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_cgroup.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/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node);
76 EXPORT_PER_CPU_SYMBOL(numa_node);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
88 int _node_numa_mem_[MAX_NUMNODES];
92 * Array of node states.
94 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
95 [N_POSSIBLE] = NODE_MASK_ALL,
96 [N_ONLINE] = { { [0] = 1UL } },
98 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 [N_HIGH_MEMORY] = { { [0] = 1UL } },
102 #ifdef CONFIG_MOVABLE_NODE
103 [N_MEMORY] = { { [0] = 1UL } },
105 [N_CPU] = { { [0] = 1UL } },
108 EXPORT_SYMBOL(node_states);
110 /* Protect totalram_pages and zone->managed_pages */
111 static DEFINE_SPINLOCK(managed_page_count_lock);
113 unsigned long totalram_pages __read_mostly;
114 unsigned long totalreserve_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;
431 static int __init debug_guardpage_minorder_setup(char *buf)
435 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
436 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
439 _debug_guardpage_minorder = res;
440 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
443 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
445 static inline void set_page_guard_flag(struct page *page)
447 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
450 static inline void clear_page_guard_flag(struct page *page)
452 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
455 static inline void set_page_guard_flag(struct page *page) { }
456 static inline void clear_page_guard_flag(struct page *page) { }
459 static inline void set_page_order(struct page *page, unsigned int order)
461 set_page_private(page, order);
462 __SetPageBuddy(page);
465 static inline void rmv_page_order(struct page *page)
467 __ClearPageBuddy(page);
468 set_page_private(page, 0);
472 * Locate the struct page for both the matching buddy in our
473 * pair (buddy1) and the combined O(n+1) page they form (page).
475 * 1) Any buddy B1 will have an order O twin B2 which satisfies
476 * the following equation:
478 * For example, if the starting buddy (buddy2) is #8 its order
480 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
482 * 2) Any buddy B will have an order O+1 parent P which
483 * satisfies the following equation:
486 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
488 static inline unsigned long
489 __find_buddy_index(unsigned long page_idx, unsigned int order)
491 return page_idx ^ (1 << order);
495 * This function checks whether a page is free && is the buddy
496 * we can do coalesce a page and its buddy if
497 * (a) the buddy is not in a hole &&
498 * (b) the buddy is in the buddy system &&
499 * (c) a page and its buddy have the same order &&
500 * (d) a page and its buddy are in the same zone.
502 * For recording whether a page is in the buddy system, we set ->_mapcount
503 * PAGE_BUDDY_MAPCOUNT_VALUE.
504 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
505 * serialized by zone->lock.
507 * For recording page's order, we use page_private(page).
509 static inline int page_is_buddy(struct page *page, struct page *buddy,
512 if (!pfn_valid_within(page_to_pfn(buddy)))
515 if (page_is_guard(buddy) && page_order(buddy) == order) {
516 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
518 if (page_zone_id(page) != page_zone_id(buddy))
524 if (PageBuddy(buddy) && page_order(buddy) == order) {
525 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
528 * zone check is done late to avoid uselessly
529 * calculating zone/node ids for pages that could
532 if (page_zone_id(page) != page_zone_id(buddy))
541 * Freeing function for a buddy system allocator.
543 * The concept of a buddy system is to maintain direct-mapped table
544 * (containing bit values) for memory blocks of various "orders".
545 * The bottom level table contains the map for the smallest allocatable
546 * units of memory (here, pages), and each level above it describes
547 * pairs of units from the levels below, hence, "buddies".
548 * At a high level, all that happens here is marking the table entry
549 * at the bottom level available, and propagating the changes upward
550 * as necessary, plus some accounting needed to play nicely with other
551 * parts of the VM system.
552 * At each level, we keep a list of pages, which are heads of continuous
553 * free pages of length of (1 << order) and marked with _mapcount
554 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
556 * So when we are allocating or freeing one, we can derive the state of the
557 * other. That is, if we allocate a small block, and both were
558 * free, the remainder of the region must be split into blocks.
559 * If a block is freed, and its buddy is also free, then this
560 * triggers coalescing into a block of larger size.
565 static inline void __free_one_page(struct page *page,
567 struct zone *zone, unsigned int order,
570 unsigned long page_idx;
571 unsigned long combined_idx;
572 unsigned long uninitialized_var(buddy_idx);
575 VM_BUG_ON(!zone_is_initialized(zone));
577 if (unlikely(PageCompound(page)))
578 if (unlikely(destroy_compound_page(page, order)))
581 VM_BUG_ON(migratetype == -1);
583 page_idx = pfn & ((1 << MAX_ORDER) - 1);
585 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
586 VM_BUG_ON_PAGE(bad_range(zone, page), page);
588 while (order < MAX_ORDER-1) {
589 buddy_idx = __find_buddy_index(page_idx, order);
590 buddy = page + (buddy_idx - page_idx);
591 if (!page_is_buddy(page, buddy, order))
594 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
595 * merge with it and move up one order.
597 if (page_is_guard(buddy)) {
598 clear_page_guard_flag(buddy);
599 set_page_private(page, 0);
600 __mod_zone_freepage_state(zone, 1 << order,
603 list_del(&buddy->lru);
604 zone->free_area[order].nr_free--;
605 rmv_page_order(buddy);
607 combined_idx = buddy_idx & page_idx;
608 page = page + (combined_idx - page_idx);
609 page_idx = combined_idx;
612 set_page_order(page, order);
615 * If this is not the largest possible page, check if the buddy
616 * of the next-highest order is free. If it is, it's possible
617 * that pages are being freed that will coalesce soon. In case,
618 * that is happening, add the free page to the tail of the list
619 * so it's less likely to be used soon and more likely to be merged
620 * as a higher order page
622 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
623 struct page *higher_page, *higher_buddy;
624 combined_idx = buddy_idx & page_idx;
625 higher_page = page + (combined_idx - page_idx);
626 buddy_idx = __find_buddy_index(combined_idx, order + 1);
627 higher_buddy = higher_page + (buddy_idx - combined_idx);
628 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
629 list_add_tail(&page->lru,
630 &zone->free_area[order].free_list[migratetype]);
635 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
637 zone->free_area[order].nr_free++;
640 static inline int free_pages_check(struct page *page)
642 const char *bad_reason = NULL;
643 unsigned long bad_flags = 0;
645 if (unlikely(page_mapcount(page)))
646 bad_reason = "nonzero mapcount";
647 if (unlikely(page->mapping != NULL))
648 bad_reason = "non-NULL mapping";
649 if (unlikely(atomic_read(&page->_count) != 0))
650 bad_reason = "nonzero _count";
651 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
652 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
653 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
655 if (unlikely(mem_cgroup_bad_page_check(page)))
656 bad_reason = "cgroup check failed";
657 if (unlikely(bad_reason)) {
658 bad_page(page, bad_reason, bad_flags);
661 page_cpupid_reset_last(page);
662 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
663 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
668 * Frees a number of pages from the PCP lists
669 * Assumes all pages on list are in same zone, and of same order.
670 * count is the number of pages to free.
672 * If the zone was previously in an "all pages pinned" state then look to
673 * see if this freeing clears that state.
675 * And clear the zone's pages_scanned counter, to hold off the "all pages are
676 * pinned" detection logic.
678 static void free_pcppages_bulk(struct zone *zone, int count,
679 struct per_cpu_pages *pcp)
684 unsigned long nr_scanned;
686 spin_lock(&zone->lock);
687 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
689 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
693 struct list_head *list;
696 * Remove pages from lists in a round-robin fashion. A
697 * batch_free count is maintained that is incremented when an
698 * empty list is encountered. This is so more pages are freed
699 * off fuller lists instead of spinning excessively around empty
704 if (++migratetype == MIGRATE_PCPTYPES)
706 list = &pcp->lists[migratetype];
707 } while (list_empty(list));
709 /* This is the only non-empty list. Free them all. */
710 if (batch_free == MIGRATE_PCPTYPES)
711 batch_free = to_free;
714 int mt; /* migratetype of the to-be-freed page */
716 page = list_entry(list->prev, struct page, lru);
717 /* must delete as __free_one_page list manipulates */
718 list_del(&page->lru);
719 mt = get_freepage_migratetype(page);
720 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
721 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
722 trace_mm_page_pcpu_drain(page, 0, mt);
723 if (likely(!is_migrate_isolate_page(page))) {
724 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
725 if (is_migrate_cma(mt))
726 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
728 } while (--to_free && --batch_free && !list_empty(list));
730 spin_unlock(&zone->lock);
733 static void free_one_page(struct zone *zone,
734 struct page *page, unsigned long pfn,
738 unsigned long nr_scanned;
739 spin_lock(&zone->lock);
740 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
742 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
744 __free_one_page(page, pfn, zone, order, migratetype);
745 if (unlikely(!is_migrate_isolate(migratetype)))
746 __mod_zone_freepage_state(zone, 1 << order, migratetype);
747 spin_unlock(&zone->lock);
750 static bool free_pages_prepare(struct page *page, unsigned int order)
755 trace_mm_page_free(page, order);
756 kmemcheck_free_shadow(page, order);
759 page->mapping = NULL;
760 for (i = 0; i < (1 << order); i++)
761 bad += free_pages_check(page + i);
765 if (!PageHighMem(page)) {
766 debug_check_no_locks_freed(page_address(page),
768 debug_check_no_obj_freed(page_address(page),
771 arch_free_page(page, order);
772 kernel_map_pages(page, 1 << order, 0);
777 static void __free_pages_ok(struct page *page, unsigned int order)
781 unsigned long pfn = page_to_pfn(page);
783 if (!free_pages_prepare(page, order))
786 migratetype = get_pfnblock_migratetype(page, pfn);
787 local_irq_save(flags);
788 __count_vm_events(PGFREE, 1 << order);
789 set_freepage_migratetype(page, migratetype);
790 free_one_page(page_zone(page), page, pfn, order, migratetype);
791 local_irq_restore(flags);
794 void __init __free_pages_bootmem(struct page *page, unsigned int order)
796 unsigned int nr_pages = 1 << order;
797 struct page *p = page;
801 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
803 __ClearPageReserved(p);
804 set_page_count(p, 0);
806 __ClearPageReserved(p);
807 set_page_count(p, 0);
809 page_zone(page)->managed_pages += nr_pages;
810 set_page_refcounted(page);
811 __free_pages(page, order);
815 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
816 void __init init_cma_reserved_pageblock(struct page *page)
818 unsigned i = pageblock_nr_pages;
819 struct page *p = page;
822 __ClearPageReserved(p);
823 set_page_count(p, 0);
826 set_pageblock_migratetype(page, MIGRATE_CMA);
828 if (pageblock_order >= MAX_ORDER) {
829 i = pageblock_nr_pages;
832 set_page_refcounted(p);
833 __free_pages(p, MAX_ORDER - 1);
834 p += MAX_ORDER_NR_PAGES;
835 } while (i -= MAX_ORDER_NR_PAGES);
837 set_page_refcounted(page);
838 __free_pages(page, pageblock_order);
841 adjust_managed_page_count(page, pageblock_nr_pages);
846 * The order of subdivision here is critical for the IO subsystem.
847 * Please do not alter this order without good reasons and regression
848 * testing. Specifically, as large blocks of memory are subdivided,
849 * the order in which smaller blocks are delivered depends on the order
850 * they're subdivided in this function. This is the primary factor
851 * influencing the order in which pages are delivered to the IO
852 * subsystem according to empirical testing, and this is also justified
853 * by considering the behavior of a buddy system containing a single
854 * large block of memory acted on by a series of small allocations.
855 * This behavior is a critical factor in sglist merging's success.
859 static inline void expand(struct zone *zone, struct page *page,
860 int low, int high, struct free_area *area,
863 unsigned long size = 1 << high;
869 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
871 #ifdef CONFIG_DEBUG_PAGEALLOC
872 if (high < debug_guardpage_minorder()) {
874 * Mark as guard pages (or page), that will allow to
875 * merge back to allocator when buddy will be freed.
876 * Corresponding page table entries will not be touched,
877 * pages will stay not present in virtual address space
879 INIT_LIST_HEAD(&page[size].lru);
880 set_page_guard_flag(&page[size]);
881 set_page_private(&page[size], high);
882 /* Guard pages are not available for any usage */
883 __mod_zone_freepage_state(zone, -(1 << high),
888 list_add(&page[size].lru, &area->free_list[migratetype]);
890 set_page_order(&page[size], high);
895 * This page is about to be returned from the page allocator
897 static inline int check_new_page(struct page *page)
899 const char *bad_reason = NULL;
900 unsigned long bad_flags = 0;
902 if (unlikely(page_mapcount(page)))
903 bad_reason = "nonzero mapcount";
904 if (unlikely(page->mapping != NULL))
905 bad_reason = "non-NULL mapping";
906 if (unlikely(atomic_read(&page->_count) != 0))
907 bad_reason = "nonzero _count";
908 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
909 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
910 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
912 if (unlikely(mem_cgroup_bad_page_check(page)))
913 bad_reason = "cgroup check failed";
914 if (unlikely(bad_reason)) {
915 bad_page(page, bad_reason, bad_flags);
921 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags)
925 for (i = 0; i < (1 << order); i++) {
926 struct page *p = page + i;
927 if (unlikely(check_new_page(p)))
931 set_page_private(page, 0);
932 set_page_refcounted(page);
934 arch_alloc_page(page, order);
935 kernel_map_pages(page, 1 << order, 1);
937 if (gfp_flags & __GFP_ZERO)
938 prep_zero_page(page, order, gfp_flags);
940 if (order && (gfp_flags & __GFP_COMP))
941 prep_compound_page(page, order);
947 * Go through the free lists for the given migratetype and remove
948 * the smallest available page from the freelists
951 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
954 unsigned int current_order;
955 struct free_area *area;
958 /* Find a page of the appropriate size in the preferred list */
959 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
960 area = &(zone->free_area[current_order]);
961 if (list_empty(&area->free_list[migratetype]))
964 page = list_entry(area->free_list[migratetype].next,
966 list_del(&page->lru);
967 rmv_page_order(page);
969 expand(zone, page, order, current_order, area, migratetype);
970 set_freepage_migratetype(page, migratetype);
979 * This array describes the order lists are fallen back to when
980 * the free lists for the desirable migrate type are depleted
982 static int fallbacks[MIGRATE_TYPES][4] = {
983 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
984 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
986 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
987 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
989 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
991 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
992 #ifdef CONFIG_MEMORY_ISOLATION
993 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
998 * Move the free pages in a range to the free lists of the requested type.
999 * Note that start_page and end_pages are not aligned on a pageblock
1000 * boundary. If alignment is required, use move_freepages_block()
1002 int move_freepages(struct zone *zone,
1003 struct page *start_page, struct page *end_page,
1007 unsigned long order;
1008 int pages_moved = 0;
1010 #ifndef CONFIG_HOLES_IN_ZONE
1012 * page_zone is not safe to call in this context when
1013 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1014 * anyway as we check zone boundaries in move_freepages_block().
1015 * Remove at a later date when no bug reports exist related to
1016 * grouping pages by mobility
1018 BUG_ON(page_zone(start_page) != page_zone(end_page));
1021 for (page = start_page; page <= end_page;) {
1022 /* Make sure we are not inadvertently changing nodes */
1023 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1025 if (!pfn_valid_within(page_to_pfn(page))) {
1030 if (!PageBuddy(page)) {
1035 order = page_order(page);
1036 list_move(&page->lru,
1037 &zone->free_area[order].free_list[migratetype]);
1038 set_freepage_migratetype(page, migratetype);
1040 pages_moved += 1 << order;
1046 int move_freepages_block(struct zone *zone, struct page *page,
1049 unsigned long start_pfn, end_pfn;
1050 struct page *start_page, *end_page;
1052 start_pfn = page_to_pfn(page);
1053 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1054 start_page = pfn_to_page(start_pfn);
1055 end_page = start_page + pageblock_nr_pages - 1;
1056 end_pfn = start_pfn + pageblock_nr_pages - 1;
1058 /* Do not cross zone boundaries */
1059 if (!zone_spans_pfn(zone, start_pfn))
1061 if (!zone_spans_pfn(zone, end_pfn))
1064 return move_freepages(zone, start_page, end_page, migratetype);
1067 static void change_pageblock_range(struct page *pageblock_page,
1068 int start_order, int migratetype)
1070 int nr_pageblocks = 1 << (start_order - pageblock_order);
1072 while (nr_pageblocks--) {
1073 set_pageblock_migratetype(pageblock_page, migratetype);
1074 pageblock_page += pageblock_nr_pages;
1079 * If breaking a large block of pages, move all free pages to the preferred
1080 * allocation list. If falling back for a reclaimable kernel allocation, be
1081 * more aggressive about taking ownership of free pages.
1083 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1084 * nor move CMA pages to different free lists. We don't want unmovable pages
1085 * to be allocated from MIGRATE_CMA areas.
1087 * Returns the new migratetype of the pageblock (or the same old migratetype
1088 * if it was unchanged).
1090 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1091 int start_type, int fallback_type)
1093 int current_order = page_order(page);
1096 * When borrowing from MIGRATE_CMA, we need to release the excess
1097 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1098 * is set to CMA so it is returned to the correct freelist in case
1099 * the page ends up being not actually allocated from the pcp lists.
