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);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node);
76 EXPORT_PER_CPU_SYMBOL(numa_node);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
91 * Array of node states.
93 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
94 [N_POSSIBLE] = NODE_MASK_ALL,
95 [N_ONLINE] = { { [0] = 1UL } },
97 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
99 [N_HIGH_MEMORY] = { { [0] = 1UL } },
101 #ifdef CONFIG_MOVABLE_NODE
102 [N_MEMORY] = { { [0] = 1UL } },
104 [N_CPU] = { { [0] = 1UL } },
107 EXPORT_SYMBOL(node_states);
109 /* Protect totalram_pages and zone->managed_pages */
110 static DEFINE_SPINLOCK(managed_page_count_lock);
112 unsigned long totalram_pages __read_mostly;
113 unsigned long totalreserve_pages __read_mostly;
115 * When calculating the number of globally allowed dirty pages, there
116 * is a certain number of per-zone reserves that should not be
117 * considered dirtyable memory. This is the sum of those reserves
118 * over all existing zones that contribute dirtyable memory.
120 unsigned long dirty_balance_reserve __read_mostly;
122 int percpu_pagelist_fraction;
123 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
125 #ifdef CONFIG_PM_SLEEP
127 * The following functions are used by the suspend/hibernate code to temporarily
128 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
129 * while devices are suspended. To avoid races with the suspend/hibernate code,
130 * they should always be called with pm_mutex held (gfp_allowed_mask also should
131 * only be modified with pm_mutex held, unless the suspend/hibernate code is
132 * guaranteed not to run in parallel with that modification).
135 static gfp_t saved_gfp_mask;
137 void pm_restore_gfp_mask(void)
139 WARN_ON(!mutex_is_locked(&pm_mutex));
140 if (saved_gfp_mask) {
141 gfp_allowed_mask = saved_gfp_mask;
146 void pm_restrict_gfp_mask(void)
148 WARN_ON(!mutex_is_locked(&pm_mutex));
149 WARN_ON(saved_gfp_mask);
150 saved_gfp_mask = gfp_allowed_mask;
151 gfp_allowed_mask &= ~GFP_IOFS;
154 bool pm_suspended_storage(void)
156 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
160 #endif /* CONFIG_PM_SLEEP */
162 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
163 int pageblock_order __read_mostly;
166 static void __free_pages_ok(struct page *page, unsigned int order);
169 * results with 256, 32 in the lowmem_reserve sysctl:
170 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
171 * 1G machine -> (16M dma, 784M normal, 224M high)
172 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
173 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
174 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
176 * TBD: should special case ZONE_DMA32 machines here - in those we normally
177 * don't need any ZONE_NORMAL reservation
179 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
180 #ifdef CONFIG_ZONE_DMA
183 #ifdef CONFIG_ZONE_DMA32
186 #ifdef CONFIG_HIGHMEM
192 EXPORT_SYMBOL(totalram_pages);
194 static char * const zone_names[MAX_NR_ZONES] = {
195 #ifdef CONFIG_ZONE_DMA
198 #ifdef CONFIG_ZONE_DMA32
202 #ifdef CONFIG_HIGHMEM
208 int min_free_kbytes = 1024;
209 int user_min_free_kbytes = -1;
211 static unsigned long __meminitdata nr_kernel_pages;
212 static unsigned long __meminitdata nr_all_pages;
213 static unsigned long __meminitdata dma_reserve;
215 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
216 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
217 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
218 static unsigned long __initdata required_kernelcore;
219 static unsigned long __initdata required_movablecore;
220 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
222 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
224 EXPORT_SYMBOL(movable_zone);
225 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
228 int nr_node_ids __read_mostly = MAX_NUMNODES;
229 int nr_online_nodes __read_mostly = 1;
230 EXPORT_SYMBOL(nr_node_ids);
231 EXPORT_SYMBOL(nr_online_nodes);
234 int page_group_by_mobility_disabled __read_mostly;
236 void set_pageblock_migratetype(struct page *page, int migratetype)
238 if (unlikely(page_group_by_mobility_disabled &&
239 migratetype < MIGRATE_PCPTYPES))
240 migratetype = MIGRATE_UNMOVABLE;
242 set_pageblock_flags_group(page, (unsigned long)migratetype,
243 PB_migrate, PB_migrate_end);
246 bool oom_killer_disabled __read_mostly;
248 #ifdef CONFIG_DEBUG_VM
249 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
253 unsigned long pfn = page_to_pfn(page);
254 unsigned long sp, start_pfn;
257 seq = zone_span_seqbegin(zone);
258 start_pfn = zone->zone_start_pfn;
259 sp = zone->spanned_pages;
260 if (!zone_spans_pfn(zone, pfn))
262 } while (zone_span_seqretry(zone, seq));
265 pr_err("page %lu outside zone [ %lu - %lu ]\n",
266 pfn, start_pfn, start_pfn + sp);
271 static int page_is_consistent(struct zone *zone, struct page *page)
273 if (!pfn_valid_within(page_to_pfn(page)))
275 if (zone != page_zone(page))
281 * Temporary debugging check for pages not lying within a given zone.
283 static int bad_range(struct zone *zone, struct page *page)
285 if (page_outside_zone_boundaries(zone, page))
287 if (!page_is_consistent(zone, page))
293 static inline int bad_range(struct zone *zone, struct page *page)
299 static void bad_page(struct page *page, char *reason, 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 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_pageblock_migratetype(page, MIGRATE_CMA);
804 if (pageblock_order >= MAX_ORDER) {
805 i = pageblock_nr_pages;
808 set_page_refcounted(p);
809 __free_pages(p, MAX_ORDER - 1);
810 p += MAX_ORDER_NR_PAGES;
811 } while (i -= MAX_ORDER_NR_PAGES);
813 set_page_refcounted(page);
814 __free_pages(page, pageblock_order);
817 adjust_managed_page_count(page, pageblock_nr_pages);
822 * The order of subdivision here is critical for the IO subsystem.
823 * Please do not alter this order without good reasons and regression
824 * testing. Specifically, as large blocks of memory are subdivided,
825 * the order in which smaller blocks are delivered depends on the order
826 * they're subdivided in this function. This is the primary factor
827 * influencing the order in which pages are delivered to the IO
828 * subsystem according to empirical testing, and this is also justified
829 * by considering the behavior of a buddy system containing a single
830 * large block of memory acted on by a series of small allocations.
831 * This behavior is a critical factor in sglist merging's success.
835 static inline void expand(struct zone *zone, struct page *page,
836 int low, int high, struct free_area *area,
839 unsigned long size = 1 << high;
845 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
847 #ifdef CONFIG_DEBUG_PAGEALLOC
848 if (high < debug_guardpage_minorder()) {
850 * Mark as guard pages (or page), that will allow to
851 * merge back to allocator when buddy will be freed.
852 * Corresponding page table entries will not be touched,
853 * pages will stay not present in virtual address space
855 INIT_LIST_HEAD(&page[size].lru);
856 set_page_guard_flag(&page[size]);
857 set_page_private(&page[size], high);
858 /* Guard pages are not available for any usage */
859 __mod_zone_freepage_state(zone, -(1 << high),
864 list_add(&page[size].lru, &area->free_list[migratetype]);
866 set_page_order(&page[size], high);
871 * This page is about to be returned from the page allocator
873 static inline int check_new_page(struct page *page)
875 char *bad_reason = NULL;
876 unsigned long bad_flags = 0;
878 if (unlikely(page_mapcount(page)))
879 bad_reason = "nonzero mapcount";
880 if (unlikely(page->mapping != NULL))
881 bad_reason = "non-NULL mapping";
882 if (unlikely(atomic_read(&page->_count) != 0))
883 bad_reason = "nonzero _count";
884 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
885 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
886 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
888 if (unlikely(mem_cgroup_bad_page_check(page)))
889 bad_reason = "cgroup check failed";
890 if (unlikely(bad_reason)) {
891 bad_page(page, bad_reason, bad_flags);
897 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
901 for (i = 0; i < (1 << order); i++) {
902 struct page *p = page + i;
903 if (unlikely(check_new_page(p)))
907 set_page_private(page, 0);
908 set_page_refcounted(page);
910 arch_alloc_page(page, order);
911 kernel_map_pages(page, 1 << order, 1);
913 if (gfp_flags & __GFP_ZERO)
914 prep_zero_page(page, order, gfp_flags);
916 if (order && (gfp_flags & __GFP_COMP))
917 prep_compound_page(page, order);
923 * Go through the free lists for the given migratetype and remove
924 * the smallest available page from the freelists
927 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
930 unsigned int current_order;
931 struct free_area *area;
934 /* Find a page of the appropriate size in the preferred list */
935 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
936 area = &(zone->free_area[current_order]);
937 if (list_empty(&area->free_list[migratetype]))
940 page = list_entry(area->free_list[migratetype].next,
942 list_del(&page->lru);
943 rmv_page_order(page);
945 expand(zone, page, order, current_order, area, migratetype);
946 set_freepage_migratetype(page, migratetype);
955 * This array describes the order lists are fallen back to when
956 * the free lists for the desirable migrate type are depleted
958 static int fallbacks[MIGRATE_TYPES][4] = {
959 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
960 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
962 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
963 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
965 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
967 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
968 #ifdef CONFIG_MEMORY_ISOLATION
969 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
974 * Move the free pages in a range to the free lists of the requested type.
975 * Note that start_page and end_pages are not aligned on a pageblock
976 * boundary. If alignment is required, use move_freepages_block()
978 int move_freepages(struct zone *zone,
979 struct page *start_page, struct page *end_page,
986 #ifndef CONFIG_HOLES_IN_ZONE
988 * page_zone is not safe to call in this context when
989 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
990 * anyway as we check zone boundaries in move_freepages_block().
991 * Remove at a later date when no bug reports exist related to
992 * grouping pages by mobility
994 BUG_ON(page_zone(start_page) != page_zone(end_page));
997 for (page = start_page; page <= end_page;) {
998 /* Make sure we are not inadvertently changing nodes */
999 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1001 if (!pfn_valid_within(page_to_pfn(page))) {
1006 if (!PageBuddy(page)) {
1011 order = page_order(page);
1012 list_move(&page->lru,
1013 &zone->free_area[order].free_list[migratetype]);
1014 set_freepage_migratetype(page, migratetype);
1016 pages_moved += 1 << order;
1022 int move_freepages_block(struct zone *zone, struct page *page,
1025 unsigned long start_pfn, end_pfn;
1026 struct page *start_page, *end_page;
1028 start_pfn = page_to_pfn(page);
1029 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1030 start_page = pfn_to_page(start_pfn);
1031 end_page = start_page + pageblock_nr_pages - 1;
1032 end_pfn = start_pfn + pageblock_nr_pages - 1;
1034 /* Do not cross zone boundaries */
1035 if (!zone_spans_pfn(zone, start_pfn))
1037 if (!zone_spans_pfn(zone, end_pfn))
1040 return move_freepages(zone, start_page, end_page, migratetype);
1043 static void change_pageblock_range(struct page *pageblock_page,
1044 int start_order, int migratetype)
1046 int nr_pageblocks = 1 << (start_order - pageblock_order);
1048 while (nr_pageblocks--) {
1049 set_pageblock_migratetype(pageblock_page, migratetype);
1050 pageblock_page += pageblock_nr_pages;
1055 * If breaking a large block of pages, move all free pages to the preferred
1056 * allocation list. If falling back for a reclaimable kernel allocation, be
1057 * more aggressive about taking ownership of free pages.
1059 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1060 * nor move CMA pages to different free lists. We don't want unmovable pages
1061 * to be allocated from MIGRATE_CMA areas.
1063 * Returns the new migratetype of the pageblock (or the same old migratetype
1064 * if it was unchanged).
1066 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1067 int start_type, int fallback_type)
1069 int current_order = page_order(page);
1072 * When borrowing from MIGRATE_CMA, we need to release the excess
1073 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1074 * is set to CMA so it is returned to the correct freelist in case
1075 * the page ends up being not actually allocated from the pcp lists.
1077 if (is_migrate_cma(fallback_type))
1078 return fallback_type;
1080 /* Take ownership for orders >= pageblock_order */
1081 if (current_order >= pageblock_order) {
1082 change_pageblock_range(page, current_order, start_type);
1086 if (current_order >= pageblock_order / 2 ||
1087 start_type == MIGRATE_RECLAIMABLE ||
1088 page_group_by_mobility_disabled) {
1091 pages = move_freepages_block(zone, page, start_type);
1093 /* Claim the whole block if over half of it is free */
1094 if (pages >= (1 << (pageblock_order-1)) ||
1095 page_group_by_mobility_disabled) {
1097 set_pageblock_migratetype(page, start_type);
1103 return fallback_type;
1106 /* Remove an element from the buddy allocator from the fallback list */
1107 static inline struct page *
1108 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1110 struct free_area *area;
1113 int migratetype, new_type, i;
1115 /* Find the largest possible block of pages in the other list */
1116 for (current_order = MAX_ORDER-1; current_order >= order;
1119 migratetype = fallbacks[start_migratetype][i];
1121 /* MIGRATE_RESERVE handled later if necessary */
1122 if (migratetype == MIGRATE_RESERVE)
1125 area = &(zone->free_area[current_order]);
1126 if (list_empty(&area->free_list[migratetype]))
1129 page = list_entry(area->free_list[migratetype].next,
1133 new_type = try_to_steal_freepages(zone, page,
1137 /* Remove the page from the freelists */
1138 list_del(&page->lru);
1139 rmv_page_order(page);
1141 expand(zone, page, order, current_order, area,
1143 /* The freepage_migratetype may differ from pageblock's
1144 * migratetype depending on the decisions in
1145 * try_to_steal_freepages. This is OK as long as it does
1146 * not differ for MIGRATE_CMA type.
1148 set_freepage_migratetype(page, new_type);
1150 trace_mm_page_alloc_extfrag(page, order, current_order,
1151 start_migratetype, migratetype, new_type);
1161 * Do the hard work of removing an element from the buddy allocator.
1162 * Call me with the zone->lock already held.
1164 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1170 page = __rmqueue_smallest(zone, order, migratetype);
1172 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1173 page = __rmqueue_fallback(zone, order, migratetype);
1176 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1177 * is used because __rmqueue_smallest is an inline function
1178 * and we want just one call site
1181 migratetype = MIGRATE_RESERVE;
1186 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1191 * Obtain a specified number of elements from the buddy allocator, all under
1192 * a single hold of the lock, for efficiency. Add them to the supplied list.
1193 * Returns the number of new pages which were placed at *list.
1195 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1196 unsigned long count, struct list_head *list,
1197 int migratetype, int cold)
1201 spin_lock(&zone->lock);
1202 for (i = 0; i < count; ++i) {
1203 struct page *page = __rmqueue(zone, order, migratetype);
1204 if (unlikely(page == NULL))
1208 * Split buddy pages returned by expand() are received here
1209 * in physical page order. The page is added to the callers and
1210 * list and the list head then moves forward. From the callers
1211 * perspective, the linked list is ordered by page number in
1212 * some conditions. This is useful for IO devices that can
1213 * merge IO requests if the physical pages are ordered
1216 if (likely(cold == 0))
1217 list_add(&page->lru, list);
1219 list_add_tail(&page->lru, list);
1221 if (is_migrate_cma(get_freepage_migratetype(page)))
1222 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1225 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1226 spin_unlock(&zone->lock);
1232 * Called from the vmstat counter updater to drain pagesets of this
1233 * currently executing processor on remote nodes after they have
1236 * Note that this function must be called with the thread pinned to
1237 * a single processor.
1239 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1241 unsigned long flags;
1243 unsigned long batch;
1245 local_irq_save(flags);
1246 batch = ACCESS_ONCE(pcp->batch);
1247 if (pcp->count >= batch)
1250 to_drain = pcp->count;
1252 free_pcppages_bulk(zone, to_drain, pcp);
1253 pcp->count -= to_drain;
1255 local_irq_restore(flags);
1260 * Drain pages of the indicated processor.
1262 * The processor must either be the current processor and the
1263 * thread pinned to the current processor or a processor that
1266 static void drain_pages(unsigned int cpu)
1268 unsigned long flags;
1271 for_each_populated_zone(zone) {
1272 struct per_cpu_pageset *pset;
1273 struct per_cpu_pages *pcp;
1275 local_irq_save(flags);
1276 pset = per_cpu_ptr(zone->pageset, cpu);
1280 free_pcppages_bulk(zone, pcp->count, pcp);
1283 local_irq_restore(flags);
1288 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1290 void drain_local_pages(void *arg)
1292 drain_pages(smp_processor_id());
1296 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1298 * Note that this code is protected against sending an IPI to an offline
1299 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1300 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1301 * nothing keeps CPUs from showing up after we populated the cpumask and
1302 * before the call to on_each_cpu_mask().
1304 void drain_all_pages(void)
1307 struct per_cpu_pageset *pcp;
1311 * Allocate in the BSS so we wont require allocation in
1312 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1314 static cpumask_t cpus_with_pcps;
1317 * We don't care about racing with CPU hotplug event
1318 * as offline notification will cause the notified
1319 * cpu to drain that CPU pcps and on_each_cpu_mask
1320 * disables preemption as part of its processing
1322 for_each_online_cpu(cpu) {
1323 bool has_pcps = false;
1324 for_each_populated_zone(zone) {
1325 pcp = per_cpu_ptr(zone->pageset, cpu);
1326 if (pcp->pcp.count) {
1332 cpumask_set_cpu(cpu, &cpus_with_pcps);
1334 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1336 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1339 #ifdef CONFIG_HIBERNATION
1341 void mark_free_pages(struct zone *zone)
1343 unsigned long pfn, max_zone_pfn;
1344 unsigned long flags;
1346 struct list_head *curr;
1348 if (zone_is_empty(zone))
1351 spin_lock_irqsave(&zone->lock, flags);
1353 max_zone_pfn = zone_end_pfn(zone);
1354 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1355 if (pfn_valid(pfn)) {
1356 struct page *page = pfn_to_page(pfn);
1358 if (!swsusp_page_is_forbidden(page))
1359 swsusp_unset_page_free(page);
1362 for_each_migratetype_order(order, t) {
1363 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1366 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1367 for (i = 0; i < (1UL << order); i++)
1368 swsusp_set_page_free(pfn_to_page(pfn + i));
1371 spin_unlock_irqrestore(&zone->lock, flags);
1373 #endif /* CONFIG_PM */
1376 * Free a 0-order page
1377 * cold == 1 ? free a cold page : free a hot page
1379 void free_hot_cold_page(struct page *page, int cold)
1381 struct zone *zone = page_zone(page);
1382 struct per_cpu_pages *pcp;
1383 unsigned long flags;
1386 if (!free_pages_prepare(page, 0))
1389 migratetype = get_pageblock_migratetype(page);
1390 set_freepage_migratetype(page, migratetype);
1391 local_irq_save(flags);
1392 __count_vm_event(PGFREE);
1395 * We only track unmovable, reclaimable and movable on pcp lists.
