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_is_guard(buddy) && page_order(buddy) == order) {
513 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
515 if (page_zone_id(page) != page_zone_id(buddy))
521 if (PageBuddy(buddy) && page_order(buddy) == order) {
522 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
525 * zone check is done late to avoid uselessly
526 * calculating zone/node ids for pages that could
529 if (page_zone_id(page) != page_zone_id(buddy))
538 * Freeing function for a buddy system allocator.
540 * The concept of a buddy system is to maintain direct-mapped table
541 * (containing bit values) for memory blocks of various "orders".
542 * The bottom level table contains the map for the smallest allocatable
543 * units of memory (here, pages), and each level above it describes
544 * pairs of units from the levels below, hence, "buddies".
545 * At a high level, all that happens here is marking the table entry
546 * at the bottom level available, and propagating the changes upward
547 * as necessary, plus some accounting needed to play nicely with other
548 * parts of the VM system.
549 * At each level, we keep a list of pages, which are heads of continuous
550 * free pages of length of (1 << order) and marked with _mapcount
551 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
553 * So when we are allocating or freeing one, we can derive the state of the
554 * other. That is, if we allocate a small block, and both were
555 * free, the remainder of the region must be split into blocks.
556 * If a block is freed, and its buddy is also free, then this
557 * triggers coalescing into a block of larger size.
562 static inline void __free_one_page(struct page *page,
563 struct zone *zone, unsigned int order,
566 unsigned long page_idx;
567 unsigned long combined_idx;
568 unsigned long uninitialized_var(buddy_idx);
571 VM_BUG_ON(!zone_is_initialized(zone));
573 if (unlikely(PageCompound(page)))
574 if (unlikely(destroy_compound_page(page, order)))
577 VM_BUG_ON(migratetype == -1);
579 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
581 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
582 VM_BUG_ON_PAGE(bad_range(zone, page), page);
584 while (order < MAX_ORDER-1) {
585 buddy_idx = __find_buddy_index(page_idx, order);
586 buddy = page + (buddy_idx - page_idx);
587 if (!page_is_buddy(page, buddy, order))
590 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
591 * merge with it and move up one order.
593 if (page_is_guard(buddy)) {
594 clear_page_guard_flag(buddy);
595 set_page_private(page, 0);
596 __mod_zone_freepage_state(zone, 1 << order,
599 list_del(&buddy->lru);
600 zone->free_area[order].nr_free--;
601 rmv_page_order(buddy);
603 combined_idx = buddy_idx & page_idx;
604 page = page + (combined_idx - page_idx);
605 page_idx = combined_idx;
608 set_page_order(page, order);
611 * If this is not the largest possible page, check if the buddy
612 * of the next-highest order is free. If it is, it's possible
613 * that pages are being freed that will coalesce soon. In case,
614 * that is happening, add the free page to the tail of the list
615 * so it's less likely to be used soon and more likely to be merged
616 * as a higher order page
618 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
619 struct page *higher_page, *higher_buddy;
620 combined_idx = buddy_idx & page_idx;
621 higher_page = page + (combined_idx - page_idx);
622 buddy_idx = __find_buddy_index(combined_idx, order + 1);
623 higher_buddy = higher_page + (buddy_idx - combined_idx);
624 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
625 list_add_tail(&page->lru,
626 &zone->free_area[order].free_list[migratetype]);
631 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
633 zone->free_area[order].nr_free++;
636 static inline int free_pages_check(struct page *page)
638 const char *bad_reason = NULL;
639 unsigned long bad_flags = 0;
641 if (unlikely(page_mapcount(page)))
642 bad_reason = "nonzero mapcount";
643 if (unlikely(page->mapping != NULL))
644 bad_reason = "non-NULL mapping";
645 if (unlikely(atomic_read(&page->_count) != 0))
646 bad_reason = "nonzero _count";
647 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
648 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
649 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
651 if (unlikely(mem_cgroup_bad_page_check(page)))
652 bad_reason = "cgroup check failed";
653 if (unlikely(bad_reason)) {
654 bad_page(page, bad_reason, bad_flags);
657 page_cpupid_reset_last(page);
658 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
659 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
664 * Frees a number of pages from the PCP lists
665 * Assumes all pages on list are in same zone, and of same order.
666 * count is the number of pages to free.
668 * If the zone was previously in an "all pages pinned" state then look to
669 * see if this freeing clears that state.
671 * And clear the zone's pages_scanned counter, to hold off the "all pages are
672 * pinned" detection logic.
674 static void free_pcppages_bulk(struct zone *zone, int count,
675 struct per_cpu_pages *pcp)
681 spin_lock(&zone->lock);
682 zone->pages_scanned = 0;
686 struct list_head *list;
689 * Remove pages from lists in a round-robin fashion. A
690 * batch_free count is maintained that is incremented when an
691 * empty list is encountered. This is so more pages are freed
692 * off fuller lists instead of spinning excessively around empty
697 if (++migratetype == MIGRATE_PCPTYPES)
699 list = &pcp->lists[migratetype];
700 } while (list_empty(list));
702 /* This is the only non-empty list. Free them all. */
703 if (batch_free == MIGRATE_PCPTYPES)
704 batch_free = to_free;
707 int mt; /* migratetype of the to-be-freed page */
709 page = list_entry(list->prev, struct page, lru);
710 /* must delete as __free_one_page list manipulates */
711 list_del(&page->lru);
712 mt = get_freepage_migratetype(page);
713 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
714 __free_one_page(page, zone, 0, mt);
715 trace_mm_page_pcpu_drain(page, 0, mt);
716 if (likely(!is_migrate_isolate_page(page))) {
717 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
718 if (is_migrate_cma(mt))
719 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
721 } while (--to_free && --batch_free && !list_empty(list));
723 spin_unlock(&zone->lock);
726 static void free_one_page(struct zone *zone, struct page *page, int order,
729 spin_lock(&zone->lock);
730 zone->pages_scanned = 0;
732 __free_one_page(page, zone, order, migratetype);
733 if (unlikely(!is_migrate_isolate(migratetype)))
734 __mod_zone_freepage_state(zone, 1 << order, migratetype);
735 spin_unlock(&zone->lock);
738 static bool free_pages_prepare(struct page *page, unsigned int order)
743 trace_mm_page_free(page, order);
744 kmemcheck_free_shadow(page, order);
747 page->mapping = NULL;
748 for (i = 0; i < (1 << order); i++)
749 bad += free_pages_check(page + i);
753 if (!PageHighMem(page)) {
754 debug_check_no_locks_freed(page_address(page),
756 debug_check_no_obj_freed(page_address(page),
759 arch_free_page(page, order);
760 kernel_map_pages(page, 1 << order, 0);
765 static void __free_pages_ok(struct page *page, unsigned int order)
770 if (!free_pages_prepare(page, order))
773 local_irq_save(flags);
774 __count_vm_events(PGFREE, 1 << order);
775 migratetype = get_pageblock_migratetype(page);
776 set_freepage_migratetype(page, migratetype);
777 free_one_page(page_zone(page), page, order, migratetype);
778 local_irq_restore(flags);
781 void __init __free_pages_bootmem(struct page *page, unsigned int order)
783 unsigned int nr_pages = 1 << order;
784 struct page *p = page;
788 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
790 __ClearPageReserved(p);
791 set_page_count(p, 0);
793 __ClearPageReserved(p);
794 set_page_count(p, 0);
796 page_zone(page)->managed_pages += nr_pages;
797 set_page_refcounted(page);
798 __free_pages(page, order);
802 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
803 void __init init_cma_reserved_pageblock(struct page *page)
805 unsigned i = pageblock_nr_pages;
806 struct page *p = page;
809 __ClearPageReserved(p);
810 set_page_count(p, 0);
813 set_page_refcounted(page);
814 set_pageblock_migratetype(page, MIGRATE_CMA);
815 __free_pages(page, pageblock_order);
816 adjust_managed_page_count(page, pageblock_nr_pages);
821 * The order of subdivision here is critical for the IO subsystem.
822 * Please do not alter this order without good reasons and regression
823 * testing. Specifically, as large blocks of memory are subdivided,
824 * the order in which smaller blocks are delivered depends on the order
825 * they're subdivided in this function. This is the primary factor
826 * influencing the order in which pages are delivered to the IO
827 * subsystem according to empirical testing, and this is also justified
828 * by considering the behavior of a buddy system containing a single
829 * large block of memory acted on by a series of small allocations.
830 * This behavior is a critical factor in sglist merging's success.
834 static inline void expand(struct zone *zone, struct page *page,
835 int low, int high, struct free_area *area,
838 unsigned long size = 1 << high;
844 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
846 #ifdef CONFIG_DEBUG_PAGEALLOC
847 if (high < debug_guardpage_minorder()) {
849 * Mark as guard pages (or page), that will allow to
850 * merge back to allocator when buddy will be freed.
851 * Corresponding page table entries will not be touched,
852 * pages will stay not present in virtual address space
854 INIT_LIST_HEAD(&page[size].lru);
855 set_page_guard_flag(&page[size]);
856 set_page_private(&page[size], high);
857 /* Guard pages are not available for any usage */
858 __mod_zone_freepage_state(zone, -(1 << high),
863 list_add(&page[size].lru, &area->free_list[migratetype]);
865 set_page_order(&page[size], high);
870 * This page is about to be returned from the page allocator
872 static inline int check_new_page(struct page *page)
874 const char *bad_reason = NULL;
875 unsigned long bad_flags = 0;
877 if (unlikely(page_mapcount(page)))
878 bad_reason = "nonzero mapcount";
879 if (unlikely(page->mapping != NULL))
880 bad_reason = "non-NULL mapping";
881 if (unlikely(atomic_read(&page->_count) != 0))
882 bad_reason = "nonzero _count";
883 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
884 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
885 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
887 if (unlikely(mem_cgroup_bad_page_check(page)))
888 bad_reason = "cgroup check failed";
889 if (unlikely(bad_reason)) {
890 bad_page(page, bad_reason, bad_flags);
896 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
900 for (i = 0; i < (1 << order); i++) {
901 struct page *p = page + i;
902 if (unlikely(check_new_page(p)))
906 set_page_private(page, 0);
907 set_page_refcounted(page);
909 arch_alloc_page(page, order);
910 kernel_map_pages(page, 1 << order, 1);
912 if (gfp_flags & __GFP_ZERO)
913 prep_zero_page(page, order, gfp_flags);
915 if (order && (gfp_flags & __GFP_COMP))
916 prep_compound_page(page, order);
922 * Go through the free lists for the given migratetype and remove
923 * the smallest available page from the freelists
926 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
929 unsigned int current_order;
930 struct free_area *area;
933 /* Find a page of the appropriate size in the preferred list */
934 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
935 area = &(zone->free_area[current_order]);
936 if (list_empty(&area->free_list[migratetype]))
939 page = list_entry(area->free_list[migratetype].next,
941 list_del(&page->lru);
942 rmv_page_order(page);
944 expand(zone, page, order, current_order, area, migratetype);
945 set_freepage_migratetype(page, migratetype);
954 * This array describes the order lists are fallen back to when
955 * the free lists for the desirable migrate type are depleted
957 static int fallbacks[MIGRATE_TYPES][4] = {
958 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
959 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
961 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
962 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
964 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
966 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
967 #ifdef CONFIG_MEMORY_ISOLATION
968 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
973 * Move the free pages in a range to the free lists of the requested type.
974 * Note that start_page and end_pages are not aligned on a pageblock
975 * boundary. If alignment is required, use move_freepages_block()
977 int move_freepages(struct zone *zone,
978 struct page *start_page, struct page *end_page,
985 #ifndef CONFIG_HOLES_IN_ZONE
987 * page_zone is not safe to call in this context when
988 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
989 * anyway as we check zone boundaries in move_freepages_block().
990 * Remove at a later date when no bug reports exist related to
991 * grouping pages by mobility
993 BUG_ON(page_zone(start_page) != page_zone(end_page));
996 for (page = start_page; page <= end_page;) {
997 /* Make sure we are not inadvertently changing nodes */
998 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1000 if (!pfn_valid_within(page_to_pfn(page))) {
1005 if (!PageBuddy(page)) {
1010 order = page_order(page);
1011 list_move(&page->lru,
1012 &zone->free_area[order].free_list[migratetype]);
1013 set_freepage_migratetype(page, migratetype);
1015 pages_moved += 1 << order;
1021 int move_freepages_block(struct zone *zone, struct page *page,
1024 unsigned long start_pfn, end_pfn;
1025 struct page *start_page, *end_page;
1027 start_pfn = page_to_pfn(page);
1028 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1029 start_page = pfn_to_page(start_pfn);
1030 end_page = start_page + pageblock_nr_pages - 1;
1031 end_pfn = start_pfn + pageblock_nr_pages - 1;
1033 /* Do not cross zone boundaries */
1034 if (!zone_spans_pfn(zone, start_pfn))
1036 if (!zone_spans_pfn(zone, end_pfn))
1039 return move_freepages(zone, start_page, end_page, migratetype);
1042 static void change_pageblock_range(struct page *pageblock_page,
1043 int start_order, int migratetype)
1045 int nr_pageblocks = 1 << (start_order - pageblock_order);
1047 while (nr_pageblocks--) {
1048 set_pageblock_migratetype(pageblock_page, migratetype);
1049 pageblock_page += pageblock_nr_pages;
1054 * If breaking a large block of pages, move all free pages to the preferred
1055 * allocation list. If falling back for a reclaimable kernel allocation, be
1056 * more aggressive about taking ownership of free pages.
1058 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1059 * nor move CMA pages to different free lists. We don't want unmovable pages
1060 * to be allocated from MIGRATE_CMA areas.
1062 * Returns the new migratetype of the pageblock (or the same old migratetype
1063 * if it was unchanged).
1065 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1066 int start_type, int fallback_type)
1068 int current_order = page_order(page);
1071 * When borrowing from MIGRATE_CMA, we need to release the excess
1072 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1073 * is set to CMA so it is returned to the correct freelist in case
1074 * the page ends up being not actually allocated from the pcp lists.
1076 if (is_migrate_cma(fallback_type))
1077 return fallback_type;
1079 /* Take ownership for orders >= pageblock_order */
1080 if (current_order >= pageblock_order) {
1081 change_pageblock_range(page, current_order, start_type);
1085 if (current_order >= pageblock_order / 2 ||
1086 start_type == MIGRATE_RECLAIMABLE ||
1087 page_group_by_mobility_disabled) {
1090 pages = move_freepages_block(zone, page, start_type);
1092 /* Claim the whole block if over half of it is free */
1093 if (pages >= (1 << (pageblock_order-1)) ||
1094 page_group_by_mobility_disabled) {
1096 set_pageblock_migratetype(page, start_type);
1102 return fallback_type;
1105 /* Remove an element from the buddy allocator from the fallback list */
1106 static inline struct page *
1107 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1109 struct free_area *area;
1112 int migratetype, new_type, i;
1114 /* Find the largest possible block of pages in the other list */
1115 for (current_order = MAX_ORDER-1; current_order >= order;
1118 migratetype = fallbacks[start_migratetype][i];
1120 /* MIGRATE_RESERVE handled later if necessary */
1121 if (migratetype == MIGRATE_RESERVE)
1124 area = &(zone->free_area[current_order]);
1125 if (list_empty(&area->free_list[migratetype]))
1128 page = list_entry(area->free_list[migratetype].next,
1132 new_type = try_to_steal_freepages(zone, page,
1136 /* Remove the page from the freelists */
1137 list_del(&page->lru);
1138 rmv_page_order(page);
1140 expand(zone, page, order, current_order, area,
1142 /* The freepage_migratetype may differ from pageblock's
1143 * migratetype depending on the decisions in
1144 * try_to_steal_freepages. This is OK as long as it does
1145 * not differ for MIGRATE_CMA type.
1147 set_freepage_migratetype(page, new_type);
1149 trace_mm_page_alloc_extfrag(page, order, current_order,
1150 start_migratetype, migratetype, new_type);
1160 * Do the hard work of removing an element from the buddy allocator.
1161 * Call me with the zone->lock already held.
1163 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1169 page = __rmqueue_smallest(zone, order, migratetype);
1171 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1172 page = __rmqueue_fallback(zone, order, migratetype);
1175 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1176 * is used because __rmqueue_smallest is an inline function
1177 * and we want just one call site
1180 migratetype = MIGRATE_RESERVE;
1185 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1190 * Obtain a specified number of elements from the buddy allocator, all under
1191 * a single hold of the lock, for efficiency. Add them to the supplied list.
1192 * Returns the number of new pages which were placed at *list.
1194 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1195 unsigned long count, struct list_head *list,
1196 int migratetype, int cold)
1200 spin_lock(&zone->lock);
1201 for (i = 0; i < count; ++i) {
1202 struct page *page = __rmqueue(zone, order, migratetype);
1203 if (unlikely(page == NULL))
1207 * Split buddy pages returned by expand() are received here
1208 * in physical page order. The page is added to the callers and
1209 * list and the list head then moves forward. From the callers
1210 * perspective, the linked list is ordered by page number in
1211 * some conditions. This is useful for IO devices that can
1212 * merge IO requests if the physical pages are ordered
1215 if (likely(cold == 0))
1216 list_add(&page->lru, list);
1218 list_add_tail(&page->lru, list);
1220 if (is_migrate_cma(get_freepage_migratetype(page)))
1221 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1224 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1225 spin_unlock(&zone->lock);
1231 * Called from the vmstat counter updater to drain pagesets of this
1232 * currently executing processor on remote nodes after they have
1235 * Note that this function must be called with the thread pinned to
1236 * a single processor.
1238 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1240 unsigned long flags;
1242 unsigned long batch;
1244 local_irq_save(flags);
1245 batch = ACCESS_ONCE(pcp->batch);
1246 if (pcp->count >= batch)
1249 to_drain = pcp->count;
1251 free_pcppages_bulk(zone, to_drain, pcp);
1252 pcp->count -= to_drain;
1254 local_irq_restore(flags);
1259 * Drain pages of the indicated processor.
