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 /* update __split_huge_page_refcount if you change this function */
385 static int destroy_compound_page(struct page *page, unsigned long order)
388 int nr_pages = 1 << order;
391 if (unlikely(compound_order(page) != order)) {
392 bad_page(page, "wrong compound order", 0);
396 __ClearPageHead(page);
398 for (i = 1; i < nr_pages; i++) {
399 struct page *p = page + i;
401 if (unlikely(!PageTail(p))) {
402 bad_page(page, "PageTail not set", 0);
404 } else if (unlikely(p->first_page != page)) {
405 bad_page(page, "first_page not consistent", 0);
414 static inline void prep_zero_page(struct page *page, unsigned int order,
420 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
421 * and __GFP_HIGHMEM from hard or soft interrupt context.
423 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
424 for (i = 0; i < (1 << order); i++)
425 clear_highpage(page + i);
428 #ifdef CONFIG_DEBUG_PAGEALLOC
429 unsigned int _debug_guardpage_minorder;
430 bool _debug_pagealloc_enabled __read_mostly;
431 bool _debug_guardpage_enabled __read_mostly;
433 static int __init early_debug_pagealloc(char *buf)
438 if (strcmp(buf, "on") == 0)
439 _debug_pagealloc_enabled = true;
443 early_param("debug_pagealloc", early_debug_pagealloc);
445 static bool need_debug_guardpage(void)
447 /* If we don't use debug_pagealloc, we don't need guard page */
448 if (!debug_pagealloc_enabled())
454 static void init_debug_guardpage(void)
456 if (!debug_pagealloc_enabled())
459 _debug_guardpage_enabled = true;
462 struct page_ext_operations debug_guardpage_ops = {
463 .need = need_debug_guardpage,
464 .init = init_debug_guardpage,
467 static int __init debug_guardpage_minorder_setup(char *buf)
471 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
472 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
475 _debug_guardpage_minorder = res;
476 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
479 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
481 static inline void set_page_guard(struct zone *zone, struct page *page,
482 unsigned int order, int migratetype)
484 struct page_ext *page_ext;
486 if (!debug_guardpage_enabled())
489 page_ext = lookup_page_ext(page);
490 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
492 INIT_LIST_HEAD(&page->lru);
493 set_page_private(page, order);
494 /* Guard pages are not available for any usage */
495 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
498 static inline void clear_page_guard(struct zone *zone, struct page *page,
499 unsigned int order, int migratetype)
501 struct page_ext *page_ext;
503 if (!debug_guardpage_enabled())
506 page_ext = lookup_page_ext(page);
507 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
509 set_page_private(page, 0);
510 if (!is_migrate_isolate(migratetype))
511 __mod_zone_freepage_state(zone, (1 << order), migratetype);
514 struct page_ext_operations debug_guardpage_ops = { NULL, };
515 static inline void set_page_guard(struct zone *zone, struct page *page,
516 unsigned int order, int migratetype) {}
517 static inline void clear_page_guard(struct zone *zone, struct page *page,
518 unsigned int order, int migratetype) {}
521 static inline void set_page_order(struct page *page, unsigned int order)
523 set_page_private(page, order);
524 __SetPageBuddy(page);
527 static inline void rmv_page_order(struct page *page)
529 __ClearPageBuddy(page);
530 set_page_private(page, 0);
534 * This function checks whether a page is free && is the buddy
535 * we can do coalesce a page and its buddy if
536 * (a) the buddy is not in a hole &&
537 * (b) the buddy is in the buddy system &&
538 * (c) a page and its buddy have the same order &&
539 * (d) a page and its buddy are in the same zone.
541 * For recording whether a page is in the buddy system, we set ->_mapcount
542 * PAGE_BUDDY_MAPCOUNT_VALUE.
543 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
544 * serialized by zone->lock.
546 * For recording page's order, we use page_private(page).
548 static inline int page_is_buddy(struct page *page, struct page *buddy,
551 if (!pfn_valid_within(page_to_pfn(buddy)))
554 if (page_is_guard(buddy) && page_order(buddy) == order) {
555 if (page_zone_id(page) != page_zone_id(buddy))
558 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
563 if (PageBuddy(buddy) && page_order(buddy) == order) {
565 * zone check is done late to avoid uselessly
566 * calculating zone/node ids for pages that could
569 if (page_zone_id(page) != page_zone_id(buddy))
572 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
580 * Freeing function for a buddy system allocator.
582 * The concept of a buddy system is to maintain direct-mapped table
583 * (containing bit values) for memory blocks of various "orders".
584 * The bottom level table contains the map for the smallest allocatable
585 * units of memory (here, pages), and each level above it describes
586 * pairs of units from the levels below, hence, "buddies".
587 * At a high level, all that happens here is marking the table entry
588 * at the bottom level available, and propagating the changes upward
589 * as necessary, plus some accounting needed to play nicely with other
590 * parts of the VM system.
591 * At each level, we keep a list of pages, which are heads of continuous
592 * free pages of length of (1 << order) and marked with _mapcount
593 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
595 * So when we are allocating or freeing one, we can derive the state of the
596 * other. That is, if we allocate a small block, and both were
597 * free, the remainder of the region must be split into blocks.
598 * If a block is freed, and its buddy is also free, then this
599 * triggers coalescing into a block of larger size.
604 static inline void __free_one_page(struct page *page,
606 struct zone *zone, unsigned int order,
609 unsigned long page_idx;
610 unsigned long combined_idx;
611 unsigned long uninitialized_var(buddy_idx);
613 int max_order = MAX_ORDER;
615 VM_BUG_ON(!zone_is_initialized(zone));
617 if (unlikely(PageCompound(page)))
618 if (unlikely(destroy_compound_page(page, order)))
621 VM_BUG_ON(migratetype == -1);
622 if (is_migrate_isolate(migratetype)) {
624 * We restrict max order of merging to prevent merge
625 * between freepages on isolate pageblock and normal
626 * pageblock. Without this, pageblock isolation
627 * could cause incorrect freepage accounting.
629 max_order = min(MAX_ORDER, pageblock_order + 1);
631 __mod_zone_freepage_state(zone, 1 << order, migratetype);
634 page_idx = pfn & ((1 << max_order) - 1);
636 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
637 VM_BUG_ON_PAGE(bad_range(zone, page), page);
639 while (order < max_order - 1) {
640 buddy_idx = __find_buddy_index(page_idx, order);
641 buddy = page + (buddy_idx - page_idx);
642 if (!page_is_buddy(page, buddy, order))
645 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
646 * merge with it and move up one order.
648 if (page_is_guard(buddy)) {
649 clear_page_guard(zone, buddy, order, migratetype);
651 list_del(&buddy->lru);
652 zone->free_area[order].nr_free--;
653 rmv_page_order(buddy);
655 combined_idx = buddy_idx & page_idx;
656 page = page + (combined_idx - page_idx);
657 page_idx = combined_idx;
660 set_page_order(page, order);
663 * If this is not the largest possible page, check if the buddy
664 * of the next-highest order is free. If it is, it's possible
665 * that pages are being freed that will coalesce soon. In case,
666 * that is happening, add the free page to the tail of the list
667 * so it's less likely to be used soon and more likely to be merged
668 * as a higher order page
670 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
671 struct page *higher_page, *higher_buddy;
672 combined_idx = buddy_idx & page_idx;
673 higher_page = page + (combined_idx - page_idx);
674 buddy_idx = __find_buddy_index(combined_idx, order + 1);
675 higher_buddy = higher_page + (buddy_idx - combined_idx);
676 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
677 list_add_tail(&page->lru,
678 &zone->free_area[order].free_list[migratetype]);
683 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
685 zone->free_area[order].nr_free++;
688 static inline int free_pages_check(struct page *page)
690 const char *bad_reason = NULL;
691 unsigned long bad_flags = 0;
693 if (unlikely(page_mapcount(page)))
694 bad_reason = "nonzero mapcount";
695 if (unlikely(page->mapping != NULL))
696 bad_reason = "non-NULL mapping";
697 if (unlikely(atomic_read(&page->_count) != 0))
698 bad_reason = "nonzero _count";
699 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
700 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
701 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
704 if (unlikely(page->mem_cgroup))
705 bad_reason = "page still charged to cgroup";
707 if (unlikely(bad_reason)) {
708 bad_page(page, bad_reason, bad_flags);
711 page_cpupid_reset_last(page);
712 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
713 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
718 * Frees a number of pages from the PCP lists
719 * Assumes all pages on list are in same zone, and of same order.
720 * count is the number of pages to free.
722 * If the zone was previously in an "all pages pinned" state then look to
723 * see if this freeing clears that state.
725 * And clear the zone's pages_scanned counter, to hold off the "all pages are
726 * pinned" detection logic.
728 static void free_pcppages_bulk(struct zone *zone, int count,
729 struct per_cpu_pages *pcp)
734 unsigned long nr_scanned;
736 spin_lock(&zone->lock);
737 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
739 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
743 struct list_head *list;
746 * Remove pages from lists in a round-robin fashion. A
747 * batch_free count is maintained that is incremented when an
748 * empty list is encountered. This is so more pages are freed
749 * off fuller lists instead of spinning excessively around empty
754 if (++migratetype == MIGRATE_PCPTYPES)
756 list = &pcp->lists[migratetype];
757 } while (list_empty(list));
759 /* This is the only non-empty list. Free them all. */
760 if (batch_free == MIGRATE_PCPTYPES)
761 batch_free = to_free;
764 int mt; /* migratetype of the to-be-freed page */
766 page = list_entry(list->prev, struct page, lru);
767 /* must delete as __free_one_page list manipulates */
768 list_del(&page->lru);
769 mt = get_freepage_migratetype(page);
770 if (unlikely(has_isolate_pageblock(zone)))
771 mt = get_pageblock_migratetype(page);
773 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
774 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
775 trace_mm_page_pcpu_drain(page, 0, mt);
776 } while (--to_free && --batch_free && !list_empty(list));
778 spin_unlock(&zone->lock);
781 static void free_one_page(struct zone *zone,
782 struct page *page, unsigned long pfn,
786 unsigned long nr_scanned;
787 spin_lock(&zone->lock);
788 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
790 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
792 if (unlikely(has_isolate_pageblock(zone) ||
793 is_migrate_isolate(migratetype))) {
794 migratetype = get_pfnblock_migratetype(page, pfn);
796 __free_one_page(page, pfn, zone, order, migratetype);
797 spin_unlock(&zone->lock);
800 static bool free_pages_prepare(struct page *page, unsigned int order)
805 VM_BUG_ON_PAGE(PageTail(page), page);
806 VM_BUG_ON_PAGE(PageHead(page) && compound_order(page) != order, page);
808 trace_mm_page_free(page, order);
809 kmemcheck_free_shadow(page, order);
812 page->mapping = NULL;
813 for (i = 0; i < (1 << order); i++)
814 bad += free_pages_check(page + i);
818 reset_page_owner(page, order);
820 if (!PageHighMem(page)) {
821 debug_check_no_locks_freed(page_address(page),
823 debug_check_no_obj_freed(page_address(page),
826 arch_free_page(page, order);
827 kernel_map_pages(page, 1 << order, 0);
832 static void __free_pages_ok(struct page *page, unsigned int order)
836 unsigned long pfn = page_to_pfn(page);
838 if (!free_pages_prepare(page, order))
841 migratetype = get_pfnblock_migratetype(page, pfn);
842 local_irq_save(flags);
843 __count_vm_events(PGFREE, 1 << order);
844 set_freepage_migratetype(page, migratetype);
845 free_one_page(page_zone(page), page, pfn, order, migratetype);
846 local_irq_restore(flags);
849 void __init __free_pages_bootmem(struct page *page, unsigned int order)
851 unsigned int nr_pages = 1 << order;
852 struct page *p = page;
856 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
858 __ClearPageReserved(p);
859 set_page_count(p, 0);
861 __ClearPageReserved(p);
862 set_page_count(p, 0);
864 page_zone(page)->managed_pages += nr_pages;
865 set_page_refcounted(page);
866 __free_pages(page, order);
870 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
871 void __init init_cma_reserved_pageblock(struct page *page)
873 unsigned i = pageblock_nr_pages;
874 struct page *p = page;
877 __ClearPageReserved(p);
878 set_page_count(p, 0);
881 set_pageblock_migratetype(page, MIGRATE_CMA);
883 if (pageblock_order >= MAX_ORDER) {
884 i = pageblock_nr_pages;
887 set_page_refcounted(p);
888 __free_pages(p, MAX_ORDER - 1);
889 p += MAX_ORDER_NR_PAGES;
890 } while (i -= MAX_ORDER_NR_PAGES);
892 set_page_refcounted(page);
893 __free_pages(page, pageblock_order);
896 adjust_managed_page_count(page, pageblock_nr_pages);
901 * The order of subdivision here is critical for the IO subsystem.
902 * Please do not alter this order without good reasons and regression
903 * testing. Specifically, as large blocks of memory are subdivided,
904 * the order in which smaller blocks are delivered depends on the order
905 * they're subdivided in this function. This is the primary factor
906 * influencing the order in which pages are delivered to the IO
907 * subsystem according to empirical testing, and this is also justified
908 * by considering the behavior of a buddy system containing a single
909 * large block of memory acted on by a series of small allocations.
910 * This behavior is a critical factor in sglist merging's success.
914 static inline void expand(struct zone *zone, struct page *page,
915 int low, int high, struct free_area *area,
918 unsigned long size = 1 << high;
924 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
926 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
927 debug_guardpage_enabled() &&
928 high < debug_guardpage_minorder()) {
930 * Mark as guard pages (or page), that will allow to
931 * merge back to allocator when buddy will be freed.
932 * Corresponding page table entries will not be touched,
933 * pages will stay not present in virtual address space
935 set_page_guard(zone, &page[size], high, migratetype);
938 list_add(&page[size].lru, &area->free_list[migratetype]);
940 set_page_order(&page[size], high);
945 * This page is about to be returned from the page allocator
947 static inline int check_new_page(struct page *page)
949 const char *bad_reason = NULL;
950 unsigned long bad_flags = 0;
952 if (unlikely(page_mapcount(page)))
953 bad_reason = "nonzero mapcount";
954 if (unlikely(page->mapping != NULL))
955 bad_reason = "non-NULL mapping";
956 if (unlikely(atomic_read(&page->_count) != 0))
957 bad_reason = "nonzero _count";
958 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
959 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
960 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
963 if (unlikely(page->mem_cgroup))
964 bad_reason = "page still charged to cgroup";
966 if (unlikely(bad_reason)) {
967 bad_page(page, bad_reason, bad_flags);
973 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
978 for (i = 0; i < (1 << order); i++) {
979 struct page *p = page + i;
980 if (unlikely(check_new_page(p)))
984 set_page_private(page, 0);
985 set_page_refcounted(page);
987 arch_alloc_page(page, order);
988 kernel_map_pages(page, 1 << order, 1);
990 if (gfp_flags & __GFP_ZERO)
991 prep_zero_page(page, order, gfp_flags);
993 if (order && (gfp_flags & __GFP_COMP))
994 prep_compound_page(page, order);
996 set_page_owner(page, order, gfp_flags);
999 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
1000 * allocate the page. The expectation is that the caller is taking
1001 * steps that will free more memory. The caller should avoid the page
1002 * being used for !PFMEMALLOC purposes.
1004 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1010 * Go through the free lists for the given migratetype and remove
1011 * the smallest available page from the freelists
1014 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1017 unsigned int current_order;
1018 struct free_area *area;
1021 /* Find a page of the appropriate size in the preferred list */
1022 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1023 area = &(zone->free_area[current_order]);
1024 if (list_empty(&area->free_list[migratetype]))
1027 page = list_entry(area->free_list[migratetype].next,
1029 list_del(&page->lru);
1030 rmv_page_order(page);
1032 expand(zone, page, order, current_order, area, migratetype);
1033 set_freepage_migratetype(page, migratetype);
1042 * This array describes the order lists are fallen back to when
1043 * the free lists for the desirable migrate type are depleted
1045 static int fallbacks[MIGRATE_TYPES][4] = {
1046 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1047 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1049 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1050 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1052 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1054 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1055 #ifdef CONFIG_MEMORY_ISOLATION
1056 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1061 * Move the free pages in a range to the free lists of the requested type.
1062 * Note that start_page and end_pages are not aligned on a pageblock
1063 * boundary. If alignment is required, use move_freepages_block()
1065 int move_freepages(struct zone *zone,
1066 struct page *start_page, struct page *end_page,
1070 unsigned long order;
1071 int pages_moved = 0;
1073 #ifndef CONFIG_HOLES_IN_ZONE
1075 * page_zone is not safe to call in this context when
1076 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1077 * anyway as we check zone boundaries in move_freepages_block().
1078 * Remove at a later date when no bug reports exist related to
1079 * grouping pages by mobility
1081 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1084 for (page = start_page; page <= end_page;) {
1085 /* Make sure we are not inadvertently changing nodes */
1086 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1088 if (!pfn_valid_within(page_to_pfn(page))) {
1093 if (!PageBuddy(page)) {
1098 order = page_order(page);
1099 list_move(&page->lru,
1100 &zone->free_area[order].free_list[migratetype]);
1101 set_freepage_migratetype(page, migratetype);
1103 pages_moved += 1 << order;
1109 int move_freepages_block(struct zone *zone, struct page *page,
1112 unsigned long start_pfn, end_pfn;
1113 struct page *start_page, *end_page;
1115 start_pfn = page_to_pfn(page);
1116 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1117 start_page = pfn_to_page(start_pfn);
1118 end_page = start_page + pageblock_nr_pages - 1;
1119 end_pfn = start_pfn + pageblock_nr_pages - 1;
1121 /* Do not cross zone boundaries */
1122 if (!zone_spans_pfn(zone, start_pfn))
1124 if (!zone_spans_pfn(zone, end_pfn))
1127 return move_freepages(zone, start_page, end_page, migratetype);
1130 static void change_pageblock_range(struct page *pageblock_page,
1131 int start_order, int migratetype)
1133 int nr_pageblocks = 1 << (start_order - pageblock_order);
1135 while (nr_pageblocks--) {
1136 set_pageblock_migratetype(pageblock_page, migratetype);
1137 pageblock_page += pageblock_nr_pages;
1142 * If breaking a large block of pages, move all free pages to the preferred
1143 * allocation list. If falling back for a reclaimable kernel allocation, be
1144 * more aggressive about taking ownership of free pages.
1146 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1147 * nor move CMA pages to different free lists. We don't want unmovable pages
1148 * to be allocated from MIGRATE_CMA areas.
1150 * Returns the new migratetype of the pageblock (or the same old migratetype
1151 * if it was unchanged).