1396 * Free ISOLATE pages back to the allocator because they are being
1397 * offlined but treat RESERVE as movable pages so we can get those
1398 * areas back if necessary. Otherwise, we may have to free
1399 * excessively into the page allocator
1401 if (migratetype >= MIGRATE_PCPTYPES) {
1402 if (unlikely(is_migrate_isolate(migratetype))) {
1403 free_one_page(zone, page, 0, migratetype);
1406 migratetype = MIGRATE_MOVABLE;
1409 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1411 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1413 list_add(&page->lru, &pcp->lists[migratetype]);
1415 if (pcp->count >= pcp->high) {
1416 unsigned long batch = ACCESS_ONCE(pcp->batch);
1417 free_pcppages_bulk(zone, batch, pcp);
1418 pcp->count -= batch;
1422 local_irq_restore(flags);
1426 * Free a list of 0-order pages
1428 void free_hot_cold_page_list(struct list_head *list, int cold)
1430 struct page *page, *next;
1432 list_for_each_entry_safe(page, next, list, lru) {
1433 trace_mm_page_free_batched(page, cold);
1434 free_hot_cold_page(page, cold);
1439 * split_page takes a non-compound higher-order page, and splits it into
1440 * n (1<<order) sub-pages: page[0..n]
1441 * Each sub-page must be freed individually.
1443 * Note: this is probably too low level an operation for use in drivers.
1444 * Please consult with lkml before using this in your driver.
1446 void split_page(struct page *page, unsigned int order)
1450 VM_BUG_ON_PAGE(PageCompound(page), page);
1451 VM_BUG_ON_PAGE(!page_count(page), page);
1453 #ifdef CONFIG_KMEMCHECK
1455 * Split shadow pages too, because free(page[0]) would
1456 * otherwise free the whole shadow.
1458 if (kmemcheck_page_is_tracked(page))
1459 split_page(virt_to_page(page[0].shadow), order);
1462 for (i = 1; i < (1 << order); i++)
1463 set_page_refcounted(page + i);
1465 EXPORT_SYMBOL_GPL(split_page);
1467 static int __isolate_free_page(struct page *page, unsigned int order)
1469 unsigned long watermark;
1473 BUG_ON(!PageBuddy(page));
1475 zone = page_zone(page);
1476 mt = get_pageblock_migratetype(page);
1478 if (!is_migrate_isolate(mt)) {
1479 /* Obey watermarks as if the page was being allocated */
1480 watermark = low_wmark_pages(zone) + (1 << order);
1481 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1484 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1487 /* Remove page from free list */
1488 list_del(&page->lru);
1489 zone->free_area[order].nr_free--;
1490 rmv_page_order(page);
1492 /* Set the pageblock if the isolated page is at least a pageblock */
1493 if (order >= pageblock_order - 1) {
1494 struct page *endpage = page + (1 << order) - 1;
1495 for (; page < endpage; page += pageblock_nr_pages) {
1496 int mt = get_pageblock_migratetype(page);
1497 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1498 set_pageblock_migratetype(page,
1503 return 1UL << order;
1507 * Similar to split_page except the page is already free. As this is only
1508 * being used for migration, the migratetype of the block also changes.
1509 * As this is called with interrupts disabled, the caller is responsible
1510 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1513 * Note: this is probably too low level an operation for use in drivers.
1514 * Please consult with lkml before using this in your driver.
1516 int split_free_page(struct page *page)
1521 order = page_order(page);
1523 nr_pages = __isolate_free_page(page, order);
1527 /* Split into individual pages */
1528 set_page_refcounted(page);
1529 split_page(page, order);
1534 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1535 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1539 struct page *buffered_rmqueue(struct zone *preferred_zone,
1540 struct zone *zone, int order, gfp_t gfp_flags,
1543 unsigned long flags;
1545 int cold = !!(gfp_flags & __GFP_COLD);
1548 if (likely(order == 0)) {
1549 struct per_cpu_pages *pcp;
1550 struct list_head *list;
1552 local_irq_save(flags);
1553 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1554 list = &pcp->lists[migratetype];
1555 if (list_empty(list)) {
1556 pcp->count += rmqueue_bulk(zone, 0,
1559 if (unlikely(list_empty(list)))
1564 page = list_entry(list->prev, struct page, lru);
1566 page = list_entry(list->next, struct page, lru);
1568 list_del(&page->lru);
1571 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1573 * __GFP_NOFAIL is not to be used in new code.
1575 * All __GFP_NOFAIL callers should be fixed so that they
1576 * properly detect and handle allocation failures.
1578 * We most definitely don't want callers attempting to
1579 * allocate greater than order-1 page units with
1582 WARN_ON_ONCE(order > 1);
1584 spin_lock_irqsave(&zone->lock, flags);
1585 page = __rmqueue(zone, order, migratetype);
1586 spin_unlock(&zone->lock);
1589 __mod_zone_freepage_state(zone, -(1 << order),
1590 get_freepage_migratetype(page));
1593 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1595 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1596 zone_statistics(preferred_zone, zone, gfp_flags);
1597 local_irq_restore(flags);
1599 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1600 if (prep_new_page(page, order, gfp_flags))
1605 local_irq_restore(flags);
1609 #ifdef CONFIG_FAIL_PAGE_ALLOC
1612 struct fault_attr attr;
1614 u32 ignore_gfp_highmem;
1615 u32 ignore_gfp_wait;
1617 } fail_page_alloc = {
1618 .attr = FAULT_ATTR_INITIALIZER,
1619 .ignore_gfp_wait = 1,
1620 .ignore_gfp_highmem = 1,
1624 static int __init setup_fail_page_alloc(char *str)
1626 return setup_fault_attr(&fail_page_alloc.attr, str);
1628 __setup("fail_page_alloc=", setup_fail_page_alloc);
1630 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1632 if (order < fail_page_alloc.min_order)
1634 if (gfp_mask & __GFP_NOFAIL)
1636 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1638 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1641 return should_fail(&fail_page_alloc.attr, 1 << order);
1644 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1646 static int __init fail_page_alloc_debugfs(void)
1648 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1651 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1652 &fail_page_alloc.attr);
1654 return PTR_ERR(dir);
1656 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1657 &fail_page_alloc.ignore_gfp_wait))
1659 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1660 &fail_page_alloc.ignore_gfp_highmem))
1662 if (!debugfs_create_u32("min-order", mode, dir,
1663 &fail_page_alloc.min_order))
1668 debugfs_remove_recursive(dir);
1673 late_initcall(fail_page_alloc_debugfs);
1675 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1677 #else /* CONFIG_FAIL_PAGE_ALLOC */
1679 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1684 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1687 * Return true if free pages are above 'mark'. This takes into account the order
1688 * of the allocation.
1690 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1691 int classzone_idx, int alloc_flags, long free_pages)
1693 /* free_pages my go negative - that's OK */
1695 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1699 free_pages -= (1 << order) - 1;
1700 if (alloc_flags & ALLOC_HIGH)
1702 if (alloc_flags & ALLOC_HARDER)
1705 /* If allocation can't use CMA areas don't use free CMA pages */
1706 if (!(alloc_flags & ALLOC_CMA))
1707 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1710 if (free_pages - free_cma <= min + lowmem_reserve)
1712 for (o = 0; o < order; o++) {
1713 /* At the next order, this order's pages become unavailable */
1714 free_pages -= z->free_area[o].nr_free << o;
1716 /* Require fewer higher order pages to be free */
1719 if (free_pages <= min)
1725 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1726 int classzone_idx, int alloc_flags)
1728 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1729 zone_page_state(z, NR_FREE_PAGES));
1732 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1733 int classzone_idx, int alloc_flags)
1735 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1737 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1738 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1740 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1746 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1747 * skip over zones that are not allowed by the cpuset, or that have
1748 * been recently (in last second) found to be nearly full. See further
1749 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1750 * that have to skip over a lot of full or unallowed zones.
1752 * If the zonelist cache is present in the passed zonelist, then
1753 * returns a pointer to the allowed node mask (either the current
1754 * tasks mems_allowed, or node_states[N_MEMORY].)
1756 * If the zonelist cache is not available for this zonelist, does
1757 * nothing and returns NULL.
1759 * If the fullzones BITMAP in the zonelist cache is stale (more than
1760 * a second since last zap'd) then we zap it out (clear its bits.)
1762 * We hold off even calling zlc_setup, until after we've checked the
1763 * first zone in the zonelist, on the theory that most allocations will
1764 * be satisfied from that first zone, so best to examine that zone as
1765 * quickly as we can.
1767 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1769 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1770 nodemask_t *allowednodes; /* zonelist_cache approximation */
1772 zlc = zonelist->zlcache_ptr;
1776 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1777 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1778 zlc->last_full_zap = jiffies;
1781 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1782 &cpuset_current_mems_allowed :
1783 &node_states[N_MEMORY];
1784 return allowednodes;
1788 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1789 * if it is worth looking at further for free memory:
1790 * 1) Check that the zone isn't thought to be full (doesn't have its
1791 * bit set in the zonelist_cache fullzones BITMAP).
1792 * 2) Check that the zones node (obtained from the zonelist_cache
1793 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1794 * Return true (non-zero) if zone is worth looking at further, or
1795 * else return false (zero) if it is not.
1797 * This check -ignores- the distinction between various watermarks,
1798 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1799 * found to be full for any variation of these watermarks, it will
1800 * be considered full for up to one second by all requests, unless
1801 * we are so low on memory on all allowed nodes that we are forced
1802 * into the second scan of the zonelist.
1804 * In the second scan we ignore this zonelist cache and exactly
1805 * apply the watermarks to all zones, even it is slower to do so.
1806 * We are low on memory in the second scan, and should leave no stone
1807 * unturned looking for a free page.
1809 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1810 nodemask_t *allowednodes)
1812 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1813 int i; /* index of *z in zonelist zones */
1814 int n; /* node that zone *z is on */
1816 zlc = zonelist->zlcache_ptr;
1820 i = z - zonelist->_zonerefs;
1823 /* This zone is worth trying if it is allowed but not full */
1824 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1828 * Given 'z' scanning a zonelist, set the corresponding bit in
1829 * zlc->fullzones, so that subsequent attempts to allocate a page
1830 * from that zone don't waste time re-examining it.
1832 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1834 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1835 int i; /* index of *z in zonelist zones */
1837 zlc = zonelist->zlcache_ptr;
1841 i = z - zonelist->_zonerefs;
1843 set_bit(i, zlc->fullzones);
1847 * clear all zones full, called after direct reclaim makes progress so that
1848 * a zone that was recently full is not skipped over for up to a second
1850 static void zlc_clear_zones_full(struct zonelist *zonelist)
1852 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1854 zlc = zonelist->zlcache_ptr;
1858 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1861 static bool zone_local(struct zone *local_zone, struct zone *zone)
1863 return local_zone->node == zone->node;
1866 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1868 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1871 static void __paginginit init_zone_allows_reclaim(int nid)
1875 for_each_node_state(i, N_MEMORY)
1876 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1877 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1879 zone_reclaim_mode = 1;
1882 #else /* CONFIG_NUMA */
1884 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1889 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1890 nodemask_t *allowednodes)
1895 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1899 static void zlc_clear_zones_full(struct zonelist *zonelist)
1903 static bool zone_local(struct zone *local_zone, struct zone *zone)
1908 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1913 static inline void init_zone_allows_reclaim(int nid)
1916 #endif /* CONFIG_NUMA */
1919 * get_page_from_freelist goes through the zonelist trying to allocate
1922 static struct page *
1923 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1924 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1925 struct zone *preferred_zone, int migratetype)
1928 struct page *page = NULL;
1931 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1932 int zlc_active = 0; /* set if using zonelist_cache */
1933 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1935 classzone_idx = zone_idx(preferred_zone);
1938 * Scan zonelist, looking for a zone with enough free.
1939 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1941 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1942 high_zoneidx, nodemask) {
1945 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1946 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1948 if ((alloc_flags & ALLOC_CPUSET) &&
1949 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1951 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1952 if (unlikely(alloc_flags & ALLOC_NO_WATERMARKS))
1955 * Distribute pages in proportion to the individual
1956 * zone size to ensure fair page aging. The zone a
1957 * page was allocated in should have no effect on the
1958 * time the page has in memory before being reclaimed.
1960 if (alloc_flags & ALLOC_FAIR) {
1961 if (!zone_local(preferred_zone, zone))
1963 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0)
1967 * When allocating a page cache page for writing, we
1968 * want to get it from a zone that is within its dirty
1969 * limit, such that no single zone holds more than its
1970 * proportional share of globally allowed dirty pages.
1971 * The dirty limits take into account the zone's
1972 * lowmem reserves and high watermark so that kswapd
1973 * should be able to balance it without having to
1974 * write pages from its LRU list.
1976 * This may look like it could increase pressure on
1977 * lower zones by failing allocations in higher zones
1978 * before they are full. But the pages that do spill
1979 * over are limited as the lower zones are protected
1980 * by this very same mechanism. It should not become
1981 * a practical burden to them.
1983 * XXX: For now, allow allocations to potentially
1984 * exceed the per-zone dirty limit in the slowpath
1985 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1986 * which is important when on a NUMA setup the allowed
1987 * zones are together not big enough to reach the
1988 * global limit. The proper fix for these situations
1989 * will require awareness of zones in the
1990 * dirty-throttling and the flusher threads.
1992 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1993 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1994 goto this_zone_full;
1996 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1997 if (!zone_watermark_ok(zone, order, mark,
1998 classzone_idx, alloc_flags)) {
2001 if (IS_ENABLED(CONFIG_NUMA) &&
2002 !did_zlc_setup && nr_online_nodes > 1) {
2004 * we do zlc_setup if there are multiple nodes
2005 * and before considering the first zone allowed
2008 allowednodes = zlc_setup(zonelist, alloc_flags);
2013 if (zone_reclaim_mode == 0 ||
2014 !zone_allows_reclaim(preferred_zone, zone))
2015 goto this_zone_full;
2018 * As we may have just activated ZLC, check if the first
2019 * eligible zone has failed zone_reclaim recently.
2021 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2022 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2025 ret = zone_reclaim(zone, gfp_mask, order);
2027 case ZONE_RECLAIM_NOSCAN:
2030 case ZONE_RECLAIM_FULL:
2031 /* scanned but unreclaimable */
2034 /* did we reclaim enough */
2035 if (zone_watermark_ok(zone, order, mark,
2036 classzone_idx, alloc_flags))
2040 * Failed to reclaim enough to meet watermark.
2041 * Only mark the zone full if checking the min
2042 * watermark or if we failed to reclaim just
2043 * 1<<order pages or else the page allocator
2044 * fastpath will prematurely mark zones full
2045 * when the watermark is between the low and
2048 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2049 ret == ZONE_RECLAIM_SOME)
2050 goto this_zone_full;
2057 page = buffered_rmqueue(preferred_zone, zone, order,
2058 gfp_mask, migratetype);
2062 if (IS_ENABLED(CONFIG_NUMA))
2063 zlc_mark_zone_full(zonelist, z);
2066 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2067 /* Disable zlc cache for second zonelist scan */
2074 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2075 * necessary to allocate the page. The expectation is
2076 * that the caller is taking steps that will free more
2077 * memory. The caller should avoid the page being used
2078 * for !PFMEMALLOC purposes.
2080 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2086 * Large machines with many possible nodes should not always dump per-node
2087 * meminfo in irq context.
2089 static inline bool should_suppress_show_mem(void)
2094 ret = in_interrupt();
2099 static DEFINE_RATELIMIT_STATE(nopage_rs,
2100 DEFAULT_RATELIMIT_INTERVAL,
2101 DEFAULT_RATELIMIT_BURST);
2103 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2105 unsigned int filter = SHOW_MEM_FILTER_NODES;
2107 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2108 debug_guardpage_minorder() > 0)
2112 * This documents exceptions given to allocations in certain
2113 * contexts that are allowed to allocate outside current's set
2116 if (!(gfp_mask & __GFP_NOMEMALLOC))
2117 if (test_thread_flag(TIF_MEMDIE) ||
2118 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2119 filter &= ~SHOW_MEM_FILTER_NODES;
2120 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2121 filter &= ~SHOW_MEM_FILTER_NODES;
2124 struct va_format vaf;
2127 va_start(args, fmt);
2132 pr_warn("%pV", &vaf);
2137 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2138 current->comm, order, gfp_mask);
2141 if (!should_suppress_show_mem())
2146 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2147 unsigned long did_some_progress,
2148 unsigned long pages_reclaimed)
2150 /* Do not loop if specifically requested */
2151 if (gfp_mask & __GFP_NORETRY)
2154 /* Always retry if specifically requested */
2155 if (gfp_mask & __GFP_NOFAIL)
2159 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2160 * making forward progress without invoking OOM. Suspend also disables
2161 * storage devices so kswapd will not help. Bail if we are suspending.
2163 if (!did_some_progress && pm_suspended_storage())
2167 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2168 * means __GFP_NOFAIL, but that may not be true in other
2171 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2175 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2176 * specified, then we retry until we no longer reclaim any pages
2177 * (above), or we've reclaimed an order of pages at least as
2178 * large as the allocation's order. In both cases, if the
2179 * allocation still fails, we stop retrying.