1261 * The processor must either be the current processor and the
1262 * thread pinned to the current processor or a processor that
1265 static void drain_pages(unsigned int cpu)
1267 unsigned long flags;
1270 for_each_populated_zone(zone) {
1271 struct per_cpu_pageset *pset;
1272 struct per_cpu_pages *pcp;
1274 local_irq_save(flags);
1275 pset = per_cpu_ptr(zone->pageset, cpu);
1279 free_pcppages_bulk(zone, pcp->count, pcp);
1282 local_irq_restore(flags);
1287 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1289 void drain_local_pages(void *arg)
1291 drain_pages(smp_processor_id());
1295 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1297 * Note that this code is protected against sending an IPI to an offline
1298 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1299 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1300 * nothing keeps CPUs from showing up after we populated the cpumask and
1301 * before the call to on_each_cpu_mask().
1303 void drain_all_pages(void)
1306 struct per_cpu_pageset *pcp;
1310 * Allocate in the BSS so we wont require allocation in
1311 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1313 static cpumask_t cpus_with_pcps;
1316 * We don't care about racing with CPU hotplug event
1317 * as offline notification will cause the notified
1318 * cpu to drain that CPU pcps and on_each_cpu_mask
1319 * disables preemption as part of its processing
1321 for_each_online_cpu(cpu) {
1322 bool has_pcps = false;
1323 for_each_populated_zone(zone) {
1324 pcp = per_cpu_ptr(zone->pageset, cpu);
1325 if (pcp->pcp.count) {
1331 cpumask_set_cpu(cpu, &cpus_with_pcps);
1333 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1335 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1338 #ifdef CONFIG_HIBERNATION
1340 void mark_free_pages(struct zone *zone)
1342 unsigned long pfn, max_zone_pfn;
1343 unsigned long flags;
1345 struct list_head *curr;
1347 if (zone_is_empty(zone))
1350 spin_lock_irqsave(&zone->lock, flags);
1352 max_zone_pfn = zone_end_pfn(zone);
1353 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1354 if (pfn_valid(pfn)) {
1355 struct page *page = pfn_to_page(pfn);
1357 if (!swsusp_page_is_forbidden(page))
1358 swsusp_unset_page_free(page);
1361 for_each_migratetype_order(order, t) {
1362 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1365 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1366 for (i = 0; i < (1UL << order); i++)
1367 swsusp_set_page_free(pfn_to_page(pfn + i));
1370 spin_unlock_irqrestore(&zone->lock, flags);
1372 #endif /* CONFIG_PM */
1375 * Free a 0-order page
1376 * cold == 1 ? free a cold page : free a hot page
1378 void free_hot_cold_page(struct page *page, int cold)
1380 struct zone *zone = page_zone(page);
1381 struct per_cpu_pages *pcp;
1382 unsigned long flags;
1385 if (!free_pages_prepare(page, 0))
1388 migratetype = get_pageblock_migratetype(page);
1389 set_freepage_migratetype(page, migratetype);
1390 local_irq_save(flags);
1391 __count_vm_event(PGFREE);
1394 * We only track unmovable, reclaimable and movable on pcp lists.
1395 * Free ISOLATE pages back to the allocator because they are being
1396 * offlined but treat RESERVE as movable pages so we can get those
1397 * areas back if necessary. Otherwise, we may have to free
1398 * excessively into the page allocator
1400 if (migratetype >= MIGRATE_PCPTYPES) {
1401 if (unlikely(is_migrate_isolate(migratetype))) {
1402 free_one_page(zone, page, 0, migratetype);
1405 migratetype = MIGRATE_MOVABLE;
1408 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1410 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1412 list_add(&page->lru, &pcp->lists[migratetype]);
1414 if (pcp->count >= pcp->high) {
1415 unsigned long batch = ACCESS_ONCE(pcp->batch);
1416 free_pcppages_bulk(zone, batch, pcp);
1417 pcp->count -= batch;
1421 local_irq_restore(flags);
1425 * Free a list of 0-order pages
1427 void free_hot_cold_page_list(struct list_head *list, int cold)
1429 struct page *page, *next;
1431 list_for_each_entry_safe(page, next, list, lru) {
1432 trace_mm_page_free_batched(page, cold);
1433 free_hot_cold_page(page, cold);
1438 * split_page takes a non-compound higher-order page, and splits it into
1439 * n (1<<order) sub-pages: page[0..n]
1440 * Each sub-page must be freed individually.
1442 * Note: this is probably too low level an operation for use in drivers.
1443 * Please consult with lkml before using this in your driver.
1445 void split_page(struct page *page, unsigned int order)
1449 VM_BUG_ON_PAGE(PageCompound(page), page);
1450 VM_BUG_ON_PAGE(!page_count(page), page);
1452 #ifdef CONFIG_KMEMCHECK
1454 * Split shadow pages too, because free(page[0]) would
1455 * otherwise free the whole shadow.
1457 if (kmemcheck_page_is_tracked(page))
1458 split_page(virt_to_page(page[0].shadow), order);
1461 for (i = 1; i < (1 << order); i++)
1462 set_page_refcounted(page + i);
1464 EXPORT_SYMBOL_GPL(split_page);
1466 static int __isolate_free_page(struct page *page, unsigned int order)
1468 unsigned long watermark;
1472 BUG_ON(!PageBuddy(page));
1474 zone = page_zone(page);
1475 mt = get_pageblock_migratetype(page);
1477 if (!is_migrate_isolate(mt)) {
1478 /* Obey watermarks as if the page was being allocated */
1479 watermark = low_wmark_pages(zone) + (1 << order);
1480 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1483 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1486 /* Remove page from free list */
1487 list_del(&page->lru);
1488 zone->free_area[order].nr_free--;
1489 rmv_page_order(page);
1491 /* Set the pageblock if the isolated page is at least a pageblock */
1492 if (order >= pageblock_order - 1) {
1493 struct page *endpage = page + (1 << order) - 1;
1494 for (; page < endpage; page += pageblock_nr_pages) {
1495 int mt = get_pageblock_migratetype(page);
1496 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1497 set_pageblock_migratetype(page,
1502 return 1UL << order;
1506 * Similar to split_page except the page is already free. As this is only
1507 * being used for migration, the migratetype of the block also changes.
1508 * As this is called with interrupts disabled, the caller is responsible
1509 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1512 * Note: this is probably too low level an operation for use in drivers.
1513 * Please consult with lkml before using this in your driver.
1515 int split_free_page(struct page *page)
1520 order = page_order(page);
1522 nr_pages = __isolate_free_page(page, order);
1526 /* Split into individual pages */
1527 set_page_refcounted(page);
1528 split_page(page, order);
1533 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1534 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1538 struct page *buffered_rmqueue(struct zone *preferred_zone,
1539 struct zone *zone, int order, gfp_t gfp_flags,
1542 unsigned long flags;
1544 int cold = !!(gfp_flags & __GFP_COLD);
1547 if (likely(order == 0)) {
1548 struct per_cpu_pages *pcp;
1549 struct list_head *list;
1551 local_irq_save(flags);
1552 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1553 list = &pcp->lists[migratetype];
1554 if (list_empty(list)) {
1555 pcp->count += rmqueue_bulk(zone, 0,
1558 if (unlikely(list_empty(list)))
1563 page = list_entry(list->prev, struct page, lru);
1565 page = list_entry(list->next, struct page, lru);
1567 list_del(&page->lru);
1570 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1572 * __GFP_NOFAIL is not to be used in new code.
1574 * All __GFP_NOFAIL callers should be fixed so that they
1575 * properly detect and handle allocation failures.
1577 * We most definitely don't want callers attempting to
1578 * allocate greater than order-1 page units with
1581 WARN_ON_ONCE(order > 1);
1583 spin_lock_irqsave(&zone->lock, flags);
1584 page = __rmqueue(zone, order, migratetype);
1585 spin_unlock(&zone->lock);
1588 __mod_zone_freepage_state(zone, -(1 << order),
1589 get_freepage_migratetype(page));
1592 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1594 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1595 zone_statistics(preferred_zone, zone, gfp_flags);
1596 local_irq_restore(flags);
1598 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1599 if (prep_new_page(page, order, gfp_flags))
1604 local_irq_restore(flags);
1608 #ifdef CONFIG_FAIL_PAGE_ALLOC
1611 struct fault_attr attr;
1613 u32 ignore_gfp_highmem;
1614 u32 ignore_gfp_wait;
1616 } fail_page_alloc = {
1617 .attr = FAULT_ATTR_INITIALIZER,
1618 .ignore_gfp_wait = 1,
1619 .ignore_gfp_highmem = 1,
1623 static int __init setup_fail_page_alloc(char *str)
1625 return setup_fault_attr(&fail_page_alloc.attr, str);
1627 __setup("fail_page_alloc=", setup_fail_page_alloc);
1629 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1631 if (order < fail_page_alloc.min_order)
1633 if (gfp_mask & __GFP_NOFAIL)
1635 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1637 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1640 return should_fail(&fail_page_alloc.attr, 1 << order);
1643 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1645 static int __init fail_page_alloc_debugfs(void)
1647 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1650 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1651 &fail_page_alloc.attr);
1653 return PTR_ERR(dir);
1655 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1656 &fail_page_alloc.ignore_gfp_wait))
1658 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1659 &fail_page_alloc.ignore_gfp_highmem))
1661 if (!debugfs_create_u32("min-order", mode, dir,
1662 &fail_page_alloc.min_order))
1667 debugfs_remove_recursive(dir);
1672 late_initcall(fail_page_alloc_debugfs);
1674 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1676 #else /* CONFIG_FAIL_PAGE_ALLOC */
1678 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1683 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1686 * Return true if free pages are above 'mark'. This takes into account the order
1687 * of the allocation.
1689 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1690 int classzone_idx, int alloc_flags, long free_pages)
1692 /* free_pages my go negative - that's OK */
1694 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1698 free_pages -= (1 << order) - 1;
1699 if (alloc_flags & ALLOC_HIGH)
1701 if (alloc_flags & ALLOC_HARDER)
1704 /* If allocation can't use CMA areas don't use free CMA pages */
1705 if (!(alloc_flags & ALLOC_CMA))
1706 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1709 if (free_pages - free_cma <= min + lowmem_reserve)
1711 for (o = 0; o < order; o++) {
1712 /* At the next order, this order's pages become unavailable */
1713 free_pages -= z->free_area[o].nr_free << o;
1715 /* Require fewer higher order pages to be free */
1718 if (free_pages <= min)
1724 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1725 int classzone_idx, int alloc_flags)
1727 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1728 zone_page_state(z, NR_FREE_PAGES));
1731 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1732 int classzone_idx, int alloc_flags)
1734 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1736 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1737 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1739 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1745 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1746 * skip over zones that are not allowed by the cpuset, or that have
1747 * been recently (in last second) found to be nearly full. See further
1748 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1749 * that have to skip over a lot of full or unallowed zones.
1751 * If the zonelist cache is present in the passed zonelist, then
1752 * returns a pointer to the allowed node mask (either the current
1753 * tasks mems_allowed, or node_states[N_MEMORY].)
1755 * If the zonelist cache is not available for this zonelist, does
1756 * nothing and returns NULL.
1758 * If the fullzones BITMAP in the zonelist cache is stale (more than
1759 * a second since last zap'd) then we zap it out (clear its bits.)
1761 * We hold off even calling zlc_setup, until after we've checked the
1762 * first zone in the zonelist, on the theory that most allocations will
1763 * be satisfied from that first zone, so best to examine that zone as
1764 * quickly as we can.
1766 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1768 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1769 nodemask_t *allowednodes; /* zonelist_cache approximation */
1771 zlc = zonelist->zlcache_ptr;
1775 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1776 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1777 zlc->last_full_zap = jiffies;
1780 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1781 &cpuset_current_mems_allowed :
1782 &node_states[N_MEMORY];
1783 return allowednodes;
1787 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1788 * if it is worth looking at further for free memory:
1789 * 1) Check that the zone isn't thought to be full (doesn't have its
1790 * bit set in the zonelist_cache fullzones BITMAP).
1791 * 2) Check that the zones node (obtained from the zonelist_cache
1792 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1793 * Return true (non-zero) if zone is worth looking at further, or
1794 * else return false (zero) if it is not.
1796 * This check -ignores- the distinction between various watermarks,
1797 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1798 * found to be full for any variation of these watermarks, it will
1799 * be considered full for up to one second by all requests, unless
1800 * we are so low on memory on all allowed nodes that we are forced
1801 * into the second scan of the zonelist.
1803 * In the second scan we ignore this zonelist cache and exactly
1804 * apply the watermarks to all zones, even it is slower to do so.
1805 * We are low on memory in the second scan, and should leave no stone
1806 * unturned looking for a free page.
1808 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1809 nodemask_t *allowednodes)
1811 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1812 int i; /* index of *z in zonelist zones */
1813 int n; /* node that zone *z is on */
1815 zlc = zonelist->zlcache_ptr;
1819 i = z - zonelist->_zonerefs;
1822 /* This zone is worth trying if it is allowed but not full */
1823 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1827 * Given 'z' scanning a zonelist, set the corresponding bit in
1828 * zlc->fullzones, so that subsequent attempts to allocate a page
1829 * from that zone don't waste time re-examining it.
1831 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1833 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1834 int i; /* index of *z in zonelist zones */
1836 zlc = zonelist->zlcache_ptr;
1840 i = z - zonelist->_zonerefs;
1842 set_bit(i, zlc->fullzones);
1846 * clear all zones full, called after direct reclaim makes progress so that
1847 * a zone that was recently full is not skipped over for up to a second
1849 static void zlc_clear_zones_full(struct zonelist *zonelist)
1851 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1853 zlc = zonelist->zlcache_ptr;
1857 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1860 static bool zone_local(struct zone *local_zone, struct zone *zone)
1862 return local_zone->node == zone->node;
1865 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1867 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1871 #else /* CONFIG_NUMA */
1873 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1878 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1879 nodemask_t *allowednodes)
1884 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1888 static void zlc_clear_zones_full(struct zonelist *zonelist)
1892 static bool zone_local(struct zone *local_zone, struct zone *zone)
1897 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1902 #endif /* CONFIG_NUMA */
1905 * get_page_from_freelist goes through the zonelist trying to allocate
1908 static struct page *
1909 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1910 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1911 struct zone *preferred_zone, int migratetype)
1914 struct page *page = NULL;
1917 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1918 int zlc_active = 0; /* set if using zonelist_cache */
1919 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1920 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
1921 (gfp_mask & __GFP_WRITE);
1923 classzone_idx = zone_idx(preferred_zone);
1926 * Scan zonelist, looking for a zone with enough free.
1927 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1929 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1930 high_zoneidx, nodemask) {
1933 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1934 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1936 if (cpusets_enabled() &&
1937 (alloc_flags & ALLOC_CPUSET) &&
1938 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1941 * Distribute pages in proportion to the individual
1942 * zone size to ensure fair page aging. The zone a
1943 * page was allocated in should have no effect on the
1944 * time the page has in memory before being reclaimed.
1946 if (alloc_flags & ALLOC_FAIR) {
1947 if (!zone_local(preferred_zone, zone))
1949 if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1953 * When allocating a page cache page for writing, we
1954 * want to get it from a zone that is within its dirty
1955 * limit, such that no single zone holds more than its
1956 * proportional share of globally allowed dirty pages.
1957 * The dirty limits take into account the zone's
1958 * lowmem reserves and high watermark so that kswapd
1959 * should be able to balance it without having to
1960 * write pages from its LRU list.
1962 * This may look like it could increase pressure on
1963 * lower zones by failing allocations in higher zones
1964 * before they are full. But the pages that do spill
1965 * over are limited as the lower zones are protected
1966 * by this very same mechanism. It should not become
1967 * a practical burden to them.
1969 * XXX: For now, allow allocations to potentially
1970 * exceed the per-zone dirty limit in the slowpath
1971 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1972 * which is important when on a NUMA setup the allowed
1973 * zones are together not big enough to reach the
1974 * global limit. The proper fix for these situations
1975 * will require awareness of zones in the
1976 * dirty-throttling and the flusher threads.
1978 if (consider_zone_dirty && !zone_dirty_ok(zone))
1981 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1982 if (!zone_watermark_ok(zone, order, mark,
1983 classzone_idx, alloc_flags)) {
1986 /* Checked here to keep the fast path fast */
1987 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1988 if (alloc_flags & ALLOC_NO_WATERMARKS)
1991 if (IS_ENABLED(CONFIG_NUMA) &&
1992 !did_zlc_setup && nr_online_nodes > 1) {
1994 * we do zlc_setup if there are multiple nodes
1995 * and before considering the first zone allowed
1998 allowednodes = zlc_setup(zonelist, alloc_flags);
2003 if (zone_reclaim_mode == 0 ||
2004 !zone_allows_reclaim(preferred_zone, zone))
2005 goto this_zone_full;
2008 * As we may have just activated ZLC, check if the first
2009 * eligible zone has failed zone_reclaim recently.
2011 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2012 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2015 ret = zone_reclaim(zone, gfp_mask, order);
2017 case ZONE_RECLAIM_NOSCAN:
2020 case ZONE_RECLAIM_FULL:
2021 /* scanned but unreclaimable */
2024 /* did we reclaim enough */
2025 if (zone_watermark_ok(zone, order, mark,
2026 classzone_idx, alloc_flags))
2030 * Failed to reclaim enough to meet watermark.
2031 * Only mark the zone full if checking the min
2032 * watermark or if we failed to reclaim just
2033 * 1<<order pages or else the page allocator
2034 * fastpath will prematurely mark zones full
2035 * when the watermark is between the low and
2038 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2039 ret == ZONE_RECLAIM_SOME)
2040 goto this_zone_full;
2047 page = buffered_rmqueue(preferred_zone, zone, order,
2048 gfp_mask, migratetype);
2052 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2053 zlc_mark_zone_full(zonelist, z);
2056 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2057 /* Disable zlc cache for second zonelist scan */
2064 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2065 * necessary to allocate the page. The expectation is
2066 * that the caller is taking steps that will free more
2067 * memory. The caller should avoid the page being used
2068 * for !PFMEMALLOC purposes.
2070 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2076 * Large machines with many possible nodes should not always dump per-node
2077 * meminfo in irq context.
