2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node);
76 EXPORT_PER_CPU_SYMBOL(numa_node);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
88 int _node_numa_mem_[MAX_NUMNODES];
92 * Array of node states.
94 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
95 [N_POSSIBLE] = NODE_MASK_ALL,
96 [N_ONLINE] = { { [0] = 1UL } },
98 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 [N_HIGH_MEMORY] = { { [0] = 1UL } },
102 #ifdef CONFIG_MOVABLE_NODE
103 [N_MEMORY] = { { [0] = 1UL } },
105 [N_CPU] = { { [0] = 1UL } },
108 EXPORT_SYMBOL(node_states);
110 /* Protect totalram_pages and zone->managed_pages */
111 static DEFINE_SPINLOCK(managed_page_count_lock);
113 unsigned long totalram_pages __read_mostly;
114 unsigned long totalreserve_pages __read_mostly;
116 * When calculating the number of globally allowed dirty pages, there
117 * is a certain number of per-zone reserves that should not be
118 * considered dirtyable memory. This is the sum of those reserves
119 * over all existing zones that contribute dirtyable memory.
121 unsigned long dirty_balance_reserve __read_mostly;
123 int percpu_pagelist_fraction;
124 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
126 #ifdef CONFIG_PM_SLEEP
128 * The following functions are used by the suspend/hibernate code to temporarily
129 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
130 * while devices are suspended. To avoid races with the suspend/hibernate code,
131 * they should always be called with pm_mutex held (gfp_allowed_mask also should
132 * only be modified with pm_mutex held, unless the suspend/hibernate code is
133 * guaranteed not to run in parallel with that modification).
136 static gfp_t saved_gfp_mask;
138 void pm_restore_gfp_mask(void)
140 WARN_ON(!mutex_is_locked(&pm_mutex));
141 if (saved_gfp_mask) {
142 gfp_allowed_mask = saved_gfp_mask;
147 void pm_restrict_gfp_mask(void)
149 WARN_ON(!mutex_is_locked(&pm_mutex));
150 WARN_ON(saved_gfp_mask);
151 saved_gfp_mask = gfp_allowed_mask;
152 gfp_allowed_mask &= ~GFP_IOFS;
155 bool pm_suspended_storage(void)
157 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
161 #endif /* CONFIG_PM_SLEEP */
163 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
164 int pageblock_order __read_mostly;
167 static void __free_pages_ok(struct page *page, unsigned int order);
170 * results with 256, 32 in the lowmem_reserve sysctl:
171 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
172 * 1G machine -> (16M dma, 784M normal, 224M high)
173 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
174 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
175 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
177 * TBD: should special case ZONE_DMA32 machines here - in those we normally
178 * don't need any ZONE_NORMAL reservation
180 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
181 #ifdef CONFIG_ZONE_DMA
184 #ifdef CONFIG_ZONE_DMA32
187 #ifdef CONFIG_HIGHMEM
193 EXPORT_SYMBOL(totalram_pages);
195 static char * const zone_names[MAX_NR_ZONES] = {
196 #ifdef CONFIG_ZONE_DMA
199 #ifdef CONFIG_ZONE_DMA32
203 #ifdef CONFIG_HIGHMEM
209 int min_free_kbytes = 1024;
210 int user_min_free_kbytes = -1;
212 static unsigned long __meminitdata nr_kernel_pages;
213 static unsigned long __meminitdata nr_all_pages;
214 static unsigned long __meminitdata dma_reserve;
216 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
217 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
218 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
219 static unsigned long __initdata required_kernelcore;
220 static unsigned long __initdata required_movablecore;
221 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
223 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
225 EXPORT_SYMBOL(movable_zone);
226 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
229 int nr_node_ids __read_mostly = MAX_NUMNODES;
230 int nr_online_nodes __read_mostly = 1;
231 EXPORT_SYMBOL(nr_node_ids);
232 EXPORT_SYMBOL(nr_online_nodes);
235 int page_group_by_mobility_disabled __read_mostly;
237 void set_pageblock_migratetype(struct page *page, int migratetype)
239 if (unlikely(page_group_by_mobility_disabled &&
240 migratetype < MIGRATE_PCPTYPES))
241 migratetype = MIGRATE_UNMOVABLE;
243 set_pageblock_flags_group(page, (unsigned long)migratetype,
244 PB_migrate, PB_migrate_end);
247 bool oom_killer_disabled __read_mostly;
249 #ifdef CONFIG_DEBUG_VM
250 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
254 unsigned long pfn = page_to_pfn(page);
255 unsigned long sp, start_pfn;
258 seq = zone_span_seqbegin(zone);
259 start_pfn = zone->zone_start_pfn;
260 sp = zone->spanned_pages;
261 if (!zone_spans_pfn(zone, pfn))
263 } while (zone_span_seqretry(zone, seq));
266 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
267 pfn, zone_to_nid(zone), zone->name,
268 start_pfn, start_pfn + sp);
273 static int page_is_consistent(struct zone *zone, struct page *page)
275 if (!pfn_valid_within(page_to_pfn(page)))
277 if (zone != page_zone(page))
283 * Temporary debugging check for pages not lying within a given zone.
285 static int bad_range(struct zone *zone, struct page *page)
287 if (page_outside_zone_boundaries(zone, page))
289 if (!page_is_consistent(zone, page))
295 static inline int bad_range(struct zone *zone, struct page *page)
301 static void bad_page(struct page *page, const char *reason,
302 unsigned long bad_flags)
304 static unsigned long resume;
305 static unsigned long nr_shown;
306 static unsigned long nr_unshown;
308 /* Don't complain about poisoned pages */
309 if (PageHWPoison(page)) {
310 page_mapcount_reset(page); /* remove PageBuddy */
315 * Allow a burst of 60 reports, then keep quiet for that minute;
316 * or allow a steady drip of one report per second.
318 if (nr_shown == 60) {
319 if (time_before(jiffies, resume)) {
325 "BUG: Bad page state: %lu messages suppressed\n",
332 resume = jiffies + 60 * HZ;
334 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
335 current->comm, page_to_pfn(page));
336 dump_page_badflags(page, reason, bad_flags);
341 /* Leave bad fields for debug, except PageBuddy could make trouble */
342 page_mapcount_reset(page); /* remove PageBuddy */
343 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
347 * Higher-order pages are called "compound pages". They are structured thusly:
349 * The first PAGE_SIZE page is called the "head page".
351 * The remaining PAGE_SIZE pages are called "tail pages".
353 * All pages have PG_compound set. All tail pages have their ->first_page
354 * pointing at the head page.
356 * The first tail page's ->lru.next holds the address of the compound page's
357 * put_page() function. Its ->lru.prev holds the order of allocation.
358 * This usage means that zero-order pages may not be compound.
361 static void free_compound_page(struct page *page)
363 __free_pages_ok(page, compound_order(page));
366 void prep_compound_page(struct page *page, unsigned long order)
369 int nr_pages = 1 << order;
371 set_compound_page_dtor(page, free_compound_page);
372 set_compound_order(page, order);
374 for (i = 1; i < nr_pages; i++) {
375 struct page *p = page + i;
376 set_page_count(p, 0);
377 p->first_page = page;
378 /* Make sure p->first_page is always valid for PageTail() */
384 /* update __split_huge_page_refcount if you change this function */
385 static int destroy_compound_page(struct page *page, unsigned long order)
388 int nr_pages = 1 << order;
391 if (unlikely(compound_order(page) != order)) {
392 bad_page(page, "wrong compound order", 0);
396 __ClearPageHead(page);
398 for (i = 1; i < nr_pages; i++) {
399 struct page *p = page + i;
401 if (unlikely(!PageTail(p))) {
402 bad_page(page, "PageTail not set", 0);
404 } else if (unlikely(p->first_page != page)) {
405 bad_page(page, "first_page not consistent", 0);
414 static inline void prep_zero_page(struct page *page, unsigned int order,
420 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
421 * and __GFP_HIGHMEM from hard or soft interrupt context.
423 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
424 for (i = 0; i < (1 << order); i++)
425 clear_highpage(page + i);
428 #ifdef CONFIG_DEBUG_PAGEALLOC
429 unsigned int _debug_guardpage_minorder;
431 static int __init debug_guardpage_minorder_setup(char *buf)
435 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
436 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
439 _debug_guardpage_minorder = res;
440 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
443 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
445 static inline void set_page_guard_flag(struct page *page)
447 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
450 static inline void clear_page_guard_flag(struct page *page)
452 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
455 static inline void set_page_guard_flag(struct page *page) { }
456 static inline void clear_page_guard_flag(struct page *page) { }
459 static inline void set_page_order(struct page *page, unsigned int order)
461 set_page_private(page, order);
462 __SetPageBuddy(page);
465 static inline void rmv_page_order(struct page *page)
467 __ClearPageBuddy(page);
468 set_page_private(page, 0);
472 * Locate the struct page for both the matching buddy in our
473 * pair (buddy1) and the combined O(n+1) page they form (page).
475 * 1) Any buddy B1 will have an order O twin B2 which satisfies
476 * the following equation:
478 * For example, if the starting buddy (buddy2) is #8 its order
480 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
482 * 2) Any buddy B will have an order O+1 parent P which
483 * satisfies the following equation:
486 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
488 static inline unsigned long
489 __find_buddy_index(unsigned long page_idx, unsigned int order)
491 return page_idx ^ (1 << order);
495 * This function checks whether a page is free && is the buddy
496 * we can do coalesce a page and its buddy if
497 * (a) the buddy is not in a hole &&
498 * (b) the buddy is in the buddy system &&
499 * (c) a page and its buddy have the same order &&
500 * (d) a page and its buddy are in the same zone.
502 * For recording whether a page is in the buddy system, we set ->_mapcount
503 * PAGE_BUDDY_MAPCOUNT_VALUE.
504 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
505 * serialized by zone->lock.
507 * For recording page's order, we use page_private(page).
509 static inline int page_is_buddy(struct page *page, struct page *buddy,
512 if (!pfn_valid_within(page_to_pfn(buddy)))
515 if (page_is_guard(buddy) && page_order(buddy) == order) {
516 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
518 if (page_zone_id(page) != page_zone_id(buddy))
524 if (PageBuddy(buddy) && page_order(buddy) == order) {
525 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
528 * zone check is done late to avoid uselessly
529 * calculating zone/node ids for pages that could
532 if (page_zone_id(page) != page_zone_id(buddy))
541 * Freeing function for a buddy system allocator.
543 * The concept of a buddy system is to maintain direct-mapped table
544 * (containing bit values) for memory blocks of various "orders".
545 * The bottom level table contains the map for the smallest allocatable
546 * units of memory (here, pages), and each level above it describes
547 * pairs of units from the levels below, hence, "buddies".
548 * At a high level, all that happens here is marking the table entry
549 * at the bottom level available, and propagating the changes upward
550 * as necessary, plus some accounting needed to play nicely with other
551 * parts of the VM system.
552 * At each level, we keep a list of pages, which are heads of continuous
553 * free pages of length of (1 << order) and marked with _mapcount
554 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
556 * So when we are allocating or freeing one, we can derive the state of the
557 * other. That is, if we allocate a small block, and both were
558 * free, the remainder of the region must be split into blocks.
559 * If a block is freed, and its buddy is also free, then this
560 * triggers coalescing into a block of larger size.
565 static inline void __free_one_page(struct page *page,
567 struct zone *zone, unsigned int order,
570 unsigned long page_idx;
571 unsigned long combined_idx;
572 unsigned long uninitialized_var(buddy_idx);
575 VM_BUG_ON(!zone_is_initialized(zone));
577 if (unlikely(PageCompound(page)))
578 if (unlikely(destroy_compound_page(page, order)))
581 VM_BUG_ON(migratetype == -1);
583 page_idx = pfn & ((1 << MAX_ORDER) - 1);
585 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
586 VM_BUG_ON_PAGE(bad_range(zone, page), page);
588 while (order < MAX_ORDER-1) {
589 buddy_idx = __find_buddy_index(page_idx, order);
590 buddy = page + (buddy_idx - page_idx);
591 if (!page_is_buddy(page, buddy, order))
594 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
595 * merge with it and move up one order.
597 if (page_is_guard(buddy)) {
598 clear_page_guard_flag(buddy);
599 set_page_private(page, 0);
600 __mod_zone_freepage_state(zone, 1 << order,
603 list_del(&buddy->lru);
604 zone->free_area[order].nr_free--;
605 rmv_page_order(buddy);
607 combined_idx = buddy_idx & page_idx;
608 page = page + (combined_idx - page_idx);
609 page_idx = combined_idx;
612 set_page_order(page, order);
615 * If this is not the largest possible page, check if the buddy
616 * of the next-highest order is free. If it is, it's possible
617 * that pages are being freed that will coalesce soon. In case,
618 * that is happening, add the free page to the tail of the list
619 * so it's less likely to be used soon and more likely to be merged
620 * as a higher order page
622 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
623 struct page *higher_page, *higher_buddy;
624 combined_idx = buddy_idx & page_idx;
625 higher_page = page + (combined_idx - page_idx);
626 buddy_idx = __find_buddy_index(combined_idx, order + 1);
627 higher_buddy = higher_page + (buddy_idx - combined_idx);
628 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
629 list_add_tail(&page->lru,
630 &zone->free_area[order].free_list[migratetype]);
635 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
637 zone->free_area[order].nr_free++;
640 static inline int free_pages_check(struct page *page)
642 const char *bad_reason = NULL;
643 unsigned long bad_flags = 0;
645 if (unlikely(page_mapcount(page)))
646 bad_reason = "nonzero mapcount";
647 if (unlikely(page->mapping != NULL))
648 bad_reason = "non-NULL mapping";
649 if (unlikely(atomic_read(&page->_count) != 0))
650 bad_reason = "nonzero _count";
651 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
652 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
653 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
655 if (unlikely(mem_cgroup_bad_page_check(page)))
656 bad_reason = "cgroup check failed";
657 if (unlikely(bad_reason)) {
658 bad_page(page, bad_reason, bad_flags);
661 page_cpupid_reset_last(page);
662 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
663 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
668 * Frees a number of pages from the PCP lists
669 * Assumes all pages on list are in same zone, and of same order.
670 * count is the number of pages to free.
672 * If the zone was previously in an "all pages pinned" state then look to
673 * see if this freeing clears that state.
675 * And clear the zone's pages_scanned counter, to hold off the "all pages are
676 * pinned" detection logic.
678 static void free_pcppages_bulk(struct zone *zone, int count,
679 struct per_cpu_pages *pcp)
684 unsigned long nr_scanned;
686 spin_lock(&zone->lock);
687 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
689 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
693 struct list_head *list;
696 * Remove pages from lists in a round-robin fashion. A
697 * batch_free count is maintained that is incremented when an
698 * empty list is encountered. This is so more pages are freed
699 * off fuller lists instead of spinning excessively around empty
704 if (++migratetype == MIGRATE_PCPTYPES)
706 list = &pcp->lists[migratetype];
707 } while (list_empty(list));
709 /* This is the only non-empty list. Free them all. */
710 if (batch_free == MIGRATE_PCPTYPES)
711 batch_free = to_free;
714 int mt; /* migratetype of the to-be-freed page */
716 page = list_entry(list->prev, struct page, lru);
717 /* must delete as __free_one_page list manipulates */
718 list_del(&page->lru);
719 mt = get_freepage_migratetype(page);
720 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
721 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
722 trace_mm_page_pcpu_drain(page, 0, mt);
723 if (likely(!is_migrate_isolate_page(page))) {
724 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
725 if (is_migrate_cma(mt))
726 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
728 } while (--to_free && --batch_free && !list_empty(list));
730 spin_unlock(&zone->lock);
733 static void free_one_page(struct zone *zone,
734 struct page *page, unsigned long pfn,
738 unsigned long nr_scanned;
739 spin_lock(&zone->lock);
740 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
742 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
744 __free_one_page(page, pfn, zone, order, migratetype);
745 if (unlikely(!is_migrate_isolate(migratetype)))
746 __mod_zone_freepage_state(zone, 1 << order, migratetype);
747 spin_unlock(&zone->lock);
750 static bool free_pages_prepare(struct page *page, unsigned int order)
755 trace_mm_page_free(page, order);
756 kmemcheck_free_shadow(page, order);
759 page->mapping = NULL;
760 for (i = 0; i < (1 << order); i++)
761 bad += free_pages_check(page + i);
765 if (!PageHighMem(page)) {
766 debug_check_no_locks_freed(page_address(page),
768 debug_check_no_obj_freed(page_address(page),
771 arch_free_page(page, order);
772 kernel_map_pages(page, 1 << order, 0);
777 static void __free_pages_ok(struct page *page, unsigned int order)
781 unsigned long pfn = page_to_pfn(page);
783 if (!free_pages_prepare(page, order))
786 migratetype = get_pfnblock_migratetype(page, pfn);
787 local_irq_save(flags);
788 __count_vm_events(PGFREE, 1 << order);
789 set_freepage_migratetype(page, migratetype);
790 free_one_page(page_zone(page), page, pfn, order, migratetype);
791 local_irq_restore(flags);
794 void __init __free_pages_bootmem(struct page *page, unsigned int order)
796 unsigned int nr_pages = 1 << order;
797 struct page *p = page;
801 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
803 __ClearPageReserved(p);
804 set_page_count(p, 0);
806 __ClearPageReserved(p);
807 set_page_count(p, 0);
809 page_zone(page)->managed_pages += nr_pages;
810 set_page_refcounted(page);
811 __free_pages(page, order);
815 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
816 void __init init_cma_reserved_pageblock(struct page *page)
818 unsigned i = pageblock_nr_pages;
819 struct page *p = page;
822 __ClearPageReserved(p);
823 set_page_count(p, 0);
826 set_pageblock_migratetype(page, MIGRATE_CMA);
828 if (pageblock_order >= MAX_ORDER) {
829 i = pageblock_nr_pages;
832 set_page_refcounted(p);
833 __free_pages(p, MAX_ORDER - 1);
834 p += MAX_ORDER_NR_PAGES;
835 } while (i -= MAX_ORDER_NR_PAGES);
837 set_page_refcounted(page);
838 __free_pages(page, pageblock_order);
841 adjust_managed_page_count(page, pageblock_nr_pages);
846 * The order of subdivision here is critical for the IO subsystem.
847 * Please do not alter this order without good reasons and regression
848 * testing. Specifically, as large blocks of memory are subdivided,
849 * the order in which smaller blocks are delivered depends on the order
850 * they're subdivided in this function. This is the primary factor
851 * influencing the order in which pages are delivered to the IO
852 * subsystem according to empirical testing, and this is also justified
853 * by considering the behavior of a buddy system containing a single
854 * large block of memory acted on by a series of small allocations.
855 * This behavior is a critical factor in sglist merging's success.
859 static inline void expand(struct zone *zone, struct page *page,
860 int low, int high, struct free_area *area,
863 unsigned long size = 1 << high;
869 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
871 #ifdef CONFIG_DEBUG_PAGEALLOC
872 if (high < debug_guardpage_minorder()) {
874 * Mark as guard pages (or page), that will allow to
875 * merge back to allocator when buddy will be freed.
