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;
114 unsigned long totalcma_pages __read_mostly;
116 * When calculating the number of globally allowed dirty pages, there
117 * is a certain number of per-zone reserves that should not be
118 * considered dirtyable memory. This is the sum of those reserves
119 * over all existing zones that contribute dirtyable memory.
121 unsigned long dirty_balance_reserve __read_mostly;
123 int percpu_pagelist_fraction;
124 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
126 #ifdef CONFIG_PM_SLEEP
128 * The following functions are used by the suspend/hibernate code to temporarily
129 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
130 * while devices are suspended. To avoid races with the suspend/hibernate code,
131 * they should always be called with pm_mutex held (gfp_allowed_mask also should
132 * only be modified with pm_mutex held, unless the suspend/hibernate code is
133 * guaranteed not to run in parallel with that modification).
136 static gfp_t saved_gfp_mask;
138 void pm_restore_gfp_mask(void)
140 WARN_ON(!mutex_is_locked(&pm_mutex));
141 if (saved_gfp_mask) {
142 gfp_allowed_mask = saved_gfp_mask;
147 void pm_restrict_gfp_mask(void)
149 WARN_ON(!mutex_is_locked(&pm_mutex));
150 WARN_ON(saved_gfp_mask);
151 saved_gfp_mask = gfp_allowed_mask;
152 gfp_allowed_mask &= ~GFP_IOFS;
155 bool pm_suspended_storage(void)
157 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
161 #endif /* CONFIG_PM_SLEEP */
163 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
164 int pageblock_order __read_mostly;
167 static void __free_pages_ok(struct page *page, unsigned int order);
170 * results with 256, 32 in the lowmem_reserve sysctl:
171 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
172 * 1G machine -> (16M dma, 784M normal, 224M high)
173 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
174 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
175 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
177 * TBD: should special case ZONE_DMA32 machines here - in those we normally
178 * don't need any ZONE_NORMAL reservation
180 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
181 #ifdef CONFIG_ZONE_DMA
184 #ifdef CONFIG_ZONE_DMA32
187 #ifdef CONFIG_HIGHMEM
193 EXPORT_SYMBOL(totalram_pages);
195 static char * const zone_names[MAX_NR_ZONES] = {
196 #ifdef CONFIG_ZONE_DMA
199 #ifdef CONFIG_ZONE_DMA32
203 #ifdef CONFIG_HIGHMEM
209 int min_free_kbytes = 1024;
210 int user_min_free_kbytes = -1;
212 static unsigned long __meminitdata nr_kernel_pages;
213 static unsigned long __meminitdata nr_all_pages;
214 static unsigned long __meminitdata dma_reserve;
216 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
217 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
218 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
219 static unsigned long __initdata required_kernelcore;
220 static unsigned long __initdata required_movablecore;
221 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
223 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
225 EXPORT_SYMBOL(movable_zone);
226 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
229 int nr_node_ids __read_mostly = MAX_NUMNODES;
230 int nr_online_nodes __read_mostly = 1;
231 EXPORT_SYMBOL(nr_node_ids);
232 EXPORT_SYMBOL(nr_online_nodes);
235 int page_group_by_mobility_disabled __read_mostly;
237 void set_pageblock_migratetype(struct page *page, int migratetype)
239 if (unlikely(page_group_by_mobility_disabled &&
240 migratetype < MIGRATE_PCPTYPES))
241 migratetype = MIGRATE_UNMOVABLE;
243 set_pageblock_flags_group(page, (unsigned long)migratetype,
244 PB_migrate, PB_migrate_end);
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
252 unsigned long pfn = page_to_pfn(page);
253 unsigned long sp, start_pfn;
256 seq = zone_span_seqbegin(zone);
257 start_pfn = zone->zone_start_pfn;
258 sp = zone->spanned_pages;
259 if (!zone_spans_pfn(zone, pfn))
261 } while (zone_span_seqretry(zone, seq));
264 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
265 pfn, zone_to_nid(zone), zone->name,
266 start_pfn, start_pfn + sp);
271 static int page_is_consistent(struct zone *zone, struct page *page)
273 if (!pfn_valid_within(page_to_pfn(page)))
275 if (zone != page_zone(page))
281 * Temporary debugging check for pages not lying within a given zone.
283 static int bad_range(struct zone *zone, struct page *page)
285 if (page_outside_zone_boundaries(zone, page))
287 if (!page_is_consistent(zone, page))
293 static inline int bad_range(struct zone *zone, struct page *page)
299 static void bad_page(struct page *page, const char *reason,
300 unsigned long bad_flags)
302 static unsigned long resume;
303 static unsigned long nr_shown;
304 static unsigned long nr_unshown;
306 /* Don't complain about poisoned pages */
307 if (PageHWPoison(page)) {
308 page_mapcount_reset(page); /* remove PageBuddy */
313 * Allow a burst of 60 reports, then keep quiet for that minute;
314 * or allow a steady drip of one report per second.
316 if (nr_shown == 60) {
317 if (time_before(jiffies, resume)) {
323 "BUG: Bad page state: %lu messages suppressed\n",
330 resume = jiffies + 60 * HZ;
332 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
333 current->comm, page_to_pfn(page));
334 dump_page_badflags(page, reason, bad_flags);
339 /* Leave bad fields for debug, except PageBuddy could make trouble */
340 page_mapcount_reset(page); /* remove PageBuddy */
341 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
345 * Higher-order pages are called "compound pages". They are structured thusly:
347 * The first PAGE_SIZE page is called the "head page".
349 * The remaining PAGE_SIZE pages are called "tail pages".
351 * All pages have PG_compound set. All tail pages have their ->first_page
352 * pointing at the head page.
354 * The first tail page's ->lru.next holds the address of the compound page's
355 * put_page() function. Its ->lru.prev holds the order of allocation.
356 * This usage means that zero-order pages may not be compound.
359 static void free_compound_page(struct page *page)
361 __free_pages_ok(page, compound_order(page));
364 void prep_compound_page(struct page *page, unsigned long order)
367 int nr_pages = 1 << order;
369 set_compound_page_dtor(page, free_compound_page);
370 set_compound_order(page, order);
372 for (i = 1; i < nr_pages; i++) {
373 struct page *p = page + i;
374 set_page_count(p, 0);
375 p->first_page = page;
376 /* Make sure p->first_page is always valid for PageTail() */
382 static inline void prep_zero_page(struct page *page, unsigned int order,
388 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
389 * and __GFP_HIGHMEM from hard or soft interrupt context.
391 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
392 for (i = 0; i < (1 << order); i++)
393 clear_highpage(page + i);
396 #ifdef CONFIG_DEBUG_PAGEALLOC
397 unsigned int _debug_guardpage_minorder;
398 bool _debug_pagealloc_enabled __read_mostly;
399 bool _debug_guardpage_enabled __read_mostly;
401 static int __init early_debug_pagealloc(char *buf)
406 if (strcmp(buf, "on") == 0)
407 _debug_pagealloc_enabled = true;
411 early_param("debug_pagealloc", early_debug_pagealloc);
413 static bool need_debug_guardpage(void)
415 /* If we don't use debug_pagealloc, we don't need guard page */
416 if (!debug_pagealloc_enabled())
422 static void init_debug_guardpage(void)
424 if (!debug_pagealloc_enabled())
427 _debug_guardpage_enabled = true;
430 struct page_ext_operations debug_guardpage_ops = {
431 .need = need_debug_guardpage,
432 .init = init_debug_guardpage,
435 static int __init debug_guardpage_minorder_setup(char *buf)
439 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
440 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
443 _debug_guardpage_minorder = res;
444 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
447 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
449 static inline void set_page_guard(struct zone *zone, struct page *page,
450 unsigned int order, int migratetype)
452 struct page_ext *page_ext;
454 if (!debug_guardpage_enabled())
457 page_ext = lookup_page_ext(page);
458 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
460 INIT_LIST_HEAD(&page->lru);
461 set_page_private(page, order);
462 /* Guard pages are not available for any usage */
463 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
466 static inline void clear_page_guard(struct zone *zone, struct page *page,
467 unsigned int order, int migratetype)
469 struct page_ext *page_ext;
471 if (!debug_guardpage_enabled())
474 page_ext = lookup_page_ext(page);
475 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
477 set_page_private(page, 0);
478 if (!is_migrate_isolate(migratetype))
479 __mod_zone_freepage_state(zone, (1 << order), migratetype);
482 struct page_ext_operations debug_guardpage_ops = { NULL, };
483 static inline void set_page_guard(struct zone *zone, struct page *page,
484 unsigned int order, int migratetype) {}
485 static inline void clear_page_guard(struct zone *zone, struct page *page,
486 unsigned int order, int migratetype) {}
489 static inline void set_page_order(struct page *page, unsigned int order)
491 set_page_private(page, order);
492 __SetPageBuddy(page);
495 static inline void rmv_page_order(struct page *page)
497 __ClearPageBuddy(page);
498 set_page_private(page, 0);
502 * This function checks whether a page is free && is the buddy
503 * we can do coalesce a page and its buddy if
504 * (a) the buddy is not in a hole &&
505 * (b) the buddy is in the buddy system &&
506 * (c) a page and its buddy have the same order &&
507 * (d) a page and its buddy are in the same zone.
509 * For recording whether a page is in the buddy system, we set ->_mapcount
510 * PAGE_BUDDY_MAPCOUNT_VALUE.
511 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
512 * serialized by zone->lock.
514 * For recording page's order, we use page_private(page).
516 static inline int page_is_buddy(struct page *page, struct page *buddy,
519 if (!pfn_valid_within(page_to_pfn(buddy)))
522 if (page_is_guard(buddy) && page_order(buddy) == order) {
523 if (page_zone_id(page) != page_zone_id(buddy))
526 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
531 if (PageBuddy(buddy) && page_order(buddy) == order) {
533 * zone check is done late to avoid uselessly
534 * calculating zone/node ids for pages that could
537 if (page_zone_id(page) != page_zone_id(buddy))
540 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
548 * Freeing function for a buddy system allocator.
550 * The concept of a buddy system is to maintain direct-mapped table
551 * (containing bit values) for memory blocks of various "orders".
552 * The bottom level table contains the map for the smallest allocatable
553 * units of memory (here, pages), and each level above it describes
554 * pairs of units from the levels below, hence, "buddies".
555 * At a high level, all that happens here is marking the table entry
556 * at the bottom level available, and propagating the changes upward
557 * as necessary, plus some accounting needed to play nicely with other
558 * parts of the VM system.
559 * At each level, we keep a list of pages, which are heads of continuous
560 * free pages of length of (1 << order) and marked with _mapcount
561 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
563 * So when we are allocating or freeing one, we can derive the state of the
564 * other. That is, if we allocate a small block, and both were
565 * free, the remainder of the region must be split into blocks.
566 * If a block is freed, and its buddy is also free, then this
567 * triggers coalescing into a block of larger size.
572 static inline void __free_one_page(struct page *page,
574 struct zone *zone, unsigned int order,
577 unsigned long page_idx;
578 unsigned long combined_idx;
579 unsigned long uninitialized_var(buddy_idx);
581 int max_order = MAX_ORDER;
583 VM_BUG_ON(!zone_is_initialized(zone));
584 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
586 VM_BUG_ON(migratetype == -1);
587 if (is_migrate_isolate(migratetype)) {
589 * We restrict max order of merging to prevent merge
590 * between freepages on isolate pageblock and normal
591 * pageblock. Without this, pageblock isolation
592 * could cause incorrect freepage accounting.
594 max_order = min(MAX_ORDER, pageblock_order + 1);
596 __mod_zone_freepage_state(zone, 1 << order, migratetype);
599 page_idx = pfn & ((1 << max_order) - 1);
601 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
602 VM_BUG_ON_PAGE(bad_range(zone, page), page);
604 while (order < max_order - 1) {
605 buddy_idx = __find_buddy_index(page_idx, order);
606 buddy = page + (buddy_idx - page_idx);
607 if (!page_is_buddy(page, buddy, order))
610 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
611 * merge with it and move up one order.
613 if (page_is_guard(buddy)) {
614 clear_page_guard(zone, buddy, order, migratetype);
616 list_del(&buddy->lru);
617 zone->free_area[order].nr_free--;
618 rmv_page_order(buddy);
620 combined_idx = buddy_idx & page_idx;
621 page = page + (combined_idx - page_idx);
622 page_idx = combined_idx;
625 set_page_order(page, order);
628 * If this is not the largest possible page, check if the buddy
629 * of the next-highest order is free. If it is, it's possible
630 * that pages are being freed that will coalesce soon. In case,
631 * that is happening, add the free page to the tail of the list
632 * so it's less likely to be used soon and more likely to be merged
633 * as a higher order page
635 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
636 struct page *higher_page, *higher_buddy;
637 combined_idx = buddy_idx & page_idx;
638 higher_page = page + (combined_idx - page_idx);
639 buddy_idx = __find_buddy_index(combined_idx, order + 1);
640 higher_buddy = higher_page + (buddy_idx - combined_idx);
641 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
642 list_add_tail(&page->lru,
643 &zone->free_area[order].free_list[migratetype]);
648 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
650 zone->free_area[order].nr_free++;
653 static inline int free_pages_check(struct page *page)
655 const char *bad_reason = NULL;
656 unsigned long bad_flags = 0;
658 if (unlikely(page_mapcount(page)))
659 bad_reason = "nonzero mapcount";
660 if (unlikely(page->mapping != NULL))
661 bad_reason = "non-NULL mapping";
662 if (unlikely(atomic_read(&page->_count) != 0))
663 bad_reason = "nonzero _count";
664 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
665 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
666 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
669 if (unlikely(page->mem_cgroup))
670 bad_reason = "page still charged to cgroup";
672 if (unlikely(bad_reason)) {
673 bad_page(page, bad_reason, bad_flags);
676 page_cpupid_reset_last(page);
677 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
678 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
683 * Frees a number of pages from the PCP lists
684 * Assumes all pages on list are in same zone, and of same order.
685 * count is the number of pages to free.
687 * If the zone was previously in an "all pages pinned" state then look to
688 * see if this freeing clears that state.
690 * And clear the zone's pages_scanned counter, to hold off the "all pages are
691 * pinned" detection logic.
693 static void free_pcppages_bulk(struct zone *zone, int count,
694 struct per_cpu_pages *pcp)
699 unsigned long nr_scanned;
701 spin_lock(&zone->lock);
702 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
704 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
708 struct list_head *list;
711 * Remove pages from lists in a round-robin fashion. A
712 * batch_free count is maintained that is incremented when an
713 * empty list is encountered. This is so more pages are freed
714 * off fuller lists instead of spinning excessively around empty
719 if (++migratetype == MIGRATE_PCPTYPES)
721 list = &pcp->lists[migratetype];
722 } while (list_empty(list));
724 /* This is the only non-empty list. Free them all. */
725 if (batch_free == MIGRATE_PCPTYPES)
726 batch_free = to_free;
729 int mt; /* migratetype of the to-be-freed page */
731 page = list_entry(list->prev, struct page, lru);
732 /* must delete as __free_one_page list manipulates */
733 list_del(&page->lru);
734 mt = get_freepage_migratetype(page);
735 if (unlikely(has_isolate_pageblock(zone)))
736 mt = get_pageblock_migratetype(page);
738 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
739 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
740 trace_mm_page_pcpu_drain(page, 0, mt);
741 } while (--to_free && --batch_free && !list_empty(list));
743 spin_unlock(&zone->lock);
746 static void free_one_page(struct zone *zone,
747 struct page *page, unsigned long pfn,
751 unsigned long nr_scanned;
752 spin_lock(&zone->lock);
753 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
755 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
757 if (unlikely(has_isolate_pageblock(zone) ||
758 is_migrate_isolate(migratetype))) {
759 migratetype = get_pfnblock_migratetype(page, pfn);
761 __free_one_page(page, pfn, zone, order, migratetype);
762 spin_unlock(&zone->lock);
765 static int free_tail_pages_check(struct page *head_page, struct page *page)
767 if (!IS_ENABLED(CONFIG_DEBUG_VM))
769 if (unlikely(!PageTail(page))) {
770 bad_page(page, "PageTail not set", 0);
773 if (unlikely(page->first_page != head_page)) {
774 bad_page(page, "first_page not consistent", 0);
780 static bool free_pages_prepare(struct page *page, unsigned int order)
782 bool compound = PageCompound(page);
785 VM_BUG_ON_PAGE(PageTail(page), page);
786 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
788 trace_mm_page_free(page, order);
789 kmemcheck_free_shadow(page, order);
792 page->mapping = NULL;
793 bad += free_pages_check(page);
794 for (i = 1; i < (1 << order); i++) {
796 bad += free_tail_pages_check(page, page + i);
797 bad += free_pages_check(page + i);
802 reset_page_owner(page, order);
804 if (!PageHighMem(page)) {
805 debug_check_no_locks_freed(page_address(page),
807 debug_check_no_obj_freed(page_address(page),
810 arch_free_page(page, order);
811 kernel_map_pages(page, 1 << order, 0);
816 static void __free_pages_ok(struct page *page, unsigned int order)
820 unsigned long pfn = page_to_pfn(page);
822 if (!free_pages_prepare(page, order))
825 migratetype = get_pfnblock_migratetype(page, pfn);
826 local_irq_save(flags);
827 __count_vm_events(PGFREE, 1 << order);
828 set_freepage_migratetype(page, migratetype);
829 free_one_page(page_zone(page), page, pfn, order, migratetype);
830 local_irq_restore(flags);
833 void __init __free_pages_bootmem(struct page *page, unsigned int order)
835 unsigned int nr_pages = 1 << order;
836 struct page *p = page;
840 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
842 __ClearPageReserved(p);
843 set_page_count(p, 0);
845 __ClearPageReserved(p);
846 set_page_count(p, 0);
848 page_zone(page)->managed_pages += nr_pages;
849 set_page_refcounted(page);
850 __free_pages(page, order);
854 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
855 void __init init_cma_reserved_pageblock(struct page *page)
857 unsigned i = pageblock_nr_pages;
858 struct page *p = page;
861 __ClearPageReserved(p);
862 set_page_count(p, 0);
865 set_pageblock_migratetype(page, MIGRATE_CMA);
867 if (pageblock_order >= MAX_ORDER) {
868 i = pageblock_nr_pages;
871 set_page_refcounted(p);
872 __free_pages(p, MAX_ORDER - 1);
873 p += MAX_ORDER_NR_PAGES;
874 } while (i -= MAX_ORDER_NR_PAGES);
876 set_page_refcounted(page);
877 __free_pages(page, pageblock_order);
880 adjust_managed_page_count(page, pageblock_nr_pages);
885 * The order of subdivision here is critical for the IO subsystem.
886 * Please do not alter this order without good reasons and regression
887 * testing. Specifically, as large blocks of memory are subdivided,
888 * the order in which smaller blocks are delivered depends on the order
889 * they're subdivided in this function. This is the primary factor
890 * influencing the order in which pages are delivered to the IO
891 * subsystem according to empirical testing, and this is also justified
892 * by considering the behavior of a buddy system containing a single
893 * large block of memory acted on by a series of small allocations.
894 * This behavior is a critical factor in sglist merging's success.