1153 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1154 int start_type, int fallback_type)
1156 int current_order = page_order(page);
1159 * When borrowing from MIGRATE_CMA, we need to release the excess
1160 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1161 * is set to CMA so it is returned to the correct freelist in case
1162 * the page ends up being not actually allocated from the pcp lists.
1164 if (is_migrate_cma(fallback_type))
1165 return fallback_type;
1167 /* Take ownership for orders >= pageblock_order */
1168 if (current_order >= pageblock_order) {
1169 change_pageblock_range(page, current_order, start_type);
1173 if (current_order >= pageblock_order / 2 ||
1174 start_type == MIGRATE_RECLAIMABLE ||
1175 page_group_by_mobility_disabled) {
1178 pages = move_freepages_block(zone, page, start_type);
1180 /* Claim the whole block if over half of it is free */
1181 if (pages >= (1 << (pageblock_order-1)) ||
1182 page_group_by_mobility_disabled) {
1184 set_pageblock_migratetype(page, start_type);
1190 return fallback_type;
1193 /* Remove an element from the buddy allocator from the fallback list */
1194 static inline struct page *
1195 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1197 struct free_area *area;
1198 unsigned int current_order;
1200 int migratetype, new_type, i;
1202 /* Find the largest possible block of pages in the other list */
1203 for (current_order = MAX_ORDER-1;
1204 current_order >= order && current_order <= MAX_ORDER-1;
1207 migratetype = fallbacks[start_migratetype][i];
1209 /* MIGRATE_RESERVE handled later if necessary */
1210 if (migratetype == MIGRATE_RESERVE)
1213 area = &(zone->free_area[current_order]);
1214 if (list_empty(&area->free_list[migratetype]))
1217 page = list_entry(area->free_list[migratetype].next,
1221 new_type = try_to_steal_freepages(zone, page,
1225 /* Remove the page from the freelists */
1226 list_del(&page->lru);
1227 rmv_page_order(page);
1229 expand(zone, page, order, current_order, area,
1231 /* The freepage_migratetype may differ from pageblock's
1232 * migratetype depending on the decisions in
1233 * try_to_steal_freepages. This is OK as long as it does
1234 * not differ for MIGRATE_CMA type.
1236 set_freepage_migratetype(page, new_type);
1238 trace_mm_page_alloc_extfrag(page, order, current_order,
1239 start_migratetype, migratetype, new_type);
1249 * Do the hard work of removing an element from the buddy allocator.
1250 * Call me with the zone->lock already held.
1252 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1258 page = __rmqueue_smallest(zone, order, migratetype);
1260 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1261 page = __rmqueue_fallback(zone, order, migratetype);
1264 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1265 * is used because __rmqueue_smallest is an inline function
1266 * and we want just one call site
1269 migratetype = MIGRATE_RESERVE;
1274 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1279 * Obtain a specified number of elements from the buddy allocator, all under
1280 * a single hold of the lock, for efficiency. Add them to the supplied list.
1281 * Returns the number of new pages which were placed at *list.
1283 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1284 unsigned long count, struct list_head *list,
1285 int migratetype, bool cold)
1289 spin_lock(&zone->lock);
1290 for (i = 0; i < count; ++i) {
1291 struct page *page = __rmqueue(zone, order, migratetype);
1292 if (unlikely(page == NULL))
1296 * Split buddy pages returned by expand() are received here
1297 * in physical page order. The page is added to the callers and
1298 * list and the list head then moves forward. From the callers
1299 * perspective, the linked list is ordered by page number in
1300 * some conditions. This is useful for IO devices that can
1301 * merge IO requests if the physical pages are ordered
1305 list_add(&page->lru, list);
1307 list_add_tail(&page->lru, list);
1309 if (is_migrate_cma(get_freepage_migratetype(page)))
1310 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1313 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1314 spin_unlock(&zone->lock);
1320 * Called from the vmstat counter updater to drain pagesets of this
1321 * currently executing processor on remote nodes after they have
1324 * Note that this function must be called with the thread pinned to
1325 * a single processor.
1327 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1329 unsigned long flags;
1330 int to_drain, batch;
1332 local_irq_save(flags);
1333 batch = ACCESS_ONCE(pcp->batch);
1334 to_drain = min(pcp->count, batch);
1336 free_pcppages_bulk(zone, to_drain, pcp);
1337 pcp->count -= to_drain;
1339 local_irq_restore(flags);
1344 * Drain pcplists of the indicated processor and zone.
1346 * The processor must either be the current processor and the
1347 * thread pinned to the current processor or a processor that
1350 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1352 unsigned long flags;
1353 struct per_cpu_pageset *pset;
1354 struct per_cpu_pages *pcp;
1356 local_irq_save(flags);
1357 pset = per_cpu_ptr(zone->pageset, cpu);
1361 free_pcppages_bulk(zone, pcp->count, pcp);
1364 local_irq_restore(flags);
1368 * Drain pcplists of all zones on the indicated processor.
1370 * The processor must either be the current processor and the
1371 * thread pinned to the current processor or a processor that
1374 static void drain_pages(unsigned int cpu)
1378 for_each_populated_zone(zone) {
1379 drain_pages_zone(cpu, zone);
1384 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1386 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1387 * the single zone's pages.
1389 void drain_local_pages(struct zone *zone)
1391 int cpu = smp_processor_id();
1394 drain_pages_zone(cpu, zone);
1400 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1402 * When zone parameter is non-NULL, spill just the single zone's pages.
1404 * Note that this code is protected against sending an IPI to an offline
1405 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1406 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1407 * nothing keeps CPUs from showing up after we populated the cpumask and
1408 * before the call to on_each_cpu_mask().
1410 void drain_all_pages(struct zone *zone)
1415 * Allocate in the BSS so we wont require allocation in
1416 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1418 static cpumask_t cpus_with_pcps;
1421 * We don't care about racing with CPU hotplug event
1422 * as offline notification will cause the notified
1423 * cpu to drain that CPU pcps and on_each_cpu_mask
1424 * disables preemption as part of its processing
1426 for_each_online_cpu(cpu) {
1427 struct per_cpu_pageset *pcp;
1429 bool has_pcps = false;
1432 pcp = per_cpu_ptr(zone->pageset, cpu);
1436 for_each_populated_zone(z) {
1437 pcp = per_cpu_ptr(z->pageset, cpu);
1438 if (pcp->pcp.count) {
1446 cpumask_set_cpu(cpu, &cpus_with_pcps);
1448 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1450 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1454 #ifdef CONFIG_HIBERNATION
1456 void mark_free_pages(struct zone *zone)
1458 unsigned long pfn, max_zone_pfn;
1459 unsigned long flags;
1460 unsigned int order, t;
1461 struct list_head *curr;
1463 if (zone_is_empty(zone))
1466 spin_lock_irqsave(&zone->lock, flags);
1468 max_zone_pfn = zone_end_pfn(zone);
1469 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1470 if (pfn_valid(pfn)) {
1471 struct page *page = pfn_to_page(pfn);
1473 if (!swsusp_page_is_forbidden(page))
1474 swsusp_unset_page_free(page);
1477 for_each_migratetype_order(order, t) {
1478 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1481 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1482 for (i = 0; i < (1UL << order); i++)
1483 swsusp_set_page_free(pfn_to_page(pfn + i));
1486 spin_unlock_irqrestore(&zone->lock, flags);
1488 #endif /* CONFIG_PM */
1491 * Free a 0-order page
1492 * cold == true ? free a cold page : free a hot page
1494 void free_hot_cold_page(struct page *page, bool cold)
1496 struct zone *zone = page_zone(page);
1497 struct per_cpu_pages *pcp;
1498 unsigned long flags;
1499 unsigned long pfn = page_to_pfn(page);
1502 if (!free_pages_prepare(page, 0))
1505 migratetype = get_pfnblock_migratetype(page, pfn);
1506 set_freepage_migratetype(page, migratetype);
1507 local_irq_save(flags);
1508 __count_vm_event(PGFREE);
1511 * We only track unmovable, reclaimable and movable on pcp lists.
1512 * Free ISOLATE pages back to the allocator because they are being
1513 * offlined but treat RESERVE as movable pages so we can get those
1514 * areas back if necessary. Otherwise, we may have to free
1515 * excessively into the page allocator
1517 if (migratetype >= MIGRATE_PCPTYPES) {
1518 if (unlikely(is_migrate_isolate(migratetype))) {
1519 free_one_page(zone, page, pfn, 0, migratetype);
1522 migratetype = MIGRATE_MOVABLE;
1525 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1527 list_add(&page->lru, &pcp->lists[migratetype]);
1529 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1531 if (pcp->count >= pcp->high) {
1532 unsigned long batch = ACCESS_ONCE(pcp->batch);
1533 free_pcppages_bulk(zone, batch, pcp);
1534 pcp->count -= batch;
1538 local_irq_restore(flags);
1542 * Free a list of 0-order pages
1544 void free_hot_cold_page_list(struct list_head *list, bool cold)
1546 struct page *page, *next;
1548 list_for_each_entry_safe(page, next, list, lru) {
1549 trace_mm_page_free_batched(page, cold);
1550 free_hot_cold_page(page, cold);
1555 * split_page takes a non-compound higher-order page, and splits it into
1556 * n (1<<order) sub-pages: page[0..n]
1557 * Each sub-page must be freed individually.
1559 * Note: this is probably too low level an operation for use in drivers.
1560 * Please consult with lkml before using this in your driver.
1562 void split_page(struct page *page, unsigned int order)
1566 VM_BUG_ON_PAGE(PageCompound(page), page);
1567 VM_BUG_ON_PAGE(!page_count(page), page);
1569 #ifdef CONFIG_KMEMCHECK
1571 * Split shadow pages too, because free(page[0]) would
1572 * otherwise free the whole shadow.
1574 if (kmemcheck_page_is_tracked(page))
1575 split_page(virt_to_page(page[0].shadow), order);
1578 set_page_owner(page, 0, 0);
1579 for (i = 1; i < (1 << order); i++) {
1580 set_page_refcounted(page + i);
1581 set_page_owner(page + i, 0, 0);
1584 EXPORT_SYMBOL_GPL(split_page);
1586 int __isolate_free_page(struct page *page, unsigned int order)
1588 unsigned long watermark;
1592 BUG_ON(!PageBuddy(page));
1594 zone = page_zone(page);
1595 mt = get_pageblock_migratetype(page);
1597 if (!is_migrate_isolate(mt)) {
1598 /* Obey watermarks as if the page was being allocated */
1599 watermark = low_wmark_pages(zone) + (1 << order);
1600 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1603 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1606 /* Remove page from free list */
1607 list_del(&page->lru);
1608 zone->free_area[order].nr_free--;
1609 rmv_page_order(page);
1611 /* Set the pageblock if the isolated page is at least a pageblock */
1612 if (order >= pageblock_order - 1) {
1613 struct page *endpage = page + (1 << order) - 1;
1614 for (; page < endpage; page += pageblock_nr_pages) {
1615 int mt = get_pageblock_migratetype(page);
1616 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1617 set_pageblock_migratetype(page,
1622 set_page_owner(page, order, 0);
1623 return 1UL << order;
1627 * Similar to split_page except the page is already free. As this is only
1628 * being used for migration, the migratetype of the block also changes.
1629 * As this is called with interrupts disabled, the caller is responsible
1630 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1633 * Note: this is probably too low level an operation for use in drivers.
1634 * Please consult with lkml before using this in your driver.
1636 int split_free_page(struct page *page)
1641 order = page_order(page);
1643 nr_pages = __isolate_free_page(page, order);
1647 /* Split into individual pages */
1648 set_page_refcounted(page);
1649 split_page(page, order);
1654 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1657 struct page *buffered_rmqueue(struct zone *preferred_zone,
1658 struct zone *zone, unsigned int order,
1659 gfp_t gfp_flags, int migratetype)
1661 unsigned long flags;
1663 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1665 if (likely(order == 0)) {
1666 struct per_cpu_pages *pcp;
1667 struct list_head *list;
1669 local_irq_save(flags);
1670 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1671 list = &pcp->lists[migratetype];
1672 if (list_empty(list)) {
1673 pcp->count += rmqueue_bulk(zone, 0,
1676 if (unlikely(list_empty(list)))
1681 page = list_entry(list->prev, struct page, lru);
1683 page = list_entry(list->next, struct page, lru);
1685 list_del(&page->lru);
1688 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1690 * __GFP_NOFAIL is not to be used in new code.
1692 * All __GFP_NOFAIL callers should be fixed so that they
1693 * properly detect and handle allocation failures.
1695 * We most definitely don't want callers attempting to
1696 * allocate greater than order-1 page units with
1699 WARN_ON_ONCE(order > 1);
1701 spin_lock_irqsave(&zone->lock, flags);
1702 page = __rmqueue(zone, order, migratetype);
1703 spin_unlock(&zone->lock);
1706 __mod_zone_freepage_state(zone, -(1 << order),
1707 get_freepage_migratetype(page));
1710 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1711 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1712 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1713 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1715 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1716 zone_statistics(preferred_zone, zone, gfp_flags);
1717 local_irq_restore(flags);
1719 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1723 local_irq_restore(flags);
1727 #ifdef CONFIG_FAIL_PAGE_ALLOC
1730 struct fault_attr attr;
1732 u32 ignore_gfp_highmem;
1733 u32 ignore_gfp_wait;
1735 } fail_page_alloc = {
1736 .attr = FAULT_ATTR_INITIALIZER,
1737 .ignore_gfp_wait = 1,
1738 .ignore_gfp_highmem = 1,
1742 static int __init setup_fail_page_alloc(char *str)
1744 return setup_fault_attr(&fail_page_alloc.attr, str);
1746 __setup("fail_page_alloc=", setup_fail_page_alloc);
1748 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1750 if (order < fail_page_alloc.min_order)
1752 if (gfp_mask & __GFP_NOFAIL)
1754 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1756 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1759 return should_fail(&fail_page_alloc.attr, 1 << order);
1762 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1764 static int __init fail_page_alloc_debugfs(void)
1766 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1769 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1770 &fail_page_alloc.attr);
1772 return PTR_ERR(dir);
1774 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1775 &fail_page_alloc.ignore_gfp_wait))
1777 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1778 &fail_page_alloc.ignore_gfp_highmem))
1780 if (!debugfs_create_u32("min-order", mode, dir,
1781 &fail_page_alloc.min_order))
1786 debugfs_remove_recursive(dir);
1791 late_initcall(fail_page_alloc_debugfs);
1793 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1795 #else /* CONFIG_FAIL_PAGE_ALLOC */
1797 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1802 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1805 * Return true if free pages are above 'mark'. This takes into account the order
1806 * of the allocation.
1808 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1809 unsigned long mark, int classzone_idx, int alloc_flags,
1812 /* free_pages may go negative - that's OK */
1817 free_pages -= (1 << order) - 1;
1818 if (alloc_flags & ALLOC_HIGH)
1820 if (alloc_flags & ALLOC_HARDER)
1823 /* If allocation can't use CMA areas don't use free CMA pages */
1824 if (!(alloc_flags & ALLOC_CMA))
1825 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1828 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1830 for (o = 0; o < order; o++) {
1831 /* At the next order, this order's pages become unavailable */
1832 free_pages -= z->free_area[o].nr_free << o;
1834 /* Require fewer higher order pages to be free */
1837 if (free_pages <= min)
1843 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1844 int classzone_idx, int alloc_flags)
1846 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1847 zone_page_state(z, NR_FREE_PAGES));
1850 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1851 unsigned long mark, int classzone_idx, int alloc_flags)
1853 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1855 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1856 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1858 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1864 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1865 * skip over zones that are not allowed by the cpuset, or that have
1866 * been recently (in last second) found to be nearly full. See further
1867 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1868 * that have to skip over a lot of full or unallowed zones.
1870 * If the zonelist cache is present in the passed zonelist, then
1871 * returns a pointer to the allowed node mask (either the current
1872 * tasks mems_allowed, or node_states[N_MEMORY].)
1874 * If the zonelist cache is not available for this zonelist, does
1875 * nothing and returns NULL.
1877 * If the fullzones BITMAP in the zonelist cache is stale (more than
1878 * a second since last zap'd) then we zap it out (clear its bits.)
1880 * We hold off even calling zlc_setup, until after we've checked the
1881 * first zone in the zonelist, on the theory that most allocations will
1882 * be satisfied from that first zone, so best to examine that zone as
1883 * quickly as we can.
1885 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1887 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1888 nodemask_t *allowednodes; /* zonelist_cache approximation */
1890 zlc = zonelist->zlcache_ptr;
1894 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1895 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1896 zlc->last_full_zap = jiffies;
1899 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1900 &cpuset_current_mems_allowed :
1901 &node_states[N_MEMORY];
1902 return allowednodes;
1906 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1907 * if it is worth looking at further for free memory:
1908 * 1) Check that the zone isn't thought to be full (doesn't have its
1909 * bit set in the zonelist_cache fullzones BITMAP).
1910 * 2) Check that the zones node (obtained from the zonelist_cache
1911 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1912 * Return true (non-zero) if zone is worth looking at further, or
1913 * else return false (zero) if it is not.
1915 * This check -ignores- the distinction between various watermarks,
1916 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1917 * found to be full for any variation of these watermarks, it will
1918 * be considered full for up to one second by all requests, unless
1919 * we are so low on memory on all allowed nodes that we are forced
1920 * into the second scan of the zonelist.
1922 * In the second scan we ignore this zonelist cache and exactly
1923 * apply the watermarks to all zones, even it is slower to do so.
1924 * We are low on memory in the second scan, and should leave no stone
1925 * unturned looking for a free page.
1927 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1928 nodemask_t *allowednodes)
1930 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1931 int i; /* index of *z in zonelist zones */
1932 int n; /* node that zone *z is on */
1934 zlc = zonelist->zlcache_ptr;
1938 i = z - zonelist->_zonerefs;
1941 /* This zone is worth trying if it is allowed but not full */
1942 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1946 * Given 'z' scanning a zonelist, set the corresponding bit in
1947 * zlc->fullzones, so that subsequent attempts to allocate a page
1948 * from that zone don't waste time re-examining it.