876 * Corresponding page table entries will not be touched,
877 * pages will stay not present in virtual address space
879 INIT_LIST_HEAD(&page[size].lru);
880 set_page_guard_flag(&page[size]);
881 set_page_private(&page[size], high);
882 /* Guard pages are not available for any usage */
883 __mod_zone_freepage_state(zone, -(1 << high),
888 list_add(&page[size].lru, &area->free_list[migratetype]);
890 set_page_order(&page[size], high);
895 * This page is about to be returned from the page allocator
897 static inline int check_new_page(struct page *page)
899 const char *bad_reason = NULL;
900 unsigned long bad_flags = 0;
902 if (unlikely(page_mapcount(page)))
903 bad_reason = "nonzero mapcount";
904 if (unlikely(page->mapping != NULL))
905 bad_reason = "non-NULL mapping";
906 if (unlikely(atomic_read(&page->_count) != 0))
907 bad_reason = "nonzero _count";
908 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
909 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
910 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
912 if (unlikely(mem_cgroup_bad_page_check(page)))
913 bad_reason = "cgroup check failed";
914 if (unlikely(bad_reason)) {
915 bad_page(page, bad_reason, bad_flags);
921 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags)
925 for (i = 0; i < (1 << order); i++) {
926 struct page *p = page + i;
927 if (unlikely(check_new_page(p)))
931 set_page_private(page, 0);
932 set_page_refcounted(page);
934 arch_alloc_page(page, order);
935 kernel_map_pages(page, 1 << order, 1);
937 if (gfp_flags & __GFP_ZERO)
938 prep_zero_page(page, order, gfp_flags);
940 if (order && (gfp_flags & __GFP_COMP))
941 prep_compound_page(page, order);
947 * Go through the free lists for the given migratetype and remove
948 * the smallest available page from the freelists
951 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
954 unsigned int current_order;
955 struct free_area *area;
958 /* Find a page of the appropriate size in the preferred list */
959 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
960 area = &(zone->free_area[current_order]);
961 if (list_empty(&area->free_list[migratetype]))
964 page = list_entry(area->free_list[migratetype].next,
966 list_del(&page->lru);
967 rmv_page_order(page);
969 expand(zone, page, order, current_order, area, migratetype);
970 set_freepage_migratetype(page, migratetype);
979 * This array describes the order lists are fallen back to when
980 * the free lists for the desirable migrate type are depleted
982 static int fallbacks[MIGRATE_TYPES][4] = {
983 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
984 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
986 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
987 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
989 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
991 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
992 #ifdef CONFIG_MEMORY_ISOLATION
993 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
998 * Move the free pages in a range to the free lists of the requested type.
999 * Note that start_page and end_pages are not aligned on a pageblock
1000 * boundary. If alignment is required, use move_freepages_block()
1002 int move_freepages(struct zone *zone,
1003 struct page *start_page, struct page *end_page,
1007 unsigned long order;
1008 int pages_moved = 0;
1010 #ifndef CONFIG_HOLES_IN_ZONE
1012 * page_zone is not safe to call in this context when
1013 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1014 * anyway as we check zone boundaries in move_freepages_block().
1015 * Remove at a later date when no bug reports exist related to
1016 * grouping pages by mobility
1018 BUG_ON(page_zone(start_page) != page_zone(end_page));
1021 for (page = start_page; page <= end_page;) {
1022 /* Make sure we are not inadvertently changing nodes */
1023 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1025 if (!pfn_valid_within(page_to_pfn(page))) {
1030 if (!PageBuddy(page)) {
1035 order = page_order(page);
1036 list_move(&page->lru,
1037 &zone->free_area[order].free_list[migratetype]);
1038 set_freepage_migratetype(page, migratetype);
1040 pages_moved += 1 << order;
1046 int move_freepages_block(struct zone *zone, struct page *page,
1049 unsigned long start_pfn, end_pfn;
1050 struct page *start_page, *end_page;
1052 start_pfn = page_to_pfn(page);
1053 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1054 start_page = pfn_to_page(start_pfn);
1055 end_page = start_page + pageblock_nr_pages - 1;
1056 end_pfn = start_pfn + pageblock_nr_pages - 1;
1058 /* Do not cross zone boundaries */
1059 if (!zone_spans_pfn(zone, start_pfn))
1061 if (!zone_spans_pfn(zone, end_pfn))
1064 return move_freepages(zone, start_page, end_page, migratetype);
1067 static void change_pageblock_range(struct page *pageblock_page,
1068 int start_order, int migratetype)
1070 int nr_pageblocks = 1 << (start_order - pageblock_order);
1072 while (nr_pageblocks--) {
1073 set_pageblock_migratetype(pageblock_page, migratetype);
1074 pageblock_page += pageblock_nr_pages;
1079 * If breaking a large block of pages, move all free pages to the preferred
1080 * allocation list. If falling back for a reclaimable kernel allocation, be
1081 * more aggressive about taking ownership of free pages.
1083 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1084 * nor move CMA pages to different free lists. We don't want unmovable pages
1085 * to be allocated from MIGRATE_CMA areas.
1087 * Returns the new migratetype of the pageblock (or the same old migratetype
1088 * if it was unchanged).
1090 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1091 int start_type, int fallback_type)
1093 int current_order = page_order(page);
1096 * When borrowing from MIGRATE_CMA, we need to release the excess
1097 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1098 * is set to CMA so it is returned to the correct freelist in case
1099 * the page ends up being not actually allocated from the pcp lists.
1101 if (is_migrate_cma(fallback_type))
1102 return fallback_type;
1104 /* Take ownership for orders >= pageblock_order */
1105 if (current_order >= pageblock_order) {
1106 change_pageblock_range(page, current_order, start_type);
1110 if (current_order >= pageblock_order / 2 ||
1111 start_type == MIGRATE_RECLAIMABLE ||
1112 page_group_by_mobility_disabled) {
1115 pages = move_freepages_block(zone, page, start_type);
1117 /* Claim the whole block if over half of it is free */
1118 if (pages >= (1 << (pageblock_order-1)) ||
1119 page_group_by_mobility_disabled) {
1121 set_pageblock_migratetype(page, start_type);
1127 return fallback_type;
1130 /* Remove an element from the buddy allocator from the fallback list */
1131 static inline struct page *
1132 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1134 struct free_area *area;
1135 unsigned int current_order;
1137 int migratetype, new_type, i;
1139 /* Find the largest possible block of pages in the other list */
1140 for (current_order = MAX_ORDER-1;
1141 current_order >= order && current_order <= MAX_ORDER-1;
1144 migratetype = fallbacks[start_migratetype][i];
1146 /* MIGRATE_RESERVE handled later if necessary */
1147 if (migratetype == MIGRATE_RESERVE)
1150 area = &(zone->free_area[current_order]);
1151 if (list_empty(&area->free_list[migratetype]))
1154 page = list_entry(area->free_list[migratetype].next,
1158 new_type = try_to_steal_freepages(zone, page,
1162 /* Remove the page from the freelists */
1163 list_del(&page->lru);
1164 rmv_page_order(page);
1166 expand(zone, page, order, current_order, area,
1168 /* The freepage_migratetype may differ from pageblock's
1169 * migratetype depending on the decisions in
1170 * try_to_steal_freepages. This is OK as long as it does
1171 * not differ for MIGRATE_CMA type.
1173 set_freepage_migratetype(page, new_type);
1175 trace_mm_page_alloc_extfrag(page, order, current_order,
1176 start_migratetype, migratetype, new_type);
1186 * Do the hard work of removing an element from the buddy allocator.
1187 * Call me with the zone->lock already held.
1189 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1195 page = __rmqueue_smallest(zone, order, migratetype);
1197 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1198 page = __rmqueue_fallback(zone, order, migratetype);
1201 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1202 * is used because __rmqueue_smallest is an inline function
1203 * and we want just one call site
1206 migratetype = MIGRATE_RESERVE;
1211 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1216 * Obtain a specified number of elements from the buddy allocator, all under
1217 * a single hold of the lock, for efficiency. Add them to the supplied list.
1218 * Returns the number of new pages which were placed at *list.
1220 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1221 unsigned long count, struct list_head *list,
1222 int migratetype, bool cold)
1226 spin_lock(&zone->lock);
1227 for (i = 0; i < count; ++i) {
1228 struct page *page = __rmqueue(zone, order, migratetype);
1229 if (unlikely(page == NULL))
1233 * Split buddy pages returned by expand() are received here
1234 * in physical page order. The page is added to the callers and
1235 * list and the list head then moves forward. From the callers
1236 * perspective, the linked list is ordered by page number in
1237 * some conditions. This is useful for IO devices that can
1238 * merge IO requests if the physical pages are ordered
1242 list_add(&page->lru, list);
1244 list_add_tail(&page->lru, list);
1246 if (is_migrate_cma(get_freepage_migratetype(page)))
1247 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1250 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1251 spin_unlock(&zone->lock);
1257 * Called from the vmstat counter updater to drain pagesets of this
1258 * currently executing processor on remote nodes after they have
1261 * Note that this function must be called with the thread pinned to
1262 * a single processor.
1264 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1266 unsigned long flags;
1267 int to_drain, batch;
1269 local_irq_save(flags);
1270 batch = ACCESS_ONCE(pcp->batch);
1271 to_drain = min(pcp->count, batch);
1273 free_pcppages_bulk(zone, to_drain, pcp);
1274 pcp->count -= to_drain;
1276 local_irq_restore(flags);
1281 * Drain pages of the indicated processor.
1283 * The processor must either be the current processor and the
1284 * thread pinned to the current processor or a processor that
1287 static void drain_pages(unsigned int cpu)
1289 unsigned long flags;
1292 for_each_populated_zone(zone) {
1293 struct per_cpu_pageset *pset;
1294 struct per_cpu_pages *pcp;
1296 local_irq_save(flags);
1297 pset = per_cpu_ptr(zone->pageset, cpu);
1301 free_pcppages_bulk(zone, pcp->count, pcp);
1304 local_irq_restore(flags);
1309 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1311 void drain_local_pages(void *arg)
1313 drain_pages(smp_processor_id());
1317 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1319 * Note that this code is protected against sending an IPI to an offline
1320 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1321 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1322 * nothing keeps CPUs from showing up after we populated the cpumask and
1323 * before the call to on_each_cpu_mask().
1325 void drain_all_pages(void)
1328 struct per_cpu_pageset *pcp;
1332 * Allocate in the BSS so we wont require allocation in
1333 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1335 static cpumask_t cpus_with_pcps;
1338 * We don't care about racing with CPU hotplug event
1339 * as offline notification will cause the notified
1340 * cpu to drain that CPU pcps and on_each_cpu_mask
1341 * disables preemption as part of its processing
1343 for_each_online_cpu(cpu) {
1344 bool has_pcps = false;
1345 for_each_populated_zone(zone) {
1346 pcp = per_cpu_ptr(zone->pageset, cpu);
1347 if (pcp->pcp.count) {
1353 cpumask_set_cpu(cpu, &cpus_with_pcps);
1355 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1357 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1360 #ifdef CONFIG_HIBERNATION
1362 void mark_free_pages(struct zone *zone)
1364 unsigned long pfn, max_zone_pfn;
1365 unsigned long flags;
1366 unsigned int order, t;
1367 struct list_head *curr;
1369 if (zone_is_empty(zone))
1372 spin_lock_irqsave(&zone->lock, flags);
1374 max_zone_pfn = zone_end_pfn(zone);
1375 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1376 if (pfn_valid(pfn)) {
1377 struct page *page = pfn_to_page(pfn);
1379 if (!swsusp_page_is_forbidden(page))
1380 swsusp_unset_page_free(page);
1383 for_each_migratetype_order(order, t) {
1384 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1387 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1388 for (i = 0; i < (1UL << order); i++)
1389 swsusp_set_page_free(pfn_to_page(pfn + i));
1392 spin_unlock_irqrestore(&zone->lock, flags);
1394 #endif /* CONFIG_PM */
1397 * Free a 0-order page
1398 * cold == true ? free a cold page : free a hot page
1400 void free_hot_cold_page(struct page *page, bool cold)
1402 struct zone *zone = page_zone(page);
1403 struct per_cpu_pages *pcp;
1404 unsigned long flags;
1405 unsigned long pfn = page_to_pfn(page);
1408 if (!free_pages_prepare(page, 0))
1411 migratetype = get_pfnblock_migratetype(page, pfn);
1412 set_freepage_migratetype(page, migratetype);
1413 local_irq_save(flags);
1414 __count_vm_event(PGFREE);
1417 * We only track unmovable, reclaimable and movable on pcp lists.
1418 * Free ISOLATE pages back to the allocator because they are being
1419 * offlined but treat RESERVE as movable pages so we can get those
1420 * areas back if necessary. Otherwise, we may have to free
1421 * excessively into the page allocator
1423 if (migratetype >= MIGRATE_PCPTYPES) {
1424 if (unlikely(is_migrate_isolate(migratetype))) {
1425 free_one_page(zone, page, pfn, 0, migratetype);
1428 migratetype = MIGRATE_MOVABLE;
1431 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1433 list_add(&page->lru, &pcp->lists[migratetype]);
1435 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1437 if (pcp->count >= pcp->high) {
1438 unsigned long batch = ACCESS_ONCE(pcp->batch);
1439 free_pcppages_bulk(zone, batch, pcp);
1440 pcp->count -= batch;
1444 local_irq_restore(flags);
1448 * Free a list of 0-order pages
1450 void free_hot_cold_page_list(struct list_head *list, bool cold)
1452 struct page *page, *next;
1454 list_for_each_entry_safe(page, next, list, lru) {
1455 trace_mm_page_free_batched(page, cold);
1456 free_hot_cold_page(page, cold);
1461 * split_page takes a non-compound higher-order page, and splits it into
1462 * n (1<<order) sub-pages: page[0..n]
1463 * Each sub-page must be freed individually.
1465 * Note: this is probably too low level an operation for use in drivers.
1466 * Please consult with lkml before using this in your driver.
1468 void split_page(struct page *page, unsigned int order)
1472 VM_BUG_ON_PAGE(PageCompound(page), page);
1473 VM_BUG_ON_PAGE(!page_count(page), page);
1475 #ifdef CONFIG_KMEMCHECK
1477 * Split shadow pages too, because free(page[0]) would
1478 * otherwise free the whole shadow.
1480 if (kmemcheck_page_is_tracked(page))
1481 split_page(virt_to_page(page[0].shadow), order);
1484 for (i = 1; i < (1 << order); i++)
1485 set_page_refcounted(page + i);
1487 EXPORT_SYMBOL_GPL(split_page);
1489 static int __isolate_free_page(struct page *page, unsigned int order)
1491 unsigned long watermark;
1495 BUG_ON(!PageBuddy(page));
1497 zone = page_zone(page);
1498 mt = get_pageblock_migratetype(page);
1500 if (!is_migrate_isolate(mt)) {
1501 /* Obey watermarks as if the page was being allocated */
1502 watermark = low_wmark_pages(zone) + (1 << order);
1503 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1506 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1509 /* Remove page from free list */
1510 list_del(&page->lru);
1511 zone->free_area[order].nr_free--;
1512 rmv_page_order(page);
1514 /* Set the pageblock if the isolated page is at least a pageblock */
1515 if (order >= pageblock_order - 1) {
1516 struct page *endpage = page + (1 << order) - 1;
1517 for (; page < endpage; page += pageblock_nr_pages) {
1518 int mt = get_pageblock_migratetype(page);
1519 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1520 set_pageblock_migratetype(page,
1525 return 1UL << order;
1529 * Similar to split_page except the page is already free. As this is only
1530 * being used for migration, the migratetype of the block also changes.
1531 * As this is called with interrupts disabled, the caller is responsible
1532 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1535 * Note: this is probably too low level an operation for use in drivers.
1536 * Please consult with lkml before using this in your driver.
1538 int split_free_page(struct page *page)
1543 order = page_order(page);
1545 nr_pages = __isolate_free_page(page, order);
1549 /* Split into individual pages */
1550 set_page_refcounted(page);
1551 split_page(page, order);
1556 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1557 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1561 struct page *buffered_rmqueue(struct zone *preferred_zone,
1562 struct zone *zone, unsigned int order,
1563 gfp_t gfp_flags, int migratetype)
1565 unsigned long flags;
1567 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1570 if (likely(order == 0)) {
1571 struct per_cpu_pages *pcp;
1572 struct list_head *list;
1574 local_irq_save(flags);
1575 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1576 list = &pcp->lists[migratetype];
1577 if (list_empty(list)) {
1578 pcp->count += rmqueue_bulk(zone, 0,
1581 if (unlikely(list_empty(list)))
1586 page = list_entry(list->prev, struct page, lru);
1588 page = list_entry(list->next, struct page, lru);
1590 list_del(&page->lru);
1593 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1595 * __GFP_NOFAIL is not to be used in new code.
1597 * All __GFP_NOFAIL callers should be fixed so that they
1598 * properly detect and handle allocation failures.
1600 * We most definitely don't want callers attempting to
1601 * allocate greater than order-1 page units with
1604 WARN_ON_ONCE(order > 1);
1606 spin_lock_irqsave(&zone->lock, flags);
1607 page = __rmqueue(zone, order, migratetype);
1608 spin_unlock(&zone->lock);
1611 __mod_zone_freepage_state(zone, -(1 << order),
1612 get_freepage_migratetype(page));
1615 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1616 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1617 !zone_is_fair_depleted(zone))
1618 zone_set_flag(zone, ZONE_FAIR_DEPLETED);
1620 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1621 zone_statistics(preferred_zone, zone, gfp_flags);
1622 local_irq_restore(flags);
1624 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1625 if (prep_new_page(page, order, gfp_flags))
1630 local_irq_restore(flags);
1634 #ifdef CONFIG_FAIL_PAGE_ALLOC
1637 struct fault_attr attr;
1639 u32 ignore_gfp_highmem;
1640 u32 ignore_gfp_wait;
1642 } fail_page_alloc = {
1643 .attr = FAULT_ATTR_INITIALIZER,
1644 .ignore_gfp_wait = 1,
1645 .ignore_gfp_highmem = 1,
1649 static int __init setup_fail_page_alloc(char *str)
1651 return setup_fault_attr(&fail_page_alloc.attr, str);
1653 __setup("fail_page_alloc=", setup_fail_page_alloc);
1655 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1657 if (order < fail_page_alloc.min_order)
1659 if (gfp_mask & __GFP_NOFAIL)
1661 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1663 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1666 return should_fail(&fail_page_alloc.attr, 1 << order);
1669 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1671 static int __init fail_page_alloc_debugfs(void)
1673 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1676 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1677 &fail_page_alloc.attr);
1679 return PTR_ERR(dir);
1681 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1682 &fail_page_alloc.ignore_gfp_wait))
1684 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1685 &fail_page_alloc.ignore_gfp_highmem))
1687 if (!debugfs_create_u32("min-order", mode, dir,
1688 &fail_page_alloc.min_order))
1693 debugfs_remove_recursive(dir);
1698 late_initcall(fail_page_alloc_debugfs);
1700 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1702 #else /* CONFIG_FAIL_PAGE_ALLOC */
1704 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1709 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1712 * Return true if free pages are above 'mark'. This takes into account the order
1713 * of the allocation.
1715 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1716 unsigned long mark, int classzone_idx, int alloc_flags,
1719 /* free_pages my go negative - that's OK */
1724 free_pages -= (1 << order) - 1;
1725 if (alloc_flags & ALLOC_HIGH)
1727 if (alloc_flags & ALLOC_HARDER)
1730 /* If allocation can't use CMA areas don't use free CMA pages */
1731 if (!(alloc_flags & ALLOC_CMA))
1732 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1735 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1737 for (o = 0; o < order; o++) {
1738 /* At the next order, this order's pages become unavailable */
1739 free_pages -= z->free_area[o].nr_free << o;
1741 /* Require fewer higher order pages to be free */
1744 if (free_pages <= min)
1750 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1751 int classzone_idx, int alloc_flags)
1753 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1754 zone_page_state(z, NR_FREE_PAGES));
1757 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1758 unsigned long mark, int classzone_idx, int alloc_flags)
1760 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1762 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1763 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1765 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1771 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1772 * skip over zones that are not allowed by the cpuset, or that have
1773 * been recently (in last second) found to be nearly full. See further
1774 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1775 * that have to skip over a lot of full or unallowed zones.
1777 * If the zonelist cache is present in the passed zonelist, then
1778 * returns a pointer to the allowed node mask (either the current
1779 * tasks mems_allowed, or node_states[N_MEMORY].)
1781 * If the zonelist cache is not available for this zonelist, does
1782 * nothing and returns NULL.
1784 * If the fullzones BITMAP in the zonelist cache is stale (more than
1785 * a second since last zap'd) then we zap it out (clear its bits.)
1787 * We hold off even calling zlc_setup, until after we've checked the
1788 * first zone in the zonelist, on the theory that most allocations will
1789 * be satisfied from that first zone, so best to examine that zone as
1790 * quickly as we can.
1792 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1794 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1795 nodemask_t *allowednodes; /* zonelist_cache approximation */
1797 zlc = zonelist->zlcache_ptr;
1801 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1802 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1803 zlc->last_full_zap = jiffies;
1806 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1807 &cpuset_current_mems_allowed :
1808 &node_states[N_MEMORY];
1809 return allowednodes;
1813 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1814 * if it is worth looking at further for free memory:
1815 * 1) Check that the zone isn't thought to be full (doesn't have its
1816 * bit set in the zonelist_cache fullzones BITMAP).