898 static inline void expand(struct zone *zone, struct page *page,
899 int low, int high, struct free_area *area,
902 unsigned long size = 1 << high;
908 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
910 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
911 debug_guardpage_enabled() &&
912 high < debug_guardpage_minorder()) {
914 * Mark as guard pages (or page), that will allow to
915 * merge back to allocator when buddy will be freed.
916 * Corresponding page table entries will not be touched,
917 * pages will stay not present in virtual address space
919 set_page_guard(zone, &page[size], high, migratetype);
922 list_add(&page[size].lru, &area->free_list[migratetype]);
924 set_page_order(&page[size], high);
929 * This page is about to be returned from the page allocator
931 static inline int check_new_page(struct page *page)
933 const char *bad_reason = NULL;
934 unsigned long bad_flags = 0;
936 if (unlikely(page_mapcount(page)))
937 bad_reason = "nonzero mapcount";
938 if (unlikely(page->mapping != NULL))
939 bad_reason = "non-NULL mapping";
940 if (unlikely(atomic_read(&page->_count) != 0))
941 bad_reason = "nonzero _count";
942 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
943 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
944 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
947 if (unlikely(page->mem_cgroup))
948 bad_reason = "page still charged to cgroup";
950 if (unlikely(bad_reason)) {
951 bad_page(page, bad_reason, bad_flags);
957 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
962 for (i = 0; i < (1 << order); i++) {
963 struct page *p = page + i;
964 if (unlikely(check_new_page(p)))
968 set_page_private(page, 0);
969 set_page_refcounted(page);
971 arch_alloc_page(page, order);
972 kernel_map_pages(page, 1 << order, 1);
974 if (gfp_flags & __GFP_ZERO)
975 prep_zero_page(page, order, gfp_flags);
977 if (order && (gfp_flags & __GFP_COMP))
978 prep_compound_page(page, order);
980 set_page_owner(page, order, gfp_flags);
983 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
984 * allocate the page. The expectation is that the caller is taking
985 * steps that will free more memory. The caller should avoid the page
986 * being used for !PFMEMALLOC purposes.
988 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
994 * Go through the free lists for the given migratetype and remove
995 * the smallest available page from the freelists
998 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1001 unsigned int current_order;
1002 struct free_area *area;
1005 /* Find a page of the appropriate size in the preferred list */
1006 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1007 area = &(zone->free_area[current_order]);
1008 if (list_empty(&area->free_list[migratetype]))
1011 page = list_entry(area->free_list[migratetype].next,
1013 list_del(&page->lru);
1014 rmv_page_order(page);
1016 expand(zone, page, order, current_order, area, migratetype);
1017 set_freepage_migratetype(page, migratetype);
1026 * This array describes the order lists are fallen back to when
1027 * the free lists for the desirable migrate type are depleted
1029 static int fallbacks[MIGRATE_TYPES][4] = {
1030 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1031 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1033 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1034 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1036 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1038 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1039 #ifdef CONFIG_MEMORY_ISOLATION
1040 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1045 * Move the free pages in a range to the free lists of the requested type.
1046 * Note that start_page and end_pages are not aligned on a pageblock
1047 * boundary. If alignment is required, use move_freepages_block()
1049 int move_freepages(struct zone *zone,
1050 struct page *start_page, struct page *end_page,
1054 unsigned long order;
1055 int pages_moved = 0;
1057 #ifndef CONFIG_HOLES_IN_ZONE
1059 * page_zone is not safe to call in this context when
1060 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1061 * anyway as we check zone boundaries in move_freepages_block().
1062 * Remove at a later date when no bug reports exist related to
1063 * grouping pages by mobility
1065 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1068 for (page = start_page; page <= end_page;) {
1069 /* Make sure we are not inadvertently changing nodes */
1070 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1072 if (!pfn_valid_within(page_to_pfn(page))) {
1077 if (!PageBuddy(page)) {
1082 order = page_order(page);
1083 list_move(&page->lru,
1084 &zone->free_area[order].free_list[migratetype]);
1085 set_freepage_migratetype(page, migratetype);
1087 pages_moved += 1 << order;
1093 int move_freepages_block(struct zone *zone, struct page *page,
1096 unsigned long start_pfn, end_pfn;
1097 struct page *start_page, *end_page;
1099 start_pfn = page_to_pfn(page);
1100 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1101 start_page = pfn_to_page(start_pfn);
1102 end_page = start_page + pageblock_nr_pages - 1;
1103 end_pfn = start_pfn + pageblock_nr_pages - 1;
1105 /* Do not cross zone boundaries */
1106 if (!zone_spans_pfn(zone, start_pfn))
1108 if (!zone_spans_pfn(zone, end_pfn))
1111 return move_freepages(zone, start_page, end_page, migratetype);
1114 static void change_pageblock_range(struct page *pageblock_page,
1115 int start_order, int migratetype)
1117 int nr_pageblocks = 1 << (start_order - pageblock_order);
1119 while (nr_pageblocks--) {
1120 set_pageblock_migratetype(pageblock_page, migratetype);
1121 pageblock_page += pageblock_nr_pages;
1126 * When we are falling back to another migratetype during allocation, try to
1127 * steal extra free pages from the same pageblocks to satisfy further
1128 * allocations, instead of polluting multiple pageblocks.
1130 * If we are stealing a relatively large buddy page, it is likely there will
1131 * be more free pages in the pageblock, so try to steal them all. For
1132 * reclaimable and unmovable allocations, we steal regardless of page size,
1133 * as fragmentation caused by those allocations polluting movable pageblocks
1134 * is worse than movable allocations stealing from unmovable and reclaimable
1137 * If we claim more than half of the pageblock, change pageblock's migratetype
1140 static void try_to_steal_freepages(struct zone *zone, struct page *page,
1141 int start_type, int fallback_type)
1143 int current_order = page_order(page);
1145 /* Take ownership for orders >= pageblock_order */
1146 if (current_order >= pageblock_order) {
1147 change_pageblock_range(page, current_order, start_type);
1151 if (current_order >= pageblock_order / 2 ||
1152 start_type == MIGRATE_RECLAIMABLE ||
1153 start_type == MIGRATE_UNMOVABLE ||
1154 page_group_by_mobility_disabled) {
1157 pages = move_freepages_block(zone, page, start_type);
1159 /* Claim the whole block if over half of it is free */
1160 if (pages >= (1 << (pageblock_order-1)) ||
1161 page_group_by_mobility_disabled)
1162 set_pageblock_migratetype(page, start_type);
1166 /* Remove an element from the buddy allocator from the fallback list */
1167 static inline struct page *
1168 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1170 struct free_area *area;
1171 unsigned int current_order;
1174 /* Find the largest possible block of pages in the other list */
1175 for (current_order = MAX_ORDER-1;
1176 current_order >= order && current_order <= MAX_ORDER-1;
1180 int migratetype = fallbacks[start_migratetype][i];
1181 int buddy_type = start_migratetype;
1183 /* MIGRATE_RESERVE handled later if necessary */
1184 if (migratetype == MIGRATE_RESERVE)
1187 area = &(zone->free_area[current_order]);
1188 if (list_empty(&area->free_list[migratetype]))
1191 page = list_entry(area->free_list[migratetype].next,
1195 if (!is_migrate_cma(migratetype)) {
1196 try_to_steal_freepages(zone, page,
1201 * When borrowing from MIGRATE_CMA, we need to
1202 * release the excess buddy pages to CMA
1203 * itself, and we do not try to steal extra
1206 buddy_type = migratetype;
1209 /* Remove the page from the freelists */
1210 list_del(&page->lru);
1211 rmv_page_order(page);
1213 expand(zone, page, order, current_order, area,
1217 * The freepage_migratetype may differ from pageblock's
1218 * migratetype depending on the decisions in
1219 * try_to_steal_freepages(). This is OK as long as it
1220 * does not differ for MIGRATE_CMA pageblocks. For CMA
1221 * we need to make sure unallocated pages flushed from
1222 * pcp lists are returned to the correct freelist.
1224 set_freepage_migratetype(page, buddy_type);
1226 trace_mm_page_alloc_extfrag(page, order, current_order,
1227 start_migratetype, migratetype);
1237 * Do the hard work of removing an element from the buddy allocator.
1238 * Call me with the zone->lock already held.
1240 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1246 page = __rmqueue_smallest(zone, order, migratetype);
1248 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1249 page = __rmqueue_fallback(zone, order, migratetype);
1252 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1253 * is used because __rmqueue_smallest is an inline function
1254 * and we want just one call site
1257 migratetype = MIGRATE_RESERVE;
1262 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1267 * Obtain a specified number of elements from the buddy allocator, all under
1268 * a single hold of the lock, for efficiency. Add them to the supplied list.
1269 * Returns the number of new pages which were placed at *list.
1271 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1272 unsigned long count, struct list_head *list,
1273 int migratetype, bool cold)
1277 spin_lock(&zone->lock);
1278 for (i = 0; i < count; ++i) {
1279 struct page *page = __rmqueue(zone, order, migratetype);
1280 if (unlikely(page == NULL))
1284 * Split buddy pages returned by expand() are received here
1285 * in physical page order. The page is added to the callers and
1286 * list and the list head then moves forward. From the callers
1287 * perspective, the linked list is ordered by page number in
1288 * some conditions. This is useful for IO devices that can
1289 * merge IO requests if the physical pages are ordered
1293 list_add(&page->lru, list);
1295 list_add_tail(&page->lru, list);
1297 if (is_migrate_cma(get_freepage_migratetype(page)))
1298 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1301 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1302 spin_unlock(&zone->lock);
1308 * Called from the vmstat counter updater to drain pagesets of this
1309 * currently executing processor on remote nodes after they have
1312 * Note that this function must be called with the thread pinned to
1313 * a single processor.
1315 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1317 unsigned long flags;
1318 int to_drain, batch;
1320 local_irq_save(flags);
1321 batch = ACCESS_ONCE(pcp->batch);
1322 to_drain = min(pcp->count, batch);
1324 free_pcppages_bulk(zone, to_drain, pcp);
1325 pcp->count -= to_drain;
1327 local_irq_restore(flags);
1332 * Drain pcplists of the indicated processor and zone.
1334 * The processor must either be the current processor and the
1335 * thread pinned to the current processor or a processor that
1338 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1340 unsigned long flags;
1341 struct per_cpu_pageset *pset;
1342 struct per_cpu_pages *pcp;
1344 local_irq_save(flags);
1345 pset = per_cpu_ptr(zone->pageset, cpu);
1349 free_pcppages_bulk(zone, pcp->count, pcp);
1352 local_irq_restore(flags);
1356 * Drain pcplists of all zones on the indicated processor.
1358 * The processor must either be the current processor and the
1359 * thread pinned to the current processor or a processor that
1362 static void drain_pages(unsigned int cpu)
1366 for_each_populated_zone(zone) {
1367 drain_pages_zone(cpu, zone);
1372 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1374 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1375 * the single zone's pages.
1377 void drain_local_pages(struct zone *zone)
1379 int cpu = smp_processor_id();
1382 drain_pages_zone(cpu, zone);
1388 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1390 * When zone parameter is non-NULL, spill just the single zone's pages.
1392 * Note that this code is protected against sending an IPI to an offline
1393 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1394 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1395 * nothing keeps CPUs from showing up after we populated the cpumask and
1396 * before the call to on_each_cpu_mask().
1398 void drain_all_pages(struct zone *zone)
1403 * Allocate in the BSS so we wont require allocation in
1404 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1406 static cpumask_t cpus_with_pcps;
1409 * We don't care about racing with CPU hotplug event
1410 * as offline notification will cause the notified
1411 * cpu to drain that CPU pcps and on_each_cpu_mask
1412 * disables preemption as part of its processing
1414 for_each_online_cpu(cpu) {
1415 struct per_cpu_pageset *pcp;
1417 bool has_pcps = false;
1420 pcp = per_cpu_ptr(zone->pageset, cpu);
1424 for_each_populated_zone(z) {
1425 pcp = per_cpu_ptr(z->pageset, cpu);
1426 if (pcp->pcp.count) {
1434 cpumask_set_cpu(cpu, &cpus_with_pcps);
1436 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1438 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1442 #ifdef CONFIG_HIBERNATION
1444 void mark_free_pages(struct zone *zone)
1446 unsigned long pfn, max_zone_pfn;
1447 unsigned long flags;
1448 unsigned int order, t;
1449 struct list_head *curr;
1451 if (zone_is_empty(zone))
1454 spin_lock_irqsave(&zone->lock, flags);
1456 max_zone_pfn = zone_end_pfn(zone);
1457 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1458 if (pfn_valid(pfn)) {
1459 struct page *page = pfn_to_page(pfn);
1461 if (!swsusp_page_is_forbidden(page))
1462 swsusp_unset_page_free(page);
1465 for_each_migratetype_order(order, t) {
1466 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1469 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1470 for (i = 0; i < (1UL << order); i++)
1471 swsusp_set_page_free(pfn_to_page(pfn + i));
1474 spin_unlock_irqrestore(&zone->lock, flags);
1476 #endif /* CONFIG_PM */
1479 * Free a 0-order page
1480 * cold == true ? free a cold page : free a hot page
1482 void free_hot_cold_page(struct page *page, bool cold)
1484 struct zone *zone = page_zone(page);
1485 struct per_cpu_pages *pcp;
1486 unsigned long flags;
1487 unsigned long pfn = page_to_pfn(page);
1490 if (!free_pages_prepare(page, 0))
1493 migratetype = get_pfnblock_migratetype(page, pfn);
1494 set_freepage_migratetype(page, migratetype);
1495 local_irq_save(flags);
1496 __count_vm_event(PGFREE);
1499 * We only track unmovable, reclaimable and movable on pcp lists.
1500 * Free ISOLATE pages back to the allocator because they are being
1501 * offlined but treat RESERVE as movable pages so we can get those
1502 * areas back if necessary. Otherwise, we may have to free
1503 * excessively into the page allocator
1505 if (migratetype >= MIGRATE_PCPTYPES) {
1506 if (unlikely(is_migrate_isolate(migratetype))) {
1507 free_one_page(zone, page, pfn, 0, migratetype);
1510 migratetype = MIGRATE_MOVABLE;
1513 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1515 list_add(&page->lru, &pcp->lists[migratetype]);
1517 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1519 if (pcp->count >= pcp->high) {
1520 unsigned long batch = ACCESS_ONCE(pcp->batch);
1521 free_pcppages_bulk(zone, batch, pcp);
1522 pcp->count -= batch;
1526 local_irq_restore(flags);
1530 * Free a list of 0-order pages
1532 void free_hot_cold_page_list(struct list_head *list, bool cold)
1534 struct page *page, *next;
1536 list_for_each_entry_safe(page, next, list, lru) {
1537 trace_mm_page_free_batched(page, cold);
1538 free_hot_cold_page(page, cold);
1543 * split_page takes a non-compound higher-order page, and splits it into
1544 * n (1<<order) sub-pages: page[0..n]
1545 * Each sub-page must be freed individually.
1547 * Note: this is probably too low level an operation for use in drivers.
1548 * Please consult with lkml before using this in your driver.
1550 void split_page(struct page *page, unsigned int order)
1554 VM_BUG_ON_PAGE(PageCompound(page), page);
1555 VM_BUG_ON_PAGE(!page_count(page), page);
1557 #ifdef CONFIG_KMEMCHECK
1559 * Split shadow pages too, because free(page[0]) would
1560 * otherwise free the whole shadow.
1562 if (kmemcheck_page_is_tracked(page))
1563 split_page(virt_to_page(page[0].shadow), order);
1566 set_page_owner(page, 0, 0);
1567 for (i = 1; i < (1 << order); i++) {
1568 set_page_refcounted(page + i);
1569 set_page_owner(page + i, 0, 0);
1572 EXPORT_SYMBOL_GPL(split_page);
1574 int __isolate_free_page(struct page *page, unsigned int order)
1576 unsigned long watermark;
1580 BUG_ON(!PageBuddy(page));
1582 zone = page_zone(page);
1583 mt = get_pageblock_migratetype(page);
1585 if (!is_migrate_isolate(mt)) {
1586 /* Obey watermarks as if the page was being allocated */
1587 watermark = low_wmark_pages(zone) + (1 << order);
1588 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1591 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1594 /* Remove page from free list */
1595 list_del(&page->lru);
1596 zone->free_area[order].nr_free--;
1597 rmv_page_order(page);
1599 /* Set the pageblock if the isolated page is at least a pageblock */
1600 if (order >= pageblock_order - 1) {
1601 struct page *endpage = page + (1 << order) - 1;
1602 for (; page < endpage; page += pageblock_nr_pages) {
1603 int mt = get_pageblock_migratetype(page);
1604 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1605 set_pageblock_migratetype(page,
1610 set_page_owner(page, order, 0);
1611 return 1UL << order;
1615 * Similar to split_page except the page is already free. As this is only
1616 * being used for migration, the migratetype of the block also changes.
1617 * As this is called with interrupts disabled, the caller is responsible
1618 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1621 * Note: this is probably too low level an operation for use in drivers.
1622 * Please consult with lkml before using this in your driver.
1624 int split_free_page(struct page *page)
1629 order = page_order(page);
1631 nr_pages = __isolate_free_page(page, order);
1635 /* Split into individual pages */
1636 set_page_refcounted(page);
1637 split_page(page, order);
1642 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1645 struct page *buffered_rmqueue(struct zone *preferred_zone,
1646 struct zone *zone, unsigned int order,
1647 gfp_t gfp_flags, int migratetype)
1649 unsigned long flags;
1651 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1653 if (likely(order == 0)) {
1654 struct per_cpu_pages *pcp;
1655 struct list_head *list;
1657 local_irq_save(flags);
1658 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1659 list = &pcp->lists[migratetype];
1660 if (list_empty(list)) {
1661 pcp->count += rmqueue_bulk(zone, 0,
1664 if (unlikely(list_empty(list)))
1669 page = list_entry(list->prev, struct page, lru);
1671 page = list_entry(list->next, struct page, lru);
1673 list_del(&page->lru);
1676 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1678 * __GFP_NOFAIL is not to be used in new code.
1680 * All __GFP_NOFAIL callers should be fixed so that they
1681 * properly detect and handle allocation failures.
1683 * We most definitely don't want callers attempting to
1684 * allocate greater than order-1 page units with
1687 WARN_ON_ONCE(order > 1);
1689 spin_lock_irqsave(&zone->lock, flags);
1690 page = __rmqueue(zone, order, migratetype);
1691 spin_unlock(&zone->lock);
1694 __mod_zone_freepage_state(zone, -(1 << order),
1695 get_freepage_migratetype(page));
1698 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1699 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1700 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1701 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1703 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1704 zone_statistics(preferred_zone, zone, gfp_flags);
1705 local_irq_restore(flags);
1707 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1711 local_irq_restore(flags);
1715 #ifdef CONFIG_FAIL_PAGE_ALLOC
1718 struct fault_attr attr;
1720 u32 ignore_gfp_highmem;
1721 u32 ignore_gfp_wait;
1723 } fail_page_alloc = {
1724 .attr = FAULT_ATTR_INITIALIZER,
1725 .ignore_gfp_wait = 1,
1726 .ignore_gfp_highmem = 1,
1730 static int __init setup_fail_page_alloc(char *str)
1732 return setup_fault_attr(&fail_page_alloc.attr, str);
1734 __setup("fail_page_alloc=", setup_fail_page_alloc);
1736 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1738 if (order < fail_page_alloc.min_order)
1740 if (gfp_mask & __GFP_NOFAIL)
1742 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1744 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1747 return should_fail(&fail_page_alloc.attr, 1 << order);
1750 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1752 static int __init fail_page_alloc_debugfs(void)
1754 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1757 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1758 &fail_page_alloc.attr);
1760 return PTR_ERR(dir);
1762 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1763 &fail_page_alloc.ignore_gfp_wait))
1765 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1766 &fail_page_alloc.ignore_gfp_highmem))
1768 if (!debugfs_create_u32("min-order", mode, dir,
1769 &fail_page_alloc.min_order))
1774 debugfs_remove_recursive(dir);
1779 late_initcall(fail_page_alloc_debugfs);
1781 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1783 #else /* CONFIG_FAIL_PAGE_ALLOC */
1785 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1790 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1793 * Return true if free pages are above 'mark'. This takes into account the order
1794 * of the allocation.
