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 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
557 if (page_zone_id(page) != page_zone_id(buddy))
563 if (PageBuddy(buddy) && page_order(buddy) == order) {
564 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
567 * zone check is done late to avoid uselessly
568 * calculating zone/node ids for pages that could
571 if (page_zone_id(page) != page_zone_id(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)
977 for (i = 0; i < (1 << order); i++) {
978 struct page *p = page + i;
979 if (unlikely(check_new_page(p)))
983 set_page_private(page, 0);
984 set_page_refcounted(page);
986 arch_alloc_page(page, order);
987 kernel_map_pages(page, 1 << order, 1);
989 if (gfp_flags & __GFP_ZERO)
990 prep_zero_page(page, order, gfp_flags);
992 if (order && (gfp_flags & __GFP_COMP))
993 prep_compound_page(page, order);
995 set_page_owner(page, order, gfp_flags);
1001 * Go through the free lists for the given migratetype and remove
1002 * the smallest available page from the freelists
1005 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1008 unsigned int current_order;
1009 struct free_area *area;
1012 /* Find a page of the appropriate size in the preferred list */
1013 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1014 area = &(zone->free_area[current_order]);
1015 if (list_empty(&area->free_list[migratetype]))
1018 page = list_entry(area->free_list[migratetype].next,
1020 list_del(&page->lru);
1021 rmv_page_order(page);
1023 expand(zone, page, order, current_order, area, migratetype);
1024 set_freepage_migratetype(page, migratetype);
1033 * This array describes the order lists are fallen back to when
1034 * the free lists for the desirable migrate type are depleted
1036 static int fallbacks[MIGRATE_TYPES][4] = {
1037 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1038 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1040 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1041 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1043 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1045 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1046 #ifdef CONFIG_MEMORY_ISOLATION
1047 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1052 * Move the free pages in a range to the free lists of the requested type.
1053 * Note that start_page and end_pages are not aligned on a pageblock
1054 * boundary. If alignment is required, use move_freepages_block()
1056 int move_freepages(struct zone *zone,
1057 struct page *start_page, struct page *end_page,
1061 unsigned long order;
1062 int pages_moved = 0;
1064 #ifndef CONFIG_HOLES_IN_ZONE
1066 * page_zone is not safe to call in this context when
1067 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1068 * anyway as we check zone boundaries in move_freepages_block().
1069 * Remove at a later date when no bug reports exist related to
1070 * grouping pages by mobility
1072 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1075 for (page = start_page; page <= end_page;) {
1076 /* Make sure we are not inadvertently changing nodes */
1077 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1079 if (!pfn_valid_within(page_to_pfn(page))) {
1084 if (!PageBuddy(page)) {
1089 order = page_order(page);
1090 list_move(&page->lru,
1091 &zone->free_area[order].free_list[migratetype]);
1092 set_freepage_migratetype(page, migratetype);
1094 pages_moved += 1 << order;
1100 int move_freepages_block(struct zone *zone, struct page *page,
1103 unsigned long start_pfn, end_pfn;
1104 struct page *start_page, *end_page;
1106 start_pfn = page_to_pfn(page);
1107 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1108 start_page = pfn_to_page(start_pfn);
1109 end_page = start_page + pageblock_nr_pages - 1;
1110 end_pfn = start_pfn + pageblock_nr_pages - 1;
1112 /* Do not cross zone boundaries */
1113 if (!zone_spans_pfn(zone, start_pfn))
1115 if (!zone_spans_pfn(zone, end_pfn))
1118 return move_freepages(zone, start_page, end_page, migratetype);
1121 static void change_pageblock_range(struct page *pageblock_page,
1122 int start_order, int migratetype)
1124 int nr_pageblocks = 1 << (start_order - pageblock_order);
1126 while (nr_pageblocks--) {
1127 set_pageblock_migratetype(pageblock_page, migratetype);
1128 pageblock_page += pageblock_nr_pages;
1133 * If breaking a large block of pages, move all free pages to the preferred
1134 * allocation list. If falling back for a reclaimable kernel allocation, be
1135 * more aggressive about taking ownership of free pages.
1137 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1138 * nor move CMA pages to different free lists. We don't want unmovable pages
1139 * to be allocated from MIGRATE_CMA areas.
1141 * Returns the new migratetype of the pageblock (or the same old migratetype
1142 * if it was unchanged).
1144 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1145 int start_type, int fallback_type)
1147 int current_order = page_order(page);
1150 * When borrowing from MIGRATE_CMA, we need to release the excess
1151 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1152 * is set to CMA so it is returned to the correct freelist in case
1153 * the page ends up being not actually allocated from the pcp lists.
1155 if (is_migrate_cma(fallback_type))
1156 return fallback_type;
1158 /* Take ownership for orders >= pageblock_order */
1159 if (current_order >= pageblock_order) {
1160 change_pageblock_range(page, current_order, start_type);
1164 if (current_order >= pageblock_order / 2 ||
1165 start_type == MIGRATE_RECLAIMABLE ||
1166 page_group_by_mobility_disabled) {
1169 pages = move_freepages_block(zone, page, start_type);
1171 /* Claim the whole block if over half of it is free */
1172 if (pages >= (1 << (pageblock_order-1)) ||
1173 page_group_by_mobility_disabled) {
1175 set_pageblock_migratetype(page, start_type);
1181 return fallback_type;
1184 /* Remove an element from the buddy allocator from the fallback list */
1185 static inline struct page *
1186 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1188 struct free_area *area;
1189 unsigned int current_order;
1191 int migratetype, new_type, i;
1193 /* Find the largest possible block of pages in the other list */
1194 for (current_order = MAX_ORDER-1;
1195 current_order >= order && current_order <= MAX_ORDER-1;
1198 migratetype = fallbacks[start_migratetype][i];
1200 /* MIGRATE_RESERVE handled later if necessary */
1201 if (migratetype == MIGRATE_RESERVE)
1204 area = &(zone->free_area[current_order]);
1205 if (list_empty(&area->free_list[migratetype]))
1208 page = list_entry(area->free_list[migratetype].next,
1212 new_type = try_to_steal_freepages(zone, page,
1216 /* Remove the page from the freelists */
1217 list_del(&page->lru);
1218 rmv_page_order(page);
1220 expand(zone, page, order, current_order, area,
1222 /* The freepage_migratetype may differ from pageblock's
1223 * migratetype depending on the decisions in
1224 * try_to_steal_freepages. This is OK as long as it does
1225 * not differ for MIGRATE_CMA type.
1227 set_freepage_migratetype(page, new_type);
1229 trace_mm_page_alloc_extfrag(page, order, current_order,
1230 start_migratetype, migratetype, new_type);
1240 * Do the hard work of removing an element from the buddy allocator.
1241 * Call me with the zone->lock already held.
1243 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1249 page = __rmqueue_smallest(zone, order, migratetype);
1251 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1252 page = __rmqueue_fallback(zone, order, migratetype);
1255 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1256 * is used because __rmqueue_smallest is an inline function
1257 * and we want just one call site
1260 migratetype = MIGRATE_RESERVE;
1265 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1270 * Obtain a specified number of elements from the buddy allocator, all under
1271 * a single hold of the lock, for efficiency. Add them to the supplied list.
1272 * Returns the number of new pages which were placed at *list.
1274 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1275 unsigned long count, struct list_head *list,
1276 int migratetype, bool cold)
1280 spin_lock(&zone->lock);
1281 for (i = 0; i < count; ++i) {
1282 struct page *page = __rmqueue(zone, order, migratetype);
1283 if (unlikely(page == NULL))
1287 * Split buddy pages returned by expand() are received here
1288 * in physical page order. The page is added to the callers and
1289 * list and the list head then moves forward. From the callers
1290 * perspective, the linked list is ordered by page number in
1291 * some conditions. This is useful for IO devices that can
1292 * merge IO requests if the physical pages are ordered
1296 list_add(&page->lru, list);
1298 list_add_tail(&page->lru, list);
1300 if (is_migrate_cma(get_freepage_migratetype(page)))
1301 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1304 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1305 spin_unlock(&zone->lock);
1311 * Called from the vmstat counter updater to drain pagesets of this
1312 * currently executing processor on remote nodes after they have
1315 * Note that this function must be called with the thread pinned to
1316 * a single processor.
1318 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1320 unsigned long flags;
1321 int to_drain, batch;
1323 local_irq_save(flags);
1324 batch = ACCESS_ONCE(pcp->batch);
1325 to_drain = min(pcp->count, batch);
1327 free_pcppages_bulk(zone, to_drain, pcp);
1328 pcp->count -= to_drain;
1330 local_irq_restore(flags);
1335 * Drain pcplists of the indicated processor and zone.
1337 * The processor must either be the current processor and the
1338 * thread pinned to the current processor or a processor that
1341 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1343 unsigned long flags;
1344 struct per_cpu_pageset *pset;
1345 struct per_cpu_pages *pcp;
1347 local_irq_save(flags);
1348 pset = per_cpu_ptr(zone->pageset, cpu);
1352 free_pcppages_bulk(zone, pcp->count, pcp);
1355 local_irq_restore(flags);
1359 * Drain pcplists of all zones on the indicated processor.
1361 * The processor must either be the current processor and the
1362 * thread pinned to the current processor or a processor that
1365 static void drain_pages(unsigned int cpu)
1369 for_each_populated_zone(zone) {
1370 drain_pages_zone(cpu, zone);
1375 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1377 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1378 * the single zone's pages.
1380 void drain_local_pages(struct zone *zone)
1382 int cpu = smp_processor_id();
1385 drain_pages_zone(cpu, zone);
1391 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1393 * When zone parameter is non-NULL, spill just the single zone's pages.
1395 * Note that this code is protected against sending an IPI to an offline
1396 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1397 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1398 * nothing keeps CPUs from showing up after we populated the cpumask and
1399 * before the call to on_each_cpu_mask().
1401 void drain_all_pages(struct zone *zone)
1406 * Allocate in the BSS so we wont require allocation in
1407 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1409 static cpumask_t cpus_with_pcps;
1412 * We don't care about racing with CPU hotplug event
1413 * as offline notification will cause the notified
1414 * cpu to drain that CPU pcps and on_each_cpu_mask
1415 * disables preemption as part of its processing
1417 for_each_online_cpu(cpu) {
1418 struct per_cpu_pageset *pcp;
1420 bool has_pcps = false;
1423 pcp = per_cpu_ptr(zone->pageset, cpu);
1427 for_each_populated_zone(z) {
1428 pcp = per_cpu_ptr(z->pageset, cpu);
1429 if (pcp->pcp.count) {
1437 cpumask_set_cpu(cpu, &cpus_with_pcps);
1439 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1441 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1445 #ifdef CONFIG_HIBERNATION
1447 void mark_free_pages(struct zone *zone)
1449 unsigned long pfn, max_zone_pfn;
1450 unsigned long flags;
1451 unsigned int order, t;
1452 struct list_head *curr;
1454 if (zone_is_empty(zone))
1457 spin_lock_irqsave(&zone->lock, flags);
1459 max_zone_pfn = zone_end_pfn(zone);
1460 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1461 if (pfn_valid(pfn)) {
1462 struct page *page = pfn_to_page(pfn);
1464 if (!swsusp_page_is_forbidden(page))
1465 swsusp_unset_page_free(page);
1468 for_each_migratetype_order(order, t) {
1469 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1472 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1473 for (i = 0; i < (1UL << order); i++)
1474 swsusp_set_page_free(pfn_to_page(pfn + i));
1477 spin_unlock_irqrestore(&zone->lock, flags);
1479 #endif /* CONFIG_PM */
1482 * Free a 0-order page
1483 * cold == true ? free a cold page : free a hot page
1485 void free_hot_cold_page(struct page *page, bool cold)
1487 struct zone *zone = page_zone(page);
1488 struct per_cpu_pages *pcp;
1489 unsigned long flags;
1490 unsigned long pfn = page_to_pfn(page);
1493 if (!free_pages_prepare(page, 0))
1496 migratetype = get_pfnblock_migratetype(page, pfn);
1497 set_freepage_migratetype(page, migratetype);
1498 local_irq_save(flags);
1499 __count_vm_event(PGFREE);
1502 * We only track unmovable, reclaimable and movable on pcp lists.
1503 * Free ISOLATE pages back to the allocator because they are being
1504 * offlined but treat RESERVE as movable pages so we can get those
1505 * areas back if necessary. Otherwise, we may have to free
1506 * excessively into the page allocator
1508 if (migratetype >= MIGRATE_PCPTYPES) {
1509 if (unlikely(is_migrate_isolate(migratetype))) {
1510 free_one_page(zone, page, pfn, 0, migratetype);
1513 migratetype = MIGRATE_MOVABLE;
1516 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1518 list_add(&page->lru, &pcp->lists[migratetype]);
1520 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1522 if (pcp->count >= pcp->high) {
1523 unsigned long batch = ACCESS_ONCE(pcp->batch);
1524 free_pcppages_bulk(zone, batch, pcp);
1525 pcp->count -= batch;
1529 local_irq_restore(flags);
1533 * Free a list of 0-order pages
1535 void free_hot_cold_page_list(struct list_head *list, bool cold)
1537 struct page *page, *next;
1539 list_for_each_entry_safe(page, next, list, lru) {
1540 trace_mm_page_free_batched(page, cold);
1541 free_hot_cold_page(page, cold);
1546 * split_page takes a non-compound higher-order page, and splits it into
1547 * n (1<<order) sub-pages: page[0..n]
1548 * Each sub-page must be freed individually.
1550 * Note: this is probably too low level an operation for use in drivers.
1551 * Please consult with lkml before using this in your driver.
1553 void split_page(struct page *page, unsigned int order)
1557 VM_BUG_ON_PAGE(PageCompound(page), page);
1558 VM_BUG_ON_PAGE(!page_count(page), page);
1560 #ifdef CONFIG_KMEMCHECK
1562 * Split shadow pages too, because free(page[0]) would
1563 * otherwise free the whole shadow.
1565 if (kmemcheck_page_is_tracked(page))
1566 split_page(virt_to_page(page[0].shadow), order);
1569 set_page_owner(page, 0, 0);
1570 for (i = 1; i < (1 << order); i++) {
1571 set_page_refcounted(page + i);
1572 set_page_owner(page + i, 0, 0);
1575 EXPORT_SYMBOL_GPL(split_page);
1577 int __isolate_free_page(struct page *page, unsigned int order)
1579 unsigned long watermark;
1583 BUG_ON(!PageBuddy(page));
1585 zone = page_zone(page);
1586 mt = get_pageblock_migratetype(page);
1588 if (!is_migrate_isolate(mt)) {
1589 /* Obey watermarks as if the page was being allocated */
1590 watermark = low_wmark_pages(zone) + (1 << order);
1591 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1594 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1597 /* Remove page from free list */
1598 list_del(&page->lru);
1599 zone->free_area[order].nr_free--;
1600 rmv_page_order(page);
1602 /* Set the pageblock if the isolated page is at least a pageblock */
1603 if (order >= pageblock_order - 1) {
1604 struct page *endpage = page + (1 << order) - 1;
1605 for (; page < endpage; page += pageblock_nr_pages) {
1606 int mt = get_pageblock_migratetype(page);
1607 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1608 set_pageblock_migratetype(page,
1613 set_page_owner(page, order, 0);
1614 return 1UL << order;
1618 * Similar to split_page except the page is already free. As this is only
1619 * being used for migration, the migratetype of the block also changes.
1620 * As this is called with interrupts disabled, the caller is responsible
1621 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1624 * Note: this is probably too low level an operation for use in drivers.
1625 * Please consult with lkml before using this in your driver.
1627 int split_free_page(struct page *page)
1632 order = page_order(page);
1634 nr_pages = __isolate_free_page(page, order);
1638 /* Split into individual pages */
1639 set_page_refcounted(page);
1640 split_page(page, order);
1645 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1646 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1650 struct page *buffered_rmqueue(struct zone *preferred_zone,
1651 struct zone *zone, unsigned int order,
1652 gfp_t gfp_flags, int migratetype)
1654 unsigned long flags;
1656 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1659 if (likely(order == 0)) {
1660 struct per_cpu_pages *pcp;
1661 struct list_head *list;
1663 local_irq_save(flags);
1664 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1665 list = &pcp->lists[migratetype];
1666 if (list_empty(list)) {
1667 pcp->count += rmqueue_bulk(zone, 0,
1670 if (unlikely(list_empty(list)))
1675 page = list_entry(list->prev, struct page, lru);
1677 page = list_entry(list->next, struct page, lru);
1679 list_del(&page->lru);
1682 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1684 * __GFP_NOFAIL is not to be used in new code.
1686 * All __GFP_NOFAIL callers should be fixed so that they
1687 * properly detect and handle allocation failures.