1817 * 2) Check that the zones node (obtained from the zonelist_cache
1818 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1819 * Return true (non-zero) if zone is worth looking at further, or
1820 * else return false (zero) if it is not.
1822 * This check -ignores- the distinction between various watermarks,
1823 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1824 * found to be full for any variation of these watermarks, it will
1825 * be considered full for up to one second by all requests, unless
1826 * we are so low on memory on all allowed nodes that we are forced
1827 * into the second scan of the zonelist.
1829 * In the second scan we ignore this zonelist cache and exactly
1830 * apply the watermarks to all zones, even it is slower to do so.
1831 * We are low on memory in the second scan, and should leave no stone
1832 * unturned looking for a free page.
1834 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1835 nodemask_t *allowednodes)
1837 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1838 int i; /* index of *z in zonelist zones */
1839 int n; /* node that zone *z is on */
1841 zlc = zonelist->zlcache_ptr;
1845 i = z - zonelist->_zonerefs;
1848 /* This zone is worth trying if it is allowed but not full */
1849 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1853 * Given 'z' scanning a zonelist, set the corresponding bit in
1854 * zlc->fullzones, so that subsequent attempts to allocate a page
1855 * from that zone don't waste time re-examining it.
1857 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1859 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1860 int i; /* index of *z in zonelist zones */
1862 zlc = zonelist->zlcache_ptr;
1866 i = z - zonelist->_zonerefs;
1868 set_bit(i, zlc->fullzones);
1872 * clear all zones full, called after direct reclaim makes progress so that
1873 * a zone that was recently full is not skipped over for up to a second
1875 static void zlc_clear_zones_full(struct zonelist *zonelist)
1877 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1879 zlc = zonelist->zlcache_ptr;
1883 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1886 static bool zone_local(struct zone *local_zone, struct zone *zone)
1888 return local_zone->node == zone->node;
1891 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1893 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1897 #else /* CONFIG_NUMA */
1899 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1904 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1905 nodemask_t *allowednodes)
1910 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1914 static void zlc_clear_zones_full(struct zonelist *zonelist)
1918 static bool zone_local(struct zone *local_zone, struct zone *zone)
1923 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1928 #endif /* CONFIG_NUMA */
1930 static void reset_alloc_batches(struct zone *preferred_zone)
1932 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
1935 mod_zone_page_state(zone, NR_ALLOC_BATCH,
1936 high_wmark_pages(zone) - low_wmark_pages(zone) -
1937 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
1938 zone_clear_flag(zone, ZONE_FAIR_DEPLETED);
1939 } while (zone++ != preferred_zone);
1943 * get_page_from_freelist goes through the zonelist trying to allocate
1946 static struct page *
1947 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1948 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1949 struct zone *preferred_zone, int classzone_idx, int migratetype)
1952 struct page *page = NULL;
1954 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1955 int zlc_active = 0; /* set if using zonelist_cache */
1956 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1957 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
1958 (gfp_mask & __GFP_WRITE);
1959 int nr_fair_skipped = 0;
1960 bool zonelist_rescan;
1963 zonelist_rescan = false;
1966 * Scan zonelist, looking for a zone with enough free.
1967 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1969 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1970 high_zoneidx, nodemask) {
1973 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1974 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1976 if (cpusets_enabled() &&
1977 (alloc_flags & ALLOC_CPUSET) &&
1978 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1981 * Distribute pages in proportion to the individual
1982 * zone size to ensure fair page aging. The zone a
1983 * page was allocated in should have no effect on the
1984 * time the page has in memory before being reclaimed.
1986 if (alloc_flags & ALLOC_FAIR) {
1987 if (!zone_local(preferred_zone, zone))
1989 if (zone_is_fair_depleted(zone)) {
1995 * When allocating a page cache page for writing, we
1996 * want to get it from a zone that is within its dirty
1997 * limit, such that no single zone holds more than its
1998 * proportional share of globally allowed dirty pages.
1999 * The dirty limits take into account the zone's
2000 * lowmem reserves and high watermark so that kswapd
2001 * should be able to balance it without having to
2002 * write pages from its LRU list.
2004 * This may look like it could increase pressure on
2005 * lower zones by failing allocations in higher zones
2006 * before they are full. But the pages that do spill
2007 * over are limited as the lower zones are protected
2008 * by this very same mechanism. It should not become
2009 * a practical burden to them.
2011 * XXX: For now, allow allocations to potentially
2012 * exceed the per-zone dirty limit in the slowpath
2013 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2014 * which is important when on a NUMA setup the allowed
2015 * zones are together not big enough to reach the
2016 * global limit. The proper fix for these situations
2017 * will require awareness of zones in the
2018 * dirty-throttling and the flusher threads.
2020 if (consider_zone_dirty && !zone_dirty_ok(zone))
2023 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2024 if (!zone_watermark_ok(zone, order, mark,
2025 classzone_idx, alloc_flags)) {
2028 /* Checked here to keep the fast path fast */
2029 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2030 if (alloc_flags & ALLOC_NO_WATERMARKS)
2033 if (IS_ENABLED(CONFIG_NUMA) &&
2034 !did_zlc_setup && nr_online_nodes > 1) {
2036 * we do zlc_setup if there are multiple nodes
2037 * and before considering the first zone allowed
2040 allowednodes = zlc_setup(zonelist, alloc_flags);
2045 if (zone_reclaim_mode == 0 ||
2046 !zone_allows_reclaim(preferred_zone, zone))
2047 goto this_zone_full;
2050 * As we may have just activated ZLC, check if the first
2051 * eligible zone has failed zone_reclaim recently.
2053 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2054 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2057 ret = zone_reclaim(zone, gfp_mask, order);
2059 case ZONE_RECLAIM_NOSCAN:
2062 case ZONE_RECLAIM_FULL:
2063 /* scanned but unreclaimable */
2066 /* did we reclaim enough */
2067 if (zone_watermark_ok(zone, order, mark,
2068 classzone_idx, alloc_flags))
2072 * Failed to reclaim enough to meet watermark.
2073 * Only mark the zone full if checking the min
2074 * watermark or if we failed to reclaim just
2075 * 1<<order pages or else the page allocator
2076 * fastpath will prematurely mark zones full
2077 * when the watermark is between the low and
2080 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2081 ret == ZONE_RECLAIM_SOME)
2082 goto this_zone_full;
2089 page = buffered_rmqueue(preferred_zone, zone, order,
2090 gfp_mask, migratetype);
2094 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2095 zlc_mark_zone_full(zonelist, z);
2100 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2101 * necessary to allocate the page. The expectation is
2102 * that the caller is taking steps that will free more
2103 * memory. The caller should avoid the page being used
2104 * for !PFMEMALLOC purposes.
2106 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2111 * The first pass makes sure allocations are spread fairly within the
2112 * local node. However, the local node might have free pages left
2113 * after the fairness batches are exhausted, and remote zones haven't
2114 * even been considered yet. Try once more without fairness, and
2115 * include remote zones now, before entering the slowpath and waking
2116 * kswapd: prefer spilling to a remote zone over swapping locally.
2118 if (alloc_flags & ALLOC_FAIR) {
2119 alloc_flags &= ~ALLOC_FAIR;
2120 if (nr_fair_skipped) {
2121 zonelist_rescan = true;
2122 reset_alloc_batches(preferred_zone);
2124 if (nr_online_nodes > 1)
2125 zonelist_rescan = true;
2128 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2129 /* Disable zlc cache for second zonelist scan */
2131 zonelist_rescan = true;
2134 if (zonelist_rescan)
2141 * Large machines with many possible nodes should not always dump per-node
2142 * meminfo in irq context.
2144 static inline bool should_suppress_show_mem(void)
2149 ret = in_interrupt();
2154 static DEFINE_RATELIMIT_STATE(nopage_rs,
2155 DEFAULT_RATELIMIT_INTERVAL,
2156 DEFAULT_RATELIMIT_BURST);
2158 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2160 unsigned int filter = SHOW_MEM_FILTER_NODES;
2162 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2163 debug_guardpage_minorder() > 0)
2167 * This documents exceptions given to allocations in certain
2168 * contexts that are allowed to allocate outside current's set
2171 if (!(gfp_mask & __GFP_NOMEMALLOC))
2172 if (test_thread_flag(TIF_MEMDIE) ||
2173 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2174 filter &= ~SHOW_MEM_FILTER_NODES;
2175 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2176 filter &= ~SHOW_MEM_FILTER_NODES;
2179 struct va_format vaf;
2182 va_start(args, fmt);
2187 pr_warn("%pV", &vaf);
2192 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2193 current->comm, order, gfp_mask);
2196 if (!should_suppress_show_mem())
2201 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2202 unsigned long did_some_progress,
2203 unsigned long pages_reclaimed)
2205 /* Do not loop if specifically requested */
2206 if (gfp_mask & __GFP_NORETRY)
2209 /* Always retry if specifically requested */
2210 if (gfp_mask & __GFP_NOFAIL)
2214 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2215 * making forward progress without invoking OOM. Suspend also disables
2216 * storage devices so kswapd will not help. Bail if we are suspending.
2218 if (!did_some_progress && pm_suspended_storage())
2222 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2223 * means __GFP_NOFAIL, but that may not be true in other
2226 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2230 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2231 * specified, then we retry until we no longer reclaim any pages
2232 * (above), or we've reclaimed an order of pages at least as
2233 * large as the allocation's order. In both cases, if the
2234 * allocation still fails, we stop retrying.
2236 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2242 static inline struct page *
2243 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2244 struct zonelist *zonelist, enum zone_type high_zoneidx,
2245 nodemask_t *nodemask, struct zone *preferred_zone,
2246 int classzone_idx, int migratetype)
2250 /* Acquire the per-zone oom lock for each zone */
2251 if (!oom_zonelist_trylock(zonelist, gfp_mask)) {
2252 schedule_timeout_uninterruptible(1);
2257 * Go through the zonelist yet one more time, keep very high watermark
2258 * here, this is only to catch a parallel oom killing, we must fail if
2259 * we're still under heavy pressure.
2261 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2262 order, zonelist, high_zoneidx,
2263 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2264 preferred_zone, classzone_idx, migratetype);
2268 if (!(gfp_mask & __GFP_NOFAIL)) {
2269 /* The OOM killer will not help higher order allocs */
2270 if (order > PAGE_ALLOC_COSTLY_ORDER)
2272 /* The OOM killer does not needlessly kill tasks for lowmem */
2273 if (high_zoneidx < ZONE_NORMAL)
2276 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2277 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2278 * The caller should handle page allocation failure by itself if
2279 * it specifies __GFP_THISNODE.
2280 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2282 if (gfp_mask & __GFP_THISNODE)
2285 /* Exhausted what can be done so it's blamo time */
2286 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2289 oom_zonelist_unlock(zonelist, gfp_mask);
2293 #ifdef CONFIG_COMPACTION
2294 /* Try memory compaction for high-order allocations before reclaim */
2295 static struct page *
2296 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2297 struct zonelist *zonelist, enum zone_type high_zoneidx,
2298 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2299 int classzone_idx, int migratetype, enum migrate_mode mode,
2300 int *contended_compaction, bool *deferred_compaction)
2302 struct zone *last_compact_zone = NULL;
2303 unsigned long compact_result;
2309 current->flags |= PF_MEMALLOC;
2310 compact_result = try_to_compact_pages(zonelist, order, gfp_mask,
2312 contended_compaction,
2313 &last_compact_zone);
2314 current->flags &= ~PF_MEMALLOC;
2316 switch (compact_result) {
2317 case COMPACT_DEFERRED:
2318 *deferred_compaction = true;
2320 case COMPACT_SKIPPED:
2327 * At least in one zone compaction wasn't deferred or skipped, so let's
2328 * count a compaction stall
2330 count_vm_event(COMPACTSTALL);
2332 /* Page migration frees to the PCP lists but we want merging */
2333 drain_pages(get_cpu());
2336 page = get_page_from_freelist(gfp_mask, nodemask,
2337 order, zonelist, high_zoneidx,
2338 alloc_flags & ~ALLOC_NO_WATERMARKS,
2339 preferred_zone, classzone_idx, migratetype);
2342 struct zone *zone = page_zone(page);
2344 zone->compact_blockskip_flush = false;
2345 compaction_defer_reset(zone, order, true);
2346 count_vm_event(COMPACTSUCCESS);
2351 * last_compact_zone is where try_to_compact_pages thought allocation
2352 * should succeed, so it did not defer compaction. But here we know
2353 * that it didn't succeed, so we do the defer.
2355 if (last_compact_zone && mode != MIGRATE_ASYNC)
2356 defer_compaction(last_compact_zone, order);
2359 * It's bad if compaction run occurs and fails. The most likely reason
2360 * is that pages exist, but not enough to satisfy watermarks.
2362 count_vm_event(COMPACTFAIL);
2369 static inline struct page *
2370 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2371 struct zonelist *zonelist, enum zone_type high_zoneidx,
2372 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2373 int classzone_idx, int migratetype, enum migrate_mode mode,
2374 int *contended_compaction, bool *deferred_compaction)
2378 #endif /* CONFIG_COMPACTION */
2380 /* Perform direct synchronous page reclaim */
2382 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2383 nodemask_t *nodemask)
2385 struct reclaim_state reclaim_state;
2390 /* We now go into synchronous reclaim */
2391 cpuset_memory_pressure_bump();
2392 current->flags |= PF_MEMALLOC;
2393 lockdep_set_current_reclaim_state(gfp_mask);
2394 reclaim_state.reclaimed_slab = 0;
2395 current->reclaim_state = &reclaim_state;
2397 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2399 current->reclaim_state = NULL;
2400 lockdep_clear_current_reclaim_state();
2401 current->flags &= ~PF_MEMALLOC;
2408 /* The really slow allocator path where we enter direct reclaim */
2409 static inline struct page *
2410 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2411 struct zonelist *zonelist, enum zone_type high_zoneidx,
2412 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2413 int classzone_idx, int migratetype, unsigned long *did_some_progress)
2415 struct page *page = NULL;
2416 bool drained = false;
2418 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2420 if (unlikely(!(*did_some_progress)))
2423 /* After successful reclaim, reconsider all zones for allocation */
2424 if (IS_ENABLED(CONFIG_NUMA))
2425 zlc_clear_zones_full(zonelist);
2428 page = get_page_from_freelist(gfp_mask, nodemask, order,
2429 zonelist, high_zoneidx,
2430 alloc_flags & ~ALLOC_NO_WATERMARKS,
2431 preferred_zone, classzone_idx,
2435 * If an allocation failed after direct reclaim, it could be because
2436 * pages are pinned on the per-cpu lists. Drain them and try again
2438 if (!page && !drained) {
2448 * This is called in the allocator slow-path if the allocation request is of
2449 * sufficient urgency to ignore watermarks and take other desperate measures
2451 static inline struct page *
2452 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2453 struct zonelist *zonelist, enum zone_type high_zoneidx,
2454 nodemask_t *nodemask, struct zone *preferred_zone,
2455 int classzone_idx, int migratetype)
2460 page = get_page_from_freelist(gfp_mask, nodemask, order,
2461 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2462 preferred_zone, classzone_idx, migratetype);
2464 if (!page && gfp_mask & __GFP_NOFAIL)
2465 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2466 } while (!page && (gfp_mask & __GFP_NOFAIL));
2471 static void wake_all_kswapds(unsigned int order,
2472 struct zonelist *zonelist,
2473 enum zone_type high_zoneidx,
2474 struct zone *preferred_zone)
2479 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2480 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2484 gfp_to_alloc_flags(gfp_t gfp_mask)
2486 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2487 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2489 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2490 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2493 * The caller may dip into page reserves a bit more if the caller
2494 * cannot run direct reclaim, or if the caller has realtime scheduling
2495 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2496 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2498 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2502 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2503 * if it can't schedule.
2505 if (!(gfp_mask & __GFP_NOMEMALLOC))
2506 alloc_flags |= ALLOC_HARDER;
2508 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2509 * comment for __cpuset_node_allowed_softwall().
2511 alloc_flags &= ~ALLOC_CPUSET;
2512 } else if (unlikely(rt_task(current)) && !in_interrupt())
2513 alloc_flags |= ALLOC_HARDER;
2515 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2516 if (gfp_mask & __GFP_MEMALLOC)
2517 alloc_flags |= ALLOC_NO_WATERMARKS;
2518 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2519 alloc_flags |= ALLOC_NO_WATERMARKS;
2520 else if (!in_interrupt() &&
2521 ((current->flags & PF_MEMALLOC) ||
2522 unlikely(test_thread_flag(TIF_MEMDIE))))
2523 alloc_flags |= ALLOC_NO_WATERMARKS;
2526 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2527 alloc_flags |= ALLOC_CMA;
2532 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2534 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2537 static inline struct page *
2538 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2539 struct zonelist *zonelist, enum zone_type high_zoneidx,
2540 nodemask_t *nodemask, struct zone *preferred_zone,
2541 int classzone_idx, int migratetype)
2543 const gfp_t wait = gfp_mask & __GFP_WAIT;
2544 struct page *page = NULL;
2546 unsigned long pages_reclaimed = 0;
2547 unsigned long did_some_progress;
2548 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2549 bool deferred_compaction = false;
2550 int contended_compaction = COMPACT_CONTENDED_NONE;
2553 * In the slowpath, we sanity check order to avoid ever trying to
2554 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2555 * be using allocators in order of preference for an area that is
2558 if (order >= MAX_ORDER) {
2559 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2564 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2565 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2566 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2567 * using a larger set of nodes after it has established that the
2568 * allowed per node queues are empty and that nodes are
2571 if (IS_ENABLED(CONFIG_NUMA) &&
2572 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2576 if (!(gfp_mask & __GFP_NO_KSWAPD))
2577 wake_all_kswapds(order, zonelist, high_zoneidx, preferred_zone);
2580 * OK, we're below the kswapd watermark and have kicked background
2581 * reclaim. Now things get more complex, so set up alloc_flags according
2582 * to how we want to proceed.
2584 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2587 * Find the true preferred zone if the allocation is unconstrained by
2590 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) {
2591 struct zoneref *preferred_zoneref;
2592 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2593 NULL, &preferred_zone);
2594 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2598 /* This is the last chance, in general, before the goto nopage. */
2599 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2600 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2601 preferred_zone, classzone_idx, migratetype);
2605 /* Allocate without watermarks if the context allows */
2606 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2608 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2609 * the allocation is high priority and these type of
2610 * allocations are system rather than user orientated
2612 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2614 page = __alloc_pages_high_priority(gfp_mask, order,
2615 zonelist, high_zoneidx, nodemask,
2616 preferred_zone, classzone_idx, migratetype);
2622 /* Atomic allocations - we can't balance anything */
2625 * All existing users of the deprecated __GFP_NOFAIL are
2626 * blockable, so warn of any new users that actually allow this
2627 * type of allocation to fail.
2629 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2633 /* Avoid recursion of direct reclaim */
2634 if (current->flags & PF_MEMALLOC)
2637 /* Avoid allocations with no watermarks from looping endlessly */
2638 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2642 * Try direct compaction. The first pass is asynchronous. Subsequent
2643 * attempts after direct reclaim are synchronous
2645 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2646 high_zoneidx, nodemask, alloc_flags,
2648 classzone_idx, migratetype,
2649 migration_mode, &contended_compaction,
2650 &deferred_compaction);
2654 /* Checks for THP-specific high-order allocations */
2655 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2657 * If compaction is deferred for high-order allocations, it is
2658 * because sync compaction recently failed. If this is the case
2659 * and the caller requested a THP allocation, we do not want
2660 * to heavily disrupt the system, so we fail the allocation
2661 * instead of entering direct reclaim.
2663 if (deferred_compaction)
2667 * In all zones where compaction was attempted (and not
2668 * deferred or skipped), lock contention has been detected.
2669 * For THP allocation we do not want to disrupt the others
2670 * so we fallback to base pages instead.
2672 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2676 * If compaction was aborted due to need_resched(), we do not
2677 * want to further increase allocation latency, unless it is
2678 * khugepaged trying to collapse.
