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 bool oom_killer_disabled __read_mostly;
249 #ifdef CONFIG_DEBUG_VM
250 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
254 unsigned long pfn = page_to_pfn(page);
255 unsigned long sp, start_pfn;
258 seq = zone_span_seqbegin(zone);
259 start_pfn = zone->zone_start_pfn;
260 sp = zone->spanned_pages;
261 if (!zone_spans_pfn(zone, pfn))
263 } while (zone_span_seqretry(zone, seq));
266 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
267 pfn, zone_to_nid(zone), zone->name,
268 start_pfn, start_pfn + sp);
273 static int page_is_consistent(struct zone *zone, struct page *page)
275 if (!pfn_valid_within(page_to_pfn(page)))
277 if (zone != page_zone(page))
283 * Temporary debugging check for pages not lying within a given zone.
285 static int bad_range(struct zone *zone, struct page *page)
287 if (page_outside_zone_boundaries(zone, page))
289 if (!page_is_consistent(zone, page))
295 static inline int bad_range(struct zone *zone, struct page *page)
301 static void bad_page(struct page *page, const char *reason,
302 unsigned long bad_flags)
304 static unsigned long resume;
305 static unsigned long nr_shown;
306 static unsigned long nr_unshown;
308 /* Don't complain about poisoned pages */
309 if (PageHWPoison(page)) {
310 page_mapcount_reset(page); /* remove PageBuddy */
315 * Allow a burst of 60 reports, then keep quiet for that minute;
316 * or allow a steady drip of one report per second.
318 if (nr_shown == 60) {
319 if (time_before(jiffies, resume)) {
325 "BUG: Bad page state: %lu messages suppressed\n",
332 resume = jiffies + 60 * HZ;
334 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
335 current->comm, page_to_pfn(page));
336 dump_page_badflags(page, reason, bad_flags);
341 /* Leave bad fields for debug, except PageBuddy could make trouble */
342 page_mapcount_reset(page); /* remove PageBuddy */
343 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
347 * Higher-order pages are called "compound pages". They are structured thusly:
349 * The first PAGE_SIZE page is called the "head page".
351 * The remaining PAGE_SIZE pages are called "tail pages".
353 * All pages have PG_compound set. All tail pages have their ->first_page
354 * pointing at the head page.
356 * The first tail page's ->lru.next holds the address of the compound page's
357 * put_page() function. Its ->lru.prev holds the order of allocation.
358 * This usage means that zero-order pages may not be compound.
361 static void free_compound_page(struct page *page)
363 __free_pages_ok(page, compound_order(page));
366 void prep_compound_page(struct page *page, unsigned long order)
369 int nr_pages = 1 << order;
371 set_compound_page_dtor(page, free_compound_page);
372 set_compound_order(page, order);
374 for (i = 1; i < nr_pages; i++) {
375 struct page *p = page + i;
376 set_page_count(p, 0);
377 p->first_page = page;
378 /* Make sure p->first_page is always valid for PageTail() */
384 static inline void prep_zero_page(struct page *page, unsigned int order,
390 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
391 * and __GFP_HIGHMEM from hard or soft interrupt context.
393 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
394 for (i = 0; i < (1 << order); i++)
395 clear_highpage(page + i);
398 #ifdef CONFIG_DEBUG_PAGEALLOC
399 unsigned int _debug_guardpage_minorder;
400 bool _debug_pagealloc_enabled __read_mostly;
401 bool _debug_guardpage_enabled __read_mostly;
403 static int __init early_debug_pagealloc(char *buf)
408 if (strcmp(buf, "on") == 0)
409 _debug_pagealloc_enabled = true;
413 early_param("debug_pagealloc", early_debug_pagealloc);
415 static bool need_debug_guardpage(void)
417 /* If we don't use debug_pagealloc, we don't need guard page */
418 if (!debug_pagealloc_enabled())
424 static void init_debug_guardpage(void)
426 if (!debug_pagealloc_enabled())
429 _debug_guardpage_enabled = true;
432 struct page_ext_operations debug_guardpage_ops = {
433 .need = need_debug_guardpage,
434 .init = init_debug_guardpage,
437 static int __init debug_guardpage_minorder_setup(char *buf)
441 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
442 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
445 _debug_guardpage_minorder = res;
446 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
449 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
451 static inline void set_page_guard(struct zone *zone, struct page *page,
452 unsigned int order, int migratetype)
454 struct page_ext *page_ext;
456 if (!debug_guardpage_enabled())
459 page_ext = lookup_page_ext(page);
460 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
462 INIT_LIST_HEAD(&page->lru);
463 set_page_private(page, order);
464 /* Guard pages are not available for any usage */
465 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
468 static inline void clear_page_guard(struct zone *zone, struct page *page,
469 unsigned int order, int migratetype)
471 struct page_ext *page_ext;
473 if (!debug_guardpage_enabled())
476 page_ext = lookup_page_ext(page);
477 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
479 set_page_private(page, 0);
480 if (!is_migrate_isolate(migratetype))
481 __mod_zone_freepage_state(zone, (1 << order), migratetype);
484 struct page_ext_operations debug_guardpage_ops = { NULL, };
485 static inline void set_page_guard(struct zone *zone, struct page *page,
486 unsigned int order, int migratetype) {}
487 static inline void clear_page_guard(struct zone *zone, struct page *page,
488 unsigned int order, int migratetype) {}
491 static inline void set_page_order(struct page *page, unsigned int order)
493 set_page_private(page, order);
494 __SetPageBuddy(page);
497 static inline void rmv_page_order(struct page *page)
499 __ClearPageBuddy(page);
500 set_page_private(page, 0);
504 * This function checks whether a page is free && is the buddy
505 * we can do coalesce a page and its buddy if
506 * (a) the buddy is not in a hole &&
507 * (b) the buddy is in the buddy system &&
508 * (c) a page and its buddy have the same order &&
509 * (d) a page and its buddy are in the same zone.
511 * For recording whether a page is in the buddy system, we set ->_mapcount
512 * PAGE_BUDDY_MAPCOUNT_VALUE.
513 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
514 * serialized by zone->lock.
516 * For recording page's order, we use page_private(page).
518 static inline int page_is_buddy(struct page *page, struct page *buddy,
521 if (!pfn_valid_within(page_to_pfn(buddy)))
524 if (page_is_guard(buddy) && page_order(buddy) == order) {
525 if (page_zone_id(page) != page_zone_id(buddy))
528 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
533 if (PageBuddy(buddy) && page_order(buddy) == order) {
535 * zone check is done late to avoid uselessly
536 * calculating zone/node ids for pages that could
539 if (page_zone_id(page) != page_zone_id(buddy))
542 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
550 * Freeing function for a buddy system allocator.
552 * The concept of a buddy system is to maintain direct-mapped table
553 * (containing bit values) for memory blocks of various "orders".
554 * The bottom level table contains the map for the smallest allocatable
555 * units of memory (here, pages), and each level above it describes
556 * pairs of units from the levels below, hence, "buddies".
557 * At a high level, all that happens here is marking the table entry
558 * at the bottom level available, and propagating the changes upward
559 * as necessary, plus some accounting needed to play nicely with other
560 * parts of the VM system.
561 * At each level, we keep a list of pages, which are heads of continuous
562 * free pages of length of (1 << order) and marked with _mapcount
563 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
565 * So when we are allocating or freeing one, we can derive the state of the
566 * other. That is, if we allocate a small block, and both were
567 * free, the remainder of the region must be split into blocks.
568 * If a block is freed, and its buddy is also free, then this
569 * triggers coalescing into a block of larger size.
574 static inline void __free_one_page(struct page *page,
576 struct zone *zone, unsigned int order,
579 unsigned long page_idx;
580 unsigned long combined_idx;
581 unsigned long uninitialized_var(buddy_idx);
583 int max_order = MAX_ORDER;
585 VM_BUG_ON(!zone_is_initialized(zone));
586 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
588 VM_BUG_ON(migratetype == -1);
589 if (is_migrate_isolate(migratetype)) {
591 * We restrict max order of merging to prevent merge
592 * between freepages on isolate pageblock and normal
593 * pageblock. Without this, pageblock isolation
594 * could cause incorrect freepage accounting.
596 max_order = min(MAX_ORDER, pageblock_order + 1);
598 __mod_zone_freepage_state(zone, 1 << order, migratetype);
601 page_idx = pfn & ((1 << max_order) - 1);
603 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
604 VM_BUG_ON_PAGE(bad_range(zone, page), page);
606 while (order < max_order - 1) {
607 buddy_idx = __find_buddy_index(page_idx, order);
608 buddy = page + (buddy_idx - page_idx);
609 if (!page_is_buddy(page, buddy, order))
612 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
613 * merge with it and move up one order.
615 if (page_is_guard(buddy)) {
616 clear_page_guard(zone, buddy, order, migratetype);
618 list_del(&buddy->lru);
619 zone->free_area[order].nr_free--;
620 rmv_page_order(buddy);
622 combined_idx = buddy_idx & page_idx;
623 page = page + (combined_idx - page_idx);
624 page_idx = combined_idx;
627 set_page_order(page, order);
630 * If this is not the largest possible page, check if the buddy
631 * of the next-highest order is free. If it is, it's possible
632 * that pages are being freed that will coalesce soon. In case,
633 * that is happening, add the free page to the tail of the list
634 * so it's less likely to be used soon and more likely to be merged
635 * as a higher order page
637 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
638 struct page *higher_page, *higher_buddy;
639 combined_idx = buddy_idx & page_idx;
640 higher_page = page + (combined_idx - page_idx);
641 buddy_idx = __find_buddy_index(combined_idx, order + 1);
642 higher_buddy = higher_page + (buddy_idx - combined_idx);
643 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
644 list_add_tail(&page->lru,
645 &zone->free_area[order].free_list[migratetype]);
650 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
652 zone->free_area[order].nr_free++;
655 static inline int free_pages_check(struct page *page)
657 const char *bad_reason = NULL;
658 unsigned long bad_flags = 0;
660 if (unlikely(page_mapcount(page)))
661 bad_reason = "nonzero mapcount";
662 if (unlikely(page->mapping != NULL))
663 bad_reason = "non-NULL mapping";
664 if (unlikely(atomic_read(&page->_count) != 0))
665 bad_reason = "nonzero _count";
666 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
667 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
668 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
671 if (unlikely(page->mem_cgroup))
672 bad_reason = "page still charged to cgroup";
674 if (unlikely(bad_reason)) {
675 bad_page(page, bad_reason, bad_flags);
678 page_cpupid_reset_last(page);
679 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
680 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
685 * Frees a number of pages from the PCP lists
686 * Assumes all pages on list are in same zone, and of same order.
687 * count is the number of pages to free.
689 * If the zone was previously in an "all pages pinned" state then look to
690 * see if this freeing clears that state.
692 * And clear the zone's pages_scanned counter, to hold off the "all pages are
693 * pinned" detection logic.
695 static void free_pcppages_bulk(struct zone *zone, int count,
696 struct per_cpu_pages *pcp)
701 unsigned long nr_scanned;
703 spin_lock(&zone->lock);
704 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
706 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
710 struct list_head *list;
713 * Remove pages from lists in a round-robin fashion. A
714 * batch_free count is maintained that is incremented when an
715 * empty list is encountered. This is so more pages are freed
716 * off fuller lists instead of spinning excessively around empty
721 if (++migratetype == MIGRATE_PCPTYPES)
723 list = &pcp->lists[migratetype];
724 } while (list_empty(list));
726 /* This is the only non-empty list. Free them all. */
727 if (batch_free == MIGRATE_PCPTYPES)
728 batch_free = to_free;
731 int mt; /* migratetype of the to-be-freed page */
733 page = list_entry(list->prev, struct page, lru);
734 /* must delete as __free_one_page list manipulates */
735 list_del(&page->lru);
736 mt = get_freepage_migratetype(page);
737 if (unlikely(has_isolate_pageblock(zone)))
738 mt = get_pageblock_migratetype(page);
740 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
741 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
742 trace_mm_page_pcpu_drain(page, 0, mt);
743 } while (--to_free && --batch_free && !list_empty(list));
745 spin_unlock(&zone->lock);
748 static void free_one_page(struct zone *zone,
749 struct page *page, unsigned long pfn,
753 unsigned long nr_scanned;
754 spin_lock(&zone->lock);
755 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
757 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
759 if (unlikely(has_isolate_pageblock(zone) ||
760 is_migrate_isolate(migratetype))) {
761 migratetype = get_pfnblock_migratetype(page, pfn);
763 __free_one_page(page, pfn, zone, order, migratetype);
764 spin_unlock(&zone->lock);
767 static int free_tail_pages_check(struct page *head_page, struct page *page)
769 if (!IS_ENABLED(CONFIG_DEBUG_VM))
771 if (unlikely(!PageTail(page))) {
772 bad_page(page, "PageTail not set", 0);
775 if (unlikely(page->first_page != head_page)) {
776 bad_page(page, "first_page not consistent", 0);
782 static bool free_pages_prepare(struct page *page, unsigned int order)
784 bool compound = PageCompound(page);
787 VM_BUG_ON_PAGE(PageTail(page), page);
788 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
790 trace_mm_page_free(page, order);
791 kmemcheck_free_shadow(page, order);
794 page->mapping = NULL;
795 bad += free_pages_check(page);
796 for (i = 1; i < (1 << order); i++) {
798 bad += free_tail_pages_check(page, page + i);
799 bad += free_pages_check(page + i);
804 reset_page_owner(page, order);
806 if (!PageHighMem(page)) {
807 debug_check_no_locks_freed(page_address(page),
809 debug_check_no_obj_freed(page_address(page),
812 arch_free_page(page, order);
813 kernel_map_pages(page, 1 << order, 0);
818 static void __free_pages_ok(struct page *page, unsigned int order)
822 unsigned long pfn = page_to_pfn(page);
824 if (!free_pages_prepare(page, order))
827 migratetype = get_pfnblock_migratetype(page, pfn);
828 local_irq_save(flags);
829 __count_vm_events(PGFREE, 1 << order);
830 set_freepage_migratetype(page, migratetype);
831 free_one_page(page_zone(page), page, pfn, order, migratetype);
832 local_irq_restore(flags);
835 void __init __free_pages_bootmem(struct page *page, unsigned int order)
837 unsigned int nr_pages = 1 << order;
838 struct page *p = page;
842 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
844 __ClearPageReserved(p);
845 set_page_count(p, 0);
847 __ClearPageReserved(p);
848 set_page_count(p, 0);
850 page_zone(page)->managed_pages += nr_pages;
851 set_page_refcounted(page);
852 __free_pages(page, order);
856 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
857 void __init init_cma_reserved_pageblock(struct page *page)
859 unsigned i = pageblock_nr_pages;
860 struct page *p = page;
863 __ClearPageReserved(p);
864 set_page_count(p, 0);
867 set_pageblock_migratetype(page, MIGRATE_CMA);
869 if (pageblock_order >= MAX_ORDER) {
870 i = pageblock_nr_pages;
873 set_page_refcounted(p);
874 __free_pages(p, MAX_ORDER - 1);
875 p += MAX_ORDER_NR_PAGES;
876 } while (i -= MAX_ORDER_NR_PAGES);
878 set_page_refcounted(page);
879 __free_pages(page, pageblock_order);
882 adjust_managed_page_count(page, pageblock_nr_pages);
887 * The order of subdivision here is critical for the IO subsystem.
888 * Please do not alter this order without good reasons and regression
889 * testing. Specifically, as large blocks of memory are subdivided,
890 * the order in which smaller blocks are delivered depends on the order
891 * they're subdivided in this function. This is the primary factor
892 * influencing the order in which pages are delivered to the IO
893 * subsystem according to empirical testing, and this is also justified
894 * by considering the behavior of a buddy system containing a single
895 * large block of memory acted on by a series of small allocations.
896 * This behavior is a critical factor in sglist merging's success.
900 static inline void expand(struct zone *zone, struct page *page,
901 int low, int high, struct free_area *area,
904 unsigned long size = 1 << high;
910 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
912 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
913 debug_guardpage_enabled() &&
914 high < debug_guardpage_minorder()) {
916 * Mark as guard pages (or page), that will allow to
917 * merge back to allocator when buddy will be freed.
918 * Corresponding page table entries will not be touched,
919 * pages will stay not present in virtual address space
921 set_page_guard(zone, &page[size], high, migratetype);
924 list_add(&page[size].lru, &area->free_list[migratetype]);
926 set_page_order(&page[size], high);
931 * This page is about to be returned from the page allocator
933 static inline int check_new_page(struct page *page)
935 const char *bad_reason = NULL;
936 unsigned long bad_flags = 0;
938 if (unlikely(page_mapcount(page)))
939 bad_reason = "nonzero mapcount";
940 if (unlikely(page->mapping != NULL))
941 bad_reason = "non-NULL mapping";
942 if (unlikely(atomic_read(&page->_count) != 0))
943 bad_reason = "nonzero _count";
944 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
945 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
946 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
949 if (unlikely(page->mem_cgroup))
950 bad_reason = "page still charged to cgroup";
952 if (unlikely(bad_reason)) {
953 bad_page(page, bad_reason, bad_flags);
959 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
964 for (i = 0; i < (1 << order); i++) {
965 struct page *p = page + i;
966 if (unlikely(check_new_page(p)))
970 set_page_private(page, 0);
971 set_page_refcounted(page);
973 arch_alloc_page(page, order);
974 kernel_map_pages(page, 1 << order, 1);
976 if (gfp_flags & __GFP_ZERO)
977 prep_zero_page(page, order, gfp_flags);
979 if (order && (gfp_flags & __GFP_COMP))
980 prep_compound_page(page, order);
982 set_page_owner(page, order, gfp_flags);
985 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
986 * allocate the page. The expectation is that the caller is taking
987 * steps that will free more memory. The caller should avoid the page
988 * being used for !PFMEMALLOC purposes.
990 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
996 * Go through the free lists for the given migratetype and remove
997 * the smallest available page from the freelists
1000 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1003 unsigned int current_order;
1004 struct free_area *area;
1007 /* Find a page of the appropriate size in the preferred list */
1008 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1009 area = &(zone->free_area[current_order]);
1010 if (list_empty(&area->free_list[migratetype]))
1013 page = list_entry(area->free_list[migratetype].next,
1015 list_del(&page->lru);
1016 rmv_page_order(page);
1018 expand(zone, page, order, current_order, area, migratetype);
1019 set_freepage_migratetype(page, migratetype);
1028 * This array describes the order lists are fallen back to when
1029 * the free lists for the desirable migrate type are depleted
1031 static int fallbacks[MIGRATE_TYPES][4] = {
1032 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1033 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1035 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1036 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1038 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1040 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1041 #ifdef CONFIG_MEMORY_ISOLATION
1042 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1047 * Move the free pages in a range to the free lists of the requested type.
