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_ext.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/prefetch.h>
57 #include <linux/mm_inline.h>
58 #include <linux/migrate.h>
59 #include <linux/page_ext.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
62 #include <linux/page_owner.h>
64 #include <asm/sections.h>
65 #include <asm/tlbflush.h>
66 #include <asm/div64.h>
69 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
70 static DEFINE_MUTEX(pcp_batch_high_lock);
71 #define MIN_PERCPU_PAGELIST_FRACTION (8)
73 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74 DEFINE_PER_CPU(int, numa_node);
75 EXPORT_PER_CPU_SYMBOL(numa_node);
78 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
85 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
86 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
87 int _node_numa_mem_[MAX_NUMNODES];
91 * Array of node states.
93 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
94 [N_POSSIBLE] = NODE_MASK_ALL,
95 [N_ONLINE] = { { [0] = 1UL } },
97 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
99 [N_HIGH_MEMORY] = { { [0] = 1UL } },
101 #ifdef CONFIG_MOVABLE_NODE
102 [N_MEMORY] = { { [0] = 1UL } },
104 [N_CPU] = { { [0] = 1UL } },
107 EXPORT_SYMBOL(node_states);
109 /* Protect totalram_pages and zone->managed_pages */
110 static DEFINE_SPINLOCK(managed_page_count_lock);
112 unsigned long totalram_pages __read_mostly;
113 unsigned long totalreserve_pages __read_mostly;
115 * When calculating the number of globally allowed dirty pages, there
116 * is a certain number of per-zone reserves that should not be
117 * considered dirtyable memory. This is the sum of those reserves
118 * over all existing zones that contribute dirtyable memory.
120 unsigned long dirty_balance_reserve __read_mostly;
122 int percpu_pagelist_fraction;
123 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
125 #ifdef CONFIG_PM_SLEEP
127 * The following functions are used by the suspend/hibernate code to temporarily
128 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
129 * while devices are suspended. To avoid races with the suspend/hibernate code,
130 * they should always be called with pm_mutex held (gfp_allowed_mask also should
131 * only be modified with pm_mutex held, unless the suspend/hibernate code is
132 * guaranteed not to run in parallel with that modification).
135 static gfp_t saved_gfp_mask;
137 void pm_restore_gfp_mask(void)
139 WARN_ON(!mutex_is_locked(&pm_mutex));
140 if (saved_gfp_mask) {
141 gfp_allowed_mask = saved_gfp_mask;
146 void pm_restrict_gfp_mask(void)
148 WARN_ON(!mutex_is_locked(&pm_mutex));
149 WARN_ON(saved_gfp_mask);
150 saved_gfp_mask = gfp_allowed_mask;
151 gfp_allowed_mask &= ~GFP_IOFS;
154 bool pm_suspended_storage(void)
156 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
160 #endif /* CONFIG_PM_SLEEP */
162 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
163 int pageblock_order __read_mostly;
166 static void __free_pages_ok(struct page *page, unsigned int order);
169 * results with 256, 32 in the lowmem_reserve sysctl:
170 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
171 * 1G machine -> (16M dma, 784M normal, 224M high)
172 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
173 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
174 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
176 * TBD: should special case ZONE_DMA32 machines here - in those we normally
177 * don't need any ZONE_NORMAL reservation
179 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
180 #ifdef CONFIG_ZONE_DMA
183 #ifdef CONFIG_ZONE_DMA32
186 #ifdef CONFIG_HIGHMEM
192 EXPORT_SYMBOL(totalram_pages);
194 static char * const zone_names[MAX_NR_ZONES] = {
195 #ifdef CONFIG_ZONE_DMA
198 #ifdef CONFIG_ZONE_DMA32
202 #ifdef CONFIG_HIGHMEM
208 int min_free_kbytes = 1024;
209 int user_min_free_kbytes = -1;
211 static unsigned long __meminitdata nr_kernel_pages;
212 static unsigned long __meminitdata nr_all_pages;
213 static unsigned long __meminitdata dma_reserve;
215 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
216 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
217 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
218 static unsigned long __initdata required_kernelcore;
219 static unsigned long __initdata required_movablecore;
220 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
222 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
224 EXPORT_SYMBOL(movable_zone);
225 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
228 int nr_node_ids __read_mostly = MAX_NUMNODES;
229 int nr_online_nodes __read_mostly = 1;
230 EXPORT_SYMBOL(nr_node_ids);
231 EXPORT_SYMBOL(nr_online_nodes);
234 int page_group_by_mobility_disabled __read_mostly;
236 void set_pageblock_migratetype(struct page *page, int migratetype)
238 if (unlikely(page_group_by_mobility_disabled &&
239 migratetype < MIGRATE_PCPTYPES))
240 migratetype = MIGRATE_UNMOVABLE;
242 set_pageblock_flags_group(page, (unsigned long)migratetype,
243 PB_migrate, PB_migrate_end);
246 bool oom_killer_disabled __read_mostly;
248 #ifdef CONFIG_DEBUG_VM
249 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
253 unsigned long pfn = page_to_pfn(page);
254 unsigned long sp, start_pfn;
257 seq = zone_span_seqbegin(zone);
258 start_pfn = zone->zone_start_pfn;
259 sp = zone->spanned_pages;
260 if (!zone_spans_pfn(zone, pfn))
262 } while (zone_span_seqretry(zone, seq));
265 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
266 pfn, zone_to_nid(zone), zone->name,
267 start_pfn, start_pfn + sp);
272 static int page_is_consistent(struct zone *zone, struct page *page)
274 if (!pfn_valid_within(page_to_pfn(page)))
276 if (zone != page_zone(page))
282 * Temporary debugging check for pages not lying within a given zone.
284 static int bad_range(struct zone *zone, struct page *page)
286 if (page_outside_zone_boundaries(zone, page))
288 if (!page_is_consistent(zone, page))
294 static inline int bad_range(struct zone *zone, struct page *page)
300 static void bad_page(struct page *page, const char *reason,
301 unsigned long bad_flags)
303 static unsigned long resume;
304 static unsigned long nr_shown;
305 static unsigned long nr_unshown;
307 /* Don't complain about poisoned pages */
308 if (PageHWPoison(page)) {
309 page_mapcount_reset(page); /* remove PageBuddy */
314 * Allow a burst of 60 reports, then keep quiet for that minute;
315 * or allow a steady drip of one report per second.
317 if (nr_shown == 60) {
318 if (time_before(jiffies, resume)) {
324 "BUG: Bad page state: %lu messages suppressed\n",
331 resume = jiffies + 60 * HZ;
333 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
334 current->comm, page_to_pfn(page));
335 dump_page_badflags(page, reason, bad_flags);
340 /* Leave bad fields for debug, except PageBuddy could make trouble */
341 page_mapcount_reset(page); /* remove PageBuddy */
342 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
346 * Higher-order pages are called "compound pages". They are structured thusly:
348 * The first PAGE_SIZE page is called the "head page".
350 * The remaining PAGE_SIZE pages are called "tail pages".
352 * All pages have PG_compound set. All tail pages have their ->first_page
353 * pointing at the head page.
355 * The first tail page's ->lru.next holds the address of the compound page's
356 * put_page() function. Its ->lru.prev holds the order of allocation.
357 * This usage means that zero-order pages may not be compound.
360 static void free_compound_page(struct page *page)
362 __free_pages_ok(page, compound_order(page));
365 void prep_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
370 set_compound_page_dtor(page, free_compound_page);
371 set_compound_order(page, order);
373 for (i = 1; i < nr_pages; i++) {
374 struct page *p = page + i;
375 set_page_count(p, 0);
376 p->first_page = page;
377 /* Make sure p->first_page is always valid for PageTail() */
383 /* update __split_huge_page_refcount if you change this function */
384 static int destroy_compound_page(struct page *page, unsigned long order)
387 int nr_pages = 1 << order;
390 if (unlikely(compound_order(page) != order)) {
391 bad_page(page, "wrong compound order", 0);
395 __ClearPageHead(page);
397 for (i = 1; i < nr_pages; i++) {
398 struct page *p = page + i;
400 if (unlikely(!PageTail(p))) {
401 bad_page(page, "PageTail not set", 0);
403 } else if (unlikely(p->first_page != page)) {
404 bad_page(page, "first_page not consistent", 0);
413 static inline void prep_zero_page(struct page *page, unsigned int order,
419 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
420 * and __GFP_HIGHMEM from hard or soft interrupt context.
422 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
423 for (i = 0; i < (1 << order); i++)
424 clear_highpage(page + i);
427 #ifdef CONFIG_DEBUG_PAGEALLOC
428 unsigned int _debug_guardpage_minorder;
429 bool _debug_pagealloc_enabled __read_mostly;
430 bool _debug_guardpage_enabled __read_mostly;
432 static int __init early_debug_pagealloc(char *buf)
437 if (strcmp(buf, "on") == 0)
438 _debug_pagealloc_enabled = true;
442 early_param("debug_pagealloc", early_debug_pagealloc);
444 static bool need_debug_guardpage(void)
446 /* If we don't use debug_pagealloc, we don't need guard page */
447 if (!debug_pagealloc_enabled())
453 static void init_debug_guardpage(void)
455 if (!debug_pagealloc_enabled())
458 _debug_guardpage_enabled = true;
461 struct page_ext_operations debug_guardpage_ops = {
462 .need = need_debug_guardpage,
463 .init = init_debug_guardpage,
466 static int __init debug_guardpage_minorder_setup(char *buf)
470 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
471 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
474 _debug_guardpage_minorder = res;
475 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
478 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
480 static inline void set_page_guard(struct zone *zone, struct page *page,
481 unsigned int order, int migratetype)
483 struct page_ext *page_ext;
485 if (!debug_guardpage_enabled())
488 page_ext = lookup_page_ext(page);
489 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
491 INIT_LIST_HEAD(&page->lru);
492 set_page_private(page, order);
493 /* Guard pages are not available for any usage */
494 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
497 static inline void clear_page_guard(struct zone *zone, struct page *page,
498 unsigned int order, int migratetype)
500 struct page_ext *page_ext;
502 if (!debug_guardpage_enabled())
505 page_ext = lookup_page_ext(page);
506 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
508 set_page_private(page, 0);
509 if (!is_migrate_isolate(migratetype))
510 __mod_zone_freepage_state(zone, (1 << order), migratetype);
513 struct page_ext_operations debug_guardpage_ops = { NULL, };
514 static inline void set_page_guard(struct zone *zone, struct page *page,
515 unsigned int order, int migratetype) {}
516 static inline void clear_page_guard(struct zone *zone, struct page *page,
517 unsigned int order, int migratetype) {}
520 static inline void set_page_order(struct page *page, unsigned int order)
522 set_page_private(page, order);
523 __SetPageBuddy(page);
526 static inline void rmv_page_order(struct page *page)
528 __ClearPageBuddy(page);
529 set_page_private(page, 0);
533 * This function checks whether a page is free && is the buddy
534 * we can do coalesce a page and its buddy if
535 * (a) the buddy is not in a hole &&
536 * (b) the buddy is in the buddy system &&
537 * (c) a page and its buddy have the same order &&
538 * (d) a page and its buddy are in the same zone.
540 * For recording whether a page is in the buddy system, we set ->_mapcount
541 * PAGE_BUDDY_MAPCOUNT_VALUE.
542 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
543 * serialized by zone->lock.
545 * For recording page's order, we use page_private(page).
547 static inline int page_is_buddy(struct page *page, struct page *buddy,
550 if (!pfn_valid_within(page_to_pfn(buddy)))
553 if (page_is_guard(buddy) && page_order(buddy) == order) {
554 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
556 if (page_zone_id(page) != page_zone_id(buddy))
562 if (PageBuddy(buddy) && page_order(buddy) == order) {
563 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
566 * zone check is done late to avoid uselessly
567 * calculating zone/node ids for pages that could
570 if (page_zone_id(page) != page_zone_id(buddy))
579 * Freeing function for a buddy system allocator.
581 * The concept of a buddy system is to maintain direct-mapped table
582 * (containing bit values) for memory blocks of various "orders".
583 * The bottom level table contains the map for the smallest allocatable
584 * units of memory (here, pages), and each level above it describes
585 * pairs of units from the levels below, hence, "buddies".
586 * At a high level, all that happens here is marking the table entry
587 * at the bottom level available, and propagating the changes upward
588 * as necessary, plus some accounting needed to play nicely with other
589 * parts of the VM system.
590 * At each level, we keep a list of pages, which are heads of continuous
591 * free pages of length of (1 << order) and marked with _mapcount
592 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
594 * So when we are allocating or freeing one, we can derive the state of the
595 * other. That is, if we allocate a small block, and both were
596 * free, the remainder of the region must be split into blocks.
597 * If a block is freed, and its buddy is also free, then this
598 * triggers coalescing into a block of larger size.
603 static inline void __free_one_page(struct page *page,
605 struct zone *zone, unsigned int order,
608 unsigned long page_idx;
609 unsigned long combined_idx;
610 unsigned long uninitialized_var(buddy_idx);
612 int max_order = MAX_ORDER;
614 VM_BUG_ON(!zone_is_initialized(zone));
616 if (unlikely(PageCompound(page)))
617 if (unlikely(destroy_compound_page(page, order)))
620 VM_BUG_ON(migratetype == -1);
621 if (is_migrate_isolate(migratetype)) {
623 * We restrict max order of merging to prevent merge
624 * between freepages on isolate pageblock and normal
625 * pageblock. Without this, pageblock isolation
626 * could cause incorrect freepage accounting.
628 max_order = min(MAX_ORDER, pageblock_order + 1);
630 __mod_zone_freepage_state(zone, 1 << order, migratetype);
633 page_idx = pfn & ((1 << max_order) - 1);
635 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
636 VM_BUG_ON_PAGE(bad_range(zone, page), page);
638 while (order < max_order - 1) {
639 buddy_idx = __find_buddy_index(page_idx, order);
640 buddy = page + (buddy_idx - page_idx);
641 if (!page_is_buddy(page, buddy, order))
644 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
645 * merge with it and move up one order.
647 if (page_is_guard(buddy)) {
648 clear_page_guard(zone, buddy, order, migratetype);
650 list_del(&buddy->lru);
651 zone->free_area[order].nr_free--;
652 rmv_page_order(buddy);
654 combined_idx = buddy_idx & page_idx;
655 page = page + (combined_idx - page_idx);
656 page_idx = combined_idx;
659 set_page_order(page, order);
662 * If this is not the largest possible page, check if the buddy
663 * of the next-highest order is free. If it is, it's possible
664 * that pages are being freed that will coalesce soon. In case,
665 * that is happening, add the free page to the tail of the list
666 * so it's less likely to be used soon and more likely to be merged
667 * as a higher order page
669 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
670 struct page *higher_page, *higher_buddy;
671 combined_idx = buddy_idx & page_idx;
672 higher_page = page + (combined_idx - page_idx);
673 buddy_idx = __find_buddy_index(combined_idx, order + 1);
674 higher_buddy = higher_page + (buddy_idx - combined_idx);
675 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
676 list_add_tail(&page->lru,
677 &zone->free_area[order].free_list[migratetype]);
682 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
684 zone->free_area[order].nr_free++;
687 static inline int free_pages_check(struct page *page)
689 const char *bad_reason = NULL;
690 unsigned long bad_flags = 0;
692 if (unlikely(page_mapcount(page)))
693 bad_reason = "nonzero mapcount";
694 if (unlikely(page->mapping != NULL))
695 bad_reason = "non-NULL mapping";
696 if (unlikely(atomic_read(&page->_count) != 0))
697 bad_reason = "nonzero _count";
698 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
699 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
700 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
703 if (unlikely(page->mem_cgroup))
704 bad_reason = "page still charged to cgroup";
706 if (unlikely(bad_reason)) {
707 bad_page(page, bad_reason, bad_flags);
710 page_cpupid_reset_last(page);
711 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
712 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
717 * Frees a number of pages from the PCP lists
718 * Assumes all pages on list are in same zone, and of same order.
719 * count is the number of pages to free.
721 * If the zone was previously in an "all pages pinned" state then look to
722 * see if this freeing clears that state.
724 * And clear the zone's pages_scanned counter, to hold off the "all pages are
725 * pinned" detection logic.
727 static void free_pcppages_bulk(struct zone *zone, int count,
728 struct per_cpu_pages *pcp)
733 unsigned long nr_scanned;
735 spin_lock(&zone->lock);
736 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
738 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
742 struct list_head *list;
745 * Remove pages from lists in a round-robin fashion. A
746 * batch_free count is maintained that is incremented when an
747 * empty list is encountered. This is so more pages are freed
748 * off fuller lists instead of spinning excessively around empty
753 if (++migratetype == MIGRATE_PCPTYPES)
755 list = &pcp->lists[migratetype];
756 } while (list_empty(list));
758 /* This is the only non-empty list. Free them all. */
759 if (batch_free == MIGRATE_PCPTYPES)
760 batch_free = to_free;
763 int mt; /* migratetype of the to-be-freed page */
765 page = list_entry(list->prev, struct page, lru);
766 /* must delete as __free_one_page list manipulates */
767 list_del(&page->lru);
768 mt = get_freepage_migratetype(page);
769 if (unlikely(has_isolate_pageblock(zone)))
770 mt = get_pageblock_migratetype(page);
772 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
773 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
774 trace_mm_page_pcpu_drain(page, 0, mt);
775 } while (--to_free && --batch_free && !list_empty(list));
777 spin_unlock(&zone->lock);
780 static void free_one_page(struct zone *zone,
781 struct page *page, unsigned long pfn,
785 unsigned long nr_scanned;
786 spin_lock(&zone->lock);
787 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
789 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
791 if (unlikely(has_isolate_pageblock(zone) ||
792 is_migrate_isolate(migratetype))) {
793 migratetype = get_pfnblock_migratetype(page, pfn);
795 __free_one_page(page, pfn, zone, order, migratetype);
796 spin_unlock(&zone->lock);
799 static bool free_pages_prepare(struct page *page, unsigned int order)
804 VM_BUG_ON_PAGE(PageTail(page), page);
805 VM_BUG_ON_PAGE(PageHead(page) && compound_order(page) != order, page);
807 trace_mm_page_free(page, order);
808 kmemcheck_free_shadow(page, order);
811 page->mapping = NULL;
812 for (i = 0; i < (1 << order); i++)
813 bad += free_pages_check(page + i);
817 reset_page_owner(page, order);
819 if (!PageHighMem(page)) {
820 debug_check_no_locks_freed(page_address(page),
822 debug_check_no_obj_freed(page_address(page),
825 arch_free_page(page, order);
826 kernel_map_pages(page, 1 << order, 0);
831 static void __free_pages_ok(struct page *page, unsigned int order)
835 unsigned long pfn = page_to_pfn(page);
837 if (!free_pages_prepare(page, order))
840 migratetype = get_pfnblock_migratetype(page, pfn);
841 local_irq_save(flags);
842 __count_vm_events(PGFREE, 1 << order);
843 set_freepage_migratetype(page, migratetype);
844 free_one_page(page_zone(page), page, pfn, order, migratetype);
845 local_irq_restore(flags);
848 void __init __free_pages_bootmem(struct page *page, unsigned int order)
850 unsigned int nr_pages = 1 << order;
851 struct page *p = page;
855 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
857 __ClearPageReserved(p);
858 set_page_count(p, 0);
860 __ClearPageReserved(p);
861 set_page_count(p, 0);
863 page_zone(page)->managed_pages += nr_pages;
864 set_page_refcounted(page);
865 __free_pages(page, order);
869 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
870 void __init init_cma_reserved_pageblock(struct page *page)
872 unsigned i = pageblock_nr_pages;
873 struct page *p = page;
876 __ClearPageReserved(p);
877 set_page_count(p, 0);
880 set_pageblock_migratetype(page, MIGRATE_CMA);
882 if (pageblock_order >= MAX_ORDER) {
883 i = pageblock_nr_pages;
886 set_page_refcounted(p);
887 __free_pages(p, MAX_ORDER - 1);
888 p += MAX_ORDER_NR_PAGES;
889 } while (i -= MAX_ORDER_NR_PAGES);
891 set_page_refcounted(page);
892 __free_pages(page, pageblock_order);
895 adjust_managed_page_count(page, pageblock_nr_pages);
900 * The order of subdivision here is critical for the IO subsystem.