2680 if (contended_compaction == COMPACT_CONTENDED_SCHED
2681 && !(current->flags & PF_KTHREAD))
2686 * It can become very expensive to allocate transparent hugepages at
2687 * fault, so use asynchronous memory compaction for THP unless it is
2688 * khugepaged trying to collapse.
2690 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2691 (current->flags & PF_KTHREAD))
2692 migration_mode = MIGRATE_SYNC_LIGHT;
2694 /* Try direct reclaim and then allocating */
2695 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2696 zonelist, high_zoneidx,
2698 alloc_flags, preferred_zone,
2699 classzone_idx, migratetype,
2700 &did_some_progress);
2705 * If we failed to make any progress reclaiming, then we are
2706 * running out of options and have to consider going OOM
2708 if (!did_some_progress) {
2709 if (oom_gfp_allowed(gfp_mask)) {
2710 if (oom_killer_disabled)
2712 /* Coredumps can quickly deplete all memory reserves */
2713 if ((current->flags & PF_DUMPCORE) &&
2714 !(gfp_mask & __GFP_NOFAIL))
2716 page = __alloc_pages_may_oom(gfp_mask, order,
2717 zonelist, high_zoneidx,
2718 nodemask, preferred_zone,
2719 classzone_idx, migratetype);
2723 if (!(gfp_mask & __GFP_NOFAIL)) {
2725 * The oom killer is not called for high-order
2726 * allocations that may fail, so if no progress
2727 * is being made, there are no other options and
2728 * retrying is unlikely to help.
2730 if (order > PAGE_ALLOC_COSTLY_ORDER)
2733 * The oom killer is not called for lowmem
2734 * allocations to prevent needlessly killing
2737 if (high_zoneidx < ZONE_NORMAL)
2745 /* Check if we should retry the allocation */
2746 pages_reclaimed += did_some_progress;
2747 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2749 /* Wait for some write requests to complete then retry */
2750 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2754 * High-order allocations do not necessarily loop after
2755 * direct reclaim and reclaim/compaction depends on compaction
2756 * being called after reclaim so call directly if necessary
2758 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2759 high_zoneidx, nodemask, alloc_flags,
2761 classzone_idx, migratetype,
2762 migration_mode, &contended_compaction,
2763 &deferred_compaction);
2769 warn_alloc_failed(gfp_mask, order, NULL);
2772 if (kmemcheck_enabled)
2773 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2779 * This is the 'heart' of the zoned buddy allocator.
2782 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2783 struct zonelist *zonelist, nodemask_t *nodemask)
2785 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2786 struct zone *preferred_zone;
2787 struct zoneref *preferred_zoneref;
2788 struct page *page = NULL;
2789 int migratetype = gfpflags_to_migratetype(gfp_mask);
2790 unsigned int cpuset_mems_cookie;
2791 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2794 gfp_mask &= gfp_allowed_mask;
2796 lockdep_trace_alloc(gfp_mask);
2798 might_sleep_if(gfp_mask & __GFP_WAIT);
2800 if (should_fail_alloc_page(gfp_mask, order))
2804 * Check the zones suitable for the gfp_mask contain at least one
2805 * valid zone. It's possible to have an empty zonelist as a result
2806 * of GFP_THISNODE and a memoryless node
2808 if (unlikely(!zonelist->_zonerefs->zone))
2811 if (IS_ENABLED(CONFIG_CMA) && migratetype == MIGRATE_MOVABLE)
2812 alloc_flags |= ALLOC_CMA;
2815 cpuset_mems_cookie = read_mems_allowed_begin();
2817 /* The preferred zone is used for statistics later */
2818 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2819 nodemask ? : &cpuset_current_mems_allowed,
2821 if (!preferred_zone)
2823 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2825 /* First allocation attempt */
2826 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2827 zonelist, high_zoneidx, alloc_flags,
2828 preferred_zone, classzone_idx, migratetype);
2829 if (unlikely(!page)) {
2831 * Runtime PM, block IO and its error handling path
2832 * can deadlock because I/O on the device might not
2835 gfp_mask = memalloc_noio_flags(gfp_mask);
2836 page = __alloc_pages_slowpath(gfp_mask, order,
2837 zonelist, high_zoneidx, nodemask,
2838 preferred_zone, classzone_idx, migratetype);
2841 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2845 * When updating a task's mems_allowed, it is possible to race with
2846 * parallel threads in such a way that an allocation can fail while
2847 * the mask is being updated. If a page allocation is about to fail,
2848 * check if the cpuset changed during allocation and if so, retry.
2850 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2855 EXPORT_SYMBOL(__alloc_pages_nodemask);
2858 * Common helper functions.
2860 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2865 * __get_free_pages() returns a 32-bit address, which cannot represent
2868 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2870 page = alloc_pages(gfp_mask, order);
2873 return (unsigned long) page_address(page);
2875 EXPORT_SYMBOL(__get_free_pages);
2877 unsigned long get_zeroed_page(gfp_t gfp_mask)
2879 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2881 EXPORT_SYMBOL(get_zeroed_page);
2883 void __free_pages(struct page *page, unsigned int order)
2885 if (put_page_testzero(page)) {
2887 free_hot_cold_page(page, false);
2889 __free_pages_ok(page, order);
2893 EXPORT_SYMBOL(__free_pages);
2895 void free_pages(unsigned long addr, unsigned int order)
2898 VM_BUG_ON(!virt_addr_valid((void *)addr));
2899 __free_pages(virt_to_page((void *)addr), order);
2903 EXPORT_SYMBOL(free_pages);
2906 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2907 * of the current memory cgroup.
2909 * It should be used when the caller would like to use kmalloc, but since the
2910 * allocation is large, it has to fall back to the page allocator.
2912 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2915 struct mem_cgroup *memcg = NULL;
2917 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2919 page = alloc_pages(gfp_mask, order);
2920 memcg_kmem_commit_charge(page, memcg, order);
2924 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2927 struct mem_cgroup *memcg = NULL;
2929 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2931 page = alloc_pages_node(nid, gfp_mask, order);
2932 memcg_kmem_commit_charge(page, memcg, order);
2937 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2940 void __free_kmem_pages(struct page *page, unsigned int order)
2942 memcg_kmem_uncharge_pages(page, order);
2943 __free_pages(page, order);
2946 void free_kmem_pages(unsigned long addr, unsigned int order)
2949 VM_BUG_ON(!virt_addr_valid((void *)addr));
2950 __free_kmem_pages(virt_to_page((void *)addr), order);
2954 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2957 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2958 unsigned long used = addr + PAGE_ALIGN(size);
2960 split_page(virt_to_page((void *)addr), order);
2961 while (used < alloc_end) {
2966 return (void *)addr;
2970 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2971 * @size: the number of bytes to allocate
2972 * @gfp_mask: GFP flags for the allocation
2974 * This function is similar to alloc_pages(), except that it allocates the
2975 * minimum number of pages to satisfy the request. alloc_pages() can only
2976 * allocate memory in power-of-two pages.
2978 * This function is also limited by MAX_ORDER.
2980 * Memory allocated by this function must be released by free_pages_exact().
2982 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2984 unsigned int order = get_order(size);
2987 addr = __get_free_pages(gfp_mask, order);
2988 return make_alloc_exact(addr, order, size);
2990 EXPORT_SYMBOL(alloc_pages_exact);
2993 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2995 * @nid: the preferred node ID where memory should be allocated
2996 * @size: the number of bytes to allocate
2997 * @gfp_mask: GFP flags for the allocation
2999 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3001 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3004 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3006 unsigned order = get_order(size);
3007 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3010 return make_alloc_exact((unsigned long)page_address(p), order, size);
3014 * free_pages_exact - release memory allocated via alloc_pages_exact()
3015 * @virt: the value returned by alloc_pages_exact.
3016 * @size: size of allocation, same value as passed to alloc_pages_exact().
3018 * Release the memory allocated by a previous call to alloc_pages_exact.
3020 void free_pages_exact(void *virt, size_t size)
3022 unsigned long addr = (unsigned long)virt;
3023 unsigned long end = addr + PAGE_ALIGN(size);
3025 while (addr < end) {
3030 EXPORT_SYMBOL(free_pages_exact);
3033 * nr_free_zone_pages - count number of pages beyond high watermark
3034 * @offset: The zone index of the highest zone
3036 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3037 * high watermark within all zones at or below a given zone index. For each
3038 * zone, the number of pages is calculated as:
3039 * managed_pages - high_pages
3041 static unsigned long nr_free_zone_pages(int offset)
3046 /* Just pick one node, since fallback list is circular */
3047 unsigned long sum = 0;
3049 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3051 for_each_zone_zonelist(zone, z, zonelist, offset) {
3052 unsigned long size = zone->managed_pages;
3053 unsigned long high = high_wmark_pages(zone);
3062 * nr_free_buffer_pages - count number of pages beyond high watermark
3064 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3065 * watermark within ZONE_DMA and ZONE_NORMAL.
3067 unsigned long nr_free_buffer_pages(void)
3069 return nr_free_zone_pages(gfp_zone(GFP_USER));
3071 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3074 * nr_free_pagecache_pages - count number of pages beyond high watermark
3076 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3077 * high watermark within all zones.
3079 unsigned long nr_free_pagecache_pages(void)
3081 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3084 static inline void show_node(struct zone *zone)
3086 if (IS_ENABLED(CONFIG_NUMA))
3087 printk("Node %d ", zone_to_nid(zone));
3090 void si_meminfo(struct sysinfo *val)
3092 val->totalram = totalram_pages;
3093 val->sharedram = global_page_state(NR_SHMEM);
3094 val->freeram = global_page_state(NR_FREE_PAGES);
3095 val->bufferram = nr_blockdev_pages();
3096 val->totalhigh = totalhigh_pages;
3097 val->freehigh = nr_free_highpages();
3098 val->mem_unit = PAGE_SIZE;
3101 EXPORT_SYMBOL(si_meminfo);
3104 void si_meminfo_node(struct sysinfo *val, int nid)
3106 int zone_type; /* needs to be signed */
3107 unsigned long managed_pages = 0;
3108 pg_data_t *pgdat = NODE_DATA(nid);
3110 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3111 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3112 val->totalram = managed_pages;
3113 val->sharedram = node_page_state(nid, NR_SHMEM);
3114 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3115 #ifdef CONFIG_HIGHMEM
3116 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3117 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3123 val->mem_unit = PAGE_SIZE;
3128 * Determine whether the node should be displayed or not, depending on whether
3129 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3131 bool skip_free_areas_node(unsigned int flags, int nid)
3134 unsigned int cpuset_mems_cookie;
3136 if (!(flags & SHOW_MEM_FILTER_NODES))
3140 cpuset_mems_cookie = read_mems_allowed_begin();
3141 ret = !node_isset(nid, cpuset_current_mems_allowed);
3142 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3147 #define K(x) ((x) << (PAGE_SHIFT-10))
3149 static void show_migration_types(unsigned char type)
3151 static const char types[MIGRATE_TYPES] = {
3152 [MIGRATE_UNMOVABLE] = 'U',
3153 [MIGRATE_RECLAIMABLE] = 'E',
3154 [MIGRATE_MOVABLE] = 'M',
3155 [MIGRATE_RESERVE] = 'R',
3157 [MIGRATE_CMA] = 'C',
3159 #ifdef CONFIG_MEMORY_ISOLATION
3160 [MIGRATE_ISOLATE] = 'I',
3163 char tmp[MIGRATE_TYPES + 1];
3167 for (i = 0; i < MIGRATE_TYPES; i++) {
3168 if (type & (1 << i))
3173 printk("(%s) ", tmp);
3177 * Show free area list (used inside shift_scroll-lock stuff)
3178 * We also calculate the percentage fragmentation. We do this by counting the
3179 * memory on each free list with the exception of the first item on the list.
3180 * Suppresses nodes that are not allowed by current's cpuset if
3181 * SHOW_MEM_FILTER_NODES is passed.
3183 void show_free_areas(unsigned int filter)
3188 for_each_populated_zone(zone) {
3189 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3192 printk("%s per-cpu:\n", zone->name);
3194 for_each_online_cpu(cpu) {
3195 struct per_cpu_pageset *pageset;
3197 pageset = per_cpu_ptr(zone->pageset, cpu);
3199 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3200 cpu, pageset->pcp.high,
3201 pageset->pcp.batch, pageset->pcp.count);
3205 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3206 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3208 " dirty:%lu writeback:%lu unstable:%lu\n"
3209 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3210 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3212 global_page_state(NR_ACTIVE_ANON),
3213 global_page_state(NR_INACTIVE_ANON),
3214 global_page_state(NR_ISOLATED_ANON),
3215 global_page_state(NR_ACTIVE_FILE),
3216 global_page_state(NR_INACTIVE_FILE),
3217 global_page_state(NR_ISOLATED_FILE),
3218 global_page_state(NR_UNEVICTABLE),
3219 global_page_state(NR_FILE_DIRTY),
3220 global_page_state(NR_WRITEBACK),
3221 global_page_state(NR_UNSTABLE_NFS),
3222 global_page_state(NR_FREE_PAGES),
3223 global_page_state(NR_SLAB_RECLAIMABLE),
3224 global_page_state(NR_SLAB_UNRECLAIMABLE),
3225 global_page_state(NR_FILE_MAPPED),
3226 global_page_state(NR_SHMEM),
3227 global_page_state(NR_PAGETABLE),
3228 global_page_state(NR_BOUNCE),
3229 global_page_state(NR_FREE_CMA_PAGES));
3231 for_each_populated_zone(zone) {
3234 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3242 " active_anon:%lukB"
3243 " inactive_anon:%lukB"
3244 " active_file:%lukB"
3245 " inactive_file:%lukB"
3246 " unevictable:%lukB"
3247 " isolated(anon):%lukB"
3248 " isolated(file):%lukB"
3256 " slab_reclaimable:%lukB"
3257 " slab_unreclaimable:%lukB"
3258 " kernel_stack:%lukB"
3263 " writeback_tmp:%lukB"
3264 " pages_scanned:%lu"
3265 " all_unreclaimable? %s"
3268 K(zone_page_state(zone, NR_FREE_PAGES)),
3269 K(min_wmark_pages(zone)),
3270 K(low_wmark_pages(zone)),
3271 K(high_wmark_pages(zone)),
3272 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3273 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3274 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3275 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3276 K(zone_page_state(zone, NR_UNEVICTABLE)),
3277 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3278 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3279 K(zone->present_pages),
3280 K(zone->managed_pages),
3281 K(zone_page_state(zone, NR_MLOCK)),
3282 K(zone_page_state(zone, NR_FILE_DIRTY)),
3283 K(zone_page_state(zone, NR_WRITEBACK)),
3284 K(zone_page_state(zone, NR_FILE_MAPPED)),
3285 K(zone_page_state(zone, NR_SHMEM)),
3286 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3287 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3288 zone_page_state(zone, NR_KERNEL_STACK) *
3290 K(zone_page_state(zone, NR_PAGETABLE)),
3291 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3292 K(zone_page_state(zone, NR_BOUNCE)),
3293 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3294 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3295 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3296 (!zone_reclaimable(zone) ? "yes" : "no")
3298 printk("lowmem_reserve[]:");
3299 for (i = 0; i < MAX_NR_ZONES; i++)
3300 printk(" %ld", zone->lowmem_reserve[i]);
3304 for_each_populated_zone(zone) {
3305 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3306 unsigned char types[MAX_ORDER];
3308 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3311 printk("%s: ", zone->name);
3313 spin_lock_irqsave(&zone->lock, flags);
3314 for (order = 0; order < MAX_ORDER; order++) {
3315 struct free_area *area = &zone->free_area[order];
3318 nr[order] = area->nr_free;
3319 total += nr[order] << order;
3322 for (type = 0; type < MIGRATE_TYPES; type++) {
3323 if (!list_empty(&area->free_list[type]))
3324 types[order] |= 1 << type;
3327 spin_unlock_irqrestore(&zone->lock, flags);
3328 for (order = 0; order < MAX_ORDER; order++) {
3329 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3331 show_migration_types(types[order]);
3333 printk("= %lukB\n", K(total));
3336 hugetlb_show_meminfo();
3338 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3340 show_swap_cache_info();
3343 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3345 zoneref->zone = zone;
3346 zoneref->zone_idx = zone_idx(zone);
3350 * Builds allocation fallback zone lists.
3352 * Add all populated zones of a node to the zonelist.
3354 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3358 enum zone_type zone_type = MAX_NR_ZONES;
3362 zone = pgdat->node_zones + zone_type;
3363 if (populated_zone(zone)) {
3364 zoneref_set_zone(zone,
3365 &zonelist->_zonerefs[nr_zones++]);
3366 check_highest_zone(zone_type);
3368 } while (zone_type);
3376 * 0 = automatic detection of better ordering.
3377 * 1 = order by ([node] distance, -zonetype)
3378 * 2 = order by (-zonetype, [node] distance)
3380 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3381 * the same zonelist. So only NUMA can configure this param.
3383 #define ZONELIST_ORDER_DEFAULT 0
3384 #define ZONELIST_ORDER_NODE 1
3385 #define ZONELIST_ORDER_ZONE 2
3387 /* zonelist order in the kernel.
3388 * set_zonelist_order() will set this to NODE or ZONE.
3390 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3391 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3395 /* The value user specified ....changed by config */
3396 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3397 /* string for sysctl */
3398 #define NUMA_ZONELIST_ORDER_LEN 16
3399 char numa_zonelist_order[16] = "default";
3402 * interface for configure zonelist ordering.
3403 * command line option "numa_zonelist_order"
3404 * = "[dD]efault - default, automatic configuration.
3405 * = "[nN]ode - order by node locality, then by zone within node
3406 * = "[zZ]one - order by zone, then by locality within zone
3409 static int __parse_numa_zonelist_order(char *s)
3411 if (*s == 'd' || *s == 'D') {
3412 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3413 } else if (*s == 'n' || *s == 'N') {
3414 user_zonelist_order = ZONELIST_ORDER_NODE;
3415 } else if (*s == 'z' || *s == 'Z') {
3416 user_zonelist_order = ZONELIST_ORDER_ZONE;
3419 "Ignoring invalid numa_zonelist_order value: "
3426 static __init int setup_numa_zonelist_order(char *s)
3433 ret = __parse_numa_zonelist_order(s);
3435 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3439 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3442 * sysctl handler for numa_zonelist_order
3444 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3445 void __user *buffer, size_t *length,
3448 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3450 static DEFINE_MUTEX(zl_order_mutex);
3452 mutex_lock(&zl_order_mutex);
3454 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3458 strcpy(saved_string, (char *)table->data);
3460 ret = proc_dostring(table, write, buffer, length, ppos);
3464 int oldval = user_zonelist_order;
3466 ret = __parse_numa_zonelist_order((char *)table->data);
3469 * bogus value. restore saved string
3471 strncpy((char *)table->data, saved_string,
3472 NUMA_ZONELIST_ORDER_LEN);
3473 user_zonelist_order = oldval;
3474 } else if (oldval != user_zonelist_order) {
3475 mutex_lock(&zonelists_mutex);
3476 build_all_zonelists(NULL, NULL);
3477 mutex_unlock(&zonelists_mutex);
3481 mutex_unlock(&zl_order_mutex);
3486 #define MAX_NODE_LOAD (nr_online_nodes)
3487 static int node_load[MAX_NUMNODES];
3490 * find_next_best_node - find the next node that should appear in a given node's fallback list
3491 * @node: node whose fallback list we're appending
3492 * @used_node_mask: nodemask_t of already used nodes
3494 * We use a number of factors to determine which is the next node that should
3495 * appear on a given node's fallback list. The node should not have appeared
3496 * already in @node's fallback list, and it should be the next closest node
3497 * according to the distance array (which contains arbitrary distance values
3498 * from each node to each node in the system), and should also prefer nodes
3499 * with no CPUs, since presumably they'll have very little allocation pressure
3500 * on them otherwise.
3501 * It returns -1 if no node is found.