1048 * Note that start_page and end_pages are not aligned on a pageblock
1049 * boundary. If alignment is required, use move_freepages_block()
1051 int move_freepages(struct zone *zone,
1052 struct page *start_page, struct page *end_page,
1056 unsigned long order;
1057 int pages_moved = 0;
1059 #ifndef CONFIG_HOLES_IN_ZONE
1061 * page_zone is not safe to call in this context when
1062 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1063 * anyway as we check zone boundaries in move_freepages_block().
1064 * Remove at a later date when no bug reports exist related to
1065 * grouping pages by mobility
1067 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1070 for (page = start_page; page <= end_page;) {
1071 /* Make sure we are not inadvertently changing nodes */
1072 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1074 if (!pfn_valid_within(page_to_pfn(page))) {
1079 if (!PageBuddy(page)) {
1084 order = page_order(page);
1085 list_move(&page->lru,
1086 &zone->free_area[order].free_list[migratetype]);
1087 set_freepage_migratetype(page, migratetype);
1089 pages_moved += 1 << order;
1095 int move_freepages_block(struct zone *zone, struct page *page,
1098 unsigned long start_pfn, end_pfn;
1099 struct page *start_page, *end_page;
1101 start_pfn = page_to_pfn(page);
1102 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1103 start_page = pfn_to_page(start_pfn);
1104 end_page = start_page + pageblock_nr_pages - 1;
1105 end_pfn = start_pfn + pageblock_nr_pages - 1;
1107 /* Do not cross zone boundaries */
1108 if (!zone_spans_pfn(zone, start_pfn))
1110 if (!zone_spans_pfn(zone, end_pfn))
1113 return move_freepages(zone, start_page, end_page, migratetype);
1116 static void change_pageblock_range(struct page *pageblock_page,
1117 int start_order, int migratetype)
1119 int nr_pageblocks = 1 << (start_order - pageblock_order);
1121 while (nr_pageblocks--) {
1122 set_pageblock_migratetype(pageblock_page, migratetype);
1123 pageblock_page += pageblock_nr_pages;
1128 * If breaking a large block of pages, move all free pages to the preferred
1129 * allocation list. If falling back for a reclaimable kernel allocation, be
1130 * more aggressive about taking ownership of free pages.
1132 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1133 * nor move CMA pages to different free lists. We don't want unmovable pages
1134 * to be allocated from MIGRATE_CMA areas.
1136 * Returns the new migratetype of the pageblock (or the same old migratetype
1137 * if it was unchanged).
1139 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1140 int start_type, int fallback_type)
1142 int current_order = page_order(page);
1145 * When borrowing from MIGRATE_CMA, we need to release the excess
1146 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1147 * is set to CMA so it is returned to the correct freelist in case
1148 * the page ends up being not actually allocated from the pcp lists.
1150 if (is_migrate_cma(fallback_type))
1151 return fallback_type;
1153 /* Take ownership for orders >= pageblock_order */
1154 if (current_order >= pageblock_order) {
1155 change_pageblock_range(page, current_order, start_type);
1159 if (current_order >= pageblock_order / 2 ||
1160 start_type == MIGRATE_RECLAIMABLE ||
1161 page_group_by_mobility_disabled) {
1164 pages = move_freepages_block(zone, page, start_type);
1166 /* Claim the whole block if over half of it is free */
1167 if (pages >= (1 << (pageblock_order-1)) ||
1168 page_group_by_mobility_disabled) {
1170 set_pageblock_migratetype(page, start_type);
1176 return fallback_type;
1179 /* Remove an element from the buddy allocator from the fallback list */
1180 static inline struct page *
1181 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1183 struct free_area *area;
1184 unsigned int current_order;
1186 int migratetype, new_type, i;
1188 /* Find the largest possible block of pages in the other list */
1189 for (current_order = MAX_ORDER-1;
1190 current_order >= order && current_order <= MAX_ORDER-1;
1193 migratetype = fallbacks[start_migratetype][i];
1195 /* MIGRATE_RESERVE handled later if necessary */
1196 if (migratetype == MIGRATE_RESERVE)
1199 area = &(zone->free_area[current_order]);
1200 if (list_empty(&area->free_list[migratetype]))
1203 page = list_entry(area->free_list[migratetype].next,
1207 new_type = try_to_steal_freepages(zone, page,
1211 /* Remove the page from the freelists */
1212 list_del(&page->lru);
1213 rmv_page_order(page);
1215 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 does
1220 * not differ for MIGRATE_CMA type.
1222 set_freepage_migratetype(page, new_type);
1224 trace_mm_page_alloc_extfrag(page, order, current_order,
1225 start_migratetype, migratetype, new_type);
1235 * Do the hard work of removing an element from the buddy allocator.
1236 * Call me with the zone->lock already held.
1238 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1244 page = __rmqueue_smallest(zone, order, migratetype);
1246 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1247 page = __rmqueue_fallback(zone, order, migratetype);
1250 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1251 * is used because __rmqueue_smallest is an inline function
1252 * and we want just one call site
1255 migratetype = MIGRATE_RESERVE;
1260 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1265 * Obtain a specified number of elements from the buddy allocator, all under
1266 * a single hold of the lock, for efficiency. Add them to the supplied list.
1267 * Returns the number of new pages which were placed at *list.
1269 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1270 unsigned long count, struct list_head *list,
1271 int migratetype, bool cold)
1275 spin_lock(&zone->lock);
1276 for (i = 0; i < count; ++i) {
1277 struct page *page = __rmqueue(zone, order, migratetype);
1278 if (unlikely(page == NULL))
1282 * Split buddy pages returned by expand() are received here
1283 * in physical page order. The page is added to the callers and
1284 * list and the list head then moves forward. From the callers
1285 * perspective, the linked list is ordered by page number in
1286 * some conditions. This is useful for IO devices that can
1287 * merge IO requests if the physical pages are ordered
1291 list_add(&page->lru, list);
1293 list_add_tail(&page->lru, list);
1295 if (is_migrate_cma(get_freepage_migratetype(page)))
1296 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1299 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1300 spin_unlock(&zone->lock);
1306 * Called from the vmstat counter updater to drain pagesets of this
1307 * currently executing processor on remote nodes after they have
1310 * Note that this function must be called with the thread pinned to
1311 * a single processor.
1313 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1315 unsigned long flags;
1316 int to_drain, batch;
1318 local_irq_save(flags);
1319 batch = ACCESS_ONCE(pcp->batch);
1320 to_drain = min(pcp->count, batch);
1322 free_pcppages_bulk(zone, to_drain, pcp);
1323 pcp->count -= to_drain;
1325 local_irq_restore(flags);
1330 * Drain pcplists of the indicated processor and zone.
1332 * The processor must either be the current processor and the
1333 * thread pinned to the current processor or a processor that
1336 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1338 unsigned long flags;
1339 struct per_cpu_pageset *pset;
1340 struct per_cpu_pages *pcp;
1342 local_irq_save(flags);
1343 pset = per_cpu_ptr(zone->pageset, cpu);
1347 free_pcppages_bulk(zone, pcp->count, pcp);
1350 local_irq_restore(flags);
1354 * Drain pcplists of all zones on the indicated processor.
1356 * The processor must either be the current processor and the
1357 * thread pinned to the current processor or a processor that
1360 static void drain_pages(unsigned int cpu)
1364 for_each_populated_zone(zone) {
1365 drain_pages_zone(cpu, zone);
1370 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1372 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1373 * the single zone's pages.
1375 void drain_local_pages(struct zone *zone)
1377 int cpu = smp_processor_id();
1380 drain_pages_zone(cpu, zone);
1386 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1388 * When zone parameter is non-NULL, spill just the single zone's pages.
1390 * Note that this code is protected against sending an IPI to an offline
1391 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1392 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1393 * nothing keeps CPUs from showing up after we populated the cpumask and
1394 * before the call to on_each_cpu_mask().
1396 void drain_all_pages(struct zone *zone)
1401 * Allocate in the BSS so we wont require allocation in
1402 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1404 static cpumask_t cpus_with_pcps;
1407 * We don't care about racing with CPU hotplug event
1408 * as offline notification will cause the notified
1409 * cpu to drain that CPU pcps and on_each_cpu_mask
1410 * disables preemption as part of its processing
1412 for_each_online_cpu(cpu) {
1413 struct per_cpu_pageset *pcp;
1415 bool has_pcps = false;
1418 pcp = per_cpu_ptr(zone->pageset, cpu);
1422 for_each_populated_zone(z) {
1423 pcp = per_cpu_ptr(z->pageset, cpu);
1424 if (pcp->pcp.count) {
1432 cpumask_set_cpu(cpu, &cpus_with_pcps);
1434 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1436 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1440 #ifdef CONFIG_HIBERNATION
1442 void mark_free_pages(struct zone *zone)
1444 unsigned long pfn, max_zone_pfn;
1445 unsigned long flags;
1446 unsigned int order, t;
1447 struct list_head *curr;
1449 if (zone_is_empty(zone))
1452 spin_lock_irqsave(&zone->lock, flags);
1454 max_zone_pfn = zone_end_pfn(zone);
1455 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1456 if (pfn_valid(pfn)) {
1457 struct page *page = pfn_to_page(pfn);
1459 if (!swsusp_page_is_forbidden(page))
1460 swsusp_unset_page_free(page);
1463 for_each_migratetype_order(order, t) {
1464 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1467 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1468 for (i = 0; i < (1UL << order); i++)
1469 swsusp_set_page_free(pfn_to_page(pfn + i));
1472 spin_unlock_irqrestore(&zone->lock, flags);
1474 #endif /* CONFIG_PM */
1477 * Free a 0-order page
1478 * cold == true ? free a cold page : free a hot page
1480 void free_hot_cold_page(struct page *page, bool cold)
1482 struct zone *zone = page_zone(page);
1483 struct per_cpu_pages *pcp;
1484 unsigned long flags;
1485 unsigned long pfn = page_to_pfn(page);
1488 if (!free_pages_prepare(page, 0))
1491 migratetype = get_pfnblock_migratetype(page, pfn);
1492 set_freepage_migratetype(page, migratetype);
1493 local_irq_save(flags);
1494 __count_vm_event(PGFREE);
1497 * We only track unmovable, reclaimable and movable on pcp lists.
1498 * Free ISOLATE pages back to the allocator because they are being
1499 * offlined but treat RESERVE as movable pages so we can get those
1500 * areas back if necessary. Otherwise, we may have to free
1501 * excessively into the page allocator
1503 if (migratetype >= MIGRATE_PCPTYPES) {
1504 if (unlikely(is_migrate_isolate(migratetype))) {
1505 free_one_page(zone, page, pfn, 0, migratetype);
1508 migratetype = MIGRATE_MOVABLE;
1511 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1513 list_add(&page->lru, &pcp->lists[migratetype]);
1515 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1517 if (pcp->count >= pcp->high) {
1518 unsigned long batch = ACCESS_ONCE(pcp->batch);
1519 free_pcppages_bulk(zone, batch, pcp);
1520 pcp->count -= batch;
1524 local_irq_restore(flags);
1528 * Free a list of 0-order pages
1530 void free_hot_cold_page_list(struct list_head *list, bool cold)
1532 struct page *page, *next;
1534 list_for_each_entry_safe(page, next, list, lru) {
1535 trace_mm_page_free_batched(page, cold);
1536 free_hot_cold_page(page, cold);
1541 * split_page takes a non-compound higher-order page, and splits it into
1542 * n (1<<order) sub-pages: page[0..n]
1543 * Each sub-page must be freed individually.
1545 * Note: this is probably too low level an operation for use in drivers.
1546 * Please consult with lkml before using this in your driver.
1548 void split_page(struct page *page, unsigned int order)
1552 VM_BUG_ON_PAGE(PageCompound(page), page);
1553 VM_BUG_ON_PAGE(!page_count(page), page);
1555 #ifdef CONFIG_KMEMCHECK
1557 * Split shadow pages too, because free(page[0]) would
1558 * otherwise free the whole shadow.
1560 if (kmemcheck_page_is_tracked(page))
1561 split_page(virt_to_page(page[0].shadow), order);
1564 set_page_owner(page, 0, 0);
1565 for (i = 1; i < (1 << order); i++) {
1566 set_page_refcounted(page + i);
1567 set_page_owner(page + i, 0, 0);
1570 EXPORT_SYMBOL_GPL(split_page);
1572 int __isolate_free_page(struct page *page, unsigned int order)
1574 unsigned long watermark;
1578 BUG_ON(!PageBuddy(page));
1580 zone = page_zone(page);
1581 mt = get_pageblock_migratetype(page);
1583 if (!is_migrate_isolate(mt)) {
1584 /* Obey watermarks as if the page was being allocated */
1585 watermark = low_wmark_pages(zone) + (1 << order);
1586 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1589 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1592 /* Remove page from free list */
1593 list_del(&page->lru);
1594 zone->free_area[order].nr_free--;
1595 rmv_page_order(page);
1597 /* Set the pageblock if the isolated page is at least a pageblock */
1598 if (order >= pageblock_order - 1) {
1599 struct page *endpage = page + (1 << order) - 1;
1600 for (; page < endpage; page += pageblock_nr_pages) {
1601 int mt = get_pageblock_migratetype(page);
1602 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1603 set_pageblock_migratetype(page,
1608 set_page_owner(page, order, 0);
1609 return 1UL << order;
1613 * Similar to split_page except the page is already free. As this is only
1614 * being used for migration, the migratetype of the block also changes.
1615 * As this is called with interrupts disabled, the caller is responsible
1616 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1619 * Note: this is probably too low level an operation for use in drivers.
1620 * Please consult with lkml before using this in your driver.
1622 int split_free_page(struct page *page)
1627 order = page_order(page);
1629 nr_pages = __isolate_free_page(page, order);
1633 /* Split into individual pages */
1634 set_page_refcounted(page);
1635 split_page(page, order);
1640 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1643 struct page *buffered_rmqueue(struct zone *preferred_zone,
1644 struct zone *zone, unsigned int order,
1645 gfp_t gfp_flags, int migratetype)
1647 unsigned long flags;
1649 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1651 if (likely(order == 0)) {
1652 struct per_cpu_pages *pcp;
1653 struct list_head *list;
1655 local_irq_save(flags);
1656 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1657 list = &pcp->lists[migratetype];
1658 if (list_empty(list)) {
1659 pcp->count += rmqueue_bulk(zone, 0,
1662 if (unlikely(list_empty(list)))
1667 page = list_entry(list->prev, struct page, lru);
1669 page = list_entry(list->next, struct page, lru);
1671 list_del(&page->lru);
1674 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1676 * __GFP_NOFAIL is not to be used in new code.
1678 * All __GFP_NOFAIL callers should be fixed so that they
1679 * properly detect and handle allocation failures.
1681 * We most definitely don't want callers attempting to
1682 * allocate greater than order-1 page units with
1685 WARN_ON_ONCE(order > 1);
1687 spin_lock_irqsave(&zone->lock, flags);
1688 page = __rmqueue(zone, order, migratetype);
1689 spin_unlock(&zone->lock);
1692 __mod_zone_freepage_state(zone, -(1 << order),
1693 get_freepage_migratetype(page));
1696 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1697 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1698 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1699 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1701 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1702 zone_statistics(preferred_zone, zone, gfp_flags);
1703 local_irq_restore(flags);
1705 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1709 local_irq_restore(flags);
1713 #ifdef CONFIG_FAIL_PAGE_ALLOC
1716 struct fault_attr attr;
1718 u32 ignore_gfp_highmem;
1719 u32 ignore_gfp_wait;
1721 } fail_page_alloc = {
1722 .attr = FAULT_ATTR_INITIALIZER,
1723 .ignore_gfp_wait = 1,
1724 .ignore_gfp_highmem = 1,
1728 static int __init setup_fail_page_alloc(char *str)
1730 return setup_fault_attr(&fail_page_alloc.attr, str);
1732 __setup("fail_page_alloc=", setup_fail_page_alloc);
1734 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1736 if (order < fail_page_alloc.min_order)
1738 if (gfp_mask & __GFP_NOFAIL)
1740 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1742 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1745 return should_fail(&fail_page_alloc.attr, 1 << order);
1748 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1750 static int __init fail_page_alloc_debugfs(void)
1752 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1755 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1756 &fail_page_alloc.attr);
1758 return PTR_ERR(dir);
1760 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1761 &fail_page_alloc.ignore_gfp_wait))
1763 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1764 &fail_page_alloc.ignore_gfp_highmem))
1766 if (!debugfs_create_u32("min-order", mode, dir,
1767 &fail_page_alloc.min_order))
1772 debugfs_remove_recursive(dir);
1777 late_initcall(fail_page_alloc_debugfs);
1779 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1781 #else /* CONFIG_FAIL_PAGE_ALLOC */
1783 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1788 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1791 * Return true if free pages are above 'mark'. This takes into account the order
1792 * of the allocation.
1794 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1795 unsigned long mark, int classzone_idx, int alloc_flags,
1798 /* free_pages may go negative - that's OK */
1803 free_pages -= (1 << order) - 1;
1804 if (alloc_flags & ALLOC_HIGH)
1806 if (alloc_flags & ALLOC_HARDER)
1809 /* If allocation can't use CMA areas don't use free CMA pages */
1810 if (!(alloc_flags & ALLOC_CMA))
1811 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1814 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1816 for (o = 0; o < order; o++) {
1817 /* At the next order, this order's pages become unavailable */
1818 free_pages -= z->free_area[o].nr_free << o;
1820 /* Require fewer higher order pages to be free */
1823 if (free_pages <= min)
1829 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1830 int classzone_idx, int alloc_flags)
1832 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1833 zone_page_state(z, NR_FREE_PAGES));
1836 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1837 unsigned long mark, int classzone_idx, int alloc_flags)
1839 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1841 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1842 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1844 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1850 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1851 * skip over zones that are not allowed by the cpuset, or that have
1852 * been recently (in last second) found to be nearly full. See further
1853 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1854 * that have to skip over a lot of full or unallowed zones.
1856 * If the zonelist cache is present in the passed zonelist, then
1857 * returns a pointer to the allowed node mask (either the current
1858 * tasks mems_allowed, or node_states[N_MEMORY].)
1860 * If the zonelist cache is not available for this zonelist, does
1861 * nothing and returns NULL.
1863 * If the fullzones BITMAP in the zonelist cache is stale (more than
1864 * a second since last zap'd) then we zap it out (clear its bits.)
1866 * We hold off even calling zlc_setup, until after we've checked the
1867 * first zone in the zonelist, on the theory that most allocations will
1868 * be satisfied from that first zone, so best to examine that zone as
1869 * quickly as we can.