1950 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1952 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1953 int i; /* index of *z in zonelist zones */
1955 zlc = zonelist->zlcache_ptr;
1959 i = z - zonelist->_zonerefs;
1961 set_bit(i, zlc->fullzones);
1965 * clear all zones full, called after direct reclaim makes progress so that
1966 * a zone that was recently full is not skipped over for up to a second
1968 static void zlc_clear_zones_full(struct zonelist *zonelist)
1970 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1972 zlc = zonelist->zlcache_ptr;
1976 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1979 static bool zone_local(struct zone *local_zone, struct zone *zone)
1981 return local_zone->node == zone->node;
1984 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1986 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1990 #else /* CONFIG_NUMA */
1992 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1997 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1998 nodemask_t *allowednodes)
2003 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2007 static void zlc_clear_zones_full(struct zonelist *zonelist)
2011 static bool zone_local(struct zone *local_zone, struct zone *zone)
2016 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2021 #endif /* CONFIG_NUMA */
2023 static void reset_alloc_batches(struct zone *preferred_zone)
2025 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2028 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2029 high_wmark_pages(zone) - low_wmark_pages(zone) -
2030 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2031 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2032 } while (zone++ != preferred_zone);
2036 * get_page_from_freelist goes through the zonelist trying to allocate
2039 static struct page *
2040 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
2041 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
2042 struct zone *preferred_zone, int classzone_idx, int migratetype)
2045 struct page *page = NULL;
2047 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2048 int zlc_active = 0; /* set if using zonelist_cache */
2049 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2050 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2051 (gfp_mask & __GFP_WRITE);
2052 int nr_fair_skipped = 0;
2053 bool zonelist_rescan;
2056 zonelist_rescan = false;
2059 * Scan zonelist, looking for a zone with enough free.
2060 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2062 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2063 high_zoneidx, nodemask) {
2066 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2067 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2069 if (cpusets_enabled() &&
2070 (alloc_flags & ALLOC_CPUSET) &&
2071 !cpuset_zone_allowed(zone, gfp_mask))
2074 * Distribute pages in proportion to the individual
2075 * zone size to ensure fair page aging. The zone a
2076 * page was allocated in should have no effect on the
2077 * time the page has in memory before being reclaimed.
2079 if (alloc_flags & ALLOC_FAIR) {
2080 if (!zone_local(preferred_zone, zone))
2082 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2088 * When allocating a page cache page for writing, we
2089 * want to get it from a zone that is within its dirty
2090 * limit, such that no single zone holds more than its
2091 * proportional share of globally allowed dirty pages.
2092 * The dirty limits take into account the zone's
2093 * lowmem reserves and high watermark so that kswapd
2094 * should be able to balance it without having to
2095 * write pages from its LRU list.
2097 * This may look like it could increase pressure on
2098 * lower zones by failing allocations in higher zones
2099 * before they are full. But the pages that do spill
2100 * over are limited as the lower zones are protected
2101 * by this very same mechanism. It should not become
2102 * a practical burden to them.
2104 * XXX: For now, allow allocations to potentially
2105 * exceed the per-zone dirty limit in the slowpath
2106 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2107 * which is important when on a NUMA setup the allowed
2108 * zones are together not big enough to reach the
2109 * global limit. The proper fix for these situations
2110 * will require awareness of zones in the
2111 * dirty-throttling and the flusher threads.
2113 if (consider_zone_dirty && !zone_dirty_ok(zone))
2116 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2117 if (!zone_watermark_ok(zone, order, mark,
2118 classzone_idx, alloc_flags)) {
2121 /* Checked here to keep the fast path fast */
2122 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2123 if (alloc_flags & ALLOC_NO_WATERMARKS)
2126 if (IS_ENABLED(CONFIG_NUMA) &&
2127 !did_zlc_setup && nr_online_nodes > 1) {
2129 * we do zlc_setup if there are multiple nodes
2130 * and before considering the first zone allowed
2133 allowednodes = zlc_setup(zonelist, alloc_flags);
2138 if (zone_reclaim_mode == 0 ||
2139 !zone_allows_reclaim(preferred_zone, zone))
2140 goto this_zone_full;
2143 * As we may have just activated ZLC, check if the first
2144 * eligible zone has failed zone_reclaim recently.
2146 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2147 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2150 ret = zone_reclaim(zone, gfp_mask, order);
2152 case ZONE_RECLAIM_NOSCAN:
2155 case ZONE_RECLAIM_FULL:
2156 /* scanned but unreclaimable */
2159 /* did we reclaim enough */
2160 if (zone_watermark_ok(zone, order, mark,
2161 classzone_idx, alloc_flags))
2165 * Failed to reclaim enough to meet watermark.
2166 * Only mark the zone full if checking the min
2167 * watermark or if we failed to reclaim just
2168 * 1<<order pages or else the page allocator
2169 * fastpath will prematurely mark zones full
2170 * when the watermark is between the low and
2173 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2174 ret == ZONE_RECLAIM_SOME)
2175 goto this_zone_full;
2182 page = buffered_rmqueue(preferred_zone, zone, order,
2183 gfp_mask, migratetype);
2185 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2190 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2191 zlc_mark_zone_full(zonelist, z);
2195 * The first pass makes sure allocations are spread fairly within the
2196 * local node. However, the local node might have free pages left
2197 * after the fairness batches are exhausted, and remote zones haven't
2198 * even been considered yet. Try once more without fairness, and
2199 * include remote zones now, before entering the slowpath and waking
2200 * kswapd: prefer spilling to a remote zone over swapping locally.
2202 if (alloc_flags & ALLOC_FAIR) {
2203 alloc_flags &= ~ALLOC_FAIR;
2204 if (nr_fair_skipped) {
2205 zonelist_rescan = true;
2206 reset_alloc_batches(preferred_zone);
2208 if (nr_online_nodes > 1)
2209 zonelist_rescan = true;
2212 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2213 /* Disable zlc cache for second zonelist scan */
2215 zonelist_rescan = true;
2218 if (zonelist_rescan)
2225 * Large machines with many possible nodes should not always dump per-node
2226 * meminfo in irq context.
2228 static inline bool should_suppress_show_mem(void)
2233 ret = in_interrupt();
2238 static DEFINE_RATELIMIT_STATE(nopage_rs,
2239 DEFAULT_RATELIMIT_INTERVAL,
2240 DEFAULT_RATELIMIT_BURST);
2242 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2244 unsigned int filter = SHOW_MEM_FILTER_NODES;
2246 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2247 debug_guardpage_minorder() > 0)
2251 * This documents exceptions given to allocations in certain
2252 * contexts that are allowed to allocate outside current's set
2255 if (!(gfp_mask & __GFP_NOMEMALLOC))
2256 if (test_thread_flag(TIF_MEMDIE) ||
2257 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2258 filter &= ~SHOW_MEM_FILTER_NODES;
2259 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2260 filter &= ~SHOW_MEM_FILTER_NODES;
2263 struct va_format vaf;
2266 va_start(args, fmt);
2271 pr_warn("%pV", &vaf);
2276 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2277 current->comm, order, gfp_mask);
2280 if (!should_suppress_show_mem())
2285 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2286 unsigned long did_some_progress,
2287 unsigned long pages_reclaimed)
2289 /* Do not loop if specifically requested */
2290 if (gfp_mask & __GFP_NORETRY)
2293 /* Always retry if specifically requested */
2294 if (gfp_mask & __GFP_NOFAIL)
2298 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2299 * making forward progress without invoking OOM. Suspend also disables
2300 * storage devices so kswapd will not help. Bail if we are suspending.
2302 if (!did_some_progress && pm_suspended_storage())
2306 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2307 * means __GFP_NOFAIL, but that may not be true in other
2310 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2314 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2315 * specified, then we retry until we no longer reclaim any pages
2316 * (above), or we've reclaimed an order of pages at least as
2317 * large as the allocation's order. In both cases, if the
2318 * allocation still fails, we stop retrying.
2320 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2326 static inline struct page *
2327 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2328 struct zonelist *zonelist, enum zone_type high_zoneidx,
2329 nodemask_t *nodemask, struct zone *preferred_zone,
2330 int classzone_idx, int migratetype, unsigned long *did_some_progress)
2334 *did_some_progress = 0;
2336 if (oom_killer_disabled)
2340 * Acquire the per-zone oom lock for each zone. If that
2341 * fails, somebody else is making progress for us.
2343 if (!oom_zonelist_trylock(zonelist, gfp_mask)) {
2344 *did_some_progress = 1;
2345 schedule_timeout_uninterruptible(1);
2350 * PM-freezer should be notified that there might be an OOM killer on
2351 * its way to kill and wake somebody up. This is too early and we might
2352 * end up not killing anything but false positives are acceptable.
2353 * See freeze_processes.
2358 * Go through the zonelist yet one more time, keep very high watermark
2359 * here, this is only to catch a parallel oom killing, we must fail if
2360 * we're still under heavy pressure.
2362 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2363 order, zonelist, high_zoneidx,
2364 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2365 preferred_zone, classzone_idx, migratetype);
2369 if (!(gfp_mask & __GFP_NOFAIL)) {
2370 /* Coredumps can quickly deplete all memory reserves */
2371 if (current->flags & PF_DUMPCORE)
2373 /* The OOM killer will not help higher order allocs */
2374 if (order > PAGE_ALLOC_COSTLY_ORDER)
2376 /* The OOM killer does not needlessly kill tasks for lowmem */
2377 if (high_zoneidx < ZONE_NORMAL)
2379 /* The OOM killer does not compensate for light reclaim */
2380 if (!(gfp_mask & __GFP_FS))
2383 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2384 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2385 * The caller should handle page allocation failure by itself if
2386 * it specifies __GFP_THISNODE.
2387 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2389 if (gfp_mask & __GFP_THISNODE)
2392 /* Exhausted what can be done so it's blamo time */
2393 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2394 *did_some_progress = 1;
2396 oom_zonelist_unlock(zonelist, gfp_mask);
2400 #ifdef CONFIG_COMPACTION
2401 /* Try memory compaction for high-order allocations before reclaim */
2402 static struct page *
2403 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2404 struct zonelist *zonelist, enum zone_type high_zoneidx,
2405 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2406 int classzone_idx, int migratetype, enum migrate_mode mode,
2407 int *contended_compaction, bool *deferred_compaction)
2409 unsigned long compact_result;
2415 current->flags |= PF_MEMALLOC;
2416 compact_result = try_to_compact_pages(zonelist, order, gfp_mask,
2418 contended_compaction,
2419 alloc_flags, classzone_idx);
2420 current->flags &= ~PF_MEMALLOC;
2422 switch (compact_result) {
2423 case COMPACT_DEFERRED:
2424 *deferred_compaction = true;
2426 case COMPACT_SKIPPED:
2433 * At least in one zone compaction wasn't deferred or skipped, so let's
2434 * count a compaction stall
2436 count_vm_event(COMPACTSTALL);
2438 page = get_page_from_freelist(gfp_mask, nodemask,
2439 order, zonelist, high_zoneidx,
2440 alloc_flags & ~ALLOC_NO_WATERMARKS,
2441 preferred_zone, classzone_idx, migratetype);
2444 struct zone *zone = page_zone(page);
2446 zone->compact_blockskip_flush = false;
2447 compaction_defer_reset(zone, order, true);
2448 count_vm_event(COMPACTSUCCESS);
2453 * It's bad if compaction run occurs and fails. The most likely reason
2454 * is that pages exist, but not enough to satisfy watermarks.
2456 count_vm_event(COMPACTFAIL);
2463 static inline struct page *
2464 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2465 struct zonelist *zonelist, enum zone_type high_zoneidx,
2466 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2467 int classzone_idx, int migratetype, enum migrate_mode mode,
2468 int *contended_compaction, bool *deferred_compaction)
2472 #endif /* CONFIG_COMPACTION */
2474 /* Perform direct synchronous page reclaim */
2476 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2477 nodemask_t *nodemask)
2479 struct reclaim_state reclaim_state;
2484 /* We now go into synchronous reclaim */
2485 cpuset_memory_pressure_bump();
2486 current->flags |= PF_MEMALLOC;
2487 lockdep_set_current_reclaim_state(gfp_mask);
2488 reclaim_state.reclaimed_slab = 0;
2489 current->reclaim_state = &reclaim_state;
2491 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2493 current->reclaim_state = NULL;
2494 lockdep_clear_current_reclaim_state();
2495 current->flags &= ~PF_MEMALLOC;
2502 /* The really slow allocator path where we enter direct reclaim */
2503 static inline struct page *
2504 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2505 struct zonelist *zonelist, enum zone_type high_zoneidx,
2506 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2507 int classzone_idx, int migratetype, unsigned long *did_some_progress)
2509 struct page *page = NULL;
2510 bool drained = false;
2512 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2514 if (unlikely(!(*did_some_progress)))
2517 /* After successful reclaim, reconsider all zones for allocation */
2518 if (IS_ENABLED(CONFIG_NUMA))
2519 zlc_clear_zones_full(zonelist);
2522 page = get_page_from_freelist(gfp_mask, nodemask, order,
2523 zonelist, high_zoneidx,
2524 alloc_flags & ~ALLOC_NO_WATERMARKS,
2525 preferred_zone, classzone_idx,
2529 * If an allocation failed after direct reclaim, it could be because
2530 * pages are pinned on the per-cpu lists. Drain them and try again
2532 if (!page && !drained) {
2533 drain_all_pages(NULL);
2542 * This is called in the allocator slow-path if the allocation request is of
2543 * sufficient urgency to ignore watermarks and take other desperate measures
2545 static inline struct page *
2546 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2547 struct zonelist *zonelist, enum zone_type high_zoneidx,
2548 nodemask_t *nodemask, struct zone *preferred_zone,
2549 int classzone_idx, int migratetype)
2554 page = get_page_from_freelist(gfp_mask, nodemask, order,
2555 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2556 preferred_zone, classzone_idx, migratetype);
2558 if (!page && gfp_mask & __GFP_NOFAIL)
2559 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2560 } while (!page && (gfp_mask & __GFP_NOFAIL));
2565 static void wake_all_kswapds(unsigned int order,
2566 struct zonelist *zonelist,
2567 enum zone_type high_zoneidx,
2568 struct zone *preferred_zone,
2569 nodemask_t *nodemask)
2574 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2575 high_zoneidx, nodemask)
2576 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2580 gfp_to_alloc_flags(gfp_t gfp_mask)
2582 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2583 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2585 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2586 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2589 * The caller may dip into page reserves a bit more if the caller
2590 * cannot run direct reclaim, or if the caller has realtime scheduling
2591 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2592 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2594 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2598 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2599 * if it can't schedule.
2601 if (!(gfp_mask & __GFP_NOMEMALLOC))
2602 alloc_flags |= ALLOC_HARDER;
2604 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2605 * comment for __cpuset_node_allowed().
2607 alloc_flags &= ~ALLOC_CPUSET;
2608 } else if (unlikely(rt_task(current)) && !in_interrupt())
2609 alloc_flags |= ALLOC_HARDER;
2611 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2612 if (gfp_mask & __GFP_MEMALLOC)
2613 alloc_flags |= ALLOC_NO_WATERMARKS;
2614 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2615 alloc_flags |= ALLOC_NO_WATERMARKS;
2616 else if (!in_interrupt() &&
2617 ((current->flags & PF_MEMALLOC) ||
2618 unlikely(test_thread_flag(TIF_MEMDIE))))
2619 alloc_flags |= ALLOC_NO_WATERMARKS;
2622 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2623 alloc_flags |= ALLOC_CMA;
2628 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2630 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2633 static inline struct page *
2634 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2635 struct zonelist *zonelist, enum zone_type high_zoneidx,
2636 nodemask_t *nodemask, struct zone *preferred_zone,
2637 int classzone_idx, int migratetype)
2639 const gfp_t wait = gfp_mask & __GFP_WAIT;
2640 struct page *page = NULL;
2642 unsigned long pages_reclaimed = 0;
2643 unsigned long did_some_progress;
2644 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2645 bool deferred_compaction = false;
2646 int contended_compaction = COMPACT_CONTENDED_NONE;
2649 * In the slowpath, we sanity check order to avoid ever trying to
2650 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2651 * be using allocators in order of preference for an area that is
2654 if (order >= MAX_ORDER) {
2655 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2660 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2661 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2662 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2663 * using a larger set of nodes after it has established that the
2664 * allowed per node queues are empty and that nodes are
2667 if (IS_ENABLED(CONFIG_NUMA) &&
2668 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2672 if (!(gfp_mask & __GFP_NO_KSWAPD))
2673 wake_all_kswapds(order, zonelist, high_zoneidx,
2674 preferred_zone, nodemask);
2677 * OK, we're below the kswapd watermark and have kicked background
2678 * reclaim. Now things get more complex, so set up alloc_flags according
2679 * to how we want to proceed.
2681 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2684 * Find the true preferred zone if the allocation is unconstrained by
2687 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) {
2688 struct zoneref *preferred_zoneref;
2689 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2690 NULL, &preferred_zone);
2691 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2694 /* This is the last chance, in general, before the goto nopage. */
2695 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2696 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2697 preferred_zone, classzone_idx, migratetype);
2701 /* Allocate without watermarks if the context allows */
2702 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2704 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2705 * the allocation is high priority and these type of
2706 * allocations are system rather than user orientated
2708 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2710 page = __alloc_pages_high_priority(gfp_mask, order,
2711 zonelist, high_zoneidx, nodemask,
2712 preferred_zone, classzone_idx, migratetype);
2718 /* Atomic allocations - we can't balance anything */
2721 * All existing users of the deprecated __GFP_NOFAIL are
2722 * blockable, so warn of any new users that actually allow this
2723 * type of allocation to fail.
2725 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2729 /* Avoid recursion of direct reclaim */
2730 if (current->flags & PF_MEMALLOC)
2733 /* Avoid allocations with no watermarks from looping endlessly */
2734 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2738 * Try direct compaction. The first pass is asynchronous. Subsequent
2739 * attempts after direct reclaim are synchronous
2741 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2742 high_zoneidx, nodemask, alloc_flags,
2744 classzone_idx, migratetype,
2745 migration_mode, &contended_compaction,
2746 &deferred_compaction);
2750 /* Checks for THP-specific high-order allocations */
2751 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2753 * If compaction is deferred for high-order allocations, it is
2754 * because sync compaction recently failed. If this is the case
2755 * and the caller requested a THP allocation, we do not want
2756 * to heavily disrupt the system, so we fail the allocation
2757 * instead of entering direct reclaim.
2759 if (deferred_compaction)
2763 * In all zones where compaction was attempted (and not
2764 * deferred or skipped), lock contention has been detected.
2765 * For THP allocation we do not want to disrupt the others
2766 * so we fallback to base pages instead.
2768 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2772 * If compaction was aborted due to need_resched(), we do not
2773 * want to further increase allocation latency, unless it is
2774 * khugepaged trying to collapse.
2776 if (contended_compaction == COMPACT_CONTENDED_SCHED
2777 && !(current->flags & PF_KTHREAD))
2782 * It can become very expensive to allocate transparent hugepages at
2783 * fault, so use asynchronous memory compaction for THP unless it is
2784 * khugepaged trying to collapse.
