up one order.
582 if (page_is_guard(buddy)) {
583 clear_page_guard_flag(buddy);
584 set_page_private(page, 0);
585 __mod_zone_freepage_state(zone, 1 << order,
588 list_del(&buddy->lru);
589 zone->free_area[order].nr_free--;
590 rmv_page_order(buddy);
592 combined_idx = buddy_idx & page_idx;
593 page = page + (combined_idx - page_idx);
594 page_idx = combined_idx;
597 set_page_order(page, order);
600 * If this is not the largest possible page, check if the buddy
601 * of the next-highest order is free. If it is, it's possible
602 * that pages are being freed that will coalesce soon. In case,
603 * that is happening, add the free page to the tail of the list
604 * so it's less likely to be used soon and more likely to be merged
605 * as a higher order page
607 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
608 struct page *higher_page, *higher_buddy;
609 combined_idx = buddy_idx & page_idx;
610 higher_page = page + (combined_idx - page_idx);
611 buddy_idx = __find_buddy_index(combined_idx, order + 1);
612 higher_buddy = higher_page + (buddy_idx - combined_idx);
613 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
614 list_add_tail(&page->lru,
615 &zone->free_area[order].free_list[migratetype]);
620 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
622 zone->free_area[order].nr_free++;
625 static inline int free_pages_check(struct page *page)
627 const char *bad_reason = NULL;
628 unsigned long bad_flags = 0;
630 if (unlikely(page_mapcount(page)))
631 bad_reason = "nonzero mapcount";
632 if (unlikely(page->mapping != NULL))
633 bad_reason = "non-NULL mapping";
634 if (unlikely(atomic_read(&page->_count) != 0))
635 bad_reason = "nonzero _count";
636 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
637 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
638 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
640 if (unlikely(mem_cgroup_bad_page_check(page)))
641 bad_reason = "cgroup check failed";
642 if (unlikely(bad_reason)) {
643 bad_page(page, bad_reason, bad_flags);
646 page_cpupid_reset_last(page);
647 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
648 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
653 * Frees a number of pages from the PCP lists
654 * Assumes all pages on list are in same zone, and of same order.
655 * count is the number of pages to free.
657 * If the zone was previously in an "all pages pinned" state then look to
658 * see if this freeing clears that state.
660 * And clear the zone's pages_scanned counter, to hold off the "all pages are
661 * pinned" detection logic.
663 static void free_pcppages_bulk(struct zone *zone, int count,
664 struct per_cpu_pages *pcp)
670 spin_lock(&zone->lock);
671 zone->pages_scanned = 0;
675 struct list_head *list;
678 * Remove pages from lists in a round-robin fashion. A
679 * batch_free count is maintained that is incremented when an
680 * empty list is encountered. This is so more pages are freed
681 * off fuller lists instead of spinning excessively around empty
686 if (++migratetype == MIGRATE_PCPTYPES)
688 list = &pcp->lists[migratetype];
689 } while (list_empty(list));
691 /* This is the only non-empty list. Free them all. */
692 if (batch_free == MIGRATE_PCPTYPES)
693 batch_free = to_free;
696 int mt; /* migratetype of the to-be-freed page */
698 page = list_entry(list->prev, struct page, lru);
699 /* must delete as __free_one_page list manipulates */
700 list_del(&page->lru);
701 mt = get_freepage_migratetype(page);
702 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
703 __free_one_page(page, zone, 0, mt);
704 trace_mm_page_pcpu_drain(page, 0, mt);
705 if (likely(!is_migrate_isolate_page(page))) {
706 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
707 if (is_migrate_cma(mt))
708 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
710 } while (--to_free && --batch_free && !list_empty(list));
712 spin_unlock(&zone->lock);
715 static void free_one_page(struct zone *zone, struct page *page, int order,
718 spin_lock(&zone->lock);
719 zone->pages_scanned = 0;
721 __free_one_page(page, zone, order, migratetype);
722 if (unlikely(!is_migrate_isolate(migratetype)))
723 __mod_zone_freepage_state(zone, 1 << order, migratetype);
724 spin_unlock(&zone->lock);
727 static bool free_pages_prepare(struct page *page, unsigned int order)
732 trace_mm_page_free(page, order);
733 kmemcheck_free_shadow(page, order);
736 page->mapping = NULL;
737 for (i = 0; i < (1 << order); i++)
738 bad += free_pages_check(page + i);
742 if (!PageHighMem(page)) {
743 debug_check_no_locks_freed(page_address(page),
745 debug_check_no_obj_freed(page_address(page),
748 arch_free_page(page, order);
749 kernel_map_pages(page, 1 << order, 0);
754 static void __free_pages_ok(struct page *page, unsigned int order)
759 if (!free_pages_prepare(page, order))
762 local_irq_save(flags);
763 __count_vm_events(PGFREE, 1 << order);
764 migratetype = get_pageblock_migratetype(page);
765 set_freepage_migratetype(page, migratetype);
766 free_one_page(page_zone(page), page, order, migratetype);
767 local_irq_restore(flags);
770 void __init __free_pages_bootmem(struct page *page, unsigned int order)
772 unsigned int nr_pages = 1 << order;
773 struct page *p = page;
777 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
779 __ClearPageReserved(p);
780 set_page_count(p, 0);
782 __ClearPageReserved(p);
783 set_page_count(p, 0);
785 page_zone(page)->managed_pages += nr_pages;
786 set_page_refcounted(page);
787 __free_pages(page, order);
791 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
792 void __init init_cma_reserved_pageblock(struct page *page)
794 unsigned i = pageblock_nr_pages;
795 struct page *p = page;
798 __ClearPageReserved(p);
799 set_page_count(p, 0);
802 set_page_refcounted(page);
803 set_pageblock_migratetype(page, MIGRATE_CMA);
804 __free_pages(page, pageblock_order);
805 adjust_managed_page_count(page, pageblock_nr_pages);
810 * The order of subdivision here is critical for the IO subsystem.
811 * Please do not alter this order without good reasons and regression
812 * testing. Specifically, as large blocks of memory are subdivided,
813 * the order in which smaller blocks are delivered depends on the order
814 * they're subdivided in this function. This is the primary factor
815 * influencing the order in which pages are delivered to the IO
816 * subsystem according to empirical testing, and this is also justified
817 * by considering the behavior of a buddy system containing a single
818 * large block of memory acted on by a series of small allocations.
819 * This behavior is a critical factor in sglist merging's success.
823 static inline void expand(struct zone *zone, struct page *page,
824 int low, int high, struct free_area *area,
827 unsigned long size = 1 << high;
833 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
835 #ifdef CONFIG_DEBUG_PAGEALLOC
836 if (high < debug_guardpage_minorder()) {
838 * Mark as guard pages (or page), that will allow to
839 * merge back to allocator when buddy will be freed.
840 * Corresponding page table entries will not be touched,
841 * pages will stay not present in virtual address space
843 INIT_LIST_HEAD(&page[size].lru);
844 set_page_guard_flag(&page[size]);
845 set_page_private(&page[size], high);
846 /* Guard pages are not available for any usage */
847 __mod_zone_freepage_state(zone, -(1 << high),
852 list_add(&page[size].lru, &area->free_list[migratetype]);
854 set_page_order(&page[size], high);
859 * This page is about to be returned from the page allocator
861 static inline int check_new_page(struct page *page)
863 const char *bad_reason = NULL;
864 unsigned long bad_flags = 0;
866 if (unlikely(page_mapcount(page)))
867 bad_reason = "nonzero mapcount";
868 if (unlikely(page->mapping != NULL))
869 bad_reason = "non-NULL mapping";
870 if (unlikely(atomic_read(&page->_count) != 0))
871 bad_reason = "nonzero _count";
872 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
873 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
874 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
876 if (unlikely(mem_cgroup_bad_page_check(page)))
877 bad_reason = "cgroup check failed";
878 if (unlikely(bad_reason)) {
879 bad_page(page, bad_reason, bad_flags);
885 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
889 for (i = 0; i < (1 << order); i++) {
890 struct page *p = page + i;
891 if (unlikely(check_new_page(p)))
895 set_page_private(page, 0);
896 set_page_refcounted(page);
898 arch_alloc_page(page, order);
899 kernel_map_pages(page, 1 << order, 1);
901 if (gfp_flags & __GFP_ZERO)
902 prep_zero_page(page, order, gfp_flags);
904 if (order && (gfp_flags & __GFP_COMP))
905 prep_compound_page(page, order);
911 * Go through the free lists for the given migratetype and remove
912 * the smallest available page from the freelists
915 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
918 unsigned int current_order;
919 struct free_area *area;
922 /* Find a page of the appropriate size in the preferred list */
923 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
924 area = &(zone->free_area[current_order]);
925 if (list_empty(&area->free_list[migratetype]))
928 page = list_entry(area->free_list[migratetype].next,
930 list_del(&page->lru);
931 rmv_page_order(page);
933 expand(zone, page, order, current_order, area, migratetype);
942 * This array describes the order lists are fallen back to when
943 * the free lists for the desirable migrate type are depleted
945 static int fallbacks[MIGRATE_TYPES][4] = {
946 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
947 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
949 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
950 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
952 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
954 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
955 #ifdef CONFIG_MEMORY_ISOLATION
956 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
961 * Move the free pages in a range to the free lists of the requested type.
962 * Note that start_page and end_pages are not aligned on a pageblock
963 * boundary. If alignment is required, use move_freepages_block()
965 int move_freepages(struct zone *zone,
966 struct page *start_page, struct page *end_page,
973 #ifndef CONFIG_HOLES_IN_ZONE
975 * page_zone is not safe to call in this context when
976 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
977 * anyway as we check zone boundaries in move_freepages_block().
978 * Remove at a later date when no bug reports exist related to
979 * grouping pages by mobility
981 BUG_ON(page_zone(start_page) != page_zone(end_page));
984 for (page = start_page; page <= end_page;) {
985 /* Make sure we are not inadvertently changing nodes */
986 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
988 if (!pfn_valid_within(page_to_pfn(page))) {
993 if (!PageBuddy(page)) {
998 order = page_order(page);
999 list_move(&page->lru,
1000 &zone->free_area[order].free_list[migratetype]);
1001 set_freepage_migratetype(page, migratetype);
1003 pages_moved += 1 << order;
1009 int move_freepages_block(struct zone *zone, struct page *page,
1012 unsigned long start_pfn, end_pfn;
1013 struct page *start_page, *end_page;
1015 start_pfn = page_to_pfn(page);
1016 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1017 start_page = pfn_to_page(start_pfn);
1018 end_page = start_page + pageblock_nr_pages - 1;
1019 end_pfn = start_pfn + pageblock_nr_pages - 1;
1021 /* Do not cross zone boundaries */
1022 if (!zone_spans_pfn(zone, start_pfn))
1024 if (!zone_spans_pfn(zone, end_pfn))
1027 return move_freepages(zone, start_page, end_page, migratetype);
1030 static void change_pageblock_range(struct page *pageblock_page,
1031 int start_order, int migratetype)
1033 int nr_pageblocks = 1 << (start_order - pageblock_order);
1035 while (nr_pageblocks--) {
1036 set_pageblock_migratetype(pageblock_page, migratetype);
1037 pageblock_page += pageblock_nr_pages;
1042 * If breaking a large block of pages, move all free pages to the preferred
1043 * allocation list. If falling back for a reclaimable kernel allocation, be
1044 * more aggressive about taking ownership of free pages.
1046 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1047 * nor move CMA pages to different free lists. We don't want unmovable pages
1048 * to be allocated from MIGRATE_CMA areas.
1050 * Returns the new migratetype of the pageblock (or the same old migratetype
1051 * if it was unchanged).
1053 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1054 int start_type, int fallback_type)
1056 int current_order = page_order(page);
1059 * When borrowing from MIGRATE_CMA, we need to release the excess
1060 * buddy pages to CMA itself.
1062 if (is_migrate_cma(fallback_type))
1063 return fallback_type;
1065 /* Take ownership for orders >= pageblock_order */
1066 if (current_order >= pageblock_order) {
1067 change_pageblock_range(page, current_order, start_type);
1071 if (current_order >= pageblock_order / 2 ||
1072 start_type == MIGRATE_RECLAIMABLE ||
1073 page_group_by_mobility_disabled) {
1076 pages = move_freepages_block(zone, page, start_type);
1078 /* Claim the whole block if over half of it is free */
1079 if (pages >= (1 << (pageblock_order-1)) ||
1080 page_group_by_mobility_disabled) {
1082 set_pageblock_migratetype(page, start_type);
1088 return fallback_type;
1091 /* Remove an element from the buddy allocator from the fallback list */
1092 static inline struct page *
1093 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1095 struct free_area *area;
1098 int migratetype, new_type, i;
1100 /* Find the largest possible block of pages in the other list */
1101 for (current_order = MAX_ORDER-1; current_order >= order;
1104 migratetype = fallbacks[start_migratetype][i];
1106 /* MIGRATE_RESERVE handled later if necessary */
1107 if (migratetype == MIGRATE_RESERVE)
1110 area = &(zone->free_area[current_order]);
1111 if (list_empty(&area->free_list[migratetype]))
1114 page = list_entry(area->free_list[migratetype].next,
1118 new_type = try_to_steal_freepages(zone, page,
1122 /* Remove the page from the freelists */
1123 list_del(&page->lru);
1124 rmv_page_order(page);
1126 expand(zone, page, order, current_order, area,
1129 trace_mm_page_alloc_extfrag(page, order, current_order,
1130 start_migratetype, migratetype, new_type);
1140 * Do the hard work of removing an element from the buddy allocator.
1141 * Call me with the zone->lock already held.
1143 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1149 page = __rmqueue_smallest(zone, order, migratetype);
1151 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1152 page = __rmqueue_fallback(zone, order, migratetype);
1155 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1156 * is used because __rmqueue_smallest is an inline function
1157 * and we want just one call site
1160 migratetype = MIGRATE_RESERVE;
1165 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1170 * Obtain a specified number of elements from the buddy allocator, all under
1171 * a single hold of the lock, for efficiency. Add them to the supplied list.
1172 * Returns the number of new pages which were placed at *list.
1174 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1175 unsigned long count, struct list_head *list,
1176 int migratetype, int cold)
1178 int mt = migratetype, i;
1180 spin_lock(&zone->lock);
1181 for (i = 0; i < count; ++i) {
1182 struct page *page = __rmqueue(zone, order, migratetype);
1183 if (unlikely(page == NULL))
1187 * Split buddy pages returned by expand() are received here
1188 * in physical page order. The page is added to the callers and
1189 * list and the list head then moves forward. From the callers
1190 * perspective, the linked list is ordered by page number in
1191 * some conditions. This is useful for IO devices that can
1192 * merge IO requests if the physical pages are ordered
1195 if (likely(cold == 0))
1196 list_add(&page->lru, list);
1198 list_add_tail(&page->lru, list);
1199 if (IS_ENABLED(CONFIG_CMA)) {
1200 mt = get_pageblock_migratetype(page);
1201 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1204 set_freepage_migratetype(page, mt);
1206 if (is_migrate_cma(mt))
1207 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1210 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1211 spin_unlock(&zone->lock);
1217 * Called from the vmstat counter updater to drain pagesets of this
1218 * currently executing processor on remote nodes after they have
1221 * Note that this function must be called with the thread pinned to
1222 * a single processor.
1224 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1226 unsigned long flags;
1228 unsigned long batch;
1230 local_irq_save(flags);
1231 batch = ACCESS_ONCE(pcp->batch);
1232 if (pcp->count >= batch)
1235 to_drain = pcp->count;
1237 free_pcppages_bulk(zone, to_drain, pcp);
1238 pcp->count -= to_drain;
1240 local_irq_restore(flags);
1242 static bool gfp_thisnode_allocation(gfp_t gfp_mask)
1244 return (gfp_mask & GFP_THISNODE) == GFP_THISNODE;
1247 static bool gfp_thisnode_allocation(gfp_t gfp_mask)
1254 * Drain pages of the indicated processor.
1256 * The processor must either be the current processor and the
1257 * thread pinned to the current processor or a processor that
1260 static void drain_pages(unsigned int cpu)
1262 unsigned long flags;
1265 for_each_populated_zone(zone) {
1266 struct per_cpu_pageset *pset;
1267 struct per_cpu_pages *pcp;
1269 local_irq_save(flags);
1270 pset = per_cpu_ptr(zone->pageset, cpu);
1274 free_pcppages_bulk(zone, pcp->count, pcp);
1277 local_irq_restore(flags);
1282 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1284 void drain_local_pages(void *arg)
1286 drain_pages(smp_processor_id());
1290 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1292 * Note that this code is protected against sending an IPI to an offline
1293 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1294 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1295 * nothing keeps CPUs from showing up after we populated the cpumask and
1296 * before the call to on_each_cpu_mask().
1298 void drain_all_pages(void)
1301 struct per_cpu_pageset *pcp;
1305 * Allocate in the BSS so we wont require allocation in
1306 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1308 static cpumask_t cpus_with_pcps;
1311 * We don't care about racing with CPU hotplug event
1312 * as offline notification will cause the notified
1313 * cpu to drain that CPU pcps and on_each_cpu_mask
1314 * disables preemption as part of its processing
1316 for_each_online_cpu(cpu) {
1317 bool has_pcps = false;
1318 for_each_populated_zone(zone) {
1319 pcp = per_cpu_ptr(zone->pageset, cpu);
1320 if (pcp->pcp.count) {
1326 cpumask_set_cpu(cpu, &cpus_with_pcps);
1328 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1330 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1333 #ifdef CONFIG_HIBERNATION
1335 void mark_free_pages(struct zone *zone)
1337 unsigned long pfn, max_zone_pfn;
1338 unsigned long flags;
1340 struct list_head *curr;
1342 if (zone_is_empty(zone))
1345 spin_lock_irqsave(&zone->lock, flags);
1347 max_zone_pfn = zone_end_pfn(zone);
1348 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1349 if (pfn_valid(pfn)) {
1350 struct page *page = pfn_to_page(pfn);
1352 if (!swsusp_page_is_forbidden(page))
1353 swsusp_unset_page_free(page);
1356 for_each_migratetype_order(order, t) {
1357 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1360 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1361 for (i = 0; i < (1UL << order); i++)
1362 swsusp_set_page_free(pfn_to_page(pfn + i));
1365 spin_unlock_irqrestore(&zone->lock, flags);
1367 #endif /* CONFIG_PM */
1370 * Free a 0-order page
1371 * cold == 1 ? free a cold page : free a hot page
1373 void free_hot_cold_page(struct page *page, int cold)
1375 struct zone *zone = page_zone(page);
1376 struct per_cpu_pages *pcp;
1377 unsigned long flags;
1380 if (!free_pages_prepare(page, 0))
1383 migratetype = get_pageblock_migratetype(page);
1384 set_freepage_migratetype(page, migratetype);
1385 local_irq_save(flags);
1386 __count_vm_event(PGFREE);
1389 * We only track unmovable, reclaimable and movable on pcp lists.