1871 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1873 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1874 nodemask_t *allowednodes; /* zonelist_cache approximation */
1876 zlc = zonelist->zlcache_ptr;
1880 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1881 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1882 zlc->last_full_zap = jiffies;
1885 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1886 &cpuset_current_mems_allowed :
1887 &node_states[N_MEMORY];
1888 return allowednodes;
1892 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1893 * if it is worth looking at further for free memory:
1894 * 1) Check that the zone isn't thought to be full (doesn't have its
1895 * bit set in the zonelist_cache fullzones BITMAP).
1896 * 2) Check that the zones node (obtained from the zonelist_cache
1897 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1898 * Return true (non-zero) if zone is worth looking at further, or
1899 * else return false (zero) if it is not.
1901 * This check -ignores- the distinction between various watermarks,
1902 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1903 * found to be full for any variation of these watermarks, it will
1904 * be considered full for up to one second by all requests, unless
1905 * we are so low on memory on all allowed nodes that we are forced
1906 * into the second scan of the zonelist.
1908 * In the second scan we ignore this zonelist cache and exactly
1909 * apply the watermarks to all zones, even it is slower to do so.
1910 * We are low on memory in the second scan, and should leave no stone
1911 * unturned looking for a free page.
1913 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1914 nodemask_t *allowednodes)
1916 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1917 int i; /* index of *z in zonelist zones */
1918 int n; /* node that zone *z is on */
1920 zlc = zonelist->zlcache_ptr;
1924 i = z - zonelist->_zonerefs;
1927 /* This zone is worth trying if it is allowed but not full */
1928 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1932 * Given 'z' scanning a zonelist, set the corresponding bit in
1933 * zlc->fullzones, so that subsequent attempts to allocate a page
1934 * from that zone don't waste time re-examining it.
1936 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1938 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1939 int i; /* index of *z in zonelist zones */
1941 zlc = zonelist->zlcache_ptr;
1945 i = z - zonelist->_zonerefs;
1947 set_bit(i, zlc->fullzones);
1951 * clear all zones full, called after direct reclaim makes progress so that
1952 * a zone that was recently full is not skipped over for up to a second
1954 static void zlc_clear_zones_full(struct zonelist *zonelist)
1956 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1958 zlc = zonelist->zlcache_ptr;
1962 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1965 static bool zone_local(struct zone *local_zone, struct zone *zone)
1967 return local_zone->node == zone->node;
1970 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1972 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1976 #else /* CONFIG_NUMA */
1978 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1983 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1984 nodemask_t *allowednodes)
1989 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1993 static void zlc_clear_zones_full(struct zonelist *zonelist)
1997 static bool zone_local(struct zone *local_zone, struct zone *zone)
2002 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2007 #endif /* CONFIG_NUMA */
2009 static void reset_alloc_batches(struct zone *preferred_zone)
2011 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2014 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2015 high_wmark_pages(zone) - low_wmark_pages(zone) -
2016 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2017 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2018 } while (zone++ != preferred_zone);
2022 * get_page_from_freelist goes through the zonelist trying to allocate
2025 static struct page *
2026 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2027 const struct alloc_context *ac)
2029 struct zonelist *zonelist = ac->zonelist;
2031 struct page *page = NULL;
2033 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2034 int zlc_active = 0; /* set if using zonelist_cache */
2035 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2036 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2037 (gfp_mask & __GFP_WRITE);
2038 int nr_fair_skipped = 0;
2039 bool zonelist_rescan;
2042 zonelist_rescan = false;
2045 * Scan zonelist, looking for a zone with enough free.
2046 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2048 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2052 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2053 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2055 if (cpusets_enabled() &&
2056 (alloc_flags & ALLOC_CPUSET) &&
2057 !cpuset_zone_allowed(zone, gfp_mask))
2060 * Distribute pages in proportion to the individual
2061 * zone size to ensure fair page aging. The zone a
2062 * page was allocated in should have no effect on the
2063 * time the page has in memory before being reclaimed.
2065 if (alloc_flags & ALLOC_FAIR) {
2066 if (!zone_local(ac->preferred_zone, zone))
2068 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2074 * When allocating a page cache page for writing, we
2075 * want to get it from a zone that is within its dirty
2076 * limit, such that no single zone holds more than its
2077 * proportional share of globally allowed dirty pages.
2078 * The dirty limits take into account the zone's
2079 * lowmem reserves and high watermark so that kswapd
2080 * should be able to balance it without having to
2081 * write pages from its LRU list.
2083 * This may look like it could increase pressure on
2084 * lower zones by failing allocations in higher zones
2085 * before they are full. But the pages that do spill
2086 * over are limited as the lower zones are protected
2087 * by this very same mechanism. It should not become
2088 * a practical burden to them.
2090 * XXX: For now, allow allocations to potentially
2091 * exceed the per-zone dirty limit in the slowpath
2092 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2093 * which is important when on a NUMA setup the allowed
2094 * zones are together not big enough to reach the
2095 * global limit. The proper fix for these situations
2096 * will require awareness of zones in the
2097 * dirty-throttling and the flusher threads.
2099 if (consider_zone_dirty && !zone_dirty_ok(zone))
2102 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2103 if (!zone_watermark_ok(zone, order, mark,
2104 ac->classzone_idx, alloc_flags)) {
2107 /* Checked here to keep the fast path fast */
2108 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2109 if (alloc_flags & ALLOC_NO_WATERMARKS)
2112 if (IS_ENABLED(CONFIG_NUMA) &&
2113 !did_zlc_setup && nr_online_nodes > 1) {
2115 * we do zlc_setup if there are multiple nodes
2116 * and before considering the first zone allowed
2119 allowednodes = zlc_setup(zonelist, alloc_flags);
2124 if (zone_reclaim_mode == 0 ||
2125 !zone_allows_reclaim(ac->preferred_zone, zone))
2126 goto this_zone_full;
2129 * As we may have just activated ZLC, check if the first
2130 * eligible zone has failed zone_reclaim recently.
2132 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2133 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2136 ret = zone_reclaim(zone, gfp_mask, order);
2138 case ZONE_RECLAIM_NOSCAN:
2141 case ZONE_RECLAIM_FULL:
2142 /* scanned but unreclaimable */
2145 /* did we reclaim enough */
2146 if (zone_watermark_ok(zone, order, mark,
2147 ac->classzone_idx, alloc_flags))
2151 * Failed to reclaim enough to meet watermark.
2152 * Only mark the zone full if checking the min
2153 * watermark or if we failed to reclaim just
2154 * 1<<order pages or else the page allocator
2155 * fastpath will prematurely mark zones full
2156 * when the watermark is between the low and
2159 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2160 ret == ZONE_RECLAIM_SOME)
2161 goto this_zone_full;
2168 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2169 gfp_mask, ac->migratetype);
2171 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2176 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2177 zlc_mark_zone_full(zonelist, z);
2181 * The first pass makes sure allocations are spread fairly within the
2182 * local node. However, the local node might have free pages left
2183 * after the fairness batches are exhausted, and remote zones haven't
2184 * even been considered yet. Try once more without fairness, and
2185 * include remote zones now, before entering the slowpath and waking
2186 * kswapd: prefer spilling to a remote zone over swapping locally.
2188 if (alloc_flags & ALLOC_FAIR) {
2189 alloc_flags &= ~ALLOC_FAIR;
2190 if (nr_fair_skipped) {
2191 zonelist_rescan = true;
2192 reset_alloc_batches(ac->preferred_zone);
2194 if (nr_online_nodes > 1)
2195 zonelist_rescan = true;
2198 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2199 /* Disable zlc cache for second zonelist scan */
2201 zonelist_rescan = true;
2204 if (zonelist_rescan)
2211 * Large machines with many possible nodes should not always dump per-node
2212 * meminfo in irq context.
2214 static inline bool should_suppress_show_mem(void)
2219 ret = in_interrupt();
2224 static DEFINE_RATELIMIT_STATE(nopage_rs,
2225 DEFAULT_RATELIMIT_INTERVAL,
2226 DEFAULT_RATELIMIT_BURST);
2228 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2230 unsigned int filter = SHOW_MEM_FILTER_NODES;
2232 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2233 debug_guardpage_minorder() > 0)
2237 * This documents exceptions given to allocations in certain
2238 * contexts that are allowed to allocate outside current's set
2241 if (!(gfp_mask & __GFP_NOMEMALLOC))
2242 if (test_thread_flag(TIF_MEMDIE) ||
2243 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2244 filter &= ~SHOW_MEM_FILTER_NODES;
2245 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2246 filter &= ~SHOW_MEM_FILTER_NODES;
2249 struct va_format vaf;
2252 va_start(args, fmt);
2257 pr_warn("%pV", &vaf);
2262 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2263 current->comm, order, gfp_mask);
2266 if (!should_suppress_show_mem())
2271 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2272 unsigned long did_some_progress,
2273 unsigned long pages_reclaimed)
2275 /* Do not loop if specifically requested */
2276 if (gfp_mask & __GFP_NORETRY)
2279 /* Always retry if specifically requested */
2280 if (gfp_mask & __GFP_NOFAIL)
2284 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2285 * making forward progress without invoking OOM. Suspend also disables
2286 * storage devices so kswapd will not help. Bail if we are suspending.
2288 if (!did_some_progress && pm_suspended_storage())
2292 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2293 * means __GFP_NOFAIL, but that may not be true in other
2296 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2300 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2301 * specified, then we retry until we no longer reclaim any pages
2302 * (above), or we've reclaimed an order of pages at least as
2303 * large as the allocation's order. In both cases, if the
2304 * allocation still fails, we stop retrying.
2306 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2312 static inline struct page *
2313 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2314 const struct alloc_context *ac, unsigned long *did_some_progress)
2318 *did_some_progress = 0;
2320 if (oom_killer_disabled)
2324 * Acquire the per-zone oom lock for each zone. If that
2325 * fails, somebody else is making progress for us.
2327 if (!oom_zonelist_trylock(ac->zonelist, gfp_mask)) {
2328 *did_some_progress = 1;
2329 schedule_timeout_uninterruptible(1);
2334 * PM-freezer should be notified that there might be an OOM killer on
2335 * its way to kill and wake somebody up. This is too early and we might
2336 * end up not killing anything but false positives are acceptable.
2337 * See freeze_processes.
2342 * Go through the zonelist yet one more time, keep very high watermark
2343 * here, this is only to catch a parallel oom killing, we must fail if
2344 * we're still under heavy pressure.
2346 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2347 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2351 if (!(gfp_mask & __GFP_NOFAIL)) {
2352 /* Coredumps can quickly deplete all memory reserves */
2353 if (current->flags & PF_DUMPCORE)
2355 /* The OOM killer will not help higher order allocs */
2356 if (order > PAGE_ALLOC_COSTLY_ORDER)
2358 /* The OOM killer does not needlessly kill tasks for lowmem */
2359 if (ac->high_zoneidx < ZONE_NORMAL)
2361 /* The OOM killer does not compensate for light reclaim */
2362 if (!(gfp_mask & __GFP_FS))
2365 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2366 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2367 * The caller should handle page allocation failure by itself if
2368 * it specifies __GFP_THISNODE.
2369 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2371 if (gfp_mask & __GFP_THISNODE)
2374 /* Exhausted what can be done so it's blamo time */
2375 out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false);
2376 *did_some_progress = 1;
2378 oom_zonelist_unlock(ac->zonelist, gfp_mask);
2382 #ifdef CONFIG_COMPACTION
2383 /* Try memory compaction for high-order allocations before reclaim */
2384 static struct page *
2385 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2386 int alloc_flags, const struct alloc_context *ac,
2387 enum migrate_mode mode, int *contended_compaction,
2388 bool *deferred_compaction)
2390 unsigned long compact_result;
2396 current->flags |= PF_MEMALLOC;
2397 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2398 mode, contended_compaction);
2399 current->flags &= ~PF_MEMALLOC;
2401 switch (compact_result) {
2402 case COMPACT_DEFERRED:
2403 *deferred_compaction = true;
2405 case COMPACT_SKIPPED:
2412 * At least in one zone compaction wasn't deferred or skipped, so let's
2413 * count a compaction stall
2415 count_vm_event(COMPACTSTALL);
2417 page = get_page_from_freelist(gfp_mask, order,
2418 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2421 struct zone *zone = page_zone(page);
2423 zone->compact_blockskip_flush = false;
2424 compaction_defer_reset(zone, order, true);
2425 count_vm_event(COMPACTSUCCESS);
2430 * It's bad if compaction run occurs and fails. The most likely reason
2431 * is that pages exist, but not enough to satisfy watermarks.
2433 count_vm_event(COMPACTFAIL);
2440 static inline struct page *
2441 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2442 int alloc_flags, const struct alloc_context *ac,
2443 enum migrate_mode mode, int *contended_compaction,
2444 bool *deferred_compaction)
2448 #endif /* CONFIG_COMPACTION */
2450 /* Perform direct synchronous page reclaim */
2452 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2453 const struct alloc_context *ac)
2455 struct reclaim_state reclaim_state;
2460 /* We now go into synchronous reclaim */
2461 cpuset_memory_pressure_bump();
2462 current->flags |= PF_MEMALLOC;
2463 lockdep_set_current_reclaim_state(gfp_mask);
2464 reclaim_state.reclaimed_slab = 0;
2465 current->reclaim_state = &reclaim_state;
2467 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2470 current->reclaim_state = NULL;
2471 lockdep_clear_current_reclaim_state();
2472 current->flags &= ~PF_MEMALLOC;
2479 /* The really slow allocator path where we enter direct reclaim */
2480 static inline struct page *
2481 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2482 int alloc_flags, const struct alloc_context *ac,
2483 unsigned long *did_some_progress)
2485 struct page *page = NULL;
2486 bool drained = false;
2488 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2489 if (unlikely(!(*did_some_progress)))
2492 /* After successful reclaim, reconsider all zones for allocation */
2493 if (IS_ENABLED(CONFIG_NUMA))
2494 zlc_clear_zones_full(ac->zonelist);
2497 page = get_page_from_freelist(gfp_mask, order,
2498 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2501 * If an allocation failed after direct reclaim, it could be because
2502 * pages are pinned on the per-cpu lists. Drain them and try again
2504 if (!page && !drained) {
2505 drain_all_pages(NULL);
2514 * This is called in the allocator slow-path if the allocation request is of
2515 * sufficient urgency to ignore watermarks and take other desperate measures
2517 static inline struct page *
2518 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2519 const struct alloc_context *ac)
2524 page = get_page_from_freelist(gfp_mask, order,
2525 ALLOC_NO_WATERMARKS, ac);
2527 if (!page && gfp_mask & __GFP_NOFAIL)
2528 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2530 } while (!page && (gfp_mask & __GFP_NOFAIL));
2535 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2540 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2541 ac->high_zoneidx, ac->nodemask)
2542 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2546 gfp_to_alloc_flags(gfp_t gfp_mask)
2548 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2549 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2551 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2552 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2555 * The caller may dip into page reserves a bit more if the caller
2556 * cannot run direct reclaim, or if the caller has realtime scheduling
2557 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2558 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2560 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2564 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2565 * if it can't schedule.
2567 if (!(gfp_mask & __GFP_NOMEMALLOC))
2568 alloc_flags |= ALLOC_HARDER;
2570 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2571 * comment for __cpuset_node_allowed().
2573 alloc_flags &= ~ALLOC_CPUSET;
2574 } else if (unlikely(rt_task(current)) && !in_interrupt())
2575 alloc_flags |= ALLOC_HARDER;
2577 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2578 if (gfp_mask & __GFP_MEMALLOC)
2579 alloc_flags |= ALLOC_NO_WATERMARKS;
2580 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2581 alloc_flags |= ALLOC_NO_WATERMARKS;
2582 else if (!in_interrupt() &&
2583 ((current->flags & PF_MEMALLOC) ||
2584 unlikely(test_thread_flag(TIF_MEMDIE))))
2585 alloc_flags |= ALLOC_NO_WATERMARKS;
2588 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2589 alloc_flags |= ALLOC_CMA;
2594 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2596 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2599 static inline struct page *
2600 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2601 struct alloc_context *ac)
2603 const gfp_t wait = gfp_mask & __GFP_WAIT;
2604 struct page *page = NULL;
2606 unsigned long pages_reclaimed = 0;
2607 unsigned long did_some_progress;
2608 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2609 bool deferred_compaction = false;
2610 int contended_compaction = COMPACT_CONTENDED_NONE;
2613 * In the slowpath, we sanity check order to avoid ever trying to
2614 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2615 * be using allocators in order of preference for an area that is
2618 if (order >= MAX_ORDER) {
2619 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2624 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2625 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2626 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2627 * using a larger set of nodes after it has established that the
2628 * allowed per node queues are empty and that nodes are
2631 if (IS_ENABLED(CONFIG_NUMA) &&
2632 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2636 if (!(gfp_mask & __GFP_NO_KSWAPD))
2637 wake_all_kswapds(order, ac);
2640 * OK, we're below the kswapd watermark and have kicked background
2641 * reclaim. Now things get more complex, so set up alloc_flags according
2642 * to how we want to proceed.
2644 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2647 * Find the true preferred zone if the allocation is unconstrained by
2650 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
2651 struct zoneref *preferred_zoneref;
2652 preferred_zoneref = first_zones_zonelist(ac->zonelist,
2653 ac->high_zoneidx, NULL, &ac->preferred_zone);
2654 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
2657 /* This is the last chance, in general, before the goto nopage. */
2658 page = get_page_from_freelist(gfp_mask, order,
2659 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2663 /* Allocate without watermarks if the context allows */
2664 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2666 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2667 * the allocation is high priority and these type of
2668 * allocations are system rather than user orientated
2670 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
2672 page = __alloc_pages_high_priority(gfp_mask, order, ac);
2679 /* Atomic allocations - we can't balance anything */
2682 * All existing users of the deprecated __GFP_NOFAIL are
2683 * blockable, so warn of any new users that actually allow this
2684 * type of allocation to fail.
2686 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2690 /* Avoid recursion of direct reclaim */
2691 if (current->flags & PF_MEMALLOC)
2694 /* Avoid allocations with no watermarks from looping endlessly */
2695 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2699 * Try direct compaction. The first pass is asynchronous. Subsequent
2700 * attempts after direct reclaim are synchronous
2702 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
2704 &contended_compaction,
2705 &deferred_compaction);
2709 /* Checks for THP-specific high-order allocations */
2710 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2712 * If compaction is deferred for high-order allocations, it is
2713 * because sync compaction recently failed. If this is the case
2714 * and the caller requested a THP allocation, we do not want
2715 * to heavily disrupt the system, so we fail the allocation
2716 * instead of entering direct reclaim.
2718 if (deferred_compaction)
2722 * In all zones where compaction was attempted (and not
2723 * deferred or skipped), lock contention has been detected.
2724 * For THP allocation we do not want to disrupt the others
2725 * so we fallback to base pages instead.