1689 * We most definitely don't want callers attempting to
1690 * allocate greater than order-1 page units with
1693 WARN_ON_ONCE(order > 1);
1695 spin_lock_irqsave(&zone->lock, flags);
1696 page = __rmqueue(zone, order, migratetype);
1697 spin_unlock(&zone->lock);
1700 __mod_zone_freepage_state(zone, -(1 << order),
1701 get_freepage_migratetype(page));
1704 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1705 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1706 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1707 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1709 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1710 zone_statistics(preferred_zone, zone, gfp_flags);
1711 local_irq_restore(flags);
1713 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1714 if (prep_new_page(page, order, gfp_flags))
1719 local_irq_restore(flags);
1723 #ifdef CONFIG_FAIL_PAGE_ALLOC
1726 struct fault_attr attr;
1728 u32 ignore_gfp_highmem;
1729 u32 ignore_gfp_wait;
1731 } fail_page_alloc = {
1732 .attr = FAULT_ATTR_INITIALIZER,
1733 .ignore_gfp_wait = 1,
1734 .ignore_gfp_highmem = 1,
1738 static int __init setup_fail_page_alloc(char *str)
1740 return setup_fault_attr(&fail_page_alloc.attr, str);
1742 __setup("fail_page_alloc=", setup_fail_page_alloc);
1744 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1746 if (order < fail_page_alloc.min_order)
1748 if (gfp_mask & __GFP_NOFAIL)
1750 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1752 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1755 return should_fail(&fail_page_alloc.attr, 1 << order);
1758 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1760 static int __init fail_page_alloc_debugfs(void)
1762 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1765 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1766 &fail_page_alloc.attr);
1768 return PTR_ERR(dir);
1770 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1771 &fail_page_alloc.ignore_gfp_wait))
1773 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1774 &fail_page_alloc.ignore_gfp_highmem))
1776 if (!debugfs_create_u32("min-order", mode, dir,
1777 &fail_page_alloc.min_order))
1782 debugfs_remove_recursive(dir);
1787 late_initcall(fail_page_alloc_debugfs);
1789 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1791 #else /* CONFIG_FAIL_PAGE_ALLOC */
1793 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1798 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1801 * Return true if free pages are above 'mark'. This takes into account the order
1802 * of the allocation.
1804 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1805 unsigned long mark, int classzone_idx, int alloc_flags,
1808 /* free_pages may go negative - that's OK */
1813 free_pages -= (1 << order) - 1;
1814 if (alloc_flags & ALLOC_HIGH)
1816 if (alloc_flags & ALLOC_HARDER)
1819 /* If allocation can't use CMA areas don't use free CMA pages */
1820 if (!(alloc_flags & ALLOC_CMA))
1821 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1824 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1826 for (o = 0; o < order; o++) {
1827 /* At the next order, this order's pages become unavailable */
1828 free_pages -= z->free_area[o].nr_free << o;
1830 /* Require fewer higher order pages to be free */
1833 if (free_pages <= min)
1839 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1840 int classzone_idx, int alloc_flags)
1842 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1843 zone_page_state(z, NR_FREE_PAGES));
1846 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1847 unsigned long mark, int classzone_idx, int alloc_flags)
1849 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1851 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1852 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1854 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1860 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1861 * skip over zones that are not allowed by the cpuset, or that have
1862 * been recently (in last second) found to be nearly full. See further
1863 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1864 * that have to skip over a lot of full or unallowed zones.
1866 * If the zonelist cache is present in the passed zonelist, then
1867 * returns a pointer to the allowed node mask (either the current
1868 * tasks mems_allowed, or node_states[N_MEMORY].)
1870 * If the zonelist cache is not available for this zonelist, does
1871 * nothing and returns NULL.
1873 * If the fullzones BITMAP in the zonelist cache is stale (more than
1874 * a second since last zap'd) then we zap it out (clear its bits.)
1876 * We hold off even calling zlc_setup, until after we've checked the
1877 * first zone in the zonelist, on the theory that most allocations will
1878 * be satisfied from that first zone, so best to examine that zone as
1879 * quickly as we can.
1881 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1883 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1884 nodemask_t *allowednodes; /* zonelist_cache approximation */
1886 zlc = zonelist->zlcache_ptr;
1890 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1891 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1892 zlc->last_full_zap = jiffies;
1895 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1896 &cpuset_current_mems_allowed :
1897 &node_states[N_MEMORY];
1898 return allowednodes;
1902 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1903 * if it is worth looking at further for free memory:
1904 * 1) Check that the zone isn't thought to be full (doesn't have its
1905 * bit set in the zonelist_cache fullzones BITMAP).
1906 * 2) Check that the zones node (obtained from the zonelist_cache
1907 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1908 * Return true (non-zero) if zone is worth looking at further, or
1909 * else return false (zero) if it is not.
1911 * This check -ignores- the distinction between various watermarks,
1912 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1913 * found to be full for any variation of these watermarks, it will
1914 * be considered full for up to one second by all requests, unless
1915 * we are so low on memory on all allowed nodes that we are forced
1916 * into the second scan of the zonelist.
1918 * In the second scan we ignore this zonelist cache and exactly
1919 * apply the watermarks to all zones, even it is slower to do so.
1920 * We are low on memory in the second scan, and should leave no stone
1921 * unturned looking for a free page.
1923 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1924 nodemask_t *allowednodes)
1926 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1927 int i; /* index of *z in zonelist zones */
1928 int n; /* node that zone *z is on */
1930 zlc = zonelist->zlcache_ptr;
1934 i = z - zonelist->_zonerefs;
1937 /* This zone is worth trying if it is allowed but not full */
1938 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1942 * Given 'z' scanning a zonelist, set the corresponding bit in
1943 * zlc->fullzones, so that subsequent attempts to allocate a page
1944 * from that zone don't waste time re-examining it.
1946 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1948 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1949 int i; /* index of *z in zonelist zones */
1951 zlc = zonelist->zlcache_ptr;
1955 i = z - zonelist->_zonerefs;
1957 set_bit(i, zlc->fullzones);
1961 * clear all zones full, called after direct reclaim makes progress so that
1962 * a zone that was recently full is not skipped over for up to a second
1964 static void zlc_clear_zones_full(struct zonelist *zonelist)
1966 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1968 zlc = zonelist->zlcache_ptr;
1972 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1975 static bool zone_local(struct zone *local_zone, struct zone *zone)
1977 return local_zone->node == zone->node;
1980 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1982 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1986 #else /* CONFIG_NUMA */
1988 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1993 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1994 nodemask_t *allowednodes)
1999 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2003 static void zlc_clear_zones_full(struct zonelist *zonelist)
2007 static bool zone_local(struct zone *local_zone, struct zone *zone)
2012 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2017 #endif /* CONFIG_NUMA */
2019 static void reset_alloc_batches(struct zone *preferred_zone)
2021 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2024 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2025 high_wmark_pages(zone) - low_wmark_pages(zone) -
2026 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2027 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2028 } while (zone++ != preferred_zone);
2032 * get_page_from_freelist goes through the zonelist trying to allocate
2035 static struct page *
2036 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
2037 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
2038 struct zone *preferred_zone, int classzone_idx, int migratetype)
2041 struct page *page = NULL;
2043 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2044 int zlc_active = 0; /* set if using zonelist_cache */
2045 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2046 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2047 (gfp_mask & __GFP_WRITE);
2048 int nr_fair_skipped = 0;
2049 bool zonelist_rescan;
2052 zonelist_rescan = false;
2055 * Scan zonelist, looking for a zone with enough free.
2056 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2058 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2059 high_zoneidx, nodemask) {
2062 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2063 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2065 if (cpusets_enabled() &&
2066 (alloc_flags & ALLOC_CPUSET) &&
2067 !cpuset_zone_allowed(zone, gfp_mask))
2070 * Distribute pages in proportion to the individual
2071 * zone size to ensure fair page aging. The zone a
2072 * page was allocated in should have no effect on the
2073 * time the page has in memory before being reclaimed.
2075 if (alloc_flags & ALLOC_FAIR) {
2076 if (!zone_local(preferred_zone, zone))
2078 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2084 * When allocating a page cache page for writing, we
2085 * want to get it from a zone that is within its dirty
2086 * limit, such that no single zone holds more than its
2087 * proportional share of globally allowed dirty pages.
2088 * The dirty limits take into account the zone's
2089 * lowmem reserves and high watermark so that kswapd
2090 * should be able to balance it without having to
2091 * write pages from its LRU list.
2093 * This may look like it could increase pressure on
2094 * lower zones by failing allocations in higher zones
2095 * before they are full. But the pages that do spill
2096 * over are limited as the lower zones are protected
2097 * by this very same mechanism. It should not become
2098 * a practical burden to them.
2100 * XXX: For now, allow allocations to potentially
2101 * exceed the per-zone dirty limit in the slowpath
2102 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2103 * which is important when on a NUMA setup the allowed
2104 * zones are together not big enough to reach the
2105 * global limit. The proper fix for these situations
2106 * will require awareness of zones in the
2107 * dirty-throttling and the flusher threads.
2109 if (consider_zone_dirty && !zone_dirty_ok(zone))
2112 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2113 if (!zone_watermark_ok(zone, order, mark,
2114 classzone_idx, alloc_flags)) {
2117 /* Checked here to keep the fast path fast */
2118 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2119 if (alloc_flags & ALLOC_NO_WATERMARKS)
2122 if (IS_ENABLED(CONFIG_NUMA) &&
2123 !did_zlc_setup && nr_online_nodes > 1) {
2125 * we do zlc_setup if there are multiple nodes
2126 * and before considering the first zone allowed
2129 allowednodes = zlc_setup(zonelist, alloc_flags);
2134 if (zone_reclaim_mode == 0 ||
2135 !zone_allows_reclaim(preferred_zone, zone))
2136 goto this_zone_full;
2139 * As we may have just activated ZLC, check if the first
2140 * eligible zone has failed zone_reclaim recently.
2142 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2143 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2146 ret = zone_reclaim(zone, gfp_mask, order);
2148 case ZONE_RECLAIM_NOSCAN:
2151 case ZONE_RECLAIM_FULL:
2152 /* scanned but unreclaimable */
2155 /* did we reclaim enough */
2156 if (zone_watermark_ok(zone, order, mark,
2157 classzone_idx, alloc_flags))
2161 * Failed to reclaim enough to meet watermark.
2162 * Only mark the zone full if checking the min
2163 * watermark or if we failed to reclaim just
2164 * 1<<order pages or else the page allocator
2165 * fastpath will prematurely mark zones full
2166 * when the watermark is between the low and
2169 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2170 ret == ZONE_RECLAIM_SOME)
2171 goto this_zone_full;
2178 page = buffered_rmqueue(preferred_zone, zone, order,
2179 gfp_mask, migratetype);
2183 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2184 zlc_mark_zone_full(zonelist, z);
2189 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2190 * necessary to allocate the page. The expectation is
2191 * that the caller is taking steps that will free more
2192 * memory. The caller should avoid the page being used
2193 * for !PFMEMALLOC purposes.
2195 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2200 * The first pass makes sure allocations are spread fairly within the
2201 * local node. However, the local node might have free pages left
2202 * after the fairness batches are exhausted, and remote zones haven't
2203 * even been considered yet. Try once more without fairness, and
2204 * include remote zones now, before entering the slowpath and waking
2205 * kswapd: prefer spilling to a remote zone over swapping locally.
2207 if (alloc_flags & ALLOC_FAIR) {
2208 alloc_flags &= ~ALLOC_FAIR;
2209 if (nr_fair_skipped) {
2210 zonelist_rescan = true;
2211 reset_alloc_batches(preferred_zone);
2213 if (nr_online_nodes > 1)
2214 zonelist_rescan = true;
2217 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2218 /* Disable zlc cache for second zonelist scan */
2220 zonelist_rescan = true;
2223 if (zonelist_rescan)
2230 * Large machines with many possible nodes should not always dump per-node
2231 * meminfo in irq context.
2233 static inline bool should_suppress_show_mem(void)
2238 ret = in_interrupt();
2243 static DEFINE_RATELIMIT_STATE(nopage_rs,
2244 DEFAULT_RATELIMIT_INTERVAL,
2245 DEFAULT_RATELIMIT_BURST);
2247 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2249 unsigned int filter = SHOW_MEM_FILTER_NODES;
2251 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2252 debug_guardpage_minorder() > 0)
2256 * This documents exceptions given to allocations in certain
2257 * contexts that are allowed to allocate outside current's set
2260 if (!(gfp_mask & __GFP_NOMEMALLOC))
2261 if (test_thread_flag(TIF_MEMDIE) ||
2262 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2263 filter &= ~SHOW_MEM_FILTER_NODES;
2264 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2265 filter &= ~SHOW_MEM_FILTER_NODES;
2268 struct va_format vaf;
2271 va_start(args, fmt);
2276 pr_warn("%pV", &vaf);
2281 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2282 current->comm, order, gfp_mask);
2285 if (!should_suppress_show_mem())
2290 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2291 unsigned long did_some_progress,
2292 unsigned long pages_reclaimed)
2294 /* Do not loop if specifically requested */
2295 if (gfp_mask & __GFP_NORETRY)
2298 /* Always retry if specifically requested */
2299 if (gfp_mask & __GFP_NOFAIL)
2303 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2304 * making forward progress without invoking OOM. Suspend also disables
2305 * storage devices so kswapd will not help. Bail if we are suspending.
2307 if (!did_some_progress && pm_suspended_storage())
2311 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2312 * means __GFP_NOFAIL, but that may not be true in other
2315 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2319 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2320 * specified, then we retry until we no longer reclaim any pages
2321 * (above), or we've reclaimed an order of pages at least as
2322 * large as the allocation's order. In both cases, if the
2323 * allocation still fails, we stop retrying.
2325 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2331 static inline struct page *
2332 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2333 struct zonelist *zonelist, enum zone_type high_zoneidx,
2334 nodemask_t *nodemask, struct zone *preferred_zone,
2335 int classzone_idx, int migratetype)
2339 /* Acquire the per-zone oom lock for each zone */
2340 if (!oom_zonelist_trylock(zonelist, gfp_mask)) {
2341 schedule_timeout_uninterruptible(1);
2346 * PM-freezer should be notified that there might be an OOM killer on
2347 * its way to kill and wake somebody up. This is too early and we might
2348 * end up not killing anything but false positives are acceptable.
2349 * See freeze_processes.
2354 * Go through the zonelist yet one more time, keep very high watermark
2355 * here, this is only to catch a parallel oom killing, we must fail if
2356 * we're still under heavy pressure.
2358 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2359 order, zonelist, high_zoneidx,
2360 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2361 preferred_zone, classzone_idx, migratetype);
2365 if (!(gfp_mask & __GFP_NOFAIL)) {
2366 /* The OOM killer will not help higher order allocs */
2367 if (order > PAGE_ALLOC_COSTLY_ORDER)
2369 /* The OOM killer does not needlessly kill tasks for lowmem */
2370 if (high_zoneidx < ZONE_NORMAL)
2373 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2374 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2375 * The caller should handle page allocation failure by itself if
2376 * it specifies __GFP_THISNODE.
2377 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2379 if (gfp_mask & __GFP_THISNODE)
2382 /* Exhausted what can be done so it's blamo time */
2383 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2386 oom_zonelist_unlock(zonelist, gfp_mask);
2390 #ifdef CONFIG_COMPACTION
2391 /* Try memory compaction for high-order allocations before reclaim */
2392 static struct page *
2393 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2394 struct zonelist *zonelist, enum zone_type high_zoneidx,
2395 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2396 int classzone_idx, int migratetype, enum migrate_mode mode,
2397 int *contended_compaction, bool *deferred_compaction)
2399 unsigned long compact_result;
2405 current->flags |= PF_MEMALLOC;
2406 compact_result = try_to_compact_pages(zonelist, order, gfp_mask,
2408 contended_compaction,
2409 alloc_flags, classzone_idx);
2410 current->flags &= ~PF_MEMALLOC;
2412 switch (compact_result) {
2413 case COMPACT_DEFERRED:
2414 *deferred_compaction = true;
2416 case COMPACT_SKIPPED:
2423 * At least in one zone compaction wasn't deferred or skipped, so let's
2424 * count a compaction stall
2426 count_vm_event(COMPACTSTALL);
2428 page = get_page_from_freelist(gfp_mask, nodemask,
2429 order, zonelist, high_zoneidx,
2430 alloc_flags & ~ALLOC_NO_WATERMARKS,
2431 preferred_zone, classzone_idx, migratetype);
2434 struct zone *zone = page_zone(page);
2436 zone->compact_blockskip_flush = false;
2437 compaction_defer_reset(zone, order, true);
2438 count_vm_event(COMPACTSUCCESS);
2443 * It's bad if compaction run occurs and fails. The most likely reason
2444 * is that pages exist, but not enough to satisfy watermarks.