1390 * Free ISOLATE pages back to the allocator because they are being
1391 * offlined but treat RESERVE as movable pages so we can get those
1392 * areas back if necessary. Otherwise, we may have to free
1393 * excessively into the page allocator
1395 if (migratetype >= MIGRATE_PCPTYPES) {
1396 if (unlikely(is_migrate_isolate(migratetype))) {
1397 free_one_page(zone, page, 0, migratetype);
1400 migratetype = MIGRATE_MOVABLE;
1403 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1405 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1407 list_add(&page->lru, &pcp->lists[migratetype]);
1409 if (pcp->count >= pcp->high) {
1410 unsigned long batch = ACCESS_ONCE(pcp->batch);
1411 free_pcppages_bulk(zone, batch, pcp);
1412 pcp->count -= batch;
1416 local_irq_restore(flags);
1420 * Free a list of 0-order pages
1422 void free_hot_cold_page_list(struct list_head *list, int cold)
1424 struct page *page, *next;
1426 list_for_each_entry_safe(page, next, list, lru) {
1427 trace_mm_page_free_batched(page, cold);
1428 free_hot_cold_page(page, cold);
1433 * split_page takes a non-compound higher-order page, and splits it into
1434 * n (1<<order) sub-pages: page[0..n]
1435 * Each sub-page must be freed individually.
1437 * Note: this is probably too low level an operation for use in drivers.
1438 * Please consult with lkml before using this in your driver.
1440 void split_page(struct page *page, unsigned int order)
1444 VM_BUG_ON_PAGE(PageCompound(page), page);
1445 VM_BUG_ON_PAGE(!page_count(page), page);
1447 #ifdef CONFIG_KMEMCHECK
1449 * Split shadow pages too, because free(page[0]) would
1450 * otherwise free the whole shadow.
1452 if (kmemcheck_page_is_tracked(page))
1453 split_page(virt_to_page(page[0].shadow), order);
1456 for (i = 1; i < (1 << order); i++)
1457 set_page_refcounted(page + i);
1459 EXPORT_SYMBOL_GPL(split_page);
1461 static int __isolate_free_page(struct page *page, unsigned int order)
1463 unsigned long watermark;
1467 BUG_ON(!PageBuddy(page));
1469 zone = page_zone(page);
1470 mt = get_pageblock_migratetype(page);
1472 if (!is_migrate_isolate(mt)) {
1473 /* Obey watermarks as if the page was being allocated */
1474 watermark = low_wmark_pages(zone) + (1 << order);
1475 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1478 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1481 /* Remove page from free list */
1482 list_del(&page->lru);
1483 zone->free_area[order].nr_free--;
1484 rmv_page_order(page);
1486 /* Set the pageblock if the isolated page is at least a pageblock */
1487 if (order >= pageblock_order - 1) {
1488 struct page *endpage = page + (1 << order) - 1;
1489 for (; page < endpage; page += pageblock_nr_pages) {
1490 int mt = get_pageblock_migratetype(page);
1491 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1492 set_pageblock_migratetype(page,
1497 return 1UL << order;
1501 * Similar to split_page except the page is already free. As this is only
1502 * being used for migration, the migratetype of the block also changes.
1503 * As this is called with interrupts disabled, the caller is responsible
1504 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1507 * Note: this is probably too low level an operation for use in drivers.
1508 * Please consult with lkml before using this in your driver.
1510 int split_free_page(struct page *page)
1515 order = page_order(page);
1517 nr_pages = __isolate_free_page(page, order);
1521 /* Split into individual pages */
1522 set_page_refcounted(page);
1523 split_page(page, order);
1528 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1529 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1533 struct page *buffered_rmqueue(struct zone *preferred_zone,
1534 struct zone *zone, int order, gfp_t gfp_flags,
1537 unsigned long flags;
1539 int cold = !!(gfp_flags & __GFP_COLD);
1542 if (likely(order == 0)) {
1543 struct per_cpu_pages *pcp;
1544 struct list_head *list;
1546 local_irq_save(flags);
1547 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1548 list = &pcp->lists[migratetype];
1549 if (list_empty(list)) {
1550 pcp->count += rmqueue_bulk(zone, 0,
1553 if (unlikely(list_empty(list)))
1558 page = list_entry(list->prev, struct page, lru);
1560 page = list_entry(list->next, struct page, lru);
1562 list_del(&page->lru);
1565 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1567 * __GFP_NOFAIL is not to be used in new code.
1569 * All __GFP_NOFAIL callers should be fixed so that they
1570 * properly detect and handle allocation failures.
1572 * We most definitely don't want callers attempting to
1573 * allocate greater than order-1 page units with
1576 WARN_ON_ONCE(order > 1);
1578 spin_lock_irqsave(&zone->lock, flags);
1579 page = __rmqueue(zone, order, migratetype);
1580 spin_unlock(&zone->lock);
1583 __mod_zone_freepage_state(zone, -(1 << order),
1584 get_pageblock_migratetype(page));
1588 * NOTE: GFP_THISNODE allocations do not partake in the kswapd
1589 * aging protocol, so they can't be fair.
1591 if (!gfp_thisnode_allocation(gfp_flags))
1592 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1594 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1595 zone_statistics(preferred_zone, zone, gfp_flags);
1596 local_irq_restore(flags);
1598 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1599 if (prep_new_page(page, order, gfp_flags))
1604 local_irq_restore(flags);
1608 #ifdef CONFIG_FAIL_PAGE_ALLOC
1611 struct fault_attr attr;
1613 u32 ignore_gfp_highmem;
1614 u32 ignore_gfp_wait;
1616 } fail_page_alloc = {
1617 .attr = FAULT_ATTR_INITIALIZER,
1618 .ignore_gfp_wait = 1,
1619 .ignore_gfp_highmem = 1,
1623 static int __init setup_fail_page_alloc(char *str)
1625 return setup_fault_attr(&fail_page_alloc.attr, str);
1627 __setup("fail_page_alloc=", setup_fail_page_alloc);
1629 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1631 if (order < fail_page_alloc.min_order)
1633 if (gfp_mask & __GFP_NOFAIL)
1635 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1637 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1640 return should_fail(&fail_page_alloc.attr, 1 << order);
1643 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1645 static int __init fail_page_alloc_debugfs(void)
1647 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1650 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1651 &fail_page_alloc.attr);
1653 return PTR_ERR(dir);
1655 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1656 &fail_page_alloc.ignore_gfp_wait))
1658 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1659 &fail_page_alloc.ignore_gfp_highmem))
1661 if (!debugfs_create_u32("min-order", mode, dir,
1662 &fail_page_alloc.min_order))
1667 debugfs_remove_recursive(dir);
1672 late_initcall(fail_page_alloc_debugfs);
1674 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1676 #else /* CONFIG_FAIL_PAGE_ALLOC */
1678 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1683 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1686 * Return true if free pages are above 'mark'. This takes into account the order
1687 * of the allocation.
1689 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1690 int classzone_idx, int alloc_flags, long free_pages)
1692 /* free_pages my go negative - that's OK */
1694 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1698 free_pages -= (1 << order) - 1;
1699 if (alloc_flags & ALLOC_HIGH)
1701 if (alloc_flags & ALLOC_HARDER)
1704 /* If allocation can't use CMA areas don't use free CMA pages */
1705 if (!(alloc_flags & ALLOC_CMA))
1706 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1709 if (free_pages - free_cma <= min + lowmem_reserve)
1711 for (o = 0; o < order; o++) {
1712 /* At the next order, this order's pages become unavailable */
1713 free_pages -= z->free_area[o].nr_free << o;
1715 /* Require fewer higher order pages to be free */
1718 if (free_pages <= min)
1724 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1725 int classzone_idx, int alloc_flags)
1727 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1728 zone_page_state(z, NR_FREE_PAGES));
1731 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1732 int classzone_idx, int alloc_flags)
1734 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1736 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1737 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1739 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1745 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1746 * skip over zones that are not allowed by the cpuset, or that have
1747 * been recently (in last second) found to be nearly full. See further
1748 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1749 * that have to skip over a lot of full or unallowed zones.
1751 * If the zonelist cache is present in the passed zonelist, then
1752 * returns a pointer to the allowed node mask (either the current
1753 * tasks mems_allowed, or node_states[N_MEMORY].)
1755 * If the zonelist cache is not available for this zonelist, does
1756 * nothing and returns NULL.
1758 * If the fullzones BITMAP in the zonelist cache is stale (more than
1759 * a second since last zap'd) then we zap it out (clear its bits.)
1761 * We hold off even calling zlc_setup, until after we've checked the
1762 * first zone in the zonelist, on the theory that most allocations will
1763 * be satisfied from that first zone, so best to examine that zone as
1764 * quickly as we can.
1766 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1768 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1769 nodemask_t *allowednodes; /* zonelist_cache approximation */
1771 zlc = zonelist->zlcache_ptr;
1775 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1776 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1777 zlc->last_full_zap = jiffies;
1780 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1781 &cpuset_current_mems_allowed :
1782 &node_states[N_MEMORY];
1783 return allowednodes;
1787 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1788 * if it is worth looking at further for free memory:
1789 * 1) Check that the zone isn't thought to be full (doesn't have its
1790 * bit set in the zonelist_cache fullzones BITMAP).
1791 * 2) Check that the zones node (obtained from the zonelist_cache
1792 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1793 * Return true (non-zero) if zone is worth looking at further, or
1794 * else return false (zero) if it is not.
1796 * This check -ignores- the distinction between various watermarks,
1797 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1798 * found to be full for any variation of these watermarks, it will
1799 * be considered full for up to one second by all requests, unless
1800 * we are so low on memory on all allowed nodes that we are forced
1801 * into the second scan of the zonelist.
1803 * In the second scan we ignore this zonelist cache and exactly
1804 * apply the watermarks to all zones, even it is slower to do so.
1805 * We are low on memory in the second scan, and should leave no stone
1806 * unturned looking for a free page.
1808 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1809 nodemask_t *allowednodes)
1811 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1812 int i; /* index of *z in zonelist zones */
1813 int n; /* node that zone *z is on */
1815 zlc = zonelist->zlcache_ptr;
1819 i = z - zonelist->_zonerefs;
1822 /* This zone is worth trying if it is allowed but not full */
1823 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1827 * Given 'z' scanning a zonelist, set the corresponding bit in
1828 * zlc->fullzones, so that subsequent attempts to allocate a page
1829 * from that zone don't waste time re-examining it.
1831 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1833 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1834 int i; /* index of *z in zonelist zones */
1836 zlc = zonelist->zlcache_ptr;
1840 i = z - zonelist->_zonerefs;
1842 set_bit(i, zlc->fullzones);
1846 * clear all zones full, called after direct reclaim makes progress so that
1847 * a zone that was recently full is not skipped over for up to a second
1849 static void zlc_clear_zones_full(struct zonelist *zonelist)
1851 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1853 zlc = zonelist->zlcache_ptr;
1857 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1860 static bool zone_local(struct zone *local_zone, struct zone *zone)
1862 return local_zone->node == zone->node;
1865 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1867 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1870 static void __paginginit init_zone_allows_reclaim(int nid)
1874 for_each_node_state(i, N_MEMORY)
1875 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1876 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1878 zone_reclaim_mode = 1;
1881 #else /* CONFIG_NUMA */
1883 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1888 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1889 nodemask_t *allowednodes)
1894 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1898 static void zlc_clear_zones_full(struct zonelist *zonelist)
1902 static bool zone_local(struct zone *local_zone, struct zone *zone)
1907 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1912 static inline void init_zone_allows_reclaim(int nid)
1915 #endif /* CONFIG_NUMA */
1918 * get_page_from_freelist goes through the zonelist trying to allocate
1921 static struct page *
1922 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1923 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1924 struct zone *preferred_zone, int migratetype)
1927 struct page *page = NULL;
1930 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1931 int zlc_active = 0; /* set if using zonelist_cache */
1932 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1934 classzone_idx = zone_idx(preferred_zone);
1937 * Scan zonelist, looking for a zone with enough free.
1938 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1940 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1941 high_zoneidx, nodemask) {
1944 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1945 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1947 if ((alloc_flags & ALLOC_CPUSET) &&
1948 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1950 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1951 if (unlikely(alloc_flags & ALLOC_NO_WATERMARKS))
1954 * Distribute pages in proportion to the individual
1955 * zone size to ensure fair page aging. The zone a
1956 * page was allocated in should have no effect on the
1957 * time the page has in memory before being reclaimed.
1959 * Try to stay in local zones in the fastpath. If
1960 * that fails, the slowpath is entered, which will do
1961 * another pass starting with the local zones, but
1962 * ultimately fall back to remote zones that do not
1963 * partake in the fairness round-robin cycle of this
1966 * NOTE: GFP_THISNODE allocations do not partake in
1967 * the kswapd aging protocol, so they can't be fair.
1969 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1970 !gfp_thisnode_allocation(gfp_mask)) {
1971 if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1973 if (!zone_local(preferred_zone, zone))
1977 * When allocating a page cache page for writing, we
1978 * want to get it from a zone that is within its dirty
1979 * limit, such that no single zone holds more than its
1980 * proportional share of globally allowed dirty pages.
1981 * The dirty limits take into account the zone's
1982 * lowmem reserves and high watermark so that kswapd
1983 * should be able to balance it without having to
1984 * write pages from its LRU list.
1986 * This may look like it could increase pressure on
1987 * lower zones by failing allocations in higher zones
1988 * before they are full. But the pages that do spill
1989 * over are limited as the lower zones are protected
1990 * by this very same mechanism. It should not become
1991 * a practical burden to them.
1993 * XXX: For now, allow allocations to potentially
1994 * exceed the per-zone dirty limit in the slowpath
1995 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1996 * which is important when on a NUMA setup the allowed
1997 * zones are together not big enough to reach the
1998 * global limit. The proper fix for these situations
1999 * will require awareness of zones in the
2000 * dirty-throttling and the flusher threads.
2002 if ((alloc_flags & ALLOC_WMARK_LOW) &&
2003 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
2004 goto this_zone_full;
2006 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2007 if (!zone_watermark_ok(zone, order, mark,
2008 classzone_idx, alloc_flags)) {
2011 if (IS_ENABLED(CONFIG_NUMA) &&
2012 !did_zlc_setup && nr_online_nodes > 1) {
2014 * we do zlc_setup if there are multiple nodes
2015 * and before considering the first zone allowed
2018 allowednodes = zlc_setup(zonelist, alloc_flags);
2023 if (zone_reclaim_mode == 0 ||
2024 !zone_allows_reclaim(preferred_zone, zone))
2025 goto this_zone_full;
2028 * As we may have just activated ZLC, check if the first
2029 * eligible zone has failed zone_reclaim recently.
2031 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2032 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2035 ret = zone_reclaim(zone, gfp_mask, order);
2037 case ZONE_RECLAIM_NOSCAN:
2040 case ZONE_RECLAIM_FULL:
2041 /* scanned but unreclaimable */
2044 /* did we reclaim enough */
2045 if (zone_watermark_ok(zone, order, mark,
2046 classzone_idx, alloc_flags))
2050 * Failed to reclaim enough to meet watermark.
2051 * Only mark the zone full if checking the min
2052 * watermark or if we failed to reclaim just
2053 * 1<<order pages or else the page allocator
2054 * fastpath will prematurely mark zones full
2055 * when the watermark is between the low and
2058 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2059 ret == ZONE_RECLAIM_SOME)
2060 goto this_zone_full;
2067 page = buffered_rmqueue(preferred_zone, zone, order,
2068 gfp_mask, migratetype);
2072 if (IS_ENABLED(CONFIG_NUMA))
2073 zlc_mark_zone_full(zonelist, z);
2076 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2077 /* Disable zlc cache for second zonelist scan */
2084 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2085 * necessary to allocate the page. The expectation is
2086 * that the caller is taking steps that will free more
2087 * memory. The caller should avoid the page being used
2088 * for !PFMEMALLOC purposes.
2090 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2096 * Large machines with many possible nodes should not always dump per-node
2097 * meminfo in irq context.
2099 static inline bool should_suppress_show_mem(void)
2104 ret = in_interrupt();
2109 static DEFINE_RATELIMIT_STATE(nopage_rs,
2110 DEFAULT_RATELIMIT_INTERVAL,
2111 DEFAULT_RATELIMIT_BURST);
2113 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2115 unsigned int filter = SHOW_MEM_FILTER_NODES;
2117 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2118 debug_guardpage_minorder() > 0)
2122 * This documents exceptions given to allocations in certain
2123 * contexts that are allowed to allocate outside current's set
2126 if (!(gfp_mask & __GFP_NOMEMALLOC))
2127 if (test_thread_flag(TIF_MEMDIE) ||
2128 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2129 filter &= ~SHOW_MEM_FILTER_NODES;
2130 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2131 filter &= ~SHOW_MEM_FILTER_NODES;
2134 struct va_format vaf;
2137 va_start(args, fmt);
2142 pr_warn("%pV", &vaf);
2147 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2148 current->comm, order, gfp_mask);
2151 if (!should_suppress_show_mem())
2156 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2157 unsigned long did_some_progress,
2158 unsigned long pages_reclaimed)
2160 /* Do not loop if specifically requested */
2161 if (gfp_mask & __GFP_NORETRY)
2164 /* Always retry if specifically requested */
2165 if (gfp_mask & __GFP_NOFAIL)
2169 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2170 * making forward progress without invoking OOM. Suspend also disables
2171 * storage devices so kswapd will not help. Bail if we are suspending.