2181 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2187 static inline struct page *
2188 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2189 struct zonelist *zonelist, enum zone_type high_zoneidx,
2190 nodemask_t *nodemask, struct zone *preferred_zone,
2195 /* Acquire the OOM killer lock for the zones in zonelist */
2196 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2197 schedule_timeout_uninterruptible(1);
2202 * PM-freezer should be notified that there might be an OOM killer on
2203 * its way to kill and wake somebody up. This is too early and we might
2204 * end up not killing anything but false positives are acceptable.
2205 * See freeze_processes.
2210 * Go through the zonelist yet one more time, keep very high watermark
2211 * here, this is only to catch a parallel oom killing, we must fail if
2212 * we're still under heavy pressure.
2214 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2215 order, zonelist, high_zoneidx,
2216 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2217 preferred_zone, migratetype);
2221 if (!(gfp_mask & __GFP_NOFAIL)) {
2222 /* The OOM killer will not help higher order allocs */
2223 if (order > PAGE_ALLOC_COSTLY_ORDER)
2225 /* The OOM killer does not needlessly kill tasks for lowmem */
2226 if (high_zoneidx < ZONE_NORMAL)
2229 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2230 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2231 * The caller should handle page allocation failure by itself if
2232 * it specifies __GFP_THISNODE.
2233 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2235 if (gfp_mask & __GFP_THISNODE)
2238 /* Exhausted what can be done so it's blamo time */
2239 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2242 clear_zonelist_oom(zonelist, gfp_mask);
2246 #ifdef CONFIG_COMPACTION
2247 /* Try memory compaction for high-order allocations before reclaim */
2248 static struct page *
2249 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2250 struct zonelist *zonelist, enum zone_type high_zoneidx,
2251 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2252 int migratetype, enum migrate_mode mode,
2253 bool *contended_compaction, bool *deferred_compaction,
2254 unsigned long *did_some_progress)
2259 if (compaction_deferred(preferred_zone, order)) {
2260 *deferred_compaction = true;
2264 current->flags |= PF_MEMALLOC;
2265 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2267 contended_compaction);
2268 current->flags &= ~PF_MEMALLOC;
2270 if (*did_some_progress != COMPACT_SKIPPED) {
2273 /* Page migration frees to the PCP lists but we want merging */
2274 drain_pages(get_cpu());
2277 page = get_page_from_freelist(gfp_mask, nodemask,
2278 order, zonelist, high_zoneidx,
2279 alloc_flags & ~ALLOC_NO_WATERMARKS,
2280 preferred_zone, migratetype);
2282 preferred_zone->compact_blockskip_flush = false;
2283 compaction_defer_reset(preferred_zone, order, true);
2284 count_vm_event(COMPACTSUCCESS);
2289 * It's bad if compaction run occurs and fails.
2290 * The most likely reason is that pages exist,
2291 * but not enough to satisfy watermarks.
2293 count_vm_event(COMPACTFAIL);
2296 * As async compaction considers a subset of pageblocks, only
2297 * defer if the failure was a sync compaction failure.
2299 if (mode != MIGRATE_ASYNC)
2300 defer_compaction(preferred_zone, order);
2308 static inline struct page *
2309 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2310 struct zonelist *zonelist, enum zone_type high_zoneidx,
2311 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2312 int migratetype, enum migrate_mode mode, bool *contended_compaction,
2313 bool *deferred_compaction, unsigned long *did_some_progress)
2317 #endif /* CONFIG_COMPACTION */
2319 /* Perform direct synchronous page reclaim */
2321 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2322 nodemask_t *nodemask)
2324 struct reclaim_state reclaim_state;
2329 /* We now go into synchronous reclaim */
2330 cpuset_memory_pressure_bump();
2331 current->flags |= PF_MEMALLOC;
2332 lockdep_set_current_reclaim_state(gfp_mask);
2333 reclaim_state.reclaimed_slab = 0;
2334 current->reclaim_state = &reclaim_state;
2336 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2338 current->reclaim_state = NULL;
2339 lockdep_clear_current_reclaim_state();
2340 current->flags &= ~PF_MEMALLOC;
2347 /* The really slow allocator path where we enter direct reclaim */
2348 static inline struct page *
2349 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2350 struct zonelist *zonelist, enum zone_type high_zoneidx,
2351 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2352 int migratetype, unsigned long *did_some_progress)
2354 struct page *page = NULL;
2355 bool drained = false;
2357 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2359 if (unlikely(!(*did_some_progress)))
2362 /* After successful reclaim, reconsider all zones for allocation */
2363 if (IS_ENABLED(CONFIG_NUMA))
2364 zlc_clear_zones_full(zonelist);
2367 page = get_page_from_freelist(gfp_mask, nodemask, order,
2368 zonelist, high_zoneidx,
2369 alloc_flags & ~ALLOC_NO_WATERMARKS,
2370 preferred_zone, migratetype);
2373 * If an allocation failed after direct reclaim, it could be because
2374 * pages are pinned on the per-cpu lists. Drain them and try again
2376 if (!page && !drained) {
2386 * This is called in the allocator slow-path if the allocation request is of
2387 * sufficient urgency to ignore watermarks and take other desperate measures
2389 static inline struct page *
2390 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2391 struct zonelist *zonelist, enum zone_type high_zoneidx,
2392 nodemask_t *nodemask, struct zone *preferred_zone,
2398 page = get_page_from_freelist(gfp_mask, nodemask, order,
2399 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2400 preferred_zone, migratetype);
2402 if (!page && gfp_mask & __GFP_NOFAIL)
2403 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2404 } while (!page && (gfp_mask & __GFP_NOFAIL));
2409 static void reset_alloc_batches(struct zonelist *zonelist,
2410 enum zone_type high_zoneidx,
2411 struct zone *preferred_zone)
2416 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2418 * Only reset the batches of zones that were actually
2419 * considered in the fairness pass, we don't want to
2420 * trash fairness information for zones that are not
2421 * actually part of this zonelist's round-robin cycle.
2423 if (!zone_local(preferred_zone, zone))
2425 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2426 high_wmark_pages(zone) - low_wmark_pages(zone) -
2427 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2431 static void wake_all_kswapds(unsigned int order,
2432 struct zonelist *zonelist,
2433 enum zone_type high_zoneidx,
2434 struct zone *preferred_zone)
2439 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2440 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2444 gfp_to_alloc_flags(gfp_t gfp_mask)
2446 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2447 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2449 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2450 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2453 * The caller may dip into page reserves a bit more if the caller
2454 * cannot run direct reclaim, or if the caller has realtime scheduling
2455 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2456 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2458 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2462 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2463 * if it can't schedule.
2465 if (!(gfp_mask & __GFP_NOMEMALLOC))
2466 alloc_flags |= ALLOC_HARDER;
2468 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2469 * comment for __cpuset_node_allowed_softwall().
2471 alloc_flags &= ~ALLOC_CPUSET;
2472 } else if (unlikely(rt_task(current)) && !in_interrupt())
2473 alloc_flags |= ALLOC_HARDER;
2475 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2476 if (gfp_mask & __GFP_MEMALLOC)
2477 alloc_flags |= ALLOC_NO_WATERMARKS;
2478 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2479 alloc_flags |= ALLOC_NO_WATERMARKS;
2480 else if (!in_interrupt() &&
2481 ((current->flags & PF_MEMALLOC) ||
2482 unlikely(test_thread_flag(TIF_MEMDIE))))
2483 alloc_flags |= ALLOC_NO_WATERMARKS;
2486 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2487 alloc_flags |= ALLOC_CMA;
2492 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2494 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2497 static inline struct page *
2498 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2499 struct zonelist *zonelist, enum zone_type high_zoneidx,
2500 nodemask_t *nodemask, struct zone *preferred_zone,
2503 const gfp_t wait = gfp_mask & __GFP_WAIT;
2504 struct page *page = NULL;
2506 unsigned long pages_reclaimed = 0;
2507 unsigned long did_some_progress;
2508 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2509 bool deferred_compaction = false;
2510 bool contended_compaction = false;
2513 * In the slowpath, we sanity check order to avoid ever trying to
2514 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2515 * be using allocators in order of preference for an area that is
2518 if (order >= MAX_ORDER) {
2519 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2524 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2525 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2526 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2527 * using a larger set of nodes after it has established that the
2528 * allowed per node queues are empty and that nodes are
2531 if (IS_ENABLED(CONFIG_NUMA) &&
2532 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2536 if (!(gfp_mask & __GFP_NO_KSWAPD))
2537 wake_all_kswapds(order, zonelist, high_zoneidx, preferred_zone);
2540 * OK, we're below the kswapd watermark and have kicked background
2541 * reclaim. Now things get more complex, so set up alloc_flags according
2542 * to how we want to proceed.
2544 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2547 * Find the true preferred zone if the allocation is unconstrained by
2550 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2551 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2555 /* This is the last chance, in general, before the goto nopage. */
2556 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2557 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2558 preferred_zone, migratetype);
2562 /* Allocate without watermarks if the context allows */
2563 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2565 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2566 * the allocation is high priority and these type of
2567 * allocations are system rather than user orientated
2569 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2571 page = __alloc_pages_high_priority(gfp_mask, order,
2572 zonelist, high_zoneidx, nodemask,
2573 preferred_zone, migratetype);
2579 /* Atomic allocations - we can't balance anything */
2582 * All existing users of the deprecated __GFP_NOFAIL are
2583 * blockable, so warn of any new users that actually allow this
2584 * type of allocation to fail.
2586 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2590 /* Avoid recursion of direct reclaim */
2591 if (current->flags & PF_MEMALLOC)
2594 /* Avoid allocations with no watermarks from looping endlessly */
2595 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2599 * Try direct compaction. The first pass is asynchronous. Subsequent
2600 * attempts after direct reclaim are synchronous
2602 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2603 high_zoneidx, nodemask, alloc_flags,
2604 preferred_zone, migratetype,
2605 migration_mode, &contended_compaction,
2606 &deferred_compaction,
2607 &did_some_progress);
2610 migration_mode = MIGRATE_SYNC_LIGHT;
2613 * If compaction is deferred for high-order allocations, it is because
2614 * sync compaction recently failed. In this is the case and the caller
2615 * requested a movable allocation that does not heavily disrupt the
2616 * system then fail the allocation instead of entering direct reclaim.
2618 if ((deferred_compaction || contended_compaction) &&
2619 (gfp_mask & __GFP_NO_KSWAPD))
2622 /* Try direct reclaim and then allocating */
2623 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2624 zonelist, high_zoneidx,
2626 alloc_flags, preferred_zone,
2627 migratetype, &did_some_progress);
2632 * If we failed to make any progress reclaiming, then we are
2633 * running out of options and have to consider going OOM
2635 if (!did_some_progress) {
2636 if (oom_gfp_allowed(gfp_mask)) {
2637 if (oom_killer_disabled)
2639 /* Coredumps can quickly deplete all memory reserves */
2640 if ((current->flags & PF_DUMPCORE) &&
2641 !(gfp_mask & __GFP_NOFAIL))
2643 page = __alloc_pages_may_oom(gfp_mask, order,
2644 zonelist, high_zoneidx,
2645 nodemask, preferred_zone,
2650 if (!(gfp_mask & __GFP_NOFAIL)) {
2652 * The oom killer is not called for high-order
2653 * allocations that may fail, so if no progress
2654 * is being made, there are no other options and
2655 * retrying is unlikely to help.
2657 if (order > PAGE_ALLOC_COSTLY_ORDER)
2660 * The oom killer is not called for lowmem
2661 * allocations to prevent needlessly killing
2664 if (high_zoneidx < ZONE_NORMAL)
2672 /* Check if we should retry the allocation */
2673 pages_reclaimed += did_some_progress;
2674 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2676 /* Wait for some write requests to complete then retry */
2677 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2681 * High-order allocations do not necessarily loop after
2682 * direct reclaim and reclaim/compaction depends on compaction
2683 * being called after reclaim so call directly if necessary
2685 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2686 high_zoneidx, nodemask, alloc_flags,
2687 preferred_zone, migratetype,
2688 migration_mode, &contended_compaction,
2689 &deferred_compaction,
2690 &did_some_progress);
2696 warn_alloc_failed(gfp_mask, order, NULL);
2699 if (kmemcheck_enabled)
2700 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2706 * This is the 'heart' of the zoned buddy allocator.
2709 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2710 struct zonelist *zonelist, nodemask_t *nodemask)
2712 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2713 struct zone *preferred_zone;
2714 struct page *page = NULL;
2715 int migratetype = allocflags_to_migratetype(gfp_mask);
2716 unsigned int cpuset_mems_cookie;
2717 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2718 struct mem_cgroup *memcg = NULL;
2720 gfp_mask &= gfp_allowed_mask;
2722 lockdep_trace_alloc(gfp_mask);
2724 might_sleep_if(gfp_mask & __GFP_WAIT);
2726 if (should_fail_alloc_page(gfp_mask, order))
2730 * Check the zones suitable for the gfp_mask contain at least one
2731 * valid zone. It's possible to have an empty zonelist as a result
2732 * of GFP_THISNODE and a memoryless node
2734 if (unlikely(!zonelist->_zonerefs->zone))
2738 * Will only have any effect when __GFP_KMEMCG is set. This is
2739 * verified in the (always inline) callee
2741 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2745 cpuset_mems_cookie = read_mems_allowed_begin();
2747 /* The preferred zone is used for statistics later */
2748 first_zones_zonelist(zonelist, high_zoneidx,
2749 nodemask ? : &cpuset_current_mems_allowed,
2751 if (!preferred_zone)
2755 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2756 alloc_flags |= ALLOC_CMA;
2759 /* First allocation attempt */
2760 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2761 zonelist, high_zoneidx, alloc_flags,
2762 preferred_zone, migratetype);
2763 if (unlikely(!page)) {
2765 * The first pass makes sure allocations are spread
2766 * fairly within the local node. However, the local
2767 * node might have free pages left after the fairness
2768 * batches are exhausted, and remote zones haven't
2769 * even been considered yet. Try once more without
2770 * fairness, and include remote zones now, before
2771 * entering the slowpath and waking kswapd: prefer
2772 * spilling to a remote zone over swapping locally.
2774 if (alloc_flags & ALLOC_FAIR) {
2775 reset_alloc_batches(zonelist, high_zoneidx,
2777 alloc_flags &= ~ALLOC_FAIR;
2781 * Runtime PM, block IO and its error handling path
2782 * can deadlock because I/O on the device might not
2785 gfp_mask = memalloc_noio_flags(gfp_mask);
2786 page = __alloc_pages_slowpath(gfp_mask, order,
2787 zonelist, high_zoneidx, nodemask,
2788 preferred_zone, migratetype);
2791 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2795 * When updating a task's mems_allowed, it is possible to race with
2796 * parallel threads in such a way that an allocation can fail while
2797 * the mask is being updated. If a page allocation is about to fail,
2798 * check if the cpuset changed during allocation and if so, retry.
2800 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2803 memcg_kmem_commit_charge(page, memcg, order);
2807 EXPORT_SYMBOL(__alloc_pages_nodemask);
2810 * Common helper functions.
2812 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2817 * __get_free_pages() returns a 32-bit address, which cannot represent
2820 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2822 page = alloc_pages(gfp_mask, order);
2825 return (unsigned long) page_address(page);
2827 EXPORT_SYMBOL(__get_free_pages);
2829 unsigned long get_zeroed_page(gfp_t gfp_mask)
2831 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2833 EXPORT_SYMBOL(get_zeroed_page);
2835 void __free_pages(struct page *page, unsigned int order)
2837 if (put_page_testzero(page)) {
2839 free_hot_cold_page(page, 0);
2841 __free_pages_ok(page, order);
2845 EXPORT_SYMBOL(__free_pages);
2847 void free_pages(unsigned long addr, unsigned int order)
2850 VM_BUG_ON(!virt_addr_valid((void *)addr));
2851 __free_pages(virt_to_page((void *)addr), order);
2855 EXPORT_SYMBOL(free_pages);
2858 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2859 * pages allocated with __GFP_KMEMCG.
2861 * Those pages are accounted to a particular memcg, embedded in the
2862 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2863 * for that information only to find out that it is NULL for users who have no
2864 * interest in that whatsoever, we provide these functions.
2866 * The caller knows better which flags it relies on.
2868 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2870 memcg_kmem_uncharge_pages(page, order);
2871 __free_pages(page, order);
2874 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2877 VM_BUG_ON(!virt_addr_valid((void *)addr));
2878 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2882 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2885 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2886 unsigned long used = addr + PAGE_ALIGN(size);
2888 split_page(virt_to_page((void *)addr), order);
2889 while (used < alloc_end) {
2894 return (void *)addr;
2898 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2899 * @size: the number of bytes to allocate
2900 * @gfp_mask: GFP flags for the allocation
2902 * This function is similar to alloc_pages(), except that it allocates the
2903 * minimum number of pages to satisfy the request. alloc_pages() can only
2904 * allocate memory in power-of-two pages.
2906 * This function is also limited by MAX_ORDER.
2908 * Memory allocated by this function must be released by free_pages_exact().
2910 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2912 unsigned int order = get_order(size);
2915 addr = __get_free_pages(gfp_mask, order);
2916 return make_alloc_exact(addr, order, size);
2918 EXPORT_SYMBOL(alloc_pages_exact);
2921 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2923 * @nid: the preferred node ID where memory should be allocated
2924 * @size: the number of bytes to allocate
2925 * @gfp_mask: GFP flags for the allocation
2927 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2929 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2932 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2934 unsigned order = get_order(size);
2935 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2938 return make_alloc_exact((unsigned long)page_address(p), order, size);
2940 EXPORT_SYMBOL(alloc_pages_exact_nid);
2943 * free_pages_exact - release memory allocated via alloc_pages_exact()
2944 * @virt: the value returned by alloc_pages_exact.
2945 * @size: size of allocation, same value as passed to alloc_pages_exact().
2947 * Release the memory allocated by a previous call to alloc_pages_exact.