1101 if (is_migrate_cma(fallback_type))
1102 return fallback_type;
1104 /* Take ownership for orders >= pageblock_order */
1105 if (current_order >= pageblock_order) {
1106 change_pageblock_range(page, current_order, start_type);
1110 if (current_order >= pageblock_order / 2 ||
1111 start_type == MIGRATE_RECLAIMABLE ||
1112 page_group_by_mobility_disabled) {
1115 pages = move_freepages_block(zone, page, start_type);
1117 /* Claim the whole block if over half of it is free */
1118 if (pages >= (1 << (pageblock_order-1)) ||
1119 page_group_by_mobility_disabled) {
1121 set_pageblock_migratetype(page, start_type);
1127 return fallback_type;
1130 /* Remove an element from the buddy allocator from the fallback list */
1131 static inline struct page *
1132 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1134 struct free_area *area;
1135 unsigned int current_order;
1137 int migratetype, new_type, i;
1139 /* Find the largest possible block of pages in the other list */
1140 for (current_order = MAX_ORDER-1;
1141 current_order >= order && current_order <= MAX_ORDER-1;
1144 migratetype = fallbacks[start_migratetype][i];
1146 /* MIGRATE_RESERVE handled later if necessary */
1147 if (migratetype == MIGRATE_RESERVE)
1150 area = &(zone->free_area[current_order]);
1151 if (list_empty(&area->free_list[migratetype]))
1154 page = list_entry(area->free_list[migratetype].next,
1158 new_type = try_to_steal_freepages(zone, page,
1162 /* Remove the page from the freelists */
1163 list_del(&page->lru);
1164 rmv_page_order(page);
1166 expand(zone, page, order, current_order, area,
1168 /* The freepage_migratetype may differ from pageblock's
1169 * migratetype depending on the decisions in
1170 * try_to_steal_freepages. This is OK as long as it does
1171 * not differ for MIGRATE_CMA type.
1173 set_freepage_migratetype(page, new_type);
1175 trace_mm_page_alloc_extfrag(page, order, current_order,
1176 start_migratetype, migratetype, new_type);
1186 * Do the hard work of removing an element from the buddy allocator.
1187 * Call me with the zone->lock already held.
1189 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1195 page = __rmqueue_smallest(zone, order, migratetype);
1197 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1198 page = __rmqueue_fallback(zone, order, migratetype);
1201 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1202 * is used because __rmqueue_smallest is an inline function
1203 * and we want just one call site
1206 migratetype = MIGRATE_RESERVE;
1211 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1216 * Obtain a specified number of elements from the buddy allocator, all under
1217 * a single hold of the lock, for efficiency. Add them to the supplied list.
1218 * Returns the number of new pages which were placed at *list.
1220 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1221 unsigned long count, struct list_head *list,
1222 int migratetype, bool cold)
1226 spin_lock(&zone->lock);
1227 for (i = 0; i < count; ++i) {
1228 struct page *page = __rmqueue(zone, order, migratetype);
1229 if (unlikely(page == NULL))
1233 * Split buddy pages returned by expand() are received here
1234 * in physical page order. The page is added to the callers and
1235 * list and the list head then moves forward. From the callers
1236 * perspective, the linked list is ordered by page number in
1237 * some conditions. This is useful for IO devices that can
1238 * merge IO requests if the physical pages are ordered
1242 list_add(&page->lru, list);
1244 list_add_tail(&page->lru, list);
1246 if (is_migrate_cma(get_freepage_migratetype(page)))
1247 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1250 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1251 spin_unlock(&zone->lock);
1257 * Called from the vmstat counter updater to drain pagesets of this
1258 * currently executing processor on remote nodes after they have
1261 * Note that this function must be called with the thread pinned to
1262 * a single processor.
1264 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1266 unsigned long flags;
1267 int to_drain, batch;
1269 local_irq_save(flags);
1270 batch = ACCESS_ONCE(pcp->batch);
1271 to_drain = min(pcp->count, batch);
1273 free_pcppages_bulk(zone, to_drain, pcp);
1274 pcp->count -= to_drain;
1276 local_irq_restore(flags);
1281 * Drain pages of the indicated processor.
1283 * The processor must either be the current processor and the
1284 * thread pinned to the current processor or a processor that
1287 static void drain_pages(unsigned int cpu)
1289 unsigned long flags;
1292 for_each_populated_zone(zone) {
1293 struct per_cpu_pageset *pset;
1294 struct per_cpu_pages *pcp;
1296 local_irq_save(flags);
1297 pset = per_cpu_ptr(zone->pageset, cpu);
1301 free_pcppages_bulk(zone, pcp->count, pcp);
1304 local_irq_restore(flags);
1309 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1311 void drain_local_pages(void *arg)
1313 drain_pages(smp_processor_id());
1317 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1319 * Note that this code is protected against sending an IPI to an offline
1320 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1321 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1322 * nothing keeps CPUs from showing up after we populated the cpumask and
1323 * before the call to on_each_cpu_mask().
1325 void drain_all_pages(void)
1328 struct per_cpu_pageset *pcp;
1332 * Allocate in the BSS so we wont require allocation in
1333 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1335 static cpumask_t cpus_with_pcps;
1338 * We don't care about racing with CPU hotplug event
1339 * as offline notification will cause the notified
1340 * cpu to drain that CPU pcps and on_each_cpu_mask
1341 * disables preemption as part of its processing
1343 for_each_online_cpu(cpu) {
1344 bool has_pcps = false;
1345 for_each_populated_zone(zone) {
1346 pcp = per_cpu_ptr(zone->pageset, cpu);
1347 if (pcp->pcp.count) {
1353 cpumask_set_cpu(cpu, &cpus_with_pcps);
1355 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1357 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1360 #ifdef CONFIG_HIBERNATION
1362 void mark_free_pages(struct zone *zone)
1364 unsigned long pfn, max_zone_pfn;
1365 unsigned long flags;
1366 unsigned int order, t;
1367 struct list_head *curr;
1369 if (zone_is_empty(zone))
1372 spin_lock_irqsave(&zone->lock, flags);
1374 max_zone_pfn = zone_end_pfn(zone);
1375 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1376 if (pfn_valid(pfn)) {
1377 struct page *page = pfn_to_page(pfn);
1379 if (!swsusp_page_is_forbidden(page))
1380 swsusp_unset_page_free(page);
1383 for_each_migratetype_order(order, t) {
1384 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1387 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1388 for (i = 0; i < (1UL << order); i++)
1389 swsusp_set_page_free(pfn_to_page(pfn + i));
1392 spin_unlock_irqrestore(&zone->lock, flags);
1394 #endif /* CONFIG_PM */
1397 * Free a 0-order page
1398 * cold == true ? free a cold page : free a hot page
1400 void free_hot_cold_page(struct page *page, bool cold)
1402 struct zone *zone = page_zone(page);
1403 struct per_cpu_pages *pcp;
1404 unsigned long flags;
1405 unsigned long pfn = page_to_pfn(page);
1408 if (!free_pages_prepare(page, 0))
1411 migratetype = get_pfnblock_migratetype(page, pfn);
1412 set_freepage_migratetype(page, migratetype);
1413 local_irq_save(flags);
1414 __count_vm_event(PGFREE);
1417 * We only track unmovable, reclaimable and movable on pcp lists.
1418 * Free ISOLATE pages back to the allocator because they are being
1419 * offlined but treat RESERVE as movable pages so we can get those
1420 * areas back if necessary. Otherwise, we may have to free
1421 * excessively into the page allocator
1423 if (migratetype >= MIGRATE_PCPTYPES) {
1424 if (unlikely(is_migrate_isolate(migratetype))) {
1425 free_one_page(zone, page, pfn, 0, migratetype);
1428 migratetype = MIGRATE_MOVABLE;
1431 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1433 list_add(&page->lru, &pcp->lists[migratetype]);
1435 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1437 if (pcp->count >= pcp->high) {
1438 unsigned long batch = ACCESS_ONCE(pcp->batch);
1439 free_pcppages_bulk(zone, batch, pcp);
1440 pcp->count -= batch;
1444 local_irq_restore(flags);
1448 * Free a list of 0-order pages
1450 void free_hot_cold_page_list(struct list_head *list, bool cold)
1452 struct page *page, *next;
1454 list_for_each_entry_safe(page, next, list, lru) {
1455 trace_mm_page_free_batched(page, cold);
1456 free_hot_cold_page(page, cold);
1461 * split_page takes a non-compound higher-order page, and splits it into
1462 * n (1<<order) sub-pages: page[0..n]
1463 * Each sub-page must be freed individually.
1465 * Note: this is probably too low level an operation for use in drivers.
1466 * Please consult with lkml before using this in your driver.
1468 void split_page(struct page *page, unsigned int order)
1472 VM_BUG_ON_PAGE(PageCompound(page), page);
1473 VM_BUG_ON_PAGE(!page_count(page), page);
1475 #ifdef CONFIG_KMEMCHECK
1477 * Split shadow pages too, because free(page[0]) would
1478 * otherwise free the whole shadow.
1480 if (kmemcheck_page_is_tracked(page))
1481 split_page(virt_to_page(page[0].shadow), order);
1484 for (i = 1; i < (1 << order); i++)
1485 set_page_refcounted(page + i);
1487 EXPORT_SYMBOL_GPL(split_page);
1489 static int __isolate_free_page(struct page *page, unsigned int order)
1491 unsigned long watermark;
1495 BUG_ON(!PageBuddy(page));
1497 zone = page_zone(page);
1498 mt = get_pageblock_migratetype(page);
1500 if (!is_migrate_isolate(mt)) {
1501 /* Obey watermarks as if the page was being allocated */
1502 watermark = low_wmark_pages(zone) + (1 << order);
1503 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1506 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1509 /* Remove page from free list */
1510 list_del(&page->lru);
1511 zone->free_area[order].nr_free--;
1512 rmv_page_order(page);
1514 /* Set the pageblock if the isolated page is at least a pageblock */
1515 if (order >= pageblock_order - 1) {
1516 struct page *endpage = page + (1 << order) - 1;
1517 for (; page < endpage; page += pageblock_nr_pages) {
1518 int mt = get_pageblock_migratetype(page);
1519 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1520 set_pageblock_migratetype(page,
1525 return 1UL << order;
1529 * Similar to split_page except the page is already free. As this is only
1530 * being used for migration, the migratetype of the block also changes.
1531 * As this is called with interrupts disabled, the caller is responsible
1532 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1535 * Note: this is probably too low level an operation for use in drivers.
1536 * Please consult with lkml before using this in your driver.
1538 int split_free_page(struct page *page)
1543 order = page_order(page);
1545 nr_pages = __isolate_free_page(page, order);
1549 /* Split into individual pages */
1550 set_page_refcounted(page);
1551 split_page(page, order);
1556 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1557 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1561 struct page *buffered_rmqueue(struct zone *preferred_zone,
1562 struct zone *zone, unsigned int order,
1563 gfp_t gfp_flags, int migratetype)
1565 unsigned long flags;
1567 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1570 if (likely(order == 0)) {
1571 struct per_cpu_pages *pcp;
1572 struct list_head *list;
1574 local_irq_save(flags);
1575 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1576 list = &pcp->lists[migratetype];
1577 if (list_empty(list)) {
1578 pcp->count += rmqueue_bulk(zone, 0,
1581 if (unlikely(list_empty(list)))
1586 page = list_entry(list->prev, struct page, lru);
1588 page = list_entry(list->next, struct page, lru);
1590 list_del(&page->lru);
1593 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1595 * __GFP_NOFAIL is not to be used in new code.
1597 * All __GFP_NOFAIL callers should be fixed so that they
1598 * properly detect and handle allocation failures.
1600 * We most definitely don't want callers attempting to
1601 * allocate greater than order-1 page units with
1604 WARN_ON_ONCE(order > 1);
1606 spin_lock_irqsave(&zone->lock, flags);
1607 page = __rmqueue(zone, order, migratetype);
1608 spin_unlock(&zone->lock);
1611 __mod_zone_freepage_state(zone, -(1 << order),
1612 get_freepage_migratetype(page));
1615 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1616 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1617 !zone_is_fair_depleted(zone))
1618 zone_set_flag(zone, ZONE_FAIR_DEPLETED);
1620 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1621 zone_statistics(preferred_zone, zone, gfp_flags);
1622 local_irq_restore(flags);
1624 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1625 if (prep_new_page(page, order, gfp_flags))
1630 local_irq_restore(flags);
1634 #ifdef CONFIG_FAIL_PAGE_ALLOC
1637 struct fault_attr attr;
1639 u32 ignore_gfp_highmem;
1640 u32 ignore_gfp_wait;
1642 } fail_page_alloc = {
1643 .attr = FAULT_ATTR_INITIALIZER,
1644 .ignore_gfp_wait = 1,
1645 .ignore_gfp_highmem = 1,
1649 static int __init setup_fail_page_alloc(char *str)
1651 return setup_fault_attr(&fail_page_alloc.attr, str);
1653 __setup("fail_page_alloc=", setup_fail_page_alloc);
1655 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1657 if (order < fail_page_alloc.min_order)
1659 if (gfp_mask & __GFP_NOFAIL)
1661 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1663 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1666 return should_fail(&fail_page_alloc.attr, 1 << order);
1669 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1671 static int __init fail_page_alloc_debugfs(void)
1673 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1676 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1677 &fail_page_alloc.attr);
1679 return PTR_ERR(dir);
1681 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1682 &fail_page_alloc.ignore_gfp_wait))
1684 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1685 &fail_page_alloc.ignore_gfp_highmem))
1687 if (!debugfs_create_u32("min-order", mode, dir,
1688 &fail_page_alloc.min_order))
1693 debugfs_remove_recursive(dir);
1698 late_initcall(fail_page_alloc_debugfs);
1700 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1702 #else /* CONFIG_FAIL_PAGE_ALLOC */
1704 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1709 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1712 * Return true if free pages are above 'mark'. This takes into account the order
1713 * of the allocation.
1715 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1716 unsigned long mark, int classzone_idx, int alloc_flags,
1719 /* free_pages my go negative - that's OK */
1724 free_pages -= (1 << order) - 1;
1725 if (alloc_flags & ALLOC_HIGH)
1727 if (alloc_flags & ALLOC_HARDER)
1730 /* If allocation can't use CMA areas don't use free CMA pages */
1731 if (!(alloc_flags & ALLOC_CMA))
1732 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1735 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1737 for (o = 0; o < order; o++) {
1738 /* At the next order, this order's pages become unavailable */
1739 free_pages -= z->free_area[o].nr_free << o;
1741 /* Require fewer higher order pages to be free */
1744 if (free_pages <= min)
1750 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1751 int classzone_idx, int alloc_flags)
1753 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1754 zone_page_state(z, NR_FREE_PAGES));
1757 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1758 unsigned long mark, int classzone_idx, int alloc_flags)
1760 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1762 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1763 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1765 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1771 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1772 * skip over zones that are not allowed by the cpuset, or that have
1773 * been recently (in last second) found to be nearly full. See further
1774 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1775 * that have to skip over a lot of full or unallowed zones.
1777 * If the zonelist cache is present in the passed zonelist, then
1778 * returns a pointer to the allowed node mask (either the current
1779 * tasks mems_allowed, or node_states[N_MEMORY].)
1781 * If the zonelist cache is not available for this zonelist, does
1782 * nothing and returns NULL.
1784 * If the fullzones BITMAP in the zonelist cache is stale (more than
1785 * a second since last zap'd) then we zap it out (clear its bits.)
1787 * We hold off even calling zlc_setup, until after we've checked the
1788 * first zone in the zonelist, on the theory that most allocations will
1789 * be satisfied from that first zone, so best to examine that zone as
1790 * quickly as we can.
1792 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1794 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1795 nodemask_t *allowednodes; /* zonelist_cache approximation */
1797 zlc = zonelist->zlcache_ptr;
1801 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1802 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1803 zlc->last_full_zap = jiffies;
1806 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1807 &cpuset_current_mems_allowed :
1808 &node_states[N_MEMORY];
1809 return allowednodes;
1813 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1814 * if it is worth looking at further for free memory:
1815 * 1) Check that the zone isn't thought to be full (doesn't have its
1816 * bit set in the zonelist_cache fullzones BITMAP).
1817 * 2) Check that the zones node (obtained from the zonelist_cache
1818 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1819 * Return true (non-zero) if zone is worth looking at further, or
1820 * else return false (zero) if it is not.
1822 * This check -ignores- the distinction between various watermarks,
1823 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1824 * found to be full for any variation of these watermarks, it will
1825 * be considered full for up to one second by all requests, unless
1826 * we are so low on memory on all allowed nodes that we are forced
1827 * into the second scan of the zonelist.
1829 * In the second scan we ignore this zonelist cache and exactly
1830 * apply the watermarks to all zones, even it is slower to do so.
1831 * We are low on memory in the second scan, and should leave no stone
1832 * unturned looking for a free page.
1834 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1835 nodemask_t *allowednodes)
1837 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1838 int i; /* index of *z in zonelist zones */
1839 int n; /* node that zone *z is on */
1841 zlc = zonelist->zlcache_ptr;
1845 i = z - zonelist->_zonerefs;
1848 /* This zone is worth trying if it is allowed but not full */
1849 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1853 * Given 'z' scanning a zonelist, set the corresponding bit in
1854 * zlc->fullzones, so that subsequent attempts to allocate a page
1855 * from that zone don't waste time re-examining it.