2173 if (!did_some_progress && pm_suspended_storage())
2177 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2178 * means __GFP_NOFAIL, but that may not be true in other
2181 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2185 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2186 * specified, then we retry until we no longer reclaim any pages
2187 * (above), or we've reclaimed an order of pages at least as
2188 * large as the allocation's order. In both cases, if the
2189 * allocation still fails, we stop retrying.
2191 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2197 static inline struct page *
2198 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2199 struct zonelist *zonelist, enum zone_type high_zoneidx,
2200 nodemask_t *nodemask, struct zone *preferred_zone,
2205 /* Acquire the OOM killer lock for the zones in zonelist */
2206 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2207 schedule_timeout_uninterruptible(1);
2212 * Go through the zonelist yet one more time, keep very high watermark
2213 * here, this is only to catch a parallel oom killing, we must fail if
2214 * we're still under heavy pressure.
2216 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2217 order, zonelist, high_zoneidx,
2218 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2219 preferred_zone, migratetype);
2223 if (!(gfp_mask & __GFP_NOFAIL)) {
2224 /* The OOM killer will not help higher order allocs */
2225 if (order > PAGE_ALLOC_COSTLY_ORDER)
2227 /* The OOM killer does not needlessly kill tasks for lowmem */
2228 if (high_zoneidx < ZONE_NORMAL)
2231 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2232 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2233 * The caller should handle page allocation failure by itself if
2234 * it specifies __GFP_THISNODE.
2235 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2237 if (gfp_mask & __GFP_THISNODE)
2240 /* Exhausted what can be done so it's blamo time */
2241 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2244 clear_zonelist_oom(zonelist, gfp_mask);
2248 #ifdef CONFIG_COMPACTION
2249 /* Try memory compaction for high-order allocations before reclaim */
2250 static struct page *
2251 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2252 struct zonelist *zonelist, enum zone_type high_zoneidx,
2253 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2254 int migratetype, bool sync_migration,
2255 bool *contended_compaction, bool *deferred_compaction,
2256 unsigned long *did_some_progress)
2261 if (compaction_deferred(preferred_zone, order)) {
2262 *deferred_compaction = true;
2266 current->flags |= PF_MEMALLOC;
2267 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2268 nodemask, sync_migration,
2269 contended_compaction);
2270 current->flags &= ~PF_MEMALLOC;
2272 if (*did_some_progress != COMPACT_SKIPPED) {
2275 /* Page migration frees to the PCP lists but we want merging */
2276 drain_pages(get_cpu());
2279 page = get_page_from_freelist(gfp_mask, nodemask,
2280 order, zonelist, high_zoneidx,
2281 alloc_flags & ~ALLOC_NO_WATERMARKS,
2282 preferred_zone, migratetype);
2284 preferred_zone->compact_blockskip_flush = false;
2285 compaction_defer_reset(preferred_zone, order, true);
2286 count_vm_event(COMPACTSUCCESS);
2291 * It's bad if compaction run occurs and fails.
2292 * The most likely reason is that pages exist,
2293 * but not enough to satisfy watermarks.
2295 count_vm_event(COMPACTFAIL);
2298 * As async compaction considers a subset of pageblocks, only
2299 * defer if the failure was a sync compaction failure.
2302 defer_compaction(preferred_zone, order);
2310 static inline struct page *
2311 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2312 struct zonelist *zonelist, enum zone_type high_zoneidx,
2313 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2314 int migratetype, bool sync_migration,
2315 bool *contended_compaction, bool *deferred_compaction,
2316 unsigned long *did_some_progress)
2320 #endif /* CONFIG_COMPACTION */
2322 /* Perform direct synchronous page reclaim */
2324 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2325 nodemask_t *nodemask)
2327 struct reclaim_state reclaim_state;
2332 /* We now go into synchronous reclaim */
2333 cpuset_memory_pressure_bump();
2334 current->flags |= PF_MEMALLOC;
2335 lockdep_set_current_reclaim_state(gfp_mask);
2336 reclaim_state.reclaimed_slab = 0;
2337 current->reclaim_state = &reclaim_state;
2339 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2341 current->reclaim_state = NULL;
2342 lockdep_clear_current_reclaim_state();
2343 current->flags &= ~PF_MEMALLOC;
2350 /* The really slow allocator path where we enter direct reclaim */
2351 static inline struct page *
2352 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2353 struct zonelist *zonelist, enum zone_type high_zoneidx,
2354 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2355 int migratetype, unsigned long *did_some_progress)
2357 struct page *page = NULL;
2358 bool drained = false;
2360 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2362 if (unlikely(!(*did_some_progress)))
2365 /* After successful reclaim, reconsider all zones for allocation */
2366 if (IS_ENABLED(CONFIG_NUMA))
2367 zlc_clear_zones_full(zonelist);
2370 page = get_page_from_freelist(gfp_mask, nodemask, order,
2371 zonelist, high_zoneidx,
2372 alloc_flags & ~ALLOC_NO_WATERMARKS,
2373 preferred_zone, migratetype);
2376 * If an allocation failed after direct reclaim, it could be because
2377 * pages are pinned on the per-cpu lists. Drain them and try again
2379 if (!page && !drained) {
2389 * This is called in the allocator slow-path if the allocation request is of
2390 * sufficient urgency to ignore watermarks and take other desperate measures
2392 static inline struct page *
2393 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2394 struct zonelist *zonelist, enum zone_type high_zoneidx,
2395 nodemask_t *nodemask, struct zone *preferred_zone,
2401 page = get_page_from_freelist(gfp_mask, nodemask, order,
2402 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2403 preferred_zone, migratetype);
2405 if (!page && gfp_mask & __GFP_NOFAIL)
2406 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2407 } while (!page && (gfp_mask & __GFP_NOFAIL));
2412 static void prepare_slowpath(gfp_t gfp_mask, unsigned int order,
2413 struct zonelist *zonelist,
2414 enum zone_type high_zoneidx,
2415 struct zone *preferred_zone)
2420 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2421 if (!(gfp_mask & __GFP_NO_KSWAPD))
2422 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2424 * Only reset the batches of zones that were actually
2425 * considered in the fast path, we don't want to
2426 * thrash fairness information for zones that are not
2427 * actually part of this zonelist's round-robin cycle.
2429 if (!zone_local(preferred_zone, zone))
2431 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2432 high_wmark_pages(zone) -
2433 low_wmark_pages(zone) -
2434 zone_page_state(zone, NR_ALLOC_BATCH));
2439 gfp_to_alloc_flags(gfp_t gfp_mask)
2441 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2442 const gfp_t wait = gfp_mask & __GFP_WAIT;
2444 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2445 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2448 * The caller may dip into page reserves a bit more if the caller
2449 * cannot run direct reclaim, or if the caller has realtime scheduling
2450 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2451 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2453 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2457 * Not worth trying to allocate harder for
2458 * __GFP_NOMEMALLOC even if it can't schedule.
2460 if (!(gfp_mask & __GFP_NOMEMALLOC))
2461 alloc_flags |= ALLOC_HARDER;
2463 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2464 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2466 alloc_flags &= ~ALLOC_CPUSET;
2467 } else if (unlikely(rt_task(current)) && !in_interrupt())
2468 alloc_flags |= ALLOC_HARDER;
2470 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2471 if (gfp_mask & __GFP_MEMALLOC)
2472 alloc_flags |= ALLOC_NO_WATERMARKS;
2473 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2474 alloc_flags |= ALLOC_NO_WATERMARKS;
2475 else if (!in_interrupt() &&
2476 ((current->flags & PF_MEMALLOC) ||
2477 unlikely(test_thread_flag(TIF_MEMDIE))))
2478 alloc_flags |= ALLOC_NO_WATERMARKS;
2481 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2482 alloc_flags |= ALLOC_CMA;
2487 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2489 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2492 static inline struct page *
2493 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2494 struct zonelist *zonelist, enum zone_type high_zoneidx,
2495 nodemask_t *nodemask, struct zone *preferred_zone,
2498 const gfp_t wait = gfp_mask & __GFP_WAIT;
2499 struct page *page = NULL;
2501 unsigned long pages_reclaimed = 0;
2502 unsigned long did_some_progress;
2503 bool sync_migration = false;
2504 bool deferred_compaction = false;
2505 bool contended_compaction = false;
2508 * In the slowpath, we sanity check order to avoid ever trying to
2509 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2510 * be using allocators in order of preference for an area that is
2513 if (order >= MAX_ORDER) {
2514 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2519 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2520 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2521 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2522 * using a larger set of nodes after it has established that the
2523 * allowed per node queues are empty and that nodes are
2526 if (gfp_thisnode_allocation(gfp_mask))
2530 prepare_slowpath(gfp_mask, order, zonelist,
2531 high_zoneidx, preferred_zone);
2534 * OK, we're below the kswapd watermark and have kicked background
2535 * reclaim. Now things get more complex, so set up alloc_flags according
2536 * to how we want to proceed.
2538 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2541 * Find the true preferred zone if the allocation is unconstrained by
2544 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2545 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2549 /* This is the last chance, in general, before the goto nopage. */
2550 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2551 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2552 preferred_zone, migratetype);
2556 /* Allocate without watermarks if the context allows */
2557 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2559 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2560 * the allocation is high priority and these type of
2561 * allocations are system rather than user orientated
2563 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2565 page = __alloc_pages_high_priority(gfp_mask, order,
2566 zonelist, high_zoneidx, nodemask,
2567 preferred_zone, migratetype);
2573 /* Atomic allocations - we can't balance anything */
2576 * All existing users of the deprecated __GFP_NOFAIL are
2577 * blockable, so warn of any new users that actually allow this
2578 * type of allocation to fail.
2580 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2584 /* Avoid recursion of direct reclaim */
2585 if (current->flags & PF_MEMALLOC)
2588 /* Avoid allocations with no watermarks from looping endlessly */
2589 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2593 * Try direct compaction. The first pass is asynchronous. Subsequent
2594 * attempts after direct reclaim are synchronous
2596 page = __alloc_pages_direct_compact(gfp_mask, order,
2597 zonelist, high_zoneidx,
2599 alloc_flags, preferred_zone,
2600 migratetype, sync_migration,
2601 &contended_compaction,
2602 &deferred_compaction,
2603 &did_some_progress);
2606 sync_migration = true;
2609 * If compaction is deferred for high-order allocations, it is because
2610 * sync compaction recently failed. In this is the case and the caller
2611 * requested a movable allocation that does not heavily disrupt the
2612 * system then fail the allocation instead of entering direct reclaim.
2614 if ((deferred_compaction || contended_compaction) &&
2615 (gfp_mask & __GFP_NO_KSWAPD))
2618 /* Try direct reclaim and then allocating */
2619 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2620 zonelist, high_zoneidx,
2622 alloc_flags, preferred_zone,
2623 migratetype, &did_some_progress);
2628 * If we failed to make any progress reclaiming, then we are
2629 * running out of options and have to consider going OOM
2631 if (!did_some_progress) {
2632 if (oom_gfp_allowed(gfp_mask)) {
2633 if (oom_killer_disabled)
2635 /* Coredumps can quickly deplete all memory reserves */
2636 if ((current->flags & PF_DUMPCORE) &&
2637 !(gfp_mask & __GFP_NOFAIL))
2639 page = __alloc_pages_may_oom(gfp_mask, order,
2640 zonelist, high_zoneidx,
2641 nodemask, preferred_zone,
2646 if (!(gfp_mask & __GFP_NOFAIL)) {
2648 * The oom killer is not called for high-order
2649 * allocations that may fail, so if no progress
2650 * is being made, there are no other options and
2651 * retrying is unlikely to help.
2653 if (order > PAGE_ALLOC_COSTLY_ORDER)
2656 * The oom killer is not called for lowmem
2657 * allocations to prevent needlessly killing
2660 if (high_zoneidx < ZONE_NORMAL)
2668 /* Check if we should retry the allocation */
2669 pages_reclaimed += did_some_progress;
2670 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2672 /* Wait for some write requests to complete then retry */
2673 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2677 * High-order allocations do not necessarily loop after
2678 * direct reclaim and reclaim/compaction depends on compaction
2679 * being called after reclaim so call directly if necessary
2681 page = __alloc_pages_direct_compact(gfp_mask, order,
2682 zonelist, high_zoneidx,
2684 alloc_flags, preferred_zone,
2685 migratetype, sync_migration,
2686 &contended_compaction,
2687 &deferred_compaction,
2688 &did_some_progress);
2694 warn_alloc_failed(gfp_mask, order, NULL);
2697 if (kmemcheck_enabled)
2698 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2704 * This is the 'heart' of the zoned buddy allocator.
2707 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2708 struct zonelist *zonelist, nodemask_t *nodemask)
2710 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2711 struct zone *preferred_zone;
2712 struct page *page = NULL;
2713 int migratetype = allocflags_to_migratetype(gfp_mask);
2714 unsigned int cpuset_mems_cookie;
2715 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2716 struct mem_cgroup *memcg = NULL;
2718 gfp_mask &= gfp_allowed_mask;
2720 lockdep_trace_alloc(gfp_mask);
2722 might_sleep_if(gfp_mask & __GFP_WAIT);
2724 if (should_fail_alloc_page(gfp_mask, order))
2728 * Check the zones suitable for the gfp_mask contain at least one
2729 * valid zone. It's possible to have an empty zonelist as a result
2730 * of GFP_THISNODE and a memoryless node
2732 if (unlikely(!zonelist->_zonerefs->zone))
2736 * Will only have any effect when __GFP_KMEMCG is set. This is
2737 * verified in the (always inline) callee
2739 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2743 cpuset_mems_cookie = read_mems_allowed_begin();
2745 /* The preferred zone is used for statistics later */
2746 first_zones_zonelist(zonelist, high_zoneidx,
2747 nodemask ? : &cpuset_current_mems_allowed,
2749 if (!preferred_zone)
2753 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2754 alloc_flags |= ALLOC_CMA;
2756 /* First allocation attempt */
2757 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2758 zonelist, high_zoneidx, alloc_flags,
2759 preferred_zone, migratetype);
2760 if (unlikely(!page)) {
2762 * Runtime PM, block IO and its error handling path
2763 * can deadlock because I/O on the device might not
2766 gfp_mask = memalloc_noio_flags(gfp_mask);
2767 page = __alloc_pages_slowpath(gfp_mask, order,
2768 zonelist, high_zoneidx, nodemask,
2769 preferred_zone, migratetype);
2772 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2776 * When updating a task's mems_allowed, it is possible to race with
2777 * parallel threads in such a way that an allocation can fail while
2778 * the mask is being updated. If a page allocation is about to fail,
2779 * check if the cpuset changed during allocation and if so, retry.
2781 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2784 memcg_kmem_commit_charge(page, memcg, order);
2788 EXPORT_SYMBOL(__alloc_pages_nodemask);
2791 * Common helper functions.
2793 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2798 * __get_free_pages() returns a 32-bit address, which cannot represent
2801 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2803 page = alloc_pages(gfp_mask, order);
2806 return (unsigned long) page_address(page);
2808 EXPORT_SYMBOL(__get_free_pages);
2810 unsigned long get_zeroed_page(gfp_t gfp_mask)
2812 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2814 EXPORT_SYMBOL(get_zeroed_page);
2816 void __free_pages(struct page *page, unsigned int order)
2818 if (put_page_testzero(page)) {
2820 free_hot_cold_page(page, 0);
2822 __free_pages_ok(page, order);
2826 EXPORT_SYMBOL(__free_pages);
2828 void free_pages(unsigned long addr, unsigned int order)
2831 VM_BUG_ON(!virt_addr_valid((void *)addr));
2832 __free_pages(virt_to_page((void *)addr), order);
2836 EXPORT_SYMBOL(free_pages);
2839 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2840 * pages allocated with __GFP_KMEMCG.
2842 * Those pages are accounted to a particular memcg, embedded in the
2843 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2844 * for that information only to find out that it is NULL for users who have no
2845 * interest in that whatsoever, we provide these functions.
2847 * The caller knows better which flags it relies on.
2849 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2851 memcg_kmem_uncharge_pages(page, order);
2852 __free_pages(page, order);
2855 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2858 VM_BUG_ON(!virt_addr_valid((void *)addr));
2859 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2863 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2866 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2867 unsigned long used = addr + PAGE_ALIGN(size);
2869 split_page(virt_to_page((void *)addr), order);
2870 while (used < alloc_end) {
2875 return (void *)addr;
2879 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2880 * @size: the number of bytes to allocate
2881 * @gfp_mask: GFP flags for the allocation
2883 * This function is similar to alloc_pages(), except that it allocates the
2884 * minimum number of pages to satisfy the request. alloc_pages() can only
2885 * allocate memory in power-of-two pages.
2887 * This function is also limited by MAX_ORDER.
2889 * Memory allocated by this function must be released by free_pages_exact().
2891 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2893 unsigned int order = get_order(size);
2896 addr = __get_free_pages(gfp_mask, order);
2897 return make_alloc_exact(addr, order, size);
2899 EXPORT_SYMBOL(alloc_pages_exact);
2902 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2904 * @nid: the preferred node ID where memory should be allocated
2905 * @size: the number of bytes to allocate
2906 * @gfp_mask: GFP flags for the allocation
2908 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2910 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2913 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2915 unsigned order = get_order(size);
2916 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2919 return make_alloc_exact((unsigned long)page_address(p), order, size);
2921 EXPORT_SYMBOL(alloc_pages_exact_nid);
2924 * free_pages_exact - release memory allocated via alloc_pages_exact()
2925 * @virt: the value returned by alloc_pages_exact.
2926 * @size: size of allocation, same value as passed to alloc_pages_exact().
2928 * Release the memory allocated by a previous call to alloc_pages_exact.