3503 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3506 int min_val = INT_MAX;
3507 int best_node = NUMA_NO_NODE;
3508 const struct cpumask *tmp = cpumask_of_node(0);
3510 /* Use the local node if we haven't already */
3511 if (!node_isset(node, *used_node_mask)) {
3512 node_set(node, *used_node_mask);
3516 for_each_node_state(n, N_MEMORY) {
3518 /* Don't want a node to appear more than once */
3519 if (node_isset(n, *used_node_mask))
3522 /* Use the distance array to find the distance */
3523 val = node_distance(node, n);
3525 /* Penalize nodes under us ("prefer the next node") */
3528 /* Give preference to headless and unused nodes */
3529 tmp = cpumask_of_node(n);
3530 if (!cpumask_empty(tmp))
3531 val += PENALTY_FOR_NODE_WITH_CPUS;
3533 /* Slight preference for less loaded node */
3534 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3535 val += node_load[n];
3537 if (val < min_val) {
3544 node_set(best_node, *used_node_mask);
3551 * Build zonelists ordered by node and zones within node.
3552 * This results in maximum locality--normal zone overflows into local
3553 * DMA zone, if any--but risks exhausting DMA zone.
3555 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3558 struct zonelist *zonelist;
3560 zonelist = &pgdat->node_zonelists[0];
3561 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3563 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3564 zonelist->_zonerefs[j].zone = NULL;
3565 zonelist->_zonerefs[j].zone_idx = 0;
3569 * Build gfp_thisnode zonelists
3571 static void build_thisnode_zonelists(pg_data_t *pgdat)
3574 struct zonelist *zonelist;
3576 zonelist = &pgdat->node_zonelists[1];
3577 j = build_zonelists_node(pgdat, zonelist, 0);
3578 zonelist->_zonerefs[j].zone = NULL;
3579 zonelist->_zonerefs[j].zone_idx = 0;
3583 * Build zonelists ordered by zone and nodes within zones.
3584 * This results in conserving DMA zone[s] until all Normal memory is
3585 * exhausted, but results in overflowing to remote node while memory
3586 * may still exist in local DMA zone.
3588 static int node_order[MAX_NUMNODES];
3590 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3593 int zone_type; /* needs to be signed */
3595 struct zonelist *zonelist;
3597 zonelist = &pgdat->node_zonelists[0];
3599 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3600 for (j = 0; j < nr_nodes; j++) {
3601 node = node_order[j];
3602 z = &NODE_DATA(node)->node_zones[zone_type];
3603 if (populated_zone(z)) {
3605 &zonelist->_zonerefs[pos++]);
3606 check_highest_zone(zone_type);
3610 zonelist->_zonerefs[pos].zone = NULL;
3611 zonelist->_zonerefs[pos].zone_idx = 0;
3614 static int default_zonelist_order(void)
3617 unsigned long low_kmem_size, total_size;
3621 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3622 * If they are really small and used heavily, the system can fall
3623 * into OOM very easily.
3624 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3626 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3629 for_each_online_node(nid) {
3630 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3631 z = &NODE_DATA(nid)->node_zones[zone_type];
3632 if (populated_zone(z)) {
3633 if (zone_type < ZONE_NORMAL)
3634 low_kmem_size += z->managed_pages;
3635 total_size += z->managed_pages;
3636 } else if (zone_type == ZONE_NORMAL) {
3638 * If any node has only lowmem, then node order
3639 * is preferred to allow kernel allocations
3640 * locally; otherwise, they can easily infringe
3641 * on other nodes when there is an abundance of
3642 * lowmem available to allocate from.
3644 return ZONELIST_ORDER_NODE;
3648 if (!low_kmem_size || /* there are no DMA area. */
3649 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3650 return ZONELIST_ORDER_NODE;
3652 * look into each node's config.
3653 * If there is a node whose DMA/DMA32 memory is very big area on
3654 * local memory, NODE_ORDER may be suitable.
3656 average_size = total_size /
3657 (nodes_weight(node_states[N_MEMORY]) + 1);
3658 for_each_online_node(nid) {
3661 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3662 z = &NODE_DATA(nid)->node_zones[zone_type];
3663 if (populated_zone(z)) {
3664 if (zone_type < ZONE_NORMAL)
3665 low_kmem_size += z->present_pages;
3666 total_size += z->present_pages;
3669 if (low_kmem_size &&
3670 total_size > average_size && /* ignore small node */
3671 low_kmem_size > total_size * 70/100)
3672 return ZONELIST_ORDER_NODE;
3674 return ZONELIST_ORDER_ZONE;
3677 static void set_zonelist_order(void)
3679 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3680 current_zonelist_order = default_zonelist_order();
3682 current_zonelist_order = user_zonelist_order;
3685 static void build_zonelists(pg_data_t *pgdat)
3689 nodemask_t used_mask;
3690 int local_node, prev_node;
3691 struct zonelist *zonelist;
3692 int order = current_zonelist_order;
3694 /* initialize zonelists */
3695 for (i = 0; i < MAX_ZONELISTS; i++) {
3696 zonelist = pgdat->node_zonelists + i;
3697 zonelist->_zonerefs[0].zone = NULL;
3698 zonelist->_zonerefs[0].zone_idx = 0;
3701 /* NUMA-aware ordering of nodes */
3702 local_node = pgdat->node_id;
3703 load = nr_online_nodes;
3704 prev_node = local_node;
3705 nodes_clear(used_mask);
3707 memset(node_order, 0, sizeof(node_order));
3710 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3712 * We don't want to pressure a particular node.
3713 * So adding penalty to the first node in same
3714 * distance group to make it round-robin.
3716 if (node_distance(local_node, node) !=
3717 node_distance(local_node, prev_node))
3718 node_load[node] = load;
3722 if (order == ZONELIST_ORDER_NODE)
3723 build_zonelists_in_node_order(pgdat, node);
3725 node_order[j++] = node; /* remember order */
3728 if (order == ZONELIST_ORDER_ZONE) {
3729 /* calculate node order -- i.e., DMA last! */
3730 build_zonelists_in_zone_order(pgdat, j);
3733 build_thisnode_zonelists(pgdat);
3736 /* Construct the zonelist performance cache - see further mmzone.h */
3737 static void build_zonelist_cache(pg_data_t *pgdat)
3739 struct zonelist *zonelist;
3740 struct zonelist_cache *zlc;
3743 zonelist = &pgdat->node_zonelists[0];
3744 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3745 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3746 for (z = zonelist->_zonerefs; z->zone; z++)
3747 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3750 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3752 * Return node id of node used for "local" allocations.
3753 * I.e., first node id of first zone in arg node's generic zonelist.
3754 * Used for initializing percpu 'numa_mem', which is used primarily
3755 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3757 int local_memory_node(int node)
3761 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3762 gfp_zone(GFP_KERNEL),
3769 #else /* CONFIG_NUMA */
3771 static void set_zonelist_order(void)
3773 current_zonelist_order = ZONELIST_ORDER_ZONE;
3776 static void build_zonelists(pg_data_t *pgdat)
3778 int node, local_node;
3780 struct zonelist *zonelist;
3782 local_node = pgdat->node_id;
3784 zonelist = &pgdat->node_zonelists[0];
3785 j = build_zonelists_node(pgdat, zonelist, 0);
3788 * Now we build the zonelist so that it contains the zones
3789 * of all the other nodes.
3790 * We don't want to pressure a particular node, so when
3791 * building the zones for node N, we make sure that the
3792 * zones coming right after the local ones are those from
3793 * node N+1 (modulo N)
3795 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3796 if (!node_online(node))
3798 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3800 for (node = 0; node < local_node; node++) {
3801 if (!node_online(node))
3803 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3806 zonelist->_zonerefs[j].zone = NULL;
3807 zonelist->_zonerefs[j].zone_idx = 0;
3810 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3811 static void build_zonelist_cache(pg_data_t *pgdat)
3813 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3816 #endif /* CONFIG_NUMA */
3819 * Boot pageset table. One per cpu which is going to be used for all
3820 * zones and all nodes. The parameters will be set in such a way
3821 * that an item put on a list will immediately be handed over to
3822 * the buddy list. This is safe since pageset manipulation is done
3823 * with interrupts disabled.
3825 * The boot_pagesets must be kept even after bootup is complete for
3826 * unused processors and/or zones. They do play a role for bootstrapping
3827 * hotplugged processors.
3829 * zoneinfo_show() and maybe other functions do
3830 * not check if the processor is online before following the pageset pointer.
3831 * Other parts of the kernel may not check if the zone is available.
3833 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3834 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3835 static void setup_zone_pageset(struct zone *zone);
3838 * Global mutex to protect against size modification of zonelists
3839 * as well as to serialize pageset setup for the new populated zone.
3841 DEFINE_MUTEX(zonelists_mutex);
3843 /* return values int ....just for stop_machine() */
3844 static int __build_all_zonelists(void *data)
3848 pg_data_t *self = data;
3851 memset(node_load, 0, sizeof(node_load));
3854 if (self && !node_online(self->node_id)) {
3855 build_zonelists(self);
3856 build_zonelist_cache(self);
3859 for_each_online_node(nid) {
3860 pg_data_t *pgdat = NODE_DATA(nid);
3862 build_zonelists(pgdat);
3863 build_zonelist_cache(pgdat);
3867 * Initialize the boot_pagesets that are going to be used
3868 * for bootstrapping processors. The real pagesets for
3869 * each zone will be allocated later when the per cpu
3870 * allocator is available.
3872 * boot_pagesets are used also for bootstrapping offline
3873 * cpus if the system is already booted because the pagesets
3874 * are needed to initialize allocators on a specific cpu too.
3875 * F.e. the percpu allocator needs the page allocator which
3876 * needs the percpu allocator in order to allocate its pagesets
3877 * (a chicken-egg dilemma).
3879 for_each_possible_cpu(cpu) {
3880 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3882 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3884 * We now know the "local memory node" for each node--
3885 * i.e., the node of the first zone in the generic zonelist.
3886 * Set up numa_mem percpu variable for on-line cpus. During
3887 * boot, only the boot cpu should be on-line; we'll init the
3888 * secondary cpus' numa_mem as they come on-line. During
3889 * node/memory hotplug, we'll fixup all on-line cpus.
3891 if (cpu_online(cpu))
3892 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3900 * Called with zonelists_mutex held always
3901 * unless system_state == SYSTEM_BOOTING.
3903 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3905 set_zonelist_order();
3907 if (system_state == SYSTEM_BOOTING) {
3908 __build_all_zonelists(NULL);
3909 mminit_verify_zonelist();
3910 cpuset_init_current_mems_allowed();
3912 #ifdef CONFIG_MEMORY_HOTPLUG
3914 setup_zone_pageset(zone);
3916 /* we have to stop all cpus to guarantee there is no user
3918 stop_machine(__build_all_zonelists, pgdat, NULL);
3919 /* cpuset refresh routine should be here */
3921 vm_total_pages = nr_free_pagecache_pages();
3923 * Disable grouping by mobility if the number of pages in the
3924 * system is too low to allow the mechanism to work. It would be
3925 * more accurate, but expensive to check per-zone. This check is
3926 * made on memory-hotadd so a system can start with mobility
3927 * disabled and enable it later
3929 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3930 page_group_by_mobility_disabled = 1;
3932 page_group_by_mobility_disabled = 0;
3934 printk("Built %i zonelists in %s order, mobility grouping %s. "
3935 "Total pages: %ld\n",
3937 zonelist_order_name[current_zonelist_order],
3938 page_group_by_mobility_disabled ? "off" : "on",
3941 printk("Policy zone: %s\n", zone_names[policy_zone]);
3946 * Helper functions to size the waitqueue hash table.
3947 * Essentially these want to choose hash table sizes sufficiently
3948 * large so that collisions trying to wait on pages are rare.
3949 * But in fact, the number of active page waitqueues on typical
3950 * systems is ridiculously low, less than 200. So this is even
3951 * conservative, even though it seems large.
3953 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3954 * waitqueues, i.e. the size of the waitq table given the number of pages.
3956 #define PAGES_PER_WAITQUEUE 256
3958 #ifndef CONFIG_MEMORY_HOTPLUG
3959 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3961 unsigned long size = 1;
3963 pages /= PAGES_PER_WAITQUEUE;
3965 while (size < pages)
3969 * Once we have dozens or even hundreds of threads sleeping
3970 * on IO we've got bigger problems than wait queue collision.
3971 * Limit the size of the wait table to a reasonable size.
3973 size = min(size, 4096UL);
3975 return max(size, 4UL);
3979 * A zone's size might be changed by hot-add, so it is not possible to determine
3980 * a suitable size for its wait_table. So we use the maximum size now.
3982 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3984 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3985 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3986 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3988 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3989 * or more by the traditional way. (See above). It equals:
3991 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3992 * ia64(16K page size) : = ( 8G + 4M)byte.
3993 * powerpc (64K page size) : = (32G +16M)byte.
3995 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4002 * This is an integer logarithm so that shifts can be used later
4003 * to extract the more random high bits from the multiplicative
4004 * hash function before the remainder is taken.
4006 static inline unsigned long wait_table_bits(unsigned long size)
4012 * Check if a pageblock contains reserved pages
4014 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4018 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4019 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4026 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4027 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4028 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4029 * higher will lead to a bigger reserve which will get freed as contiguous
4030 * blocks as reclaim kicks in
4032 static void setup_zone_migrate_reserve(struct zone *zone)
4034 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4036 unsigned long block_migratetype;
4041 * Get the start pfn, end pfn and the number of blocks to reserve
4042 * We have to be careful to be aligned to pageblock_nr_pages to
4043 * make sure that we always check pfn_valid for the first page in
4046 start_pfn = zone->zone_start_pfn;
4047 end_pfn = zone_end_pfn(zone);
4048 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4049 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4053 * Reserve blocks are generally in place to help high-order atomic
4054 * allocations that are short-lived. A min_free_kbytes value that
4055 * would result in more than 2 reserve blocks for atomic allocations
4056 * is assumed to be in place to help anti-fragmentation for the
4057 * future allocation of hugepages at runtime.
4059 reserve = min(2, reserve);
4060 old_reserve = zone->nr_migrate_reserve_block;
4062 /* When memory hot-add, we almost always need to do nothing */
4063 if (reserve == old_reserve)
4065 zone->nr_migrate_reserve_block = reserve;
4067 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4068 if (!pfn_valid(pfn))
4070 page = pfn_to_page(pfn);
4072 /* Watch out for overlapping nodes */
4073 if (page_to_nid(page) != zone_to_nid(zone))
4076 block_migratetype = get_pageblock_migratetype(page);
4078 /* Only test what is necessary when the reserves are not met */
4081 * Blocks with reserved pages will never free, skip
4084 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4085 if (pageblock_is_reserved(pfn, block_end_pfn))
4088 /* If this block is reserved, account for it */
4089 if (block_migratetype == MIGRATE_RESERVE) {
4094 /* Suitable for reserving if this block is movable */
4095 if (block_migratetype == MIGRATE_MOVABLE) {
4096 set_pageblock_migratetype(page,
4098 move_freepages_block(zone, page,
4103 } else if (!old_reserve) {
4105 * At boot time we don't need to scan the whole zone
4106 * for turning off MIGRATE_RESERVE.
4112 * If the reserve is met and this is a previous reserved block,
4115 if (block_migratetype == MIGRATE_RESERVE) {
4116 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4117 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4123 * Initially all pages are reserved - free ones are freed
4124 * up by free_all_bootmem() once the early boot process is
4125 * done. Non-atomic initialization, single-pass.
4127 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4128 unsigned long start_pfn, enum memmap_context context)
4131 unsigned long end_pfn = start_pfn + size;
4135 if (highest_memmap_pfn < end_pfn - 1)
4136 highest_memmap_pfn = end_pfn - 1;
4138 z = &NODE_DATA(nid)->node_zones[zone];
4139 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4141 * There can be holes in boot-time mem_map[]s
4142 * handed to this function. They do not
4143 * exist on hotplugged memory.
4145 if (context == MEMMAP_EARLY) {
4146 if (!early_pfn_valid(pfn))
4148 if (!early_pfn_in_nid(pfn, nid))
4151 page = pfn_to_page(pfn);
4152 set_page_links(page, zone, nid, pfn);
4153 mminit_verify_page_links(page, zone, nid, pfn);
4154 init_page_count(page);
4155 page_mapcount_reset(page);
4156 page_cpupid_reset_last(page);
4157 SetPageReserved(page);
4159 * Mark the block movable so that blocks are reserved for
4160 * movable at startup. This will force kernel allocations
4161 * to reserve their blocks rather than leaking throughout
4162 * the address space during boot when many long-lived
4163 * kernel allocations are made. Later some blocks near
4164 * the start are marked MIGRATE_RESERVE by
4165 * setup_zone_migrate_reserve()
4167 * bitmap is created for zone's valid pfn range. but memmap
4168 * can be created for invalid pages (for alignment)
4169 * check here not to call set_pageblock_migratetype() against
4172 if ((z->zone_start_pfn <= pfn)
4173 && (pfn < zone_end_pfn(z))
4174 && !(pfn & (pageblock_nr_pages - 1)))
4175 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4177 INIT_LIST_HEAD(&page->lru);
4178 #ifdef WANT_PAGE_VIRTUAL
4179 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4180 if (!is_highmem_idx(zone))
4181 set_page_address(page, __va(pfn << PAGE_SHIFT));
4186 static void __meminit zone_init_free_lists(struct zone *zone)
4188 unsigned int order, t;
4189 for_each_migratetype_order(order, t) {
4190 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4191 zone->free_area[order].nr_free = 0;
4195 #ifndef __HAVE_ARCH_MEMMAP_INIT
4196 #define memmap_init(size, nid, zone, start_pfn) \
4197 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4200 static int zone_batchsize(struct zone *zone)
4206 * The per-cpu-pages pools are set to around 1000th of the
4207 * size of the zone. But no more than 1/2 of a meg.
4209 * OK, so we don't know how big the cache is. So guess.
4211 batch = zone->managed_pages / 1024;
4212 if (batch * PAGE_SIZE > 512 * 1024)
4213 batch = (512 * 1024) / PAGE_SIZE;
4214 batch /= 4; /* We effectively *= 4 below */
4219 * Clamp the batch to a 2^n - 1 value. Having a power
4220 * of 2 value was found to be more likely to have
4221 * suboptimal cache aliasing properties in some cases.
4223 * For example if 2 tasks are alternately allocating
4224 * batches of pages, one task can end up with a lot
4225 * of pages of one half of the possible page colors
4226 * and the other with pages of the other colors.
4228 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4233 /* The deferral and batching of frees should be suppressed under NOMMU
4236 * The problem is that NOMMU needs to be able to allocate large chunks
4237 * of contiguous memory as there's no hardware page translation to
4238 * assemble apparent contiguous memory from discontiguous pages.
4240 * Queueing large contiguous runs of pages for batching, however,
4241 * causes the pages to actually be freed in smaller chunks. As there
4242 * can be a significant delay between the individual batches being
4243 * recycled, this leads to the once large chunks of space being
4244 * fragmented and becoming unavailable for high-order allocations.
4251 * pcp->high and pcp->batch values are related and dependent on one another:
4252 * ->batch must never be higher then ->high.
4253 * The following function updates them in a safe manner without read side
4256 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4257 * those fields changing asynchronously (acording the the above rule).
4259 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4260 * outside of boot time (or some other assurance that no concurrent updaters
4263 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4264 unsigned long batch)
4266 /* start with a fail safe value for batch */
4270 /* Update high, then batch, in order */
4277 /* a companion to pageset_set_high() */
4278 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4280 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4283 static void pageset_init(struct per_cpu_pageset *p)
4285 struct per_cpu_pages *pcp;
4288 memset(p, 0, sizeof(*p));
4292 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4293 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4296 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4299 pageset_set_batch(p, batch);
4303 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4304 * to the value high for the pageset p.
4306 static void pageset_set_high(struct per_cpu_pageset *p,
4309 unsigned long batch = max(1UL, high / 4);
4310 if ((high / 4) > (PAGE_SHIFT * 8))
4311 batch = PAGE_SHIFT * 8;
4313 pageset_update(&p->pcp, high, batch);
4316 static void pageset_set_high_and_batch(struct zone *zone,
4317 struct per_cpu_pageset *pcp)
4319 if (percpu_pagelist_fraction)
4320 pageset_set_high(pcp,
4321 (zone->managed_pages /
4322 percpu_pagelist_fraction));
4324 pageset_set_batch(pcp, zone_batchsize(zone));
4327 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4329 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4332 pageset_set_high_and_batch(zone, pcp);
4335 static void __meminit setup_zone_pageset(struct zone *zone)
4338 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4339 for_each_possible_cpu(cpu)
4340 zone_pageset_init(zone, cpu);
4344 * Allocate per cpu pagesets and initialize them.