1796 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1797 unsigned long mark, int classzone_idx, int alloc_flags,
1800 /* free_pages may go negative - that's OK */
1805 free_pages -= (1 << order) - 1;
1806 if (alloc_flags & ALLOC_HIGH)
1808 if (alloc_flags & ALLOC_HARDER)
1811 /* If allocation can't use CMA areas don't use free CMA pages */
1812 if (!(alloc_flags & ALLOC_CMA))
1813 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1816 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1818 for (o = 0; o < order; o++) {
1819 /* At the next order, this order's pages become unavailable */
1820 free_pages -= z->free_area[o].nr_free << o;
1822 /* Require fewer higher order pages to be free */
1825 if (free_pages <= min)
1831 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1832 int classzone_idx, int alloc_flags)
1834 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1835 zone_page_state(z, NR_FREE_PAGES));
1838 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1839 unsigned long mark, int classzone_idx, int alloc_flags)
1841 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1843 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1844 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1846 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1852 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1853 * skip over zones that are not allowed by the cpuset, or that have
1854 * been recently (in last second) found to be nearly full. See further
1855 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1856 * that have to skip over a lot of full or unallowed zones.
1858 * If the zonelist cache is present in the passed zonelist, then
1859 * returns a pointer to the allowed node mask (either the current
1860 * tasks mems_allowed, or node_states[N_MEMORY].)
1862 * If the zonelist cache is not available for this zonelist, does
1863 * nothing and returns NULL.
1865 * If the fullzones BITMAP in the zonelist cache is stale (more than
1866 * a second since last zap'd) then we zap it out (clear its bits.)
1868 * We hold off even calling zlc_setup, until after we've checked the
1869 * first zone in the zonelist, on the theory that most allocations will
1870 * be satisfied from that first zone, so best to examine that zone as
1871 * quickly as we can.
1873 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1875 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1876 nodemask_t *allowednodes; /* zonelist_cache approximation */
1878 zlc = zonelist->zlcache_ptr;
1882 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1883 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1884 zlc->last_full_zap = jiffies;
1887 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1888 &cpuset_current_mems_allowed :
1889 &node_states[N_MEMORY];
1890 return allowednodes;
1894 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1895 * if it is worth looking at further for free memory:
1896 * 1) Check that the zone isn't thought to be full (doesn't have its
1897 * bit set in the zonelist_cache fullzones BITMAP).
1898 * 2) Check that the zones node (obtained from the zonelist_cache
1899 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1900 * Return true (non-zero) if zone is worth looking at further, or
1901 * else return false (zero) if it is not.
1903 * This check -ignores- the distinction between various watermarks,
1904 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1905 * found to be full for any variation of these watermarks, it will
1906 * be considered full for up to one second by all requests, unless
1907 * we are so low on memory on all allowed nodes that we are forced
1908 * into the second scan of the zonelist.
1910 * In the second scan we ignore this zonelist cache and exactly
1911 * apply the watermarks to all zones, even it is slower to do so.
1912 * We are low on memory in the second scan, and should leave no stone
1913 * unturned looking for a free page.
1915 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1916 nodemask_t *allowednodes)
1918 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1919 int i; /* index of *z in zonelist zones */
1920 int n; /* node that zone *z is on */
1922 zlc = zonelist->zlcache_ptr;
1926 i = z - zonelist->_zonerefs;
1929 /* This zone is worth trying if it is allowed but not full */
1930 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1934 * Given 'z' scanning a zonelist, set the corresponding bit in
1935 * zlc->fullzones, so that subsequent attempts to allocate a page
1936 * from that zone don't waste time re-examining it.
1938 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1940 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1941 int i; /* index of *z in zonelist zones */
1943 zlc = zonelist->zlcache_ptr;
1947 i = z - zonelist->_zonerefs;
1949 set_bit(i, zlc->fullzones);
1953 * clear all zones full, called after direct reclaim makes progress so that
1954 * a zone that was recently full is not skipped over for up to a second
1956 static void zlc_clear_zones_full(struct zonelist *zonelist)
1958 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1960 zlc = zonelist->zlcache_ptr;
1964 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1967 static bool zone_local(struct zone *local_zone, struct zone *zone)
1969 return local_zone->node == zone->node;
1972 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1974 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1978 #else /* CONFIG_NUMA */
1980 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1985 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1986 nodemask_t *allowednodes)
1991 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1995 static void zlc_clear_zones_full(struct zonelist *zonelist)
1999 static bool zone_local(struct zone *local_zone, struct zone *zone)
2004 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2009 #endif /* CONFIG_NUMA */
2011 static void reset_alloc_batches(struct zone *preferred_zone)
2013 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2016 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2017 high_wmark_pages(zone) - low_wmark_pages(zone) -
2018 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2019 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2020 } while (zone++ != preferred_zone);
2024 * get_page_from_freelist goes through the zonelist trying to allocate
2027 static struct page *
2028 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2029 const struct alloc_context *ac)
2031 struct zonelist *zonelist = ac->zonelist;
2033 struct page *page = NULL;
2035 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2036 int zlc_active = 0; /* set if using zonelist_cache */
2037 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2038 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2039 (gfp_mask & __GFP_WRITE);
2040 int nr_fair_skipped = 0;
2041 bool zonelist_rescan;
2044 zonelist_rescan = false;
2047 * Scan zonelist, looking for a zone with enough free.
2048 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2050 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2054 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2055 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2057 if (cpusets_enabled() &&
2058 (alloc_flags & ALLOC_CPUSET) &&
2059 !cpuset_zone_allowed(zone, gfp_mask))
2062 * Distribute pages in proportion to the individual
2063 * zone size to ensure fair page aging. The zone a
2064 * page was allocated in should have no effect on the
2065 * time the page has in memory before being reclaimed.
2067 if (alloc_flags & ALLOC_FAIR) {
2068 if (!zone_local(ac->preferred_zone, zone))
2070 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2076 * When allocating a page cache page for writing, we
2077 * want to get it from a zone that is within its dirty
2078 * limit, such that no single zone holds more than its
2079 * proportional share of globally allowed dirty pages.
2080 * The dirty limits take into account the zone's
2081 * lowmem reserves and high watermark so that kswapd
2082 * should be able to balance it without having to
2083 * write pages from its LRU list.
2085 * This may look like it could increase pressure on
2086 * lower zones by failing allocations in higher zones
2087 * before they are full. But the pages that do spill
2088 * over are limited as the lower zones are protected
2089 * by this very same mechanism. It should not become
2090 * a practical burden to them.
2092 * XXX: For now, allow allocations to potentially
2093 * exceed the per-zone dirty limit in the slowpath
2094 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2095 * which is important when on a NUMA setup the allowed
2096 * zones are together not big enough to reach the
2097 * global limit. The proper fix for these situations
2098 * will require awareness of zones in the
2099 * dirty-throttling and the flusher threads.
2101 if (consider_zone_dirty && !zone_dirty_ok(zone))
2104 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2105 if (!zone_watermark_ok(zone, order, mark,
2106 ac->classzone_idx, alloc_flags)) {
2109 /* Checked here to keep the fast path fast */
2110 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2111 if (alloc_flags & ALLOC_NO_WATERMARKS)
2114 if (IS_ENABLED(CONFIG_NUMA) &&
2115 !did_zlc_setup && nr_online_nodes > 1) {
2117 * we do zlc_setup if there are multiple nodes
2118 * and before considering the first zone allowed
2121 allowednodes = zlc_setup(zonelist, alloc_flags);
2126 if (zone_reclaim_mode == 0 ||
2127 !zone_allows_reclaim(ac->preferred_zone, zone))
2128 goto this_zone_full;
2131 * As we may have just activated ZLC, check if the first
2132 * eligible zone has failed zone_reclaim recently.
2134 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2135 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2138 ret = zone_reclaim(zone, gfp_mask, order);
2140 case ZONE_RECLAIM_NOSCAN:
2143 case ZONE_RECLAIM_FULL:
2144 /* scanned but unreclaimable */
2147 /* did we reclaim enough */
2148 if (zone_watermark_ok(zone, order, mark,
2149 ac->classzone_idx, alloc_flags))
2153 * Failed to reclaim enough to meet watermark.
2154 * Only mark the zone full if checking the min
2155 * watermark or if we failed to reclaim just
2156 * 1<<order pages or else the page allocator
2157 * fastpath will prematurely mark zones full
2158 * when the watermark is between the low and
2161 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2162 ret == ZONE_RECLAIM_SOME)
2163 goto this_zone_full;
2170 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2171 gfp_mask, ac->migratetype);
2173 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2178 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2179 zlc_mark_zone_full(zonelist, z);
2183 * The first pass makes sure allocations are spread fairly within the
2184 * local node. However, the local node might have free pages left
2185 * after the fairness batches are exhausted, and remote zones haven't
2186 * even been considered yet. Try once more without fairness, and
2187 * include remote zones now, before entering the slowpath and waking
2188 * kswapd: prefer spilling to a remote zone over swapping locally.
2190 if (alloc_flags & ALLOC_FAIR) {
2191 alloc_flags &= ~ALLOC_FAIR;
2192 if (nr_fair_skipped) {
2193 zonelist_rescan = true;
2194 reset_alloc_batches(ac->preferred_zone);
2196 if (nr_online_nodes > 1)
2197 zonelist_rescan = true;
2200 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2201 /* Disable zlc cache for second zonelist scan */
2203 zonelist_rescan = true;
2206 if (zonelist_rescan)
2213 * Large machines with many possible nodes should not always dump per-node
2214 * meminfo in irq context.
2216 static inline bool should_suppress_show_mem(void)
2221 ret = in_interrupt();
2226 static DEFINE_RATELIMIT_STATE(nopage_rs,
2227 DEFAULT_RATELIMIT_INTERVAL,
2228 DEFAULT_RATELIMIT_BURST);
2230 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2232 unsigned int filter = SHOW_MEM_FILTER_NODES;
2234 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2235 debug_guardpage_minorder() > 0)
2239 * This documents exceptions given to allocations in certain
2240 * contexts that are allowed to allocate outside current's set
2243 if (!(gfp_mask & __GFP_NOMEMALLOC))
2244 if (test_thread_flag(TIF_MEMDIE) ||
2245 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2246 filter &= ~SHOW_MEM_FILTER_NODES;
2247 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2248 filter &= ~SHOW_MEM_FILTER_NODES;
2251 struct va_format vaf;
2254 va_start(args, fmt);
2259 pr_warn("%pV", &vaf);
2264 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2265 current->comm, order, gfp_mask);
2268 if (!should_suppress_show_mem())
2273 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2274 unsigned long did_some_progress,
2275 unsigned long pages_reclaimed)
2277 /* Do not loop if specifically requested */
2278 if (gfp_mask & __GFP_NORETRY)
2281 /* Always retry if specifically requested */
2282 if (gfp_mask & __GFP_NOFAIL)
2286 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2287 * making forward progress without invoking OOM. Suspend also disables
2288 * storage devices so kswapd will not help. Bail if we are suspending.
2290 if (!did_some_progress && pm_suspended_storage())
2294 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2295 * means __GFP_NOFAIL, but that may not be true in other
2298 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2302 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2303 * specified, then we retry until we no longer reclaim any pages
2304 * (above), or we've reclaimed an order of pages at least as
2305 * large as the allocation's order. In both cases, if the
2306 * allocation still fails, we stop retrying.
2308 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2314 static inline struct page *
2315 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2316 const struct alloc_context *ac, unsigned long *did_some_progress)
2320 *did_some_progress = 0;
2323 * Acquire the per-zone oom lock for each zone. If that
2324 * fails, somebody else is making progress for us.
2326 if (!oom_zonelist_trylock(ac->zonelist, gfp_mask)) {
2327 *did_some_progress = 1;
2328 schedule_timeout_uninterruptible(1);
2333 * Go through the zonelist yet one more time, keep very high watermark
2334 * here, this is only to catch a parallel oom killing, we must fail if
2335 * we're still under heavy pressure.
2337 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2338 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2342 if (!(gfp_mask & __GFP_NOFAIL)) {
2343 /* Coredumps can quickly deplete all memory reserves */
2344 if (current->flags & PF_DUMPCORE)
2346 /* The OOM killer will not help higher order allocs */
2347 if (order > PAGE_ALLOC_COSTLY_ORDER)
2349 /* The OOM killer does not needlessly kill tasks for lowmem */
2350 if (ac->high_zoneidx < ZONE_NORMAL)
2352 /* The OOM killer does not compensate for light reclaim */
2353 if (!(gfp_mask & __GFP_FS))
2356 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2357 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2358 * The caller should handle page allocation failure by itself if
2359 * it specifies __GFP_THISNODE.
2360 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2362 if (gfp_mask & __GFP_THISNODE)
2365 /* Exhausted what can be done so it's blamo time */
2366 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false))
2367 *did_some_progress = 1;
2369 oom_zonelist_unlock(ac->zonelist, gfp_mask);
2373 #ifdef CONFIG_COMPACTION
2374 /* Try memory compaction for high-order allocations before reclaim */
2375 static struct page *
2376 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2377 int alloc_flags, const struct alloc_context *ac,
2378 enum migrate_mode mode, int *contended_compaction,
2379 bool *deferred_compaction)
2381 unsigned long compact_result;
2387 current->flags |= PF_MEMALLOC;
2388 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2389 mode, contended_compaction);
2390 current->flags &= ~PF_MEMALLOC;
2392 switch (compact_result) {
2393 case COMPACT_DEFERRED:
2394 *deferred_compaction = true;
2396 case COMPACT_SKIPPED:
2403 * At least in one zone compaction wasn't deferred or skipped, so let's
2404 * count a compaction stall
2406 count_vm_event(COMPACTSTALL);
2408 page = get_page_from_freelist(gfp_mask, order,
2409 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2412 struct zone *zone = page_zone(page);
2414 zone->compact_blockskip_flush = false;
2415 compaction_defer_reset(zone, order, true);
2416 count_vm_event(COMPACTSUCCESS);
2421 * It's bad if compaction run occurs and fails. The most likely reason
2422 * is that pages exist, but not enough to satisfy watermarks.
2424 count_vm_event(COMPACTFAIL);
2431 static inline struct page *
2432 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2433 int alloc_flags, const struct alloc_context *ac,
2434 enum migrate_mode mode, int *contended_compaction,
2435 bool *deferred_compaction)
2439 #endif /* CONFIG_COMPACTION */
2441 /* Perform direct synchronous page reclaim */
2443 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2444 const struct alloc_context *ac)
2446 struct reclaim_state reclaim_state;
2451 /* We now go into synchronous reclaim */
2452 cpuset_memory_pressure_bump();
2453 current->flags |= PF_MEMALLOC;
2454 lockdep_set_current_reclaim_state(gfp_mask);
2455 reclaim_state.reclaimed_slab = 0;
2456 current->reclaim_state = &reclaim_state;
2458 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2461 current->reclaim_state = NULL;
2462 lockdep_clear_current_reclaim_state();
2463 current->flags &= ~PF_MEMALLOC;
2470 /* The really slow allocator path where we enter direct reclaim */
2471 static inline struct page *
2472 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2473 int alloc_flags, const struct alloc_context *ac,
2474 unsigned long *did_some_progress)
2476 struct page *page = NULL;
2477 bool drained = false;
2479 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2480 if (unlikely(!(*did_some_progress)))
2483 /* After successful reclaim, reconsider all zones for allocation */
2484 if (IS_ENABLED(CONFIG_NUMA))
2485 zlc_clear_zones_full(ac->zonelist);
2488 page = get_page_from_freelist(gfp_mask, order,
2489 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2492 * If an allocation failed after direct reclaim, it could be because
2493 * pages are pinned on the per-cpu lists. Drain them and try again
2495 if (!page && !drained) {
2496 drain_all_pages(NULL);
2505 * This is called in the allocator slow-path if the allocation request is of
2506 * sufficient urgency to ignore watermarks and take other desperate measures
2508 static inline struct page *
2509 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2510 const struct alloc_context *ac)
2515 page = get_page_from_freelist(gfp_mask, order,
2516 ALLOC_NO_WATERMARKS, ac);
2518 if (!page && gfp_mask & __GFP_NOFAIL)
2519 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2521 } while (!page && (gfp_mask & __GFP_NOFAIL));
2526 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2531 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2532 ac->high_zoneidx, ac->nodemask)
2533 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2537 gfp_to_alloc_flags(gfp_t gfp_mask)
2539 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2540 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2542 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2543 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2546 * The caller may dip into page reserves a bit more if the caller
2547 * cannot run direct reclaim, or if the caller has realtime scheduling
2548 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2549 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2551 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2555 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2556 * if it can't schedule.
2558 if (!(gfp_mask & __GFP_NOMEMALLOC))
2559 alloc_flags |= ALLOC_HARDER;
2561 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2562 * comment for __cpuset_node_allowed().
2564 alloc_flags &= ~ALLOC_CPUSET;
2565 } else if (unlikely(rt_task(current)) && !in_interrupt())
2566 alloc_flags |= ALLOC_HARDER;
2568 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2569 if (gfp_mask & __GFP_MEMALLOC)
2570 alloc_flags |= ALLOC_NO_WATERMARKS;
2571 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2572 alloc_flags |= ALLOC_NO_WATERMARKS;
2573 else if (!in_interrupt() &&
2574 ((current->flags & PF_MEMALLOC) ||
2575 unlikely(test_thread_flag(TIF_MEMDIE))))
2576 alloc_flags |= ALLOC_NO_WATERMARKS;
2579 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2580 alloc_flags |= ALLOC_CMA;
2585 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2587 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2590 static inline struct page *
2591 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2592 struct alloc_context *ac)
2594 const gfp_t wait = gfp_mask & __GFP_WAIT;
2595 struct page *page = NULL;
2597 unsigned long pages_reclaimed = 0;
2598 unsigned long did_some_progress;
2599 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2600 bool deferred_compaction = false;
2601 int contended_compaction = COMPACT_CONTENDED_NONE;
2604 * In the slowpath, we sanity check order to avoid ever trying to
2605 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2606 * be using allocators in order of preference for an area that is
2609 if (order >= MAX_ORDER) {
2610 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2615 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2616 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2617 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2618 * using a larger set of nodes after it has established that the
2619 * allowed per node queues are empty and that nodes are
2622 if (IS_ENABLED(CONFIG_NUMA) &&
2623 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2627 if (!(gfp_mask & __GFP_NO_KSWAPD))
2628 wake_all_kswapds(order, ac);
2631 * OK, we're below the kswapd watermark and have kicked background
2632 * reclaim. Now things get more complex, so set up alloc_flags according
2633 * to how we want to proceed.