2727 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2731 * If compaction was aborted due to need_resched(), we do not
2732 * want to further increase allocation latency, unless it is
2733 * khugepaged trying to collapse.
2735 if (contended_compaction == COMPACT_CONTENDED_SCHED
2736 && !(current->flags & PF_KTHREAD))
2741 * It can become very expensive to allocate transparent hugepages at
2742 * fault, so use asynchronous memory compaction for THP unless it is
2743 * khugepaged trying to collapse.
2745 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2746 (current->flags & PF_KTHREAD))
2747 migration_mode = MIGRATE_SYNC_LIGHT;
2749 /* Try direct reclaim and then allocating */
2750 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
2751 &did_some_progress);
2755 /* Check if we should retry the allocation */
2756 pages_reclaimed += did_some_progress;
2757 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2760 * If we fail to make progress by freeing individual
2761 * pages, but the allocation wants us to keep going,
2762 * start OOM killing tasks.
2764 if (!did_some_progress) {
2765 page = __alloc_pages_may_oom(gfp_mask, order, ac,
2766 &did_some_progress);
2769 if (!did_some_progress)
2772 /* Wait for some write requests to complete then retry */
2773 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
2777 * High-order allocations do not necessarily loop after
2778 * direct reclaim and reclaim/compaction depends on compaction
2779 * being called after reclaim so call directly if necessary
2781 page = __alloc_pages_direct_compact(gfp_mask, order,
2782 alloc_flags, ac, migration_mode,
2783 &contended_compaction,
2784 &deferred_compaction);
2790 warn_alloc_failed(gfp_mask, order, NULL);
2796 * This is the 'heart' of the zoned buddy allocator.
2799 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2800 struct zonelist *zonelist, nodemask_t *nodemask)
2802 struct zoneref *preferred_zoneref;
2803 struct page *page = NULL;
2804 unsigned int cpuset_mems_cookie;
2805 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2806 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
2807 struct alloc_context ac = {
2808 .high_zoneidx = gfp_zone(gfp_mask),
2809 .nodemask = nodemask,
2810 .migratetype = gfpflags_to_migratetype(gfp_mask),
2813 gfp_mask &= gfp_allowed_mask;
2815 lockdep_trace_alloc(gfp_mask);
2817 might_sleep_if(gfp_mask & __GFP_WAIT);
2819 if (should_fail_alloc_page(gfp_mask, order))
2823 * Check the zones suitable for the gfp_mask contain at least one
2824 * valid zone. It's possible to have an empty zonelist as a result
2825 * of GFP_THISNODE and a memoryless node
2827 if (unlikely(!zonelist->_zonerefs->zone))
2830 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
2831 alloc_flags |= ALLOC_CMA;
2834 cpuset_mems_cookie = read_mems_allowed_begin();
2836 /* We set it here, as __alloc_pages_slowpath might have changed it */
2837 ac.zonelist = zonelist;
2838 /* The preferred zone is used for statistics later */
2839 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
2840 ac.nodemask ? : &cpuset_current_mems_allowed,
2841 &ac.preferred_zone);
2842 if (!ac.preferred_zone)
2844 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
2846 /* First allocation attempt */
2847 alloc_mask = gfp_mask|__GFP_HARDWALL;
2848 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
2849 if (unlikely(!page)) {
2851 * Runtime PM, block IO and its error handling path
2852 * can deadlock because I/O on the device might not
2855 alloc_mask = memalloc_noio_flags(gfp_mask);
2857 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
2860 if (kmemcheck_enabled && page)
2861 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2863 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
2867 * When updating a task's mems_allowed, it is possible to race with
2868 * parallel threads in such a way that an allocation can fail while
2869 * the mask is being updated. If a page allocation is about to fail,
2870 * check if the cpuset changed during allocation and if so, retry.
2872 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2877 EXPORT_SYMBOL(__alloc_pages_nodemask);
2880 * Common helper functions.
2882 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2887 * __get_free_pages() returns a 32-bit address, which cannot represent
2890 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2892 page = alloc_pages(gfp_mask, order);
2895 return (unsigned long) page_address(page);
2897 EXPORT_SYMBOL(__get_free_pages);
2899 unsigned long get_zeroed_page(gfp_t gfp_mask)
2901 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2903 EXPORT_SYMBOL(get_zeroed_page);
2905 void __free_pages(struct page *page, unsigned int order)
2907 if (put_page_testzero(page)) {
2909 free_hot_cold_page(page, false);
2911 __free_pages_ok(page, order);
2915 EXPORT_SYMBOL(__free_pages);
2917 void free_pages(unsigned long addr, unsigned int order)
2920 VM_BUG_ON(!virt_addr_valid((void *)addr));
2921 __free_pages(virt_to_page((void *)addr), order);
2925 EXPORT_SYMBOL(free_pages);
2928 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2929 * of the current memory cgroup.
2931 * It should be used when the caller would like to use kmalloc, but since the
2932 * allocation is large, it has to fall back to the page allocator.
2934 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2937 struct mem_cgroup *memcg = NULL;
2939 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2941 page = alloc_pages(gfp_mask, order);
2942 memcg_kmem_commit_charge(page, memcg, order);
2946 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2949 struct mem_cgroup *memcg = NULL;
2951 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2953 page = alloc_pages_node(nid, gfp_mask, order);
2954 memcg_kmem_commit_charge(page, memcg, order);
2959 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2962 void __free_kmem_pages(struct page *page, unsigned int order)
2964 memcg_kmem_uncharge_pages(page, order);
2965 __free_pages(page, order);
2968 void free_kmem_pages(unsigned long addr, unsigned int order)
2971 VM_BUG_ON(!virt_addr_valid((void *)addr));
2972 __free_kmem_pages(virt_to_page((void *)addr), order);
2976 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2979 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2980 unsigned long used = addr + PAGE_ALIGN(size);
2982 split_page(virt_to_page((void *)addr), order);
2983 while (used < alloc_end) {
2988 return (void *)addr;
2992 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2993 * @size: the number of bytes to allocate
2994 * @gfp_mask: GFP flags for the allocation
2996 * This function is similar to alloc_pages(), except that it allocates the
2997 * minimum number of pages to satisfy the request. alloc_pages() can only
2998 * allocate memory in power-of-two pages.
3000 * This function is also limited by MAX_ORDER.
3002 * Memory allocated by this function must be released by free_pages_exact().
3004 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3006 unsigned int order = get_order(size);
3009 addr = __get_free_pages(gfp_mask, order);
3010 return make_alloc_exact(addr, order, size);
3012 EXPORT_SYMBOL(alloc_pages_exact);
3015 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3017 * @nid: the preferred node ID where memory should be allocated
3018 * @size: the number of bytes to allocate
3019 * @gfp_mask: GFP flags for the allocation
3021 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3023 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3026 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3028 unsigned order = get_order(size);
3029 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3032 return make_alloc_exact((unsigned long)page_address(p), order, size);
3036 * free_pages_exact - release memory allocated via alloc_pages_exact()
3037 * @virt: the value returned by alloc_pages_exact.
3038 * @size: size of allocation, same value as passed to alloc_pages_exact().
3040 * Release the memory allocated by a previous call to alloc_pages_exact.
3042 void free_pages_exact(void *virt, size_t size)
3044 unsigned long addr = (unsigned long)virt;
3045 unsigned long end = addr + PAGE_ALIGN(size);
3047 while (addr < end) {
3052 EXPORT_SYMBOL(free_pages_exact);
3055 * nr_free_zone_pages - count number of pages beyond high watermark
3056 * @offset: The zone index of the highest zone
3058 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3059 * high watermark within all zones at or below a given zone index. For each
3060 * zone, the number of pages is calculated as:
3061 * managed_pages - high_pages
3063 static unsigned long nr_free_zone_pages(int offset)
3068 /* Just pick one node, since fallback list is circular */
3069 unsigned long sum = 0;
3071 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3073 for_each_zone_zonelist(zone, z, zonelist, offset) {
3074 unsigned long size = zone->managed_pages;
3075 unsigned long high = high_wmark_pages(zone);
3084 * nr_free_buffer_pages - count number of pages beyond high watermark
3086 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3087 * watermark within ZONE_DMA and ZONE_NORMAL.
3089 unsigned long nr_free_buffer_pages(void)
3091 return nr_free_zone_pages(gfp_zone(GFP_USER));
3093 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3096 * nr_free_pagecache_pages - count number of pages beyond high watermark
3098 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3099 * high watermark within all zones.
3101 unsigned long nr_free_pagecache_pages(void)
3103 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3106 static inline void show_node(struct zone *zone)
3108 if (IS_ENABLED(CONFIG_NUMA))
3109 printk("Node %d ", zone_to_nid(zone));
3112 void si_meminfo(struct sysinfo *val)
3114 val->totalram = totalram_pages;
3115 val->sharedram = global_page_state(NR_SHMEM);
3116 val->freeram = global_page_state(NR_FREE_PAGES);
3117 val->bufferram = nr_blockdev_pages();
3118 val->totalhigh = totalhigh_pages;
3119 val->freehigh = nr_free_highpages();
3120 val->mem_unit = PAGE_SIZE;
3123 EXPORT_SYMBOL(si_meminfo);
3126 void si_meminfo_node(struct sysinfo *val, int nid)
3128 int zone_type; /* needs to be signed */
3129 unsigned long managed_pages = 0;
3130 pg_data_t *pgdat = NODE_DATA(nid);
3132 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3133 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3134 val->totalram = managed_pages;
3135 val->sharedram = node_page_state(nid, NR_SHMEM);
3136 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3137 #ifdef CONFIG_HIGHMEM
3138 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3139 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3145 val->mem_unit = PAGE_SIZE;
3150 * Determine whether the node should be displayed or not, depending on whether
3151 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3153 bool skip_free_areas_node(unsigned int flags, int nid)
3156 unsigned int cpuset_mems_cookie;
3158 if (!(flags & SHOW_MEM_FILTER_NODES))
3162 cpuset_mems_cookie = read_mems_allowed_begin();
3163 ret = !node_isset(nid, cpuset_current_mems_allowed);
3164 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3169 #define K(x) ((x) << (PAGE_SHIFT-10))
3171 static void show_migration_types(unsigned char type)
3173 static const char types[MIGRATE_TYPES] = {
3174 [MIGRATE_UNMOVABLE] = 'U',
3175 [MIGRATE_RECLAIMABLE] = 'E',
3176 [MIGRATE_MOVABLE] = 'M',
3177 [MIGRATE_RESERVE] = 'R',
3179 [MIGRATE_CMA] = 'C',
3181 #ifdef CONFIG_MEMORY_ISOLATION
3182 [MIGRATE_ISOLATE] = 'I',
3185 char tmp[MIGRATE_TYPES + 1];
3189 for (i = 0; i < MIGRATE_TYPES; i++) {
3190 if (type & (1 << i))
3195 printk("(%s) ", tmp);
3199 * Show free area list (used inside shift_scroll-lock stuff)
3200 * We also calculate the percentage fragmentation. We do this by counting the
3201 * memory on each free list with the exception of the first item on the list.
3202 * Suppresses nodes that are not allowed by current's cpuset if
3203 * SHOW_MEM_FILTER_NODES is passed.
3205 void show_free_areas(unsigned int filter)
3210 for_each_populated_zone(zone) {
3211 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3214 printk("%s per-cpu:\n", zone->name);
3216 for_each_online_cpu(cpu) {
3217 struct per_cpu_pageset *pageset;
3219 pageset = per_cpu_ptr(zone->pageset, cpu);
3221 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3222 cpu, pageset->pcp.high,
3223 pageset->pcp.batch, pageset->pcp.count);
3227 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3228 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3230 " dirty:%lu writeback:%lu unstable:%lu\n"
3231 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3232 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3234 global_page_state(NR_ACTIVE_ANON),
3235 global_page_state(NR_INACTIVE_ANON),
3236 global_page_state(NR_ISOLATED_ANON),
3237 global_page_state(NR_ACTIVE_FILE),
3238 global_page_state(NR_INACTIVE_FILE),
3239 global_page_state(NR_ISOLATED_FILE),
3240 global_page_state(NR_UNEVICTABLE),
3241 global_page_state(NR_FILE_DIRTY),
3242 global_page_state(NR_WRITEBACK),
3243 global_page_state(NR_UNSTABLE_NFS),
3244 global_page_state(NR_FREE_PAGES),
3245 global_page_state(NR_SLAB_RECLAIMABLE),
3246 global_page_state(NR_SLAB_UNRECLAIMABLE),
3247 global_page_state(NR_FILE_MAPPED),
3248 global_page_state(NR_SHMEM),
3249 global_page_state(NR_PAGETABLE),
3250 global_page_state(NR_BOUNCE),
3251 global_page_state(NR_FREE_CMA_PAGES));
3253 for_each_populated_zone(zone) {
3256 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3264 " active_anon:%lukB"
3265 " inactive_anon:%lukB"
3266 " active_file:%lukB"
3267 " inactive_file:%lukB"
3268 " unevictable:%lukB"
3269 " isolated(anon):%lukB"
3270 " isolated(file):%lukB"
3278 " slab_reclaimable:%lukB"
3279 " slab_unreclaimable:%lukB"
3280 " kernel_stack:%lukB"
3285 " writeback_tmp:%lukB"
3286 " pages_scanned:%lu"
3287 " all_unreclaimable? %s"
3290 K(zone_page_state(zone, NR_FREE_PAGES)),
3291 K(min_wmark_pages(zone)),
3292 K(low_wmark_pages(zone)),
3293 K(high_wmark_pages(zone)),
3294 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3295 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3296 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3297 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3298 K(zone_page_state(zone, NR_UNEVICTABLE)),
3299 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3300 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3301 K(zone->present_pages),
3302 K(zone->managed_pages),
3303 K(zone_page_state(zone, NR_MLOCK)),
3304 K(zone_page_state(zone, NR_FILE_DIRTY)),
3305 K(zone_page_state(zone, NR_WRITEBACK)),
3306 K(zone_page_state(zone, NR_FILE_MAPPED)),
3307 K(zone_page_state(zone, NR_SHMEM)),
3308 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3309 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3310 zone_page_state(zone, NR_KERNEL_STACK) *
3312 K(zone_page_state(zone, NR_PAGETABLE)),
3313 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3314 K(zone_page_state(zone, NR_BOUNCE)),
3315 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3316 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3317 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3318 (!zone_reclaimable(zone) ? "yes" : "no")
3320 printk("lowmem_reserve[]:");
3321 for (i = 0; i < MAX_NR_ZONES; i++)
3322 printk(" %ld", zone->lowmem_reserve[i]);
3326 for_each_populated_zone(zone) {
3327 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3328 unsigned char types[MAX_ORDER];
3330 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3333 printk("%s: ", zone->name);
3335 spin_lock_irqsave(&zone->lock, flags);
3336 for (order = 0; order < MAX_ORDER; order++) {
3337 struct free_area *area = &zone->free_area[order];
3340 nr[order] = area->nr_free;
3341 total += nr[order] << order;
3344 for (type = 0; type < MIGRATE_TYPES; type++) {
3345 if (!list_empty(&area->free_list[type]))
3346 types[order] |= 1 << type;
3349 spin_unlock_irqrestore(&zone->lock, flags);
3350 for (order = 0; order < MAX_ORDER; order++) {
3351 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3353 show_migration_types(types[order]);
3355 printk("= %lukB\n", K(total));
3358 hugetlb_show_meminfo();
3360 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3362 show_swap_cache_info();
3365 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3367 zoneref->zone = zone;
3368 zoneref->zone_idx = zone_idx(zone);
3372 * Builds allocation fallback zone lists.
3374 * Add all populated zones of a node to the zonelist.
3376 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3380 enum zone_type zone_type = MAX_NR_ZONES;
3384 zone = pgdat->node_zones + zone_type;
3385 if (populated_zone(zone)) {
3386 zoneref_set_zone(zone,
3387 &zonelist->_zonerefs[nr_zones++]);
3388 check_highest_zone(zone_type);
3390 } while (zone_type);
3398 * 0 = automatic detection of better ordering.
3399 * 1 = order by ([node] distance, -zonetype)
3400 * 2 = order by (-zonetype, [node] distance)
3402 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3403 * the same zonelist. So only NUMA can configure this param.
3405 #define ZONELIST_ORDER_DEFAULT 0
3406 #define ZONELIST_ORDER_NODE 1
3407 #define ZONELIST_ORDER_ZONE 2
3409 /* zonelist order in the kernel.
3410 * set_zonelist_order() will set this to NODE or ZONE.
3412 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3413 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3417 /* The value user specified ....changed by config */
3418 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3419 /* string for sysctl */
3420 #define NUMA_ZONELIST_ORDER_LEN 16
3421 char numa_zonelist_order[16] = "default";
3424 * interface for configure zonelist ordering.
3425 * command line option "numa_zonelist_order"
3426 * = "[dD]efault - default, automatic configuration.
3427 * = "[nN]ode - order by node locality, then by zone within node
3428 * = "[zZ]one - order by zone, then by locality within zone
3431 static int __parse_numa_zonelist_order(char *s)
3433 if (*s == 'd' || *s == 'D') {
3434 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3435 } else if (*s == 'n' || *s == 'N') {
3436 user_zonelist_order = ZONELIST_ORDER_NODE;
3437 } else if (*s == 'z' || *s == 'Z') {
3438 user_zonelist_order = ZONELIST_ORDER_ZONE;
3441 "Ignoring invalid numa_zonelist_order value: "
3448 static __init int setup_numa_zonelist_order(char *s)
3455 ret = __parse_numa_zonelist_order(s);
3457 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3461 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3464 * sysctl handler for numa_zonelist_order
3466 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3467 void __user *buffer, size_t *length,
3470 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3472 static DEFINE_MUTEX(zl_order_mutex);
3474 mutex_lock(&zl_order_mutex);
3476 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3480 strcpy(saved_string, (char *)table->data);
3482 ret = proc_dostring(table, write, buffer, length, ppos);
3486 int oldval = user_zonelist_order;
3488 ret = __parse_numa_zonelist_order((char *)table->data);
3491 * bogus value. restore saved string
3493 strncpy((char *)table->data, saved_string,
3494 NUMA_ZONELIST_ORDER_LEN);
3495 user_zonelist_order = oldval;
3496 } else if (oldval != user_zonelist_order) {
3497 mutex_lock(&zonelists_mutex);
3498 build_all_zonelists(NULL, NULL);
3499 mutex_unlock(&zonelists_mutex);
3503 mutex_unlock(&zl_order_mutex);
3508 #define MAX_NODE_LOAD (nr_online_nodes)
3509 static int node_load[MAX_NUMNODES];
3512 * find_next_best_node - find the next node that should appear in a given node's fallback list
3513 * @node: node whose fallback list we're appending
3514 * @used_node_mask: nodemask_t of already used nodes
3516 * We use a number of factors to determine which is the next node that should
3517 * appear on a given node's fallback list. The node should not have appeared
3518 * already in @node's fallback list, and it should be the next closest node
3519 * according to the distance array (which contains arbitrary distance values
3520 * from each node to each node in the system), and should also prefer nodes
3521 * with no CPUs, since presumably they'll have very little allocation pressure
3522 * on them otherwise.