2446 count_vm_event(COMPACTFAIL);
2453 static inline struct page *
2454 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2455 struct zonelist *zonelist, enum zone_type high_zoneidx,
2456 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2457 int classzone_idx, int migratetype, enum migrate_mode mode,
2458 int *contended_compaction, bool *deferred_compaction)
2462 #endif /* CONFIG_COMPACTION */
2464 /* Perform direct synchronous page reclaim */
2466 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2467 nodemask_t *nodemask)
2469 struct reclaim_state reclaim_state;
2474 /* We now go into synchronous reclaim */
2475 cpuset_memory_pressure_bump();
2476 current->flags |= PF_MEMALLOC;
2477 lockdep_set_current_reclaim_state(gfp_mask);
2478 reclaim_state.reclaimed_slab = 0;
2479 current->reclaim_state = &reclaim_state;
2481 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2483 current->reclaim_state = NULL;
2484 lockdep_clear_current_reclaim_state();
2485 current->flags &= ~PF_MEMALLOC;
2492 /* The really slow allocator path where we enter direct reclaim */
2493 static inline struct page *
2494 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2495 struct zonelist *zonelist, enum zone_type high_zoneidx,
2496 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2497 int classzone_idx, int migratetype, unsigned long *did_some_progress)
2499 struct page *page = NULL;
2500 bool drained = false;
2502 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2504 if (unlikely(!(*did_some_progress)))
2507 /* After successful reclaim, reconsider all zones for allocation */
2508 if (IS_ENABLED(CONFIG_NUMA))
2509 zlc_clear_zones_full(zonelist);
2512 page = get_page_from_freelist(gfp_mask, nodemask, order,
2513 zonelist, high_zoneidx,
2514 alloc_flags & ~ALLOC_NO_WATERMARKS,
2515 preferred_zone, classzone_idx,
2519 * If an allocation failed after direct reclaim, it could be because
2520 * pages are pinned on the per-cpu lists. Drain them and try again
2522 if (!page && !drained) {
2523 drain_all_pages(NULL);
2532 * This is called in the allocator slow-path if the allocation request is of
2533 * sufficient urgency to ignore watermarks and take other desperate measures
2535 static inline struct page *
2536 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2537 struct zonelist *zonelist, enum zone_type high_zoneidx,
2538 nodemask_t *nodemask, struct zone *preferred_zone,
2539 int classzone_idx, int migratetype)
2544 page = get_page_from_freelist(gfp_mask, nodemask, order,
2545 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2546 preferred_zone, classzone_idx, migratetype);
2548 if (!page && gfp_mask & __GFP_NOFAIL)
2549 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2550 } while (!page && (gfp_mask & __GFP_NOFAIL));
2555 static void wake_all_kswapds(unsigned int order,
2556 struct zonelist *zonelist,
2557 enum zone_type high_zoneidx,
2558 struct zone *preferred_zone,
2559 nodemask_t *nodemask)
2564 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2565 high_zoneidx, nodemask)
2566 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2570 gfp_to_alloc_flags(gfp_t gfp_mask)
2572 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2573 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2575 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2576 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2579 * The caller may dip into page reserves a bit more if the caller
2580 * cannot run direct reclaim, or if the caller has realtime scheduling
2581 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2582 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2584 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2588 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2589 * if it can't schedule.
2591 if (!(gfp_mask & __GFP_NOMEMALLOC))
2592 alloc_flags |= ALLOC_HARDER;
2594 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2595 * comment for __cpuset_node_allowed().
2597 alloc_flags &= ~ALLOC_CPUSET;
2598 } else if (unlikely(rt_task(current)) && !in_interrupt())
2599 alloc_flags |= ALLOC_HARDER;
2601 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2602 if (gfp_mask & __GFP_MEMALLOC)
2603 alloc_flags |= ALLOC_NO_WATERMARKS;
2604 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2605 alloc_flags |= ALLOC_NO_WATERMARKS;
2606 else if (!in_interrupt() &&
2607 ((current->flags & PF_MEMALLOC) ||
2608 unlikely(test_thread_flag(TIF_MEMDIE))))
2609 alloc_flags |= ALLOC_NO_WATERMARKS;
2612 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2613 alloc_flags |= ALLOC_CMA;
2618 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2620 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2623 static inline struct page *
2624 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2625 struct zonelist *zonelist, enum zone_type high_zoneidx,
2626 nodemask_t *nodemask, struct zone *preferred_zone,
2627 int classzone_idx, int migratetype)
2629 const gfp_t wait = gfp_mask & __GFP_WAIT;
2630 struct page *page = NULL;
2632 unsigned long pages_reclaimed = 0;
2633 unsigned long did_some_progress;
2634 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2635 bool deferred_compaction = false;
2636 int contended_compaction = COMPACT_CONTENDED_NONE;
2639 * In the slowpath, we sanity check order to avoid ever trying to
2640 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2641 * be using allocators in order of preference for an area that is
2644 if (order >= MAX_ORDER) {
2645 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2650 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2651 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2652 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2653 * using a larger set of nodes after it has established that the
2654 * allowed per node queues are empty and that nodes are
2657 if (IS_ENABLED(CONFIG_NUMA) &&
2658 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2662 if (!(gfp_mask & __GFP_NO_KSWAPD))
2663 wake_all_kswapds(order, zonelist, high_zoneidx,
2664 preferred_zone, nodemask);
2667 * OK, we're below the kswapd watermark and have kicked background
2668 * reclaim. Now things get more complex, so set up alloc_flags according
2669 * to how we want to proceed.
2671 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2674 * Find the true preferred zone if the allocation is unconstrained by
2677 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) {
2678 struct zoneref *preferred_zoneref;
2679 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2680 NULL, &preferred_zone);
2681 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2685 /* This is the last chance, in general, before the goto nopage. */
2686 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2687 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2688 preferred_zone, classzone_idx, migratetype);
2692 /* Allocate without watermarks if the context allows */
2693 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2695 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2696 * the allocation is high priority and these type of
2697 * allocations are system rather than user orientated
2699 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2701 page = __alloc_pages_high_priority(gfp_mask, order,
2702 zonelist, high_zoneidx, nodemask,
2703 preferred_zone, classzone_idx, migratetype);
2709 /* Atomic allocations - we can't balance anything */
2712 * All existing users of the deprecated __GFP_NOFAIL are
2713 * blockable, so warn of any new users that actually allow this
2714 * type of allocation to fail.
2716 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2720 /* Avoid recursion of direct reclaim */
2721 if (current->flags & PF_MEMALLOC)
2724 /* Avoid allocations with no watermarks from looping endlessly */
2725 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2729 * Try direct compaction. The first pass is asynchronous. Subsequent
2730 * attempts after direct reclaim are synchronous
2732 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2733 high_zoneidx, nodemask, alloc_flags,
2735 classzone_idx, migratetype,
2736 migration_mode, &contended_compaction,
2737 &deferred_compaction);
2741 /* Checks for THP-specific high-order allocations */
2742 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2744 * If compaction is deferred for high-order allocations, it is
2745 * because sync compaction recently failed. If this is the case
2746 * and the caller requested a THP allocation, we do not want
2747 * to heavily disrupt the system, so we fail the allocation
2748 * instead of entering direct reclaim.
2750 if (deferred_compaction)
2754 * In all zones where compaction was attempted (and not
2755 * deferred or skipped), lock contention has been detected.
2756 * For THP allocation we do not want to disrupt the others
2757 * so we fallback to base pages instead.
2759 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2763 * If compaction was aborted due to need_resched(), we do not
2764 * want to further increase allocation latency, unless it is
2765 * khugepaged trying to collapse.
2767 if (contended_compaction == COMPACT_CONTENDED_SCHED
2768 && !(current->flags & PF_KTHREAD))
2773 * It can become very expensive to allocate transparent hugepages at
2774 * fault, so use asynchronous memory compaction for THP unless it is
2775 * khugepaged trying to collapse.
2777 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2778 (current->flags & PF_KTHREAD))
2779 migration_mode = MIGRATE_SYNC_LIGHT;
2781 /* Try direct reclaim and then allocating */
2782 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2783 zonelist, high_zoneidx,
2785 alloc_flags, preferred_zone,
2786 classzone_idx, migratetype,
2787 &did_some_progress);
2792 * If we failed to make any progress reclaiming, then we are
2793 * running out of options and have to consider going OOM
2795 if (!did_some_progress) {
2796 if (oom_gfp_allowed(gfp_mask)) {
2797 if (oom_killer_disabled)
2799 /* Coredumps can quickly deplete all memory reserves */
2800 if ((current->flags & PF_DUMPCORE) &&
2801 !(gfp_mask & __GFP_NOFAIL))
2803 page = __alloc_pages_may_oom(gfp_mask, order,
2804 zonelist, high_zoneidx,
2805 nodemask, preferred_zone,
2806 classzone_idx, migratetype);
2810 if (!(gfp_mask & __GFP_NOFAIL)) {
2812 * The oom killer is not called for high-order
2813 * allocations that may fail, so if no progress
2814 * is being made, there are no other options and
2815 * retrying is unlikely to help.
2817 if (order > PAGE_ALLOC_COSTLY_ORDER)
2820 * The oom killer is not called for lowmem
2821 * allocations to prevent needlessly killing
2824 if (high_zoneidx < ZONE_NORMAL)
2832 /* Check if we should retry the allocation */
2833 pages_reclaimed += did_some_progress;
2834 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2836 /* Wait for some write requests to complete then retry */
2837 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2841 * High-order allocations do not necessarily loop after
2842 * direct reclaim and reclaim/compaction depends on compaction
2843 * being called after reclaim so call directly if necessary
2845 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2846 high_zoneidx, nodemask, alloc_flags,
2848 classzone_idx, migratetype,
2849 migration_mode, &contended_compaction,
2850 &deferred_compaction);
2856 warn_alloc_failed(gfp_mask, order, NULL);
2859 if (kmemcheck_enabled)
2860 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2866 * This is the 'heart' of the zoned buddy allocator.
2869 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2870 struct zonelist *zonelist, nodemask_t *nodemask)
2872 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2873 struct zone *preferred_zone;
2874 struct zoneref *preferred_zoneref;
2875 struct page *page = NULL;
2876 int migratetype = gfpflags_to_migratetype(gfp_mask);
2877 unsigned int cpuset_mems_cookie;
2878 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2881 gfp_mask &= gfp_allowed_mask;
2883 lockdep_trace_alloc(gfp_mask);
2885 might_sleep_if(gfp_mask & __GFP_WAIT);
2887 if (should_fail_alloc_page(gfp_mask, order))
2891 * Check the zones suitable for the gfp_mask contain at least one
2892 * valid zone. It's possible to have an empty zonelist as a result
2893 * of GFP_THISNODE and a memoryless node
2895 if (unlikely(!zonelist->_zonerefs->zone))
2898 if (IS_ENABLED(CONFIG_CMA) && migratetype == MIGRATE_MOVABLE)
2899 alloc_flags |= ALLOC_CMA;
2902 cpuset_mems_cookie = read_mems_allowed_begin();
2904 /* The preferred zone is used for statistics later */
2905 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2906 nodemask ? : &cpuset_current_mems_allowed,
2908 if (!preferred_zone)
2910 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2912 /* First allocation attempt */
2913 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2914 zonelist, high_zoneidx, alloc_flags,
2915 preferred_zone, classzone_idx, migratetype);
2916 if (unlikely(!page)) {
2918 * Runtime PM, block IO and its error handling path
2919 * can deadlock because I/O on the device might not
2922 gfp_mask = memalloc_noio_flags(gfp_mask);
2923 page = __alloc_pages_slowpath(gfp_mask, order,
2924 zonelist, high_zoneidx, nodemask,
2925 preferred_zone, classzone_idx, migratetype);
2928 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2932 * When updating a task's mems_allowed, it is possible to race with
2933 * parallel threads in such a way that an allocation can fail while
2934 * the mask is being updated. If a page allocation is about to fail,
2935 * check if the cpuset changed during allocation and if so, retry.
2937 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2942 EXPORT_SYMBOL(__alloc_pages_nodemask);
2945 * Common helper functions.
2947 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2952 * __get_free_pages() returns a 32-bit address, which cannot represent
2955 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2957 page = alloc_pages(gfp_mask, order);
2960 return (unsigned long) page_address(page);
2962 EXPORT_SYMBOL(__get_free_pages);
2964 unsigned long get_zeroed_page(gfp_t gfp_mask)
2966 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2968 EXPORT_SYMBOL(get_zeroed_page);
2970 void __free_pages(struct page *page, unsigned int order)
2972 if (put_page_testzero(page)) {
2974 free_hot_cold_page(page, false);
2976 __free_pages_ok(page, order);
2980 EXPORT_SYMBOL(__free_pages);
2982 void free_pages(unsigned long addr, unsigned int order)
2985 VM_BUG_ON(!virt_addr_valid((void *)addr));
2986 __free_pages(virt_to_page((void *)addr), order);
2990 EXPORT_SYMBOL(free_pages);
2993 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2994 * of the current memory cgroup.
2996 * It should be used when the caller would like to use kmalloc, but since the
2997 * allocation is large, it has to fall back to the page allocator.
2999 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3002 struct mem_cgroup *memcg = NULL;
3004 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3006 page = alloc_pages(gfp_mask, order);
3007 memcg_kmem_commit_charge(page, memcg, order);
3011 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3014 struct mem_cgroup *memcg = NULL;
3016 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3018 page = alloc_pages_node(nid, gfp_mask, order);
3019 memcg_kmem_commit_charge(page, memcg, order);
3024 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3027 void __free_kmem_pages(struct page *page, unsigned int order)
3029 memcg_kmem_uncharge_pages(page, order);
3030 __free_pages(page, order);
3033 void free_kmem_pages(unsigned long addr, unsigned int order)
3036 VM_BUG_ON(!virt_addr_valid((void *)addr));
3037 __free_kmem_pages(virt_to_page((void *)addr), order);
3041 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3044 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3045 unsigned long used = addr + PAGE_ALIGN(size);
3047 split_page(virt_to_page((void *)addr), order);
3048 while (used < alloc_end) {
3053 return (void *)addr;
3057 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3058 * @size: the number of bytes to allocate
3059 * @gfp_mask: GFP flags for the allocation
3061 * This function is similar to alloc_pages(), except that it allocates the
3062 * minimum number of pages to satisfy the request. alloc_pages() can only
3063 * allocate memory in power-of-two pages.
3065 * This function is also limited by MAX_ORDER.
3067 * Memory allocated by this function must be released by free_pages_exact().
3069 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3071 unsigned int order = get_order(size);
3074 addr = __get_free_pages(gfp_mask, order);
3075 return make_alloc_exact(addr, order, size);
3077 EXPORT_SYMBOL(alloc_pages_exact);
3080 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3082 * @nid: the preferred node ID where memory should be allocated
3083 * @size: the number of bytes to allocate
3084 * @gfp_mask: GFP flags for the allocation
3086 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3088 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3091 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3093 unsigned order = get_order(size);
3094 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3097 return make_alloc_exact((unsigned long)page_address(p), order, size);
3101 * free_pages_exact - release memory allocated via alloc_pages_exact()
3102 * @virt: the value returned by alloc_pages_exact.
3103 * @size: size of allocation, same value as passed to alloc_pages_exact().
3105 * Release the memory allocated by a previous call to alloc_pages_exact.
3107 void free_pages_exact(void *virt, size_t size)
3109 unsigned long addr = (unsigned long)virt;
3110 unsigned long end = addr + PAGE_ALIGN(size);
3112 while (addr < end) {
3117 EXPORT_SYMBOL(free_pages_exact);
3120 * nr_free_zone_pages - count number of pages beyond high watermark
3121 * @offset: The zone index of the highest zone
3123 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3124 * high watermark within all zones at or below a given zone index. For each
3125 * zone, the number of pages is calculated as:
3126 * managed_pages - high_pages
3128 static unsigned long nr_free_zone_pages(int offset)
3133 /* Just pick one node, since fallback list is circular */
3134 unsigned long sum = 0;
3136 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3138 for_each_zone_zonelist(zone, z, zonelist, offset) {
3139 unsigned long size = zone->managed_pages;
3140 unsigned long high = high_wmark_pages(zone);
3149 * nr_free_buffer_pages - count number of pages beyond high watermark
3151 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3152 * watermark within ZONE_DMA and ZONE_NORMAL.
3154 unsigned long nr_free_buffer_pages(void)
3156 return nr_free_zone_pages(gfp_zone(GFP_USER));
3158 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3161 * nr_free_pagecache_pages - count number of pages beyond high watermark
3163 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3164 * high watermark within all zones.