901 * Please do not alter this order without good reasons and regression
902 * testing. Specifically, as large blocks of memory are subdivided,
903 * the order in which smaller blocks are delivered depends on the order
904 * they're subdivided in this function. This is the primary factor
905 * influencing the order in which pages are delivered to the IO
906 * subsystem according to empirical testing, and this is also justified
907 * by considering the behavior of a buddy system containing a single
908 * large block of memory acted on by a series of small allocations.
909 * This behavior is a critical factor in sglist merging's success.
913 static inline void expand(struct zone *zone, struct page *page,
914 int low, int high, struct free_area *area,
917 unsigned long size = 1 << high;
923 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
925 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
926 debug_guardpage_enabled() &&
927 high < debug_guardpage_minorder()) {
929 * Mark as guard pages (or page), that will allow to
930 * merge back to allocator when buddy will be freed.
931 * Corresponding page table entries will not be touched,
932 * pages will stay not present in virtual address space
934 set_page_guard(zone, &page[size], high, migratetype);
937 list_add(&page[size].lru, &area->free_list[migratetype]);
939 set_page_order(&page[size], high);
944 * This page is about to be returned from the page allocator
946 static inline int check_new_page(struct page *page)
948 const char *bad_reason = NULL;
949 unsigned long bad_flags = 0;
951 if (unlikely(page_mapcount(page)))
952 bad_reason = "nonzero mapcount";
953 if (unlikely(page->mapping != NULL))
954 bad_reason = "non-NULL mapping";
955 if (unlikely(atomic_read(&page->_count) != 0))
956 bad_reason = "nonzero _count";
957 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
958 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
959 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
962 if (unlikely(page->mem_cgroup))
963 bad_reason = "page still charged to cgroup";
965 if (unlikely(bad_reason)) {
966 bad_page(page, bad_reason, bad_flags);
972 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags)
976 for (i = 0; i < (1 << order); i++) {
977 struct page *p = page + i;
978 if (unlikely(check_new_page(p)))
982 set_page_private(page, 0);
983 set_page_refcounted(page);
985 arch_alloc_page(page, order);
986 kernel_map_pages(page, 1 << order, 1);
988 if (gfp_flags & __GFP_ZERO)
989 prep_zero_page(page, order, gfp_flags);
991 if (order && (gfp_flags & __GFP_COMP))
992 prep_compound_page(page, order);
994 set_page_owner(page, order, gfp_flags);
1000 * Go through the free lists for the given migratetype and remove
1001 * the smallest available page from the freelists
1004 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1007 unsigned int current_order;
1008 struct free_area *area;
1011 /* Find a page of the appropriate size in the preferred list */
1012 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1013 area = &(zone->free_area[current_order]);
1014 if (list_empty(&area->free_list[migratetype]))
1017 page = list_entry(area->free_list[migratetype].next,
1019 list_del(&page->lru);
1020 rmv_page_order(page);
1022 expand(zone, page, order, current_order, area, migratetype);
1023 set_freepage_migratetype(page, migratetype);
1032 * This array describes the order lists are fallen back to when
1033 * the free lists for the desirable migrate type are depleted
1035 static int fallbacks[MIGRATE_TYPES][4] = {
1036 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1037 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1039 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1040 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1042 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1044 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1045 #ifdef CONFIG_MEMORY_ISOLATION
1046 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1051 * Move the free pages in a range to the free lists of the requested type.
1052 * Note that start_page and end_pages are not aligned on a pageblock
1053 * boundary. If alignment is required, use move_freepages_block()
1055 int move_freepages(struct zone *zone,
1056 struct page *start_page, struct page *end_page,
1060 unsigned long order;
1061 int pages_moved = 0;
1063 #ifndef CONFIG_HOLES_IN_ZONE
1065 * page_zone is not safe to call in this context when
1066 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1067 * anyway as we check zone boundaries in move_freepages_block().
1068 * Remove at a later date when no bug reports exist related to
1069 * grouping pages by mobility
1071 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1074 for (page = start_page; page <= end_page;) {
1075 /* Make sure we are not inadvertently changing nodes */
1076 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1078 if (!pfn_valid_within(page_to_pfn(page))) {
1083 if (!PageBuddy(page)) {
1088 order = page_order(page);
1089 list_move(&page->lru,
1090 &zone->free_area[order].free_list[migratetype]);
1091 set_freepage_migratetype(page, migratetype);
1093 pages_moved += 1 << order;
1099 int move_freepages_block(struct zone *zone, struct page *page,
1102 unsigned long start_pfn, end_pfn;
1103 struct page *start_page, *end_page;
1105 start_pfn = page_to_pfn(page);
1106 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1107 start_page = pfn_to_page(start_pfn);
1108 end_page = start_page + pageblock_nr_pages - 1;
1109 end_pfn = start_pfn + pageblock_nr_pages - 1;
1111 /* Do not cross zone boundaries */
1112 if (!zone_spans_pfn(zone, start_pfn))
1114 if (!zone_spans_pfn(zone, end_pfn))
1117 return move_freepages(zone, start_page, end_page, migratetype);
1120 static void change_pageblock_range(struct page *pageblock_page,
1121 int start_order, int migratetype)
1123 int nr_pageblocks = 1 << (start_order - pageblock_order);
1125 while (nr_pageblocks--) {
1126 set_pageblock_migratetype(pageblock_page, migratetype);
1127 pageblock_page += pageblock_nr_pages;
1132 * If breaking a large block of pages, move all free pages to the preferred
1133 * allocation list. If falling back for a reclaimable kernel allocation, be
1134 * more aggressive about taking ownership of free pages.
1136 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1137 * nor move CMA pages to different free lists. We don't want unmovable pages
1138 * to be allocated from MIGRATE_CMA areas.
1140 * Returns the new migratetype of the pageblock (or the same old migratetype
1141 * if it was unchanged).
1143 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1144 int start_type, int fallback_type)
1146 int current_order = page_order(page);
1149 * When borrowing from MIGRATE_CMA, we need to release the excess
1150 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1151 * is set to CMA so it is returned to the correct freelist in case
1152 * the page ends up being not actually allocated from the pcp lists.
1154 if (is_migrate_cma(fallback_type))
1155 return fallback_type;
1157 /* Take ownership for orders >= pageblock_order */
1158 if (current_order >= pageblock_order) {
1159 change_pageblock_range(page, current_order, start_type);
1163 if (current_order >= pageblock_order / 2 ||
1164 start_type == MIGRATE_RECLAIMABLE ||
1165 page_group_by_mobility_disabled) {
1168 pages = move_freepages_block(zone, page, start_type);
1170 /* Claim the whole block if over half of it is free */
1171 if (pages >= (1 << (pageblock_order-1)) ||
1172 page_group_by_mobility_disabled) {
1174 set_pageblock_migratetype(page, start_type);
1180 return fallback_type;
1183 /* Remove an element from the buddy allocator from the fallback list */
1184 static inline struct page *
1185 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1187 struct free_area *area;
1188 unsigned int current_order;
1190 int migratetype, new_type, i;
1192 /* Find the largest possible block of pages in the other list */
1193 for (current_order = MAX_ORDER-1;
1194 current_order >= order && current_order <= MAX_ORDER-1;
1197 migratetype = fallbacks[start_migratetype][i];
1199 /* MIGRATE_RESERVE handled later if necessary */
1200 if (migratetype == MIGRATE_RESERVE)
1203 area = &(zone->free_area[current_order]);
1204 if (list_empty(&area->free_list[migratetype]))
1207 page = list_entry(area->free_list[migratetype].next,
1211 new_type = try_to_steal_freepages(zone, page,
1215 /* Remove the page from the freelists */
1216 list_del(&page->lru);
1217 rmv_page_order(page);
1219 expand(zone, page, order, current_order, area,
1221 /* The freepage_migratetype may differ from pageblock's
1222 * migratetype depending on the decisions in
1223 * try_to_steal_freepages. This is OK as long as it does
1224 * not differ for MIGRATE_CMA type.
1226 set_freepage_migratetype(page, new_type);
1228 trace_mm_page_alloc_extfrag(page, order, current_order,
1229 start_migratetype, migratetype, new_type);
1239 * Do the hard work of removing an element from the buddy allocator.
1240 * Call me with the zone->lock already held.
1242 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1248 page = __rmqueue_smallest(zone, order, migratetype);
1250 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1251 page = __rmqueue_fallback(zone, order, migratetype);
1254 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1255 * is used because __rmqueue_smallest is an inline function
1256 * and we want just one call site
1259 migratetype = MIGRATE_RESERVE;
1264 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1269 * Obtain a specified number of elements from the buddy allocator, all under
1270 * a single hold of the lock, for efficiency. Add them to the supplied list.
1271 * Returns the number of new pages which were placed at *list.
1273 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1274 unsigned long count, struct list_head *list,
1275 int migratetype, bool cold)
1279 spin_lock(&zone->lock);
1280 for (i = 0; i < count; ++i) {
1281 struct page *page = __rmqueue(zone, order, migratetype);
1282 if (unlikely(page == NULL))
1286 * Split buddy pages returned by expand() are received here
1287 * in physical page order. The page is added to the callers and
1288 * list and the list head then moves forward. From the callers
1289 * perspective, the linked list is ordered by page number in
1290 * some conditions. This is useful for IO devices that can
1291 * merge IO requests if the physical pages are ordered
1295 list_add(&page->lru, list);
1297 list_add_tail(&page->lru, list);
1299 if (is_migrate_cma(get_freepage_migratetype(page)))
1300 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1303 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1304 spin_unlock(&zone->lock);
1310 * Called from the vmstat counter updater to drain pagesets of this
1311 * currently executing processor on remote nodes after they have
1314 * Note that this function must be called with the thread pinned to
1315 * a single processor.
1317 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1319 unsigned long flags;
1320 int to_drain, batch;
1322 local_irq_save(flags);
1323 batch = ACCESS_ONCE(pcp->batch);
1324 to_drain = min(pcp->count, batch);
1326 free_pcppages_bulk(zone, to_drain, pcp);
1327 pcp->count -= to_drain;
1329 local_irq_restore(flags);
1334 * Drain pcplists of the indicated processor and zone.
1336 * The processor must either be the current processor and the
1337 * thread pinned to the current processor or a processor that
1340 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1342 unsigned long flags;
1343 struct per_cpu_pageset *pset;
1344 struct per_cpu_pages *pcp;
1346 local_irq_save(flags);
1347 pset = per_cpu_ptr(zone->pageset, cpu);
1351 free_pcppages_bulk(zone, pcp->count, pcp);
1354 local_irq_restore(flags);
1358 * Drain pcplists of all zones on the indicated processor.
1360 * The processor must either be the current processor and the
1361 * thread pinned to the current processor or a processor that
1364 static void drain_pages(unsigned int cpu)
1368 for_each_populated_zone(zone) {
1369 drain_pages_zone(cpu, zone);
1374 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1376 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1377 * the single zone's pages.
1379 void drain_local_pages(struct zone *zone)
1381 int cpu = smp_processor_id();
1384 drain_pages_zone(cpu, zone);
1390 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1392 * When zone parameter is non-NULL, spill just the single zone's pages.
1394 * Note that this code is protected against sending an IPI to an offline
1395 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1396 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1397 * nothing keeps CPUs from showing up after we populated the cpumask and
1398 * before the call to on_each_cpu_mask().
1400 void drain_all_pages(struct zone *zone)
1405 * Allocate in the BSS so we wont require allocation in
1406 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1408 static cpumask_t cpus_with_pcps;
1411 * We don't care about racing with CPU hotplug event
1412 * as offline notification will cause the notified
1413 * cpu to drain that CPU pcps and on_each_cpu_mask
1414 * disables preemption as part of its processing
1416 for_each_online_cpu(cpu) {
1417 struct per_cpu_pageset *pcp;
1419 bool has_pcps = false;
1422 pcp = per_cpu_ptr(zone->pageset, cpu);
1426 for_each_populated_zone(z) {
1427 pcp = per_cpu_ptr(z->pageset, cpu);
1428 if (pcp->pcp.count) {
1436 cpumask_set_cpu(cpu, &cpus_with_pcps);
1438 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1440 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1444 #ifdef CONFIG_HIBERNATION
1446 void mark_free_pages(struct zone *zone)
1448 unsigned long pfn, max_zone_pfn;
1449 unsigned long flags;
1450 unsigned int order, t;
1451 struct list_head *curr;
1453 if (zone_is_empty(zone))
1456 spin_lock_irqsave(&zone->lock, flags);
1458 max_zone_pfn = zone_end_pfn(zone);
1459 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1460 if (pfn_valid(pfn)) {
1461 struct page *page = pfn_to_page(pfn);
1463 if (!swsusp_page_is_forbidden(page))
1464 swsusp_unset_page_free(page);
1467 for_each_migratetype_order(order, t) {
1468 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1471 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1472 for (i = 0; i < (1UL << order); i++)
1473 swsusp_set_page_free(pfn_to_page(pfn + i));
1476 spin_unlock_irqrestore(&zone->lock, flags);
1478 #endif /* CONFIG_PM */
1481 * Free a 0-order page
1482 * cold == true ? free a cold page : free a hot page
1484 void free_hot_cold_page(struct page *page, bool cold)
1486 struct zone *zone = page_zone(page);
1487 struct per_cpu_pages *pcp;
1488 unsigned long flags;
1489 unsigned long pfn = page_to_pfn(page);
1492 if (!free_pages_prepare(page, 0))
1495 migratetype = get_pfnblock_migratetype(page, pfn);
1496 set_freepage_migratetype(page, migratetype);
1497 local_irq_save(flags);
1498 __count_vm_event(PGFREE);
1501 * We only track unmovable, reclaimable and movable on pcp lists.
1502 * Free ISOLATE pages back to the allocator because they are being
1503 * offlined but treat RESERVE as movable pages so we can get those
1504 * areas back if necessary. Otherwise, we may have to free
1505 * excessively into the page allocator
1507 if (migratetype >= MIGRATE_PCPTYPES) {
1508 if (unlikely(is_migrate_isolate(migratetype))) {
1509 free_one_page(zone, page, pfn, 0, migratetype);
1512 migratetype = MIGRATE_MOVABLE;
1515 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1517 list_add(&page->lru, &pcp->lists[migratetype]);
1519 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1521 if (pcp->count >= pcp->high) {
1522 unsigned long batch = ACCESS_ONCE(pcp->batch);
1523 free_pcppages_bulk(zone, batch, pcp);
1524 pcp->count -= batch;
1528 local_irq_restore(flags);
1532 * Free a list of 0-order pages
1534 void free_hot_cold_page_list(struct list_head *list, bool cold)
1536 struct page *page, *next;
1538 list_for_each_entry_safe(page, next, list, lru) {
1539 trace_mm_page_free_batched(page, cold);
1540 free_hot_cold_page(page, cold);
1545 * split_page takes a non-compound higher-order page, and splits it into
1546 * n (1<<order) sub-pages: page[0..n]
1547 * Each sub-page must be freed individually.
1549 * Note: this is probably too low level an operation for use in drivers.
1550 * Please consult with lkml before using this in your driver.
1552 void split_page(struct page *page, unsigned int order)
1556 VM_BUG_ON_PAGE(PageCompound(page), page);
1557 VM_BUG_ON_PAGE(!page_count(page), page);
1559 #ifdef CONFIG_KMEMCHECK
1561 * Split shadow pages too, because free(page[0]) would
1562 * otherwise free the whole shadow.
1564 if (kmemcheck_page_is_tracked(page))
1565 split_page(virt_to_page(page[0].shadow), order);
1568 set_page_owner(page, 0, 0);
1569 for (i = 1; i < (1 << order); i++) {
1570 set_page_refcounted(page + i);
1571 set_page_owner(page + i, 0, 0);
1574 EXPORT_SYMBOL_GPL(split_page);
1576 int __isolate_free_page(struct page *page, unsigned int order)
1578 unsigned long watermark;
1582 BUG_ON(!PageBuddy(page));
1584 zone = page_zone(page);
1585 mt = get_pageblock_migratetype(page);
1587 if (!is_migrate_isolate(mt)) {
1588 /* Obey watermarks as if the page was being allocated */
1589 watermark = low_wmark_pages(zone) + (1 << order);
1590 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1593 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1596 /* Remove page from free list */
1597 list_del(&page->lru);
1598 zone->free_area[order].nr_free--;
1599 rmv_page_order(page);
1601 /* Set the pageblock if the isolated page is at least a pageblock */
1602 if (order >= pageblock_order - 1) {
1603 struct page *endpage = page + (1 << order) - 1;
1604 for (; page < endpage; page += pageblock_nr_pages) {
1605 int mt = get_pageblock_migratetype(page);
1606 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1607 set_pageblock_migratetype(page,
1612 set_page_owner(page, order, 0);
1613 return 1UL << order;
1617 * Similar to split_page except the page is already free. As this is only
1618 * being used for migration, the migratetype of the block also changes.
1619 * As this is called with interrupts disabled, the caller is responsible
1620 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1623 * Note: this is probably too low level an operation for use in drivers.
1624 * Please consult with lkml before using this in your driver.
1626 int split_free_page(struct page *page)
1631 order = page_order(page);
1633 nr_pages = __isolate_free_page(page, order);
1637 /* Split into individual pages */
1638 set_page_refcounted(page);
1639 split_page(page, order);
1644 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1645 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1649 struct page *buffered_rmqueue(struct zone *preferred_zone,
1650 struct zone *zone, unsigned int order,
1651 gfp_t gfp_flags, int migratetype)
1653 unsigned long flags;
1655 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1658 if (likely(order == 0)) {
1659 struct per_cpu_pages *pcp;
1660 struct list_head *list;
1662 local_irq_save(flags);
1663 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1664 list = &pcp->lists[migratetype];
1665 if (list_empty(list)) {
1666 pcp->count += rmqueue_bulk(zone, 0,
1669 if (unlikely(list_empty(list)))
1674 page = list_entry(list->prev, struct page, lru);
1676 page = list_entry(list->next, struct page, lru);
1678 list_del(&page->lru);
1681 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1683 * __GFP_NOFAIL is not to be used in new code.
1685 * All __GFP_NOFAIL callers should be fixed so that they
1686 * properly detect and handle allocation failures.