2635 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2638 * Find the true preferred zone if the allocation is unconstrained by
2641 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
2642 struct zoneref *preferred_zoneref;
2643 preferred_zoneref = first_zones_zonelist(ac->zonelist,
2644 ac->high_zoneidx, NULL, &ac->preferred_zone);
2645 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
2648 /* This is the last chance, in general, before the goto nopage. */
2649 page = get_page_from_freelist(gfp_mask, order,
2650 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2654 /* Allocate without watermarks if the context allows */
2655 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2657 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2658 * the allocation is high priority and these type of
2659 * allocations are system rather than user orientated
2661 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
2663 page = __alloc_pages_high_priority(gfp_mask, order, ac);
2670 /* Atomic allocations - we can't balance anything */
2673 * All existing users of the deprecated __GFP_NOFAIL are
2674 * blockable, so warn of any new users that actually allow this
2675 * type of allocation to fail.
2677 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2681 /* Avoid recursion of direct reclaim */
2682 if (current->flags & PF_MEMALLOC)
2685 /* Avoid allocations with no watermarks from looping endlessly */
2686 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2690 * Try direct compaction. The first pass is asynchronous. Subsequent
2691 * attempts after direct reclaim are synchronous
2693 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
2695 &contended_compaction,
2696 &deferred_compaction);
2700 /* Checks for THP-specific high-order allocations */
2701 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2703 * If compaction is deferred for high-order allocations, it is
2704 * because sync compaction recently failed. If this is the case
2705 * and the caller requested a THP allocation, we do not want
2706 * to heavily disrupt the system, so we fail the allocation
2707 * instead of entering direct reclaim.
2709 if (deferred_compaction)
2713 * In all zones where compaction was attempted (and not
2714 * deferred or skipped), lock contention has been detected.
2715 * For THP allocation we do not want to disrupt the others
2716 * so we fallback to base pages instead.
2718 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2722 * If compaction was aborted due to need_resched(), we do not
2723 * want to further increase allocation latency, unless it is
2724 * khugepaged trying to collapse.
2726 if (contended_compaction == COMPACT_CONTENDED_SCHED
2727 && !(current->flags & PF_KTHREAD))
2732 * It can become very expensive to allocate transparent hugepages at
2733 * fault, so use asynchronous memory compaction for THP unless it is
2734 * khugepaged trying to collapse.
2736 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2737 (current->flags & PF_KTHREAD))
2738 migration_mode = MIGRATE_SYNC_LIGHT;
2740 /* Try direct reclaim and then allocating */
2741 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
2742 &did_some_progress);
2746 /* Check if we should retry the allocation */
2747 pages_reclaimed += did_some_progress;
2748 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2751 * If we fail to make progress by freeing individual
2752 * pages, but the allocation wants us to keep going,
2753 * start OOM killing tasks.
2755 if (!did_some_progress) {
2756 page = __alloc_pages_may_oom(gfp_mask, order, ac,
2757 &did_some_progress);
2760 if (!did_some_progress)
2763 /* Wait for some write requests to complete then retry */
2764 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
2768 * High-order allocations do not necessarily loop after
2769 * direct reclaim and reclaim/compaction depends on compaction
2770 * being called after reclaim so call directly if necessary
2772 page = __alloc_pages_direct_compact(gfp_mask, order,
2773 alloc_flags, ac, migration_mode,
2774 &contended_compaction,
2775 &deferred_compaction);
2781 warn_alloc_failed(gfp_mask, order, NULL);
2787 * This is the 'heart' of the zoned buddy allocator.
2790 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2791 struct zonelist *zonelist, nodemask_t *nodemask)
2793 struct zoneref *preferred_zoneref;
2794 struct page *page = NULL;
2795 unsigned int cpuset_mems_cookie;
2796 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2797 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
2798 struct alloc_context ac = {
2799 .high_zoneidx = gfp_zone(gfp_mask),
2800 .nodemask = nodemask,
2801 .migratetype = gfpflags_to_migratetype(gfp_mask),
2804 gfp_mask &= gfp_allowed_mask;
2806 lockdep_trace_alloc(gfp_mask);
2808 might_sleep_if(gfp_mask & __GFP_WAIT);
2810 if (should_fail_alloc_page(gfp_mask, order))
2814 * Check the zones suitable for the gfp_mask contain at least one
2815 * valid zone. It's possible to have an empty zonelist as a result
2816 * of GFP_THISNODE and a memoryless node
2818 if (unlikely(!zonelist->_zonerefs->zone))
2821 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
2822 alloc_flags |= ALLOC_CMA;
2825 cpuset_mems_cookie = read_mems_allowed_begin();
2827 /* We set it here, as __alloc_pages_slowpath might have changed it */
2828 ac.zonelist = zonelist;
2829 /* The preferred zone is used for statistics later */
2830 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
2831 ac.nodemask ? : &cpuset_current_mems_allowed,
2832 &ac.preferred_zone);
2833 if (!ac.preferred_zone)
2835 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
2837 /* First allocation attempt */
2838 alloc_mask = gfp_mask|__GFP_HARDWALL;
2839 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
2840 if (unlikely(!page)) {
2842 * Runtime PM, block IO and its error handling path
2843 * can deadlock because I/O on the device might not
2846 alloc_mask = memalloc_noio_flags(gfp_mask);
2848 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
2851 if (kmemcheck_enabled && page)
2852 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2854 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
2858 * When updating a task's mems_allowed, it is possible to race with
2859 * parallel threads in such a way that an allocation can fail while
2860 * the mask is being updated. If a page allocation is about to fail,
2861 * check if the cpuset changed during allocation and if so, retry.
2863 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2868 EXPORT_SYMBOL(__alloc_pages_nodemask);
2871 * Common helper functions.
2873 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2878 * __get_free_pages() returns a 32-bit address, which cannot represent
2881 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2883 page = alloc_pages(gfp_mask, order);
2886 return (unsigned long) page_address(page);
2888 EXPORT_SYMBOL(__get_free_pages);
2890 unsigned long get_zeroed_page(gfp_t gfp_mask)
2892 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2894 EXPORT_SYMBOL(get_zeroed_page);
2896 void __free_pages(struct page *page, unsigned int order)
2898 if (put_page_testzero(page)) {
2900 free_hot_cold_page(page, false);
2902 __free_pages_ok(page, order);
2906 EXPORT_SYMBOL(__free_pages);
2908 void free_pages(unsigned long addr, unsigned int order)
2911 VM_BUG_ON(!virt_addr_valid((void *)addr));
2912 __free_pages(virt_to_page((void *)addr), order);
2916 EXPORT_SYMBOL(free_pages);
2919 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2920 * of the current memory cgroup.
2922 * It should be used when the caller would like to use kmalloc, but since the
2923 * allocation is large, it has to fall back to the page allocator.
2925 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2928 struct mem_cgroup *memcg = NULL;
2930 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2932 page = alloc_pages(gfp_mask, order);
2933 memcg_kmem_commit_charge(page, memcg, order);
2937 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2940 struct mem_cgroup *memcg = NULL;
2942 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2944 page = alloc_pages_node(nid, gfp_mask, order);
2945 memcg_kmem_commit_charge(page, memcg, order);
2950 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2953 void __free_kmem_pages(struct page *page, unsigned int order)
2955 memcg_kmem_uncharge_pages(page, order);
2956 __free_pages(page, order);
2959 void free_kmem_pages(unsigned long addr, unsigned int order)
2962 VM_BUG_ON(!virt_addr_valid((void *)addr));
2963 __free_kmem_pages(virt_to_page((void *)addr), order);
2967 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2970 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2971 unsigned long used = addr + PAGE_ALIGN(size);
2973 split_page(virt_to_page((void *)addr), order);
2974 while (used < alloc_end) {
2979 return (void *)addr;
2983 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2984 * @size: the number of bytes to allocate
2985 * @gfp_mask: GFP flags for the allocation
2987 * This function is similar to alloc_pages(), except that it allocates the
2988 * minimum number of pages to satisfy the request. alloc_pages() can only
2989 * allocate memory in power-of-two pages.
2991 * This function is also limited by MAX_ORDER.
2993 * Memory allocated by this function must be released by free_pages_exact().
2995 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2997 unsigned int order = get_order(size);
3000 addr = __get_free_pages(gfp_mask, order);
3001 return make_alloc_exact(addr, order, size);
3003 EXPORT_SYMBOL(alloc_pages_exact);
3006 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3008 * @nid: the preferred node ID where memory should be allocated
3009 * @size: the number of bytes to allocate
3010 * @gfp_mask: GFP flags for the allocation
3012 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3014 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3017 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3019 unsigned order = get_order(size);
3020 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3023 return make_alloc_exact((unsigned long)page_address(p), order, size);
3027 * free_pages_exact - release memory allocated via alloc_pages_exact()
3028 * @virt: the value returned by alloc_pages_exact.
3029 * @size: size of allocation, same value as passed to alloc_pages_exact().
3031 * Release the memory allocated by a previous call to alloc_pages_exact.
3033 void free_pages_exact(void *virt, size_t size)
3035 unsigned long addr = (unsigned long)virt;
3036 unsigned long end = addr + PAGE_ALIGN(size);
3038 while (addr < end) {
3043 EXPORT_SYMBOL(free_pages_exact);
3046 * nr_free_zone_pages - count number of pages beyond high watermark
3047 * @offset: The zone index of the highest zone
3049 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3050 * high watermark within all zones at or below a given zone index. For each
3051 * zone, the number of pages is calculated as:
3052 * managed_pages - high_pages
3054 static unsigned long nr_free_zone_pages(int offset)
3059 /* Just pick one node, since fallback list is circular */
3060 unsigned long sum = 0;
3062 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3064 for_each_zone_zonelist(zone, z, zonelist, offset) {
3065 unsigned long size = zone->managed_pages;
3066 unsigned long high = high_wmark_pages(zone);
3075 * nr_free_buffer_pages - count number of pages beyond high watermark
3077 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3078 * watermark within ZONE_DMA and ZONE_NORMAL.
3080 unsigned long nr_free_buffer_pages(void)
3082 return nr_free_zone_pages(gfp_zone(GFP_USER));
3084 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3087 * nr_free_pagecache_pages - count number of pages beyond high watermark
3089 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3090 * high watermark within all zones.
3092 unsigned long nr_free_pagecache_pages(void)
3094 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3097 static inline void show_node(struct zone *zone)
3099 if (IS_ENABLED(CONFIG_NUMA))
3100 printk("Node %d ", zone_to_nid(zone));
3103 void si_meminfo(struct sysinfo *val)
3105 val->totalram = totalram_pages;
3106 val->sharedram = global_page_state(NR_SHMEM);
3107 val->freeram = global_page_state(NR_FREE_PAGES);
3108 val->bufferram = nr_blockdev_pages();
3109 val->totalhigh = totalhigh_pages;
3110 val->freehigh = nr_free_highpages();
3111 val->mem_unit = PAGE_SIZE;
3114 EXPORT_SYMBOL(si_meminfo);
3117 void si_meminfo_node(struct sysinfo *val, int nid)
3119 int zone_type; /* needs to be signed */
3120 unsigned long managed_pages = 0;
3121 pg_data_t *pgdat = NODE_DATA(nid);
3123 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3124 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3125 val->totalram = managed_pages;
3126 val->sharedram = node_page_state(nid, NR_SHMEM);
3127 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3128 #ifdef CONFIG_HIGHMEM
3129 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3130 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3136 val->mem_unit = PAGE_SIZE;
3141 * Determine whether the node should be displayed or not, depending on whether
3142 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3144 bool skip_free_areas_node(unsigned int flags, int nid)
3147 unsigned int cpuset_mems_cookie;
3149 if (!(flags & SHOW_MEM_FILTER_NODES))
3153 cpuset_mems_cookie = read_mems_allowed_begin();
3154 ret = !node_isset(nid, cpuset_current_mems_allowed);
3155 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3160 #define K(x) ((x) << (PAGE_SHIFT-10))
3162 static void show_migration_types(unsigned char type)
3164 static const char types[MIGRATE_TYPES] = {
3165 [MIGRATE_UNMOVABLE] = 'U',
3166 [MIGRATE_RECLAIMABLE] = 'E',
3167 [MIGRATE_MOVABLE] = 'M',
3168 [MIGRATE_RESERVE] = 'R',
3170 [MIGRATE_CMA] = 'C',
3172 #ifdef CONFIG_MEMORY_ISOLATION
3173 [MIGRATE_ISOLATE] = 'I',
3176 char tmp[MIGRATE_TYPES + 1];
3180 for (i = 0; i < MIGRATE_TYPES; i++) {
3181 if (type & (1 << i))
3186 printk("(%s) ", tmp);
3190 * Show free area list (used inside shift_scroll-lock stuff)
3191 * We also calculate the percentage fragmentation. We do this by counting the
3192 * memory on each free list with the exception of the first item on the list.
3193 * Suppresses nodes that are not allowed by current's cpuset if
3194 * SHOW_MEM_FILTER_NODES is passed.
3196 void show_free_areas(unsigned int filter)
3201 for_each_populated_zone(zone) {
3202 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3205 printk("%s per-cpu:\n", zone->name);
3207 for_each_online_cpu(cpu) {
3208 struct per_cpu_pageset *pageset;
3210 pageset = per_cpu_ptr(zone->pageset, cpu);
3212 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3213 cpu, pageset->pcp.high,
3214 pageset->pcp.batch, pageset->pcp.count);
3218 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3219 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3221 " dirty:%lu writeback:%lu unstable:%lu\n"
3222 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3223 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3225 global_page_state(NR_ACTIVE_ANON),
3226 global_page_state(NR_INACTIVE_ANON),
3227 global_page_state(NR_ISOLATED_ANON),
3228 global_page_state(NR_ACTIVE_FILE),
3229 global_page_state(NR_INACTIVE_FILE),
3230 global_page_state(NR_ISOLATED_FILE),
3231 global_page_state(NR_UNEVICTABLE),
3232 global_page_state(NR_FILE_DIRTY),
3233 global_page_state(NR_WRITEBACK),
3234 global_page_state(NR_UNSTABLE_NFS),
3235 global_page_state(NR_FREE_PAGES),
3236 global_page_state(NR_SLAB_RECLAIMABLE),
3237 global_page_state(NR_SLAB_UNRECLAIMABLE),
3238 global_page_state(NR_FILE_MAPPED),
3239 global_page_state(NR_SHMEM),
3240 global_page_state(NR_PAGETABLE),
3241 global_page_state(NR_BOUNCE),
3242 global_page_state(NR_FREE_CMA_PAGES));
3244 for_each_populated_zone(zone) {
3247 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3255 " active_anon:%lukB"
3256 " inactive_anon:%lukB"
3257 " active_file:%lukB"
3258 " inactive_file:%lukB"
3259 " unevictable:%lukB"
3260 " isolated(anon):%lukB"
3261 " isolated(file):%lukB"
3269 " slab_reclaimable:%lukB"
3270 " slab_unreclaimable:%lukB"
3271 " kernel_stack:%lukB"
3276 " writeback_tmp:%lukB"
3277 " pages_scanned:%lu"
3278 " all_unreclaimable? %s"
3281 K(zone_page_state(zone, NR_FREE_PAGES)),
3282 K(min_wmark_pages(zone)),
3283 K(low_wmark_pages(zone)),
3284 K(high_wmark_pages(zone)),
3285 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3286 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3287 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3288 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3289 K(zone_page_state(zone, NR_UNEVICTABLE)),
3290 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3291 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3292 K(zone->present_pages),
3293 K(zone->managed_pages),
3294 K(zone_page_state(zone, NR_MLOCK)),
3295 K(zone_page_state(zone, NR_FILE_DIRTY)),
3296 K(zone_page_state(zone, NR_WRITEBACK)),
3297 K(zone_page_state(zone, NR_FILE_MAPPED)),
3298 K(zone_page_state(zone, NR_SHMEM)),
3299 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3300 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3301 zone_page_state(zone, NR_KERNEL_STACK) *
3303 K(zone_page_state(zone, NR_PAGETABLE)),
3304 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3305 K(zone_page_state(zone, NR_BOUNCE)),
3306 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3307 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3308 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3309 (!zone_reclaimable(zone) ? "yes" : "no")
3311 printk("lowmem_reserve[]:");
3312 for (i = 0; i < MAX_NR_ZONES; i++)
3313 printk(" %ld", zone->lowmem_reserve[i]);
3317 for_each_populated_zone(zone) {
3318 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3319 unsigned char types[MAX_ORDER];
3321 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3324 printk("%s: ", zone->name);
3326 spin_lock_irqsave(&zone->lock, flags);
3327 for (order = 0; order < MAX_ORDER; order++) {
3328 struct free_area *area = &zone->free_area[order];
3331 nr[order] = area->nr_free;
3332 total += nr[order] << order;
3335 for (type = 0; type < MIGRATE_TYPES; type++) {
3336 if (!list_empty(&area->free_list[type]))
3337 types[order] |= 1 << type;
3340 spin_unlock_irqrestore(&zone->lock, flags);
3341 for (order = 0; order < MAX_ORDER; order++) {
3342 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3344 show_migration_types(types[order]);
3346 printk("= %lukB\n", K(total));
3349 hugetlb_show_meminfo();
3351 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3353 show_swap_cache_info();
3356 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3358 zoneref->zone = zone;
3359 zoneref->zone_idx = zone_idx(zone);
3363 * Builds allocation fallback zone lists.
3365 * Add all populated zones of a node to the zonelist.
3367 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3371 enum zone_type zone_type = MAX_NR_ZONES;
3375 zone = pgdat->node_zones + zone_type;
3376 if (populated_zone(zone)) {
3377 zoneref_set_zone(zone,
3378 &zonelist->_zonerefs[nr_zones++]);
3379 check_highest_zone(zone_type);
3381 } while (zone_type);
3389 * 0 = automatic detection of better ordering.
3390 * 1 = order by ([node] distance, -zonetype)
3391 * 2 = order by (-zonetype, [node] distance)
3393 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3394 * the same zonelist. So only NUMA can configure this param.
3396 #define ZONELIST_ORDER_DEFAULT 0
3397 #define ZONELIST_ORDER_NODE 1
3398 #define ZONELIST_ORDER_ZONE 2
3400 /* zonelist order in the kernel.
3401 * set_zonelist_order() will set this to NODE or ZONE.
3403 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3404 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3408 /* The value user specified ....changed by config */
3409 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3410 /* string for sysctl */
3411 #define NUMA_ZONELIST_ORDER_LEN 16
3412 char numa_zonelist_order[16] = "default";
3415 * interface for configure zonelist ordering.
3416 * command line option "numa_zonelist_order"
3417 * = "[dD]efault - default, automatic configuration.
