2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74 DEFINE_PER_CPU(int, numa_node);
75 EXPORT_PER_CPU_SYMBOL(numa_node);
78 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
85 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
86 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
90 * Array of node states.
92 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
93 [N_POSSIBLE] = NODE_MASK_ALL,
94 [N_ONLINE] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY] = { { [0] = 1UL } },
103 [N_CPU] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock);
111 unsigned long totalram_pages __read_mostly;
112 unsigned long totalreserve_pages __read_mostly;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly;
121 int percpu_pagelist_fraction;
122 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex));
139 if (saved_gfp_mask) {
140 gfp_allowed_mask = saved_gfp_mask;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex));
148 WARN_ON(saved_gfp_mask);
149 saved_gfp_mask = gfp_allowed_mask;
150 gfp_allowed_mask &= ~GFP_IOFS;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly;
165 static void __free_pages_ok(struct page *page, unsigned int order);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages);
193 static char * const zone_names[MAX_NR_ZONES] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes = 1024;
208 int user_min_free_kbytes = -1;
210 static unsigned long __meminitdata nr_kernel_pages;
211 static unsigned long __meminitdata nr_all_pages;
212 static unsigned long __meminitdata dma_reserve;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
217 static unsigned long __initdata required_kernelcore;
218 static unsigned long __initdata required_movablecore;
219 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly = MAX_NUMNODES;
228 int nr_online_nodes __read_mostly = 1;
229 EXPORT_SYMBOL(nr_node_ids);
230 EXPORT_SYMBOL(nr_online_nodes);
233 int page_group_by_mobility_disabled __read_mostly;
235 void set_pageblock_migratetype(struct page *page, int migratetype)
237 if (unlikely(page_group_by_mobility_disabled &&
238 migratetype < MIGRATE_PCPTYPES))
239 migratetype = MIGRATE_UNMOVABLE;
241 set_pageblock_flags_group(page, (unsigned long)migratetype,
242 PB_migrate, PB_migrate_end);
245 bool oom_killer_disabled __read_mostly;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
252 unsigned long pfn = page_to_pfn(page);
253 unsigned long sp, start_pfn;
256 seq = zone_span_seqbegin(zone);
257 start_pfn = zone->zone_start_pfn;
258 sp = zone->spanned_pages;
259 if (!zone_spans_pfn(zone, pfn))
261 } while (zone_span_seqretry(zone, seq));
264 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
265 pfn, zone_to_nid(zone), zone->name,
266 start_pfn, start_pfn + sp);
271 static int page_is_consistent(struct zone *zone, struct page *page)
273 if (!pfn_valid_within(page_to_pfn(page)))
275 if (zone != page_zone(page))
281 * Temporary debugging check for pages not lying within a given zone.
283 static int bad_range(struct zone *zone, struct page *page)
285 if (page_outside_zone_boundaries(zone, page))
287 if (!page_is_consistent(zone, page))
293 static inline int bad_range(struct zone *zone, struct page *page)
299 static void bad_page(struct page *page, const char *reason,
300 unsigned long bad_flags)
302 static unsigned long resume;
303 static unsigned long nr_shown;
304 static unsigned long nr_unshown;
306 /* Don't complain about poisoned pages */
307 if (PageHWPoison(page)) {
308 page_mapcount_reset(page); /* remove PageBuddy */
313 * Allow a burst of 60 reports, then keep quiet for that minute;
314 * or allow a steady drip of one report per second.
316 if (nr_shown == 60) {
317 if (time_before(jiffies, resume)) {
323 "BUG: Bad page state: %lu messages suppressed\n",
330 resume = jiffies + 60 * HZ;
332 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
333 current->comm, page_to_pfn(page));
334 dump_page_badflags(page, reason, bad_flags);
339 /* Leave bad fields for debug, except PageBuddy could make trouble */
340 page_mapcount_reset(page); /* remove PageBuddy */
341 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
345 * Higher-order pages are called "compound pages". They are structured thusly:
347 * The first PAGE_SIZE page is called the "head page".
349 * The remaining PAGE_SIZE pages are called "tail pages".
351 * All pages have PG_compound set. All tail pages have their ->first_page
352 * pointing at the head page.
354 * The first tail page's ->lru.next holds the address of the compound page's
355 * put_page() function. Its ->lru.prev holds the order of allocation.
356 * This usage means that zero-order pages may not be compound.
359 static void free_compound_page(struct page *page)
361 __free_pages_ok(page, compound_order(page));
364 void prep_compound_page(struct page *page, unsigned long order)
367 int nr_pages = 1 << order;
369 set_compound_page_dtor(page, free_compound_page);
370 set_compound_order(page, order);
372 for (i = 1; i < nr_pages; i++) {
373 struct page *p = page + i;
374 set_page_count(p, 0);
375 p->first_page = page;
376 /* Make sure p->first_page is always valid for PageTail() */
382 /* update __split_huge_page_refcount if you change this function */
383 static int destroy_compound_page(struct page *page, unsigned long order)
386 int nr_pages = 1 << order;
389 if (unlikely(compound_order(page) != order)) {
390 bad_page(page, "wrong compound order", 0);
394 __ClearPageHead(page);
396 for (i = 1; i < nr_pages; i++) {
397 struct page *p = page + i;
399 if (unlikely(!PageTail(p))) {
400 bad_page(page, "PageTail not set", 0);
402 } else if (unlikely(p->first_page != page)) {
403 bad_page(page, "first_page not consistent", 0);
412 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
417 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
418 * and __GFP_HIGHMEM from hard or soft interrupt context.
420 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
421 for (i = 0; i < (1 << order); i++)
422 clear_highpage(page + i);
425 #ifdef CONFIG_DEBUG_PAGEALLOC
426 unsigned int _debug_guardpage_minorder;
428 static int __init debug_guardpage_minorder_setup(char *buf)
432 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
433 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
436 _debug_guardpage_minorder = res;
437 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
440 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
442 static inline void set_page_guard_flag(struct page *page)
444 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
447 static inline void clear_page_guard_flag(struct page *page)
449 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
452 static inline void set_page_guard_flag(struct page *page) { }
453 static inline void clear_page_guard_flag(struct page *page) { }
456 static inline void set_page_order(struct page *page, int order)
458 set_page_private(page, order);
459 __SetPageBuddy(page);
462 static inline void rmv_page_order(struct page *page)
464 __ClearPageBuddy(page);
465 set_page_private(page, 0);
469 * Locate the struct page for both the matching buddy in our
470 * pair (buddy1) and the combined O(n+1) page they form (page).
472 * 1) Any buddy B1 will have an order O twin B2 which satisfies
473 * the following equation:
475 * For example, if the starting buddy (buddy2) is #8 its order
477 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
479 * 2) Any buddy B will have an order O+1 parent P which
480 * satisfies the following equation:
483 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
485 static inline unsigned long
486 __find_buddy_index(unsigned long page_idx, unsigned int order)
488 return page_idx ^ (1 << order);
492 * This function checks whether a page is free && is the buddy
493 * we can do coalesce a page and its buddy if
494 * (a) the buddy is not in a hole &&
495 * (b) the buddy is in the buddy system &&
496 * (c) a page and its buddy have the same order &&
497 * (d) a page and its buddy are in the same zone.
499 * For recording whether a page is in the buddy system, we set ->_mapcount
500 * PAGE_BUDDY_MAPCOUNT_VALUE.
501 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
502 * serialized by zone->lock.
504 * For recording page's order, we use page_private(page).
506 static inline int page_is_buddy(struct page *page, struct page *buddy,
509 if (!pfn_valid_within(page_to_pfn(buddy)))
512 if (page_zone_id(page) != page_zone_id(buddy))
515 if (page_is_guard(buddy) && page_order(buddy) == order) {
516 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
520 if (PageBuddy(buddy) && page_order(buddy) == order) {
521 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
528 * Freeing function for a buddy system allocator.
530 * The concept of a buddy system is to maintain direct-mapped table
531 * (containing bit values) for memory blocks of various "orders".
532 * The bottom level table contains the map for the smallest allocatable
533 * units of memory (here, pages), and each level above it describes
534 * pairs of units from the levels below, hence, "buddies".
535 * At a high level, all that happens here is marking the table entry
536 * at the bottom level available, and propagating the changes upward
537 * as necessary, plus some accounting needed to play nicely with other
538 * parts of the VM system.
539 * At each level, we keep a list of pages, which are heads of continuous
540 * free pages of length of (1 << order) and marked with _mapcount
541 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
543 * So when we are allocating or freeing one, we can derive the state of the
544 * other. That is, if we allocate a small block, and both were
545 * free, the remainder of the region must be split into blocks.
546 * If a block is freed, and its buddy is also free, then this
547 * triggers coalescing into a block of larger size.
552 static inline void __free_one_page(struct page *page,
553 struct zone *zone, unsigned int order,
556 unsigned long page_idx;
557 unsigned long combined_idx;
558 unsigned long uninitialized_var(buddy_idx);
561 VM_BUG_ON(!zone_is_initialized(zone));
563 if (unlikely(PageCompound(page)))
564 if (unlikely(destroy_compound_page(page, order)))
567 VM_BUG_ON(migratetype == -1);
569 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
571 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
572 VM_BUG_ON_PAGE(bad_range(zone, page), page);
574 while (order < MAX_ORDER-1) {
575 buddy_idx = __find_buddy_index(page_idx, order);
576 buddy = page + (buddy_idx - page_idx);
577 if (!page_is_buddy(page, buddy, order))
580 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
581 * merge with it and move up one order.
583 if (page_is_guard(buddy)) {
584 clear_page_guard_flag(buddy);
585 set_page_private(page, 0);
586 __mod_zone_freepage_state(zone, 1 << order,
589 list_del(&buddy->lru);
590 zone->free_area[order].nr_free--;
591 rmv_page_order(buddy);
593 combined_idx = buddy_idx & page_idx;
594 page = page + (combined_idx - page_idx);
595 page_idx = combined_idx;
598 set_page_order(page, order);
601 * If this is not the largest possible page, check if the buddy
602 * of the next-highest order is free. If it is, it's possible
603 * that pages are being freed that will coalesce soon. In case,
604 * that is happening, add the free page to the tail of the list
605 * so it's less likely to be used soon and more likely to be merged
606 * as a higher order page
608 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
609 struct page *higher_page, *higher_buddy;
610 combined_idx = buddy_idx & page_idx;
611 higher_page = page + (combined_idx - page_idx);
612 buddy_idx = __find_buddy_index(combined_idx, order + 1);
613 higher_buddy = higher_page + (buddy_idx - combined_idx);
614 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
615 list_add_tail(&page->lru,
616 &zone->free_area[order].free_list[migratetype]);
621 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
623 zone->free_area[order].nr_free++;
626 static inline int free_pages_check(struct page *page)
628 const char *bad_reason = NULL;
629 unsigned long bad_flags = 0;
631 if (unlikely(page_mapcount(page)))
632 bad_reason = "nonzero mapcount";
633 if (unlikely(page->mapping != NULL))
634 bad_reason = "non-NULL mapping";
635 if (unlikely(atomic_read(&page->_count) != 0))
636 bad_reason = "nonzero _count";
637 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
638 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
639 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
641 if (unlikely(mem_cgroup_bad_page_check(page)))
642 bad_reason = "cgroup check failed";
643 if (unlikely(bad_reason)) {
644 bad_page(page, bad_reason, bad_flags);
647 page_cpupid_reset_last(page);
648 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
649 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
654 * Frees a number of pages from the PCP lists
655 * Assumes all pages on list are in same zone, and of same order.
656 * count is the number of pages to free.
658 * If the zone was previously in an "all pages pinned" state then look to
659 * see if this freeing clears that state.
661 * And clear the zone's pages_scanned counter, to hold off the "all pages are
662 * pinned" detection logic.
664 static void free_pcppages_bulk(struct zone *zone, int count,
665 struct per_cpu_pages *pcp)
671 spin_lock(&zone->lock);
672 zone->pages_scanned = 0;
676 struct list_head *list;
679 * Remove pages from lists in a round-robin fashion. A
680 * batch_free count is maintained that is incremented when an
681 * empty list is encountered. This is so more pages are freed
682 * off fuller lists instead of spinning excessively around empty
687 if (++migratetype == MIGRATE_PCPTYPES)
689 list = &pcp->lists[migratetype];
690 } while (list_empty(list));
692 /* This is the only non-empty list. Free them all. */
693 if (batch_free == MIGRATE_PCPTYPES)
694 batch_free = to_free;
697 int mt; /* migratetype of the to-be-freed page */
699 page = list_entry(list->prev, struct page, lru);
700 /* must delete as __free_one_page list manipulates */
701 list_del(&page->lru);
702 mt = get_freepage_migratetype(page);
703 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
704 __free_one_page(page, zone, 0, mt);
705 trace_mm_page_pcpu_drain(page, 0, mt);
706 if (likely(!is_migrate_isolate_page(page))) {
707 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
708 if (is_migrate_cma(mt))
709 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
711 } while (--to_free && --batch_free && !list_empty(list));
713 spin_unlock(&zone->lock);
716 static void free_one_page(struct zone *zone, struct page *page, int order,
719 spin_lock(&zone->lock);
720 zone->pages_scanned = 0;
722 __free_one_page(page, zone, order, migratetype);
723 if (unlikely(!is_migrate_isolate(migratetype)))
724 __mod_zone_freepage_state(zone, 1 << order, migratetype);
725 spin_unlock(&zone->lock);
728 static bool free_pages_prepare(struct page *page, unsigned int order)
733 trace_mm_page_free(page, order);
734 kmemcheck_free_shadow(page, order);
737 page->mapping = NULL;
738 for (i = 0; i < (1 << order); i++)
739 bad += free_pages_check(page + i);
743 if (!PageHighMem(page)) {
744 debug_check_no_locks_freed(page_address(page),
746 debug_check_no_obj_freed(page_address(page),
749 arch_free_page(page, order);
750 kernel_map_pages(page, 1 << order, 0);
755 static void __free_pages_ok(struct page *page, unsigned int order)
760 if (!free_pages_prepare(page, order))
763 local_irq_save(flags);
764 __count_vm_events(PGFREE, 1 << order);
765 migratetype = get_pageblock_migratetype(page);
766 set_freepage_migratetype(page, migratetype);
767 free_one_page(page_zone(page), page, order, migratetype);
768 local_irq_restore(flags);
771 void __init __free_pages_bootmem(struct page *page, unsigned int order)
773 unsigned int nr_pages = 1 << order;
774 struct page *p = page;
778 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
780 __ClearPageReserved(p);
781 set_page_count(p, 0);
783 __ClearPageReserved(p);
784 set_page_count(p, 0);
786 page_zone(page)->managed_pages += nr_pages;
787 set_page_refcounted(page);
788 __free_pages(page, order);
792 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
793 void __init init_cma_reserved_pageblock(struct page *page)
795 unsigned i = pageblock_nr_pages;
796 struct page *p = page;
799 __ClearPageReserved(p);
800 set_page_count(p, 0);
803 set_page_refcounted(page);
804 set_pageblock_migratetype(page, MIGRATE_CMA);
805 __free_pages(page, pageblock_order);
806 adjust_managed_page_count(page, pageblock_nr_pages);
811 * The order of subdivision here is critical for the IO subsystem.
812 * Please do not alter this order without good reasons and regression
813 * testing. Specifically, as large blocks of memory are subdivided,
814 * the order in which smaller blocks are delivered depends on the order
815 * they're subdivided in this function. This is the primary factor
816 * influencing the order in which pages are delivered to the IO
817 * subsystem according to empirical testing, and this is also justified
818 * by considering the behavior of a buddy system containing a single
819 * large block of memory acted on by a series of small allocations.
820 * This behavior is a critical factor in sglist merging's success.
824 static inline void expand(struct zone *zone, struct page *page,
825 int low, int high, struct free_area *area,
828 unsigned long size = 1 << high;
834 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
836 #ifdef CONFIG_DEBUG_PAGEALLOC
837 if (high < debug_guardpage_minorder()) {
839 * Mark as guard pages (or page), that will allow to
840 * merge back to allocator when buddy will be freed.
841 * Corresponding page table entries will not be touched,
842 * pages will stay not present in virtual address space
844 INIT_LIST_HEAD(&page[size].lru);
845 set_page_guard_flag(&page[size]);
846 set_page_private(&page[size], high);
847 /* Guard pages are not available for any usage */
848 __mod_zone_freepage_state(zone, -(1 << high),
853 list_add(&page[size].lru, &area->free_list[migratetype]);
855 set_page_order(&page[size], high);
860 * This page is about to be returned from the page allocator
862 static inline int check_new_page(struct page *page)
864 const char *bad_reason = NULL;
865 unsigned long bad_flags = 0;
867 if (unlikely(page_mapcount(page)))
868 bad_reason = "nonzero mapcount";
869 if (unlikely(page->mapping != NULL))
870 bad_reason = "non-NULL mapping";
871 if (unlikely(atomic_read(&page->_count) != 0))
872 bad_reason = "nonzero _count";
873 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
874 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
875 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
877 if (unlikely(mem_cgroup_bad_page_check(page)))
878 bad_reason = "cgroup check failed";
879 if (unlikely(bad_reason)) {
880 bad_page(page, bad_reason, bad_flags);
886 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
890 for (i = 0; i < (1 << order); i++) {
891 struct page *p = page + i;
892 if (unlikely(check_new_page(p)))
896 set_page_private(page, 0);
897 set_page_refcounted(page);
899 arch_alloc_page(page, order);
900 kernel_map_pages(page, 1 << order, 1);
902 if (gfp_flags & __GFP_ZERO)
903 prep_zero_page(page, order, gfp_flags);
905 if (order && (gfp_flags & __GFP_COMP))
906 prep_compound_page(page, order);
912 * Go through the free lists for the given migratetype and remove
913 * the smallest available page from the freelists
916 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
919 unsigned int current_order;
920 struct free_area *area;
923 /* Find a page of the appropriate size in the preferred list */
924 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
925 area = &(zone->free_area[current_order]);
926 if (list_empty(&area->free_list[migratetype]))
929 page = list_entry(area->free_list[migratetype].next,
931 list_del(&page->lru);
932 rmv_page_order(page);
934 expand(zone, page, order, current_order, area, migratetype);
935 set_freepage_migratetype(page, migratetype);
944 * This array describes the order lists are fallen back to when
945 * the free lists for the desirable migrate type are depleted
947 static int fallbacks[MIGRATE_TYPES][4] = {
948 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
949 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
951 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
952 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
954 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
956 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
957 #ifdef CONFIG_MEMORY_ISOLATION
958 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
963 * Move the free pages in a range to the free lists of the requested type.
964 * Note that start_page and end_pages are not aligned on a pageblock
965 * boundary. If alignment is required, use move_freepages_block()
967 int move_freepages(struct zone *zone,
968 struct page *start_page, struct page *end_page,
975 #ifndef CONFIG_HOLES_IN_ZONE
977 * page_zone is not safe to call in this context when
978 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
979 * anyway as we check zone boundaries in move_freepages_block().