1688 * We most definitely don't want callers attempting to
1689 * allocate greater than order-1 page units with
1692 WARN_ON_ONCE(order > 1);
1694 spin_lock_irqsave(&zone->lock, flags);
1695 page = __rmqueue(zone, order, migratetype);
1696 spin_unlock(&zone->lock);
1699 __mod_zone_freepage_state(zone, -(1 << order),
1700 get_freepage_migratetype(page));
1703 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1704 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1705 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1706 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1708 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1709 zone_statistics(preferred_zone, zone, gfp_flags);
1710 local_irq_restore(flags);
1712 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1713 if (prep_new_page(page, order, gfp_flags))
1718 local_irq_restore(flags);
1722 #ifdef CONFIG_FAIL_PAGE_ALLOC
1725 struct fault_attr attr;
1727 u32 ignore_gfp_highmem;
1728 u32 ignore_gfp_wait;
1730 } fail_page_alloc = {
1731 .attr = FAULT_ATTR_INITIALIZER,
1732 .ignore_gfp_wait = 1,
1733 .ignore_gfp_highmem = 1,
1737 static int __init setup_fail_page_alloc(char *str)
1739 return setup_fault_attr(&fail_page_alloc.attr, str);
1741 __setup("fail_page_alloc=", setup_fail_page_alloc);
1743 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1745 if (order < fail_page_alloc.min_order)
1747 if (gfp_mask & __GFP_NOFAIL)
1749 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1751 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1754 return should_fail(&fail_page_alloc.attr, 1 << order);
1757 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1759 static int __init fail_page_alloc_debugfs(void)
1761 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1764 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1765 &fail_page_alloc.attr);
1767 return PTR_ERR(dir);
1769 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1770 &fail_page_alloc.ignore_gfp_wait))
1772 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1773 &fail_page_alloc.ignore_gfp_highmem))
1775 if (!debugfs_create_u32("min-order", mode, dir,
1776 &fail_page_alloc.min_order))
1781 debugfs_remove_recursive(dir);
1786 late_initcall(fail_page_alloc_debugfs);
1788 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1790 #else /* CONFIG_FAIL_PAGE_ALLOC */
1792 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1797 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1800 * Return true if free pages are above 'mark'. This takes into account the order
1801 * of the allocation.
1803 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1804 unsigned long mark, int classzone_idx, int alloc_flags,
1807 /* free_pages may go negative - that's OK */
1812 free_pages -= (1 << order) - 1;
1813 if (alloc_flags & ALLOC_HIGH)
1815 if (alloc_flags & ALLOC_HARDER)
1818 /* If allocation can't use CMA areas don't use free CMA pages */
1819 if (!(alloc_flags & ALLOC_CMA))
1820 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1823 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1825 for (o = 0; o < order; o++) {
1826 /* At the next order, this order's pages become unavailable */
1827 free_pages -= z->free_area[o].nr_free << o;
1829 /* Require fewer higher order pages to be free */
1832 if (free_pages <= min)
1838 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1839 int classzone_idx, int alloc_flags)
1841 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1842 zone_page_state(z, NR_FREE_PAGES));
1845 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1846 unsigned long mark, int classzone_idx, int alloc_flags)
1848 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1850 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1851 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1853 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1859 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1860 * skip over zones that are not allowed by the cpuset, or that have
1861 * been recently (in last second) found to be nearly full. See further
1862 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1863 * that have to skip over a lot of full or unallowed zones.
1865 * If the zonelist cache is present in the passed zonelist, then
1866 * returns a pointer to the allowed node mask (either the current
1867 * tasks mems_allowed, or node_states[N_MEMORY].)
1869 * If the zonelist cache is not available for this zonelist, does
1870 * nothing and returns NULL.
1872 * If the fullzones BITMAP in the zonelist cache is stale (more than
1873 * a second since last zap'd) then we zap it out (clear its bits.)
1875 * We hold off even calling zlc_setup, until after we've checked the
1876 * first zone in the zonelist, on the theory that most allocations will
1877 * be satisfied from that first zone, so best to examine that zone as
1878 * quickly as we can.
1880 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1882 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1883 nodemask_t *allowednodes; /* zonelist_cache approximation */
1885 zlc = zonelist->zlcache_ptr;
1889 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1890 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1891 zlc->last_full_zap = jiffies;
1894 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1895 &cpuset_current_mems_allowed :
1896 &node_states[N_MEMORY];
1897 return allowednodes;
1901 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1902 * if it is worth looking at further for free memory:
1903 * 1) Check that the zone isn't thought to be full (doesn't have its
1904 * bit set in the zonelist_cache fullzones BITMAP).
1905 * 2) Check that the zones node (obtained from the zonelist_cache
1906 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1907 * Return true (non-zero) if zone is worth looking at further, or
1908 * else return false (zero) if it is not.
1910 * This check -ignores- the distinction between various watermarks,
1911 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1912 * found to be full for any variation of these watermarks, it will
1913 * be considered full for up to one second by all requests, unless
1914 * we are so low on memory on all allowed nodes that we are forced
1915 * into the second scan of the zonelist.
1917 * In the second scan we ignore this zonelist cache and exactly
1918 * apply the watermarks to all zones, even it is slower to do so.
1919 * We are low on memory in the second scan, and should leave no stone
1920 * unturned looking for a free page.
1922 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1923 nodemask_t *allowednodes)
1925 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1926 int i; /* index of *z in zonelist zones */
1927 int n; /* node that zone *z is on */
1929 zlc = zonelist->zlcache_ptr;
1933 i = z - zonelist->_zonerefs;
1936 /* This zone is worth trying if it is allowed but not full */
1937 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1941 * Given 'z' scanning a zonelist, set the corresponding bit in
1942 * zlc->fullzones, so that subsequent attempts to allocate a page
1943 * from that zone don't waste time re-examining it.
1945 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1947 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1948 int i; /* index of *z in zonelist zones */
1950 zlc = zonelist->zlcache_ptr;
1954 i = z - zonelist->_zonerefs;
1956 set_bit(i, zlc->fullzones);
1960 * clear all zones full, called after direct reclaim makes progress so that
1961 * a zone that was recently full is not skipped over for up to a second
1963 static void zlc_clear_zones_full(struct zonelist *zonelist)
1965 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1967 zlc = zonelist->zlcache_ptr;
1971 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1974 static bool zone_local(struct zone *local_zone, struct zone *zone)
1976 return local_zone->node == zone->node;
1979 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1981 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1985 #else /* CONFIG_NUMA */
1987 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1992 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1993 nodemask_t *allowednodes)
1998 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2002 static void zlc_clear_zones_full(struct zonelist *zonelist)
2006 static bool zone_local(struct zone *local_zone, struct zone *zone)
2011 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2016 #endif /* CONFIG_NUMA */
2018 static void reset_alloc_batches(struct zone *preferred_zone)
2020 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2023 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2024 high_wmark_pages(zone) - low_wmark_pages(zone) -
2025 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2026 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2027 } while (zone++ != preferred_zone);
2031 * get_page_from_freelist goes through the zonelist trying to allocate
2034 static struct page *
2035 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
2036 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
2037 struct zone *preferred_zone, int classzone_idx, int migratetype)
2040 struct page *page = NULL;
2042 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2043 int zlc_active = 0; /* set if using zonelist_cache */
2044 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2045 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2046 (gfp_mask & __GFP_WRITE);
2047 int nr_fair_skipped = 0;
2048 bool zonelist_rescan;
2051 zonelist_rescan = false;
2054 * Scan zonelist, looking for a zone with enough free.
2055 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2057 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2058 high_zoneidx, nodemask) {
2061 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2062 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2064 if (cpusets_enabled() &&
2065 (alloc_flags & ALLOC_CPUSET) &&
2066 !cpuset_zone_allowed(zone, gfp_mask))
2069 * Distribute pages in proportion to the individual
2070 * zone size to ensure fair page aging. The zone a
2071 * page was allocated in should have no effect on the
2072 * time the page has in memory before being reclaimed.
2074 if (alloc_flags & ALLOC_FAIR) {
2075 if (!zone_local(preferred_zone, zone))
2077 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2083 * When allocating a page cache page for writing, we
2084 * want to get it from a zone that is within its dirty
2085 * limit, such that no single zone holds more than its
2086 * proportional share of globally allowed dirty pages.
2087 * The dirty limits take into account the zone's
2088 * lowmem reserves and high watermark so that kswapd
2089 * should be able to balance it without having to
2090 * write pages from its LRU list.
2092 * This may look like it could increase pressure on
2093 * lower zones by failing allocations in higher zones
2094 * before they are full. But the pages that do spill
2095 * over are limited as the lower zones are protected
2096 * by this very same mechanism. It should not become
2097 * a practical burden to them.
2099 * XXX: For now, allow allocations to potentially
2100 * exceed the per-zone dirty limit in the slowpath
2101 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2102 * which is important when on a NUMA setup the allowed
2103 * zones are together not big enough to reach the
2104 * global limit. The proper fix for these situations
2105 * will require awareness of zones in the
2106 * dirty-throttling and the flusher threads.
2108 if (consider_zone_dirty && !zone_dirty_ok(zone))
2111 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2112 if (!zone_watermark_ok(zone, order, mark,
2113 classzone_idx, alloc_flags)) {
2116 /* Checked here to keep the fast path fast */
2117 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2118 if (alloc_flags & ALLOC_NO_WATERMARKS)
2121 if (IS_ENABLED(CONFIG_NUMA) &&
2122 !did_zlc_setup && nr_online_nodes > 1) {
2124 * we do zlc_setup if there are multiple nodes
2125 * and before considering the first zone allowed
2128 allowednodes = zlc_setup(zonelist, alloc_flags);
2133 if (zone_reclaim_mode == 0 ||
2134 !zone_allows_reclaim(preferred_zone, zone))
2135 goto this_zone_full;
2138 * As we may have just activated ZLC, check if the first
2139 * eligible zone has failed zone_reclaim recently.
2141 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2142 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2145 ret = zone_reclaim(zone, gfp_mask, order);
2147 case ZONE_RECLAIM_NOSCAN:
2150 case ZONE_RECLAIM_FULL:
2151 /* scanned but unreclaimable */
2154 /* did we reclaim enough */
2155 if (zone_watermark_ok(zone, order, mark,
2156 classzone_idx, alloc_flags))
2160 * Failed to reclaim enough to meet watermark.
2161 * Only mark the zone full if checking the min
2162 * watermark or if we failed to reclaim just
2163 * 1<<order pages or else the page allocator
2164 * fastpath will prematurely mark zones full
2165 * when the watermark is between the low and
2168 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2169 ret == ZONE_RECLAIM_SOME)
2170 goto this_zone_full;
2177 page = buffered_rmqueue(preferred_zone, zone, order,
2178 gfp_mask, migratetype);
2182 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2183 zlc_mark_zone_full(zonelist, z);
2188 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2189 * necessary to allocate the page. The expectation is
2190 * that the caller is taking steps that will free more
2191 * memory. The caller should avoid the page being used
2192 * for !PFMEMALLOC purposes.
2194 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2199 * The first pass makes sure allocations are spread fairly within the
2200 * local node. However, the local node might have free pages left
2201 * after the fairness batches are exhausted, and remote zones haven't
2202 * even been considered yet. Try once more without fairness, and
2203 * include remote zones now, before entering the slowpath and waking
2204 * kswapd: prefer spilling to a remote zone over swapping locally.
2206 if (alloc_flags & ALLOC_FAIR) {
2207 alloc_flags &= ~ALLOC_FAIR;
2208 if (nr_fair_skipped) {
2209 zonelist_rescan = true;
2210 reset_alloc_batches(preferred_zone);
2212 if (nr_online_nodes > 1)
2213 zonelist_rescan = true;
2216 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2217 /* Disable zlc cache for second zonelist scan */
2219 zonelist_rescan = true;
2222 if (zonelist_rescan)
2229 * Large machines with many possible nodes should not always dump per-node
2230 * meminfo in irq context.
2232 static inline bool should_suppress_show_mem(void)
2237 ret = in_interrupt();
2242 static DEFINE_RATELIMIT_STATE(nopage_rs,
2243 DEFAULT_RATELIMIT_INTERVAL,
2244 DEFAULT_RATELIMIT_BURST);
2246 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2248 unsigned int filter = SHOW_MEM_FILTER_NODES;
2250 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2251 debug_guardpage_minorder() > 0)
2255 * This documents exceptions given to allocations in certain
2256 * contexts that are allowed to allocate outside current's set
2259 if (!(gfp_mask & __GFP_NOMEMALLOC))
2260 if (test_thread_flag(TIF_MEMDIE) ||
2261 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2262 filter &= ~SHOW_MEM_FILTER_NODES;
2263 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2264 filter &= ~SHOW_MEM_FILTER_NODES;
2267 struct va_format vaf;
2270 va_start(args, fmt);
2275 pr_warn("%pV", &vaf);
2280 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2281 current->comm, order, gfp_mask);
2284 if (!should_suppress_show_mem())
2289 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2290 unsigned long did_some_progress,
2291 unsigned long pages_reclaimed)
2293 /* Do not loop if specifically requested */
2294 if (gfp_mask & __GFP_NORETRY)
2297 /* Always retry if specifically requested */
2298 if (gfp_mask & __GFP_NOFAIL)
2302 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2303 * making forward progress without invoking OOM. Suspend also disables
2304 * storage devices so kswapd will not help. Bail if we are suspending.
2306 if (!did_some_progress && pm_suspended_storage())
2310 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2311 * means __GFP_NOFAIL, but that may not be true in other
2314 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2318 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2319 * specified, then we retry until we no longer reclaim any pages
2320 * (above), or we've reclaimed an order of pages at least as
2321 * large as the allocation's order. In both cases, if the
2322 * allocation still fails, we stop retrying.
2324 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2330 static inline struct page *
2331 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2332 struct zonelist *zonelist, enum zone_type high_zoneidx,
2333 nodemask_t *nodemask, struct zone *preferred_zone,
2334 int classzone_idx, int migratetype)
2338 /* Acquire the per-zone oom lock for each zone */
2339 if (!oom_zonelist_trylock(zonelist, gfp_mask)) {
2340 schedule_timeout_uninterruptible(1);
2345 * PM-freezer should be notified that there might be an OOM killer on
2346 * its way to kill and wake somebody up. This is too early and we might
2347 * end up not killing anything but false positives are acceptable.
2348 * See freeze_processes.
2353 * Go through the zonelist yet one more time, keep very high watermark
2354 * here, this is only to catch a parallel oom killing, we must fail if
2355 * we're still under heavy pressure.
2357 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2358 order, zonelist, high_zoneidx,
2359 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2360 preferred_zone, classzone_idx, migratetype);
2364 if (!(gfp_mask & __GFP_NOFAIL)) {
2365 /* The OOM killer will not help higher order allocs */
2366 if (order > PAGE_ALLOC_COSTLY_ORDER)
2368 /* The OOM killer does not needlessly kill tasks for lowmem */
2369 if (high_zoneidx < ZONE_NORMAL)
2372 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2373 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2374 * The caller should handle page allocation failure by itself if
2375 * it specifies __GFP_THISNODE.
2376 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2378 if (gfp_mask & __GFP_THISNODE)
2381 /* Exhausted what can be done so it's blamo time */
2382 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2385 oom_zonelist_unlock(zonelist, gfp_mask);
2389 #ifdef CONFIG_COMPACTION
2390 /* Try memory compaction for high-order allocations before reclaim */
2391 static struct page *
2392 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2393 struct zonelist *zonelist, enum zone_type high_zoneidx,
2394 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2395 int classzone_idx, int migratetype, enum migrate_mode mode,
2396 int *contended_compaction, bool *deferred_compaction)
2398 unsigned long compact_result;
2404 current->flags |= PF_MEMALLOC;
2405 compact_result = try_to_compact_pages(zonelist, order, gfp_mask,
2407 contended_compaction,
2408 alloc_flags, classzone_idx);
2409 current->flags &= ~PF_MEMALLOC;
2411 switch (compact_result) {
2412 case COMPACT_DEFERRED:
2413 *deferred_compaction = true;
2415 case COMPACT_SKIPPED:
2422 * At least in one zone compaction wasn't deferred or skipped, so let's
2423 * count a compaction stall
2425 count_vm_event(COMPACTSTALL);
2427 page = get_page_from_freelist(gfp_mask, nodemask,
2428 order, zonelist, high_zoneidx,
2429 alloc_flags & ~ALLOC_NO_WATERMARKS,
2430 preferred_zone, classzone_idx, migratetype);
2433 struct zone *zone = page_zone(page);
2435 zone->compact_blockskip_flush = false;
2436 compaction_defer_reset(zone, order, true);
2437 count_vm_event(COMPACTSUCCESS);
2442 * It's bad if compaction run occurs and fails. The most likely reason
2443 * is that pages exist, but not enough to satisfy watermarks.
2445 count_vm_event(COMPACTFAIL);
2452 static inline struct page *
2453 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2454 struct zonelist *zonelist, enum zone_type high_zoneidx,
2455 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2456 int classzone_idx, int migratetype, enum migrate_mode mode,
2457 int *contended_compaction, bool *deferred_compaction)
2461 #endif /* CONFIG_COMPACTION */
2463 /* Perform direct synchronous page reclaim */
2465 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2466 nodemask_t *nodemask)
2468 struct reclaim_state reclaim_state;
2473 /* We now go into synchronous reclaim */
2474 cpuset_memory_pressure_bump();
2475 current->flags |= PF_MEMALLOC;
2476 lockdep_set_current_reclaim_state(gfp_mask);
2477 reclaim_state.reclaimed_slab = 0;
2478 current->reclaim_state = &reclaim_state;
2480 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2482 current->reclaim_state = NULL;
2483 lockdep_clear_current_reclaim_state();
2484 current->flags &= ~PF_MEMALLOC;
2491 /* The really slow allocator path where we enter direct reclaim */
2492 static inline struct page *
2493 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2494 struct zonelist *zonelist, enum zone_type high_zoneidx,
2495 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2496 int classzone_idx, int migratetype, unsigned long *did_some_progress)
2498 struct page *page = NULL;
2499 bool drained = false;
2501 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2503 if (unlikely(!(*did_some_progress)))
2506 /* After successful reclaim, reconsider all zones for allocation */
2507 if (IS_ENABLED(CONFIG_NUMA))
2508 zlc_clear_zones_full(zonelist);
2511 page = get_page_from_freelist(gfp_mask, nodemask, order,
2512 zonelist, high_zoneidx,
2513 alloc_flags & ~ALLOC_NO_WATERMARKS,
2514 preferred_zone, classzone_idx,
2518 * If an allocation failed after direct reclaim, it could be because
2519 * pages are pinned on the per-cpu lists. Drain them and try again
2521 if (!page && !drained) {
2522 drain_all_pages(NULL);
2531 * This is called in the allocator slow-path if the allocation request is of
2532 * sufficient urgency to ignore watermarks and take other desperate measures
2534 static inline struct page *
2535 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2536 struct zonelist *zonelist, enum zone_type high_zoneidx,
2537 nodemask_t *nodemask, struct zone *preferred_zone,
2538 int classzone_idx, int migratetype)
2543 page = get_page_from_freelist(gfp_mask, nodemask, order,
2544 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2545 preferred_zone, classzone_idx, migratetype);
2547 if (!page && gfp_mask & __GFP_NOFAIL)
2548 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2549 } while (!page && (gfp_mask & __GFP_NOFAIL));
2554 static void wake_all_kswapds(unsigned int order,
2555 struct zonelist *zonelist,
2556 enum zone_type high_zoneidx,
2557 struct zone *preferred_zone,
2558 nodemask_t *nodemask)
2563 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2564 high_zoneidx, nodemask)
2565 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2569 gfp_to_alloc_flags(gfp_t gfp_mask)
2571 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2572 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2574 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2575 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2578 * The caller may dip into page reserves a bit more if the caller
2579 * cannot run direct reclaim, or if the caller has realtime scheduling
2580 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2581 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2583 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2587 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2588 * if it can't schedule.
2590 if (!(gfp_mask & __GFP_NOMEMALLOC))
2591 alloc_flags |= ALLOC_HARDER;
2593 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2594 * comment for __cpuset_node_allowed().