1857 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1859 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1860 int i; /* index of *z in zonelist zones */
1862 zlc = zonelist->zlcache_ptr;
1866 i = z - zonelist->_zonerefs;
1868 set_bit(i, zlc->fullzones);
1872 * clear all zones full, called after direct reclaim makes progress so that
1873 * a zone that was recently full is not skipped over for up to a second
1875 static void zlc_clear_zones_full(struct zonelist *zonelist)
1877 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1879 zlc = zonelist->zlcache_ptr;
1883 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1886 static bool zone_local(struct zone *local_zone, struct zone *zone)
1888 return local_zone->node == zone->node;
1891 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1893 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1897 #else /* CONFIG_NUMA */
1899 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1904 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1905 nodemask_t *allowednodes)
1910 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1914 static void zlc_clear_zones_full(struct zonelist *zonelist)
1918 static bool zone_local(struct zone *local_zone, struct zone *zone)
1923 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1928 #endif /* CONFIG_NUMA */
1930 static void reset_alloc_batches(struct zone *preferred_zone)
1932 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
1935 mod_zone_page_state(zone, NR_ALLOC_BATCH,
1936 high_wmark_pages(zone) - low_wmark_pages(zone) -
1937 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
1938 zone_clear_flag(zone, ZONE_FAIR_DEPLETED);
1939 } while (zone++ != preferred_zone);
1943 * get_page_from_freelist goes through the zonelist trying to allocate
1946 static struct page *
1947 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1948 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1949 struct zone *preferred_zone, int classzone_idx, int migratetype)
1952 struct page *page = NULL;
1954 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1955 int zlc_active = 0; /* set if using zonelist_cache */
1956 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1957 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
1958 (gfp_mask & __GFP_WRITE);
1959 int nr_fair_skipped = 0;
1960 bool zonelist_rescan;
1963 zonelist_rescan = false;
1966 * Scan zonelist, looking for a zone with enough free.
1967 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1969 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1970 high_zoneidx, nodemask) {
1973 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1974 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1976 if (cpusets_enabled() &&
1977 (alloc_flags & ALLOC_CPUSET) &&
1978 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1981 * Distribute pages in proportion to the individual
1982 * zone size to ensure fair page aging. The zone a
1983 * page was allocated in should have no effect on the
1984 * time the page has in memory before being reclaimed.
1986 if (alloc_flags & ALLOC_FAIR) {
1987 if (!zone_local(preferred_zone, zone))
1989 if (zone_is_fair_depleted(zone)) {
1995 * When allocating a page cache page for writing, we
1996 * want to get it from a zone that is within its dirty
1997 * limit, such that no single zone holds more than its
1998 * proportional share of globally allowed dirty pages.
1999 * The dirty limits take into account the zone's
2000 * lowmem reserves and high watermark so that kswapd
2001 * should be able to balance it without having to
2002 * write pages from its LRU list.
2004 * This may look like it could increase pressure on
2005 * lower zones by failing allocations in higher zones
2006 * before they are full. But the pages that do spill
2007 * over are limited as the lower zones are protected
2008 * by this very same mechanism. It should not become
2009 * a practical burden to them.
2011 * XXX: For now, allow allocations to potentially
2012 * exceed the per-zone dirty limit in the slowpath
2013 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2014 * which is important when on a NUMA setup the allowed
2015 * zones are together not big enough to reach the
2016 * global limit. The proper fix for these situations
2017 * will require awareness of zones in the
2018 * dirty-throttling and the flusher threads.
2020 if (consider_zone_dirty && !zone_dirty_ok(zone))
2023 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2024 if (!zone_watermark_ok(zone, order, mark,
2025 classzone_idx, alloc_flags)) {
2028 /* Checked here to keep the fast path fast */
2029 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2030 if (alloc_flags & ALLOC_NO_WATERMARKS)
2033 if (IS_ENABLED(CONFIG_NUMA) &&
2034 !did_zlc_setup && nr_online_nodes > 1) {
2036 * we do zlc_setup if there are multiple nodes
2037 * and before considering the first zone allowed
2040 allowednodes = zlc_setup(zonelist, alloc_flags);
2045 if (zone_reclaim_mode == 0 ||
2046 !zone_allows_reclaim(preferred_zone, zone))
2047 goto this_zone_full;
2050 * As we may have just activated ZLC, check if the first
2051 * eligible zone has failed zone_reclaim recently.
2053 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2054 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2057 ret = zone_reclaim(zone, gfp_mask, order);
2059 case ZONE_RECLAIM_NOSCAN:
2062 case ZONE_RECLAIM_FULL:
2063 /* scanned but unreclaimable */
2066 /* did we reclaim enough */
2067 if (zone_watermark_ok(zone, order, mark,
2068 classzone_idx, alloc_flags))
2072 * Failed to reclaim enough to meet watermark.
2073 * Only mark the zone full if checking the min
2074 * watermark or if we failed to reclaim just
2075 * 1<<order pages or else the page allocator
2076 * fastpath will prematurely mark zones full
2077 * when the watermark is between the low and
2080 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2081 ret == ZONE_RECLAIM_SOME)
2082 goto this_zone_full;
2089 page = buffered_rmqueue(preferred_zone, zone, order,
2090 gfp_mask, migratetype);
2094 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2095 zlc_mark_zone_full(zonelist, z);
2100 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2101 * necessary to allocate the page. The expectation is
2102 * that the caller is taking steps that will free more
2103 * memory. The caller should avoid the page being used
2104 * for !PFMEMALLOC purposes.
2106 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2111 * The first pass makes sure allocations are spread fairly within the
2112 * local node. However, the local node might have free pages left
2113 * after the fairness batches are exhausted, and remote zones haven't
2114 * even been considered yet. Try once more without fairness, and
2115 * include remote zones now, before entering the slowpath and waking
2116 * kswapd: prefer spilling to a remote zone over swapping locally.
2118 if (alloc_flags & ALLOC_FAIR) {
2119 alloc_flags &= ~ALLOC_FAIR;
2120 if (nr_fair_skipped) {
2121 zonelist_rescan = true;
2122 reset_alloc_batches(preferred_zone);
2124 if (nr_online_nodes > 1)
2125 zonelist_rescan = true;
2128 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2129 /* Disable zlc cache for second zonelist scan */
2131 zonelist_rescan = true;
2134 if (zonelist_rescan)
2141 * Large machines with many possible nodes should not always dump per-node
2142 * meminfo in irq context.
2144 static inline bool should_suppress_show_mem(void)
2149 ret = in_interrupt();
2154 static DEFINE_RATELIMIT_STATE(nopage_rs,
2155 DEFAULT_RATELIMIT_INTERVAL,
2156 DEFAULT_RATELIMIT_BURST);
2158 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2160 unsigned int filter = SHOW_MEM_FILTER_NODES;
2162 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2163 debug_guardpage_minorder() > 0)
2167 * This documents exceptions given to allocations in certain
2168 * contexts that are allowed to allocate outside current's set
2171 if (!(gfp_mask & __GFP_NOMEMALLOC))
2172 if (test_thread_flag(TIF_MEMDIE) ||
2173 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2174 filter &= ~SHOW_MEM_FILTER_NODES;
2175 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2176 filter &= ~SHOW_MEM_FILTER_NODES;
2179 struct va_format vaf;
2182 va_start(args, fmt);
2187 pr_warn("%pV", &vaf);
2192 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2193 current->comm, order, gfp_mask);
2196 if (!should_suppress_show_mem())
2201 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2202 unsigned long did_some_progress,
2203 unsigned long pages_reclaimed)
2205 /* Do not loop if specifically requested */
2206 if (gfp_mask & __GFP_NORETRY)
2209 /* Always retry if specifically requested */
2210 if (gfp_mask & __GFP_NOFAIL)
2214 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2215 * making forward progress without invoking OOM. Suspend also disables
2216 * storage devices so kswapd will not help. Bail if we are suspending.
2218 if (!did_some_progress && pm_suspended_storage())
2222 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2223 * means __GFP_NOFAIL, but that may not be true in other
2226 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2230 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2231 * specified, then we retry until we no longer reclaim any pages
2232 * (above), or we've reclaimed an order of pages at least as
2233 * large as the allocation's order. In both cases, if the
2234 * allocation still fails, we stop retrying.
2236 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2242 static inline struct page *
2243 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2244 struct zonelist *zonelist, enum zone_type high_zoneidx,
2245 nodemask_t *nodemask, struct zone *preferred_zone,
2246 int classzone_idx, int migratetype)
2250 /* Acquire the per-zone oom lock for each zone */
2251 if (!oom_zonelist_trylock(zonelist, gfp_mask)) {
2252 schedule_timeout_uninterruptible(1);
2257 * Go through the zonelist yet one more time, keep very high watermark
2258 * here, this is only to catch a parallel oom killing, we must fail if
2259 * we're still under heavy pressure.
2261 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2262 order, zonelist, high_zoneidx,
2263 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2264 preferred_zone, classzone_idx, migratetype);
2268 if (!(gfp_mask & __GFP_NOFAIL)) {
2269 /* The OOM killer will not help higher order allocs */
2270 if (order > PAGE_ALLOC_COSTLY_ORDER)
2272 /* The OOM killer does not needlessly kill tasks for lowmem */
2273 if (high_zoneidx < ZONE_NORMAL)
2276 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2277 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2278 * The caller should handle page allocation failure by itself if
2279 * it specifies __GFP_THISNODE.
2280 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2282 if (gfp_mask & __GFP_THISNODE)
2285 /* Exhausted what can be done so it's blamo time */
2286 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2289 oom_zonelist_unlock(zonelist, gfp_mask);
2293 #ifdef CONFIG_COMPACTION
2294 /* Try memory compaction for high-order allocations before reclaim */
2295 static struct page *
2296 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2297 struct zonelist *zonelist, enum zone_type high_zoneidx,
2298 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2299 int classzone_idx, int migratetype, enum migrate_mode mode,
2300 bool *contended_compaction, bool *deferred_compaction,
2301 unsigned long *did_some_progress)
2306 if (compaction_deferred(preferred_zone, order)) {
2307 *deferred_compaction = true;
2311 current->flags |= PF_MEMALLOC;
2312 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2314 contended_compaction);
2315 current->flags &= ~PF_MEMALLOC;
2317 if (*did_some_progress != COMPACT_SKIPPED) {
2320 /* Page migration frees to the PCP lists but we want merging */
2321 drain_pages(get_cpu());
2324 page = get_page_from_freelist(gfp_mask, nodemask,
2325 order, zonelist, high_zoneidx,
2326 alloc_flags & ~ALLOC_NO_WATERMARKS,
2327 preferred_zone, classzone_idx, migratetype);
2329 preferred_zone->compact_blockskip_flush = false;
2330 compaction_defer_reset(preferred_zone, order, true);
2331 count_vm_event(COMPACTSUCCESS);
2336 * It's bad if compaction run occurs and fails.
2337 * The most likely reason is that pages exist,
2338 * but not enough to satisfy watermarks.
2340 count_vm_event(COMPACTFAIL);
2343 * As async compaction considers a subset of pageblocks, only
2344 * defer if the failure was a sync compaction failure.
2346 if (mode != MIGRATE_ASYNC)
2347 defer_compaction(preferred_zone, order);
2355 static inline struct page *
2356 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2357 struct zonelist *zonelist, enum zone_type high_zoneidx,
2358 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2359 int classzone_idx, int migratetype,
2360 enum migrate_mode mode, bool *contended_compaction,
2361 bool *deferred_compaction, unsigned long *did_some_progress)
2365 #endif /* CONFIG_COMPACTION */
2367 /* Perform direct synchronous page reclaim */
2369 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2370 nodemask_t *nodemask)
2372 struct reclaim_state reclaim_state;
2377 /* We now go into synchronous reclaim */
2378 cpuset_memory_pressure_bump();
2379 current->flags |= PF_MEMALLOC;
2380 lockdep_set_current_reclaim_state(gfp_mask);
2381 reclaim_state.reclaimed_slab = 0;
2382 current->reclaim_state = &reclaim_state;
2384 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2386 current->reclaim_state = NULL;
2387 lockdep_clear_current_reclaim_state();
2388 current->flags &= ~PF_MEMALLOC;
2395 /* The really slow allocator path where we enter direct reclaim */
2396 static inline struct page *
2397 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2398 struct zonelist *zonelist, enum zone_type high_zoneidx,
2399 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2400 int classzone_idx, int migratetype, unsigned long *did_some_progress)
2402 struct page *page = NULL;
2403 bool drained = false;
2405 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2407 if (unlikely(!(*did_some_progress)))
2410 /* After successful reclaim, reconsider all zones for allocation */
2411 if (IS_ENABLED(CONFIG_NUMA))
2412 zlc_clear_zones_full(zonelist);
2415 page = get_page_from_freelist(gfp_mask, nodemask, order,
2416 zonelist, high_zoneidx,
2417 alloc_flags & ~ALLOC_NO_WATERMARKS,
2418 preferred_zone, classzone_idx,
2422 * If an allocation failed after direct reclaim, it could be because
2423 * pages are pinned on the per-cpu lists. Drain them and try again
2425 if (!page && !drained) {
2435 * This is called in the allocator slow-path if the allocation request is of
2436 * sufficient urgency to ignore watermarks and take other desperate measures
2438 static inline struct page *
2439 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2440 struct zonelist *zonelist, enum zone_type high_zoneidx,
2441 nodemask_t *nodemask, struct zone *preferred_zone,
2442 int classzone_idx, int migratetype)
2447 page = get_page_from_freelist(gfp_mask, nodemask, order,
2448 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2449 preferred_zone, classzone_idx, migratetype);
2451 if (!page && gfp_mask & __GFP_NOFAIL)
2452 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2453 } while (!page && (gfp_mask & __GFP_NOFAIL));
2458 static void wake_all_kswapds(unsigned int order,
2459 struct zonelist *zonelist,
2460 enum zone_type high_zoneidx,
2461 struct zone *preferred_zone)
2466 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2467 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2471 gfp_to_alloc_flags(gfp_t gfp_mask)
2473 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2474 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2476 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2477 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2480 * The caller may dip into page reserves a bit more if the caller
2481 * cannot run direct reclaim, or if the caller has realtime scheduling
2482 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2483 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2485 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2489 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2490 * if it can't schedule.
2492 if (!(gfp_mask & __GFP_NOMEMALLOC))
2493 alloc_flags |= ALLOC_HARDER;
2495 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2496 * comment for __cpuset_node_allowed_softwall().
2498 alloc_flags &= ~ALLOC_CPUSET;
2499 } else if (unlikely(rt_task(current)) && !in_interrupt())
2500 alloc_flags |= ALLOC_HARDER;
2502 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2503 if (gfp_mask & __GFP_MEMALLOC)
2504 alloc_flags |= ALLOC_NO_WATERMARKS;
2505 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2506 alloc_flags |= ALLOC_NO_WATERMARKS;
2507 else if (!in_interrupt() &&
2508 ((current->flags & PF_MEMALLOC) ||
2509 unlikely(test_thread_flag(TIF_MEMDIE))))
2510 alloc_flags |= ALLOC_NO_WATERMARKS;
2513 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2514 alloc_flags |= ALLOC_CMA;
2519 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2521 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2524 static inline struct page *
2525 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2526 struct zonelist *zonelist, enum zone_type high_zoneidx,
2527 nodemask_t *nodemask, struct zone *preferred_zone,
2528 int classzone_idx, int migratetype)
2530 const gfp_t wait = gfp_mask & __GFP_WAIT;
2531 struct page *page = NULL;
2533 unsigned long pages_reclaimed = 0;
2534 unsigned long did_some_progress;
2535 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2536 bool deferred_compaction = false;
2537 bool contended_compaction = false;
2540 * In the slowpath, we sanity check order to avoid ever trying to
2541 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2542 * be using allocators in order of preference for an area that is
2545 if (order >= MAX_ORDER) {
2546 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2551 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2552 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2553 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2554 * using a larger set of nodes after it has established that the
2555 * allowed per node queues are empty and that nodes are
2558 if (IS_ENABLED(CONFIG_NUMA) &&
2559 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2563 if (!(gfp_mask & __GFP_NO_KSWAPD))
2564 wake_all_kswapds(order, zonelist, high_zoneidx, preferred_zone);
2567 * OK, we're below the kswapd watermark and have kicked background
2568 * reclaim. Now things get more complex, so set up alloc_flags according
2569 * to how we want to proceed.
2571 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2574 * Find the true preferred zone if the allocation is unconstrained by
2577 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) {
2578 struct zoneref *preferred_zoneref;
2579 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2580 NULL, &preferred_zone);
2581 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2585 /* This is the last chance, in general, before the goto nopage. */
2586 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2587 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2588 preferred_zone, classzone_idx, migratetype);
2592 /* Allocate without watermarks if the context allows */
2593 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2595 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2596 * the allocation is high priority and these type of
2597 * allocations are system rather than user orientated
2599 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2601 page = __alloc_pages_high_priority(gfp_mask, order,
2602 zonelist, high_zoneidx, nodemask,
2603 preferred_zone, classzone_idx, migratetype);
2609 /* Atomic allocations - we can't balance anything */
2612 * All existing users of the deprecated __GFP_NOFAIL are
2613 * blockable, so warn of any new users that actually allow this
2614 * type of allocation to fail.
2616 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2620 /* Avoid recursion of direct reclaim */
2621 if (current->flags & PF_MEMALLOC)
2624 /* Avoid allocations with no watermarks from looping endlessly */
2625 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2629 * Try direct compaction. The first pass is asynchronous. Subsequent
2630 * attempts after direct reclaim are synchronous
2632 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2633 high_zoneidx, nodemask, alloc_flags,
2635 classzone_idx, migratetype,
2636 migration_mode, &contended_compaction,
2637 &deferred_compaction,
2638 &did_some_progress);
2643 * If compaction is deferred for high-order allocations, it is because
2644 * sync compaction recently failed. In this is the case and the caller
2645 * requested a movable allocation that does not heavily disrupt the
2646 * system then fail the allocation instead of entering direct reclaim.
2648 if ((deferred_compaction || contended_compaction) &&
2649 (gfp_mask & __GFP_NO_KSWAPD))
2653 * It can become very expensive to allocate transparent hugepages at
2654 * fault, so use asynchronous memory compaction for THP unless it is
2655 * khugepaged trying to collapse.