980 * Remove at a later date when no bug reports exist related to
981 * grouping pages by mobility
983 BUG_ON(page_zone(start_page) != page_zone(end_page));
986 for (page = start_page; page <= end_page;) {
987 /* Make sure we are not inadvertently changing nodes */
988 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
990 if (!pfn_valid_within(page_to_pfn(page))) {
995 if (!PageBuddy(page)) {
1000 order = page_order(page);
1001 list_move(&page->lru,
1002 &zone->free_area[order].free_list[migratetype]);
1003 set_freepage_migratetype(page, migratetype);
1005 pages_moved += 1 << order;
1011 int move_freepages_block(struct zone *zone, struct page *page,
1014 unsigned long start_pfn, end_pfn;
1015 struct page *start_page, *end_page;
1017 start_pfn = page_to_pfn(page);
1018 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1019 start_page = pfn_to_page(start_pfn);
1020 end_page = start_page + pageblock_nr_pages - 1;
1021 end_pfn = start_pfn + pageblock_nr_pages - 1;
1023 /* Do not cross zone boundaries */
1024 if (!zone_spans_pfn(zone, start_pfn))
1026 if (!zone_spans_pfn(zone, end_pfn))
1029 return move_freepages(zone, start_page, end_page, migratetype);
1032 static void change_pageblock_range(struct page *pageblock_page,
1033 int start_order, int migratetype)
1035 int nr_pageblocks = 1 << (start_order - pageblock_order);
1037 while (nr_pageblocks--) {
1038 set_pageblock_migratetype(pageblock_page, migratetype);
1039 pageblock_page += pageblock_nr_pages;
1044 * If breaking a large block of pages, move all free pages to the preferred
1045 * allocation list. If falling back for a reclaimable kernel allocation, be
1046 * more aggressive about taking ownership of free pages.
1048 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1049 * nor move CMA pages to different free lists. We don't want unmovable pages
1050 * to be allocated from MIGRATE_CMA areas.
1052 * Returns the new migratetype of the pageblock (or the same old migratetype
1053 * if it was unchanged).
1055 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1056 int start_type, int fallback_type)
1058 int current_order = page_order(page);
1061 * When borrowing from MIGRATE_CMA, we need to release the excess
1062 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1063 * is set to CMA so it is returned to the correct freelist in case
1064 * the page ends up being not actually allocated from the pcp lists.
1066 if (is_migrate_cma(fallback_type))
1067 return fallback_type;
1069 /* Take ownership for orders >= pageblock_order */
1070 if (current_order >= pageblock_order) {
1071 change_pageblock_range(page, current_order, start_type);
1075 if (current_order >= pageblock_order / 2 ||
1076 start_type == MIGRATE_RECLAIMABLE ||
1077 page_group_by_mobility_disabled) {
1080 pages = move_freepages_block(zone, page, start_type);
1082 /* Claim the whole block if over half of it is free */
1083 if (pages >= (1 << (pageblock_order-1)) ||
1084 page_group_by_mobility_disabled) {
1086 set_pageblock_migratetype(page, start_type);
1092 return fallback_type;
1095 /* Remove an element from the buddy allocator from the fallback list */
1096 static inline struct page *
1097 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1099 struct free_area *area;
1102 int migratetype, new_type, i;
1104 /* Find the largest possible block of pages in the other list */
1105 for (current_order = MAX_ORDER-1; current_order >= order;
1108 migratetype = fallbacks[start_migratetype][i];
1110 /* MIGRATE_RESERVE handled later if necessary */
1111 if (migratetype == MIGRATE_RESERVE)
1114 area = &(zone->free_area[current_order]);
1115 if (list_empty(&area->free_list[migratetype]))
1118 page = list_entry(area->free_list[migratetype].next,
1122 new_type = try_to_steal_freepages(zone, page,
1126 /* Remove the page from the freelists */
1127 list_del(&page->lru);
1128 rmv_page_order(page);
1130 expand(zone, page, order, current_order, area,
1132 /* The freepage_migratetype may differ from pageblock's
1133 * migratetype depending on the decisions in
1134 * try_to_steal_freepages. This is OK as long as it does
1135 * not differ for MIGRATE_CMA type.
1137 set_freepage_migratetype(page, new_type);
1139 trace_mm_page_alloc_extfrag(page, order, current_order,
1140 start_migratetype, migratetype, new_type);
1150 * Do the hard work of removing an element from the buddy allocator.
1151 * Call me with the zone->lock already held.
1153 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1159 page = __rmqueue_smallest(zone, order, migratetype);
1161 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1162 page = __rmqueue_fallback(zone, order, migratetype);
1165 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1166 * is used because __rmqueue_smallest is an inline function
1167 * and we want just one call site
1170 migratetype = MIGRATE_RESERVE;
1175 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1180 * Obtain a specified number of elements from the buddy allocator, all under
1181 * a single hold of the lock, for efficiency. Add them to the supplied list.
1182 * Returns the number of new pages which were placed at *list.
1184 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1185 unsigned long count, struct list_head *list,
1186 int migratetype, int cold)
1190 spin_lock(&zone->lock);
1191 for (i = 0; i < count; ++i) {
1192 struct page *page = __rmqueue(zone, order, migratetype);
1193 if (unlikely(page == NULL))
1197 * Split buddy pages returned by expand() are received here
1198 * in physical page order. The page is added to the callers and
1199 * list and the list head then moves forward. From the callers
1200 * perspective, the linked list is ordered by page number in
1201 * some conditions. This is useful for IO devices that can
1202 * merge IO requests if the physical pages are ordered
1205 if (likely(cold == 0))
1206 list_add(&page->lru, list);
1208 list_add_tail(&page->lru, list);
1210 if (is_migrate_cma(get_freepage_migratetype(page)))
1211 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1214 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1215 spin_unlock(&zone->lock);
1221 * Called from the vmstat counter updater to drain pagesets of this
1222 * currently executing processor on remote nodes after they have
1225 * Note that this function must be called with the thread pinned to
1226 * a single processor.
1228 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1230 unsigned long flags;
1232 unsigned long batch;
1234 local_irq_save(flags);
1235 batch = ACCESS_ONCE(pcp->batch);
1236 if (pcp->count >= batch)
1239 to_drain = pcp->count;
1241 free_pcppages_bulk(zone, to_drain, pcp);
1242 pcp->count -= to_drain;
1244 local_irq_restore(flags);
1249 * Drain pages of the indicated processor.
1251 * The processor must either be the current processor and the
1252 * thread pinned to the current processor or a processor that
1255 static void drain_pages(unsigned int cpu)
1257 unsigned long flags;
1260 for_each_populated_zone(zone) {
1261 struct per_cpu_pageset *pset;
1262 struct per_cpu_pages *pcp;
1264 local_irq_save(flags);
1265 pset = per_cpu_ptr(zone->pageset, cpu);
1269 free_pcppages_bulk(zone, pcp->count, pcp);
1272 local_irq_restore(flags);
1277 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1279 void drain_local_pages(void *arg)
1281 drain_pages(smp_processor_id());
1285 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1287 * Note that this code is protected against sending an IPI to an offline
1288 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1289 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1290 * nothing keeps CPUs from showing up after we populated the cpumask and
1291 * before the call to on_each_cpu_mask().
1293 void drain_all_pages(void)
1296 struct per_cpu_pageset *pcp;
1300 * Allocate in the BSS so we wont require allocation in
1301 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1303 static cpumask_t cpus_with_pcps;
1306 * We don't care about racing with CPU hotplug event
1307 * as offline notification will cause the notified
1308 * cpu to drain that CPU pcps and on_each_cpu_mask
1309 * disables preemption as part of its processing
1311 for_each_online_cpu(cpu) {
1312 bool has_pcps = false;
1313 for_each_populated_zone(zone) {
1314 pcp = per_cpu_ptr(zone->pageset, cpu);
1315 if (pcp->pcp.count) {
1321 cpumask_set_cpu(cpu, &cpus_with_pcps);
1323 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1325 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1328 #ifdef CONFIG_HIBERNATION
1330 void mark_free_pages(struct zone *zone)
1332 unsigned long pfn, max_zone_pfn;
1333 unsigned long flags;
1335 struct list_head *curr;
1337 if (zone_is_empty(zone))
1340 spin_lock_irqsave(&zone->lock, flags);
1342 max_zone_pfn = zone_end_pfn(zone);
1343 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1344 if (pfn_valid(pfn)) {
1345 struct page *page = pfn_to_page(pfn);
1347 if (!swsusp_page_is_forbidden(page))
1348 swsusp_unset_page_free(page);
1351 for_each_migratetype_order(order, t) {
1352 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1355 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1356 for (i = 0; i < (1UL << order); i++)
1357 swsusp_set_page_free(pfn_to_page(pfn + i));
1360 spin_unlock_irqrestore(&zone->lock, flags);
1362 #endif /* CONFIG_PM */
1365 * Free a 0-order page
1366 * cold == 1 ? free a cold page : free a hot page
1368 void free_hot_cold_page(struct page *page, int cold)
1370 struct zone *zone = page_zone(page);
1371 struct per_cpu_pages *pcp;
1372 unsigned long flags;
1375 if (!free_pages_prepare(page, 0))
1378 migratetype = get_pageblock_migratetype(page);
1379 set_freepage_migratetype(page, migratetype);
1380 local_irq_save(flags);
1381 __count_vm_event(PGFREE);
1384 * We only track unmovable, reclaimable and movable on pcp lists.
1385 * Free ISOLATE pages back to the allocator because they are being
1386 * offlined but treat RESERVE as movable pages so we can get those
1387 * areas back if necessary. Otherwise, we may have to free
1388 * excessively into the page allocator
1390 if (migratetype >= MIGRATE_PCPTYPES) {
1391 if (unlikely(is_migrate_isolate(migratetype))) {
1392 free_one_page(zone, page, 0, migratetype);
1395 migratetype = MIGRATE_MOVABLE;
1398 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1400 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1402 list_add(&page->lru, &pcp->lists[migratetype]);
1404 if (pcp->count >= pcp->high) {
1405 unsigned long batch = ACCESS_ONCE(pcp->batch);
1406 free_pcppages_bulk(zone, batch, pcp);
1407 pcp->count -= batch;
1411 local_irq_restore(flags);
1415 * Free a list of 0-order pages
1417 void free_hot_cold_page_list(struct list_head *list, int cold)
1419 struct page *page, *next;
1421 list_for_each_entry_safe(page, next, list, lru) {
1422 trace_mm_page_free_batched(page, cold);
1423 free_hot_cold_page(page, cold);
1428 * split_page takes a non-compound higher-order page, and splits it into
1429 * n (1<<order) sub-pages: page[0..n]
1430 * Each sub-page must be freed individually.
1432 * Note: this is probably too low level an operation for use in drivers.
1433 * Please consult with lkml before using this in your driver.
1435 void split_page(struct page *page, unsigned int order)
1439 VM_BUG_ON_PAGE(PageCompound(page), page);
1440 VM_BUG_ON_PAGE(!page_count(page), page);
1442 #ifdef CONFIG_KMEMCHECK
1444 * Split shadow pages too, because free(page[0]) would
1445 * otherwise free the whole shadow.
1447 if (kmemcheck_page_is_tracked(page))
1448 split_page(virt_to_page(page[0].shadow), order);
1451 for (i = 1; i < (1 << order); i++)
1452 set_page_refcounted(page + i);
1454 EXPORT_SYMBOL_GPL(split_page);
1456 static int __isolate_free_page(struct page *page, unsigned int order)
1458 unsigned long watermark;
1462 BUG_ON(!PageBuddy(page));
1464 zone = page_zone(page);
1465 mt = get_pageblock_migratetype(page);
1467 if (!is_migrate_isolate(mt)) {
1468 /* Obey watermarks as if the page was being allocated */
1469 watermark = low_wmark_pages(zone) + (1 << order);
1470 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1473 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1476 /* Remove page from free list */
1477 list_del(&page->lru);
1478 zone->free_area[order].nr_free--;
1479 rmv_page_order(page);
1481 /* Set the pageblock if the isolated page is at least a pageblock */
1482 if (order >= pageblock_order - 1) {
1483 struct page *endpage = page + (1 << order) - 1;
1484 for (; page < endpage; page += pageblock_nr_pages) {
1485 int mt = get_pageblock_migratetype(page);
1486 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1487 set_pageblock_migratetype(page,
1492 return 1UL << order;
1496 * Similar to split_page except the page is already free. As this is only
1497 * being used for migration, the migratetype of the block also changes.
1498 * As this is called with interrupts disabled, the caller is responsible
1499 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1502 * Note: this is probably too low level an operation for use in drivers.
1503 * Please consult with lkml before using this in your driver.
1505 int split_free_page(struct page *page)
1510 order = page_order(page);
1512 nr_pages = __isolate_free_page(page, order);
1516 /* Split into individual pages */
1517 set_page_refcounted(page);
1518 split_page(page, order);
1523 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1524 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1528 struct page *buffered_rmqueue(struct zone *preferred_zone,
1529 struct zone *zone, int order, gfp_t gfp_flags,
1532 unsigned long flags;
1534 int cold = !!(gfp_flags & __GFP_COLD);
1537 if (likely(order == 0)) {
1538 struct per_cpu_pages *pcp;
1539 struct list_head *list;
1541 local_irq_save(flags);
1542 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1543 list = &pcp->lists[migratetype];
1544 if (list_empty(list)) {
1545 pcp->count += rmqueue_bulk(zone, 0,
1548 if (unlikely(list_empty(list)))
1553 page = list_entry(list->prev, struct page, lru);
1555 page = list_entry(list->next, struct page, lru);
1557 list_del(&page->lru);
1560 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1562 * __GFP_NOFAIL is not to be used in new code.
1564 * All __GFP_NOFAIL callers should be fixed so that they
1565 * properly detect and handle allocation failures.
1567 * We most definitely don't want callers attempting to
1568 * allocate greater than order-1 page units with
1571 WARN_ON_ONCE(order > 1);
1573 spin_lock_irqsave(&zone->lock, flags);
1574 page = __rmqueue(zone, order, migratetype);
1575 spin_unlock(&zone->lock);
1578 __mod_zone_freepage_state(zone, -(1 << order),
1579 get_freepage_migratetype(page));
1582 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1584 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1585 zone_statistics(preferred_zone, zone, gfp_flags);
1586 local_irq_restore(flags);
1588 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1589 if (prep_new_page(page, order, gfp_flags))
1594 local_irq_restore(flags);
1598 #ifdef CONFIG_FAIL_PAGE_ALLOC
1601 struct fault_attr attr;
1603 u32 ignore_gfp_highmem;
1604 u32 ignore_gfp_wait;
1606 } fail_page_alloc = {
1607 .attr = FAULT_ATTR_INITIALIZER,
1608 .ignore_gfp_wait = 1,
1609 .ignore_gfp_highmem = 1,
1613 static int __init setup_fail_page_alloc(char *str)
1615 return setup_fault_attr(&fail_page_alloc.attr, str);
1617 __setup("fail_page_alloc=", setup_fail_page_alloc);
1619 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1621 if (order < fail_page_alloc.min_order)
1623 if (gfp_mask & __GFP_NOFAIL)
1625 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1627 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1630 return should_fail(&fail_page_alloc.attr, 1 << order);
1633 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1635 static int __init fail_page_alloc_debugfs(void)
1637 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1640 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1641 &fail_page_alloc.attr);
1643 return PTR_ERR(dir);
1645 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1646 &fail_page_alloc.ignore_gfp_wait))
1648 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1649 &fail_page_alloc.ignore_gfp_highmem))
1651 if (!debugfs_create_u32("min-order", mode, dir,
1652 &fail_page_alloc.min_order))
1657 debugfs_remove_recursive(dir);
1662 late_initcall(fail_page_alloc_debugfs);
1664 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1666 #else /* CONFIG_FAIL_PAGE_ALLOC */
1668 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1673 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1676 * Return true if free pages are above 'mark'. This takes into account the order
1677 * of the allocation.
1679 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1680 int classzone_idx, int alloc_flags, long free_pages)
1682 /* free_pages my go negative - that's OK */
1684 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1688 free_pages -= (1 << order) - 1;
1689 if (alloc_flags & ALLOC_HIGH)
1691 if (alloc_flags & ALLOC_HARDER)
1694 /* If allocation can't use CMA areas don't use free CMA pages */
1695 if (!(alloc_flags & ALLOC_CMA))
1696 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1699 if (free_pages - free_cma <= min + lowmem_reserve)
1701 for (o = 0; o < order; o++) {
1702 /* At the next order, this order's pages become unavailable */
1703 free_pages -= z->free_area[o].nr_free << o;
1705 /* Require fewer higher order pages to be free */
1708 if (free_pages <= min)
1714 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1715 int classzone_idx, int alloc_flags)
1717 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1718 zone_page_state(z, NR_FREE_PAGES));
1721 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1722 int classzone_idx, int alloc_flags)
1724 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1726 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1727 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1729 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1735 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1736 * skip over zones that are not allowed by the cpuset, or that have
1737 * been recently (in last second) found to be nearly full. See further
1738 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1739 * that have to skip over a lot of full or unallowed zones.
1741 * If the zonelist cache is present in the passed zonelist, then
1742 * returns a pointer to the allowed node mask (either the current
1743 * tasks mems_allowed, or node_states[N_MEMORY].)
1745 * If the zonelist cache is not available for this zonelist, does
1746 * nothing and returns NULL.
1748 * If the fullzones BITMAP in the zonelist cache is stale (more than
1749 * a second since last zap'd) then we zap it out (clear its bits.)
1751 * We hold off even calling zlc_setup, until after we've checked the
1752 * first zone in the zonelist, on the theory that most allocations will
1753 * be satisfied from that first zone, so best to examine that zone as
1754 * quickly as we can.
1756 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1758 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1759 nodemask_t *allowednodes; /* zonelist_cache approximation */
1761 zlc = zonelist->zlcache_ptr;
1765 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1766 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1767 zlc->last_full_zap = jiffies;
1770 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1771 &cpuset_current_mems_allowed :
1772 &node_states[N_MEMORY];
1773 return allowednodes;
1777 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1778 * if it is worth looking at further for free memory:
1779 * 1) Check that the zone isn't thought to be full (doesn't have its
1780 * bit set in the zonelist_cache fullzones BITMAP).
1781 * 2) Check that the zones node (obtained from the zonelist_cache
1782 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1783 * Return true (non-zero) if zone is worth looking at further, or
1784 * else return false (zero) if it is not.
1786 * This check -ignores- the distinction between various watermarks,
1787 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1788 * found to be full for any variation of these watermarks, it will
1789 * be considered full for up to one second by all requests, unless
1790 * we are so low on memory on all allowed nodes that we are forced
1791 * into the second scan of the zonelist.
1793 * In the second scan we ignore this zonelist cache and exactly
1794 * apply the watermarks to all zones, even it is slower to do so.
1795 * We are low on memory in the second scan, and should leave no stone
1796 * unturned looking for a free page.
1798 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1799 nodemask_t *allowednodes)
1801 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1802 int i; /* index of *z in zonelist zones */
1803 int n; /* node that zone *z is on */
1805 zlc = zonelist->zlcache_ptr;
1809 i = z - zonelist->_zonerefs;
1812 /* This zone is worth trying if it is allowed but not full */
1813 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1817 * Given 'z' scanning a zonelist, set the corresponding bit in
1818 * zlc->fullzones, so that subsequent attempts to allocate a page
1819 * from that zone don't waste time re-examining it.