3418 * = "[nN]ode - order by node locality, then by zone within node
3419 * = "[zZ]one - order by zone, then by locality within zone
3422 static int __parse_numa_zonelist_order(char *s)
3424 if (*s == 'd' || *s == 'D') {
3425 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3426 } else if (*s == 'n' || *s == 'N') {
3427 user_zonelist_order = ZONELIST_ORDER_NODE;
3428 } else if (*s == 'z' || *s == 'Z') {
3429 user_zonelist_order = ZONELIST_ORDER_ZONE;
3432 "Ignoring invalid numa_zonelist_order value: "
3439 static __init int setup_numa_zonelist_order(char *s)
3446 ret = __parse_numa_zonelist_order(s);
3448 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3452 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3455 * sysctl handler for numa_zonelist_order
3457 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3458 void __user *buffer, size_t *length,
3461 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3463 static DEFINE_MUTEX(zl_order_mutex);
3465 mutex_lock(&zl_order_mutex);
3467 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3471 strcpy(saved_string, (char *)table->data);
3473 ret = proc_dostring(table, write, buffer, length, ppos);
3477 int oldval = user_zonelist_order;
3479 ret = __parse_numa_zonelist_order((char *)table->data);
3482 * bogus value. restore saved string
3484 strncpy((char *)table->data, saved_string,
3485 NUMA_ZONELIST_ORDER_LEN);
3486 user_zonelist_order = oldval;
3487 } else if (oldval != user_zonelist_order) {
3488 mutex_lock(&zonelists_mutex);
3489 build_all_zonelists(NULL, NULL);
3490 mutex_unlock(&zonelists_mutex);
3494 mutex_unlock(&zl_order_mutex);
3499 #define MAX_NODE_LOAD (nr_online_nodes)
3500 static int node_load[MAX_NUMNODES];
3503 * find_next_best_node - find the next node that should appear in a given node's fallback list
3504 * @node: node whose fallback list we're appending
3505 * @used_node_mask: nodemask_t of already used nodes
3507 * We use a number of factors to determine which is the next node that should
3508 * appear on a given node's fallback list. The node should not have appeared
3509 * already in @node's fallback list, and it should be the next closest node
3510 * according to the distance array (which contains arbitrary distance values
3511 * from each node to each node in the system), and should also prefer nodes
3512 * with no CPUs, since presumably they'll have very little allocation pressure
3513 * on them otherwise.
3514 * It returns -1 if no node is found.
3516 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3519 int min_val = INT_MAX;
3520 int best_node = NUMA_NO_NODE;
3521 const struct cpumask *tmp = cpumask_of_node(0);
3523 /* Use the local node if we haven't already */
3524 if (!node_isset(node, *used_node_mask)) {
3525 node_set(node, *used_node_mask);
3529 for_each_node_state(n, N_MEMORY) {
3531 /* Don't want a node to appear more than once */
3532 if (node_isset(n, *used_node_mask))
3535 /* Use the distance array to find the distance */
3536 val = node_distance(node, n);
3538 /* Penalize nodes under us ("prefer the next node") */
3541 /* Give preference to headless and unused nodes */
3542 tmp = cpumask_of_node(n);
3543 if (!cpumask_empty(tmp))
3544 val += PENALTY_FOR_NODE_WITH_CPUS;
3546 /* Slight preference for less loaded node */
3547 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3548 val += node_load[n];
3550 if (val < min_val) {
3557 node_set(best_node, *used_node_mask);
3564 * Build zonelists ordered by node and zones within node.
3565 * This results in maximum locality--normal zone overflows into local
3566 * DMA zone, if any--but risks exhausting DMA zone.
3568 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3571 struct zonelist *zonelist;
3573 zonelist = &pgdat->node_zonelists[0];
3574 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3576 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3577 zonelist->_zonerefs[j].zone = NULL;
3578 zonelist->_zonerefs[j].zone_idx = 0;
3582 * Build gfp_thisnode zonelists
3584 static void build_thisnode_zonelists(pg_data_t *pgdat)
3587 struct zonelist *zonelist;
3589 zonelist = &pgdat->node_zonelists[1];
3590 j = build_zonelists_node(pgdat, zonelist, 0);
3591 zonelist->_zonerefs[j].zone = NULL;
3592 zonelist->_zonerefs[j].zone_idx = 0;
3596 * Build zonelists ordered by zone and nodes within zones.
3597 * This results in conserving DMA zone[s] until all Normal memory is
3598 * exhausted, but results in overflowing to remote node while memory
3599 * may still exist in local DMA zone.
3601 static int node_order[MAX_NUMNODES];
3603 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3606 int zone_type; /* needs to be signed */
3608 struct zonelist *zonelist;
3610 zonelist = &pgdat->node_zonelists[0];
3612 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3613 for (j = 0; j < nr_nodes; j++) {
3614 node = node_order[j];
3615 z = &NODE_DATA(node)->node_zones[zone_type];
3616 if (populated_zone(z)) {
3618 &zonelist->_zonerefs[pos++]);
3619 check_highest_zone(zone_type);
3623 zonelist->_zonerefs[pos].zone = NULL;
3624 zonelist->_zonerefs[pos].zone_idx = 0;
3627 #if defined(CONFIG_64BIT)
3629 * Devices that require DMA32/DMA are relatively rare and do not justify a
3630 * penalty to every machine in case the specialised case applies. Default
3631 * to Node-ordering on 64-bit NUMA machines
3633 static int default_zonelist_order(void)
3635 return ZONELIST_ORDER_NODE;
3639 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3640 * by the kernel. If processes running on node 0 deplete the low memory zone
3641 * then reclaim will occur more frequency increasing stalls and potentially
3642 * be easier to OOM if a large percentage of the zone is under writeback or
3643 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3644 * Hence, default to zone ordering on 32-bit.
3646 static int default_zonelist_order(void)
3648 return ZONELIST_ORDER_ZONE;
3650 #endif /* CONFIG_64BIT */
3652 static void set_zonelist_order(void)
3654 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3655 current_zonelist_order = default_zonelist_order();
3657 current_zonelist_order = user_zonelist_order;
3660 static void build_zonelists(pg_data_t *pgdat)
3664 nodemask_t used_mask;
3665 int local_node, prev_node;
3666 struct zonelist *zonelist;
3667 int order = current_zonelist_order;
3669 /* initialize zonelists */
3670 for (i = 0; i < MAX_ZONELISTS; i++) {
3671 zonelist = pgdat->node_zonelists + i;
3672 zonelist->_zonerefs[0].zone = NULL;
3673 zonelist->_zonerefs[0].zone_idx = 0;
3676 /* NUMA-aware ordering of nodes */
3677 local_node = pgdat->node_id;
3678 load = nr_online_nodes;
3679 prev_node = local_node;
3680 nodes_clear(used_mask);
3682 memset(node_order, 0, sizeof(node_order));
3685 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3687 * We don't want to pressure a particular node.
3688 * So adding penalty to the first node in same
3689 * distance group to make it round-robin.
3691 if (node_distance(local_node, node) !=
3692 node_distance(local_node, prev_node))
3693 node_load[node] = load;
3697 if (order == ZONELIST_ORDER_NODE)
3698 build_zonelists_in_node_order(pgdat, node);
3700 node_order[j++] = node; /* remember order */
3703 if (order == ZONELIST_ORDER_ZONE) {
3704 /* calculate node order -- i.e., DMA last! */
3705 build_zonelists_in_zone_order(pgdat, j);
3708 build_thisnode_zonelists(pgdat);
3711 /* Construct the zonelist performance cache - see further mmzone.h */
3712 static void build_zonelist_cache(pg_data_t *pgdat)
3714 struct zonelist *zonelist;
3715 struct zonelist_cache *zlc;
3718 zonelist = &pgdat->node_zonelists[0];
3719 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3720 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3721 for (z = zonelist->_zonerefs; z->zone; z++)
3722 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3725 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3727 * Return node id of node used for "local" allocations.
3728 * I.e., first node id of first zone in arg node's generic zonelist.
3729 * Used for initializing percpu 'numa_mem', which is used primarily
3730 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3732 int local_memory_node(int node)
3736 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3737 gfp_zone(GFP_KERNEL),
3744 #else /* CONFIG_NUMA */
3746 static void set_zonelist_order(void)
3748 current_zonelist_order = ZONELIST_ORDER_ZONE;
3751 static void build_zonelists(pg_data_t *pgdat)
3753 int node, local_node;
3755 struct zonelist *zonelist;
3757 local_node = pgdat->node_id;
3759 zonelist = &pgdat->node_zonelists[0];
3760 j = build_zonelists_node(pgdat, zonelist, 0);
3763 * Now we build the zonelist so that it contains the zones
3764 * of all the other nodes.
3765 * We don't want to pressure a particular node, so when
3766 * building the zones for node N, we make sure that the
3767 * zones coming right after the local ones are those from
3768 * node N+1 (modulo N)
3770 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3771 if (!node_online(node))
3773 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3775 for (node = 0; node < local_node; node++) {
3776 if (!node_online(node))
3778 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3781 zonelist->_zonerefs[j].zone = NULL;
3782 zonelist->_zonerefs[j].zone_idx = 0;
3785 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3786 static void build_zonelist_cache(pg_data_t *pgdat)
3788 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3791 #endif /* CONFIG_NUMA */
3794 * Boot pageset table. One per cpu which is going to be used for all
3795 * zones and all nodes. The parameters will be set in such a way
3796 * that an item put on a list will immediately be handed over to
3797 * the buddy list. This is safe since pageset manipulation is done
3798 * with interrupts disabled.
3800 * The boot_pagesets must be kept even after bootup is complete for
3801 * unused processors and/or zones. They do play a role for bootstrapping
3802 * hotplugged processors.
3804 * zoneinfo_show() and maybe other functions do
3805 * not check if the processor is online before following the pageset pointer.
3806 * Other parts of the kernel may not check if the zone is available.
3808 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3809 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3810 static void setup_zone_pageset(struct zone *zone);
3813 * Global mutex to protect against size modification of zonelists
3814 * as well as to serialize pageset setup for the new populated zone.
3816 DEFINE_MUTEX(zonelists_mutex);
3818 /* return values int ....just for stop_machine() */
3819 static int __build_all_zonelists(void *data)
3823 pg_data_t *self = data;
3826 memset(node_load, 0, sizeof(node_load));
3829 if (self && !node_online(self->node_id)) {
3830 build_zonelists(self);
3831 build_zonelist_cache(self);
3834 for_each_online_node(nid) {
3835 pg_data_t *pgdat = NODE_DATA(nid);
3837 build_zonelists(pgdat);
3838 build_zonelist_cache(pgdat);
3842 * Initialize the boot_pagesets that are going to be used
3843 * for bootstrapping processors. The real pagesets for
3844 * each zone will be allocated later when the per cpu
3845 * allocator is available.
3847 * boot_pagesets are used also for bootstrapping offline
3848 * cpus if the system is already booted because the pagesets
3849 * are needed to initialize allocators on a specific cpu too.
3850 * F.e. the percpu allocator needs the page allocator which
3851 * needs the percpu allocator in order to allocate its pagesets
3852 * (a chicken-egg dilemma).
3854 for_each_possible_cpu(cpu) {
3855 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3857 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3859 * We now know the "local memory node" for each node--
3860 * i.e., the node of the first zone in the generic zonelist.
3861 * Set up numa_mem percpu variable for on-line cpus. During
3862 * boot, only the boot cpu should be on-line; we'll init the
3863 * secondary cpus' numa_mem as they come on-line. During
3864 * node/memory hotplug, we'll fixup all on-line cpus.
3866 if (cpu_online(cpu))
3867 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3875 * Called with zonelists_mutex held always
3876 * unless system_state == SYSTEM_BOOTING.
3878 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3880 set_zonelist_order();
3882 if (system_state == SYSTEM_BOOTING) {
3883 __build_all_zonelists(NULL);
3884 mminit_verify_zonelist();
3885 cpuset_init_current_mems_allowed();
3887 #ifdef CONFIG_MEMORY_HOTPLUG
3889 setup_zone_pageset(zone);
3891 /* we have to stop all cpus to guarantee there is no user
3893 stop_machine(__build_all_zonelists, pgdat, NULL);
3894 /* cpuset refresh routine should be here */
3896 vm_total_pages = nr_free_pagecache_pages();
3898 * Disable grouping by mobility if the number of pages in the
3899 * system is too low to allow the mechanism to work. It would be
3900 * more accurate, but expensive to check per-zone. This check is
3901 * made on memory-hotadd so a system can start with mobility
3902 * disabled and enable it later
3904 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3905 page_group_by_mobility_disabled = 1;
3907 page_group_by_mobility_disabled = 0;
3909 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3910 "Total pages: %ld\n",
3912 zonelist_order_name[current_zonelist_order],
3913 page_group_by_mobility_disabled ? "off" : "on",
3916 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
3921 * Helper functions to size the waitqueue hash table.
3922 * Essentially these want to choose hash table sizes sufficiently
3923 * large so that collisions trying to wait on pages are rare.
3924 * But in fact, the number of active page waitqueues on typical
3925 * systems is ridiculously low, less than 200. So this is even
3926 * conservative, even though it seems large.
3928 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3929 * waitqueues, i.e. the size of the waitq table given the number of pages.
3931 #define PAGES_PER_WAITQUEUE 256
3933 #ifndef CONFIG_MEMORY_HOTPLUG
3934 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3936 unsigned long size = 1;
3938 pages /= PAGES_PER_WAITQUEUE;
3940 while (size < pages)
3944 * Once we have dozens or even hundreds of threads sleeping
3945 * on IO we've got bigger problems than wait queue collision.
3946 * Limit the size of the wait table to a reasonable size.
3948 size = min(size, 4096UL);
3950 return max(size, 4UL);
3954 * A zone's size might be changed by hot-add, so it is not possible to determine
3955 * a suitable size for its wait_table. So we use the maximum size now.
3957 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3959 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3960 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3961 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3963 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3964 * or more by the traditional way. (See above). It equals:
3966 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3967 * ia64(16K page size) : = ( 8G + 4M)byte.
3968 * powerpc (64K page size) : = (32G +16M)byte.
3970 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3977 * This is an integer logarithm so that shifts can be used later
3978 * to extract the more random high bits from the multiplicative
3979 * hash function before the remainder is taken.
3981 static inline unsigned long wait_table_bits(unsigned long size)
3987 * Check if a pageblock contains reserved pages
3989 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3993 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3994 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4001 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4002 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4003 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4004 * higher will lead to a bigger reserve which will get freed as contiguous
4005 * blocks as reclaim kicks in
4007 static void setup_zone_migrate_reserve(struct zone *zone)
4009 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4011 unsigned long block_migratetype;
4016 * Get the start pfn, end pfn and the number of blocks to reserve
4017 * We have to be careful to be aligned to pageblock_nr_pages to
4018 * make sure that we always check pfn_valid for the first page in
4021 start_pfn = zone->zone_start_pfn;
4022 end_pfn = zone_end_pfn(zone);
4023 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4024 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4028 * Reserve blocks are generally in place to help high-order atomic
4029 * allocations that are short-lived. A min_free_kbytes value that
4030 * would result in more than 2 reserve blocks for atomic allocations
4031 * is assumed to be in place to help anti-fragmentation for the
4032 * future allocation of hugepages at runtime.
4034 reserve = min(2, reserve);
4035 old_reserve = zone->nr_migrate_reserve_block;
4037 /* When memory hot-add, we almost always need to do nothing */
4038 if (reserve == old_reserve)
4040 zone->nr_migrate_reserve_block = reserve;
4042 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4043 if (!pfn_valid(pfn))
4045 page = pfn_to_page(pfn);
4047 /* Watch out for overlapping nodes */
4048 if (page_to_nid(page) != zone_to_nid(zone))
4051 block_migratetype = get_pageblock_migratetype(page);
4053 /* Only test what is necessary when the reserves are not met */
4056 * Blocks with reserved pages will never free, skip
4059 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4060 if (pageblock_is_reserved(pfn, block_end_pfn))
4063 /* If this block is reserved, account for it */
4064 if (block_migratetype == MIGRATE_RESERVE) {
4069 /* Suitable for reserving if this block is movable */
4070 if (block_migratetype == MIGRATE_MOVABLE) {
4071 set_pageblock_migratetype(page,
4073 move_freepages_block(zone, page,
4078 } else if (!old_reserve) {
4080 * At boot time we don't need to scan the whole zone
4081 * for turning off MIGRATE_RESERVE.
4087 * If the reserve is met and this is a previous reserved block,
4090 if (block_migratetype == MIGRATE_RESERVE) {
4091 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4092 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4098 * Initially all pages are reserved - free ones are freed
4099 * up by free_all_bootmem() once the early boot process is
4100 * done. Non-atomic initialization, single-pass.
4102 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4103 unsigned long start_pfn, enum memmap_context context)
4106 unsigned long end_pfn = start_pfn + size;
4110 if (highest_memmap_pfn < end_pfn - 1)
4111 highest_memmap_pfn = end_pfn - 1;
4113 z = &NODE_DATA(nid)->node_zones[zone];
4114 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4116 * There can be holes in boot-time mem_map[]s
4117 * handed to this function. They do not
4118 * exist on hotplugged memory.
4120 if (context == MEMMAP_EARLY) {
4121 if (!early_pfn_valid(pfn))
4123 if (!early_pfn_in_nid(pfn, nid))
4126 page = pfn_to_page(pfn);
4127 set_page_links(page, zone, nid, pfn);
4128 mminit_verify_page_links(page, zone, nid, pfn);
4129 init_page_count(page);
4130 page_mapcount_reset(page);
4131 page_cpupid_reset_last(page);
4132 SetPageReserved(page);
4134 * Mark the block movable so that blocks are reserved for
4135 * movable at startup. This will force kernel allocations
4136 * to reserve their blocks rather than leaking throughout
4137 * the address space during boot when many long-lived
4138 * kernel allocations are made. Later some blocks near
4139 * the start are marked MIGRATE_RESERVE by
4140 * setup_zone_migrate_reserve()
4142 * bitmap is created for zone's valid pfn range. but memmap
4143 * can be created for invalid pages (for alignment)
4144 * check here not to call set_pageblock_migratetype() against
4147 if ((z->zone_start_pfn <= pfn)
4148 && (pfn < zone_end_pfn(z))
4149 && !(pfn & (pageblock_nr_pages - 1)))
4150 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4152 INIT_LIST_HEAD(&page->lru);
4153 #ifdef WANT_PAGE_VIRTUAL
4154 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4155 if (!is_highmem_idx(zone))
4156 set_page_address(page, __va(pfn << PAGE_SHIFT));
4161 static void __meminit zone_init_free_lists(struct zone *zone)
4163 unsigned int order, t;
4164 for_each_migratetype_order(order, t) {
4165 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4166 zone->free_area[order].nr_free = 0;
4170 #ifndef __HAVE_ARCH_MEMMAP_INIT
4171 #define memmap_init(size, nid, zone, start_pfn) \
4172 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4175 static int zone_batchsize(struct zone *zone)
4181 * The per-cpu-pages pools are set to around 1000th of the
4182 * size of the zone. But no more than 1/2 of a meg.
4184 * OK, so we don't know how big the cache is. So guess.