2949 void free_pages_exact(void *virt, size_t size)
2951 unsigned long addr = (unsigned long)virt;
2952 unsigned long end = addr + PAGE_ALIGN(size);
2954 while (addr < end) {
2959 EXPORT_SYMBOL(free_pages_exact);
2962 * nr_free_zone_pages - count number of pages beyond high watermark
2963 * @offset: The zone index of the highest zone
2965 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2966 * high watermark within all zones at or below a given zone index. For each
2967 * zone, the number of pages is calculated as:
2968 * managed_pages - high_pages
2970 static unsigned long nr_free_zone_pages(int offset)
2975 /* Just pick one node, since fallback list is circular */
2976 unsigned long sum = 0;
2978 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2980 for_each_zone_zonelist(zone, z, zonelist, offset) {
2981 unsigned long size = zone->managed_pages;
2982 unsigned long high = high_wmark_pages(zone);
2991 * nr_free_buffer_pages - count number of pages beyond high watermark
2993 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2994 * watermark within ZONE_DMA and ZONE_NORMAL.
2996 unsigned long nr_free_buffer_pages(void)
2998 return nr_free_zone_pages(gfp_zone(GFP_USER));
3000 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3003 * nr_free_pagecache_pages - count number of pages beyond high watermark
3005 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3006 * high watermark within all zones.
3008 unsigned long nr_free_pagecache_pages(void)
3010 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3013 static inline void show_node(struct zone *zone)
3015 if (IS_ENABLED(CONFIG_NUMA))
3016 printk("Node %d ", zone_to_nid(zone));
3019 void si_meminfo(struct sysinfo *val)
3021 val->totalram = totalram_pages;
3023 val->freeram = global_page_state(NR_FREE_PAGES);
3024 val->bufferram = nr_blockdev_pages();
3025 val->totalhigh = totalhigh_pages;
3026 val->freehigh = nr_free_highpages();
3027 val->mem_unit = PAGE_SIZE;
3030 EXPORT_SYMBOL(si_meminfo);
3033 void si_meminfo_node(struct sysinfo *val, int nid)
3035 int zone_type; /* needs to be signed */
3036 unsigned long managed_pages = 0;
3037 pg_data_t *pgdat = NODE_DATA(nid);
3039 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3040 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3041 val->totalram = managed_pages;
3042 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3043 #ifdef CONFIG_HIGHMEM
3044 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3045 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3051 val->mem_unit = PAGE_SIZE;
3056 * Determine whether the node should be displayed or not, depending on whether
3057 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3059 bool skip_free_areas_node(unsigned int flags, int nid)
3062 unsigned int cpuset_mems_cookie;
3064 if (!(flags & SHOW_MEM_FILTER_NODES))
3068 cpuset_mems_cookie = read_mems_allowed_begin();
3069 ret = !node_isset(nid, cpuset_current_mems_allowed);
3070 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3075 #define K(x) ((x) << (PAGE_SHIFT-10))
3077 static void show_migration_types(unsigned char type)
3079 static const char types[MIGRATE_TYPES] = {
3080 [MIGRATE_UNMOVABLE] = 'U',
3081 [MIGRATE_RECLAIMABLE] = 'E',
3082 [MIGRATE_MOVABLE] = 'M',
3083 [MIGRATE_RESERVE] = 'R',
3085 [MIGRATE_CMA] = 'C',
3087 #ifdef CONFIG_MEMORY_ISOLATION
3088 [MIGRATE_ISOLATE] = 'I',
3091 char tmp[MIGRATE_TYPES + 1];
3095 for (i = 0; i < MIGRATE_TYPES; i++) {
3096 if (type & (1 << i))
3101 printk("(%s) ", tmp);
3105 * Show free area list (used inside shift_scroll-lock stuff)
3106 * We also calculate the percentage fragmentation. We do this by counting the
3107 * memory on each free list with the exception of the first item on the list.
3108 * Suppresses nodes that are not allowed by current's cpuset if
3109 * SHOW_MEM_FILTER_NODES is passed.
3111 void show_free_areas(unsigned int filter)
3116 for_each_populated_zone(zone) {
3117 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3120 printk("%s per-cpu:\n", zone->name);
3122 for_each_online_cpu(cpu) {
3123 struct per_cpu_pageset *pageset;
3125 pageset = per_cpu_ptr(zone->pageset, cpu);
3127 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3128 cpu, pageset->pcp.high,
3129 pageset->pcp.batch, pageset->pcp.count);
3133 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3134 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3136 " dirty:%lu writeback:%lu unstable:%lu\n"
3137 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3138 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3140 global_page_state(NR_ACTIVE_ANON),
3141 global_page_state(NR_INACTIVE_ANON),
3142 global_page_state(NR_ISOLATED_ANON),
3143 global_page_state(NR_ACTIVE_FILE),
3144 global_page_state(NR_INACTIVE_FILE),
3145 global_page_state(NR_ISOLATED_FILE),
3146 global_page_state(NR_UNEVICTABLE),
3147 global_page_state(NR_FILE_DIRTY),
3148 global_page_state(NR_WRITEBACK),
3149 global_page_state(NR_UNSTABLE_NFS),
3150 global_page_state(NR_FREE_PAGES),
3151 global_page_state(NR_SLAB_RECLAIMABLE),
3152 global_page_state(NR_SLAB_UNRECLAIMABLE),
3153 global_page_state(NR_FILE_MAPPED),
3154 global_page_state(NR_SHMEM),
3155 global_page_state(NR_PAGETABLE),
3156 global_page_state(NR_BOUNCE),
3157 global_page_state(NR_FREE_CMA_PAGES));
3159 for_each_populated_zone(zone) {
3162 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3170 " active_anon:%lukB"
3171 " inactive_anon:%lukB"
3172 " active_file:%lukB"
3173 " inactive_file:%lukB"
3174 " unevictable:%lukB"
3175 " isolated(anon):%lukB"
3176 " isolated(file):%lukB"
3184 " slab_reclaimable:%lukB"
3185 " slab_unreclaimable:%lukB"
3186 " kernel_stack:%lukB"
3191 " writeback_tmp:%lukB"
3192 " pages_scanned:%lu"
3193 " all_unreclaimable? %s"
3196 K(zone_page_state(zone, NR_FREE_PAGES)),
3197 K(min_wmark_pages(zone)),
3198 K(low_wmark_pages(zone)),
3199 K(high_wmark_pages(zone)),
3200 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3201 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3202 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3203 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3204 K(zone_page_state(zone, NR_UNEVICTABLE)),
3205 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3206 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3207 K(zone->present_pages),
3208 K(zone->managed_pages),
3209 K(zone_page_state(zone, NR_MLOCK)),
3210 K(zone_page_state(zone, NR_FILE_DIRTY)),
3211 K(zone_page_state(zone, NR_WRITEBACK)),
3212 K(zone_page_state(zone, NR_FILE_MAPPED)),
3213 K(zone_page_state(zone, NR_SHMEM)),
3214 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3215 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3216 zone_page_state(zone, NR_KERNEL_STACK) *
3218 K(zone_page_state(zone, NR_PAGETABLE)),
3219 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3220 K(zone_page_state(zone, NR_BOUNCE)),
3221 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3222 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3223 zone->pages_scanned,
3224 (!zone_reclaimable(zone) ? "yes" : "no")
3226 printk("lowmem_reserve[]:");
3227 for (i = 0; i < MAX_NR_ZONES; i++)
3228 printk(" %lu", zone->lowmem_reserve[i]);
3232 for_each_populated_zone(zone) {
3233 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3234 unsigned char types[MAX_ORDER];
3236 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3239 printk("%s: ", zone->name);
3241 spin_lock_irqsave(&zone->lock, flags);
3242 for (order = 0; order < MAX_ORDER; order++) {
3243 struct free_area *area = &zone->free_area[order];
3246 nr[order] = area->nr_free;
3247 total += nr[order] << order;
3250 for (type = 0; type < MIGRATE_TYPES; type++) {
3251 if (!list_empty(&area->free_list[type]))
3252 types[order] |= 1 << type;
3255 spin_unlock_irqrestore(&zone->lock, flags);
3256 for (order = 0; order < MAX_ORDER; order++) {
3257 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3259 show_migration_types(types[order]);
3261 printk("= %lukB\n", K(total));
3264 hugetlb_show_meminfo();
3266 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3268 show_swap_cache_info();
3271 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3273 zoneref->zone = zone;
3274 zoneref->zone_idx = zone_idx(zone);
3278 * Builds allocation fallback zone lists.
3280 * Add all populated zones of a node to the zonelist.
3282 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3286 enum zone_type zone_type = MAX_NR_ZONES;
3290 zone = pgdat->node_zones + zone_type;
3291 if (populated_zone(zone)) {
3292 zoneref_set_zone(zone,
3293 &zonelist->_zonerefs[nr_zones++]);
3294 check_highest_zone(zone_type);
3296 } while (zone_type);
3304 * 0 = automatic detection of better ordering.
3305 * 1 = order by ([node] distance, -zonetype)
3306 * 2 = order by (-zonetype, [node] distance)
3308 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3309 * the same zonelist. So only NUMA can configure this param.
3311 #define ZONELIST_ORDER_DEFAULT 0
3312 #define ZONELIST_ORDER_NODE 1
3313 #define ZONELIST_ORDER_ZONE 2
3315 /* zonelist order in the kernel.
3316 * set_zonelist_order() will set this to NODE or ZONE.
3318 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3319 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3323 /* The value user specified ....changed by config */
3324 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3325 /* string for sysctl */
3326 #define NUMA_ZONELIST_ORDER_LEN 16
3327 char numa_zonelist_order[16] = "default";
3330 * interface for configure zonelist ordering.
3331 * command line option "numa_zonelist_order"
3332 * = "[dD]efault - default, automatic configuration.
3333 * = "[nN]ode - order by node locality, then by zone within node
3334 * = "[zZ]one - order by zone, then by locality within zone
3337 static int __parse_numa_zonelist_order(char *s)
3339 if (*s == 'd' || *s == 'D') {
3340 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3341 } else if (*s == 'n' || *s == 'N') {
3342 user_zonelist_order = ZONELIST_ORDER_NODE;
3343 } else if (*s == 'z' || *s == 'Z') {
3344 user_zonelist_order = ZONELIST_ORDER_ZONE;
3347 "Ignoring invalid numa_zonelist_order value: "
3354 static __init int setup_numa_zonelist_order(char *s)
3361 ret = __parse_numa_zonelist_order(s);
3363 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3367 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3370 * sysctl handler for numa_zonelist_order
3372 int numa_zonelist_order_handler(ctl_table *table, int write,
3373 void __user *buffer, size_t *length,
3376 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3378 static DEFINE_MUTEX(zl_order_mutex);
3380 mutex_lock(&zl_order_mutex);
3382 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3386 strcpy(saved_string, (char *)table->data);
3388 ret = proc_dostring(table, write, buffer, length, ppos);
3392 int oldval = user_zonelist_order;
3394 ret = __parse_numa_zonelist_order((char *)table->data);
3397 * bogus value. restore saved string
3399 strncpy((char *)table->data, saved_string,
3400 NUMA_ZONELIST_ORDER_LEN);
3401 user_zonelist_order = oldval;
3402 } else if (oldval != user_zonelist_order) {
3403 mutex_lock(&zonelists_mutex);
3404 build_all_zonelists(NULL, NULL);
3405 mutex_unlock(&zonelists_mutex);
3409 mutex_unlock(&zl_order_mutex);
3414 #define MAX_NODE_LOAD (nr_online_nodes)
3415 static int node_load[MAX_NUMNODES];
3418 * find_next_best_node - find the next node that should appear in a given node's fallback list
3419 * @node: node whose fallback list we're appending
3420 * @used_node_mask: nodemask_t of already used nodes
3422 * We use a number of factors to determine which is the next node that should
3423 * appear on a given node's fallback list. The node should not have appeared
3424 * already in @node's fallback list, and it should be the next closest node
3425 * according to the distance array (which contains arbitrary distance values
3426 * from each node to each node in the system), and should also prefer nodes
3427 * with no CPUs, since presumably they'll have very little allocation pressure
3428 * on them otherwise.
3429 * It returns -1 if no node is found.
3431 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3434 int min_val = INT_MAX;
3435 int best_node = NUMA_NO_NODE;
3436 const struct cpumask *tmp = cpumask_of_node(0);
3438 /* Use the local node if we haven't already */
3439 if (!node_isset(node, *used_node_mask)) {
3440 node_set(node, *used_node_mask);
3444 for_each_node_state(n, N_MEMORY) {
3446 /* Don't want a node to appear more than once */
3447 if (node_isset(n, *used_node_mask))
3450 /* Use the distance array to find the distance */
3451 val = node_distance(node, n);
3453 /* Penalize nodes under us ("prefer the next node") */
3456 /* Give preference to headless and unused nodes */
3457 tmp = cpumask_of_node(n);
3458 if (!cpumask_empty(tmp))
3459 val += PENALTY_FOR_NODE_WITH_CPUS;
3461 /* Slight preference for less loaded node */
3462 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3463 val += node_load[n];
3465 if (val < min_val) {
3472 node_set(best_node, *used_node_mask);
3479 * Build zonelists ordered by node and zones within node.
3480 * This results in maximum locality--normal zone overflows into local
3481 * DMA zone, if any--but risks exhausting DMA zone.
3483 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3486 struct zonelist *zonelist;
3488 zonelist = &pgdat->node_zonelists[0];
3489 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3491 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3492 zonelist->_zonerefs[j].zone = NULL;
3493 zonelist->_zonerefs[j].zone_idx = 0;
3497 * Build gfp_thisnode zonelists
3499 static void build_thisnode_zonelists(pg_data_t *pgdat)
3502 struct zonelist *zonelist;
3504 zonelist = &pgdat->node_zonelists[1];
3505 j = build_zonelists_node(pgdat, zonelist, 0);
3506 zonelist->_zonerefs[j].zone = NULL;
3507 zonelist->_zonerefs[j].zone_idx = 0;
3511 * Build zonelists ordered by zone and nodes within zones.
3512 * This results in conserving DMA zone[s] until all Normal memory is
3513 * exhausted, but results in overflowing to remote node while memory
3514 * may still exist in local DMA zone.
3516 static int node_order[MAX_NUMNODES];
3518 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3521 int zone_type; /* needs to be signed */
3523 struct zonelist *zonelist;
3525 zonelist = &pgdat->node_zonelists[0];
3527 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3528 for (j = 0; j < nr_nodes; j++) {
3529 node = node_order[j];
3530 z = &NODE_DATA(node)->node_zones[zone_type];
3531 if (populated_zone(z)) {
3533 &zonelist->_zonerefs[pos++]);
3534 check_highest_zone(zone_type);
3538 zonelist->_zonerefs[pos].zone = NULL;
3539 zonelist->_zonerefs[pos].zone_idx = 0;
3542 static int default_zonelist_order(void)
3545 unsigned long low_kmem_size, total_size;
3549 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3550 * If they are really small and used heavily, the system can fall
3551 * into OOM very easily.
3552 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3554 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3557 for_each_online_node(nid) {
3558 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3559 z = &NODE_DATA(nid)->node_zones[zone_type];
3560 if (populated_zone(z)) {
3561 if (zone_type < ZONE_NORMAL)
3562 low_kmem_size += z->managed_pages;
3563 total_size += z->managed_pages;
3564 } else if (zone_type == ZONE_NORMAL) {
3566 * If any node has only lowmem, then node order
3567 * is preferred to allow kernel allocations
3568 * locally; otherwise, they can easily infringe
3569 * on other nodes when there is an abundance of
3570 * lowmem available to allocate from.
3572 return ZONELIST_ORDER_NODE;
3576 if (!low_kmem_size || /* there are no DMA area. */
3577 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3578 return ZONELIST_ORDER_NODE;
3580 * look into each node's config.
3581 * If there is a node whose DMA/DMA32 memory is very big area on
3582 * local memory, NODE_ORDER may be suitable.
3584 average_size = total_size /
3585 (nodes_weight(node_states[N_MEMORY]) + 1);
3586 for_each_online_node(nid) {
3589 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3590 z = &NODE_DATA(nid)->node_zones[zone_type];
3591 if (populated_zone(z)) {
3592 if (zone_type < ZONE_NORMAL)
3593 low_kmem_size += z->present_pages;
3594 total_size += z->present_pages;
3597 if (low_kmem_size &&
3598 total_size > average_size && /* ignore small node */
3599 low_kmem_size > total_size * 70/100)
3600 return ZONELIST_ORDER_NODE;
3602 return ZONELIST_ORDER_ZONE;
3605 static void set_zonelist_order(void)
3607 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3608 current_zonelist_order = default_zonelist_order();
3610 current_zonelist_order = user_zonelist_order;
3613 static void build_zonelists(pg_data_t *pgdat)
3617 nodemask_t used_mask;
3618 int local_node, prev_node;
3619 struct zonelist *zonelist;
3620 int order = current_zonelist_order;
3622 /* initialize zonelists */
3623 for (i = 0; i < MAX_ZONELISTS; i++) {
3624 zonelist = pgdat->node_zonelists + i;
3625 zonelist->_zonerefs[0].zone = NULL;
3626 zonelist->_zonerefs[0].zone_idx = 0;
3629 /* NUMA-aware ordering of nodes */
3630 local_node = pgdat->node_id;
3631 load = nr_online_nodes;
3632 prev_node = local_node;
3633 nodes_clear(used_mask);
3635 memset(node_order, 0, sizeof(node_order));
3638 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3640 * We don't want to pressure a particular node.
3641 * So adding penalty to the first node in same
3642 * distance group to make it round-robin.
3644 if (node_distance(local_node, node) !=
3645 node_distance(local_node, prev_node))
3646 node_load[node] = load;
3650 if (order == ZONELIST_ORDER_NODE)
3651 build_zonelists_in_node_order(pgdat, node);
3653 node_order[j++] = node; /* remember order */
3656 if (order == ZONELIST_ORDER_ZONE) {
3657 /* calculate node order -- i.e., DMA last! */
3658 build_zonelists_in_zone_order(pgdat, j);
3661 build_thisnode_zonelists(pgdat);
3664 /* Construct the zonelist performance cache - see further mmzone.h */
3665 static void build_zonelist_cache(pg_data_t *pgdat)
3667 struct zonelist *zonelist;
3668 struct zonelist_cache *zlc;
3671 zonelist = &pgdat->node_zonelists[0];
3672 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3673 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3674 for (z = zonelist->_zonerefs; z->zone; z++)
3675 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3678 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3680 * Return node id of node used for "local" allocations.