2079 static inline bool should_suppress_show_mem(void)
2084 ret = in_interrupt();
2089 static DEFINE_RATELIMIT_STATE(nopage_rs,
2090 DEFAULT_RATELIMIT_INTERVAL,
2091 DEFAULT_RATELIMIT_BURST);
2093 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2095 unsigned int filter = SHOW_MEM_FILTER_NODES;
2097 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2098 debug_guardpage_minorder() > 0)
2102 * This documents exceptions given to allocations in certain
2103 * contexts that are allowed to allocate outside current's set
2106 if (!(gfp_mask & __GFP_NOMEMALLOC))
2107 if (test_thread_flag(TIF_MEMDIE) ||
2108 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2109 filter &= ~SHOW_MEM_FILTER_NODES;
2110 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2111 filter &= ~SHOW_MEM_FILTER_NODES;
2114 struct va_format vaf;
2117 va_start(args, fmt);
2122 pr_warn("%pV", &vaf);
2127 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2128 current->comm, order, gfp_mask);
2131 if (!should_suppress_show_mem())
2136 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2137 unsigned long did_some_progress,
2138 unsigned long pages_reclaimed)
2140 /* Do not loop if specifically requested */
2141 if (gfp_mask & __GFP_NORETRY)
2144 /* Always retry if specifically requested */
2145 if (gfp_mask & __GFP_NOFAIL)
2149 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2150 * making forward progress without invoking OOM. Suspend also disables
2151 * storage devices so kswapd will not help. Bail if we are suspending.
2153 if (!did_some_progress && pm_suspended_storage())
2157 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2158 * means __GFP_NOFAIL, but that may not be true in other
2161 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2165 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2166 * specified, then we retry until we no longer reclaim any pages
2167 * (above), or we've reclaimed an order of pages at least as
2168 * large as the allocation's order. In both cases, if the
2169 * allocation still fails, we stop retrying.
2171 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2177 static inline struct page *
2178 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2179 struct zonelist *zonelist, enum zone_type high_zoneidx,
2180 nodemask_t *nodemask, struct zone *preferred_zone,
2185 /* Acquire the OOM killer lock for the zones in zonelist */
2186 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2187 schedule_timeout_uninterruptible(1);
2192 * Go through the zonelist yet one more time, keep very high watermark
2193 * here, this is only to catch a parallel oom killing, we must fail if
2194 * we're still under heavy pressure.
2196 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2197 order, zonelist, high_zoneidx,
2198 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2199 preferred_zone, migratetype);
2203 if (!(gfp_mask & __GFP_NOFAIL)) {
2204 /* The OOM killer will not help higher order allocs */
2205 if (order > PAGE_ALLOC_COSTLY_ORDER)
2207 /* The OOM killer does not needlessly kill tasks for lowmem */
2208 if (high_zoneidx < ZONE_NORMAL)
2211 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2212 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2213 * The caller should handle page allocation failure by itself if
2214 * it specifies __GFP_THISNODE.
2215 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2217 if (gfp_mask & __GFP_THISNODE)
2220 /* Exhausted what can be done so it's blamo time */
2221 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2224 clear_zonelist_oom(zonelist, gfp_mask);
2228 #ifdef CONFIG_COMPACTION
2229 /* Try memory compaction for high-order allocations before reclaim */
2230 static struct page *
2231 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2232 struct zonelist *zonelist, enum zone_type high_zoneidx,
2233 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2234 int migratetype, enum migrate_mode mode,
2235 bool *contended_compaction, bool *deferred_compaction,
2236 unsigned long *did_some_progress)
2241 if (compaction_deferred(preferred_zone, order)) {
2242 *deferred_compaction = true;
2246 current->flags |= PF_MEMALLOC;
2247 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2249 contended_compaction);
2250 current->flags &= ~PF_MEMALLOC;
2252 if (*did_some_progress != COMPACT_SKIPPED) {
2255 /* Page migration frees to the PCP lists but we want merging */
2256 drain_pages(get_cpu());
2259 page = get_page_from_freelist(gfp_mask, nodemask,
2260 order, zonelist, high_zoneidx,
2261 alloc_flags & ~ALLOC_NO_WATERMARKS,
2262 preferred_zone, migratetype);
2264 preferred_zone->compact_blockskip_flush = false;
2265 compaction_defer_reset(preferred_zone, order, true);
2266 count_vm_event(COMPACTSUCCESS);
2271 * It's bad if compaction run occurs and fails.
2272 * The most likely reason is that pages exist,
2273 * but not enough to satisfy watermarks.
2275 count_vm_event(COMPACTFAIL);
2278 * As async compaction considers a subset of pageblocks, only
2279 * defer if the failure was a sync compaction failure.
2281 if (mode != MIGRATE_ASYNC)
2282 defer_compaction(preferred_zone, order);
2290 static inline struct page *
2291 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2292 struct zonelist *zonelist, enum zone_type high_zoneidx,
2293 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2294 int migratetype, enum migrate_mode mode, bool *contended_compaction,
2295 bool *deferred_compaction, unsigned long *did_some_progress)
2299 #endif /* CONFIG_COMPACTION */
2301 /* Perform direct synchronous page reclaim */
2303 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2304 nodemask_t *nodemask)
2306 struct reclaim_state reclaim_state;
2311 /* We now go into synchronous reclaim */
2312 cpuset_memory_pressure_bump();
2313 current->flags |= PF_MEMALLOC;
2314 lockdep_set_current_reclaim_state(gfp_mask);
2315 reclaim_state.reclaimed_slab = 0;
2316 current->reclaim_state = &reclaim_state;
2318 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2320 current->reclaim_state = NULL;
2321 lockdep_clear_current_reclaim_state();
2322 current->flags &= ~PF_MEMALLOC;
2329 /* The really slow allocator path where we enter direct reclaim */
2330 static inline struct page *
2331 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2332 struct zonelist *zonelist, enum zone_type high_zoneidx,
2333 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2334 int migratetype, unsigned long *did_some_progress)
2336 struct page *page = NULL;
2337 bool drained = false;
2339 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2341 if (unlikely(!(*did_some_progress)))
2344 /* After successful reclaim, reconsider all zones for allocation */
2345 if (IS_ENABLED(CONFIG_NUMA))
2346 zlc_clear_zones_full(zonelist);
2349 page = get_page_from_freelist(gfp_mask, nodemask, order,
2350 zonelist, high_zoneidx,
2351 alloc_flags & ~ALLOC_NO_WATERMARKS,
2352 preferred_zone, migratetype);
2355 * If an allocation failed after direct reclaim, it could be because
2356 * pages are pinned on the per-cpu lists. Drain them and try again
2358 if (!page && !drained) {
2368 * This is called in the allocator slow-path if the allocation request is of
2369 * sufficient urgency to ignore watermarks and take other desperate measures
2371 static inline struct page *
2372 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2373 struct zonelist *zonelist, enum zone_type high_zoneidx,
2374 nodemask_t *nodemask, struct zone *preferred_zone,
2380 page = get_page_from_freelist(gfp_mask, nodemask, order,
2381 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2382 preferred_zone, migratetype);
2384 if (!page && gfp_mask & __GFP_NOFAIL)
2385 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2386 } while (!page && (gfp_mask & __GFP_NOFAIL));
2391 static void reset_alloc_batches(struct zonelist *zonelist,
2392 enum zone_type high_zoneidx,
2393 struct zone *preferred_zone)
2398 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2400 * Only reset the batches of zones that were actually
2401 * considered in the fairness pass, we don't want to
2402 * trash fairness information for zones that are not
2403 * actually part of this zonelist's round-robin cycle.
2405 if (!zone_local(preferred_zone, zone))
2407 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2408 high_wmark_pages(zone) - low_wmark_pages(zone) -
2409 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2413 static void wake_all_kswapds(unsigned int order,
2414 struct zonelist *zonelist,
2415 enum zone_type high_zoneidx,
2416 struct zone *preferred_zone)
2421 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2422 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2426 gfp_to_alloc_flags(gfp_t gfp_mask)
2428 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2429 const gfp_t wait = gfp_mask & __GFP_WAIT;
2431 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2432 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2435 * The caller may dip into page reserves a bit more if the caller
2436 * cannot run direct reclaim, or if the caller has realtime scheduling
2437 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2438 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2440 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2444 * Not worth trying to allocate harder for
2445 * __GFP_NOMEMALLOC even if it can't schedule.
2447 if (!(gfp_mask & __GFP_NOMEMALLOC))
2448 alloc_flags |= ALLOC_HARDER;
2450 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2451 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2453 alloc_flags &= ~ALLOC_CPUSET;
2454 } else if (unlikely(rt_task(current)) && !in_interrupt())
2455 alloc_flags |= ALLOC_HARDER;
2457 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2458 if (gfp_mask & __GFP_MEMALLOC)
2459 alloc_flags |= ALLOC_NO_WATERMARKS;
2460 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2461 alloc_flags |= ALLOC_NO_WATERMARKS;
2462 else if (!in_interrupt() &&
2463 ((current->flags & PF_MEMALLOC) ||
2464 unlikely(test_thread_flag(TIF_MEMDIE))))
2465 alloc_flags |= ALLOC_NO_WATERMARKS;
2468 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2469 alloc_flags |= ALLOC_CMA;
2474 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2476 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2479 static inline struct page *
2480 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2481 struct zonelist *zonelist, enum zone_type high_zoneidx,
2482 nodemask_t *nodemask, struct zone *preferred_zone,
2485 const gfp_t wait = gfp_mask & __GFP_WAIT;
2486 struct page *page = NULL;
2488 unsigned long pages_reclaimed = 0;
2489 unsigned long did_some_progress;
2490 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2491 bool deferred_compaction = false;
2492 bool contended_compaction = false;
2495 * In the slowpath, we sanity check order to avoid ever trying to
2496 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2497 * be using allocators in order of preference for an area that is
2500 if (order >= MAX_ORDER) {
2501 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2506 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2507 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2508 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2509 * using a larger set of nodes after it has established that the
2510 * allowed per node queues are empty and that nodes are
2513 if (IS_ENABLED(CONFIG_NUMA) &&
2514 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2518 if (!(gfp_mask & __GFP_NO_KSWAPD))
2519 wake_all_kswapds(order, zonelist, high_zoneidx, preferred_zone);
2522 * OK, we're below the kswapd watermark and have kicked background
2523 * reclaim. Now things get more complex, so set up alloc_flags according
2524 * to how we want to proceed.
2526 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2529 * Find the true preferred zone if the allocation is unconstrained by
2532 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2533 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2537 /* This is the last chance, in general, before the goto nopage. */
2538 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2539 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2540 preferred_zone, migratetype);
2544 /* Allocate without watermarks if the context allows */
2545 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2547 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2548 * the allocation is high priority and these type of
2549 * allocations are system rather than user orientated
2551 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2553 page = __alloc_pages_high_priority(gfp_mask, order,
2554 zonelist, high_zoneidx, nodemask,
2555 preferred_zone, migratetype);
2561 /* Atomic allocations - we can't balance anything */
2564 * All existing users of the deprecated __GFP_NOFAIL are
2565 * blockable, so warn of any new users that actually allow this
2566 * type of allocation to fail.
2568 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2572 /* Avoid recursion of direct reclaim */
2573 if (current->flags & PF_MEMALLOC)
2576 /* Avoid allocations with no watermarks from looping endlessly */
2577 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2581 * Try direct compaction. The first pass is asynchronous. Subsequent
2582 * attempts after direct reclaim are synchronous
2584 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2585 high_zoneidx, nodemask, alloc_flags,
2586 preferred_zone, migratetype,
2587 migration_mode, &contended_compaction,
2588 &deferred_compaction,
2589 &did_some_progress);
2594 * It can become very expensive to allocate transparent hugepages at
2595 * fault, so use asynchronous memory compaction for THP unless it is
2596 * khugepaged trying to collapse.
2598 if (!(gfp_mask & __GFP_NO_KSWAPD) || (current->flags & PF_KTHREAD))
2599 migration_mode = MIGRATE_SYNC_LIGHT;
2602 * If compaction is deferred for high-order allocations, it is because
2603 * sync compaction recently failed. In this is the case and the caller
2604 * requested a movable allocation that does not heavily disrupt the
2605 * system then fail the allocation instead of entering direct reclaim.
2607 if ((deferred_compaction || contended_compaction) &&
2608 (gfp_mask & __GFP_NO_KSWAPD))
2611 /* Try direct reclaim and then allocating */
2612 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2613 zonelist, high_zoneidx,
2615 alloc_flags, preferred_zone,
2616 migratetype, &did_some_progress);
2621 * If we failed to make any progress reclaiming, then we are
2622 * running out of options and have to consider going OOM
2624 if (!did_some_progress) {
2625 if (oom_gfp_allowed(gfp_mask)) {
2626 if (oom_killer_disabled)
2628 /* Coredumps can quickly deplete all memory reserves */
2629 if ((current->flags & PF_DUMPCORE) &&
2630 !(gfp_mask & __GFP_NOFAIL))
2632 page = __alloc_pages_may_oom(gfp_mask, order,
2633 zonelist, high_zoneidx,
2634 nodemask, preferred_zone,
2639 if (!(gfp_mask & __GFP_NOFAIL)) {
2641 * The oom killer is not called for high-order
2642 * allocations that may fail, so if no progress
2643 * is being made, there are no other options and
2644 * retrying is unlikely to help.
2646 if (order > PAGE_ALLOC_COSTLY_ORDER)
2649 * The oom killer is not called for lowmem
2650 * allocations to prevent needlessly killing
2653 if (high_zoneidx < ZONE_NORMAL)
2661 /* Check if we should retry the allocation */
2662 pages_reclaimed += did_some_progress;
2663 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2665 /* Wait for some write requests to complete then retry */
2666 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2670 * High-order allocations do not necessarily loop after
2671 * direct reclaim and reclaim/compaction depends on compaction
2672 * being called after reclaim so call directly if necessary
2674 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2675 high_zoneidx, nodemask, alloc_flags,
2676 preferred_zone, migratetype,
2677 migration_mode, &contended_compaction,
2678 &deferred_compaction,
2679 &did_some_progress);
2685 warn_alloc_failed(gfp_mask, order, NULL);
2688 if (kmemcheck_enabled)
2689 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2695 * This is the 'heart' of the zoned buddy allocator.
2698 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2699 struct zonelist *zonelist, nodemask_t *nodemask)
2701 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2702 struct zone *preferred_zone;
2703 struct page *page = NULL;
2704 int migratetype = allocflags_to_migratetype(gfp_mask);
2705 unsigned int cpuset_mems_cookie;
2706 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2708 gfp_mask &= gfp_allowed_mask;
2710 lockdep_trace_alloc(gfp_mask);
2712 might_sleep_if(gfp_mask & __GFP_WAIT);
2714 if (should_fail_alloc_page(gfp_mask, order))
2718 * Check the zones suitable for the gfp_mask contain at least one
2719 * valid zone. It's possible to have an empty zonelist as a result
2720 * of GFP_THISNODE and a memoryless node
2722 if (unlikely(!zonelist->_zonerefs->zone))
2726 cpuset_mems_cookie = read_mems_allowed_begin();
2728 /* The preferred zone is used for statistics later */
2729 first_zones_zonelist(zonelist, high_zoneidx,
2730 nodemask ? : &cpuset_current_mems_allowed,
2732 if (!preferred_zone)
2736 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2737 alloc_flags |= ALLOC_CMA;
2740 /* First allocation attempt */
2741 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2742 zonelist, high_zoneidx, alloc_flags,
2743 preferred_zone, migratetype);
2744 if (unlikely(!page)) {
2746 * The first pass makes sure allocations are spread
2747 * fairly within the local node. However, the local
2748 * node might have free pages left after the fairness
2749 * batches are exhausted, and remote zones haven't
2750 * even been considered yet. Try once more without
2751 * fairness, and include remote zones now, before
2752 * entering the slowpath and waking kswapd: prefer
2753 * spilling to a remote zone over swapping locally.
2755 if (alloc_flags & ALLOC_FAIR) {
2756 reset_alloc_batches(zonelist, high_zoneidx,
2758 alloc_flags &= ~ALLOC_FAIR;
2762 * Runtime PM, block IO and its error handling path
2763 * can deadlock because I/O on the device might not
2766 gfp_mask = memalloc_noio_flags(gfp_mask);
2767 page = __alloc_pages_slowpath(gfp_mask, order,
2768 zonelist, high_zoneidx, nodemask,
2769 preferred_zone, migratetype);
2772 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2776 * When updating a task's mems_allowed, it is possible to race with
2777 * parallel threads in such a way that an allocation can fail while
2778 * the mask is being updated. If a page allocation is about to fail,
2779 * check if the cpuset changed during allocation and if so, retry.
2781 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2786 EXPORT_SYMBOL(__alloc_pages_nodemask);
2789 * Common helper functions.
2791 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2796 * __get_free_pages() returns a 32-bit address, which cannot represent
2799 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2801 page = alloc_pages(gfp_mask, order);
2804 return (unsigned long) page_address(page);
2806 EXPORT_SYMBOL(__get_free_pages);
2808 unsigned long get_zeroed_page(gfp_t gfp_mask)
2810 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2812 EXPORT_SYMBOL(get_zeroed_page);
2814 void __free_pages(struct page *page, unsigned int order)
2816 if (put_page_testzero(page)) {
2818 free_hot_cold_page(page, 0);
2820 __free_pages_ok(page, order);
2824 EXPORT_SYMBOL(__free_pages);
2826 void free_pages(unsigned long addr, unsigned int order)
2829 VM_BUG_ON(!virt_addr_valid((void *)addr));
2830 __free_pages(virt_to_page((void *)addr), order);
2834 EXPORT_SYMBOL(free_pages);
2837 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2838 * of the current memory cgroup.
2840 * It should be used when the caller would like to use kmalloc, but since the
2841 * allocation is large, it has to fall back to the page allocator.
2843 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2846 struct mem_cgroup *memcg = NULL;
2848 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2850 page = alloc_pages(gfp_mask, order);
2851 memcg_kmem_commit_charge(page, memcg, order);
2855 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2858 struct mem_cgroup *memcg = NULL;
2860 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2862 page = alloc_pages_node(nid, gfp_mask, order);
2863 memcg_kmem_commit_charge(page, memcg, order);
2868 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2871 void __free_kmem_pages(struct page *page, unsigned int order)
2873 memcg_kmem_uncharge_pages(page, order);
2874 __free_pages(page, order);
2877 void free_kmem_pages(unsigned long addr, unsigned int order)
2880 VM_BUG_ON(!virt_addr_valid((void *)addr));
2881 __free_kmem_pages(virt_to_page((void *)addr), order);
2885 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2888 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2889 unsigned long used = addr + PAGE_ALIGN(size);
2891 split_page(virt_to_page((void *)addr), order);
2892 while (used < alloc_end) {
2897 return (void *)addr;
2901 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2902 * @size: the number of bytes to allocate
2903 * @gfp_mask: GFP flags for the allocation
2905 * This function is similar to alloc_pages(), except that it allocates the
2906 * minimum number of pages to satisfy the request. alloc_pages() can only
2907 * allocate memory in power-of-two pages.