4186 batch = zone->managed_pages / 1024;
4187 if (batch * PAGE_SIZE > 512 * 1024)
4188 batch = (512 * 1024) / PAGE_SIZE;
4189 batch /= 4; /* We effectively *= 4 below */
4194 * Clamp the batch to a 2^n - 1 value. Having a power
4195 * of 2 value was found to be more likely to have
4196 * suboptimal cache aliasing properties in some cases.
4198 * For example if 2 tasks are alternately allocating
4199 * batches of pages, one task can end up with a lot
4200 * of pages of one half of the possible page colors
4201 * and the other with pages of the other colors.
4203 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4208 /* The deferral and batching of frees should be suppressed under NOMMU
4211 * The problem is that NOMMU needs to be able to allocate large chunks
4212 * of contiguous memory as there's no hardware page translation to
4213 * assemble apparent contiguous memory from discontiguous pages.
4215 * Queueing large contiguous runs of pages for batching, however,
4216 * causes the pages to actually be freed in smaller chunks. As there
4217 * can be a significant delay between the individual batches being
4218 * recycled, this leads to the once large chunks of space being
4219 * fragmented and becoming unavailable for high-order allocations.
4226 * pcp->high and pcp->batch values are related and dependent on one another:
4227 * ->batch must never be higher then ->high.
4228 * The following function updates them in a safe manner without read side
4231 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4232 * those fields changing asynchronously (acording the the above rule).
4234 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4235 * outside of boot time (or some other assurance that no concurrent updaters
4238 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4239 unsigned long batch)
4241 /* start with a fail safe value for batch */
4245 /* Update high, then batch, in order */
4252 /* a companion to pageset_set_high() */
4253 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4255 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4258 static void pageset_init(struct per_cpu_pageset *p)
4260 struct per_cpu_pages *pcp;
4263 memset(p, 0, sizeof(*p));
4267 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4268 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4271 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4274 pageset_set_batch(p, batch);
4278 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4279 * to the value high for the pageset p.
4281 static void pageset_set_high(struct per_cpu_pageset *p,
4284 unsigned long batch = max(1UL, high / 4);
4285 if ((high / 4) > (PAGE_SHIFT * 8))
4286 batch = PAGE_SHIFT * 8;
4288 pageset_update(&p->pcp, high, batch);
4291 static void pageset_set_high_and_batch(struct zone *zone,
4292 struct per_cpu_pageset *pcp)
4294 if (percpu_pagelist_fraction)
4295 pageset_set_high(pcp,
4296 (zone->managed_pages /
4297 percpu_pagelist_fraction));
4299 pageset_set_batch(pcp, zone_batchsize(zone));
4302 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4304 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4307 pageset_set_high_and_batch(zone, pcp);
4310 static void __meminit setup_zone_pageset(struct zone *zone)
4313 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4314 for_each_possible_cpu(cpu)
4315 zone_pageset_init(zone, cpu);
4319 * Allocate per cpu pagesets and initialize them.
4320 * Before this call only boot pagesets were available.
4322 void __init setup_per_cpu_pageset(void)
4326 for_each_populated_zone(zone)
4327 setup_zone_pageset(zone);
4330 static noinline __init_refok
4331 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4337 * The per-page waitqueue mechanism uses hashed waitqueues
4340 zone->wait_table_hash_nr_entries =
4341 wait_table_hash_nr_entries(zone_size_pages);
4342 zone->wait_table_bits =
4343 wait_table_bits(zone->wait_table_hash_nr_entries);
4344 alloc_size = zone->wait_table_hash_nr_entries
4345 * sizeof(wait_queue_head_t);
4347 if (!slab_is_available()) {
4348 zone->wait_table = (wait_queue_head_t *)
4349 memblock_virt_alloc_node_nopanic(
4350 alloc_size, zone->zone_pgdat->node_id);
4353 * This case means that a zone whose size was 0 gets new memory
4354 * via memory hot-add.
4355 * But it may be the case that a new node was hot-added. In
4356 * this case vmalloc() will not be able to use this new node's
4357 * memory - this wait_table must be initialized to use this new
4358 * node itself as well.
4359 * To use this new node's memory, further consideration will be
4362 zone->wait_table = vmalloc(alloc_size);
4364 if (!zone->wait_table)
4367 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4368 init_waitqueue_head(zone->wait_table + i);
4373 static __meminit void zone_pcp_init(struct zone *zone)
4376 * per cpu subsystem is not up at this point. The following code
4377 * relies on the ability of the linker to provide the
4378 * offset of a (static) per cpu variable into the per cpu area.
4380 zone->pageset = &boot_pageset;
4382 if (populated_zone(zone))
4383 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4384 zone->name, zone->present_pages,
4385 zone_batchsize(zone));
4388 int __meminit init_currently_empty_zone(struct zone *zone,
4389 unsigned long zone_start_pfn,
4391 enum memmap_context context)
4393 struct pglist_data *pgdat = zone->zone_pgdat;
4395 ret = zone_wait_table_init(zone, size);
4398 pgdat->nr_zones = zone_idx(zone) + 1;
4400 zone->zone_start_pfn = zone_start_pfn;
4402 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4403 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4405 (unsigned long)zone_idx(zone),
4406 zone_start_pfn, (zone_start_pfn + size));
4408 zone_init_free_lists(zone);
4413 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4414 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4416 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4418 int __meminit __early_pfn_to_nid(unsigned long pfn)
4420 unsigned long start_pfn, end_pfn;
4423 * NOTE: The following SMP-unsafe globals are only used early in boot
4424 * when the kernel is running single-threaded.
4426 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4427 static int __meminitdata last_nid;
4429 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4432 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4434 last_start_pfn = start_pfn;
4435 last_end_pfn = end_pfn;
4441 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4443 int __meminit early_pfn_to_nid(unsigned long pfn)
4447 nid = __early_pfn_to_nid(pfn);
4450 /* just returns 0 */
4454 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4455 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4459 nid = __early_pfn_to_nid(pfn);
4460 if (nid >= 0 && nid != node)
4467 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4468 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4469 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4471 * If an architecture guarantees that all ranges registered contain no holes
4472 * and may be freed, this this function may be used instead of calling
4473 * memblock_free_early_nid() manually.
4475 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4477 unsigned long start_pfn, end_pfn;
4480 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4481 start_pfn = min(start_pfn, max_low_pfn);
4482 end_pfn = min(end_pfn, max_low_pfn);
4484 if (start_pfn < end_pfn)
4485 memblock_free_early_nid(PFN_PHYS(start_pfn),
4486 (end_pfn - start_pfn) << PAGE_SHIFT,
4492 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4493 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4495 * If an architecture guarantees that all ranges registered contain no holes and may
4496 * be freed, this function may be used instead of calling memory_present() manually.
4498 void __init sparse_memory_present_with_active_regions(int nid)
4500 unsigned long start_pfn, end_pfn;
4503 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4504 memory_present(this_nid, start_pfn, end_pfn);
4508 * get_pfn_range_for_nid - Return the start and end page frames for a node
4509 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4510 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4511 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4513 * It returns the start and end page frame of a node based on information
4514 * provided by memblock_set_node(). If called for a node
4515 * with no available memory, a warning is printed and the start and end
4518 void __meminit get_pfn_range_for_nid(unsigned int nid,
4519 unsigned long *start_pfn, unsigned long *end_pfn)
4521 unsigned long this_start_pfn, this_end_pfn;
4527 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4528 *start_pfn = min(*start_pfn, this_start_pfn);
4529 *end_pfn = max(*end_pfn, this_end_pfn);
4532 if (*start_pfn == -1UL)
4537 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4538 * assumption is made that zones within a node are ordered in monotonic
4539 * increasing memory addresses so that the "highest" populated zone is used
4541 static void __init find_usable_zone_for_movable(void)
4544 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4545 if (zone_index == ZONE_MOVABLE)
4548 if (arch_zone_highest_possible_pfn[zone_index] >
4549 arch_zone_lowest_possible_pfn[zone_index])
4553 VM_BUG_ON(zone_index == -1);
4554 movable_zone = zone_index;
4558 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4559 * because it is sized independent of architecture. Unlike the other zones,
4560 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4561 * in each node depending on the size of each node and how evenly kernelcore
4562 * is distributed. This helper function adjusts the zone ranges
4563 * provided by the architecture for a given node by using the end of the
4564 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4565 * zones within a node are in order of monotonic increases memory addresses
4567 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4568 unsigned long zone_type,
4569 unsigned long node_start_pfn,
4570 unsigned long node_end_pfn,
4571 unsigned long *zone_start_pfn,
4572 unsigned long *zone_end_pfn)
4574 /* Only adjust if ZONE_MOVABLE is on this node */
4575 if (zone_movable_pfn[nid]) {
4576 /* Size ZONE_MOVABLE */
4577 if (zone_type == ZONE_MOVABLE) {
4578 *zone_start_pfn = zone_movable_pfn[nid];
4579 *zone_end_pfn = min(node_end_pfn,
4580 arch_zone_highest_possible_pfn[movable_zone]);
4582 /* Adjust for ZONE_MOVABLE starting within this range */
4583 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4584 *zone_end_pfn > zone_movable_pfn[nid]) {
4585 *zone_end_pfn = zone_movable_pfn[nid];
4587 /* Check if this whole range is within ZONE_MOVABLE */
4588 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4589 *zone_start_pfn = *zone_end_pfn;
4594 * Return the number of pages a zone spans in a node, including holes
4595 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4597 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4598 unsigned long zone_type,
4599 unsigned long node_start_pfn,
4600 unsigned long node_end_pfn,
4601 unsigned long *ignored)
4603 unsigned long zone_start_pfn, zone_end_pfn;
4605 /* Get the start and end of the zone */
4606 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4607 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4608 adjust_zone_range_for_zone_movable(nid, zone_type,
4609 node_start_pfn, node_end_pfn,
4610 &zone_start_pfn, &zone_end_pfn);
4612 /* Check that this node has pages within the zone's required range */
4613 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4616 /* Move the zone boundaries inside the node if necessary */
4617 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4618 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4620 /* Return the spanned pages */
4621 return zone_end_pfn - zone_start_pfn;
4625 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4626 * then all holes in the requested range will be accounted for.
4628 unsigned long __meminit __absent_pages_in_range(int nid,
4629 unsigned long range_start_pfn,
4630 unsigned long range_end_pfn)
4632 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4633 unsigned long start_pfn, end_pfn;
4636 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4637 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4638 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4639 nr_absent -= end_pfn - start_pfn;
4645 * absent_pages_in_range - Return number of page frames in holes within a range
4646 * @start_pfn: The start PFN to start searching for holes
4647 * @end_pfn: The end PFN to stop searching for holes
4649 * It returns the number of pages frames in memory holes within a range.
4651 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4652 unsigned long end_pfn)
4654 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4657 /* Return the number of page frames in holes in a zone on a node */
4658 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4659 unsigned long zone_type,
4660 unsigned long node_start_pfn,
4661 unsigned long node_end_pfn,
4662 unsigned long *ignored)
4664 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4665 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4666 unsigned long zone_start_pfn, zone_end_pfn;
4668 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4669 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4671 adjust_zone_range_for_zone_movable(nid, zone_type,
4672 node_start_pfn, node_end_pfn,
4673 &zone_start_pfn, &zone_end_pfn);
4674 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4677 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4678 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4679 unsigned long zone_type,
4680 unsigned long node_start_pfn,
4681 unsigned long node_end_pfn,
4682 unsigned long *zones_size)
4684 return zones_size[zone_type];
4687 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4688 unsigned long zone_type,
4689 unsigned long node_start_pfn,
4690 unsigned long node_end_pfn,
4691 unsigned long *zholes_size)
4696 return zholes_size[zone_type];
4699 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4701 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4702 unsigned long node_start_pfn,
4703 unsigned long node_end_pfn,
4704 unsigned long *zones_size,
4705 unsigned long *zholes_size)
4707 unsigned long realtotalpages, totalpages = 0;
4710 for (i = 0; i < MAX_NR_ZONES; i++)
4711 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4715 pgdat->node_spanned_pages = totalpages;
4717 realtotalpages = totalpages;
4718 for (i = 0; i < MAX_NR_ZONES; i++)
4720 zone_absent_pages_in_node(pgdat->node_id, i,
4721 node_start_pfn, node_end_pfn,
4723 pgdat->node_present_pages = realtotalpages;
4724 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4728 #ifndef CONFIG_SPARSEMEM
4730 * Calculate the size of the zone->blockflags rounded to an unsigned long
4731 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4732 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4733 * round what is now in bits to nearest long in bits, then return it in
4736 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4738 unsigned long usemapsize;
4740 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4741 usemapsize = roundup(zonesize, pageblock_nr_pages);
4742 usemapsize = usemapsize >> pageblock_order;
4743 usemapsize *= NR_PAGEBLOCK_BITS;
4744 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4746 return usemapsize / 8;
4749 static void __init setup_usemap(struct pglist_data *pgdat,
4751 unsigned long zone_start_pfn,
4752 unsigned long zonesize)
4754 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4755 zone->pageblock_flags = NULL;
4757 zone->pageblock_flags =
4758 memblock_virt_alloc_node_nopanic(usemapsize,
4762 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4763 unsigned long zone_start_pfn, unsigned long zonesize) {}
4764 #endif /* CONFIG_SPARSEMEM */
4766 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4768 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4769 void __paginginit set_pageblock_order(void)
4773 /* Check that pageblock_nr_pages has not already been setup */
4774 if (pageblock_order)
4777 if (HPAGE_SHIFT > PAGE_SHIFT)
4778 order = HUGETLB_PAGE_ORDER;
4780 order = MAX_ORDER - 1;
4783 * Assume the largest contiguous order of interest is a huge page.
4784 * This value may be variable depending on boot parameters on IA64 and
4787 pageblock_order = order;
4789 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4792 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4793 * is unused as pageblock_order is set at compile-time. See
4794 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4797 void __paginginit set_pageblock_order(void)
4801 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4803 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4804 unsigned long present_pages)
4806 unsigned long pages = spanned_pages;
4809 * Provide a more accurate estimation if there are holes within
4810 * the zone and SPARSEMEM is in use. If there are holes within the
4811 * zone, each populated memory region may cost us one or two extra
4812 * memmap pages due to alignment because memmap pages for each
4813 * populated regions may not naturally algined on page boundary.
4814 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4816 if (spanned_pages > present_pages + (present_pages >> 4) &&
4817 IS_ENABLED(CONFIG_SPARSEMEM))
4818 pages = present_pages;
4820 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4824 * Set up the zone data structures:
4825 * - mark all pages reserved
4826 * - mark all memory queues empty
4827 * - clear the memory bitmaps
4829 * NOTE: pgdat should get zeroed by caller.
4831 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4832 unsigned long node_start_pfn, unsigned long node_end_pfn,
4833 unsigned long *zones_size, unsigned long *zholes_size)
4836 int nid = pgdat->node_id;
4837 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4840 pgdat_resize_init(pgdat);
4841 #ifdef CONFIG_NUMA_BALANCING
4842 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4843 pgdat->numabalancing_migrate_nr_pages = 0;
4844 pgdat->numabalancing_migrate_next_window = jiffies;
4846 init_waitqueue_head(&pgdat->kswapd_wait);
4847 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4848 pgdat_page_ext_init(pgdat);
4850 for (j = 0; j < MAX_NR_ZONES; j++) {
4851 struct zone *zone = pgdat->node_zones + j;
4852 unsigned long size, realsize, freesize, memmap_pages;
4854 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4855 node_end_pfn, zones_size);
4856 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4862 * Adjust freesize so that it accounts for how much memory
4863 * is used by this zone for memmap. This affects the watermark
4864 * and per-cpu initialisations
4866 memmap_pages = calc_memmap_size(size, realsize);
4867 if (!is_highmem_idx(j)) {
4868 if (freesize >= memmap_pages) {
4869 freesize -= memmap_pages;
4872 " %s zone: %lu pages used for memmap\n",
4873 zone_names[j], memmap_pages);
4876 " %s zone: %lu pages exceeds freesize %lu\n",
4877 zone_names[j], memmap_pages, freesize);
4880 /* Account for reserved pages */
4881 if (j == 0 && freesize > dma_reserve) {
4882 freesize -= dma_reserve;
4883 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4884 zone_names[0], dma_reserve);
4887 if (!is_highmem_idx(j))
4888 nr_kernel_pages += freesize;
4889 /* Charge for highmem memmap if there are enough kernel pages */
4890 else if (nr_kernel_pages > memmap_pages * 2)
4891 nr_kernel_pages -= memmap_pages;
4892 nr_all_pages += freesize;
4894 zone->spanned_pages = size;
4895 zone->present_pages = realsize;
4897 * Set an approximate value for lowmem here, it will be adjusted
4898 * when the bootmem allocator frees pages into the buddy system.
4899 * And all highmem pages will be managed by the buddy system.
4901 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4904 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4906 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4908 zone->name = zone_names[j];
4909 spin_lock_init(&zone->lock);
4910 spin_lock_init(&zone->lru_lock);
4911 zone_seqlock_init(zone);
4912 zone->zone_pgdat = pgdat;
4913 zone_pcp_init(zone);
4915 /* For bootup, initialized properly in watermark setup */
4916 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4918 lruvec_init(&zone->lruvec);
4922 set_pageblock_order();
4923 setup_usemap(pgdat, zone, zone_start_pfn, size);
4924 ret = init_currently_empty_zone(zone, zone_start_pfn,
4925 size, MEMMAP_EARLY);
4927 memmap_init(size, nid, j, zone_start_pfn);
4928 zone_start_pfn += size;
4932 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4934 /* Skip empty nodes */
4935 if (!pgdat->node_spanned_pages)
4938 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4939 /* ia64 gets its own node_mem_map, before this, without bootmem */
4940 if (!pgdat->node_mem_map) {
4941 unsigned long size, start, end;
4945 * The zone's endpoints aren't required to be MAX_ORDER
4946 * aligned but the node_mem_map endpoints must be in order
4947 * for the buddy allocator to function correctly.
4949 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4950 end = pgdat_end_pfn(pgdat);
4951 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4952 size = (end - start) * sizeof(struct page);
4953 map = alloc_remap(pgdat->node_id, size);
4955 map = memblock_virt_alloc_node_nopanic(size,
4957 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4959 #ifndef CONFIG_NEED_MULTIPLE_NODES
4961 * With no DISCONTIG, the global mem_map is just set as node 0's
4963 if (pgdat == NODE_DATA(0)) {
4964 mem_map = NODE_DATA(0)->node_mem_map;
4965 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4966 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4967 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4968 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4971 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4974 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4975 unsigned long node_start_pfn, unsigned long *zholes_size)
4977 pg_data_t *pgdat = NODE_DATA(nid);
4978 unsigned long start_pfn = 0;
4979 unsigned long end_pfn = 0;
4981 /* pg_data_t should be reset to zero when it's allocated */
4982 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4984 pgdat->node_id = nid;
4985 pgdat->node_start_pfn = node_start_pfn;
4986 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4987 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4988 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
4989 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
4991 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4992 zones_size, zholes_size);
4994 alloc_node_mem_map(pgdat);
4995 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4996 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4997 nid, (unsigned long)pgdat,
4998 (unsigned long)pgdat->node_mem_map);
5001 free_area_init_core(pgdat, start_pfn, end_pfn,
5002 zones_size, zholes_size);
5005 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5007 #if MAX_NUMNODES > 1
5009 * Figure out the number of possible node ids.