2657 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2658 (current->flags & PF_KTHREAD))
2659 migration_mode = MIGRATE_SYNC_LIGHT;
2661 /* Try direct reclaim and then allocating */
2662 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2663 zonelist, high_zoneidx,
2665 alloc_flags, preferred_zone,
2666 classzone_idx, migratetype,
2667 &did_some_progress);
2672 * If we failed to make any progress reclaiming, then we are
2673 * running out of options and have to consider going OOM
2675 if (!did_some_progress) {
2676 if (oom_gfp_allowed(gfp_mask)) {
2677 if (oom_killer_disabled)
2679 /* Coredumps can quickly deplete all memory reserves */
2680 if ((current->flags & PF_DUMPCORE) &&
2681 !(gfp_mask & __GFP_NOFAIL))
2683 page = __alloc_pages_may_oom(gfp_mask, order,
2684 zonelist, high_zoneidx,
2685 nodemask, preferred_zone,
2686 classzone_idx, migratetype);
2690 if (!(gfp_mask & __GFP_NOFAIL)) {
2692 * The oom killer is not called for high-order
2693 * allocations that may fail, so if no progress
2694 * is being made, there are no other options and
2695 * retrying is unlikely to help.
2697 if (order > PAGE_ALLOC_COSTLY_ORDER)
2700 * The oom killer is not called for lowmem
2701 * allocations to prevent needlessly killing
2704 if (high_zoneidx < ZONE_NORMAL)
2712 /* Check if we should retry the allocation */
2713 pages_reclaimed += did_some_progress;
2714 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2716 /* Wait for some write requests to complete then retry */
2717 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2721 * High-order allocations do not necessarily loop after
2722 * direct reclaim and reclaim/compaction depends on compaction
2723 * being called after reclaim so call directly if necessary
2725 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2726 high_zoneidx, nodemask, alloc_flags,
2728 classzone_idx, migratetype,
2729 migration_mode, &contended_compaction,
2730 &deferred_compaction,
2731 &did_some_progress);
2737 warn_alloc_failed(gfp_mask, order, NULL);
2740 if (kmemcheck_enabled)
2741 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2747 * This is the 'heart' of the zoned buddy allocator.
2750 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2751 struct zonelist *zonelist, nodemask_t *nodemask)
2753 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2754 struct zone *preferred_zone;
2755 struct zoneref *preferred_zoneref;
2756 struct page *page = NULL;
2757 int migratetype = allocflags_to_migratetype(gfp_mask);
2758 unsigned int cpuset_mems_cookie;
2759 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2762 gfp_mask &= gfp_allowed_mask;
2764 lockdep_trace_alloc(gfp_mask);
2766 might_sleep_if(gfp_mask & __GFP_WAIT);
2768 if (should_fail_alloc_page(gfp_mask, order))
2772 * Check the zones suitable for the gfp_mask contain at least one
2773 * valid zone. It's possible to have an empty zonelist as a result
2774 * of GFP_THISNODE and a memoryless node
2776 if (unlikely(!zonelist->_zonerefs->zone))
2780 cpuset_mems_cookie = read_mems_allowed_begin();
2782 /* The preferred zone is used for statistics later */
2783 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2784 nodemask ? : &cpuset_current_mems_allowed,
2786 if (!preferred_zone)
2788 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2791 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2792 alloc_flags |= ALLOC_CMA;
2794 /* First allocation attempt */
2795 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2796 zonelist, high_zoneidx, alloc_flags,
2797 preferred_zone, classzone_idx, migratetype);
2798 if (unlikely(!page)) {
2800 * Runtime PM, block IO and its error handling path
2801 * can deadlock because I/O on the device might not
2804 gfp_mask = memalloc_noio_flags(gfp_mask);
2805 page = __alloc_pages_slowpath(gfp_mask, order,
2806 zonelist, high_zoneidx, nodemask,
2807 preferred_zone, classzone_idx, migratetype);
2810 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2814 * When updating a task's mems_allowed, it is possible to race with
2815 * parallel threads in such a way that an allocation can fail while
2816 * the mask is being updated. If a page allocation is about to fail,
2817 * check if the cpuset changed during allocation and if so, retry.
2819 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2824 EXPORT_SYMBOL(__alloc_pages_nodemask);
2827 * Common helper functions.
2829 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2834 * __get_free_pages() returns a 32-bit address, which cannot represent
2837 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2839 page = alloc_pages(gfp_mask, order);
2842 return (unsigned long) page_address(page);
2844 EXPORT_SYMBOL(__get_free_pages);
2846 unsigned long get_zeroed_page(gfp_t gfp_mask)
2848 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2850 EXPORT_SYMBOL(get_zeroed_page);
2852 void __free_pages(struct page *page, unsigned int order)
2854 if (put_page_testzero(page)) {
2856 free_hot_cold_page(page, false);
2858 __free_pages_ok(page, order);
2862 EXPORT_SYMBOL(__free_pages);
2864 void free_pages(unsigned long addr, unsigned int order)
2867 VM_BUG_ON(!virt_addr_valid((void *)addr));
2868 __free_pages(virt_to_page((void *)addr), order);
2872 EXPORT_SYMBOL(free_pages);
2875 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2876 * of the current memory cgroup.
2878 * It should be used when the caller would like to use kmalloc, but since the
2879 * allocation is large, it has to fall back to the page allocator.
2881 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2884 struct mem_cgroup *memcg = NULL;
2886 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2888 page = alloc_pages(gfp_mask, order);
2889 memcg_kmem_commit_charge(page, memcg, order);
2893 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2896 struct mem_cgroup *memcg = NULL;
2898 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2900 page = alloc_pages_node(nid, gfp_mask, order);
2901 memcg_kmem_commit_charge(page, memcg, order);
2906 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2909 void __free_kmem_pages(struct page *page, unsigned int order)
2911 memcg_kmem_uncharge_pages(page, order);
2912 __free_pages(page, order);
2915 void free_kmem_pages(unsigned long addr, unsigned int order)
2918 VM_BUG_ON(!virt_addr_valid((void *)addr));
2919 __free_kmem_pages(virt_to_page((void *)addr), order);
2923 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2926 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2927 unsigned long used = addr + PAGE_ALIGN(size);
2929 split_page(virt_to_page((void *)addr), order);
2930 while (used < alloc_end) {
2935 return (void *)addr;
2939 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2940 * @size: the number of bytes to allocate
2941 * @gfp_mask: GFP flags for the allocation
2943 * This function is similar to alloc_pages(), except that it allocates the
2944 * minimum number of pages to satisfy the request. alloc_pages() can only
2945 * allocate memory in power-of-two pages.
2947 * This function is also limited by MAX_ORDER.
2949 * Memory allocated by this function must be released by free_pages_exact().
2951 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2953 unsigned int order = get_order(size);
2956 addr = __get_free_pages(gfp_mask, order);
2957 return make_alloc_exact(addr, order, size);
2959 EXPORT_SYMBOL(alloc_pages_exact);
2962 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2964 * @nid: the preferred node ID where memory should be allocated
2965 * @size: the number of bytes to allocate
2966 * @gfp_mask: GFP flags for the allocation
2968 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2970 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2973 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2975 unsigned order = get_order(size);
2976 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2979 return make_alloc_exact((unsigned long)page_address(p), order, size);
2983 * free_pages_exact - release memory allocated via alloc_pages_exact()
2984 * @virt: the value returned by alloc_pages_exact.
2985 * @size: size of allocation, same value as passed to alloc_pages_exact().
2987 * Release the memory allocated by a previous call to alloc_pages_exact.
2989 void free_pages_exact(void *virt, size_t size)
2991 unsigned long addr = (unsigned long)virt;
2992 unsigned long end = addr + PAGE_ALIGN(size);
2994 while (addr < end) {
2999 EXPORT_SYMBOL(free_pages_exact);
3002 * nr_free_zone_pages - count number of pages beyond high watermark
3003 * @offset: The zone index of the highest zone
3005 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3006 * high watermark within all zones at or below a given zone index. For each
3007 * zone, the number of pages is calculated as:
3008 * managed_pages - high_pages
3010 static unsigned long nr_free_zone_pages(int offset)
3015 /* Just pick one node, since fallback list is circular */
3016 unsigned long sum = 0;
3018 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3020 for_each_zone_zonelist(zone, z, zonelist, offset) {
3021 unsigned long size = zone->managed_pages;
3022 unsigned long high = high_wmark_pages(zone);
3031 * nr_free_buffer_pages - count number of pages beyond high watermark
3033 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3034 * watermark within ZONE_DMA and ZONE_NORMAL.
3036 unsigned long nr_free_buffer_pages(void)
3038 return nr_free_zone_pages(gfp_zone(GFP_USER));
3040 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3043 * nr_free_pagecache_pages - count number of pages beyond high watermark
3045 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3046 * high watermark within all zones.
3048 unsigned long nr_free_pagecache_pages(void)
3050 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3053 static inline void show_node(struct zone *zone)
3055 if (IS_ENABLED(CONFIG_NUMA))
3056 printk("Node %d ", zone_to_nid(zone));
3059 void si_meminfo(struct sysinfo *val)
3061 val->totalram = totalram_pages;
3062 val->sharedram = global_page_state(NR_SHMEM);
3063 val->freeram = global_page_state(NR_FREE_PAGES);
3064 val->bufferram = nr_blockdev_pages();
3065 val->totalhigh = totalhigh_pages;
3066 val->freehigh = nr_free_highpages();
3067 val->mem_unit = PAGE_SIZE;
3070 EXPORT_SYMBOL(si_meminfo);
3073 void si_meminfo_node(struct sysinfo *val, int nid)
3075 int zone_type; /* needs to be signed */
3076 unsigned long managed_pages = 0;
3077 pg_data_t *pgdat = NODE_DATA(nid);
3079 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3080 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3081 val->totalram = managed_pages;
3082 val->sharedram = node_page_state(nid, NR_SHMEM);
3083 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3084 #ifdef CONFIG_HIGHMEM
3085 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3086 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3092 val->mem_unit = PAGE_SIZE;
3097 * Determine whether the node should be displayed or not, depending on whether
3098 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3100 bool skip_free_areas_node(unsigned int flags, int nid)
3103 unsigned int cpuset_mems_cookie;
3105 if (!(flags & SHOW_MEM_FILTER_NODES))
3109 cpuset_mems_cookie = read_mems_allowed_begin();
3110 ret = !node_isset(nid, cpuset_current_mems_allowed);
3111 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3116 #define K(x) ((x) << (PAGE_SHIFT-10))
3118 static void show_migration_types(unsigned char type)
3120 static const char types[MIGRATE_TYPES] = {
3121 [MIGRATE_UNMOVABLE] = 'U',
3122 [MIGRATE_RECLAIMABLE] = 'E',
3123 [MIGRATE_MOVABLE] = 'M',
3124 [MIGRATE_RESERVE] = 'R',
3126 [MIGRATE_CMA] = 'C',
3128 #ifdef CONFIG_MEMORY_ISOLATION
3129 [MIGRATE_ISOLATE] = 'I',
3132 char tmp[MIGRATE_TYPES + 1];
3136 for (i = 0; i < MIGRATE_TYPES; i++) {
3137 if (type & (1 << i))
3142 printk("(%s) ", tmp);
3146 * Show free area list (used inside shift_scroll-lock stuff)
3147 * We also calculate the percentage fragmentation. We do this by counting the
3148 * memory on each free list with the exception of the first item on the list.
3149 * Suppresses nodes that are not allowed by current's cpuset if
3150 * SHOW_MEM_FILTER_NODES is passed.
3152 void show_free_areas(unsigned int filter)
3157 for_each_populated_zone(zone) {
3158 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3161 printk("%s per-cpu:\n", zone->name);
3163 for_each_online_cpu(cpu) {
3164 struct per_cpu_pageset *pageset;
3166 pageset = per_cpu_ptr(zone->pageset, cpu);
3168 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3169 cpu, pageset->pcp.high,
3170 pageset->pcp.batch, pageset->pcp.count);
3174 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3175 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3177 " dirty:%lu writeback:%lu unstable:%lu\n"
3178 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3179 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3181 global_page_state(NR_ACTIVE_ANON),
3182 global_page_state(NR_INACTIVE_ANON),
3183 global_page_state(NR_ISOLATED_ANON),
3184 global_page_state(NR_ACTIVE_FILE),
3185 global_page_state(NR_INACTIVE_FILE),
3186 global_page_state(NR_ISOLATED_FILE),
3187 global_page_state(NR_UNEVICTABLE),
3188 global_page_state(NR_FILE_DIRTY),
3189 global_page_state(NR_WRITEBACK),
3190 global_page_state(NR_UNSTABLE_NFS),
3191 global_page_state(NR_FREE_PAGES),
3192 global_page_state(NR_SLAB_RECLAIMABLE),
3193 global_page_state(NR_SLAB_UNRECLAIMABLE),
3194 global_page_state(NR_FILE_MAPPED),
3195 global_page_state(NR_SHMEM),
3196 global_page_state(NR_PAGETABLE),
3197 global_page_state(NR_BOUNCE),
3198 global_page_state(NR_FREE_CMA_PAGES));
3200 for_each_populated_zone(zone) {
3203 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3211 " active_anon:%lukB"
3212 " inactive_anon:%lukB"
3213 " active_file:%lukB"
3214 " inactive_file:%lukB"
3215 " unevictable:%lukB"
3216 " isolated(anon):%lukB"
3217 " isolated(file):%lukB"
3225 " slab_reclaimable:%lukB"
3226 " slab_unreclaimable:%lukB"
3227 " kernel_stack:%lukB"
3232 " writeback_tmp:%lukB"
3233 " pages_scanned:%lu"
3234 " all_unreclaimable? %s"
3237 K(zone_page_state(zone, NR_FREE_PAGES)),
3238 K(min_wmark_pages(zone)),
3239 K(low_wmark_pages(zone)),
3240 K(high_wmark_pages(zone)),
3241 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3242 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3243 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3244 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3245 K(zone_page_state(zone, NR_UNEVICTABLE)),
3246 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3247 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3248 K(zone->present_pages),
3249 K(zone->managed_pages),
3250 K(zone_page_state(zone, NR_MLOCK)),
3251 K(zone_page_state(zone, NR_FILE_DIRTY)),
3252 K(zone_page_state(zone, NR_WRITEBACK)),
3253 K(zone_page_state(zone, NR_FILE_MAPPED)),
3254 K(zone_page_state(zone, NR_SHMEM)),
3255 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3256 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3257 zone_page_state(zone, NR_KERNEL_STACK) *
3259 K(zone_page_state(zone, NR_PAGETABLE)),
3260 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3261 K(zone_page_state(zone, NR_BOUNCE)),
3262 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3263 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3264 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3265 (!zone_reclaimable(zone) ? "yes" : "no")
3267 printk("lowmem_reserve[]:");
3268 for (i = 0; i < MAX_NR_ZONES; i++)
3269 printk(" %ld", zone->lowmem_reserve[i]);
3273 for_each_populated_zone(zone) {
3274 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3275 unsigned char types[MAX_ORDER];
3277 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3280 printk("%s: ", zone->name);
3282 spin_lock_irqsave(&zone->lock, flags);
3283 for (order = 0; order < MAX_ORDER; order++) {
3284 struct free_area *area = &zone->free_area[order];
3287 nr[order] = area->nr_free;
3288 total += nr[order] << order;
3291 for (type = 0; type < MIGRATE_TYPES; type++) {
3292 if (!list_empty(&area->free_list[type]))
3293 types[order] |= 1 << type;
3296 spin_unlock_irqrestore(&zone->lock, flags);
3297 for (order = 0; order < MAX_ORDER; order++) {
3298 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3300 show_migration_types(types[order]);
3302 printk("= %lukB\n", K(total));
3305 hugetlb_show_meminfo();
3307 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3309 show_swap_cache_info();
3312 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3314 zoneref->zone = zone;
3315 zoneref->zone_idx = zone_idx(zone);
3319 * Builds allocation fallback zone lists.
3321 * Add all populated zones of a node to the zonelist.
3323 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3327 enum zone_type zone_type = MAX_NR_ZONES;
3331 zone = pgdat->node_zones + zone_type;
3332 if (populated_zone(zone)) {
3333 zoneref_set_zone(zone,
3334 &zonelist->_zonerefs[nr_zones++]);
3335 check_highest_zone(zone_type);
3337 } while (zone_type);
3345 * 0 = automatic detection of better ordering.
3346 * 1 = order by ([node] distance, -zonetype)
3347 * 2 = order by (-zonetype, [node] distance)
3349 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3350 * the same zonelist. So only NUMA can configure this param.
3352 #define ZONELIST_ORDER_DEFAULT 0
3353 #define ZONELIST_ORDER_NODE 1
3354 #define ZONELIST_ORDER_ZONE 2
3356 /* zonelist order in the kernel.
3357 * set_zonelist_order() will set this to NODE or ZONE.
3359 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3360 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3364 /* The value user specified ....changed by config */
3365 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3366 /* string for sysctl */
3367 #define NUMA_ZONELIST_ORDER_LEN 16
3368 char numa_zonelist_order[16] = "default";
3371 * interface for configure zonelist ordering.
3372 * command line option "numa_zonelist_order"
3373 * = "[dD]efault - default, automatic configuration.