2930 void free_pages_exact(void *virt, size_t size)
2932 unsigned long addr = (unsigned long)virt;
2933 unsigned long end = addr + PAGE_ALIGN(size);
2935 while (addr < end) {
2940 EXPORT_SYMBOL(free_pages_exact);
2943 * nr_free_zone_pages - count number of pages beyond high watermark
2944 * @offset: The zone index of the highest zone
2946 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2947 * high watermark within all zones at or below a given zone index. For each
2948 * zone, the number of pages is calculated as:
2949 * managed_pages - high_pages
2951 static unsigned long nr_free_zone_pages(int offset)
2956 /* Just pick one node, since fallback list is circular */
2957 unsigned long sum = 0;
2959 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2961 for_each_zone_zonelist(zone, z, zonelist, offset) {
2962 unsigned long size = zone->managed_pages;
2963 unsigned long high = high_wmark_pages(zone);
2972 * nr_free_buffer_pages - count number of pages beyond high watermark
2974 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2975 * watermark within ZONE_DMA and ZONE_NORMAL.
2977 unsigned long nr_free_buffer_pages(void)
2979 return nr_free_zone_pages(gfp_zone(GFP_USER));
2981 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2984 * nr_free_pagecache_pages - count number of pages beyond high watermark
2986 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2987 * high watermark within all zones.
2989 unsigned long nr_free_pagecache_pages(void)
2991 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2994 static inline void show_node(struct zone *zone)
2996 if (IS_ENABLED(CONFIG_NUMA))
2997 printk("Node %d ", zone_to_nid(zone));
3000 void si_meminfo(struct sysinfo *val)
3002 val->totalram = totalram_pages;
3004 val->freeram = global_page_state(NR_FREE_PAGES);
3005 val->bufferram = nr_blockdev_pages();
3006 val->totalhigh = totalhigh_pages;
3007 val->freehigh = nr_free_highpages();
3008 val->mem_unit = PAGE_SIZE;
3011 EXPORT_SYMBOL(si_meminfo);
3014 void si_meminfo_node(struct sysinfo *val, int nid)
3016 int zone_type; /* needs to be signed */
3017 unsigned long managed_pages = 0;
3018 pg_data_t *pgdat = NODE_DATA(nid);
3020 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3021 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3022 val->totalram = managed_pages;
3023 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3024 #ifdef CONFIG_HIGHMEM
3025 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3026 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3032 val->mem_unit = PAGE_SIZE;
3037 * Determine whether the node should be displayed or not, depending on whether
3038 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3040 bool skip_free_areas_node(unsigned int flags, int nid)
3043 unsigned int cpuset_mems_cookie;
3045 if (!(flags & SHOW_MEM_FILTER_NODES))
3049 cpuset_mems_cookie = read_mems_allowed_begin();
3050 ret = !node_isset(nid, cpuset_current_mems_allowed);
3051 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3056 #define K(x) ((x) << (PAGE_SHIFT-10))
3058 static void show_migration_types(unsigned char type)
3060 static const char types[MIGRATE_TYPES] = {
3061 [MIGRATE_UNMOVABLE] = 'U',
3062 [MIGRATE_RECLAIMABLE] = 'E',
3063 [MIGRATE_MOVABLE] = 'M',
3064 [MIGRATE_RESERVE] = 'R',
3066 [MIGRATE_CMA] = 'C',
3068 #ifdef CONFIG_MEMORY_ISOLATION
3069 [MIGRATE_ISOLATE] = 'I',
3072 char tmp[MIGRATE_TYPES + 1];
3076 for (i = 0; i < MIGRATE_TYPES; i++) {
3077 if (type & (1 << i))
3082 printk("(%s) ", tmp);
3086 * Show free area list (used inside shift_scroll-lock stuff)
3087 * We also calculate the percentage fragmentation. We do this by counting the
3088 * memory on each free list with the exception of the first item on the list.
3089 * Suppresses nodes that are not allowed by current's cpuset if
3090 * SHOW_MEM_FILTER_NODES is passed.
3092 void show_free_areas(unsigned int filter)
3097 for_each_populated_zone(zone) {
3098 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3101 printk("%s per-cpu:\n", zone->name);
3103 for_each_online_cpu(cpu) {
3104 struct per_cpu_pageset *pageset;
3106 pageset = per_cpu_ptr(zone->pageset, cpu);
3108 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3109 cpu, pageset->pcp.high,
3110 pageset->pcp.batch, pageset->pcp.count);
3114 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3115 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3117 " dirty:%lu writeback:%lu unstable:%lu\n"
3118 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3119 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3121 global_page_state(NR_ACTIVE_ANON),
3122 global_page_state(NR_INACTIVE_ANON),
3123 global_page_state(NR_ISOLATED_ANON),
3124 global_page_state(NR_ACTIVE_FILE),
3125 global_page_state(NR_INACTIVE_FILE),
3126 global_page_state(NR_ISOLATED_FILE),
3127 global_page_state(NR_UNEVICTABLE),
3128 global_page_state(NR_FILE_DIRTY),
3129 global_page_state(NR_WRITEBACK),
3130 global_page_state(NR_UNSTABLE_NFS),
3131 global_page_state(NR_FREE_PAGES),
3132 global_page_state(NR_SLAB_RECLAIMABLE),
3133 global_page_state(NR_SLAB_UNRECLAIMABLE),
3134 global_page_state(NR_FILE_MAPPED),
3135 global_page_state(NR_SHMEM),
3136 global_page_state(NR_PAGETABLE),
3137 global_page_state(NR_BOUNCE),
3138 global_page_state(NR_FREE_CMA_PAGES));
3140 for_each_populated_zone(zone) {
3143 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3151 " active_anon:%lukB"
3152 " inactive_anon:%lukB"
3153 " active_file:%lukB"
3154 " inactive_file:%lukB"
3155 " unevictable:%lukB"
3156 " isolated(anon):%lukB"
3157 " isolated(file):%lukB"
3165 " slab_reclaimable:%lukB"
3166 " slab_unreclaimable:%lukB"
3167 " kernel_stack:%lukB"
3172 " writeback_tmp:%lukB"
3173 " pages_scanned:%lu"
3174 " all_unreclaimable? %s"
3177 K(zone_page_state(zone, NR_FREE_PAGES)),
3178 K(min_wmark_pages(zone)),
3179 K(low_wmark_pages(zone)),
3180 K(high_wmark_pages(zone)),
3181 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3182 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3183 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3184 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3185 K(zone_page_state(zone, NR_UNEVICTABLE)),
3186 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3187 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3188 K(zone->present_pages),
3189 K(zone->managed_pages),
3190 K(zone_page_state(zone, NR_MLOCK)),
3191 K(zone_page_state(zone, NR_FILE_DIRTY)),
3192 K(zone_page_state(zone, NR_WRITEBACK)),
3193 K(zone_page_state(zone, NR_FILE_MAPPED)),
3194 K(zone_page_state(zone, NR_SHMEM)),
3195 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3196 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3197 zone_page_state(zone, NR_KERNEL_STACK) *
3199 K(zone_page_state(zone, NR_PAGETABLE)),
3200 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3201 K(zone_page_state(zone, NR_BOUNCE)),
3202 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3203 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3204 zone->pages_scanned,
3205 (!zone_reclaimable(zone) ? "yes" : "no")
3207 printk("lowmem_reserve[]:");
3208 for (i = 0; i < MAX_NR_ZONES; i++)
3209 printk(" %lu", zone->lowmem_reserve[i]);
3213 for_each_populated_zone(zone) {
3214 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3215 unsigned char types[MAX_ORDER];
3217 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3220 printk("%s: ", zone->name);
3222 spin_lock_irqsave(&zone->lock, flags);
3223 for (order = 0; order < MAX_ORDER; order++) {
3224 struct free_area *area = &zone->free_area[order];
3227 nr[order] = area->nr_free;
3228 total += nr[order] << order;
3231 for (type = 0; type < MIGRATE_TYPES; type++) {
3232 if (!list_empty(&area->free_list[type]))
3233 types[order] |= 1 << type;
3236 spin_unlock_irqrestore(&zone->lock, flags);
3237 for (order = 0; order < MAX_ORDER; order++) {
3238 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3240 show_migration_types(types[order]);
3242 printk("= %lukB\n", K(total));
3245 hugetlb_show_meminfo();
3247 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3249 show_swap_cache_info();
3252 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3254 zoneref->zone = zone;
3255 zoneref->zone_idx = zone_idx(zone);
3259 * Builds allocation fallback zone lists.
3261 * Add all populated zones of a node to the zonelist.
3263 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3267 enum zone_type zone_type = MAX_NR_ZONES;
3271 zone = pgdat->node_zones + zone_type;
3272 if (populated_zone(zone)) {
3273 zoneref_set_zone(zone,
3274 &zonelist->_zonerefs[nr_zones++]);
3275 check_highest_zone(zone_type);
3277 } while (zone_type);
3285 * 0 = automatic detection of better ordering.
3286 * 1 = order by ([node] distance, -zonetype)
3287 * 2 = order by (-zonetype, [node] distance)
3289 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3290 * the same zonelist. So only NUMA can configure this param.
3292 #define ZONELIST_ORDER_DEFAULT 0
3293 #define ZONELIST_ORDER_NODE 1
3294 #define ZONELIST_ORDER_ZONE 2
3296 /* zonelist order in the kernel.
3297 * set_zonelist_order() will set this to NODE or ZONE.
3299 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3300 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3304 /* The value user specified ....changed by config */
3305 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3306 /* string for sysctl */
3307 #define NUMA_ZONELIST_ORDER_LEN 16
3308 char numa_zonelist_order[16] = "default";
3311 * interface for configure zonelist ordering.
3312 * command line option "numa_zonelist_order"
3313 * = "[dD]efault - default, automatic configuration.
3314 * = "[nN]ode - order by node locality, then by zone within node
3315 * = "[zZ]one - order by zone, then by locality within zone
3318 static int __parse_numa_zonelist_order(char *s)
3320 if (*s == 'd' || *s == 'D') {
3321 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3322 } else if (*s == 'n' || *s == 'N') {
3323 user_zonelist_order = ZONELIST_ORDER_NODE;
3324 } else if (*s == 'z' || *s == 'Z') {
3325 user_zonelist_order = ZONELIST_ORDER_ZONE;
3328 "Ignoring invalid numa_zonelist_order value: "
3335 static __init int setup_numa_zonelist_order(char *s)
3342 ret = __parse_numa_zonelist_order(s);
3344 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3348 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3351 * sysctl handler for numa_zonelist_order
3353 int numa_zonelist_order_handler(ctl_table *table, int write,
3354 void __user *buffer, size_t *length,
3357 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3359 static DEFINE_MUTEX(zl_order_mutex);
3361 mutex_lock(&zl_order_mutex);
3363 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3367 strcpy(saved_string, (char *)table->data);
3369 ret = proc_dostring(table, write, buffer, length, ppos);
3373 int oldval = user_zonelist_order;
3375 ret = __parse_numa_zonelist_order((char *)table->data);
3378 * bogus value. restore saved string
3380 strncpy((char *)table->data, saved_string,
3381 NUMA_ZONELIST_ORDER_LEN);
3382 user_zonelist_order = oldval;
3383 } else if (oldval != user_zonelist_order) {
3384 mutex_lock(&zonelists_mutex);
3385 build_all_zonelists(NULL, NULL);
3386 mutex_unlock(&zonelists_mutex);
3390 mutex_unlock(&zl_order_mutex);
3395 #define MAX_NODE_LOAD (nr_online_nodes)
3396 static int node_load[MAX_NUMNODES];
3399 * find_next_best_node - find the next node that should appear in a given node's fallback list
3400 * @node: node whose fallback list we're appending
3401 * @used_node_mask: nodemask_t of already used nodes
3403 * We use a number of factors to determine which is the next node that should
3404 * appear on a given node's fallback list. The node should not have appeared
3405 * already in @node's fallback list, and it should be the next closest node
3406 * according to the distance array (which contains arbitrary distance values
3407 * from each node to each node in the system), and should also prefer nodes
3408 * with no CPUs, since presumably they'll have very little allocation pressure
3409 * on them otherwise.
3410 * It returns -1 if no node is found.
3412 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3415 int min_val = INT_MAX;
3416 int best_node = NUMA_NO_NODE;
3417 const struct cpumask *tmp = cpumask_of_node(0);
3419 /* Use the local node if we haven't already */
3420 if (!node_isset(node, *used_node_mask)) {
3421 node_set(node, *used_node_mask);
3425 for_each_node_state(n, N_MEMORY) {
3427 /* Don't want a node to appear more than once */
3428 if (node_isset(n, *used_node_mask))
3431 /* Use the distance array to find the distance */
3432 val = node_distance(node, n);
3434 /* Penalize nodes under us ("prefer the next node") */
3437 /* Give preference to headless and unused nodes */
3438 tmp = cpumask_of_node(n);
3439 if (!cpumask_empty(tmp))
3440 val += PENALTY_FOR_NODE_WITH_CPUS;
3442 /* Slight preference for less loaded node */
3443 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3444 val += node_load[n];
3446 if (val < min_val) {
3453 node_set(best_node, *used_node_mask);
3460 * Build zonelists ordered by node and zones within node.
3461 * This results in maximum locality--normal zone overflows into local
3462 * DMA zone, if any--but risks exhausting DMA zone.
3464 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3467 struct zonelist *zonelist;
3469 zonelist = &pgdat->node_zonelists[0];
3470 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3472 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3473 zonelist->_zonerefs[j].zone = NULL;
3474 zonelist->_zonerefs[j].zone_idx = 0;
3478 * Build gfp_thisnode zonelists
3480 static void build_thisnode_zonelists(pg_data_t *pgdat)
3483 struct zonelist *zonelist;
3485 zonelist = &pgdat->node_zonelists[1];
3486 j = build_zonelists_node(pgdat, zonelist, 0);
3487 zonelist->_zonerefs[j].zone = NULL;
3488 zonelist->_zonerefs[j].zone_idx = 0;
3492 * Build zonelists ordered by zone and nodes within zones.
3493 * This results in conserving DMA zone[s] until all Normal memory is
3494 * exhausted, but results in overflowing to remote node while memory
3495 * may still exist in local DMA zone.
3497 static int node_order[MAX_NUMNODES];
3499 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3502 int zone_type; /* needs to be signed */
3504 struct zonelist *zonelist;
3506 zonelist = &pgdat->node_zonelists[0];
3508 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3509 for (j = 0; j < nr_nodes; j++) {
3510 node = node_order[j];
3511 z = &NODE_DATA(node)->node_zones[zone_type];
3512 if (populated_zone(z)) {
3514 &zonelist->_zonerefs[pos++]);
3515 check_highest_zone(zone_type);
3519 zonelist->_zonerefs[pos].zone = NULL;
3520 zonelist->_zonerefs[pos].zone_idx = 0;
3523 static int default_zonelist_order(void)
3526 unsigned long low_kmem_size, total_size;
3530 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3531 * If they are really small and used heavily, the system can fall
3532 * into OOM very easily.
3533 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3535 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3538 for_each_online_node(nid) {
3539 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3540 z = &NODE_DATA(nid)->node_zones[zone_type];
3541 if (populated_zone(z)) {
3542 if (zone_type < ZONE_NORMAL)
3543 low_kmem_size += z->managed_pages;
3544 total_size += z->managed_pages;
3545 } else if (zone_type == ZONE_NORMAL) {
3547 * If any node has only lowmem, then node order
3548 * is preferred to allow kernel allocations
3549 * locally; otherwise, they can easily infringe
3550 * on other nodes when there is an abundance of
3551 * lowmem available to allocate from.
3553 return ZONELIST_ORDER_NODE;
3557 if (!low_kmem_size || /* there are no DMA area. */
3558 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3559 return ZONELIST_ORDER_NODE;
3561 * look into each node's config.
3562 * If there is a node whose DMA/DMA32 memory is very big area on
3563 * local memory, NODE_ORDER may be suitable.
3565 average_size = total_size /
3566 (nodes_weight(node_states[N_MEMORY]) + 1);
3567 for_each_online_node(nid) {
3570 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3571 z = &NODE_DATA(nid)->node_zones[zone_type];
3572 if (populated_zone(z)) {
3573 if (zone_type < ZONE_NORMAL)
3574 low_kmem_size += z->present_pages;
3575 total_size += z->present_pages;
3578 if (low_kmem_size &&
3579 total_size > average_size && /* ignore small node */
3580 low_kmem_size > total_size * 70/100)
3581 return ZONELIST_ORDER_NODE;
3583 return ZONELIST_ORDER_ZONE;
3586 static void set_zonelist_order(void)
3588 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3589 current_zonelist_order = default_zonelist_order();
3591 current_zonelist_order = user_zonelist_order;
3594 static void build_zonelists(pg_data_t *pgdat)
3598 nodemask_t used_mask;
3599 int local_node, prev_node;
3600 struct zonelist *zonelist;
3601 int order = current_zonelist_order;
3603 /* initialize zonelists */
3604 for (i = 0; i < MAX_ZONELISTS; i++) {
3605 zonelist = pgdat->node_zonelists + i;
3606 zonelist->_zonerefs[0].zone = NULL;
3607 zonelist->_zonerefs[0].zone_idx = 0;
3610 /* NUMA-aware ordering of nodes */
3611 local_node = pgdat->node_id;
3612 load = nr_online_nodes;
3613 prev_node = local_node;
3614 nodes_clear(used_mask);
3616 memset(node_order, 0, sizeof(node_order));
3619 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3621 * We don't want to pressure a particular node.
3622 * So adding penalty to the first node in same
3623 * distance group to make it round-robin.