2786 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2787 (current->flags & PF_KTHREAD))
2788 migration_mode = MIGRATE_SYNC_LIGHT;
2790 /* Try direct reclaim and then allocating */
2791 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2792 zonelist, high_zoneidx,
2794 alloc_flags, preferred_zone,
2795 classzone_idx, migratetype,
2796 &did_some_progress);
2800 /* Check if we should retry the allocation */
2801 pages_reclaimed += did_some_progress;
2802 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2805 * If we fail to make progress by freeing individual
2806 * pages, but the allocation wants us to keep going,
2807 * start OOM killing tasks.
2809 if (!did_some_progress) {
2810 page = __alloc_pages_may_oom(gfp_mask, order, zonelist,
2811 high_zoneidx, nodemask,
2812 preferred_zone, classzone_idx,
2813 migratetype,&did_some_progress);
2816 if (!did_some_progress)
2819 /* Wait for some write requests to complete then retry */
2820 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2824 * High-order allocations do not necessarily loop after
2825 * direct reclaim and reclaim/compaction depends on compaction
2826 * being called after reclaim so call directly if necessary
2828 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2829 high_zoneidx, nodemask, alloc_flags,
2831 classzone_idx, migratetype,
2832 migration_mode, &contended_compaction,
2833 &deferred_compaction);
2839 warn_alloc_failed(gfp_mask, order, NULL);
2845 * This is the 'heart' of the zoned buddy allocator.
2848 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2849 struct zonelist *zonelist, nodemask_t *nodemask)
2851 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2852 struct zone *preferred_zone;
2853 struct zoneref *preferred_zoneref;
2854 struct page *page = NULL;
2855 int migratetype = gfpflags_to_migratetype(gfp_mask);
2856 unsigned int cpuset_mems_cookie;
2857 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2859 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
2861 gfp_mask &= gfp_allowed_mask;
2863 lockdep_trace_alloc(gfp_mask);
2865 might_sleep_if(gfp_mask & __GFP_WAIT);
2867 if (should_fail_alloc_page(gfp_mask, order))
2871 * Check the zones suitable for the gfp_mask contain at least one
2872 * valid zone. It's possible to have an empty zonelist as a result
2873 * of GFP_THISNODE and a memoryless node
2875 if (unlikely(!zonelist->_zonerefs->zone))
2878 if (IS_ENABLED(CONFIG_CMA) && migratetype == MIGRATE_MOVABLE)
2879 alloc_flags |= ALLOC_CMA;
2882 cpuset_mems_cookie = read_mems_allowed_begin();
2884 /* The preferred zone is used for statistics later */
2885 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2886 nodemask ? : &cpuset_current_mems_allowed,
2888 if (!preferred_zone)
2890 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2892 /* First allocation attempt */
2893 alloc_mask = gfp_mask|__GFP_HARDWALL;
2894 page = get_page_from_freelist(alloc_mask, nodemask, order, zonelist,
2895 high_zoneidx, alloc_flags, preferred_zone,
2896 classzone_idx, migratetype);
2897 if (unlikely(!page)) {
2899 * Runtime PM, block IO and its error handling path
2900 * can deadlock because I/O on the device might not
2903 alloc_mask = memalloc_noio_flags(gfp_mask);
2905 page = __alloc_pages_slowpath(alloc_mask, order,
2906 zonelist, high_zoneidx, nodemask,
2907 preferred_zone, classzone_idx, migratetype);
2910 if (kmemcheck_enabled && page)
2911 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2913 trace_mm_page_alloc(page, order, alloc_mask, migratetype);
2917 * When updating a task's mems_allowed, it is possible to race with
2918 * parallel threads in such a way that an allocation can fail while
2919 * the mask is being updated. If a page allocation is about to fail,
2920 * check if the cpuset changed during allocation and if so, retry.
2922 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2927 EXPORT_SYMBOL(__alloc_pages_nodemask);
2930 * Common helper functions.
2932 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2937 * __get_free_pages() returns a 32-bit address, which cannot represent
2940 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2942 page = alloc_pages(gfp_mask, order);
2945 return (unsigned long) page_address(page);
2947 EXPORT_SYMBOL(__get_free_pages);
2949 unsigned long get_zeroed_page(gfp_t gfp_mask)
2951 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2953 EXPORT_SYMBOL(get_zeroed_page);
2955 void __free_pages(struct page *page, unsigned int order)
2957 if (put_page_testzero(page)) {
2959 free_hot_cold_page(page, false);
2961 __free_pages_ok(page, order);
2965 EXPORT_SYMBOL(__free_pages);
2967 void free_pages(unsigned long addr, unsigned int order)
2970 VM_BUG_ON(!virt_addr_valid((void *)addr));
2971 __free_pages(virt_to_page((void *)addr), order);
2975 EXPORT_SYMBOL(free_pages);
2978 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2979 * of the current memory cgroup.
2981 * It should be used when the caller would like to use kmalloc, but since the
2982 * allocation is large, it has to fall back to the page allocator.
2984 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2987 struct mem_cgroup *memcg = NULL;
2989 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2991 page = alloc_pages(gfp_mask, order);
2992 memcg_kmem_commit_charge(page, memcg, order);
2996 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2999 struct mem_cgroup *memcg = NULL;
3001 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3003 page = alloc_pages_node(nid, gfp_mask, order);
3004 memcg_kmem_commit_charge(page, memcg, order);
3009 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3012 void __free_kmem_pages(struct page *page, unsigned int order)
3014 memcg_kmem_uncharge_pages(page, order);
3015 __free_pages(page, order);
3018 void free_kmem_pages(unsigned long addr, unsigned int order)
3021 VM_BUG_ON(!virt_addr_valid((void *)addr));
3022 __free_kmem_pages(virt_to_page((void *)addr), order);
3026 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3029 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3030 unsigned long used = addr + PAGE_ALIGN(size);
3032 split_page(virt_to_page((void *)addr), order);
3033 while (used < alloc_end) {
3038 return (void *)addr;
3042 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3043 * @size: the number of bytes to allocate
3044 * @gfp_mask: GFP flags for the allocation
3046 * This function is similar to alloc_pages(), except that it allocates the
3047 * minimum number of pages to satisfy the request. alloc_pages() can only
3048 * allocate memory in power-of-two pages.
3050 * This function is also limited by MAX_ORDER.
3052 * Memory allocated by this function must be released by free_pages_exact().
3054 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3056 unsigned int order = get_order(size);
3059 addr = __get_free_pages(gfp_mask, order);
3060 return make_alloc_exact(addr, order, size);
3062 EXPORT_SYMBOL(alloc_pages_exact);
3065 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3067 * @nid: the preferred node ID where memory should be allocated
3068 * @size: the number of bytes to allocate
3069 * @gfp_mask: GFP flags for the allocation
3071 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3073 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3076 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3078 unsigned order = get_order(size);
3079 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3082 return make_alloc_exact((unsigned long)page_address(p), order, size);
3086 * free_pages_exact - release memory allocated via alloc_pages_exact()
3087 * @virt: the value returned by alloc_pages_exact.
3088 * @size: size of allocation, same value as passed to alloc_pages_exact().
3090 * Release the memory allocated by a previous call to alloc_pages_exact.
3092 void free_pages_exact(void *virt, size_t size)
3094 unsigned long addr = (unsigned long)virt;
3095 unsigned long end = addr + PAGE_ALIGN(size);
3097 while (addr < end) {
3102 EXPORT_SYMBOL(free_pages_exact);
3105 * nr_free_zone_pages - count number of pages beyond high watermark
3106 * @offset: The zone index of the highest zone
3108 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3109 * high watermark within all zones at or below a given zone index. For each
3110 * zone, the number of pages is calculated as:
3111 * managed_pages - high_pages
3113 static unsigned long nr_free_zone_pages(int offset)
3118 /* Just pick one node, since fallback list is circular */
3119 unsigned long sum = 0;
3121 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3123 for_each_zone_zonelist(zone, z, zonelist, offset) {
3124 unsigned long size = zone->managed_pages;
3125 unsigned long high = high_wmark_pages(zone);
3134 * nr_free_buffer_pages - count number of pages beyond high watermark
3136 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3137 * watermark within ZONE_DMA and ZONE_NORMAL.
3139 unsigned long nr_free_buffer_pages(void)
3141 return nr_free_zone_pages(gfp_zone(GFP_USER));
3143 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3146 * nr_free_pagecache_pages - count number of pages beyond high watermark
3148 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3149 * high watermark within all zones.
3151 unsigned long nr_free_pagecache_pages(void)
3153 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3156 static inline void show_node(struct zone *zone)
3158 if (IS_ENABLED(CONFIG_NUMA))
3159 printk("Node %d ", zone_to_nid(zone));
3162 void si_meminfo(struct sysinfo *val)
3164 val->totalram = totalram_pages;
3165 val->sharedram = global_page_state(NR_SHMEM);
3166 val->freeram = global_page_state(NR_FREE_PAGES);
3167 val->bufferram = nr_blockdev_pages();
3168 val->totalhigh = totalhigh_pages;
3169 val->freehigh = nr_free_highpages();
3170 val->mem_unit = PAGE_SIZE;
3173 EXPORT_SYMBOL(si_meminfo);
3176 void si_meminfo_node(struct sysinfo *val, int nid)
3178 int zone_type; /* needs to be signed */
3179 unsigned long managed_pages = 0;
3180 pg_data_t *pgdat = NODE_DATA(nid);
3182 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3183 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3184 val->totalram = managed_pages;
3185 val->sharedram = node_page_state(nid, NR_SHMEM);
3186 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3187 #ifdef CONFIG_HIGHMEM
3188 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3189 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3195 val->mem_unit = PAGE_SIZE;
3200 * Determine whether the node should be displayed or not, depending on whether
3201 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3203 bool skip_free_areas_node(unsigned int flags, int nid)
3206 unsigned int cpuset_mems_cookie;
3208 if (!(flags & SHOW_MEM_FILTER_NODES))
3212 cpuset_mems_cookie = read_mems_allowed_begin();
3213 ret = !node_isset(nid, cpuset_current_mems_allowed);
3214 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3219 #define K(x) ((x) << (PAGE_SHIFT-10))
3221 static void show_migration_types(unsigned char type)
3223 static const char types[MIGRATE_TYPES] = {
3224 [MIGRATE_UNMOVABLE] = 'U',
3225 [MIGRATE_RECLAIMABLE] = 'E',
3226 [MIGRATE_MOVABLE] = 'M',
3227 [MIGRATE_RESERVE] = 'R',
3229 [MIGRATE_CMA] = 'C',
3231 #ifdef CONFIG_MEMORY_ISOLATION
3232 [MIGRATE_ISOLATE] = 'I',
3235 char tmp[MIGRATE_TYPES + 1];
3239 for (i = 0; i < MIGRATE_TYPES; i++) {
3240 if (type & (1 << i))
3245 printk("(%s) ", tmp);
3249 * Show free area list (used inside shift_scroll-lock stuff)
3250 * We also calculate the percentage fragmentation. We do this by counting the
3251 * memory on each free list with the exception of the first item on the list.
3252 * Suppresses nodes that are not allowed by current's cpuset if
3253 * SHOW_MEM_FILTER_NODES is passed.
3255 void show_free_areas(unsigned int filter)
3260 for_each_populated_zone(zone) {
3261 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3264 printk("%s per-cpu:\n", zone->name);
3266 for_each_online_cpu(cpu) {
3267 struct per_cpu_pageset *pageset;
3269 pageset = per_cpu_ptr(zone->pageset, cpu);
3271 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3272 cpu, pageset->pcp.high,
3273 pageset->pcp.batch, pageset->pcp.count);
3277 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3278 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3280 " dirty:%lu writeback:%lu unstable:%lu\n"
3281 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3282 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3284 global_page_state(NR_ACTIVE_ANON),
3285 global_page_state(NR_INACTIVE_ANON),
3286 global_page_state(NR_ISOLATED_ANON),
3287 global_page_state(NR_ACTIVE_FILE),
3288 global_page_state(NR_INACTIVE_FILE),
3289 global_page_state(NR_ISOLATED_FILE),
3290 global_page_state(NR_UNEVICTABLE),
3291 global_page_state(NR_FILE_DIRTY),
3292 global_page_state(NR_WRITEBACK),
3293 global_page_state(NR_UNSTABLE_NFS),
3294 global_page_state(NR_FREE_PAGES),
3295 global_page_state(NR_SLAB_RECLAIMABLE),
3296 global_page_state(NR_SLAB_UNRECLAIMABLE),
3297 global_page_state(NR_FILE_MAPPED),
3298 global_page_state(NR_SHMEM),
3299 global_page_state(NR_PAGETABLE),
3300 global_page_state(NR_BOUNCE),
3301 global_page_state(NR_FREE_CMA_PAGES));
3303 for_each_populated_zone(zone) {
3306 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3314 " active_anon:%lukB"
3315 " inactive_anon:%lukB"
3316 " active_file:%lukB"
3317 " inactive_file:%lukB"
3318 " unevictable:%lukB"
3319 " isolated(anon):%lukB"
3320 " isolated(file):%lukB"
3328 " slab_reclaimable:%lukB"
3329 " slab_unreclaimable:%lukB"
3330 " kernel_stack:%lukB"
3335 " writeback_tmp:%lukB"
3336 " pages_scanned:%lu"
3337 " all_unreclaimable? %s"
3340 K(zone_page_state(zone, NR_FREE_PAGES)),
3341 K(min_wmark_pages(zone)),
3342 K(low_wmark_pages(zone)),
3343 K(high_wmark_pages(zone)),
3344 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3345 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3346 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3347 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3348 K(zone_page_state(zone, NR_UNEVICTABLE)),
3349 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3350 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3351 K(zone->present_pages),
3352 K(zone->managed_pages),
3353 K(zone_page_state(zone, NR_MLOCK)),
3354 K(zone_page_state(zone, NR_FILE_DIRTY)),
3355 K(zone_page_state(zone, NR_WRITEBACK)),
3356 K(zone_page_state(zone, NR_FILE_MAPPED)),
3357 K(zone_page_state(zone, NR_SHMEM)),
3358 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3359 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3360 zone_page_state(zone, NR_KERNEL_STACK) *
3362 K(zone_page_state(zone, NR_PAGETABLE)),
3363 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3364 K(zone_page_state(zone, NR_BOUNCE)),
3365 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3366 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3367 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3368 (!zone_reclaimable(zone) ? "yes" : "no")
3370 printk("lowmem_reserve[]:");
3371 for (i = 0; i < MAX_NR_ZONES; i++)
3372 printk(" %ld", zone->lowmem_reserve[i]);
3376 for_each_populated_zone(zone) {
3377 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3378 unsigned char types[MAX_ORDER];
3380 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3383 printk("%s: ", zone->name);
3385 spin_lock_irqsave(&zone->lock, flags);
3386 for (order = 0; order < MAX_ORDER; order++) {
3387 struct free_area *area = &zone->free_area[order];
3390 nr[order] = area->nr_free;
3391 total += nr[order] << order;
3394 for (type = 0; type < MIGRATE_TYPES; type++) {
3395 if (!list_empty(&area->free_list[type]))
3396 types[order] |= 1 << type;
3399 spin_unlock_irqrestore(&zone->lock, flags);
3400 for (order = 0; order < MAX_ORDER; order++) {
3401 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3403 show_migration_types(types[order]);
3405 printk("= %lukB\n", K(total));
3408 hugetlb_show_meminfo();
3410 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3412 show_swap_cache_info();
3415 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3417 zoneref->zone = zone;
3418 zoneref->zone_idx = zone_idx(zone);
3422 * Builds allocation fallback zone lists.
3424 * Add all populated zones of a node to the zonelist.
3426 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3430 enum zone_type zone_type = MAX_NR_ZONES;
3434 zone = pgdat->node_zones + zone_type;
3435 if (populated_zone(zone)) {
3436 zoneref_set_zone(zone,
3437 &zonelist->_zonerefs[nr_zones++]);
3438 check_highest_zone(zone_type);
3440 } while (zone_type);
3448 * 0 = automatic detection of better ordering.
3449 * 1 = order by ([node] distance, -zonetype)
3450 * 2 = order by (-zonetype, [node] distance)
3452 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3453 * the same zonelist. So only NUMA can configure this param.
3455 #define ZONELIST_ORDER_DEFAULT 0
3456 #define ZONELIST_ORDER_NODE 1
3457 #define ZONELIST_ORDER_ZONE 2
3459 /* zonelist order in the kernel.
3460 * set_zonelist_order() will set this to NODE or ZONE.
3462 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3463 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3467 /* The value user specified ....changed by config */
3468 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3469 /* string for sysctl */
3470 #define NUMA_ZONELIST_ORDER_LEN 16
3471 char numa_zonelist_order[16] = "default";
3474 * interface for configure zonelist ordering.
3475 * command line option "numa_zonelist_order"
3476 * = "[dD]efault - default, automatic configuration.
3477 * = "[nN]ode - order by node locality, then by zone within node
3478 * = "[zZ]one - order by zone, then by locality within zone
3481 static int __parse_numa_zonelist_order(char *s)
3483 if (*s == 'd' || *s == 'D') {
3484 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3485 } else if (*s == 'n' || *s == 'N') {
3486 user_zonelist_order = ZONELIST_ORDER_NODE;
3487 } else if (*s == 'z' || *s == 'Z') {
3488 user_zonelist_order = ZONELIST_ORDER_ZONE;
3491 "Ignoring invalid numa_zonelist_order value: "
3498 static __init int setup_numa_zonelist_order(char *s)
3505 ret = __parse_numa_zonelist_order(s);
3507 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3511 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3514 * sysctl handler for numa_zonelist_order
3516 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3517 void __user *buffer, size_t *length,
3520 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3522 static DEFINE_MUTEX(zl_order_mutex);
3524 mutex_lock(&zl_order_mutex);
3526 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3530 strcpy(saved_string, (char *)table->data);
3532 ret = proc_dostring(table, write, buffer, length, ppos);
3536 int oldval = user_zonelist_order;
3538 ret = __parse_numa_zonelist_order((char *)table->data);
3541 * bogus value. restore saved string
3543 strncpy((char *)table->data, saved_string,
3544 NUMA_ZONELIST_ORDER_LEN);
3545 user_zonelist_order = oldval;
3546 } else if (oldval != user_zonelist_order) {
3547 mutex_lock(&zonelists_mutex);
3548 build_all_zonelists(NULL, NULL);
3549 mutex_unlock(&zonelists_mutex);
3553 mutex_unlock(&zl_order_mutex);
3558 #define MAX_NODE_LOAD (nr_online_nodes)
3559 static int node_load[MAX_NUMNODES];
3562 * find_next_best_node - find the next node that should appear in a given node's fallback list
3563 * @node: node whose fallback list we're appending
3564 * @used_node_mask: nodemask_t of already used nodes
3566 * We use a number of factors to determine which is the next node that should
3567 * appear on a given node's fallback list. The node should not have appeared
3568 * already in @node's fallback list, and it should be the next closest node
3569 * according to the distance array (which contains arbitrary distance values
3570 * from each node to each node in the system), and should also prefer nodes
3571 * with no CPUs, since presumably they'll have very little allocation pressure
3572 * on them otherwise.