3681 * I.e., first node id of first zone in arg node's generic zonelist.
3682 * Used for initializing percpu 'numa_mem', which is used primarily
3683 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3685 int local_memory_node(int node)
3689 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3690 gfp_zone(GFP_KERNEL),
3697 #else /* CONFIG_NUMA */
3699 static void set_zonelist_order(void)
3701 current_zonelist_order = ZONELIST_ORDER_ZONE;
3704 static void build_zonelists(pg_data_t *pgdat)
3706 int node, local_node;
3708 struct zonelist *zonelist;
3710 local_node = pgdat->node_id;
3712 zonelist = &pgdat->node_zonelists[0];
3713 j = build_zonelists_node(pgdat, zonelist, 0);
3716 * Now we build the zonelist so that it contains the zones
3717 * of all the other nodes.
3718 * We don't want to pressure a particular node, so when
3719 * building the zones for node N, we make sure that the
3720 * zones coming right after the local ones are those from
3721 * node N+1 (modulo N)
3723 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3724 if (!node_online(node))
3726 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3728 for (node = 0; node < local_node; node++) {
3729 if (!node_online(node))
3731 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3734 zonelist->_zonerefs[j].zone = NULL;
3735 zonelist->_zonerefs[j].zone_idx = 0;
3738 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3739 static void build_zonelist_cache(pg_data_t *pgdat)
3741 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3744 #endif /* CONFIG_NUMA */
3747 * Boot pageset table. One per cpu which is going to be used for all
3748 * zones and all nodes. The parameters will be set in such a way
3749 * that an item put on a list will immediately be handed over to
3750 * the buddy list. This is safe since pageset manipulation is done
3751 * with interrupts disabled.
3753 * The boot_pagesets must be kept even after bootup is complete for
3754 * unused processors and/or zones. They do play a role for bootstrapping
3755 * hotplugged processors.
3757 * zoneinfo_show() and maybe other functions do
3758 * not check if the processor is online before following the pageset pointer.
3759 * Other parts of the kernel may not check if the zone is available.
3761 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3762 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3763 static void setup_zone_pageset(struct zone *zone);
3766 * Global mutex to protect against size modification of zonelists
3767 * as well as to serialize pageset setup for the new populated zone.
3769 DEFINE_MUTEX(zonelists_mutex);
3771 /* return values int ....just for stop_machine() */
3772 static int __build_all_zonelists(void *data)
3776 pg_data_t *self = data;
3779 memset(node_load, 0, sizeof(node_load));
3782 if (self && !node_online(self->node_id)) {
3783 build_zonelists(self);
3784 build_zonelist_cache(self);
3787 for_each_online_node(nid) {
3788 pg_data_t *pgdat = NODE_DATA(nid);
3790 build_zonelists(pgdat);
3791 build_zonelist_cache(pgdat);
3795 * Initialize the boot_pagesets that are going to be used
3796 * for bootstrapping processors. The real pagesets for
3797 * each zone will be allocated later when the per cpu
3798 * allocator is available.
3800 * boot_pagesets are used also for bootstrapping offline
3801 * cpus if the system is already booted because the pagesets
3802 * are needed to initialize allocators on a specific cpu too.
3803 * F.e. the percpu allocator needs the page allocator which
3804 * needs the percpu allocator in order to allocate its pagesets
3805 * (a chicken-egg dilemma).
3807 for_each_possible_cpu(cpu) {
3808 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3810 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3812 * We now know the "local memory node" for each node--
3813 * i.e., the node of the first zone in the generic zonelist.
3814 * Set up numa_mem percpu variable for on-line cpus. During
3815 * boot, only the boot cpu should be on-line; we'll init the
3816 * secondary cpus' numa_mem as they come on-line. During
3817 * node/memory hotplug, we'll fixup all on-line cpus.
3819 if (cpu_online(cpu))
3820 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3828 * Called with zonelists_mutex held always
3829 * unless system_state == SYSTEM_BOOTING.
3831 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3833 set_zonelist_order();
3835 if (system_state == SYSTEM_BOOTING) {
3836 __build_all_zonelists(NULL);
3837 mminit_verify_zonelist();
3838 cpuset_init_current_mems_allowed();
3840 #ifdef CONFIG_MEMORY_HOTPLUG
3842 setup_zone_pageset(zone);
3844 /* we have to stop all cpus to guarantee there is no user
3846 stop_machine(__build_all_zonelists, pgdat, NULL);
3847 /* cpuset refresh routine should be here */
3849 vm_total_pages = nr_free_pagecache_pages();
3851 * Disable grouping by mobility if the number of pages in the
3852 * system is too low to allow the mechanism to work. It would be
3853 * more accurate, but expensive to check per-zone. This check is
3854 * made on memory-hotadd so a system can start with mobility
3855 * disabled and enable it later
3857 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3858 page_group_by_mobility_disabled = 1;
3860 page_group_by_mobility_disabled = 0;
3862 printk("Built %i zonelists in %s order, mobility grouping %s. "
3863 "Total pages: %ld\n",
3865 zonelist_order_name[current_zonelist_order],
3866 page_group_by_mobility_disabled ? "off" : "on",
3869 printk("Policy zone: %s\n", zone_names[policy_zone]);
3874 * Helper functions to size the waitqueue hash table.
3875 * Essentially these want to choose hash table sizes sufficiently
3876 * large so that collisions trying to wait on pages are rare.
3877 * But in fact, the number of active page waitqueues on typical
3878 * systems is ridiculously low, less than 200. So this is even
3879 * conservative, even though it seems large.
3881 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3882 * waitqueues, i.e. the size of the waitq table given the number of pages.
3884 #define PAGES_PER_WAITQUEUE 256
3886 #ifndef CONFIG_MEMORY_HOTPLUG
3887 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3889 unsigned long size = 1;
3891 pages /= PAGES_PER_WAITQUEUE;
3893 while (size < pages)
3897 * Once we have dozens or even hundreds of threads sleeping
3898 * on IO we've got bigger problems than wait queue collision.
3899 * Limit the size of the wait table to a reasonable size.
3901 size = min(size, 4096UL);
3903 return max(size, 4UL);
3907 * A zone's size might be changed by hot-add, so it is not possible to determine
3908 * a suitable size for its wait_table. So we use the maximum size now.
3910 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3912 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3913 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3914 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3916 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3917 * or more by the traditional way. (See above). It equals:
3919 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3920 * ia64(16K page size) : = ( 8G + 4M)byte.
3921 * powerpc (64K page size) : = (32G +16M)byte.
3923 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3930 * This is an integer logarithm so that shifts can be used later
3931 * to extract the more random high bits from the multiplicative
3932 * hash function before the remainder is taken.
3934 static inline unsigned long wait_table_bits(unsigned long size)
3940 * Check if a pageblock contains reserved pages
3942 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3946 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3947 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3954 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3955 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3956 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3957 * higher will lead to a bigger reserve which will get freed as contiguous
3958 * blocks as reclaim kicks in
3960 static void setup_zone_migrate_reserve(struct zone *zone)
3962 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3964 unsigned long block_migratetype;
3969 * Get the start pfn, end pfn and the number of blocks to reserve
3970 * We have to be careful to be aligned to pageblock_nr_pages to
3971 * make sure that we always check pfn_valid for the first page in
3974 start_pfn = zone->zone_start_pfn;
3975 end_pfn = zone_end_pfn(zone);
3976 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3977 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3981 * Reserve blocks are generally in place to help high-order atomic
3982 * allocations that are short-lived. A min_free_kbytes value that
3983 * would result in more than 2 reserve blocks for atomic allocations
3984 * is assumed to be in place to help anti-fragmentation for the
3985 * future allocation of hugepages at runtime.
3987 reserve = min(2, reserve);
3988 old_reserve = zone->nr_migrate_reserve_block;
3990 /* When memory hot-add, we almost always need to do nothing */
3991 if (reserve == old_reserve)
3993 zone->nr_migrate_reserve_block = reserve;
3995 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3996 if (!pfn_valid(pfn))
3998 page = pfn_to_page(pfn);
4000 /* Watch out for overlapping nodes */
4001 if (page_to_nid(page) != zone_to_nid(zone))
4004 block_migratetype = get_pageblock_migratetype(page);
4006 /* Only test what is necessary when the reserves are not met */
4009 * Blocks with reserved pages will never free, skip
4012 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4013 if (pageblock_is_reserved(pfn, block_end_pfn))
4016 /* If this block is reserved, account for it */
4017 if (block_migratetype == MIGRATE_RESERVE) {
4022 /* Suitable for reserving if this block is movable */
4023 if (block_migratetype == MIGRATE_MOVABLE) {
4024 set_pageblock_migratetype(page,
4026 move_freepages_block(zone, page,
4031 } else if (!old_reserve) {
4033 * At boot time we don't need to scan the whole zone
4034 * for turning off MIGRATE_RESERVE.
4040 * If the reserve is met and this is a previous reserved block,
4043 if (block_migratetype == MIGRATE_RESERVE) {
4044 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4045 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4051 * Initially all pages are reserved - free ones are freed
4052 * up by free_all_bootmem() once the early boot process is
4053 * done. Non-atomic initialization, single-pass.
4055 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4056 unsigned long start_pfn, enum memmap_context context)
4059 unsigned long end_pfn = start_pfn + size;
4063 if (highest_memmap_pfn < end_pfn - 1)
4064 highest_memmap_pfn = end_pfn - 1;
4066 z = &NODE_DATA(nid)->node_zones[zone];
4067 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4069 * There can be holes in boot-time mem_map[]s
4070 * handed to this function. They do not
4071 * exist on hotplugged memory.
4073 if (context == MEMMAP_EARLY) {
4074 if (!early_pfn_valid(pfn))
4076 if (!early_pfn_in_nid(pfn, nid))
4079 page = pfn_to_page(pfn);
4080 set_page_links(page, zone, nid, pfn);
4081 mminit_verify_page_links(page, zone, nid, pfn);
4082 init_page_count(page);
4083 page_mapcount_reset(page);
4084 page_cpupid_reset_last(page);
4085 SetPageReserved(page);
4087 * Mark the block movable so that blocks are reserved for
4088 * movable at startup. This will force kernel allocations
4089 * to reserve their blocks rather than leaking throughout
4090 * the address space during boot when many long-lived
4091 * kernel allocations are made. Later some blocks near
4092 * the start are marked MIGRATE_RESERVE by
4093 * setup_zone_migrate_reserve()
4095 * bitmap is created for zone's valid pfn range. but memmap
4096 * can be created for invalid pages (for alignment)
4097 * check here not to call set_pageblock_migratetype() against
4100 if ((z->zone_start_pfn <= pfn)
4101 && (pfn < zone_end_pfn(z))
4102 && !(pfn & (pageblock_nr_pages - 1)))
4103 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4105 INIT_LIST_HEAD(&page->lru);
4106 #ifdef WANT_PAGE_VIRTUAL
4107 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4108 if (!is_highmem_idx(zone))
4109 set_page_address(page, __va(pfn << PAGE_SHIFT));
4114 static void __meminit zone_init_free_lists(struct zone *zone)
4117 for_each_migratetype_order(order, t) {
4118 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4119 zone->free_area[order].nr_free = 0;
4123 #ifndef __HAVE_ARCH_MEMMAP_INIT
4124 #define memmap_init(size, nid, zone, start_pfn) \
4125 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4128 static int zone_batchsize(struct zone *zone)
4134 * The per-cpu-pages pools are set to around 1000th of the
4135 * size of the zone. But no more than 1/2 of a meg.
4137 * OK, so we don't know how big the cache is. So guess.
4139 batch = zone->managed_pages / 1024;
4140 if (batch * PAGE_SIZE > 512 * 1024)
4141 batch = (512 * 1024) / PAGE_SIZE;
4142 batch /= 4; /* We effectively *= 4 below */
4147 * Clamp the batch to a 2^n - 1 value. Having a power
4148 * of 2 value was found to be more likely to have
4149 * suboptimal cache aliasing properties in some cases.
4151 * For example if 2 tasks are alternately allocating
4152 * batches of pages, one task can end up with a lot
4153 * of pages of one half of the possible page colors
4154 * and the other with pages of the other colors.
4156 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4161 /* The deferral and batching of frees should be suppressed under NOMMU
4164 * The problem is that NOMMU needs to be able to allocate large chunks
4165 * of contiguous memory as there's no hardware page translation to
4166 * assemble apparent contiguous memory from discontiguous pages.
4168 * Queueing large contiguous runs of pages for batching, however,
4169 * causes the pages to actually be freed in smaller chunks. As there
4170 * can be a significant delay between the individual batches being
4171 * recycled, this leads to the once large chunks of space being
4172 * fragmented and becoming unavailable for high-order allocations.
4179 * pcp->high and pcp->batch values are related and dependent on one another:
4180 * ->batch must never be higher then ->high.
4181 * The following function updates them in a safe manner without read side
4184 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4185 * those fields changing asynchronously (acording the the above rule).
4187 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4188 * outside of boot time (or some other assurance that no concurrent updaters
4191 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4192 unsigned long batch)
4194 /* start with a fail safe value for batch */
4198 /* Update high, then batch, in order */
4205 /* a companion to pageset_set_high() */
4206 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4208 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4211 static void pageset_init(struct per_cpu_pageset *p)
4213 struct per_cpu_pages *pcp;
4216 memset(p, 0, sizeof(*p));
4220 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4221 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4224 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4227 pageset_set_batch(p, batch);
4231 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4232 * to the value high for the pageset p.
4234 static void pageset_set_high(struct per_cpu_pageset *p,
4237 unsigned long batch = max(1UL, high / 4);
4238 if ((high / 4) > (PAGE_SHIFT * 8))
4239 batch = PAGE_SHIFT * 8;
4241 pageset_update(&p->pcp, high, batch);
4244 static void pageset_set_high_and_batch(struct zone *zone,
4245 struct per_cpu_pageset *pcp)
4247 if (percpu_pagelist_fraction)
4248 pageset_set_high(pcp,
4249 (zone->managed_pages /
4250 percpu_pagelist_fraction));
4252 pageset_set_batch(pcp, zone_batchsize(zone));
4255 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4257 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4260 pageset_set_high_and_batch(zone, pcp);
4263 static void __meminit setup_zone_pageset(struct zone *zone)
4266 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4267 for_each_possible_cpu(cpu)
4268 zone_pageset_init(zone, cpu);
4272 * Allocate per cpu pagesets and initialize them.
4273 * Before this call only boot pagesets were available.
4275 void __init setup_per_cpu_pageset(void)
4279 for_each_populated_zone(zone)
4280 setup_zone_pageset(zone);
4283 static noinline __init_refok
4284 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4290 * The per-page waitqueue mechanism uses hashed waitqueues
4293 zone->wait_table_hash_nr_entries =
4294 wait_table_hash_nr_entries(zone_size_pages);
4295 zone->wait_table_bits =
4296 wait_table_bits(zone->wait_table_hash_nr_entries);
4297 alloc_size = zone->wait_table_hash_nr_entries
4298 * sizeof(wait_queue_head_t);
4300 if (!slab_is_available()) {
4301 zone->wait_table = (wait_queue_head_t *)
4302 memblock_virt_alloc_node_nopanic(
4303 alloc_size, zone->zone_pgdat->node_id);
4306 * This case means that a zone whose size was 0 gets new memory
4307 * via memory hot-add.
4308 * But it may be the case that a new node was hot-added. In
4309 * this case vmalloc() will not be able to use this new node's
4310 * memory - this wait_table must be initialized to use this new
4311 * node itself as well.
4312 * To use this new node's memory, further consideration will be
4315 zone->wait_table = vmalloc(alloc_size);
4317 if (!zone->wait_table)
4320 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4321 init_waitqueue_head(zone->wait_table + i);
4326 static __meminit void zone_pcp_init(struct zone *zone)
4329 * per cpu subsystem is not up at this point. The following code
4330 * relies on the ability of the linker to provide the
4331 * offset of a (static) per cpu variable into the per cpu area.
4333 zone->pageset = &boot_pageset;
4335 if (populated_zone(zone))
4336 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4337 zone->name, zone->present_pages,
4338 zone_batchsize(zone));
4341 int __meminit init_currently_empty_zone(struct zone *zone,
4342 unsigned long zone_start_pfn,
4344 enum memmap_context context)
4346 struct pglist_data *pgdat = zone->zone_pgdat;
4348 ret = zone_wait_table_init(zone, size);
4351 pgdat->nr_zones = zone_idx(zone) + 1;
4353 zone->zone_start_pfn = zone_start_pfn;
4355 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4356 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4358 (unsigned long)zone_idx(zone),
4359 zone_start_pfn, (zone_start_pfn + size));
4361 zone_init_free_lists(zone);
4366 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4367 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4369 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4370 * Architectures may implement their own version but if add_active_range()
4371 * was used and there are no special requirements, this is a convenient
4374 int __meminit __early_pfn_to_nid(unsigned long pfn)
4376 unsigned long start_pfn, end_pfn;
4379 * NOTE: The following SMP-unsafe globals are only used early in boot
4380 * when the kernel is running single-threaded.
4382 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4383 static int __meminitdata last_nid;
4385 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4388 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4390 last_start_pfn = start_pfn;
4391 last_end_pfn = end_pfn;
4397 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4399 int __meminit early_pfn_to_nid(unsigned long pfn)
4403 nid = __early_pfn_to_nid(pfn);
4406 /* just returns 0 */
4410 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4411 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4415 nid = __early_pfn_to_nid(pfn);
4416 if (nid >= 0 && nid != node)
4423 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4424 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4425 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4427 * If an architecture guarantees that all ranges registered with
4428 * add_active_ranges() contain no holes and may be freed, this
4429 * this function may be used instead of calling memblock_free_early_nid()
4432 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4434 unsigned long start_pfn, end_pfn;
4437 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4438 start_pfn = min(start_pfn, max_low_pfn);
4439 end_pfn = min(end_pfn, max_low_pfn);
4441 if (start_pfn < end_pfn)
4442 memblock_free_early_nid(PFN_PHYS(start_pfn),
4443 (end_pfn - start_pfn) << PAGE_SHIFT,
4449 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4450 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4452 * If an architecture guarantees that all ranges registered with
4453 * add_active_ranges() contain no holes and may be freed, this
4454 * function may be used instead of calling memory_present() manually.