3374 * = "[nN]ode - order by node locality, then by zone within node
3375 * = "[zZ]one - order by zone, then by locality within zone
3378 static int __parse_numa_zonelist_order(char *s)
3380 if (*s == 'd' || *s == 'D') {
3381 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3382 } else if (*s == 'n' || *s == 'N') {
3383 user_zonelist_order = ZONELIST_ORDER_NODE;
3384 } else if (*s == 'z' || *s == 'Z') {
3385 user_zonelist_order = ZONELIST_ORDER_ZONE;
3388 "Ignoring invalid numa_zonelist_order value: "
3395 static __init int setup_numa_zonelist_order(char *s)
3402 ret = __parse_numa_zonelist_order(s);
3404 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3408 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3411 * sysctl handler for numa_zonelist_order
3413 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3414 void __user *buffer, size_t *length,
3417 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3419 static DEFINE_MUTEX(zl_order_mutex);
3421 mutex_lock(&zl_order_mutex);
3423 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3427 strcpy(saved_string, (char *)table->data);
3429 ret = proc_dostring(table, write, buffer, length, ppos);
3433 int oldval = user_zonelist_order;
3435 ret = __parse_numa_zonelist_order((char *)table->data);
3438 * bogus value. restore saved string
3440 strncpy((char *)table->data, saved_string,
3441 NUMA_ZONELIST_ORDER_LEN);
3442 user_zonelist_order = oldval;
3443 } else if (oldval != user_zonelist_order) {
3444 mutex_lock(&zonelists_mutex);
3445 build_all_zonelists(NULL, NULL);
3446 mutex_unlock(&zonelists_mutex);
3450 mutex_unlock(&zl_order_mutex);
3455 #define MAX_NODE_LOAD (nr_online_nodes)
3456 static int node_load[MAX_NUMNODES];
3459 * find_next_best_node - find the next node that should appear in a given node's fallback list
3460 * @node: node whose fallback list we're appending
3461 * @used_node_mask: nodemask_t of already used nodes
3463 * We use a number of factors to determine which is the next node that should
3464 * appear on a given node's fallback list. The node should not have appeared
3465 * already in @node's fallback list, and it should be the next closest node
3466 * according to the distance array (which contains arbitrary distance values
3467 * from each node to each node in the system), and should also prefer nodes
3468 * with no CPUs, since presumably they'll have very little allocation pressure
3469 * on them otherwise.
3470 * It returns -1 if no node is found.
3472 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3475 int min_val = INT_MAX;
3476 int best_node = NUMA_NO_NODE;
3477 const struct cpumask *tmp = cpumask_of_node(0);
3479 /* Use the local node if we haven't already */
3480 if (!node_isset(node, *used_node_mask)) {
3481 node_set(node, *used_node_mask);
3485 for_each_node_state(n, N_MEMORY) {
3487 /* Don't want a node to appear more than once */
3488 if (node_isset(n, *used_node_mask))
3491 /* Use the distance array to find the distance */
3492 val = node_distance(node, n);
3494 /* Penalize nodes under us ("prefer the next node") */
3497 /* Give preference to headless and unused nodes */
3498 tmp = cpumask_of_node(n);
3499 if (!cpumask_empty(tmp))
3500 val += PENALTY_FOR_NODE_WITH_CPUS;
3502 /* Slight preference for less loaded node */
3503 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3504 val += node_load[n];
3506 if (val < min_val) {
3513 node_set(best_node, *used_node_mask);
3520 * Build zonelists ordered by node and zones within node.
3521 * This results in maximum locality--normal zone overflows into local
3522 * DMA zone, if any--but risks exhausting DMA zone.
3524 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3527 struct zonelist *zonelist;
3529 zonelist = &pgdat->node_zonelists[0];
3530 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3532 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3533 zonelist->_zonerefs[j].zone = NULL;
3534 zonelist->_zonerefs[j].zone_idx = 0;
3538 * Build gfp_thisnode zonelists
3540 static void build_thisnode_zonelists(pg_data_t *pgdat)
3543 struct zonelist *zonelist;
3545 zonelist = &pgdat->node_zonelists[1];
3546 j = build_zonelists_node(pgdat, zonelist, 0);
3547 zonelist->_zonerefs[j].zone = NULL;
3548 zonelist->_zonerefs[j].zone_idx = 0;
3552 * Build zonelists ordered by zone and nodes within zones.
3553 * This results in conserving DMA zone[s] until all Normal memory is
3554 * exhausted, but results in overflowing to remote node while memory
3555 * may still exist in local DMA zone.
3557 static int node_order[MAX_NUMNODES];
3559 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3562 int zone_type; /* needs to be signed */
3564 struct zonelist *zonelist;
3566 zonelist = &pgdat->node_zonelists[0];
3568 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3569 for (j = 0; j < nr_nodes; j++) {
3570 node = node_order[j];
3571 z = &NODE_DATA(node)->node_zones[zone_type];
3572 if (populated_zone(z)) {
3574 &zonelist->_zonerefs[pos++]);
3575 check_highest_zone(zone_type);
3579 zonelist->_zonerefs[pos].zone = NULL;
3580 zonelist->_zonerefs[pos].zone_idx = 0;
3583 static int default_zonelist_order(void)
3586 unsigned long low_kmem_size, total_size;
3590 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3591 * If they are really small and used heavily, the system can fall
3592 * into OOM very easily.
3593 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3595 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3598 for_each_online_node(nid) {
3599 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3600 z = &NODE_DATA(nid)->node_zones[zone_type];
3601 if (populated_zone(z)) {
3602 if (zone_type < ZONE_NORMAL)
3603 low_kmem_size += z->managed_pages;
3604 total_size += z->managed_pages;
3605 } else if (zone_type == ZONE_NORMAL) {
3607 * If any node has only lowmem, then node order
3608 * is preferred to allow kernel allocations
3609 * locally; otherwise, they can easily infringe
3610 * on other nodes when there is an abundance of
3611 * lowmem available to allocate from.
3613 return ZONELIST_ORDER_NODE;
3617 if (!low_kmem_size || /* there are no DMA area. */
3618 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3619 return ZONELIST_ORDER_NODE;
3621 * look into each node's config.
3622 * If there is a node whose DMA/DMA32 memory is very big area on
3623 * local memory, NODE_ORDER may be suitable.
3625 average_size = total_size /
3626 (nodes_weight(node_states[N_MEMORY]) + 1);
3627 for_each_online_node(nid) {
3630 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3631 z = &NODE_DATA(nid)->node_zones[zone_type];
3632 if (populated_zone(z)) {
3633 if (zone_type < ZONE_NORMAL)
3634 low_kmem_size += z->present_pages;
3635 total_size += z->present_pages;
3638 if (low_kmem_size &&
3639 total_size > average_size && /* ignore small node */
3640 low_kmem_size > total_size * 70/100)
3641 return ZONELIST_ORDER_NODE;
3643 return ZONELIST_ORDER_ZONE;
3646 static void set_zonelist_order(void)
3648 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3649 current_zonelist_order = default_zonelist_order();
3651 current_zonelist_order = user_zonelist_order;
3654 static void build_zonelists(pg_data_t *pgdat)
3658 nodemask_t used_mask;
3659 int local_node, prev_node;
3660 struct zonelist *zonelist;
3661 int order = current_zonelist_order;
3663 /* initialize zonelists */
3664 for (i = 0; i < MAX_ZONELISTS; i++) {
3665 zonelist = pgdat->node_zonelists + i;
3666 zonelist->_zonerefs[0].zone = NULL;
3667 zonelist->_zonerefs[0].zone_idx = 0;
3670 /* NUMA-aware ordering of nodes */
3671 local_node = pgdat->node_id;
3672 load = nr_online_nodes;
3673 prev_node = local_node;
3674 nodes_clear(used_mask);
3676 memset(node_order, 0, sizeof(node_order));
3679 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3681 * We don't want to pressure a particular node.
3682 * So adding penalty to the first node in same
3683 * distance group to make it round-robin.
3685 if (node_distance(local_node, node) !=
3686 node_distance(local_node, prev_node))
3687 node_load[node] = load;
3691 if (order == ZONELIST_ORDER_NODE)
3692 build_zonelists_in_node_order(pgdat, node);
3694 node_order[j++] = node; /* remember order */
3697 if (order == ZONELIST_ORDER_ZONE) {
3698 /* calculate node order -- i.e., DMA last! */
3699 build_zonelists_in_zone_order(pgdat, j);
3702 build_thisnode_zonelists(pgdat);
3705 /* Construct the zonelist performance cache - see further mmzone.h */
3706 static void build_zonelist_cache(pg_data_t *pgdat)
3708 struct zonelist *zonelist;
3709 struct zonelist_cache *zlc;
3712 zonelist = &pgdat->node_zonelists[0];
3713 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3714 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3715 for (z = zonelist->_zonerefs; z->zone; z++)
3716 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3719 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3721 * Return node id of node used for "local" allocations.
3722 * I.e., first node id of first zone in arg node's generic zonelist.
3723 * Used for initializing percpu 'numa_mem', which is used primarily
3724 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3726 int local_memory_node(int node)
3730 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3731 gfp_zone(GFP_KERNEL),
3738 #else /* CONFIG_NUMA */
3740 static void set_zonelist_order(void)
3742 current_zonelist_order = ZONELIST_ORDER_ZONE;
3745 static void build_zonelists(pg_data_t *pgdat)
3747 int node, local_node;
3749 struct zonelist *zonelist;
3751 local_node = pgdat->node_id;
3753 zonelist = &pgdat->node_zonelists[0];
3754 j = build_zonelists_node(pgdat, zonelist, 0);
3757 * Now we build the zonelist so that it contains the zones
3758 * of all the other nodes.
3759 * We don't want to pressure a particular node, so when
3760 * building the zones for node N, we make sure that the
3761 * zones coming right after the local ones are those from
3762 * node N+1 (modulo N)
3764 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3765 if (!node_online(node))
3767 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3769 for (node = 0; node < local_node; node++) {
3770 if (!node_online(node))
3772 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3775 zonelist->_zonerefs[j].zone = NULL;
3776 zonelist->_zonerefs[j].zone_idx = 0;
3779 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3780 static void build_zonelist_cache(pg_data_t *pgdat)
3782 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3785 #endif /* CONFIG_NUMA */
3788 * Boot pageset table. One per cpu which is going to be used for all
3789 * zones and all nodes. The parameters will be set in such a way
3790 * that an item put on a list will immediately be handed over to
3791 * the buddy list. This is safe since pageset manipulation is done
3792 * with interrupts disabled.
3794 * The boot_pagesets must be kept even after bootup is complete for
3795 * unused processors and/or zones. They do play a role for bootstrapping
3796 * hotplugged processors.
3798 * zoneinfo_show() and maybe other functions do
3799 * not check if the processor is online before following the pageset pointer.
3800 * Other parts of the kernel may not check if the zone is available.
3802 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3803 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3804 static void setup_zone_pageset(struct zone *zone);
3807 * Global mutex to protect against size modification of zonelists
3808 * as well as to serialize pageset setup for the new populated zone.
3810 DEFINE_MUTEX(zonelists_mutex);
3812 /* return values int ....just for stop_machine() */
3813 static int __build_all_zonelists(void *data)
3817 pg_data_t *self = data;
3820 memset(node_load, 0, sizeof(node_load));
3823 if (self && !node_online(self->node_id)) {
3824 build_zonelists(self);
3825 build_zonelist_cache(self);
3828 for_each_online_node(nid) {
3829 pg_data_t *pgdat = NODE_DATA(nid);
3831 build_zonelists(pgdat);
3832 build_zonelist_cache(pgdat);
3836 * Initialize the boot_pagesets that are going to be used
3837 * for bootstrapping processors. The real pagesets for
3838 * each zone will be allocated later when the per cpu
3839 * allocator is available.
3841 * boot_pagesets are used also for bootstrapping offline
3842 * cpus if the system is already booted because the pagesets
3843 * are needed to initialize allocators on a specific cpu too.
3844 * F.e. the percpu allocator needs the page allocator which
3845 * needs the percpu allocator in order to allocate its pagesets
3846 * (a chicken-egg dilemma).
3848 for_each_possible_cpu(cpu) {
3849 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3851 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3853 * We now know the "local memory node" for each node--
3854 * i.e., the node of the first zone in the generic zonelist.
3855 * Set up numa_mem percpu variable for on-line cpus. During
3856 * boot, only the boot cpu should be on-line; we'll init the
3857 * secondary cpus' numa_mem as they come on-line. During
3858 * node/memory hotplug, we'll fixup all on-line cpus.
3860 if (cpu_online(cpu))
3861 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3869 * Called with zonelists_mutex held always
3870 * unless system_state == SYSTEM_BOOTING.
3872 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3874 set_zonelist_order();
3876 if (system_state == SYSTEM_BOOTING) {
3877 __build_all_zonelists(NULL);
3878 mminit_verify_zonelist();
3879 cpuset_init_current_mems_allowed();
3881 #ifdef CONFIG_MEMORY_HOTPLUG
3883 setup_zone_pageset(zone);
3885 /* we have to stop all cpus to guarantee there is no user
3887 stop_machine(__build_all_zonelists, pgdat, NULL);
3888 /* cpuset refresh routine should be here */
3890 vm_total_pages = nr_free_pagecache_pages();
3892 * Disable grouping by mobility if the number of pages in the
3893 * system is too low to allow the mechanism to work. It would be
3894 * more accurate, but expensive to check per-zone. This check is
3895 * made on memory-hotadd so a system can start with mobility
3896 * disabled and enable it later
3898 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3899 page_group_by_mobility_disabled = 1;
3901 page_group_by_mobility_disabled = 0;
3903 printk("Built %i zonelists in %s order, mobility grouping %s. "
3904 "Total pages: %ld\n",
3906 zonelist_order_name[current_zonelist_order],
3907 page_group_by_mobility_disabled ? "off" : "on",
3910 printk("Policy zone: %s\n", zone_names[policy_zone]);
3915 * Helper functions to size the waitqueue hash table.
3916 * Essentially these want to choose hash table sizes sufficiently
3917 * large so that collisions trying to wait on pages are rare.
3918 * But in fact, the number of active page waitqueues on typical
3919 * systems is ridiculously low, less than 200. So this is even
3920 * conservative, even though it seems large.
3922 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3923 * waitqueues, i.e. the size of the waitq table given the number of pages.
3925 #define PAGES_PER_WAITQUEUE 256
3927 #ifndef CONFIG_MEMORY_HOTPLUG
3928 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3930 unsigned long size = 1;
3932 pages /= PAGES_PER_WAITQUEUE;
3934 while (size < pages)
3938 * Once we have dozens or even hundreds of threads sleeping
3939 * on IO we've got bigger problems than wait queue collision.
3940 * Limit the size of the wait table to a reasonable size.
3942 size = min(size, 4096UL);
3944 return max(size, 4UL);
3948 * A zone's size might be changed by hot-add, so it is not possible to determine
3949 * a suitable size for its wait_table. So we use the maximum size now.
3951 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3953 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3954 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3955 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3957 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3958 * or more by the traditional way. (See above). It equals:
3960 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3961 * ia64(16K page size) : = ( 8G + 4M)byte.
3962 * powerpc (64K page size) : = (32G +16M)byte.
3964 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3971 * This is an integer logarithm so that shifts can be used later
3972 * to extract the more random high bits from the multiplicative
3973 * hash function before the remainder is taken.
3975 static inline unsigned long wait_table_bits(unsigned long size)
3981 * Check if a pageblock contains reserved pages
3983 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3987 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3988 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3995 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3996 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3997 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3998 * higher will lead to a bigger reserve which will get freed as contiguous
3999 * blocks as reclaim kicks in
4001 static void setup_zone_migrate_reserve(struct zone *zone)
4003 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4005 unsigned long block_migratetype;
4010 * Get the start pfn, end pfn and the number of blocks to reserve
4011 * We have to be careful to be aligned to pageblock_nr_pages to
4012 * make sure that we always check pfn_valid for the first page in
4015 start_pfn = zone->zone_start_pfn;
4016 end_pfn = zone_end_pfn(zone);
4017 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4018 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4022 * Reserve blocks are generally in place to help high-order atomic
4023 * allocations that are short-lived. A min_free_kbytes value that
4024 * would result in more than 2 reserve blocks for atomic allocations
4025 * is assumed to be in place to help anti-fragmentation for the
4026 * future allocation of hugepages at runtime.
4028 reserve = min(2, reserve);
4029 old_reserve = zone->nr_migrate_reserve_block;
4031 /* When memory hot-add, we almost always need to do nothing */
4032 if (reserve == old_reserve)
4034 zone->nr_migrate_reserve_block = reserve;
4036 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4037 if (!pfn_valid(pfn))
4039 page = pfn_to_page(pfn);
4041 /* Watch out for overlapping nodes */
4042 if (page_to_nid(page) != zone_to_nid(zone))
4045 block_migratetype = get_pageblock_migratetype(page);
4047 /* Only test what is necessary when the reserves are not met */
4050 * Blocks with reserved pages will never free, skip
4053 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4054 if (pageblock_is_reserved(pfn, block_end_pfn))
4057 /* If this block is reserved, account for it */
4058 if (block_migratetype == MIGRATE_RESERVE) {
4063 /* Suitable for reserving if this block is movable */
4064 if (block_migratetype == MIGRATE_MOVABLE) {
4065 set_pageblock_migratetype(page,
4067 move_freepages_block(zone, page,
4072 } else if (!old_reserve) {
4074 * At boot time we don't need to scan the whole zone
4075 * for turning off MIGRATE_RESERVE.
4081 * If the reserve is met and this is a previous reserved block,
4084 if (block_migratetype == MIGRATE_RESERVE) {
4085 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4086 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4092 * Initially all pages are reserved - free ones are freed
4093 * up by free_all_bootmem() once the early boot process is
4094 * done. Non-atomic initialization, single-pass.