2596 alloc_flags &= ~ALLOC_CPUSET;
2597 } else if (unlikely(rt_task(current)) && !in_interrupt())
2598 alloc_flags |= ALLOC_HARDER;
2600 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2601 if (gfp_mask & __GFP_MEMALLOC)
2602 alloc_flags |= ALLOC_NO_WATERMARKS;
2603 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2604 alloc_flags |= ALLOC_NO_WATERMARKS;
2605 else if (!in_interrupt() &&
2606 ((current->flags & PF_MEMALLOC) ||
2607 unlikely(test_thread_flag(TIF_MEMDIE))))
2608 alloc_flags |= ALLOC_NO_WATERMARKS;
2611 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2612 alloc_flags |= ALLOC_CMA;
2617 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2619 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2622 static inline struct page *
2623 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2624 struct zonelist *zonelist, enum zone_type high_zoneidx,
2625 nodemask_t *nodemask, struct zone *preferred_zone,
2626 int classzone_idx, int migratetype)
2628 const gfp_t wait = gfp_mask & __GFP_WAIT;
2629 struct page *page = NULL;
2631 unsigned long pages_reclaimed = 0;
2632 unsigned long did_some_progress;
2633 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2634 bool deferred_compaction = false;
2635 int contended_compaction = COMPACT_CONTENDED_NONE;
2638 * In the slowpath, we sanity check order to avoid ever trying to
2639 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2640 * be using allocators in order of preference for an area that is
2643 if (order >= MAX_ORDER) {
2644 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2649 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2650 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2651 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2652 * using a larger set of nodes after it has established that the
2653 * allowed per node queues are empty and that nodes are
2656 if (IS_ENABLED(CONFIG_NUMA) &&
2657 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2661 if (!(gfp_mask & __GFP_NO_KSWAPD))
2662 wake_all_kswapds(order, zonelist, high_zoneidx,
2663 preferred_zone, nodemask);
2666 * OK, we're below the kswapd watermark and have kicked background
2667 * reclaim. Now things get more complex, so set up alloc_flags according
2668 * to how we want to proceed.
2670 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2673 * Find the true preferred zone if the allocation is unconstrained by
2676 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) {
2677 struct zoneref *preferred_zoneref;
2678 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2679 NULL, &preferred_zone);
2680 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2684 /* This is the last chance, in general, before the goto nopage. */
2685 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2686 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2687 preferred_zone, classzone_idx, migratetype);
2691 /* Allocate without watermarks if the context allows */
2692 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2694 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2695 * the allocation is high priority and these type of
2696 * allocations are system rather than user orientated
2698 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2700 page = __alloc_pages_high_priority(gfp_mask, order,
2701 zonelist, high_zoneidx, nodemask,
2702 preferred_zone, classzone_idx, migratetype);
2708 /* Atomic allocations - we can't balance anything */
2711 * All existing users of the deprecated __GFP_NOFAIL are
2712 * blockable, so warn of any new users that actually allow this
2713 * type of allocation to fail.
2715 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2719 /* Avoid recursion of direct reclaim */
2720 if (current->flags & PF_MEMALLOC)
2723 /* Avoid allocations with no watermarks from looping endlessly */
2724 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2728 * Try direct compaction. The first pass is asynchronous. Subsequent
2729 * attempts after direct reclaim are synchronous
2731 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2732 high_zoneidx, nodemask, alloc_flags,
2734 classzone_idx, migratetype,
2735 migration_mode, &contended_compaction,
2736 &deferred_compaction);
2740 /* Checks for THP-specific high-order allocations */
2741 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2743 * If compaction is deferred for high-order allocations, it is
2744 * because sync compaction recently failed. If this is the case
2745 * and the caller requested a THP allocation, we do not want
2746 * to heavily disrupt the system, so we fail the allocation
2747 * instead of entering direct reclaim.
2749 if (deferred_compaction)
2753 * In all zones where compaction was attempted (and not
2754 * deferred or skipped), lock contention has been detected.
2755 * For THP allocation we do not want to disrupt the others
2756 * so we fallback to base pages instead.
2758 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2762 * If compaction was aborted due to need_resched(), we do not
2763 * want to further increase allocation latency, unless it is
2764 * khugepaged trying to collapse.
2766 if (contended_compaction == COMPACT_CONTENDED_SCHED
2767 && !(current->flags & PF_KTHREAD))
2772 * It can become very expensive to allocate transparent hugepages at
2773 * fault, so use asynchronous memory compaction for THP unless it is
2774 * khugepaged trying to collapse.
2776 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2777 (current->flags & PF_KTHREAD))
2778 migration_mode = MIGRATE_SYNC_LIGHT;
2780 /* Try direct reclaim and then allocating */
2781 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2782 zonelist, high_zoneidx,
2784 alloc_flags, preferred_zone,
2785 classzone_idx, migratetype,
2786 &did_some_progress);
2791 * If we failed to make any progress reclaiming, then we are
2792 * running out of options and have to consider going OOM
2794 if (!did_some_progress) {
2795 if (oom_gfp_allowed(gfp_mask)) {
2796 if (oom_killer_disabled)
2798 /* Coredumps can quickly deplete all memory reserves */
2799 if ((current->flags & PF_DUMPCORE) &&
2800 !(gfp_mask & __GFP_NOFAIL))
2802 page = __alloc_pages_may_oom(gfp_mask, order,
2803 zonelist, high_zoneidx,
2804 nodemask, preferred_zone,
2805 classzone_idx, migratetype);
2809 if (!(gfp_mask & __GFP_NOFAIL)) {
2811 * The oom killer is not called for high-order
2812 * allocations that may fail, so if no progress
2813 * is being made, there are no other options and
2814 * retrying is unlikely to help.
2816 if (order > PAGE_ALLOC_COSTLY_ORDER)
2819 * The oom killer is not called for lowmem
2820 * allocations to prevent needlessly killing
2823 if (high_zoneidx < ZONE_NORMAL)
2831 /* Check if we should retry the allocation */
2832 pages_reclaimed += did_some_progress;
2833 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2835 /* Wait for some write requests to complete then retry */
2836 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2840 * High-order allocations do not necessarily loop after
2841 * direct reclaim and reclaim/compaction depends on compaction
2842 * being called after reclaim so call directly if necessary
2844 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2845 high_zoneidx, nodemask, alloc_flags,
2847 classzone_idx, migratetype,
2848 migration_mode, &contended_compaction,
2849 &deferred_compaction);
2855 warn_alloc_failed(gfp_mask, order, NULL);
2858 if (kmemcheck_enabled)
2859 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2865 * This is the 'heart' of the zoned buddy allocator.
2868 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2869 struct zonelist *zonelist, nodemask_t *nodemask)
2871 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2872 struct zone *preferred_zone;
2873 struct zoneref *preferred_zoneref;
2874 struct page *page = NULL;
2875 int migratetype = gfpflags_to_migratetype(gfp_mask);
2876 unsigned int cpuset_mems_cookie;
2877 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2880 gfp_mask &= gfp_allowed_mask;
2882 lockdep_trace_alloc(gfp_mask);
2884 might_sleep_if(gfp_mask & __GFP_WAIT);
2886 if (should_fail_alloc_page(gfp_mask, order))
2890 * Check the zones suitable for the gfp_mask contain at least one
2891 * valid zone. It's possible to have an empty zonelist as a result
2892 * of GFP_THISNODE and a memoryless node
2894 if (unlikely(!zonelist->_zonerefs->zone))
2897 if (IS_ENABLED(CONFIG_CMA) && migratetype == MIGRATE_MOVABLE)
2898 alloc_flags |= ALLOC_CMA;
2901 cpuset_mems_cookie = read_mems_allowed_begin();
2903 /* The preferred zone is used for statistics later */
2904 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2905 nodemask ? : &cpuset_current_mems_allowed,
2907 if (!preferred_zone)
2909 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2911 /* First allocation attempt */
2912 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2913 zonelist, high_zoneidx, alloc_flags,
2914 preferred_zone, classzone_idx, migratetype);
2915 if (unlikely(!page)) {
2917 * Runtime PM, block IO and its error handling path
2918 * can deadlock because I/O on the device might not
2921 gfp_mask = memalloc_noio_flags(gfp_mask);
2922 page = __alloc_pages_slowpath(gfp_mask, order,
2923 zonelist, high_zoneidx, nodemask,
2924 preferred_zone, classzone_idx, migratetype);
2927 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2931 * When updating a task's mems_allowed, it is possible to race with
2932 * parallel threads in such a way that an allocation can fail while
2933 * the mask is being updated. If a page allocation is about to fail,
2934 * check if the cpuset changed during allocation and if so, retry.
2936 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2941 EXPORT_SYMBOL(__alloc_pages_nodemask);
2944 * Common helper functions.
2946 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2951 * __get_free_pages() returns a 32-bit address, which cannot represent
2954 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2956 page = alloc_pages(gfp_mask, order);
2959 return (unsigned long) page_address(page);
2961 EXPORT_SYMBOL(__get_free_pages);
2963 unsigned long get_zeroed_page(gfp_t gfp_mask)
2965 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2967 EXPORT_SYMBOL(get_zeroed_page);
2969 void __free_pages(struct page *page, unsigned int order)
2971 if (put_page_testzero(page)) {
2973 free_hot_cold_page(page, false);
2975 __free_pages_ok(page, order);
2979 EXPORT_SYMBOL(__free_pages);
2981 void free_pages(unsigned long addr, unsigned int order)
2984 VM_BUG_ON(!virt_addr_valid((void *)addr));
2985 __free_pages(virt_to_page((void *)addr), order);
2989 EXPORT_SYMBOL(free_pages);
2992 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2993 * of the current memory cgroup.
2995 * It should be used when the caller would like to use kmalloc, but since the
2996 * allocation is large, it has to fall back to the page allocator.
2998 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3001 struct mem_cgroup *memcg = NULL;
3003 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3005 page = alloc_pages(gfp_mask, order);
3006 memcg_kmem_commit_charge(page, memcg, order);
3010 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3013 struct mem_cgroup *memcg = NULL;
3015 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3017 page = alloc_pages_node(nid, gfp_mask, order);
3018 memcg_kmem_commit_charge(page, memcg, order);
3023 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3026 void __free_kmem_pages(struct page *page, unsigned int order)
3028 memcg_kmem_uncharge_pages(page, order);
3029 __free_pages(page, order);
3032 void free_kmem_pages(unsigned long addr, unsigned int order)
3035 VM_BUG_ON(!virt_addr_valid((void *)addr));
3036 __free_kmem_pages(virt_to_page((void *)addr), order);
3040 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3043 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3044 unsigned long used = addr + PAGE_ALIGN(size);
3046 split_page(virt_to_page((void *)addr), order);
3047 while (used < alloc_end) {
3052 return (void *)addr;
3056 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3057 * @size: the number of bytes to allocate
3058 * @gfp_mask: GFP flags for the allocation
3060 * This function is similar to alloc_pages(), except that it allocates the
3061 * minimum number of pages to satisfy the request. alloc_pages() can only
3062 * allocate memory in power-of-two pages.
3064 * This function is also limited by MAX_ORDER.
3066 * Memory allocated by this function must be released by free_pages_exact().
3068 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3070 unsigned int order = get_order(size);
3073 addr = __get_free_pages(gfp_mask, order);
3074 return make_alloc_exact(addr, order, size);
3076 EXPORT_SYMBOL(alloc_pages_exact);
3079 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3081 * @nid: the preferred node ID where memory should be allocated
3082 * @size: the number of bytes to allocate
3083 * @gfp_mask: GFP flags for the allocation
3085 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3087 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3090 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3092 unsigned order = get_order(size);
3093 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3096 return make_alloc_exact((unsigned long)page_address(p), order, size);
3100 * free_pages_exact - release memory allocated via alloc_pages_exact()
3101 * @virt: the value returned by alloc_pages_exact.
3102 * @size: size of allocation, same value as passed to alloc_pages_exact().
3104 * Release the memory allocated by a previous call to alloc_pages_exact.
3106 void free_pages_exact(void *virt, size_t size)
3108 unsigned long addr = (unsigned long)virt;
3109 unsigned long end = addr + PAGE_ALIGN(size);
3111 while (addr < end) {
3116 EXPORT_SYMBOL(free_pages_exact);
3119 * nr_free_zone_pages - count number of pages beyond high watermark
3120 * @offset: The zone index of the highest zone
3122 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3123 * high watermark within all zones at or below a given zone index. For each
3124 * zone, the number of pages is calculated as:
3125 * managed_pages - high_pages
3127 static unsigned long nr_free_zone_pages(int offset)
3132 /* Just pick one node, since fallback list is circular */
3133 unsigned long sum = 0;
3135 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3137 for_each_zone_zonelist(zone, z, zonelist, offset) {
3138 unsigned long size = zone->managed_pages;
3139 unsigned long high = high_wmark_pages(zone);
3148 * nr_free_buffer_pages - count number of pages beyond high watermark
3150 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3151 * watermark within ZONE_DMA and ZONE_NORMAL.
3153 unsigned long nr_free_buffer_pages(void)
3155 return nr_free_zone_pages(gfp_zone(GFP_USER));
3157 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3160 * nr_free_pagecache_pages - count number of pages beyond high watermark
3162 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3163 * high watermark within all zones.
3165 unsigned long nr_free_pagecache_pages(void)
3167 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3170 static inline void show_node(struct zone *zone)
3172 if (IS_ENABLED(CONFIG_NUMA))
3173 printk("Node %d ", zone_to_nid(zone));
3176 void si_meminfo(struct sysinfo *val)
3178 val->totalram = totalram_pages;
3179 val->sharedram = global_page_state(NR_SHMEM);
3180 val->freeram = global_page_state(NR_FREE_PAGES);
3181 val->bufferram = nr_blockdev_pages();
3182 val->totalhigh = totalhigh_pages;
3183 val->freehigh = nr_free_highpages();
3184 val->mem_unit = PAGE_SIZE;
3187 EXPORT_SYMBOL(si_meminfo);
3190 void si_meminfo_node(struct sysinfo *val, int nid)
3192 int zone_type; /* needs to be signed */
3193 unsigned long managed_pages = 0;
3194 pg_data_t *pgdat = NODE_DATA(nid);
3196 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3197 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3198 val->totalram = managed_pages;
3199 val->sharedram = node_page_state(nid, NR_SHMEM);
3200 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3201 #ifdef CONFIG_HIGHMEM
3202 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3203 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3209 val->mem_unit = PAGE_SIZE;
3214 * Determine whether the node should be displayed or not, depending on whether
3215 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3217 bool skip_free_areas_node(unsigned int flags, int nid)
3220 unsigned int cpuset_mems_cookie;
3222 if (!(flags & SHOW_MEM_FILTER_NODES))
3226 cpuset_mems_cookie = read_mems_allowed_begin();
3227 ret = !node_isset(nid, cpuset_current_mems_allowed);
3228 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3233 #define K(x) ((x) << (PAGE_SHIFT-10))
3235 static void show_migration_types(unsigned char type)
3237 static const char types[MIGRATE_TYPES] = {
3238 [MIGRATE_UNMOVABLE] = 'U',
3239 [MIGRATE_RECLAIMABLE] = 'E',
3240 [MIGRATE_MOVABLE] = 'M',
3241 [MIGRATE_RESERVE] = 'R',
3243 [MIGRATE_CMA] = 'C',
3245 #ifdef CONFIG_MEMORY_ISOLATION
3246 [MIGRATE_ISOLATE] = 'I',
3249 char tmp[MIGRATE_TYPES + 1];
3253 for (i = 0; i < MIGRATE_TYPES; i++) {
3254 if (type & (1 << i))
3259 printk("(%s) ", tmp);
3263 * Show free area list (used inside shift_scroll-lock stuff)
3264 * We also calculate the percentage fragmentation. We do this by counting the
3265 * memory on each free list with the exception of the first item on the list.
3266 * Suppresses nodes that are not allowed by current's cpuset if
3267 * SHOW_MEM_FILTER_NODES is passed.
3269 void show_free_areas(unsigned int filter)
3274 for_each_populated_zone(zone) {
3275 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3278 printk("%s per-cpu:\n", zone->name);
3280 for_each_online_cpu(cpu) {
3281 struct per_cpu_pageset *pageset;
3283 pageset = per_cpu_ptr(zone->pageset, cpu);
3285 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3286 cpu, pageset->pcp.high,
3287 pageset->pcp.batch, pageset->pcp.count);
3291 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3292 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3294 " dirty:%lu writeback:%lu unstable:%lu\n"
3295 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3296 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3298 global_page_state(NR_ACTIVE_ANON),
3299 global_page_state(NR_INACTIVE_ANON),
3300 global_page_state(NR_ISOLATED_ANON),
3301 global_page_state(NR_ACTIVE_FILE),
3302 global_page_state(NR_INACTIVE_FILE),
3303 global_page_state(NR_ISOLATED_FILE),
3304 global_page_state(NR_UNEVICTABLE),
3305 global_page_state(NR_FILE_DIRTY),
3306 global_page_state(NR_WRITEBACK),
3307 global_page_state(NR_UNSTABLE_NFS),
3308 global_page_state(NR_FREE_PAGES),
3309 global_page_state(NR_SLAB_RECLAIMABLE),
3310 global_page_state(NR_SLAB_UNRECLAIMABLE),
3311 global_page_state(NR_FILE_MAPPED),
3312 global_page_state(NR_SHMEM),
3313 global_page_state(NR_PAGETABLE),
3314 global_page_state(NR_BOUNCE),
3315 global_page_state(NR_FREE_CMA_PAGES));
3317 for_each_populated_zone(zone) {
3320 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3328 " active_anon:%lukB"
3329 " inactive_anon:%lukB"
3330 " active_file:%lukB"
3331 " inactive_file:%lukB"
3332 " unevictable:%lukB"
3333 " isolated(anon):%lukB"
3334 " isolated(file):%lukB"
3342 " slab_reclaimable:%lukB"
3343 " slab_unreclaimable:%lukB"
3344 " kernel_stack:%lukB"
3349 " writeback_tmp:%lukB"
3350 " pages_scanned:%lu"
3351 " all_unreclaimable? %s"
3354 K(zone_page_state(zone, NR_FREE_PAGES)),
3355 K(min_wmark_pages(zone)),
3356 K(low_wmark_pages(zone)),
3357 K(high_wmark_pages(zone)),
3358 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3359 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3360 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3361 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3362 K(zone_page_state(zone, NR_UNEVICTABLE)),
3363 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3364 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3365 K(zone->present_pages),
3366 K(zone->managed_pages),
3367 K(zone_page_state(zone, NR_MLOCK)),
3368 K(zone_page_state(zone, NR_FILE_DIRTY)),
3369 K(zone_page_state(zone, NR_WRITEBACK)),
3370 K(zone_page_state(zone, NR_FILE_MAPPED)),
3371 K(zone_page_state(zone, NR_SHMEM)),
3372 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3373 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3374 zone_page_state(zone, NR_KERNEL_STACK) *
3376 K(zone_page_state(zone, NR_PAGETABLE)),
3377 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3378 K(zone_page_state(zone, NR_BOUNCE)),
3379 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3380 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3381 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3382 (!zone_reclaimable(zone) ? "yes" : "no")
3384 printk("lowmem_reserve[]:");
3385 for (i = 0; i < MAX_NR_ZONES; i++)
3386 printk(" %ld", zone->lowmem_reserve[i]);
3390 for_each_populated_zone(zone) {
3391 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3392 unsigned char types[MAX_ORDER];
3394 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3397 printk("%s: ", zone->name);
3399 spin_lock_irqsave(&zone->lock, flags);
3400 for (order = 0; order < MAX_ORDER; order++) {
3401 struct free_area *area = &zone->free_area[order];
3404 nr[order] = area->nr_free;
3405 total += nr[order] << order;
3408 for (type = 0; type < MIGRATE_TYPES; type++) {
3409 if (!list_empty(&area->free_list[type]))
3410 types[order] |= 1 << type;
3413 spin_unlock_irqrestore(&zone->lock, flags);
3414 for (order = 0; order < MAX_ORDER; order++) {
3415 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3417 show_migration_types(types[order]);
3419 printk("= %lukB\n", K(total));
3422 hugetlb_show_meminfo();
3424 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3426 show_swap_cache_info();
3429 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3431 zoneref->zone = zone;
3432 zoneref->zone_idx = zone_idx(zone);
3436 * Builds allocation fallback zone lists.