3523 * It returns -1 if no node is found.
3525 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3528 int min_val = INT_MAX;
3529 int best_node = NUMA_NO_NODE;
3530 const struct cpumask *tmp = cpumask_of_node(0);
3532 /* Use the local node if we haven't already */
3533 if (!node_isset(node, *used_node_mask)) {
3534 node_set(node, *used_node_mask);
3538 for_each_node_state(n, N_MEMORY) {
3540 /* Don't want a node to appear more than once */
3541 if (node_isset(n, *used_node_mask))
3544 /* Use the distance array to find the distance */
3545 val = node_distance(node, n);
3547 /* Penalize nodes under us ("prefer the next node") */
3550 /* Give preference to headless and unused nodes */
3551 tmp = cpumask_of_node(n);
3552 if (!cpumask_empty(tmp))
3553 val += PENALTY_FOR_NODE_WITH_CPUS;
3555 /* Slight preference for less loaded node */
3556 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3557 val += node_load[n];
3559 if (val < min_val) {
3566 node_set(best_node, *used_node_mask);
3573 * Build zonelists ordered by node and zones within node.
3574 * This results in maximum locality--normal zone overflows into local
3575 * DMA zone, if any--but risks exhausting DMA zone.
3577 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3580 struct zonelist *zonelist;
3582 zonelist = &pgdat->node_zonelists[0];
3583 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3585 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3586 zonelist->_zonerefs[j].zone = NULL;
3587 zonelist->_zonerefs[j].zone_idx = 0;
3591 * Build gfp_thisnode zonelists
3593 static void build_thisnode_zonelists(pg_data_t *pgdat)
3596 struct zonelist *zonelist;
3598 zonelist = &pgdat->node_zonelists[1];
3599 j = build_zonelists_node(pgdat, zonelist, 0);
3600 zonelist->_zonerefs[j].zone = NULL;
3601 zonelist->_zonerefs[j].zone_idx = 0;
3605 * Build zonelists ordered by zone and nodes within zones.
3606 * This results in conserving DMA zone[s] until all Normal memory is
3607 * exhausted, but results in overflowing to remote node while memory
3608 * may still exist in local DMA zone.
3610 static int node_order[MAX_NUMNODES];
3612 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3615 int zone_type; /* needs to be signed */
3617 struct zonelist *zonelist;
3619 zonelist = &pgdat->node_zonelists[0];
3621 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3622 for (j = 0; j < nr_nodes; j++) {
3623 node = node_order[j];
3624 z = &NODE_DATA(node)->node_zones[zone_type];
3625 if (populated_zone(z)) {
3627 &zonelist->_zonerefs[pos++]);
3628 check_highest_zone(zone_type);
3632 zonelist->_zonerefs[pos].zone = NULL;
3633 zonelist->_zonerefs[pos].zone_idx = 0;
3636 #if defined(CONFIG_64BIT)
3638 * Devices that require DMA32/DMA are relatively rare and do not justify a
3639 * penalty to every machine in case the specialised case applies. Default
3640 * to Node-ordering on 64-bit NUMA machines
3642 static int default_zonelist_order(void)
3644 return ZONELIST_ORDER_NODE;
3648 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3649 * by the kernel. If processes running on node 0 deplete the low memory zone
3650 * then reclaim will occur more frequency increasing stalls and potentially
3651 * be easier to OOM if a large percentage of the zone is under writeback or
3652 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3653 * Hence, default to zone ordering on 32-bit.
3655 static int default_zonelist_order(void)
3657 return ZONELIST_ORDER_ZONE;
3659 #endif /* CONFIG_64BIT */
3661 static void set_zonelist_order(void)
3663 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3664 current_zonelist_order = default_zonelist_order();
3666 current_zonelist_order = user_zonelist_order;
3669 static void build_zonelists(pg_data_t *pgdat)
3673 nodemask_t used_mask;
3674 int local_node, prev_node;
3675 struct zonelist *zonelist;
3676 int order = current_zonelist_order;
3678 /* initialize zonelists */
3679 for (i = 0; i < MAX_ZONELISTS; i++) {
3680 zonelist = pgdat->node_zonelists + i;
3681 zonelist->_zonerefs[0].zone = NULL;
3682 zonelist->_zonerefs[0].zone_idx = 0;
3685 /* NUMA-aware ordering of nodes */
3686 local_node = pgdat->node_id;
3687 load = nr_online_nodes;
3688 prev_node = local_node;
3689 nodes_clear(used_mask);
3691 memset(node_order, 0, sizeof(node_order));
3694 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3696 * We don't want to pressure a particular node.
3697 * So adding penalty to the first node in same
3698 * distance group to make it round-robin.
3700 if (node_distance(local_node, node) !=
3701 node_distance(local_node, prev_node))
3702 node_load[node] = load;
3706 if (order == ZONELIST_ORDER_NODE)
3707 build_zonelists_in_node_order(pgdat, node);
3709 node_order[j++] = node; /* remember order */
3712 if (order == ZONELIST_ORDER_ZONE) {
3713 /* calculate node order -- i.e., DMA last! */
3714 build_zonelists_in_zone_order(pgdat, j);
3717 build_thisnode_zonelists(pgdat);
3720 /* Construct the zonelist performance cache - see further mmzone.h */
3721 static void build_zonelist_cache(pg_data_t *pgdat)
3723 struct zonelist *zonelist;
3724 struct zonelist_cache *zlc;
3727 zonelist = &pgdat->node_zonelists[0];
3728 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3729 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3730 for (z = zonelist->_zonerefs; z->zone; z++)
3731 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3734 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3736 * Return node id of node used for "local" allocations.
3737 * I.e., first node id of first zone in arg node's generic zonelist.
3738 * Used for initializing percpu 'numa_mem', which is used primarily
3739 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3741 int local_memory_node(int node)
3745 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3746 gfp_zone(GFP_KERNEL),
3753 #else /* CONFIG_NUMA */
3755 static void set_zonelist_order(void)
3757 current_zonelist_order = ZONELIST_ORDER_ZONE;
3760 static void build_zonelists(pg_data_t *pgdat)
3762 int node, local_node;
3764 struct zonelist *zonelist;
3766 local_node = pgdat->node_id;
3768 zonelist = &pgdat->node_zonelists[0];
3769 j = build_zonelists_node(pgdat, zonelist, 0);
3772 * Now we build the zonelist so that it contains the zones
3773 * of all the other nodes.
3774 * We don't want to pressure a particular node, so when
3775 * building the zones for node N, we make sure that the
3776 * zones coming right after the local ones are those from
3777 * node N+1 (modulo N)
3779 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3780 if (!node_online(node))
3782 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3784 for (node = 0; node < local_node; node++) {
3785 if (!node_online(node))
3787 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3790 zonelist->_zonerefs[j].zone = NULL;
3791 zonelist->_zonerefs[j].zone_idx = 0;
3794 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3795 static void build_zonelist_cache(pg_data_t *pgdat)
3797 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3800 #endif /* CONFIG_NUMA */
3803 * Boot pageset table. One per cpu which is going to be used for all
3804 * zones and all nodes. The parameters will be set in such a way
3805 * that an item put on a list will immediately be handed over to
3806 * the buddy list. This is safe since pageset manipulation is done
3807 * with interrupts disabled.
3809 * The boot_pagesets must be kept even after bootup is complete for
3810 * unused processors and/or zones. They do play a role for bootstrapping
3811 * hotplugged processors.
3813 * zoneinfo_show() and maybe other functions do
3814 * not check if the processor is online before following the pageset pointer.
3815 * Other parts of the kernel may not check if the zone is available.
3817 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3818 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3819 static void setup_zone_pageset(struct zone *zone);
3822 * Global mutex to protect against size modification of zonelists
3823 * as well as to serialize pageset setup for the new populated zone.
3825 DEFINE_MUTEX(zonelists_mutex);
3827 /* return values int ....just for stop_machine() */
3828 static int __build_all_zonelists(void *data)
3832 pg_data_t *self = data;
3835 memset(node_load, 0, sizeof(node_load));
3838 if (self && !node_online(self->node_id)) {
3839 build_zonelists(self);
3840 build_zonelist_cache(self);
3843 for_each_online_node(nid) {
3844 pg_data_t *pgdat = NODE_DATA(nid);
3846 build_zonelists(pgdat);
3847 build_zonelist_cache(pgdat);
3851 * Initialize the boot_pagesets that are going to be used
3852 * for bootstrapping processors. The real pagesets for
3853 * each zone will be allocated later when the per cpu
3854 * allocator is available.
3856 * boot_pagesets are used also for bootstrapping offline
3857 * cpus if the system is already booted because the pagesets
3858 * are needed to initialize allocators on a specific cpu too.
3859 * F.e. the percpu allocator needs the page allocator which
3860 * needs the percpu allocator in order to allocate its pagesets
3861 * (a chicken-egg dilemma).
3863 for_each_possible_cpu(cpu) {
3864 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3866 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3868 * We now know the "local memory node" for each node--
3869 * i.e., the node of the first zone in the generic zonelist.
3870 * Set up numa_mem percpu variable for on-line cpus. During
3871 * boot, only the boot cpu should be on-line; we'll init the
3872 * secondary cpus' numa_mem as they come on-line. During
3873 * node/memory hotplug, we'll fixup all on-line cpus.
3875 if (cpu_online(cpu))
3876 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3884 * Called with zonelists_mutex held always
3885 * unless system_state == SYSTEM_BOOTING.
3887 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3889 set_zonelist_order();
3891 if (system_state == SYSTEM_BOOTING) {
3892 __build_all_zonelists(NULL);
3893 mminit_verify_zonelist();
3894 cpuset_init_current_mems_allowed();
3896 #ifdef CONFIG_MEMORY_HOTPLUG
3898 setup_zone_pageset(zone);
3900 /* we have to stop all cpus to guarantee there is no user
3902 stop_machine(__build_all_zonelists, pgdat, NULL);
3903 /* cpuset refresh routine should be here */
3905 vm_total_pages = nr_free_pagecache_pages();
3907 * Disable grouping by mobility if the number of pages in the
3908 * system is too low to allow the mechanism to work. It would be
3909 * more accurate, but expensive to check per-zone. This check is
3910 * made on memory-hotadd so a system can start with mobility
3911 * disabled and enable it later
3913 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3914 page_group_by_mobility_disabled = 1;
3916 page_group_by_mobility_disabled = 0;
3918 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3919 "Total pages: %ld\n",
3921 zonelist_order_name[current_zonelist_order],
3922 page_group_by_mobility_disabled ? "off" : "on",
3925 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
3930 * Helper functions to size the waitqueue hash table.
3931 * Essentially these want to choose hash table sizes sufficiently
3932 * large so that collisions trying to wait on pages are rare.
3933 * But in fact, the number of active page waitqueues on typical
3934 * systems is ridiculously low, less than 200. So this is even
3935 * conservative, even though it seems large.
3937 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3938 * waitqueues, i.e. the size of the waitq table given the number of pages.
3940 #define PAGES_PER_WAITQUEUE 256
3942 #ifndef CONFIG_MEMORY_HOTPLUG
3943 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3945 unsigned long size = 1;
3947 pages /= PAGES_PER_WAITQUEUE;
3949 while (size < pages)
3953 * Once we have dozens or even hundreds of threads sleeping
3954 * on IO we've got bigger problems than wait queue collision.
3955 * Limit the size of the wait table to a reasonable size.
3957 size = min(size, 4096UL);
3959 return max(size, 4UL);
3963 * A zone's size might be changed by hot-add, so it is not possible to determine
3964 * a suitable size for its wait_table. So we use the maximum size now.
3966 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3968 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3969 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3970 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3972 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3973 * or more by the traditional way. (See above). It equals:
3975 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3976 * ia64(16K page size) : = ( 8G + 4M)byte.
3977 * powerpc (64K page size) : = (32G +16M)byte.
3979 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3986 * This is an integer logarithm so that shifts can be used later
3987 * to extract the more random high bits from the multiplicative
3988 * hash function before the remainder is taken.
3990 static inline unsigned long wait_table_bits(unsigned long size)
3996 * Check if a pageblock contains reserved pages
3998 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4002 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4003 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4010 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4011 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4012 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4013 * higher will lead to a bigger reserve which will get freed as contiguous
4014 * blocks as reclaim kicks in
4016 static void setup_zone_migrate_reserve(struct zone *zone)
4018 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4020 unsigned long block_migratetype;
4025 * Get the start pfn, end pfn and the number of blocks to reserve
4026 * We have to be careful to be aligned to pageblock_nr_pages to
4027 * make sure that we always check pfn_valid for the first page in
4030 start_pfn = zone->zone_start_pfn;
4031 end_pfn = zone_end_pfn(zone);
4032 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4033 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4037 * Reserve blocks are generally in place to help high-order atomic
4038 * allocations that are short-lived. A min_free_kbytes value that
4039 * would result in more than 2 reserve blocks for atomic allocations
4040 * is assumed to be in place to help anti-fragmentation for the
4041 * future allocation of hugepages at runtime.
4043 reserve = min(2, reserve);
4044 old_reserve = zone->nr_migrate_reserve_block;
4046 /* When memory hot-add, we almost always need to do nothing */
4047 if (reserve == old_reserve)
4049 zone->nr_migrate_reserve_block = reserve;
4051 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4052 if (!pfn_valid(pfn))
4054 page = pfn_to_page(pfn);
4056 /* Watch out for overlapping nodes */
4057 if (page_to_nid(page) != zone_to_nid(zone))
4060 block_migratetype = get_pageblock_migratetype(page);
4062 /* Only test what is necessary when the reserves are not met */
4065 * Blocks with reserved pages will never free, skip
4068 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4069 if (pageblock_is_reserved(pfn, block_end_pfn))
4072 /* If this block is reserved, account for it */
4073 if (block_migratetype == MIGRATE_RESERVE) {
4078 /* Suitable for reserving if this block is movable */
4079 if (block_migratetype == MIGRATE_MOVABLE) {
4080 set_pageblock_migratetype(page,
4082 move_freepages_block(zone, page,
4087 } else if (!old_reserve) {
4089 * At boot time we don't need to scan the whole zone
4090 * for turning off MIGRATE_RESERVE.
4096 * If the reserve is met and this is a previous reserved block,
4099 if (block_migratetype == MIGRATE_RESERVE) {
4100 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4101 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4107 * Initially all pages are reserved - free ones are freed
4108 * up by free_all_bootmem() once the early boot process is
4109 * done. Non-atomic initialization, single-pass.
4111 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4112 unsigned long start_pfn, enum memmap_context context)
4115 unsigned long end_pfn = start_pfn + size;
4119 if (highest_memmap_pfn < end_pfn - 1)
4120 highest_memmap_pfn = end_pfn - 1;
4122 z = &NODE_DATA(nid)->node_zones[zone];
4123 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4125 * There can be holes in boot-time mem_map[]s
4126 * handed to this function. They do not
4127 * exist on hotplugged memory.
4129 if (context == MEMMAP_EARLY) {
4130 if (!early_pfn_valid(pfn))
4132 if (!early_pfn_in_nid(pfn, nid))
4135 page = pfn_to_page(pfn);
4136 set_page_links(page, zone, nid, pfn);
4137 mminit_verify_page_links(page, zone, nid, pfn);
4138 init_page_count(page);
4139 page_mapcount_reset(page);
4140 page_cpupid_reset_last(page);
4141 SetPageReserved(page);
4143 * Mark the block movable so that blocks are reserved for
4144 * movable at startup. This will force kernel allocations
4145 * to reserve their blocks rather than leaking throughout
4146 * the address space during boot when many long-lived
4147 * kernel allocations are made. Later some blocks near
4148 * the start are marked MIGRATE_RESERVE by
4149 * setup_zone_migrate_reserve()
4151 * bitmap is created for zone's valid pfn range. but memmap
4152 * can be created for invalid pages (for alignment)
4153 * check here not to call set_pageblock_migratetype() against
4156 if ((z->zone_start_pfn <= pfn)
4157 && (pfn < zone_end_pfn(z))
4158 && !(pfn & (pageblock_nr_pages - 1)))
4159 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4161 INIT_LIST_HEAD(&page->lru);
4162 #ifdef WANT_PAGE_VIRTUAL
4163 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4164 if (!is_highmem_idx(zone))
4165 set_page_address(page, __va(pfn << PAGE_SHIFT));
4170 static void __meminit zone_init_free_lists(struct zone *zone)
4172 unsigned int order, t;
4173 for_each_migratetype_order(order, t) {
4174 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4175 zone->free_area[order].nr_free = 0;
4179 #ifndef __HAVE_ARCH_MEMMAP_INIT
4180 #define memmap_init(size, nid, zone, start_pfn) \
4181 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4184 static int zone_batchsize(struct zone *zone)
4190 * The per-cpu-pages pools are set to around 1000th of the
4191 * size of the zone. But no more than 1/2 of a meg.
4193 * OK, so we don't know how big the cache is. So guess.
4195 batch = zone->managed_pages / 1024;
4196 if (batch * PAGE_SIZE > 512 * 1024)
4197 batch = (512 * 1024) / PAGE_SIZE;
4198 batch /= 4; /* We effectively *= 4 below */
4203 * Clamp the batch to a 2^n - 1 value. Having a power
4204 * of 2 value was found to be more likely to have
4205 * suboptimal cache aliasing properties in some cases.
4207 * For example if 2 tasks are alternately allocating
4208 * batches of pages, one task can end up with a lot
4209 * of pages of one half of the possible page colors
4210 * and the other with pages of the other colors.
4212 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4217 /* The deferral and batching of frees should be suppressed under NOMMU
4220 * The problem is that NOMMU needs to be able to allocate large chunks
4221 * of contiguous memory as there's no hardware page translation to
4222 * assemble apparent contiguous memory from discontiguous pages.
4224 * Queueing large contiguous runs of pages for batching, however,
4225 * causes the pages to actually be freed in smaller chunks. As there
4226 * can be a significant delay between the individual batches being
4227 * recycled, this leads to the once large chunks of space being
4228 * fragmented and becoming unavailable for high-order allocations.
4235 * pcp->high and pcp->batch values are related and dependent on one another:
4236 * ->batch must never be higher then ->high.
4237 * The following function updates them in a safe manner without read side
4240 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4241 * those fields changing asynchronously (acording the the above rule).