4096 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4097 unsigned long start_pfn, enum memmap_context context)
4100 unsigned long end_pfn = start_pfn + size;
4104 if (highest_memmap_pfn < end_pfn - 1)
4105 highest_memmap_pfn = end_pfn - 1;
4107 z = &NODE_DATA(nid)->node_zones[zone];
4108 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4110 * There can be holes in boot-time mem_map[]s
4111 * handed to this function. They do not
4112 * exist on hotplugged memory.
4114 if (context == MEMMAP_EARLY) {
4115 if (!early_pfn_valid(pfn))
4117 if (!early_pfn_in_nid(pfn, nid))
4120 page = pfn_to_page(pfn);
4121 set_page_links(page, zone, nid, pfn);
4122 mminit_verify_page_links(page, zone, nid, pfn);
4123 init_page_count(page);
4124 page_mapcount_reset(page);
4125 page_cpupid_reset_last(page);
4126 SetPageReserved(page);
4128 * Mark the block movable so that blocks are reserved for
4129 * movable at startup. This will force kernel allocations
4130 * to reserve their blocks rather than leaking throughout
4131 * the address space during boot when many long-lived
4132 * kernel allocations are made. Later some blocks near
4133 * the start are marked MIGRATE_RESERVE by
4134 * setup_zone_migrate_reserve()
4136 * bitmap is created for zone's valid pfn range. but memmap
4137 * can be created for invalid pages (for alignment)
4138 * check here not to call set_pageblock_migratetype() against
4141 if ((z->zone_start_pfn <= pfn)
4142 && (pfn < zone_end_pfn(z))
4143 && !(pfn & (pageblock_nr_pages - 1)))
4144 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4146 INIT_LIST_HEAD(&page->lru);
4147 #ifdef WANT_PAGE_VIRTUAL
4148 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4149 if (!is_highmem_idx(zone))
4150 set_page_address(page, __va(pfn << PAGE_SHIFT));
4155 static void __meminit zone_init_free_lists(struct zone *zone)
4157 unsigned int order, t;
4158 for_each_migratetype_order(order, t) {
4159 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4160 zone->free_area[order].nr_free = 0;
4164 #ifndef __HAVE_ARCH_MEMMAP_INIT
4165 #define memmap_init(size, nid, zone, start_pfn) \
4166 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4169 static int zone_batchsize(struct zone *zone)
4175 * The per-cpu-pages pools are set to around 1000th of the
4176 * size of the zone. But no more than 1/2 of a meg.
4178 * OK, so we don't know how big the cache is. So guess.
4180 batch = zone->managed_pages / 1024;
4181 if (batch * PAGE_SIZE > 512 * 1024)
4182 batch = (512 * 1024) / PAGE_SIZE;
4183 batch /= 4; /* We effectively *= 4 below */
4188 * Clamp the batch to a 2^n - 1 value. Having a power
4189 * of 2 value was found to be more likely to have
4190 * suboptimal cache aliasing properties in some cases.
4192 * For example if 2 tasks are alternately allocating
4193 * batches of pages, one task can end up with a lot
4194 * of pages of one half of the possible page colors
4195 * and the other with pages of the other colors.
4197 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4202 /* The deferral and batching of frees should be suppressed under NOMMU
4205 * The problem is that NOMMU needs to be able to allocate large chunks
4206 * of contiguous memory as there's no hardware page translation to
4207 * assemble apparent contiguous memory from discontiguous pages.
4209 * Queueing large contiguous runs of pages for batching, however,
4210 * causes the pages to actually be freed in smaller chunks. As there
4211 * can be a significant delay between the individual batches being
4212 * recycled, this leads to the once large chunks of space being
4213 * fragmented and becoming unavailable for high-order allocations.
4220 * pcp->high and pcp->batch values are related and dependent on one another:
4221 * ->batch must never be higher then ->high.
4222 * The following function updates them in a safe manner without read side
4225 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4226 * those fields changing asynchronously (acording the the above rule).
4228 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4229 * outside of boot time (or some other assurance that no concurrent updaters
4232 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4233 unsigned long batch)
4235 /* start with a fail safe value for batch */
4239 /* Update high, then batch, in order */
4246 /* a companion to pageset_set_high() */
4247 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4249 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4252 static void pageset_init(struct per_cpu_pageset *p)
4254 struct per_cpu_pages *pcp;
4257 memset(p, 0, sizeof(*p));
4261 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4262 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4265 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4268 pageset_set_batch(p, batch);
4272 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4273 * to the value high for the pageset p.
4275 static void pageset_set_high(struct per_cpu_pageset *p,
4278 unsigned long batch = max(1UL, high / 4);
4279 if ((high / 4) > (PAGE_SHIFT * 8))
4280 batch = PAGE_SHIFT * 8;
4282 pageset_update(&p->pcp, high, batch);
4285 static void pageset_set_high_and_batch(struct zone *zone,
4286 struct per_cpu_pageset *pcp)
4288 if (percpu_pagelist_fraction)
4289 pageset_set_high(pcp,
4290 (zone->managed_pages /
4291 percpu_pagelist_fraction));
4293 pageset_set_batch(pcp, zone_batchsize(zone));
4296 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4298 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4301 pageset_set_high_and_batch(zone, pcp);
4304 static void __meminit setup_zone_pageset(struct zone *zone)
4307 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4308 for_each_possible_cpu(cpu)
4309 zone_pageset_init(zone, cpu);
4313 * Allocate per cpu pagesets and initialize them.
4314 * Before this call only boot pagesets were available.
4316 void __init setup_per_cpu_pageset(void)
4320 for_each_populated_zone(zone)
4321 setup_zone_pageset(zone);
4324 static noinline __init_refok
4325 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4331 * The per-page waitqueue mechanism uses hashed waitqueues
4334 zone->wait_table_hash_nr_entries =
4335 wait_table_hash_nr_entries(zone_size_pages);
4336 zone->wait_table_bits =
4337 wait_table_bits(zone->wait_table_hash_nr_entries);
4338 alloc_size = zone->wait_table_hash_nr_entries
4339 * sizeof(wait_queue_head_t);
4341 if (!slab_is_available()) {
4342 zone->wait_table = (wait_queue_head_t *)
4343 memblock_virt_alloc_node_nopanic(
4344 alloc_size, zone->zone_pgdat->node_id);
4347 * This case means that a zone whose size was 0 gets new memory
4348 * via memory hot-add.
4349 * But it may be the case that a new node was hot-added. In
4350 * this case vmalloc() will not be able to use this new node's
4351 * memory - this wait_table must be initialized to use this new
4352 * node itself as well.
4353 * To use this new node's memory, further consideration will be
4356 zone->wait_table = vmalloc(alloc_size);
4358 if (!zone->wait_table)
4361 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4362 init_waitqueue_head(zone->wait_table + i);
4367 static __meminit void zone_pcp_init(struct zone *zone)
4370 * per cpu subsystem is not up at this point. The following code
4371 * relies on the ability of the linker to provide the
4372 * offset of a (static) per cpu variable into the per cpu area.
4374 zone->pageset = &boot_pageset;
4376 if (populated_zone(zone))
4377 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4378 zone->name, zone->present_pages,
4379 zone_batchsize(zone));
4382 int __meminit init_currently_empty_zone(struct zone *zone,
4383 unsigned long zone_start_pfn,
4385 enum memmap_context context)
4387 struct pglist_data *pgdat = zone->zone_pgdat;
4389 ret = zone_wait_table_init(zone, size);
4392 pgdat->nr_zones = zone_idx(zone) + 1;
4394 zone->zone_start_pfn = zone_start_pfn;
4396 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4397 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4399 (unsigned long)zone_idx(zone),
4400 zone_start_pfn, (zone_start_pfn + size));
4402 zone_init_free_lists(zone);
4407 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4408 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4410 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4412 int __meminit __early_pfn_to_nid(unsigned long pfn)
4414 unsigned long start_pfn, end_pfn;
4417 * NOTE: The following SMP-unsafe globals are only used early in boot
4418 * when the kernel is running single-threaded.
4420 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4421 static int __meminitdata last_nid;
4423 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4426 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4428 last_start_pfn = start_pfn;
4429 last_end_pfn = end_pfn;
4435 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4437 int __meminit early_pfn_to_nid(unsigned long pfn)
4441 nid = __early_pfn_to_nid(pfn);
4444 /* just returns 0 */
4448 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4449 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4453 nid = __early_pfn_to_nid(pfn);
4454 if (nid >= 0 && nid != node)
4461 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4462 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4463 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4465 * If an architecture guarantees that all ranges registered contain no holes
4466 * and may be freed, this this function may be used instead of calling
4467 * memblock_free_early_nid() manually.
4469 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4471 unsigned long start_pfn, end_pfn;
4474 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4475 start_pfn = min(start_pfn, max_low_pfn);
4476 end_pfn = min(end_pfn, max_low_pfn);
4478 if (start_pfn < end_pfn)
4479 memblock_free_early_nid(PFN_PHYS(start_pfn),
4480 (end_pfn - start_pfn) << PAGE_SHIFT,
4486 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4487 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4489 * If an architecture guarantees that all ranges registered contain no holes and may
4490 * be freed, this function may be used instead of calling memory_present() manually.
4492 void __init sparse_memory_present_with_active_regions(int nid)
4494 unsigned long start_pfn, end_pfn;
4497 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4498 memory_present(this_nid, start_pfn, end_pfn);
4502 * get_pfn_range_for_nid - Return the start and end page frames for a node
4503 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4504 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4505 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4507 * It returns the start and end page frame of a node based on information
4508 * provided by memblock_set_node(). If called for a node
4509 * with no available memory, a warning is printed and the start and end
4512 void __meminit get_pfn_range_for_nid(unsigned int nid,
4513 unsigned long *start_pfn, unsigned long *end_pfn)
4515 unsigned long this_start_pfn, this_end_pfn;
4521 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4522 *start_pfn = min(*start_pfn, this_start_pfn);
4523 *end_pfn = max(*end_pfn, this_end_pfn);
4526 if (*start_pfn == -1UL)
4531 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4532 * assumption is made that zones within a node are ordered in monotonic
4533 * increasing memory addresses so that the "highest" populated zone is used
4535 static void __init find_usable_zone_for_movable(void)
4538 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4539 if (zone_index == ZONE_MOVABLE)
4542 if (arch_zone_highest_possible_pfn[zone_index] >
4543 arch_zone_lowest_possible_pfn[zone_index])
4547 VM_BUG_ON(zone_index == -1);
4548 movable_zone = zone_index;
4552 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4553 * because it is sized independent of architecture. Unlike the other zones,
4554 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4555 * in each node depending on the size of each node and how evenly kernelcore
4556 * is distributed. This helper function adjusts the zone ranges
4557 * provided by the architecture for a given node by using the end of the
4558 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4559 * zones within a node are in order of monotonic increases memory addresses
4561 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4562 unsigned long zone_type,
4563 unsigned long node_start_pfn,
4564 unsigned long node_end_pfn,
4565 unsigned long *zone_start_pfn,
4566 unsigned long *zone_end_pfn)
4568 /* Only adjust if ZONE_MOVABLE is on this node */
4569 if (zone_movable_pfn[nid]) {
4570 /* Size ZONE_MOVABLE */
4571 if (zone_type == ZONE_MOVABLE) {
4572 *zone_start_pfn = zone_movable_pfn[nid];
4573 *zone_end_pfn = min(node_end_pfn,
4574 arch_zone_highest_possible_pfn[movable_zone]);
4576 /* Adjust for ZONE_MOVABLE starting within this range */
4577 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4578 *zone_end_pfn > zone_movable_pfn[nid]) {
4579 *zone_end_pfn = zone_movable_pfn[nid];
4581 /* Check if this whole range is within ZONE_MOVABLE */
4582 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4583 *zone_start_pfn = *zone_end_pfn;
4588 * Return the number of pages a zone spans in a node, including holes
4589 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4591 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4592 unsigned long zone_type,
4593 unsigned long node_start_pfn,
4594 unsigned long node_end_pfn,
4595 unsigned long *ignored)
4597 unsigned long zone_start_pfn, zone_end_pfn;
4599 /* Get the start and end of the zone */
4600 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4601 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4602 adjust_zone_range_for_zone_movable(nid, zone_type,
4603 node_start_pfn, node_end_pfn,
4604 &zone_start_pfn, &zone_end_pfn);
4606 /* Check that this node has pages within the zone's required range */
4607 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4610 /* Move the zone boundaries inside the node if necessary */
4611 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4612 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4614 /* Return the spanned pages */
4615 return zone_end_pfn - zone_start_pfn;
4619 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4620 * then all holes in the requested range will be accounted for.
4622 unsigned long __meminit __absent_pages_in_range(int nid,
4623 unsigned long range_start_pfn,
4624 unsigned long range_end_pfn)
4626 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4627 unsigned long start_pfn, end_pfn;
4630 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4631 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4632 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4633 nr_absent -= end_pfn - start_pfn;
4639 * absent_pages_in_range - Return number of page frames in holes within a range
4640 * @start_pfn: The start PFN to start searching for holes
4641 * @end_pfn: The end PFN to stop searching for holes
4643 * It returns the number of pages frames in memory holes within a range.
4645 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4646 unsigned long end_pfn)
4648 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4651 /* Return the number of page frames in holes in a zone on a node */
4652 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4653 unsigned long zone_type,
4654 unsigned long node_start_pfn,
4655 unsigned long node_end_pfn,
4656 unsigned long *ignored)
4658 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4659 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4660 unsigned long zone_start_pfn, zone_end_pfn;
4662 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4663 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4665 adjust_zone_range_for_zone_movable(nid, zone_type,
4666 node_start_pfn, node_end_pfn,
4667 &zone_start_pfn, &zone_end_pfn);
4668 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4671 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4672 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4673 unsigned long zone_type,
4674 unsigned long node_start_pfn,
4675 unsigned long node_end_pfn,
4676 unsigned long *zones_size)
4678 return zones_size[zone_type];
4681 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4682 unsigned long zone_type,
4683 unsigned long node_start_pfn,
4684 unsigned long node_end_pfn,
4685 unsigned long *zholes_size)
4690 return zholes_size[zone_type];
4693 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4695 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4696 unsigned long node_start_pfn,
4697 unsigned long node_end_pfn,
4698 unsigned long *zones_size,
4699 unsigned long *zholes_size)
4701 unsigned long realtotalpages, totalpages = 0;
4704 for (i = 0; i < MAX_NR_ZONES; i++)
4705 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4709 pgdat->node_spanned_pages = totalpages;
4711 realtotalpages = totalpages;
4712 for (i = 0; i < MAX_NR_ZONES; i++)
4714 zone_absent_pages_in_node(pgdat->node_id, i,
4715 node_start_pfn, node_end_pfn,
4717 pgdat->node_present_pages = realtotalpages;
4718 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4722 #ifndef CONFIG_SPARSEMEM
4724 * Calculate the size of the zone->blockflags rounded to an unsigned long
4725 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4726 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4727 * round what is now in bits to nearest long in bits, then return it in
4730 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4732 unsigned long usemapsize;
4734 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4735 usemapsize = roundup(zonesize, pageblock_nr_pages);
4736 usemapsize = usemapsize >> pageblock_order;
4737 usemapsize *= NR_PAGEBLOCK_BITS;
4738 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4740 return usemapsize / 8;
4743 static void __init setup_usemap(struct pglist_data *pgdat,
4745 unsigned long zone_start_pfn,
4746 unsigned long zonesize)
4748 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4749 zone->pageblock_flags = NULL;
4751 zone->pageblock_flags =
4752 memblock_virt_alloc_node_nopanic(usemapsize,
4756 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4757 unsigned long zone_start_pfn, unsigned long zonesize) {}
4758 #endif /* CONFIG_SPARSEMEM */
4760 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4762 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4763 void __paginginit set_pageblock_order(void)
4767 /* Check that pageblock_nr_pages has not already been setup */
4768 if (pageblock_order)
4771 if (HPAGE_SHIFT > PAGE_SHIFT)
4772 order = HUGETLB_PAGE_ORDER;
4774 order = MAX_ORDER - 1;
4777 * Assume the largest contiguous order of interest is a huge page.
4778 * This value may be variable depending on boot parameters on IA64 and
4781 pageblock_order = order;
4783 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4786 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4787 * is unused as pageblock_order is set at compile-time. See
4788 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4791 void __paginginit set_pageblock_order(void)
4795 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4797 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4798 unsigned long present_pages)
4800 unsigned long pages = spanned_pages;
4803 * Provide a more accurate estimation if there are holes within
4804 * the zone and SPARSEMEM is in use. If there are holes within the
4805 * zone, each populated memory region may cost us one or two extra
4806 * memmap pages due to alignment because memmap pages for each
4807 * populated regions may not naturally algined on page boundary.
4808 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4810 if (spanned_pages > present_pages + (present_pages >> 4) &&
4811 IS_ENABLED(CONFIG_SPARSEMEM))
4812 pages = present_pages;
4814 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4818 * Set up the zone data structures:
4819 * - mark all pages reserved
4820 * - mark all memory queues empty
4821 * - clear the memory bitmaps
4823 * NOTE: pgdat should get zeroed by caller.