5011 void __init setup_nr_node_ids(void)
5014 unsigned int highest = 0;
5016 for_each_node_mask(node, node_possible_map)
5018 nr_node_ids = highest + 1;
5023 * node_map_pfn_alignment - determine the maximum internode alignment
5025 * This function should be called after node map is populated and sorted.
5026 * It calculates the maximum power of two alignment which can distinguish
5029 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5030 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5031 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5032 * shifted, 1GiB is enough and this function will indicate so.
5034 * This is used to test whether pfn -> nid mapping of the chosen memory
5035 * model has fine enough granularity to avoid incorrect mapping for the
5036 * populated node map.
5038 * Returns the determined alignment in pfn's. 0 if there is no alignment
5039 * requirement (single node).
5041 unsigned long __init node_map_pfn_alignment(void)
5043 unsigned long accl_mask = 0, last_end = 0;
5044 unsigned long start, end, mask;
5048 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5049 if (!start || last_nid < 0 || last_nid == nid) {
5056 * Start with a mask granular enough to pin-point to the
5057 * start pfn and tick off bits one-by-one until it becomes
5058 * too coarse to separate the current node from the last.
5060 mask = ~((1 << __ffs(start)) - 1);
5061 while (mask && last_end <= (start & (mask << 1)))
5064 /* accumulate all internode masks */
5068 /* convert mask to number of pages */
5069 return ~accl_mask + 1;
5072 /* Find the lowest pfn for a node */
5073 static unsigned long __init find_min_pfn_for_node(int nid)
5075 unsigned long min_pfn = ULONG_MAX;
5076 unsigned long start_pfn;
5079 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5080 min_pfn = min(min_pfn, start_pfn);
5082 if (min_pfn == ULONG_MAX) {
5084 "Could not find start_pfn for node %d\n", nid);
5092 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5094 * It returns the minimum PFN based on information provided via
5095 * memblock_set_node().
5097 unsigned long __init find_min_pfn_with_active_regions(void)
5099 return find_min_pfn_for_node(MAX_NUMNODES);
5103 * early_calculate_totalpages()
5104 * Sum pages in active regions for movable zone.
5105 * Populate N_MEMORY for calculating usable_nodes.
5107 static unsigned long __init early_calculate_totalpages(void)
5109 unsigned long totalpages = 0;
5110 unsigned long start_pfn, end_pfn;
5113 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5114 unsigned long pages = end_pfn - start_pfn;
5116 totalpages += pages;
5118 node_set_state(nid, N_MEMORY);
5124 * Find the PFN the Movable zone begins in each node. Kernel memory
5125 * is spread evenly between nodes as long as the nodes have enough
5126 * memory. When they don't, some nodes will have more kernelcore than
5129 static void __init find_zone_movable_pfns_for_nodes(void)
5132 unsigned long usable_startpfn;
5133 unsigned long kernelcore_node, kernelcore_remaining;
5134 /* save the state before borrow the nodemask */
5135 nodemask_t saved_node_state = node_states[N_MEMORY];
5136 unsigned long totalpages = early_calculate_totalpages();
5137 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5138 struct memblock_region *r;
5140 /* Need to find movable_zone earlier when movable_node is specified. */
5141 find_usable_zone_for_movable();
5144 * If movable_node is specified, ignore kernelcore and movablecore
5147 if (movable_node_is_enabled()) {
5148 for_each_memblock(memory, r) {
5149 if (!memblock_is_hotpluggable(r))
5154 usable_startpfn = PFN_DOWN(r->base);
5155 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5156 min(usable_startpfn, zone_movable_pfn[nid]) :
5164 * If movablecore=nn[KMG] was specified, calculate what size of
5165 * kernelcore that corresponds so that memory usable for
5166 * any allocation type is evenly spread. If both kernelcore
5167 * and movablecore are specified, then the value of kernelcore
5168 * will be used for required_kernelcore if it's greater than
5169 * what movablecore would have allowed.
5171 if (required_movablecore) {
5172 unsigned long corepages;
5175 * Round-up so that ZONE_MOVABLE is at least as large as what
5176 * was requested by the user
5178 required_movablecore =
5179 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5180 corepages = totalpages - required_movablecore;
5182 required_kernelcore = max(required_kernelcore, corepages);
5185 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5186 if (!required_kernelcore)
5189 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5190 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5193 /* Spread kernelcore memory as evenly as possible throughout nodes */
5194 kernelcore_node = required_kernelcore / usable_nodes;
5195 for_each_node_state(nid, N_MEMORY) {
5196 unsigned long start_pfn, end_pfn;
5199 * Recalculate kernelcore_node if the division per node
5200 * now exceeds what is necessary to satisfy the requested
5201 * amount of memory for the kernel
5203 if (required_kernelcore < kernelcore_node)
5204 kernelcore_node = required_kernelcore / usable_nodes;
5207 * As the map is walked, we track how much memory is usable
5208 * by the kernel using kernelcore_remaining. When it is
5209 * 0, the rest of the node is usable by ZONE_MOVABLE
5211 kernelcore_remaining = kernelcore_node;
5213 /* Go through each range of PFNs within this node */
5214 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5215 unsigned long size_pages;
5217 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5218 if (start_pfn >= end_pfn)
5221 /* Account for what is only usable for kernelcore */
5222 if (start_pfn < usable_startpfn) {
5223 unsigned long kernel_pages;
5224 kernel_pages = min(end_pfn, usable_startpfn)
5227 kernelcore_remaining -= min(kernel_pages,
5228 kernelcore_remaining);
5229 required_kernelcore -= min(kernel_pages,
5230 required_kernelcore);
5232 /* Continue if range is now fully accounted */
5233 if (end_pfn <= usable_startpfn) {
5236 * Push zone_movable_pfn to the end so
5237 * that if we have to rebalance
5238 * kernelcore across nodes, we will
5239 * not double account here
5241 zone_movable_pfn[nid] = end_pfn;
5244 start_pfn = usable_startpfn;
5248 * The usable PFN range for ZONE_MOVABLE is from
5249 * start_pfn->end_pfn. Calculate size_pages as the
5250 * number of pages used as kernelcore
5252 size_pages = end_pfn - start_pfn;
5253 if (size_pages > kernelcore_remaining)
5254 size_pages = kernelcore_remaining;
5255 zone_movable_pfn[nid] = start_pfn + size_pages;
5258 * Some kernelcore has been met, update counts and
5259 * break if the kernelcore for this node has been
5262 required_kernelcore -= min(required_kernelcore,
5264 kernelcore_remaining -= size_pages;
5265 if (!kernelcore_remaining)
5271 * If there is still required_kernelcore, we do another pass with one
5272 * less node in the count. This will push zone_movable_pfn[nid] further
5273 * along on the nodes that still have memory until kernelcore is
5277 if (usable_nodes && required_kernelcore > usable_nodes)
5281 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5282 for (nid = 0; nid < MAX_NUMNODES; nid++)
5283 zone_movable_pfn[nid] =
5284 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5287 /* restore the node_state */
5288 node_states[N_MEMORY] = saved_node_state;
5291 /* Any regular or high memory on that node ? */
5292 static void check_for_memory(pg_data_t *pgdat, int nid)
5294 enum zone_type zone_type;
5296 if (N_MEMORY == N_NORMAL_MEMORY)
5299 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5300 struct zone *zone = &pgdat->node_zones[zone_type];
5301 if (populated_zone(zone)) {
5302 node_set_state(nid, N_HIGH_MEMORY);
5303 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5304 zone_type <= ZONE_NORMAL)
5305 node_set_state(nid, N_NORMAL_MEMORY);
5312 * free_area_init_nodes - Initialise all pg_data_t and zone data
5313 * @max_zone_pfn: an array of max PFNs for each zone
5315 * This will call free_area_init_node() for each active node in the system.
5316 * Using the page ranges provided by memblock_set_node(), the size of each
5317 * zone in each node and their holes is calculated. If the maximum PFN
5318 * between two adjacent zones match, it is assumed that the zone is empty.
5319 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5320 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5321 * starts where the previous one ended. For example, ZONE_DMA32 starts
5322 * at arch_max_dma_pfn.
5324 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5326 unsigned long start_pfn, end_pfn;
5329 /* Record where the zone boundaries are */
5330 memset(arch_zone_lowest_possible_pfn, 0,
5331 sizeof(arch_zone_lowest_possible_pfn));
5332 memset(arch_zone_highest_possible_pfn, 0,
5333 sizeof(arch_zone_highest_possible_pfn));
5334 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5335 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5336 for (i = 1; i < MAX_NR_ZONES; i++) {
5337 if (i == ZONE_MOVABLE)
5339 arch_zone_lowest_possible_pfn[i] =
5340 arch_zone_highest_possible_pfn[i-1];
5341 arch_zone_highest_possible_pfn[i] =
5342 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5344 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5345 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5347 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5348 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5349 find_zone_movable_pfns_for_nodes();
5351 /* Print out the zone ranges */
5352 pr_info("Zone ranges:\n");
5353 for (i = 0; i < MAX_NR_ZONES; i++) {
5354 if (i == ZONE_MOVABLE)
5356 pr_info(" %-8s ", zone_names[i]);
5357 if (arch_zone_lowest_possible_pfn[i] ==
5358 arch_zone_highest_possible_pfn[i])
5361 pr_cont("[mem %#018Lx-%#018Lx]\n",
5362 (u64)arch_zone_lowest_possible_pfn[i]
5364 ((u64)arch_zone_highest_possible_pfn[i]
5365 << PAGE_SHIFT) - 1);
5368 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5369 pr_info("Movable zone start for each node\n");
5370 for (i = 0; i < MAX_NUMNODES; i++) {
5371 if (zone_movable_pfn[i])
5372 pr_info(" Node %d: %#018Lx\n", i,
5373 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5376 /* Print out the early node map */
5377 pr_info("Early memory node ranges\n");
5378 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5379 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5380 (u64)start_pfn << PAGE_SHIFT,
5381 ((u64)end_pfn << PAGE_SHIFT) - 1);
5383 /* Initialise every node */
5384 mminit_verify_pageflags_layout();
5385 setup_nr_node_ids();
5386 for_each_online_node(nid) {
5387 pg_data_t *pgdat = NODE_DATA(nid);
5388 free_area_init_node(nid, NULL,
5389 find_min_pfn_for_node(nid), NULL);
5391 /* Any memory on that node */
5392 if (pgdat->node_present_pages)
5393 node_set_state(nid, N_MEMORY);
5394 check_for_memory(pgdat, nid);
5398 static int __init cmdline_parse_core(char *p, unsigned long *core)
5400 unsigned long long coremem;
5404 coremem = memparse(p, &p);
5405 *core = coremem >> PAGE_SHIFT;
5407 /* Paranoid check that UL is enough for the coremem value */
5408 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5414 * kernelcore=size sets the amount of memory for use for allocations that
5415 * cannot be reclaimed or migrated.
5417 static int __init cmdline_parse_kernelcore(char *p)
5419 return cmdline_parse_core(p, &required_kernelcore);
5423 * movablecore=size sets the amount of memory for use for allocations that
5424 * can be reclaimed or migrated.
5426 static int __init cmdline_parse_movablecore(char *p)
5428 return cmdline_parse_core(p, &required_movablecore);
5431 early_param("kernelcore", cmdline_parse_kernelcore);
5432 early_param("movablecore", cmdline_parse_movablecore);
5434 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5436 void adjust_managed_page_count(struct page *page, long count)
5438 spin_lock(&managed_page_count_lock);
5439 page_zone(page)->managed_pages += count;
5440 totalram_pages += count;
5441 #ifdef CONFIG_HIGHMEM
5442 if (PageHighMem(page))
5443 totalhigh_pages += count;
5445 spin_unlock(&managed_page_count_lock);
5447 EXPORT_SYMBOL(adjust_managed_page_count);
5449 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5452 unsigned long pages = 0;
5454 start = (void *)PAGE_ALIGN((unsigned long)start);
5455 end = (void *)((unsigned long)end & PAGE_MASK);
5456 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5457 if ((unsigned int)poison <= 0xFF)
5458 memset(pos, poison, PAGE_SIZE);
5459 free_reserved_page(virt_to_page(pos));
5463 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5464 s, pages << (PAGE_SHIFT - 10), start, end);
5468 EXPORT_SYMBOL(free_reserved_area);
5470 #ifdef CONFIG_HIGHMEM
5471 void free_highmem_page(struct page *page)
5473 __free_reserved_page(page);
5475 page_zone(page)->managed_pages++;
5481 void __init mem_init_print_info(const char *str)
5483 unsigned long physpages, codesize, datasize, rosize, bss_size;
5484 unsigned long init_code_size, init_data_size;
5486 physpages = get_num_physpages();
5487 codesize = _etext - _stext;
5488 datasize = _edata - _sdata;
5489 rosize = __end_rodata - __start_rodata;
5490 bss_size = __bss_stop - __bss_start;
5491 init_data_size = __init_end - __init_begin;
5492 init_code_size = _einittext - _sinittext;
5495 * Detect special cases and adjust section sizes accordingly:
5496 * 1) .init.* may be embedded into .data sections
5497 * 2) .init.text.* may be out of [__init_begin, __init_end],
5498 * please refer to arch/tile/kernel/vmlinux.lds.S.
5499 * 3) .rodata.* may be embedded into .text or .data sections.
5501 #define adj_init_size(start, end, size, pos, adj) \
5503 if (start <= pos && pos < end && size > adj) \
5507 adj_init_size(__init_begin, __init_end, init_data_size,
5508 _sinittext, init_code_size);
5509 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5510 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5511 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5512 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5514 #undef adj_init_size
5516 pr_info("Memory: %luK/%luK available "
5517 "(%luK kernel code, %luK rwdata, %luK rodata, "
5518 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5519 #ifdef CONFIG_HIGHMEM
5523 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5524 codesize >> 10, datasize >> 10, rosize >> 10,
5525 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5526 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5527 totalcma_pages << (PAGE_SHIFT-10),
5528 #ifdef CONFIG_HIGHMEM
5529 totalhigh_pages << (PAGE_SHIFT-10),
5531 str ? ", " : "", str ? str : "");
5535 * set_dma_reserve - set the specified number of pages reserved in the first zone
5536 * @new_dma_reserve: The number of pages to mark reserved
5538 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5539 * In the DMA zone, a significant percentage may be consumed by kernel image
5540 * and other unfreeable allocations which can skew the watermarks badly. This
5541 * function may optionally be used to account for unfreeable pages in the
5542 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5543 * smaller per-cpu batchsize.
5545 void __init set_dma_reserve(unsigned long new_dma_reserve)
5547 dma_reserve = new_dma_reserve;
5550 void __init free_area_init(unsigned long *zones_size)
5552 free_area_init_node(0, zones_size,
5553 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5556 static int page_alloc_cpu_notify(struct notifier_block *self,
5557 unsigned long action, void *hcpu)
5559 int cpu = (unsigned long)hcpu;
5561 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5562 lru_add_drain_cpu(cpu);
5566 * Spill the event counters of the dead processor
5567 * into the current processors event counters.
5568 * This artificially elevates the count of the current
5571 vm_events_fold_cpu(cpu);
5574 * Zero the differential counters of the dead processor
5575 * so that the vm statistics are consistent.
5577 * This is only okay since the processor is dead and cannot
5578 * race with what we are doing.
5580 cpu_vm_stats_fold(cpu);
5585 void __init page_alloc_init(void)
5587 hotcpu_notifier(page_alloc_cpu_notify, 0);
5591 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5592 * or min_free_kbytes changes.
5594 static void calculate_totalreserve_pages(void)
5596 struct pglist_data *pgdat;
5597 unsigned long reserve_pages = 0;
5598 enum zone_type i, j;
5600 for_each_online_pgdat(pgdat) {
5601 for (i = 0; i < MAX_NR_ZONES; i++) {
5602 struct zone *zone = pgdat->node_zones + i;
5605 /* Find valid and maximum lowmem_reserve in the zone */
5606 for (j = i; j < MAX_NR_ZONES; j++) {
5607 if (zone->lowmem_reserve[j] > max)
5608 max = zone->lowmem_reserve[j];
5611 /* we treat the high watermark as reserved pages. */
5612 max += high_wmark_pages(zone);
5614 if (max > zone->managed_pages)
5615 max = zone->managed_pages;
5616 reserve_pages += max;
5618 * Lowmem reserves are not available to
5619 * GFP_HIGHUSER page cache allocations and
5620 * kswapd tries to balance zones to their high
5621 * watermark. As a result, neither should be
5622 * regarded as dirtyable memory, to prevent a
5623 * situation where reclaim has to clean pages
5624 * in order to balance the zones.
5626 zone->dirty_balance_reserve = max;
5629 dirty_balance_reserve = reserve_pages;
5630 totalreserve_pages = reserve_pages;
5634 * setup_per_zone_lowmem_reserve - called whenever
5635 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5636 * has a correct pages reserved value, so an adequate number of
5637 * pages are left in the zone after a successful __alloc_pages().
5639 static void setup_per_zone_lowmem_reserve(void)
5641 struct pglist_data *pgdat;
5642 enum zone_type j, idx;
5644 for_each_online_pgdat(pgdat) {
5645 for (j = 0; j < MAX_NR_ZONES; j++) {
5646 struct zone *zone = pgdat->node_zones + j;
5647 unsigned long managed_pages = zone->managed_pages;
5649 zone->lowmem_reserve[j] = 0;
5653 struct zone *lower_zone;
5657 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5658 sysctl_lowmem_reserve_ratio[idx] = 1;
5660 lower_zone = pgdat->node_zones + idx;
5661 lower_zone->lowmem_reserve[j] = managed_pages /
5662 sysctl_lowmem_reserve_ratio[idx];
5663 managed_pages += lower_zone->managed_pages;
5668 /* update totalreserve_pages */
5669 calculate_totalreserve_pages();
5672 static void __setup_per_zone_wmarks(void)
5674 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5675 unsigned long lowmem_pages = 0;
5677 unsigned long flags;
5679 /* Calculate total number of !ZONE_HIGHMEM pages */
5680 for_each_zone(zone) {
5681 if (!is_highmem(zone))
5682 lowmem_pages += zone->managed_pages;
5685 for_each_zone(zone) {
5688 spin_lock_irqsave(&zone->lock, flags);
5689 tmp = (u64)pages_min * zone->managed_pages;
5690 do_div(tmp, lowmem_pages);
5691 if (is_highmem(zone)) {
5693 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5694 * need highmem pages, so cap pages_min to a small
5697 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5698 * deltas controls asynch page reclaim, and so should
5699 * not be capped for highmem.
5701 unsigned long min_pages;
5703 min_pages = zone->managed_pages / 1024;
5704 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5705 zone->watermark[WMARK_MIN] = min_pages;
5708 * If it's a lowmem zone, reserve a number of pages
5709 * proportionate to the zone's size.
5711 zone->watermark[WMARK_MIN] = tmp;
5714 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5715 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5717 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5718 high_wmark_pages(zone) - low_wmark_pages(zone) -
5719 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5721 setup_zone_migrate_reserve(zone);
5722 spin_unlock_irqrestore(&zone->lock, flags);
5725 /* update totalreserve_pages */
5726 calculate_totalreserve_pages();
5730 * setup_per_zone_wmarks - called when min_free_kbytes changes
5731 * or when memory is hot-{added|removed}
5733 * Ensures that the watermark[min,low,high] values for each zone are set
5734 * correctly with respect to min_free_kbytes.