3166 unsigned long nr_free_pagecache_pages(void)
3168 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3171 static inline void show_node(struct zone *zone)
3173 if (IS_ENABLED(CONFIG_NUMA))
3174 printk("Node %d ", zone_to_nid(zone));
3177 void si_meminfo(struct sysinfo *val)
3179 val->totalram = totalram_pages;
3180 val->sharedram = global_page_state(NR_SHMEM);
3181 val->freeram = global_page_state(NR_FREE_PAGES);
3182 val->bufferram = nr_blockdev_pages();
3183 val->totalhigh = totalhigh_pages;
3184 val->freehigh = nr_free_highpages();
3185 val->mem_unit = PAGE_SIZE;
3188 EXPORT_SYMBOL(si_meminfo);
3191 void si_meminfo_node(struct sysinfo *val, int nid)
3193 int zone_type; /* needs to be signed */
3194 unsigned long managed_pages = 0;
3195 pg_data_t *pgdat = NODE_DATA(nid);
3197 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3198 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3199 val->totalram = managed_pages;
3200 val->sharedram = node_page_state(nid, NR_SHMEM);
3201 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3202 #ifdef CONFIG_HIGHMEM
3203 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3204 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3210 val->mem_unit = PAGE_SIZE;
3215 * Determine whether the node should be displayed or not, depending on whether
3216 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3218 bool skip_free_areas_node(unsigned int flags, int nid)
3221 unsigned int cpuset_mems_cookie;
3223 if (!(flags & SHOW_MEM_FILTER_NODES))
3227 cpuset_mems_cookie = read_mems_allowed_begin();
3228 ret = !node_isset(nid, cpuset_current_mems_allowed);
3229 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3234 #define K(x) ((x) << (PAGE_SHIFT-10))
3236 static void show_migration_types(unsigned char type)
3238 static const char types[MIGRATE_TYPES] = {
3239 [MIGRATE_UNMOVABLE] = 'U',
3240 [MIGRATE_RECLAIMABLE] = 'E',
3241 [MIGRATE_MOVABLE] = 'M',
3242 [MIGRATE_RESERVE] = 'R',
3244 [MIGRATE_CMA] = 'C',
3246 #ifdef CONFIG_MEMORY_ISOLATION
3247 [MIGRATE_ISOLATE] = 'I',
3250 char tmp[MIGRATE_TYPES + 1];
3254 for (i = 0; i < MIGRATE_TYPES; i++) {
3255 if (type & (1 << i))
3260 printk("(%s) ", tmp);
3264 * Show free area list (used inside shift_scroll-lock stuff)
3265 * We also calculate the percentage fragmentation. We do this by counting the
3266 * memory on each free list with the exception of the first item on the list.
3267 * Suppresses nodes that are not allowed by current's cpuset if
3268 * SHOW_MEM_FILTER_NODES is passed.
3270 void show_free_areas(unsigned int filter)
3275 for_each_populated_zone(zone) {
3276 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3279 printk("%s per-cpu:\n", zone->name);
3281 for_each_online_cpu(cpu) {
3282 struct per_cpu_pageset *pageset;
3284 pageset = per_cpu_ptr(zone->pageset, cpu);
3286 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3287 cpu, pageset->pcp.high,
3288 pageset->pcp.batch, pageset->pcp.count);
3292 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3293 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3295 " dirty:%lu writeback:%lu unstable:%lu\n"
3296 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3297 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3299 global_page_state(NR_ACTIVE_ANON),
3300 global_page_state(NR_INACTIVE_ANON),
3301 global_page_state(NR_ISOLATED_ANON),
3302 global_page_state(NR_ACTIVE_FILE),
3303 global_page_state(NR_INACTIVE_FILE),
3304 global_page_state(NR_ISOLATED_FILE),
3305 global_page_state(NR_UNEVICTABLE),
3306 global_page_state(NR_FILE_DIRTY),
3307 global_page_state(NR_WRITEBACK),
3308 global_page_state(NR_UNSTABLE_NFS),
3309 global_page_state(NR_FREE_PAGES),
3310 global_page_state(NR_SLAB_RECLAIMABLE),
3311 global_page_state(NR_SLAB_UNRECLAIMABLE),
3312 global_page_state(NR_FILE_MAPPED),
3313 global_page_state(NR_SHMEM),
3314 global_page_state(NR_PAGETABLE),
3315 global_page_state(NR_BOUNCE),
3316 global_page_state(NR_FREE_CMA_PAGES));
3318 for_each_populated_zone(zone) {
3321 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3329 " active_anon:%lukB"
3330 " inactive_anon:%lukB"
3331 " active_file:%lukB"
3332 " inactive_file:%lukB"
3333 " unevictable:%lukB"
3334 " isolated(anon):%lukB"
3335 " isolated(file):%lukB"
3343 " slab_reclaimable:%lukB"
3344 " slab_unreclaimable:%lukB"
3345 " kernel_stack:%lukB"
3350 " writeback_tmp:%lukB"
3351 " pages_scanned:%lu"
3352 " all_unreclaimable? %s"
3355 K(zone_page_state(zone, NR_FREE_PAGES)),
3356 K(min_wmark_pages(zone)),
3357 K(low_wmark_pages(zone)),
3358 K(high_wmark_pages(zone)),
3359 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3360 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3361 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3362 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3363 K(zone_page_state(zone, NR_UNEVICTABLE)),
3364 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3365 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3366 K(zone->present_pages),
3367 K(zone->managed_pages),
3368 K(zone_page_state(zone, NR_MLOCK)),
3369 K(zone_page_state(zone, NR_FILE_DIRTY)),
3370 K(zone_page_state(zone, NR_WRITEBACK)),
3371 K(zone_page_state(zone, NR_FILE_MAPPED)),
3372 K(zone_page_state(zone, NR_SHMEM)),
3373 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3374 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3375 zone_page_state(zone, NR_KERNEL_STACK) *
3377 K(zone_page_state(zone, NR_PAGETABLE)),
3378 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3379 K(zone_page_state(zone, NR_BOUNCE)),
3380 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3381 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3382 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3383 (!zone_reclaimable(zone) ? "yes" : "no")
3385 printk("lowmem_reserve[]:");
3386 for (i = 0; i < MAX_NR_ZONES; i++)
3387 printk(" %ld", zone->lowmem_reserve[i]);
3391 for_each_populated_zone(zone) {
3392 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3393 unsigned char types[MAX_ORDER];
3395 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3398 printk("%s: ", zone->name);
3400 spin_lock_irqsave(&zone->lock, flags);
3401 for (order = 0; order < MAX_ORDER; order++) {
3402 struct free_area *area = &zone->free_area[order];
3405 nr[order] = area->nr_free;
3406 total += nr[order] << order;
3409 for (type = 0; type < MIGRATE_TYPES; type++) {
3410 if (!list_empty(&area->free_list[type]))
3411 types[order] |= 1 << type;
3414 spin_unlock_irqrestore(&zone->lock, flags);
3415 for (order = 0; order < MAX_ORDER; order++) {
3416 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3418 show_migration_types(types[order]);
3420 printk("= %lukB\n", K(total));
3423 hugetlb_show_meminfo();
3425 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3427 show_swap_cache_info();
3430 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3432 zoneref->zone = zone;
3433 zoneref->zone_idx = zone_idx(zone);
3437 * Builds allocation fallback zone lists.
3439 * Add all populated zones of a node to the zonelist.
3441 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3445 enum zone_type zone_type = MAX_NR_ZONES;
3449 zone = pgdat->node_zones + zone_type;
3450 if (populated_zone(zone)) {
3451 zoneref_set_zone(zone,
3452 &zonelist->_zonerefs[nr_zones++]);
3453 check_highest_zone(zone_type);
3455 } while (zone_type);
3463 * 0 = automatic detection of better ordering.
3464 * 1 = order by ([node] distance, -zonetype)
3465 * 2 = order by (-zonetype, [node] distance)
3467 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3468 * the same zonelist. So only NUMA can configure this param.
3470 #define ZONELIST_ORDER_DEFAULT 0
3471 #define ZONELIST_ORDER_NODE 1
3472 #define ZONELIST_ORDER_ZONE 2
3474 /* zonelist order in the kernel.
3475 * set_zonelist_order() will set this to NODE or ZONE.
3477 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3478 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3482 /* The value user specified ....changed by config */
3483 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3484 /* string for sysctl */
3485 #define NUMA_ZONELIST_ORDER_LEN 16
3486 char numa_zonelist_order[16] = "default";
3489 * interface for configure zonelist ordering.
3490 * command line option "numa_zonelist_order"
3491 * = "[dD]efault - default, automatic configuration.
3492 * = "[nN]ode - order by node locality, then by zone within node
3493 * = "[zZ]one - order by zone, then by locality within zone
3496 static int __parse_numa_zonelist_order(char *s)
3498 if (*s == 'd' || *s == 'D') {
3499 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3500 } else if (*s == 'n' || *s == 'N') {
3501 user_zonelist_order = ZONELIST_ORDER_NODE;
3502 } else if (*s == 'z' || *s == 'Z') {
3503 user_zonelist_order = ZONELIST_ORDER_ZONE;
3506 "Ignoring invalid numa_zonelist_order value: "
3513 static __init int setup_numa_zonelist_order(char *s)
3520 ret = __parse_numa_zonelist_order(s);
3522 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3526 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3529 * sysctl handler for numa_zonelist_order
3531 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3532 void __user *buffer, size_t *length,
3535 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3537 static DEFINE_MUTEX(zl_order_mutex);
3539 mutex_lock(&zl_order_mutex);
3541 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3545 strcpy(saved_string, (char *)table->data);
3547 ret = proc_dostring(table, write, buffer, length, ppos);
3551 int oldval = user_zonelist_order;
3553 ret = __parse_numa_zonelist_order((char *)table->data);
3556 * bogus value. restore saved string
3558 strncpy((char *)table->data, saved_string,
3559 NUMA_ZONELIST_ORDER_LEN);
3560 user_zonelist_order = oldval;
3561 } else if (oldval != user_zonelist_order) {
3562 mutex_lock(&zonelists_mutex);
3563 build_all_zonelists(NULL, NULL);
3564 mutex_unlock(&zonelists_mutex);
3568 mutex_unlock(&zl_order_mutex);
3573 #define MAX_NODE_LOAD (nr_online_nodes)
3574 static int node_load[MAX_NUMNODES];
3577 * find_next_best_node - find the next node that should appear in a given node's fallback list
3578 * @node: node whose fallback list we're appending
3579 * @used_node_mask: nodemask_t of already used nodes
3581 * We use a number of factors to determine which is the next node that should
3582 * appear on a given node's fallback list. The node should not have appeared
3583 * already in @node's fallback list, and it should be the next closest node
3584 * according to the distance array (which contains arbitrary distance values
3585 * from each node to each node in the system), and should also prefer nodes
3586 * with no CPUs, since presumably they'll have very little allocation pressure
3587 * on them otherwise.
3588 * It returns -1 if no node is found.
3590 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3593 int min_val = INT_MAX;
3594 int best_node = NUMA_NO_NODE;
3595 const struct cpumask *tmp = cpumask_of_node(0);
3597 /* Use the local node if we haven't already */
3598 if (!node_isset(node, *used_node_mask)) {
3599 node_set(node, *used_node_mask);
3603 for_each_node_state(n, N_MEMORY) {
3605 /* Don't want a node to appear more than once */
3606 if (node_isset(n, *used_node_mask))
3609 /* Use the distance array to find the distance */
3610 val = node_distance(node, n);
3612 /* Penalize nodes under us ("prefer the next node") */
3615 /* Give preference to headless and unused nodes */
3616 tmp = cpumask_of_node(n);
3617 if (!cpumask_empty(tmp))
3618 val += PENALTY_FOR_NODE_WITH_CPUS;
3620 /* Slight preference for less loaded node */
3621 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3622 val += node_load[n];
3624 if (val < min_val) {
3631 node_set(best_node, *used_node_mask);
3638 * Build zonelists ordered by node and zones within node.
3639 * This results in maximum locality--normal zone overflows into local
3640 * DMA zone, if any--but risks exhausting DMA zone.
3642 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3645 struct zonelist *zonelist;
3647 zonelist = &pgdat->node_zonelists[0];
3648 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3650 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3651 zonelist->_zonerefs[j].zone = NULL;
3652 zonelist->_zonerefs[j].zone_idx = 0;
3656 * Build gfp_thisnode zonelists
3658 static void build_thisnode_zonelists(pg_data_t *pgdat)
3661 struct zonelist *zonelist;
3663 zonelist = &pgdat->node_zonelists[1];
3664 j = build_zonelists_node(pgdat, zonelist, 0);
3665 zonelist->_zonerefs[j].zone = NULL;
3666 zonelist->_zonerefs[j].zone_idx = 0;
3670 * Build zonelists ordered by zone and nodes within zones.
3671 * This results in conserving DMA zone[s] until all Normal memory is
3672 * exhausted, but results in overflowing to remote node while memory
3673 * may still exist in local DMA zone.
3675 static int node_order[MAX_NUMNODES];
3677 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3680 int zone_type; /* needs to be signed */
3682 struct zonelist *zonelist;
3684 zonelist = &pgdat->node_zonelists[0];
3686 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3687 for (j = 0; j < nr_nodes; j++) {
3688 node = node_order[j];
3689 z = &NODE_DATA(node)->node_zones[zone_type];
3690 if (populated_zone(z)) {
3692 &zonelist->_zonerefs[pos++]);
3693 check_highest_zone(zone_type);
3697 zonelist->_zonerefs[pos].zone = NULL;
3698 zonelist->_zonerefs[pos].zone_idx = 0;
3701 #if defined(CONFIG_64BIT)
3703 * Devices that require DMA32/DMA are relatively rare and do not justify a
3704 * penalty to every machine in case the specialised case applies. Default
3705 * to Node-ordering on 64-bit NUMA machines
3707 static int default_zonelist_order(void)
3709 return ZONELIST_ORDER_NODE;
3713 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3714 * by the kernel. If processes running on node 0 deplete the low memory zone
3715 * then reclaim will occur more frequency increasing stalls and potentially
3716 * be easier to OOM if a large percentage of the zone is under writeback or
3717 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3718 * Hence, default to zone ordering on 32-bit.
3720 static int default_zonelist_order(void)
3722 return ZONELIST_ORDER_ZONE;
3724 #endif /* CONFIG_64BIT */
3726 static void set_zonelist_order(void)
3728 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3729 current_zonelist_order = default_zonelist_order();
3731 current_zonelist_order = user_zonelist_order;
3734 static void build_zonelists(pg_data_t *pgdat)
3738 nodemask_t used_mask;
3739 int local_node, prev_node;
3740 struct zonelist *zonelist;
3741 int order = current_zonelist_order;
3743 /* initialize zonelists */
3744 for (i = 0; i < MAX_ZONELISTS; i++) {
3745 zonelist = pgdat->node_zonelists + i;
3746 zonelist->_zonerefs[0].zone = NULL;
3747 zonelist->_zonerefs[0].zone_idx = 0;
3750 /* NUMA-aware ordering of nodes */
3751 local_node = pgdat->node_id;
3752 load = nr_online_nodes;
3753 prev_node = local_node;
3754 nodes_clear(used_mask);
3756 memset(node_order, 0, sizeof(node_order));
3759 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3761 * We don't want to pressure a particular node.
3762 * So adding penalty to the first node in same
3763 * distance group to make it round-robin.
3765 if (node_distance(local_node, node) !=
3766 node_distance(local_node, prev_node))
3767 node_load[node] = load;
3771 if (order == ZONELIST_ORDER_NODE)
3772 build_zonelists_in_node_order(pgdat, node);
3774 node_order[j++] = node; /* remember order */
3777 if (order == ZONELIST_ORDER_ZONE) {
3778 /* calculate node order -- i.e., DMA last! */
3779 build_zonelists_in_zone_order(pgdat, j);
3782 build_thisnode_zonelists(pgdat);
3785 /* Construct the zonelist performance cache - see further mmzone.h */
3786 static void build_zonelist_cache(pg_data_t *pgdat)
3788 struct zonelist *zonelist;
3789 struct zonelist_cache *zlc;
3792 zonelist = &pgdat->node_zonelists[0];
3793 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3794 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3795 for (z = zonelist->_zonerefs; z->zone; z++)
3796 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3799 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3801 * Return node id of node used for "local" allocations.
3802 * I.e., first node id of first zone in arg node's generic zonelist.
3803 * Used for initializing percpu 'numa_mem', which is used primarily
3804 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3806 int local_memory_node(int node)
3810 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3811 gfp_zone(GFP_KERNEL),
3818 #else /* CONFIG_NUMA */
3820 static void set_zonelist_order(void)
3822 current_zonelist_order = ZONELIST_ORDER_ZONE;
3825 static void build_zonelists(pg_data_t *pgdat)
3827 int node, local_node;
3829 struct zonelist *zonelist;
3831 local_node = pgdat->node_id;
3833 zonelist = &pgdat->node_zonelists[0];
3834 j = build_zonelists_node(pgdat, zonelist, 0);
3837 * Now we build the zonelist so that it contains the zones
3838 * of all the other nodes.
3839 * We don't want to pressure a particular node, so when
3840 * building the zones for node N, we make sure that the
3841 * zones coming right after the local ones are those from
3842 * node N+1 (modulo N)
3844 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3845 if (!node_online(node))
3847 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3849 for (node = 0; node < local_node; node++) {
3850 if (!node_online(node))
3852 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3855 zonelist->_zonerefs[j].zone = NULL;
3856 zonelist->_zonerefs[j].zone_idx = 0;
3859 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3860 static void build_zonelist_cache(pg_data_t *pgdat)
3862 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3865 #endif /* CONFIG_NUMA */
3868 * Boot pageset table. One per cpu which is going to be used for all
3869 * zones and all nodes. The parameters will be set in such a way
3870 * that an item put on a list will immediately be handed over to
3871 * the buddy list. This is safe since pageset manipulation is done
3872 * with interrupts disabled.