4243 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4244 * outside of boot time (or some other assurance that no concurrent updaters
4247 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4248 unsigned long batch)
4250 /* start with a fail safe value for batch */
4254 /* Update high, then batch, in order */
4261 /* a companion to pageset_set_high() */
4262 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4264 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4267 static void pageset_init(struct per_cpu_pageset *p)
4269 struct per_cpu_pages *pcp;
4272 memset(p, 0, sizeof(*p));
4276 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4277 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4280 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4283 pageset_set_batch(p, batch);
4287 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4288 * to the value high for the pageset p.
4290 static void pageset_set_high(struct per_cpu_pageset *p,
4293 unsigned long batch = max(1UL, high / 4);
4294 if ((high / 4) > (PAGE_SHIFT * 8))
4295 batch = PAGE_SHIFT * 8;
4297 pageset_update(&p->pcp, high, batch);
4300 static void pageset_set_high_and_batch(struct zone *zone,
4301 struct per_cpu_pageset *pcp)
4303 if (percpu_pagelist_fraction)
4304 pageset_set_high(pcp,
4305 (zone->managed_pages /
4306 percpu_pagelist_fraction));
4308 pageset_set_batch(pcp, zone_batchsize(zone));
4311 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4313 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4316 pageset_set_high_and_batch(zone, pcp);
4319 static void __meminit setup_zone_pageset(struct zone *zone)
4322 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4323 for_each_possible_cpu(cpu)
4324 zone_pageset_init(zone, cpu);
4328 * Allocate per cpu pagesets and initialize them.
4329 * Before this call only boot pagesets were available.
4331 void __init setup_per_cpu_pageset(void)
4335 for_each_populated_zone(zone)
4336 setup_zone_pageset(zone);
4339 static noinline __init_refok
4340 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4346 * The per-page waitqueue mechanism uses hashed waitqueues
4349 zone->wait_table_hash_nr_entries =
4350 wait_table_hash_nr_entries(zone_size_pages);
4351 zone->wait_table_bits =
4352 wait_table_bits(zone->wait_table_hash_nr_entries);
4353 alloc_size = zone->wait_table_hash_nr_entries
4354 * sizeof(wait_queue_head_t);
4356 if (!slab_is_available()) {
4357 zone->wait_table = (wait_queue_head_t *)
4358 memblock_virt_alloc_node_nopanic(
4359 alloc_size, zone->zone_pgdat->node_id);
4362 * This case means that a zone whose size was 0 gets new memory
4363 * via memory hot-add.
4364 * But it may be the case that a new node was hot-added. In
4365 * this case vmalloc() will not be able to use this new node's
4366 * memory - this wait_table must be initialized to use this new
4367 * node itself as well.
4368 * To use this new node's memory, further consideration will be
4371 zone->wait_table = vmalloc(alloc_size);
4373 if (!zone->wait_table)
4376 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4377 init_waitqueue_head(zone->wait_table + i);
4382 static __meminit void zone_pcp_init(struct zone *zone)
4385 * per cpu subsystem is not up at this point. The following code
4386 * relies on the ability of the linker to provide the
4387 * offset of a (static) per cpu variable into the per cpu area.
4389 zone->pageset = &boot_pageset;
4391 if (populated_zone(zone))
4392 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4393 zone->name, zone->present_pages,
4394 zone_batchsize(zone));
4397 int __meminit init_currently_empty_zone(struct zone *zone,
4398 unsigned long zone_start_pfn,
4400 enum memmap_context context)
4402 struct pglist_data *pgdat = zone->zone_pgdat;
4404 ret = zone_wait_table_init(zone, size);
4407 pgdat->nr_zones = zone_idx(zone) + 1;
4409 zone->zone_start_pfn = zone_start_pfn;
4411 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4412 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4414 (unsigned long)zone_idx(zone),
4415 zone_start_pfn, (zone_start_pfn + size));
4417 zone_init_free_lists(zone);
4422 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4423 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4425 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4427 int __meminit __early_pfn_to_nid(unsigned long pfn)
4429 unsigned long start_pfn, end_pfn;
4432 * NOTE: The following SMP-unsafe globals are only used early in boot
4433 * when the kernel is running single-threaded.
4435 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4436 static int __meminitdata last_nid;
4438 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4441 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4443 last_start_pfn = start_pfn;
4444 last_end_pfn = end_pfn;
4450 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4452 int __meminit early_pfn_to_nid(unsigned long pfn)
4456 nid = __early_pfn_to_nid(pfn);
4459 /* just returns 0 */
4463 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4464 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4468 nid = __early_pfn_to_nid(pfn);
4469 if (nid >= 0 && nid != node)
4476 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4477 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4478 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4480 * If an architecture guarantees that all ranges registered contain no holes
4481 * and may be freed, this this function may be used instead of calling
4482 * memblock_free_early_nid() manually.
4484 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4486 unsigned long start_pfn, end_pfn;
4489 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4490 start_pfn = min(start_pfn, max_low_pfn);
4491 end_pfn = min(end_pfn, max_low_pfn);
4493 if (start_pfn < end_pfn)
4494 memblock_free_early_nid(PFN_PHYS(start_pfn),
4495 (end_pfn - start_pfn) << PAGE_SHIFT,
4501 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4502 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4504 * If an architecture guarantees that all ranges registered contain no holes and may
4505 * be freed, this function may be used instead of calling memory_present() manually.
4507 void __init sparse_memory_present_with_active_regions(int nid)
4509 unsigned long start_pfn, end_pfn;
4512 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4513 memory_present(this_nid, start_pfn, end_pfn);
4517 * get_pfn_range_for_nid - Return the start and end page frames for a node
4518 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4519 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4520 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4522 * It returns the start and end page frame of a node based on information
4523 * provided by memblock_set_node(). If called for a node
4524 * with no available memory, a warning is printed and the start and end
4527 void __meminit get_pfn_range_for_nid(unsigned int nid,
4528 unsigned long *start_pfn, unsigned long *end_pfn)
4530 unsigned long this_start_pfn, this_end_pfn;
4536 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4537 *start_pfn = min(*start_pfn, this_start_pfn);
4538 *end_pfn = max(*end_pfn, this_end_pfn);
4541 if (*start_pfn == -1UL)
4546 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4547 * assumption is made that zones within a node are ordered in monotonic
4548 * increasing memory addresses so that the "highest" populated zone is used
4550 static void __init find_usable_zone_for_movable(void)
4553 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4554 if (zone_index == ZONE_MOVABLE)
4557 if (arch_zone_highest_possible_pfn[zone_index] >
4558 arch_zone_lowest_possible_pfn[zone_index])
4562 VM_BUG_ON(zone_index == -1);
4563 movable_zone = zone_index;
4567 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4568 * because it is sized independent of architecture. Unlike the other zones,
4569 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4570 * in each node depending on the size of each node and how evenly kernelcore
4571 * is distributed. This helper function adjusts the zone ranges
4572 * provided by the architecture for a given node by using the end of the
4573 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4574 * zones within a node are in order of monotonic increases memory addresses
4576 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4577 unsigned long zone_type,
4578 unsigned long node_start_pfn,
4579 unsigned long node_end_pfn,
4580 unsigned long *zone_start_pfn,
4581 unsigned long *zone_end_pfn)
4583 /* Only adjust if ZONE_MOVABLE is on this node */
4584 if (zone_movable_pfn[nid]) {
4585 /* Size ZONE_MOVABLE */
4586 if (zone_type == ZONE_MOVABLE) {
4587 *zone_start_pfn = zone_movable_pfn[nid];
4588 *zone_end_pfn = min(node_end_pfn,
4589 arch_zone_highest_possible_pfn[movable_zone]);
4591 /* Adjust for ZONE_MOVABLE starting within this range */
4592 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4593 *zone_end_pfn > zone_movable_pfn[nid]) {
4594 *zone_end_pfn = zone_movable_pfn[nid];
4596 /* Check if this whole range is within ZONE_MOVABLE */
4597 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4598 *zone_start_pfn = *zone_end_pfn;
4603 * Return the number of pages a zone spans in a node, including holes
4604 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4606 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4607 unsigned long zone_type,
4608 unsigned long node_start_pfn,
4609 unsigned long node_end_pfn,
4610 unsigned long *ignored)
4612 unsigned long zone_start_pfn, zone_end_pfn;
4614 /* Get the start and end of the zone */
4615 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4616 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4617 adjust_zone_range_for_zone_movable(nid, zone_type,
4618 node_start_pfn, node_end_pfn,
4619 &zone_start_pfn, &zone_end_pfn);
4621 /* Check that this node has pages within the zone's required range */
4622 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4625 /* Move the zone boundaries inside the node if necessary */
4626 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4627 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4629 /* Return the spanned pages */
4630 return zone_end_pfn - zone_start_pfn;
4634 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4635 * then all holes in the requested range will be accounted for.
4637 unsigned long __meminit __absent_pages_in_range(int nid,
4638 unsigned long range_start_pfn,
4639 unsigned long range_end_pfn)
4641 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4642 unsigned long start_pfn, end_pfn;
4645 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4646 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4647 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4648 nr_absent -= end_pfn - start_pfn;
4654 * absent_pages_in_range - Return number of page frames in holes within a range
4655 * @start_pfn: The start PFN to start searching for holes
4656 * @end_pfn: The end PFN to stop searching for holes
4658 * It returns the number of pages frames in memory holes within a range.
4660 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4661 unsigned long end_pfn)
4663 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4666 /* Return the number of page frames in holes in a zone on a node */
4667 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4668 unsigned long zone_type,
4669 unsigned long node_start_pfn,
4670 unsigned long node_end_pfn,
4671 unsigned long *ignored)
4673 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4674 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4675 unsigned long zone_start_pfn, zone_end_pfn;
4677 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4678 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4680 adjust_zone_range_for_zone_movable(nid, zone_type,
4681 node_start_pfn, node_end_pfn,
4682 &zone_start_pfn, &zone_end_pfn);
4683 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4686 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4687 static inline unsigned long __meminit zone_spanned_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 *zones_size)
4693 return zones_size[zone_type];
4696 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4697 unsigned long zone_type,
4698 unsigned long node_start_pfn,
4699 unsigned long node_end_pfn,
4700 unsigned long *zholes_size)
4705 return zholes_size[zone_type];
4708 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4710 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4711 unsigned long node_start_pfn,
4712 unsigned long node_end_pfn,
4713 unsigned long *zones_size,
4714 unsigned long *zholes_size)
4716 unsigned long realtotalpages, totalpages = 0;
4719 for (i = 0; i < MAX_NR_ZONES; i++)
4720 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4724 pgdat->node_spanned_pages = totalpages;
4726 realtotalpages = totalpages;
4727 for (i = 0; i < MAX_NR_ZONES; i++)
4729 zone_absent_pages_in_node(pgdat->node_id, i,
4730 node_start_pfn, node_end_pfn,
4732 pgdat->node_present_pages = realtotalpages;
4733 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4737 #ifndef CONFIG_SPARSEMEM
4739 * Calculate the size of the zone->blockflags rounded to an unsigned long
4740 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4741 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4742 * round what is now in bits to nearest long in bits, then return it in
4745 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4747 unsigned long usemapsize;
4749 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4750 usemapsize = roundup(zonesize, pageblock_nr_pages);
4751 usemapsize = usemapsize >> pageblock_order;
4752 usemapsize *= NR_PAGEBLOCK_BITS;
4753 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4755 return usemapsize / 8;
4758 static void __init setup_usemap(struct pglist_data *pgdat,
4760 unsigned long zone_start_pfn,
4761 unsigned long zonesize)
4763 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4764 zone->pageblock_flags = NULL;
4766 zone->pageblock_flags =
4767 memblock_virt_alloc_node_nopanic(usemapsize,
4771 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4772 unsigned long zone_start_pfn, unsigned long zonesize) {}
4773 #endif /* CONFIG_SPARSEMEM */
4775 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4777 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4778 void __paginginit set_pageblock_order(void)
4782 /* Check that pageblock_nr_pages has not already been setup */
4783 if (pageblock_order)
4786 if (HPAGE_SHIFT > PAGE_SHIFT)
4787 order = HUGETLB_PAGE_ORDER;
4789 order = MAX_ORDER - 1;
4792 * Assume the largest contiguous order of interest is a huge page.
4793 * This value may be variable depending on boot parameters on IA64 and
4796 pageblock_order = order;
4798 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4801 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4802 * is unused as pageblock_order is set at compile-time. See
4803 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4806 void __paginginit set_pageblock_order(void)
4810 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4812 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4813 unsigned long present_pages)
4815 unsigned long pages = spanned_pages;
4818 * Provide a more accurate estimation if there are holes within
4819 * the zone and SPARSEMEM is in use. If there are holes within the
4820 * zone, each populated memory region may cost us one or two extra
4821 * memmap pages due to alignment because memmap pages for each
4822 * populated regions may not naturally algined on page boundary.
4823 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4825 if (spanned_pages > present_pages + (present_pages >> 4) &&
4826 IS_ENABLED(CONFIG_SPARSEMEM))
4827 pages = present_pages;
4829 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4833 * Set up the zone data structures:
4834 * - mark all pages reserved
4835 * - mark all memory queues empty
4836 * - clear the memory bitmaps
4838 * NOTE: pgdat should get zeroed by caller.
4840 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4841 unsigned long node_start_pfn, unsigned long node_end_pfn,
4842 unsigned long *zones_size, unsigned long *zholes_size)
4845 int nid = pgdat->node_id;
4846 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4849 pgdat_resize_init(pgdat);
4850 #ifdef CONFIG_NUMA_BALANCING
4851 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4852 pgdat->numabalancing_migrate_nr_pages = 0;
4853 pgdat->numabalancing_migrate_next_window = jiffies;
4855 init_waitqueue_head(&pgdat->kswapd_wait);
4856 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4857 pgdat_page_ext_init(pgdat);
4859 for (j = 0; j < MAX_NR_ZONES; j++) {
4860 struct zone *zone = pgdat->node_zones + j;
4861 unsigned long size, realsize, freesize, memmap_pages;
4863 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4864 node_end_pfn, zones_size);
4865 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4871 * Adjust freesize so that it accounts for how much memory
4872 * is used by this zone for memmap. This affects the watermark
4873 * and per-cpu initialisations
4875 memmap_pages = calc_memmap_size(size, realsize);
4876 if (!is_highmem_idx(j)) {
4877 if (freesize >= memmap_pages) {
4878 freesize -= memmap_pages;
4881 " %s zone: %lu pages used for memmap\n",
4882 zone_names[j], memmap_pages);
4885 " %s zone: %lu pages exceeds freesize %lu\n",
4886 zone_names[j], memmap_pages, freesize);
4889 /* Account for reserved pages */
4890 if (j == 0 && freesize > dma_reserve) {
4891 freesize -= dma_reserve;
4892 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4893 zone_names[0], dma_reserve);
4896 if (!is_highmem_idx(j))
4897 nr_kernel_pages += freesize;
4898 /* Charge for highmem memmap if there are enough kernel pages */
4899 else if (nr_kernel_pages > memmap_pages * 2)
4900 nr_kernel_pages -= memmap_pages;
4901 nr_all_pages += freesize;
4903 zone->spanned_pages = size;
4904 zone->present_pages = realsize;
4906 * Set an approximate value for lowmem here, it will be adjusted
4907 * when the bootmem allocator frees pages into the buddy system.
4908 * And all highmem pages will be managed by the buddy system.
4910 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4913 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4915 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4917 zone->name = zone_names[j];
4918 spin_lock_init(&zone->lock);
4919 spin_lock_init(&zone->lru_lock);
4920 zone_seqlock_init(zone);
4921 zone->zone_pgdat = pgdat;
4922 zone_pcp_init(zone);
4924 /* For bootup, initialized properly in watermark setup */
4925 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4927 lruvec_init(&zone->lruvec);
4931 set_pageblock_order();
4932 setup_usemap(pgdat, zone, zone_start_pfn, size);
4933 ret = init_currently_empty_zone(zone, zone_start_pfn,
4934 size, MEMMAP_EARLY);
4936 memmap_init(size, nid, j, zone_start_pfn);
4937 zone_start_pfn += size;
4941 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4943 /* Skip empty nodes */
4944 if (!pgdat->node_spanned_pages)
4947 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4948 /* ia64 gets its own node_mem_map, before this, without bootmem */
4949 if (!pgdat->node_mem_map) {
4950 unsigned long size, start, end;
4954 * The zone's endpoints aren't required to be MAX_ORDER
4955 * aligned but the node_mem_map endpoints must be in order
4956 * for the buddy allocator to function correctly.
4958 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4959 end = pgdat_end_pfn(pgdat);
4960 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4961 size = (end - start) * sizeof(struct page);
4962 map = alloc_remap(pgdat->node_id, size);
4964 map = memblock_virt_alloc_node_nopanic(size,
4966 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4968 #ifndef CONFIG_NEED_MULTIPLE_NODES
4970 * With no DISCONTIG, the global mem_map is just set as node 0's
4972 if (pgdat == NODE_DATA(0)) {
4973 mem_map = NODE_DATA(0)->node_mem_map;
4974 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4975 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4976 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4977 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4980 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4983 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4984 unsigned long node_start_pfn, unsigned long *zholes_size)
4986 pg_data_t *pgdat = NODE_DATA(nid);
4987 unsigned long start_pfn = 0;
4988 unsigned long end_pfn = 0;
4990 /* pg_data_t should be reset to zero when it's allocated */
4991 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4993 pgdat->node_id = nid;
4994 pgdat->node_start_pfn = node_start_pfn;
4995 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4996 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4997 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
4998 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
5000 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5001 zones_size, zholes_size);
5003 alloc_node_mem_map(pgdat);
5004 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5005 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5006 nid, (unsigned long)pgdat,
5007 (unsigned long)pgdat->node_mem_map);
5010 free_area_init_core(pgdat, start_pfn, end_pfn,
5011 zones_size, zholes_size);
5014 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5016 #if MAX_NUMNODES > 1
5018 * Figure out the number of possible node ids.
5020 void __init setup_nr_node_ids(void)
5023 unsigned int highest = 0;
5025 for_each_node_mask(node, node_possible_map)
5027 nr_node_ids = highest + 1;
5032 * node_map_pfn_alignment - determine the maximum internode alignment
5034 * This function should be called after node map is populated and sorted.
5035 * It calculates the maximum power of two alignment which can distinguish
5038 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5039 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5040 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5041 * shifted, 1GiB is enough and this function will indicate so.
5043 * This is used to test whether pfn -> nid mapping of the chosen memory
5044 * model has fine enough granularity to avoid incorrect mapping for the
5045 * populated node map.
5047 * Returns the determined alignment in pfn's. 0 if there is no alignment
5048 * requirement (single node).