3625 if (node_distance(local_node, node) !=
3626 node_distance(local_node, prev_node))
3627 node_load[node] = load;
3631 if (order == ZONELIST_ORDER_NODE)
3632 build_zonelists_in_node_order(pgdat, node);
3634 node_order[j++] = node; /* remember order */
3637 if (order == ZONELIST_ORDER_ZONE) {
3638 /* calculate node order -- i.e., DMA last! */
3639 build_zonelists_in_zone_order(pgdat, j);
3642 build_thisnode_zonelists(pgdat);
3645 /* Construct the zonelist performance cache - see further mmzone.h */
3646 static void build_zonelist_cache(pg_data_t *pgdat)
3648 struct zonelist *zonelist;
3649 struct zonelist_cache *zlc;
3652 zonelist = &pgdat->node_zonelists[0];
3653 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3654 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3655 for (z = zonelist->_zonerefs; z->zone; z++)
3656 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3659 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3661 * Return node id of node used for "local" allocations.
3662 * I.e., first node id of first zone in arg node's generic zonelist.
3663 * Used for initializing percpu 'numa_mem', which is used primarily
3664 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3666 int local_memory_node(int node)
3670 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3671 gfp_zone(GFP_KERNEL),
3678 #else /* CONFIG_NUMA */
3680 static void set_zonelist_order(void)
3682 current_zonelist_order = ZONELIST_ORDER_ZONE;
3685 static void build_zonelists(pg_data_t *pgdat)
3687 int node, local_node;
3689 struct zonelist *zonelist;
3691 local_node = pgdat->node_id;
3693 zonelist = &pgdat->node_zonelists[0];
3694 j = build_zonelists_node(pgdat, zonelist, 0);
3697 * Now we build the zonelist so that it contains the zones
3698 * of all the other nodes.
3699 * We don't want to pressure a particular node, so when
3700 * building the zones for node N, we make sure that the
3701 * zones coming right after the local ones are those from
3702 * node N+1 (modulo N)
3704 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3705 if (!node_online(node))
3707 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3709 for (node = 0; node < local_node; node++) {
3710 if (!node_online(node))
3712 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3715 zonelist->_zonerefs[j].zone = NULL;
3716 zonelist->_zonerefs[j].zone_idx = 0;
3719 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3720 static void build_zonelist_cache(pg_data_t *pgdat)
3722 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3725 #endif /* CONFIG_NUMA */
3728 * Boot pageset table. One per cpu which is going to be used for all
3729 * zones and all nodes. The parameters will be set in such a way
3730 * that an item put on a list will immediately be handed over to
3731 * the buddy list. This is safe since pageset manipulation is done
3732 * with interrupts disabled.
3734 * The boot_pagesets must be kept even after bootup is complete for
3735 * unused processors and/or zones. They do play a role for bootstrapping
3736 * hotplugged processors.
3738 * zoneinfo_show() and maybe other functions do
3739 * not check if the processor is online before following the pageset pointer.
3740 * Other parts of the kernel may not check if the zone is available.
3742 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3743 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3744 static void setup_zone_pageset(struct zone *zone);
3747 * Global mutex to protect against size modification of zonelists
3748 * as well as to serialize pageset setup for the new populated zone.
3750 DEFINE_MUTEX(zonelists_mutex);
3752 /* return values int ....just for stop_machine() */
3753 static int __build_all_zonelists(void *data)
3757 pg_data_t *self = data;
3760 memset(node_load, 0, sizeof(node_load));
3763 if (self && !node_online(self->node_id)) {
3764 build_zonelists(self);
3765 build_zonelist_cache(self);
3768 for_each_online_node(nid) {
3769 pg_data_t *pgdat = NODE_DATA(nid);
3771 build_zonelists(pgdat);
3772 build_zonelist_cache(pgdat);
3776 * Initialize the boot_pagesets that are going to be used
3777 * for bootstrapping processors. The real pagesets for
3778 * each zone will be allocated later when the per cpu
3779 * allocator is available.
3781 * boot_pagesets are used also for bootstrapping offline
3782 * cpus if the system is already booted because the pagesets
3783 * are needed to initialize allocators on a specific cpu too.
3784 * F.e. the percpu allocator needs the page allocator which
3785 * needs the percpu allocator in order to allocate its pagesets
3786 * (a chicken-egg dilemma).
3788 for_each_possible_cpu(cpu) {
3789 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3791 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3793 * We now know the "local memory node" for each node--
3794 * i.e., the node of the first zone in the generic zonelist.
3795 * Set up numa_mem percpu variable for on-line cpus. During
3796 * boot, only the boot cpu should be on-line; we'll init the
3797 * secondary cpus' numa_mem as they come on-line. During
3798 * node/memory hotplug, we'll fixup all on-line cpus.
3800 if (cpu_online(cpu))
3801 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3809 * Called with zonelists_mutex held always
3810 * unless system_state == SYSTEM_BOOTING.
3812 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3814 set_zonelist_order();
3816 if (system_state == SYSTEM_BOOTING) {
3817 __build_all_zonelists(NULL);
3818 mminit_verify_zonelist();
3819 cpuset_init_current_mems_allowed();
3821 #ifdef CONFIG_MEMORY_HOTPLUG
3823 setup_zone_pageset(zone);
3825 /* we have to stop all cpus to guarantee there is no user
3827 stop_machine(__build_all_zonelists, pgdat, NULL);
3828 /* cpuset refresh routine should be here */
3830 vm_total_pages = nr_free_pagecache_pages();
3832 * Disable grouping by mobility if the number of pages in the
3833 * system is too low to allow the mechanism to work. It would be
3834 * more accurate, but expensive to check per-zone. This check is
3835 * made on memory-hotadd so a system can start with mobility
3836 * disabled and enable it later
3838 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3839 page_group_by_mobility_disabled = 1;
3841 page_group_by_mobility_disabled = 0;
3843 printk("Built %i zonelists in %s order, mobility grouping %s. "
3844 "Total pages: %ld\n",
3846 zonelist_order_name[current_zonelist_order],
3847 page_group_by_mobility_disabled ? "off" : "on",
3850 printk("Policy zone: %s\n", zone_names[policy_zone]);
3855 * Helper functions to size the waitqueue hash table.
3856 * Essentially these want to choose hash table sizes sufficiently
3857 * large so that collisions trying to wait on pages are rare.
3858 * But in fact, the number of active page waitqueues on typical
3859 * systems is ridiculously low, less than 200. So this is even
3860 * conservative, even though it seems large.
3862 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3863 * waitqueues, i.e. the size of the waitq table given the number of pages.
3865 #define PAGES_PER_WAITQUEUE 256
3867 #ifndef CONFIG_MEMORY_HOTPLUG
3868 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3870 unsigned long size = 1;
3872 pages /= PAGES_PER_WAITQUEUE;
3874 while (size < pages)
3878 * Once we have dozens or even hundreds of threads sleeping
3879 * on IO we've got bigger problems than wait queue collision.
3880 * Limit the size of the wait table to a reasonable size.
3882 size = min(size, 4096UL);
3884 return max(size, 4UL);
3888 * A zone's size might be changed by hot-add, so it is not possible to determine
3889 * a suitable size for its wait_table. So we use the maximum size now.
3891 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3893 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3894 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3895 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3897 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3898 * or more by the traditional way. (See above). It equals:
3900 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3901 * ia64(16K page size) : = ( 8G + 4M)byte.
3902 * powerpc (64K page size) : = (32G +16M)byte.
3904 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3911 * This is an integer logarithm so that shifts can be used later
3912 * to extract the more random high bits from the multiplicative
3913 * hash function before the remainder is taken.
3915 static inline unsigned long wait_table_bits(unsigned long size)
3921 * Check if a pageblock contains reserved pages
3923 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3927 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3928 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3935 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3936 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3937 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3938 * higher will lead to a bigger reserve which will get freed as contiguous
3939 * blocks as reclaim kicks in
3941 static void setup_zone_migrate_reserve(struct zone *zone)
3943 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3945 unsigned long block_migratetype;
3950 * Get the start pfn, end pfn and the number of blocks to reserve
3951 * We have to be careful to be aligned to pageblock_nr_pages to
3952 * make sure that we always check pfn_valid for the first page in
3955 start_pfn = zone->zone_start_pfn;
3956 end_pfn = zone_end_pfn(zone);
3957 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3958 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3962 * Reserve blocks are generally in place to help high-order atomic
3963 * allocations that are short-lived. A min_free_kbytes value that
3964 * would result in more than 2 reserve blocks for atomic allocations
3965 * is assumed to be in place to help anti-fragmentation for the
3966 * future allocation of hugepages at runtime.
3968 reserve = min(2, reserve);
3969 old_reserve = zone->nr_migrate_reserve_block;
3971 /* When memory hot-add, we almost always need to do nothing */
3972 if (reserve == old_reserve)
3974 zone->nr_migrate_reserve_block = reserve;
3976 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3977 if (!pfn_valid(pfn))
3979 page = pfn_to_page(pfn);
3981 /* Watch out for overlapping nodes */
3982 if (page_to_nid(page) != zone_to_nid(zone))
3985 block_migratetype = get_pageblock_migratetype(page);
3987 /* Only test what is necessary when the reserves are not met */
3990 * Blocks with reserved pages will never free, skip
3993 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3994 if (pageblock_is_reserved(pfn, block_end_pfn))
3997 /* If this block is reserved, account for it */
3998 if (block_migratetype == MIGRATE_RESERVE) {
4003 /* Suitable for reserving if this block is movable */
4004 if (block_migratetype == MIGRATE_MOVABLE) {
4005 set_pageblock_migratetype(page,
4007 move_freepages_block(zone, page,
4012 } else if (!old_reserve) {
4014 * At boot time we don't need to scan the whole zone
4015 * for turning off MIGRATE_RESERVE.
4021 * If the reserve is met and this is a previous reserved block,
4024 if (block_migratetype == MIGRATE_RESERVE) {
4025 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4026 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4032 * Initially all pages are reserved - free ones are freed
4033 * up by free_all_bootmem() once the early boot process is
4034 * done. Non-atomic initialization, single-pass.
4036 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4037 unsigned long start_pfn, enum memmap_context context)
4040 unsigned long end_pfn = start_pfn + size;
4044 if (highest_memmap_pfn < end_pfn - 1)
4045 highest_memmap_pfn = end_pfn - 1;
4047 z = &NODE_DATA(nid)->node_zones[zone];
4048 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4050 * There can be holes in boot-time mem_map[]s
4051 * handed to this function. They do not
4052 * exist on hotplugged memory.
4054 if (context == MEMMAP_EARLY) {
4055 if (!early_pfn_valid(pfn))
4057 if (!early_pfn_in_nid(pfn, nid))
4060 page = pfn_to_page(pfn);
4061 set_page_links(page, zone, nid, pfn);
4062 mminit_verify_page_links(page, zone, nid, pfn);
4063 init_page_count(page);
4064 page_mapcount_reset(page);
4065 page_cpupid_reset_last(page);
4066 SetPageReserved(page);
4068 * Mark the block movable so that blocks are reserved for
4069 * movable at startup. This will force kernel allocations
4070 * to reserve their blocks rather than leaking throughout
4071 * the address space during boot when many long-lived
4072 * kernel allocations are made. Later some blocks near
4073 * the start are marked MIGRATE_RESERVE by
4074 * setup_zone_migrate_reserve()
4076 * bitmap is created for zone's valid pfn range. but memmap
4077 * can be created for invalid pages (for alignment)
4078 * check here not to call set_pageblock_migratetype() against
4081 if ((z->zone_start_pfn <= pfn)
4082 && (pfn < zone_end_pfn(z))
4083 && !(pfn & (pageblock_nr_pages - 1)))
4084 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4086 INIT_LIST_HEAD(&page->lru);
4087 #ifdef WANT_PAGE_VIRTUAL
4088 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4089 if (!is_highmem_idx(zone))
4090 set_page_address(page, __va(pfn << PAGE_SHIFT));
4095 static void __meminit zone_init_free_lists(struct zone *zone)
4098 for_each_migratetype_order(order, t) {
4099 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4100 zone->free_area[order].nr_free = 0;
4104 #ifndef __HAVE_ARCH_MEMMAP_INIT
4105 #define memmap_init(size, nid, zone, start_pfn) \
4106 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4109 static int __meminit zone_batchsize(struct zone *zone)
4115 * The per-cpu-pages pools are set to around 1000th of the
4116 * size of the zone. But no more than 1/2 of a meg.
4118 * OK, so we don't know how big the cache is. So guess.
4120 batch = zone->managed_pages / 1024;
4121 if (batch * PAGE_SIZE > 512 * 1024)
4122 batch = (512 * 1024) / PAGE_SIZE;
4123 batch /= 4; /* We effectively *= 4 below */
4128 * Clamp the batch to a 2^n - 1 value. Having a power
4129 * of 2 value was found to be more likely to have
4130 * suboptimal cache aliasing properties in some cases.
4132 * For example if 2 tasks are alternately allocating
4133 * batches of pages, one task can end up with a lot
4134 * of pages of one half of the possible page colors
4135 * and the other with pages of the other colors.
4137 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4142 /* The deferral and batching of frees should be suppressed under NOMMU
4145 * The problem is that NOMMU needs to be able to allocate large chunks
4146 * of contiguous memory as there's no hardware page translation to
4147 * assemble apparent contiguous memory from discontiguous pages.
4149 * Queueing large contiguous runs of pages for batching, however,
4150 * causes the pages to actually be freed in smaller chunks. As there
4151 * can be a significant delay between the individual batches being
4152 * recycled, this leads to the once large chunks of space being
4153 * fragmented and becoming unavailable for high-order allocations.
4160 * pcp->high and pcp->batch values are related and dependent on one another:
4161 * ->batch must never be higher then ->high.
4162 * The following function updates them in a safe manner without read side
4165 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4166 * those fields changing asynchronously (acording the the above rule).
4168 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4169 * outside of boot time (or some other assurance that no concurrent updaters
4172 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4173 unsigned long batch)
4175 /* start with a fail safe value for batch */
4179 /* Update high, then batch, in order */
4186 /* a companion to pageset_set_high() */
4187 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4189 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4192 static void pageset_init(struct per_cpu_pageset *p)
4194 struct per_cpu_pages *pcp;
4197 memset(p, 0, sizeof(*p));
4201 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4202 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4205 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4208 pageset_set_batch(p, batch);
4212 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4213 * to the value high for the pageset p.
4215 static void pageset_set_high(struct per_cpu_pageset *p,
4218 unsigned long batch = max(1UL, high / 4);
4219 if ((high / 4) > (PAGE_SHIFT * 8))
4220 batch = PAGE_SHIFT * 8;
4222 pageset_update(&p->pcp, high, batch);
4225 static void __meminit pageset_set_high_and_batch(struct zone *zone,
4226 struct per_cpu_pageset *pcp)
4228 if (percpu_pagelist_fraction)
4229 pageset_set_high(pcp,
4230 (zone->managed_pages /
4231 percpu_pagelist_fraction));
4233 pageset_set_batch(pcp, zone_batchsize(zone));
4236 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4238 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4241 pageset_set_high_and_batch(zone, pcp);
4244 static void __meminit setup_zone_pageset(struct zone *zone)
4247 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4248 for_each_possible_cpu(cpu)
4249 zone_pageset_init(zone, cpu);
4253 * Allocate per cpu pagesets and initialize them.
4254 * Before this call only boot pagesets were available.
4256 void __init setup_per_cpu_pageset(void)
4260 for_each_populated_zone(zone)
4261 setup_zone_pageset(zone);
4264 static noinline __init_refok
4265 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4271 * The per-page waitqueue mechanism uses hashed waitqueues
4274 zone->wait_table_hash_nr_entries =
4275 wait_table_hash_nr_entries(zone_size_pages);
4276 zone->wait_table_bits =
4277 wait_table_bits(zone->wait_table_hash_nr_entries);
4278 alloc_size = zone->wait_table_hash_nr_entries
4279 * sizeof(wait_queue_head_t);
4281 if (!slab_is_available()) {
4282 zone->wait_table = (wait_queue_head_t *)
4283 memblock_virt_alloc_node_nopanic(
4284 alloc_size, zone->zone_pgdat->node_id);
4287 * This case means that a zone whose size was 0 gets new memory
4288 * via memory hot-add.
4289 * But it may be the case that a new node was hot-added. In
4290 * this case vmalloc() will not be able to use this new node's
4291 * memory - this wait_table must be initialized to use this new
4292 * node itself as well.
4293 * To use this new node's memory, further consideration will be
4296 zone->wait_table = vmalloc(alloc_size);
4298 if (!zone->wait_table)
4301 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4302 init_waitqueue_head(zone->wait_table + i);
4307 static __meminit void zone_pcp_init(struct zone *zone)
4310 * per cpu subsystem is not up at this point. The following code
4311 * relies on the ability of the linker to provide the
4312 * offset of a (static) per cpu variable into the per cpu area.
4314 zone->pageset = &boot_pageset;
4316 if (populated_zone(zone))
4317 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4318 zone->name, zone->present_pages,
4319 zone_batchsize(zone));
4322 int __meminit init_currently_empty_zone(struct zone *zone,
4323 unsigned long zone_start_pfn,
4325 enum memmap_context context)
4327 struct pglist_data *pgdat = zone->zone_pgdat;
4329 ret = zone_wait_table_init(zone, size);
4332 pgdat->nr_zones = zone_idx(zone) + 1;
4334 zone->zone_start_pfn = zone_start_pfn;
4336 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4337 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4339 (unsigned long)zone_idx(zone),
4340 zone_start_pfn, (zone_start_pfn + size));
4342 zone_init_free_lists(zone);
4347 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4348 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4350 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4351 * Architectures may implement their own version but if add_active_range()
4352 * was used and there are no special requirements, this is a convenient
4355 int __meminit __early_pfn_to_nid(unsigned long pfn)
4357 unsigned long start_pfn, end_pfn;
4360 * NOTE: The following SMP-unsafe globals are only used early in boot
4361 * when the kernel is running single-threaded.