3874 * The boot_pagesets must be kept even after bootup is complete for
3875 * unused processors and/or zones. They do play a role for bootstrapping
3876 * hotplugged processors.
3878 * zoneinfo_show() and maybe other functions do
3879 * not check if the processor is online before following the pageset pointer.
3880 * Other parts of the kernel may not check if the zone is available.
3882 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3883 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3884 static void setup_zone_pageset(struct zone *zone);
3887 * Global mutex to protect against size modification of zonelists
3888 * as well as to serialize pageset setup for the new populated zone.
3890 DEFINE_MUTEX(zonelists_mutex);
3892 /* return values int ....just for stop_machine() */
3893 static int __build_all_zonelists(void *data)
3897 pg_data_t *self = data;
3900 memset(node_load, 0, sizeof(node_load));
3903 if (self && !node_online(self->node_id)) {
3904 build_zonelists(self);
3905 build_zonelist_cache(self);
3908 for_each_online_node(nid) {
3909 pg_data_t *pgdat = NODE_DATA(nid);
3911 build_zonelists(pgdat);
3912 build_zonelist_cache(pgdat);
3916 * Initialize the boot_pagesets that are going to be used
3917 * for bootstrapping processors. The real pagesets for
3918 * each zone will be allocated later when the per cpu
3919 * allocator is available.
3921 * boot_pagesets are used also for bootstrapping offline
3922 * cpus if the system is already booted because the pagesets
3923 * are needed to initialize allocators on a specific cpu too.
3924 * F.e. the percpu allocator needs the page allocator which
3925 * needs the percpu allocator in order to allocate its pagesets
3926 * (a chicken-egg dilemma).
3928 for_each_possible_cpu(cpu) {
3929 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3931 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3933 * We now know the "local memory node" for each node--
3934 * i.e., the node of the first zone in the generic zonelist.
3935 * Set up numa_mem percpu variable for on-line cpus. During
3936 * boot, only the boot cpu should be on-line; we'll init the
3937 * secondary cpus' numa_mem as they come on-line. During
3938 * node/memory hotplug, we'll fixup all on-line cpus.
3940 if (cpu_online(cpu))
3941 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3949 * Called with zonelists_mutex held always
3950 * unless system_state == SYSTEM_BOOTING.
3952 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3954 set_zonelist_order();
3956 if (system_state == SYSTEM_BOOTING) {
3957 __build_all_zonelists(NULL);
3958 mminit_verify_zonelist();
3959 cpuset_init_current_mems_allowed();
3961 #ifdef CONFIG_MEMORY_HOTPLUG
3963 setup_zone_pageset(zone);
3965 /* we have to stop all cpus to guarantee there is no user
3967 stop_machine(__build_all_zonelists, pgdat, NULL);
3968 /* cpuset refresh routine should be here */
3970 vm_total_pages = nr_free_pagecache_pages();
3972 * Disable grouping by mobility if the number of pages in the
3973 * system is too low to allow the mechanism to work. It would be
3974 * more accurate, but expensive to check per-zone. This check is
3975 * made on memory-hotadd so a system can start with mobility
3976 * disabled and enable it later
3978 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3979 page_group_by_mobility_disabled = 1;
3981 page_group_by_mobility_disabled = 0;
3983 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3984 "Total pages: %ld\n",
3986 zonelist_order_name[current_zonelist_order],
3987 page_group_by_mobility_disabled ? "off" : "on",
3990 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
3995 * Helper functions to size the waitqueue hash table.
3996 * Essentially these want to choose hash table sizes sufficiently
3997 * large so that collisions trying to wait on pages are rare.
3998 * But in fact, the number of active page waitqueues on typical
3999 * systems is ridiculously low, less than 200. So this is even
4000 * conservative, even though it seems large.
4002 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4003 * waitqueues, i.e. the size of the waitq table given the number of pages.
4005 #define PAGES_PER_WAITQUEUE 256
4007 #ifndef CONFIG_MEMORY_HOTPLUG
4008 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4010 unsigned long size = 1;
4012 pages /= PAGES_PER_WAITQUEUE;
4014 while (size < pages)
4018 * Once we have dozens or even hundreds of threads sleeping
4019 * on IO we've got bigger problems than wait queue collision.
4020 * Limit the size of the wait table to a reasonable size.
4022 size = min(size, 4096UL);
4024 return max(size, 4UL);
4028 * A zone's size might be changed by hot-add, so it is not possible to determine
4029 * a suitable size for its wait_table. So we use the maximum size now.
4031 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4033 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4034 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4035 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4037 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4038 * or more by the traditional way. (See above). It equals:
4040 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4041 * ia64(16K page size) : = ( 8G + 4M)byte.
4042 * powerpc (64K page size) : = (32G +16M)byte.
4044 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4051 * This is an integer logarithm so that shifts can be used later
4052 * to extract the more random high bits from the multiplicative
4053 * hash function before the remainder is taken.
4055 static inline unsigned long wait_table_bits(unsigned long size)
4061 * Check if a pageblock contains reserved pages
4063 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4067 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4068 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4075 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4076 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4077 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4078 * higher will lead to a bigger reserve which will get freed as contiguous
4079 * blocks as reclaim kicks in
4081 static void setup_zone_migrate_reserve(struct zone *zone)
4083 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4085 unsigned long block_migratetype;
4090 * Get the start pfn, end pfn and the number of blocks to reserve
4091 * We have to be careful to be aligned to pageblock_nr_pages to
4092 * make sure that we always check pfn_valid for the first page in
4095 start_pfn = zone->zone_start_pfn;
4096 end_pfn = zone_end_pfn(zone);
4097 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4098 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4102 * Reserve blocks are generally in place to help high-order atomic
4103 * allocations that are short-lived. A min_free_kbytes value that
4104 * would result in more than 2 reserve blocks for atomic allocations
4105 * is assumed to be in place to help anti-fragmentation for the
4106 * future allocation of hugepages at runtime.
4108 reserve = min(2, reserve);
4109 old_reserve = zone->nr_migrate_reserve_block;
4111 /* When memory hot-add, we almost always need to do nothing */
4112 if (reserve == old_reserve)
4114 zone->nr_migrate_reserve_block = reserve;
4116 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4117 if (!pfn_valid(pfn))
4119 page = pfn_to_page(pfn);
4121 /* Watch out for overlapping nodes */
4122 if (page_to_nid(page) != zone_to_nid(zone))
4125 block_migratetype = get_pageblock_migratetype(page);
4127 /* Only test what is necessary when the reserves are not met */
4130 * Blocks with reserved pages will never free, skip
4133 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4134 if (pageblock_is_reserved(pfn, block_end_pfn))
4137 /* If this block is reserved, account for it */
4138 if (block_migratetype == MIGRATE_RESERVE) {
4143 /* Suitable for reserving if this block is movable */
4144 if (block_migratetype == MIGRATE_MOVABLE) {
4145 set_pageblock_migratetype(page,
4147 move_freepages_block(zone, page,
4152 } else if (!old_reserve) {
4154 * At boot time we don't need to scan the whole zone
4155 * for turning off MIGRATE_RESERVE.
4161 * If the reserve is met and this is a previous reserved block,
4164 if (block_migratetype == MIGRATE_RESERVE) {
4165 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4166 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4172 * Initially all pages are reserved - free ones are freed
4173 * up by free_all_bootmem() once the early boot process is
4174 * done. Non-atomic initialization, single-pass.
4176 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4177 unsigned long start_pfn, enum memmap_context context)
4180 unsigned long end_pfn = start_pfn + size;
4184 if (highest_memmap_pfn < end_pfn - 1)
4185 highest_memmap_pfn = end_pfn - 1;
4187 z = &NODE_DATA(nid)->node_zones[zone];
4188 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4190 * There can be holes in boot-time mem_map[]s
4191 * handed to this function. They do not
4192 * exist on hotplugged memory.
4194 if (context == MEMMAP_EARLY) {
4195 if (!early_pfn_valid(pfn))
4197 if (!early_pfn_in_nid(pfn, nid))
4200 page = pfn_to_page(pfn);
4201 set_page_links(page, zone, nid, pfn);
4202 mminit_verify_page_links(page, zone, nid, pfn);
4203 init_page_count(page);
4204 page_mapcount_reset(page);
4205 page_cpupid_reset_last(page);
4206 SetPageReserved(page);
4208 * Mark the block movable so that blocks are reserved for
4209 * movable at startup. This will force kernel allocations
4210 * to reserve their blocks rather than leaking throughout
4211 * the address space during boot when many long-lived
4212 * kernel allocations are made. Later some blocks near
4213 * the start are marked MIGRATE_RESERVE by
4214 * setup_zone_migrate_reserve()
4216 * bitmap is created for zone's valid pfn range. but memmap
4217 * can be created for invalid pages (for alignment)
4218 * check here not to call set_pageblock_migratetype() against
4221 if ((z->zone_start_pfn <= pfn)
4222 && (pfn < zone_end_pfn(z))
4223 && !(pfn & (pageblock_nr_pages - 1)))
4224 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4226 INIT_LIST_HEAD(&page->lru);
4227 #ifdef WANT_PAGE_VIRTUAL
4228 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4229 if (!is_highmem_idx(zone))
4230 set_page_address(page, __va(pfn << PAGE_SHIFT));
4235 static void __meminit zone_init_free_lists(struct zone *zone)
4237 unsigned int order, t;
4238 for_each_migratetype_order(order, t) {
4239 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4240 zone->free_area[order].nr_free = 0;
4244 #ifndef __HAVE_ARCH_MEMMAP_INIT
4245 #define memmap_init(size, nid, zone, start_pfn) \
4246 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4249 static int zone_batchsize(struct zone *zone)
4255 * The per-cpu-pages pools are set to around 1000th of the
4256 * size of the zone. But no more than 1/2 of a meg.
4258 * OK, so we don't know how big the cache is. So guess.
4260 batch = zone->managed_pages / 1024;
4261 if (batch * PAGE_SIZE > 512 * 1024)
4262 batch = (512 * 1024) / PAGE_SIZE;
4263 batch /= 4; /* We effectively *= 4 below */
4268 * Clamp the batch to a 2^n - 1 value. Having a power
4269 * of 2 value was found to be more likely to have
4270 * suboptimal cache aliasing properties in some cases.
4272 * For example if 2 tasks are alternately allocating
4273 * batches of pages, one task can end up with a lot
4274 * of pages of one half of the possible page colors
4275 * and the other with pages of the other colors.
4277 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4282 /* The deferral and batching of frees should be suppressed under NOMMU
4285 * The problem is that NOMMU needs to be able to allocate large chunks
4286 * of contiguous memory as there's no hardware page translation to
4287 * assemble apparent contiguous memory from discontiguous pages.
4289 * Queueing large contiguous runs of pages for batching, however,
4290 * causes the pages to actually be freed in smaller chunks. As there
4291 * can be a significant delay between the individual batches being
4292 * recycled, this leads to the once large chunks of space being
4293 * fragmented and becoming unavailable for high-order allocations.
4300 * pcp->high and pcp->batch values are related and dependent on one another:
4301 * ->batch must never be higher then ->high.
4302 * The following function updates them in a safe manner without read side
4305 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4306 * those fields changing asynchronously (acording the the above rule).
4308 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4309 * outside of boot time (or some other assurance that no concurrent updaters
4312 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4313 unsigned long batch)
4315 /* start with a fail safe value for batch */
4319 /* Update high, then batch, in order */
4326 /* a companion to pageset_set_high() */
4327 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4329 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4332 static void pageset_init(struct per_cpu_pageset *p)
4334 struct per_cpu_pages *pcp;
4337 memset(p, 0, sizeof(*p));
4341 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4342 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4345 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4348 pageset_set_batch(p, batch);
4352 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4353 * to the value high for the pageset p.
4355 static void pageset_set_high(struct per_cpu_pageset *p,
4358 unsigned long batch = max(1UL, high / 4);
4359 if ((high / 4) > (PAGE_SHIFT * 8))
4360 batch = PAGE_SHIFT * 8;
4362 pageset_update(&p->pcp, high, batch);
4365 static void pageset_set_high_and_batch(struct zone *zone,
4366 struct per_cpu_pageset *pcp)
4368 if (percpu_pagelist_fraction)
4369 pageset_set_high(pcp,
4370 (zone->managed_pages /
4371 percpu_pagelist_fraction));
4373 pageset_set_batch(pcp, zone_batchsize(zone));
4376 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4378 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4381 pageset_set_high_and_batch(zone, pcp);
4384 static void __meminit setup_zone_pageset(struct zone *zone)
4387 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4388 for_each_possible_cpu(cpu)
4389 zone_pageset_init(zone, cpu);
4393 * Allocate per cpu pagesets and initialize them.
4394 * Before this call only boot pagesets were available.
4396 void __init setup_per_cpu_pageset(void)
4400 for_each_populated_zone(zone)
4401 setup_zone_pageset(zone);
4404 static noinline __init_refok
4405 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4411 * The per-page waitqueue mechanism uses hashed waitqueues
4414 zone->wait_table_hash_nr_entries =
4415 wait_table_hash_nr_entries(zone_size_pages);
4416 zone->wait_table_bits =
4417 wait_table_bits(zone->wait_table_hash_nr_entries);
4418 alloc_size = zone->wait_table_hash_nr_entries
4419 * sizeof(wait_queue_head_t);
4421 if (!slab_is_available()) {
4422 zone->wait_table = (wait_queue_head_t *)
4423 memblock_virt_alloc_node_nopanic(
4424 alloc_size, zone->zone_pgdat->node_id);
4427 * This case means that a zone whose size was 0 gets new memory
4428 * via memory hot-add.
4429 * But it may be the case that a new node was hot-added. In
4430 * this case vmalloc() will not be able to use this new node's
4431 * memory - this wait_table must be initialized to use this new
4432 * node itself as well.
4433 * To use this new node's memory, further consideration will be
4436 zone->wait_table = vmalloc(alloc_size);
4438 if (!zone->wait_table)
4441 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4442 init_waitqueue_head(zone->wait_table + i);
4447 static __meminit void zone_pcp_init(struct zone *zone)
4450 * per cpu subsystem is not up at this point. The following code
4451 * relies on the ability of the linker to provide the
4452 * offset of a (static) per cpu variable into the per cpu area.
4454 zone->pageset = &boot_pageset;
4456 if (populated_zone(zone))
4457 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4458 zone->name, zone->present_pages,
4459 zone_batchsize(zone));
4462 int __meminit init_currently_empty_zone(struct zone *zone,
4463 unsigned long zone_start_pfn,
4465 enum memmap_context context)
4467 struct pglist_data *pgdat = zone->zone_pgdat;
4469 ret = zone_wait_table_init(zone, size);
4472 pgdat->nr_zones = zone_idx(zone) + 1;
4474 zone->zone_start_pfn = zone_start_pfn;
4476 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4477 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4479 (unsigned long)zone_idx(zone),
4480 zone_start_pfn, (zone_start_pfn + size));
4482 zone_init_free_lists(zone);
4487 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4488 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4490 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4492 int __meminit __early_pfn_to_nid(unsigned long pfn)
4494 unsigned long start_pfn, end_pfn;
4497 * NOTE: The following SMP-unsafe globals are only used early in boot
4498 * when the kernel is running single-threaded.
4500 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4501 static int __meminitdata last_nid;
4503 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4506 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4508 last_start_pfn = start_pfn;
4509 last_end_pfn = end_pfn;
4515 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4517 int __meminit early_pfn_to_nid(unsigned long pfn)
4521 nid = __early_pfn_to_nid(pfn);
4524 /* just returns 0 */
4528 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4529 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4533 nid = __early_pfn_to_nid(pfn);
4534 if (nid >= 0 && nid != node)
4541 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4542 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4543 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4545 * If an architecture guarantees that all ranges registered contain no holes
4546 * and may be freed, this this function may be used instead of calling
4547 * memblock_free_early_nid() manually.
4549 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4551 unsigned long start_pfn, end_pfn;
4554 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4555 start_pfn = min(start_pfn, max_low_pfn);
4556 end_pfn = min(end_pfn, max_low_pfn);
4558 if (start_pfn < end_pfn)
4559 memblock_free_early_nid(PFN_PHYS(start_pfn),
4560 (end_pfn - start_pfn) << PAGE_SHIFT,
4566 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4567 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4569 * If an architecture guarantees that all ranges registered contain no holes and may
4570 * be freed, this function may be used instead of calling memory_present() manually.