5050 unsigned long __init node_map_pfn_alignment(void)
5052 unsigned long accl_mask = 0, last_end = 0;
5053 unsigned long start, end, mask;
5057 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5058 if (!start || last_nid < 0 || last_nid == nid) {
5065 * Start with a mask granular enough to pin-point to the
5066 * start pfn and tick off bits one-by-one until it becomes
5067 * too coarse to separate the current node from the last.
5069 mask = ~((1 << __ffs(start)) - 1);
5070 while (mask && last_end <= (start & (mask << 1)))
5073 /* accumulate all internode masks */
5077 /* convert mask to number of pages */
5078 return ~accl_mask + 1;
5081 /* Find the lowest pfn for a node */
5082 static unsigned long __init find_min_pfn_for_node(int nid)
5084 unsigned long min_pfn = ULONG_MAX;
5085 unsigned long start_pfn;
5088 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5089 min_pfn = min(min_pfn, start_pfn);
5091 if (min_pfn == ULONG_MAX) {
5093 "Could not find start_pfn for node %d\n", nid);
5101 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5103 * It returns the minimum PFN based on information provided via
5104 * memblock_set_node().
5106 unsigned long __init find_min_pfn_with_active_regions(void)
5108 return find_min_pfn_for_node(MAX_NUMNODES);
5112 * early_calculate_totalpages()
5113 * Sum pages in active regions for movable zone.
5114 * Populate N_MEMORY for calculating usable_nodes.
5116 static unsigned long __init early_calculate_totalpages(void)
5118 unsigned long totalpages = 0;
5119 unsigned long start_pfn, end_pfn;
5122 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5123 unsigned long pages = end_pfn - start_pfn;
5125 totalpages += pages;
5127 node_set_state(nid, N_MEMORY);
5133 * Find the PFN the Movable zone begins in each node. Kernel memory
5134 * is spread evenly between nodes as long as the nodes have enough
5135 * memory. When they don't, some nodes will have more kernelcore than
5138 static void __init find_zone_movable_pfns_for_nodes(void)
5141 unsigned long usable_startpfn;
5142 unsigned long kernelcore_node, kernelcore_remaining;
5143 /* save the state before borrow the nodemask */
5144 nodemask_t saved_node_state = node_states[N_MEMORY];
5145 unsigned long totalpages = early_calculate_totalpages();
5146 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5147 struct memblock_region *r;
5149 /* Need to find movable_zone earlier when movable_node is specified. */
5150 find_usable_zone_for_movable();
5153 * If movable_node is specified, ignore kernelcore and movablecore
5156 if (movable_node_is_enabled()) {
5157 for_each_memblock(memory, r) {
5158 if (!memblock_is_hotpluggable(r))
5163 usable_startpfn = PFN_DOWN(r->base);
5164 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5165 min(usable_startpfn, zone_movable_pfn[nid]) :
5173 * If movablecore=nn[KMG] was specified, calculate what size of
5174 * kernelcore that corresponds so that memory usable for
5175 * any allocation type is evenly spread. If both kernelcore
5176 * and movablecore are specified, then the value of kernelcore
5177 * will be used for required_kernelcore if it's greater than
5178 * what movablecore would have allowed.
5180 if (required_movablecore) {
5181 unsigned long corepages;
5184 * Round-up so that ZONE_MOVABLE is at least as large as what
5185 * was requested by the user
5187 required_movablecore =
5188 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5189 corepages = totalpages - required_movablecore;
5191 required_kernelcore = max(required_kernelcore, corepages);
5194 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5195 if (!required_kernelcore)
5198 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5199 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5202 /* Spread kernelcore memory as evenly as possible throughout nodes */
5203 kernelcore_node = required_kernelcore / usable_nodes;
5204 for_each_node_state(nid, N_MEMORY) {
5205 unsigned long start_pfn, end_pfn;
5208 * Recalculate kernelcore_node if the division per node
5209 * now exceeds what is necessary to satisfy the requested
5210 * amount of memory for the kernel
5212 if (required_kernelcore < kernelcore_node)
5213 kernelcore_node = required_kernelcore / usable_nodes;
5216 * As the map is walked, we track how much memory is usable
5217 * by the kernel using kernelcore_remaining. When it is
5218 * 0, the rest of the node is usable by ZONE_MOVABLE
5220 kernelcore_remaining = kernelcore_node;
5222 /* Go through each range of PFNs within this node */
5223 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5224 unsigned long size_pages;
5226 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5227 if (start_pfn >= end_pfn)
5230 /* Account for what is only usable for kernelcore */
5231 if (start_pfn < usable_startpfn) {
5232 unsigned long kernel_pages;
5233 kernel_pages = min(end_pfn, usable_startpfn)
5236 kernelcore_remaining -= min(kernel_pages,
5237 kernelcore_remaining);
5238 required_kernelcore -= min(kernel_pages,
5239 required_kernelcore);
5241 /* Continue if range is now fully accounted */
5242 if (end_pfn <= usable_startpfn) {
5245 * Push zone_movable_pfn to the end so
5246 * that if we have to rebalance
5247 * kernelcore across nodes, we will
5248 * not double account here
5250 zone_movable_pfn[nid] = end_pfn;
5253 start_pfn = usable_startpfn;
5257 * The usable PFN range for ZONE_MOVABLE is from
5258 * start_pfn->end_pfn. Calculate size_pages as the
5259 * number of pages used as kernelcore
5261 size_pages = end_pfn - start_pfn;
5262 if (size_pages > kernelcore_remaining)
5263 size_pages = kernelcore_remaining;
5264 zone_movable_pfn[nid] = start_pfn + size_pages;
5267 * Some kernelcore has been met, update counts and
5268 * break if the kernelcore for this node has been
5271 required_kernelcore -= min(required_kernelcore,
5273 kernelcore_remaining -= size_pages;
5274 if (!kernelcore_remaining)
5280 * If there is still required_kernelcore, we do another pass with one
5281 * less node in the count. This will push zone_movable_pfn[nid] further
5282 * along on the nodes that still have memory until kernelcore is
5286 if (usable_nodes && required_kernelcore > usable_nodes)
5290 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5291 for (nid = 0; nid < MAX_NUMNODES; nid++)
5292 zone_movable_pfn[nid] =
5293 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5296 /* restore the node_state */
5297 node_states[N_MEMORY] = saved_node_state;
5300 /* Any regular or high memory on that node ? */
5301 static void check_for_memory(pg_data_t *pgdat, int nid)
5303 enum zone_type zone_type;
5305 if (N_MEMORY == N_NORMAL_MEMORY)
5308 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5309 struct zone *zone = &pgdat->node_zones[zone_type];
5310 if (populated_zone(zone)) {
5311 node_set_state(nid, N_HIGH_MEMORY);
5312 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5313 zone_type <= ZONE_NORMAL)
5314 node_set_state(nid, N_NORMAL_MEMORY);
5321 * free_area_init_nodes - Initialise all pg_data_t and zone data
5322 * @max_zone_pfn: an array of max PFNs for each zone
5324 * This will call free_area_init_node() for each active node in the system.
5325 * Using the page ranges provided by memblock_set_node(), the size of each
5326 * zone in each node and their holes is calculated. If the maximum PFN
5327 * between two adjacent zones match, it is assumed that the zone is empty.
5328 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5329 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5330 * starts where the previous one ended. For example, ZONE_DMA32 starts
5331 * at arch_max_dma_pfn.
5333 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5335 unsigned long start_pfn, end_pfn;
5338 /* Record where the zone boundaries are */
5339 memset(arch_zone_lowest_possible_pfn, 0,
5340 sizeof(arch_zone_lowest_possible_pfn));
5341 memset(arch_zone_highest_possible_pfn, 0,
5342 sizeof(arch_zone_highest_possible_pfn));
5343 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5344 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5345 for (i = 1; i < MAX_NR_ZONES; i++) {
5346 if (i == ZONE_MOVABLE)
5348 arch_zone_lowest_possible_pfn[i] =
5349 arch_zone_highest_possible_pfn[i-1];
5350 arch_zone_highest_possible_pfn[i] =
5351 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5353 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5354 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5356 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5357 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5358 find_zone_movable_pfns_for_nodes();
5360 /* Print out the zone ranges */
5361 pr_info("Zone ranges:\n");
5362 for (i = 0; i < MAX_NR_ZONES; i++) {
5363 if (i == ZONE_MOVABLE)
5365 pr_info(" %-8s ", zone_names[i]);
5366 if (arch_zone_lowest_possible_pfn[i] ==
5367 arch_zone_highest_possible_pfn[i])
5370 pr_cont("[mem %#018Lx-%#018Lx]\n",
5371 (u64)arch_zone_lowest_possible_pfn[i]
5373 ((u64)arch_zone_highest_possible_pfn[i]
5374 << PAGE_SHIFT) - 1);
5377 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5378 pr_info("Movable zone start for each node\n");
5379 for (i = 0; i < MAX_NUMNODES; i++) {
5380 if (zone_movable_pfn[i])
5381 pr_info(" Node %d: %#018Lx\n", i,
5382 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5385 /* Print out the early node map */
5386 pr_info("Early memory node ranges\n");
5387 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5388 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5389 (u64)start_pfn << PAGE_SHIFT,
5390 ((u64)end_pfn << PAGE_SHIFT) - 1);
5392 /* Initialise every node */
5393 mminit_verify_pageflags_layout();
5394 setup_nr_node_ids();
5395 for_each_online_node(nid) {
5396 pg_data_t *pgdat = NODE_DATA(nid);
5397 free_area_init_node(nid, NULL,
5398 find_min_pfn_for_node(nid), NULL);
5400 /* Any memory on that node */
5401 if (pgdat->node_present_pages)
5402 node_set_state(nid, N_MEMORY);
5403 check_for_memory(pgdat, nid);
5407 static int __init cmdline_parse_core(char *p, unsigned long *core)
5409 unsigned long long coremem;
5413 coremem = memparse(p, &p);
5414 *core = coremem >> PAGE_SHIFT;
5416 /* Paranoid check that UL is enough for the coremem value */
5417 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5423 * kernelcore=size sets the amount of memory for use for allocations that
5424 * cannot be reclaimed or migrated.
5426 static int __init cmdline_parse_kernelcore(char *p)
5428 return cmdline_parse_core(p, &required_kernelcore);
5432 * movablecore=size sets the amount of memory for use for allocations that
5433 * can be reclaimed or migrated.
5435 static int __init cmdline_parse_movablecore(char *p)
5437 return cmdline_parse_core(p, &required_movablecore);
5440 early_param("kernelcore", cmdline_parse_kernelcore);
5441 early_param("movablecore", cmdline_parse_movablecore);
5443 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5445 void adjust_managed_page_count(struct page *page, long count)
5447 spin_lock(&managed_page_count_lock);
5448 page_zone(page)->managed_pages += count;
5449 totalram_pages += count;
5450 #ifdef CONFIG_HIGHMEM
5451 if (PageHighMem(page))
5452 totalhigh_pages += count;
5454 spin_unlock(&managed_page_count_lock);
5456 EXPORT_SYMBOL(adjust_managed_page_count);
5458 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5461 unsigned long pages = 0;
5463 start = (void *)PAGE_ALIGN((unsigned long)start);
5464 end = (void *)((unsigned long)end & PAGE_MASK);
5465 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5466 if ((unsigned int)poison <= 0xFF)
5467 memset(pos, poison, PAGE_SIZE);
5468 free_reserved_page(virt_to_page(pos));
5472 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5473 s, pages << (PAGE_SHIFT - 10), start, end);
5477 EXPORT_SYMBOL(free_reserved_area);
5479 #ifdef CONFIG_HIGHMEM
5480 void free_highmem_page(struct page *page)
5482 __free_reserved_page(page);
5484 page_zone(page)->managed_pages++;
5490 void __init mem_init_print_info(const char *str)
5492 unsigned long physpages, codesize, datasize, rosize, bss_size;
5493 unsigned long init_code_size, init_data_size;
5495 physpages = get_num_physpages();
5496 codesize = _etext - _stext;
5497 datasize = _edata - _sdata;
5498 rosize = __end_rodata - __start_rodata;
5499 bss_size = __bss_stop - __bss_start;
5500 init_data_size = __init_end - __init_begin;
5501 init_code_size = _einittext - _sinittext;
5504 * Detect special cases and adjust section sizes accordingly:
5505 * 1) .init.* may be embedded into .data sections
5506 * 2) .init.text.* may be out of [__init_begin, __init_end],
5507 * please refer to arch/tile/kernel/vmlinux.lds.S.
5508 * 3) .rodata.* may be embedded into .text or .data sections.
5510 #define adj_init_size(start, end, size, pos, adj) \
5512 if (start <= pos && pos < end && size > adj) \
5516 adj_init_size(__init_begin, __init_end, init_data_size,
5517 _sinittext, init_code_size);
5518 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5519 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5520 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5521 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5523 #undef adj_init_size
5525 pr_info("Memory: %luK/%luK available "
5526 "(%luK kernel code, %luK rwdata, %luK rodata, "
5527 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5528 #ifdef CONFIG_HIGHMEM
5532 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5533 codesize >> 10, datasize >> 10, rosize >> 10,
5534 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5535 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5536 totalcma_pages << (PAGE_SHIFT-10),
5537 #ifdef CONFIG_HIGHMEM
5538 totalhigh_pages << (PAGE_SHIFT-10),
5540 str ? ", " : "", str ? str : "");
5544 * set_dma_reserve - set the specified number of pages reserved in the first zone
5545 * @new_dma_reserve: The number of pages to mark reserved
5547 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5548 * In the DMA zone, a significant percentage may be consumed by kernel image
5549 * and other unfreeable allocations which can skew the watermarks badly. This
5550 * function may optionally be used to account for unfreeable pages in the
5551 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5552 * smaller per-cpu batchsize.
5554 void __init set_dma_reserve(unsigned long new_dma_reserve)
5556 dma_reserve = new_dma_reserve;
5559 void __init free_area_init(unsigned long *zones_size)
5561 free_area_init_node(0, zones_size,
5562 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5565 static int page_alloc_cpu_notify(struct notifier_block *self,
5566 unsigned long action, void *hcpu)
5568 int cpu = (unsigned long)hcpu;
5570 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5571 lru_add_drain_cpu(cpu);
5575 * Spill the event counters of the dead processor
5576 * into the current processors event counters.
5577 * This artificially elevates the count of the current
5580 vm_events_fold_cpu(cpu);
5583 * Zero the differential counters of the dead processor
5584 * so that the vm statistics are consistent.
5586 * This is only okay since the processor is dead and cannot
5587 * race with what we are doing.
5589 cpu_vm_stats_fold(cpu);
5594 void __init page_alloc_init(void)
5596 hotcpu_notifier(page_alloc_cpu_notify, 0);
5600 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5601 * or min_free_kbytes changes.
5603 static void calculate_totalreserve_pages(void)
5605 struct pglist_data *pgdat;
5606 unsigned long reserve_pages = 0;
5607 enum zone_type i, j;
5609 for_each_online_pgdat(pgdat) {
5610 for (i = 0; i < MAX_NR_ZONES; i++) {
5611 struct zone *zone = pgdat->node_zones + i;
5614 /* Find valid and maximum lowmem_reserve in the zone */
5615 for (j = i; j < MAX_NR_ZONES; j++) {
5616 if (zone->lowmem_reserve[j] > max)
5617 max = zone->lowmem_reserve[j];
5620 /* we treat the high watermark as reserved pages. */
5621 max += high_wmark_pages(zone);
5623 if (max > zone->managed_pages)
5624 max = zone->managed_pages;
5625 reserve_pages += max;
5627 * Lowmem reserves are not available to
5628 * GFP_HIGHUSER page cache allocations and
5629 * kswapd tries to balance zones to their high
5630 * watermark. As a result, neither should be
5631 * regarded as dirtyable memory, to prevent a
5632 * situation where reclaim has to clean pages
5633 * in order to balance the zones.
5635 zone->dirty_balance_reserve = max;
5638 dirty_balance_reserve = reserve_pages;
5639 totalreserve_pages = reserve_pages;
5643 * setup_per_zone_lowmem_reserve - called whenever
5644 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5645 * has a correct pages reserved value, so an adequate number of
5646 * pages are left in the zone after a successful __alloc_pages().
5648 static void setup_per_zone_lowmem_reserve(void)
5650 struct pglist_data *pgdat;
5651 enum zone_type j, idx;
5653 for_each_online_pgdat(pgdat) {
5654 for (j = 0; j < MAX_NR_ZONES; j++) {
5655 struct zone *zone = pgdat->node_zones + j;
5656 unsigned long managed_pages = zone->managed_pages;
5658 zone->lowmem_reserve[j] = 0;
5662 struct zone *lower_zone;
5666 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5667 sysctl_lowmem_reserve_ratio[idx] = 1;
5669 lower_zone = pgdat->node_zones + idx;
5670 lower_zone->lowmem_reserve[j] = managed_pages /
5671 sysctl_lowmem_reserve_ratio[idx];
5672 managed_pages += lower_zone->managed_pages;
5677 /* update totalreserve_pages */
5678 calculate_totalreserve_pages();
5681 static void __setup_per_zone_wmarks(void)
5683 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5684 unsigned long lowmem_pages = 0;
5686 unsigned long flags;
5688 /* Calculate total number of !ZONE_HIGHMEM pages */
5689 for_each_zone(zone) {
5690 if (!is_highmem(zone))
5691 lowmem_pages += zone->managed_pages;
5694 for_each_zone(zone) {
5697 spin_lock_irqsave(&zone->lock, flags);
5698 tmp = (u64)pages_min * zone->managed_pages;
5699 do_div(tmp, lowmem_pages);
5700 if (is_highmem(zone)) {
5702 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5703 * need highmem pages, so cap pages_min to a small
5706 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5707 * deltas controls asynch page reclaim, and so should
5708 * not be capped for highmem.
5710 unsigned long min_pages;
5712 min_pages = zone->managed_pages / 1024;
5713 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5714 zone->watermark[WMARK_MIN] = min_pages;
5717 * If it's a lowmem zone, reserve a number of pages
5718 * proportionate to the zone's size.
5720 zone->watermark[WMARK_MIN] = tmp;
5723 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5724 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5726 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5727 high_wmark_pages(zone) - low_wmark_pages(zone) -
5728 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5730 setup_zone_migrate_reserve(zone);
5731 spin_unlock_irqrestore(&zone->lock, flags);
5734 /* update totalreserve_pages */
5735 calculate_totalreserve_pages();
5739 * setup_per_zone_wmarks - called when min_free_kbytes changes
5740 * or when memory is hot-{added|removed}
5742 * Ensures that the watermark[min,low,high] values for each zone are set
5743 * correctly with respect to min_free_kbytes.
5745 void setup_per_zone_wmarks(void)
5747 mutex_lock(&zonelists_mutex);
5748 __setup_per_zone_wmarks();
5749 mutex_unlock(&zonelists_mutex);
5753 * The inactive anon list should be small enough that the VM never has to
5754 * do too much work, but large enough that each inactive page has a chance
5755 * to be referenced again before it is swapped out.