3573 * It returns -1 if no node is found.
3575 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3578 int min_val = INT_MAX;
3579 int best_node = NUMA_NO_NODE;
3580 const struct cpumask *tmp = cpumask_of_node(0);
3582 /* Use the local node if we haven't already */
3583 if (!node_isset(node, *used_node_mask)) {
3584 node_set(node, *used_node_mask);
3588 for_each_node_state(n, N_MEMORY) {
3590 /* Don't want a node to appear more than once */
3591 if (node_isset(n, *used_node_mask))
3594 /* Use the distance array to find the distance */
3595 val = node_distance(node, n);
3597 /* Penalize nodes under us ("prefer the next node") */
3600 /* Give preference to headless and unused nodes */
3601 tmp = cpumask_of_node(n);
3602 if (!cpumask_empty(tmp))
3603 val += PENALTY_FOR_NODE_WITH_CPUS;
3605 /* Slight preference for less loaded node */
3606 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3607 val += node_load[n];
3609 if (val < min_val) {
3616 node_set(best_node, *used_node_mask);
3623 * Build zonelists ordered by node and zones within node.
3624 * This results in maximum locality--normal zone overflows into local
3625 * DMA zone, if any--but risks exhausting DMA zone.
3627 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3630 struct zonelist *zonelist;
3632 zonelist = &pgdat->node_zonelists[0];
3633 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3635 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3636 zonelist->_zonerefs[j].zone = NULL;
3637 zonelist->_zonerefs[j].zone_idx = 0;
3641 * Build gfp_thisnode zonelists
3643 static void build_thisnode_zonelists(pg_data_t *pgdat)
3646 struct zonelist *zonelist;
3648 zonelist = &pgdat->node_zonelists[1];
3649 j = build_zonelists_node(pgdat, zonelist, 0);
3650 zonelist->_zonerefs[j].zone = NULL;
3651 zonelist->_zonerefs[j].zone_idx = 0;
3655 * Build zonelists ordered by zone and nodes within zones.
3656 * This results in conserving DMA zone[s] until all Normal memory is
3657 * exhausted, but results in overflowing to remote node while memory
3658 * may still exist in local DMA zone.
3660 static int node_order[MAX_NUMNODES];
3662 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3665 int zone_type; /* needs to be signed */
3667 struct zonelist *zonelist;
3669 zonelist = &pgdat->node_zonelists[0];
3671 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3672 for (j = 0; j < nr_nodes; j++) {
3673 node = node_order[j];
3674 z = &NODE_DATA(node)->node_zones[zone_type];
3675 if (populated_zone(z)) {
3677 &zonelist->_zonerefs[pos++]);
3678 check_highest_zone(zone_type);
3682 zonelist->_zonerefs[pos].zone = NULL;
3683 zonelist->_zonerefs[pos].zone_idx = 0;
3686 #if defined(CONFIG_64BIT)
3688 * Devices that require DMA32/DMA are relatively rare and do not justify a
3689 * penalty to every machine in case the specialised case applies. Default
3690 * to Node-ordering on 64-bit NUMA machines
3692 static int default_zonelist_order(void)
3694 return ZONELIST_ORDER_NODE;
3698 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3699 * by the kernel. If processes running on node 0 deplete the low memory zone
3700 * then reclaim will occur more frequency increasing stalls and potentially
3701 * be easier to OOM if a large percentage of the zone is under writeback or
3702 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3703 * Hence, default to zone ordering on 32-bit.
3705 static int default_zonelist_order(void)
3707 return ZONELIST_ORDER_ZONE;
3709 #endif /* CONFIG_64BIT */
3711 static void set_zonelist_order(void)
3713 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3714 current_zonelist_order = default_zonelist_order();
3716 current_zonelist_order = user_zonelist_order;
3719 static void build_zonelists(pg_data_t *pgdat)
3723 nodemask_t used_mask;
3724 int local_node, prev_node;
3725 struct zonelist *zonelist;
3726 int order = current_zonelist_order;
3728 /* initialize zonelists */
3729 for (i = 0; i < MAX_ZONELISTS; i++) {
3730 zonelist = pgdat->node_zonelists + i;
3731 zonelist->_zonerefs[0].zone = NULL;
3732 zonelist->_zonerefs[0].zone_idx = 0;
3735 /* NUMA-aware ordering of nodes */
3736 local_node = pgdat->node_id;
3737 load = nr_online_nodes;
3738 prev_node = local_node;
3739 nodes_clear(used_mask);
3741 memset(node_order, 0, sizeof(node_order));
3744 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3746 * We don't want to pressure a particular node.
3747 * So adding penalty to the first node in same
3748 * distance group to make it round-robin.
3750 if (node_distance(local_node, node) !=
3751 node_distance(local_node, prev_node))
3752 node_load[node] = load;
3756 if (order == ZONELIST_ORDER_NODE)
3757 build_zonelists_in_node_order(pgdat, node);
3759 node_order[j++] = node; /* remember order */
3762 if (order == ZONELIST_ORDER_ZONE) {
3763 /* calculate node order -- i.e., DMA last! */
3764 build_zonelists_in_zone_order(pgdat, j);
3767 build_thisnode_zonelists(pgdat);
3770 /* Construct the zonelist performance cache - see further mmzone.h */
3771 static void build_zonelist_cache(pg_data_t *pgdat)
3773 struct zonelist *zonelist;
3774 struct zonelist_cache *zlc;
3777 zonelist = &pgdat->node_zonelists[0];
3778 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3779 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3780 for (z = zonelist->_zonerefs; z->zone; z++)
3781 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3784 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3786 * Return node id of node used for "local" allocations.
3787 * I.e., first node id of first zone in arg node's generic zonelist.
3788 * Used for initializing percpu 'numa_mem', which is used primarily
3789 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3791 int local_memory_node(int node)
3795 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3796 gfp_zone(GFP_KERNEL),
3803 #else /* CONFIG_NUMA */
3805 static void set_zonelist_order(void)
3807 current_zonelist_order = ZONELIST_ORDER_ZONE;
3810 static void build_zonelists(pg_data_t *pgdat)
3812 int node, local_node;
3814 struct zonelist *zonelist;
3816 local_node = pgdat->node_id;
3818 zonelist = &pgdat->node_zonelists[0];
3819 j = build_zonelists_node(pgdat, zonelist, 0);
3822 * Now we build the zonelist so that it contains the zones
3823 * of all the other nodes.
3824 * We don't want to pressure a particular node, so when
3825 * building the zones for node N, we make sure that the
3826 * zones coming right after the local ones are those from
3827 * node N+1 (modulo N)
3829 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3830 if (!node_online(node))
3832 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3834 for (node = 0; node < local_node; node++) {
3835 if (!node_online(node))
3837 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3840 zonelist->_zonerefs[j].zone = NULL;
3841 zonelist->_zonerefs[j].zone_idx = 0;
3844 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3845 static void build_zonelist_cache(pg_data_t *pgdat)
3847 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3850 #endif /* CONFIG_NUMA */
3853 * Boot pageset table. One per cpu which is going to be used for all
3854 * zones and all nodes. The parameters will be set in such a way
3855 * that an item put on a list will immediately be handed over to
3856 * the buddy list. This is safe since pageset manipulation is done
3857 * with interrupts disabled.
3859 * The boot_pagesets must be kept even after bootup is complete for
3860 * unused processors and/or zones. They do play a role for bootstrapping
3861 * hotplugged processors.
3863 * zoneinfo_show() and maybe other functions do
3864 * not check if the processor is online before following the pageset pointer.
3865 * Other parts of the kernel may not check if the zone is available.
3867 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3868 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3869 static void setup_zone_pageset(struct zone *zone);
3872 * Global mutex to protect against size modification of zonelists
3873 * as well as to serialize pageset setup for the new populated zone.
3875 DEFINE_MUTEX(zonelists_mutex);
3877 /* return values int ....just for stop_machine() */
3878 static int __build_all_zonelists(void *data)
3882 pg_data_t *self = data;
3885 memset(node_load, 0, sizeof(node_load));
3888 if (self && !node_online(self->node_id)) {
3889 build_zonelists(self);
3890 build_zonelist_cache(self);
3893 for_each_online_node(nid) {
3894 pg_data_t *pgdat = NODE_DATA(nid);
3896 build_zonelists(pgdat);
3897 build_zonelist_cache(pgdat);
3901 * Initialize the boot_pagesets that are going to be used
3902 * for bootstrapping processors. The real pagesets for
3903 * each zone will be allocated later when the per cpu
3904 * allocator is available.
3906 * boot_pagesets are used also for bootstrapping offline
3907 * cpus if the system is already booted because the pagesets
3908 * are needed to initialize allocators on a specific cpu too.
3909 * F.e. the percpu allocator needs the page allocator which
3910 * needs the percpu allocator in order to allocate its pagesets
3911 * (a chicken-egg dilemma).
3913 for_each_possible_cpu(cpu) {
3914 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3916 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3918 * We now know the "local memory node" for each node--
3919 * i.e., the node of the first zone in the generic zonelist.
3920 * Set up numa_mem percpu variable for on-line cpus. During
3921 * boot, only the boot cpu should be on-line; we'll init the
3922 * secondary cpus' numa_mem as they come on-line. During
3923 * node/memory hotplug, we'll fixup all on-line cpus.
3925 if (cpu_online(cpu))
3926 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3934 * Called with zonelists_mutex held always
3935 * unless system_state == SYSTEM_BOOTING.
3937 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3939 set_zonelist_order();
3941 if (system_state == SYSTEM_BOOTING) {
3942 __build_all_zonelists(NULL);
3943 mminit_verify_zonelist();
3944 cpuset_init_current_mems_allowed();
3946 #ifdef CONFIG_MEMORY_HOTPLUG
3948 setup_zone_pageset(zone);
3950 /* we have to stop all cpus to guarantee there is no user
3952 stop_machine(__build_all_zonelists, pgdat, NULL);
3953 /* cpuset refresh routine should be here */
3955 vm_total_pages = nr_free_pagecache_pages();
3957 * Disable grouping by mobility if the number of pages in the
3958 * system is too low to allow the mechanism to work. It would be
3959 * more accurate, but expensive to check per-zone. This check is
3960 * made on memory-hotadd so a system can start with mobility
3961 * disabled and enable it later
3963 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3964 page_group_by_mobility_disabled = 1;
3966 page_group_by_mobility_disabled = 0;
3968 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3969 "Total pages: %ld\n",
3971 zonelist_order_name[current_zonelist_order],
3972 page_group_by_mobility_disabled ? "off" : "on",
3975 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
3980 * Helper functions to size the waitqueue hash table.
3981 * Essentially these want to choose hash table sizes sufficiently
3982 * large so that collisions trying to wait on pages are rare.
3983 * But in fact, the number of active page waitqueues on typical
3984 * systems is ridiculously low, less than 200. So this is even
3985 * conservative, even though it seems large.
3987 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3988 * waitqueues, i.e. the size of the waitq table given the number of pages.
3990 #define PAGES_PER_WAITQUEUE 256
3992 #ifndef CONFIG_MEMORY_HOTPLUG
3993 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3995 unsigned long size = 1;
3997 pages /= PAGES_PER_WAITQUEUE;
3999 while (size < pages)
4003 * Once we have dozens or even hundreds of threads sleeping
4004 * on IO we've got bigger problems than wait queue collision.
4005 * Limit the size of the wait table to a reasonable size.
4007 size = min(size, 4096UL);
4009 return max(size, 4UL);
4013 * A zone's size might be changed by hot-add, so it is not possible to determine
4014 * a suitable size for its wait_table. So we use the maximum size now.
4016 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4018 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4019 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4020 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4022 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4023 * or more by the traditional way. (See above). It equals:
4025 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4026 * ia64(16K page size) : = ( 8G + 4M)byte.
4027 * powerpc (64K page size) : = (32G +16M)byte.
4029 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4036 * This is an integer logarithm so that shifts can be used later
4037 * to extract the more random high bits from the multiplicative
4038 * hash function before the remainder is taken.
4040 static inline unsigned long wait_table_bits(unsigned long size)
4046 * Check if a pageblock contains reserved pages
4048 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4052 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4053 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4060 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4061 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4062 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4063 * higher will lead to a bigger reserve which will get freed as contiguous
4064 * blocks as reclaim kicks in
4066 static void setup_zone_migrate_reserve(struct zone *zone)
4068 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4070 unsigned long block_migratetype;
4075 * Get the start pfn, end pfn and the number of blocks to reserve
4076 * We have to be careful to be aligned to pageblock_nr_pages to
4077 * make sure that we always check pfn_valid for the first page in
4080 start_pfn = zone->zone_start_pfn;
4081 end_pfn = zone_end_pfn(zone);
4082 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4083 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4087 * Reserve blocks are generally in place to help high-order atomic
4088 * allocations that are short-lived. A min_free_kbytes value that
4089 * would result in more than 2 reserve blocks for atomic allocations
4090 * is assumed to be in place to help anti-fragmentation for the
4091 * future allocation of hugepages at runtime.
4093 reserve = min(2, reserve);
4094 old_reserve = zone->nr_migrate_reserve_block;
4096 /* When memory hot-add, we almost always need to do nothing */
4097 if (reserve == old_reserve)
4099 zone->nr_migrate_reserve_block = reserve;
4101 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4102 if (!pfn_valid(pfn))
4104 page = pfn_to_page(pfn);
4106 /* Watch out for overlapping nodes */
4107 if (page_to_nid(page) != zone_to_nid(zone))
4110 block_migratetype = get_pageblock_migratetype(page);
4112 /* Only test what is necessary when the reserves are not met */
4115 * Blocks with reserved pages will never free, skip
4118 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4119 if (pageblock_is_reserved(pfn, block_end_pfn))
4122 /* If this block is reserved, account for it */
4123 if (block_migratetype == MIGRATE_RESERVE) {
4128 /* Suitable for reserving if this block is movable */
4129 if (block_migratetype == MIGRATE_MOVABLE) {
4130 set_pageblock_migratetype(page,
4132 move_freepages_block(zone, page,
4137 } else if (!old_reserve) {
4139 * At boot time we don't need to scan the whole zone
4140 * for turning off MIGRATE_RESERVE.
4146 * If the reserve is met and this is a previous reserved block,
4149 if (block_migratetype == MIGRATE_RESERVE) {
4150 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4151 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4157 * Initially all pages are reserved - free ones are freed
4158 * up by free_all_bootmem() once the early boot process is
4159 * done. Non-atomic initialization, single-pass.
4161 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4162 unsigned long start_pfn, enum memmap_context context)
4165 unsigned long end_pfn = start_pfn + size;
4169 if (highest_memmap_pfn < end_pfn - 1)
4170 highest_memmap_pfn = end_pfn - 1;
4172 z = &NODE_DATA(nid)->node_zones[zone];
4173 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4175 * There can be holes in boot-time mem_map[]s
4176 * handed to this function. They do not
4177 * exist on hotplugged memory.
4179 if (context == MEMMAP_EARLY) {
4180 if (!early_pfn_valid(pfn))
4182 if (!early_pfn_in_nid(pfn, nid))
4185 page = pfn_to_page(pfn);
4186 set_page_links(page, zone, nid, pfn);
4187 mminit_verify_page_links(page, zone, nid, pfn);
4188 init_page_count(page);
4189 page_mapcount_reset(page);
4190 page_cpupid_reset_last(page);
4191 SetPageReserved(page);
4193 * Mark the block movable so that blocks are reserved for
4194 * movable at startup. This will force kernel allocations
4195 * to reserve their blocks rather than leaking throughout
4196 * the address space during boot when many long-lived
4197 * kernel allocations are made. Later some blocks near
4198 * the start are marked MIGRATE_RESERVE by
4199 * setup_zone_migrate_reserve()
4201 * bitmap is created for zone's valid pfn range. but memmap
4202 * can be created for invalid pages (for alignment)
4203 * check here not to call set_pageblock_migratetype() against
4206 if ((z->zone_start_pfn <= pfn)
4207 && (pfn < zone_end_pfn(z))
4208 && !(pfn & (pageblock_nr_pages - 1)))
4209 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4211 INIT_LIST_HEAD(&page->lru);
4212 #ifdef WANT_PAGE_VIRTUAL
4213 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4214 if (!is_highmem_idx(zone))
4215 set_page_address(page, __va(pfn << PAGE_SHIFT));
4220 static void __meminit zone_init_free_lists(struct zone *zone)
4222 unsigned int order, t;
4223 for_each_migratetype_order(order, t) {
4224 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4225 zone->free_area[order].nr_free = 0;
4229 #ifndef __HAVE_ARCH_MEMMAP_INIT
4230 #define memmap_init(size, nid, zone, start_pfn) \
4231 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4234 static int zone_batchsize(struct zone *zone)
4240 * The per-cpu-pages pools are set to around 1000th of the
4241 * size of the zone. But no more than 1/2 of a meg.
4243 * OK, so we don't know how big the cache is. So guess.
4245 batch = zone->managed_pages / 1024;
4246 if (batch * PAGE_SIZE > 512 * 1024)
4247 batch = (512 * 1024) / PAGE_SIZE;
4248 batch /= 4; /* We effectively *= 4 below */
4253 * Clamp the batch to a 2^n - 1 value. Having a power
4254 * of 2 value was found to be more likely to have
4255 * suboptimal cache aliasing properties in some cases.
4257 * For example if 2 tasks are alternately allocating
4258 * batches of pages, one task can end up with a lot
4259 * of pages of one half of the possible page colors
4260 * and the other with pages of the other colors.
4262 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4267 /* The deferral and batching of frees should be suppressed under NOMMU
4270 * The problem is that NOMMU needs to be able to allocate large chunks
4271 * of contiguous memory as there's no hardware page translation to
4272 * assemble apparent contiguous memory from discontiguous pages.
4274 * Queueing large contiguous runs of pages for batching, however,
4275 * causes the pages to actually be freed in smaller chunks. As there
4276 * can be a significant delay between the individual batches being
4277 * recycled, this leads to the once large chunks of space being
4278 * fragmented and becoming unavailable for high-order allocations.
4285 * pcp->high and pcp->batch values are related and dependent on one another:
4286 * ->batch must never be higher then ->high.
4287 * The following function updates them in a safe manner without read side
4290 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4291 * those fields changing asynchronously (acording the the above rule).
4293 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4294 * outside of boot time (or some other assurance that no concurrent updaters
4297 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4298 unsigned long batch)
4300 /* start with a fail safe value for batch */
4304 /* Update high, then batch, in order */
4311 /* a companion to pageset_set_high() */
4312 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4314 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4317 static void pageset_init(struct per_cpu_pageset *p)
4319 struct per_cpu_pages *pcp;
4322 memset(p, 0, sizeof(*p));
4326 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4327 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4330 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4333 pageset_set_batch(p, batch);
4337 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4338 * to the value high for the pageset p.