4456 void __init sparse_memory_present_with_active_regions(int nid)
4458 unsigned long start_pfn, end_pfn;
4461 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4462 memory_present(this_nid, start_pfn, end_pfn);
4466 * get_pfn_range_for_nid - Return the start and end page frames for a node
4467 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4468 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4469 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4471 * It returns the start and end page frame of a node based on information
4472 * provided by an arch calling add_active_range(). If called for a node
4473 * with no available memory, a warning is printed and the start and end
4476 void __meminit get_pfn_range_for_nid(unsigned int nid,
4477 unsigned long *start_pfn, unsigned long *end_pfn)
4479 unsigned long this_start_pfn, this_end_pfn;
4485 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4486 *start_pfn = min(*start_pfn, this_start_pfn);
4487 *end_pfn = max(*end_pfn, this_end_pfn);
4490 if (*start_pfn == -1UL)
4495 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4496 * assumption is made that zones within a node are ordered in monotonic
4497 * increasing memory addresses so that the "highest" populated zone is used
4499 static void __init find_usable_zone_for_movable(void)
4502 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4503 if (zone_index == ZONE_MOVABLE)
4506 if (arch_zone_highest_possible_pfn[zone_index] >
4507 arch_zone_lowest_possible_pfn[zone_index])
4511 VM_BUG_ON(zone_index == -1);
4512 movable_zone = zone_index;
4516 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4517 * because it is sized independent of architecture. Unlike the other zones,
4518 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4519 * in each node depending on the size of each node and how evenly kernelcore
4520 * is distributed. This helper function adjusts the zone ranges
4521 * provided by the architecture for a given node by using the end of the
4522 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4523 * zones within a node are in order of monotonic increases memory addresses
4525 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4526 unsigned long zone_type,
4527 unsigned long node_start_pfn,
4528 unsigned long node_end_pfn,
4529 unsigned long *zone_start_pfn,
4530 unsigned long *zone_end_pfn)
4532 /* Only adjust if ZONE_MOVABLE is on this node */
4533 if (zone_movable_pfn[nid]) {
4534 /* Size ZONE_MOVABLE */
4535 if (zone_type == ZONE_MOVABLE) {
4536 *zone_start_pfn = zone_movable_pfn[nid];
4537 *zone_end_pfn = min(node_end_pfn,
4538 arch_zone_highest_possible_pfn[movable_zone]);
4540 /* Adjust for ZONE_MOVABLE starting within this range */
4541 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4542 *zone_end_pfn > zone_movable_pfn[nid]) {
4543 *zone_end_pfn = zone_movable_pfn[nid];
4545 /* Check if this whole range is within ZONE_MOVABLE */
4546 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4547 *zone_start_pfn = *zone_end_pfn;
4552 * Return the number of pages a zone spans in a node, including holes
4553 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4555 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4556 unsigned long zone_type,
4557 unsigned long node_start_pfn,
4558 unsigned long node_end_pfn,
4559 unsigned long *ignored)
4561 unsigned long zone_start_pfn, zone_end_pfn;
4563 /* Get the start and end of the zone */
4564 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4565 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4566 adjust_zone_range_for_zone_movable(nid, zone_type,
4567 node_start_pfn, node_end_pfn,
4568 &zone_start_pfn, &zone_end_pfn);
4570 /* Check that this node has pages within the zone's required range */
4571 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4574 /* Move the zone boundaries inside the node if necessary */
4575 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4576 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4578 /* Return the spanned pages */
4579 return zone_end_pfn - zone_start_pfn;
4583 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4584 * then all holes in the requested range will be accounted for.
4586 unsigned long __meminit __absent_pages_in_range(int nid,
4587 unsigned long range_start_pfn,
4588 unsigned long range_end_pfn)
4590 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4591 unsigned long start_pfn, end_pfn;
4594 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4595 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4596 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4597 nr_absent -= end_pfn - start_pfn;
4603 * absent_pages_in_range - Return number of page frames in holes within a range
4604 * @start_pfn: The start PFN to start searching for holes
4605 * @end_pfn: The end PFN to stop searching for holes
4607 * It returns the number of pages frames in memory holes within a range.
4609 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4610 unsigned long end_pfn)
4612 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4615 /* Return the number of page frames in holes in a zone on a node */
4616 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4617 unsigned long zone_type,
4618 unsigned long node_start_pfn,
4619 unsigned long node_end_pfn,
4620 unsigned long *ignored)
4622 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4623 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4624 unsigned long zone_start_pfn, zone_end_pfn;
4626 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4627 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4629 adjust_zone_range_for_zone_movable(nid, zone_type,
4630 node_start_pfn, node_end_pfn,
4631 &zone_start_pfn, &zone_end_pfn);
4632 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4635 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4636 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4637 unsigned long zone_type,
4638 unsigned long node_start_pfn,
4639 unsigned long node_end_pfn,
4640 unsigned long *zones_size)
4642 return zones_size[zone_type];
4645 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4646 unsigned long zone_type,
4647 unsigned long node_start_pfn,
4648 unsigned long node_end_pfn,
4649 unsigned long *zholes_size)
4654 return zholes_size[zone_type];
4657 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4659 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4660 unsigned long node_start_pfn,
4661 unsigned long node_end_pfn,
4662 unsigned long *zones_size,
4663 unsigned long *zholes_size)
4665 unsigned long realtotalpages, totalpages = 0;
4668 for (i = 0; i < MAX_NR_ZONES; i++)
4669 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4673 pgdat->node_spanned_pages = totalpages;
4675 realtotalpages = totalpages;
4676 for (i = 0; i < MAX_NR_ZONES; i++)
4678 zone_absent_pages_in_node(pgdat->node_id, i,
4679 node_start_pfn, node_end_pfn,
4681 pgdat->node_present_pages = realtotalpages;
4682 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4686 #ifndef CONFIG_SPARSEMEM
4688 * Calculate the size of the zone->blockflags rounded to an unsigned long
4689 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4690 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4691 * round what is now in bits to nearest long in bits, then return it in
4694 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4696 unsigned long usemapsize;
4698 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4699 usemapsize = roundup(zonesize, pageblock_nr_pages);
4700 usemapsize = usemapsize >> pageblock_order;
4701 usemapsize *= NR_PAGEBLOCK_BITS;
4702 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4704 return usemapsize / 8;
4707 static void __init setup_usemap(struct pglist_data *pgdat,
4709 unsigned long zone_start_pfn,
4710 unsigned long zonesize)
4712 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4713 zone->pageblock_flags = NULL;
4715 zone->pageblock_flags =
4716 memblock_virt_alloc_node_nopanic(usemapsize,
4720 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4721 unsigned long zone_start_pfn, unsigned long zonesize) {}
4722 #endif /* CONFIG_SPARSEMEM */
4724 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4726 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4727 void __paginginit set_pageblock_order(void)
4731 /* Check that pageblock_nr_pages has not already been setup */
4732 if (pageblock_order)
4735 if (HPAGE_SHIFT > PAGE_SHIFT)
4736 order = HUGETLB_PAGE_ORDER;
4738 order = MAX_ORDER - 1;
4741 * Assume the largest contiguous order of interest is a huge page.
4742 * This value may be variable depending on boot parameters on IA64 and
4745 pageblock_order = order;
4747 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4750 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4751 * is unused as pageblock_order is set at compile-time. See
4752 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4755 void __paginginit set_pageblock_order(void)
4759 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4761 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4762 unsigned long present_pages)
4764 unsigned long pages = spanned_pages;
4767 * Provide a more accurate estimation if there are holes within
4768 * the zone and SPARSEMEM is in use. If there are holes within the
4769 * zone, each populated memory region may cost us one or two extra
4770 * memmap pages due to alignment because memmap pages for each
4771 * populated regions may not naturally algined on page boundary.
4772 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4774 if (spanned_pages > present_pages + (present_pages >> 4) &&
4775 IS_ENABLED(CONFIG_SPARSEMEM))
4776 pages = present_pages;
4778 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4782 * Set up the zone data structures:
4783 * - mark all pages reserved
4784 * - mark all memory queues empty
4785 * - clear the memory bitmaps
4787 * NOTE: pgdat should get zeroed by caller.
4789 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4790 unsigned long node_start_pfn, unsigned long node_end_pfn,
4791 unsigned long *zones_size, unsigned long *zholes_size)
4794 int nid = pgdat->node_id;
4795 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4798 pgdat_resize_init(pgdat);
4799 #ifdef CONFIG_NUMA_BALANCING
4800 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4801 pgdat->numabalancing_migrate_nr_pages = 0;
4802 pgdat->numabalancing_migrate_next_window = jiffies;
4804 init_waitqueue_head(&pgdat->kswapd_wait);
4805 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4806 pgdat_page_cgroup_init(pgdat);
4808 for (j = 0; j < MAX_NR_ZONES; j++) {
4809 struct zone *zone = pgdat->node_zones + j;
4810 unsigned long size, realsize, freesize, memmap_pages;
4812 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4813 node_end_pfn, zones_size);
4814 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4820 * Adjust freesize so that it accounts for how much memory
4821 * is used by this zone for memmap. This affects the watermark
4822 * and per-cpu initialisations
4824 memmap_pages = calc_memmap_size(size, realsize);
4825 if (freesize >= memmap_pages) {
4826 freesize -= memmap_pages;
4829 " %s zone: %lu pages used for memmap\n",
4830 zone_names[j], memmap_pages);
4833 " %s zone: %lu pages exceeds freesize %lu\n",
4834 zone_names[j], memmap_pages, freesize);
4836 /* Account for reserved pages */
4837 if (j == 0 && freesize > dma_reserve) {
4838 freesize -= dma_reserve;
4839 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4840 zone_names[0], dma_reserve);
4843 if (!is_highmem_idx(j))
4844 nr_kernel_pages += freesize;
4845 /* Charge for highmem memmap if there are enough kernel pages */
4846 else if (nr_kernel_pages > memmap_pages * 2)
4847 nr_kernel_pages -= memmap_pages;
4848 nr_all_pages += freesize;
4850 zone->spanned_pages = size;
4851 zone->present_pages = realsize;
4853 * Set an approximate value for lowmem here, it will be adjusted
4854 * when the bootmem allocator frees pages into the buddy system.
4855 * And all highmem pages will be managed by the buddy system.
4857 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4860 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4862 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4864 zone->name = zone_names[j];
4865 spin_lock_init(&zone->lock);
4866 spin_lock_init(&zone->lru_lock);
4867 zone_seqlock_init(zone);
4868 zone->zone_pgdat = pgdat;
4869 zone_pcp_init(zone);
4871 /* For bootup, initialized properly in watermark setup */
4872 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4874 lruvec_init(&zone->lruvec);
4878 set_pageblock_order();
4879 setup_usemap(pgdat, zone, zone_start_pfn, size);
4880 ret = init_currently_empty_zone(zone, zone_start_pfn,
4881 size, MEMMAP_EARLY);
4883 memmap_init(size, nid, j, zone_start_pfn);
4884 zone_start_pfn += size;
4888 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4890 /* Skip empty nodes */
4891 if (!pgdat->node_spanned_pages)
4894 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4895 /* ia64 gets its own node_mem_map, before this, without bootmem */
4896 if (!pgdat->node_mem_map) {
4897 unsigned long size, start, end;
4901 * The zone's endpoints aren't required to be MAX_ORDER
4902 * aligned but the node_mem_map endpoints must be in order
4903 * for the buddy allocator to function correctly.
4905 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4906 end = pgdat_end_pfn(pgdat);
4907 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4908 size = (end - start) * sizeof(struct page);
4909 map = alloc_remap(pgdat->node_id, size);
4911 map = memblock_virt_alloc_node_nopanic(size,
4913 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4915 #ifndef CONFIG_NEED_MULTIPLE_NODES
4917 * With no DISCONTIG, the global mem_map is just set as node 0's
4919 if (pgdat == NODE_DATA(0)) {
4920 mem_map = NODE_DATA(0)->node_mem_map;
4921 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4922 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4923 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4924 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4927 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4930 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4931 unsigned long node_start_pfn, unsigned long *zholes_size)
4933 pg_data_t *pgdat = NODE_DATA(nid);
4934 unsigned long start_pfn = 0;
4935 unsigned long end_pfn = 0;
4937 /* pg_data_t should be reset to zero when it's allocated */
4938 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4940 pgdat->node_id = nid;
4941 pgdat->node_start_pfn = node_start_pfn;
4942 if (node_state(nid, N_MEMORY))
4943 init_zone_allows_reclaim(nid);
4944 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4945 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4947 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4948 zones_size, zholes_size);
4950 alloc_node_mem_map(pgdat);
4951 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4952 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4953 nid, (unsigned long)pgdat,
4954 (unsigned long)pgdat->node_mem_map);
4957 free_area_init_core(pgdat, start_pfn, end_pfn,
4958 zones_size, zholes_size);
4961 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4963 #if MAX_NUMNODES > 1
4965 * Figure out the number of possible node ids.
4967 void __init setup_nr_node_ids(void)
4970 unsigned int highest = 0;
4972 for_each_node_mask(node, node_possible_map)
4974 nr_node_ids = highest + 1;
4979 * node_map_pfn_alignment - determine the maximum internode alignment
4981 * This function should be called after node map is populated and sorted.
4982 * It calculates the maximum power of two alignment which can distinguish
4985 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4986 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4987 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4988 * shifted, 1GiB is enough and this function will indicate so.
4990 * This is used to test whether pfn -> nid mapping of the chosen memory
4991 * model has fine enough granularity to avoid incorrect mapping for the
4992 * populated node map.
4994 * Returns the determined alignment in pfn's. 0 if there is no alignment
4995 * requirement (single node).
4997 unsigned long __init node_map_pfn_alignment(void)
4999 unsigned long accl_mask = 0, last_end = 0;
5000 unsigned long start, end, mask;
5004 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5005 if (!start || last_nid < 0 || last_nid == nid) {
5012 * Start with a mask granular enough to pin-point to the
5013 * start pfn and tick off bits one-by-one until it becomes
5014 * too coarse to separate the current node from the last.
5016 mask = ~((1 << __ffs(start)) - 1);
5017 while (mask && last_end <= (start & (mask << 1)))
5020 /* accumulate all internode masks */
5024 /* convert mask to number of pages */
5025 return ~accl_mask + 1;
5028 /* Find the lowest pfn for a node */
5029 static unsigned long __init find_min_pfn_for_node(int nid)
5031 unsigned long min_pfn = ULONG_MAX;
5032 unsigned long start_pfn;
5035 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5036 min_pfn = min(min_pfn, start_pfn);
5038 if (min_pfn == ULONG_MAX) {
5040 "Could not find start_pfn for node %d\n", nid);
5048 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5050 * It returns the minimum PFN based on information provided via
5051 * add_active_range().
5053 unsigned long __init find_min_pfn_with_active_regions(void)
5055 return find_min_pfn_for_node(MAX_NUMNODES);
5059 * early_calculate_totalpages()
5060 * Sum pages in active regions for movable zone.
5061 * Populate N_MEMORY for calculating usable_nodes.
5063 static unsigned long __init early_calculate_totalpages(void)
5065 unsigned long totalpages = 0;
5066 unsigned long start_pfn, end_pfn;
5069 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5070 unsigned long pages = end_pfn - start_pfn;
5072 totalpages += pages;
5074 node_set_state(nid, N_MEMORY);
5080 * Find the PFN the Movable zone begins in each node. Kernel memory
5081 * is spread evenly between nodes as long as the nodes have enough
5082 * memory. When they don't, some nodes will have more kernelcore than
5085 static void __init find_zone_movable_pfns_for_nodes(void)
5088 unsigned long usable_startpfn;
5089 unsigned long kernelcore_node, kernelcore_remaining;
5090 /* save the state before borrow the nodemask */
5091 nodemask_t saved_node_state = node_states[N_MEMORY];
5092 unsigned long totalpages = early_calculate_totalpages();
5093 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5094 struct memblock_type *type = &memblock.memory;
5096 /* Need to find movable_zone earlier when movable_node is specified. */
5097 find_usable_zone_for_movable();
5100 * If movable_node is specified, ignore kernelcore and movablecore
5103 if (movable_node_is_enabled()) {
5104 for (i = 0; i < type->cnt; i++) {
5105 if (!memblock_is_hotpluggable(&type->regions[i]))
5108 nid = type->regions[i].nid;
5110 usable_startpfn = PFN_DOWN(type->regions[i].base);
5111 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5112 min(usable_startpfn, zone_movable_pfn[nid]) :
5120 * If movablecore=nn[KMG] was specified, calculate what size of
5121 * kernelcore that corresponds so that memory usable for
5122 * any allocation type is evenly spread. If both kernelcore
5123 * and movablecore are specified, then the value of kernelcore
5124 * will be used for required_kernelcore if it's greater than
5125 * what movablecore would have allowed.