5757 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5758 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5759 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5760 * the anonymous pages are kept on the inactive list.
5763 * memory ratio inactive anon
5764 * -------------------------------------
5773 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5775 unsigned int gb, ratio;
5777 /* Zone size in gigabytes */
5778 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5780 ratio = int_sqrt(10 * gb);
5784 zone->inactive_ratio = ratio;
5787 static void __meminit setup_per_zone_inactive_ratio(void)
5792 calculate_zone_inactive_ratio(zone);
5796 * Initialise min_free_kbytes.
5798 * For small machines we want it small (128k min). For large machines
5799 * we want it large (64MB max). But it is not linear, because network
5800 * bandwidth does not increase linearly with machine size. We use
5802 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5803 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5819 int __meminit init_per_zone_wmark_min(void)
5821 unsigned long lowmem_kbytes;
5822 int new_min_free_kbytes;
5824 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5825 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5827 if (new_min_free_kbytes > user_min_free_kbytes) {
5828 min_free_kbytes = new_min_free_kbytes;
5829 if (min_free_kbytes < 128)
5830 min_free_kbytes = 128;
5831 if (min_free_kbytes > 65536)
5832 min_free_kbytes = 65536;
5834 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5835 new_min_free_kbytes, user_min_free_kbytes);
5837 setup_per_zone_wmarks();
5838 refresh_zone_stat_thresholds();
5839 setup_per_zone_lowmem_reserve();
5840 setup_per_zone_inactive_ratio();
5843 module_init(init_per_zone_wmark_min)
5846 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5847 * that we can call two helper functions whenever min_free_kbytes
5850 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5851 void __user *buffer, size_t *length, loff_t *ppos)
5855 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5860 user_min_free_kbytes = min_free_kbytes;
5861 setup_per_zone_wmarks();
5867 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5868 void __user *buffer, size_t *length, loff_t *ppos)
5873 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5878 zone->min_unmapped_pages = (zone->managed_pages *
5879 sysctl_min_unmapped_ratio) / 100;
5883 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5884 void __user *buffer, size_t *length, loff_t *ppos)
5889 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5894 zone->min_slab_pages = (zone->managed_pages *
5895 sysctl_min_slab_ratio) / 100;
5901 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5902 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5903 * whenever sysctl_lowmem_reserve_ratio changes.
5905 * The reserve ratio obviously has absolutely no relation with the
5906 * minimum watermarks. The lowmem reserve ratio can only make sense
5907 * if in function of the boot time zone sizes.
5909 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5910 void __user *buffer, size_t *length, loff_t *ppos)
5912 proc_dointvec_minmax(table, write, buffer, length, ppos);
5913 setup_per_zone_lowmem_reserve();
5918 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5919 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5920 * pagelist can have before it gets flushed back to buddy allocator.
5922 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5923 void __user *buffer, size_t *length, loff_t *ppos)
5926 int old_percpu_pagelist_fraction;
5929 mutex_lock(&pcp_batch_high_lock);
5930 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5932 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5933 if (!write || ret < 0)
5936 /* Sanity checking to avoid pcp imbalance */
5937 if (percpu_pagelist_fraction &&
5938 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
5939 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
5945 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
5948 for_each_populated_zone(zone) {
5951 for_each_possible_cpu(cpu)
5952 pageset_set_high_and_batch(zone,
5953 per_cpu_ptr(zone->pageset, cpu));
5956 mutex_unlock(&pcp_batch_high_lock);
5960 int hashdist = HASHDIST_DEFAULT;
5963 static int __init set_hashdist(char *str)
5967 hashdist = simple_strtoul(str, &str, 0);
5970 __setup("hashdist=", set_hashdist);
5974 * allocate a large system hash table from bootmem
5975 * - it is assumed that the hash table must contain an exact power-of-2
5976 * quantity of entries
5977 * - limit is the number of hash buckets, not the total allocation size
5979 void *__init alloc_large_system_hash(const char *tablename,
5980 unsigned long bucketsize,
5981 unsigned long numentries,
5984 unsigned int *_hash_shift,
5985 unsigned int *_hash_mask,
5986 unsigned long low_limit,
5987 unsigned long high_limit)
5989 unsigned long long max = high_limit;
5990 unsigned long log2qty, size;
5993 /* allow the kernel cmdline to have a say */
5995 /* round applicable memory size up to nearest megabyte */
5996 numentries = nr_kernel_pages;
5998 /* It isn't necessary when PAGE_SIZE >= 1MB */
5999 if (PAGE_SHIFT < 20)
6000 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6002 /* limit to 1 bucket per 2^scale bytes of low memory */
6003 if (scale > PAGE_SHIFT)
6004 numentries >>= (scale - PAGE_SHIFT);
6006 numentries <<= (PAGE_SHIFT - scale);
6008 /* Make sure we've got at least a 0-order allocation.. */
6009 if (unlikely(flags & HASH_SMALL)) {
6010 /* Makes no sense without HASH_EARLY */
6011 WARN_ON(!(flags & HASH_EARLY));
6012 if (!(numentries >> *_hash_shift)) {
6013 numentries = 1UL << *_hash_shift;
6014 BUG_ON(!numentries);
6016 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6017 numentries = PAGE_SIZE / bucketsize;
6019 numentries = roundup_pow_of_two(numentries);
6021 /* limit allocation size to 1/16 total memory by default */
6023 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6024 do_div(max, bucketsize);
6026 max = min(max, 0x80000000ULL);
6028 if (numentries < low_limit)
6029 numentries = low_limit;
6030 if (numentries > max)
6033 log2qty = ilog2(numentries);
6036 size = bucketsize << log2qty;
6037 if (flags & HASH_EARLY)
6038 table = memblock_virt_alloc_nopanic(size, 0);
6040 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6043 * If bucketsize is not a power-of-two, we may free
6044 * some pages at the end of hash table which
6045 * alloc_pages_exact() automatically does
6047 if (get_order(size) < MAX_ORDER) {
6048 table = alloc_pages_exact(size, GFP_ATOMIC);
6049 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6052 } while (!table && size > PAGE_SIZE && --log2qty);
6055 panic("Failed to allocate %s hash table\n", tablename);
6057 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6060 ilog2(size) - PAGE_SHIFT,
6064 *_hash_shift = log2qty;
6066 *_hash_mask = (1 << log2qty) - 1;
6071 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6072 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6075 #ifdef CONFIG_SPARSEMEM
6076 return __pfn_to_section(pfn)->pageblock_flags;
6078 return zone->pageblock_flags;
6079 #endif /* CONFIG_SPARSEMEM */
6082 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6084 #ifdef CONFIG_SPARSEMEM
6085 pfn &= (PAGES_PER_SECTION-1);
6086 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6088 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6089 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6090 #endif /* CONFIG_SPARSEMEM */
6094 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6095 * @page: The page within the block of interest
6096 * @pfn: The target page frame number
6097 * @end_bitidx: The last bit of interest to retrieve
6098 * @mask: mask of bits that the caller is interested in
6100 * Return: pageblock_bits flags
6102 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6103 unsigned long end_bitidx,
6107 unsigned long *bitmap;
6108 unsigned long bitidx, word_bitidx;
6111 zone = page_zone(page);
6112 bitmap = get_pageblock_bitmap(zone, pfn);
6113 bitidx = pfn_to_bitidx(zone, pfn);
6114 word_bitidx = bitidx / BITS_PER_LONG;
6115 bitidx &= (BITS_PER_LONG-1);
6117 word = bitmap[word_bitidx];
6118 bitidx += end_bitidx;
6119 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6123 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6124 * @page: The page within the block of interest
6125 * @flags: The flags to set
6126 * @pfn: The target page frame number
6127 * @end_bitidx: The last bit of interest
6128 * @mask: mask of bits that the caller is interested in
6130 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6132 unsigned long end_bitidx,
6136 unsigned long *bitmap;
6137 unsigned long bitidx, word_bitidx;
6138 unsigned long old_word, word;
6140 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6142 zone = page_zone(page);
6143 bitmap = get_pageblock_bitmap(zone, pfn);
6144 bitidx = pfn_to_bitidx(zone, pfn);
6145 word_bitidx = bitidx / BITS_PER_LONG;
6146 bitidx &= (BITS_PER_LONG-1);
6148 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6150 bitidx += end_bitidx;
6151 mask <<= (BITS_PER_LONG - bitidx - 1);
6152 flags <<= (BITS_PER_LONG - bitidx - 1);
6154 word = ACCESS_ONCE(bitmap[word_bitidx]);
6156 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6157 if (word == old_word)
6164 * This function checks whether pageblock includes unmovable pages or not.
6165 * If @count is not zero, it is okay to include less @count unmovable pages
6167 * PageLRU check without isolation or lru_lock could race so that
6168 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6169 * expect this function should be exact.
6171 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6172 bool skip_hwpoisoned_pages)
6174 unsigned long pfn, iter, found;
6178 * For avoiding noise data, lru_add_drain_all() should be called
6179 * If ZONE_MOVABLE, the zone never contains unmovable pages
6181 if (zone_idx(zone) == ZONE_MOVABLE)
6183 mt = get_pageblock_migratetype(page);
6184 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6187 pfn = page_to_pfn(page);
6188 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6189 unsigned long check = pfn + iter;
6191 if (!pfn_valid_within(check))
6194 page = pfn_to_page(check);
6197 * Hugepages are not in LRU lists, but they're movable.
6198 * We need not scan over tail pages bacause we don't
6199 * handle each tail page individually in migration.
6201 if (PageHuge(page)) {
6202 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6207 * We can't use page_count without pin a page
6208 * because another CPU can free compound page.
6209 * This check already skips compound tails of THP
6210 * because their page->_count is zero at all time.
6212 if (!atomic_read(&page->_count)) {
6213 if (PageBuddy(page))
6214 iter += (1 << page_order(page)) - 1;
6219 * The HWPoisoned page may be not in buddy system, and
6220 * page_count() is not 0.
6222 if (skip_hwpoisoned_pages && PageHWPoison(page))
6228 * If there are RECLAIMABLE pages, we need to check
6229 * it. But now, memory offline itself doesn't call
6230 * shrink_node_slabs() and it still to be fixed.
6233 * If the page is not RAM, page_count()should be 0.
6234 * we don't need more check. This is an _used_ not-movable page.
6236 * The problematic thing here is PG_reserved pages. PG_reserved
6237 * is set to both of a memory hole page and a _used_ kernel
6246 bool is_pageblock_removable_nolock(struct page *page)
6252 * We have to be careful here because we are iterating over memory
6253 * sections which are not zone aware so we might end up outside of
6254 * the zone but still within the section.
6255 * We have to take care about the node as well. If the node is offline
6256 * its NODE_DATA will be NULL - see page_zone.
6258 if (!node_online(page_to_nid(page)))
6261 zone = page_zone(page);
6262 pfn = page_to_pfn(page);
6263 if (!zone_spans_pfn(zone, pfn))
6266 return !has_unmovable_pages(zone, page, 0, true);
6271 static unsigned long pfn_max_align_down(unsigned long pfn)
6273 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6274 pageblock_nr_pages) - 1);
6277 static unsigned long pfn_max_align_up(unsigned long pfn)
6279 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6280 pageblock_nr_pages));
6283 /* [start, end) must belong to a single zone. */
6284 static int __alloc_contig_migrate_range(struct compact_control *cc,
6285 unsigned long start, unsigned long end)
6287 /* This function is based on compact_zone() from compaction.c. */
6288 unsigned long nr_reclaimed;
6289 unsigned long pfn = start;
6290 unsigned int tries = 0;
6295 while (pfn < end || !list_empty(&cc->migratepages)) {
6296 if (fatal_signal_pending(current)) {
6301 if (list_empty(&cc->migratepages)) {
6302 cc->nr_migratepages = 0;
6303 pfn = isolate_migratepages_range(cc, pfn, end);
6309 } else if (++tries == 5) {
6310 ret = ret < 0 ? ret : -EBUSY;
6314 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6316 cc->nr_migratepages -= nr_reclaimed;
6318 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6319 NULL, 0, cc->mode, MR_CMA);
6322 putback_movable_pages(&cc->migratepages);
6329 * alloc_contig_range() -- tries to allocate given range of pages
6330 * @start: start PFN to allocate
6331 * @end: one-past-the-last PFN to allocate
6332 * @migratetype: migratetype of the underlaying pageblocks (either
6333 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6334 * in range must have the same migratetype and it must
6335 * be either of the two.
6337 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6338 * aligned, however it's the caller's responsibility to guarantee that
6339 * we are the only thread that changes migrate type of pageblocks the
6342 * The PFN range must belong to a single zone.
6344 * Returns zero on success or negative error code. On success all
6345 * pages which PFN is in [start, end) are allocated for the caller and
6346 * need to be freed with free_contig_range().
6348 int alloc_contig_range(unsigned long start, unsigned long end,
6349 unsigned migratetype)
6351 unsigned long outer_start, outer_end;
6354 struct compact_control cc = {
6355 .nr_migratepages = 0,
6357 .zone = page_zone(pfn_to_page(start)),
6358 .mode = MIGRATE_SYNC,
6359 .ignore_skip_hint = true,
6361 INIT_LIST_HEAD(&cc.migratepages);
6364 * What we do here is we mark all pageblocks in range as
6365 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6366 * have different sizes, and due to the way page allocator
6367 * work, we align the range to biggest of the two pages so
6368 * that page allocator won't try to merge buddies from
6369 * different pageblocks and change MIGRATE_ISOLATE to some
6370 * other migration type.
6372 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6373 * migrate the pages from an unaligned range (ie. pages that
6374 * we are interested in). This will put all the pages in
6375 * range back to page allocator as MIGRATE_ISOLATE.
6377 * When this is done, we take the pages in range from page
6378 * allocator removing them from the buddy system. This way
6379 * page allocator will never consider using them.
6381 * This lets us mark the pageblocks back as
6382 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6383 * aligned range but not in the unaligned, original range are
6384 * put back to page allocator so that buddy can use them.
6387 ret = start_isolate_page_range(pfn_max_align_down(start),
6388 pfn_max_align_up(end), migratetype,
6393 ret = __alloc_contig_migrate_range(&cc, start, end);
6398 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6399 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6400 * more, all pages in [start, end) are free in page allocator.
6401 * What we are going to do is to allocate all pages from
6402 * [start, end) (that is remove them from page allocator).
6404 * The only problem is that pages at the beginning and at the
6405 * end of interesting range may be not aligned with pages that
6406 * page allocator holds, ie. they can be part of higher order
6407 * pages. Because of this, we reserve the bigger range and
6408 * once this is done free the pages we are not interested in.
6410 * We don't have to hold zone->lock here because the pages are
6411 * isolated thus they won't get removed from buddy.
6414 lru_add_drain_all();
6415 drain_all_pages(cc.zone);
6418 outer_start = start;
6419 while (!PageBuddy(pfn_to_page(outer_start))) {
6420 if (++order >= MAX_ORDER) {
6424 outer_start &= ~0UL << order;
6427 /* Make sure the range is really isolated. */
6428 if (test_pages_isolated(outer_start, end, false)) {
6429 pr_info("%s: [%lx, %lx) PFNs busy\n",
6430 __func__, outer_start, end);
6435 /* Grab isolated pages from freelists. */
6436 outer_end = isolate_freepages_range(&cc, outer_start, end);
6442 /* Free head and tail (if any) */
6443 if (start != outer_start)
6444 free_contig_range(outer_start, start - outer_start);
6445 if (end != outer_end)
6446 free_contig_range(end, outer_end - end);
6449 undo_isolate_page_range(pfn_max_align_down(start),
6450 pfn_max_align_up(end), migratetype);
6454 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6456 unsigned int count = 0;
6458 for (; nr_pages--; pfn++) {
6459 struct page *page = pfn_to_page(pfn);
6461 count += page_count(page) != 1;
6464 WARN(count != 0, "%d pages are still in use!\n", count);
6468 #ifdef CONFIG_MEMORY_HOTPLUG
6470 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6471 * page high values need to be recalulated.
6473 void __meminit zone_pcp_update(struct zone *zone)
6476 mutex_lock(&pcp_batch_high_lock);
6477 for_each_possible_cpu(cpu)
6478 pageset_set_high_and_batch(zone,
6479 per_cpu_ptr(zone->pageset, cpu));
6480 mutex_unlock(&pcp_batch_high_lock);
6484 void zone_pcp_reset(struct zone *zone)
6486 unsigned long flags;
6488 struct per_cpu_pageset *pset;
6490 /* avoid races with drain_pages() */
6491 local_irq_save(flags);
6492 if (zone->pageset != &boot_pageset) {
6493 for_each_online_cpu(cpu) {
6494 pset = per_cpu_ptr(zone->pageset, cpu);
6495 drain_zonestat(zone, pset);
6497 free_percpu(zone->pageset);
6498 zone->pageset = &boot_pageset;
6500 local_irq_restore(flags);
6503 #ifdef CONFIG_MEMORY_HOTREMOVE
6505 * All pages in the range must be isolated before calling this.
6508 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6512 unsigned int order, i;
6514 unsigned long flags;
6515 /* find the first valid pfn */
6516 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6521 zone = page_zone(pfn_to_page(pfn));
6522 spin_lock_irqsave(&zone->lock, flags);
6524 while (pfn < end_pfn) {
6525 if (!pfn_valid(pfn)) {
6529 page = pfn_to_page(pfn);
6531 * The HWPoisoned page may be not in buddy system, and
6532 * page_count() is not 0.
6534 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6536 SetPageReserved(page);
6540 BUG_ON(page_count(page));
6541 BUG_ON(!PageBuddy(page));
6542 order = page_order(page);
6543 #ifdef CONFIG_DEBUG_VM
6544 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6545 pfn, 1 << order, end_pfn);
6547 list_del(&page->lru);
6548 rmv_page_order(page);
6549 zone->free_area[order].nr_free--;
6550 for (i = 0; i < (1 << order); i++)
6551 SetPageReserved((page+i));
6552 pfn += (1 << order);
6554 spin_unlock_irqrestore(&zone->lock, flags);
6558 #ifdef CONFIG_MEMORY_FAILURE
6559 bool is_free_buddy_page(struct page *page)
6561 struct zone *zone = page_zone(page);
6562 unsigned long pfn = page_to_pfn(page);
6563 unsigned long flags;
6566 spin_lock_irqsave(&zone->lock, flags);
6567 for (order = 0; order < MAX_ORDER; order++) {
6568 struct page *page_head = page - (pfn & ((1 << order) - 1));
6570 if (PageBuddy(page_head) && page_order(page_head) >= order)
6573 spin_unlock_irqrestore(&zone->lock, flags);
6575 return order < MAX_ORDER;