4340 static void pageset_set_high(struct per_cpu_pageset *p,
4343 unsigned long batch = max(1UL, high / 4);
4344 if ((high / 4) > (PAGE_SHIFT * 8))
4345 batch = PAGE_SHIFT * 8;
4347 pageset_update(&p->pcp, high, batch);
4350 static void pageset_set_high_and_batch(struct zone *zone,
4351 struct per_cpu_pageset *pcp)
4353 if (percpu_pagelist_fraction)
4354 pageset_set_high(pcp,
4355 (zone->managed_pages /
4356 percpu_pagelist_fraction));
4358 pageset_set_batch(pcp, zone_batchsize(zone));
4361 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4363 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4366 pageset_set_high_and_batch(zone, pcp);
4369 static void __meminit setup_zone_pageset(struct zone *zone)
4372 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4373 for_each_possible_cpu(cpu)
4374 zone_pageset_init(zone, cpu);
4378 * Allocate per cpu pagesets and initialize them.
4379 * Before this call only boot pagesets were available.
4381 void __init setup_per_cpu_pageset(void)
4385 for_each_populated_zone(zone)
4386 setup_zone_pageset(zone);
4389 static noinline __init_refok
4390 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4396 * The per-page waitqueue mechanism uses hashed waitqueues
4399 zone->wait_table_hash_nr_entries =
4400 wait_table_hash_nr_entries(zone_size_pages);
4401 zone->wait_table_bits =
4402 wait_table_bits(zone->wait_table_hash_nr_entries);
4403 alloc_size = zone->wait_table_hash_nr_entries
4404 * sizeof(wait_queue_head_t);
4406 if (!slab_is_available()) {
4407 zone->wait_table = (wait_queue_head_t *)
4408 memblock_virt_alloc_node_nopanic(
4409 alloc_size, zone->zone_pgdat->node_id);
4412 * This case means that a zone whose size was 0 gets new memory
4413 * via memory hot-add.
4414 * But it may be the case that a new node was hot-added. In
4415 * this case vmalloc() will not be able to use this new node's
4416 * memory - this wait_table must be initialized to use this new
4417 * node itself as well.
4418 * To use this new node's memory, further consideration will be
4421 zone->wait_table = vmalloc(alloc_size);
4423 if (!zone->wait_table)
4426 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4427 init_waitqueue_head(zone->wait_table + i);
4432 static __meminit void zone_pcp_init(struct zone *zone)
4435 * per cpu subsystem is not up at this point. The following code
4436 * relies on the ability of the linker to provide the
4437 * offset of a (static) per cpu variable into the per cpu area.
4439 zone->pageset = &boot_pageset;
4441 if (populated_zone(zone))
4442 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4443 zone->name, zone->present_pages,
4444 zone_batchsize(zone));
4447 int __meminit init_currently_empty_zone(struct zone *zone,
4448 unsigned long zone_start_pfn,
4450 enum memmap_context context)
4452 struct pglist_data *pgdat = zone->zone_pgdat;
4454 ret = zone_wait_table_init(zone, size);
4457 pgdat->nr_zones = zone_idx(zone) + 1;
4459 zone->zone_start_pfn = zone_start_pfn;
4461 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4462 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4464 (unsigned long)zone_idx(zone),
4465 zone_start_pfn, (zone_start_pfn + size));
4467 zone_init_free_lists(zone);
4472 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4473 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4475 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4477 int __meminit __early_pfn_to_nid(unsigned long pfn)
4479 unsigned long start_pfn, end_pfn;
4482 * NOTE: The following SMP-unsafe globals are only used early in boot
4483 * when the kernel is running single-threaded.
4485 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4486 static int __meminitdata last_nid;
4488 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4491 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4493 last_start_pfn = start_pfn;
4494 last_end_pfn = end_pfn;
4500 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4502 int __meminit early_pfn_to_nid(unsigned long pfn)
4506 nid = __early_pfn_to_nid(pfn);
4509 /* just returns 0 */
4513 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4514 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4518 nid = __early_pfn_to_nid(pfn);
4519 if (nid >= 0 && nid != node)
4526 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4527 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4528 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4530 * If an architecture guarantees that all ranges registered contain no holes
4531 * and may be freed, this this function may be used instead of calling
4532 * memblock_free_early_nid() manually.
4534 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4536 unsigned long start_pfn, end_pfn;
4539 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4540 start_pfn = min(start_pfn, max_low_pfn);
4541 end_pfn = min(end_pfn, max_low_pfn);
4543 if (start_pfn < end_pfn)
4544 memblock_free_early_nid(PFN_PHYS(start_pfn),
4545 (end_pfn - start_pfn) << PAGE_SHIFT,
4551 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4552 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4554 * If an architecture guarantees that all ranges registered contain no holes and may
4555 * be freed, this function may be used instead of calling memory_present() manually.
4557 void __init sparse_memory_present_with_active_regions(int nid)
4559 unsigned long start_pfn, end_pfn;
4562 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4563 memory_present(this_nid, start_pfn, end_pfn);
4567 * get_pfn_range_for_nid - Return the start and end page frames for a node
4568 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4569 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4570 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4572 * It returns the start and end page frame of a node based on information
4573 * provided by memblock_set_node(). If called for a node
4574 * with no available memory, a warning is printed and the start and end
4577 void __meminit get_pfn_range_for_nid(unsigned int nid,
4578 unsigned long *start_pfn, unsigned long *end_pfn)
4580 unsigned long this_start_pfn, this_end_pfn;
4586 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4587 *start_pfn = min(*start_pfn, this_start_pfn);
4588 *end_pfn = max(*end_pfn, this_end_pfn);
4591 if (*start_pfn == -1UL)
4596 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4597 * assumption is made that zones within a node are ordered in monotonic
4598 * increasing memory addresses so that the "highest" populated zone is used
4600 static void __init find_usable_zone_for_movable(void)
4603 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4604 if (zone_index == ZONE_MOVABLE)
4607 if (arch_zone_highest_possible_pfn[zone_index] >
4608 arch_zone_lowest_possible_pfn[zone_index])
4612 VM_BUG_ON(zone_index == -1);
4613 movable_zone = zone_index;
4617 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4618 * because it is sized independent of architecture. Unlike the other zones,
4619 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4620 * in each node depending on the size of each node and how evenly kernelcore
4621 * is distributed. This helper function adjusts the zone ranges
4622 * provided by the architecture for a given node by using the end of the
4623 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4624 * zones within a node are in order of monotonic increases memory addresses
4626 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4627 unsigned long zone_type,
4628 unsigned long node_start_pfn,
4629 unsigned long node_end_pfn,
4630 unsigned long *zone_start_pfn,
4631 unsigned long *zone_end_pfn)
4633 /* Only adjust if ZONE_MOVABLE is on this node */
4634 if (zone_movable_pfn[nid]) {
4635 /* Size ZONE_MOVABLE */
4636 if (zone_type == ZONE_MOVABLE) {
4637 *zone_start_pfn = zone_movable_pfn[nid];
4638 *zone_end_pfn = min(node_end_pfn,
4639 arch_zone_highest_possible_pfn[movable_zone]);
4641 /* Adjust for ZONE_MOVABLE starting within this range */
4642 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4643 *zone_end_pfn > zone_movable_pfn[nid]) {
4644 *zone_end_pfn = zone_movable_pfn[nid];
4646 /* Check if this whole range is within ZONE_MOVABLE */
4647 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4648 *zone_start_pfn = *zone_end_pfn;
4653 * Return the number of pages a zone spans in a node, including holes
4654 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4656 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4657 unsigned long zone_type,
4658 unsigned long node_start_pfn,
4659 unsigned long node_end_pfn,
4660 unsigned long *ignored)
4662 unsigned long zone_start_pfn, zone_end_pfn;
4664 /* Get the start and end of the zone */
4665 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4666 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4667 adjust_zone_range_for_zone_movable(nid, zone_type,
4668 node_start_pfn, node_end_pfn,
4669 &zone_start_pfn, &zone_end_pfn);
4671 /* Check that this node has pages within the zone's required range */
4672 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4675 /* Move the zone boundaries inside the node if necessary */
4676 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4677 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4679 /* Return the spanned pages */
4680 return zone_end_pfn - zone_start_pfn;
4684 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4685 * then all holes in the requested range will be accounted for.
4687 unsigned long __meminit __absent_pages_in_range(int nid,
4688 unsigned long range_start_pfn,
4689 unsigned long range_end_pfn)
4691 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4692 unsigned long start_pfn, end_pfn;
4695 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4696 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4697 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4698 nr_absent -= end_pfn - start_pfn;
4704 * absent_pages_in_range - Return number of page frames in holes within a range
4705 * @start_pfn: The start PFN to start searching for holes
4706 * @end_pfn: The end PFN to stop searching for holes
4708 * It returns the number of pages frames in memory holes within a range.
4710 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4711 unsigned long end_pfn)
4713 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4716 /* Return the number of page frames in holes in a zone on a node */
4717 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4718 unsigned long zone_type,
4719 unsigned long node_start_pfn,
4720 unsigned long node_end_pfn,
4721 unsigned long *ignored)
4723 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4724 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4725 unsigned long zone_start_pfn, zone_end_pfn;
4727 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4728 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4730 adjust_zone_range_for_zone_movable(nid, zone_type,
4731 node_start_pfn, node_end_pfn,
4732 &zone_start_pfn, &zone_end_pfn);
4733 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4736 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4737 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4738 unsigned long zone_type,
4739 unsigned long node_start_pfn,
4740 unsigned long node_end_pfn,
4741 unsigned long *zones_size)
4743 return zones_size[zone_type];
4746 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4747 unsigned long zone_type,
4748 unsigned long node_start_pfn,
4749 unsigned long node_end_pfn,
4750 unsigned long *zholes_size)
4755 return zholes_size[zone_type];
4758 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4760 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4761 unsigned long node_start_pfn,
4762 unsigned long node_end_pfn,
4763 unsigned long *zones_size,
4764 unsigned long *zholes_size)
4766 unsigned long realtotalpages, totalpages = 0;
4769 for (i = 0; i < MAX_NR_ZONES; i++)
4770 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4774 pgdat->node_spanned_pages = totalpages;
4776 realtotalpages = totalpages;
4777 for (i = 0; i < MAX_NR_ZONES; i++)
4779 zone_absent_pages_in_node(pgdat->node_id, i,
4780 node_start_pfn, node_end_pfn,
4782 pgdat->node_present_pages = realtotalpages;
4783 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4787 #ifndef CONFIG_SPARSEMEM
4789 * Calculate the size of the zone->blockflags rounded to an unsigned long
4790 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4791 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4792 * round what is now in bits to nearest long in bits, then return it in
4795 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4797 unsigned long usemapsize;
4799 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4800 usemapsize = roundup(zonesize, pageblock_nr_pages);
4801 usemapsize = usemapsize >> pageblock_order;
4802 usemapsize *= NR_PAGEBLOCK_BITS;
4803 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4805 return usemapsize / 8;
4808 static void __init setup_usemap(struct pglist_data *pgdat,
4810 unsigned long zone_start_pfn,
4811 unsigned long zonesize)
4813 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4814 zone->pageblock_flags = NULL;
4816 zone->pageblock_flags =
4817 memblock_virt_alloc_node_nopanic(usemapsize,
4821 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4822 unsigned long zone_start_pfn, unsigned long zonesize) {}
4823 #endif /* CONFIG_SPARSEMEM */
4825 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4827 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4828 void __paginginit set_pageblock_order(void)
4832 /* Check that pageblock_nr_pages has not already been setup */
4833 if (pageblock_order)
4836 if (HPAGE_SHIFT > PAGE_SHIFT)
4837 order = HUGETLB_PAGE_ORDER;
4839 order = MAX_ORDER - 1;
4842 * Assume the largest contiguous order of interest is a huge page.
4843 * This value may be variable depending on boot parameters on IA64 and
4846 pageblock_order = order;
4848 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4851 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4852 * is unused as pageblock_order is set at compile-time. See
4853 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4856 void __paginginit set_pageblock_order(void)
4860 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4862 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4863 unsigned long present_pages)
4865 unsigned long pages = spanned_pages;
4868 * Provide a more accurate estimation if there are holes within
4869 * the zone and SPARSEMEM is in use. If there are holes within the
4870 * zone, each populated memory region may cost us one or two extra
4871 * memmap pages due to alignment because memmap pages for each
4872 * populated regions may not naturally algined on page boundary.
4873 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4875 if (spanned_pages > present_pages + (present_pages >> 4) &&
4876 IS_ENABLED(CONFIG_SPARSEMEM))
4877 pages = present_pages;
4879 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4883 * Set up the zone data structures:
4884 * - mark all pages reserved
4885 * - mark all memory queues empty
4886 * - clear the memory bitmaps
4888 * NOTE: pgdat should get zeroed by caller.
4890 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4891 unsigned long node_start_pfn, unsigned long node_end_pfn,
4892 unsigned long *zones_size, unsigned long *zholes_size)
4895 int nid = pgdat->node_id;
4896 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4899 pgdat_resize_init(pgdat);
4900 #ifdef CONFIG_NUMA_BALANCING
4901 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4902 pgdat->numabalancing_migrate_nr_pages = 0;
4903 pgdat->numabalancing_migrate_next_window = jiffies;
4905 init_waitqueue_head(&pgdat->kswapd_wait);
4906 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4907 pgdat_page_ext_init(pgdat);
4909 for (j = 0; j < MAX_NR_ZONES; j++) {
4910 struct zone *zone = pgdat->node_zones + j;
4911 unsigned long size, realsize, freesize, memmap_pages;
4913 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4914 node_end_pfn, zones_size);
4915 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4921 * Adjust freesize so that it accounts for how much memory
4922 * is used by this zone for memmap. This affects the watermark
4923 * and per-cpu initialisations
4925 memmap_pages = calc_memmap_size(size, realsize);
4926 if (!is_highmem_idx(j)) {
4927 if (freesize >= memmap_pages) {
4928 freesize -= memmap_pages;
4931 " %s zone: %lu pages used for memmap\n",
4932 zone_names[j], memmap_pages);
4935 " %s zone: %lu pages exceeds freesize %lu\n",
4936 zone_names[j], memmap_pages, freesize);
4939 /* Account for reserved pages */
4940 if (j == 0 && freesize > dma_reserve) {
4941 freesize -= dma_reserve;
4942 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4943 zone_names[0], dma_reserve);
4946 if (!is_highmem_idx(j))
4947 nr_kernel_pages += freesize;
4948 /* Charge for highmem memmap if there are enough kernel pages */
4949 else if (nr_kernel_pages > memmap_pages * 2)
4950 nr_kernel_pages -= memmap_pages;
4951 nr_all_pages += freesize;
4953 zone->spanned_pages = size;
4954 zone->present_pages = realsize;
4956 * Set an approximate value for lowmem here, it will be adjusted
4957 * when the bootmem allocator frees pages into the buddy system.
4958 * And all highmem pages will be managed by the buddy system.
4960 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4963 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4965 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4967 zone->name = zone_names[j];
4968 spin_lock_init(&zone->lock);
4969 spin_lock_init(&zone->lru_lock);
4970 zone_seqlock_init(zone);
4971 zone->zone_pgdat = pgdat;
4972 zone_pcp_init(zone);
4974 /* For bootup, initialized properly in watermark setup */
4975 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4977 lruvec_init(&zone->lruvec);
4981 set_pageblock_order();
4982 setup_usemap(pgdat, zone, zone_start_pfn, size);
4983 ret = init_currently_empty_zone(zone, zone_start_pfn,
4984 size, MEMMAP_EARLY);
4986 memmap_init(size, nid, j, zone_start_pfn);
4987 zone_start_pfn += size;
4991 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4993 /* Skip empty nodes */
4994 if (!pgdat->node_spanned_pages)
4997 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4998 /* ia64 gets its own node_mem_map, before this, without bootmem */
4999 if (!pgdat->node_mem_map) {
5000 unsigned long size, start, end;
5004 * The zone's endpoints aren't required to be MAX_ORDER
5005 * aligned but the node_mem_map endpoints must be in order
5006 * for the buddy allocator to function correctly.
5008 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5009 end = pgdat_end_pfn(pgdat);
5010 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5011 size = (end - start) * sizeof(struct page);
5012 map = alloc_remap(pgdat->node_id, size);
5014 map = memblock_virt_alloc_node_nopanic(size,
5016 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5018 #ifndef CONFIG_NEED_MULTIPLE_NODES
5020 * With no DISCONTIG, the global mem_map is just set as node 0's
5022 if (pgdat == NODE_DATA(0)) {
5023 mem_map = NODE_DATA(0)->node_mem_map;
5024 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5025 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5026 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5027 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5030 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5033 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5034 unsigned long node_start_pfn, unsigned long *zholes_size)
5036 pg_data_t *pgdat = NODE_DATA(nid);
5037 unsigned long start_pfn = 0;
5038 unsigned long end_pfn = 0;
5040 /* pg_data_t should be reset to zero when it's allocated */
5041 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5043 pgdat->node_id = nid;
5044 pgdat->node_start_pfn = node_start_pfn;
5045 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5046 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5047 printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid,
5048 (u64) start_pfn << PAGE_SHIFT, (u64) (end_pfn << PAGE_SHIFT) - 1);
5050 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5051 zones_size, zholes_size);
5053 alloc_node_mem_map(pgdat);
5054 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5055 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5056 nid, (unsigned long)pgdat,
5057 (unsigned long)pgdat->node_mem_map);
5060 free_area_init_core(pgdat, start_pfn, end_pfn,
5061 zones_size, zholes_size);
5064 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5066 #if MAX_NUMNODES > 1
5068 * Figure out the number of possible node ids.
5070 void __init setup_nr_node_ids(void)
5073 unsigned int highest = 0;
5075 for_each_node_mask(node, node_possible_map)
5077 nr_node_ids = highest + 1;
5082 * node_map_pfn_alignment - determine the maximum internode alignment
5084 * This function should be called after node map is populated and sorted.
5085 * It calculates the maximum power of two alignment which can distinguish
5088 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5089 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5090 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5091 * shifted, 1GiB is enough and this function will indicate so.
5093 * This is used to test whether pfn -> nid mapping of the chosen memory
5094 * model has fine enough granularity to avoid incorrect mapping for the
5095 * populated node map.
5097 * Returns the determined alignment in pfn's. 0 if there is no alignment
5098 * requirement (single node).