5127 if (required_movablecore) {
5128 unsigned long corepages;
5131 * Round-up so that ZONE_MOVABLE is at least as large as what
5132 * was requested by the user
5134 required_movablecore =
5135 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5136 corepages = totalpages - required_movablecore;
5138 required_kernelcore = max(required_kernelcore, corepages);
5141 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5142 if (!required_kernelcore)
5145 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5146 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5149 /* Spread kernelcore memory as evenly as possible throughout nodes */
5150 kernelcore_node = required_kernelcore / usable_nodes;
5151 for_each_node_state(nid, N_MEMORY) {
5152 unsigned long start_pfn, end_pfn;
5155 * Recalculate kernelcore_node if the division per node
5156 * now exceeds what is necessary to satisfy the requested
5157 * amount of memory for the kernel
5159 if (required_kernelcore < kernelcore_node)
5160 kernelcore_node = required_kernelcore / usable_nodes;
5163 * As the map is walked, we track how much memory is usable
5164 * by the kernel using kernelcore_remaining. When it is
5165 * 0, the rest of the node is usable by ZONE_MOVABLE
5167 kernelcore_remaining = kernelcore_node;
5169 /* Go through each range of PFNs within this node */
5170 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5171 unsigned long size_pages;
5173 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5174 if (start_pfn >= end_pfn)
5177 /* Account for what is only usable for kernelcore */
5178 if (start_pfn < usable_startpfn) {
5179 unsigned long kernel_pages;
5180 kernel_pages = min(end_pfn, usable_startpfn)
5183 kernelcore_remaining -= min(kernel_pages,
5184 kernelcore_remaining);
5185 required_kernelcore -= min(kernel_pages,
5186 required_kernelcore);
5188 /* Continue if range is now fully accounted */
5189 if (end_pfn <= usable_startpfn) {
5192 * Push zone_movable_pfn to the end so
5193 * that if we have to rebalance
5194 * kernelcore across nodes, we will
5195 * not double account here
5197 zone_movable_pfn[nid] = end_pfn;
5200 start_pfn = usable_startpfn;
5204 * The usable PFN range for ZONE_MOVABLE is from
5205 * start_pfn->end_pfn. Calculate size_pages as the
5206 * number of pages used as kernelcore
5208 size_pages = end_pfn - start_pfn;
5209 if (size_pages > kernelcore_remaining)
5210 size_pages = kernelcore_remaining;
5211 zone_movable_pfn[nid] = start_pfn + size_pages;
5214 * Some kernelcore has been met, update counts and
5215 * break if the kernelcore for this node has been
5218 required_kernelcore -= min(required_kernelcore,
5220 kernelcore_remaining -= size_pages;
5221 if (!kernelcore_remaining)
5227 * If there is still required_kernelcore, we do another pass with one
5228 * less node in the count. This will push zone_movable_pfn[nid] further
5229 * along on the nodes that still have memory until kernelcore is
5233 if (usable_nodes && required_kernelcore > usable_nodes)
5237 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5238 for (nid = 0; nid < MAX_NUMNODES; nid++)
5239 zone_movable_pfn[nid] =
5240 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5243 /* restore the node_state */
5244 node_states[N_MEMORY] = saved_node_state;
5247 /* Any regular or high memory on that node ? */
5248 static void check_for_memory(pg_data_t *pgdat, int nid)
5250 enum zone_type zone_type;
5252 if (N_MEMORY == N_NORMAL_MEMORY)
5255 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5256 struct zone *zone = &pgdat->node_zones[zone_type];
5257 if (populated_zone(zone)) {
5258 node_set_state(nid, N_HIGH_MEMORY);
5259 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5260 zone_type <= ZONE_NORMAL)
5261 node_set_state(nid, N_NORMAL_MEMORY);
5268 * free_area_init_nodes - Initialise all pg_data_t and zone data
5269 * @max_zone_pfn: an array of max PFNs for each zone
5271 * This will call free_area_init_node() for each active node in the system.
5272 * Using the page ranges provided by add_active_range(), the size of each
5273 * zone in each node and their holes is calculated. If the maximum PFN
5274 * between two adjacent zones match, it is assumed that the zone is empty.
5275 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5276 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5277 * starts where the previous one ended. For example, ZONE_DMA32 starts
5278 * at arch_max_dma_pfn.
5280 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5282 unsigned long start_pfn, end_pfn;
5285 /* Record where the zone boundaries are */
5286 memset(arch_zone_lowest_possible_pfn, 0,
5287 sizeof(arch_zone_lowest_possible_pfn));
5288 memset(arch_zone_highest_possible_pfn, 0,
5289 sizeof(arch_zone_highest_possible_pfn));
5290 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5291 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5292 for (i = 1; i < MAX_NR_ZONES; i++) {
5293 if (i == ZONE_MOVABLE)
5295 arch_zone_lowest_possible_pfn[i] =
5296 arch_zone_highest_possible_pfn[i-1];
5297 arch_zone_highest_possible_pfn[i] =
5298 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5300 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5301 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5303 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5304 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5305 find_zone_movable_pfns_for_nodes();
5307 /* Print out the zone ranges */
5308 printk("Zone ranges:\n");
5309 for (i = 0; i < MAX_NR_ZONES; i++) {
5310 if (i == ZONE_MOVABLE)
5312 printk(KERN_CONT " %-8s ", zone_names[i]);
5313 if (arch_zone_lowest_possible_pfn[i] ==
5314 arch_zone_highest_possible_pfn[i])
5315 printk(KERN_CONT "empty\n");
5317 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5318 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5319 (arch_zone_highest_possible_pfn[i]
5320 << PAGE_SHIFT) - 1);
5323 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5324 printk("Movable zone start for each node\n");
5325 for (i = 0; i < MAX_NUMNODES; i++) {
5326 if (zone_movable_pfn[i])
5327 printk(" Node %d: %#010lx\n", i,
5328 zone_movable_pfn[i] << PAGE_SHIFT);
5331 /* Print out the early node map */
5332 printk("Early memory node ranges\n");
5333 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5334 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5335 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5337 /* Initialise every node */
5338 mminit_verify_pageflags_layout();
5339 setup_nr_node_ids();
5340 for_each_online_node(nid) {
5341 pg_data_t *pgdat = NODE_DATA(nid);
5342 free_area_init_node(nid, NULL,
5343 find_min_pfn_for_node(nid), NULL);
5345 /* Any memory on that node */
5346 if (pgdat->node_present_pages)
5347 node_set_state(nid, N_MEMORY);
5348 check_for_memory(pgdat, nid);
5352 static int __init cmdline_parse_core(char *p, unsigned long *core)
5354 unsigned long long coremem;
5358 coremem = memparse(p, &p);
5359 *core = coremem >> PAGE_SHIFT;
5361 /* Paranoid check that UL is enough for the coremem value */
5362 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5368 * kernelcore=size sets the amount of memory for use for allocations that
5369 * cannot be reclaimed or migrated.
5371 static int __init cmdline_parse_kernelcore(char *p)
5373 return cmdline_parse_core(p, &required_kernelcore);
5377 * movablecore=size sets the amount of memory for use for allocations that
5378 * can be reclaimed or migrated.
5380 static int __init cmdline_parse_movablecore(char *p)
5382 return cmdline_parse_core(p, &required_movablecore);
5385 early_param("kernelcore", cmdline_parse_kernelcore);
5386 early_param("movablecore", cmdline_parse_movablecore);
5388 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5390 void adjust_managed_page_count(struct page *page, long count)
5392 spin_lock(&managed_page_count_lock);
5393 page_zone(page)->managed_pages += count;
5394 totalram_pages += count;
5395 #ifdef CONFIG_HIGHMEM
5396 if (PageHighMem(page))
5397 totalhigh_pages += count;
5399 spin_unlock(&managed_page_count_lock);
5401 EXPORT_SYMBOL(adjust_managed_page_count);
5403 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5406 unsigned long pages = 0;
5408 start = (void *)PAGE_ALIGN((unsigned long)start);
5409 end = (void *)((unsigned long)end & PAGE_MASK);
5410 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5411 if ((unsigned int)poison <= 0xFF)
5412 memset(pos, poison, PAGE_SIZE);
5413 free_reserved_page(virt_to_page(pos));
5417 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5418 s, pages << (PAGE_SHIFT - 10), start, end);
5422 EXPORT_SYMBOL(free_reserved_area);
5424 #ifdef CONFIG_HIGHMEM
5425 void free_highmem_page(struct page *page)
5427 __free_reserved_page(page);
5429 page_zone(page)->managed_pages++;
5435 void __init mem_init_print_info(const char *str)
5437 unsigned long physpages, codesize, datasize, rosize, bss_size;
5438 unsigned long init_code_size, init_data_size;
5440 physpages = get_num_physpages();
5441 codesize = _etext - _stext;
5442 datasize = _edata - _sdata;
5443 rosize = __end_rodata - __start_rodata;
5444 bss_size = __bss_stop - __bss_start;
5445 init_data_size = __init_end - __init_begin;
5446 init_code_size = _einittext - _sinittext;
5449 * Detect special cases and adjust section sizes accordingly:
5450 * 1) .init.* may be embedded into .data sections
5451 * 2) .init.text.* may be out of [__init_begin, __init_end],
5452 * please refer to arch/tile/kernel/vmlinux.lds.S.
5453 * 3) .rodata.* may be embedded into .text or .data sections.
5455 #define adj_init_size(start, end, size, pos, adj) \
5457 if (start <= pos && pos < end && size > adj) \
5461 adj_init_size(__init_begin, __init_end, init_data_size,
5462 _sinittext, init_code_size);
5463 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5464 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5465 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5466 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5468 #undef adj_init_size
5470 printk("Memory: %luK/%luK available "
5471 "(%luK kernel code, %luK rwdata, %luK rodata, "
5472 "%luK init, %luK bss, %luK reserved"
5473 #ifdef CONFIG_HIGHMEM
5477 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5478 codesize >> 10, datasize >> 10, rosize >> 10,
5479 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5480 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5481 #ifdef CONFIG_HIGHMEM
5482 totalhigh_pages << (PAGE_SHIFT-10),
5484 str ? ", " : "", str ? str : "");
5488 * set_dma_reserve - set the specified number of pages reserved in the first zone
5489 * @new_dma_reserve: The number of pages to mark reserved
5491 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5492 * In the DMA zone, a significant percentage may be consumed by kernel image
5493 * and other unfreeable allocations which can skew the watermarks badly. This
5494 * function may optionally be used to account for unfreeable pages in the
5495 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5496 * smaller per-cpu batchsize.
5498 void __init set_dma_reserve(unsigned long new_dma_reserve)
5500 dma_reserve = new_dma_reserve;
5503 void __init free_area_init(unsigned long *zones_size)
5505 free_area_init_node(0, zones_size,
5506 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5509 static int page_alloc_cpu_notify(struct notifier_block *self,
5510 unsigned long action, void *hcpu)
5512 int cpu = (unsigned long)hcpu;
5514 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5515 lru_add_drain_cpu(cpu);
5519 * Spill the event counters of the dead processor
5520 * into the current processors event counters.
5521 * This artificially elevates the count of the current
5524 vm_events_fold_cpu(cpu);
5527 * Zero the differential counters of the dead processor
5528 * so that the vm statistics are consistent.
5530 * This is only okay since the processor is dead and cannot
5531 * race with what we are doing.
5533 cpu_vm_stats_fold(cpu);
5538 void __init page_alloc_init(void)
5540 hotcpu_notifier(page_alloc_cpu_notify, 0);
5544 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5545 * or min_free_kbytes changes.
5547 static void calculate_totalreserve_pages(void)
5549 struct pglist_data *pgdat;
5550 unsigned long reserve_pages = 0;
5551 enum zone_type i, j;
5553 for_each_online_pgdat(pgdat) {
5554 for (i = 0; i < MAX_NR_ZONES; i++) {
5555 struct zone *zone = pgdat->node_zones + i;
5556 unsigned long max = 0;
5558 /* Find valid and maximum lowmem_reserve in the zone */
5559 for (j = i; j < MAX_NR_ZONES; j++) {
5560 if (zone->lowmem_reserve[j] > max)
5561 max = zone->lowmem_reserve[j];
5564 /* we treat the high watermark as reserved pages. */
5565 max += high_wmark_pages(zone);
5567 if (max > zone->managed_pages)
5568 max = zone->managed_pages;
5569 reserve_pages += max;
5571 * Lowmem reserves are not available to
5572 * GFP_HIGHUSER page cache allocations and
5573 * kswapd tries to balance zones to their high
5574 * watermark. As a result, neither should be
5575 * regarded as dirtyable memory, to prevent a
5576 * situation where reclaim has to clean pages
5577 * in order to balance the zones.
5579 zone->dirty_balance_reserve = max;
5582 dirty_balance_reserve = reserve_pages;
5583 totalreserve_pages = reserve_pages;
5587 * setup_per_zone_lowmem_reserve - called whenever
5588 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5589 * has a correct pages reserved value, so an adequate number of
5590 * pages are left in the zone after a successful __alloc_pages().
5592 static void setup_per_zone_lowmem_reserve(void)
5594 struct pglist_data *pgdat;
5595 enum zone_type j, idx;
5597 for_each_online_pgdat(pgdat) {
5598 for (j = 0; j < MAX_NR_ZONES; j++) {
5599 struct zone *zone = pgdat->node_zones + j;
5600 unsigned long managed_pages = zone->managed_pages;
5602 zone->lowmem_reserve[j] = 0;
5606 struct zone *lower_zone;
5610 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5611 sysctl_lowmem_reserve_ratio[idx] = 1;
5613 lower_zone = pgdat->node_zones + idx;
5614 lower_zone->lowmem_reserve[j] = managed_pages /
5615 sysctl_lowmem_reserve_ratio[idx];
5616 managed_pages += lower_zone->managed_pages;
5621 /* update totalreserve_pages */
5622 calculate_totalreserve_pages();
5625 static void __setup_per_zone_wmarks(void)
5627 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5628 unsigned long lowmem_pages = 0;
5630 unsigned long flags;
5632 /* Calculate total number of !ZONE_HIGHMEM pages */
5633 for_each_zone(zone) {
5634 if (!is_highmem(zone))
5635 lowmem_pages += zone->managed_pages;
5638 for_each_zone(zone) {
5641 spin_lock_irqsave(&zone->lock, flags);
5642 tmp = (u64)pages_min * zone->managed_pages;
5643 do_div(tmp, lowmem_pages);
5644 if (is_highmem(zone)) {
5646 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5647 * need highmem pages, so cap pages_min to a small
5650 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5651 * deltas controls asynch page reclaim, and so should
5652 * not be capped for highmem.
5654 unsigned long min_pages;
5656 min_pages = zone->managed_pages / 1024;
5657 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5658 zone->watermark[WMARK_MIN] = min_pages;
5661 * If it's a lowmem zone, reserve a number of pages
5662 * proportionate to the zone's size.
5664 zone->watermark[WMARK_MIN] = tmp;
5667 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5668 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5670 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5671 high_wmark_pages(zone) - low_wmark_pages(zone) -
5672 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5674 setup_zone_migrate_reserve(zone);
5675 spin_unlock_irqrestore(&zone->lock, flags);
5678 /* update totalreserve_pages */
5679 calculate_totalreserve_pages();
5683 * setup_per_zone_wmarks - called when min_free_kbytes changes
5684 * or when memory is hot-{added|removed}
5686 * Ensures that the watermark[min,low,high] values for each zone are set
5687 * correctly with respect to min_free_kbytes.
5689 void setup_per_zone_wmarks(void)
5691 mutex_lock(&zonelists_mutex);
5692 __setup_per_zone_wmarks();
5693 mutex_unlock(&zonelists_mutex);
5697 * The inactive anon list should be small enough that the VM never has to
5698 * do too much work, but large enough that each inactive page has a chance
5699 * to be referenced again before it is swapped out.
5701 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5702 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5703 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5704 * the anonymous pages are kept on the inactive list.
5707 * memory ratio inactive anon
5708 * -------------------------------------
5717 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5719 unsigned int gb, ratio;
5721 /* Zone size in gigabytes */
5722 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5724 ratio = int_sqrt(10 * gb);
5728 zone->inactive_ratio = ratio;
5731 static void __meminit setup_per_zone_inactive_ratio(void)
5736 calculate_zone_inactive_ratio(zone);
5740 * Initialise min_free_kbytes.
5742 * For small machines we want it small (128k min). For large machines
5743 * we want it large (64MB max). But it is not linear, because network
5744 * bandwidth does not increase linearly with machine size. We use
5746 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5747 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5763 int __meminit init_per_zone_wmark_min(void)
5765 unsigned long lowmem_kbytes;
5766 int new_min_free_kbytes;
5768 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5769 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5771 if (new_min_free_kbytes > user_min_free_kbytes) {
5772 min_free_kbytes = new_min_free_kbytes;
5773 if (min_free_kbytes < 128)
5774 min_free_kbytes = 128;
5775 if (min_free_kbytes > 65536)
5776 min_free_kbytes = 65536;
5778 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5779 new_min_free_kbytes, user_min_free_kbytes);
5781 setup_per_zone_wmarks();
5782 refresh_zone_stat_thresholds();
5783 setup_per_zone_lowmem_reserve();
5784 setup_per_zone_inactive_ratio();
5787 module_init(init_per_zone_wmark_min)
5790 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5791 * that we can call two helper functions whenever min_free_kbytes
5794 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5795 void __user *buffer, size_t *length, loff_t *ppos)
5799 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5804 user_min_free_kbytes = min_free_kbytes;
5805 setup_per_zone_wmarks();
5811 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5812 void __user *buffer, size_t *length, loff_t *ppos)
5817 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5822 zone->min_unmapped_pages = (zone->managed_pages *
5823 sysctl_min_unmapped_ratio) / 100;
5827 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5828 void __user *buffer, size_t *length, loff_t *ppos)
5833 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5838 zone->min_slab_pages = (zone->managed_pages *
5839 sysctl_min_slab_ratio) / 100;
5845 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5846 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5847 * whenever sysctl_lowmem_reserve_ratio changes.
5849 * The reserve ratio obviously has absolutely no relation with the
5850 * minimum watermarks. The lowmem reserve ratio can only make sense
5851 * if in function of the boot time zone sizes.
5853 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5854 void __user *buffer, size_t *length, loff_t *ppos)
5856 proc_dointvec_minmax(table, write, buffer, length, ppos);
5857 setup_per_zone_lowmem_reserve();
5862 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5863 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5864 * pagelist can have before it gets flushed back to buddy allocator.