1821 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1823 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1824 int i; /* index of *z in zonelist zones */
1826 zlc = zonelist->zlcache_ptr;
1830 i = z - zonelist->_zonerefs;
1832 set_bit(i, zlc->fullzones);
1836 * clear all zones full, called after direct reclaim makes progress so that
1837 * a zone that was recently full is not skipped over for up to a second
1839 static void zlc_clear_zones_full(struct zonelist *zonelist)
1841 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1843 zlc = zonelist->zlcache_ptr;
1847 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1850 static bool zone_local(struct zone *local_zone, struct zone *zone)
1852 return local_zone->node == zone->node;
1855 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1857 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1861 #else /* CONFIG_NUMA */
1863 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1868 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1869 nodemask_t *allowednodes)
1874 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1878 static void zlc_clear_zones_full(struct zonelist *zonelist)
1882 static bool zone_local(struct zone *local_zone, struct zone *zone)
1887 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1892 #endif /* CONFIG_NUMA */
1895 * get_page_from_freelist goes through the zonelist trying to allocate
1898 static struct page *
1899 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1900 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1901 struct zone *preferred_zone, int migratetype)
1904 struct page *page = NULL;
1907 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1908 int zlc_active = 0; /* set if using zonelist_cache */
1909 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1911 classzone_idx = zone_idx(preferred_zone);
1914 * Scan zonelist, looking for a zone with enough free.
1915 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1917 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1918 high_zoneidx, nodemask) {
1921 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1922 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1924 if ((alloc_flags & ALLOC_CPUSET) &&
1925 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1927 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1928 if (unlikely(alloc_flags & ALLOC_NO_WATERMARKS))
1931 * Distribute pages in proportion to the individual
1932 * zone size to ensure fair page aging. The zone a
1933 * page was allocated in should have no effect on the
1934 * time the page has in memory before being reclaimed.
1936 if (alloc_flags & ALLOC_FAIR) {
1937 if (!zone_local(preferred_zone, zone))
1939 if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1943 * When allocating a page cache page for writing, we
1944 * want to get it from a zone that is within its dirty
1945 * limit, such that no single zone holds more than its
1946 * proportional share of globally allowed dirty pages.
1947 * The dirty limits take into account the zone's
1948 * lowmem reserves and high watermark so that kswapd
1949 * should be able to balance it without having to
1950 * write pages from its LRU list.
1952 * This may look like it could increase pressure on
1953 * lower zones by failing allocations in higher zones
1954 * before they are full. But the pages that do spill
1955 * over are limited as the lower zones are protected
1956 * by this very same mechanism. It should not become
1957 * a practical burden to them.
1959 * XXX: For now, allow allocations to potentially
1960 * exceed the per-zone dirty limit in the slowpath
1961 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1962 * which is important when on a NUMA setup the allowed
1963 * zones are together not big enough to reach the
1964 * global limit. The proper fix for these situations
1965 * will require awareness of zones in the
1966 * dirty-throttling and the flusher threads.
1968 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1969 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1970 goto this_zone_full;
1972 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1973 if (!zone_watermark_ok(zone, order, mark,
1974 classzone_idx, alloc_flags)) {
1977 if (IS_ENABLED(CONFIG_NUMA) &&
1978 !did_zlc_setup && nr_online_nodes > 1) {
1980 * we do zlc_setup if there are multiple nodes
1981 * and before considering the first zone allowed
1984 allowednodes = zlc_setup(zonelist, alloc_flags);
1989 if (zone_reclaim_mode == 0 ||
1990 !zone_allows_reclaim(preferred_zone, zone))
1991 goto this_zone_full;
1994 * As we may have just activated ZLC, check if the first
1995 * eligible zone has failed zone_reclaim recently.
1997 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1998 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2001 ret = zone_reclaim(zone, gfp_mask, order);
2003 case ZONE_RECLAIM_NOSCAN:
2006 case ZONE_RECLAIM_FULL:
2007 /* scanned but unreclaimable */
2010 /* did we reclaim enough */
2011 if (zone_watermark_ok(zone, order, mark,
2012 classzone_idx, alloc_flags))
2016 * Failed to reclaim enough to meet watermark.
2017 * Only mark the zone full if checking the min
2018 * watermark or if we failed to reclaim just
2019 * 1<<order pages or else the page allocator
2020 * fastpath will prematurely mark zones full
2021 * when the watermark is between the low and
2024 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2025 ret == ZONE_RECLAIM_SOME)
2026 goto this_zone_full;
2033 page = buffered_rmqueue(preferred_zone, zone, order,
2034 gfp_mask, migratetype);
2038 if (IS_ENABLED(CONFIG_NUMA))
2039 zlc_mark_zone_full(zonelist, z);
2042 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2043 /* Disable zlc cache for second zonelist scan */
2050 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2051 * necessary to allocate the page. The expectation is
2052 * that the caller is taking steps that will free more
2053 * memory. The caller should avoid the page being used
2054 * for !PFMEMALLOC purposes.
2056 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2062 * Large machines with many possible nodes should not always dump per-node
2063 * meminfo in irq context.
2065 static inline bool should_suppress_show_mem(void)
2070 ret = in_interrupt();
2075 static DEFINE_RATELIMIT_STATE(nopage_rs,
2076 DEFAULT_RATELIMIT_INTERVAL,
2077 DEFAULT_RATELIMIT_BURST);
2079 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2081 unsigned int filter = SHOW_MEM_FILTER_NODES;
2083 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2084 debug_guardpage_minorder() > 0)
2088 * This documents exceptions given to allocations in certain
2089 * contexts that are allowed to allocate outside current's set
2092 if (!(gfp_mask & __GFP_NOMEMALLOC))
2093 if (test_thread_flag(TIF_MEMDIE) ||
2094 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2095 filter &= ~SHOW_MEM_FILTER_NODES;
2096 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2097 filter &= ~SHOW_MEM_FILTER_NODES;
2100 struct va_format vaf;
2103 va_start(args, fmt);
2108 pr_warn("%pV", &vaf);
2113 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2114 current->comm, order, gfp_mask);
2117 if (!should_suppress_show_mem())
2122 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2123 unsigned long did_some_progress,
2124 unsigned long pages_reclaimed)
2126 /* Do not loop if specifically requested */
2127 if (gfp_mask & __GFP_NORETRY)
2130 /* Always retry if specifically requested */
2131 if (gfp_mask & __GFP_NOFAIL)
2135 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2136 * making forward progress without invoking OOM. Suspend also disables
2137 * storage devices so kswapd will not help. Bail if we are suspending.
2139 if (!did_some_progress && pm_suspended_storage())
2143 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2144 * means __GFP_NOFAIL, but that may not be true in other
2147 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2151 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2152 * specified, then we retry until we no longer reclaim any pages
2153 * (above), or we've reclaimed an order of pages at least as
2154 * large as the allocation's order. In both cases, if the
2155 * allocation still fails, we stop retrying.
2157 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2163 static inline struct page *
2164 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2165 struct zonelist *zonelist, enum zone_type high_zoneidx,
2166 nodemask_t *nodemask, struct zone *preferred_zone,
2171 /* Acquire the OOM killer lock for the zones in zonelist */
2172 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2173 schedule_timeout_uninterruptible(1);
2178 * Go through the zonelist yet one more time, keep very high watermark
2179 * here, this is only to catch a parallel oom killing, we must fail if
2180 * we're still under heavy pressure.
2182 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2183 order, zonelist, high_zoneidx,
2184 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2185 preferred_zone, migratetype);
2189 if (!(gfp_mask & __GFP_NOFAIL)) {
2190 /* The OOM killer will not help higher order allocs */
2191 if (order > PAGE_ALLOC_COSTLY_ORDER)
2193 /* The OOM killer does not needlessly kill tasks for lowmem */
2194 if (high_zoneidx < ZONE_NORMAL)
2197 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2198 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2199 * The caller should handle page allocation failure by itself if
2200 * it specifies __GFP_THISNODE.
2201 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2203 if (gfp_mask & __GFP_THISNODE)
2206 /* Exhausted what can be done so it's blamo time */
2207 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2210 clear_zonelist_oom(zonelist, gfp_mask);
2214 #ifdef CONFIG_COMPACTION
2215 /* Try memory compaction for high-order allocations before reclaim */
2216 static struct page *
2217 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2218 struct zonelist *zonelist, enum zone_type high_zoneidx,
2219 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2220 int migratetype, bool sync_migration,
2221 bool *contended_compaction, bool *deferred_compaction,
2222 unsigned long *did_some_progress)
2227 if (compaction_deferred(preferred_zone, order)) {
2228 *deferred_compaction = true;
2232 current->flags |= PF_MEMALLOC;
2233 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2234 nodemask, sync_migration,
2235 contended_compaction);
2236 current->flags &= ~PF_MEMALLOC;
2238 if (*did_some_progress != COMPACT_SKIPPED) {
2241 /* Page migration frees to the PCP lists but we want merging */
2242 drain_pages(get_cpu());
2245 page = get_page_from_freelist(gfp_mask, nodemask,
2246 order, zonelist, high_zoneidx,
2247 alloc_flags & ~ALLOC_NO_WATERMARKS,
2248 preferred_zone, migratetype);
2250 preferred_zone->compact_blockskip_flush = false;
2251 compaction_defer_reset(preferred_zone, order, true);
2252 count_vm_event(COMPACTSUCCESS);
2257 * It's bad if compaction run occurs and fails.
2258 * The most likely reason is that pages exist,
2259 * but not enough to satisfy watermarks.
2261 count_vm_event(COMPACTFAIL);
2264 * As async compaction considers a subset of pageblocks, only
2265 * defer if the failure was a sync compaction failure.
2268 defer_compaction(preferred_zone, order);
2276 static inline struct page *
2277 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2278 struct zonelist *zonelist, enum zone_type high_zoneidx,
2279 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2280 int migratetype, bool sync_migration,
2281 bool *contended_compaction, bool *deferred_compaction,
2282 unsigned long *did_some_progress)
2286 #endif /* CONFIG_COMPACTION */
2288 /* Perform direct synchronous page reclaim */
2290 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2291 nodemask_t *nodemask)
2293 struct reclaim_state reclaim_state;
2298 /* We now go into synchronous reclaim */
2299 cpuset_memory_pressure_bump();
2300 current->flags |= PF_MEMALLOC;
2301 lockdep_set_current_reclaim_state(gfp_mask);
2302 reclaim_state.reclaimed_slab = 0;
2303 current->reclaim_state = &reclaim_state;
2305 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2307 current->reclaim_state = NULL;
2308 lockdep_clear_current_reclaim_state();
2309 current->flags &= ~PF_MEMALLOC;
2316 /* The really slow allocator path where we enter direct reclaim */
2317 static inline struct page *
2318 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2319 struct zonelist *zonelist, enum zone_type high_zoneidx,
2320 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2321 int migratetype, unsigned long *did_some_progress)
2323 struct page *page = NULL;
2324 bool drained = false;
2326 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2328 if (unlikely(!(*did_some_progress)))
2331 /* After successful reclaim, reconsider all zones for allocation */
2332 if (IS_ENABLED(CONFIG_NUMA))
2333 zlc_clear_zones_full(zonelist);
2336 page = get_page_from_freelist(gfp_mask, nodemask, order,
2337 zonelist, high_zoneidx,
2338 alloc_flags & ~ALLOC_NO_WATERMARKS,
2339 preferred_zone, migratetype);
2342 * If an allocation failed after direct reclaim, it could be because
2343 * pages are pinned on the per-cpu lists. Drain them and try again
2345 if (!page && !drained) {
2355 * This is called in the allocator slow-path if the allocation request is of
2356 * sufficient urgency to ignore watermarks and take other desperate measures
2358 static inline struct page *
2359 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2360 struct zonelist *zonelist, enum zone_type high_zoneidx,
2361 nodemask_t *nodemask, struct zone *preferred_zone,
2367 page = get_page_from_freelist(gfp_mask, nodemask, order,
2368 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2369 preferred_zone, migratetype);
2371 if (!page && gfp_mask & __GFP_NOFAIL)
2372 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2373 } while (!page && (gfp_mask & __GFP_NOFAIL));
2378 static void reset_alloc_batches(struct zonelist *zonelist,
2379 enum zone_type high_zoneidx,
2380 struct zone *preferred_zone)
2385 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2387 * Only reset the batches of zones that were actually
2388 * considered in the fairness pass, we don't want to
2389 * trash fairness information for zones that are not
2390 * actually part of this zonelist's round-robin cycle.
2392 if (!zone_local(preferred_zone, zone))
2394 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2395 high_wmark_pages(zone) - low_wmark_pages(zone) -
2396 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2400 static void wake_all_kswapds(unsigned int order,
2401 struct zonelist *zonelist,
2402 enum zone_type high_zoneidx,
2403 struct zone *preferred_zone)
2408 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2409 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2413 gfp_to_alloc_flags(gfp_t gfp_mask)
2415 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2416 const gfp_t wait = gfp_mask & __GFP_WAIT;
2418 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2419 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2422 * The caller may dip into page reserves a bit more if the caller
2423 * cannot run direct reclaim, or if the caller has realtime scheduling
2424 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2425 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2427 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2431 * Not worth trying to allocate harder for
2432 * __GFP_NOMEMALLOC even if it can't schedule.
2434 if (!(gfp_mask & __GFP_NOMEMALLOC))
2435 alloc_flags |= ALLOC_HARDER;
2437 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2438 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2440 alloc_flags &= ~ALLOC_CPUSET;
2441 } else if (unlikely(rt_task(current)) && !in_interrupt())
2442 alloc_flags |= ALLOC_HARDER;
2444 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2445 if (gfp_mask & __GFP_MEMALLOC)
2446 alloc_flags |= ALLOC_NO_WATERMARKS;
2447 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2448 alloc_flags |= ALLOC_NO_WATERMARKS;
2449 else if (!in_interrupt() &&
2450 ((current->flags & PF_MEMALLOC) ||
2451 unlikely(test_thread_flag(TIF_MEMDIE))))
2452 alloc_flags |= ALLOC_NO_WATERMARKS;
2455 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2456 alloc_flags |= ALLOC_CMA;
2461 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2463 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2466 static inline struct page *
2467 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2468 struct zonelist *zonelist, enum zone_type high_zoneidx,
2469 nodemask_t *nodemask, struct zone *preferred_zone,
2472 const gfp_t wait = gfp_mask & __GFP_WAIT;
2473 struct page *page = NULL;
2475 unsigned long pages_reclaimed = 0;
2476 unsigned long did_some_progress;
2477 bool sync_migration = false;
2478 bool deferred_compaction = false;
2479 bool contended_compaction = false;
2482 * In the slowpath, we sanity check order to avoid ever trying to
2483 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2484 * be using allocators in order of preference for an area that is
2487 if (order >= MAX_ORDER) {
2488 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2493 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2494 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2495 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2496 * using a larger set of nodes after it has established that the
2497 * allowed per node queues are empty and that nodes are
2500 if (IS_ENABLED(CONFIG_NUMA) &&
2501 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2505 if (!(gfp_mask & __GFP_NO_KSWAPD))
2506 wake_all_kswapds(order, zonelist, high_zoneidx, preferred_zone);
2509 * OK, we're below the kswapd watermark and have kicked background
2510 * reclaim. Now things get more complex, so set up alloc_flags according
2511 * to how we want to proceed.
2513 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2516 * Find the true preferred zone if the allocation is unconstrained by
2519 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2520 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2524 /* This is the last chance, in general, before the goto nopage. */
2525 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2526 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2527 preferred_zone, migratetype);
2531 /* Allocate without watermarks if the context allows */
2532 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2534 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2535 * the allocation is high priority and these type of
2536 * allocations are system rather than user orientated
2538 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2540 page = __alloc_pages_high_priority(gfp_mask, order,
2541 zonelist, high_zoneidx, nodemask,
2542 preferred_zone, migratetype);
2548 /* Atomic allocations - we can't balance anything */
2551 * All existing users of the deprecated __GFP_NOFAIL are
2552 * blockable, so warn of any new users that actually allow this
2553 * type of allocation to fail.
2555 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2559 /* Avoid recursion of direct reclaim */
2560 if (current->flags & PF_MEMALLOC)
2563 /* Avoid allocations with no watermarks from looping endlessly */
2564 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2568 * Try direct compaction. The first pass is asynchronous. Subsequent
2569 * attempts after direct reclaim are synchronous
2571 page = __alloc_pages_direct_compact(gfp_mask, order,
2572 zonelist, high_zoneidx,
2574 alloc_flags, preferred_zone,
2575 migratetype, sync_migration,
2576 &contended_compaction,
2577 &deferred_compaction,
2578 &did_some_progress);
2581 sync_migration = true;
2584 * If compaction is deferred for high-order allocations, it is because
2585 * sync compaction recently failed. In this is the case and the caller
2586 * requested a movable allocation that does not heavily disrupt the
2587 * system then fail the allocation instead of entering direct reclaim.
2589 if ((deferred_compaction || contended_compaction) &&
2590 (gfp_mask & __GFP_NO_KSWAPD))
2593 /* Try direct reclaim and then allocating */
2594 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2595 zonelist, high_zoneidx,
2597 alloc_flags, preferred_zone,
2598 migratetype, &did_some_progress);
2603 * If we failed to make any progress reclaiming, then we are
2604 * running out of options and have to consider going OOM
2606 if (!did_some_progress) {
2607 if (oom_gfp_allowed(gfp_mask)) {
2608 if (oom_killer_disabled)
2610 /* Coredumps can quickly deplete all memory reserves */
2611 if ((current->flags & PF_DUMPCORE) &&
2612 !(gfp_mask & __GFP_NOFAIL))
2614 page = __alloc_pages_may_oom(gfp_mask, order,
2615 zonelist, high_zoneidx,
2616 nodemask, preferred_zone,
2621 if (!(gfp_mask & __GFP_NOFAIL)) {
2623 * The oom killer is not called for high-order
2624 * allocations that may fail, so if no progress
2625 * is being made, there are no other options and
2626 * retrying is unlikely to help.
2628 if (order > PAGE_ALLOC_COSTLY_ORDER)
2631 * The oom killer is not called for lowmem
2632 * allocations to prevent needlessly killing
2635 if (high_zoneidx < ZONE_NORMAL)
2643 /* Check if we should retry the allocation */
2644 pages_reclaimed += did_some_progress;
2645 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2647 /* Wait for some write requests to complete then retry */
2648 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2652 * High-order allocations do not necessarily loop after
2653 * direct reclaim and reclaim/compaction depends on compaction
2654 * being called after reclaim so call directly if necessary
2656 page = __alloc_pages_direct_compact(gfp_mask, order,
2657 zonelist, high_zoneidx,
2659 alloc_flags, preferred_zone,
2660 migratetype, sync_migration,
2661 &contended_compaction,
2662 &deferred_compaction,
2663 &did_some_progress);
2669 warn_alloc_failed(gfp_mask, order, NULL);
2672 if (kmemcheck_enabled)
2673 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2679 * This is the 'heart' of the zoned buddy allocator.