4345 * Before this call only boot pagesets were available.
4347 void __init setup_per_cpu_pageset(void)
4351 for_each_populated_zone(zone)
4352 setup_zone_pageset(zone);
4355 static noinline __init_refok
4356 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4362 * The per-page waitqueue mechanism uses hashed waitqueues
4365 zone->wait_table_hash_nr_entries =
4366 wait_table_hash_nr_entries(zone_size_pages);
4367 zone->wait_table_bits =
4368 wait_table_bits(zone->wait_table_hash_nr_entries);
4369 alloc_size = zone->wait_table_hash_nr_entries
4370 * sizeof(wait_queue_head_t);
4372 if (!slab_is_available()) {
4373 zone->wait_table = (wait_queue_head_t *)
4374 memblock_virt_alloc_node_nopanic(
4375 alloc_size, zone->zone_pgdat->node_id);
4378 * This case means that a zone whose size was 0 gets new memory
4379 * via memory hot-add.
4380 * But it may be the case that a new node was hot-added. In
4381 * this case vmalloc() will not be able to use this new node's
4382 * memory - this wait_table must be initialized to use this new
4383 * node itself as well.
4384 * To use this new node's memory, further consideration will be
4387 zone->wait_table = vmalloc(alloc_size);
4389 if (!zone->wait_table)
4392 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4393 init_waitqueue_head(zone->wait_table + i);
4398 static __meminit void zone_pcp_init(struct zone *zone)
4401 * per cpu subsystem is not up at this point. The following code
4402 * relies on the ability of the linker to provide the
4403 * offset of a (static) per cpu variable into the per cpu area.
4405 zone->pageset = &boot_pageset;
4407 if (populated_zone(zone))
4408 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4409 zone->name, zone->present_pages,
4410 zone_batchsize(zone));
4413 int __meminit init_currently_empty_zone(struct zone *zone,
4414 unsigned long zone_start_pfn,
4416 enum memmap_context context)
4418 struct pglist_data *pgdat = zone->zone_pgdat;
4420 ret = zone_wait_table_init(zone, size);
4423 pgdat->nr_zones = zone_idx(zone) + 1;
4425 zone->zone_start_pfn = zone_start_pfn;
4427 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4428 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4430 (unsigned long)zone_idx(zone),
4431 zone_start_pfn, (zone_start_pfn + size));
4433 zone_init_free_lists(zone);
4438 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4439 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4441 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4443 int __meminit __early_pfn_to_nid(unsigned long pfn)
4445 unsigned long start_pfn, end_pfn;
4448 * NOTE: The following SMP-unsafe globals are only used early in boot
4449 * when the kernel is running single-threaded.
4451 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4452 static int __meminitdata last_nid;
4454 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4457 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4459 last_start_pfn = start_pfn;
4460 last_end_pfn = end_pfn;
4466 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4468 int __meminit early_pfn_to_nid(unsigned long pfn)
4472 nid = __early_pfn_to_nid(pfn);
4475 /* just returns 0 */
4479 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4480 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4484 nid = __early_pfn_to_nid(pfn);
4485 if (nid >= 0 && nid != node)
4492 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4493 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4494 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4496 * If an architecture guarantees that all ranges registered contain no holes
4497 * and may be freed, this this function may be used instead of calling
4498 * memblock_free_early_nid() manually.
4500 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4502 unsigned long start_pfn, end_pfn;
4505 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4506 start_pfn = min(start_pfn, max_low_pfn);
4507 end_pfn = min(end_pfn, max_low_pfn);
4509 if (start_pfn < end_pfn)
4510 memblock_free_early_nid(PFN_PHYS(start_pfn),
4511 (end_pfn - start_pfn) << PAGE_SHIFT,
4517 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4518 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4520 * If an architecture guarantees that all ranges registered contain no holes and may
4521 * be freed, this function may be used instead of calling memory_present() manually.
4523 void __init sparse_memory_present_with_active_regions(int nid)
4525 unsigned long start_pfn, end_pfn;
4528 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4529 memory_present(this_nid, start_pfn, end_pfn);
4533 * get_pfn_range_for_nid - Return the start and end page frames for a node
4534 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4535 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4536 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4538 * It returns the start and end page frame of a node based on information
4539 * provided by memblock_set_node(). If called for a node
4540 * with no available memory, a warning is printed and the start and end
4543 void __meminit get_pfn_range_for_nid(unsigned int nid,
4544 unsigned long *start_pfn, unsigned long *end_pfn)
4546 unsigned long this_start_pfn, this_end_pfn;
4552 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4553 *start_pfn = min(*start_pfn, this_start_pfn);
4554 *end_pfn = max(*end_pfn, this_end_pfn);
4557 if (*start_pfn == -1UL)
4562 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4563 * assumption is made that zones within a node are ordered in monotonic
4564 * increasing memory addresses so that the "highest" populated zone is used
4566 static void __init find_usable_zone_for_movable(void)
4569 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4570 if (zone_index == ZONE_MOVABLE)
4573 if (arch_zone_highest_possible_pfn[zone_index] >
4574 arch_zone_lowest_possible_pfn[zone_index])
4578 VM_BUG_ON(zone_index == -1);
4579 movable_zone = zone_index;
4583 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4584 * because it is sized independent of architecture. Unlike the other zones,
4585 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4586 * in each node depending on the size of each node and how evenly kernelcore
4587 * is distributed. This helper function adjusts the zone ranges
4588 * provided by the architecture for a given node by using the end of the
4589 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4590 * zones within a node are in order of monotonic increases memory addresses
4592 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4593 unsigned long zone_type,
4594 unsigned long node_start_pfn,
4595 unsigned long node_end_pfn,
4596 unsigned long *zone_start_pfn,
4597 unsigned long *zone_end_pfn)
4599 /* Only adjust if ZONE_MOVABLE is on this node */
4600 if (zone_movable_pfn[nid]) {
4601 /* Size ZONE_MOVABLE */
4602 if (zone_type == ZONE_MOVABLE) {
4603 *zone_start_pfn = zone_movable_pfn[nid];
4604 *zone_end_pfn = min(node_end_pfn,
4605 arch_zone_highest_possible_pfn[movable_zone]);
4607 /* Adjust for ZONE_MOVABLE starting within this range */
4608 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4609 *zone_end_pfn > zone_movable_pfn[nid]) {
4610 *zone_end_pfn = zone_movable_pfn[nid];
4612 /* Check if this whole range is within ZONE_MOVABLE */
4613 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4614 *zone_start_pfn = *zone_end_pfn;
4619 * Return the number of pages a zone spans in a node, including holes
4620 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4622 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4623 unsigned long zone_type,
4624 unsigned long node_start_pfn,
4625 unsigned long node_end_pfn,
4626 unsigned long *ignored)
4628 unsigned long zone_start_pfn, zone_end_pfn;
4630 /* Get the start and end of the zone */
4631 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4632 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4633 adjust_zone_range_for_zone_movable(nid, zone_type,
4634 node_start_pfn, node_end_pfn,
4635 &zone_start_pfn, &zone_end_pfn);
4637 /* Check that this node has pages within the zone's required range */
4638 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4641 /* Move the zone boundaries inside the node if necessary */
4642 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4643 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4645 /* Return the spanned pages */
4646 return zone_end_pfn - zone_start_pfn;
4650 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4651 * then all holes in the requested range will be accounted for.
4653 unsigned long __meminit __absent_pages_in_range(int nid,
4654 unsigned long range_start_pfn,
4655 unsigned long range_end_pfn)
4657 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4658 unsigned long start_pfn, end_pfn;
4661 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4662 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4663 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4664 nr_absent -= end_pfn - start_pfn;
4670 * absent_pages_in_range - Return number of page frames in holes within a range
4671 * @start_pfn: The start PFN to start searching for holes
4672 * @end_pfn: The end PFN to stop searching for holes
4674 * It returns the number of pages frames in memory holes within a range.
4676 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4677 unsigned long end_pfn)
4679 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4682 /* Return the number of page frames in holes in a zone on a node */
4683 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4684 unsigned long zone_type,
4685 unsigned long node_start_pfn,
4686 unsigned long node_end_pfn,
4687 unsigned long *ignored)
4689 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4690 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4691 unsigned long zone_start_pfn, zone_end_pfn;
4693 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4694 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4696 adjust_zone_range_for_zone_movable(nid, zone_type,
4697 node_start_pfn, node_end_pfn,
4698 &zone_start_pfn, &zone_end_pfn);
4699 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4702 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4703 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4704 unsigned long zone_type,
4705 unsigned long node_start_pfn,
4706 unsigned long node_end_pfn,
4707 unsigned long *zones_size)
4709 return zones_size[zone_type];
4712 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4713 unsigned long zone_type,
4714 unsigned long node_start_pfn,
4715 unsigned long node_end_pfn,
4716 unsigned long *zholes_size)
4721 return zholes_size[zone_type];
4724 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4726 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4727 unsigned long node_start_pfn,
4728 unsigned long node_end_pfn,
4729 unsigned long *zones_size,
4730 unsigned long *zholes_size)
4732 unsigned long realtotalpages, totalpages = 0;
4735 for (i = 0; i < MAX_NR_ZONES; i++)
4736 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4740 pgdat->node_spanned_pages = totalpages;
4742 realtotalpages = totalpages;
4743 for (i = 0; i < MAX_NR_ZONES; i++)
4745 zone_absent_pages_in_node(pgdat->node_id, i,
4746 node_start_pfn, node_end_pfn,
4748 pgdat->node_present_pages = realtotalpages;
4749 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4753 #ifndef CONFIG_SPARSEMEM
4755 * Calculate the size of the zone->blockflags rounded to an unsigned long
4756 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4757 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4758 * round what is now in bits to nearest long in bits, then return it in
4761 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4763 unsigned long usemapsize;
4765 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4766 usemapsize = roundup(zonesize, pageblock_nr_pages);
4767 usemapsize = usemapsize >> pageblock_order;
4768 usemapsize *= NR_PAGEBLOCK_BITS;
4769 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4771 return usemapsize / 8;
4774 static void __init setup_usemap(struct pglist_data *pgdat,
4776 unsigned long zone_start_pfn,
4777 unsigned long zonesize)
4779 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4780 zone->pageblock_flags = NULL;
4782 zone->pageblock_flags =
4783 memblock_virt_alloc_node_nopanic(usemapsize,
4787 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4788 unsigned long zone_start_pfn, unsigned long zonesize) {}
4789 #endif /* CONFIG_SPARSEMEM */
4791 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4793 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4794 void __paginginit set_pageblock_order(void)
4798 /* Check that pageblock_nr_pages has not already been setup */
4799 if (pageblock_order)
4802 if (HPAGE_SHIFT > PAGE_SHIFT)
4803 order = HUGETLB_PAGE_ORDER;
4805 order = MAX_ORDER - 1;
4808 * Assume the largest contiguous order of interest is a huge page.
4809 * This value may be variable depending on boot parameters on IA64 and
4812 pageblock_order = order;
4814 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4817 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4818 * is unused as pageblock_order is set at compile-time. See
4819 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4822 void __paginginit set_pageblock_order(void)
4826 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4828 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4829 unsigned long present_pages)
4831 unsigned long pages = spanned_pages;
4834 * Provide a more accurate estimation if there are holes within
4835 * the zone and SPARSEMEM is in use. If there are holes within the
4836 * zone, each populated memory region may cost us one or two extra
4837 * memmap pages due to alignment because memmap pages for each
4838 * populated regions may not naturally algined on page boundary.
4839 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4841 if (spanned_pages > present_pages + (present_pages >> 4) &&
4842 IS_ENABLED(CONFIG_SPARSEMEM))
4843 pages = present_pages;
4845 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4849 * Set up the zone data structures:
4850 * - mark all pages reserved
4851 * - mark all memory queues empty
4852 * - clear the memory bitmaps
4854 * NOTE: pgdat should get zeroed by caller.
4856 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4857 unsigned long node_start_pfn, unsigned long node_end_pfn,
4858 unsigned long *zones_size, unsigned long *zholes_size)
4861 int nid = pgdat->node_id;
4862 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4865 pgdat_resize_init(pgdat);
4866 #ifdef CONFIG_NUMA_BALANCING
4867 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4868 pgdat->numabalancing_migrate_nr_pages = 0;
4869 pgdat->numabalancing_migrate_next_window = jiffies;
4871 init_waitqueue_head(&pgdat->kswapd_wait);
4872 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4873 pgdat_page_cgroup_init(pgdat);
4875 for (j = 0; j < MAX_NR_ZONES; j++) {
4876 struct zone *zone = pgdat->node_zones + j;
4877 unsigned long size, realsize, freesize, memmap_pages;
4879 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4880 node_end_pfn, zones_size);
4881 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4887 * Adjust freesize so that it accounts for how much memory
4888 * is used by this zone for memmap. This affects the watermark
4889 * and per-cpu initialisations
4891 memmap_pages = calc_memmap_size(size, realsize);
4892 if (freesize >= memmap_pages) {
4893 freesize -= memmap_pages;
4896 " %s zone: %lu pages used for memmap\n",
4897 zone_names[j], memmap_pages);
4900 " %s zone: %lu pages exceeds freesize %lu\n",
4901 zone_names[j], memmap_pages, freesize);
4903 /* Account for reserved pages */
4904 if (j == 0 && freesize > dma_reserve) {
4905 freesize -= dma_reserve;
4906 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4907 zone_names[0], dma_reserve);
4910 if (!is_highmem_idx(j))
4911 nr_kernel_pages += freesize;
4912 /* Charge for highmem memmap if there are enough kernel pages */
4913 else if (nr_kernel_pages > memmap_pages * 2)
4914 nr_kernel_pages -= memmap_pages;
4915 nr_all_pages += freesize;
4917 zone->spanned_pages = size;
4918 zone->present_pages = realsize;
4920 * Set an approximate value for lowmem here, it will be adjusted
4921 * when the bootmem allocator frees pages into the buddy system.
4922 * And all highmem pages will be managed by the buddy system.
4924 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4927 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4929 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4931 zone->name = zone_names[j];
4932 spin_lock_init(&zone->lock);
4933 spin_lock_init(&zone->lru_lock);
4934 zone_seqlock_init(zone);
4935 zone->zone_pgdat = pgdat;
4936 zone_pcp_init(zone);
4938 /* For bootup, initialized properly in watermark setup */
4939 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4941 lruvec_init(&zone->lruvec);
4945 set_pageblock_order();
4946 setup_usemap(pgdat, zone, zone_start_pfn, size);
4947 ret = init_currently_empty_zone(zone, zone_start_pfn,
4948 size, MEMMAP_EARLY);
4950 memmap_init(size, nid, j, zone_start_pfn);
4951 zone_start_pfn += size;
4955 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4957 /* Skip empty nodes */
4958 if (!pgdat->node_spanned_pages)
4961 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4962 /* ia64 gets its own node_mem_map, before this, without bootmem */
4963 if (!pgdat->node_mem_map) {
4964 unsigned long size, start, end;
4968 * The zone's endpoints aren't required to be MAX_ORDER
4969 * aligned but the node_mem_map endpoints must be in order
4970 * for the buddy allocator to function correctly.
4972 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4973 end = pgdat_end_pfn(pgdat);
4974 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4975 size = (end - start) * sizeof(struct page);
4976 map = alloc_remap(pgdat->node_id, size);
4978 map = memblock_virt_alloc_node_nopanic(size,
4980 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4982 #ifndef CONFIG_NEED_MULTIPLE_NODES
4984 * With no DISCONTIG, the global mem_map is just set as node 0's
4986 if (pgdat == NODE_DATA(0)) {
4987 mem_map = NODE_DATA(0)->node_mem_map;
4988 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4989 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4990 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4991 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4994 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4997 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4998 unsigned long node_start_pfn, unsigned long *zholes_size)
5000 pg_data_t *pgdat = NODE_DATA(nid);
5001 unsigned long start_pfn = 0;
5002 unsigned long end_pfn = 0;
5004 /* pg_data_t should be reset to zero when it's allocated */
5005 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5007 pgdat->node_id = nid;
5008 pgdat->node_start_pfn = node_start_pfn;
5009 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5010 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5012 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5013 zones_size, zholes_size);
5015 alloc_node_mem_map(pgdat);
5016 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5017 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5018 nid, (unsigned long)pgdat,
5019 (unsigned long)pgdat->node_mem_map);
5022 free_area_init_core(pgdat, start_pfn, end_pfn,
5023 zones_size, zholes_size);
5026 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5028 #if MAX_NUMNODES > 1
5030 * Figure out the number of possible node ids.
5032 void __init setup_nr_node_ids(void)
5035 unsigned int highest = 0;
5037 for_each_node_mask(node, node_possible_map)
5039 nr_node_ids = highest + 1;
5044 * node_map_pfn_alignment - determine the maximum internode alignment
5046 * This function should be called after node map is populated and sorted.
5047 * It calculates the maximum power of two alignment which can distinguish
5050 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5051 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5052 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5053 * shifted, 1GiB is enough and this function will indicate so.
5055 * This is used to test whether pfn -> nid mapping of the chosen memory
5056 * model has fine enough granularity to avoid incorrect mapping for the
5057 * populated node map.
5059 * Returns the determined alignment in pfn's. 0 if there is no alignment
5060 * requirement (single node).
5062 unsigned long __init node_map_pfn_alignment(void)
5064 unsigned long accl_mask = 0, last_end = 0;
5065 unsigned long start, end, mask;
5069 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5070 if (!start || last_nid < 0 || last_nid == nid) {
5077 * Start with a mask granular enough to pin-point to the
5078 * start pfn and tick off bits one-by-one until it becomes
5079 * too coarse to separate the current node from the last.
5081 mask = ~((1 << __ffs(start)) - 1);
5082 while (mask && last_end <= (start & (mask << 1)))
5085 /* accumulate all internode masks */
5089 /* convert mask to number of pages */
5090 return ~accl_mask + 1;
5093 /* Find the lowest pfn for a node */
5094 static unsigned long __init find_min_pfn_for_node(int nid)
5096 unsigned long min_pfn = ULONG_MAX;
5097 unsigned long start_pfn;
5100 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5101 min_pfn = min(min_pfn, start_pfn);
5103 if (min_pfn == ULONG_MAX) {
5105 "Could not find start_pfn for node %d\n", nid);
5113 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5115 * It returns the minimum PFN based on information provided via
5116 * memblock_set_node().
5118 unsigned long __init find_min_pfn_with_active_regions(void)
5120 return find_min_pfn_for_node(MAX_NUMNODES);
5124 * early_calculate_totalpages()
5125 * Sum pages in active regions for movable zone.
5126 * Populate N_MEMORY for calculating usable_nodes.
5128 static unsigned long __init early_calculate_totalpages(void)
5130 unsigned long totalpages = 0;
5131 unsigned long start_pfn, end_pfn;
5134 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5135 unsigned long pages = end_pfn - start_pfn;
5137 totalpages += pages;
5139 node_set_state(nid, N_MEMORY);
5145 * Find the PFN the Movable zone begins in each node. Kernel memory
5146 * is spread evenly between nodes as long as the nodes have enough
5147 * memory. When they don't, some nodes will have more kernelcore than
5150 static void __init find_zone_movable_pfns_for_nodes(void)
5153 unsigned long usable_startpfn;
5154 unsigned long kernelcore_node, kernelcore_remaining;
5155 /* save the state before borrow the nodemask */
5156 nodemask_t saved_node_state = node_states[N_MEMORY];
5157 unsigned long totalpages = early_calculate_totalpages();
5158 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5159 struct memblock_region *r;
5161 /* Need to find movable_zone earlier when movable_node is specified. */
5162 find_usable_zone_for_movable();
5165 * If movable_node is specified, ignore kernelcore and movablecore
5168 if (movable_node_is_enabled()) {
5169 for_each_memblock(memory, r) {
5170 if (!memblock_is_hotpluggable(r))
5175 usable_startpfn = PFN_DOWN(r->base);
5176 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5177 min(usable_startpfn, zone_movable_pfn[nid]) :
5185 * If movablecore=nn[KMG] was specified, calculate what size of
5186 * kernelcore that corresponds so that memory usable for
5187 * any allocation type is evenly spread. If both kernelcore
5188 * and movablecore are specified, then the value of kernelcore
5189 * will be used for required_kernelcore if it's greater than
5190 * what movablecore would have allowed.