4363 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4364 static int __meminitdata last_nid;
4366 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4369 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4371 last_start_pfn = start_pfn;
4372 last_end_pfn = end_pfn;
4378 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4380 int __meminit early_pfn_to_nid(unsigned long pfn)
4384 nid = __early_pfn_to_nid(pfn);
4387 /* just returns 0 */
4391 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4392 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4396 nid = __early_pfn_to_nid(pfn);
4397 if (nid >= 0 && nid != node)
4404 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4405 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4406 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4408 * If an architecture guarantees that all ranges registered with
4409 * add_active_ranges() contain no holes and may be freed, this
4410 * this function may be used instead of calling memblock_free_early_nid()
4413 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4415 unsigned long start_pfn, end_pfn;
4418 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4419 start_pfn = min(start_pfn, max_low_pfn);
4420 end_pfn = min(end_pfn, max_low_pfn);
4422 if (start_pfn < end_pfn)
4423 memblock_free_early_nid(PFN_PHYS(start_pfn),
4424 (end_pfn - start_pfn) << PAGE_SHIFT,
4430 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4431 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4433 * If an architecture guarantees that all ranges registered with
4434 * add_active_ranges() contain no holes and may be freed, this
4435 * function may be used instead of calling memory_present() manually.
4437 void __init sparse_memory_present_with_active_regions(int nid)
4439 unsigned long start_pfn, end_pfn;
4442 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4443 memory_present(this_nid, start_pfn, end_pfn);
4447 * get_pfn_range_for_nid - Return the start and end page frames for a node
4448 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4449 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4450 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4452 * It returns the start and end page frame of a node based on information
4453 * provided by an arch calling add_active_range(). If called for a node
4454 * with no available memory, a warning is printed and the start and end
4457 void __meminit get_pfn_range_for_nid(unsigned int nid,
4458 unsigned long *start_pfn, unsigned long *end_pfn)
4460 unsigned long this_start_pfn, this_end_pfn;
4466 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4467 *start_pfn = min(*start_pfn, this_start_pfn);
4468 *end_pfn = max(*end_pfn, this_end_pfn);
4471 if (*start_pfn == -1UL)
4476 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4477 * assumption is made that zones within a node are ordered in monotonic
4478 * increasing memory addresses so that the "highest" populated zone is used
4480 static void __init find_usable_zone_for_movable(void)
4483 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4484 if (zone_index == ZONE_MOVABLE)
4487 if (arch_zone_highest_possible_pfn[zone_index] >
4488 arch_zone_lowest_possible_pfn[zone_index])
4492 VM_BUG_ON(zone_index == -1);
4493 movable_zone = zone_index;
4497 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4498 * because it is sized independent of architecture. Unlike the other zones,
4499 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4500 * in each node depending on the size of each node and how evenly kernelcore
4501 * is distributed. This helper function adjusts the zone ranges
4502 * provided by the architecture for a given node by using the end of the
4503 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4504 * zones within a node are in order of monotonic increases memory addresses
4506 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4507 unsigned long zone_type,
4508 unsigned long node_start_pfn,
4509 unsigned long node_end_pfn,
4510 unsigned long *zone_start_pfn,
4511 unsigned long *zone_end_pfn)
4513 /* Only adjust if ZONE_MOVABLE is on this node */
4514 if (zone_movable_pfn[nid]) {
4515 /* Size ZONE_MOVABLE */
4516 if (zone_type == ZONE_MOVABLE) {
4517 *zone_start_pfn = zone_movable_pfn[nid];
4518 *zone_end_pfn = min(node_end_pfn,
4519 arch_zone_highest_possible_pfn[movable_zone]);
4521 /* Adjust for ZONE_MOVABLE starting within this range */
4522 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4523 *zone_end_pfn > zone_movable_pfn[nid]) {
4524 *zone_end_pfn = zone_movable_pfn[nid];
4526 /* Check if this whole range is within ZONE_MOVABLE */
4527 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4528 *zone_start_pfn = *zone_end_pfn;
4533 * Return the number of pages a zone spans in a node, including holes
4534 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4536 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4537 unsigned long zone_type,
4538 unsigned long node_start_pfn,
4539 unsigned long node_end_pfn,
4540 unsigned long *ignored)
4542 unsigned long zone_start_pfn, zone_end_pfn;
4544 /* Get the start and end of the zone */
4545 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4546 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4547 adjust_zone_range_for_zone_movable(nid, zone_type,
4548 node_start_pfn, node_end_pfn,
4549 &zone_start_pfn, &zone_end_pfn);
4551 /* Check that this node has pages within the zone's required range */
4552 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4555 /* Move the zone boundaries inside the node if necessary */
4556 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4557 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4559 /* Return the spanned pages */
4560 return zone_end_pfn - zone_start_pfn;
4564 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4565 * then all holes in the requested range will be accounted for.
4567 unsigned long __meminit __absent_pages_in_range(int nid,
4568 unsigned long range_start_pfn,
4569 unsigned long range_end_pfn)
4571 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4572 unsigned long start_pfn, end_pfn;
4575 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4576 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4577 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4578 nr_absent -= end_pfn - start_pfn;
4584 * absent_pages_in_range - Return number of page frames in holes within a range
4585 * @start_pfn: The start PFN to start searching for holes
4586 * @end_pfn: The end PFN to stop searching for holes
4588 * It returns the number of pages frames in memory holes within a range.
4590 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4591 unsigned long end_pfn)
4593 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4596 /* Return the number of page frames in holes in a zone on a node */
4597 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4598 unsigned long zone_type,
4599 unsigned long node_start_pfn,
4600 unsigned long node_end_pfn,
4601 unsigned long *ignored)
4603 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4604 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4605 unsigned long zone_start_pfn, zone_end_pfn;
4607 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4608 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4610 adjust_zone_range_for_zone_movable(nid, zone_type,
4611 node_start_pfn, node_end_pfn,
4612 &zone_start_pfn, &zone_end_pfn);
4613 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4616 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4617 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4618 unsigned long zone_type,
4619 unsigned long node_start_pfn,
4620 unsigned long node_end_pfn,
4621 unsigned long *zones_size)
4623 return zones_size[zone_type];
4626 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4627 unsigned long zone_type,
4628 unsigned long node_start_pfn,
4629 unsigned long node_end_pfn,
4630 unsigned long *zholes_size)
4635 return zholes_size[zone_type];
4638 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4640 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4641 unsigned long node_start_pfn,
4642 unsigned long node_end_pfn,
4643 unsigned long *zones_size,
4644 unsigned long *zholes_size)
4646 unsigned long realtotalpages, totalpages = 0;
4649 for (i = 0; i < MAX_NR_ZONES; i++)
4650 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4654 pgdat->node_spanned_pages = totalpages;
4656 realtotalpages = totalpages;
4657 for (i = 0; i < MAX_NR_ZONES; i++)
4659 zone_absent_pages_in_node(pgdat->node_id, i,
4660 node_start_pfn, node_end_pfn,
4662 pgdat->node_present_pages = realtotalpages;
4663 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4667 #ifndef CONFIG_SPARSEMEM
4669 * Calculate the size of the zone->blockflags rounded to an unsigned long
4670 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4671 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4672 * round what is now in bits to nearest long in bits, then return it in
4675 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4677 unsigned long usemapsize;
4679 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4680 usemapsize = roundup(zonesize, pageblock_nr_pages);
4681 usemapsize = usemapsize >> pageblock_order;
4682 usemapsize *= NR_PAGEBLOCK_BITS;
4683 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4685 return usemapsize / 8;
4688 static void __init setup_usemap(struct pglist_data *pgdat,
4690 unsigned long zone_start_pfn,
4691 unsigned long zonesize)
4693 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4694 zone->pageblock_flags = NULL;
4696 zone->pageblock_flags =
4697 memblock_virt_alloc_node_nopanic(usemapsize,
4701 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4702 unsigned long zone_start_pfn, unsigned long zonesize) {}
4703 #endif /* CONFIG_SPARSEMEM */
4705 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4707 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4708 void __paginginit set_pageblock_order(void)
4712 /* Check that pageblock_nr_pages has not already been setup */
4713 if (pageblock_order)
4716 if (HPAGE_SHIFT > PAGE_SHIFT)
4717 order = HUGETLB_PAGE_ORDER;
4719 order = MAX_ORDER - 1;
4722 * Assume the largest contiguous order of interest is a huge page.
4723 * This value may be variable depending on boot parameters on IA64 and
4726 pageblock_order = order;
4728 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4731 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4732 * is unused as pageblock_order is set at compile-time. See
4733 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4736 void __paginginit set_pageblock_order(void)
4740 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4742 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4743 unsigned long present_pages)
4745 unsigned long pages = spanned_pages;
4748 * Provide a more accurate estimation if there are holes within
4749 * the zone and SPARSEMEM is in use. If there are holes within the
4750 * zone, each populated memory region may cost us one or two extra
4751 * memmap pages due to alignment because memmap pages for each
4752 * populated regions may not naturally algined on page boundary.
4753 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4755 if (spanned_pages > present_pages + (present_pages >> 4) &&
4756 IS_ENABLED(CONFIG_SPARSEMEM))
4757 pages = present_pages;
4759 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4763 * Set up the zone data structures:
4764 * - mark all pages reserved
4765 * - mark all memory queues empty
4766 * - clear the memory bitmaps
4768 * NOTE: pgdat should get zeroed by caller.
4770 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4771 unsigned long node_start_pfn, unsigned long node_end_pfn,
4772 unsigned long *zones_size, unsigned long *zholes_size)
4775 int nid = pgdat->node_id;
4776 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4779 pgdat_resize_init(pgdat);
4780 #ifdef CONFIG_NUMA_BALANCING
4781 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4782 pgdat->numabalancing_migrate_nr_pages = 0;
4783 pgdat->numabalancing_migrate_next_window = jiffies;
4785 init_waitqueue_head(&pgdat->kswapd_wait);
4786 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4787 pgdat_page_cgroup_init(pgdat);
4789 for (j = 0; j < MAX_NR_ZONES; j++) {
4790 struct zone *zone = pgdat->node_zones + j;
4791 unsigned long size, realsize, freesize, memmap_pages;
4793 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4794 node_end_pfn, zones_size);
4795 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4801 * Adjust freesize so that it accounts for how much memory
4802 * is used by this zone for memmap. This affects the watermark
4803 * and per-cpu initialisations
4805 memmap_pages = calc_memmap_size(size, realsize);
4806 if (freesize >= memmap_pages) {
4807 freesize -= memmap_pages;
4810 " %s zone: %lu pages used for memmap\n",
4811 zone_names[j], memmap_pages);
4814 " %s zone: %lu pages exceeds freesize %lu\n",
4815 zone_names[j], memmap_pages, freesize);
4817 /* Account for reserved pages */
4818 if (j == 0 && freesize > dma_reserve) {
4819 freesize -= dma_reserve;
4820 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4821 zone_names[0], dma_reserve);
4824 if (!is_highmem_idx(j))
4825 nr_kernel_pages += freesize;
4826 /* Charge for highmem memmap if there are enough kernel pages */
4827 else if (nr_kernel_pages > memmap_pages * 2)
4828 nr_kernel_pages -= memmap_pages;
4829 nr_all_pages += freesize;
4831 zone->spanned_pages = size;
4832 zone->present_pages = realsize;
4834 * Set an approximate value for lowmem here, it will be adjusted
4835 * when the bootmem allocator frees pages into the buddy system.
4836 * And all highmem pages will be managed by the buddy system.
4838 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4841 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4843 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4845 zone->name = zone_names[j];
4846 spin_lock_init(&zone->lock);
4847 spin_lock_init(&zone->lru_lock);
4848 zone_seqlock_init(zone);
4849 zone->zone_pgdat = pgdat;
4850 zone_pcp_init(zone);
4852 /* For bootup, initialized properly in watermark setup */
4853 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4855 lruvec_init(&zone->lruvec);
4859 set_pageblock_order();
4860 setup_usemap(pgdat, zone, zone_start_pfn, size);
4861 ret = init_currently_empty_zone(zone, zone_start_pfn,
4862 size, MEMMAP_EARLY);
4864 memmap_init(size, nid, j, zone_start_pfn);
4865 zone_start_pfn += size;
4869 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4871 /* Skip empty nodes */
4872 if (!pgdat->node_spanned_pages)
4875 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4876 /* ia64 gets its own node_mem_map, before this, without bootmem */
4877 if (!pgdat->node_mem_map) {
4878 unsigned long size, start, end;
4882 * The zone's endpoints aren't required to be MAX_ORDER
4883 * aligned but the node_mem_map endpoints must be in order
4884 * for the buddy allocator to function correctly.
4886 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4887 end = pgdat_end_pfn(pgdat);
4888 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4889 size = (end - start) * sizeof(struct page);
4890 map = alloc_remap(pgdat->node_id, size);
4892 map = memblock_virt_alloc_node_nopanic(size,
4894 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4896 #ifndef CONFIG_NEED_MULTIPLE_NODES
4898 * With no DISCONTIG, the global mem_map is just set as node 0's
4900 if (pgdat == NODE_DATA(0)) {
4901 mem_map = NODE_DATA(0)->node_mem_map;
4902 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4903 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4904 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4905 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4908 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4911 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4912 unsigned long node_start_pfn, unsigned long *zholes_size)
4914 pg_data_t *pgdat = NODE_DATA(nid);
4915 unsigned long start_pfn = 0;
4916 unsigned long end_pfn = 0;
4918 /* pg_data_t should be reset to zero when it's allocated */
4919 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4921 pgdat->node_id = nid;
4922 pgdat->node_start_pfn = node_start_pfn;
4923 if (node_state(nid, N_MEMORY))
4924 init_zone_allows_reclaim(nid);
4925 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4926 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4928 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4929 zones_size, zholes_size);
4931 alloc_node_mem_map(pgdat);
4932 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4933 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4934 nid, (unsigned long)pgdat,
4935 (unsigned long)pgdat->node_mem_map);
4938 free_area_init_core(pgdat, start_pfn, end_pfn,
4939 zones_size, zholes_size);
4942 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4944 #if MAX_NUMNODES > 1
4946 * Figure out the number of possible node ids.
4948 void __init setup_nr_node_ids(void)
4951 unsigned int highest = 0;
4953 for_each_node_mask(node, node_possible_map)
4955 nr_node_ids = highest + 1;
4960 * node_map_pfn_alignment - determine the maximum internode alignment
4962 * This function should be called after node map is populated and sorted.
4963 * It calculates the maximum power of two alignment which can distinguish
4966 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4967 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4968 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4969 * shifted, 1GiB is enough and this function will indicate so.
4971 * This is used to test whether pfn -> nid mapping of the chosen memory
4972 * model has fine enough granularity to avoid incorrect mapping for the
4973 * populated node map.
4975 * Returns the determined alignment in pfn's. 0 if there is no alignment
4976 * requirement (single node).
4978 unsigned long __init node_map_pfn_alignment(void)
4980 unsigned long accl_mask = 0, last_end = 0;
4981 unsigned long start, end, mask;
4985 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4986 if (!start || last_nid < 0 || last_nid == nid) {
4993 * Start with a mask granular enough to pin-point to the
4994 * start pfn and tick off bits one-by-one until it becomes
4995 * too coarse to separate the current node from the last.
4997 mask = ~((1 << __ffs(start)) - 1);
4998 while (mask && last_end <= (start & (mask << 1)))
5001 /* accumulate all internode masks */
5005 /* convert mask to number of pages */
5006 return ~accl_mask + 1;
5009 /* Find the lowest pfn for a node */
5010 static unsigned long __init find_min_pfn_for_node(int nid)
5012 unsigned long min_pfn = ULONG_MAX;
5013 unsigned long start_pfn;
5016 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5017 min_pfn = min(min_pfn, start_pfn);
5019 if (min_pfn == ULONG_MAX) {
5021 "Could not find start_pfn for node %d\n", nid);
5029 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5031 * It returns the minimum PFN based on information provided via
5032 * add_active_range().
5034 unsigned long __init find_min_pfn_with_active_regions(void)
5036 return find_min_pfn_for_node(MAX_NUMNODES);
5040 * early_calculate_totalpages()
5041 * Sum pages in active regions for movable zone.
5042 * Populate N_MEMORY for calculating usable_nodes.
5044 static unsigned long __init early_calculate_totalpages(void)
5046 unsigned long totalpages = 0;
5047 unsigned long start_pfn, end_pfn;
5050 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5051 unsigned long pages = end_pfn - start_pfn;
5053 totalpages += pages;
5055 node_set_state(nid, N_MEMORY);
5061 * Find the PFN the Movable zone begins in each node. Kernel memory
5062 * is spread evenly between nodes as long as the nodes have enough
5063 * memory. When they don't, some nodes will have more kernelcore than
5066 static void __init find_zone_movable_pfns_for_nodes(void)
5069 unsigned long usable_startpfn;
5070 unsigned long kernelcore_node, kernelcore_remaining;
5071 /* save the state before borrow the nodemask */
5072 nodemask_t saved_node_state = node_states[N_MEMORY];
5073 unsigned long totalpages = early_calculate_totalpages();
5074 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5075 struct memblock_type *type = &memblock.memory;
5077 /* Need to find movable_zone earlier when movable_node is specified. */
5078 find_usable_zone_for_movable();
5081 * If movable_node is specified, ignore kernelcore and movablecore
5084 if (movable_node_is_enabled()) {
5085 for (i = 0; i < type->cnt; i++) {
5086 if (!memblock_is_hotpluggable(&type->regions[i]))
5089 nid = type->regions[i].nid;
5091 usable_startpfn = PFN_DOWN(type->regions[i].base);
5092 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5093 min(usable_startpfn, zone_movable_pfn[nid]) :
5101 * If movablecore=nn[KMG] was specified, calculate what size of
5102 * kernelcore that corresponds so that memory usable for
5103 * any allocation type is evenly spread. If both kernelcore
5104 * and movablecore are specified, then the value of kernelcore
5105 * will be used for required_kernelcore if it's greater than
5106 * what movablecore would have allowed.