2909 * This function is also limited by MAX_ORDER.
2911 * Memory allocated by this function must be released by free_pages_exact().
2913 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2915 unsigned int order = get_order(size);
2918 addr = __get_free_pages(gfp_mask, order);
2919 return make_alloc_exact(addr, order, size);
2921 EXPORT_SYMBOL(alloc_pages_exact);
2924 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2926 * @nid: the preferred node ID where memory should be allocated
2927 * @size: the number of bytes to allocate
2928 * @gfp_mask: GFP flags for the allocation
2930 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2932 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2935 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2937 unsigned order = get_order(size);
2938 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2941 return make_alloc_exact((unsigned long)page_address(p), order, size);
2943 EXPORT_SYMBOL(alloc_pages_exact_nid);
2946 * free_pages_exact - release memory allocated via alloc_pages_exact()
2947 * @virt: the value returned by alloc_pages_exact.
2948 * @size: size of allocation, same value as passed to alloc_pages_exact().
2950 * Release the memory allocated by a previous call to alloc_pages_exact.
2952 void free_pages_exact(void *virt, size_t size)
2954 unsigned long addr = (unsigned long)virt;
2955 unsigned long end = addr + PAGE_ALIGN(size);
2957 while (addr < end) {
2962 EXPORT_SYMBOL(free_pages_exact);
2965 * nr_free_zone_pages - count number of pages beyond high watermark
2966 * @offset: The zone index of the highest zone
2968 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2969 * high watermark within all zones at or below a given zone index. For each
2970 * zone, the number of pages is calculated as:
2971 * managed_pages - high_pages
2973 static unsigned long nr_free_zone_pages(int offset)
2978 /* Just pick one node, since fallback list is circular */
2979 unsigned long sum = 0;
2981 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2983 for_each_zone_zonelist(zone, z, zonelist, offset) {
2984 unsigned long size = zone->managed_pages;
2985 unsigned long high = high_wmark_pages(zone);
2994 * nr_free_buffer_pages - count number of pages beyond high watermark
2996 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2997 * watermark within ZONE_DMA and ZONE_NORMAL.
2999 unsigned long nr_free_buffer_pages(void)
3001 return nr_free_zone_pages(gfp_zone(GFP_USER));
3003 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3006 * nr_free_pagecache_pages - count number of pages beyond high watermark
3008 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3009 * high watermark within all zones.
3011 unsigned long nr_free_pagecache_pages(void)
3013 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3016 static inline void show_node(struct zone *zone)
3018 if (IS_ENABLED(CONFIG_NUMA))
3019 printk("Node %d ", zone_to_nid(zone));
3022 void si_meminfo(struct sysinfo *val)
3024 val->totalram = totalram_pages;
3026 val->freeram = global_page_state(NR_FREE_PAGES);
3027 val->bufferram = nr_blockdev_pages();
3028 val->totalhigh = totalhigh_pages;
3029 val->freehigh = nr_free_highpages();
3030 val->mem_unit = PAGE_SIZE;
3033 EXPORT_SYMBOL(si_meminfo);
3036 void si_meminfo_node(struct sysinfo *val, int nid)
3038 int zone_type; /* needs to be signed */
3039 unsigned long managed_pages = 0;
3040 pg_data_t *pgdat = NODE_DATA(nid);
3042 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3043 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3044 val->totalram = managed_pages;
3045 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3046 #ifdef CONFIG_HIGHMEM
3047 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3048 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3054 val->mem_unit = PAGE_SIZE;
3059 * Determine whether the node should be displayed or not, depending on whether
3060 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3062 bool skip_free_areas_node(unsigned int flags, int nid)
3065 unsigned int cpuset_mems_cookie;
3067 if (!(flags & SHOW_MEM_FILTER_NODES))
3071 cpuset_mems_cookie = read_mems_allowed_begin();
3072 ret = !node_isset(nid, cpuset_current_mems_allowed);
3073 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3078 #define K(x) ((x) << (PAGE_SHIFT-10))
3080 static void show_migration_types(unsigned char type)
3082 static const char types[MIGRATE_TYPES] = {
3083 [MIGRATE_UNMOVABLE] = 'U',
3084 [MIGRATE_RECLAIMABLE] = 'E',
3085 [MIGRATE_MOVABLE] = 'M',
3086 [MIGRATE_RESERVE] = 'R',
3088 [MIGRATE_CMA] = 'C',
3090 #ifdef CONFIG_MEMORY_ISOLATION
3091 [MIGRATE_ISOLATE] = 'I',
3094 char tmp[MIGRATE_TYPES + 1];
3098 for (i = 0; i < MIGRATE_TYPES; i++) {
3099 if (type & (1 << i))
3104 printk("(%s) ", tmp);
3108 * Show free area list (used inside shift_scroll-lock stuff)
3109 * We also calculate the percentage fragmentation. We do this by counting the
3110 * memory on each free list with the exception of the first item on the list.
3111 * Suppresses nodes that are not allowed by current's cpuset if
3112 * SHOW_MEM_FILTER_NODES is passed.
3114 void show_free_areas(unsigned int filter)
3119 for_each_populated_zone(zone) {
3120 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3123 printk("%s per-cpu:\n", zone->name);
3125 for_each_online_cpu(cpu) {
3126 struct per_cpu_pageset *pageset;
3128 pageset = per_cpu_ptr(zone->pageset, cpu);
3130 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3131 cpu, pageset->pcp.high,
3132 pageset->pcp.batch, pageset->pcp.count);
3136 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3137 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3139 " dirty:%lu writeback:%lu unstable:%lu\n"
3140 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3141 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3143 global_page_state(NR_ACTIVE_ANON),
3144 global_page_state(NR_INACTIVE_ANON),
3145 global_page_state(NR_ISOLATED_ANON),
3146 global_page_state(NR_ACTIVE_FILE),
3147 global_page_state(NR_INACTIVE_FILE),
3148 global_page_state(NR_ISOLATED_FILE),
3149 global_page_state(NR_UNEVICTABLE),
3150 global_page_state(NR_FILE_DIRTY),
3151 global_page_state(NR_WRITEBACK),
3152 global_page_state(NR_UNSTABLE_NFS),
3153 global_page_state(NR_FREE_PAGES),
3154 global_page_state(NR_SLAB_RECLAIMABLE),
3155 global_page_state(NR_SLAB_UNRECLAIMABLE),
3156 global_page_state(NR_FILE_MAPPED),
3157 global_page_state(NR_SHMEM),
3158 global_page_state(NR_PAGETABLE),
3159 global_page_state(NR_BOUNCE),
3160 global_page_state(NR_FREE_CMA_PAGES));
3162 for_each_populated_zone(zone) {
3165 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3173 " active_anon:%lukB"
3174 " inactive_anon:%lukB"
3175 " active_file:%lukB"
3176 " inactive_file:%lukB"
3177 " unevictable:%lukB"
3178 " isolated(anon):%lukB"
3179 " isolated(file):%lukB"
3187 " slab_reclaimable:%lukB"
3188 " slab_unreclaimable:%lukB"
3189 " kernel_stack:%lukB"
3194 " writeback_tmp:%lukB"
3195 " pages_scanned:%lu"
3196 " all_unreclaimable? %s"
3199 K(zone_page_state(zone, NR_FREE_PAGES)),
3200 K(min_wmark_pages(zone)),
3201 K(low_wmark_pages(zone)),
3202 K(high_wmark_pages(zone)),
3203 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3204 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3205 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3206 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3207 K(zone_page_state(zone, NR_UNEVICTABLE)),
3208 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3209 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3210 K(zone->present_pages),
3211 K(zone->managed_pages),
3212 K(zone_page_state(zone, NR_MLOCK)),
3213 K(zone_page_state(zone, NR_FILE_DIRTY)),
3214 K(zone_page_state(zone, NR_WRITEBACK)),
3215 K(zone_page_state(zone, NR_FILE_MAPPED)),
3216 K(zone_page_state(zone, NR_SHMEM)),
3217 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3218 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3219 zone_page_state(zone, NR_KERNEL_STACK) *
3221 K(zone_page_state(zone, NR_PAGETABLE)),
3222 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3223 K(zone_page_state(zone, NR_BOUNCE)),
3224 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3225 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3226 zone->pages_scanned,
3227 (!zone_reclaimable(zone) ? "yes" : "no")
3229 printk("lowmem_reserve[]:");
3230 for (i = 0; i < MAX_NR_ZONES; i++)
3231 printk(" %lu", zone->lowmem_reserve[i]);
3235 for_each_populated_zone(zone) {
3236 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3237 unsigned char types[MAX_ORDER];
3239 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3242 printk("%s: ", zone->name);
3244 spin_lock_irqsave(&zone->lock, flags);
3245 for (order = 0; order < MAX_ORDER; order++) {
3246 struct free_area *area = &zone->free_area[order];
3249 nr[order] = area->nr_free;
3250 total += nr[order] << order;
3253 for (type = 0; type < MIGRATE_TYPES; type++) {
3254 if (!list_empty(&area->free_list[type]))
3255 types[order] |= 1 << type;
3258 spin_unlock_irqrestore(&zone->lock, flags);
3259 for (order = 0; order < MAX_ORDER; order++) {
3260 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3262 show_migration_types(types[order]);
3264 printk("= %lukB\n", K(total));
3267 hugetlb_show_meminfo();
3269 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3271 show_swap_cache_info();
3274 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3276 zoneref->zone = zone;
3277 zoneref->zone_idx = zone_idx(zone);
3281 * Builds allocation fallback zone lists.
3283 * Add all populated zones of a node to the zonelist.
3285 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3289 enum zone_type zone_type = MAX_NR_ZONES;
3293 zone = pgdat->node_zones + zone_type;
3294 if (populated_zone(zone)) {
3295 zoneref_set_zone(zone,
3296 &zonelist->_zonerefs[nr_zones++]);
3297 check_highest_zone(zone_type);
3299 } while (zone_type);
3307 * 0 = automatic detection of better ordering.
3308 * 1 = order by ([node] distance, -zonetype)
3309 * 2 = order by (-zonetype, [node] distance)
3311 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3312 * the same zonelist. So only NUMA can configure this param.
3314 #define ZONELIST_ORDER_DEFAULT 0
3315 #define ZONELIST_ORDER_NODE 1
3316 #define ZONELIST_ORDER_ZONE 2
3318 /* zonelist order in the kernel.
3319 * set_zonelist_order() will set this to NODE or ZONE.
3321 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3322 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3326 /* The value user specified ....changed by config */
3327 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3328 /* string for sysctl */
3329 #define NUMA_ZONELIST_ORDER_LEN 16
3330 char numa_zonelist_order[16] = "default";
3333 * interface for configure zonelist ordering.
3334 * command line option "numa_zonelist_order"
3335 * = "[dD]efault - default, automatic configuration.
3336 * = "[nN]ode - order by node locality, then by zone within node
3337 * = "[zZ]one - order by zone, then by locality within zone
3340 static int __parse_numa_zonelist_order(char *s)
3342 if (*s == 'd' || *s == 'D') {
3343 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3344 } else if (*s == 'n' || *s == 'N') {
3345 user_zonelist_order = ZONELIST_ORDER_NODE;
3346 } else if (*s == 'z' || *s == 'Z') {
3347 user_zonelist_order = ZONELIST_ORDER_ZONE;
3350 "Ignoring invalid numa_zonelist_order value: "
3357 static __init int setup_numa_zonelist_order(char *s)
3364 ret = __parse_numa_zonelist_order(s);
3366 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3370 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3373 * sysctl handler for numa_zonelist_order
3375 int numa_zonelist_order_handler(ctl_table *table, int write,
3376 void __user *buffer, size_t *length,
3379 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3381 static DEFINE_MUTEX(zl_order_mutex);
3383 mutex_lock(&zl_order_mutex);
3385 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3389 strcpy(saved_string, (char *)table->data);
3391 ret = proc_dostring(table, write, buffer, length, ppos);
3395 int oldval = user_zonelist_order;
3397 ret = __parse_numa_zonelist_order((char *)table->data);
3400 * bogus value. restore saved string
3402 strncpy((char *)table->data, saved_string,
3403 NUMA_ZONELIST_ORDER_LEN);
3404 user_zonelist_order = oldval;
3405 } else if (oldval != user_zonelist_order) {
3406 mutex_lock(&zonelists_mutex);
3407 build_all_zonelists(NULL, NULL);
3408 mutex_unlock(&zonelists_mutex);
3412 mutex_unlock(&zl_order_mutex);
3417 #define MAX_NODE_LOAD (nr_online_nodes)
3418 static int node_load[MAX_NUMNODES];
3421 * find_next_best_node - find the next node that should appear in a given node's fallback list
3422 * @node: node whose fallback list we're appending
3423 * @used_node_mask: nodemask_t of already used nodes
3425 * We use a number of factors to determine which is the next node that should
3426 * appear on a given node's fallback list. The node should not have appeared
3427 * already in @node's fallback list, and it should be the next closest node
3428 * according to the distance array (which contains arbitrary distance values
3429 * from each node to each node in the system), and should also prefer nodes
3430 * with no CPUs, since presumably they'll have very little allocation pressure
3431 * on them otherwise.
3432 * It returns -1 if no node is found.
3434 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3437 int min_val = INT_MAX;
3438 int best_node = NUMA_NO_NODE;
3439 const struct cpumask *tmp = cpumask_of_node(0);
3441 /* Use the local node if we haven't already */
3442 if (!node_isset(node, *used_node_mask)) {
3443 node_set(node, *used_node_mask);
3447 for_each_node_state(n, N_MEMORY) {
3449 /* Don't want a node to appear more than once */
3450 if (node_isset(n, *used_node_mask))
3453 /* Use the distance array to find the distance */
3454 val = node_distance(node, n);
3456 /* Penalize nodes under us ("prefer the next node") */
3459 /* Give preference to headless and unused nodes */
3460 tmp = cpumask_of_node(n);
3461 if (!cpumask_empty(tmp))
3462 val += PENALTY_FOR_NODE_WITH_CPUS;
3464 /* Slight preference for less loaded node */
3465 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3466 val += node_load[n];
3468 if (val < min_val) {
3475 node_set(best_node, *used_node_mask);
3482 * Build zonelists ordered by node and zones within node.
3483 * This results in maximum locality--normal zone overflows into local
3484 * DMA zone, if any--but risks exhausting DMA zone.
3486 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3489 struct zonelist *zonelist;
3491 zonelist = &pgdat->node_zonelists[0];
3492 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3494 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3495 zonelist->_zonerefs[j].zone = NULL;
3496 zonelist->_zonerefs[j].zone_idx = 0;
3500 * Build gfp_thisnode zonelists
3502 static void build_thisnode_zonelists(pg_data_t *pgdat)
3505 struct zonelist *zonelist;
3507 zonelist = &pgdat->node_zonelists[1];
3508 j = build_zonelists_node(pgdat, zonelist, 0);
3509 zonelist->_zonerefs[j].zone = NULL;
3510 zonelist->_zonerefs[j].zone_idx = 0;
3514 * Build zonelists ordered by zone and nodes within zones.
3515 * This results in conserving DMA zone[s] until all Normal memory is
3516 * exhausted, but results in overflowing to remote node while memory
3517 * may still exist in local DMA zone.
3519 static int node_order[MAX_NUMNODES];
3521 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3524 int zone_type; /* needs to be signed */
3526 struct zonelist *zonelist;
3528 zonelist = &pgdat->node_zonelists[0];
3530 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3531 for (j = 0; j < nr_nodes; j++) {
3532 node = node_order[j];
3533 z = &NODE_DATA(node)->node_zones[zone_type];
3534 if (populated_zone(z)) {
3536 &zonelist->_zonerefs[pos++]);
3537 check_highest_zone(zone_type);
3541 zonelist->_zonerefs[pos].zone = NULL;
3542 zonelist->_zonerefs[pos].zone_idx = 0;
3545 static int default_zonelist_order(void)
3548 unsigned long low_kmem_size, total_size;
3552 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3553 * If they are really small and used heavily, the system can fall
3554 * into OOM very easily.
3555 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3557 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3560 for_each_online_node(nid) {
3561 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3562 z = &NODE_DATA(nid)->node_zones[zone_type];
3563 if (populated_zone(z)) {
3564 if (zone_type < ZONE_NORMAL)
3565 low_kmem_size += z->managed_pages;
3566 total_size += z->managed_pages;
3567 } else if (zone_type == ZONE_NORMAL) {
3569 * If any node has only lowmem, then node order
3570 * is preferred to allow kernel allocations
3571 * locally; otherwise, they can easily infringe
3572 * on other nodes when there is an abundance of
3573 * lowmem available to allocate from.
3575 return ZONELIST_ORDER_NODE;
3579 if (!low_kmem_size || /* there are no DMA area. */
3580 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3581 return ZONELIST_ORDER_NODE;
3583 * look into each node's config.
3584 * If there is a node whose DMA/DMA32 memory is very big area on
3585 * local memory, NODE_ORDER may be suitable.
3587 average_size = total_size /
3588 (nodes_weight(node_states[N_MEMORY]) + 1);
3589 for_each_online_node(nid) {
3592 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3593 z = &NODE_DATA(nid)->node_zones[zone_type];
3594 if (populated_zone(z)) {
3595 if (zone_type < ZONE_NORMAL)
3596 low_kmem_size += z->present_pages;
3597 total_size += z->present_pages;
3600 if (low_kmem_size &&
3601 total_size > average_size && /* ignore small node */
3602 low_kmem_size > total_size * 70/100)
3603 return ZONELIST_ORDER_NODE;
3605 return ZONELIST_ORDER_ZONE;
3608 static void set_zonelist_order(void)
3610 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3611 current_zonelist_order = default_zonelist_order();
3613 current_zonelist_order = user_zonelist_order;
3616 static void build_zonelists(pg_data_t *pgdat)
3620 nodemask_t used_mask;
3621 int local_node, prev_node;
3622 struct zonelist *zonelist;
3623 int order = current_zonelist_order;
3625 /* initialize zonelists */
3626 for (i = 0; i < MAX_ZONELISTS; i++) {
3627 zonelist = pgdat->node_zonelists + i;
3628 zonelist->_zonerefs[0].zone = NULL;
3629 zonelist->_zonerefs[0].zone_idx = 0;
3632 /* NUMA-aware ordering of nodes */
3633 local_node = pgdat->node_id;
3634 load = nr_online_nodes;
3635 prev_node = local_node;
3636 nodes_clear(used_mask);
3638 memset(node_order, 0, sizeof(node_order));
3641 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3643 * We don't want to pressure a particular node.