3438 * Add all populated zones of a node to the zonelist.
3440 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3444 enum zone_type zone_type = MAX_NR_ZONES;
3448 zone = pgdat->node_zones + zone_type;
3449 if (populated_zone(zone)) {
3450 zoneref_set_zone(zone,
3451 &zonelist->_zonerefs[nr_zones++]);
3452 check_highest_zone(zone_type);
3454 } while (zone_type);
3462 * 0 = automatic detection of better ordering.
3463 * 1 = order by ([node] distance, -zonetype)
3464 * 2 = order by (-zonetype, [node] distance)
3466 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3467 * the same zonelist. So only NUMA can configure this param.
3469 #define ZONELIST_ORDER_DEFAULT 0
3470 #define ZONELIST_ORDER_NODE 1
3471 #define ZONELIST_ORDER_ZONE 2
3473 /* zonelist order in the kernel.
3474 * set_zonelist_order() will set this to NODE or ZONE.
3476 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3477 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3481 /* The value user specified ....changed by config */
3482 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3483 /* string for sysctl */
3484 #define NUMA_ZONELIST_ORDER_LEN 16
3485 char numa_zonelist_order[16] = "default";
3488 * interface for configure zonelist ordering.
3489 * command line option "numa_zonelist_order"
3490 * = "[dD]efault - default, automatic configuration.
3491 * = "[nN]ode - order by node locality, then by zone within node
3492 * = "[zZ]one - order by zone, then by locality within zone
3495 static int __parse_numa_zonelist_order(char *s)
3497 if (*s == 'd' || *s == 'D') {
3498 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3499 } else if (*s == 'n' || *s == 'N') {
3500 user_zonelist_order = ZONELIST_ORDER_NODE;
3501 } else if (*s == 'z' || *s == 'Z') {
3502 user_zonelist_order = ZONELIST_ORDER_ZONE;
3505 "Ignoring invalid numa_zonelist_order value: "
3512 static __init int setup_numa_zonelist_order(char *s)
3519 ret = __parse_numa_zonelist_order(s);
3521 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3525 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3528 * sysctl handler for numa_zonelist_order
3530 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3531 void __user *buffer, size_t *length,
3534 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3536 static DEFINE_MUTEX(zl_order_mutex);
3538 mutex_lock(&zl_order_mutex);
3540 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3544 strcpy(saved_string, (char *)table->data);
3546 ret = proc_dostring(table, write, buffer, length, ppos);
3550 int oldval = user_zonelist_order;
3552 ret = __parse_numa_zonelist_order((char *)table->data);
3555 * bogus value. restore saved string
3557 strncpy((char *)table->data, saved_string,
3558 NUMA_ZONELIST_ORDER_LEN);
3559 user_zonelist_order = oldval;
3560 } else if (oldval != user_zonelist_order) {
3561 mutex_lock(&zonelists_mutex);
3562 build_all_zonelists(NULL, NULL);
3563 mutex_unlock(&zonelists_mutex);
3567 mutex_unlock(&zl_order_mutex);
3572 #define MAX_NODE_LOAD (nr_online_nodes)
3573 static int node_load[MAX_NUMNODES];
3576 * find_next_best_node - find the next node that should appear in a given node's fallback list
3577 * @node: node whose fallback list we're appending
3578 * @used_node_mask: nodemask_t of already used nodes
3580 * We use a number of factors to determine which is the next node that should
3581 * appear on a given node's fallback list. The node should not have appeared
3582 * already in @node's fallback list, and it should be the next closest node
3583 * according to the distance array (which contains arbitrary distance values
3584 * from each node to each node in the system), and should also prefer nodes
3585 * with no CPUs, since presumably they'll have very little allocation pressure
3586 * on them otherwise.
3587 * It returns -1 if no node is found.
3589 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3592 int min_val = INT_MAX;
3593 int best_node = NUMA_NO_NODE;
3594 const struct cpumask *tmp = cpumask_of_node(0);
3596 /* Use the local node if we haven't already */
3597 if (!node_isset(node, *used_node_mask)) {
3598 node_set(node, *used_node_mask);
3602 for_each_node_state(n, N_MEMORY) {
3604 /* Don't want a node to appear more than once */
3605 if (node_isset(n, *used_node_mask))
3608 /* Use the distance array to find the distance */
3609 val = node_distance(node, n);
3611 /* Penalize nodes under us ("prefer the next node") */
3614 /* Give preference to headless and unused nodes */
3615 tmp = cpumask_of_node(n);
3616 if (!cpumask_empty(tmp))
3617 val += PENALTY_FOR_NODE_WITH_CPUS;
3619 /* Slight preference for less loaded node */
3620 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3621 val += node_load[n];
3623 if (val < min_val) {
3630 node_set(best_node, *used_node_mask);
3637 * Build zonelists ordered by node and zones within node.
3638 * This results in maximum locality--normal zone overflows into local
3639 * DMA zone, if any--but risks exhausting DMA zone.
3641 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3644 struct zonelist *zonelist;
3646 zonelist = &pgdat->node_zonelists[0];
3647 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3649 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3650 zonelist->_zonerefs[j].zone = NULL;
3651 zonelist->_zonerefs[j].zone_idx = 0;
3655 * Build gfp_thisnode zonelists
3657 static void build_thisnode_zonelists(pg_data_t *pgdat)
3660 struct zonelist *zonelist;
3662 zonelist = &pgdat->node_zonelists[1];
3663 j = build_zonelists_node(pgdat, zonelist, 0);
3664 zonelist->_zonerefs[j].zone = NULL;
3665 zonelist->_zonerefs[j].zone_idx = 0;
3669 * Build zonelists ordered by zone and nodes within zones.
3670 * This results in conserving DMA zone[s] until all Normal memory is
3671 * exhausted, but results in overflowing to remote node while memory
3672 * may still exist in local DMA zone.
3674 static int node_order[MAX_NUMNODES];
3676 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3679 int zone_type; /* needs to be signed */
3681 struct zonelist *zonelist;
3683 zonelist = &pgdat->node_zonelists[0];
3685 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3686 for (j = 0; j < nr_nodes; j++) {
3687 node = node_order[j];
3688 z = &NODE_DATA(node)->node_zones[zone_type];
3689 if (populated_zone(z)) {
3691 &zonelist->_zonerefs[pos++]);
3692 check_highest_zone(zone_type);
3696 zonelist->_zonerefs[pos].zone = NULL;
3697 zonelist->_zonerefs[pos].zone_idx = 0;
3700 #if defined(CONFIG_64BIT)
3702 * Devices that require DMA32/DMA are relatively rare and do not justify a
3703 * penalty to every machine in case the specialised case applies. Default
3704 * to Node-ordering on 64-bit NUMA machines
3706 static int default_zonelist_order(void)
3708 return ZONELIST_ORDER_NODE;
3712 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3713 * by the kernel. If processes running on node 0 deplete the low memory zone
3714 * then reclaim will occur more frequency increasing stalls and potentially
3715 * be easier to OOM if a large percentage of the zone is under writeback or
3716 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3717 * Hence, default to zone ordering on 32-bit.
3719 static int default_zonelist_order(void)
3721 return ZONELIST_ORDER_ZONE;
3723 #endif /* CONFIG_64BIT */
3725 static void set_zonelist_order(void)
3727 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3728 current_zonelist_order = default_zonelist_order();
3730 current_zonelist_order = user_zonelist_order;
3733 static void build_zonelists(pg_data_t *pgdat)
3737 nodemask_t used_mask;
3738 int local_node, prev_node;
3739 struct zonelist *zonelist;
3740 int order = current_zonelist_order;
3742 /* initialize zonelists */
3743 for (i = 0; i < MAX_ZONELISTS; i++) {
3744 zonelist = pgdat->node_zonelists + i;
3745 zonelist->_zonerefs[0].zone = NULL;
3746 zonelist->_zonerefs[0].zone_idx = 0;
3749 /* NUMA-aware ordering of nodes */
3750 local_node = pgdat->node_id;
3751 load = nr_online_nodes;
3752 prev_node = local_node;
3753 nodes_clear(used_mask);
3755 memset(node_order, 0, sizeof(node_order));
3758 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3760 * We don't want to pressure a particular node.
3761 * So adding penalty to the first node in same
3762 * distance group to make it round-robin.
3764 if (node_distance(local_node, node) !=
3765 node_distance(local_node, prev_node))
3766 node_load[node] = load;
3770 if (order == ZONELIST_ORDER_NODE)
3771 build_zonelists_in_node_order(pgdat, node);
3773 node_order[j++] = node; /* remember order */
3776 if (order == ZONELIST_ORDER_ZONE) {
3777 /* calculate node order -- i.e., DMA last! */
3778 build_zonelists_in_zone_order(pgdat, j);
3781 build_thisnode_zonelists(pgdat);
3784 /* Construct the zonelist performance cache - see further mmzone.h */
3785 static void build_zonelist_cache(pg_data_t *pgdat)
3787 struct zonelist *zonelist;
3788 struct zonelist_cache *zlc;
3791 zonelist = &pgdat->node_zonelists[0];
3792 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3793 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3794 for (z = zonelist->_zonerefs; z->zone; z++)
3795 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3798 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3800 * Return node id of node used for "local" allocations.
3801 * I.e., first node id of first zone in arg node's generic zonelist.
3802 * Used for initializing percpu 'numa_mem', which is used primarily
3803 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3805 int local_memory_node(int node)
3809 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3810 gfp_zone(GFP_KERNEL),
3817 #else /* CONFIG_NUMA */
3819 static void set_zonelist_order(void)
3821 current_zonelist_order = ZONELIST_ORDER_ZONE;
3824 static void build_zonelists(pg_data_t *pgdat)
3826 int node, local_node;
3828 struct zonelist *zonelist;
3830 local_node = pgdat->node_id;
3832 zonelist = &pgdat->node_zonelists[0];
3833 j = build_zonelists_node(pgdat, zonelist, 0);
3836 * Now we build the zonelist so that it contains the zones
3837 * of all the other nodes.
3838 * We don't want to pressure a particular node, so when
3839 * building the zones for node N, we make sure that the
3840 * zones coming right after the local ones are those from
3841 * node N+1 (modulo N)
3843 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3844 if (!node_online(node))
3846 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3848 for (node = 0; node < local_node; node++) {
3849 if (!node_online(node))
3851 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3854 zonelist->_zonerefs[j].zone = NULL;
3855 zonelist->_zonerefs[j].zone_idx = 0;
3858 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3859 static void build_zonelist_cache(pg_data_t *pgdat)
3861 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3864 #endif /* CONFIG_NUMA */
3867 * Boot pageset table. One per cpu which is going to be used for all
3868 * zones and all nodes. The parameters will be set in such a way
3869 * that an item put on a list will immediately be handed over to
3870 * the buddy list. This is safe since pageset manipulation is done
3871 * with interrupts disabled.
3873 * The boot_pagesets must be kept even after bootup is complete for
3874 * unused processors and/or zones. They do play a role for bootstrapping
3875 * hotplugged processors.
3877 * zoneinfo_show() and maybe other functions do
3878 * not check if the processor is online before following the pageset pointer.
3879 * Other parts of the kernel may not check if the zone is available.
3881 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3882 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3883 static void setup_zone_pageset(struct zone *zone);
3886 * Global mutex to protect against size modification of zonelists
3887 * as well as to serialize pageset setup for the new populated zone.
3889 DEFINE_MUTEX(zonelists_mutex);
3891 /* return values int ....just for stop_machine() */
3892 static int __build_all_zonelists(void *data)
3896 pg_data_t *self = data;
3899 memset(node_load, 0, sizeof(node_load));
3902 if (self && !node_online(self->node_id)) {
3903 build_zonelists(self);
3904 build_zonelist_cache(self);
3907 for_each_online_node(nid) {
3908 pg_data_t *pgdat = NODE_DATA(nid);
3910 build_zonelists(pgdat);
3911 build_zonelist_cache(pgdat);
3915 * Initialize the boot_pagesets that are going to be used
3916 * for bootstrapping processors. The real pagesets for
3917 * each zone will be allocated later when the per cpu
3918 * allocator is available.
3920 * boot_pagesets are used also for bootstrapping offline
3921 * cpus if the system is already booted because the pagesets
3922 * are needed to initialize allocators on a specific cpu too.
3923 * F.e. the percpu allocator needs the page allocator which
3924 * needs the percpu allocator in order to allocate its pagesets
3925 * (a chicken-egg dilemma).
3927 for_each_possible_cpu(cpu) {
3928 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3930 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3932 * We now know the "local memory node" for each node--
3933 * i.e., the node of the first zone in the generic zonelist.
3934 * Set up numa_mem percpu variable for on-line cpus. During
3935 * boot, only the boot cpu should be on-line; we'll init the
3936 * secondary cpus' numa_mem as they come on-line. During
3937 * node/memory hotplug, we'll fixup all on-line cpus.
3939 if (cpu_online(cpu))
3940 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3948 * Called with zonelists_mutex held always
3949 * unless system_state == SYSTEM_BOOTING.
3951 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3953 set_zonelist_order();
3955 if (system_state == SYSTEM_BOOTING) {
3956 __build_all_zonelists(NULL);
3957 mminit_verify_zonelist();
3958 cpuset_init_current_mems_allowed();
3960 #ifdef CONFIG_MEMORY_HOTPLUG
3962 setup_zone_pageset(zone);
3964 /* we have to stop all cpus to guarantee there is no user
3966 stop_machine(__build_all_zonelists, pgdat, NULL);
3967 /* cpuset refresh routine should be here */
3969 vm_total_pages = nr_free_pagecache_pages();
3971 * Disable grouping by mobility if the number of pages in the
3972 * system is too low to allow the mechanism to work. It would be
3973 * more accurate, but expensive to check per-zone. This check is
3974 * made on memory-hotadd so a system can start with mobility
3975 * disabled and enable it later
3977 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3978 page_group_by_mobility_disabled = 1;
3980 page_group_by_mobility_disabled = 0;
3982 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3983 "Total pages: %ld\n",
3985 zonelist_order_name[current_zonelist_order],
3986 page_group_by_mobility_disabled ? "off" : "on",
3989 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
3994 * Helper functions to size the waitqueue hash table.
3995 * Essentially these want to choose hash table sizes sufficiently
3996 * large so that collisions trying to wait on pages are rare.
3997 * But in fact, the number of active page waitqueues on typical
3998 * systems is ridiculously low, less than 200. So this is even
3999 * conservative, even though it seems large.
4001 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4002 * waitqueues, i.e. the size of the waitq table given the number of pages.
4004 #define PAGES_PER_WAITQUEUE 256
4006 #ifndef CONFIG_MEMORY_HOTPLUG
4007 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4009 unsigned long size = 1;
4011 pages /= PAGES_PER_WAITQUEUE;
4013 while (size < pages)
4017 * Once we have dozens or even hundreds of threads sleeping
4018 * on IO we've got bigger problems than wait queue collision.
4019 * Limit the size of the wait table to a reasonable size.
4021 size = min(size, 4096UL);
4023 return max(size, 4UL);
4027 * A zone's size might be changed by hot-add, so it is not possible to determine
4028 * a suitable size for its wait_table. So we use the maximum size now.
4030 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4032 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4033 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4034 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4036 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4037 * or more by the traditional way. (See above). It equals:
4039 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4040 * ia64(16K page size) : = ( 8G + 4M)byte.
4041 * powerpc (64K page size) : = (32G +16M)byte.
4043 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4050 * This is an integer logarithm so that shifts can be used later
4051 * to extract the more random high bits from the multiplicative
4052 * hash function before the remainder is taken.
4054 static inline unsigned long wait_table_bits(unsigned long size)
4060 * Check if a pageblock contains reserved pages
4062 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4066 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4067 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4074 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4075 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4076 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4077 * higher will lead to a bigger reserve which will get freed as contiguous
4078 * blocks as reclaim kicks in
4080 static void setup_zone_migrate_reserve(struct zone *zone)
4082 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4084 unsigned long block_migratetype;
4089 * Get the start pfn, end pfn and the number of blocks to reserve
4090 * We have to be careful to be aligned to pageblock_nr_pages to
4091 * make sure that we always check pfn_valid for the first page in
4094 start_pfn = zone->zone_start_pfn;
4095 end_pfn = zone_end_pfn(zone);
4096 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4097 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4101 * Reserve blocks are generally in place to help high-order atomic
4102 * allocations that are short-lived. A min_free_kbytes value that
4103 * would result in more than 2 reserve blocks for atomic allocations
4104 * is assumed to be in place to help anti-fragmentation for the
4105 * future allocation of hugepages at runtime.
4107 reserve = min(2, reserve);
4108 old_reserve = zone->nr_migrate_reserve_block;
4110 /* When memory hot-add, we almost always need to do nothing */
4111 if (reserve == old_reserve)
4113 zone->nr_migrate_reserve_block = reserve;
4115 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4116 if (!pfn_valid(pfn))
4118 page = pfn_to_page(pfn);
4120 /* Watch out for overlapping nodes */
4121 if (page_to_nid(page) != zone_to_nid(zone))
4124 block_migratetype = get_pageblock_migratetype(page);
4126 /* Only test what is necessary when the reserves are not met */
4129 * Blocks with reserved pages will never free, skip
4132 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4133 if (pageblock_is_reserved(pfn, block_end_pfn))
4136 /* If this block is reserved, account for it */
4137 if (block_migratetype == MIGRATE_RESERVE) {
4142 /* Suitable for reserving if this block is movable */
4143 if (block_migratetype == MIGRATE_MOVABLE) {
4144 set_pageblock_migratetype(page,
4146 move_freepages_block(zone, page,
4151 } else if (!old_reserve) {
4153 * At boot time we don't need to scan the whole zone
4154 * for turning off MIGRATE_RESERVE.
4160 * If the reserve is met and this is a previous reserved block,
4163 if (block_migratetype == MIGRATE_RESERVE) {
4164 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4165 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4171 * Initially all pages are reserved - free ones are freed
4172 * up by free_all_bootmem() once the early boot process is
4173 * done. Non-atomic initialization, single-pass.
4175 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4176 unsigned long start_pfn, enum memmap_context context)
4179 unsigned long end_pfn = start_pfn + size;
4183 if (highest_memmap_pfn < end_pfn - 1)
4184 highest_memmap_pfn = end_pfn - 1;
4186 z = &NODE_DATA(nid)->node_zones[zone];
4187 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4189 * There can be holes in boot-time mem_map[]s
4190 * handed to this function. They do not
4191 * exist on hotplugged memory.