4572 void __init sparse_memory_present_with_active_regions(int nid)
4574 unsigned long start_pfn, end_pfn;
4577 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4578 memory_present(this_nid, start_pfn, end_pfn);
4582 * get_pfn_range_for_nid - Return the start and end page frames for a node
4583 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4584 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4585 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4587 * It returns the start and end page frame of a node based on information
4588 * provided by memblock_set_node(). If called for a node
4589 * with no available memory, a warning is printed and the start and end
4592 void __meminit get_pfn_range_for_nid(unsigned int nid,
4593 unsigned long *start_pfn, unsigned long *end_pfn)
4595 unsigned long this_start_pfn, this_end_pfn;
4601 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4602 *start_pfn = min(*start_pfn, this_start_pfn);
4603 *end_pfn = max(*end_pfn, this_end_pfn);
4606 if (*start_pfn == -1UL)
4611 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4612 * assumption is made that zones within a node are ordered in monotonic
4613 * increasing memory addresses so that the "highest" populated zone is used
4615 static void __init find_usable_zone_for_movable(void)
4618 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4619 if (zone_index == ZONE_MOVABLE)
4622 if (arch_zone_highest_possible_pfn[zone_index] >
4623 arch_zone_lowest_possible_pfn[zone_index])
4627 VM_BUG_ON(zone_index == -1);
4628 movable_zone = zone_index;
4632 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4633 * because it is sized independent of architecture. Unlike the other zones,
4634 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4635 * in each node depending on the size of each node and how evenly kernelcore
4636 * is distributed. This helper function adjusts the zone ranges
4637 * provided by the architecture for a given node by using the end of the
4638 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4639 * zones within a node are in order of monotonic increases memory addresses
4641 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4642 unsigned long zone_type,
4643 unsigned long node_start_pfn,
4644 unsigned long node_end_pfn,
4645 unsigned long *zone_start_pfn,
4646 unsigned long *zone_end_pfn)
4648 /* Only adjust if ZONE_MOVABLE is on this node */
4649 if (zone_movable_pfn[nid]) {
4650 /* Size ZONE_MOVABLE */
4651 if (zone_type == ZONE_MOVABLE) {
4652 *zone_start_pfn = zone_movable_pfn[nid];
4653 *zone_end_pfn = min(node_end_pfn,
4654 arch_zone_highest_possible_pfn[movable_zone]);
4656 /* Adjust for ZONE_MOVABLE starting within this range */
4657 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4658 *zone_end_pfn > zone_movable_pfn[nid]) {
4659 *zone_end_pfn = zone_movable_pfn[nid];
4661 /* Check if this whole range is within ZONE_MOVABLE */
4662 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4663 *zone_start_pfn = *zone_end_pfn;
4668 * Return the number of pages a zone spans in a node, including holes
4669 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4671 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4672 unsigned long zone_type,
4673 unsigned long node_start_pfn,
4674 unsigned long node_end_pfn,
4675 unsigned long *ignored)
4677 unsigned long zone_start_pfn, zone_end_pfn;
4679 /* Get the start and end of the zone */
4680 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4681 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4682 adjust_zone_range_for_zone_movable(nid, zone_type,
4683 node_start_pfn, node_end_pfn,
4684 &zone_start_pfn, &zone_end_pfn);
4686 /* Check that this node has pages within the zone's required range */
4687 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4690 /* Move the zone boundaries inside the node if necessary */
4691 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4692 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4694 /* Return the spanned pages */
4695 return zone_end_pfn - zone_start_pfn;
4699 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4700 * then all holes in the requested range will be accounted for.
4702 unsigned long __meminit __absent_pages_in_range(int nid,
4703 unsigned long range_start_pfn,
4704 unsigned long range_end_pfn)
4706 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4707 unsigned long start_pfn, end_pfn;
4710 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4711 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4712 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4713 nr_absent -= end_pfn - start_pfn;
4719 * absent_pages_in_range - Return number of page frames in holes within a range
4720 * @start_pfn: The start PFN to start searching for holes
4721 * @end_pfn: The end PFN to stop searching for holes
4723 * It returns the number of pages frames in memory holes within a range.
4725 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4726 unsigned long end_pfn)
4728 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4731 /* Return the number of page frames in holes in a zone on a node */
4732 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4733 unsigned long zone_type,
4734 unsigned long node_start_pfn,
4735 unsigned long node_end_pfn,
4736 unsigned long *ignored)
4738 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4739 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4740 unsigned long zone_start_pfn, zone_end_pfn;
4742 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4743 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4745 adjust_zone_range_for_zone_movable(nid, zone_type,
4746 node_start_pfn, node_end_pfn,
4747 &zone_start_pfn, &zone_end_pfn);
4748 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4751 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4752 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4753 unsigned long zone_type,
4754 unsigned long node_start_pfn,
4755 unsigned long node_end_pfn,
4756 unsigned long *zones_size)
4758 return zones_size[zone_type];
4761 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4762 unsigned long zone_type,
4763 unsigned long node_start_pfn,
4764 unsigned long node_end_pfn,
4765 unsigned long *zholes_size)
4770 return zholes_size[zone_type];
4773 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4775 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4776 unsigned long node_start_pfn,
4777 unsigned long node_end_pfn,
4778 unsigned long *zones_size,
4779 unsigned long *zholes_size)
4781 unsigned long realtotalpages, totalpages = 0;
4784 for (i = 0; i < MAX_NR_ZONES; i++)
4785 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4789 pgdat->node_spanned_pages = totalpages;
4791 realtotalpages = totalpages;
4792 for (i = 0; i < MAX_NR_ZONES; i++)
4794 zone_absent_pages_in_node(pgdat->node_id, i,
4795 node_start_pfn, node_end_pfn,
4797 pgdat->node_present_pages = realtotalpages;
4798 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4802 #ifndef CONFIG_SPARSEMEM
4804 * Calculate the size of the zone->blockflags rounded to an unsigned long
4805 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4806 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4807 * round what is now in bits to nearest long in bits, then return it in
4810 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4812 unsigned long usemapsize;
4814 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4815 usemapsize = roundup(zonesize, pageblock_nr_pages);
4816 usemapsize = usemapsize >> pageblock_order;
4817 usemapsize *= NR_PAGEBLOCK_BITS;
4818 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4820 return usemapsize / 8;
4823 static void __init setup_usemap(struct pglist_data *pgdat,
4825 unsigned long zone_start_pfn,
4826 unsigned long zonesize)
4828 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4829 zone->pageblock_flags = NULL;
4831 zone->pageblock_flags =
4832 memblock_virt_alloc_node_nopanic(usemapsize,
4836 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4837 unsigned long zone_start_pfn, unsigned long zonesize) {}
4838 #endif /* CONFIG_SPARSEMEM */
4840 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4842 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4843 void __paginginit set_pageblock_order(void)
4847 /* Check that pageblock_nr_pages has not already been setup */
4848 if (pageblock_order)
4851 if (HPAGE_SHIFT > PAGE_SHIFT)
4852 order = HUGETLB_PAGE_ORDER;
4854 order = MAX_ORDER - 1;
4857 * Assume the largest contiguous order of interest is a huge page.
4858 * This value may be variable depending on boot parameters on IA64 and
4861 pageblock_order = order;
4863 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4866 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4867 * is unused as pageblock_order is set at compile-time. See
4868 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4871 void __paginginit set_pageblock_order(void)
4875 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4877 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4878 unsigned long present_pages)
4880 unsigned long pages = spanned_pages;
4883 * Provide a more accurate estimation if there are holes within
4884 * the zone and SPARSEMEM is in use. If there are holes within the
4885 * zone, each populated memory region may cost us one or two extra
4886 * memmap pages due to alignment because memmap pages for each
4887 * populated regions may not naturally algined on page boundary.
4888 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4890 if (spanned_pages > present_pages + (present_pages >> 4) &&
4891 IS_ENABLED(CONFIG_SPARSEMEM))
4892 pages = present_pages;
4894 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4898 * Set up the zone data structures:
4899 * - mark all pages reserved
4900 * - mark all memory queues empty
4901 * - clear the memory bitmaps
4903 * NOTE: pgdat should get zeroed by caller.
4905 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4906 unsigned long node_start_pfn, unsigned long node_end_pfn,
4907 unsigned long *zones_size, unsigned long *zholes_size)
4910 int nid = pgdat->node_id;
4911 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4914 pgdat_resize_init(pgdat);
4915 #ifdef CONFIG_NUMA_BALANCING
4916 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4917 pgdat->numabalancing_migrate_nr_pages = 0;
4918 pgdat->numabalancing_migrate_next_window = jiffies;
4920 init_waitqueue_head(&pgdat->kswapd_wait);
4921 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4922 pgdat_page_ext_init(pgdat);
4924 for (j = 0; j < MAX_NR_ZONES; j++) {
4925 struct zone *zone = pgdat->node_zones + j;
4926 unsigned long size, realsize, freesize, memmap_pages;
4928 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4929 node_end_pfn, zones_size);
4930 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4936 * Adjust freesize so that it accounts for how much memory
4937 * is used by this zone for memmap. This affects the watermark
4938 * and per-cpu initialisations
4940 memmap_pages = calc_memmap_size(size, realsize);
4941 if (!is_highmem_idx(j)) {
4942 if (freesize >= memmap_pages) {
4943 freesize -= memmap_pages;
4946 " %s zone: %lu pages used for memmap\n",
4947 zone_names[j], memmap_pages);
4950 " %s zone: %lu pages exceeds freesize %lu\n",
4951 zone_names[j], memmap_pages, freesize);
4954 /* Account for reserved pages */
4955 if (j == 0 && freesize > dma_reserve) {
4956 freesize -= dma_reserve;
4957 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4958 zone_names[0], dma_reserve);
4961 if (!is_highmem_idx(j))
4962 nr_kernel_pages += freesize;
4963 /* Charge for highmem memmap if there are enough kernel pages */
4964 else if (nr_kernel_pages > memmap_pages * 2)
4965 nr_kernel_pages -= memmap_pages;
4966 nr_all_pages += freesize;
4968 zone->spanned_pages = size;
4969 zone->present_pages = realsize;
4971 * Set an approximate value for lowmem here, it will be adjusted
4972 * when the bootmem allocator frees pages into the buddy system.
4973 * And all highmem pages will be managed by the buddy system.
4975 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4978 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4980 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4982 zone->name = zone_names[j];
4983 spin_lock_init(&zone->lock);
4984 spin_lock_init(&zone->lru_lock);
4985 zone_seqlock_init(zone);
4986 zone->zone_pgdat = pgdat;
4987 zone_pcp_init(zone);
4989 /* For bootup, initialized properly in watermark setup */
4990 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4992 lruvec_init(&zone->lruvec);
4996 set_pageblock_order();
4997 setup_usemap(pgdat, zone, zone_start_pfn, size);
4998 ret = init_currently_empty_zone(zone, zone_start_pfn,
4999 size, MEMMAP_EARLY);
5001 memmap_init(size, nid, j, zone_start_pfn);
5002 zone_start_pfn += size;
5006 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5008 /* Skip empty nodes */
5009 if (!pgdat->node_spanned_pages)
5012 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5013 /* ia64 gets its own node_mem_map, before this, without bootmem */
5014 if (!pgdat->node_mem_map) {
5015 unsigned long size, start, end;
5019 * The zone's endpoints aren't required to be MAX_ORDER
5020 * aligned but the node_mem_map endpoints must be in order
5021 * for the buddy allocator to function correctly.
5023 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5024 end = pgdat_end_pfn(pgdat);
5025 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5026 size = (end - start) * sizeof(struct page);
5027 map = alloc_remap(pgdat->node_id, size);
5029 map = memblock_virt_alloc_node_nopanic(size,
5031 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5033 #ifndef CONFIG_NEED_MULTIPLE_NODES
5035 * With no DISCONTIG, the global mem_map is just set as node 0's
5037 if (pgdat == NODE_DATA(0)) {
5038 mem_map = NODE_DATA(0)->node_mem_map;
5039 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5040 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5041 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5042 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5045 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5048 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5049 unsigned long node_start_pfn, unsigned long *zholes_size)
5051 pg_data_t *pgdat = NODE_DATA(nid);
5052 unsigned long start_pfn = 0;
5053 unsigned long end_pfn = 0;
5055 /* pg_data_t should be reset to zero when it's allocated */
5056 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5058 pgdat->node_id = nid;
5059 pgdat->node_start_pfn = node_start_pfn;
5060 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5061 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5062 printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid,
5063 (u64) start_pfn << PAGE_SHIFT, (u64) (end_pfn << PAGE_SHIFT) - 1);
5065 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5066 zones_size, zholes_size);
5068 alloc_node_mem_map(pgdat);
5069 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5070 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5071 nid, (unsigned long)pgdat,
5072 (unsigned long)pgdat->node_mem_map);
5075 free_area_init_core(pgdat, start_pfn, end_pfn,
5076 zones_size, zholes_size);
5079 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5081 #if MAX_NUMNODES > 1
5083 * Figure out the number of possible node ids.
5085 void __init setup_nr_node_ids(void)
5088 unsigned int highest = 0;
5090 for_each_node_mask(node, node_possible_map)
5092 nr_node_ids = highest + 1;
5097 * node_map_pfn_alignment - determine the maximum internode alignment
5099 * This function should be called after node map is populated and sorted.
5100 * It calculates the maximum power of two alignment which can distinguish
5103 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5104 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5105 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5106 * shifted, 1GiB is enough and this function will indicate so.
5108 * This is used to test whether pfn -> nid mapping of the chosen memory
5109 * model has fine enough granularity to avoid incorrect mapping for the
5110 * populated node map.
5112 * Returns the determined alignment in pfn's. 0 if there is no alignment
5113 * requirement (single node).
5115 unsigned long __init node_map_pfn_alignment(void)
5117 unsigned long accl_mask = 0, last_end = 0;
5118 unsigned long start, end, mask;
5122 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5123 if (!start || last_nid < 0 || last_nid == nid) {
5130 * Start with a mask granular enough to pin-point to the
5131 * start pfn and tick off bits one-by-one until it becomes
5132 * too coarse to separate the current node from the last.
5134 mask = ~((1 << __ffs(start)) - 1);
5135 while (mask && last_end <= (start & (mask << 1)))
5138 /* accumulate all internode masks */
5142 /* convert mask to number of pages */
5143 return ~accl_mask + 1;
5146 /* Find the lowest pfn for a node */
5147 static unsigned long __init find_min_pfn_for_node(int nid)
5149 unsigned long min_pfn = ULONG_MAX;
5150 unsigned long start_pfn;
5153 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5154 min_pfn = min(min_pfn, start_pfn);
5156 if (min_pfn == ULONG_MAX) {
5158 "Could not find start_pfn for node %d\n", nid);
5166 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5168 * It returns the minimum PFN based on information provided via
5169 * memblock_set_node().
5171 unsigned long __init find_min_pfn_with_active_regions(void)
5173 return find_min_pfn_for_node(MAX_NUMNODES);
5177 * early_calculate_totalpages()
5178 * Sum pages in active regions for movable zone.
5179 * Populate N_MEMORY for calculating usable_nodes.
5181 static unsigned long __init early_calculate_totalpages(void)
5183 unsigned long totalpages = 0;
5184 unsigned long start_pfn, end_pfn;
5187 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5188 unsigned long pages = end_pfn - start_pfn;
5190 totalpages += pages;
5192 node_set_state(nid, N_MEMORY);
5198 * Find the PFN the Movable zone begins in each node. Kernel memory
5199 * is spread evenly between nodes as long as the nodes have enough
5200 * memory. When they don't, some nodes will have more kernelcore than
5203 static void __init find_zone_movable_pfns_for_nodes(void)
5206 unsigned long usable_startpfn;
5207 unsigned long kernelcore_node, kernelcore_remaining;
5208 /* save the state before borrow the nodemask */
5209 nodemask_t saved_node_state = node_states[N_MEMORY];
5210 unsigned long totalpages = early_calculate_totalpages();
5211 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5212 struct memblock_region *r;
5214 /* Need to find movable_zone earlier when movable_node is specified. */
5215 find_usable_zone_for_movable();
5218 * If movable_node is specified, ignore kernelcore and movablecore
5221 if (movable_node_is_enabled()) {
5222 for_each_memblock(memory, r) {
5223 if (!memblock_is_hotpluggable(r))
5228 usable_startpfn = PFN_DOWN(r->base);
5229 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5230 min(usable_startpfn, zone_movable_pfn[nid]) :
5238 * If movablecore=nn[KMG] was specified, calculate what size of
5239 * kernelcore that corresponds so that memory usable for
5240 * any allocation type is evenly spread. If both kernelcore
5241 * and movablecore are specified, then the value of kernelcore
5242 * will be used for required_kernelcore if it's greater than
5243 * what movablecore would have allowed.