5108 if (required_movablecore) {
5109 unsigned long corepages;
5112 * Round-up so that ZONE_MOVABLE is at least as large as what
5113 * was requested by the user
5115 required_movablecore =
5116 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5117 corepages = totalpages - required_movablecore;
5119 required_kernelcore = max(required_kernelcore, corepages);
5122 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5123 if (!required_kernelcore)
5126 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5127 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5130 /* Spread kernelcore memory as evenly as possible throughout nodes */
5131 kernelcore_node = required_kernelcore / usable_nodes;
5132 for_each_node_state(nid, N_MEMORY) {
5133 unsigned long start_pfn, end_pfn;
5136 * Recalculate kernelcore_node if the division per node
5137 * now exceeds what is necessary to satisfy the requested
5138 * amount of memory for the kernel
5140 if (required_kernelcore < kernelcore_node)
5141 kernelcore_node = required_kernelcore / usable_nodes;
5144 * As the map is walked, we track how much memory is usable
5145 * by the kernel using kernelcore_remaining. When it is
5146 * 0, the rest of the node is usable by ZONE_MOVABLE
5148 kernelcore_remaining = kernelcore_node;
5150 /* Go through each range of PFNs within this node */
5151 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5152 unsigned long size_pages;
5154 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5155 if (start_pfn >= end_pfn)
5158 /* Account for what is only usable for kernelcore */
5159 if (start_pfn < usable_startpfn) {
5160 unsigned long kernel_pages;
5161 kernel_pages = min(end_pfn, usable_startpfn)
5164 kernelcore_remaining -= min(kernel_pages,
5165 kernelcore_remaining);
5166 required_kernelcore -= min(kernel_pages,
5167 required_kernelcore);
5169 /* Continue if range is now fully accounted */
5170 if (end_pfn <= usable_startpfn) {
5173 * Push zone_movable_pfn to the end so
5174 * that if we have to rebalance
5175 * kernelcore across nodes, we will
5176 * not double account here
5178 zone_movable_pfn[nid] = end_pfn;
5181 start_pfn = usable_startpfn;
5185 * The usable PFN range for ZONE_MOVABLE is from
5186 * start_pfn->end_pfn. Calculate size_pages as the
5187 * number of pages used as kernelcore
5189 size_pages = end_pfn - start_pfn;
5190 if (size_pages > kernelcore_remaining)
5191 size_pages = kernelcore_remaining;
5192 zone_movable_pfn[nid] = start_pfn + size_pages;
5195 * Some kernelcore has been met, update counts and
5196 * break if the kernelcore for this node has been
5199 required_kernelcore -= min(required_kernelcore,
5201 kernelcore_remaining -= size_pages;
5202 if (!kernelcore_remaining)
5208 * If there is still required_kernelcore, we do another pass with one
5209 * less node in the count. This will push zone_movable_pfn[nid] further
5210 * along on the nodes that still have memory until kernelcore is
5214 if (usable_nodes && required_kernelcore > usable_nodes)
5218 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5219 for (nid = 0; nid < MAX_NUMNODES; nid++)
5220 zone_movable_pfn[nid] =
5221 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5224 /* restore the node_state */
5225 node_states[N_MEMORY] = saved_node_state;
5228 /* Any regular or high memory on that node ? */
5229 static void check_for_memory(pg_data_t *pgdat, int nid)
5231 enum zone_type zone_type;
5233 if (N_MEMORY == N_NORMAL_MEMORY)
5236 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5237 struct zone *zone = &pgdat->node_zones[zone_type];
5238 if (populated_zone(zone)) {
5239 node_set_state(nid, N_HIGH_MEMORY);
5240 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5241 zone_type <= ZONE_NORMAL)
5242 node_set_state(nid, N_NORMAL_MEMORY);
5249 * free_area_init_nodes - Initialise all pg_data_t and zone data
5250 * @max_zone_pfn: an array of max PFNs for each zone
5252 * This will call free_area_init_node() for each active node in the system.
5253 * Using the page ranges provided by add_active_range(), the size of each
5254 * zone in each node and their holes is calculated. If the maximum PFN
5255 * between two adjacent zones match, it is assumed that the zone is empty.
5256 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5257 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5258 * starts where the previous one ended. For example, ZONE_DMA32 starts
5259 * at arch_max_dma_pfn.
5261 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5263 unsigned long start_pfn, end_pfn;
5266 /* Record where the zone boundaries are */
5267 memset(arch_zone_lowest_possible_pfn, 0,
5268 sizeof(arch_zone_lowest_possible_pfn));
5269 memset(arch_zone_highest_possible_pfn, 0,
5270 sizeof(arch_zone_highest_possible_pfn));
5271 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5272 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5273 for (i = 1; i < MAX_NR_ZONES; i++) {
5274 if (i == ZONE_MOVABLE)
5276 arch_zone_lowest_possible_pfn[i] =
5277 arch_zone_highest_possible_pfn[i-1];
5278 arch_zone_highest_possible_pfn[i] =
5279 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5281 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5282 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5284 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5285 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5286 find_zone_movable_pfns_for_nodes();
5288 /* Print out the zone ranges */
5289 printk("Zone ranges:\n");
5290 for (i = 0; i < MAX_NR_ZONES; i++) {
5291 if (i == ZONE_MOVABLE)
5293 printk(KERN_CONT " %-8s ", zone_names[i]);
5294 if (arch_zone_lowest_possible_pfn[i] ==
5295 arch_zone_highest_possible_pfn[i])
5296 printk(KERN_CONT "empty\n");
5298 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5299 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5300 (arch_zone_highest_possible_pfn[i]
5301 << PAGE_SHIFT) - 1);
5304 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5305 printk("Movable zone start for each node\n");
5306 for (i = 0; i < MAX_NUMNODES; i++) {
5307 if (zone_movable_pfn[i])
5308 printk(" Node %d: %#010lx\n", i,
5309 zone_movable_pfn[i] << PAGE_SHIFT);
5312 /* Print out the early node map */
5313 printk("Early memory node ranges\n");
5314 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5315 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5316 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5318 /* Initialise every node */
5319 mminit_verify_pageflags_layout();
5320 setup_nr_node_ids();
5321 for_each_online_node(nid) {
5322 pg_data_t *pgdat = NODE_DATA(nid);
5323 free_area_init_node(nid, NULL,
5324 find_min_pfn_for_node(nid), NULL);
5326 /* Any memory on that node */
5327 if (pgdat->node_present_pages)
5328 node_set_state(nid, N_MEMORY);
5329 check_for_memory(pgdat, nid);
5333 static int __init cmdline_parse_core(char *p, unsigned long *core)
5335 unsigned long long coremem;
5339 coremem = memparse(p, &p);
5340 *core = coremem >> PAGE_SHIFT;
5342 /* Paranoid check that UL is enough for the coremem value */
5343 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5349 * kernelcore=size sets the amount of memory for use for allocations that
5350 * cannot be reclaimed or migrated.
5352 static int __init cmdline_parse_kernelcore(char *p)
5354 return cmdline_parse_core(p, &required_kernelcore);
5358 * movablecore=size sets the amount of memory for use for allocations that
5359 * can be reclaimed or migrated.
5361 static int __init cmdline_parse_movablecore(char *p)
5363 return cmdline_parse_core(p, &required_movablecore);
5366 early_param("kernelcore", cmdline_parse_kernelcore);
5367 early_param("movablecore", cmdline_parse_movablecore);
5369 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5371 void adjust_managed_page_count(struct page *page, long count)
5373 spin_lock(&managed_page_count_lock);
5374 page_zone(page)->managed_pages += count;
5375 totalram_pages += count;
5376 #ifdef CONFIG_HIGHMEM
5377 if (PageHighMem(page))
5378 totalhigh_pages += count;
5380 spin_unlock(&managed_page_count_lock);
5382 EXPORT_SYMBOL(adjust_managed_page_count);
5384 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5387 unsigned long pages = 0;
5389 start = (void *)PAGE_ALIGN((unsigned long)start);
5390 end = (void *)((unsigned long)end & PAGE_MASK);
5391 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5392 if ((unsigned int)poison <= 0xFF)
5393 memset(pos, poison, PAGE_SIZE);
5394 free_reserved_page(virt_to_page(pos));
5398 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5399 s, pages << (PAGE_SHIFT - 10), start, end);
5403 EXPORT_SYMBOL(free_reserved_area);
5405 #ifdef CONFIG_HIGHMEM
5406 void free_highmem_page(struct page *page)
5408 __free_reserved_page(page);
5410 page_zone(page)->managed_pages++;
5416 void __init mem_init_print_info(const char *str)
5418 unsigned long physpages, codesize, datasize, rosize, bss_size;
5419 unsigned long init_code_size, init_data_size;
5421 physpages = get_num_physpages();
5422 codesize = _etext - _stext;
5423 datasize = _edata - _sdata;
5424 rosize = __end_rodata - __start_rodata;
5425 bss_size = __bss_stop - __bss_start;
5426 init_data_size = __init_end - __init_begin;
5427 init_code_size = _einittext - _sinittext;
5430 * Detect special cases and adjust section sizes accordingly:
5431 * 1) .init.* may be embedded into .data sections
5432 * 2) .init.text.* may be out of [__init_begin, __init_end],
5433 * please refer to arch/tile/kernel/vmlinux.lds.S.
5434 * 3) .rodata.* may be embedded into .text or .data sections.
5436 #define adj_init_size(start, end, size, pos, adj) \
5438 if (start <= pos && pos < end && size > adj) \
5442 adj_init_size(__init_begin, __init_end, init_data_size,
5443 _sinittext, init_code_size);
5444 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5445 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5446 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5447 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5449 #undef adj_init_size
5451 printk("Memory: %luK/%luK available "
5452 "(%luK kernel code, %luK rwdata, %luK rodata, "
5453 "%luK init, %luK bss, %luK reserved"
5454 #ifdef CONFIG_HIGHMEM
5458 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5459 codesize >> 10, datasize >> 10, rosize >> 10,
5460 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5461 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5462 #ifdef CONFIG_HIGHMEM
5463 totalhigh_pages << (PAGE_SHIFT-10),
5465 str ? ", " : "", str ? str : "");
5469 * set_dma_reserve - set the specified number of pages reserved in the first zone
5470 * @new_dma_reserve: The number of pages to mark reserved
5472 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5473 * In the DMA zone, a significant percentage may be consumed by kernel image
5474 * and other unfreeable allocations which can skew the watermarks badly. This
5475 * function may optionally be used to account for unfreeable pages in the
5476 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5477 * smaller per-cpu batchsize.
5479 void __init set_dma_reserve(unsigned long new_dma_reserve)
5481 dma_reserve = new_dma_reserve;
5484 void __init free_area_init(unsigned long *zones_size)
5486 free_area_init_node(0, zones_size,
5487 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5490 static int page_alloc_cpu_notify(struct notifier_block *self,
5491 unsigned long action, void *hcpu)
5493 int cpu = (unsigned long)hcpu;
5495 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5496 lru_add_drain_cpu(cpu);
5500 * Spill the event counters of the dead processor
5501 * into the current processors event counters.
5502 * This artificially elevates the count of the current
5505 vm_events_fold_cpu(cpu);
5508 * Zero the differential counters of the dead processor
5509 * so that the vm statistics are consistent.
5511 * This is only okay since the processor is dead and cannot
5512 * race with what we are doing.
5514 cpu_vm_stats_fold(cpu);
5519 void __init page_alloc_init(void)
5521 hotcpu_notifier(page_alloc_cpu_notify, 0);
5525 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5526 * or min_free_kbytes changes.
5528 static void calculate_totalreserve_pages(void)
5530 struct pglist_data *pgdat;
5531 unsigned long reserve_pages = 0;
5532 enum zone_type i, j;
5534 for_each_online_pgdat(pgdat) {
5535 for (i = 0; i < MAX_NR_ZONES; i++) {
5536 struct zone *zone = pgdat->node_zones + i;
5537 unsigned long max = 0;
5539 /* Find valid and maximum lowmem_reserve in the zone */
5540 for (j = i; j < MAX_NR_ZONES; j++) {
5541 if (zone->lowmem_reserve[j] > max)
5542 max = zone->lowmem_reserve[j];
5545 /* we treat the high watermark as reserved pages. */
5546 max += high_wmark_pages(zone);
5548 if (max > zone->managed_pages)
5549 max = zone->managed_pages;
5550 reserve_pages += max;
5552 * Lowmem reserves are not available to
5553 * GFP_HIGHUSER page cache allocations and
5554 * kswapd tries to balance zones to their high
5555 * watermark. As a result, neither should be
5556 * regarded as dirtyable memory, to prevent a
5557 * situation where reclaim has to clean pages
5558 * in order to balance the zones.
5560 zone->dirty_balance_reserve = max;
5563 dirty_balance_reserve = reserve_pages;
5564 totalreserve_pages = reserve_pages;
5568 * setup_per_zone_lowmem_reserve - called whenever
5569 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5570 * has a correct pages reserved value, so an adequate number of
5571 * pages are left in the zone after a successful __alloc_pages().
5573 static void setup_per_zone_lowmem_reserve(void)
5575 struct pglist_data *pgdat;
5576 enum zone_type j, idx;
5578 for_each_online_pgdat(pgdat) {
5579 for (j = 0; j < MAX_NR_ZONES; j++) {
5580 struct zone *zone = pgdat->node_zones + j;
5581 unsigned long managed_pages = zone->managed_pages;
5583 zone->lowmem_reserve[j] = 0;
5587 struct zone *lower_zone;
5591 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5592 sysctl_lowmem_reserve_ratio[idx] = 1;
5594 lower_zone = pgdat->node_zones + idx;
5595 lower_zone->lowmem_reserve[j] = managed_pages /
5596 sysctl_lowmem_reserve_ratio[idx];
5597 managed_pages += lower_zone->managed_pages;
5602 /* update totalreserve_pages */
5603 calculate_totalreserve_pages();
5606 static void __setup_per_zone_wmarks(void)
5608 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5609 unsigned long lowmem_pages = 0;
5611 unsigned long flags;
5613 /* Calculate total number of !ZONE_HIGHMEM pages */
5614 for_each_zone(zone) {
5615 if (!is_highmem(zone))
5616 lowmem_pages += zone->managed_pages;
5619 for_each_zone(zone) {
5622 spin_lock_irqsave(&zone->lock, flags);
5623 tmp = (u64)pages_min * zone->managed_pages;
5624 do_div(tmp, lowmem_pages);
5625 if (is_highmem(zone)) {
5627 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5628 * need highmem pages, so cap pages_min to a small
5631 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5632 * deltas controls asynch page reclaim, and so should
5633 * not be capped for highmem.
5635 unsigned long min_pages;
5637 min_pages = zone->managed_pages / 1024;
5638 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5639 zone->watermark[WMARK_MIN] = min_pages;
5642 * If it's a lowmem zone, reserve a number of pages
5643 * proportionate to the zone's size.
5645 zone->watermark[WMARK_MIN] = tmp;
5648 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5649 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5651 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5652 high_wmark_pages(zone) -
5653 low_wmark_pages(zone) -
5654 zone_page_state(zone, NR_ALLOC_BATCH));
5656 setup_zone_migrate_reserve(zone);
5657 spin_unlock_irqrestore(&zone->lock, flags);
5660 /* update totalreserve_pages */
5661 calculate_totalreserve_pages();
5665 * setup_per_zone_wmarks - called when min_free_kbytes changes
5666 * or when memory is hot-{added|removed}
5668 * Ensures that the watermark[min,low,high] values for each zone are set
5669 * correctly with respect to min_free_kbytes.
5671 void setup_per_zone_wmarks(void)
5673 mutex_lock(&zonelists_mutex);
5674 __setup_per_zone_wmarks();
5675 mutex_unlock(&zonelists_mutex);
5679 * The inactive anon list should be small enough that the VM never has to
5680 * do too much work, but large enough that each inactive page has a chance
5681 * to be referenced again before it is swapped out.
5683 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5684 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5685 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5686 * the anonymous pages are kept on the inactive list.
5689 * memory ratio inactive anon
5690 * -------------------------------------
5699 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5701 unsigned int gb, ratio;
5703 /* Zone size in gigabytes */
5704 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5706 ratio = int_sqrt(10 * gb);
5710 zone->inactive_ratio = ratio;
5713 static void __meminit setup_per_zone_inactive_ratio(void)
5718 calculate_zone_inactive_ratio(zone);
5722 * Initialise min_free_kbytes.
5724 * For small machines we want it small (128k min). For large machines
5725 * we want it large (64MB max). But it is not linear, because network
5726 * bandwidth does not increase linearly with machine size. We use
5728 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5729 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5745 int __meminit init_per_zone_wmark_min(void)
5747 unsigned long lowmem_kbytes;
5748 int new_min_free_kbytes;
5750 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5751 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5753 if (new_min_free_kbytes > user_min_free_kbytes) {
5754 min_free_kbytes = new_min_free_kbytes;
5755 if (min_free_kbytes < 128)
5756 min_free_kbytes = 128;
5757 if (min_free_kbytes > 65536)
5758 min_free_kbytes = 65536;
5760 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5761 new_min_free_kbytes, user_min_free_kbytes);
5763 setup_per_zone_wmarks();
5764 refresh_zone_stat_thresholds();
5765 setup_per_zone_lowmem_reserve();
5766 setup_per_zone_inactive_ratio();
5769 module_init(init_per_zone_wmark_min)
5772 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5773 * that we can call two helper functions whenever min_free_kbytes
5776 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5777 void __user *buffer, size_t *length, loff_t *ppos)
5781 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5786 user_min_free_kbytes = min_free_kbytes;
5787 setup_per_zone_wmarks();
5793 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5794 void __user *buffer, size_t *length, loff_t *ppos)
5799 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5804 zone->min_unmapped_pages = (zone->managed_pages *
5805 sysctl_min_unmapped_ratio) / 100;
5809 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5810 void __user *buffer, size_t *length, loff_t *ppos)
5815 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5820 zone->min_slab_pages = (zone->managed_pages *
5821 sysctl_min_slab_ratio) / 100;
5827 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5828 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5829 * whenever sysctl_lowmem_reserve_ratio changes.
5831 * The reserve ratio obviously has absolutely no relation with the
5832 * minimum watermarks. The lowmem reserve ratio can only make sense
5833 * if in function of the boot time zone sizes.
5835 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5836 void __user *buffer, size_t *length, loff_t *ppos)
5838 proc_dointvec_minmax(table, write, buffer, length, ppos);
5839 setup_per_zone_lowmem_reserve();
5844 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5845 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5846 * pagelist can have before it gets flushed back to buddy allocator.