3644 * So adding penalty to the first node in same
3645 * distance group to make it round-robin.
3647 if (node_distance(local_node, node) !=
3648 node_distance(local_node, prev_node))
3649 node_load[node] = load;
3653 if (order == ZONELIST_ORDER_NODE)
3654 build_zonelists_in_node_order(pgdat, node);
3656 node_order[j++] = node; /* remember order */
3659 if (order == ZONELIST_ORDER_ZONE) {
3660 /* calculate node order -- i.e., DMA last! */
3661 build_zonelists_in_zone_order(pgdat, j);
3664 build_thisnode_zonelists(pgdat);
3667 /* Construct the zonelist performance cache - see further mmzone.h */
3668 static void build_zonelist_cache(pg_data_t *pgdat)
3670 struct zonelist *zonelist;
3671 struct zonelist_cache *zlc;
3674 zonelist = &pgdat->node_zonelists[0];
3675 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3676 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3677 for (z = zonelist->_zonerefs; z->zone; z++)
3678 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3681 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3683 * Return node id of node used for "local" allocations.
3684 * I.e., first node id of first zone in arg node's generic zonelist.
3685 * Used for initializing percpu 'numa_mem', which is used primarily
3686 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3688 int local_memory_node(int node)
3692 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3693 gfp_zone(GFP_KERNEL),
3700 #else /* CONFIG_NUMA */
3702 static void set_zonelist_order(void)
3704 current_zonelist_order = ZONELIST_ORDER_ZONE;
3707 static void build_zonelists(pg_data_t *pgdat)
3709 int node, local_node;
3711 struct zonelist *zonelist;
3713 local_node = pgdat->node_id;
3715 zonelist = &pgdat->node_zonelists[0];
3716 j = build_zonelists_node(pgdat, zonelist, 0);
3719 * Now we build the zonelist so that it contains the zones
3720 * of all the other nodes.
3721 * We don't want to pressure a particular node, so when
3722 * building the zones for node N, we make sure that the
3723 * zones coming right after the local ones are those from
3724 * node N+1 (modulo N)
3726 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3727 if (!node_online(node))
3729 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3731 for (node = 0; node < local_node; node++) {
3732 if (!node_online(node))
3734 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3737 zonelist->_zonerefs[j].zone = NULL;
3738 zonelist->_zonerefs[j].zone_idx = 0;
3741 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3742 static void build_zonelist_cache(pg_data_t *pgdat)
3744 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3747 #endif /* CONFIG_NUMA */
3750 * Boot pageset table. One per cpu which is going to be used for all
3751 * zones and all nodes. The parameters will be set in such a way
3752 * that an item put on a list will immediately be handed over to
3753 * the buddy list. This is safe since pageset manipulation is done
3754 * with interrupts disabled.
3756 * The boot_pagesets must be kept even after bootup is complete for
3757 * unused processors and/or zones. They do play a role for bootstrapping
3758 * hotplugged processors.
3760 * zoneinfo_show() and maybe other functions do
3761 * not check if the processor is online before following the pageset pointer.
3762 * Other parts of the kernel may not check if the zone is available.
3764 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3765 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3766 static void setup_zone_pageset(struct zone *zone);
3769 * Global mutex to protect against size modification of zonelists
3770 * as well as to serialize pageset setup for the new populated zone.
3772 DEFINE_MUTEX(zonelists_mutex);
3774 /* return values int ....just for stop_machine() */
3775 static int __build_all_zonelists(void *data)
3779 pg_data_t *self = data;
3782 memset(node_load, 0, sizeof(node_load));
3785 if (self && !node_online(self->node_id)) {
3786 build_zonelists(self);
3787 build_zonelist_cache(self);
3790 for_each_online_node(nid) {
3791 pg_data_t *pgdat = NODE_DATA(nid);
3793 build_zonelists(pgdat);
3794 build_zonelist_cache(pgdat);
3798 * Initialize the boot_pagesets that are going to be used
3799 * for bootstrapping processors. The real pagesets for
3800 * each zone will be allocated later when the per cpu
3801 * allocator is available.
3803 * boot_pagesets are used also for bootstrapping offline
3804 * cpus if the system is already booted because the pagesets
3805 * are needed to initialize allocators on a specific cpu too.
3806 * F.e. the percpu allocator needs the page allocator which
3807 * needs the percpu allocator in order to allocate its pagesets
3808 * (a chicken-egg dilemma).
3810 for_each_possible_cpu(cpu) {
3811 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3813 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3815 * We now know the "local memory node" for each node--
3816 * i.e., the node of the first zone in the generic zonelist.
3817 * Set up numa_mem percpu variable for on-line cpus. During
3818 * boot, only the boot cpu should be on-line; we'll init the
3819 * secondary cpus' numa_mem as they come on-line. During
3820 * node/memory hotplug, we'll fixup all on-line cpus.
3822 if (cpu_online(cpu))
3823 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3831 * Called with zonelists_mutex held always
3832 * unless system_state == SYSTEM_BOOTING.
3834 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3836 set_zonelist_order();
3838 if (system_state == SYSTEM_BOOTING) {
3839 __build_all_zonelists(NULL);
3840 mminit_verify_zonelist();
3841 cpuset_init_current_mems_allowed();
3843 #ifdef CONFIG_MEMORY_HOTPLUG
3845 setup_zone_pageset(zone);
3847 /* we have to stop all cpus to guarantee there is no user
3849 stop_machine(__build_all_zonelists, pgdat, NULL);
3850 /* cpuset refresh routine should be here */
3852 vm_total_pages = nr_free_pagecache_pages();
3854 * Disable grouping by mobility if the number of pages in the
3855 * system is too low to allow the mechanism to work. It would be
3856 * more accurate, but expensive to check per-zone. This check is
3857 * made on memory-hotadd so a system can start with mobility
3858 * disabled and enable it later
3860 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3861 page_group_by_mobility_disabled = 1;
3863 page_group_by_mobility_disabled = 0;
3865 printk("Built %i zonelists in %s order, mobility grouping %s. "
3866 "Total pages: %ld\n",
3868 zonelist_order_name[current_zonelist_order],
3869 page_group_by_mobility_disabled ? "off" : "on",
3872 printk("Policy zone: %s\n", zone_names[policy_zone]);
3877 * Helper functions to size the waitqueue hash table.
3878 * Essentially these want to choose hash table sizes sufficiently
3879 * large so that collisions trying to wait on pages are rare.
3880 * But in fact, the number of active page waitqueues on typical
3881 * systems is ridiculously low, less than 200. So this is even
3882 * conservative, even though it seems large.
3884 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3885 * waitqueues, i.e. the size of the waitq table given the number of pages.
3887 #define PAGES_PER_WAITQUEUE 256
3889 #ifndef CONFIG_MEMORY_HOTPLUG
3890 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3892 unsigned long size = 1;
3894 pages /= PAGES_PER_WAITQUEUE;
3896 while (size < pages)
3900 * Once we have dozens or even hundreds of threads sleeping
3901 * on IO we've got bigger problems than wait queue collision.
3902 * Limit the size of the wait table to a reasonable size.
3904 size = min(size, 4096UL);
3906 return max(size, 4UL);
3910 * A zone's size might be changed by hot-add, so it is not possible to determine
3911 * a suitable size for its wait_table. So we use the maximum size now.
3913 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3915 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3916 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3917 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3919 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3920 * or more by the traditional way. (See above). It equals:
3922 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3923 * ia64(16K page size) : = ( 8G + 4M)byte.
3924 * powerpc (64K page size) : = (32G +16M)byte.
3926 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3933 * This is an integer logarithm so that shifts can be used later
3934 * to extract the more random high bits from the multiplicative
3935 * hash function before the remainder is taken.
3937 static inline unsigned long wait_table_bits(unsigned long size)
3943 * Check if a pageblock contains reserved pages
3945 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3949 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3950 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3957 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3958 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3959 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3960 * higher will lead to a bigger reserve which will get freed as contiguous
3961 * blocks as reclaim kicks in
3963 static void setup_zone_migrate_reserve(struct zone *zone)
3965 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3967 unsigned long block_migratetype;
3972 * Get the start pfn, end pfn and the number of blocks to reserve
3973 * We have to be careful to be aligned to pageblock_nr_pages to
3974 * make sure that we always check pfn_valid for the first page in
3977 start_pfn = zone->zone_start_pfn;
3978 end_pfn = zone_end_pfn(zone);
3979 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3980 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3984 * Reserve blocks are generally in place to help high-order atomic
3985 * allocations that are short-lived. A min_free_kbytes value that
3986 * would result in more than 2 reserve blocks for atomic allocations
3987 * is assumed to be in place to help anti-fragmentation for the
3988 * future allocation of hugepages at runtime.
3990 reserve = min(2, reserve);
3991 old_reserve = zone->nr_migrate_reserve_block;
3993 /* When memory hot-add, we almost always need to do nothing */
3994 if (reserve == old_reserve)
3996 zone->nr_migrate_reserve_block = reserve;
3998 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3999 if (!pfn_valid(pfn))
4001 page = pfn_to_page(pfn);
4003 /* Watch out for overlapping nodes */
4004 if (page_to_nid(page) != zone_to_nid(zone))
4007 block_migratetype = get_pageblock_migratetype(page);
4009 /* Only test what is necessary when the reserves are not met */
4012 * Blocks with reserved pages will never free, skip
4015 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4016 if (pageblock_is_reserved(pfn, block_end_pfn))
4019 /* If this block is reserved, account for it */
4020 if (block_migratetype == MIGRATE_RESERVE) {
4025 /* Suitable for reserving if this block is movable */
4026 if (block_migratetype == MIGRATE_MOVABLE) {
4027 set_pageblock_migratetype(page,
4029 move_freepages_block(zone, page,
4034 } else if (!old_reserve) {
4036 * At boot time we don't need to scan the whole zone
4037 * for turning off MIGRATE_RESERVE.
4043 * If the reserve is met and this is a previous reserved block,
4046 if (block_migratetype == MIGRATE_RESERVE) {
4047 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4048 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4054 * Initially all pages are reserved - free ones are freed
4055 * up by free_all_bootmem() once the early boot process is
4056 * done. Non-atomic initialization, single-pass.
4058 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4059 unsigned long start_pfn, enum memmap_context context)
4062 unsigned long end_pfn = start_pfn + size;
4066 if (highest_memmap_pfn < end_pfn - 1)
4067 highest_memmap_pfn = end_pfn - 1;
4069 z = &NODE_DATA(nid)->node_zones[zone];
4070 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4072 * There can be holes in boot-time mem_map[]s
4073 * handed to this function. They do not
4074 * exist on hotplugged memory.
4076 if (context == MEMMAP_EARLY) {
4077 if (!early_pfn_valid(pfn))
4079 if (!early_pfn_in_nid(pfn, nid))
4082 page = pfn_to_page(pfn);
4083 set_page_links(page, zone, nid, pfn);
4084 mminit_verify_page_links(page, zone, nid, pfn);
4085 init_page_count(page);
4086 page_mapcount_reset(page);
4087 page_cpupid_reset_last(page);
4088 SetPageReserved(page);
4090 * Mark the block movable so that blocks are reserved for
4091 * movable at startup. This will force kernel allocations
4092 * to reserve their blocks rather than leaking throughout
4093 * the address space during boot when many long-lived
4094 * kernel allocations are made. Later some blocks near
4095 * the start are marked MIGRATE_RESERVE by
4096 * setup_zone_migrate_reserve()
4098 * bitmap is created for zone's valid pfn range. but memmap
4099 * can be created for invalid pages (for alignment)
4100 * check here not to call set_pageblock_migratetype() against
4103 if ((z->zone_start_pfn <= pfn)
4104 && (pfn < zone_end_pfn(z))
4105 && !(pfn & (pageblock_nr_pages - 1)))
4106 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4108 INIT_LIST_HEAD(&page->lru);
4109 #ifdef WANT_PAGE_VIRTUAL
4110 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4111 if (!is_highmem_idx(zone))
4112 set_page_address(page, __va(pfn << PAGE_SHIFT));
4117 static void __meminit zone_init_free_lists(struct zone *zone)
4120 for_each_migratetype_order(order, t) {
4121 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4122 zone->free_area[order].nr_free = 0;
4126 #ifndef __HAVE_ARCH_MEMMAP_INIT
4127 #define memmap_init(size, nid, zone, start_pfn) \
4128 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4131 static int __meminit zone_batchsize(struct zone *zone)
4137 * The per-cpu-pages pools are set to around 1000th of the
4138 * size of the zone. But no more than 1/2 of a meg.
4140 * OK, so we don't know how big the cache is. So guess.
4142 batch = zone->managed_pages / 1024;
4143 if (batch * PAGE_SIZE > 512 * 1024)
4144 batch = (512 * 1024) / PAGE_SIZE;
4145 batch /= 4; /* We effectively *= 4 below */
4150 * Clamp the batch to a 2^n - 1 value. Having a power
4151 * of 2 value was found to be more likely to have
4152 * suboptimal cache aliasing properties in some cases.
4154 * For example if 2 tasks are alternately allocating
4155 * batches of pages, one task can end up with a lot
4156 * of pages of one half of the possible page colors
4157 * and the other with pages of the other colors.
4159 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4164 /* The deferral and batching of frees should be suppressed under NOMMU
4167 * The problem is that NOMMU needs to be able to allocate large chunks
4168 * of contiguous memory as there's no hardware page translation to
4169 * assemble apparent contiguous memory from discontiguous pages.
4171 * Queueing large contiguous runs of pages for batching, however,
4172 * causes the pages to actually be freed in smaller chunks. As there
4173 * can be a significant delay between the individual batches being
4174 * recycled, this leads to the once large chunks of space being
4175 * fragmented and becoming unavailable for high-order allocations.
4182 * pcp->high and pcp->batch values are related and dependent on one another:
4183 * ->batch must never be higher then ->high.
4184 * The following function updates them in a safe manner without read side
4187 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4188 * those fields changing asynchronously (acording the the above rule).
4190 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4191 * outside of boot time (or some other assurance that no concurrent updaters
4194 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4195 unsigned long batch)
4197 /* start with a fail safe value for batch */
4201 /* Update high, then batch, in order */
4208 /* a companion to pageset_set_high() */
4209 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4211 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4214 static void pageset_init(struct per_cpu_pageset *p)
4216 struct per_cpu_pages *pcp;
4219 memset(p, 0, sizeof(*p));
4223 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4224 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4227 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4230 pageset_set_batch(p, batch);
4234 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4235 * to the value high for the pageset p.
4237 static void pageset_set_high(struct per_cpu_pageset *p,
4240 unsigned long batch = max(1UL, high / 4);
4241 if ((high / 4) > (PAGE_SHIFT * 8))
4242 batch = PAGE_SHIFT * 8;
4244 pageset_update(&p->pcp, high, batch);
4247 static void __meminit pageset_set_high_and_batch(struct zone *zone,
4248 struct per_cpu_pageset *pcp)
4250 if (percpu_pagelist_fraction)
4251 pageset_set_high(pcp,
4252 (zone->managed_pages /
4253 percpu_pagelist_fraction));
4255 pageset_set_batch(pcp, zone_batchsize(zone));
4258 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4260 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4263 pageset_set_high_and_batch(zone, pcp);
4266 static void __meminit setup_zone_pageset(struct zone *zone)
4269 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4270 for_each_possible_cpu(cpu)
4271 zone_pageset_init(zone, cpu);
4275 * Allocate per cpu pagesets and initialize them.
4276 * Before this call only boot pagesets were available.
4278 void __init setup_per_cpu_pageset(void)
4282 for_each_populated_zone(zone)
4283 setup_zone_pageset(zone);
4286 static noinline __init_refok
4287 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4293 * The per-page waitqueue mechanism uses hashed waitqueues
4296 zone->wait_table_hash_nr_entries =
4297 wait_table_hash_nr_entries(zone_size_pages);
4298 zone->wait_table_bits =
4299 wait_table_bits(zone->wait_table_hash_nr_entries);
4300 alloc_size = zone->wait_table_hash_nr_entries
4301 * sizeof(wait_queue_head_t);
4303 if (!slab_is_available()) {
4304 zone->wait_table = (wait_queue_head_t *)
4305 memblock_virt_alloc_node_nopanic(
4306 alloc_size, zone->zone_pgdat->node_id);
4309 * This case means that a zone whose size was 0 gets new memory
4310 * via memory hot-add.
4311 * But it may be the case that a new node was hot-added. In
4312 * this case vmalloc() will not be able to use this new node's
4313 * memory - this wait_table must be initialized to use this new
4314 * node itself as well.
4315 * To use this new node's memory, further consideration will be
4318 zone->wait_table = vmalloc(alloc_size);
4320 if (!zone->wait_table)
4323 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4324 init_waitqueue_head(zone->wait_table + i);
4329 static __meminit void zone_pcp_init(struct zone *zone)
4332 * per cpu subsystem is not up at this point. The following code
4333 * relies on the ability of the linker to provide the
4334 * offset of a (static) per cpu variable into the per cpu area.
4336 zone->pageset = &boot_pageset;
4338 if (populated_zone(zone))
4339 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4340 zone->name, zone->present_pages,
4341 zone_batchsize(zone));
4344 int __meminit init_currently_empty_zone(struct zone *zone,
4345 unsigned long zone_start_pfn,
4347 enum memmap_context context)
4349 struct pglist_data *pgdat = zone->zone_pgdat;
4351 ret = zone_wait_table_init(zone, size);
4354 pgdat->nr_zones = zone_idx(zone) + 1;
4356 zone->zone_start_pfn = zone_start_pfn;
4358 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4359 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4361 (unsigned long)zone_idx(zone),
4362 zone_start_pfn, (zone_start_pfn + size));
4364 zone_init_free_lists(zone);
4369 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4370 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4372 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4373 * Architectures may implement their own version but if add_active_range()
4374 * was used and there are no special requirements, this is a convenient
4377 int __meminit __early_pfn_to_nid(unsigned long pfn)
4379 unsigned long start_pfn, end_pfn;
4382 * NOTE: The following SMP-unsafe globals are only used early in boot
4383 * when the kernel is running single-threaded.