5245 if (required_movablecore) {
5246 unsigned long corepages;
5249 * Round-up so that ZONE_MOVABLE is at least as large as what
5250 * was requested by the user
5252 required_movablecore =
5253 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5254 corepages = totalpages - required_movablecore;
5256 required_kernelcore = max(required_kernelcore, corepages);
5259 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5260 if (!required_kernelcore)
5263 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5264 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5267 /* Spread kernelcore memory as evenly as possible throughout nodes */
5268 kernelcore_node = required_kernelcore / usable_nodes;
5269 for_each_node_state(nid, N_MEMORY) {
5270 unsigned long start_pfn, end_pfn;
5273 * Recalculate kernelcore_node if the division per node
5274 * now exceeds what is necessary to satisfy the requested
5275 * amount of memory for the kernel
5277 if (required_kernelcore < kernelcore_node)
5278 kernelcore_node = required_kernelcore / usable_nodes;
5281 * As the map is walked, we track how much memory is usable
5282 * by the kernel using kernelcore_remaining. When it is
5283 * 0, the rest of the node is usable by ZONE_MOVABLE
5285 kernelcore_remaining = kernelcore_node;
5287 /* Go through each range of PFNs within this node */
5288 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5289 unsigned long size_pages;
5291 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5292 if (start_pfn >= end_pfn)
5295 /* Account for what is only usable for kernelcore */
5296 if (start_pfn < usable_startpfn) {
5297 unsigned long kernel_pages;
5298 kernel_pages = min(end_pfn, usable_startpfn)
5301 kernelcore_remaining -= min(kernel_pages,
5302 kernelcore_remaining);
5303 required_kernelcore -= min(kernel_pages,
5304 required_kernelcore);
5306 /* Continue if range is now fully accounted */
5307 if (end_pfn <= usable_startpfn) {
5310 * Push zone_movable_pfn to the end so
5311 * that if we have to rebalance
5312 * kernelcore across nodes, we will
5313 * not double account here
5315 zone_movable_pfn[nid] = end_pfn;
5318 start_pfn = usable_startpfn;
5322 * The usable PFN range for ZONE_MOVABLE is from
5323 * start_pfn->end_pfn. Calculate size_pages as the
5324 * number of pages used as kernelcore
5326 size_pages = end_pfn - start_pfn;
5327 if (size_pages > kernelcore_remaining)
5328 size_pages = kernelcore_remaining;
5329 zone_movable_pfn[nid] = start_pfn + size_pages;
5332 * Some kernelcore has been met, update counts and
5333 * break if the kernelcore for this node has been
5336 required_kernelcore -= min(required_kernelcore,
5338 kernelcore_remaining -= size_pages;
5339 if (!kernelcore_remaining)
5345 * If there is still required_kernelcore, we do another pass with one
5346 * less node in the count. This will push zone_movable_pfn[nid] further
5347 * along on the nodes that still have memory until kernelcore is
5351 if (usable_nodes && required_kernelcore > usable_nodes)
5355 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5356 for (nid = 0; nid < MAX_NUMNODES; nid++)
5357 zone_movable_pfn[nid] =
5358 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5361 /* restore the node_state */
5362 node_states[N_MEMORY] = saved_node_state;
5365 /* Any regular or high memory on that node ? */
5366 static void check_for_memory(pg_data_t *pgdat, int nid)
5368 enum zone_type zone_type;
5370 if (N_MEMORY == N_NORMAL_MEMORY)
5373 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5374 struct zone *zone = &pgdat->node_zones[zone_type];
5375 if (populated_zone(zone)) {
5376 node_set_state(nid, N_HIGH_MEMORY);
5377 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5378 zone_type <= ZONE_NORMAL)
5379 node_set_state(nid, N_NORMAL_MEMORY);
5386 * free_area_init_nodes - Initialise all pg_data_t and zone data
5387 * @max_zone_pfn: an array of max PFNs for each zone
5389 * This will call free_area_init_node() for each active node in the system.
5390 * Using the page ranges provided by memblock_set_node(), the size of each
5391 * zone in each node and their holes is calculated. If the maximum PFN
5392 * between two adjacent zones match, it is assumed that the zone is empty.
5393 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5394 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5395 * starts where the previous one ended. For example, ZONE_DMA32 starts
5396 * at arch_max_dma_pfn.
5398 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5400 unsigned long start_pfn, end_pfn;
5403 /* Record where the zone boundaries are */
5404 memset(arch_zone_lowest_possible_pfn, 0,
5405 sizeof(arch_zone_lowest_possible_pfn));
5406 memset(arch_zone_highest_possible_pfn, 0,
5407 sizeof(arch_zone_highest_possible_pfn));
5408 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5409 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5410 for (i = 1; i < MAX_NR_ZONES; i++) {
5411 if (i == ZONE_MOVABLE)
5413 arch_zone_lowest_possible_pfn[i] =
5414 arch_zone_highest_possible_pfn[i-1];
5415 arch_zone_highest_possible_pfn[i] =
5416 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5418 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5419 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5421 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5422 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5423 find_zone_movable_pfns_for_nodes();
5425 /* Print out the zone ranges */
5426 pr_info("Zone ranges:\n");
5427 for (i = 0; i < MAX_NR_ZONES; i++) {
5428 if (i == ZONE_MOVABLE)
5430 pr_info(" %-8s ", zone_names[i]);
5431 if (arch_zone_lowest_possible_pfn[i] ==
5432 arch_zone_highest_possible_pfn[i])
5435 pr_cont("[mem %0#10lx-%0#10lx]\n",
5436 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5437 (arch_zone_highest_possible_pfn[i]
5438 << PAGE_SHIFT) - 1);
5441 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5442 pr_info("Movable zone start for each node\n");
5443 for (i = 0; i < MAX_NUMNODES; i++) {
5444 if (zone_movable_pfn[i])
5445 pr_info(" Node %d: %#010lx\n", i,
5446 zone_movable_pfn[i] << PAGE_SHIFT);
5449 /* Print out the early node map */
5450 pr_info("Early memory node ranges\n");
5451 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5452 pr_info(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5453 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5455 /* Initialise every node */
5456 mminit_verify_pageflags_layout();
5457 setup_nr_node_ids();
5458 for_each_online_node(nid) {
5459 pg_data_t *pgdat = NODE_DATA(nid);
5460 free_area_init_node(nid, NULL,
5461 find_min_pfn_for_node(nid), NULL);
5463 /* Any memory on that node */
5464 if (pgdat->node_present_pages)
5465 node_set_state(nid, N_MEMORY);
5466 check_for_memory(pgdat, nid);
5470 static int __init cmdline_parse_core(char *p, unsigned long *core)
5472 unsigned long long coremem;
5476 coremem = memparse(p, &p);
5477 *core = coremem >> PAGE_SHIFT;
5479 /* Paranoid check that UL is enough for the coremem value */
5480 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5486 * kernelcore=size sets the amount of memory for use for allocations that
5487 * cannot be reclaimed or migrated.
5489 static int __init cmdline_parse_kernelcore(char *p)
5491 return cmdline_parse_core(p, &required_kernelcore);
5495 * movablecore=size sets the amount of memory for use for allocations that
5496 * can be reclaimed or migrated.
5498 static int __init cmdline_parse_movablecore(char *p)
5500 return cmdline_parse_core(p, &required_movablecore);
5503 early_param("kernelcore", cmdline_parse_kernelcore);
5504 early_param("movablecore", cmdline_parse_movablecore);
5506 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5508 void adjust_managed_page_count(struct page *page, long count)
5510 spin_lock(&managed_page_count_lock);
5511 page_zone(page)->managed_pages += count;
5512 totalram_pages += count;
5513 #ifdef CONFIG_HIGHMEM
5514 if (PageHighMem(page))
5515 totalhigh_pages += count;
5517 spin_unlock(&managed_page_count_lock);
5519 EXPORT_SYMBOL(adjust_managed_page_count);
5521 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5524 unsigned long pages = 0;
5526 start = (void *)PAGE_ALIGN((unsigned long)start);
5527 end = (void *)((unsigned long)end & PAGE_MASK);
5528 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5529 if ((unsigned int)poison <= 0xFF)
5530 memset(pos, poison, PAGE_SIZE);
5531 free_reserved_page(virt_to_page(pos));
5535 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5536 s, pages << (PAGE_SHIFT - 10), start, end);
5540 EXPORT_SYMBOL(free_reserved_area);
5542 #ifdef CONFIG_HIGHMEM
5543 void free_highmem_page(struct page *page)
5545 __free_reserved_page(page);
5547 page_zone(page)->managed_pages++;
5553 void __init mem_init_print_info(const char *str)
5555 unsigned long physpages, codesize, datasize, rosize, bss_size;
5556 unsigned long init_code_size, init_data_size;
5558 physpages = get_num_physpages();
5559 codesize = _etext - _stext;
5560 datasize = _edata - _sdata;
5561 rosize = __end_rodata - __start_rodata;
5562 bss_size = __bss_stop - __bss_start;
5563 init_data_size = __init_end - __init_begin;
5564 init_code_size = _einittext - _sinittext;
5567 * Detect special cases and adjust section sizes accordingly:
5568 * 1) .init.* may be embedded into .data sections
5569 * 2) .init.text.* may be out of [__init_begin, __init_end],
5570 * please refer to arch/tile/kernel/vmlinux.lds.S.
5571 * 3) .rodata.* may be embedded into .text or .data sections.
5573 #define adj_init_size(start, end, size, pos, adj) \
5575 if (start <= pos && pos < end && size > adj) \
5579 adj_init_size(__init_begin, __init_end, init_data_size,
5580 _sinittext, init_code_size);
5581 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5582 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5583 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5584 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5586 #undef adj_init_size
5588 pr_info("Memory: %luK/%luK available "
5589 "(%luK kernel code, %luK rwdata, %luK rodata, "
5590 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5591 #ifdef CONFIG_HIGHMEM
5595 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5596 codesize >> 10, datasize >> 10, rosize >> 10,
5597 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5598 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5599 totalcma_pages << (PAGE_SHIFT-10),
5600 #ifdef CONFIG_HIGHMEM
5601 totalhigh_pages << (PAGE_SHIFT-10),
5603 str ? ", " : "", str ? str : "");
5607 * set_dma_reserve - set the specified number of pages reserved in the first zone
5608 * @new_dma_reserve: The number of pages to mark reserved
5610 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5611 * In the DMA zone, a significant percentage may be consumed by kernel image
5612 * and other unfreeable allocations which can skew the watermarks badly. This
5613 * function may optionally be used to account for unfreeable pages in the
5614 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5615 * smaller per-cpu batchsize.
5617 void __init set_dma_reserve(unsigned long new_dma_reserve)
5619 dma_reserve = new_dma_reserve;
5622 void __init free_area_init(unsigned long *zones_size)
5624 free_area_init_node(0, zones_size,
5625 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5628 static int page_alloc_cpu_notify(struct notifier_block *self,
5629 unsigned long action, void *hcpu)
5631 int cpu = (unsigned long)hcpu;
5633 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5634 lru_add_drain_cpu(cpu);
5638 * Spill the event counters of the dead processor
5639 * into the current processors event counters.
5640 * This artificially elevates the count of the current
5643 vm_events_fold_cpu(cpu);
5646 * Zero the differential counters of the dead processor
5647 * so that the vm statistics are consistent.
5649 * This is only okay since the processor is dead and cannot
5650 * race with what we are doing.
5652 cpu_vm_stats_fold(cpu);
5657 void __init page_alloc_init(void)
5659 hotcpu_notifier(page_alloc_cpu_notify, 0);
5663 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5664 * or min_free_kbytes changes.
5666 static void calculate_totalreserve_pages(void)
5668 struct pglist_data *pgdat;
5669 unsigned long reserve_pages = 0;
5670 enum zone_type i, j;
5672 for_each_online_pgdat(pgdat) {
5673 for (i = 0; i < MAX_NR_ZONES; i++) {
5674 struct zone *zone = pgdat->node_zones + i;
5677 /* Find valid and maximum lowmem_reserve in the zone */
5678 for (j = i; j < MAX_NR_ZONES; j++) {
5679 if (zone->lowmem_reserve[j] > max)
5680 max = zone->lowmem_reserve[j];
5683 /* we treat the high watermark as reserved pages. */
5684 max += high_wmark_pages(zone);
5686 if (max > zone->managed_pages)
5687 max = zone->managed_pages;
5688 reserve_pages += max;
5690 * Lowmem reserves are not available to
5691 * GFP_HIGHUSER page cache allocations and
5692 * kswapd tries to balance zones to their high
5693 * watermark. As a result, neither should be
5694 * regarded as dirtyable memory, to prevent a
5695 * situation where reclaim has to clean pages
5696 * in order to balance the zones.
5698 zone->dirty_balance_reserve = max;
5701 dirty_balance_reserve = reserve_pages;
5702 totalreserve_pages = reserve_pages;
5706 * setup_per_zone_lowmem_reserve - called whenever
5707 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5708 * has a correct pages reserved value, so an adequate number of
5709 * pages are left in the zone after a successful __alloc_pages().
5711 static void setup_per_zone_lowmem_reserve(void)
5713 struct pglist_data *pgdat;
5714 enum zone_type j, idx;
5716 for_each_online_pgdat(pgdat) {
5717 for (j = 0; j < MAX_NR_ZONES; j++) {
5718 struct zone *zone = pgdat->node_zones + j;
5719 unsigned long managed_pages = zone->managed_pages;
5721 zone->lowmem_reserve[j] = 0;
5725 struct zone *lower_zone;
5729 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5730 sysctl_lowmem_reserve_ratio[idx] = 1;
5732 lower_zone = pgdat->node_zones + idx;
5733 lower_zone->lowmem_reserve[j] = managed_pages /
5734 sysctl_lowmem_reserve_ratio[idx];
5735 managed_pages += lower_zone->managed_pages;
5740 /* update totalreserve_pages */
5741 calculate_totalreserve_pages();
5744 static void __setup_per_zone_wmarks(void)
5746 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5747 unsigned long lowmem_pages = 0;
5749 unsigned long flags;
5751 /* Calculate total number of !ZONE_HIGHMEM pages */
5752 for_each_zone(zone) {
5753 if (!is_highmem(zone))
5754 lowmem_pages += zone->managed_pages;
5757 for_each_zone(zone) {
5760 spin_lock_irqsave(&zone->lock, flags);
5761 tmp = (u64)pages_min * zone->managed_pages;
5762 do_div(tmp, lowmem_pages);
5763 if (is_highmem(zone)) {
5765 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5766 * need highmem pages, so cap pages_min to a small
5769 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5770 * deltas controls asynch page reclaim, and so should
5771 * not be capped for highmem.
5773 unsigned long min_pages;
5775 min_pages = zone->managed_pages / 1024;
5776 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5777 zone->watermark[WMARK_MIN] = min_pages;
5780 * If it's a lowmem zone, reserve a number of pages
5781 * proportionate to the zone's size.
5783 zone->watermark[WMARK_MIN] = tmp;
5786 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5787 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5789 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5790 high_wmark_pages(zone) - low_wmark_pages(zone) -
5791 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5793 setup_zone_migrate_reserve(zone);
5794 spin_unlock_irqrestore(&zone->lock, flags);
5797 /* update totalreserve_pages */
5798 calculate_totalreserve_pages();
5802 * setup_per_zone_wmarks - called when min_free_kbytes changes
5803 * or when memory is hot-{added|removed}
5805 * Ensures that the watermark[min,low,high] values for each zone are set
5806 * correctly with respect to min_free_kbytes.
5808 void setup_per_zone_wmarks(void)
5810 mutex_lock(&zonelists_mutex);
5811 __setup_per_zone_wmarks();
5812 mutex_unlock(&zonelists_mutex);
5816 * The inactive anon list should be small enough that the VM never has to
5817 * do too much work, but large enough that each inactive page has a chance
5818 * to be referenced again before it is swapped out.
5820 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5821 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5822 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5823 * the anonymous pages are kept on the inactive list.
5826 * memory ratio inactive anon
5827 * -------------------------------------
5836 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5838 unsigned int gb, ratio;
5840 /* Zone size in gigabytes */
5841 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5843 ratio = int_sqrt(10 * gb);
5847 zone->inactive_ratio = ratio;
5850 static void __meminit setup_per_zone_inactive_ratio(void)
5855 calculate_zone_inactive_ratio(zone);
5859 * Initialise min_free_kbytes.
5861 * For small machines we want it small (128k min). For large machines
5862 * we want it large (64MB max). But it is not linear, because network
5863 * bandwidth does not increase linearly with machine size. We use
5865 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5866 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5882 int __meminit init_per_zone_wmark_min(void)
5884 unsigned long lowmem_kbytes;
5885 int new_min_free_kbytes;
5887 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5888 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5890 if (new_min_free_kbytes > user_min_free_kbytes) {
5891 min_free_kbytes = new_min_free_kbytes;
5892 if (min_free_kbytes < 128)
5893 min_free_kbytes = 128;
5894 if (min_free_kbytes > 65536)
5895 min_free_kbytes = 65536;
5897 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5898 new_min_free_kbytes, user_min_free_kbytes);
5900 setup_per_zone_wmarks();
5901 refresh_zone_stat_thresholds();
5902 setup_per_zone_lowmem_reserve();
5903 setup_per_zone_inactive_ratio();
5906 module_init(init_per_zone_wmark_min)
5909 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5910 * that we can call two helper functions whenever min_free_kbytes
5913 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5914 void __user *buffer, size_t *length, loff_t *ppos)
5918 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5923 user_min_free_kbytes = min_free_kbytes;
5924 setup_per_zone_wmarks();
5930 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5931 void __user *buffer, size_t *length, loff_t *ppos)
5936 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5941 zone->min_unmapped_pages = (zone->managed_pages *
5942 sysctl_min_unmapped_ratio) / 100;
5946 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5947 void __user *buffer, size_t *length, loff_t *ppos)
5952 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5957 zone->min_slab_pages = (zone->managed_pages *
5958 sysctl_min_slab_ratio) / 100;
5964 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5965 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5966 * whenever sysctl_lowmem_reserve_ratio changes.