4193 if (context == MEMMAP_EARLY) {
4194 if (!early_pfn_valid(pfn))
4196 if (!early_pfn_in_nid(pfn, nid))
4199 page = pfn_to_page(pfn);
4200 set_page_links(page, zone, nid, pfn);
4201 mminit_verify_page_links(page, zone, nid, pfn);
4202 init_page_count(page);
4203 page_mapcount_reset(page);
4204 page_cpupid_reset_last(page);
4205 SetPageReserved(page);
4207 * Mark the block movable so that blocks are reserved for
4208 * movable at startup. This will force kernel allocations
4209 * to reserve their blocks rather than leaking throughout
4210 * the address space during boot when many long-lived
4211 * kernel allocations are made. Later some blocks near
4212 * the start are marked MIGRATE_RESERVE by
4213 * setup_zone_migrate_reserve()
4215 * bitmap is created for zone's valid pfn range. but memmap
4216 * can be created for invalid pages (for alignment)
4217 * check here not to call set_pageblock_migratetype() against
4220 if ((z->zone_start_pfn <= pfn)
4221 && (pfn < zone_end_pfn(z))
4222 && !(pfn & (pageblock_nr_pages - 1)))
4223 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4225 INIT_LIST_HEAD(&page->lru);
4226 #ifdef WANT_PAGE_VIRTUAL
4227 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4228 if (!is_highmem_idx(zone))
4229 set_page_address(page, __va(pfn << PAGE_SHIFT));
4234 static void __meminit zone_init_free_lists(struct zone *zone)
4236 unsigned int order, t;
4237 for_each_migratetype_order(order, t) {
4238 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4239 zone->free_area[order].nr_free = 0;
4243 #ifndef __HAVE_ARCH_MEMMAP_INIT
4244 #define memmap_init(size, nid, zone, start_pfn) \
4245 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4248 static int zone_batchsize(struct zone *zone)
4254 * The per-cpu-pages pools are set to around 1000th of the
4255 * size of the zone. But no more than 1/2 of a meg.
4257 * OK, so we don't know how big the cache is. So guess.
4259 batch = zone->managed_pages / 1024;
4260 if (batch * PAGE_SIZE > 512 * 1024)
4261 batch = (512 * 1024) / PAGE_SIZE;
4262 batch /= 4; /* We effectively *= 4 below */
4267 * Clamp the batch to a 2^n - 1 value. Having a power
4268 * of 2 value was found to be more likely to have
4269 * suboptimal cache aliasing properties in some cases.
4271 * For example if 2 tasks are alternately allocating
4272 * batches of pages, one task can end up with a lot
4273 * of pages of one half of the possible page colors
4274 * and the other with pages of the other colors.
4276 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4281 /* The deferral and batching of frees should be suppressed under NOMMU
4284 * The problem is that NOMMU needs to be able to allocate large chunks
4285 * of contiguous memory as there's no hardware page translation to
4286 * assemble apparent contiguous memory from discontiguous pages.
4288 * Queueing large contiguous runs of pages for batching, however,
4289 * causes the pages to actually be freed in smaller chunks. As there
4290 * can be a significant delay between the individual batches being
4291 * recycled, this leads to the once large chunks of space being
4292 * fragmented and becoming unavailable for high-order allocations.
4299 * pcp->high and pcp->batch values are related and dependent on one another:
4300 * ->batch must never be higher then ->high.
4301 * The following function updates them in a safe manner without read side
4304 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4305 * those fields changing asynchronously (acording the the above rule).
4307 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4308 * outside of boot time (or some other assurance that no concurrent updaters
4311 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4312 unsigned long batch)
4314 /* start with a fail safe value for batch */
4318 /* Update high, then batch, in order */
4325 /* a companion to pageset_set_high() */
4326 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4328 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4331 static void pageset_init(struct per_cpu_pageset *p)
4333 struct per_cpu_pages *pcp;
4336 memset(p, 0, sizeof(*p));
4340 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4341 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4344 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4347 pageset_set_batch(p, batch);
4351 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4352 * to the value high for the pageset p.
4354 static void pageset_set_high(struct per_cpu_pageset *p,
4357 unsigned long batch = max(1UL, high / 4);
4358 if ((high / 4) > (PAGE_SHIFT * 8))
4359 batch = PAGE_SHIFT * 8;
4361 pageset_update(&p->pcp, high, batch);
4364 static void pageset_set_high_and_batch(struct zone *zone,
4365 struct per_cpu_pageset *pcp)
4367 if (percpu_pagelist_fraction)
4368 pageset_set_high(pcp,
4369 (zone->managed_pages /
4370 percpu_pagelist_fraction));
4372 pageset_set_batch(pcp, zone_batchsize(zone));
4375 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4377 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4380 pageset_set_high_and_batch(zone, pcp);
4383 static void __meminit setup_zone_pageset(struct zone *zone)
4386 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4387 for_each_possible_cpu(cpu)
4388 zone_pageset_init(zone, cpu);
4392 * Allocate per cpu pagesets and initialize them.
4393 * Before this call only boot pagesets were available.
4395 void __init setup_per_cpu_pageset(void)
4399 for_each_populated_zone(zone)
4400 setup_zone_pageset(zone);
4403 static noinline __init_refok
4404 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4410 * The per-page waitqueue mechanism uses hashed waitqueues
4413 zone->wait_table_hash_nr_entries =
4414 wait_table_hash_nr_entries(zone_size_pages);
4415 zone->wait_table_bits =
4416 wait_table_bits(zone->wait_table_hash_nr_entries);
4417 alloc_size = zone->wait_table_hash_nr_entries
4418 * sizeof(wait_queue_head_t);
4420 if (!slab_is_available()) {
4421 zone->wait_table = (wait_queue_head_t *)
4422 memblock_virt_alloc_node_nopanic(
4423 alloc_size, zone->zone_pgdat->node_id);
4426 * This case means that a zone whose size was 0 gets new memory
4427 * via memory hot-add.
4428 * But it may be the case that a new node was hot-added. In
4429 * this case vmalloc() will not be able to use this new node's
4430 * memory - this wait_table must be initialized to use this new
4431 * node itself as well.
4432 * To use this new node's memory, further consideration will be
4435 zone->wait_table = vmalloc(alloc_size);
4437 if (!zone->wait_table)
4440 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4441 init_waitqueue_head(zone->wait_table + i);
4446 static __meminit void zone_pcp_init(struct zone *zone)
4449 * per cpu subsystem is not up at this point. The following code
4450 * relies on the ability of the linker to provide the
4451 * offset of a (static) per cpu variable into the per cpu area.
4453 zone->pageset = &boot_pageset;
4455 if (populated_zone(zone))
4456 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4457 zone->name, zone->present_pages,
4458 zone_batchsize(zone));
4461 int __meminit init_currently_empty_zone(struct zone *zone,
4462 unsigned long zone_start_pfn,
4464 enum memmap_context context)
4466 struct pglist_data *pgdat = zone->zone_pgdat;
4468 ret = zone_wait_table_init(zone, size);
4471 pgdat->nr_zones = zone_idx(zone) + 1;
4473 zone->zone_start_pfn = zone_start_pfn;
4475 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4476 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4478 (unsigned long)zone_idx(zone),
4479 zone_start_pfn, (zone_start_pfn + size));
4481 zone_init_free_lists(zone);
4486 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4487 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4489 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4491 int __meminit __early_pfn_to_nid(unsigned long pfn)
4493 unsigned long start_pfn, end_pfn;
4496 * NOTE: The following SMP-unsafe globals are only used early in boot
4497 * when the kernel is running single-threaded.
4499 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4500 static int __meminitdata last_nid;
4502 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4505 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4507 last_start_pfn = start_pfn;
4508 last_end_pfn = end_pfn;
4514 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4516 int __meminit early_pfn_to_nid(unsigned long pfn)
4520 nid = __early_pfn_to_nid(pfn);
4523 /* just returns 0 */
4527 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4528 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4532 nid = __early_pfn_to_nid(pfn);
4533 if (nid >= 0 && nid != node)
4540 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4541 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4542 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4544 * If an architecture guarantees that all ranges registered contain no holes
4545 * and may be freed, this this function may be used instead of calling
4546 * memblock_free_early_nid() manually.
4548 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4550 unsigned long start_pfn, end_pfn;
4553 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4554 start_pfn = min(start_pfn, max_low_pfn);
4555 end_pfn = min(end_pfn, max_low_pfn);
4557 if (start_pfn < end_pfn)
4558 memblock_free_early_nid(PFN_PHYS(start_pfn),
4559 (end_pfn - start_pfn) << PAGE_SHIFT,
4565 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4566 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4568 * If an architecture guarantees that all ranges registered contain no holes and may
4569 * be freed, this function may be used instead of calling memory_present() manually.
4571 void __init sparse_memory_present_with_active_regions(int nid)
4573 unsigned long start_pfn, end_pfn;
4576 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4577 memory_present(this_nid, start_pfn, end_pfn);
4581 * get_pfn_range_for_nid - Return the start and end page frames for a node
4582 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4583 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4584 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4586 * It returns the start and end page frame of a node based on information
4587 * provided by memblock_set_node(). If called for a node
4588 * with no available memory, a warning is printed and the start and end
4591 void __meminit get_pfn_range_for_nid(unsigned int nid,
4592 unsigned long *start_pfn, unsigned long *end_pfn)
4594 unsigned long this_start_pfn, this_end_pfn;
4600 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4601 *start_pfn = min(*start_pfn, this_start_pfn);
4602 *end_pfn = max(*end_pfn, this_end_pfn);
4605 if (*start_pfn == -1UL)
4610 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4611 * assumption is made that zones within a node are ordered in monotonic
4612 * increasing memory addresses so that the "highest" populated zone is used
4614 static void __init find_usable_zone_for_movable(void)
4617 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4618 if (zone_index == ZONE_MOVABLE)
4621 if (arch_zone_highest_possible_pfn[zone_index] >
4622 arch_zone_lowest_possible_pfn[zone_index])
4626 VM_BUG_ON(zone_index == -1);
4627 movable_zone = zone_index;
4631 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4632 * because it is sized independent of architecture. Unlike the other zones,
4633 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4634 * in each node depending on the size of each node and how evenly kernelcore
4635 * is distributed. This helper function adjusts the zone ranges
4636 * provided by the architecture for a given node by using the end of the
4637 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4638 * zones within a node are in order of monotonic increases memory addresses
4640 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4641 unsigned long zone_type,
4642 unsigned long node_start_pfn,
4643 unsigned long node_end_pfn,
4644 unsigned long *zone_start_pfn,
4645 unsigned long *zone_end_pfn)
4647 /* Only adjust if ZONE_MOVABLE is on this node */
4648 if (zone_movable_pfn[nid]) {
4649 /* Size ZONE_MOVABLE */
4650 if (zone_type == ZONE_MOVABLE) {
4651 *zone_start_pfn = zone_movable_pfn[nid];
4652 *zone_end_pfn = min(node_end_pfn,
4653 arch_zone_highest_possible_pfn[movable_zone]);
4655 /* Adjust for ZONE_MOVABLE starting within this range */
4656 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4657 *zone_end_pfn > zone_movable_pfn[nid]) {
4658 *zone_end_pfn = zone_movable_pfn[nid];
4660 /* Check if this whole range is within ZONE_MOVABLE */
4661 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4662 *zone_start_pfn = *zone_end_pfn;
4667 * Return the number of pages a zone spans in a node, including holes
4668 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4670 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4671 unsigned long zone_type,
4672 unsigned long node_start_pfn,
4673 unsigned long node_end_pfn,
4674 unsigned long *ignored)
4676 unsigned long zone_start_pfn, zone_end_pfn;
4678 /* Get the start and end of the zone */
4679 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4680 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4681 adjust_zone_range_for_zone_movable(nid, zone_type,
4682 node_start_pfn, node_end_pfn,
4683 &zone_start_pfn, &zone_end_pfn);
4685 /* Check that this node has pages within the zone's required range */
4686 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4689 /* Move the zone boundaries inside the node if necessary */
4690 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4691 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4693 /* Return the spanned pages */
4694 return zone_end_pfn - zone_start_pfn;
4698 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4699 * then all holes in the requested range will be accounted for.
4701 unsigned long __meminit __absent_pages_in_range(int nid,
4702 unsigned long range_start_pfn,
4703 unsigned long range_end_pfn)
4705 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4706 unsigned long start_pfn, end_pfn;
4709 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4710 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4711 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4712 nr_absent -= end_pfn - start_pfn;
4718 * absent_pages_in_range - Return number of page frames in holes within a range
4719 * @start_pfn: The start PFN to start searching for holes
4720 * @end_pfn: The end PFN to stop searching for holes
4722 * It returns the number of pages frames in memory holes within a range.
4724 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4725 unsigned long end_pfn)
4727 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4730 /* Return the number of page frames in holes in a zone on a node */
4731 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4732 unsigned long zone_type,
4733 unsigned long node_start_pfn,
4734 unsigned long node_end_pfn,
4735 unsigned long *ignored)
4737 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4738 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4739 unsigned long zone_start_pfn, zone_end_pfn;
4741 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4742 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4744 adjust_zone_range_for_zone_movable(nid, zone_type,
4745 node_start_pfn, node_end_pfn,
4746 &zone_start_pfn, &zone_end_pfn);
4747 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4750 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4751 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4752 unsigned long zone_type,
4753 unsigned long node_start_pfn,
4754 unsigned long node_end_pfn,
4755 unsigned long *zones_size)
4757 return zones_size[zone_type];
4760 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4761 unsigned long zone_type,
4762 unsigned long node_start_pfn,
4763 unsigned long node_end_pfn,
4764 unsigned long *zholes_size)
4769 return zholes_size[zone_type];
4772 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4774 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4775 unsigned long node_start_pfn,
4776 unsigned long node_end_pfn,
4777 unsigned long *zones_size,
4778 unsigned long *zholes_size)
4780 unsigned long realtotalpages, totalpages = 0;
4783 for (i = 0; i < MAX_NR_ZONES; i++)
4784 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4788 pgdat->node_spanned_pages = totalpages;
4790 realtotalpages = totalpages;
4791 for (i = 0; i < MAX_NR_ZONES; i++)
4793 zone_absent_pages_in_node(pgdat->node_id, i,
4794 node_start_pfn, node_end_pfn,
4796 pgdat->node_present_pages = realtotalpages;
4797 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4801 #ifndef CONFIG_SPARSEMEM
4803 * Calculate the size of the zone->blockflags rounded to an unsigned long
4804 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4805 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4806 * round what is now in bits to nearest long in bits, then return it in
4809 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4811 unsigned long usemapsize;
4813 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4814 usemapsize = roundup(zonesize, pageblock_nr_pages);
4815 usemapsize = usemapsize >> pageblock_order;
4816 usemapsize *= NR_PAGEBLOCK_BITS;
4817 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4819 return usemapsize / 8;
4822 static void __init setup_usemap(struct pglist_data *pgdat,
4824 unsigned long zone_start_pfn,
4825 unsigned long zonesize)
4827 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4828 zone->pageblock_flags = NULL;
4830 zone->pageblock_flags =
4831 memblock_virt_alloc_node_nopanic(usemapsize,
4835 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4836 unsigned long zone_start_pfn, unsigned long zonesize) {}
4837 #endif /* CONFIG_SPARSEMEM */
4839 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4841 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4842 void __paginginit set_pageblock_order(void)
4846 /* Check that pageblock_nr_pages has not already been setup */
4847 if (pageblock_order)
4850 if (HPAGE_SHIFT > PAGE_SHIFT)
4851 order = HUGETLB_PAGE_ORDER;
4853 order = MAX_ORDER - 1;
4856 * Assume the largest contiguous order of interest is a huge page.
4857 * This value may be variable depending on boot parameters on IA64 and
4860 pageblock_order = order;
4862 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4865 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4866 * is unused as pageblock_order is set at compile-time. See
4867 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4870 void __paginginit set_pageblock_order(void)
4874 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4876 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4877 unsigned long present_pages)
4879 unsigned long pages = spanned_pages;
4882 * Provide a more accurate estimation if there are holes within
4883 * the zone and SPARSEMEM is in use. If there are holes within the
4884 * zone, each populated memory region may cost us one or two extra
4885 * memmap pages due to alignment because memmap pages for each
4886 * populated regions may not naturally algined on page boundary.
4887 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4889 if (spanned_pages > present_pages + (present_pages >> 4) &&
4890 IS_ENABLED(CONFIG_SPARSEMEM))
4891 pages = present_pages;
4893 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4897 * Set up the zone data structures:
4898 * - mark all pages reserved
4899 * - mark all memory queues empty
4900 * - clear the memory bitmaps
4902 * NOTE: pgdat should get zeroed by caller.
4904 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4905 unsigned long node_start_pfn, unsigned long node_end_pfn,
4906 unsigned long *zones_size, unsigned long *zholes_size)
4909 int nid = pgdat->node_id;
4910 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4913 pgdat_resize_init(pgdat);
4914 #ifdef CONFIG_NUMA_BALANCING
4915 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4916 pgdat->numabalancing_migrate_nr_pages = 0;
4917 pgdat->numabalancing_migrate_next_window = jiffies;
4919 init_waitqueue_head(&pgdat->kswapd_wait);
4920 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4921 pgdat_page_ext_init(pgdat);
4923 for (j = 0; j < MAX_NR_ZONES; j++) {
4924 struct zone *zone = pgdat->node_zones + j;
4925 unsigned long size, realsize, freesize, memmap_pages;
4927 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4928 node_end_pfn, zones_size);
4929 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4935 * Adjust freesize so that it accounts for how much memory
4936 * is used by this zone for memmap. This affects the watermark
4937 * and per-cpu initialisations
4939 memmap_pages = calc_memmap_size(size, realsize);
4940 if (!is_highmem_idx(j)) {
4941 if (freesize >= memmap_pages) {
4942 freesize -= memmap_pages;
4945 " %s zone: %lu pages used for memmap\n",
4946 zone_names[j], memmap_pages);
4949 " %s zone: %lu pages exceeds freesize %lu\n",
4950 zone_names[j], memmap_pages, freesize);
4953 /* Account for reserved pages */
4954 if (j == 0 && freesize > dma_reserve) {
4955 freesize -= dma_reserve;
4956 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4957 zone_names[0], dma_reserve);
4960 if (!is_highmem_idx(j))
4961 nr_kernel_pages += freesize;
4962 /* Charge for highmem memmap if there are enough kernel pages */
4963 else if (nr_kernel_pages > memmap_pages * 2)
4964 nr_kernel_pages -= memmap_pages;
4965 nr_all_pages += freesize;
4967 zone->spanned_pages = size;
4968 zone->present_pages = realsize;
4970 * Set an approximate value for lowmem here, it will be adjusted
4971 * when the bootmem allocator frees pages into the buddy system.
4972 * And all highmem pages will be managed by the buddy system.
4974 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4977 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4979 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4981 zone->name = zone_names[j];
4982 spin_lock_init(&zone->lock);
4983 spin_lock_init(&zone->lru_lock);
4984 zone_seqlock_init(zone);
4985 zone->zone_pgdat = pgdat;
4986 zone_pcp_init(zone);
4988 /* For bootup, initialized properly in watermark setup */
4989 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4991 lruvec_init(&zone->lruvec);
4995 set_pageblock_order();
4996 setup_usemap(pgdat, zone, zone_start_pfn, size);
4997 ret = init_currently_empty_zone(zone, zone_start_pfn,
4998 size, MEMMAP_EARLY);
5000 memmap_init(size, nid, j, zone_start_pfn);
5001 zone_start_pfn += size;
5005 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5007 /* Skip empty nodes */
5008 if (!pgdat->node_spanned_pages)
5011 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5012 /* ia64 gets its own node_mem_map, before this, without bootmem */
5013 if (!pgdat->node_mem_map) {
5014 unsigned long size, start, end;
5018 * The zone's endpoints aren't required to be MAX_ORDER
5019 * aligned but the node_mem_map endpoints must be in order
5020 * for the buddy allocator to function correctly.