5100 unsigned long __init node_map_pfn_alignment(void)
5102 unsigned long accl_mask = 0, last_end = 0;
5103 unsigned long start, end, mask;
5107 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5108 if (!start || last_nid < 0 || last_nid == nid) {
5115 * Start with a mask granular enough to pin-point to the
5116 * start pfn and tick off bits one-by-one until it becomes
5117 * too coarse to separate the current node from the last.
5119 mask = ~((1 << __ffs(start)) - 1);
5120 while (mask && last_end <= (start & (mask << 1)))
5123 /* accumulate all internode masks */
5127 /* convert mask to number of pages */
5128 return ~accl_mask + 1;
5131 /* Find the lowest pfn for a node */
5132 static unsigned long __init find_min_pfn_for_node(int nid)
5134 unsigned long min_pfn = ULONG_MAX;
5135 unsigned long start_pfn;
5138 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5139 min_pfn = min(min_pfn, start_pfn);
5141 if (min_pfn == ULONG_MAX) {
5143 "Could not find start_pfn for node %d\n", nid);
5151 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5153 * It returns the minimum PFN based on information provided via
5154 * memblock_set_node().
5156 unsigned long __init find_min_pfn_with_active_regions(void)
5158 return find_min_pfn_for_node(MAX_NUMNODES);
5162 * early_calculate_totalpages()
5163 * Sum pages in active regions for movable zone.
5164 * Populate N_MEMORY for calculating usable_nodes.
5166 static unsigned long __init early_calculate_totalpages(void)
5168 unsigned long totalpages = 0;
5169 unsigned long start_pfn, end_pfn;
5172 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5173 unsigned long pages = end_pfn - start_pfn;
5175 totalpages += pages;
5177 node_set_state(nid, N_MEMORY);
5183 * Find the PFN the Movable zone begins in each node. Kernel memory
5184 * is spread evenly between nodes as long as the nodes have enough
5185 * memory. When they don't, some nodes will have more kernelcore than
5188 static void __init find_zone_movable_pfns_for_nodes(void)
5191 unsigned long usable_startpfn;
5192 unsigned long kernelcore_node, kernelcore_remaining;
5193 /* save the state before borrow the nodemask */
5194 nodemask_t saved_node_state = node_states[N_MEMORY];
5195 unsigned long totalpages = early_calculate_totalpages();
5196 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5197 struct memblock_region *r;
5199 /* Need to find movable_zone earlier when movable_node is specified. */
5200 find_usable_zone_for_movable();
5203 * If movable_node is specified, ignore kernelcore and movablecore
5206 if (movable_node_is_enabled()) {
5207 for_each_memblock(memory, r) {
5208 if (!memblock_is_hotpluggable(r))
5213 usable_startpfn = PFN_DOWN(r->base);
5214 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5215 min(usable_startpfn, zone_movable_pfn[nid]) :
5223 * If movablecore=nn[KMG] was specified, calculate what size of
5224 * kernelcore that corresponds so that memory usable for
5225 * any allocation type is evenly spread. If both kernelcore
5226 * and movablecore are specified, then the value of kernelcore
5227 * will be used for required_kernelcore if it's greater than
5228 * what movablecore would have allowed.
5230 if (required_movablecore) {
5231 unsigned long corepages;
5234 * Round-up so that ZONE_MOVABLE is at least as large as what
5235 * was requested by the user
5237 required_movablecore =
5238 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5239 corepages = totalpages - required_movablecore;
5241 required_kernelcore = max(required_kernelcore, corepages);
5244 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5245 if (!required_kernelcore)
5248 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5249 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5252 /* Spread kernelcore memory as evenly as possible throughout nodes */
5253 kernelcore_node = required_kernelcore / usable_nodes;
5254 for_each_node_state(nid, N_MEMORY) {
5255 unsigned long start_pfn, end_pfn;
5258 * Recalculate kernelcore_node if the division per node
5259 * now exceeds what is necessary to satisfy the requested
5260 * amount of memory for the kernel
5262 if (required_kernelcore < kernelcore_node)
5263 kernelcore_node = required_kernelcore / usable_nodes;
5266 * As the map is walked, we track how much memory is usable
5267 * by the kernel using kernelcore_remaining. When it is
5268 * 0, the rest of the node is usable by ZONE_MOVABLE
5270 kernelcore_remaining = kernelcore_node;
5272 /* Go through each range of PFNs within this node */
5273 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5274 unsigned long size_pages;
5276 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5277 if (start_pfn >= end_pfn)
5280 /* Account for what is only usable for kernelcore */
5281 if (start_pfn < usable_startpfn) {
5282 unsigned long kernel_pages;
5283 kernel_pages = min(end_pfn, usable_startpfn)
5286 kernelcore_remaining -= min(kernel_pages,
5287 kernelcore_remaining);
5288 required_kernelcore -= min(kernel_pages,
5289 required_kernelcore);
5291 /* Continue if range is now fully accounted */
5292 if (end_pfn <= usable_startpfn) {
5295 * Push zone_movable_pfn to the end so
5296 * that if we have to rebalance
5297 * kernelcore across nodes, we will
5298 * not double account here
5300 zone_movable_pfn[nid] = end_pfn;
5303 start_pfn = usable_startpfn;
5307 * The usable PFN range for ZONE_MOVABLE is from
5308 * start_pfn->end_pfn. Calculate size_pages as the
5309 * number of pages used as kernelcore
5311 size_pages = end_pfn - start_pfn;
5312 if (size_pages > kernelcore_remaining)
5313 size_pages = kernelcore_remaining;
5314 zone_movable_pfn[nid] = start_pfn + size_pages;
5317 * Some kernelcore has been met, update counts and
5318 * break if the kernelcore for this node has been
5321 required_kernelcore -= min(required_kernelcore,
5323 kernelcore_remaining -= size_pages;
5324 if (!kernelcore_remaining)
5330 * If there is still required_kernelcore, we do another pass with one
5331 * less node in the count. This will push zone_movable_pfn[nid] further
5332 * along on the nodes that still have memory until kernelcore is
5336 if (usable_nodes && required_kernelcore > usable_nodes)
5340 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5341 for (nid = 0; nid < MAX_NUMNODES; nid++)
5342 zone_movable_pfn[nid] =
5343 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5346 /* restore the node_state */
5347 node_states[N_MEMORY] = saved_node_state;
5350 /* Any regular or high memory on that node ? */
5351 static void check_for_memory(pg_data_t *pgdat, int nid)
5353 enum zone_type zone_type;
5355 if (N_MEMORY == N_NORMAL_MEMORY)
5358 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5359 struct zone *zone = &pgdat->node_zones[zone_type];
5360 if (populated_zone(zone)) {
5361 node_set_state(nid, N_HIGH_MEMORY);
5362 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5363 zone_type <= ZONE_NORMAL)
5364 node_set_state(nid, N_NORMAL_MEMORY);
5371 * free_area_init_nodes - Initialise all pg_data_t and zone data
5372 * @max_zone_pfn: an array of max PFNs for each zone
5374 * This will call free_area_init_node() for each active node in the system.
5375 * Using the page ranges provided by memblock_set_node(), the size of each
5376 * zone in each node and their holes is calculated. If the maximum PFN
5377 * between two adjacent zones match, it is assumed that the zone is empty.
5378 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5379 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5380 * starts where the previous one ended. For example, ZONE_DMA32 starts
5381 * at arch_max_dma_pfn.
5383 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5385 unsigned long start_pfn, end_pfn;
5388 /* Record where the zone boundaries are */
5389 memset(arch_zone_lowest_possible_pfn, 0,
5390 sizeof(arch_zone_lowest_possible_pfn));
5391 memset(arch_zone_highest_possible_pfn, 0,
5392 sizeof(arch_zone_highest_possible_pfn));
5393 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5394 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5395 for (i = 1; i < MAX_NR_ZONES; i++) {
5396 if (i == ZONE_MOVABLE)
5398 arch_zone_lowest_possible_pfn[i] =
5399 arch_zone_highest_possible_pfn[i-1];
5400 arch_zone_highest_possible_pfn[i] =
5401 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5403 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5404 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5406 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5407 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5408 find_zone_movable_pfns_for_nodes();
5410 /* Print out the zone ranges */
5411 pr_info("Zone ranges:\n");
5412 for (i = 0; i < MAX_NR_ZONES; i++) {
5413 if (i == ZONE_MOVABLE)
5415 pr_info(" %-8s ", zone_names[i]);
5416 if (arch_zone_lowest_possible_pfn[i] ==
5417 arch_zone_highest_possible_pfn[i])
5420 pr_cont("[mem %0#10lx-%0#10lx]\n",
5421 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5422 (arch_zone_highest_possible_pfn[i]
5423 << PAGE_SHIFT) - 1);
5426 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5427 pr_info("Movable zone start for each node\n");
5428 for (i = 0; i < MAX_NUMNODES; i++) {
5429 if (zone_movable_pfn[i])
5430 pr_info(" Node %d: %#010lx\n", i,
5431 zone_movable_pfn[i] << PAGE_SHIFT);
5434 /* Print out the early node map */
5435 pr_info("Early memory node ranges\n");
5436 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5437 pr_info(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5438 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5440 /* Initialise every node */
5441 mminit_verify_pageflags_layout();
5442 setup_nr_node_ids();
5443 for_each_online_node(nid) {
5444 pg_data_t *pgdat = NODE_DATA(nid);
5445 free_area_init_node(nid, NULL,
5446 find_min_pfn_for_node(nid), NULL);
5448 /* Any memory on that node */
5449 if (pgdat->node_present_pages)
5450 node_set_state(nid, N_MEMORY);
5451 check_for_memory(pgdat, nid);
5455 static int __init cmdline_parse_core(char *p, unsigned long *core)
5457 unsigned long long coremem;
5461 coremem = memparse(p, &p);
5462 *core = coremem >> PAGE_SHIFT;
5464 /* Paranoid check that UL is enough for the coremem value */
5465 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5471 * kernelcore=size sets the amount of memory for use for allocations that
5472 * cannot be reclaimed or migrated.
5474 static int __init cmdline_parse_kernelcore(char *p)
5476 return cmdline_parse_core(p, &required_kernelcore);
5480 * movablecore=size sets the amount of memory for use for allocations that
5481 * can be reclaimed or migrated.
5483 static int __init cmdline_parse_movablecore(char *p)
5485 return cmdline_parse_core(p, &required_movablecore);
5488 early_param("kernelcore", cmdline_parse_kernelcore);
5489 early_param("movablecore", cmdline_parse_movablecore);
5491 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5493 void adjust_managed_page_count(struct page *page, long count)
5495 spin_lock(&managed_page_count_lock);
5496 page_zone(page)->managed_pages += count;
5497 totalram_pages += count;
5498 #ifdef CONFIG_HIGHMEM
5499 if (PageHighMem(page))
5500 totalhigh_pages += count;
5502 spin_unlock(&managed_page_count_lock);
5504 EXPORT_SYMBOL(adjust_managed_page_count);
5506 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5509 unsigned long pages = 0;
5511 start = (void *)PAGE_ALIGN((unsigned long)start);
5512 end = (void *)((unsigned long)end & PAGE_MASK);
5513 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5514 if ((unsigned int)poison <= 0xFF)
5515 memset(pos, poison, PAGE_SIZE);
5516 free_reserved_page(virt_to_page(pos));
5520 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5521 s, pages << (PAGE_SHIFT - 10), start, end);
5525 EXPORT_SYMBOL(free_reserved_area);
5527 #ifdef CONFIG_HIGHMEM
5528 void free_highmem_page(struct page *page)
5530 __free_reserved_page(page);
5532 page_zone(page)->managed_pages++;
5538 void __init mem_init_print_info(const char *str)
5540 unsigned long physpages, codesize, datasize, rosize, bss_size;
5541 unsigned long init_code_size, init_data_size;
5543 physpages = get_num_physpages();
5544 codesize = _etext - _stext;
5545 datasize = _edata - _sdata;
5546 rosize = __end_rodata - __start_rodata;
5547 bss_size = __bss_stop - __bss_start;
5548 init_data_size = __init_end - __init_begin;
5549 init_code_size = _einittext - _sinittext;
5552 * Detect special cases and adjust section sizes accordingly:
5553 * 1) .init.* may be embedded into .data sections
5554 * 2) .init.text.* may be out of [__init_begin, __init_end],
5555 * please refer to arch/tile/kernel/vmlinux.lds.S.
5556 * 3) .rodata.* may be embedded into .text or .data sections.
5558 #define adj_init_size(start, end, size, pos, adj) \
5560 if (start <= pos && pos < end && size > adj) \
5564 adj_init_size(__init_begin, __init_end, init_data_size,
5565 _sinittext, init_code_size);
5566 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5567 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5568 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5569 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5571 #undef adj_init_size
5573 pr_info("Memory: %luK/%luK available "
5574 "(%luK kernel code, %luK rwdata, %luK rodata, "
5575 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5576 #ifdef CONFIG_HIGHMEM
5580 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5581 codesize >> 10, datasize >> 10, rosize >> 10,
5582 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5583 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5584 totalcma_pages << (PAGE_SHIFT-10),
5585 #ifdef CONFIG_HIGHMEM
5586 totalhigh_pages << (PAGE_SHIFT-10),
5588 str ? ", " : "", str ? str : "");
5592 * set_dma_reserve - set the specified number of pages reserved in the first zone
5593 * @new_dma_reserve: The number of pages to mark reserved
5595 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5596 * In the DMA zone, a significant percentage may be consumed by kernel image
5597 * and other unfreeable allocations which can skew the watermarks badly. This
5598 * function may optionally be used to account for unfreeable pages in the
5599 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5600 * smaller per-cpu batchsize.
5602 void __init set_dma_reserve(unsigned long new_dma_reserve)
5604 dma_reserve = new_dma_reserve;
5607 void __init free_area_init(unsigned long *zones_size)
5609 free_area_init_node(0, zones_size,
5610 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5613 static int page_alloc_cpu_notify(struct notifier_block *self,
5614 unsigned long action, void *hcpu)
5616 int cpu = (unsigned long)hcpu;
5618 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5619 lru_add_drain_cpu(cpu);
5623 * Spill the event counters of the dead processor
5624 * into the current processors event counters.
5625 * This artificially elevates the count of the current
5628 vm_events_fold_cpu(cpu);
5631 * Zero the differential counters of the dead processor
5632 * so that the vm statistics are consistent.
5634 * This is only okay since the processor is dead and cannot
5635 * race with what we are doing.
5637 cpu_vm_stats_fold(cpu);
5642 void __init page_alloc_init(void)
5644 hotcpu_notifier(page_alloc_cpu_notify, 0);
5648 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5649 * or min_free_kbytes changes.
5651 static void calculate_totalreserve_pages(void)
5653 struct pglist_data *pgdat;
5654 unsigned long reserve_pages = 0;
5655 enum zone_type i, j;
5657 for_each_online_pgdat(pgdat) {
5658 for (i = 0; i < MAX_NR_ZONES; i++) {
5659 struct zone *zone = pgdat->node_zones + i;
5662 /* Find valid and maximum lowmem_reserve in the zone */
5663 for (j = i; j < MAX_NR_ZONES; j++) {
5664 if (zone->lowmem_reserve[j] > max)
5665 max = zone->lowmem_reserve[j];
5668 /* we treat the high watermark as reserved pages. */
5669 max += high_wmark_pages(zone);
5671 if (max > zone->managed_pages)
5672 max = zone->managed_pages;
5673 reserve_pages += max;
5675 * Lowmem reserves are not available to
5676 * GFP_HIGHUSER page cache allocations and
5677 * kswapd tries to balance zones to their high
5678 * watermark. As a result, neither should be
5679 * regarded as dirtyable memory, to prevent a
5680 * situation where reclaim has to clean pages
5681 * in order to balance the zones.
5683 zone->dirty_balance_reserve = max;
5686 dirty_balance_reserve = reserve_pages;
5687 totalreserve_pages = reserve_pages;
5691 * setup_per_zone_lowmem_reserve - called whenever
5692 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5693 * has a correct pages reserved value, so an adequate number of
5694 * pages are left in the zone after a successful __alloc_pages().
5696 static void setup_per_zone_lowmem_reserve(void)
5698 struct pglist_data *pgdat;
5699 enum zone_type j, idx;
5701 for_each_online_pgdat(pgdat) {
5702 for (j = 0; j < MAX_NR_ZONES; j++) {
5703 struct zone *zone = pgdat->node_zones + j;
5704 unsigned long managed_pages = zone->managed_pages;
5706 zone->lowmem_reserve[j] = 0;
5710 struct zone *lower_zone;
5714 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5715 sysctl_lowmem_reserve_ratio[idx] = 1;
5717 lower_zone = pgdat->node_zones + idx;
5718 lower_zone->lowmem_reserve[j] = managed_pages /
5719 sysctl_lowmem_reserve_ratio[idx];
5720 managed_pages += lower_zone->managed_pages;
5725 /* update totalreserve_pages */
5726 calculate_totalreserve_pages();
5729 static void __setup_per_zone_wmarks(void)
5731 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5732 unsigned long lowmem_pages = 0;
5734 unsigned long flags;
5736 /* Calculate total number of !ZONE_HIGHMEM pages */
5737 for_each_zone(zone) {
5738 if (!is_highmem(zone))
5739 lowmem_pages += zone->managed_pages;
5742 for_each_zone(zone) {
5745 spin_lock_irqsave(&zone->lock, flags);
5746 tmp = (u64)pages_min * zone->managed_pages;
5747 do_div(tmp, lowmem_pages);
5748 if (is_highmem(zone)) {
5750 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5751 * need highmem pages, so cap pages_min to a small
5754 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5755 * deltas controls asynch page reclaim, and so should
5756 * not be capped for highmem.
5758 unsigned long min_pages;
5760 min_pages = zone->managed_pages / 1024;
5761 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5762 zone->watermark[WMARK_MIN] = min_pages;
5765 * If it's a lowmem zone, reserve a number of pages
5766 * proportionate to the zone's size.
5768 zone->watermark[WMARK_MIN] = tmp;
5771 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5772 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5774 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5775 high_wmark_pages(zone) - low_wmark_pages(zone) -
5776 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5778 setup_zone_migrate_reserve(zone);
5779 spin_unlock_irqrestore(&zone->lock, flags);
5782 /* update totalreserve_pages */
5783 calculate_totalreserve_pages();
5787 * setup_per_zone_wmarks - called when min_free_kbytes changes
5788 * or when memory is hot-{added|removed}
5790 * Ensures that the watermark[min,low,high] values for each zone are set
5791 * correctly with respect to min_free_kbytes.
5793 void setup_per_zone_wmarks(void)
5795 mutex_lock(&zonelists_mutex);
5796 __setup_per_zone_wmarks();
5797 mutex_unlock(&zonelists_mutex);
5801 * The inactive anon list should be small enough that the VM never has to
5802 * do too much work, but large enough that each inactive page has a chance
5803 * to be referenced again before it is swapped out.