5192 if (required_movablecore) {
5193 unsigned long corepages;
5196 * Round-up so that ZONE_MOVABLE is at least as large as what
5197 * was requested by the user
5199 required_movablecore =
5200 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5201 corepages = totalpages - required_movablecore;
5203 required_kernelcore = max(required_kernelcore, corepages);
5206 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5207 if (!required_kernelcore)
5210 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5211 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5214 /* Spread kernelcore memory as evenly as possible throughout nodes */
5215 kernelcore_node = required_kernelcore / usable_nodes;
5216 for_each_node_state(nid, N_MEMORY) {
5217 unsigned long start_pfn, end_pfn;
5220 * Recalculate kernelcore_node if the division per node
5221 * now exceeds what is necessary to satisfy the requested
5222 * amount of memory for the kernel
5224 if (required_kernelcore < kernelcore_node)
5225 kernelcore_node = required_kernelcore / usable_nodes;
5228 * As the map is walked, we track how much memory is usable
5229 * by the kernel using kernelcore_remaining. When it is
5230 * 0, the rest of the node is usable by ZONE_MOVABLE
5232 kernelcore_remaining = kernelcore_node;
5234 /* Go through each range of PFNs within this node */
5235 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5236 unsigned long size_pages;
5238 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5239 if (start_pfn >= end_pfn)
5242 /* Account for what is only usable for kernelcore */
5243 if (start_pfn < usable_startpfn) {
5244 unsigned long kernel_pages;
5245 kernel_pages = min(end_pfn, usable_startpfn)
5248 kernelcore_remaining -= min(kernel_pages,
5249 kernelcore_remaining);
5250 required_kernelcore -= min(kernel_pages,
5251 required_kernelcore);
5253 /* Continue if range is now fully accounted */
5254 if (end_pfn <= usable_startpfn) {
5257 * Push zone_movable_pfn to the end so
5258 * that if we have to rebalance
5259 * kernelcore across nodes, we will
5260 * not double account here
5262 zone_movable_pfn[nid] = end_pfn;
5265 start_pfn = usable_startpfn;
5269 * The usable PFN range for ZONE_MOVABLE is from
5270 * start_pfn->end_pfn. Calculate size_pages as the
5271 * number of pages used as kernelcore
5273 size_pages = end_pfn - start_pfn;
5274 if (size_pages > kernelcore_remaining)
5275 size_pages = kernelcore_remaining;
5276 zone_movable_pfn[nid] = start_pfn + size_pages;
5279 * Some kernelcore has been met, update counts and
5280 * break if the kernelcore for this node has been
5283 required_kernelcore -= min(required_kernelcore,
5285 kernelcore_remaining -= size_pages;
5286 if (!kernelcore_remaining)
5292 * If there is still required_kernelcore, we do another pass with one
5293 * less node in the count. This will push zone_movable_pfn[nid] further
5294 * along on the nodes that still have memory until kernelcore is
5298 if (usable_nodes && required_kernelcore > usable_nodes)
5302 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5303 for (nid = 0; nid < MAX_NUMNODES; nid++)
5304 zone_movable_pfn[nid] =
5305 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5308 /* restore the node_state */
5309 node_states[N_MEMORY] = saved_node_state;
5312 /* Any regular or high memory on that node ? */
5313 static void check_for_memory(pg_data_t *pgdat, int nid)
5315 enum zone_type zone_type;
5317 if (N_MEMORY == N_NORMAL_MEMORY)
5320 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5321 struct zone *zone = &pgdat->node_zones[zone_type];
5322 if (populated_zone(zone)) {
5323 node_set_state(nid, N_HIGH_MEMORY);
5324 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5325 zone_type <= ZONE_NORMAL)
5326 node_set_state(nid, N_NORMAL_MEMORY);
5333 * free_area_init_nodes - Initialise all pg_data_t and zone data
5334 * @max_zone_pfn: an array of max PFNs for each zone
5336 * This will call free_area_init_node() for each active node in the system.
5337 * Using the page ranges provided by memblock_set_node(), the size of each
5338 * zone in each node and their holes is calculated. If the maximum PFN
5339 * between two adjacent zones match, it is assumed that the zone is empty.
5340 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5341 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5342 * starts where the previous one ended. For example, ZONE_DMA32 starts
5343 * at arch_max_dma_pfn.
5345 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5347 unsigned long start_pfn, end_pfn;
5350 /* Record where the zone boundaries are */
5351 memset(arch_zone_lowest_possible_pfn, 0,
5352 sizeof(arch_zone_lowest_possible_pfn));
5353 memset(arch_zone_highest_possible_pfn, 0,
5354 sizeof(arch_zone_highest_possible_pfn));
5355 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5356 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5357 for (i = 1; i < MAX_NR_ZONES; i++) {
5358 if (i == ZONE_MOVABLE)
5360 arch_zone_lowest_possible_pfn[i] =
5361 arch_zone_highest_possible_pfn[i-1];
5362 arch_zone_highest_possible_pfn[i] =
5363 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5365 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5366 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5368 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5369 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5370 find_zone_movable_pfns_for_nodes();
5372 /* Print out the zone ranges */
5373 printk("Zone ranges:\n");
5374 for (i = 0; i < MAX_NR_ZONES; i++) {
5375 if (i == ZONE_MOVABLE)
5377 printk(KERN_CONT " %-8s ", zone_names[i]);
5378 if (arch_zone_lowest_possible_pfn[i] ==
5379 arch_zone_highest_possible_pfn[i])
5380 printk(KERN_CONT "empty\n");
5382 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5383 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5384 (arch_zone_highest_possible_pfn[i]
5385 << PAGE_SHIFT) - 1);
5388 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5389 printk("Movable zone start for each node\n");
5390 for (i = 0; i < MAX_NUMNODES; i++) {
5391 if (zone_movable_pfn[i])
5392 printk(" Node %d: %#010lx\n", i,
5393 zone_movable_pfn[i] << PAGE_SHIFT);
5396 /* Print out the early node map */
5397 printk("Early memory node ranges\n");
5398 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5399 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5400 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5402 /* Initialise every node */
5403 mminit_verify_pageflags_layout();
5404 setup_nr_node_ids();
5405 for_each_online_node(nid) {
5406 pg_data_t *pgdat = NODE_DATA(nid);
5407 free_area_init_node(nid, NULL,
5408 find_min_pfn_for_node(nid), NULL);
5410 /* Any memory on that node */
5411 if (pgdat->node_present_pages)
5412 node_set_state(nid, N_MEMORY);
5413 check_for_memory(pgdat, nid);
5417 static int __init cmdline_parse_core(char *p, unsigned long *core)
5419 unsigned long long coremem;
5423 coremem = memparse(p, &p);
5424 *core = coremem >> PAGE_SHIFT;
5426 /* Paranoid check that UL is enough for the coremem value */
5427 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5433 * kernelcore=size sets the amount of memory for use for allocations that
5434 * cannot be reclaimed or migrated.
5436 static int __init cmdline_parse_kernelcore(char *p)
5438 return cmdline_parse_core(p, &required_kernelcore);
5442 * movablecore=size sets the amount of memory for use for allocations that
5443 * can be reclaimed or migrated.
5445 static int __init cmdline_parse_movablecore(char *p)
5447 return cmdline_parse_core(p, &required_movablecore);
5450 early_param("kernelcore", cmdline_parse_kernelcore);
5451 early_param("movablecore", cmdline_parse_movablecore);
5453 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5455 void adjust_managed_page_count(struct page *page, long count)
5457 spin_lock(&managed_page_count_lock);
5458 page_zone(page)->managed_pages += count;
5459 totalram_pages += count;
5460 #ifdef CONFIG_HIGHMEM
5461 if (PageHighMem(page))
5462 totalhigh_pages += count;
5464 spin_unlock(&managed_page_count_lock);
5466 EXPORT_SYMBOL(adjust_managed_page_count);
5468 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5471 unsigned long pages = 0;
5473 start = (void *)PAGE_ALIGN((unsigned long)start);
5474 end = (void *)((unsigned long)end & PAGE_MASK);
5475 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5476 if ((unsigned int)poison <= 0xFF)
5477 memset(pos, poison, PAGE_SIZE);
5478 free_reserved_page(virt_to_page(pos));
5482 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5483 s, pages << (PAGE_SHIFT - 10), start, end);
5487 EXPORT_SYMBOL(free_reserved_area);
5489 #ifdef CONFIG_HIGHMEM
5490 void free_highmem_page(struct page *page)
5492 __free_reserved_page(page);
5494 page_zone(page)->managed_pages++;
5500 void __init mem_init_print_info(const char *str)
5502 unsigned long physpages, codesize, datasize, rosize, bss_size;
5503 unsigned long init_code_size, init_data_size;
5505 physpages = get_num_physpages();
5506 codesize = _etext - _stext;
5507 datasize = _edata - _sdata;
5508 rosize = __end_rodata - __start_rodata;
5509 bss_size = __bss_stop - __bss_start;
5510 init_data_size = __init_end - __init_begin;
5511 init_code_size = _einittext - _sinittext;
5514 * Detect special cases and adjust section sizes accordingly:
5515 * 1) .init.* may be embedded into .data sections
5516 * 2) .init.text.* may be out of [__init_begin, __init_end],
5517 * please refer to arch/tile/kernel/vmlinux.lds.S.
5518 * 3) .rodata.* may be embedded into .text or .data sections.
5520 #define adj_init_size(start, end, size, pos, adj) \
5522 if (start <= pos && pos < end && size > adj) \
5526 adj_init_size(__init_begin, __init_end, init_data_size,
5527 _sinittext, init_code_size);
5528 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5529 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5530 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5531 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5533 #undef adj_init_size
5535 printk("Memory: %luK/%luK available "
5536 "(%luK kernel code, %luK rwdata, %luK rodata, "
5537 "%luK init, %luK bss, %luK reserved"
5538 #ifdef CONFIG_HIGHMEM
5542 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5543 codesize >> 10, datasize >> 10, rosize >> 10,
5544 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5545 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5546 #ifdef CONFIG_HIGHMEM
5547 totalhigh_pages << (PAGE_SHIFT-10),
5549 str ? ", " : "", str ? str : "");
5553 * set_dma_reserve - set the specified number of pages reserved in the first zone
5554 * @new_dma_reserve: The number of pages to mark reserved
5556 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5557 * In the DMA zone, a significant percentage may be consumed by kernel image
5558 * and other unfreeable allocations which can skew the watermarks badly. This
5559 * function may optionally be used to account for unfreeable pages in the
5560 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5561 * smaller per-cpu batchsize.
5563 void __init set_dma_reserve(unsigned long new_dma_reserve)
5565 dma_reserve = new_dma_reserve;
5568 void __init free_area_init(unsigned long *zones_size)
5570 free_area_init_node(0, zones_size,
5571 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5574 static int page_alloc_cpu_notify(struct notifier_block *self,
5575 unsigned long action, void *hcpu)
5577 int cpu = (unsigned long)hcpu;
5579 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5580 lru_add_drain_cpu(cpu);
5584 * Spill the event counters of the dead processor
5585 * into the current processors event counters.
5586 * This artificially elevates the count of the current
5589 vm_events_fold_cpu(cpu);
5592 * Zero the differential counters of the dead processor
5593 * so that the vm statistics are consistent.
5595 * This is only okay since the processor is dead and cannot
5596 * race with what we are doing.
5598 cpu_vm_stats_fold(cpu);
5603 void __init page_alloc_init(void)
5605 hotcpu_notifier(page_alloc_cpu_notify, 0);
5609 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5610 * or min_free_kbytes changes.
5612 static void calculate_totalreserve_pages(void)
5614 struct pglist_data *pgdat;
5615 unsigned long reserve_pages = 0;
5616 enum zone_type i, j;
5618 for_each_online_pgdat(pgdat) {
5619 for (i = 0; i < MAX_NR_ZONES; i++) {
5620 struct zone *zone = pgdat->node_zones + i;
5623 /* Find valid and maximum lowmem_reserve in the zone */
5624 for (j = i; j < MAX_NR_ZONES; j++) {
5625 if (zone->lowmem_reserve[j] > max)
5626 max = zone->lowmem_reserve[j];
5629 /* we treat the high watermark as reserved pages. */
5630 max += high_wmark_pages(zone);
5632 if (max > zone->managed_pages)
5633 max = zone->managed_pages;
5634 reserve_pages += max;
5636 * Lowmem reserves are not available to
5637 * GFP_HIGHUSER page cache allocations and
5638 * kswapd tries to balance zones to their high
5639 * watermark. As a result, neither should be
5640 * regarded as dirtyable memory, to prevent a
5641 * situation where reclaim has to clean pages
5642 * in order to balance the zones.
5644 zone->dirty_balance_reserve = max;
5647 dirty_balance_reserve = reserve_pages;
5648 totalreserve_pages = reserve_pages;
5652 * setup_per_zone_lowmem_reserve - called whenever
5653 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5654 * has a correct pages reserved value, so an adequate number of
5655 * pages are left in the zone after a successful __alloc_pages().
5657 static void setup_per_zone_lowmem_reserve(void)
5659 struct pglist_data *pgdat;
5660 enum zone_type j, idx;
5662 for_each_online_pgdat(pgdat) {
5663 for (j = 0; j < MAX_NR_ZONES; j++) {
5664 struct zone *zone = pgdat->node_zones + j;
5665 unsigned long managed_pages = zone->managed_pages;
5667 zone->lowmem_reserve[j] = 0;
5671 struct zone *lower_zone;
5675 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5676 sysctl_lowmem_reserve_ratio[idx] = 1;
5678 lower_zone = pgdat->node_zones + idx;
5679 lower_zone->lowmem_reserve[j] = managed_pages /
5680 sysctl_lowmem_reserve_ratio[idx];
5681 managed_pages += lower_zone->managed_pages;
5686 /* update totalreserve_pages */
5687 calculate_totalreserve_pages();
5690 static void __setup_per_zone_wmarks(void)
5692 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5693 unsigned long lowmem_pages = 0;
5695 unsigned long flags;
5697 /* Calculate total number of !ZONE_HIGHMEM pages */
5698 for_each_zone(zone) {
5699 if (!is_highmem(zone))
5700 lowmem_pages += zone->managed_pages;
5703 for_each_zone(zone) {
5706 spin_lock_irqsave(&zone->lock, flags);
5707 tmp = (u64)pages_min * zone->managed_pages;
5708 do_div(tmp, lowmem_pages);
5709 if (is_highmem(zone)) {
5711 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5712 * need highmem pages, so cap pages_min to a small
5715 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5716 * deltas controls asynch page reclaim, and so should
5717 * not be capped for highmem.
5719 unsigned long min_pages;
5721 min_pages = zone->managed_pages / 1024;
5722 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5723 zone->watermark[WMARK_MIN] = min_pages;
5726 * If it's a lowmem zone, reserve a number of pages
5727 * proportionate to the zone's size.
5729 zone->watermark[WMARK_MIN] = tmp;
5732 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5733 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5735 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5736 high_wmark_pages(zone) - low_wmark_pages(zone) -
5737 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5739 setup_zone_migrate_reserve(zone);
5740 spin_unlock_irqrestore(&zone->lock, flags);
5743 /* update totalreserve_pages */
5744 calculate_totalreserve_pages();
5748 * setup_per_zone_wmarks - called when min_free_kbytes changes
5749 * or when memory is hot-{added|removed}
5751 * Ensures that the watermark[min,low,high] values for each zone are set
5752 * correctly with respect to min_free_kbytes.
5754 void setup_per_zone_wmarks(void)
5756 mutex_lock(&zonelists_mutex);
5757 __setup_per_zone_wmarks();
5758 mutex_unlock(&zonelists_mutex);
5762 * The inactive anon list should be small enough that the VM never has to
5763 * do too much work, but large enough that each inactive page has a chance
5764 * to be referenced again before it is swapped out.
5766 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5767 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5768 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5769 * the anonymous pages are kept on the inactive list.
5772 * memory ratio inactive anon
5773 * -------------------------------------
5782 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5784 unsigned int gb, ratio;
5786 /* Zone size in gigabytes */
5787 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5789 ratio = int_sqrt(10 * gb);
5793 zone->inactive_ratio = ratio;
5796 static void __meminit setup_per_zone_inactive_ratio(void)
5801 calculate_zone_inactive_ratio(zone);
5805 * Initialise min_free_kbytes.
5807 * For small machines we want it small (128k min). For large machines
5808 * we want it large (64MB max). But it is not linear, because network
5809 * bandwidth does not increase linearly with machine size. We use
5811 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5812 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5828 int __meminit init_per_zone_wmark_min(void)
5830 unsigned long lowmem_kbytes;
5831 int new_min_free_kbytes;
5833 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5834 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5836 if (new_min_free_kbytes > user_min_free_kbytes) {
5837 min_free_kbytes = new_min_free_kbytes;
5838 if (min_free_kbytes < 128)
5839 min_free_kbytes = 128;
5840 if (min_free_kbytes > 65536)
5841 min_free_kbytes = 65536;
5843 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5844 new_min_free_kbytes, user_min_free_kbytes);
5846 setup_per_zone_wmarks();
5847 refresh_zone_stat_thresholds();
5848 setup_per_zone_lowmem_reserve();
5849 setup_per_zone_inactive_ratio();
5852 module_init(init_per_zone_wmark_min)
5855 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5856 * that we can call two helper functions whenever min_free_kbytes
5859 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5860 void __user *buffer, size_t *length, loff_t *ppos)
5864 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5869 user_min_free_kbytes = min_free_kbytes;
5870 setup_per_zone_wmarks();
5876 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5877 void __user *buffer, size_t *length, loff_t *ppos)
5882 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5887 zone->min_unmapped_pages = (zone->managed_pages *
5888 sysctl_min_unmapped_ratio) / 100;
5892 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5893 void __user *buffer, size_t *length, loff_t *ppos)
5898 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5903 zone->min_slab_pages = (zone->managed_pages *
5904 sysctl_min_slab_ratio) / 100;
5910 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5911 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5912 * whenever sysctl_lowmem_reserve_ratio changes.
5914 * The reserve ratio obviously has absolutely no relation with the
5915 * minimum watermarks. The lowmem reserve ratio can only make sense
5916 * if in function of the boot time zone sizes.
5918 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5919 void __user *buffer, size_t *length, loff_t *ppos)
5921 proc_dointvec_minmax(table, write, buffer, length, ppos);
5922 setup_per_zone_lowmem_reserve();
5927 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5928 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5929 * pagelist can have before it gets flushed back to buddy allocator.