2682 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2683 struct zonelist *zonelist, nodemask_t *nodemask)
2685 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2686 struct zone *preferred_zone;
2687 struct page *page = NULL;
2688 int migratetype = allocflags_to_migratetype(gfp_mask);
2689 unsigned int cpuset_mems_cookie;
2690 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2692 gfp_mask &= gfp_allowed_mask;
2694 lockdep_trace_alloc(gfp_mask);
2696 might_sleep_if(gfp_mask & __GFP_WAIT);
2698 if (should_fail_alloc_page(gfp_mask, order))
2702 * Check the zones suitable for the gfp_mask contain at least one
2703 * valid zone. It's possible to have an empty zonelist as a result
2704 * of GFP_THISNODE and a memoryless node
2706 if (unlikely(!zonelist->_zonerefs->zone))
2710 cpuset_mems_cookie = read_mems_allowed_begin();
2712 /* The preferred zone is used for statistics later */
2713 first_zones_zonelist(zonelist, high_zoneidx,
2714 nodemask ? : &cpuset_current_mems_allowed,
2716 if (!preferred_zone)
2720 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2721 alloc_flags |= ALLOC_CMA;
2724 /* First allocation attempt */
2725 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2726 zonelist, high_zoneidx, alloc_flags,
2727 preferred_zone, migratetype);
2728 if (unlikely(!page)) {
2730 * The first pass makes sure allocations are spread
2731 * fairly within the local node. However, the local
2732 * node might have free pages left after the fairness
2733 * batches are exhausted, and remote zones haven't
2734 * even been considered yet. Try once more without
2735 * fairness, and include remote zones now, before
2736 * entering the slowpath and waking kswapd: prefer
2737 * spilling to a remote zone over swapping locally.
2739 if (alloc_flags & ALLOC_FAIR) {
2740 reset_alloc_batches(zonelist, high_zoneidx,
2742 alloc_flags &= ~ALLOC_FAIR;
2746 * Runtime PM, block IO and its error handling path
2747 * can deadlock because I/O on the device might not
2750 gfp_mask = memalloc_noio_flags(gfp_mask);
2751 page = __alloc_pages_slowpath(gfp_mask, order,
2752 zonelist, high_zoneidx, nodemask,
2753 preferred_zone, migratetype);
2756 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2760 * When updating a task's mems_allowed, it is possible to race with
2761 * parallel threads in such a way that an allocation can fail while
2762 * the mask is being updated. If a page allocation is about to fail,
2763 * check if the cpuset changed during allocation and if so, retry.
2765 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2770 EXPORT_SYMBOL(__alloc_pages_nodemask);
2773 * Common helper functions.
2775 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2780 * __get_free_pages() returns a 32-bit address, which cannot represent
2783 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2785 page = alloc_pages(gfp_mask, order);
2788 return (unsigned long) page_address(page);
2790 EXPORT_SYMBOL(__get_free_pages);
2792 unsigned long get_zeroed_page(gfp_t gfp_mask)
2794 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2796 EXPORT_SYMBOL(get_zeroed_page);
2798 void __free_pages(struct page *page, unsigned int order)
2800 if (put_page_testzero(page)) {
2802 free_hot_cold_page(page, 0);
2804 __free_pages_ok(page, order);
2808 EXPORT_SYMBOL(__free_pages);
2810 void free_pages(unsigned long addr, unsigned int order)
2813 VM_BUG_ON(!virt_addr_valid((void *)addr));
2814 __free_pages(virt_to_page((void *)addr), order);
2818 EXPORT_SYMBOL(free_pages);
2821 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2822 * of the current memory cgroup.
2824 * It should be used when the caller would like to use kmalloc, but since the
2825 * allocation is large, it has to fall back to the page allocator.
2827 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2830 struct mem_cgroup *memcg = NULL;
2832 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2834 page = alloc_pages(gfp_mask, order);
2835 memcg_kmem_commit_charge(page, memcg, order);
2839 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2842 struct mem_cgroup *memcg = NULL;
2844 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2846 page = alloc_pages_node(nid, gfp_mask, order);
2847 memcg_kmem_commit_charge(page, memcg, order);
2852 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2855 void __free_kmem_pages(struct page *page, unsigned int order)
2857 memcg_kmem_uncharge_pages(page, order);
2858 __free_pages(page, order);
2861 void free_kmem_pages(unsigned long addr, unsigned int order)
2864 VM_BUG_ON(!virt_addr_valid((void *)addr));
2865 __free_kmem_pages(virt_to_page((void *)addr), order);
2869 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2872 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2873 unsigned long used = addr + PAGE_ALIGN(size);
2875 split_page(virt_to_page((void *)addr), order);
2876 while (used < alloc_end) {
2881 return (void *)addr;
2885 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2886 * @size: the number of bytes to allocate
2887 * @gfp_mask: GFP flags for the allocation
2889 * This function is similar to alloc_pages(), except that it allocates the
2890 * minimum number of pages to satisfy the request. alloc_pages() can only
2891 * allocate memory in power-of-two pages.
2893 * This function is also limited by MAX_ORDER.
2895 * Memory allocated by this function must be released by free_pages_exact().
2897 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2899 unsigned int order = get_order(size);
2902 addr = __get_free_pages(gfp_mask, order);
2903 return make_alloc_exact(addr, order, size);
2905 EXPORT_SYMBOL(alloc_pages_exact);
2908 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2910 * @nid: the preferred node ID where memory should be allocated
2911 * @size: the number of bytes to allocate
2912 * @gfp_mask: GFP flags for the allocation
2914 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2916 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2919 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2921 unsigned order = get_order(size);
2922 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2925 return make_alloc_exact((unsigned long)page_address(p), order, size);
2927 EXPORT_SYMBOL(alloc_pages_exact_nid);
2930 * free_pages_exact - release memory allocated via alloc_pages_exact()
2931 * @virt: the value returned by alloc_pages_exact.
2932 * @size: size of allocation, same value as passed to alloc_pages_exact().
2934 * Release the memory allocated by a previous call to alloc_pages_exact.
2936 void free_pages_exact(void *virt, size_t size)
2938 unsigned long addr = (unsigned long)virt;
2939 unsigned long end = addr + PAGE_ALIGN(size);
2941 while (addr < end) {
2946 EXPORT_SYMBOL(free_pages_exact);
2949 * nr_free_zone_pages - count number of pages beyond high watermark
2950 * @offset: The zone index of the highest zone
2952 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2953 * high watermark within all zones at or below a given zone index. For each
2954 * zone, the number of pages is calculated as:
2955 * managed_pages - high_pages
2957 static unsigned long nr_free_zone_pages(int offset)
2962 /* Just pick one node, since fallback list is circular */
2963 unsigned long sum = 0;
2965 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2967 for_each_zone_zonelist(zone, z, zonelist, offset) {
2968 unsigned long size = zone->managed_pages;
2969 unsigned long high = high_wmark_pages(zone);
2978 * nr_free_buffer_pages - count number of pages beyond high watermark
2980 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2981 * watermark within ZONE_DMA and ZONE_NORMAL.
2983 unsigned long nr_free_buffer_pages(void)
2985 return nr_free_zone_pages(gfp_zone(GFP_USER));
2987 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2990 * nr_free_pagecache_pages - count number of pages beyond high watermark
2992 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2993 * high watermark within all zones.
2995 unsigned long nr_free_pagecache_pages(void)
2997 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3000 static inline void show_node(struct zone *zone)
3002 if (IS_ENABLED(CONFIG_NUMA))
3003 printk("Node %d ", zone_to_nid(zone));
3006 void si_meminfo(struct sysinfo *val)
3008 val->totalram = totalram_pages;
3010 val->freeram = global_page_state(NR_FREE_PAGES);
3011 val->bufferram = nr_blockdev_pages();
3012 val->totalhigh = totalhigh_pages;
3013 val->freehigh = nr_free_highpages();
3014 val->mem_unit = PAGE_SIZE;
3017 EXPORT_SYMBOL(si_meminfo);
3020 void si_meminfo_node(struct sysinfo *val, int nid)
3022 int zone_type; /* needs to be signed */
3023 unsigned long managed_pages = 0;
3024 pg_data_t *pgdat = NODE_DATA(nid);
3026 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3027 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3028 val->totalram = managed_pages;
3029 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3030 #ifdef CONFIG_HIGHMEM
3031 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3032 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3038 val->mem_unit = PAGE_SIZE;
3043 * Determine whether the node should be displayed or not, depending on whether
3044 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3046 bool skip_free_areas_node(unsigned int flags, int nid)
3049 unsigned int cpuset_mems_cookie;
3051 if (!(flags & SHOW_MEM_FILTER_NODES))
3055 cpuset_mems_cookie = read_mems_allowed_begin();
3056 ret = !node_isset(nid, cpuset_current_mems_allowed);
3057 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3062 #define K(x) ((x) << (PAGE_SHIFT-10))
3064 static void show_migration_types(unsigned char type)
3066 static const char types[MIGRATE_TYPES] = {
3067 [MIGRATE_UNMOVABLE] = 'U',
3068 [MIGRATE_RECLAIMABLE] = 'E',
3069 [MIGRATE_MOVABLE] = 'M',
3070 [MIGRATE_RESERVE] = 'R',
3072 [MIGRATE_CMA] = 'C',
3074 #ifdef CONFIG_MEMORY_ISOLATION
3075 [MIGRATE_ISOLATE] = 'I',
3078 char tmp[MIGRATE_TYPES + 1];
3082 for (i = 0; i < MIGRATE_TYPES; i++) {
3083 if (type & (1 << i))
3088 printk("(%s) ", tmp);
3092 * Show free area list (used inside shift_scroll-lock stuff)
3093 * We also calculate the percentage fragmentation. We do this by counting the
3094 * memory on each free list with the exception of the first item on the list.
3095 * Suppresses nodes that are not allowed by current's cpuset if
3096 * SHOW_MEM_FILTER_NODES is passed.
3098 void show_free_areas(unsigned int filter)
3103 for_each_populated_zone(zone) {
3104 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3107 printk("%s per-cpu:\n", zone->name);
3109 for_each_online_cpu(cpu) {
3110 struct per_cpu_pageset *pageset;
3112 pageset = per_cpu_ptr(zone->pageset, cpu);
3114 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3115 cpu, pageset->pcp.high,
3116 pageset->pcp.batch, pageset->pcp.count);
3120 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3121 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3123 " dirty:%lu writeback:%lu unstable:%lu\n"
3124 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3125 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3127 global_page_state(NR_ACTIVE_ANON),
3128 global_page_state(NR_INACTIVE_ANON),
3129 global_page_state(NR_ISOLATED_ANON),
3130 global_page_state(NR_ACTIVE_FILE),
3131 global_page_state(NR_INACTIVE_FILE),
3132 global_page_state(NR_ISOLATED_FILE),
3133 global_page_state(NR_UNEVICTABLE),
3134 global_page_state(NR_FILE_DIRTY),
3135 global_page_state(NR_WRITEBACK),
3136 global_page_state(NR_UNSTABLE_NFS),
3137 global_page_state(NR_FREE_PAGES),
3138 global_page_state(NR_SLAB_RECLAIMABLE),
3139 global_page_state(NR_SLAB_UNRECLAIMABLE),
3140 global_page_state(NR_FILE_MAPPED),
3141 global_page_state(NR_SHMEM),
3142 global_page_state(NR_PAGETABLE),
3143 global_page_state(NR_BOUNCE),
3144 global_page_state(NR_FREE_CMA_PAGES));
3146 for_each_populated_zone(zone) {
3149 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3157 " active_anon:%lukB"
3158 " inactive_anon:%lukB"
3159 " active_file:%lukB"
3160 " inactive_file:%lukB"
3161 " unevictable:%lukB"
3162 " isolated(anon):%lukB"
3163 " isolated(file):%lukB"
3171 " slab_reclaimable:%lukB"
3172 " slab_unreclaimable:%lukB"
3173 " kernel_stack:%lukB"
3178 " writeback_tmp:%lukB"
3179 " pages_scanned:%lu"
3180 " all_unreclaimable? %s"
3183 K(zone_page_state(zone, NR_FREE_PAGES)),
3184 K(min_wmark_pages(zone)),
3185 K(low_wmark_pages(zone)),
3186 K(high_wmark_pages(zone)),
3187 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3188 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3189 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3190 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3191 K(zone_page_state(zone, NR_UNEVICTABLE)),
3192 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3193 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3194 K(zone->present_pages),
3195 K(zone->managed_pages),
3196 K(zone_page_state(zone, NR_MLOCK)),
3197 K(zone_page_state(zone, NR_FILE_DIRTY)),
3198 K(zone_page_state(zone, NR_WRITEBACK)),
3199 K(zone_page_state(zone, NR_FILE_MAPPED)),
3200 K(zone_page_state(zone, NR_SHMEM)),
3201 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3202 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3203 zone_page_state(zone, NR_KERNEL_STACK) *
3205 K(zone_page_state(zone, NR_PAGETABLE)),
3206 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3207 K(zone_page_state(zone, NR_BOUNCE)),
3208 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3209 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3210 zone->pages_scanned,
3211 (!zone_reclaimable(zone) ? "yes" : "no")
3213 printk("lowmem_reserve[]:");
3214 for (i = 0; i < MAX_NR_ZONES; i++)
3215 printk(" %lu", zone->lowmem_reserve[i]);
3219 for_each_populated_zone(zone) {
3220 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3221 unsigned char types[MAX_ORDER];
3223 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3226 printk("%s: ", zone->name);
3228 spin_lock_irqsave(&zone->lock, flags);
3229 for (order = 0; order < MAX_ORDER; order++) {
3230 struct free_area *area = &zone->free_area[order];
3233 nr[order] = area->nr_free;
3234 total += nr[order] << order;
3237 for (type = 0; type < MIGRATE_TYPES; type++) {
3238 if (!list_empty(&area->free_list[type]))
3239 types[order] |= 1 << type;
3242 spin_unlock_irqrestore(&zone->lock, flags);
3243 for (order = 0; order < MAX_ORDER; order++) {
3244 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3246 show_migration_types(types[order]);
3248 printk("= %lukB\n", K(total));
3251 hugetlb_show_meminfo();
3253 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3255 show_swap_cache_info();
3258 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3260 zoneref->zone = zone;
3261 zoneref->zone_idx = zone_idx(zone);
3265 * Builds allocation fallback zone lists.
3267 * Add all populated zones of a node to the zonelist.
3269 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3273 enum zone_type zone_type = MAX_NR_ZONES;
3277 zone = pgdat->node_zones + zone_type;
3278 if (populated_zone(zone)) {
3279 zoneref_set_zone(zone,
3280 &zonelist->_zonerefs[nr_zones++]);
3281 check_highest_zone(zone_type);
3283 } while (zone_type);
3291 * 0 = automatic detection of better ordering.
3292 * 1 = order by ([node] distance, -zonetype)
3293 * 2 = order by (-zonetype, [node] distance)
3295 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3296 * the same zonelist. So only NUMA can configure this param.
3298 #define ZONELIST_ORDER_DEFAULT 0
3299 #define ZONELIST_ORDER_NODE 1
3300 #define ZONELIST_ORDER_ZONE 2
3302 /* zonelist order in the kernel.
3303 * set_zonelist_order() will set this to NODE or ZONE.
3305 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3306 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3310 /* The value user specified ....changed by config */
3311 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3312 /* string for sysctl */
3313 #define NUMA_ZONELIST_ORDER_LEN 16
3314 char numa_zonelist_order[16] = "default";
3317 * interface for configure zonelist ordering.
3318 * command line option "numa_zonelist_order"
3319 * = "[dD]efault - default, automatic configuration.
3320 * = "[nN]ode - order by node locality, then by zone within node
3321 * = "[zZ]one - order by zone, then by locality within zone
3324 static int __parse_numa_zonelist_order(char *s)
3326 if (*s == 'd' || *s == 'D') {
3327 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3328 } else if (*s == 'n' || *s == 'N') {
3329 user_zonelist_order = ZONELIST_ORDER_NODE;
3330 } else if (*s == 'z' || *s == 'Z') {
3331 user_zonelist_order = ZONELIST_ORDER_ZONE;
3334 "Ignoring invalid numa_zonelist_order value: "
3341 static __init int setup_numa_zonelist_order(char *s)
3348 ret = __parse_numa_zonelist_order(s);
3350 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3354 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3357 * sysctl handler for numa_zonelist_order
3359 int numa_zonelist_order_handler(ctl_table *table, int write,
3360 void __user *buffer, size_t *length,
3363 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3365 static DEFINE_MUTEX(zl_order_mutex);
3367 mutex_lock(&zl_order_mutex);
3369 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3373 strcpy(saved_string, (char *)table->data);
3375 ret = proc_dostring(table, write, buffer, length, ppos);
3379 int oldval = user_zonelist_order;
3381 ret = __parse_numa_zonelist_order((char *)table->data);
3384 * bogus value. restore saved string
3386 strncpy((char *)table->data, saved_string,
3387 NUMA_ZONELIST_ORDER_LEN);
3388 user_zonelist_order = oldval;
3389 } else if (oldval != user_zonelist_order) {
3390 mutex_lock(&zonelists_mutex);
3391 build_all_zonelists(NULL, NULL);
3392 mutex_unlock(&zonelists_mutex);
3396 mutex_unlock(&zl_order_mutex);
3401 #define MAX_NODE_LOAD (nr_online_nodes)
3402 static int node_load[MAX_NUMNODES];
3405 * find_next_best_node - find the next node that should appear in a given node's fallback list
3406 * @node: node whose fallback list we're appending
3407 * @used_node_mask: nodemask_t of already used nodes
3409 * We use a number of factors to determine which is the next node that should
3410 * appear on a given node's fallback list. The node should not have appeared
3411 * already in @node's fallback list, and it should be the next closest node
3412 * according to the distance array (which contains arbitrary distance values
3413 * from each node to each node in the system), and should also prefer nodes
3414 * with no CPUs, since presumably they'll have very little allocation pressure
3415 * on them otherwise.
3416 * It returns -1 if no node is found.
3418 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3421 int min_val = INT_MAX;
3422 int best_node = NUMA_NO_NODE;
3423 const struct cpumask *tmp = cpumask_of_node(0);
3425 /* Use the local node if we haven't already */
3426 if (!node_isset(node, *used_node_mask)) {
3427 node_set(node, *used_node_mask);
3431 for_each_node_state(n, N_MEMORY) {
3433 /* Don't want a node to appear more than once */
3434 if (node_isset(n, *used_node_mask))
3437 /* Use the distance array to find the distance */
3438 val = node_distance(node, n);
3440 /* Penalize nodes under us ("prefer the next node") */
3443 /* Give preference to headless and unused nodes */
3444 tmp = cpumask_of_node(n);
3445 if (!cpumask_empty(tmp))
3446 val += PENALTY_FOR_NODE_WITH_CPUS;
3448 /* Slight preference for less loaded node */
3449 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3450 val += node_load[n];
3452 if (val < min_val) {
3459 node_set(best_node, *used_node_mask);
3466 * Build zonelists ordered by node and zones within node.
3467 * This results in maximum locality--normal zone overflows into local
3468 * DMA zone, if any--but risks exhausting DMA zone.