4825 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4826 unsigned long node_start_pfn, unsigned long node_end_pfn,
4827 unsigned long *zones_size, unsigned long *zholes_size)
4830 int nid = pgdat->node_id;
4831 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4834 pgdat_resize_init(pgdat);
4835 #ifdef CONFIG_NUMA_BALANCING
4836 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4837 pgdat->numabalancing_migrate_nr_pages = 0;
4838 pgdat->numabalancing_migrate_next_window = jiffies;
4840 init_waitqueue_head(&pgdat->kswapd_wait);
4841 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4842 pgdat_page_cgroup_init(pgdat);
4844 for (j = 0; j < MAX_NR_ZONES; j++) {
4845 struct zone *zone = pgdat->node_zones + j;
4846 unsigned long size, realsize, freesize, memmap_pages;
4848 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4849 node_end_pfn, zones_size);
4850 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4856 * Adjust freesize so that it accounts for how much memory
4857 * is used by this zone for memmap. This affects the watermark
4858 * and per-cpu initialisations
4860 memmap_pages = calc_memmap_size(size, realsize);
4861 if (freesize >= memmap_pages) {
4862 freesize -= memmap_pages;
4865 " %s zone: %lu pages used for memmap\n",
4866 zone_names[j], memmap_pages);
4869 " %s zone: %lu pages exceeds freesize %lu\n",
4870 zone_names[j], memmap_pages, freesize);
4872 /* Account for reserved pages */
4873 if (j == 0 && freesize > dma_reserve) {
4874 freesize -= dma_reserve;
4875 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4876 zone_names[0], dma_reserve);
4879 if (!is_highmem_idx(j))
4880 nr_kernel_pages += freesize;
4881 /* Charge for highmem memmap if there are enough kernel pages */
4882 else if (nr_kernel_pages > memmap_pages * 2)
4883 nr_kernel_pages -= memmap_pages;
4884 nr_all_pages += freesize;
4886 zone->spanned_pages = size;
4887 zone->present_pages = realsize;
4889 * Set an approximate value for lowmem here, it will be adjusted
4890 * when the bootmem allocator frees pages into the buddy system.
4891 * And all highmem pages will be managed by the buddy system.
4893 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4896 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4898 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4900 zone->name = zone_names[j];
4901 spin_lock_init(&zone->lock);
4902 spin_lock_init(&zone->lru_lock);
4903 zone_seqlock_init(zone);
4904 zone->zone_pgdat = pgdat;
4905 zone_pcp_init(zone);
4907 /* For bootup, initialized properly in watermark setup */
4908 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4910 lruvec_init(&zone->lruvec);
4914 set_pageblock_order();
4915 setup_usemap(pgdat, zone, zone_start_pfn, size);
4916 ret = init_currently_empty_zone(zone, zone_start_pfn,
4917 size, MEMMAP_EARLY);
4919 memmap_init(size, nid, j, zone_start_pfn);
4920 zone_start_pfn += size;
4924 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4926 /* Skip empty nodes */
4927 if (!pgdat->node_spanned_pages)
4930 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4931 /* ia64 gets its own node_mem_map, before this, without bootmem */
4932 if (!pgdat->node_mem_map) {
4933 unsigned long size, start, end;
4937 * The zone's endpoints aren't required to be MAX_ORDER
4938 * aligned but the node_mem_map endpoints must be in order
4939 * for the buddy allocator to function correctly.
4941 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4942 end = pgdat_end_pfn(pgdat);
4943 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4944 size = (end - start) * sizeof(struct page);
4945 map = alloc_remap(pgdat->node_id, size);
4947 map = memblock_virt_alloc_node_nopanic(size,
4949 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4951 #ifndef CONFIG_NEED_MULTIPLE_NODES
4953 * With no DISCONTIG, the global mem_map is just set as node 0's
4955 if (pgdat == NODE_DATA(0)) {
4956 mem_map = NODE_DATA(0)->node_mem_map;
4957 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4958 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4959 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4960 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4963 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4966 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4967 unsigned long node_start_pfn, unsigned long *zholes_size)
4969 pg_data_t *pgdat = NODE_DATA(nid);
4970 unsigned long start_pfn = 0;
4971 unsigned long end_pfn = 0;
4973 /* pg_data_t should be reset to zero when it's allocated */
4974 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4976 pgdat->node_id = nid;
4977 pgdat->node_start_pfn = node_start_pfn;
4978 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4979 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4981 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4982 zones_size, zholes_size);
4984 alloc_node_mem_map(pgdat);
4985 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4986 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4987 nid, (unsigned long)pgdat,
4988 (unsigned long)pgdat->node_mem_map);
4991 free_area_init_core(pgdat, start_pfn, end_pfn,
4992 zones_size, zholes_size);
4995 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4997 #if MAX_NUMNODES > 1
4999 * Figure out the number of possible node ids.
5001 void __init setup_nr_node_ids(void)
5004 unsigned int highest = 0;
5006 for_each_node_mask(node, node_possible_map)
5008 nr_node_ids = highest + 1;
5013 * node_map_pfn_alignment - determine the maximum internode alignment
5015 * This function should be called after node map is populated and sorted.
5016 * It calculates the maximum power of two alignment which can distinguish
5019 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5020 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5021 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5022 * shifted, 1GiB is enough and this function will indicate so.
5024 * This is used to test whether pfn -> nid mapping of the chosen memory
5025 * model has fine enough granularity to avoid incorrect mapping for the
5026 * populated node map.
5028 * Returns the determined alignment in pfn's. 0 if there is no alignment
5029 * requirement (single node).
5031 unsigned long __init node_map_pfn_alignment(void)
5033 unsigned long accl_mask = 0, last_end = 0;
5034 unsigned long start, end, mask;
5038 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5039 if (!start || last_nid < 0 || last_nid == nid) {
5046 * Start with a mask granular enough to pin-point to the
5047 * start pfn and tick off bits one-by-one until it becomes
5048 * too coarse to separate the current node from the last.
5050 mask = ~((1 << __ffs(start)) - 1);
5051 while (mask && last_end <= (start & (mask << 1)))
5054 /* accumulate all internode masks */
5058 /* convert mask to number of pages */
5059 return ~accl_mask + 1;
5062 /* Find the lowest pfn for a node */
5063 static unsigned long __init find_min_pfn_for_node(int nid)
5065 unsigned long min_pfn = ULONG_MAX;
5066 unsigned long start_pfn;
5069 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5070 min_pfn = min(min_pfn, start_pfn);
5072 if (min_pfn == ULONG_MAX) {
5074 "Could not find start_pfn for node %d\n", nid);
5082 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5084 * It returns the minimum PFN based on information provided via
5085 * memblock_set_node().
5087 unsigned long __init find_min_pfn_with_active_regions(void)
5089 return find_min_pfn_for_node(MAX_NUMNODES);
5093 * early_calculate_totalpages()
5094 * Sum pages in active regions for movable zone.
5095 * Populate N_MEMORY for calculating usable_nodes.
5097 static unsigned long __init early_calculate_totalpages(void)
5099 unsigned long totalpages = 0;
5100 unsigned long start_pfn, end_pfn;
5103 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5104 unsigned long pages = end_pfn - start_pfn;
5106 totalpages += pages;
5108 node_set_state(nid, N_MEMORY);
5114 * Find the PFN the Movable zone begins in each node. Kernel memory
5115 * is spread evenly between nodes as long as the nodes have enough
5116 * memory. When they don't, some nodes will have more kernelcore than
5119 static void __init find_zone_movable_pfns_for_nodes(void)
5122 unsigned long usable_startpfn;
5123 unsigned long kernelcore_node, kernelcore_remaining;
5124 /* save the state before borrow the nodemask */
5125 nodemask_t saved_node_state = node_states[N_MEMORY];
5126 unsigned long totalpages = early_calculate_totalpages();
5127 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5128 struct memblock_region *r;
5130 /* Need to find movable_zone earlier when movable_node is specified. */
5131 find_usable_zone_for_movable();
5134 * If movable_node is specified, ignore kernelcore and movablecore
5137 if (movable_node_is_enabled()) {
5138 for_each_memblock(memory, r) {
5139 if (!memblock_is_hotpluggable(r))
5144 usable_startpfn = PFN_DOWN(r->base);
5145 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5146 min(usable_startpfn, zone_movable_pfn[nid]) :
5154 * If movablecore=nn[KMG] was specified, calculate what size of
5155 * kernelcore that corresponds so that memory usable for
5156 * any allocation type is evenly spread. If both kernelcore
5157 * and movablecore are specified, then the value of kernelcore
5158 * will be used for required_kernelcore if it's greater than
5159 * what movablecore would have allowed.
5161 if (required_movablecore) {
5162 unsigned long corepages;
5165 * Round-up so that ZONE_MOVABLE is at least as large as what
5166 * was requested by the user
5168 required_movablecore =
5169 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5170 corepages = totalpages - required_movablecore;
5172 required_kernelcore = max(required_kernelcore, corepages);
5175 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5176 if (!required_kernelcore)
5179 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5180 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5183 /* Spread kernelcore memory as evenly as possible throughout nodes */
5184 kernelcore_node = required_kernelcore / usable_nodes;
5185 for_each_node_state(nid, N_MEMORY) {
5186 unsigned long start_pfn, end_pfn;
5189 * Recalculate kernelcore_node if the division per node
5190 * now exceeds what is necessary to satisfy the requested
5191 * amount of memory for the kernel
5193 if (required_kernelcore < kernelcore_node)
5194 kernelcore_node = required_kernelcore / usable_nodes;
5197 * As the map is walked, we track how much memory is usable
5198 * by the kernel using kernelcore_remaining. When it is
5199 * 0, the rest of the node is usable by ZONE_MOVABLE
5201 kernelcore_remaining = kernelcore_node;
5203 /* Go through each range of PFNs within this node */
5204 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5205 unsigned long size_pages;
5207 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5208 if (start_pfn >= end_pfn)
5211 /* Account for what is only usable for kernelcore */
5212 if (start_pfn < usable_startpfn) {
5213 unsigned long kernel_pages;
5214 kernel_pages = min(end_pfn, usable_startpfn)
5217 kernelcore_remaining -= min(kernel_pages,
5218 kernelcore_remaining);
5219 required_kernelcore -= min(kernel_pages,
5220 required_kernelcore);
5222 /* Continue if range is now fully accounted */
5223 if (end_pfn <= usable_startpfn) {
5226 * Push zone_movable_pfn to the end so
5227 * that if we have to rebalance
5228 * kernelcore across nodes, we will
5229 * not double account here
5231 zone_movable_pfn[nid] = end_pfn;
5234 start_pfn = usable_startpfn;
5238 * The usable PFN range for ZONE_MOVABLE is from
5239 * start_pfn->end_pfn. Calculate size_pages as the
5240 * number of pages used as kernelcore
5242 size_pages = end_pfn - start_pfn;
5243 if (size_pages > kernelcore_remaining)
5244 size_pages = kernelcore_remaining;
5245 zone_movable_pfn[nid] = start_pfn + size_pages;
5248 * Some kernelcore has been met, update counts and
5249 * break if the kernelcore for this node has been
5252 required_kernelcore -= min(required_kernelcore,
5254 kernelcore_remaining -= size_pages;
5255 if (!kernelcore_remaining)
5261 * If there is still required_kernelcore, we do another pass with one
5262 * less node in the count. This will push zone_movable_pfn[nid] further
5263 * along on the nodes that still have memory until kernelcore is
5267 if (usable_nodes && required_kernelcore > usable_nodes)
5271 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5272 for (nid = 0; nid < MAX_NUMNODES; nid++)
5273 zone_movable_pfn[nid] =
5274 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5277 /* restore the node_state */
5278 node_states[N_MEMORY] = saved_node_state;
5281 /* Any regular or high memory on that node ? */
5282 static void check_for_memory(pg_data_t *pgdat, int nid)
5284 enum zone_type zone_type;
5286 if (N_MEMORY == N_NORMAL_MEMORY)
5289 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5290 struct zone *zone = &pgdat->node_zones[zone_type];
5291 if (populated_zone(zone)) {
5292 node_set_state(nid, N_HIGH_MEMORY);
5293 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5294 zone_type <= ZONE_NORMAL)
5295 node_set_state(nid, N_NORMAL_MEMORY);
5302 * free_area_init_nodes - Initialise all pg_data_t and zone data
5303 * @max_zone_pfn: an array of max PFNs for each zone
5305 * This will call free_area_init_node() for each active node in the system.
5306 * Using the page ranges provided by memblock_set_node(), the size of each
5307 * zone in each node and their holes is calculated. If the maximum PFN
5308 * between two adjacent zones match, it is assumed that the zone is empty.
5309 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5310 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5311 * starts where the previous one ended. For example, ZONE_DMA32 starts
5312 * at arch_max_dma_pfn.
5314 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5316 unsigned long start_pfn, end_pfn;
5319 /* Record where the zone boundaries are */
5320 memset(arch_zone_lowest_possible_pfn, 0,
5321 sizeof(arch_zone_lowest_possible_pfn));
5322 memset(arch_zone_highest_possible_pfn, 0,
5323 sizeof(arch_zone_highest_possible_pfn));
5324 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5325 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5326 for (i = 1; i < MAX_NR_ZONES; i++) {
5327 if (i == ZONE_MOVABLE)
5329 arch_zone_lowest_possible_pfn[i] =
5330 arch_zone_highest_possible_pfn[i-1];
5331 arch_zone_highest_possible_pfn[i] =
5332 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5334 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5335 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5337 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5338 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5339 find_zone_movable_pfns_for_nodes();
5341 /* Print out the zone ranges */
5342 printk("Zone ranges:\n");
5343 for (i = 0; i < MAX_NR_ZONES; i++) {
5344 if (i == ZONE_MOVABLE)
5346 printk(KERN_CONT " %-8s ", zone_names[i]);
5347 if (arch_zone_lowest_possible_pfn[i] ==
5348 arch_zone_highest_possible_pfn[i])
5349 printk(KERN_CONT "empty\n");
5351 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5352 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5353 (arch_zone_highest_possible_pfn[i]
5354 << PAGE_SHIFT) - 1);
5357 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5358 printk("Movable zone start for each node\n");
5359 for (i = 0; i < MAX_NUMNODES; i++) {
5360 if (zone_movable_pfn[i])
5361 printk(" Node %d: %#010lx\n", i,
5362 zone_movable_pfn[i] << PAGE_SHIFT);
5365 /* Print out the early node map */
5366 printk("Early memory node ranges\n");
5367 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5368 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5369 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5371 /* Initialise every node */
5372 mminit_verify_pageflags_layout();
5373 setup_nr_node_ids();
5374 for_each_online_node(nid) {
5375 pg_data_t *pgdat = NODE_DATA(nid);
5376 free_area_init_node(nid, NULL,
5377 find_min_pfn_for_node(nid), NULL);
5379 /* Any memory on that node */
5380 if (pgdat->node_present_pages)
5381 node_set_state(nid, N_MEMORY);
5382 check_for_memory(pgdat, nid);
5386 static int __init cmdline_parse_core(char *p, unsigned long *core)
5388 unsigned long long coremem;
5392 coremem = memparse(p, &p);
5393 *core = coremem >> PAGE_SHIFT;
5395 /* Paranoid check that UL is enough for the coremem value */
5396 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5402 * kernelcore=size sets the amount of memory for use for allocations that
5403 * cannot be reclaimed or migrated.
5405 static int __init cmdline_parse_kernelcore(char *p)
5407 return cmdline_parse_core(p, &required_kernelcore);
5411 * movablecore=size sets the amount of memory for use for allocations that
5412 * can be reclaimed or migrated.
5414 static int __init cmdline_parse_movablecore(char *p)
5416 return cmdline_parse_core(p, &required_movablecore);
5419 early_param("kernelcore", cmdline_parse_kernelcore);
5420 early_param("movablecore", cmdline_parse_movablecore);
5422 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5424 void adjust_managed_page_count(struct page *page, long count)
5426 spin_lock(&managed_page_count_lock);
5427 page_zone(page)->managed_pages += count;
5428 totalram_pages += count;
5429 #ifdef CONFIG_HIGHMEM
5430 if (PageHighMem(page))
5431 totalhigh_pages += count;
5433 spin_unlock(&managed_page_count_lock);
5435 EXPORT_SYMBOL(adjust_managed_page_count);
5437 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5440 unsigned long pages = 0;
5442 start = (void *)PAGE_ALIGN((unsigned long)start);
5443 end = (void *)((unsigned long)end & PAGE_MASK);
5444 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5445 if ((unsigned int)poison <= 0xFF)
5446 memset(pos, poison, PAGE_SIZE);
5447 free_reserved_page(virt_to_page(pos));
5451 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5452 s, pages << (PAGE_SHIFT - 10), start, end);
5456 EXPORT_SYMBOL(free_reserved_area);
5458 #ifdef CONFIG_HIGHMEM
5459 void free_highmem_page(struct page *page)
5461 __free_reserved_page(page);
5463 page_zone(page)->managed_pages++;
5469 void __init mem_init_print_info(const char *str)
5471 unsigned long physpages, codesize, datasize, rosize, bss_size;
5472 unsigned long init_code_size, init_data_size;
5474 physpages = get_num_physpages();
5475 codesize = _etext - _stext;
5476 datasize = _edata - _sdata;
5477 rosize = __end_rodata - __start_rodata;
5478 bss_size = __bss_stop - __bss_start;
5479 init_data_size = __init_end - __init_begin;
5480 init_code_size = _einittext - _sinittext;
5483 * Detect special cases and adjust section sizes accordingly:
5484 * 1) .init.* may be embedded into .data sections
5485 * 2) .init.text.* may be out of [__init_begin, __init_end],
5486 * please refer to arch/tile/kernel/vmlinux.lds.S.
5487 * 3) .rodata.* may be embedded into .text or .data sections.