5736 void setup_per_zone_wmarks(void)
5738 mutex_lock(&zonelists_mutex);
5739 __setup_per_zone_wmarks();
5740 mutex_unlock(&zonelists_mutex);
5744 * The inactive anon list should be small enough that the VM never has to
5745 * do too much work, but large enough that each inactive page has a chance
5746 * to be referenced again before it is swapped out.
5748 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5749 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5750 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5751 * the anonymous pages are kept on the inactive list.
5754 * memory ratio inactive anon
5755 * -------------------------------------
5764 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5766 unsigned int gb, ratio;
5768 /* Zone size in gigabytes */
5769 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5771 ratio = int_sqrt(10 * gb);
5775 zone->inactive_ratio = ratio;
5778 static void __meminit setup_per_zone_inactive_ratio(void)
5783 calculate_zone_inactive_ratio(zone);
5787 * Initialise min_free_kbytes.
5789 * For small machines we want it small (128k min). For large machines
5790 * we want it large (64MB max). But it is not linear, because network
5791 * bandwidth does not increase linearly with machine size. We use
5793 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5794 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5810 int __meminit init_per_zone_wmark_min(void)
5812 unsigned long lowmem_kbytes;
5813 int new_min_free_kbytes;
5815 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5816 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5818 if (new_min_free_kbytes > user_min_free_kbytes) {
5819 min_free_kbytes = new_min_free_kbytes;
5820 if (min_free_kbytes < 128)
5821 min_free_kbytes = 128;
5822 if (min_free_kbytes > 65536)
5823 min_free_kbytes = 65536;
5825 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5826 new_min_free_kbytes, user_min_free_kbytes);
5828 setup_per_zone_wmarks();
5829 refresh_zone_stat_thresholds();
5830 setup_per_zone_lowmem_reserve();
5831 setup_per_zone_inactive_ratio();
5834 module_init(init_per_zone_wmark_min)
5837 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5838 * that we can call two helper functions whenever min_free_kbytes
5841 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5842 void __user *buffer, size_t *length, loff_t *ppos)
5846 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5851 user_min_free_kbytes = min_free_kbytes;
5852 setup_per_zone_wmarks();
5858 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5859 void __user *buffer, size_t *length, loff_t *ppos)
5864 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5869 zone->min_unmapped_pages = (zone->managed_pages *
5870 sysctl_min_unmapped_ratio) / 100;
5874 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5875 void __user *buffer, size_t *length, loff_t *ppos)
5880 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5885 zone->min_slab_pages = (zone->managed_pages *
5886 sysctl_min_slab_ratio) / 100;
5892 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5893 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5894 * whenever sysctl_lowmem_reserve_ratio changes.
5896 * The reserve ratio obviously has absolutely no relation with the
5897 * minimum watermarks. The lowmem reserve ratio can only make sense
5898 * if in function of the boot time zone sizes.
5900 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5901 void __user *buffer, size_t *length, loff_t *ppos)
5903 proc_dointvec_minmax(table, write, buffer, length, ppos);
5904 setup_per_zone_lowmem_reserve();
5909 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5910 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5911 * pagelist can have before it gets flushed back to buddy allocator.
5913 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5914 void __user *buffer, size_t *length, loff_t *ppos)
5917 int old_percpu_pagelist_fraction;
5920 mutex_lock(&pcp_batch_high_lock);
5921 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5923 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5924 if (!write || ret < 0)
5927 /* Sanity checking to avoid pcp imbalance */
5928 if (percpu_pagelist_fraction &&
5929 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
5930 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
5936 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
5939 for_each_populated_zone(zone) {
5942 for_each_possible_cpu(cpu)
5943 pageset_set_high_and_batch(zone,
5944 per_cpu_ptr(zone->pageset, cpu));
5947 mutex_unlock(&pcp_batch_high_lock);
5951 int hashdist = HASHDIST_DEFAULT;
5954 static int __init set_hashdist(char *str)
5958 hashdist = simple_strtoul(str, &str, 0);
5961 __setup("hashdist=", set_hashdist);
5965 * allocate a large system hash table from bootmem
5966 * - it is assumed that the hash table must contain an exact power-of-2
5967 * quantity of entries
5968 * - limit is the number of hash buckets, not the total allocation size
5970 void *__init alloc_large_system_hash(const char *tablename,
5971 unsigned long bucketsize,
5972 unsigned long numentries,
5975 unsigned int *_hash_shift,
5976 unsigned int *_hash_mask,
5977 unsigned long low_limit,
5978 unsigned long high_limit)
5980 unsigned long long max = high_limit;
5981 unsigned long log2qty, size;
5984 /* allow the kernel cmdline to have a say */
5986 /* round applicable memory size up to nearest megabyte */
5987 numentries = nr_kernel_pages;
5989 /* It isn't necessary when PAGE_SIZE >= 1MB */
5990 if (PAGE_SHIFT < 20)
5991 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5993 /* limit to 1 bucket per 2^scale bytes of low memory */
5994 if (scale > PAGE_SHIFT)
5995 numentries >>= (scale - PAGE_SHIFT);
5997 numentries <<= (PAGE_SHIFT - scale);
5999 /* Make sure we've got at least a 0-order allocation.. */
6000 if (unlikely(flags & HASH_SMALL)) {
6001 /* Makes no sense without HASH_EARLY */
6002 WARN_ON(!(flags & HASH_EARLY));
6003 if (!(numentries >> *_hash_shift)) {
6004 numentries = 1UL << *_hash_shift;
6005 BUG_ON(!numentries);
6007 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6008 numentries = PAGE_SIZE / bucketsize;
6010 numentries = roundup_pow_of_two(numentries);
6012 /* limit allocation size to 1/16 total memory by default */
6014 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6015 do_div(max, bucketsize);
6017 max = min(max, 0x80000000ULL);
6019 if (numentries < low_limit)
6020 numentries = low_limit;
6021 if (numentries > max)
6024 log2qty = ilog2(numentries);
6027 size = bucketsize << log2qty;
6028 if (flags & HASH_EARLY)
6029 table = memblock_virt_alloc_nopanic(size, 0);
6031 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6034 * If bucketsize is not a power-of-two, we may free
6035 * some pages at the end of hash table which
6036 * alloc_pages_exact() automatically does
6038 if (get_order(size) < MAX_ORDER) {
6039 table = alloc_pages_exact(size, GFP_ATOMIC);
6040 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6043 } while (!table && size > PAGE_SIZE && --log2qty);
6046 panic("Failed to allocate %s hash table\n", tablename);
6048 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6051 ilog2(size) - PAGE_SHIFT,
6055 *_hash_shift = log2qty;
6057 *_hash_mask = (1 << log2qty) - 1;
6062 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6063 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6066 #ifdef CONFIG_SPARSEMEM
6067 return __pfn_to_section(pfn)->pageblock_flags;
6069 return zone->pageblock_flags;
6070 #endif /* CONFIG_SPARSEMEM */
6073 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6075 #ifdef CONFIG_SPARSEMEM
6076 pfn &= (PAGES_PER_SECTION-1);
6077 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6079 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6080 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6081 #endif /* CONFIG_SPARSEMEM */
6085 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6086 * @page: The page within the block of interest
6087 * @pfn: The target page frame number
6088 * @end_bitidx: The last bit of interest to retrieve
6089 * @mask: mask of bits that the caller is interested in
6091 * Return: pageblock_bits flags
6093 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6094 unsigned long end_bitidx,
6098 unsigned long *bitmap;
6099 unsigned long bitidx, word_bitidx;
6102 zone = page_zone(page);
6103 bitmap = get_pageblock_bitmap(zone, pfn);
6104 bitidx = pfn_to_bitidx(zone, pfn);
6105 word_bitidx = bitidx / BITS_PER_LONG;
6106 bitidx &= (BITS_PER_LONG-1);
6108 word = bitmap[word_bitidx];
6109 bitidx += end_bitidx;
6110 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6114 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6115 * @page: The page within the block of interest
6116 * @flags: The flags to set
6117 * @pfn: The target page frame number
6118 * @end_bitidx: The last bit of interest
6119 * @mask: mask of bits that the caller is interested in
6121 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6123 unsigned long end_bitidx,
6127 unsigned long *bitmap;
6128 unsigned long bitidx, word_bitidx;
6129 unsigned long old_word, word;
6131 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6133 zone = page_zone(page);
6134 bitmap = get_pageblock_bitmap(zone, pfn);
6135 bitidx = pfn_to_bitidx(zone, pfn);
6136 word_bitidx = bitidx / BITS_PER_LONG;
6137 bitidx &= (BITS_PER_LONG-1);
6139 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6141 bitidx += end_bitidx;
6142 mask <<= (BITS_PER_LONG - bitidx - 1);
6143 flags <<= (BITS_PER_LONG - bitidx - 1);
6145 word = ACCESS_ONCE(bitmap[word_bitidx]);
6147 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6148 if (word == old_word)
6155 * This function checks whether pageblock includes unmovable pages or not.
6156 * If @count is not zero, it is okay to include less @count unmovable pages
6158 * PageLRU check without isolation or lru_lock could race so that
6159 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6160 * expect this function should be exact.
6162 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6163 bool skip_hwpoisoned_pages)
6165 unsigned long pfn, iter, found;
6169 * For avoiding noise data, lru_add_drain_all() should be called
6170 * If ZONE_MOVABLE, the zone never contains unmovable pages
6172 if (zone_idx(zone) == ZONE_MOVABLE)
6174 mt = get_pageblock_migratetype(page);
6175 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6178 pfn = page_to_pfn(page);
6179 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6180 unsigned long check = pfn + iter;
6182 if (!pfn_valid_within(check))
6185 page = pfn_to_page(check);
6188 * Hugepages are not in LRU lists, but they're movable.
6189 * We need not scan over tail pages bacause we don't
6190 * handle each tail page individually in migration.
6192 if (PageHuge(page)) {
6193 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6198 * We can't use page_count without pin a page
6199 * because another CPU can free compound page.
6200 * This check already skips compound tails of THP
6201 * because their page->_count is zero at all time.
6203 if (!atomic_read(&page->_count)) {
6204 if (PageBuddy(page))
6205 iter += (1 << page_order(page)) - 1;
6210 * The HWPoisoned page may be not in buddy system, and
6211 * page_count() is not 0.
6213 if (skip_hwpoisoned_pages && PageHWPoison(page))
6219 * If there are RECLAIMABLE pages, we need to check
6220 * it. But now, memory offline itself doesn't call
6221 * shrink_node_slabs() and it still to be fixed.
6224 * If the page is not RAM, page_count()should be 0.
6225 * we don't need more check. This is an _used_ not-movable page.
6227 * The problematic thing here is PG_reserved pages. PG_reserved
6228 * is set to both of a memory hole page and a _used_ kernel
6237 bool is_pageblock_removable_nolock(struct page *page)
6243 * We have to be careful here because we are iterating over memory
6244 * sections which are not zone aware so we might end up outside of
6245 * the zone but still within the section.
6246 * We have to take care about the node as well. If the node is offline
6247 * its NODE_DATA will be NULL - see page_zone.
6249 if (!node_online(page_to_nid(page)))
6252 zone = page_zone(page);
6253 pfn = page_to_pfn(page);
6254 if (!zone_spans_pfn(zone, pfn))
6257 return !has_unmovable_pages(zone, page, 0, true);
6262 static unsigned long pfn_max_align_down(unsigned long pfn)
6264 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6265 pageblock_nr_pages) - 1);
6268 static unsigned long pfn_max_align_up(unsigned long pfn)
6270 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6271 pageblock_nr_pages));
6274 /* [start, end) must belong to a single zone. */
6275 static int __alloc_contig_migrate_range(struct compact_control *cc,
6276 unsigned long start, unsigned long end)
6278 /* This function is based on compact_zone() from compaction.c. */
6279 unsigned long nr_reclaimed;
6280 unsigned long pfn = start;
6281 unsigned int tries = 0;
6286 while (pfn < end || !list_empty(&cc->migratepages)) {
6287 if (fatal_signal_pending(current)) {
6292 if (list_empty(&cc->migratepages)) {
6293 cc->nr_migratepages = 0;
6294 pfn = isolate_migratepages_range(cc, pfn, end);
6300 } else if (++tries == 5) {
6301 ret = ret < 0 ? ret : -EBUSY;
6305 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6307 cc->nr_migratepages -= nr_reclaimed;
6309 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6310 NULL, 0, cc->mode, MR_CMA);
6313 putback_movable_pages(&cc->migratepages);
6320 * alloc_contig_range() -- tries to allocate given range of pages
6321 * @start: start PFN to allocate
6322 * @end: one-past-the-last PFN to allocate
6323 * @migratetype: migratetype of the underlaying pageblocks (either
6324 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6325 * in range must have the same migratetype and it must
6326 * be either of the two.
6328 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6329 * aligned, however it's the caller's responsibility to guarantee that
6330 * we are the only thread that changes migrate type of pageblocks the
6333 * The PFN range must belong to a single zone.
6335 * Returns zero on success or negative error code. On success all
6336 * pages which PFN is in [start, end) are allocated for the caller and
6337 * need to be freed with free_contig_range().
6339 int alloc_contig_range(unsigned long start, unsigned long end,
6340 unsigned migratetype)
6342 unsigned long outer_start, outer_end;
6345 struct compact_control cc = {
6346 .nr_migratepages = 0,
6348 .zone = page_zone(pfn_to_page(start)),
6349 .mode = MIGRATE_SYNC,
6350 .ignore_skip_hint = true,
6352 INIT_LIST_HEAD(&cc.migratepages);
6355 * What we do here is we mark all pageblocks in range as
6356 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6357 * have different sizes, and due to the way page allocator
6358 * work, we align the range to biggest of the two pages so
6359 * that page allocator won't try to merge buddies from
6360 * different pageblocks and change MIGRATE_ISOLATE to some
6361 * other migration type.
6363 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6364 * migrate the pages from an unaligned range (ie. pages that
6365 * we are interested in). This will put all the pages in
6366 * range back to page allocator as MIGRATE_ISOLATE.
6368 * When this is done, we take the pages in range from page
6369 * allocator removing them from the buddy system. This way
6370 * page allocator will never consider using them.
6372 * This lets us mark the pageblocks back as
6373 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6374 * aligned range but not in the unaligned, original range are
6375 * put back to page allocator so that buddy can use them.
6378 ret = start_isolate_page_range(pfn_max_align_down(start),
6379 pfn_max_align_up(end), migratetype,
6384 ret = __alloc_contig_migrate_range(&cc, start, end);
6389 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6390 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6391 * more, all pages in [start, end) are free in page allocator.
6392 * What we are going to do is to allocate all pages from
6393 * [start, end) (that is remove them from page allocator).
6395 * The only problem is that pages at the beginning and at the
6396 * end of interesting range may be not aligned with pages that
6397 * page allocator holds, ie. they can be part of higher order
6398 * pages. Because of this, we reserve the bigger range and
6399 * once this is done free the pages we are not interested in.
6401 * We don't have to hold zone->lock here because the pages are
6402 * isolated thus they won't get removed from buddy.
6405 lru_add_drain_all();
6406 drain_all_pages(cc.zone);
6409 outer_start = start;
6410 while (!PageBuddy(pfn_to_page(outer_start))) {
6411 if (++order >= MAX_ORDER) {
6415 outer_start &= ~0UL << order;
6418 /* Make sure the range is really isolated. */
6419 if (test_pages_isolated(outer_start, end, false)) {
6420 pr_info("%s: [%lx, %lx) PFNs busy\n",
6421 __func__, outer_start, end);
6426 /* Grab isolated pages from freelists. */
6427 outer_end = isolate_freepages_range(&cc, outer_start, end);
6433 /* Free head and tail (if any) */
6434 if (start != outer_start)
6435 free_contig_range(outer_start, start - outer_start);
6436 if (end != outer_end)
6437 free_contig_range(end, outer_end - end);
6440 undo_isolate_page_range(pfn_max_align_down(start),
6441 pfn_max_align_up(end), migratetype);
6445 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6447 unsigned int count = 0;
6449 for (; nr_pages--; pfn++) {
6450 struct page *page = pfn_to_page(pfn);
6452 count += page_count(page) != 1;
6455 WARN(count != 0, "%d pages are still in use!\n", count);
6459 #ifdef CONFIG_MEMORY_HOTPLUG
6461 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6462 * page high values need to be recalulated.
6464 void __meminit zone_pcp_update(struct zone *zone)
6467 mutex_lock(&pcp_batch_high_lock);
6468 for_each_possible_cpu(cpu)
6469 pageset_set_high_and_batch(zone,
6470 per_cpu_ptr(zone->pageset, cpu));
6471 mutex_unlock(&pcp_batch_high_lock);
6475 void zone_pcp_reset(struct zone *zone)
6477 unsigned long flags;
6479 struct per_cpu_pageset *pset;
6481 /* avoid races with drain_pages() */
6482 local_irq_save(flags);
6483 if (zone->pageset != &boot_pageset) {
6484 for_each_online_cpu(cpu) {
6485 pset = per_cpu_ptr(zone->pageset, cpu);
6486 drain_zonestat(zone, pset);
6488 free_percpu(zone->pageset);
6489 zone->pageset = &boot_pageset;
6491 local_irq_restore(flags);
6494 #ifdef CONFIG_MEMORY_HOTREMOVE
6496 * All pages in the range must be isolated before calling this.
6499 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6503 unsigned int order, i;
6505 unsigned long flags;
6506 /* find the first valid pfn */
6507 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6512 zone = page_zone(pfn_to_page(pfn));
6513 spin_lock_irqsave(&zone->lock, flags);
6515 while (pfn < end_pfn) {
6516 if (!pfn_valid(pfn)) {
6520 page = pfn_to_page(pfn);
6522 * The HWPoisoned page may be not in buddy system, and
6523 * page_count() is not 0.
6525 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6527 SetPageReserved(page);
6531 BUG_ON(page_count(page));
6532 BUG_ON(!PageBuddy(page));
6533 order = page_order(page);
6534 #ifdef CONFIG_DEBUG_VM
6535 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6536 pfn, 1 << order, end_pfn);
6538 list_del(&page->lru);
6539 rmv_page_order(page);
6540 zone->free_area[order].nr_free--;
6541 for (i = 0; i < (1 << order); i++)
6542 SetPageReserved((page+i));
6543 pfn += (1 << order);
6545 spin_unlock_irqrestore(&zone->lock, flags);
6549 #ifdef CONFIG_MEMORY_FAILURE
6550 bool is_free_buddy_page(struct page *page)
6552 struct zone *zone = page_zone(page);
6553 unsigned long pfn = page_to_pfn(page);
6554 unsigned long flags;
6557 spin_lock_irqsave(&zone->lock, flags);
6558 for (order = 0; order < MAX_ORDER; order++) {
6559 struct page *page_head = page - (pfn & ((1 << order) - 1));
6561 if (PageBuddy(page_head) && page_order(page_head) >= order)
6564 spin_unlock_irqrestore(&zone->lock, flags);
6566 return order < MAX_ORDER;