3470 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3473 struct zonelist *zonelist;
3475 zonelist = &pgdat->node_zonelists[0];
3476 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3478 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3479 zonelist->_zonerefs[j].zone = NULL;
3480 zonelist->_zonerefs[j].zone_idx = 0;
3484 * Build gfp_thisnode zonelists
3486 static void build_thisnode_zonelists(pg_data_t *pgdat)
3489 struct zonelist *zonelist;
3491 zonelist = &pgdat->node_zonelists[1];
3492 j = build_zonelists_node(pgdat, zonelist, 0);
3493 zonelist->_zonerefs[j].zone = NULL;
3494 zonelist->_zonerefs[j].zone_idx = 0;
3498 * Build zonelists ordered by zone and nodes within zones.
3499 * This results in conserving DMA zone[s] until all Normal memory is
3500 * exhausted, but results in overflowing to remote node while memory
3501 * may still exist in local DMA zone.
3503 static int node_order[MAX_NUMNODES];
3505 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3508 int zone_type; /* needs to be signed */
3510 struct zonelist *zonelist;
3512 zonelist = &pgdat->node_zonelists[0];
3514 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3515 for (j = 0; j < nr_nodes; j++) {
3516 node = node_order[j];
3517 z = &NODE_DATA(node)->node_zones[zone_type];
3518 if (populated_zone(z)) {
3520 &zonelist->_zonerefs[pos++]);
3521 check_highest_zone(zone_type);
3525 zonelist->_zonerefs[pos].zone = NULL;
3526 zonelist->_zonerefs[pos].zone_idx = 0;
3529 static int default_zonelist_order(void)
3532 unsigned long low_kmem_size, total_size;
3536 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3537 * If they are really small and used heavily, the system can fall
3538 * into OOM very easily.
3539 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3541 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3544 for_each_online_node(nid) {
3545 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3546 z = &NODE_DATA(nid)->node_zones[zone_type];
3547 if (populated_zone(z)) {
3548 if (zone_type < ZONE_NORMAL)
3549 low_kmem_size += z->managed_pages;
3550 total_size += z->managed_pages;
3551 } else if (zone_type == ZONE_NORMAL) {
3553 * If any node has only lowmem, then node order
3554 * is preferred to allow kernel allocations
3555 * locally; otherwise, they can easily infringe
3556 * on other nodes when there is an abundance of
3557 * lowmem available to allocate from.
3559 return ZONELIST_ORDER_NODE;
3563 if (!low_kmem_size || /* there are no DMA area. */
3564 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3565 return ZONELIST_ORDER_NODE;
3567 * look into each node's config.
3568 * If there is a node whose DMA/DMA32 memory is very big area on
3569 * local memory, NODE_ORDER may be suitable.
3571 average_size = total_size /
3572 (nodes_weight(node_states[N_MEMORY]) + 1);
3573 for_each_online_node(nid) {
3576 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3577 z = &NODE_DATA(nid)->node_zones[zone_type];
3578 if (populated_zone(z)) {
3579 if (zone_type < ZONE_NORMAL)
3580 low_kmem_size += z->present_pages;
3581 total_size += z->present_pages;
3584 if (low_kmem_size &&
3585 total_size > average_size && /* ignore small node */
3586 low_kmem_size > total_size * 70/100)
3587 return ZONELIST_ORDER_NODE;
3589 return ZONELIST_ORDER_ZONE;
3592 static void set_zonelist_order(void)
3594 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3595 current_zonelist_order = default_zonelist_order();
3597 current_zonelist_order = user_zonelist_order;
3600 static void build_zonelists(pg_data_t *pgdat)
3604 nodemask_t used_mask;
3605 int local_node, prev_node;
3606 struct zonelist *zonelist;
3607 int order = current_zonelist_order;
3609 /* initialize zonelists */
3610 for (i = 0; i < MAX_ZONELISTS; i++) {
3611 zonelist = pgdat->node_zonelists + i;
3612 zonelist->_zonerefs[0].zone = NULL;
3613 zonelist->_zonerefs[0].zone_idx = 0;
3616 /* NUMA-aware ordering of nodes */
3617 local_node = pgdat->node_id;
3618 load = nr_online_nodes;
3619 prev_node = local_node;
3620 nodes_clear(used_mask);
3622 memset(node_order, 0, sizeof(node_order));
3625 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3627 * We don't want to pressure a particular node.
3628 * So adding penalty to the first node in same
3629 * distance group to make it round-robin.
3631 if (node_distance(local_node, node) !=
3632 node_distance(local_node, prev_node))
3633 node_load[node] = load;
3637 if (order == ZONELIST_ORDER_NODE)
3638 build_zonelists_in_node_order(pgdat, node);
3640 node_order[j++] = node; /* remember order */
3643 if (order == ZONELIST_ORDER_ZONE) {
3644 /* calculate node order -- i.e., DMA last! */
3645 build_zonelists_in_zone_order(pgdat, j);
3648 build_thisnode_zonelists(pgdat);
3651 /* Construct the zonelist performance cache - see further mmzone.h */
3652 static void build_zonelist_cache(pg_data_t *pgdat)
3654 struct zonelist *zonelist;
3655 struct zonelist_cache *zlc;
3658 zonelist = &pgdat->node_zonelists[0];
3659 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3660 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3661 for (z = zonelist->_zonerefs; z->zone; z++)
3662 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3665 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3667 * Return node id of node used for "local" allocations.
3668 * I.e., first node id of first zone in arg node's generic zonelist.
3669 * Used for initializing percpu 'numa_mem', which is used primarily
3670 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3672 int local_memory_node(int node)
3676 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3677 gfp_zone(GFP_KERNEL),
3684 #else /* CONFIG_NUMA */
3686 static void set_zonelist_order(void)
3688 current_zonelist_order = ZONELIST_ORDER_ZONE;
3691 static void build_zonelists(pg_data_t *pgdat)
3693 int node, local_node;
3695 struct zonelist *zonelist;
3697 local_node = pgdat->node_id;
3699 zonelist = &pgdat->node_zonelists[0];
3700 j = build_zonelists_node(pgdat, zonelist, 0);
3703 * Now we build the zonelist so that it contains the zones
3704 * of all the other nodes.
3705 * We don't want to pressure a particular node, so when
3706 * building the zones for node N, we make sure that the
3707 * zones coming right after the local ones are those from
3708 * node N+1 (modulo N)
3710 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3711 if (!node_online(node))
3713 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3715 for (node = 0; node < local_node; node++) {
3716 if (!node_online(node))
3718 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3721 zonelist->_zonerefs[j].zone = NULL;
3722 zonelist->_zonerefs[j].zone_idx = 0;
3725 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3726 static void build_zonelist_cache(pg_data_t *pgdat)
3728 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3731 #endif /* CONFIG_NUMA */
3734 * Boot pageset table. One per cpu which is going to be used for all
3735 * zones and all nodes. The parameters will be set in such a way
3736 * that an item put on a list will immediately be handed over to
3737 * the buddy list. This is safe since pageset manipulation is done
3738 * with interrupts disabled.
3740 * The boot_pagesets must be kept even after bootup is complete for
3741 * unused processors and/or zones. They do play a role for bootstrapping
3742 * hotplugged processors.
3744 * zoneinfo_show() and maybe other functions do
3745 * not check if the processor is online before following the pageset pointer.
3746 * Other parts of the kernel may not check if the zone is available.
3748 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3749 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3750 static void setup_zone_pageset(struct zone *zone);
3753 * Global mutex to protect against size modification of zonelists
3754 * as well as to serialize pageset setup for the new populated zone.
3756 DEFINE_MUTEX(zonelists_mutex);
3758 /* return values int ....just for stop_machine() */
3759 static int __build_all_zonelists(void *data)
3763 pg_data_t *self = data;
3766 memset(node_load, 0, sizeof(node_load));
3769 if (self && !node_online(self->node_id)) {
3770 build_zonelists(self);
3771 build_zonelist_cache(self);
3774 for_each_online_node(nid) {
3775 pg_data_t *pgdat = NODE_DATA(nid);
3777 build_zonelists(pgdat);
3778 build_zonelist_cache(pgdat);
3782 * Initialize the boot_pagesets that are going to be used
3783 * for bootstrapping processors. The real pagesets for
3784 * each zone will be allocated later when the per cpu
3785 * allocator is available.
3787 * boot_pagesets are used also for bootstrapping offline
3788 * cpus if the system is already booted because the pagesets
3789 * are needed to initialize allocators on a specific cpu too.
3790 * F.e. the percpu allocator needs the page allocator which
3791 * needs the percpu allocator in order to allocate its pagesets
3792 * (a chicken-egg dilemma).
3794 for_each_possible_cpu(cpu) {
3795 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3797 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3799 * We now know the "local memory node" for each node--
3800 * i.e., the node of the first zone in the generic zonelist.
3801 * Set up numa_mem percpu variable for on-line cpus. During
3802 * boot, only the boot cpu should be on-line; we'll init the
3803 * secondary cpus' numa_mem as they come on-line. During
3804 * node/memory hotplug, we'll fixup all on-line cpus.
3806 if (cpu_online(cpu))
3807 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3815 * Called with zonelists_mutex held always
3816 * unless system_state == SYSTEM_BOOTING.
3818 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3820 set_zonelist_order();
3822 if (system_state == SYSTEM_BOOTING) {
3823 __build_all_zonelists(NULL);
3824 mminit_verify_zonelist();
3825 cpuset_init_current_mems_allowed();
3827 #ifdef CONFIG_MEMORY_HOTPLUG
3829 setup_zone_pageset(zone);
3831 /* we have to stop all cpus to guarantee there is no user
3833 stop_machine(__build_all_zonelists, pgdat, NULL);
3834 /* cpuset refresh routine should be here */
3836 vm_total_pages = nr_free_pagecache_pages();
3838 * Disable grouping by mobility if the number of pages in the
3839 * system is too low to allow the mechanism to work. It would be
3840 * more accurate, but expensive to check per-zone. This check is
3841 * made on memory-hotadd so a system can start with mobility
3842 * disabled and enable it later
3844 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3845 page_group_by_mobility_disabled = 1;
3847 page_group_by_mobility_disabled = 0;
3849 printk("Built %i zonelists in %s order, mobility grouping %s. "
3850 "Total pages: %ld\n",
3852 zonelist_order_name[current_zonelist_order],
3853 page_group_by_mobility_disabled ? "off" : "on",
3856 printk("Policy zone: %s\n", zone_names[policy_zone]);
3861 * Helper functions to size the waitqueue hash table.
3862 * Essentially these want to choose hash table sizes sufficiently
3863 * large so that collisions trying to wait on pages are rare.
3864 * But in fact, the number of active page waitqueues on typical
3865 * systems is ridiculously low, less than 200. So this is even
3866 * conservative, even though it seems large.
3868 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3869 * waitqueues, i.e. the size of the waitq table given the number of pages.
3871 #define PAGES_PER_WAITQUEUE 256
3873 #ifndef CONFIG_MEMORY_HOTPLUG
3874 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3876 unsigned long size = 1;
3878 pages /= PAGES_PER_WAITQUEUE;
3880 while (size < pages)
3884 * Once we have dozens or even hundreds of threads sleeping
3885 * on IO we've got bigger problems than wait queue collision.
3886 * Limit the size of the wait table to a reasonable size.
3888 size = min(size, 4096UL);
3890 return max(size, 4UL);
3894 * A zone's size might be changed by hot-add, so it is not possible to determine
3895 * a suitable size for its wait_table. So we use the maximum size now.
3897 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3899 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3900 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3901 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3903 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3904 * or more by the traditional way. (See above). It equals:
3906 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3907 * ia64(16K page size) : = ( 8G + 4M)byte.
3908 * powerpc (64K page size) : = (32G +16M)byte.
3910 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3917 * This is an integer logarithm so that shifts can be used later
3918 * to extract the more random high bits from the multiplicative
3919 * hash function before the remainder is taken.
3921 static inline unsigned long wait_table_bits(unsigned long size)
3927 * Check if a pageblock contains reserved pages
3929 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3933 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3934 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3941 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3942 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3943 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3944 * higher will lead to a bigger reserve which will get freed as contiguous
3945 * blocks as reclaim kicks in
3947 static void setup_zone_migrate_reserve(struct zone *zone)
3949 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3951 unsigned long block_migratetype;
3956 * Get the start pfn, end pfn and the number of blocks to reserve
3957 * We have to be careful to be aligned to pageblock_nr_pages to
3958 * make sure that we always check pfn_valid for the first page in
3961 start_pfn = zone->zone_start_pfn;
3962 end_pfn = zone_end_pfn(zone);
3963 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3964 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3968 * Reserve blocks are generally in place to help high-order atomic
3969 * allocations that are short-lived. A min_free_kbytes value that
3970 * would result in more than 2 reserve blocks for atomic allocations
3971 * is assumed to be in place to help anti-fragmentation for the
3972 * future allocation of hugepages at runtime.
3974 reserve = min(2, reserve);
3975 old_reserve = zone->nr_migrate_reserve_block;
3977 /* When memory hot-add, we almost always need to do nothing */
3978 if (reserve == old_reserve)
3980 zone->nr_migrate_reserve_block = reserve;
3982 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3983 if (!pfn_valid(pfn))
3985 page = pfn_to_page(pfn);
3987 /* Watch out for overlapping nodes */
3988 if (page_to_nid(page) != zone_to_nid(zone))
3991 block_migratetype = get_pageblock_migratetype(page);
3993 /* Only test what is necessary when the reserves are not met */
3996 * Blocks with reserved pages will never free, skip
3999 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4000 if (pageblock_is_reserved(pfn, block_end_pfn))
4003 /* If this block is reserved, account for it */
4004 if (block_migratetype == MIGRATE_RESERVE) {
4009 /* Suitable for reserving if this block is movable */
4010 if (block_migratetype == MIGRATE_MOVABLE) {
4011 set_pageblock_migratetype(page,
4013 move_freepages_block(zone, page,
4018 } else if (!old_reserve) {
4020 * At boot time we don't need to scan the whole zone
4021 * for turning off MIGRATE_RESERVE.
4027 * If the reserve is met and this is a previous reserved block,
4030 if (block_migratetype == MIGRATE_RESERVE) {
4031 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4032 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4038 * Initially all pages are reserved - free ones are freed
4039 * up by free_all_bootmem() once the early boot process is
4040 * done. Non-atomic initialization, single-pass.
4042 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4043 unsigned long start_pfn, enum memmap_context context)
4046 unsigned long end_pfn = start_pfn + size;
4050 if (highest_memmap_pfn < end_pfn - 1)
4051 highest_memmap_pfn = end_pfn - 1;
4053 z = &NODE_DATA(nid)->node_zones[zone];
4054 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4056 * There can be holes in boot-time mem_map[]s
4057 * handed to this function. They do not
4058 * exist on hotplugged memory.
4060 if (context == MEMMAP_EARLY) {
4061 if (!early_pfn_valid(pfn))
4063 if (!early_pfn_in_nid(pfn, nid))
4066 page = pfn_to_page(pfn);
4067 set_page_links(page, zone, nid, pfn);
4068 mminit_verify_page_links(page, zone, nid, pfn);
4069 init_page_count(page);
4070 page_mapcount_reset(page);
4071 page_cpupid_reset_last(page);
4072 SetPageReserved(page);
4074 * Mark the block movable so that blocks are reserved for
4075 * movable at startup. This will force kernel allocations
4076 * to reserve their blocks rather than leaking throughout
4077 * the address space during boot when many long-lived
4078 * kernel allocations are made. Later some blocks near
4079 * the start are marked MIGRATE_RESERVE by
4080 * setup_zone_migrate_reserve()
4082 * bitmap is created for zone's valid pfn range. but memmap
4083 * can be created for invalid pages (for alignment)
4084 * check here not to call set_pageblock_migratetype() against
4087 if ((z->zone_start_pfn <= pfn)
4088 && (pfn < zone_end_pfn(z))
4089 && !(pfn & (pageblock_nr_pages - 1)))
4090 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4092 INIT_LIST_HEAD(&page->lru);
4093 #ifdef WANT_PAGE_VIRTUAL
4094 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4095 if (!is_highmem_idx(zone))
4096 set_page_address(page, __va(pfn << PAGE_SHIFT));
4101 static void __meminit zone_init_free_lists(struct zone *zone)
4104 for_each_migratetype_order(order, t) {
4105 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4106 zone->free_area[order].nr_free = 0;
4110 #ifndef __HAVE_ARCH_MEMMAP_INIT
4111 #define memmap_init(size, nid, zone, start_pfn) \
4112 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4115 static int __meminit zone_batchsize(struct zone *zone)
4121 * The per-cpu-pages pools are set to around 1000th of the
4122 * size of the zone. But no more than 1/2 of a meg.
4124 * OK, so we don't know how big the cache is. So guess.
4126 batch = zone->managed_pages / 1024;
4127 if (batch * PAGE_SIZE > 512 * 1024)
4128 batch = (512 * 1024) / PAGE_SIZE;
4129 batch /= 4; /* We effectively *= 4 below */
4134 * Clamp the batch to a 2^n - 1 value. Having a power
4135 * of 2 value was found to be more likely to have
4136 * suboptimal cache aliasing properties in some cases.
4138 * For example if 2 tasks are alternately allocating
4139 * batches of pages, one task can end up with a lot
4140 * of pages of one half of the possible page colors
4141 * and the other with pages of the other colors.
4143 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4148 /* The deferral and batching of frees should be suppressed under NOMMU
4151 * The problem is that NOMMU needs to be able to allocate large chunks
4152 * of contiguous memory as there's no hardware page translation to
4153 * assemble apparent contiguous memory from discontiguous pages.
4155 * Queueing large contiguous runs of pages for batching, however,
4156 * causes the pages to actually be freed in smaller chunks. As there
4157 * can be a significant delay between the individual batches being
4158 * recycled, this leads to the once large chunks of space being
4159 * fragmented and becoming unavailable for high-order allocations.
4166 * pcp->high and pcp->batch values are related and dependent on one another:
4167 * ->batch must never be higher then ->high.
4168 * The following function updates them in a safe manner without read side
4171 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4172 * those fields changing asynchronously (acording the the above rule).
4174 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4175 * outside of boot time (or some other assurance that no concurrent updaters
4178 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4179 unsigned long batch)
4181 /* start with a fail safe value for batch */
4185 /* Update high, then batch, in order */
4192 /* a companion to pageset_set_high() */
4193 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4195 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4198 static void pageset_init(struct per_cpu_pageset *p)
4200 struct per_cpu_pages *pcp;
4203 memset(p, 0, sizeof(*p));
4207 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4208 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4211 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4214 pageset_set_batch(p, batch);
4218 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4219 * to the value high for the pageset p.
4221 static void pageset_set_high(struct per_cpu_pageset *p,
4224 unsigned long batch = max(1UL, high / 4);
4225 if ((high / 4) > (PAGE_SHIFT * 8))
4226 batch = PAGE_SHIFT * 8;
4228 pageset_update(&p->pcp, high, batch);
4231 static void __meminit pageset_set_high_and_batch(struct zone *zone,
4232 struct per_cpu_pageset *pcp)
4234 if (percpu_pagelist_fraction)
4235 pageset_set_high(pcp,
4236 (zone->managed_pages /
4237 percpu_pagelist_fraction));
4239 pageset_set_batch(pcp, zone_batchsize(zone));
4242 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4244 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4247 pageset_set_high_and_batch(zone, pcp);
4250 static void __meminit setup_zone_pageset(struct zone *zone)
4253 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4254 for_each_possible_cpu(cpu)
4255 zone_pageset_init(zone, cpu);
4259 * Allocate per cpu pagesets and initialize them.