5848 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5849 void __user *buffer, size_t *length, loff_t *ppos)
5855 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5856 if (!write || (ret < 0))
5859 mutex_lock(&pcp_batch_high_lock);
5860 for_each_populated_zone(zone) {
5862 high = zone->managed_pages / percpu_pagelist_fraction;
5863 for_each_possible_cpu(cpu)
5864 pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5867 mutex_unlock(&pcp_batch_high_lock);
5871 int hashdist = HASHDIST_DEFAULT;
5874 static int __init set_hashdist(char *str)
5878 hashdist = simple_strtoul(str, &str, 0);
5881 __setup("hashdist=", set_hashdist);
5885 * allocate a large system hash table from bootmem
5886 * - it is assumed that the hash table must contain an exact power-of-2
5887 * quantity of entries
5888 * - limit is the number of hash buckets, not the total allocation size
5890 void *__init alloc_large_system_hash(const char *tablename,
5891 unsigned long bucketsize,
5892 unsigned long numentries,
5895 unsigned int *_hash_shift,
5896 unsigned int *_hash_mask,
5897 unsigned long low_limit,
5898 unsigned long high_limit)
5900 unsigned long long max = high_limit;
5901 unsigned long log2qty, size;
5904 /* allow the kernel cmdline to have a say */
5906 /* round applicable memory size up to nearest megabyte */
5907 numentries = nr_kernel_pages;
5909 /* It isn't necessary when PAGE_SIZE >= 1MB */
5910 if (PAGE_SHIFT < 20)
5911 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5913 /* limit to 1 bucket per 2^scale bytes of low memory */
5914 if (scale > PAGE_SHIFT)
5915 numentries >>= (scale - PAGE_SHIFT);
5917 numentries <<= (PAGE_SHIFT - scale);
5919 /* Make sure we've got at least a 0-order allocation.. */
5920 if (unlikely(flags & HASH_SMALL)) {
5921 /* Makes no sense without HASH_EARLY */
5922 WARN_ON(!(flags & HASH_EARLY));
5923 if (!(numentries >> *_hash_shift)) {
5924 numentries = 1UL << *_hash_shift;
5925 BUG_ON(!numentries);
5927 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5928 numentries = PAGE_SIZE / bucketsize;
5930 numentries = roundup_pow_of_two(numentries);
5932 /* limit allocation size to 1/16 total memory by default */
5934 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5935 do_div(max, bucketsize);
5937 max = min(max, 0x80000000ULL);
5939 if (numentries < low_limit)
5940 numentries = low_limit;
5941 if (numentries > max)
5944 log2qty = ilog2(numentries);
5947 size = bucketsize << log2qty;
5948 if (flags & HASH_EARLY)
5949 table = memblock_virt_alloc_nopanic(size, 0);
5951 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5954 * If bucketsize is not a power-of-two, we may free
5955 * some pages at the end of hash table which
5956 * alloc_pages_exact() automatically does
5958 if (get_order(size) < MAX_ORDER) {
5959 table = alloc_pages_exact(size, GFP_ATOMIC);
5960 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5963 } while (!table && size > PAGE_SIZE && --log2qty);
5966 panic("Failed to allocate %s hash table\n", tablename);
5968 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5971 ilog2(size) - PAGE_SHIFT,
5975 *_hash_shift = log2qty;
5977 *_hash_mask = (1 << log2qty) - 1;
5982 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5983 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5986 #ifdef CONFIG_SPARSEMEM
5987 return __pfn_to_section(pfn)->pageblock_flags;
5989 return zone->pageblock_flags;
5990 #endif /* CONFIG_SPARSEMEM */
5993 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5995 #ifdef CONFIG_SPARSEMEM
5996 pfn &= (PAGES_PER_SECTION-1);
5997 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5999 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6000 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6001 #endif /* CONFIG_SPARSEMEM */
6005 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
6006 * @page: The page within the block of interest
6007 * @start_bitidx: The first bit of interest to retrieve
6008 * @end_bitidx: The last bit of interest
6009 * returns pageblock_bits flags
6011 unsigned long get_pageblock_flags_group(struct page *page,
6012 int start_bitidx, int end_bitidx)
6015 unsigned long *bitmap;
6016 unsigned long pfn, bitidx;
6017 unsigned long flags = 0;
6018 unsigned long value = 1;
6020 zone = page_zone(page);
6021 pfn = page_to_pfn(page);
6022 bitmap = get_pageblock_bitmap(zone, pfn);
6023 bitidx = pfn_to_bitidx(zone, pfn);
6025 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6026 if (test_bit(bitidx + start_bitidx, bitmap))
6033 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
6034 * @page: The page within the block of interest
6035 * @start_bitidx: The first bit of interest
6036 * @end_bitidx: The last bit of interest
6037 * @flags: The flags to set
6039 void set_pageblock_flags_group(struct page *page, unsigned long flags,
6040 int start_bitidx, int end_bitidx)
6043 unsigned long *bitmap;
6044 unsigned long pfn, bitidx;
6045 unsigned long value = 1;
6047 zone = page_zone(page);
6048 pfn = page_to_pfn(page);
6049 bitmap = get_pageblock_bitmap(zone, pfn);
6050 bitidx = pfn_to_bitidx(zone, pfn);
6051 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6053 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6055 __set_bit(bitidx + start_bitidx, bitmap);
6057 __clear_bit(bitidx + start_bitidx, bitmap);
6061 * This function checks whether pageblock includes unmovable pages or not.
6062 * If @count is not zero, it is okay to include less @count unmovable pages
6064 * PageLRU check without isolation or lru_lock could race so that
6065 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6066 * expect this function should be exact.
6068 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6069 bool skip_hwpoisoned_pages)
6071 unsigned long pfn, iter, found;
6075 * For avoiding noise data, lru_add_drain_all() should be called
6076 * If ZONE_MOVABLE, the zone never contains unmovable pages
6078 if (zone_idx(zone) == ZONE_MOVABLE)
6080 mt = get_pageblock_migratetype(page);
6081 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6084 pfn = page_to_pfn(page);
6085 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6086 unsigned long check = pfn + iter;
6088 if (!pfn_valid_within(check))
6091 page = pfn_to_page(check);
6094 * Hugepages are not in LRU lists, but they're movable.
6095 * We need not scan over tail pages bacause we don't
6096 * handle each tail page individually in migration.
6098 if (PageHuge(page)) {
6099 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6104 * We can't use page_count without pin a page
6105 * because another CPU can free compound page.
6106 * This check already skips compound tails of THP
6107 * because their page->_count is zero at all time.
6109 if (!atomic_read(&page->_count)) {
6110 if (PageBuddy(page))
6111 iter += (1 << page_order(page)) - 1;
6116 * The HWPoisoned page may be not in buddy system, and
6117 * page_count() is not 0.
6119 if (skip_hwpoisoned_pages && PageHWPoison(page))
6125 * If there are RECLAIMABLE pages, we need to check it.
6126 * But now, memory offline itself doesn't call shrink_slab()
6127 * and it still to be fixed.
6130 * If the page is not RAM, page_count()should be 0.
6131 * we don't need more check. This is an _used_ not-movable page.
6133 * The problematic thing here is PG_reserved pages. PG_reserved
6134 * is set to both of a memory hole page and a _used_ kernel
6143 bool is_pageblock_removable_nolock(struct page *page)
6149 * We have to be careful here because we are iterating over memory
6150 * sections which are not zone aware so we might end up outside of
6151 * the zone but still within the section.
6152 * We have to take care about the node as well. If the node is offline
6153 * its NODE_DATA will be NULL - see page_zone.
6155 if (!node_online(page_to_nid(page)))
6158 zone = page_zone(page);
6159 pfn = page_to_pfn(page);
6160 if (!zone_spans_pfn(zone, pfn))
6163 return !has_unmovable_pages(zone, page, 0, true);
6168 static unsigned long pfn_max_align_down(unsigned long pfn)
6170 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6171 pageblock_nr_pages) - 1);
6174 static unsigned long pfn_max_align_up(unsigned long pfn)
6176 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6177 pageblock_nr_pages));
6180 /* [start, end) must belong to a single zone. */
6181 static int __alloc_contig_migrate_range(struct compact_control *cc,
6182 unsigned long start, unsigned long end)
6184 /* This function is based on compact_zone() from compaction.c. */
6185 unsigned long nr_reclaimed;
6186 unsigned long pfn = start;
6187 unsigned int tries = 0;
6192 while (pfn < end || !list_empty(&cc->migratepages)) {
6193 if (fatal_signal_pending(current)) {
6198 if (list_empty(&cc->migratepages)) {
6199 cc->nr_migratepages = 0;
6200 pfn = isolate_migratepages_range(cc->zone, cc,
6207 } else if (++tries == 5) {
6208 ret = ret < 0 ? ret : -EBUSY;
6212 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6214 cc->nr_migratepages -= nr_reclaimed;
6216 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6217 0, MIGRATE_SYNC, MR_CMA);
6220 putback_movable_pages(&cc->migratepages);
6227 * alloc_contig_range() -- tries to allocate given range of pages
6228 * @start: start PFN to allocate
6229 * @end: one-past-the-last PFN to allocate
6230 * @migratetype: migratetype of the underlaying pageblocks (either
6231 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6232 * in range must have the same migratetype and it must
6233 * be either of the two.
6235 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6236 * aligned, however it's the caller's responsibility to guarantee that
6237 * we are the only thread that changes migrate type of pageblocks the
6240 * The PFN range must belong to a single zone.
6242 * Returns zero on success or negative error code. On success all
6243 * pages which PFN is in [start, end) are allocated for the caller and
6244 * need to be freed with free_contig_range().
6246 int alloc_contig_range(unsigned long start, unsigned long end,
6247 unsigned migratetype)
6249 unsigned long outer_start, outer_end;
6252 struct compact_control cc = {
6253 .nr_migratepages = 0,
6255 .zone = page_zone(pfn_to_page(start)),
6257 .ignore_skip_hint = true,
6259 INIT_LIST_HEAD(&cc.migratepages);
6262 * What we do here is we mark all pageblocks in range as
6263 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6264 * have different sizes, and due to the way page allocator
6265 * work, we align the range to biggest of the two pages so
6266 * that page allocator won't try to merge buddies from
6267 * different pageblocks and change MIGRATE_ISOLATE to some
6268 * other migration type.
6270 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6271 * migrate the pages from an unaligned range (ie. pages that
6272 * we are interested in). This will put all the pages in
6273 * range back to page allocator as MIGRATE_ISOLATE.
6275 * When this is done, we take the pages in range from page
6276 * allocator removing them from the buddy system. This way
6277 * page allocator will never consider using them.
6279 * This lets us mark the pageblocks back as
6280 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6281 * aligned range but not in the unaligned, original range are
6282 * put back to page allocator so that buddy can use them.
6285 ret = start_isolate_page_range(pfn_max_align_down(start),
6286 pfn_max_align_up(end), migratetype,
6291 ret = __alloc_contig_migrate_range(&cc, start, end);
6296 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6297 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6298 * more, all pages in [start, end) are free in page allocator.
6299 * What we are going to do is to allocate all pages from
6300 * [start, end) (that is remove them from page allocator).
6302 * The only problem is that pages at the beginning and at the
6303 * end of interesting range may be not aligned with pages that
6304 * page allocator holds, ie. they can be part of higher order
6305 * pages. Because of this, we reserve the bigger range and
6306 * once this is done free the pages we are not interested in.
6308 * We don't have to hold zone->lock here because the pages are
6309 * isolated thus they won't get removed from buddy.
6312 lru_add_drain_all();
6316 outer_start = start;
6317 while (!PageBuddy(pfn_to_page(outer_start))) {
6318 if (++order >= MAX_ORDER) {
6322 outer_start &= ~0UL << order;
6325 /* Make sure the range is really isolated. */
6326 if (test_pages_isolated(outer_start, end, false)) {
6327 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6334 /* Grab isolated pages from freelists. */
6335 outer_end = isolate_freepages_range(&cc, outer_start, end);
6341 /* Free head and tail (if any) */
6342 if (start != outer_start)
6343 free_contig_range(outer_start, start - outer_start);
6344 if (end != outer_end)
6345 free_contig_range(end, outer_end - end);
6348 undo_isolate_page_range(pfn_max_align_down(start),
6349 pfn_max_align_up(end), migratetype);
6353 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6355 unsigned int count = 0;
6357 for (; nr_pages--; pfn++) {
6358 struct page *page = pfn_to_page(pfn);
6360 count += page_count(page) != 1;
6363 WARN(count != 0, "%d pages are still in use!\n", count);
6367 #ifdef CONFIG_MEMORY_HOTPLUG
6369 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6370 * page high values need to be recalulated.
6372 void __meminit zone_pcp_update(struct zone *zone)
6375 mutex_lock(&pcp_batch_high_lock);
6376 for_each_possible_cpu(cpu)
6377 pageset_set_high_and_batch(zone,
6378 per_cpu_ptr(zone->pageset, cpu));
6379 mutex_unlock(&pcp_batch_high_lock);
6383 void zone_pcp_reset(struct zone *zone)
6385 unsigned long flags;
6387 struct per_cpu_pageset *pset;
6389 /* avoid races with drain_pages() */
6390 local_irq_save(flags);
6391 if (zone->pageset != &boot_pageset) {
6392 for_each_online_cpu(cpu) {
6393 pset = per_cpu_ptr(zone->pageset, cpu);
6394 drain_zonestat(zone, pset);
6396 free_percpu(zone->pageset);
6397 zone->pageset = &boot_pageset;
6399 local_irq_restore(flags);
6402 #ifdef CONFIG_MEMORY_HOTREMOVE
6404 * All pages in the range must be isolated before calling this.
6407 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6413 unsigned long flags;
6414 /* find the first valid pfn */
6415 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6420 zone = page_zone(pfn_to_page(pfn));
6421 spin_lock_irqsave(&zone->lock, flags);
6423 while (pfn < end_pfn) {
6424 if (!pfn_valid(pfn)) {
6428 page = pfn_to_page(pfn);
6430 * The HWPoisoned page may be not in buddy system, and
6431 * page_count() is not 0.
6433 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6435 SetPageReserved(page);
6439 BUG_ON(page_count(page));
6440 BUG_ON(!PageBuddy(page));
6441 order = page_order(page);
6442 #ifdef CONFIG_DEBUG_VM
6443 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6444 pfn, 1 << order, end_pfn);
6446 list_del(&page->lru);
6447 rmv_page_order(page);
6448 zone->free_area[order].nr_free--;
6449 for (i = 0; i < (1 << order); i++)
6450 SetPageReserved((page+i));
6451 pfn += (1 << order);
6453 spin_unlock_irqrestore(&zone->lock, flags);
6457 #ifdef CONFIG_MEMORY_FAILURE
6458 bool is_free_buddy_page(struct page *page)
6460 struct zone *zone = page_zone(page);
6461 unsigned long pfn = page_to_pfn(page);
6462 unsigned long flags;
6465 spin_lock_irqsave(&zone->lock, flags);
6466 for (order = 0; order < MAX_ORDER; order++) {
6467 struct page *page_head = page - (pfn & ((1 << order) - 1));
6469 if (PageBuddy(page_head) && page_order(page_head) >= order)
6472 spin_unlock_irqrestore(&zone->lock, flags);
6474 return order < MAX_ORDER;
6478 static const struct trace_print_flags pageflag_names[] = {
6479 {1UL << PG_locked, "locked" },
6480 {1UL << PG_error, "error" },
6481 {1UL << PG_referenced, "referenced" },
6482 {1UL << PG_uptodate, "uptodate" },
6483 {1UL << PG_dirty, "dirty" },
6484 {1UL << PG_lru, "lru" },
6485 {1UL << PG_active, "active" },
6486 {1UL << PG_slab, "slab" },
6487 {1UL << PG_owner_priv_1, "owner_priv_1" },
6488 {1UL << PG_arch_1, "arch_1" },
6489 {1UL << PG_reserved, "reserved" },
6490 {1UL << PG_private, "private" },
6491 {1UL << PG_private_2, "private_2" },
6492 {1UL << PG_writeback, "writeback" },
6493 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6494 {1UL << PG_head, "head" },
6495 {1UL << PG_tail, "tail" },
6497 {1UL << PG_compound, "compound" },
6499 {1UL << PG_swapcache, "swapcache" },
6500 {1UL << PG_mappedtodisk, "mappedtodisk" },
6501 {1UL << PG_reclaim, "reclaim" },
6502 {1UL << PG_swapbacked, "swapbacked" },
6503 {1UL << PG_unevictable, "unevictable" },
6505 {1UL << PG_mlocked, "mlocked" },
6507 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6508 {1UL << PG_uncached, "uncached" },
6510 #ifdef CONFIG_MEMORY_FAILURE
6511 {1UL << PG_hwpoison, "hwpoison" },
6513 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6514 {1UL << PG_compound_lock, "compound_lock" },
6518 static void dump_page_flags(unsigned long flags)
6520 const char *delim = "";
6524 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6526 printk(KERN_ALERT "page flags: %#lx(", flags);
6528 /* remove zone id */
6529 flags &= (1UL << NR_PAGEFLAGS) - 1;
6531 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6533 mask = pageflag_names[i].mask;
6534 if ((flags & mask) != mask)
6538 printk("%s%s", delim, pageflag_names[i].name);
6542 /* check for left over flags */
6544 printk("%s%#lx", delim, flags);
6549 void dump_page_badflags(struct page *page, const char *reason,
6550 unsigned long badflags)
6553 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6554 page, atomic_read(&page->_count), page_mapcount(page),
6555 page->mapping, page->index);
6556 dump_page_flags(page->flags);
6558 pr_alert("page dumped because: %s\n", reason);
6559 if (page->flags & badflags) {
6560 pr_alert("bad because of flags:\n");
6561 dump_page_flags(page->flags & badflags);
6563 mem_cgroup_print_bad_page(page);
6566 void dump_page(struct page *page, const char *reason)
6568 dump_page_badflags(page, reason, 0);
6570 EXPORT_SYMBOL_GPL(dump_page);