5931 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5932 void __user *buffer, size_t *length, loff_t *ppos)
5935 int old_percpu_pagelist_fraction;
5938 mutex_lock(&pcp_batch_high_lock);
5939 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5941 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5942 if (!write || ret < 0)
5945 /* Sanity checking to avoid pcp imbalance */
5946 if (percpu_pagelist_fraction &&
5947 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
5948 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
5954 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
5957 for_each_populated_zone(zone) {
5960 for_each_possible_cpu(cpu)
5961 pageset_set_high_and_batch(zone,
5962 per_cpu_ptr(zone->pageset, cpu));
5965 mutex_unlock(&pcp_batch_high_lock);
5969 int hashdist = HASHDIST_DEFAULT;
5972 static int __init set_hashdist(char *str)
5976 hashdist = simple_strtoul(str, &str, 0);
5979 __setup("hashdist=", set_hashdist);
5983 * allocate a large system hash table from bootmem
5984 * - it is assumed that the hash table must contain an exact power-of-2
5985 * quantity of entries
5986 * - limit is the number of hash buckets, not the total allocation size
5988 void *__init alloc_large_system_hash(const char *tablename,
5989 unsigned long bucketsize,
5990 unsigned long numentries,
5993 unsigned int *_hash_shift,
5994 unsigned int *_hash_mask,
5995 unsigned long low_limit,
5996 unsigned long high_limit)
5998 unsigned long long max = high_limit;
5999 unsigned long log2qty, size;
6002 /* allow the kernel cmdline to have a say */
6004 /* round applicable memory size up to nearest megabyte */
6005 numentries = nr_kernel_pages;
6007 /* It isn't necessary when PAGE_SIZE >= 1MB */
6008 if (PAGE_SHIFT < 20)
6009 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6011 /* limit to 1 bucket per 2^scale bytes of low memory */
6012 if (scale > PAGE_SHIFT)
6013 numentries >>= (scale - PAGE_SHIFT);
6015 numentries <<= (PAGE_SHIFT - scale);
6017 /* Make sure we've got at least a 0-order allocation.. */
6018 if (unlikely(flags & HASH_SMALL)) {
6019 /* Makes no sense without HASH_EARLY */
6020 WARN_ON(!(flags & HASH_EARLY));
6021 if (!(numentries >> *_hash_shift)) {
6022 numentries = 1UL << *_hash_shift;
6023 BUG_ON(!numentries);
6025 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6026 numentries = PAGE_SIZE / bucketsize;
6028 numentries = roundup_pow_of_two(numentries);
6030 /* limit allocation size to 1/16 total memory by default */
6032 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6033 do_div(max, bucketsize);
6035 max = min(max, 0x80000000ULL);
6037 if (numentries < low_limit)
6038 numentries = low_limit;
6039 if (numentries > max)
6042 log2qty = ilog2(numentries);
6045 size = bucketsize << log2qty;
6046 if (flags & HASH_EARLY)
6047 table = memblock_virt_alloc_nopanic(size, 0);
6049 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6052 * If bucketsize is not a power-of-two, we may free
6053 * some pages at the end of hash table which
6054 * alloc_pages_exact() automatically does
6056 if (get_order(size) < MAX_ORDER) {
6057 table = alloc_pages_exact(size, GFP_ATOMIC);
6058 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6061 } while (!table && size > PAGE_SIZE && --log2qty);
6064 panic("Failed to allocate %s hash table\n", tablename);
6066 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6069 ilog2(size) - PAGE_SHIFT,
6073 *_hash_shift = log2qty;
6075 *_hash_mask = (1 << log2qty) - 1;
6080 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6081 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6084 #ifdef CONFIG_SPARSEMEM
6085 return __pfn_to_section(pfn)->pageblock_flags;
6087 return zone->pageblock_flags;
6088 #endif /* CONFIG_SPARSEMEM */
6091 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6093 #ifdef CONFIG_SPARSEMEM
6094 pfn &= (PAGES_PER_SECTION-1);
6095 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6097 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6098 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6099 #endif /* CONFIG_SPARSEMEM */
6103 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6104 * @page: The page within the block of interest
6105 * @pfn: The target page frame number
6106 * @end_bitidx: The last bit of interest to retrieve
6107 * @mask: mask of bits that the caller is interested in
6109 * Return: pageblock_bits flags
6111 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6112 unsigned long end_bitidx,
6116 unsigned long *bitmap;
6117 unsigned long bitidx, word_bitidx;
6120 zone = page_zone(page);
6121 bitmap = get_pageblock_bitmap(zone, pfn);
6122 bitidx = pfn_to_bitidx(zone, pfn);
6123 word_bitidx = bitidx / BITS_PER_LONG;
6124 bitidx &= (BITS_PER_LONG-1);
6126 word = bitmap[word_bitidx];
6127 bitidx += end_bitidx;
6128 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6132 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6133 * @page: The page within the block of interest
6134 * @flags: The flags to set
6135 * @pfn: The target page frame number
6136 * @end_bitidx: The last bit of interest
6137 * @mask: mask of bits that the caller is interested in
6139 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6141 unsigned long end_bitidx,
6145 unsigned long *bitmap;
6146 unsigned long bitidx, word_bitidx;
6147 unsigned long old_word, word;
6149 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6151 zone = page_zone(page);
6152 bitmap = get_pageblock_bitmap(zone, pfn);
6153 bitidx = pfn_to_bitidx(zone, pfn);
6154 word_bitidx = bitidx / BITS_PER_LONG;
6155 bitidx &= (BITS_PER_LONG-1);
6157 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6159 bitidx += end_bitidx;
6160 mask <<= (BITS_PER_LONG - bitidx - 1);
6161 flags <<= (BITS_PER_LONG - bitidx - 1);
6163 word = ACCESS_ONCE(bitmap[word_bitidx]);
6165 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6166 if (word == old_word)
6173 * This function checks whether pageblock includes unmovable pages or not.
6174 * If @count is not zero, it is okay to include less @count unmovable pages
6176 * PageLRU check without isolation or lru_lock could race so that
6177 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6178 * expect this function should be exact.
6180 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6181 bool skip_hwpoisoned_pages)
6183 unsigned long pfn, iter, found;
6187 * For avoiding noise data, lru_add_drain_all() should be called
6188 * If ZONE_MOVABLE, the zone never contains unmovable pages
6190 if (zone_idx(zone) == ZONE_MOVABLE)
6192 mt = get_pageblock_migratetype(page);
6193 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6196 pfn = page_to_pfn(page);
6197 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6198 unsigned long check = pfn + iter;
6200 if (!pfn_valid_within(check))
6203 page = pfn_to_page(check);
6206 * Hugepages are not in LRU lists, but they're movable.
6207 * We need not scan over tail pages bacause we don't
6208 * handle each tail page individually in migration.
6210 if (PageHuge(page)) {
6211 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6216 * We can't use page_count without pin a page
6217 * because another CPU can free compound page.
6218 * This check already skips compound tails of THP
6219 * because their page->_count is zero at all time.
6221 if (!atomic_read(&page->_count)) {
6222 if (PageBuddy(page))
6223 iter += (1 << page_order(page)) - 1;
6228 * The HWPoisoned page may be not in buddy system, and
6229 * page_count() is not 0.
6231 if (skip_hwpoisoned_pages && PageHWPoison(page))
6237 * If there are RECLAIMABLE pages, we need to check it.
6238 * But now, memory offline itself doesn't call shrink_slab()
6239 * and it still to be fixed.
6242 * If the page is not RAM, page_count()should be 0.
6243 * we don't need more check. This is an _used_ not-movable page.
6245 * The problematic thing here is PG_reserved pages. PG_reserved
6246 * is set to both of a memory hole page and a _used_ kernel
6255 bool is_pageblock_removable_nolock(struct page *page)
6261 * We have to be careful here because we are iterating over memory
6262 * sections which are not zone aware so we might end up outside of
6263 * the zone but still within the section.
6264 * We have to take care about the node as well. If the node is offline
6265 * its NODE_DATA will be NULL - see page_zone.
6267 if (!node_online(page_to_nid(page)))
6270 zone = page_zone(page);
6271 pfn = page_to_pfn(page);
6272 if (!zone_spans_pfn(zone, pfn))
6275 return !has_unmovable_pages(zone, page, 0, true);
6280 static unsigned long pfn_max_align_down(unsigned long pfn)
6282 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6283 pageblock_nr_pages) - 1);
6286 static unsigned long pfn_max_align_up(unsigned long pfn)
6288 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6289 pageblock_nr_pages));
6292 /* [start, end) must belong to a single zone. */
6293 static int __alloc_contig_migrate_range(struct compact_control *cc,
6294 unsigned long start, unsigned long end)
6296 /* This function is based on compact_zone() from compaction.c. */
6297 unsigned long nr_reclaimed;
6298 unsigned long pfn = start;
6299 unsigned int tries = 0;
6304 while (pfn < end || !list_empty(&cc->migratepages)) {
6305 if (fatal_signal_pending(current)) {
6310 if (list_empty(&cc->migratepages)) {
6311 cc->nr_migratepages = 0;
6312 pfn = isolate_migratepages_range(cc, pfn, end);
6318 } else if (++tries == 5) {
6319 ret = ret < 0 ? ret : -EBUSY;
6323 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6325 cc->nr_migratepages -= nr_reclaimed;
6327 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6328 NULL, 0, cc->mode, MR_CMA);
6331 putback_movable_pages(&cc->migratepages);
6338 * alloc_contig_range() -- tries to allocate given range of pages
6339 * @start: start PFN to allocate
6340 * @end: one-past-the-last PFN to allocate
6341 * @migratetype: migratetype of the underlaying pageblocks (either
6342 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6343 * in range must have the same migratetype and it must
6344 * be either of the two.
6346 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6347 * aligned, however it's the caller's responsibility to guarantee that
6348 * we are the only thread that changes migrate type of pageblocks the
6351 * The PFN range must belong to a single zone.
6353 * Returns zero on success or negative error code. On success all
6354 * pages which PFN is in [start, end) are allocated for the caller and
6355 * need to be freed with free_contig_range().
6357 int alloc_contig_range(unsigned long start, unsigned long end,
6358 unsigned migratetype)
6360 unsigned long outer_start, outer_end;
6363 struct compact_control cc = {
6364 .nr_migratepages = 0,
6366 .zone = page_zone(pfn_to_page(start)),
6367 .mode = MIGRATE_SYNC,
6368 .ignore_skip_hint = true,
6370 INIT_LIST_HEAD(&cc.migratepages);
6373 * What we do here is we mark all pageblocks in range as
6374 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6375 * have different sizes, and due to the way page allocator
6376 * work, we align the range to biggest of the two pages so
6377 * that page allocator won't try to merge buddies from
6378 * different pageblocks and change MIGRATE_ISOLATE to some
6379 * other migration type.
6381 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6382 * migrate the pages from an unaligned range (ie. pages that
6383 * we are interested in). This will put all the pages in
6384 * range back to page allocator as MIGRATE_ISOLATE.
6386 * When this is done, we take the pages in range from page
6387 * allocator removing them from the buddy system. This way
6388 * page allocator will never consider using them.
6390 * This lets us mark the pageblocks back as
6391 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6392 * aligned range but not in the unaligned, original range are
6393 * put back to page allocator so that buddy can use them.
6396 ret = start_isolate_page_range(pfn_max_align_down(start),
6397 pfn_max_align_up(end), migratetype,
6402 ret = __alloc_contig_migrate_range(&cc, start, end);
6407 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6408 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6409 * more, all pages in [start, end) are free in page allocator.
6410 * What we are going to do is to allocate all pages from
6411 * [start, end) (that is remove them from page allocator).
6413 * The only problem is that pages at the beginning and at the
6414 * end of interesting range may be not aligned with pages that
6415 * page allocator holds, ie. they can be part of higher order
6416 * pages. Because of this, we reserve the bigger range and
6417 * once this is done free the pages we are not interested in.
6419 * We don't have to hold zone->lock here because the pages are
6420 * isolated thus they won't get removed from buddy.
6423 lru_add_drain_all();
6427 outer_start = start;
6428 while (!PageBuddy(pfn_to_page(outer_start))) {
6429 if (++order >= MAX_ORDER) {
6433 outer_start &= ~0UL << order;
6436 /* Make sure the range is really isolated. */
6437 if (test_pages_isolated(outer_start, end, false)) {
6438 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6445 /* Grab isolated pages from freelists. */
6446 outer_end = isolate_freepages_range(&cc, outer_start, end);
6452 /* Free head and tail (if any) */
6453 if (start != outer_start)
6454 free_contig_range(outer_start, start - outer_start);
6455 if (end != outer_end)
6456 free_contig_range(end, outer_end - end);
6459 undo_isolate_page_range(pfn_max_align_down(start),
6460 pfn_max_align_up(end), migratetype);
6464 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6466 unsigned int count = 0;
6468 for (; nr_pages--; pfn++) {
6469 struct page *page = pfn_to_page(pfn);
6471 count += page_count(page) != 1;
6474 WARN(count != 0, "%d pages are still in use!\n", count);
6478 #ifdef CONFIG_MEMORY_HOTPLUG
6480 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6481 * page high values need to be recalulated.
6483 void __meminit zone_pcp_update(struct zone *zone)
6486 mutex_lock(&pcp_batch_high_lock);
6487 for_each_possible_cpu(cpu)
6488 pageset_set_high_and_batch(zone,
6489 per_cpu_ptr(zone->pageset, cpu));
6490 mutex_unlock(&pcp_batch_high_lock);
6494 void zone_pcp_reset(struct zone *zone)
6496 unsigned long flags;
6498 struct per_cpu_pageset *pset;
6500 /* avoid races with drain_pages() */
6501 local_irq_save(flags);
6502 if (zone->pageset != &boot_pageset) {
6503 for_each_online_cpu(cpu) {
6504 pset = per_cpu_ptr(zone->pageset, cpu);
6505 drain_zonestat(zone, pset);
6507 free_percpu(zone->pageset);
6508 zone->pageset = &boot_pageset;
6510 local_irq_restore(flags);
6513 #ifdef CONFIG_MEMORY_HOTREMOVE
6515 * All pages in the range must be isolated before calling this.
6518 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6522 unsigned int order, i;
6524 unsigned long flags;
6525 /* find the first valid pfn */
6526 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6531 zone = page_zone(pfn_to_page(pfn));
6532 spin_lock_irqsave(&zone->lock, flags);
6534 while (pfn < end_pfn) {
6535 if (!pfn_valid(pfn)) {
6539 page = pfn_to_page(pfn);
6541 * The HWPoisoned page may be not in buddy system, and
6542 * page_count() is not 0.
6544 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6546 SetPageReserved(page);
6550 BUG_ON(page_count(page));
6551 BUG_ON(!PageBuddy(page));
6552 order = page_order(page);
6553 #ifdef CONFIG_DEBUG_VM
6554 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6555 pfn, 1 << order, end_pfn);
6557 list_del(&page->lru);
6558 rmv_page_order(page);
6559 zone->free_area[order].nr_free--;
6560 for (i = 0; i < (1 << order); i++)
6561 SetPageReserved((page+i));
6562 pfn += (1 << order);
6564 spin_unlock_irqrestore(&zone->lock, flags);
6568 #ifdef CONFIG_MEMORY_FAILURE
6569 bool is_free_buddy_page(struct page *page)
6571 struct zone *zone = page_zone(page);
6572 unsigned long pfn = page_to_pfn(page);
6573 unsigned long flags;
6576 spin_lock_irqsave(&zone->lock, flags);
6577 for (order = 0; order < MAX_ORDER; order++) {
6578 struct page *page_head = page - (pfn & ((1 << order) - 1));
6580 if (PageBuddy(page_head) && page_order(page_head) >= order)
6583 spin_unlock_irqrestore(&zone->lock, flags);
6585 return order < MAX_ORDER;
6589 static const struct trace_print_flags pageflag_names[] = {
6590 {1UL << PG_locked, "locked" },
6591 {1UL << PG_error, "error" },
6592 {1UL << PG_referenced, "referenced" },
6593 {1UL << PG_uptodate, "uptodate" },
6594 {1UL << PG_dirty, "dirty" },
6595 {1UL << PG_lru, "lru" },
6596 {1UL << PG_active, "active" },
6597 {1UL << PG_slab, "slab" },
6598 {1UL << PG_owner_priv_1, "owner_priv_1" },
6599 {1UL << PG_arch_1, "arch_1" },
6600 {1UL << PG_reserved, "reserved" },
6601 {1UL << PG_private, "private" },
6602 {1UL << PG_private_2, "private_2" },
6603 {1UL << PG_writeback, "writeback" },
6604 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6605 {1UL << PG_head, "head" },
6606 {1UL << PG_tail, "tail" },
6608 {1UL << PG_compound, "compound" },
6610 {1UL << PG_swapcache, "swapcache" },
6611 {1UL << PG_mappedtodisk, "mappedtodisk" },
6612 {1UL << PG_reclaim, "reclaim" },
6613 {1UL << PG_swapbacked, "swapbacked" },
6614 {1UL << PG_unevictable, "unevictable" },
6616 {1UL << PG_mlocked, "mlocked" },
6618 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6619 {1UL << PG_uncached, "uncached" },
6621 #ifdef CONFIG_MEMORY_FAILURE
6622 {1UL << PG_hwpoison, "hwpoison" },
6624 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6625 {1UL << PG_compound_lock, "compound_lock" },
6629 static void dump_flags(unsigned long flags,
6630 const struct trace_print_flags *names, int count)
6632 const char *delim = "";
6636 printk(KERN_ALERT "flags: %#lx(", flags);
6638 /* remove zone id */
6639 flags &= (1UL << NR_PAGEFLAGS) - 1;
6641 for (i = 0; i < count && flags; i++) {
6643 mask = names[i].mask;
6644 if ((flags & mask) != mask)
6648 printk("%s%s", delim, names[i].name);
6652 /* check for left over flags */
6654 printk("%s%#lx", delim, flags);
6659 void dump_page_badflags(struct page *page, const char *reason,
6660 unsigned long badflags)
6663 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6664 page, atomic_read(&page->_count), page_mapcount(page),
6665 page->mapping, page->index);
6666 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6667 dump_flags(page->flags, pageflag_names, ARRAY_SIZE(pageflag_names));
6669 pr_alert("page dumped because: %s\n", reason);
6670 if (page->flags & badflags) {
6671 pr_alert("bad because of flags:\n");
6672 dump_flags(page->flags & badflags,
6673 pageflag_names, ARRAY_SIZE(pageflag_names));
6675 mem_cgroup_print_bad_page(page);
6678 void dump_page(struct page *page, const char *reason)
6680 dump_page_badflags(page, reason, 0);
6682 EXPORT_SYMBOL(dump_page);
6684 #ifdef CONFIG_DEBUG_VM
6686 static const struct trace_print_flags vmaflags_names[] = {
6688 {VM_WRITE, "write" },
6690 {VM_SHARED, "shared" },
6691 {VM_MAYREAD, "mayread" },
6692 {VM_MAYWRITE, "maywrite" },
6693 {VM_MAYEXEC, "mayexec" },
6694 {VM_MAYSHARE, "mayshare" },
6695 {VM_GROWSDOWN, "growsdown" },
6696 {VM_PFNMAP, "pfnmap" },
6697 {VM_DENYWRITE, "denywrite" },
6698 {VM_LOCKED, "locked" },
6700 {VM_SEQ_READ, "seqread" },
6701 {VM_RAND_READ, "randread" },
6702 {VM_DONTCOPY, "dontcopy" },
6703 {VM_DONTEXPAND, "dontexpand" },
6704 {VM_ACCOUNT, "account" },
6705 {VM_NORESERVE, "noreserve" },
6706 {VM_HUGETLB, "hugetlb" },
6707 {VM_NONLINEAR, "nonlinear" },
6708 #if defined(CONFIG_X86)
6710 #elif defined(CONFIG_PPC)
6712 #elif defined(CONFIG_PARISC) || defined(CONFIG_METAG) || defined(CONFIG_IA64)
6713 {VM_GROWSUP, "growsup" },
6714 #elif !defined(CONFIG_MMU)
6715 {VM_MAPPED_COPY, "mappedcopy" },
6717 {VM_ARCH_1, "arch_1" },
6719 {VM_DONTDUMP, "dontdump" },
6720 #ifdef CONFIG_MEM_SOFT_DIRTY
6721 {VM_SOFTDIRTY, "softdirty" },
6723 {VM_MIXEDMAP, "mixedmap" },
6724 {VM_HUGEPAGE, "hugepage" },
6725 {VM_NOHUGEPAGE, "nohugepage" },
6726 {VM_MERGEABLE, "mergeable" },
6729 void dump_vma(const struct vm_area_struct *vma)
6732 "vma %p start %p end %p\n"
6733 "next %p prev %p mm %p\n"
6734 "prot %lx anon_vma %p vm_ops %p\n"
6735 "pgoff %lx file %p private_data %p\n",
6736 vma, (void *)vma->vm_start, (void *)vma->vm_end, vma->vm_next,
6737 vma->vm_prev, vma->vm_mm,
6738 (unsigned long)pgprot_val(vma->vm_page_prot),
6739 vma->anon_vma, vma->vm_ops, vma->vm_pgoff,
6740 vma->vm_file, vma->vm_private_data);
6741 dump_flags(vma->vm_flags, vmaflags_names, ARRAY_SIZE(vmaflags_names));
6743 EXPORT_SYMBOL(dump_vma);
6745 #endif /* CONFIG_DEBUG_VM */