4260 * Before this call only boot pagesets were available.
4262 void __init setup_per_cpu_pageset(void)
4266 for_each_populated_zone(zone)
4267 setup_zone_pageset(zone);
4270 static noinline __init_refok
4271 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4277 * The per-page waitqueue mechanism uses hashed waitqueues
4280 zone->wait_table_hash_nr_entries =
4281 wait_table_hash_nr_entries(zone_size_pages);
4282 zone->wait_table_bits =
4283 wait_table_bits(zone->wait_table_hash_nr_entries);
4284 alloc_size = zone->wait_table_hash_nr_entries
4285 * sizeof(wait_queue_head_t);
4287 if (!slab_is_available()) {
4288 zone->wait_table = (wait_queue_head_t *)
4289 memblock_virt_alloc_node_nopanic(
4290 alloc_size, zone->zone_pgdat->node_id);
4293 * This case means that a zone whose size was 0 gets new memory
4294 * via memory hot-add.
4295 * But it may be the case that a new node was hot-added. In
4296 * this case vmalloc() will not be able to use this new node's
4297 * memory - this wait_table must be initialized to use this new
4298 * node itself as well.
4299 * To use this new node's memory, further consideration will be
4302 zone->wait_table = vmalloc(alloc_size);
4304 if (!zone->wait_table)
4307 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4308 init_waitqueue_head(zone->wait_table + i);
4313 static __meminit void zone_pcp_init(struct zone *zone)
4316 * per cpu subsystem is not up at this point. The following code
4317 * relies on the ability of the linker to provide the
4318 * offset of a (static) per cpu variable into the per cpu area.
4320 zone->pageset = &boot_pageset;
4322 if (populated_zone(zone))
4323 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4324 zone->name, zone->present_pages,
4325 zone_batchsize(zone));
4328 int __meminit init_currently_empty_zone(struct zone *zone,
4329 unsigned long zone_start_pfn,
4331 enum memmap_context context)
4333 struct pglist_data *pgdat = zone->zone_pgdat;
4335 ret = zone_wait_table_init(zone, size);
4338 pgdat->nr_zones = zone_idx(zone) + 1;
4340 zone->zone_start_pfn = zone_start_pfn;
4342 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4343 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4345 (unsigned long)zone_idx(zone),
4346 zone_start_pfn, (zone_start_pfn + size));
4348 zone_init_free_lists(zone);
4353 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4354 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4356 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4357 * Architectures may implement their own version but if add_active_range()
4358 * was used and there are no special requirements, this is a convenient
4361 int __meminit __early_pfn_to_nid(unsigned long pfn)
4363 unsigned long start_pfn, end_pfn;
4366 * NOTE: The following SMP-unsafe globals are only used early in boot
4367 * when the kernel is running single-threaded.
4369 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4370 static int __meminitdata last_nid;
4372 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4375 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4377 last_start_pfn = start_pfn;
4378 last_end_pfn = end_pfn;
4384 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4386 int __meminit early_pfn_to_nid(unsigned long pfn)
4390 nid = __early_pfn_to_nid(pfn);
4393 /* just returns 0 */
4397 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4398 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4402 nid = __early_pfn_to_nid(pfn);
4403 if (nid >= 0 && nid != node)
4410 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4411 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4412 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4414 * If an architecture guarantees that all ranges registered with
4415 * add_active_ranges() contain no holes and may be freed, this
4416 * this function may be used instead of calling memblock_free_early_nid()
4419 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4421 unsigned long start_pfn, end_pfn;
4424 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4425 start_pfn = min(start_pfn, max_low_pfn);
4426 end_pfn = min(end_pfn, max_low_pfn);
4428 if (start_pfn < end_pfn)
4429 memblock_free_early_nid(PFN_PHYS(start_pfn),
4430 (end_pfn - start_pfn) << PAGE_SHIFT,
4436 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4437 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4439 * If an architecture guarantees that all ranges registered with
4440 * add_active_ranges() contain no holes and may be freed, this
4441 * function may be used instead of calling memory_present() manually.
4443 void __init sparse_memory_present_with_active_regions(int nid)
4445 unsigned long start_pfn, end_pfn;
4448 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4449 memory_present(this_nid, start_pfn, end_pfn);
4453 * get_pfn_range_for_nid - Return the start and end page frames for a node
4454 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4455 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4456 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4458 * It returns the start and end page frame of a node based on information
4459 * provided by an arch calling add_active_range(). If called for a node
4460 * with no available memory, a warning is printed and the start and end
4463 void __meminit get_pfn_range_for_nid(unsigned int nid,
4464 unsigned long *start_pfn, unsigned long *end_pfn)
4466 unsigned long this_start_pfn, this_end_pfn;
4472 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4473 *start_pfn = min(*start_pfn, this_start_pfn);
4474 *end_pfn = max(*end_pfn, this_end_pfn);
4477 if (*start_pfn == -1UL)
4482 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4483 * assumption is made that zones within a node are ordered in monotonic
4484 * increasing memory addresses so that the "highest" populated zone is used
4486 static void __init find_usable_zone_for_movable(void)
4489 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4490 if (zone_index == ZONE_MOVABLE)
4493 if (arch_zone_highest_possible_pfn[zone_index] >
4494 arch_zone_lowest_possible_pfn[zone_index])
4498 VM_BUG_ON(zone_index == -1);
4499 movable_zone = zone_index;
4503 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4504 * because it is sized independent of architecture. Unlike the other zones,
4505 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4506 * in each node depending on the size of each node and how evenly kernelcore
4507 * is distributed. This helper function adjusts the zone ranges
4508 * provided by the architecture for a given node by using the end of the
4509 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4510 * zones within a node are in order of monotonic increases memory addresses
4512 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4513 unsigned long zone_type,
4514 unsigned long node_start_pfn,
4515 unsigned long node_end_pfn,
4516 unsigned long *zone_start_pfn,
4517 unsigned long *zone_end_pfn)
4519 /* Only adjust if ZONE_MOVABLE is on this node */
4520 if (zone_movable_pfn[nid]) {
4521 /* Size ZONE_MOVABLE */
4522 if (zone_type == ZONE_MOVABLE) {
4523 *zone_start_pfn = zone_movable_pfn[nid];
4524 *zone_end_pfn = min(node_end_pfn,
4525 arch_zone_highest_possible_pfn[movable_zone]);
4527 /* Adjust for ZONE_MOVABLE starting within this range */
4528 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4529 *zone_end_pfn > zone_movable_pfn[nid]) {
4530 *zone_end_pfn = zone_movable_pfn[nid];
4532 /* Check if this whole range is within ZONE_MOVABLE */
4533 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4534 *zone_start_pfn = *zone_end_pfn;
4539 * Return the number of pages a zone spans in a node, including holes
4540 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4542 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4543 unsigned long zone_type,
4544 unsigned long node_start_pfn,
4545 unsigned long node_end_pfn,
4546 unsigned long *ignored)
4548 unsigned long zone_start_pfn, zone_end_pfn;
4550 /* Get the start and end of the zone */
4551 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4552 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4553 adjust_zone_range_for_zone_movable(nid, zone_type,
4554 node_start_pfn, node_end_pfn,
4555 &zone_start_pfn, &zone_end_pfn);
4557 /* Check that this node has pages within the zone's required range */
4558 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4561 /* Move the zone boundaries inside the node if necessary */
4562 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4563 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4565 /* Return the spanned pages */
4566 return zone_end_pfn - zone_start_pfn;
4570 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4571 * then all holes in the requested range will be accounted for.
4573 unsigned long __meminit __absent_pages_in_range(int nid,
4574 unsigned long range_start_pfn,
4575 unsigned long range_end_pfn)
4577 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4578 unsigned long start_pfn, end_pfn;
4581 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4582 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4583 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4584 nr_absent -= end_pfn - start_pfn;
4590 * absent_pages_in_range - Return number of page frames in holes within a range
4591 * @start_pfn: The start PFN to start searching for holes
4592 * @end_pfn: The end PFN to stop searching for holes
4594 * It returns the number of pages frames in memory holes within a range.
4596 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4597 unsigned long end_pfn)
4599 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4602 /* Return the number of page frames in holes in a zone on a node */
4603 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4604 unsigned long zone_type,
4605 unsigned long node_start_pfn,
4606 unsigned long node_end_pfn,
4607 unsigned long *ignored)
4609 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4610 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4611 unsigned long zone_start_pfn, zone_end_pfn;
4613 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4614 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4616 adjust_zone_range_for_zone_movable(nid, zone_type,
4617 node_start_pfn, node_end_pfn,
4618 &zone_start_pfn, &zone_end_pfn);
4619 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4622 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4623 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4624 unsigned long zone_type,
4625 unsigned long node_start_pfn,
4626 unsigned long node_end_pfn,
4627 unsigned long *zones_size)
4629 return zones_size[zone_type];
4632 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4633 unsigned long zone_type,
4634 unsigned long node_start_pfn,
4635 unsigned long node_end_pfn,
4636 unsigned long *zholes_size)
4641 return zholes_size[zone_type];
4644 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4646 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4647 unsigned long node_start_pfn,
4648 unsigned long node_end_pfn,
4649 unsigned long *zones_size,
4650 unsigned long *zholes_size)
4652 unsigned long realtotalpages, totalpages = 0;
4655 for (i = 0; i < MAX_NR_ZONES; i++)
4656 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4660 pgdat->node_spanned_pages = totalpages;
4662 realtotalpages = totalpages;
4663 for (i = 0; i < MAX_NR_ZONES; i++)
4665 zone_absent_pages_in_node(pgdat->node_id, i,
4666 node_start_pfn, node_end_pfn,
4668 pgdat->node_present_pages = realtotalpages;
4669 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4673 #ifndef CONFIG_SPARSEMEM
4675 * Calculate the size of the zone->blockflags rounded to an unsigned long
4676 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4677 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4678 * round what is now in bits to nearest long in bits, then return it in
4681 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4683 unsigned long usemapsize;
4685 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4686 usemapsize = roundup(zonesize, pageblock_nr_pages);
4687 usemapsize = usemapsize >> pageblock_order;
4688 usemapsize *= NR_PAGEBLOCK_BITS;
4689 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4691 return usemapsize / 8;
4694 static void __init setup_usemap(struct pglist_data *pgdat,
4696 unsigned long zone_start_pfn,
4697 unsigned long zonesize)
4699 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4700 zone->pageblock_flags = NULL;
4702 zone->pageblock_flags =
4703 memblock_virt_alloc_node_nopanic(usemapsize,
4707 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4708 unsigned long zone_start_pfn, unsigned long zonesize) {}
4709 #endif /* CONFIG_SPARSEMEM */
4711 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4713 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4714 void __paginginit set_pageblock_order(void)
4718 /* Check that pageblock_nr_pages has not already been setup */
4719 if (pageblock_order)
4722 if (HPAGE_SHIFT > PAGE_SHIFT)
4723 order = HUGETLB_PAGE_ORDER;
4725 order = MAX_ORDER - 1;
4728 * Assume the largest contiguous order of interest is a huge page.
4729 * This value may be variable depending on boot parameters on IA64 and
4732 pageblock_order = order;
4734 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4737 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4738 * is unused as pageblock_order is set at compile-time. See
4739 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4742 void __paginginit set_pageblock_order(void)
4746 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4748 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4749 unsigned long present_pages)
4751 unsigned long pages = spanned_pages;
4754 * Provide a more accurate estimation if there are holes within
4755 * the zone and SPARSEMEM is in use. If there are holes within the
4756 * zone, each populated memory region may cost us one or two extra
4757 * memmap pages due to alignment because memmap pages for each
4758 * populated regions may not naturally algined on page boundary.
4759 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4761 if (spanned_pages > present_pages + (present_pages >> 4) &&
4762 IS_ENABLED(CONFIG_SPARSEMEM))
4763 pages = present_pages;
4765 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4769 * Set up the zone data structures:
4770 * - mark all pages reserved
4771 * - mark all memory queues empty
4772 * - clear the memory bitmaps
4774 * NOTE: pgdat should get zeroed by caller.
4776 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4777 unsigned long node_start_pfn, unsigned long node_end_pfn,
4778 unsigned long *zones_size, unsigned long *zholes_size)
4781 int nid = pgdat->node_id;
4782 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4785 pgdat_resize_init(pgdat);
4786 #ifdef CONFIG_NUMA_BALANCING
4787 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4788 pgdat->numabalancing_migrate_nr_pages = 0;
4789 pgdat->numabalancing_migrate_next_window = jiffies;
4791 init_waitqueue_head(&pgdat->kswapd_wait);
4792 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4793 pgdat_page_cgroup_init(pgdat);
4795 for (j = 0; j < MAX_NR_ZONES; j++) {
4796 struct zone *zone = pgdat->node_zones + j;
4797 unsigned long size, realsize, freesize, memmap_pages;
4799 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4800 node_end_pfn, zones_size);
4801 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4807 * Adjust freesize so that it accounts for how much memory
4808 * is used by this zone for memmap. This affects the watermark
4809 * and per-cpu initialisations
4811 memmap_pages = calc_memmap_size(size, realsize);
4812 if (freesize >= memmap_pages) {
4813 freesize -= memmap_pages;
4816 " %s zone: %lu pages used for memmap\n",
4817 zone_names[j], memmap_pages);
4820 " %s zone: %lu pages exceeds freesize %lu\n",
4821 zone_names[j], memmap_pages, freesize);
4823 /* Account for reserved pages */
4824 if (j == 0 && freesize > dma_reserve) {
4825 freesize -= dma_reserve;
4826 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4827 zone_names[0], dma_reserve);
4830 if (!is_highmem_idx(j))
4831 nr_kernel_pages += freesize;
4832 /* Charge for highmem memmap if there are enough kernel pages */
4833 else if (nr_kernel_pages > memmap_pages * 2)
4834 nr_kernel_pages -= memmap_pages;
4835 nr_all_pages += freesize;
4837 zone->spanned_pages = size;
4838 zone->present_pages = realsize;
4840 * Set an approximate value for lowmem here, it will be adjusted
4841 * when the bootmem allocator frees pages into the buddy system.
4842 * And all highmem pages will be managed by the buddy system.
4844 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4847 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4849 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4851 zone->name = zone_names[j];
4852 spin_lock_init(&zone->lock);
4853 spin_lock_init(&zone->lru_lock);
4854 zone_seqlock_init(zone);
4855 zone->zone_pgdat = pgdat;
4856 zone_pcp_init(zone);
4858 /* For bootup, initialized properly in watermark setup */
4859 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4861 lruvec_init(&zone->lruvec);
4865 set_pageblock_order();
4866 setup_usemap(pgdat, zone, zone_start_pfn, size);
4867 ret = init_currently_empty_zone(zone, zone_start_pfn,
4868 size, MEMMAP_EARLY);
4870 memmap_init(size, nid, j, zone_start_pfn);
4871 zone_start_pfn += size;
4875 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4877 /* Skip empty nodes */
4878 if (!pgdat->node_spanned_pages)
4881 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4882 /* ia64 gets its own node_mem_map, before this, without bootmem */
4883 if (!pgdat->node_mem_map) {
4884 unsigned long size, start, end;
4888 * The zone's endpoints aren't required to be MAX_ORDER
4889 * aligned but the node_mem_map endpoints must be in order
4890 * for the buddy allocator to function correctly.
4892 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4893 end = pgdat_end_pfn(pgdat);
4894 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4895 size = (end - start) * sizeof(struct page);
4896 map = alloc_remap(pgdat->node_id, size);
4898 map = memblock_virt_alloc_node_nopanic(size,
4900 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4902 #ifndef CONFIG_NEED_MULTIPLE_NODES
4904 * With no DISCONTIG, the global mem_map is just set as node 0's
4906 if (pgdat == NODE_DATA(0)) {
4907 mem_map = NODE_DATA(0)->node_mem_map;
4908 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4909 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4910 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4911 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4914 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4917 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4918 unsigned long node_start_pfn, unsigned long *zholes_size)
4920 pg_data_t *pgdat = NODE_DATA(nid);
4921 unsigned long start_pfn = 0;
4922 unsigned long end_pfn = 0;
4924 /* pg_data_t should be reset to zero when it's allocated */
4925 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4927 pgdat->node_id = nid;
4928 pgdat->node_start_pfn = node_start_pfn;
4929 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4930 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4932 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4933 zones_size, zholes_size);
4935 alloc_node_mem_map(pgdat);
4936 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4937 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4938 nid, (unsigned long)pgdat,
4939 (unsigned long)pgdat->node_mem_map);
4942 free_area_init_core(pgdat, start_pfn, end_pfn,
4943 zones_size, zholes_size);
4946 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4948 #if MAX_NUMNODES > 1
4950 * Figure out the number of possible node ids.
4952 void __init setup_nr_node_ids(void)
4955 unsigned int highest = 0;
4957 for_each_node_mask(node, node_possible_map)
4959 nr_node_ids = highest + 1;
4964 * node_map_pfn_alignment - determine the maximum internode alignment
4966 * This function should be called after node map is populated and sorted.
4967 * It calculates the maximum power of two alignment which can distinguish
4970 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4971 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4972 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4973 * shifted, 1GiB is enough and this function will indicate so.
4975 * This is used to test whether pfn -> nid mapping of the chosen memory
4976 * model has fine enough granularity to avoid incorrect mapping for the
4977 * populated node map.
4979 * Returns the determined alignment in pfn's. 0 if there is no alignment
4980 * requirement (single node).
4982 unsigned long __init node_map_pfn_alignment(void)
4984 unsigned long accl_mask = 0, last_end = 0;
4985 unsigned long start, end, mask;
4989 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4990 if (!start || last_nid < 0 || last_nid == nid) {
4997 * Start with a mask granular enough to pin-point to the
4998 * start pfn and tick off bits one-by-one until it becomes
4999 * too coarse to separate the current node from the last.
5001 mask = ~((1 << __ffs(start)) - 1);
5002 while (mask && last_end <= (start & (mask << 1)))
5005 /* accumulate all internode masks */
5009 /* convert mask to number of pages */
5010 return ~accl_mask + 1;
5013 /* Find the lowest pfn for a node */
5014 static unsigned long __init find_min_pfn_for_node(int nid)
5016 unsigned long min_pfn = ULONG_MAX;
5017 unsigned long start_pfn;
5020 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5021 min_pfn = min(min_pfn, start_pfn);
5023 if (min_pfn == ULONG_MAX) {
5025 "Could not find start_pfn for node %d\n", nid);
5033 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5035 * It returns the minimum PFN based on information provided via
5036 * add_active_range().
5038 unsigned long __init find_min_pfn_with_active_regions(void)
5040 return find_min_pfn_for_node(MAX_NUMNODES);
5044 * early_calculate_totalpages()
5045 * Sum pages in active regions for movable zone.