5968 * The reserve ratio obviously has absolutely no relation with the
5969 * minimum watermarks. The lowmem reserve ratio can only make sense
5970 * if in function of the boot time zone sizes.
5972 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5973 void __user *buffer, size_t *length, loff_t *ppos)
5975 proc_dointvec_minmax(table, write, buffer, length, ppos);
5976 setup_per_zone_lowmem_reserve();
5981 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5982 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5983 * pagelist can have before it gets flushed back to buddy allocator.
5985 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5986 void __user *buffer, size_t *length, loff_t *ppos)
5989 int old_percpu_pagelist_fraction;
5992 mutex_lock(&pcp_batch_high_lock);
5993 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5995 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5996 if (!write || ret < 0)
5999 /* Sanity checking to avoid pcp imbalance */
6000 if (percpu_pagelist_fraction &&
6001 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6002 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6008 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6011 for_each_populated_zone(zone) {
6014 for_each_possible_cpu(cpu)
6015 pageset_set_high_and_batch(zone,
6016 per_cpu_ptr(zone->pageset, cpu));
6019 mutex_unlock(&pcp_batch_high_lock);
6023 int hashdist = HASHDIST_DEFAULT;
6026 static int __init set_hashdist(char *str)
6030 hashdist = simple_strtoul(str, &str, 0);
6033 __setup("hashdist=", set_hashdist);
6037 * allocate a large system hash table from bootmem
6038 * - it is assumed that the hash table must contain an exact power-of-2
6039 * quantity of entries
6040 * - limit is the number of hash buckets, not the total allocation size
6042 void *__init alloc_large_system_hash(const char *tablename,
6043 unsigned long bucketsize,
6044 unsigned long numentries,
6047 unsigned int *_hash_shift,
6048 unsigned int *_hash_mask,
6049 unsigned long low_limit,
6050 unsigned long high_limit)
6052 unsigned long long max = high_limit;
6053 unsigned long log2qty, size;
6056 /* allow the kernel cmdline to have a say */
6058 /* round applicable memory size up to nearest megabyte */
6059 numentries = nr_kernel_pages;
6061 /* It isn't necessary when PAGE_SIZE >= 1MB */
6062 if (PAGE_SHIFT < 20)
6063 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6065 /* limit to 1 bucket per 2^scale bytes of low memory */
6066 if (scale > PAGE_SHIFT)
6067 numentries >>= (scale - PAGE_SHIFT);
6069 numentries <<= (PAGE_SHIFT - scale);
6071 /* Make sure we've got at least a 0-order allocation.. */
6072 if (unlikely(flags & HASH_SMALL)) {
6073 /* Makes no sense without HASH_EARLY */
6074 WARN_ON(!(flags & HASH_EARLY));
6075 if (!(numentries >> *_hash_shift)) {
6076 numentries = 1UL << *_hash_shift;
6077 BUG_ON(!numentries);
6079 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6080 numentries = PAGE_SIZE / bucketsize;
6082 numentries = roundup_pow_of_two(numentries);
6084 /* limit allocation size to 1/16 total memory by default */
6086 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6087 do_div(max, bucketsize);
6089 max = min(max, 0x80000000ULL);
6091 if (numentries < low_limit)
6092 numentries = low_limit;
6093 if (numentries > max)
6096 log2qty = ilog2(numentries);
6099 size = bucketsize << log2qty;
6100 if (flags & HASH_EARLY)
6101 table = memblock_virt_alloc_nopanic(size, 0);
6103 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6106 * If bucketsize is not a power-of-two, we may free
6107 * some pages at the end of hash table which
6108 * alloc_pages_exact() automatically does
6110 if (get_order(size) < MAX_ORDER) {
6111 table = alloc_pages_exact(size, GFP_ATOMIC);
6112 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6115 } while (!table && size > PAGE_SIZE && --log2qty);
6118 panic("Failed to allocate %s hash table\n", tablename);
6120 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6123 ilog2(size) - PAGE_SHIFT,
6127 *_hash_shift = log2qty;
6129 *_hash_mask = (1 << log2qty) - 1;
6134 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6135 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6138 #ifdef CONFIG_SPARSEMEM
6139 return __pfn_to_section(pfn)->pageblock_flags;
6141 return zone->pageblock_flags;
6142 #endif /* CONFIG_SPARSEMEM */
6145 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6147 #ifdef CONFIG_SPARSEMEM
6148 pfn &= (PAGES_PER_SECTION-1);
6149 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6151 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6152 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6153 #endif /* CONFIG_SPARSEMEM */
6157 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6158 * @page: The page within the block of interest
6159 * @pfn: The target page frame number
6160 * @end_bitidx: The last bit of interest to retrieve
6161 * @mask: mask of bits that the caller is interested in
6163 * Return: pageblock_bits flags
6165 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6166 unsigned long end_bitidx,
6170 unsigned long *bitmap;
6171 unsigned long bitidx, word_bitidx;
6174 zone = page_zone(page);
6175 bitmap = get_pageblock_bitmap(zone, pfn);
6176 bitidx = pfn_to_bitidx(zone, pfn);
6177 word_bitidx = bitidx / BITS_PER_LONG;
6178 bitidx &= (BITS_PER_LONG-1);
6180 word = bitmap[word_bitidx];
6181 bitidx += end_bitidx;
6182 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6186 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6187 * @page: The page within the block of interest
6188 * @flags: The flags to set
6189 * @pfn: The target page frame number
6190 * @end_bitidx: The last bit of interest
6191 * @mask: mask of bits that the caller is interested in
6193 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6195 unsigned long end_bitidx,
6199 unsigned long *bitmap;
6200 unsigned long bitidx, word_bitidx;
6201 unsigned long old_word, word;
6203 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6205 zone = page_zone(page);
6206 bitmap = get_pageblock_bitmap(zone, pfn);
6207 bitidx = pfn_to_bitidx(zone, pfn);
6208 word_bitidx = bitidx / BITS_PER_LONG;
6209 bitidx &= (BITS_PER_LONG-1);
6211 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6213 bitidx += end_bitidx;
6214 mask <<= (BITS_PER_LONG - bitidx - 1);
6215 flags <<= (BITS_PER_LONG - bitidx - 1);
6217 word = ACCESS_ONCE(bitmap[word_bitidx]);
6219 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6220 if (word == old_word)
6227 * This function checks whether pageblock includes unmovable pages or not.
6228 * If @count is not zero, it is okay to include less @count unmovable pages
6230 * PageLRU check without isolation or lru_lock could race so that
6231 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6232 * expect this function should be exact.
6234 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6235 bool skip_hwpoisoned_pages)
6237 unsigned long pfn, iter, found;
6241 * For avoiding noise data, lru_add_drain_all() should be called
6242 * If ZONE_MOVABLE, the zone never contains unmovable pages
6244 if (zone_idx(zone) == ZONE_MOVABLE)
6246 mt = get_pageblock_migratetype(page);
6247 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6250 pfn = page_to_pfn(page);
6251 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6252 unsigned long check = pfn + iter;
6254 if (!pfn_valid_within(check))
6257 page = pfn_to_page(check);
6260 * Hugepages are not in LRU lists, but they're movable.
6261 * We need not scan over tail pages bacause we don't
6262 * handle each tail page individually in migration.
6264 if (PageHuge(page)) {
6265 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6270 * We can't use page_count without pin a page
6271 * because another CPU can free compound page.
6272 * This check already skips compound tails of THP
6273 * because their page->_count is zero at all time.
6275 if (!atomic_read(&page->_count)) {
6276 if (PageBuddy(page))
6277 iter += (1 << page_order(page)) - 1;
6282 * The HWPoisoned page may be not in buddy system, and
6283 * page_count() is not 0.
6285 if (skip_hwpoisoned_pages && PageHWPoison(page))
6291 * If there are RECLAIMABLE pages, we need to check
6292 * it. But now, memory offline itself doesn't call
6293 * shrink_node_slabs() and it still to be fixed.
6296 * If the page is not RAM, page_count()should be 0.
6297 * we don't need more check. This is an _used_ not-movable page.
6299 * The problematic thing here is PG_reserved pages. PG_reserved
6300 * is set to both of a memory hole page and a _used_ kernel
6309 bool is_pageblock_removable_nolock(struct page *page)
6315 * We have to be careful here because we are iterating over memory
6316 * sections which are not zone aware so we might end up outside of
6317 * the zone but still within the section.
6318 * We have to take care about the node as well. If the node is offline
6319 * its NODE_DATA will be NULL - see page_zone.
6321 if (!node_online(page_to_nid(page)))
6324 zone = page_zone(page);
6325 pfn = page_to_pfn(page);
6326 if (!zone_spans_pfn(zone, pfn))
6329 return !has_unmovable_pages(zone, page, 0, true);
6334 static unsigned long pfn_max_align_down(unsigned long pfn)
6336 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6337 pageblock_nr_pages) - 1);
6340 static unsigned long pfn_max_align_up(unsigned long pfn)
6342 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6343 pageblock_nr_pages));
6346 /* [start, end) must belong to a single zone. */
6347 static int __alloc_contig_migrate_range(struct compact_control *cc,
6348 unsigned long start, unsigned long end)
6350 /* This function is based on compact_zone() from compaction.c. */
6351 unsigned long nr_reclaimed;
6352 unsigned long pfn = start;
6353 unsigned int tries = 0;
6358 while (pfn < end || !list_empty(&cc->migratepages)) {
6359 if (fatal_signal_pending(current)) {
6364 if (list_empty(&cc->migratepages)) {
6365 cc->nr_migratepages = 0;
6366 pfn = isolate_migratepages_range(cc, pfn, end);
6372 } else if (++tries == 5) {
6373 ret = ret < 0 ? ret : -EBUSY;
6377 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6379 cc->nr_migratepages -= nr_reclaimed;
6381 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6382 NULL, 0, cc->mode, MR_CMA);
6385 putback_movable_pages(&cc->migratepages);
6392 * alloc_contig_range() -- tries to allocate given range of pages
6393 * @start: start PFN to allocate
6394 * @end: one-past-the-last PFN to allocate
6395 * @migratetype: migratetype of the underlaying pageblocks (either
6396 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6397 * in range must have the same migratetype and it must
6398 * be either of the two.
6400 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6401 * aligned, however it's the caller's responsibility to guarantee that
6402 * we are the only thread that changes migrate type of pageblocks the
6405 * The PFN range must belong to a single zone.
6407 * Returns zero on success or negative error code. On success all
6408 * pages which PFN is in [start, end) are allocated for the caller and
6409 * need to be freed with free_contig_range().
6411 int alloc_contig_range(unsigned long start, unsigned long end,
6412 unsigned migratetype)
6414 unsigned long outer_start, outer_end;
6417 struct compact_control cc = {
6418 .nr_migratepages = 0,
6420 .zone = page_zone(pfn_to_page(start)),
6421 .mode = MIGRATE_SYNC,
6422 .ignore_skip_hint = true,
6424 INIT_LIST_HEAD(&cc.migratepages);
6427 * What we do here is we mark all pageblocks in range as
6428 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6429 * have different sizes, and due to the way page allocator
6430 * work, we align the range to biggest of the two pages so
6431 * that page allocator won't try to merge buddies from
6432 * different pageblocks and change MIGRATE_ISOLATE to some
6433 * other migration type.
6435 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6436 * migrate the pages from an unaligned range (ie. pages that
6437 * we are interested in). This will put all the pages in
6438 * range back to page allocator as MIGRATE_ISOLATE.
6440 * When this is done, we take the pages in range from page
6441 * allocator removing them from the buddy system. This way
6442 * page allocator will never consider using them.
6444 * This lets us mark the pageblocks back as
6445 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6446 * aligned range but not in the unaligned, original range are
6447 * put back to page allocator so that buddy can use them.
6450 ret = start_isolate_page_range(pfn_max_align_down(start),
6451 pfn_max_align_up(end), migratetype,
6456 ret = __alloc_contig_migrate_range(&cc, start, end);
6461 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6462 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6463 * more, all pages in [start, end) are free in page allocator.
6464 * What we are going to do is to allocate all pages from
6465 * [start, end) (that is remove them from page allocator).
6467 * The only problem is that pages at the beginning and at the
6468 * end of interesting range may be not aligned with pages that
6469 * page allocator holds, ie. they can be part of higher order
6470 * pages. Because of this, we reserve the bigger range and
6471 * once this is done free the pages we are not interested in.
6473 * We don't have to hold zone->lock here because the pages are
6474 * isolated thus they won't get removed from buddy.
6477 lru_add_drain_all();
6478 drain_all_pages(cc.zone);
6481 outer_start = start;
6482 while (!PageBuddy(pfn_to_page(outer_start))) {
6483 if (++order >= MAX_ORDER) {
6487 outer_start &= ~0UL << order;
6490 /* Make sure the range is really isolated. */
6491 if (test_pages_isolated(outer_start, end, false)) {
6492 pr_info("%s: [%lx, %lx) PFNs busy\n",
6493 __func__, outer_start, end);
6498 /* Grab isolated pages from freelists. */
6499 outer_end = isolate_freepages_range(&cc, outer_start, end);
6505 /* Free head and tail (if any) */
6506 if (start != outer_start)
6507 free_contig_range(outer_start, start - outer_start);
6508 if (end != outer_end)
6509 free_contig_range(end, outer_end - end);
6512 undo_isolate_page_range(pfn_max_align_down(start),
6513 pfn_max_align_up(end), migratetype);
6517 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6519 unsigned int count = 0;
6521 for (; nr_pages--; pfn++) {
6522 struct page *page = pfn_to_page(pfn);
6524 count += page_count(page) != 1;
6527 WARN(count != 0, "%d pages are still in use!\n", count);
6531 #ifdef CONFIG_MEMORY_HOTPLUG
6533 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6534 * page high values need to be recalulated.
6536 void __meminit zone_pcp_update(struct zone *zone)
6539 mutex_lock(&pcp_batch_high_lock);
6540 for_each_possible_cpu(cpu)
6541 pageset_set_high_and_batch(zone,
6542 per_cpu_ptr(zone->pageset, cpu));
6543 mutex_unlock(&pcp_batch_high_lock);
6547 void zone_pcp_reset(struct zone *zone)
6549 unsigned long flags;
6551 struct per_cpu_pageset *pset;
6553 /* avoid races with drain_pages() */
6554 local_irq_save(flags);
6555 if (zone->pageset != &boot_pageset) {
6556 for_each_online_cpu(cpu) {
6557 pset = per_cpu_ptr(zone->pageset, cpu);
6558 drain_zonestat(zone, pset);
6560 free_percpu(zone->pageset);
6561 zone->pageset = &boot_pageset;
6563 local_irq_restore(flags);
6566 #ifdef CONFIG_MEMORY_HOTREMOVE
6568 * All pages in the range must be isolated before calling this.
6571 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6575 unsigned int order, i;
6577 unsigned long flags;
6578 /* find the first valid pfn */
6579 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6584 zone = page_zone(pfn_to_page(pfn));
6585 spin_lock_irqsave(&zone->lock, flags);
6587 while (pfn < end_pfn) {
6588 if (!pfn_valid(pfn)) {
6592 page = pfn_to_page(pfn);
6594 * The HWPoisoned page may be not in buddy system, and
6595 * page_count() is not 0.
6597 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6599 SetPageReserved(page);
6603 BUG_ON(page_count(page));
6604 BUG_ON(!PageBuddy(page));
6605 order = page_order(page);
6606 #ifdef CONFIG_DEBUG_VM
6607 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6608 pfn, 1 << order, end_pfn);
6610 list_del(&page->lru);
6611 rmv_page_order(page);
6612 zone->free_area[order].nr_free--;
6613 for (i = 0; i < (1 << order); i++)
6614 SetPageReserved((page+i));
6615 pfn += (1 << order);
6617 spin_unlock_irqrestore(&zone->lock, flags);
6621 #ifdef CONFIG_MEMORY_FAILURE
6622 bool is_free_buddy_page(struct page *page)
6624 struct zone *zone = page_zone(page);
6625 unsigned long pfn = page_to_pfn(page);
6626 unsigned long flags;
6629 spin_lock_irqsave(&zone->lock, flags);
6630 for (order = 0; order < MAX_ORDER; order++) {
6631 struct page *page_head = page - (pfn & ((1 << order) - 1));
6633 if (PageBuddy(page_head) && page_order(page_head) >= order)
6636 spin_unlock_irqrestore(&zone->lock, flags);
6638 return order < MAX_ORDER;