5866 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5867 void __user *buffer, size_t *length, loff_t *ppos)
5870 int old_percpu_pagelist_fraction;
5873 mutex_lock(&pcp_batch_high_lock);
5874 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5876 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5877 if (!write || ret < 0)
5880 /* Sanity checking to avoid pcp imbalance */
5881 if (percpu_pagelist_fraction &&
5882 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
5883 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
5889 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
5892 for_each_populated_zone(zone) {
5895 for_each_possible_cpu(cpu)
5896 pageset_set_high_and_batch(zone,
5897 per_cpu_ptr(zone->pageset, cpu));
5900 mutex_unlock(&pcp_batch_high_lock);
5904 int hashdist = HASHDIST_DEFAULT;
5907 static int __init set_hashdist(char *str)
5911 hashdist = simple_strtoul(str, &str, 0);
5914 __setup("hashdist=", set_hashdist);
5918 * allocate a large system hash table from bootmem
5919 * - it is assumed that the hash table must contain an exact power-of-2
5920 * quantity of entries
5921 * - limit is the number of hash buckets, not the total allocation size
5923 void *__init alloc_large_system_hash(const char *tablename,
5924 unsigned long bucketsize,
5925 unsigned long numentries,
5928 unsigned int *_hash_shift,
5929 unsigned int *_hash_mask,
5930 unsigned long low_limit,
5931 unsigned long high_limit)
5933 unsigned long long max = high_limit;
5934 unsigned long log2qty, size;
5937 /* allow the kernel cmdline to have a say */
5939 /* round applicable memory size up to nearest megabyte */
5940 numentries = nr_kernel_pages;
5942 /* It isn't necessary when PAGE_SIZE >= 1MB */
5943 if (PAGE_SHIFT < 20)
5944 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5946 /* limit to 1 bucket per 2^scale bytes of low memory */
5947 if (scale > PAGE_SHIFT)
5948 numentries >>= (scale - PAGE_SHIFT);
5950 numentries <<= (PAGE_SHIFT - scale);
5952 /* Make sure we've got at least a 0-order allocation.. */
5953 if (unlikely(flags & HASH_SMALL)) {
5954 /* Makes no sense without HASH_EARLY */
5955 WARN_ON(!(flags & HASH_EARLY));
5956 if (!(numentries >> *_hash_shift)) {
5957 numentries = 1UL << *_hash_shift;
5958 BUG_ON(!numentries);
5960 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5961 numentries = PAGE_SIZE / bucketsize;
5963 numentries = roundup_pow_of_two(numentries);
5965 /* limit allocation size to 1/16 total memory by default */
5967 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5968 do_div(max, bucketsize);
5970 max = min(max, 0x80000000ULL);
5972 if (numentries < low_limit)
5973 numentries = low_limit;
5974 if (numentries > max)
5977 log2qty = ilog2(numentries);
5980 size = bucketsize << log2qty;
5981 if (flags & HASH_EARLY)
5982 table = memblock_virt_alloc_nopanic(size, 0);
5984 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5987 * If bucketsize is not a power-of-two, we may free
5988 * some pages at the end of hash table which
5989 * alloc_pages_exact() automatically does
5991 if (get_order(size) < MAX_ORDER) {
5992 table = alloc_pages_exact(size, GFP_ATOMIC);
5993 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5996 } while (!table && size > PAGE_SIZE && --log2qty);
5999 panic("Failed to allocate %s hash table\n", tablename);
6001 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6004 ilog2(size) - PAGE_SHIFT,
6008 *_hash_shift = log2qty;
6010 *_hash_mask = (1 << log2qty) - 1;
6015 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6016 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6019 #ifdef CONFIG_SPARSEMEM
6020 return __pfn_to_section(pfn)->pageblock_flags;
6022 return zone->pageblock_flags;
6023 #endif /* CONFIG_SPARSEMEM */
6026 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6028 #ifdef CONFIG_SPARSEMEM
6029 pfn &= (PAGES_PER_SECTION-1);
6030 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6032 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6033 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6034 #endif /* CONFIG_SPARSEMEM */
6038 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
6039 * @page: The page within the block of interest
6040 * @start_bitidx: The first bit of interest to retrieve
6041 * @end_bitidx: The last bit of interest
6042 * returns pageblock_bits flags
6044 unsigned long get_pageblock_flags_mask(struct page *page,
6045 unsigned long end_bitidx,
6049 unsigned long *bitmap;
6050 unsigned long pfn, bitidx, word_bitidx;
6053 zone = page_zone(page);
6054 pfn = page_to_pfn(page);
6055 bitmap = get_pageblock_bitmap(zone, pfn);
6056 bitidx = pfn_to_bitidx(zone, pfn);
6057 word_bitidx = bitidx / BITS_PER_LONG;
6058 bitidx &= (BITS_PER_LONG-1);
6060 word = bitmap[word_bitidx];
6061 bitidx += end_bitidx;
6062 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6066 * set_pageblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6067 * @page: The page within the block of interest
6068 * @start_bitidx: The first bit of interest
6069 * @end_bitidx: The last bit of interest
6070 * @flags: The flags to set
6072 void set_pageblock_flags_mask(struct page *page, unsigned long flags,
6073 unsigned long end_bitidx,
6077 unsigned long *bitmap;
6078 unsigned long pfn, bitidx, word_bitidx;
6079 unsigned long old_word, word;
6081 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6083 zone = page_zone(page);
6084 pfn = page_to_pfn(page);
6085 bitmap = get_pageblock_bitmap(zone, pfn);
6086 bitidx = pfn_to_bitidx(zone, pfn);
6087 word_bitidx = bitidx / BITS_PER_LONG;
6088 bitidx &= (BITS_PER_LONG-1);
6090 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6092 bitidx += end_bitidx;
6093 mask <<= (BITS_PER_LONG - bitidx - 1);
6094 flags <<= (BITS_PER_LONG - bitidx - 1);
6096 word = ACCESS_ONCE(bitmap[word_bitidx]);
6098 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6099 if (word == old_word)
6106 * This function checks whether pageblock includes unmovable pages or not.
6107 * If @count is not zero, it is okay to include less @count unmovable pages
6109 * PageLRU check without isolation or lru_lock could race so that
6110 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6111 * expect this function should be exact.
6113 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6114 bool skip_hwpoisoned_pages)
6116 unsigned long pfn, iter, found;
6120 * For avoiding noise data, lru_add_drain_all() should be called
6121 * If ZONE_MOVABLE, the zone never contains unmovable pages
6123 if (zone_idx(zone) == ZONE_MOVABLE)
6125 mt = get_pageblock_migratetype(page);
6126 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6129 pfn = page_to_pfn(page);
6130 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6131 unsigned long check = pfn + iter;
6133 if (!pfn_valid_within(check))
6136 page = pfn_to_page(check);
6139 * Hugepages are not in LRU lists, but they're movable.
6140 * We need not scan over tail pages bacause we don't
6141 * handle each tail page individually in migration.
6143 if (PageHuge(page)) {
6144 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6149 * We can't use page_count without pin a page
6150 * because another CPU can free compound page.
6151 * This check already skips compound tails of THP
6152 * because their page->_count is zero at all time.
6154 if (!atomic_read(&page->_count)) {
6155 if (PageBuddy(page))
6156 iter += (1 << page_order(page)) - 1;
6161 * The HWPoisoned page may be not in buddy system, and
6162 * page_count() is not 0.
6164 if (skip_hwpoisoned_pages && PageHWPoison(page))
6170 * If there are RECLAIMABLE pages, we need to check it.
6171 * But now, memory offline itself doesn't call shrink_slab()
6172 * and it still to be fixed.
6175 * If the page is not RAM, page_count()should be 0.
6176 * we don't need more check. This is an _used_ not-movable page.
6178 * The problematic thing here is PG_reserved pages. PG_reserved
6179 * is set to both of a memory hole page and a _used_ kernel
6188 bool is_pageblock_removable_nolock(struct page *page)
6194 * We have to be careful here because we are iterating over memory
6195 * sections which are not zone aware so we might end up outside of
6196 * the zone but still within the section.
6197 * We have to take care about the node as well. If the node is offline
6198 * its NODE_DATA will be NULL - see page_zone.
6200 if (!node_online(page_to_nid(page)))
6203 zone = page_zone(page);
6204 pfn = page_to_pfn(page);
6205 if (!zone_spans_pfn(zone, pfn))
6208 return !has_unmovable_pages(zone, page, 0, true);
6213 static unsigned long pfn_max_align_down(unsigned long pfn)
6215 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6216 pageblock_nr_pages) - 1);
6219 static unsigned long pfn_max_align_up(unsigned long pfn)
6221 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6222 pageblock_nr_pages));
6225 /* [start, end) must belong to a single zone. */
6226 static int __alloc_contig_migrate_range(struct compact_control *cc,
6227 unsigned long start, unsigned long end)
6229 /* This function is based on compact_zone() from compaction.c. */
6230 unsigned long nr_reclaimed;
6231 unsigned long pfn = start;
6232 unsigned int tries = 0;
6237 while (pfn < end || !list_empty(&cc->migratepages)) {
6238 if (fatal_signal_pending(current)) {
6243 if (list_empty(&cc->migratepages)) {
6244 cc->nr_migratepages = 0;
6245 pfn = isolate_migratepages_range(cc->zone, cc,
6252 } else if (++tries == 5) {
6253 ret = ret < 0 ? ret : -EBUSY;
6257 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6259 cc->nr_migratepages -= nr_reclaimed;
6261 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6262 NULL, 0, cc->mode, MR_CMA);
6265 putback_movable_pages(&cc->migratepages);
6272 * alloc_contig_range() -- tries to allocate given range of pages
6273 * @start: start PFN to allocate
6274 * @end: one-past-the-last PFN to allocate
6275 * @migratetype: migratetype of the underlaying pageblocks (either
6276 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6277 * in range must have the same migratetype and it must
6278 * be either of the two.
6280 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6281 * aligned, however it's the caller's responsibility to guarantee that
6282 * we are the only thread that changes migrate type of pageblocks the
6285 * The PFN range must belong to a single zone.
6287 * Returns zero on success or negative error code. On success all
6288 * pages which PFN is in [start, end) are allocated for the caller and
6289 * need to be freed with free_contig_range().
6291 int alloc_contig_range(unsigned long start, unsigned long end,
6292 unsigned migratetype)
6294 unsigned long outer_start, outer_end;
6297 struct compact_control cc = {
6298 .nr_migratepages = 0,
6300 .zone = page_zone(pfn_to_page(start)),
6301 .mode = MIGRATE_SYNC,
6302 .ignore_skip_hint = true,
6304 INIT_LIST_HEAD(&cc.migratepages);
6307 * What we do here is we mark all pageblocks in range as
6308 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6309 * have different sizes, and due to the way page allocator
6310 * work, we align the range to biggest of the two pages so
6311 * that page allocator won't try to merge buddies from
6312 * different pageblocks and change MIGRATE_ISOLATE to some
6313 * other migration type.
6315 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6316 * migrate the pages from an unaligned range (ie. pages that
6317 * we are interested in). This will put all the pages in
6318 * range back to page allocator as MIGRATE_ISOLATE.
6320 * When this is done, we take the pages in range from page
6321 * allocator removing them from the buddy system. This way
6322 * page allocator will never consider using them.
6324 * This lets us mark the pageblocks back as
6325 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6326 * aligned range but not in the unaligned, original range are
6327 * put back to page allocator so that buddy can use them.
6330 ret = start_isolate_page_range(pfn_max_align_down(start),
6331 pfn_max_align_up(end), migratetype,
6336 ret = __alloc_contig_migrate_range(&cc, start, end);
6341 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6342 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6343 * more, all pages in [start, end) are free in page allocator.
6344 * What we are going to do is to allocate all pages from
6345 * [start, end) (that is remove them from page allocator).
6347 * The only problem is that pages at the beginning and at the
6348 * end of interesting range may be not aligned with pages that
6349 * page allocator holds, ie. they can be part of higher order
6350 * pages. Because of this, we reserve the bigger range and
6351 * once this is done free the pages we are not interested in.
6353 * We don't have to hold zone->lock here because the pages are
6354 * isolated thus they won't get removed from buddy.
6357 lru_add_drain_all();
6361 outer_start = start;
6362 while (!PageBuddy(pfn_to_page(outer_start))) {
6363 if (++order >= MAX_ORDER) {
6367 outer_start &= ~0UL << order;
6370 /* Make sure the range is really isolated. */
6371 if (test_pages_isolated(outer_start, end, false)) {
6372 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6379 /* Grab isolated pages from freelists. */
6380 outer_end = isolate_freepages_range(&cc, outer_start, end);
6386 /* Free head and tail (if any) */
6387 if (start != outer_start)
6388 free_contig_range(outer_start, start - outer_start);
6389 if (end != outer_end)
6390 free_contig_range(end, outer_end - end);
6393 undo_isolate_page_range(pfn_max_align_down(start),
6394 pfn_max_align_up(end), migratetype);
6398 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6400 unsigned int count = 0;
6402 for (; nr_pages--; pfn++) {
6403 struct page *page = pfn_to_page(pfn);
6405 count += page_count(page) != 1;
6408 WARN(count != 0, "%d pages are still in use!\n", count);
6412 #ifdef CONFIG_MEMORY_HOTPLUG
6414 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6415 * page high values need to be recalulated.
6417 void __meminit zone_pcp_update(struct zone *zone)
6420 mutex_lock(&pcp_batch_high_lock);
6421 for_each_possible_cpu(cpu)
6422 pageset_set_high_and_batch(zone,
6423 per_cpu_ptr(zone->pageset, cpu));
6424 mutex_unlock(&pcp_batch_high_lock);
6428 void zone_pcp_reset(struct zone *zone)
6430 unsigned long flags;
6432 struct per_cpu_pageset *pset;
6434 /* avoid races with drain_pages() */
6435 local_irq_save(flags);
6436 if (zone->pageset != &boot_pageset) {
6437 for_each_online_cpu(cpu) {
6438 pset = per_cpu_ptr(zone->pageset, cpu);
6439 drain_zonestat(zone, pset);
6441 free_percpu(zone->pageset);
6442 zone->pageset = &boot_pageset;
6444 local_irq_restore(flags);
6447 #ifdef CONFIG_MEMORY_HOTREMOVE
6449 * All pages in the range must be isolated before calling this.
6452 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6458 unsigned long flags;
6459 /* find the first valid pfn */
6460 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6465 zone = page_zone(pfn_to_page(pfn));
6466 spin_lock_irqsave(&zone->lock, flags);
6468 while (pfn < end_pfn) {
6469 if (!pfn_valid(pfn)) {
6473 page = pfn_to_page(pfn);
6475 * The HWPoisoned page may be not in buddy system, and
6476 * page_count() is not 0.
6478 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6480 SetPageReserved(page);
6484 BUG_ON(page_count(page));
6485 BUG_ON(!PageBuddy(page));
6486 order = page_order(page);
6487 #ifdef CONFIG_DEBUG_VM
6488 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6489 pfn, 1 << order, end_pfn);
6491 list_del(&page->lru);
6492 rmv_page_order(page);
6493 zone->free_area[order].nr_free--;
6494 for (i = 0; i < (1 << order); i++)
6495 SetPageReserved((page+i));
6496 pfn += (1 << order);
6498 spin_unlock_irqrestore(&zone->lock, flags);
6502 #ifdef CONFIG_MEMORY_FAILURE
6503 bool is_free_buddy_page(struct page *page)
6505 struct zone *zone = page_zone(page);
6506 unsigned long pfn = page_to_pfn(page);
6507 unsigned long flags;
6510 spin_lock_irqsave(&zone->lock, flags);
6511 for (order = 0; order < MAX_ORDER; order++) {
6512 struct page *page_head = page - (pfn & ((1 << order) - 1));
6514 if (PageBuddy(page_head) && page_order(page_head) >= order)
6517 spin_unlock_irqrestore(&zone->lock, flags);
6519 return order < MAX_ORDER;
6523 static const struct trace_print_flags pageflag_names[] = {
6524 {1UL << PG_locked, "locked" },
6525 {1UL << PG_error, "error" },
6526 {1UL << PG_referenced, "referenced" },
6527 {1UL << PG_uptodate, "uptodate" },
6528 {1UL << PG_dirty, "dirty" },
6529 {1UL << PG_lru, "lru" },
6530 {1UL << PG_active, "active" },
6531 {1UL << PG_slab, "slab" },
6532 {1UL << PG_owner_priv_1, "owner_priv_1" },
6533 {1UL << PG_arch_1, "arch_1" },
6534 {1UL << PG_reserved, "reserved" },
6535 {1UL << PG_private, "private" },
6536 {1UL << PG_private_2, "private_2" },
6537 {1UL << PG_writeback, "writeback" },
6538 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6539 {1UL << PG_head, "head" },
6540 {1UL << PG_tail, "tail" },
6542 {1UL << PG_compound, "compound" },
6544 {1UL << PG_swapcache, "swapcache" },
6545 {1UL << PG_mappedtodisk, "mappedtodisk" },
6546 {1UL << PG_reclaim, "reclaim" },
6547 {1UL << PG_swapbacked, "swapbacked" },
6548 {1UL << PG_unevictable, "unevictable" },
6550 {1UL << PG_mlocked, "mlocked" },
6552 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6553 {1UL << PG_uncached, "uncached" },
6555 #ifdef CONFIG_MEMORY_FAILURE
6556 {1UL << PG_hwpoison, "hwpoison" },
6558 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6559 {1UL << PG_compound_lock, "compound_lock" },
6563 static void dump_page_flags(unsigned long flags)
6565 const char *delim = "";
6569 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6571 printk(KERN_ALERT "page flags: %#lx(", flags);
6573 /* remove zone id */
6574 flags &= (1UL << NR_PAGEFLAGS) - 1;
6576 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6578 mask = pageflag_names[i].mask;
6579 if ((flags & mask) != mask)
6583 printk("%s%s", delim, pageflag_names[i].name);
6587 /* check for left over flags */
6589 printk("%s%#lx", delim, flags);
6594 void dump_page_badflags(struct page *page, char *reason, unsigned long badflags)
6597 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6598 page, atomic_read(&page->_count), page_mapcount(page),
6599 page->mapping, page->index);
6600 dump_page_flags(page->flags);
6602 pr_alert("page dumped because: %s\n", reason);
6603 if (page->flags & badflags) {
6604 pr_alert("bad because of flags:\n");
6605 dump_page_flags(page->flags & badflags);
6607 mem_cgroup_print_bad_page(page);
6610 void dump_page(struct page *page, char *reason)
6612 dump_page_badflags(page, reason, 0);
6614 EXPORT_SYMBOL_GPL(dump_page);