5046 * Populate N_MEMORY for calculating usable_nodes.
5048 static unsigned long __init early_calculate_totalpages(void)
5050 unsigned long totalpages = 0;
5051 unsigned long start_pfn, end_pfn;
5054 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5055 unsigned long pages = end_pfn - start_pfn;
5057 totalpages += pages;
5059 node_set_state(nid, N_MEMORY);
5065 * Find the PFN the Movable zone begins in each node. Kernel memory
5066 * is spread evenly between nodes as long as the nodes have enough
5067 * memory. When they don't, some nodes will have more kernelcore than
5070 static void __init find_zone_movable_pfns_for_nodes(void)
5073 unsigned long usable_startpfn;
5074 unsigned long kernelcore_node, kernelcore_remaining;
5075 /* save the state before borrow the nodemask */
5076 nodemask_t saved_node_state = node_states[N_MEMORY];
5077 unsigned long totalpages = early_calculate_totalpages();
5078 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5079 struct memblock_region *r;
5081 /* Need to find movable_zone earlier when movable_node is specified. */
5082 find_usable_zone_for_movable();
5085 * If movable_node is specified, ignore kernelcore and movablecore
5088 if (movable_node_is_enabled()) {
5089 for_each_memblock(memory, r) {
5090 if (!memblock_is_hotpluggable(r))
5095 usable_startpfn = PFN_DOWN(r->base);
5096 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5097 min(usable_startpfn, zone_movable_pfn[nid]) :
5105 * If movablecore=nn[KMG] was specified, calculate what size of
5106 * kernelcore that corresponds so that memory usable for
5107 * any allocation type is evenly spread. If both kernelcore
5108 * and movablecore are specified, then the value of kernelcore
5109 * will be used for required_kernelcore if it's greater than
5110 * what movablecore would have allowed.
5112 if (required_movablecore) {
5113 unsigned long corepages;
5116 * Round-up so that ZONE_MOVABLE is at least as large as what
5117 * was requested by the user
5119 required_movablecore =
5120 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5121 corepages = totalpages - required_movablecore;
5123 required_kernelcore = max(required_kernelcore, corepages);
5126 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5127 if (!required_kernelcore)
5130 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5131 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5134 /* Spread kernelcore memory as evenly as possible throughout nodes */
5135 kernelcore_node = required_kernelcore / usable_nodes;
5136 for_each_node_state(nid, N_MEMORY) {
5137 unsigned long start_pfn, end_pfn;
5140 * Recalculate kernelcore_node if the division per node
5141 * now exceeds what is necessary to satisfy the requested
5142 * amount of memory for the kernel
5144 if (required_kernelcore < kernelcore_node)
5145 kernelcore_node = required_kernelcore / usable_nodes;
5148 * As the map is walked, we track how much memory is usable
5149 * by the kernel using kernelcore_remaining. When it is
5150 * 0, the rest of the node is usable by ZONE_MOVABLE
5152 kernelcore_remaining = kernelcore_node;
5154 /* Go through each range of PFNs within this node */
5155 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5156 unsigned long size_pages;
5158 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5159 if (start_pfn >= end_pfn)
5162 /* Account for what is only usable for kernelcore */
5163 if (start_pfn < usable_startpfn) {
5164 unsigned long kernel_pages;
5165 kernel_pages = min(end_pfn, usable_startpfn)
5168 kernelcore_remaining -= min(kernel_pages,
5169 kernelcore_remaining);
5170 required_kernelcore -= min(kernel_pages,
5171 required_kernelcore);
5173 /* Continue if range is now fully accounted */
5174 if (end_pfn <= usable_startpfn) {
5177 * Push zone_movable_pfn to the end so
5178 * that if we have to rebalance
5179 * kernelcore across nodes, we will
5180 * not double account here
5182 zone_movable_pfn[nid] = end_pfn;
5185 start_pfn = usable_startpfn;
5189 * The usable PFN range for ZONE_MOVABLE is from
5190 * start_pfn->end_pfn. Calculate size_pages as the
5191 * number of pages used as kernelcore
5193 size_pages = end_pfn - start_pfn;
5194 if (size_pages > kernelcore_remaining)
5195 size_pages = kernelcore_remaining;
5196 zone_movable_pfn[nid] = start_pfn + size_pages;
5199 * Some kernelcore has been met, update counts and
5200 * break if the kernelcore for this node has been
5203 required_kernelcore -= min(required_kernelcore,
5205 kernelcore_remaining -= size_pages;
5206 if (!kernelcore_remaining)
5212 * If there is still required_kernelcore, we do another pass with one
5213 * less node in the count. This will push zone_movable_pfn[nid] further
5214 * along on the nodes that still have memory until kernelcore is
5218 if (usable_nodes && required_kernelcore > usable_nodes)
5222 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5223 for (nid = 0; nid < MAX_NUMNODES; nid++)
5224 zone_movable_pfn[nid] =
5225 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5228 /* restore the node_state */
5229 node_states[N_MEMORY] = saved_node_state;
5232 /* Any regular or high memory on that node ? */
5233 static void check_for_memory(pg_data_t *pgdat, int nid)
5235 enum zone_type zone_type;
5237 if (N_MEMORY == N_NORMAL_MEMORY)
5240 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5241 struct zone *zone = &pgdat->node_zones[zone_type];
5242 if (populated_zone(zone)) {
5243 node_set_state(nid, N_HIGH_MEMORY);
5244 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5245 zone_type <= ZONE_NORMAL)
5246 node_set_state(nid, N_NORMAL_MEMORY);
5253 * free_area_init_nodes - Initialise all pg_data_t and zone data
5254 * @max_zone_pfn: an array of max PFNs for each zone
5256 * This will call free_area_init_node() for each active node in the system.
5257 * Using the page ranges provided by add_active_range(), the size of each
5258 * zone in each node and their holes is calculated. If the maximum PFN
5259 * between two adjacent zones match, it is assumed that the zone is empty.
5260 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5261 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5262 * starts where the previous one ended. For example, ZONE_DMA32 starts
5263 * at arch_max_dma_pfn.
5265 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5267 unsigned long start_pfn, end_pfn;
5270 /* Record where the zone boundaries are */
5271 memset(arch_zone_lowest_possible_pfn, 0,
5272 sizeof(arch_zone_lowest_possible_pfn));
5273 memset(arch_zone_highest_possible_pfn, 0,
5274 sizeof(arch_zone_highest_possible_pfn));
5275 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5276 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5277 for (i = 1; i < MAX_NR_ZONES; i++) {
5278 if (i == ZONE_MOVABLE)
5280 arch_zone_lowest_possible_pfn[i] =
5281 arch_zone_highest_possible_pfn[i-1];
5282 arch_zone_highest_possible_pfn[i] =
5283 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5285 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5286 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5288 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5289 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5290 find_zone_movable_pfns_for_nodes();
5292 /* Print out the zone ranges */
5293 printk("Zone ranges:\n");
5294 for (i = 0; i < MAX_NR_ZONES; i++) {
5295 if (i == ZONE_MOVABLE)
5297 printk(KERN_CONT " %-8s ", zone_names[i]);
5298 if (arch_zone_lowest_possible_pfn[i] ==
5299 arch_zone_highest_possible_pfn[i])
5300 printk(KERN_CONT "empty\n");
5302 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5303 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5304 (arch_zone_highest_possible_pfn[i]
5305 << PAGE_SHIFT) - 1);
5308 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5309 printk("Movable zone start for each node\n");
5310 for (i = 0; i < MAX_NUMNODES; i++) {
5311 if (zone_movable_pfn[i])
5312 printk(" Node %d: %#010lx\n", i,
5313 zone_movable_pfn[i] << PAGE_SHIFT);
5316 /* Print out the early node map */
5317 printk("Early memory node ranges\n");
5318 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5319 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5320 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5322 /* Initialise every node */
5323 mminit_verify_pageflags_layout();
5324 setup_nr_node_ids();
5325 for_each_online_node(nid) {
5326 pg_data_t *pgdat = NODE_DATA(nid);
5327 free_area_init_node(nid, NULL,
5328 find_min_pfn_for_node(nid), NULL);
5330 /* Any memory on that node */
5331 if (pgdat->node_present_pages)
5332 node_set_state(nid, N_MEMORY);
5333 check_for_memory(pgdat, nid);
5337 static int __init cmdline_parse_core(char *p, unsigned long *core)
5339 unsigned long long coremem;
5343 coremem = memparse(p, &p);
5344 *core = coremem >> PAGE_SHIFT;
5346 /* Paranoid check that UL is enough for the coremem value */
5347 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5353 * kernelcore=size sets the amount of memory for use for allocations that
5354 * cannot be reclaimed or migrated.
5356 static int __init cmdline_parse_kernelcore(char *p)
5358 return cmdline_parse_core(p, &required_kernelcore);
5362 * movablecore=size sets the amount of memory for use for allocations that
5363 * can be reclaimed or migrated.
5365 static int __init cmdline_parse_movablecore(char *p)
5367 return cmdline_parse_core(p, &required_movablecore);
5370 early_param("kernelcore", cmdline_parse_kernelcore);
5371 early_param("movablecore", cmdline_parse_movablecore);
5373 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5375 void adjust_managed_page_count(struct page *page, long count)
5377 spin_lock(&managed_page_count_lock);
5378 page_zone(page)->managed_pages += count;
5379 totalram_pages += count;
5380 #ifdef CONFIG_HIGHMEM
5381 if (PageHighMem(page))
5382 totalhigh_pages += count;
5384 spin_unlock(&managed_page_count_lock);
5386 EXPORT_SYMBOL(adjust_managed_page_count);
5388 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5391 unsigned long pages = 0;
5393 start = (void *)PAGE_ALIGN((unsigned long)start);
5394 end = (void *)((unsigned long)end & PAGE_MASK);
5395 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5396 if ((unsigned int)poison <= 0xFF)
5397 memset(pos, poison, PAGE_SIZE);
5398 free_reserved_page(virt_to_page(pos));
5402 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5403 s, pages << (PAGE_SHIFT - 10), start, end);
5407 EXPORT_SYMBOL(free_reserved_area);
5409 #ifdef CONFIG_HIGHMEM
5410 void free_highmem_page(struct page *page)
5412 __free_reserved_page(page);
5414 page_zone(page)->managed_pages++;
5420 void __init mem_init_print_info(const char *str)
5422 unsigned long physpages, codesize, datasize, rosize, bss_size;
5423 unsigned long init_code_size, init_data_size;
5425 physpages = get_num_physpages();
5426 codesize = _etext - _stext;
5427 datasize = _edata - _sdata;
5428 rosize = __end_rodata - __start_rodata;
5429 bss_size = __bss_stop - __bss_start;
5430 init_data_size = __init_end - __init_begin;
5431 init_code_size = _einittext - _sinittext;
5434 * Detect special cases and adjust section sizes accordingly:
5435 * 1) .init.* may be embedded into .data sections
5436 * 2) .init.text.* may be out of [__init_begin, __init_end],
5437 * please refer to arch/tile/kernel/vmlinux.lds.S.
5438 * 3) .rodata.* may be embedded into .text or .data sections.
5440 #define adj_init_size(start, end, size, pos, adj) \
5442 if (start <= pos && pos < end && size > adj) \
5446 adj_init_size(__init_begin, __init_end, init_data_size,
5447 _sinittext, init_code_size);
5448 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5449 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5450 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5451 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5453 #undef adj_init_size
5455 printk("Memory: %luK/%luK available "
5456 "(%luK kernel code, %luK rwdata, %luK rodata, "
5457 "%luK init, %luK bss, %luK reserved"
5458 #ifdef CONFIG_HIGHMEM
5462 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5463 codesize >> 10, datasize >> 10, rosize >> 10,
5464 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5465 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5466 #ifdef CONFIG_HIGHMEM
5467 totalhigh_pages << (PAGE_SHIFT-10),
5469 str ? ", " : "", str ? str : "");
5473 * set_dma_reserve - set the specified number of pages reserved in the first zone
5474 * @new_dma_reserve: The number of pages to mark reserved
5476 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5477 * In the DMA zone, a significant percentage may be consumed by kernel image
5478 * and other unfreeable allocations which can skew the watermarks badly. This
5479 * function may optionally be used to account for unfreeable pages in the
5480 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5481 * smaller per-cpu batchsize.
5483 void __init set_dma_reserve(unsigned long new_dma_reserve)
5485 dma_reserve = new_dma_reserve;
5488 void __init free_area_init(unsigned long *zones_size)
5490 free_area_init_node(0, zones_size,
5491 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5494 static int page_alloc_cpu_notify(struct notifier_block *self,
5495 unsigned long action, void *hcpu)
5497 int cpu = (unsigned long)hcpu;
5499 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5500 lru_add_drain_cpu(cpu);
5504 * Spill the event counters of the dead processor
5505 * into the current processors event counters.
5506 * This artificially elevates the count of the current
5509 vm_events_fold_cpu(cpu);
5512 * Zero the differential counters of the dead processor
5513 * so that the vm statistics are consistent.
5515 * This is only okay since the processor is dead and cannot
5516 * race with what we are doing.
5518 cpu_vm_stats_fold(cpu);
5523 void __init page_alloc_init(void)
5525 hotcpu_notifier(page_alloc_cpu_notify, 0);
5529 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5530 * or min_free_kbytes changes.
5532 static void calculate_totalreserve_pages(void)
5534 struct pglist_data *pgdat;
5535 unsigned long reserve_pages = 0;
5536 enum zone_type i, j;
5538 for_each_online_pgdat(pgdat) {
5539 for (i = 0; i < MAX_NR_ZONES; i++) {
5540 struct zone *zone = pgdat->node_zones + i;
5541 unsigned long max = 0;
5543 /* Find valid and maximum lowmem_reserve in the zone */
5544 for (j = i; j < MAX_NR_ZONES; j++) {
5545 if (zone->lowmem_reserve[j] > max)
5546 max = zone->lowmem_reserve[j];
5549 /* we treat the high watermark as reserved pages. */
5550 max += high_wmark_pages(zone);
5552 if (max > zone->managed_pages)
5553 max = zone->managed_pages;
5554 reserve_pages += max;
5556 * Lowmem reserves are not available to
5557 * GFP_HIGHUSER page cache allocations and
5558 * kswapd tries to balance zones to their high
5559 * watermark. As a result, neither should be
5560 * regarded as dirtyable memory, to prevent a
5561 * situation where reclaim has to clean pages
5562 * in order to balance the zones.
5564 zone->dirty_balance_reserve = max;
5567 dirty_balance_reserve = reserve_pages;
5568 totalreserve_pages = reserve_pages;
5572 * setup_per_zone_lowmem_reserve - called whenever
5573 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5574 * has a correct pages reserved value, so an adequate number of
5575 * pages are left in the zone after a successful __alloc_pages().
5577 static void setup_per_zone_lowmem_reserve(void)
5579 struct pglist_data *pgdat;
5580 enum zone_type j, idx;
5582 for_each_online_pgdat(pgdat) {
5583 for (j = 0; j < MAX_NR_ZONES; j++) {
5584 struct zone *zone = pgdat->node_zones + j;
5585 unsigned long managed_pages = zone->managed_pages;
5587 zone->lowmem_reserve[j] = 0;
5591 struct zone *lower_zone;
5595 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5596 sysctl_lowmem_reserve_ratio[idx] = 1;
5598 lower_zone = pgdat->node_zones + idx;
5599 lower_zone->lowmem_reserve[j] = managed_pages /
5600 sysctl_lowmem_reserve_ratio[idx];
5601 managed_pages += lower_zone->managed_pages;
5606 /* update totalreserve_pages */
5607 calculate_totalreserve_pages();
5610 static void __setup_per_zone_wmarks(void)
5612 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5613 unsigned long lowmem_pages = 0;
5615 unsigned long flags;
5617 /* Calculate total number of !ZONE_HIGHMEM pages */
5618 for_each_zone(zone) {
5619 if (!is_highmem(zone))
5620 lowmem_pages += zone->managed_pages;
5623 for_each_zone(zone) {
5626 spin_lock_irqsave(&zone->lock, flags);
5627 tmp = (u64)pages_min * zone->managed_pages;
5628 do_div(tmp, lowmem_pages);
5629 if (is_highmem(zone)) {
5631 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5632 * need highmem pages, so cap pages_min to a small
5635 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5636 * deltas controls asynch page reclaim, and so should
5637 * not be capped for highmem.
5639 unsigned long min_pages;
5641 min_pages = zone->managed_pages / 1024;
5642 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5643 zone->watermark[WMARK_MIN] = min_pages;
5646 * If it's a lowmem zone, reserve a number of pages
5647 * proportionate to the zone's size.
5649 zone->watermark[WMARK_MIN] = tmp;
5652 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5653 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5655 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5656 high_wmark_pages(zone) -
5657 low_wmark_pages(zone) -
5658 zone_page_state(zone, NR_ALLOC_BATCH));
5660 setup_zone_migrate_reserve(zone);
5661 spin_unlock_irqrestore(&zone->lock, flags);
5664 /* update totalreserve_pages */
5665 calculate_totalreserve_pages();
5669 * setup_per_zone_wmarks - called when min_free_kbytes changes
5670 * or when memory is hot-{added|removed}
5672 * Ensures that the watermark[min,low,high] values for each zone are set
5673 * correctly with respect to min_free_kbytes.
5675 void setup_per_zone_wmarks(void)
5677 mutex_lock(&zonelists_mutex);
5678 __setup_per_zone_wmarks();
5679 mutex_unlock(&zonelists_mutex);
5683 * The inactive anon list should be small enough that the VM never has to
5684 * do too much work, but large enough that each inactive page has a chance
5685 * to be referenced again before it is swapped out.
5687 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5688 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5689 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5690 * the anonymous pages are kept on the inactive list.
5693 * memory ratio inactive anon
5694 * -------------------------------------
5703 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5705 unsigned int gb, ratio;
5707 /* Zone size in gigabytes */
5708 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5710 ratio = int_sqrt(10 * gb);
5714 zone->inactive_ratio = ratio;
5717 static void __meminit setup_per_zone_inactive_ratio(void)
5722 calculate_zone_inactive_ratio(zone);
5726 * Initialise min_free_kbytes.
5728 * For small machines we want it small (128k min). For large machines
5729 * we want it large (64MB max). But it is not linear, because network
5730 * bandwidth does not increase linearly with machine size. We use
5732 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5733 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5749 int __meminit init_per_zone_wmark_min(void)
5751 unsigned long lowmem_kbytes;
5752 int new_min_free_kbytes;
5754 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5755 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5757 if (new_min_free_kbytes > user_min_free_kbytes) {
5758 min_free_kbytes = new_min_free_kbytes;
5759 if (min_free_kbytes < 128)
5760 min_free_kbytes = 128;
5761 if (min_free_kbytes > 65536)
5762 min_free_kbytes = 65536;
5764 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5765 new_min_free_kbytes, user_min_free_kbytes);
5767 setup_per_zone_wmarks();
5768 refresh_zone_stat_thresholds();
5769 setup_per_zone_lowmem_reserve();
5770 setup_per_zone_inactive_ratio();
5773 module_init(init_per_zone_wmark_min)
5776 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5777 * that we can call two helper functions whenever min_free_kbytes
5780 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5781 void __user *buffer, size_t *length, loff_t *ppos)
5785 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5790 user_min_free_kbytes = min_free_kbytes;
5791 setup_per_zone_wmarks();
5797 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5798 void __user *buffer, size_t *length, loff_t *ppos)
5803 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5808 zone->min_unmapped_pages = (zone->managed_pages *
5809 sysctl_min_unmapped_ratio) / 100;
5813 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5814 void __user *buffer, size_t *length, loff_t *ppos)
5819 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5824 zone->min_slab_pages = (zone->managed_pages *
5825 sysctl_min_slab_ratio) / 100;
5831 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5832 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5833 * whenever sysctl_lowmem_reserve_ratio changes.