5022 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5023 end = pgdat_end_pfn(pgdat);
5024 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5025 size = (end - start) * sizeof(struct page);
5026 map = alloc_remap(pgdat->node_id, size);
5028 map = memblock_virt_alloc_node_nopanic(size,
5030 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5032 #ifndef CONFIG_NEED_MULTIPLE_NODES
5034 * With no DISCONTIG, the global mem_map is just set as node 0's
5036 if (pgdat == NODE_DATA(0)) {
5037 mem_map = NODE_DATA(0)->node_mem_map;
5038 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5039 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5040 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5041 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5044 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5047 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5048 unsigned long node_start_pfn, unsigned long *zholes_size)
5050 pg_data_t *pgdat = NODE_DATA(nid);
5051 unsigned long start_pfn = 0;
5052 unsigned long end_pfn = 0;
5054 /* pg_data_t should be reset to zero when it's allocated */
5055 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5057 pgdat->node_id = nid;
5058 pgdat->node_start_pfn = node_start_pfn;
5059 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5060 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5061 printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid,
5062 (u64) start_pfn << PAGE_SHIFT, (u64) (end_pfn << PAGE_SHIFT) - 1);
5064 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5065 zones_size, zholes_size);
5067 alloc_node_mem_map(pgdat);
5068 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5069 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5070 nid, (unsigned long)pgdat,
5071 (unsigned long)pgdat->node_mem_map);
5074 free_area_init_core(pgdat, start_pfn, end_pfn,
5075 zones_size, zholes_size);
5078 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5080 #if MAX_NUMNODES > 1
5082 * Figure out the number of possible node ids.
5084 void __init setup_nr_node_ids(void)
5087 unsigned int highest = 0;
5089 for_each_node_mask(node, node_possible_map)
5091 nr_node_ids = highest + 1;
5096 * node_map_pfn_alignment - determine the maximum internode alignment
5098 * This function should be called after node map is populated and sorted.
5099 * It calculates the maximum power of two alignment which can distinguish
5102 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5103 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5104 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5105 * shifted, 1GiB is enough and this function will indicate so.
5107 * This is used to test whether pfn -> nid mapping of the chosen memory
5108 * model has fine enough granularity to avoid incorrect mapping for the
5109 * populated node map.
5111 * Returns the determined alignment in pfn's. 0 if there is no alignment
5112 * requirement (single node).
5114 unsigned long __init node_map_pfn_alignment(void)
5116 unsigned long accl_mask = 0, last_end = 0;
5117 unsigned long start, end, mask;
5121 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5122 if (!start || last_nid < 0 || last_nid == nid) {
5129 * Start with a mask granular enough to pin-point to the
5130 * start pfn and tick off bits one-by-one until it becomes
5131 * too coarse to separate the current node from the last.
5133 mask = ~((1 << __ffs(start)) - 1);
5134 while (mask && last_end <= (start & (mask << 1)))
5137 /* accumulate all internode masks */
5141 /* convert mask to number of pages */
5142 return ~accl_mask + 1;
5145 /* Find the lowest pfn for a node */
5146 static unsigned long __init find_min_pfn_for_node(int nid)
5148 unsigned long min_pfn = ULONG_MAX;
5149 unsigned long start_pfn;
5152 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5153 min_pfn = min(min_pfn, start_pfn);
5155 if (min_pfn == ULONG_MAX) {
5157 "Could not find start_pfn for node %d\n", nid);
5165 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5167 * It returns the minimum PFN based on information provided via
5168 * memblock_set_node().
5170 unsigned long __init find_min_pfn_with_active_regions(void)
5172 return find_min_pfn_for_node(MAX_NUMNODES);
5176 * early_calculate_totalpages()
5177 * Sum pages in active regions for movable zone.
5178 * Populate N_MEMORY for calculating usable_nodes.
5180 static unsigned long __init early_calculate_totalpages(void)
5182 unsigned long totalpages = 0;
5183 unsigned long start_pfn, end_pfn;
5186 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5187 unsigned long pages = end_pfn - start_pfn;
5189 totalpages += pages;
5191 node_set_state(nid, N_MEMORY);
5197 * Find the PFN the Movable zone begins in each node. Kernel memory
5198 * is spread evenly between nodes as long as the nodes have enough
5199 * memory. When they don't, some nodes will have more kernelcore than
5202 static void __init find_zone_movable_pfns_for_nodes(void)
5205 unsigned long usable_startpfn;
5206 unsigned long kernelcore_node, kernelcore_remaining;
5207 /* save the state before borrow the nodemask */
5208 nodemask_t saved_node_state = node_states[N_MEMORY];
5209 unsigned long totalpages = early_calculate_totalpages();
5210 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5211 struct memblock_region *r;
5213 /* Need to find movable_zone earlier when movable_node is specified. */
5214 find_usable_zone_for_movable();
5217 * If movable_node is specified, ignore kernelcore and movablecore
5220 if (movable_node_is_enabled()) {
5221 for_each_memblock(memory, r) {
5222 if (!memblock_is_hotpluggable(r))
5227 usable_startpfn = PFN_DOWN(r->base);
5228 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5229 min(usable_startpfn, zone_movable_pfn[nid]) :
5237 * If movablecore=nn[KMG] was specified, calculate what size of
5238 * kernelcore that corresponds so that memory usable for
5239 * any allocation type is evenly spread. If both kernelcore
5240 * and movablecore are specified, then the value of kernelcore
5241 * will be used for required_kernelcore if it's greater than
5242 * what movablecore would have allowed.
5244 if (required_movablecore) {
5245 unsigned long corepages;
5248 * Round-up so that ZONE_MOVABLE is at least as large as what
5249 * was requested by the user
5251 required_movablecore =
5252 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5253 corepages = totalpages - required_movablecore;
5255 required_kernelcore = max(required_kernelcore, corepages);
5258 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5259 if (!required_kernelcore)
5262 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5263 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5266 /* Spread kernelcore memory as evenly as possible throughout nodes */
5267 kernelcore_node = required_kernelcore / usable_nodes;
5268 for_each_node_state(nid, N_MEMORY) {
5269 unsigned long start_pfn, end_pfn;
5272 * Recalculate kernelcore_node if the division per node
5273 * now exceeds what is necessary to satisfy the requested
5274 * amount of memory for the kernel
5276 if (required_kernelcore < kernelcore_node)
5277 kernelcore_node = required_kernelcore / usable_nodes;
5280 * As the map is walked, we track how much memory is usable
5281 * by the kernel using kernelcore_remaining. When it is
5282 * 0, the rest of the node is usable by ZONE_MOVABLE
5284 kernelcore_remaining = kernelcore_node;
5286 /* Go through each range of PFNs within this node */
5287 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5288 unsigned long size_pages;
5290 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5291 if (start_pfn >= end_pfn)
5294 /* Account for what is only usable for kernelcore */
5295 if (start_pfn < usable_startpfn) {
5296 unsigned long kernel_pages;
5297 kernel_pages = min(end_pfn, usable_startpfn)
5300 kernelcore_remaining -= min(kernel_pages,
5301 kernelcore_remaining);
5302 required_kernelcore -= min(kernel_pages,
5303 required_kernelcore);
5305 /* Continue if range is now fully accounted */
5306 if (end_pfn <= usable_startpfn) {
5309 * Push zone_movable_pfn to the end so
5310 * that if we have to rebalance
5311 * kernelcore across nodes, we will
5312 * not double account here
5314 zone_movable_pfn[nid] = end_pfn;
5317 start_pfn = usable_startpfn;
5321 * The usable PFN range for ZONE_MOVABLE is from
5322 * start_pfn->end_pfn. Calculate size_pages as the
5323 * number of pages used as kernelcore
5325 size_pages = end_pfn - start_pfn;
5326 if (size_pages > kernelcore_remaining)
5327 size_pages = kernelcore_remaining;
5328 zone_movable_pfn[nid] = start_pfn + size_pages;
5331 * Some kernelcore has been met, update counts and
5332 * break if the kernelcore for this node has been
5335 required_kernelcore -= min(required_kernelcore,
5337 kernelcore_remaining -= size_pages;
5338 if (!kernelcore_remaining)
5344 * If there is still required_kernelcore, we do another pass with one
5345 * less node in the count. This will push zone_movable_pfn[nid] further
5346 * along on the nodes that still have memory until kernelcore is
5350 if (usable_nodes && required_kernelcore > usable_nodes)
5354 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5355 for (nid = 0; nid < MAX_NUMNODES; nid++)
5356 zone_movable_pfn[nid] =
5357 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5360 /* restore the node_state */
5361 node_states[N_MEMORY] = saved_node_state;
5364 /* Any regular or high memory on that node ? */
5365 static void check_for_memory(pg_data_t *pgdat, int nid)
5367 enum zone_type zone_type;
5369 if (N_MEMORY == N_NORMAL_MEMORY)
5372 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5373 struct zone *zone = &pgdat->node_zones[zone_type];
5374 if (populated_zone(zone)) {
5375 node_set_state(nid, N_HIGH_MEMORY);
5376 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5377 zone_type <= ZONE_NORMAL)
5378 node_set_state(nid, N_NORMAL_MEMORY);
5385 * free_area_init_nodes - Initialise all pg_data_t and zone data
5386 * @max_zone_pfn: an array of max PFNs for each zone
5388 * This will call free_area_init_node() for each active node in the system.
5389 * Using the page ranges provided by memblock_set_node(), the size of each
5390 * zone in each node and their holes is calculated. If the maximum PFN
5391 * between two adjacent zones match, it is assumed that the zone is empty.
5392 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5393 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5394 * starts where the previous one ended. For example, ZONE_DMA32 starts
5395 * at arch_max_dma_pfn.
5397 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5399 unsigned long start_pfn, end_pfn;
5402 /* Record where the zone boundaries are */
5403 memset(arch_zone_lowest_possible_pfn, 0,
5404 sizeof(arch_zone_lowest_possible_pfn));
5405 memset(arch_zone_highest_possible_pfn, 0,
5406 sizeof(arch_zone_highest_possible_pfn));
5407 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5408 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5409 for (i = 1; i < MAX_NR_ZONES; i++) {
5410 if (i == ZONE_MOVABLE)
5412 arch_zone_lowest_possible_pfn[i] =
5413 arch_zone_highest_possible_pfn[i-1];
5414 arch_zone_highest_possible_pfn[i] =
5415 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5417 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5418 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5420 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5421 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5422 find_zone_movable_pfns_for_nodes();
5424 /* Print out the zone ranges */
5425 pr_info("Zone ranges:\n");
5426 for (i = 0; i < MAX_NR_ZONES; i++) {
5427 if (i == ZONE_MOVABLE)
5429 pr_info(" %-8s ", zone_names[i]);
5430 if (arch_zone_lowest_possible_pfn[i] ==
5431 arch_zone_highest_possible_pfn[i])
5434 pr_cont("[mem %0#10lx-%0#10lx]\n",
5435 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5436 (arch_zone_highest_possible_pfn[i]
5437 << PAGE_SHIFT) - 1);
5440 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5441 pr_info("Movable zone start for each node\n");
5442 for (i = 0; i < MAX_NUMNODES; i++) {
5443 if (zone_movable_pfn[i])
5444 pr_info(" Node %d: %#010lx\n", i,
5445 zone_movable_pfn[i] << PAGE_SHIFT);
5448 /* Print out the early node map */
5449 pr_info("Early memory node ranges\n");
5450 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5451 pr_info(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5452 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5454 /* Initialise every node */
5455 mminit_verify_pageflags_layout();
5456 setup_nr_node_ids();
5457 for_each_online_node(nid) {
5458 pg_data_t *pgdat = NODE_DATA(nid);
5459 free_area_init_node(nid, NULL,
5460 find_min_pfn_for_node(nid), NULL);
5462 /* Any memory on that node */
5463 if (pgdat->node_present_pages)
5464 node_set_state(nid, N_MEMORY);
5465 check_for_memory(pgdat, nid);
5469 static int __init cmdline_parse_core(char *p, unsigned long *core)
5471 unsigned long long coremem;
5475 coremem = memparse(p, &p);
5476 *core = coremem >> PAGE_SHIFT;
5478 /* Paranoid check that UL is enough for the coremem value */
5479 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5485 * kernelcore=size sets the amount of memory for use for allocations that
5486 * cannot be reclaimed or migrated.
5488 static int __init cmdline_parse_kernelcore(char *p)
5490 return cmdline_parse_core(p, &required_kernelcore);
5494 * movablecore=size sets the amount of memory for use for allocations that
5495 * can be reclaimed or migrated.
5497 static int __init cmdline_parse_movablecore(char *p)
5499 return cmdline_parse_core(p, &required_movablecore);
5502 early_param("kernelcore", cmdline_parse_kernelcore);
5503 early_param("movablecore", cmdline_parse_movablecore);
5505 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5507 void adjust_managed_page_count(struct page *page, long count)
5509 spin_lock(&managed_page_count_lock);
5510 page_zone(page)->managed_pages += count;
5511 totalram_pages += count;
5512 #ifdef CONFIG_HIGHMEM
5513 if (PageHighMem(page))
5514 totalhigh_pages += count;
5516 spin_unlock(&managed_page_count_lock);
5518 EXPORT_SYMBOL(adjust_managed_page_count);
5520 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5523 unsigned long pages = 0;
5525 start = (void *)PAGE_ALIGN((unsigned long)start);
5526 end = (void *)((unsigned long)end & PAGE_MASK);
5527 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5528 if ((unsigned int)poison <= 0xFF)
5529 memset(pos, poison, PAGE_SIZE);
5530 free_reserved_page(virt_to_page(pos));
5534 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5535 s, pages << (PAGE_SHIFT - 10), start, end);
5539 EXPORT_SYMBOL(free_reserved_area);
5541 #ifdef CONFIG_HIGHMEM
5542 void free_highmem_page(struct page *page)
5544 __free_reserved_page(page);
5546 page_zone(page)->managed_pages++;
5552 void __init mem_init_print_info(const char *str)
5554 unsigned long physpages, codesize, datasize, rosize, bss_size;
5555 unsigned long init_code_size, init_data_size;
5557 physpages = get_num_physpages();
5558 codesize = _etext - _stext;
5559 datasize = _edata - _sdata;
5560 rosize = __end_rodata - __start_rodata;
5561 bss_size = __bss_stop - __bss_start;
5562 init_data_size = __init_end - __init_begin;
5563 init_code_size = _einittext - _sinittext;
5566 * Detect special cases and adjust section sizes accordingly:
5567 * 1) .init.* may be embedded into .data sections
5568 * 2) .init.text.* may be out of [__init_begin, __init_end],
5569 * please refer to arch/tile/kernel/vmlinux.lds.S.
5570 * 3) .rodata.* may be embedded into .text or .data sections.
5572 #define adj_init_size(start, end, size, pos, adj) \
5574 if (start <= pos && pos < end && size > adj) \
5578 adj_init_size(__init_begin, __init_end, init_data_size,
5579 _sinittext, init_code_size);
5580 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5581 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5582 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5583 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5585 #undef adj_init_size
5587 pr_info("Memory: %luK/%luK available "
5588 "(%luK kernel code, %luK rwdata, %luK rodata, "
5589 "%luK init, %luK bss, %luK reserved"
5590 #ifdef CONFIG_HIGHMEM
5594 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5595 codesize >> 10, datasize >> 10, rosize >> 10,
5596 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5597 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5598 #ifdef CONFIG_HIGHMEM
5599 totalhigh_pages << (PAGE_SHIFT-10),
5601 str ? ", " : "", str ? str : "");
5605 * set_dma_reserve - set the specified number of pages reserved in the first zone
5606 * @new_dma_reserve: The number of pages to mark reserved
5608 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5609 * In the DMA zone, a significant percentage may be consumed by kernel image
5610 * and other unfreeable allocations which can skew the watermarks badly. This
5611 * function may optionally be used to account for unfreeable pages in the
5612 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5613 * smaller per-cpu batchsize.
5615 void __init set_dma_reserve(unsigned long new_dma_reserve)
5617 dma_reserve = new_dma_reserve;
5620 void __init free_area_init(unsigned long *zones_size)
5622 free_area_init_node(0, zones_size,
5623 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5626 static int page_alloc_cpu_notify(struct notifier_block *self,
5627 unsigned long action, void *hcpu)
5629 int cpu = (unsigned long)hcpu;
5631 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5632 lru_add_drain_cpu(cpu);
5636 * Spill the event counters of the dead processor
5637 * into the current processors event counters.
5638 * This artificially elevates the count of the current
5641 vm_events_fold_cpu(cpu);
5644 * Zero the differential counters of the dead processor
5645 * so that the vm statistics are consistent.
5647 * This is only okay since the processor is dead and cannot
5648 * race with what we are doing.
5650 cpu_vm_stats_fold(cpu);
5655 void __init page_alloc_init(void)
5657 hotcpu_notifier(page_alloc_cpu_notify, 0);
5661 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5662 * or min_free_kbytes changes.
5664 static void calculate_totalreserve_pages(void)
5666 struct pglist_data *pgdat;
5667 unsigned long reserve_pages = 0;
5668 enum zone_type i, j;
5670 for_each_online_pgdat(pgdat) {
5671 for (i = 0; i < MAX_NR_ZONES; i++) {
5672 struct zone *zone = pgdat->node_zones + i;
5675 /* Find valid and maximum lowmem_reserve in the zone */
5676 for (j = i; j < MAX_NR_ZONES; j++) {
5677 if (zone->lowmem_reserve[j] > max)
5678 max = zone->lowmem_reserve[j];
5681 /* we treat the high watermark as reserved pages. */
5682 max += high_wmark_pages(zone);
5684 if (max > zone->managed_pages)
5685 max = zone->managed_pages;
5686 reserve_pages += max;
5688 * Lowmem reserves are not available to
5689 * GFP_HIGHUSER page cache allocations and
5690 * kswapd tries to balance zones to their high
5691 * watermark. As a result, neither should be
5692 * regarded as dirtyable memory, to prevent a
5693 * situation where reclaim has to clean pages
5694 * in order to balance the zones.
5696 zone->dirty_balance_reserve = max;
5699 dirty_balance_reserve = reserve_pages;
5700 totalreserve_pages = reserve_pages;
5704 * setup_per_zone_lowmem_reserve - called whenever
5705 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5706 * has a correct pages reserved value, so an adequate number of
5707 * pages are left in the zone after a successful __alloc_pages().
5709 static void setup_per_zone_lowmem_reserve(void)
5711 struct pglist_data *pgdat;
5712 enum zone_type j, idx;
5714 for_each_online_pgdat(pgdat) {
5715 for (j = 0; j < MAX_NR_ZONES; j++) {
5716 struct zone *zone = pgdat->node_zones + j;
5717 unsigned long managed_pages = zone->managed_pages;
5719 zone->lowmem_reserve[j] = 0;
5723 struct zone *lower_zone;
5727 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5728 sysctl_lowmem_reserve_ratio[idx] = 1;
5730 lower_zone = pgdat->node_zones + idx;
5731 lower_zone->lowmem_reserve[j] = managed_pages /
5732 sysctl_lowmem_reserve_ratio[idx];
5733 managed_pages += lower_zone->managed_pages;
5738 /* update totalreserve_pages */
5739 calculate_totalreserve_pages();
5742 static void __setup_per_zone_wmarks(void)
5744 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5745 unsigned long lowmem_pages = 0;
5747 unsigned long flags;
5749 /* Calculate total number of !ZONE_HIGHMEM pages */
5750 for_each_zone(zone) {
5751 if (!is_highmem(zone))
5752 lowmem_pages += zone->managed_pages;
5755 for_each_zone(zone) {
5758 spin_lock_irqsave(&zone->lock, flags);
5759 tmp = (u64)pages_min * zone->managed_pages;
5760 do_div(tmp, lowmem_pages);
5761 if (is_highmem(zone)) {
5763 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5764 * need highmem pages, so cap pages_min to a small
5767 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5768 * deltas controls asynch page reclaim, and so should
5769 * not be capped for highmem.
5771 unsigned long min_pages;
5773 min_pages = zone->managed_pages / 1024;
5774 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5775 zone->watermark[WMARK_MIN] = min_pages;
5778 * If it's a lowmem zone, reserve a number of pages
5779 * proportionate to the zone's size.