4385 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4386 static int __meminitdata last_nid;
4388 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4391 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4393 last_start_pfn = start_pfn;
4394 last_end_pfn = end_pfn;
4400 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4402 int __meminit early_pfn_to_nid(unsigned long pfn)
4406 nid = __early_pfn_to_nid(pfn);
4409 /* just returns 0 */
4413 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4414 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4418 nid = __early_pfn_to_nid(pfn);
4419 if (nid >= 0 && nid != node)
4426 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4427 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4428 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4430 * If an architecture guarantees that all ranges registered with
4431 * add_active_ranges() contain no holes and may be freed, this
4432 * this function may be used instead of calling memblock_free_early_nid()
4435 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4437 unsigned long start_pfn, end_pfn;
4440 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4441 start_pfn = min(start_pfn, max_low_pfn);
4442 end_pfn = min(end_pfn, max_low_pfn);
4444 if (start_pfn < end_pfn)
4445 memblock_free_early_nid(PFN_PHYS(start_pfn),
4446 (end_pfn - start_pfn) << PAGE_SHIFT,
4452 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4453 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4455 * If an architecture guarantees that all ranges registered with
4456 * add_active_ranges() contain no holes and may be freed, this
4457 * function may be used instead of calling memory_present() manually.
4459 void __init sparse_memory_present_with_active_regions(int nid)
4461 unsigned long start_pfn, end_pfn;
4464 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4465 memory_present(this_nid, start_pfn, end_pfn);
4469 * get_pfn_range_for_nid - Return the start and end page frames for a node
4470 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4471 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4472 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4474 * It returns the start and end page frame of a node based on information
4475 * provided by an arch calling add_active_range(). If called for a node
4476 * with no available memory, a warning is printed and the start and end
4479 void __meminit get_pfn_range_for_nid(unsigned int nid,
4480 unsigned long *start_pfn, unsigned long *end_pfn)
4482 unsigned long this_start_pfn, this_end_pfn;
4488 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4489 *start_pfn = min(*start_pfn, this_start_pfn);
4490 *end_pfn = max(*end_pfn, this_end_pfn);
4493 if (*start_pfn == -1UL)
4498 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4499 * assumption is made that zones within a node are ordered in monotonic
4500 * increasing memory addresses so that the "highest" populated zone is used
4502 static void __init find_usable_zone_for_movable(void)
4505 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4506 if (zone_index == ZONE_MOVABLE)
4509 if (arch_zone_highest_possible_pfn[zone_index] >
4510 arch_zone_lowest_possible_pfn[zone_index])
4514 VM_BUG_ON(zone_index == -1);
4515 movable_zone = zone_index;
4519 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4520 * because it is sized independent of architecture. Unlike the other zones,
4521 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4522 * in each node depending on the size of each node and how evenly kernelcore
4523 * is distributed. This helper function adjusts the zone ranges
4524 * provided by the architecture for a given node by using the end of the
4525 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4526 * zones within a node are in order of monotonic increases memory addresses
4528 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4529 unsigned long zone_type,
4530 unsigned long node_start_pfn,
4531 unsigned long node_end_pfn,
4532 unsigned long *zone_start_pfn,
4533 unsigned long *zone_end_pfn)
4535 /* Only adjust if ZONE_MOVABLE is on this node */
4536 if (zone_movable_pfn[nid]) {
4537 /* Size ZONE_MOVABLE */
4538 if (zone_type == ZONE_MOVABLE) {
4539 *zone_start_pfn = zone_movable_pfn[nid];
4540 *zone_end_pfn = min(node_end_pfn,
4541 arch_zone_highest_possible_pfn[movable_zone]);
4543 /* Adjust for ZONE_MOVABLE starting within this range */
4544 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4545 *zone_end_pfn > zone_movable_pfn[nid]) {
4546 *zone_end_pfn = zone_movable_pfn[nid];
4548 /* Check if this whole range is within ZONE_MOVABLE */
4549 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4550 *zone_start_pfn = *zone_end_pfn;
4555 * Return the number of pages a zone spans in a node, including holes
4556 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4558 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4559 unsigned long zone_type,
4560 unsigned long node_start_pfn,
4561 unsigned long node_end_pfn,
4562 unsigned long *ignored)
4564 unsigned long zone_start_pfn, zone_end_pfn;
4566 /* Get the start and end of the zone */
4567 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4568 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4569 adjust_zone_range_for_zone_movable(nid, zone_type,
4570 node_start_pfn, node_end_pfn,
4571 &zone_start_pfn, &zone_end_pfn);
4573 /* Check that this node has pages within the zone's required range */
4574 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4577 /* Move the zone boundaries inside the node if necessary */
4578 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4579 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4581 /* Return the spanned pages */
4582 return zone_end_pfn - zone_start_pfn;
4586 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4587 * then all holes in the requested range will be accounted for.
4589 unsigned long __meminit __absent_pages_in_range(int nid,
4590 unsigned long range_start_pfn,
4591 unsigned long range_end_pfn)
4593 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4594 unsigned long start_pfn, end_pfn;
4597 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4598 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4599 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4600 nr_absent -= end_pfn - start_pfn;
4606 * absent_pages_in_range - Return number of page frames in holes within a range
4607 * @start_pfn: The start PFN to start searching for holes
4608 * @end_pfn: The end PFN to stop searching for holes
4610 * It returns the number of pages frames in memory holes within a range.
4612 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4613 unsigned long end_pfn)
4615 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4618 /* Return the number of page frames in holes in a zone on a node */
4619 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4620 unsigned long zone_type,
4621 unsigned long node_start_pfn,
4622 unsigned long node_end_pfn,
4623 unsigned long *ignored)
4625 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4626 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4627 unsigned long zone_start_pfn, zone_end_pfn;
4629 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4630 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4632 adjust_zone_range_for_zone_movable(nid, zone_type,
4633 node_start_pfn, node_end_pfn,
4634 &zone_start_pfn, &zone_end_pfn);
4635 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4638 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4639 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4640 unsigned long zone_type,
4641 unsigned long node_start_pfn,
4642 unsigned long node_end_pfn,
4643 unsigned long *zones_size)
4645 return zones_size[zone_type];
4648 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4649 unsigned long zone_type,
4650 unsigned long node_start_pfn,
4651 unsigned long node_end_pfn,
4652 unsigned long *zholes_size)
4657 return zholes_size[zone_type];
4660 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4662 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4663 unsigned long node_start_pfn,
4664 unsigned long node_end_pfn,
4665 unsigned long *zones_size,
4666 unsigned long *zholes_size)
4668 unsigned long realtotalpages, totalpages = 0;
4671 for (i = 0; i < MAX_NR_ZONES; i++)
4672 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4676 pgdat->node_spanned_pages = totalpages;
4678 realtotalpages = totalpages;
4679 for (i = 0; i < MAX_NR_ZONES; i++)
4681 zone_absent_pages_in_node(pgdat->node_id, i,
4682 node_start_pfn, node_end_pfn,
4684 pgdat->node_present_pages = realtotalpages;
4685 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4689 #ifndef CONFIG_SPARSEMEM
4691 * Calculate the size of the zone->blockflags rounded to an unsigned long
4692 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4693 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4694 * round what is now in bits to nearest long in bits, then return it in
4697 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4699 unsigned long usemapsize;
4701 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4702 usemapsize = roundup(zonesize, pageblock_nr_pages);
4703 usemapsize = usemapsize >> pageblock_order;
4704 usemapsize *= NR_PAGEBLOCK_BITS;
4705 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4707 return usemapsize / 8;
4710 static void __init setup_usemap(struct pglist_data *pgdat,
4712 unsigned long zone_start_pfn,
4713 unsigned long zonesize)
4715 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4716 zone->pageblock_flags = NULL;
4718 zone->pageblock_flags =
4719 memblock_virt_alloc_node_nopanic(usemapsize,
4723 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4724 unsigned long zone_start_pfn, unsigned long zonesize) {}
4725 #endif /* CONFIG_SPARSEMEM */
4727 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4729 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4730 void __paginginit set_pageblock_order(void)
4734 /* Check that pageblock_nr_pages has not already been setup */
4735 if (pageblock_order)
4738 if (HPAGE_SHIFT > PAGE_SHIFT)
4739 order = HUGETLB_PAGE_ORDER;
4741 order = MAX_ORDER - 1;
4744 * Assume the largest contiguous order of interest is a huge page.
4745 * This value may be variable depending on boot parameters on IA64 and
4748 pageblock_order = order;
4750 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4753 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4754 * is unused as pageblock_order is set at compile-time. See
4755 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4758 void __paginginit set_pageblock_order(void)
4762 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4764 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4765 unsigned long present_pages)
4767 unsigned long pages = spanned_pages;
4770 * Provide a more accurate estimation if there are holes within
4771 * the zone and SPARSEMEM is in use. If there are holes within the
4772 * zone, each populated memory region may cost us one or two extra
4773 * memmap pages due to alignment because memmap pages for each
4774 * populated regions may not naturally algined on page boundary.
4775 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4777 if (spanned_pages > present_pages + (present_pages >> 4) &&
4778 IS_ENABLED(CONFIG_SPARSEMEM))
4779 pages = present_pages;
4781 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4785 * Set up the zone data structures:
4786 * - mark all pages reserved
4787 * - mark all memory queues empty
4788 * - clear the memory bitmaps
4790 * NOTE: pgdat should get zeroed by caller.
4792 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4793 unsigned long node_start_pfn, unsigned long node_end_pfn,
4794 unsigned long *zones_size, unsigned long *zholes_size)
4797 int nid = pgdat->node_id;
4798 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4801 pgdat_resize_init(pgdat);
4802 #ifdef CONFIG_NUMA_BALANCING
4803 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4804 pgdat->numabalancing_migrate_nr_pages = 0;
4805 pgdat->numabalancing_migrate_next_window = jiffies;
4807 init_waitqueue_head(&pgdat->kswapd_wait);
4808 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4809 pgdat_page_cgroup_init(pgdat);
4811 for (j = 0; j < MAX_NR_ZONES; j++) {
4812 struct zone *zone = pgdat->node_zones + j;
4813 unsigned long size, realsize, freesize, memmap_pages;
4815 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4816 node_end_pfn, zones_size);
4817 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4823 * Adjust freesize so that it accounts for how much memory
4824 * is used by this zone for memmap. This affects the watermark
4825 * and per-cpu initialisations
4827 memmap_pages = calc_memmap_size(size, realsize);
4828 if (freesize >= memmap_pages) {
4829 freesize -= memmap_pages;
4832 " %s zone: %lu pages used for memmap\n",
4833 zone_names[j], memmap_pages);
4836 " %s zone: %lu pages exceeds freesize %lu\n",
4837 zone_names[j], memmap_pages, freesize);
4839 /* Account for reserved pages */
4840 if (j == 0 && freesize > dma_reserve) {
4841 freesize -= dma_reserve;
4842 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4843 zone_names[0], dma_reserve);
4846 if (!is_highmem_idx(j))
4847 nr_kernel_pages += freesize;
4848 /* Charge for highmem memmap if there are enough kernel pages */
4849 else if (nr_kernel_pages > memmap_pages * 2)
4850 nr_kernel_pages -= memmap_pages;
4851 nr_all_pages += freesize;
4853 zone->spanned_pages = size;
4854 zone->present_pages = realsize;
4856 * Set an approximate value for lowmem here, it will be adjusted
4857 * when the bootmem allocator frees pages into the buddy system.
4858 * And all highmem pages will be managed by the buddy system.
4860 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4863 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4865 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4867 zone->name = zone_names[j];
4868 spin_lock_init(&zone->lock);
4869 spin_lock_init(&zone->lru_lock);
4870 zone_seqlock_init(zone);
4871 zone->zone_pgdat = pgdat;
4872 zone_pcp_init(zone);
4874 /* For bootup, initialized properly in watermark setup */
4875 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4877 lruvec_init(&zone->lruvec);
4881 set_pageblock_order();
4882 setup_usemap(pgdat, zone, zone_start_pfn, size);
4883 ret = init_currently_empty_zone(zone, zone_start_pfn,
4884 size, MEMMAP_EARLY);
4886 memmap_init(size, nid, j, zone_start_pfn);
4887 zone_start_pfn += size;
4891 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4893 /* Skip empty nodes */
4894 if (!pgdat->node_spanned_pages)
4897 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4898 /* ia64 gets its own node_mem_map, before this, without bootmem */
4899 if (!pgdat->node_mem_map) {
4900 unsigned long size, start, end;
4904 * The zone's endpoints aren't required to be MAX_ORDER
4905 * aligned but the node_mem_map endpoints must be in order
4906 * for the buddy allocator to function correctly.
4908 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4909 end = pgdat_end_pfn(pgdat);
4910 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4911 size = (end - start) * sizeof(struct page);
4912 map = alloc_remap(pgdat->node_id, size);
4914 map = memblock_virt_alloc_node_nopanic(size,
4916 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4918 #ifndef CONFIG_NEED_MULTIPLE_NODES
4920 * With no DISCONTIG, the global mem_map is just set as node 0's
4922 if (pgdat == NODE_DATA(0)) {
4923 mem_map = NODE_DATA(0)->node_mem_map;
4924 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4925 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4926 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4927 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4930 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4933 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4934 unsigned long node_start_pfn, unsigned long *zholes_size)
4936 pg_data_t *pgdat = NODE_DATA(nid);
4937 unsigned long start_pfn = 0;
4938 unsigned long end_pfn = 0;
4940 /* pg_data_t should be reset to zero when it's allocated */
4941 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4943 pgdat->node_id = nid;
4944 pgdat->node_start_pfn = node_start_pfn;
4945 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4946 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4948 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4949 zones_size, zholes_size);
4951 alloc_node_mem_map(pgdat);
4952 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4953 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4954 nid, (unsigned long)pgdat,
4955 (unsigned long)pgdat->node_mem_map);
4958 free_area_init_core(pgdat, start_pfn, end_pfn,
4959 zones_size, zholes_size);
4962 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4964 #if MAX_NUMNODES > 1
4966 * Figure out the number of possible node ids.
4968 void __init setup_nr_node_ids(void)
4971 unsigned int highest = 0;
4973 for_each_node_mask(node, node_possible_map)
4975 nr_node_ids = highest + 1;
4980 * node_map_pfn_alignment - determine the maximum internode alignment
4982 * This function should be called after node map is populated and sorted.
4983 * It calculates the maximum power of two alignment which can distinguish
4986 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4987 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4988 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4989 * shifted, 1GiB is enough and this function will indicate so.
4991 * This is used to test whether pfn -> nid mapping of the chosen memory
4992 * model has fine enough granularity to avoid incorrect mapping for the
4993 * populated node map.
4995 * Returns the determined alignment in pfn's. 0 if there is no alignment
4996 * requirement (single node).
4998 unsigned long __init node_map_pfn_alignment(void)
5000 unsigned long accl_mask = 0, last_end = 0;
5001 unsigned long start, end, mask;
5005 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5006 if (!start || last_nid < 0 || last_nid == nid) {
5013 * Start with a mask granular enough to pin-point to the
5014 * start pfn and tick off bits one-by-one until it becomes
5015 * too coarse to separate the current node from the last.
5017 mask = ~((1 << __ffs(start)) - 1);
5018 while (mask && last_end <= (start & (mask << 1)))
5021 /* accumulate all internode masks */
5025 /* convert mask to number of pages */
5026 return ~accl_mask + 1;
5029 /* Find the lowest pfn for a node */
5030 static unsigned long __init find_min_pfn_for_node(int nid)
5032 unsigned long min_pfn = ULONG_MAX;
5033 unsigned long start_pfn;
5036 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5037 min_pfn = min(min_pfn, start_pfn);
5039 if (min_pfn == ULONG_MAX) {
5041 "Could not find start_pfn for node %d\n", nid);
5049 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5051 * It returns the minimum PFN based on information provided via
5052 * add_active_range().
5054 unsigned long __init find_min_pfn_with_active_regions(void)
5056 return find_min_pfn_for_node(MAX_NUMNODES);
5060 * early_calculate_totalpages()
5061 * Sum pages in active regions for movable zone.
5062 * Populate N_MEMORY for calculating usable_nodes.
5064 static unsigned long __init early_calculate_totalpages(void)
5066 unsigned long totalpages = 0;
5067 unsigned long start_pfn, end_pfn;
5070 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5071 unsigned long pages = end_pfn - start_pfn;
5073 totalpages += pages;
5075 node_set_state(nid, N_MEMORY);
5081 * Find the PFN the Movable zone begins in each node. Kernel memory
5082 * is spread evenly between nodes as long as the nodes have enough
5083 * memory. When they don't, some nodes will have more kernelcore than
5086 static void __init find_zone_movable_pfns_for_nodes(void)
5089 unsigned long usable_startpfn;
5090 unsigned long kernelcore_node, kernelcore_remaining;
5091 /* save the state before borrow the nodemask */
5092 nodemask_t saved_node_state = node_states[N_MEMORY];
5093 unsigned long totalpages = early_calculate_totalpages();
5094 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5095 struct memblock_region *r;
5097 /* Need to find movable_zone earlier when movable_node is specified. */
5098 find_usable_zone_for_movable();
5101 * If movable_node is specified, ignore kernelcore and movablecore
5104 if (movable_node_is_enabled()) {
5105 for_each_memblock(memory, r) {
5106 if (!memblock_is_hotpluggable(r))
5111 usable_startpfn = PFN_DOWN(r->base);
5112 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5113 min(usable_startpfn, zone_movable_pfn[nid]) :
5121 * If movablecore=nn[KMG] was specified, calculate what size of
5122 * kernelcore that corresponds so that memory usable for
5123 * any allocation type is evenly spread. If both kernelcore
5124 * and movablecore are specified, then the value of kernelcore
5125 * will be used for required_kernelcore if it's greater than
5126 * what movablecore would have allowed.