5835 * The reserve ratio obviously has absolutely no relation with the
5836 * minimum watermarks. The lowmem reserve ratio can only make sense
5837 * if in function of the boot time zone sizes.
5839 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5840 void __user *buffer, size_t *length, loff_t *ppos)
5842 proc_dointvec_minmax(table, write, buffer, length, ppos);
5843 setup_per_zone_lowmem_reserve();
5848 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5849 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5850 * pagelist can have before it gets flushed back to buddy allocator.
5852 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5853 void __user *buffer, size_t *length, loff_t *ppos)
5859 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5860 if (!write || (ret < 0))
5863 mutex_lock(&pcp_batch_high_lock);
5864 for_each_populated_zone(zone) {
5866 high = zone->managed_pages / percpu_pagelist_fraction;
5867 for_each_possible_cpu(cpu)
5868 pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5871 mutex_unlock(&pcp_batch_high_lock);
5875 int hashdist = HASHDIST_DEFAULT;
5878 static int __init set_hashdist(char *str)
5882 hashdist = simple_strtoul(str, &str, 0);
5885 __setup("hashdist=", set_hashdist);
5889 * allocate a large system hash table from bootmem
5890 * - it is assumed that the hash table must contain an exact power-of-2
5891 * quantity of entries
5892 * - limit is the number of hash buckets, not the total allocation size
5894 void *__init alloc_large_system_hash(const char *tablename,
5895 unsigned long bucketsize,
5896 unsigned long numentries,
5899 unsigned int *_hash_shift,
5900 unsigned int *_hash_mask,
5901 unsigned long low_limit,
5902 unsigned long high_limit)
5904 unsigned long long max = high_limit;
5905 unsigned long log2qty, size;
5908 /* allow the kernel cmdline to have a say */
5910 /* round applicable memory size up to nearest megabyte */
5911 numentries = nr_kernel_pages;
5913 /* It isn't necessary when PAGE_SIZE >= 1MB */
5914 if (PAGE_SHIFT < 20)
5915 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5917 /* limit to 1 bucket per 2^scale bytes of low memory */
5918 if (scale > PAGE_SHIFT)
5919 numentries >>= (scale - PAGE_SHIFT);
5921 numentries <<= (PAGE_SHIFT - scale);
5923 /* Make sure we've got at least a 0-order allocation.. */
5924 if (unlikely(flags & HASH_SMALL)) {
5925 /* Makes no sense without HASH_EARLY */
5926 WARN_ON(!(flags & HASH_EARLY));
5927 if (!(numentries >> *_hash_shift)) {
5928 numentries = 1UL << *_hash_shift;
5929 BUG_ON(!numentries);
5931 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5932 numentries = PAGE_SIZE / bucketsize;
5934 numentries = roundup_pow_of_two(numentries);
5936 /* limit allocation size to 1/16 total memory by default */
5938 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5939 do_div(max, bucketsize);
5941 max = min(max, 0x80000000ULL);
5943 if (numentries < low_limit)
5944 numentries = low_limit;
5945 if (numentries > max)
5948 log2qty = ilog2(numentries);
5951 size = bucketsize << log2qty;
5952 if (flags & HASH_EARLY)
5953 table = memblock_virt_alloc_nopanic(size, 0);
5955 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5958 * If bucketsize is not a power-of-two, we may free
5959 * some pages at the end of hash table which
5960 * alloc_pages_exact() automatically does
5962 if (get_order(size) < MAX_ORDER) {
5963 table = alloc_pages_exact(size, GFP_ATOMIC);
5964 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5967 } while (!table && size > PAGE_SIZE && --log2qty);
5970 panic("Failed to allocate %s hash table\n", tablename);
5972 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5975 ilog2(size) - PAGE_SHIFT,
5979 *_hash_shift = log2qty;
5981 *_hash_mask = (1 << log2qty) - 1;
5986 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5987 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5990 #ifdef CONFIG_SPARSEMEM
5991 return __pfn_to_section(pfn)->pageblock_flags;
5993 return zone->pageblock_flags;
5994 #endif /* CONFIG_SPARSEMEM */
5997 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5999 #ifdef CONFIG_SPARSEMEM
6000 pfn &= (PAGES_PER_SECTION-1);
6001 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6003 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6004 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6005 #endif /* CONFIG_SPARSEMEM */
6009 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
6010 * @page: The page within the block of interest
6011 * @start_bitidx: The first bit of interest to retrieve
6012 * @end_bitidx: The last bit of interest
6013 * returns pageblock_bits flags
6015 unsigned long get_pageblock_flags_group(struct page *page,
6016 int start_bitidx, int end_bitidx)
6019 unsigned long *bitmap;
6020 unsigned long pfn, bitidx;
6021 unsigned long flags = 0;
6022 unsigned long value = 1;
6024 zone = page_zone(page);
6025 pfn = page_to_pfn(page);
6026 bitmap = get_pageblock_bitmap(zone, pfn);
6027 bitidx = pfn_to_bitidx(zone, pfn);
6029 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6030 if (test_bit(bitidx + start_bitidx, bitmap))
6037 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
6038 * @page: The page within the block of interest
6039 * @start_bitidx: The first bit of interest
6040 * @end_bitidx: The last bit of interest
6041 * @flags: The flags to set
6043 void set_pageblock_flags_group(struct page *page, unsigned long flags,
6044 int start_bitidx, int end_bitidx)
6047 unsigned long *bitmap;
6048 unsigned long pfn, bitidx;
6049 unsigned long value = 1;
6051 zone = page_zone(page);
6052 pfn = page_to_pfn(page);
6053 bitmap = get_pageblock_bitmap(zone, pfn);
6054 bitidx = pfn_to_bitidx(zone, pfn);
6055 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6057 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6059 __set_bit(bitidx + start_bitidx, bitmap);
6061 __clear_bit(bitidx + start_bitidx, bitmap);
6065 * This function checks whether pageblock includes unmovable pages or not.
6066 * If @count is not zero, it is okay to include less @count unmovable pages
6068 * PageLRU check without isolation or lru_lock could race so that
6069 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6070 * expect this function should be exact.
6072 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6073 bool skip_hwpoisoned_pages)
6075 unsigned long pfn, iter, found;
6079 * For avoiding noise data, lru_add_drain_all() should be called
6080 * If ZONE_MOVABLE, the zone never contains unmovable pages
6082 if (zone_idx(zone) == ZONE_MOVABLE)
6084 mt = get_pageblock_migratetype(page);
6085 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6088 pfn = page_to_pfn(page);
6089 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6090 unsigned long check = pfn + iter;
6092 if (!pfn_valid_within(check))
6095 page = pfn_to_page(check);
6098 * Hugepages are not in LRU lists, but they're movable.
6099 * We need not scan over tail pages bacause we don't
6100 * handle each tail page individually in migration.
6102 if (PageHuge(page)) {
6103 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6108 * We can't use page_count without pin a page
6109 * because another CPU can free compound page.
6110 * This check already skips compound tails of THP
6111 * because their page->_count is zero at all time.
6113 if (!atomic_read(&page->_count)) {
6114 if (PageBuddy(page))
6115 iter += (1 << page_order(page)) - 1;
6120 * The HWPoisoned page may be not in buddy system, and
6121 * page_count() is not 0.
6123 if (skip_hwpoisoned_pages && PageHWPoison(page))
6129 * If there are RECLAIMABLE pages, we need to check it.
6130 * But now, memory offline itself doesn't call shrink_slab()
6131 * and it still to be fixed.
6134 * If the page is not RAM, page_count()should be 0.
6135 * we don't need more check. This is an _used_ not-movable page.
6137 * The problematic thing here is PG_reserved pages. PG_reserved
6138 * is set to both of a memory hole page and a _used_ kernel
6147 bool is_pageblock_removable_nolock(struct page *page)
6153 * We have to be careful here because we are iterating over memory
6154 * sections which are not zone aware so we might end up outside of
6155 * the zone but still within the section.
6156 * We have to take care about the node as well. If the node is offline
6157 * its NODE_DATA will be NULL - see page_zone.
6159 if (!node_online(page_to_nid(page)))
6162 zone = page_zone(page);
6163 pfn = page_to_pfn(page);
6164 if (!zone_spans_pfn(zone, pfn))
6167 return !has_unmovable_pages(zone, page, 0, true);
6172 static unsigned long pfn_max_align_down(unsigned long pfn)
6174 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6175 pageblock_nr_pages) - 1);
6178 static unsigned long pfn_max_align_up(unsigned long pfn)
6180 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6181 pageblock_nr_pages));
6184 /* [start, end) must belong to a single zone. */
6185 static int __alloc_contig_migrate_range(struct compact_control *cc,
6186 unsigned long start, unsigned long end)
6188 /* This function is based on compact_zone() from compaction.c. */
6189 unsigned long nr_reclaimed;
6190 unsigned long pfn = start;
6191 unsigned int tries = 0;
6196 while (pfn < end || !list_empty(&cc->migratepages)) {
6197 if (fatal_signal_pending(current)) {
6202 if (list_empty(&cc->migratepages)) {
6203 cc->nr_migratepages = 0;
6204 pfn = isolate_migratepages_range(cc->zone, cc,
6211 } else if (++tries == 5) {
6212 ret = ret < 0 ? ret : -EBUSY;
6216 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6218 cc->nr_migratepages -= nr_reclaimed;
6220 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6221 0, MIGRATE_SYNC, MR_CMA);
6224 putback_movable_pages(&cc->migratepages);
6231 * alloc_contig_range() -- tries to allocate given range of pages
6232 * @start: start PFN to allocate
6233 * @end: one-past-the-last PFN to allocate
6234 * @migratetype: migratetype of the underlaying pageblocks (either
6235 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6236 * in range must have the same migratetype and it must
6237 * be either of the two.
6239 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6240 * aligned, however it's the caller's responsibility to guarantee that
6241 * we are the only thread that changes migrate type of pageblocks the
6244 * The PFN range must belong to a single zone.
6246 * Returns zero on success or negative error code. On success all
6247 * pages which PFN is in [start, end) are allocated for the caller and
6248 * need to be freed with free_contig_range().
6250 int alloc_contig_range(unsigned long start, unsigned long end,
6251 unsigned migratetype)
6253 unsigned long outer_start, outer_end;
6256 struct compact_control cc = {
6257 .nr_migratepages = 0,
6259 .zone = page_zone(pfn_to_page(start)),
6261 .ignore_skip_hint = true,
6263 INIT_LIST_HEAD(&cc.migratepages);
6266 * What we do here is we mark all pageblocks in range as
6267 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6268 * have different sizes, and due to the way page allocator
6269 * work, we align the range to biggest of the two pages so
6270 * that page allocator won't try to merge buddies from
6271 * different pageblocks and change MIGRATE_ISOLATE to some
6272 * other migration type.
6274 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6275 * migrate the pages from an unaligned range (ie. pages that
6276 * we are interested in). This will put all the pages in
6277 * range back to page allocator as MIGRATE_ISOLATE.
6279 * When this is done, we take the pages in range from page
6280 * allocator removing them from the buddy system. This way
6281 * page allocator will never consider using them.
6283 * This lets us mark the pageblocks back as
6284 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6285 * aligned range but not in the unaligned, original range are
6286 * put back to page allocator so that buddy can use them.
6289 ret = start_isolate_page_range(pfn_max_align_down(start),
6290 pfn_max_align_up(end), migratetype,
6295 ret = __alloc_contig_migrate_range(&cc, start, end);
6300 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6301 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6302 * more, all pages in [start, end) are free in page allocator.
6303 * What we are going to do is to allocate all pages from
6304 * [start, end) (that is remove them from page allocator).
6306 * The only problem is that pages at the beginning and at the
6307 * end of interesting range may be not aligned with pages that
6308 * page allocator holds, ie. they can be part of higher order
6309 * pages. Because of this, we reserve the bigger range and
6310 * once this is done free the pages we are not interested in.
6312 * We don't have to hold zone->lock here because the pages are
6313 * isolated thus they won't get removed from buddy.
6316 lru_add_drain_all();
6320 outer_start = start;
6321 while (!PageBuddy(pfn_to_page(outer_start))) {
6322 if (++order >= MAX_ORDER) {
6326 outer_start &= ~0UL << order;
6329 /* Make sure the range is really isolated. */
6330 if (test_pages_isolated(outer_start, end, false)) {
6331 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6338 /* Grab isolated pages from freelists. */
6339 outer_end = isolate_freepages_range(&cc, outer_start, end);
6345 /* Free head and tail (if any) */
6346 if (start != outer_start)
6347 free_contig_range(outer_start, start - outer_start);
6348 if (end != outer_end)
6349 free_contig_range(end, outer_end - end);
6352 undo_isolate_page_range(pfn_max_align_down(start),
6353 pfn_max_align_up(end), migratetype);
6357 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6359 unsigned int count = 0;
6361 for (; nr_pages--; pfn++) {
6362 struct page *page = pfn_to_page(pfn);
6364 count += page_count(page) != 1;
6367 WARN(count != 0, "%d pages are still in use!\n", count);
6371 #ifdef CONFIG_MEMORY_HOTPLUG
6373 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6374 * page high values need to be recalulated.
6376 void __meminit zone_pcp_update(struct zone *zone)
6379 mutex_lock(&pcp_batch_high_lock);
6380 for_each_possible_cpu(cpu)
6381 pageset_set_high_and_batch(zone,
6382 per_cpu_ptr(zone->pageset, cpu));
6383 mutex_unlock(&pcp_batch_high_lock);
6387 void zone_pcp_reset(struct zone *zone)
6389 unsigned long flags;
6391 struct per_cpu_pageset *pset;
6393 /* avoid races with drain_pages() */
6394 local_irq_save(flags);
6395 if (zone->pageset != &boot_pageset) {
6396 for_each_online_cpu(cpu) {
6397 pset = per_cpu_ptr(zone->pageset, cpu);
6398 drain_zonestat(zone, pset);
6400 free_percpu(zone->pageset);
6401 zone->pageset = &boot_pageset;
6403 local_irq_restore(flags);
6406 #ifdef CONFIG_MEMORY_HOTREMOVE
6408 * All pages in the range must be isolated before calling this.
6411 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6417 unsigned long flags;
6418 /* find the first valid pfn */
6419 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6424 zone = page_zone(pfn_to_page(pfn));
6425 spin_lock_irqsave(&zone->lock, flags);
6427 while (pfn < end_pfn) {
6428 if (!pfn_valid(pfn)) {
6432 page = pfn_to_page(pfn);
6434 * The HWPoisoned page may be not in buddy system, and
6435 * page_count() is not 0.
6437 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6439 SetPageReserved(page);
6443 BUG_ON(page_count(page));
6444 BUG_ON(!PageBuddy(page));
6445 order = page_order(page);
6446 #ifdef CONFIG_DEBUG_VM
6447 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6448 pfn, 1 << order, end_pfn);
6450 list_del(&page->lru);
6451 rmv_page_order(page);
6452 zone->free_area[order].nr_free--;
6453 for (i = 0; i < (1 << order); i++)
6454 SetPageReserved((page+i));
6455 pfn += (1 << order);
6457 spin_unlock_irqrestore(&zone->lock, flags);
6461 #ifdef CONFIG_MEMORY_FAILURE
6462 bool is_free_buddy_page(struct page *page)
6464 struct zone *zone = page_zone(page);
6465 unsigned long pfn = page_to_pfn(page);
6466 unsigned long flags;
6469 spin_lock_irqsave(&zone->lock, flags);
6470 for (order = 0; order < MAX_ORDER; order++) {
6471 struct page *page_head = page - (pfn & ((1 << order) - 1));
6473 if (PageBuddy(page_head) && page_order(page_head) >= order)
6476 spin_unlock_irqrestore(&zone->lock, flags);
6478 return order < MAX_ORDER;
6482 static const struct trace_print_flags pageflag_names[] = {
6483 {1UL << PG_locked, "locked" },
6484 {1UL << PG_error, "error" },
6485 {1UL << PG_referenced, "referenced" },
6486 {1UL << PG_uptodate, "uptodate" },
6487 {1UL << PG_dirty, "dirty" },
6488 {1UL << PG_lru, "lru" },
6489 {1UL << PG_active, "active" },
6490 {1UL << PG_slab, "slab" },
6491 {1UL << PG_owner_priv_1, "owner_priv_1" },
6492 {1UL << PG_arch_1, "arch_1" },
6493 {1UL << PG_reserved, "reserved" },
6494 {1UL << PG_private, "private" },
6495 {1UL << PG_private_2, "private_2" },
6496 {1UL << PG_writeback, "writeback" },
6497 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6498 {1UL << PG_head, "head" },
6499 {1UL << PG_tail, "tail" },
6501 {1UL << PG_compound, "compound" },
6503 {1UL << PG_swapcache, "swapcache" },
6504 {1UL << PG_mappedtodisk, "mappedtodisk" },
6505 {1UL << PG_reclaim, "reclaim" },
6506 {1UL << PG_swapbacked, "swapbacked" },
6507 {1UL << PG_unevictable, "unevictable" },
6509 {1UL << PG_mlocked, "mlocked" },
6511 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6512 {1UL << PG_uncached, "uncached" },
6514 #ifdef CONFIG_MEMORY_FAILURE
6515 {1UL << PG_hwpoison, "hwpoison" },
6517 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6518 {1UL << PG_compound_lock, "compound_lock" },
6522 static void dump_page_flags(unsigned long flags)
6524 const char *delim = "";
6528 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6530 printk(KERN_ALERT "page flags: %#lx(", flags);
6532 /* remove zone id */
6533 flags &= (1UL << NR_PAGEFLAGS) - 1;
6535 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6537 mask = pageflag_names[i].mask;
6538 if ((flags & mask) != mask)
6542 printk("%s%s", delim, pageflag_names[i].name);
6546 /* check for left over flags */
6548 printk("%s%#lx", delim, flags);
6553 void dump_page_badflags(struct page *page, const char *reason,
6554 unsigned long badflags)
6557 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6558 page, atomic_read(&page->_count), page_mapcount(page),
6559 page->mapping, page->index);
6560 dump_page_flags(page->flags);
6562 pr_alert("page dumped because: %s\n", reason);
6563 if (page->flags & badflags) {
6564 pr_alert("bad because of flags:\n");
6565 dump_page_flags(page->flags & badflags);
6567 mem_cgroup_print_bad_page(page);
6570 void dump_page(struct page *page, const char *reason)
6572 dump_page_badflags(page, reason, 0);
6574 EXPORT_SYMBOL(dump_page);