5781 zone->watermark[WMARK_MIN] = tmp;
5784 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5785 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5787 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5788 high_wmark_pages(zone) - low_wmark_pages(zone) -
5789 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5791 setup_zone_migrate_reserve(zone);
5792 spin_unlock_irqrestore(&zone->lock, flags);
5795 /* update totalreserve_pages */
5796 calculate_totalreserve_pages();
5800 * setup_per_zone_wmarks - called when min_free_kbytes changes
5801 * or when memory is hot-{added|removed}
5803 * Ensures that the watermark[min,low,high] values for each zone are set
5804 * correctly with respect to min_free_kbytes.
5806 void setup_per_zone_wmarks(void)
5808 mutex_lock(&zonelists_mutex);
5809 __setup_per_zone_wmarks();
5810 mutex_unlock(&zonelists_mutex);
5814 * The inactive anon list should be small enough that the VM never has to
5815 * do too much work, but large enough that each inactive page has a chance
5816 * to be referenced again before it is swapped out.
5818 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5819 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5820 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5821 * the anonymous pages are kept on the inactive list.
5824 * memory ratio inactive anon
5825 * -------------------------------------
5834 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5836 unsigned int gb, ratio;
5838 /* Zone size in gigabytes */
5839 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5841 ratio = int_sqrt(10 * gb);
5845 zone->inactive_ratio = ratio;
5848 static void __meminit setup_per_zone_inactive_ratio(void)
5853 calculate_zone_inactive_ratio(zone);
5857 * Initialise min_free_kbytes.
5859 * For small machines we want it small (128k min). For large machines
5860 * we want it large (64MB max). But it is not linear, because network
5861 * bandwidth does not increase linearly with machine size. We use
5863 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5864 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5880 int __meminit init_per_zone_wmark_min(void)
5882 unsigned long lowmem_kbytes;
5883 int new_min_free_kbytes;
5885 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5886 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5888 if (new_min_free_kbytes > user_min_free_kbytes) {
5889 min_free_kbytes = new_min_free_kbytes;
5890 if (min_free_kbytes < 128)
5891 min_free_kbytes = 128;
5892 if (min_free_kbytes > 65536)
5893 min_free_kbytes = 65536;
5895 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5896 new_min_free_kbytes, user_min_free_kbytes);
5898 setup_per_zone_wmarks();
5899 refresh_zone_stat_thresholds();
5900 setup_per_zone_lowmem_reserve();
5901 setup_per_zone_inactive_ratio();
5904 module_init(init_per_zone_wmark_min)
5907 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5908 * that we can call two helper functions whenever min_free_kbytes
5911 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5912 void __user *buffer, size_t *length, loff_t *ppos)
5916 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5921 user_min_free_kbytes = min_free_kbytes;
5922 setup_per_zone_wmarks();
5928 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5929 void __user *buffer, size_t *length, loff_t *ppos)
5934 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5939 zone->min_unmapped_pages = (zone->managed_pages *
5940 sysctl_min_unmapped_ratio) / 100;
5944 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5945 void __user *buffer, size_t *length, loff_t *ppos)
5950 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5955 zone->min_slab_pages = (zone->managed_pages *
5956 sysctl_min_slab_ratio) / 100;
5962 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5963 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5964 * whenever sysctl_lowmem_reserve_ratio changes.
5966 * The reserve ratio obviously has absolutely no relation with the
5967 * minimum watermarks. The lowmem reserve ratio can only make sense
5968 * if in function of the boot time zone sizes.
5970 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5971 void __user *buffer, size_t *length, loff_t *ppos)
5973 proc_dointvec_minmax(table, write, buffer, length, ppos);
5974 setup_per_zone_lowmem_reserve();
5979 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5980 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5981 * pagelist can have before it gets flushed back to buddy allocator.
5983 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5984 void __user *buffer, size_t *length, loff_t *ppos)
5987 int old_percpu_pagelist_fraction;
5990 mutex_lock(&pcp_batch_high_lock);
5991 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5993 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5994 if (!write || ret < 0)
5997 /* Sanity checking to avoid pcp imbalance */
5998 if (percpu_pagelist_fraction &&
5999 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6000 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6006 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6009 for_each_populated_zone(zone) {
6012 for_each_possible_cpu(cpu)
6013 pageset_set_high_and_batch(zone,
6014 per_cpu_ptr(zone->pageset, cpu));
6017 mutex_unlock(&pcp_batch_high_lock);
6021 int hashdist = HASHDIST_DEFAULT;
6024 static int __init set_hashdist(char *str)
6028 hashdist = simple_strtoul(str, &str, 0);
6031 __setup("hashdist=", set_hashdist);
6035 * allocate a large system hash table from bootmem
6036 * - it is assumed that the hash table must contain an exact power-of-2
6037 * quantity of entries
6038 * - limit is the number of hash buckets, not the total allocation size
6040 void *__init alloc_large_system_hash(const char *tablename,
6041 unsigned long bucketsize,
6042 unsigned long numentries,
6045 unsigned int *_hash_shift,
6046 unsigned int *_hash_mask,
6047 unsigned long low_limit,
6048 unsigned long high_limit)
6050 unsigned long long max = high_limit;
6051 unsigned long log2qty, size;
6054 /* allow the kernel cmdline to have a say */
6056 /* round applicable memory size up to nearest megabyte */
6057 numentries = nr_kernel_pages;
6059 /* It isn't necessary when PAGE_SIZE >= 1MB */
6060 if (PAGE_SHIFT < 20)
6061 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6063 /* limit to 1 bucket per 2^scale bytes of low memory */
6064 if (scale > PAGE_SHIFT)
6065 numentries >>= (scale - PAGE_SHIFT);
6067 numentries <<= (PAGE_SHIFT - scale);
6069 /* Make sure we've got at least a 0-order allocation.. */
6070 if (unlikely(flags & HASH_SMALL)) {
6071 /* Makes no sense without HASH_EARLY */
6072 WARN_ON(!(flags & HASH_EARLY));
6073 if (!(numentries >> *_hash_shift)) {
6074 numentries = 1UL << *_hash_shift;
6075 BUG_ON(!numentries);
6077 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6078 numentries = PAGE_SIZE / bucketsize;
6080 numentries = roundup_pow_of_two(numentries);
6082 /* limit allocation size to 1/16 total memory by default */
6084 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6085 do_div(max, bucketsize);
6087 max = min(max, 0x80000000ULL);
6089 if (numentries < low_limit)
6090 numentries = low_limit;
6091 if (numentries > max)
6094 log2qty = ilog2(numentries);
6097 size = bucketsize << log2qty;
6098 if (flags & HASH_EARLY)
6099 table = memblock_virt_alloc_nopanic(size, 0);
6101 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6104 * If bucketsize is not a power-of-two, we may free
6105 * some pages at the end of hash table which
6106 * alloc_pages_exact() automatically does
6108 if (get_order(size) < MAX_ORDER) {
6109 table = alloc_pages_exact(size, GFP_ATOMIC);
6110 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6113 } while (!table && size > PAGE_SIZE && --log2qty);
6116 panic("Failed to allocate %s hash table\n", tablename);
6118 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6121 ilog2(size) - PAGE_SHIFT,
6125 *_hash_shift = log2qty;
6127 *_hash_mask = (1 << log2qty) - 1;
6132 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6133 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6136 #ifdef CONFIG_SPARSEMEM
6137 return __pfn_to_section(pfn)->pageblock_flags;
6139 return zone->pageblock_flags;
6140 #endif /* CONFIG_SPARSEMEM */
6143 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6145 #ifdef CONFIG_SPARSEMEM
6146 pfn &= (PAGES_PER_SECTION-1);
6147 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6149 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6150 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6151 #endif /* CONFIG_SPARSEMEM */
6155 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6156 * @page: The page within the block of interest
6157 * @pfn: The target page frame number
6158 * @end_bitidx: The last bit of interest to retrieve
6159 * @mask: mask of bits that the caller is interested in
6161 * Return: pageblock_bits flags
6163 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6164 unsigned long end_bitidx,
6168 unsigned long *bitmap;
6169 unsigned long bitidx, word_bitidx;
6172 zone = page_zone(page);
6173 bitmap = get_pageblock_bitmap(zone, pfn);
6174 bitidx = pfn_to_bitidx(zone, pfn);
6175 word_bitidx = bitidx / BITS_PER_LONG;
6176 bitidx &= (BITS_PER_LONG-1);
6178 word = bitmap[word_bitidx];
6179 bitidx += end_bitidx;
6180 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6184 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6185 * @page: The page within the block of interest
6186 * @flags: The flags to set
6187 * @pfn: The target page frame number
6188 * @end_bitidx: The last bit of interest
6189 * @mask: mask of bits that the caller is interested in
6191 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6193 unsigned long end_bitidx,
6197 unsigned long *bitmap;
6198 unsigned long bitidx, word_bitidx;
6199 unsigned long old_word, word;
6201 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6203 zone = page_zone(page);
6204 bitmap = get_pageblock_bitmap(zone, pfn);
6205 bitidx = pfn_to_bitidx(zone, pfn);
6206 word_bitidx = bitidx / BITS_PER_LONG;
6207 bitidx &= (BITS_PER_LONG-1);
6209 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6211 bitidx += end_bitidx;
6212 mask <<= (BITS_PER_LONG - bitidx - 1);
6213 flags <<= (BITS_PER_LONG - bitidx - 1);
6215 word = ACCESS_ONCE(bitmap[word_bitidx]);
6217 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6218 if (word == old_word)
6225 * This function checks whether pageblock includes unmovable pages or not.
6226 * If @count is not zero, it is okay to include less @count unmovable pages
6228 * PageLRU check without isolation or lru_lock could race so that
6229 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6230 * expect this function should be exact.
6232 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6233 bool skip_hwpoisoned_pages)
6235 unsigned long pfn, iter, found;
6239 * For avoiding noise data, lru_add_drain_all() should be called
6240 * If ZONE_MOVABLE, the zone never contains unmovable pages
6242 if (zone_idx(zone) == ZONE_MOVABLE)
6244 mt = get_pageblock_migratetype(page);
6245 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6248 pfn = page_to_pfn(page);
6249 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6250 unsigned long check = pfn + iter;
6252 if (!pfn_valid_within(check))
6255 page = pfn_to_page(check);
6258 * Hugepages are not in LRU lists, but they're movable.
6259 * We need not scan over tail pages bacause we don't
6260 * handle each tail page individually in migration.
6262 if (PageHuge(page)) {
6263 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6268 * We can't use page_count without pin a page
6269 * because another CPU can free compound page.
6270 * This check already skips compound tails of THP
6271 * because their page->_count is zero at all time.
6273 if (!atomic_read(&page->_count)) {
6274 if (PageBuddy(page))
6275 iter += (1 << page_order(page)) - 1;
6280 * The HWPoisoned page may be not in buddy system, and
6281 * page_count() is not 0.
6283 if (skip_hwpoisoned_pages && PageHWPoison(page))
6289 * If there are RECLAIMABLE pages, we need to check
6290 * it. But now, memory offline itself doesn't call
6291 * shrink_node_slabs() and it still to be fixed.
6294 * If the page is not RAM, page_count()should be 0.
6295 * we don't need more check. This is an _used_ not-movable page.
6297 * The problematic thing here is PG_reserved pages. PG_reserved
6298 * is set to both of a memory hole page and a _used_ kernel
6307 bool is_pageblock_removable_nolock(struct page *page)
6313 * We have to be careful here because we are iterating over memory
6314 * sections which are not zone aware so we might end up outside of
6315 * the zone but still within the section.
6316 * We have to take care about the node as well. If the node is offline
6317 * its NODE_DATA will be NULL - see page_zone.
6319 if (!node_online(page_to_nid(page)))
6322 zone = page_zone(page);
6323 pfn = page_to_pfn(page);
6324 if (!zone_spans_pfn(zone, pfn))
6327 return !has_unmovable_pages(zone, page, 0, true);
6332 static unsigned long pfn_max_align_down(unsigned long pfn)
6334 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6335 pageblock_nr_pages) - 1);
6338 static unsigned long pfn_max_align_up(unsigned long pfn)
6340 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6341 pageblock_nr_pages));
6344 /* [start, end) must belong to a single zone. */
6345 static int __alloc_contig_migrate_range(struct compact_control *cc,
6346 unsigned long start, unsigned long end)
6348 /* This function is based on compact_zone() from compaction.c. */
6349 unsigned long nr_reclaimed;
6350 unsigned long pfn = start;
6351 unsigned int tries = 0;
6356 while (pfn < end || !list_empty(&cc->migratepages)) {
6357 if (fatal_signal_pending(current)) {
6362 if (list_empty(&cc->migratepages)) {
6363 cc->nr_migratepages = 0;
6364 pfn = isolate_migratepages_range(cc, pfn, end);
6370 } else if (++tries == 5) {
6371 ret = ret < 0 ? ret : -EBUSY;
6375 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6377 cc->nr_migratepages -= nr_reclaimed;
6379 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6380 NULL, 0, cc->mode, MR_CMA);
6383 putback_movable_pages(&cc->migratepages);
6390 * alloc_contig_range() -- tries to allocate given range of pages
6391 * @start: start PFN to allocate
6392 * @end: one-past-the-last PFN to allocate
6393 * @migratetype: migratetype of the underlaying pageblocks (either
6394 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6395 * in range must have the same migratetype and it must
6396 * be either of the two.
6398 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6399 * aligned, however it's the caller's responsibility to guarantee that
6400 * we are the only thread that changes migrate type of pageblocks the
6403 * The PFN range must belong to a single zone.
6405 * Returns zero on success or negative error code. On success all
6406 * pages which PFN is in [start, end) are allocated for the caller and
6407 * need to be freed with free_contig_range().
6409 int alloc_contig_range(unsigned long start, unsigned long end,
6410 unsigned migratetype)
6412 unsigned long outer_start, outer_end;
6415 struct compact_control cc = {
6416 .nr_migratepages = 0,
6418 .zone = page_zone(pfn_to_page(start)),
6419 .mode = MIGRATE_SYNC,
6420 .ignore_skip_hint = true,
6422 INIT_LIST_HEAD(&cc.migratepages);
6425 * What we do here is we mark all pageblocks in range as
6426 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6427 * have different sizes, and due to the way page allocator
6428 * work, we align the range to biggest of the two pages so
6429 * that page allocator won't try to merge buddies from
6430 * different pageblocks and change MIGRATE_ISOLATE to some
6431 * other migration type.
6433 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6434 * migrate the pages from an unaligned range (ie. pages that
6435 * we are interested in). This will put all the pages in
6436 * range back to page allocator as MIGRATE_ISOLATE.
6438 * When this is done, we take the pages in range from page
6439 * allocator removing them from the buddy system. This way
6440 * page allocator will never consider using them.
6442 * This lets us mark the pageblocks back as
6443 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6444 * aligned range but not in the unaligned, original range are
6445 * put back to page allocator so that buddy can use them.
6448 ret = start_isolate_page_range(pfn_max_align_down(start),
6449 pfn_max_align_up(end), migratetype,
6454 ret = __alloc_contig_migrate_range(&cc, start, end);
6459 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6460 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6461 * more, all pages in [start, end) are free in page allocator.
6462 * What we are going to do is to allocate all pages from
6463 * [start, end) (that is remove them from page allocator).
6465 * The only problem is that pages at the beginning and at the
6466 * end of interesting range may be not aligned with pages that
6467 * page allocator holds, ie. they can be part of higher order
6468 * pages. Because of this, we reserve the bigger range and
6469 * once this is done free the pages we are not interested in.
6471 * We don't have to hold zone->lock here because the pages are
6472 * isolated thus they won't get removed from buddy.
6475 lru_add_drain_all();
6476 drain_all_pages(cc.zone);
6479 outer_start = start;
6480 while (!PageBuddy(pfn_to_page(outer_start))) {
6481 if (++order >= MAX_ORDER) {
6485 outer_start &= ~0UL << order;
6488 /* Make sure the range is really isolated. */
6489 if (test_pages_isolated(outer_start, end, false)) {
6490 pr_info("%s: [%lx, %lx) PFNs busy\n",
6491 __func__, outer_start, end);
6496 /* Grab isolated pages from freelists. */
6497 outer_end = isolate_freepages_range(&cc, outer_start, end);
6503 /* Free head and tail (if any) */
6504 if (start != outer_start)
6505 free_contig_range(outer_start, start - outer_start);
6506 if (end != outer_end)
6507 free_contig_range(end, outer_end - end);
6510 undo_isolate_page_range(pfn_max_align_down(start),
6511 pfn_max_align_up(end), migratetype);
6515 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6517 unsigned int count = 0;
6519 for (; nr_pages--; pfn++) {
6520 struct page *page = pfn_to_page(pfn);
6522 count += page_count(page) != 1;
6525 WARN(count != 0, "%d pages are still in use!\n", count);
6529 #ifdef CONFIG_MEMORY_HOTPLUG
6531 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6532 * page high values need to be recalulated.
6534 void __meminit zone_pcp_update(struct zone *zone)
6537 mutex_lock(&pcp_batch_high_lock);
6538 for_each_possible_cpu(cpu)
6539 pageset_set_high_and_batch(zone,
6540 per_cpu_ptr(zone->pageset, cpu));
6541 mutex_unlock(&pcp_batch_high_lock);
6545 void zone_pcp_reset(struct zone *zone)
6547 unsigned long flags;
6549 struct per_cpu_pageset *pset;
6551 /* avoid races with drain_pages() */
6552 local_irq_save(flags);
6553 if (zone->pageset != &boot_pageset) {
6554 for_each_online_cpu(cpu) {
6555 pset = per_cpu_ptr(zone->pageset, cpu);
6556 drain_zonestat(zone, pset);
6558 free_percpu(zone->pageset);
6559 zone->pageset = &boot_pageset;
6561 local_irq_restore(flags);
6564 #ifdef CONFIG_MEMORY_HOTREMOVE
6566 * All pages in the range must be isolated before calling this.
6569 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6573 unsigned int order, i;
6575 unsigned long flags;
6576 /* find the first valid pfn */
6577 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6582 zone = page_zone(pfn_to_page(pfn));
6583 spin_lock_irqsave(&zone->lock, flags);
6585 while (pfn < end_pfn) {
6586 if (!pfn_valid(pfn)) {
6590 page = pfn_to_page(pfn);
6592 * The HWPoisoned page may be not in buddy system, and
6593 * page_count() is not 0.
6595 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6597 SetPageReserved(page);
6601 BUG_ON(page_count(page));
6602 BUG_ON(!PageBuddy(page));
6603 order = page_order(page);
6604 #ifdef CONFIG_DEBUG_VM
6605 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6606 pfn, 1 << order, end_pfn);
6608 list_del(&page->lru);
6609 rmv_page_order(page);
6610 zone->free_area[order].nr_free--;
6611 for (i = 0; i < (1 << order); i++)
6612 SetPageReserved((page+i));
6613 pfn += (1 << order);
6615 spin_unlock_irqrestore(&zone->lock, flags);
6619 #ifdef CONFIG_MEMORY_FAILURE
6620 bool is_free_buddy_page(struct page *page)
6622 struct zone *zone = page_zone(page);
6623 unsigned long pfn = page_to_pfn(page);
6624 unsigned long flags;
6627 spin_lock_irqsave(&zone->lock, flags);
6628 for (order = 0; order < MAX_ORDER; order++) {
6629 struct page *page_head = page - (pfn & ((1 << order) - 1));
6631 if (PageBuddy(page_head) && page_order(page_head) >= order)
6634 spin_unlock_irqrestore(&zone->lock, flags);
6636 return order < MAX_ORDER;