5805 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5806 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5807 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5808 * the anonymous pages are kept on the inactive list.
5811 * memory ratio inactive anon
5812 * -------------------------------------
5821 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5823 unsigned int gb, ratio;
5825 /* Zone size in gigabytes */
5826 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5828 ratio = int_sqrt(10 * gb);
5832 zone->inactive_ratio = ratio;
5835 static void __meminit setup_per_zone_inactive_ratio(void)
5840 calculate_zone_inactive_ratio(zone);
5844 * Initialise min_free_kbytes.
5846 * For small machines we want it small (128k min). For large machines
5847 * we want it large (64MB max). But it is not linear, because network
5848 * bandwidth does not increase linearly with machine size. We use
5850 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5851 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5867 int __meminit init_per_zone_wmark_min(void)
5869 unsigned long lowmem_kbytes;
5870 int new_min_free_kbytes;
5872 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5873 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5875 if (new_min_free_kbytes > user_min_free_kbytes) {
5876 min_free_kbytes = new_min_free_kbytes;
5877 if (min_free_kbytes < 128)
5878 min_free_kbytes = 128;
5879 if (min_free_kbytes > 65536)
5880 min_free_kbytes = 65536;
5882 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5883 new_min_free_kbytes, user_min_free_kbytes);
5885 setup_per_zone_wmarks();
5886 refresh_zone_stat_thresholds();
5887 setup_per_zone_lowmem_reserve();
5888 setup_per_zone_inactive_ratio();
5891 module_init(init_per_zone_wmark_min)
5894 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5895 * that we can call two helper functions whenever min_free_kbytes
5898 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5899 void __user *buffer, size_t *length, loff_t *ppos)
5903 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5908 user_min_free_kbytes = min_free_kbytes;
5909 setup_per_zone_wmarks();
5915 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5916 void __user *buffer, size_t *length, loff_t *ppos)
5921 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5926 zone->min_unmapped_pages = (zone->managed_pages *
5927 sysctl_min_unmapped_ratio) / 100;
5931 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5932 void __user *buffer, size_t *length, loff_t *ppos)
5937 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5942 zone->min_slab_pages = (zone->managed_pages *
5943 sysctl_min_slab_ratio) / 100;
5949 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5950 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5951 * whenever sysctl_lowmem_reserve_ratio changes.
5953 * The reserve ratio obviously has absolutely no relation with the
5954 * minimum watermarks. The lowmem reserve ratio can only make sense
5955 * if in function of the boot time zone sizes.
5957 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5958 void __user *buffer, size_t *length, loff_t *ppos)
5960 proc_dointvec_minmax(table, write, buffer, length, ppos);
5961 setup_per_zone_lowmem_reserve();
5966 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5967 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5968 * pagelist can have before it gets flushed back to buddy allocator.
5970 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5971 void __user *buffer, size_t *length, loff_t *ppos)
5974 int old_percpu_pagelist_fraction;
5977 mutex_lock(&pcp_batch_high_lock);
5978 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5980 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5981 if (!write || ret < 0)
5984 /* Sanity checking to avoid pcp imbalance */
5985 if (percpu_pagelist_fraction &&
5986 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
5987 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
5993 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
5996 for_each_populated_zone(zone) {
5999 for_each_possible_cpu(cpu)
6000 pageset_set_high_and_batch(zone,
6001 per_cpu_ptr(zone->pageset, cpu));
6004 mutex_unlock(&pcp_batch_high_lock);
6008 int hashdist = HASHDIST_DEFAULT;
6011 static int __init set_hashdist(char *str)
6015 hashdist = simple_strtoul(str, &str, 0);
6018 __setup("hashdist=", set_hashdist);
6022 * allocate a large system hash table from bootmem
6023 * - it is assumed that the hash table must contain an exact power-of-2
6024 * quantity of entries
6025 * - limit is the number of hash buckets, not the total allocation size
6027 void *__init alloc_large_system_hash(const char *tablename,
6028 unsigned long bucketsize,
6029 unsigned long numentries,
6032 unsigned int *_hash_shift,
6033 unsigned int *_hash_mask,
6034 unsigned long low_limit,
6035 unsigned long high_limit)
6037 unsigned long long max = high_limit;
6038 unsigned long log2qty, size;
6041 /* allow the kernel cmdline to have a say */
6043 /* round applicable memory size up to nearest megabyte */
6044 numentries = nr_kernel_pages;
6046 /* It isn't necessary when PAGE_SIZE >= 1MB */
6047 if (PAGE_SHIFT < 20)
6048 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6050 /* limit to 1 bucket per 2^scale bytes of low memory */
6051 if (scale > PAGE_SHIFT)
6052 numentries >>= (scale - PAGE_SHIFT);
6054 numentries <<= (PAGE_SHIFT - scale);
6056 /* Make sure we've got at least a 0-order allocation.. */
6057 if (unlikely(flags & HASH_SMALL)) {
6058 /* Makes no sense without HASH_EARLY */
6059 WARN_ON(!(flags & HASH_EARLY));
6060 if (!(numentries >> *_hash_shift)) {
6061 numentries = 1UL << *_hash_shift;
6062 BUG_ON(!numentries);
6064 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6065 numentries = PAGE_SIZE / bucketsize;
6067 numentries = roundup_pow_of_two(numentries);
6069 /* limit allocation size to 1/16 total memory by default */
6071 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6072 do_div(max, bucketsize);
6074 max = min(max, 0x80000000ULL);
6076 if (numentries < low_limit)
6077 numentries = low_limit;
6078 if (numentries > max)
6081 log2qty = ilog2(numentries);
6084 size = bucketsize << log2qty;
6085 if (flags & HASH_EARLY)
6086 table = memblock_virt_alloc_nopanic(size, 0);
6088 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6091 * If bucketsize is not a power-of-two, we may free
6092 * some pages at the end of hash table which
6093 * alloc_pages_exact() automatically does
6095 if (get_order(size) < MAX_ORDER) {
6096 table = alloc_pages_exact(size, GFP_ATOMIC);
6097 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6100 } while (!table && size > PAGE_SIZE && --log2qty);
6103 panic("Failed to allocate %s hash table\n", tablename);
6105 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6108 ilog2(size) - PAGE_SHIFT,
6112 *_hash_shift = log2qty;
6114 *_hash_mask = (1 << log2qty) - 1;
6119 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6120 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6123 #ifdef CONFIG_SPARSEMEM
6124 return __pfn_to_section(pfn)->pageblock_flags;
6126 return zone->pageblock_flags;
6127 #endif /* CONFIG_SPARSEMEM */
6130 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6132 #ifdef CONFIG_SPARSEMEM
6133 pfn &= (PAGES_PER_SECTION-1);
6134 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6136 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6137 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6138 #endif /* CONFIG_SPARSEMEM */
6142 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6143 * @page: The page within the block of interest
6144 * @pfn: The target page frame number
6145 * @end_bitidx: The last bit of interest to retrieve
6146 * @mask: mask of bits that the caller is interested in
6148 * Return: pageblock_bits flags
6150 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6151 unsigned long end_bitidx,
6155 unsigned long *bitmap;
6156 unsigned long bitidx, word_bitidx;
6159 zone = page_zone(page);
6160 bitmap = get_pageblock_bitmap(zone, pfn);
6161 bitidx = pfn_to_bitidx(zone, pfn);
6162 word_bitidx = bitidx / BITS_PER_LONG;
6163 bitidx &= (BITS_PER_LONG-1);
6165 word = bitmap[word_bitidx];
6166 bitidx += end_bitidx;
6167 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6171 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6172 * @page: The page within the block of interest
6173 * @flags: The flags to set
6174 * @pfn: The target page frame number
6175 * @end_bitidx: The last bit of interest
6176 * @mask: mask of bits that the caller is interested in
6178 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6180 unsigned long end_bitidx,
6184 unsigned long *bitmap;
6185 unsigned long bitidx, word_bitidx;
6186 unsigned long old_word, word;
6188 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6190 zone = page_zone(page);
6191 bitmap = get_pageblock_bitmap(zone, pfn);
6192 bitidx = pfn_to_bitidx(zone, pfn);
6193 word_bitidx = bitidx / BITS_PER_LONG;
6194 bitidx &= (BITS_PER_LONG-1);
6196 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6198 bitidx += end_bitidx;
6199 mask <<= (BITS_PER_LONG - bitidx - 1);
6200 flags <<= (BITS_PER_LONG - bitidx - 1);
6202 word = ACCESS_ONCE(bitmap[word_bitidx]);
6204 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6205 if (word == old_word)
6212 * This function checks whether pageblock includes unmovable pages or not.
6213 * If @count is not zero, it is okay to include less @count unmovable pages
6215 * PageLRU check without isolation or lru_lock could race so that
6216 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6217 * expect this function should be exact.
6219 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6220 bool skip_hwpoisoned_pages)
6222 unsigned long pfn, iter, found;
6226 * For avoiding noise data, lru_add_drain_all() should be called
6227 * If ZONE_MOVABLE, the zone never contains unmovable pages
6229 if (zone_idx(zone) == ZONE_MOVABLE)
6231 mt = get_pageblock_migratetype(page);
6232 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6235 pfn = page_to_pfn(page);
6236 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6237 unsigned long check = pfn + iter;
6239 if (!pfn_valid_within(check))
6242 page = pfn_to_page(check);
6245 * Hugepages are not in LRU lists, but they're movable.
6246 * We need not scan over tail pages bacause we don't
6247 * handle each tail page individually in migration.
6249 if (PageHuge(page)) {
6250 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6255 * We can't use page_count without pin a page
6256 * because another CPU can free compound page.
6257 * This check already skips compound tails of THP
6258 * because their page->_count is zero at all time.
6260 if (!atomic_read(&page->_count)) {
6261 if (PageBuddy(page))
6262 iter += (1 << page_order(page)) - 1;
6267 * The HWPoisoned page may be not in buddy system, and
6268 * page_count() is not 0.
6270 if (skip_hwpoisoned_pages && PageHWPoison(page))
6276 * If there are RECLAIMABLE pages, we need to check
6277 * it. But now, memory offline itself doesn't call
6278 * shrink_node_slabs() and it still to be fixed.
6281 * If the page is not RAM, page_count()should be 0.
6282 * we don't need more check. This is an _used_ not-movable page.
6284 * The problematic thing here is PG_reserved pages. PG_reserved
6285 * is set to both of a memory hole page and a _used_ kernel
6294 bool is_pageblock_removable_nolock(struct page *page)
6300 * We have to be careful here because we are iterating over memory
6301 * sections which are not zone aware so we might end up outside of
6302 * the zone but still within the section.
6303 * We have to take care about the node as well. If the node is offline
6304 * its NODE_DATA will be NULL - see page_zone.
6306 if (!node_online(page_to_nid(page)))
6309 zone = page_zone(page);
6310 pfn = page_to_pfn(page);
6311 if (!zone_spans_pfn(zone, pfn))
6314 return !has_unmovable_pages(zone, page, 0, true);
6319 static unsigned long pfn_max_align_down(unsigned long pfn)
6321 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6322 pageblock_nr_pages) - 1);
6325 static unsigned long pfn_max_align_up(unsigned long pfn)
6327 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6328 pageblock_nr_pages));
6331 /* [start, end) must belong to a single zone. */
6332 static int __alloc_contig_migrate_range(struct compact_control *cc,
6333 unsigned long start, unsigned long end)
6335 /* This function is based on compact_zone() from compaction.c. */
6336 unsigned long nr_reclaimed;
6337 unsigned long pfn = start;
6338 unsigned int tries = 0;
6343 while (pfn < end || !list_empty(&cc->migratepages)) {
6344 if (fatal_signal_pending(current)) {
6349 if (list_empty(&cc->migratepages)) {
6350 cc->nr_migratepages = 0;
6351 pfn = isolate_migratepages_range(cc, pfn, end);
6357 } else if (++tries == 5) {
6358 ret = ret < 0 ? ret : -EBUSY;
6362 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6364 cc->nr_migratepages -= nr_reclaimed;
6366 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6367 NULL, 0, cc->mode, MR_CMA);
6370 putback_movable_pages(&cc->migratepages);
6377 * alloc_contig_range() -- tries to allocate given range of pages
6378 * @start: start PFN to allocate
6379 * @end: one-past-the-last PFN to allocate
6380 * @migratetype: migratetype of the underlaying pageblocks (either
6381 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6382 * in range must have the same migratetype and it must
6383 * be either of the two.
6385 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6386 * aligned, however it's the caller's responsibility to guarantee that
6387 * we are the only thread that changes migrate type of pageblocks the
6390 * The PFN range must belong to a single zone.
6392 * Returns zero on success or negative error code. On success all
6393 * pages which PFN is in [start, end) are allocated for the caller and
6394 * need to be freed with free_contig_range().
6396 int alloc_contig_range(unsigned long start, unsigned long end,
6397 unsigned migratetype)
6399 unsigned long outer_start, outer_end;
6402 struct compact_control cc = {
6403 .nr_migratepages = 0,
6405 .zone = page_zone(pfn_to_page(start)),
6406 .mode = MIGRATE_SYNC,
6407 .ignore_skip_hint = true,
6409 INIT_LIST_HEAD(&cc.migratepages);
6412 * What we do here is we mark all pageblocks in range as
6413 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6414 * have different sizes, and due to the way page allocator
6415 * work, we align the range to biggest of the two pages so
6416 * that page allocator won't try to merge buddies from
6417 * different pageblocks and change MIGRATE_ISOLATE to some
6418 * other migration type.
6420 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6421 * migrate the pages from an unaligned range (ie. pages that
6422 * we are interested in). This will put all the pages in
6423 * range back to page allocator as MIGRATE_ISOLATE.
6425 * When this is done, we take the pages in range from page
6426 * allocator removing them from the buddy system. This way
6427 * page allocator will never consider using them.
6429 * This lets us mark the pageblocks back as
6430 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6431 * aligned range but not in the unaligned, original range are
6432 * put back to page allocator so that buddy can use them.
6435 ret = start_isolate_page_range(pfn_max_align_down(start),
6436 pfn_max_align_up(end), migratetype,
6441 ret = __alloc_contig_migrate_range(&cc, start, end);
6446 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6447 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6448 * more, all pages in [start, end) are free in page allocator.
6449 * What we are going to do is to allocate all pages from
6450 * [start, end) (that is remove them from page allocator).
6452 * The only problem is that pages at the beginning and at the
6453 * end of interesting range may be not aligned with pages that
6454 * page allocator holds, ie. they can be part of higher order
6455 * pages. Because of this, we reserve the bigger range and
6456 * once this is done free the pages we are not interested in.
6458 * We don't have to hold zone->lock here because the pages are
6459 * isolated thus they won't get removed from buddy.
6462 lru_add_drain_all();
6463 drain_all_pages(cc.zone);
6466 outer_start = start;
6467 while (!PageBuddy(pfn_to_page(outer_start))) {
6468 if (++order >= MAX_ORDER) {
6472 outer_start &= ~0UL << order;
6475 /* Make sure the range is really isolated. */
6476 if (test_pages_isolated(outer_start, end, false)) {
6477 pr_info("%s: [%lx, %lx) PFNs busy\n",
6478 __func__, outer_start, end);
6483 /* Grab isolated pages from freelists. */
6484 outer_end = isolate_freepages_range(&cc, outer_start, end);
6490 /* Free head and tail (if any) */
6491 if (start != outer_start)
6492 free_contig_range(outer_start, start - outer_start);
6493 if (end != outer_end)
6494 free_contig_range(end, outer_end - end);
6497 undo_isolate_page_range(pfn_max_align_down(start),
6498 pfn_max_align_up(end), migratetype);
6502 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6504 unsigned int count = 0;
6506 for (; nr_pages--; pfn++) {
6507 struct page *page = pfn_to_page(pfn);
6509 count += page_count(page) != 1;
6512 WARN(count != 0, "%d pages are still in use!\n", count);
6516 #ifdef CONFIG_MEMORY_HOTPLUG
6518 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6519 * page high values need to be recalulated.
6521 void __meminit zone_pcp_update(struct zone *zone)
6524 mutex_lock(&pcp_batch_high_lock);
6525 for_each_possible_cpu(cpu)
6526 pageset_set_high_and_batch(zone,
6527 per_cpu_ptr(zone->pageset, cpu));
6528 mutex_unlock(&pcp_batch_high_lock);
6532 void zone_pcp_reset(struct zone *zone)
6534 unsigned long flags;
6536 struct per_cpu_pageset *pset;
6538 /* avoid races with drain_pages() */
6539 local_irq_save(flags);
6540 if (zone->pageset != &boot_pageset) {
6541 for_each_online_cpu(cpu) {
6542 pset = per_cpu_ptr(zone->pageset, cpu);
6543 drain_zonestat(zone, pset);
6545 free_percpu(zone->pageset);
6546 zone->pageset = &boot_pageset;
6548 local_irq_restore(flags);
6551 #ifdef CONFIG_MEMORY_HOTREMOVE
6553 * All pages in the range must be isolated before calling this.
6556 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6560 unsigned int order, i;
6562 unsigned long flags;
6563 /* find the first valid pfn */
6564 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6569 zone = page_zone(pfn_to_page(pfn));
6570 spin_lock_irqsave(&zone->lock, flags);
6572 while (pfn < end_pfn) {
6573 if (!pfn_valid(pfn)) {
6577 page = pfn_to_page(pfn);
6579 * The HWPoisoned page may be not in buddy system, and
6580 * page_count() is not 0.
6582 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6584 SetPageReserved(page);
6588 BUG_ON(page_count(page));
6589 BUG_ON(!PageBuddy(page));
6590 order = page_order(page);
6591 #ifdef CONFIG_DEBUG_VM
6592 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6593 pfn, 1 << order, end_pfn);
6595 list_del(&page->lru);
6596 rmv_page_order(page);
6597 zone->free_area[order].nr_free--;
6598 for (i = 0; i < (1 << order); i++)
6599 SetPageReserved((page+i));
6600 pfn += (1 << order);
6602 spin_unlock_irqrestore(&zone->lock, flags);
6606 #ifdef CONFIG_MEMORY_FAILURE
6607 bool is_free_buddy_page(struct page *page)
6609 struct zone *zone = page_zone(page);
6610 unsigned long pfn = page_to_pfn(page);
6611 unsigned long flags;
6614 spin_lock_irqsave(&zone->lock, flags);
6615 for (order = 0; order < MAX_ORDER; order++) {
6616 struct page *page_head = page - (pfn & ((1 << order) - 1));
6618 if (PageBuddy(page_head) && page_order(page_head) >= order)
6621 spin_unlock_irqrestore(&zone->lock, flags);
6623 return order < MAX_ORDER;