5128 if (required_movablecore) {
5129 unsigned long corepages;
5132 * Round-up so that ZONE_MOVABLE is at least as large as what
5133 * was requested by the user
5135 required_movablecore =
5136 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5137 corepages = totalpages - required_movablecore;
5139 required_kernelcore = max(required_kernelcore, corepages);
5142 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5143 if (!required_kernelcore)
5146 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5147 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5150 /* Spread kernelcore memory as evenly as possible throughout nodes */
5151 kernelcore_node = required_kernelcore / usable_nodes;
5152 for_each_node_state(nid, N_MEMORY) {
5153 unsigned long start_pfn, end_pfn;
5156 * Recalculate kernelcore_node if the division per node
5157 * now exceeds what is necessary to satisfy the requested
5158 * amount of memory for the kernel
5160 if (required_kernelcore < kernelcore_node)
5161 kernelcore_node = required_kernelcore / usable_nodes;
5164 * As the map is walked, we track how much memory is usable
5165 * by the kernel using kernelcore_remaining. When it is
5166 * 0, the rest of the node is usable by ZONE_MOVABLE
5168 kernelcore_remaining = kernelcore_node;
5170 /* Go through each range of PFNs within this node */
5171 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5172 unsigned long size_pages;
5174 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5175 if (start_pfn >= end_pfn)
5178 /* Account for what is only usable for kernelcore */
5179 if (start_pfn < usable_startpfn) {
5180 unsigned long kernel_pages;
5181 kernel_pages = min(end_pfn, usable_startpfn)
5184 kernelcore_remaining -= min(kernel_pages,
5185 kernelcore_remaining);
5186 required_kernelcore -= min(kernel_pages,
5187 required_kernelcore);
5189 /* Continue if range is now fully accounted */
5190 if (end_pfn <= usable_startpfn) {
5193 * Push zone_movable_pfn to the end so
5194 * that if we have to rebalance
5195 * kernelcore across nodes, we will
5196 * not double account here
5198 zone_movable_pfn[nid] = end_pfn;
5201 start_pfn = usable_startpfn;
5205 * The usable PFN range for ZONE_MOVABLE is from
5206 * start_pfn->end_pfn. Calculate size_pages as the
5207 * number of pages used as kernelcore
5209 size_pages = end_pfn - start_pfn;
5210 if (size_pages > kernelcore_remaining)
5211 size_pages = kernelcore_remaining;
5212 zone_movable_pfn[nid] = start_pfn + size_pages;
5215 * Some kernelcore has been met, update counts and
5216 * break if the kernelcore for this node has been
5219 required_kernelcore -= min(required_kernelcore,
5221 kernelcore_remaining -= size_pages;
5222 if (!kernelcore_remaining)
5228 * If there is still required_kernelcore, we do another pass with one
5229 * less node in the count. This will push zone_movable_pfn[nid] further
5230 * along on the nodes that still have memory until kernelcore is
5234 if (usable_nodes && required_kernelcore > usable_nodes)
5238 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5239 for (nid = 0; nid < MAX_NUMNODES; nid++)
5240 zone_movable_pfn[nid] =
5241 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5244 /* restore the node_state */
5245 node_states[N_MEMORY] = saved_node_state;
5248 /* Any regular or high memory on that node ? */
5249 static void check_for_memory(pg_data_t *pgdat, int nid)
5251 enum zone_type zone_type;
5253 if (N_MEMORY == N_NORMAL_MEMORY)
5256 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5257 struct zone *zone = &pgdat->node_zones[zone_type];
5258 if (populated_zone(zone)) {
5259 node_set_state(nid, N_HIGH_MEMORY);
5260 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5261 zone_type <= ZONE_NORMAL)
5262 node_set_state(nid, N_NORMAL_MEMORY);
5269 * free_area_init_nodes - Initialise all pg_data_t and zone data
5270 * @max_zone_pfn: an array of max PFNs for each zone
5272 * This will call free_area_init_node() for each active node in the system.
5273 * Using the page ranges provided by add_active_range(), the size of each
5274 * zone in each node and their holes is calculated. If the maximum PFN
5275 * between two adjacent zones match, it is assumed that the zone is empty.
5276 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5277 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5278 * starts where the previous one ended. For example, ZONE_DMA32 starts
5279 * at arch_max_dma_pfn.
5281 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5283 unsigned long start_pfn, end_pfn;
5286 /* Record where the zone boundaries are */
5287 memset(arch_zone_lowest_possible_pfn, 0,
5288 sizeof(arch_zone_lowest_possible_pfn));
5289 memset(arch_zone_highest_possible_pfn, 0,
5290 sizeof(arch_zone_highest_possible_pfn));
5291 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5292 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5293 for (i = 1; i < MAX_NR_ZONES; i++) {
5294 if (i == ZONE_MOVABLE)
5296 arch_zone_lowest_possible_pfn[i] =
5297 arch_zone_highest_possible_pfn[i-1];
5298 arch_zone_highest_possible_pfn[i] =
5299 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5301 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5302 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5304 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5305 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5306 find_zone_movable_pfns_for_nodes();
5308 /* Print out the zone ranges */
5309 printk("Zone ranges:\n");
5310 for (i = 0; i < MAX_NR_ZONES; i++) {
5311 if (i == ZONE_MOVABLE)
5313 printk(KERN_CONT " %-8s ", zone_names[i]);
5314 if (arch_zone_lowest_possible_pfn[i] ==
5315 arch_zone_highest_possible_pfn[i])
5316 printk(KERN_CONT "empty\n");
5318 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5319 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5320 (arch_zone_highest_possible_pfn[i]
5321 << PAGE_SHIFT) - 1);
5324 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5325 printk("Movable zone start for each node\n");
5326 for (i = 0; i < MAX_NUMNODES; i++) {
5327 if (zone_movable_pfn[i])
5328 printk(" Node %d: %#010lx\n", i,
5329 zone_movable_pfn[i] << PAGE_SHIFT);
5332 /* Print out the early node map */
5333 printk("Early memory node ranges\n");
5334 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5335 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5336 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5338 /* Initialise every node */
5339 mminit_verify_pageflags_layout();
5340 setup_nr_node_ids();
5341 for_each_online_node(nid) {
5342 pg_data_t *pgdat = NODE_DATA(nid);
5343 free_area_init_node(nid, NULL,
5344 find_min_pfn_for_node(nid), NULL);
5346 /* Any memory on that node */
5347 if (pgdat->node_present_pages)
5348 node_set_state(nid, N_MEMORY);
5349 check_for_memory(pgdat, nid);
5353 static int __init cmdline_parse_core(char *p, unsigned long *core)
5355 unsigned long long coremem;
5359 coremem = memparse(p, &p);
5360 *core = coremem >> PAGE_SHIFT;
5362 /* Paranoid check that UL is enough for the coremem value */
5363 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5369 * kernelcore=size sets the amount of memory for use for allocations that
5370 * cannot be reclaimed or migrated.
5372 static int __init cmdline_parse_kernelcore(char *p)
5374 return cmdline_parse_core(p, &required_kernelcore);
5378 * movablecore=size sets the amount of memory for use for allocations that
5379 * can be reclaimed or migrated.
5381 static int __init cmdline_parse_movablecore(char *p)
5383 return cmdline_parse_core(p, &required_movablecore);
5386 early_param("kernelcore", cmdline_parse_kernelcore);
5387 early_param("movablecore", cmdline_parse_movablecore);
5389 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5391 void adjust_managed_page_count(struct page *page, long count)
5393 spin_lock(&managed_page_count_lock);
5394 page_zone(page)->managed_pages += count;
5395 totalram_pages += count;
5396 #ifdef CONFIG_HIGHMEM
5397 if (PageHighMem(page))
5398 totalhigh_pages += count;
5400 spin_unlock(&managed_page_count_lock);
5402 EXPORT_SYMBOL(adjust_managed_page_count);
5404 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5407 unsigned long pages = 0;
5409 start = (void *)PAGE_ALIGN((unsigned long)start);
5410 end = (void *)((unsigned long)end & PAGE_MASK);
5411 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5412 if ((unsigned int)poison <= 0xFF)
5413 memset(pos, poison, PAGE_SIZE);
5414 free_reserved_page(virt_to_page(pos));
5418 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5419 s, pages << (PAGE_SHIFT - 10), start, end);
5423 EXPORT_SYMBOL(free_reserved_area);
5425 #ifdef CONFIG_HIGHMEM
5426 void free_highmem_page(struct page *page)
5428 __free_reserved_page(page);
5430 page_zone(page)->managed_pages++;
5436 void __init mem_init_print_info(const char *str)
5438 unsigned long physpages, codesize, datasize, rosize, bss_size;
5439 unsigned long init_code_size, init_data_size;
5441 physpages = get_num_physpages();
5442 codesize = _etext - _stext;
5443 datasize = _edata - _sdata;
5444 rosize = __end_rodata - __start_rodata;
5445 bss_size = __bss_stop - __bss_start;
5446 init_data_size = __init_end - __init_begin;
5447 init_code_size = _einittext - _sinittext;
5450 * Detect special cases and adjust section sizes accordingly:
5451 * 1) .init.* may be embedded into .data sections
5452 * 2) .init.text.* may be out of [__init_begin, __init_end],
5453 * please refer to arch/tile/kernel/vmlinux.lds.S.
5454 * 3) .rodata.* may be embedded into .text or .data sections.
5456 #define adj_init_size(start, end, size, pos, adj) \
5458 if (start <= pos && pos < end && size > adj) \
5462 adj_init_size(__init_begin, __init_end, init_data_size,
5463 _sinittext, init_code_size);
5464 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5465 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5466 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5467 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5469 #undef adj_init_size
5471 printk("Memory: %luK/%luK available "
5472 "(%luK kernel code, %luK rwdata, %luK rodata, "
5473 "%luK init, %luK bss, %luK reserved"
5474 #ifdef CONFIG_HIGHMEM
5478 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5479 codesize >> 10, datasize >> 10, rosize >> 10,
5480 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5481 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5482 #ifdef CONFIG_HIGHMEM
5483 totalhigh_pages << (PAGE_SHIFT-10),
5485 str ? ", " : "", str ? str : "");
5489 * set_dma_reserve - set the specified number of pages reserved in the first zone
5490 * @new_dma_reserve: The number of pages to mark reserved
5492 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5493 * In the DMA zone, a significant percentage may be consumed by kernel image
5494 * and other unfreeable allocations which can skew the watermarks badly. This
5495 * function may optionally be used to account for unfreeable pages in the
5496 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5497 * smaller per-cpu batchsize.
5499 void __init set_dma_reserve(unsigned long new_dma_reserve)
5501 dma_reserve = new_dma_reserve;
5504 void __init free_area_init(unsigned long *zones_size)
5506 free_area_init_node(0, zones_size,
5507 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5510 static int page_alloc_cpu_notify(struct notifier_block *self,
5511 unsigned long action, void *hcpu)
5513 int cpu = (unsigned long)hcpu;
5515 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5516 lru_add_drain_cpu(cpu);
5520 * Spill the event counters of the dead processor
5521 * into the current processors event counters.
5522 * This artificially elevates the count of the current
5525 vm_events_fold_cpu(cpu);
5528 * Zero the differential counters of the dead processor
5529 * so that the vm statistics are consistent.
5531 * This is only okay since the processor is dead and cannot
5532 * race with what we are doing.
5534 cpu_vm_stats_fold(cpu);
5539 void __init page_alloc_init(void)
5541 hotcpu_notifier(page_alloc_cpu_notify, 0);
5545 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5546 * or min_free_kbytes changes.
5548 static void calculate_totalreserve_pages(void)
5550 struct pglist_data *pgdat;
5551 unsigned long reserve_pages = 0;
5552 enum zone_type i, j;
5554 for_each_online_pgdat(pgdat) {
5555 for (i = 0; i < MAX_NR_ZONES; i++) {
5556 struct zone *zone = pgdat->node_zones + i;
5557 unsigned long max = 0;
5559 /* Find valid and maximum lowmem_reserve in the zone */
5560 for (j = i; j < MAX_NR_ZONES; j++) {
5561 if (zone->lowmem_reserve[j] > max)
5562 max = zone->lowmem_reserve[j];
5565 /* we treat the high watermark as reserved pages. */
5566 max += high_wmark_pages(zone);
5568 if (max > zone->managed_pages)
5569 max = zone->managed_pages;
5570 reserve_pages += max;
5572 * Lowmem reserves are not available to
5573 * GFP_HIGHUSER page cache allocations and
5574 * kswapd tries to balance zones to their high
5575 * watermark. As a result, neither should be
5576 * regarded as dirtyable memory, to prevent a
5577 * situation where reclaim has to clean pages
5578 * in order to balance the zones.
5580 zone->dirty_balance_reserve = max;
5583 dirty_balance_reserve = reserve_pages;
5584 totalreserve_pages = reserve_pages;
5588 * setup_per_zone_lowmem_reserve - called whenever
5589 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5590 * has a correct pages reserved value, so an adequate number of
5591 * pages are left in the zone after a successful __alloc_pages().
5593 static void setup_per_zone_lowmem_reserve(void)
5595 struct pglist_data *pgdat;
5596 enum zone_type j, idx;
5598 for_each_online_pgdat(pgdat) {
5599 for (j = 0; j < MAX_NR_ZONES; j++) {
5600 struct zone *zone = pgdat->node_zones + j;
5601 unsigned long managed_pages = zone->managed_pages;
5603 zone->lowmem_reserve[j] = 0;
5607 struct zone *lower_zone;
5611 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5612 sysctl_lowmem_reserve_ratio[idx] = 1;
5614 lower_zone = pgdat->node_zones + idx;
5615 lower_zone->lowmem_reserve[j] = managed_pages /
5616 sysctl_lowmem_reserve_ratio[idx];
5617 managed_pages += lower_zone->managed_pages;
5622 /* update totalreserve_pages */
5623 calculate_totalreserve_pages();
5626 static void __setup_per_zone_wmarks(void)
5628 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5629 unsigned long lowmem_pages = 0;
5631 unsigned long flags;
5633 /* Calculate total number of !ZONE_HIGHMEM pages */
5634 for_each_zone(zone) {
5635 if (!is_highmem(zone))
5636 lowmem_pages += zone->managed_pages;
5639 for_each_zone(zone) {
5642 spin_lock_irqsave(&zone->lock, flags);
5643 tmp = (u64)pages_min * zone->managed_pages;
5644 do_div(tmp, lowmem_pages);
5645 if (is_highmem(zone)) {
5647 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5648 * need highmem pages, so cap pages_min to a small
5651 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5652 * deltas controls asynch page reclaim, and so should
5653 * not be capped for highmem.
5655 unsigned long min_pages;
5657 min_pages = zone->managed_pages / 1024;
5658 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5659 zone->watermark[WMARK_MIN] = min_pages;
5662 * If it's a lowmem zone, reserve a number of pages
5663 * proportionate to the zone's size.
5665 zone->watermark[WMARK_MIN] = tmp;
5668 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5669 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5671 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5672 high_wmark_pages(zone) -
5673 low_wmark_pages(zone) -
5674 zone_page_state(zone, NR_ALLOC_BATCH));
5676 setup_zone_migrate_reserve(zone);
5677 spin_unlock_irqrestore(&zone->lock, flags);
5680 /* update totalreserve_pages */
5681 calculate_totalreserve_pages();
5685 * setup_per_zone_wmarks - called when min_free_kbytes changes
5686 * or when memory is hot-{added|removed}
5688 * Ensures that the watermark[min,low,high] values for each zone are set
5689 * correctly with respect to min_free_kbytes.
5691 void setup_per_zone_wmarks(void)
5693 mutex_lock(&zonelists_mutex);
5694 __setup_per_zone_wmarks();
5695 mutex_unlock(&zonelists_mutex);
5699 * The inactive anon list should be small enough that the VM never has to
5700 * do too much work, but large enough that each inactive page has a chance
5701 * to be referenced again before it is swapped out.
5703 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5704 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5705 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5706 * the anonymous pages are kept on the inactive list.
5709 * memory ratio inactive anon
5710 * -------------------------------------
5719 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5721 unsigned int gb, ratio;
5723 /* Zone size in gigabytes */
5724 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5726 ratio = int_sqrt(10 * gb);
5730 zone->inactive_ratio = ratio;
5733 static void __meminit setup_per_zone_inactive_ratio(void)
5738 calculate_zone_inactive_ratio(zone);
5742 * Initialise min_free_kbytes.
5744 * For small machines we want it small (128k min). For large machines
5745 * we want it large (64MB max). But it is not linear, because network
5746 * bandwidth does not increase linearly with machine size. We use
5748 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5749 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5765 int __meminit init_per_zone_wmark_min(void)
5767 unsigned long lowmem_kbytes;
5768 int new_min_free_kbytes;
5770 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5771 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5773 if (new_min_free_kbytes > user_min_free_kbytes) {
5774 min_free_kbytes = new_min_free_kbytes;
5775 if (min_free_kbytes < 128)
5776 min_free_kbytes = 128;
5777 if (min_free_kbytes > 65536)
5778 min_free_kbytes = 65536;
5780 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5781 new_min_free_kbytes, user_min_free_kbytes);
5783 setup_per_zone_wmarks();
5784 refresh_zone_stat_thresholds();
5785 setup_per_zone_lowmem_reserve();
5786 setup_per_zone_inactive_ratio();
5789 module_init(init_per_zone_wmark_min)
5792 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5793 * that we can call two helper functions whenever min_free_kbytes
5796 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5797 void __user *buffer, size_t *length, loff_t *ppos)
5801 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5806 user_min_free_kbytes = min_free_kbytes;
5807 setup_per_zone_wmarks();
5813 int sysctl_min_unmapped_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_unmapped_pages = (zone->managed_pages *
5825 sysctl_min_unmapped_ratio) / 100;
5829 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5830 void __user *buffer, size_t *length, loff_t *ppos)
5835 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5840 zone->min_slab_pages = (zone->managed_pages *
5841 sysctl_min_slab_ratio) / 100;
5847 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5848 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5849 * whenever sysctl_lowmem_reserve_ratio changes.
5851 * The reserve ratio obviously has absolutely no relation with the
5852 * minimum watermarks. The lowmem reserve ratio can only make sense
5853 * if in function of the boot time zone sizes.
5855 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5856 void __user *buffer, size_t *length, loff_t *ppos)
5858 proc_dointvec_minmax(table, write, buffer, length, ppos);
5859 setup_per_zone_lowmem_reserve();
5864 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5865 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5866 * pagelist can have before it gets flushed back to buddy allocator.
5868 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5869 void __user *buffer, size_t *length, loff_t *ppos)
5875 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5876 if (!write || (ret < 0))
5879 mutex_lock(&pcp_batch_high_lock);
5880 for_each_populated_zone(zone) {
5882 high = zone->managed_pages / percpu_pagelist_fraction;
5883 for_each_possible_cpu(cpu)
5884 pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5887 mutex_unlock(&pcp_batch_high_lock);
5891 int hashdist = HASHDIST_DEFAULT;
5894 static int __init set_hashdist(char *str)
5898 hashdist = simple_strtoul(str, &str, 0);
5901 __setup("hashdist=", set_hashdist);
5905 * allocate a large system hash table from bootmem
5906 * - it is assumed that the hash table must contain an exact power-of-2
5907 * quantity of entries
5908 * - limit is the number of hash buckets, not the total allocation size
5910 void *__init alloc_large_system_hash(const char *tablename,
5911 unsigned long bucketsize,
5912 unsigned long numentries,
5915 unsigned int *_hash_shift,
5916 unsigned int *_hash_mask,
5917 unsigned long low_limit,
5918 unsigned long high_limit)
5920 unsigned long long max = high_limit;
5921 unsigned long log2qty, size;
5924 /* allow the kernel cmdline to have a say */
5926 /* round applicable memory size up to nearest megabyte */
5927 numentries = nr_kernel_pages;
5929 /* It isn't necessary when PAGE_SIZE >= 1MB */
5930 if (PAGE_SHIFT < 20)
5931 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5933 /* limit to 1 bucket per 2^scale bytes of low memory */
5934 if (scale > PAGE_SHIFT)
5935 numentries >>= (scale - PAGE_SHIFT);
5937 numentries <<= (PAGE_SHIFT - scale);
5939 /* Make sure we've got at least a 0-order allocation.. */
5940 if (unlikely(flags & HASH_SMALL)) {
5941 /* Makes no sense without HASH_EARLY */
5942 WARN_ON(!(flags & HASH_EARLY));
5943 if (!(numentries >> *_hash_shift)) {
5944 numentries = 1UL << *_hash_shift;
5945 BUG_ON(!numentries);
5947 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5948 numentries = PAGE_SIZE / bucketsize;
5950 numentries = roundup_pow_of_two(numentries);
5952 /* limit allocation size to 1/16 total memory by default */
5954 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5955 do_div(max, bucketsize);
5957 max = min(max, 0x80000000ULL);
5959 if (numentries < low_limit)
5960 numentries = low_limit;
5961 if (numentries > max)
5964 log2qty = ilog2(numentries);
5967 size = bucketsize << log2qty;
5968 if (flags & HASH_EARLY)
5969 table = memblock_virt_alloc_nopanic(size, 0);
5971 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5974 * If bucketsize is not a power-of-two, we may free
5975 * some pages at the end of hash table which
5976 * alloc_pages_exact() automatically does
5978 if (get_order(size) < MAX_ORDER) {
5979 table = alloc_pages_exact(size, GFP_ATOMIC);
5980 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5983 } while (!table && size > PAGE_SIZE && --log2qty);
5986 panic("Failed to allocate %s hash table\n", tablename);
5988 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5991 ilog2(size) - PAGE_SHIFT,
5995 *_hash_shift = log2qty;
5997 *_hash_mask = (1 << log2qty) - 1;
6002 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6003 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6006 #ifdef CONFIG_SPARSEMEM
6007 return __pfn_to_section(pfn)->pageblock_flags;
6009 return zone->pageblock_flags;
6010 #endif /* CONFIG_SPARSEMEM */
6013 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6015 #ifdef CONFIG_SPARSEMEM
6016 pfn &= (PAGES_PER_SECTION-1);
6017 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6019 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6020 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6021 #endif /* CONFIG_SPARSEMEM */
6025 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
6026 * @page: The page within the block of interest
6027 * @start_bitidx: The first bit of interest to retrieve
6028 * @end_bitidx: The last bit of interest
6029 * returns pageblock_bits flags
6031 unsigned long get_pageblock_flags_mask(struct page *page,
6032 unsigned long end_bitidx,
6036 unsigned long *bitmap;
6037 unsigned long pfn, bitidx, word_bitidx;
6040 zone = page_zone(page);
6041 pfn = page_to_pfn(page);
6042 bitmap = get_pageblock_bitmap(zone, pfn);
6043 bitidx = pfn_to_bitidx(zone, pfn);
6044 word_bitidx = bitidx / BITS_PER_LONG;
6045 bitidx &= (BITS_PER_LONG-1);
6047 word = bitmap[word_bitidx];
6048 bitidx += end_bitidx;
6049 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6053 * set_pageblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6054 * @page: The page within the block of interest
6055 * @start_bitidx: The first bit of interest
6056 * @end_bitidx: The last bit of interest
6057 * @flags: The flags to set
6059 void set_pageblock_flags_mask(struct page *page, unsigned long flags,
6060 unsigned long end_bitidx,
6064 unsigned long *bitmap;
6065 unsigned long pfn, bitidx, word_bitidx;
6066 unsigned long old_word, word;
6068 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6070 zone = page_zone(page);
6071 pfn = page_to_pfn(page);
6072 bitmap = get_pageblock_bitmap(zone, pfn);
6073 bitidx = pfn_to_bitidx(zone, pfn);
6074 word_bitidx = bitidx / BITS_PER_LONG;
6075 bitidx &= (BITS_PER_LONG-1);
6077 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6079 bitidx += end_bitidx;
6080 mask <<= (BITS_PER_LONG - bitidx - 1);
6081 flags <<= (BITS_PER_LONG - bitidx - 1);
6083 word = ACCESS_ONCE(bitmap[word_bitidx]);
6085 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6086 if (word == old_word)
6093 * This function checks whether pageblock includes unmovable pages or not.
6094 * If @count is not zero, it is okay to include less @count unmovable pages
6096 * PageLRU check without isolation or lru_lock could race so that
6097 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6098 * expect this function should be exact.
6100 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6101 bool skip_hwpoisoned_pages)
6103 unsigned long pfn, iter, found;
6107 * For avoiding noise data, lru_add_drain_all() should be called
6108 * If ZONE_MOVABLE, the zone never contains unmovable pages
6110 if (zone_idx(zone) == ZONE_MOVABLE)
6112 mt = get_pageblock_migratetype(page);
6113 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6116 pfn = page_to_pfn(page);
6117 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6118 unsigned long check = pfn + iter;
6120 if (!pfn_valid_within(check))
6123 page = pfn_to_page(check);
6126 * Hugepages are not in LRU lists, but they're movable.
6127 * We need not scan over tail pages bacause we don't
6128 * handle each tail page individually in migration.
6130 if (PageHuge(page)) {
6131 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6136 * We can't use page_count without pin a page
6137 * because another CPU can free compound page.
6138 * This check already skips compound tails of THP
6139 * because their page->_count is zero at all time.
6141 if (!atomic_read(&page->_count)) {
6142 if (PageBuddy(page))
6143 iter += (1 << page_order(page)) - 1;
6148 * The HWPoisoned page may be not in buddy system, and
6149 * page_count() is not 0.
6151 if (skip_hwpoisoned_pages && PageHWPoison(page))
6157 * If there are RECLAIMABLE pages, we need to check it.
6158 * But now, memory offline itself doesn't call shrink_slab()
6159 * and it still to be fixed.
6162 * If the page is not RAM, page_count()should be 0.
6163 * we don't need more check. This is an _used_ not-movable page.
6165 * The problematic thing here is PG_reserved pages. PG_reserved
6166 * is set to both of a memory hole page and a _used_ kernel
6175 bool is_pageblock_removable_nolock(struct page *page)
6181 * We have to be careful here because we are iterating over memory
6182 * sections which are not zone aware so we might end up outside of
6183 * the zone but still within the section.
6184 * We have to take care about the node as well. If the node is offline
6185 * its NODE_DATA will be NULL - see page_zone.
6187 if (!node_online(page_to_nid(page)))
6190 zone = page_zone(page);
6191 pfn = page_to_pfn(page);
6192 if (!zone_spans_pfn(zone, pfn))
6195 return !has_unmovable_pages(zone, page, 0, true);
6200 static unsigned long pfn_max_align_down(unsigned long pfn)
6202 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6203 pageblock_nr_pages) - 1);
6206 static unsigned long pfn_max_align_up(unsigned long pfn)
6208 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6209 pageblock_nr_pages));
6212 /* [start, end) must belong to a single zone. */
6213 static int __alloc_contig_migrate_range(struct compact_control *cc,
6214 unsigned long start, unsigned long end)
6216 /* This function is based on compact_zone() from compaction.c. */
6217 unsigned long nr_reclaimed;
6218 unsigned long pfn = start;
6219 unsigned int tries = 0;
6224 while (pfn < end || !list_empty(&cc->migratepages)) {
6225 if (fatal_signal_pending(current)) {
6230 if (list_empty(&cc->migratepages)) {
6231 cc->nr_migratepages = 0;
6232 pfn = isolate_migratepages_range(cc->zone, cc,
6239 } else if (++tries == 5) {
6240 ret = ret < 0 ? ret : -EBUSY;
6244 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6246 cc->nr_migratepages -= nr_reclaimed;
6248 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6249 NULL, 0, cc->mode, MR_CMA);
6252 putback_movable_pages(&cc->migratepages);
6259 * alloc_contig_range() -- tries to allocate given range of pages
6260 * @start: start PFN to allocate
6261 * @end: one-past-the-last PFN to allocate
6262 * @migratetype: migratetype of the underlaying pageblocks (either
6263 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6264 * in range must have the same migratetype and it must
6265 * be either of the two.
6267 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6268 * aligned, however it's the caller's responsibility to guarantee that
6269 * we are the only thread that changes migrate type of pageblocks the
6272 * The PFN range must belong to a single zone.
6274 * Returns zero on success or negative error code. On success all
6275 * pages which PFN is in [start, end) are allocated for the caller and
6276 * need to be freed with free_contig_range().
6278 int alloc_contig_range(unsigned long start, unsigned long end,
6279 unsigned migratetype)
6281 unsigned long outer_start, outer_end;
6284 struct compact_control cc = {
6285 .nr_migratepages = 0,
6287 .zone = page_zone(pfn_to_page(start)),
6288 .mode = MIGRATE_SYNC,
6289 .ignore_skip_hint = true,
6291 INIT_LIST_HEAD(&cc.migratepages);
6294 * What we do here is we mark all pageblocks in range as
6295 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6296 * have different sizes, and due to the way page allocator
6297 * work, we align the range to biggest of the two pages so
6298 * that page allocator won't try to merge buddies from
6299 * different pageblocks and change MIGRATE_ISOLATE to some
6300 * other migration type.
6302 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6303 * migrate the pages from an unaligned range (ie. pages that
6304 * we are interested in). This will put all the pages in
6305 * range back to page allocator as MIGRATE_ISOLATE.
6307 * When this is done, we take the pages in range from page
6308 * allocator removing them from the buddy system. This way
6309 * page allocator will never consider using them.
6311 * This lets us mark the pageblocks back as
6312 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6313 * aligned range but not in the unaligned, original range are
6314 * put back to page allocator so that buddy can use them.
6317 ret = start_isolate_page_range(pfn_max_align_down(start),
6318 pfn_max_align_up(end), migratetype,
6323 ret = __alloc_contig_migrate_range(&cc, start, end);
6328 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6329 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6330 * more, all pages in [start, end) are free in page allocator.
6331 * What we are going to do is to allocate all pages from
6332 * [start, end) (that is remove them from page allocator).
6334 * The only problem is that pages at the beginning and at the
6335 * end of interesting range may be not aligned with pages that
6336 * page allocator holds, ie. they can be part of higher order
6337 * pages. Because of this, we reserve the bigger range and
6338 * once this is done free the pages we are not interested in.
6340 * We don't have to hold zone->lock here because the pages are
6341 * isolated thus they won't get removed from buddy.
6344 lru_add_drain_all();
6348 outer_start = start;
6349 while (!PageBuddy(pfn_to_page(outer_start))) {
6350 if (++order >= MAX_ORDER) {
6354 outer_start &= ~0UL << order;
6357 /* Make sure the range is really isolated. */
6358 if (test_pages_isolated(outer_start, end, false)) {
6359 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6366 /* Grab isolated pages from freelists. */
6367 outer_end = isolate_freepages_range(&cc, outer_start, end);
6373 /* Free head and tail (if any) */
6374 if (start != outer_start)
6375 free_contig_range(outer_start, start - outer_start);
6376 if (end != outer_end)
6377 free_contig_range(end, outer_end - end);
6380 undo_isolate_page_range(pfn_max_align_down(start),
6381 pfn_max_align_up(end), migratetype);
6385 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6387 unsigned int count = 0;
6389 for (; nr_pages--; pfn++) {
6390 struct page *page = pfn_to_page(pfn);
6392 count += page_count(page) != 1;
6395 WARN(count != 0, "%d pages are still in use!\n", count);
6399 #ifdef CONFIG_MEMORY_HOTPLUG
6401 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6402 * page high values need to be recalulated.
6404 void __meminit zone_pcp_update(struct zone *zone)
6407 mutex_lock(&pcp_batch_high_lock);
6408 for_each_possible_cpu(cpu)
6409 pageset_set_high_and_batch(zone,
6410 per_cpu_ptr(zone->pageset, cpu));
6411 mutex_unlock(&pcp_batch_high_lock);
6415 void zone_pcp_reset(struct zone *zone)
6417 unsigned long flags;
6419 struct per_cpu_pageset *pset;
6421 /* avoid races with drain_pages() */
6422 local_irq_save(flags);
6423 if (zone->pageset != &boot_pageset) {
6424 for_each_online_cpu(cpu) {
6425 pset = per_cpu_ptr(zone->pageset, cpu);
6426 drain_zonestat(zone, pset);
6428 free_percpu(zone->pageset);
6429 zone->pageset = &boot_pageset;
6431 local_irq_restore(flags);
6434 #ifdef CONFIG_MEMORY_HOTREMOVE
6436 * All pages in the range must be isolated before calling this.
6439 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6445 unsigned long flags;
6446 /* find the first valid pfn */
6447 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6452 zone = page_zone(pfn_to_page(pfn));
6453 spin_lock_irqsave(&zone->lock, flags);
6455 while (pfn < end_pfn) {
6456 if (!pfn_valid(pfn)) {
6460 page = pfn_to_page(pfn);
6462 * The HWPoisoned page may be not in buddy system, and
6463 * page_count() is not 0.
6465 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6467 SetPageReserved(page);
6471 BUG_ON(page_count(page));
6472 BUG_ON(!PageBuddy(page));
6473 order = page_order(page);
6474 #ifdef CONFIG_DEBUG_VM
6475 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6476 pfn, 1 << order, end_pfn);
6478 list_del(&page->lru);
6479 rmv_page_order(page);
6480 zone->free_area[order].nr_free--;
6481 for (i = 0; i < (1 << order); i++)
6482 SetPageReserved((page+i));
6483 pfn += (1 << order);
6485 spin_unlock_irqrestore(&zone->lock, flags);
6489 #ifdef CONFIG_MEMORY_FAILURE
6490 bool is_free_buddy_page(struct page *page)
6492 struct zone *zone = page_zone(page);
6493 unsigned long pfn = page_to_pfn(page);
6494 unsigned long flags;
6497 spin_lock_irqsave(&zone->lock, flags);
6498 for (order = 0; order < MAX_ORDER; order++) {
6499 struct page *page_head = page - (pfn & ((1 << order) - 1));
6501 if (PageBuddy(page_head) && page_order(page_head) >= order)
6504 spin_unlock_irqrestore(&zone->lock, flags);
6506 return order < MAX_ORDER;
6510 static const struct trace_print_flags pageflag_names[] = {
6511 {1UL << PG_locked, "locked" },
6512 {1UL << PG_error, "error" },
6513 {1UL << PG_referenced, "referenced" },
6514 {1UL << PG_uptodate, "uptodate" },
6515 {1UL << PG_dirty, "dirty" },
6516 {1UL << PG_lru, "lru" },
6517 {1UL << PG_active, "active" },
6518 {1UL << PG_slab, "slab" },
6519 {1UL << PG_owner_priv_1, "owner_priv_1" },
6520 {1UL << PG_arch_1, "arch_1" },
6521 {1UL << PG_reserved, "reserved" },
6522 {1UL << PG_private, "private" },
6523 {1UL << PG_private_2, "private_2" },
6524 {1UL << PG_writeback, "writeback" },
6525 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6526 {1UL << PG_head, "head" },
6527 {1UL << PG_tail, "tail" },
6529 {1UL << PG_compound, "compound" },
6531 {1UL << PG_swapcache, "swapcache" },
6532 {1UL << PG_mappedtodisk, "mappedtodisk" },
6533 {1UL << PG_reclaim, "reclaim" },
6534 {1UL << PG_swapbacked, "swapbacked" },
6535 {1UL << PG_unevictable, "unevictable" },
6537 {1UL << PG_mlocked, "mlocked" },
6539 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6540 {1UL << PG_uncached, "uncached" },
6542 #ifdef CONFIG_MEMORY_FAILURE
6543 {1UL << PG_hwpoison, "hwpoison" },
6545 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6546 {1UL << PG_compound_lock, "compound_lock" },
6550 static void dump_page_flags(unsigned long flags)
6552 const char *delim = "";
6556 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6558 printk(KERN_ALERT "page flags: %#lx(", flags);
6560 /* remove zone id */
6561 flags &= (1UL << NR_PAGEFLAGS) - 1;
6563 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6565 mask = pageflag_names[i].mask;
6566 if ((flags & mask) != mask)
6570 printk("%s%s", delim, pageflag_names[i].name);
6574 /* check for left over flags */
6576 printk("%s%#lx", delim, flags);
6581 void dump_page_badflags(struct page *page, const char *reason,
6582 unsigned long badflags)
6585 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6586 page, atomic_read(&page->_count), page_mapcount(page),
6587 page->mapping, page->index);
6588 dump_page_flags(page->flags);
6590 pr_alert("page dumped because: %s\n", reason);
6591 if (page->flags & badflags) {
6592 pr_alert("bad because of flags:\n");
6593 dump_page_flags(page->flags & badflags);
6595 mem_cgroup_print_bad_page(page);
6598 void dump_page(struct page *page, const char *reason)
6600 dump_page_badflags(page, reason, 0);
6602 EXPORT_SYMBOL(dump_page);