5489 #define adj_init_size(start, end, size, pos, adj) \
5491 if (start <= pos && pos < end && size > adj) \
5495 adj_init_size(__init_begin, __init_end, init_data_size,
5496 _sinittext, init_code_size);
5497 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5498 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5499 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5500 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5502 #undef adj_init_size
5504 printk("Memory: %luK/%luK available "
5505 "(%luK kernel code, %luK rwdata, %luK rodata, "
5506 "%luK init, %luK bss, %luK reserved"
5507 #ifdef CONFIG_HIGHMEM
5511 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5512 codesize >> 10, datasize >> 10, rosize >> 10,
5513 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5514 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5515 #ifdef CONFIG_HIGHMEM
5516 totalhigh_pages << (PAGE_SHIFT-10),
5518 str ? ", " : "", str ? str : "");
5522 * set_dma_reserve - set the specified number of pages reserved in the first zone
5523 * @new_dma_reserve: The number of pages to mark reserved
5525 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5526 * In the DMA zone, a significant percentage may be consumed by kernel image
5527 * and other unfreeable allocations which can skew the watermarks badly. This
5528 * function may optionally be used to account for unfreeable pages in the
5529 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5530 * smaller per-cpu batchsize.
5532 void __init set_dma_reserve(unsigned long new_dma_reserve)
5534 dma_reserve = new_dma_reserve;
5537 void __init free_area_init(unsigned long *zones_size)
5539 free_area_init_node(0, zones_size,
5540 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5543 static int page_alloc_cpu_notify(struct notifier_block *self,
5544 unsigned long action, void *hcpu)
5546 int cpu = (unsigned long)hcpu;
5548 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5549 lru_add_drain_cpu(cpu);
5553 * Spill the event counters of the dead processor
5554 * into the current processors event counters.
5555 * This artificially elevates the count of the current
5558 vm_events_fold_cpu(cpu);
5561 * Zero the differential counters of the dead processor
5562 * so that the vm statistics are consistent.
5564 * This is only okay since the processor is dead and cannot
5565 * race with what we are doing.
5567 cpu_vm_stats_fold(cpu);
5572 void __init page_alloc_init(void)
5574 hotcpu_notifier(page_alloc_cpu_notify, 0);
5578 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5579 * or min_free_kbytes changes.
5581 static void calculate_totalreserve_pages(void)
5583 struct pglist_data *pgdat;
5584 unsigned long reserve_pages = 0;
5585 enum zone_type i, j;
5587 for_each_online_pgdat(pgdat) {
5588 for (i = 0; i < MAX_NR_ZONES; i++) {
5589 struct zone *zone = pgdat->node_zones + i;
5592 /* Find valid and maximum lowmem_reserve in the zone */
5593 for (j = i; j < MAX_NR_ZONES; j++) {
5594 if (zone->lowmem_reserve[j] > max)
5595 max = zone->lowmem_reserve[j];
5598 /* we treat the high watermark as reserved pages. */
5599 max += high_wmark_pages(zone);
5601 if (max > zone->managed_pages)
5602 max = zone->managed_pages;
5603 reserve_pages += max;
5605 * Lowmem reserves are not available to
5606 * GFP_HIGHUSER page cache allocations and
5607 * kswapd tries to balance zones to their high
5608 * watermark. As a result, neither should be
5609 * regarded as dirtyable memory, to prevent a
5610 * situation where reclaim has to clean pages
5611 * in order to balance the zones.
5613 zone->dirty_balance_reserve = max;
5616 dirty_balance_reserve = reserve_pages;
5617 totalreserve_pages = reserve_pages;
5621 * setup_per_zone_lowmem_reserve - called whenever
5622 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5623 * has a correct pages reserved value, so an adequate number of
5624 * pages are left in the zone after a successful __alloc_pages().
5626 static void setup_per_zone_lowmem_reserve(void)
5628 struct pglist_data *pgdat;
5629 enum zone_type j, idx;
5631 for_each_online_pgdat(pgdat) {
5632 for (j = 0; j < MAX_NR_ZONES; j++) {
5633 struct zone *zone = pgdat->node_zones + j;
5634 unsigned long managed_pages = zone->managed_pages;
5636 zone->lowmem_reserve[j] = 0;
5640 struct zone *lower_zone;
5644 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5645 sysctl_lowmem_reserve_ratio[idx] = 1;
5647 lower_zone = pgdat->node_zones + idx;
5648 lower_zone->lowmem_reserve[j] = managed_pages /
5649 sysctl_lowmem_reserve_ratio[idx];
5650 managed_pages += lower_zone->managed_pages;
5655 /* update totalreserve_pages */
5656 calculate_totalreserve_pages();
5659 static void __setup_per_zone_wmarks(void)
5661 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5662 unsigned long lowmem_pages = 0;
5664 unsigned long flags;
5666 /* Calculate total number of !ZONE_HIGHMEM pages */
5667 for_each_zone(zone) {
5668 if (!is_highmem(zone))
5669 lowmem_pages += zone->managed_pages;
5672 for_each_zone(zone) {
5675 spin_lock_irqsave(&zone->lock, flags);
5676 tmp = (u64)pages_min * zone->managed_pages;
5677 do_div(tmp, lowmem_pages);
5678 if (is_highmem(zone)) {
5680 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5681 * need highmem pages, so cap pages_min to a small
5684 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5685 * deltas controls asynch page reclaim, and so should
5686 * not be capped for highmem.
5688 unsigned long min_pages;
5690 min_pages = zone->managed_pages / 1024;
5691 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5692 zone->watermark[WMARK_MIN] = min_pages;
5695 * If it's a lowmem zone, reserve a number of pages
5696 * proportionate to the zone's size.
5698 zone->watermark[WMARK_MIN] = tmp;
5701 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5702 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5704 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5705 high_wmark_pages(zone) - low_wmark_pages(zone) -
5706 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5708 setup_zone_migrate_reserve(zone);
5709 spin_unlock_irqrestore(&zone->lock, flags);
5712 /* update totalreserve_pages */
5713 calculate_totalreserve_pages();
5717 * setup_per_zone_wmarks - called when min_free_kbytes changes
5718 * or when memory is hot-{added|removed}
5720 * Ensures that the watermark[min,low,high] values for each zone are set
5721 * correctly with respect to min_free_kbytes.
5723 void setup_per_zone_wmarks(void)
5725 mutex_lock(&zonelists_mutex);
5726 __setup_per_zone_wmarks();
5727 mutex_unlock(&zonelists_mutex);
5731 * The inactive anon list should be small enough that the VM never has to
5732 * do too much work, but large enough that each inactive page has a chance
5733 * to be referenced again before it is swapped out.
5735 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5736 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5737 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5738 * the anonymous pages are kept on the inactive list.
5741 * memory ratio inactive anon
5742 * -------------------------------------
5751 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5753 unsigned int gb, ratio;
5755 /* Zone size in gigabytes */
5756 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5758 ratio = int_sqrt(10 * gb);
5762 zone->inactive_ratio = ratio;
5765 static void __meminit setup_per_zone_inactive_ratio(void)
5770 calculate_zone_inactive_ratio(zone);
5774 * Initialise min_free_kbytes.
5776 * For small machines we want it small (128k min). For large machines
5777 * we want it large (64MB max). But it is not linear, because network
5778 * bandwidth does not increase linearly with machine size. We use
5780 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5781 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5797 int __meminit init_per_zone_wmark_min(void)
5799 unsigned long lowmem_kbytes;
5800 int new_min_free_kbytes;
5802 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5803 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5805 if (new_min_free_kbytes > user_min_free_kbytes) {
5806 min_free_kbytes = new_min_free_kbytes;
5807 if (min_free_kbytes < 128)
5808 min_free_kbytes = 128;
5809 if (min_free_kbytes > 65536)
5810 min_free_kbytes = 65536;
5812 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5813 new_min_free_kbytes, user_min_free_kbytes);
5815 setup_per_zone_wmarks();
5816 refresh_zone_stat_thresholds();
5817 setup_per_zone_lowmem_reserve();
5818 setup_per_zone_inactive_ratio();
5821 module_init(init_per_zone_wmark_min)
5824 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5825 * that we can call two helper functions whenever min_free_kbytes
5828 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5829 void __user *buffer, size_t *length, loff_t *ppos)
5833 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5838 user_min_free_kbytes = min_free_kbytes;
5839 setup_per_zone_wmarks();
5845 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5846 void __user *buffer, size_t *length, loff_t *ppos)
5851 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5856 zone->min_unmapped_pages = (zone->managed_pages *
5857 sysctl_min_unmapped_ratio) / 100;
5861 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5862 void __user *buffer, size_t *length, loff_t *ppos)
5867 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5872 zone->min_slab_pages = (zone->managed_pages *
5873 sysctl_min_slab_ratio) / 100;
5879 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5880 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5881 * whenever sysctl_lowmem_reserve_ratio changes.
5883 * The reserve ratio obviously has absolutely no relation with the
5884 * minimum watermarks. The lowmem reserve ratio can only make sense
5885 * if in function of the boot time zone sizes.
5887 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5888 void __user *buffer, size_t *length, loff_t *ppos)
5890 proc_dointvec_minmax(table, write, buffer, length, ppos);
5891 setup_per_zone_lowmem_reserve();
5896 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5897 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5898 * pagelist can have before it gets flushed back to buddy allocator.
5900 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5901 void __user *buffer, size_t *length, loff_t *ppos)
5904 int old_percpu_pagelist_fraction;
5907 mutex_lock(&pcp_batch_high_lock);
5908 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5910 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5911 if (!write || ret < 0)
5914 /* Sanity checking to avoid pcp imbalance */
5915 if (percpu_pagelist_fraction &&
5916 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
5917 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
5923 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
5926 for_each_populated_zone(zone) {
5929 for_each_possible_cpu(cpu)
5930 pageset_set_high_and_batch(zone,
5931 per_cpu_ptr(zone->pageset, cpu));
5934 mutex_unlock(&pcp_batch_high_lock);
5938 int hashdist = HASHDIST_DEFAULT;
5941 static int __init set_hashdist(char *str)
5945 hashdist = simple_strtoul(str, &str, 0);
5948 __setup("hashdist=", set_hashdist);
5952 * allocate a large system hash table from bootmem
5953 * - it is assumed that the hash table must contain an exact power-of-2
5954 * quantity of entries
5955 * - limit is the number of hash buckets, not the total allocation size
5957 void *__init alloc_large_system_hash(const char *tablename,
5958 unsigned long bucketsize,
5959 unsigned long numentries,
5962 unsigned int *_hash_shift,
5963 unsigned int *_hash_mask,
5964 unsigned long low_limit,
5965 unsigned long high_limit)
5967 unsigned long long max = high_limit;
5968 unsigned long log2qty, size;
5971 /* allow the kernel cmdline to have a say */
5973 /* round applicable memory size up to nearest megabyte */
5974 numentries = nr_kernel_pages;
5976 /* It isn't necessary when PAGE_SIZE >= 1MB */
5977 if (PAGE_SHIFT < 20)
5978 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5980 /* limit to 1 bucket per 2^scale bytes of low memory */
5981 if (scale > PAGE_SHIFT)
5982 numentries >>= (scale - PAGE_SHIFT);
5984 numentries <<= (PAGE_SHIFT - scale);
5986 /* Make sure we've got at least a 0-order allocation.. */
5987 if (unlikely(flags & HASH_SMALL)) {
5988 /* Makes no sense without HASH_EARLY */
5989 WARN_ON(!(flags & HASH_EARLY));
5990 if (!(numentries >> *_hash_shift)) {
5991 numentries = 1UL << *_hash_shift;
5992 BUG_ON(!numentries);
5994 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5995 numentries = PAGE_SIZE / bucketsize;
5997 numentries = roundup_pow_of_two(numentries);
5999 /* limit allocation size to 1/16 total memory by default */
6001 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6002 do_div(max, bucketsize);
6004 max = min(max, 0x80000000ULL);
6006 if (numentries < low_limit)
6007 numentries = low_limit;
6008 if (numentries > max)
6011 log2qty = ilog2(numentries);
6014 size = bucketsize << log2qty;
6015 if (flags & HASH_EARLY)
6016 table = memblock_virt_alloc_nopanic(size, 0);
6018 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6021 * If bucketsize is not a power-of-two, we may free
6022 * some pages at the end of hash table which
6023 * alloc_pages_exact() automatically does
6025 if (get_order(size) < MAX_ORDER) {
6026 table = alloc_pages_exact(size, GFP_ATOMIC);
6027 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6030 } while (!table && size > PAGE_SIZE && --log2qty);
6033 panic("Failed to allocate %s hash table\n", tablename);
6035 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6038 ilog2(size) - PAGE_SHIFT,
6042 *_hash_shift = log2qty;
6044 *_hash_mask = (1 << log2qty) - 1;
6049 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6050 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6053 #ifdef CONFIG_SPARSEMEM
6054 return __pfn_to_section(pfn)->pageblock_flags;
6056 return zone->pageblock_flags;
6057 #endif /* CONFIG_SPARSEMEM */
6060 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6062 #ifdef CONFIG_SPARSEMEM
6063 pfn &= (PAGES_PER_SECTION-1);
6064 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6066 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6067 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6068 #endif /* CONFIG_SPARSEMEM */
6072 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6073 * @page: The page within the block of interest
6074 * @pfn: The target page frame number
6075 * @end_bitidx: The last bit of interest to retrieve
6076 * @mask: mask of bits that the caller is interested in
6078 * Return: pageblock_bits flags
6080 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6081 unsigned long end_bitidx,
6085 unsigned long *bitmap;
6086 unsigned long bitidx, word_bitidx;
6089 zone = page_zone(page);
6090 bitmap = get_pageblock_bitmap(zone, pfn);
6091 bitidx = pfn_to_bitidx(zone, pfn);
6092 word_bitidx = bitidx / BITS_PER_LONG;
6093 bitidx &= (BITS_PER_LONG-1);
6095 word = bitmap[word_bitidx];
6096 bitidx += end_bitidx;
6097 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6101 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6102 * @page: The page within the block of interest
6103 * @flags: The flags to set
6104 * @pfn: The target page frame number
6105 * @end_bitidx: The last bit of interest
6106 * @mask: mask of bits that the caller is interested in
6108 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6110 unsigned long end_bitidx,
6114 unsigned long *bitmap;
6115 unsigned long bitidx, word_bitidx;
6116 unsigned long old_word, word;
6118 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6120 zone = page_zone(page);
6121 bitmap = get_pageblock_bitmap(zone, pfn);
6122 bitidx = pfn_to_bitidx(zone, pfn);
6123 word_bitidx = bitidx / BITS_PER_LONG;
6124 bitidx &= (BITS_PER_LONG-1);
6126 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6128 bitidx += end_bitidx;
6129 mask <<= (BITS_PER_LONG - bitidx - 1);
6130 flags <<= (BITS_PER_LONG - bitidx - 1);
6132 word = ACCESS_ONCE(bitmap[word_bitidx]);
6134 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6135 if (word == old_word)
6142 * This function checks whether pageblock includes unmovable pages or not.
6143 * If @count is not zero, it is okay to include less @count unmovable pages
6145 * PageLRU check without isolation or lru_lock could race so that
6146 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6147 * expect this function should be exact.
6149 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6150 bool skip_hwpoisoned_pages)
6152 unsigned long pfn, iter, found;
6156 * For avoiding noise data, lru_add_drain_all() should be called
6157 * If ZONE_MOVABLE, the zone never contains unmovable pages
6159 if (zone_idx(zone) == ZONE_MOVABLE)
6161 mt = get_pageblock_migratetype(page);
6162 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6165 pfn = page_to_pfn(page);
6166 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6167 unsigned long check = pfn + iter;
6169 if (!pfn_valid_within(check))
6172 page = pfn_to_page(check);
6175 * Hugepages are not in LRU lists, but they're movable.
6176 * We need not scan over tail pages bacause we don't
6177 * handle each tail page individually in migration.
6179 if (PageHuge(page)) {
6180 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6185 * We can't use page_count without pin a page
6186 * because another CPU can free compound page.
6187 * This check already skips compound tails of THP
6188 * because their page->_count is zero at all time.
6190 if (!atomic_read(&page->_count)) {
6191 if (PageBuddy(page))
6192 iter += (1 << page_order(page)) - 1;
6197 * The HWPoisoned page may be not in buddy system, and
6198 * page_count() is not 0.
6200 if (skip_hwpoisoned_pages && PageHWPoison(page))
6206 * If there are RECLAIMABLE pages, we need to check it.
6207 * But now, memory offline itself doesn't call shrink_slab()
6208 * and it still to be fixed.
6211 * If the page is not RAM, page_count()should be 0.
6212 * we don't need more check. This is an _used_ not-movable page.
6214 * The problematic thing here is PG_reserved pages. PG_reserved
6215 * is set to both of a memory hole page and a _used_ kernel
6224 bool is_pageblock_removable_nolock(struct page *page)
6230 * We have to be careful here because we are iterating over memory
6231 * sections which are not zone aware so we might end up outside of
6232 * the zone but still within the section.
6233 * We have to take care about the node as well. If the node is offline
6234 * its NODE_DATA will be NULL - see page_zone.
6236 if (!node_online(page_to_nid(page)))
6239 zone = page_zone(page);
6240 pfn = page_to_pfn(page);
6241 if (!zone_spans_pfn(zone, pfn))
6244 return !has_unmovable_pages(zone, page, 0, true);
6249 static unsigned long pfn_max_align_down(unsigned long pfn)
6251 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6252 pageblock_nr_pages) - 1);
6255 static unsigned long pfn_max_align_up(unsigned long pfn)
6257 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6258 pageblock_nr_pages));
6261 /* [start, end) must belong to a single zone. */
6262 static int __alloc_contig_migrate_range(struct compact_control *cc,
6263 unsigned long start, unsigned long end)
6265 /* This function is based on compact_zone() from compaction.c. */
6266 unsigned long nr_reclaimed;
6267 unsigned long pfn = start;
6268 unsigned int tries = 0;
6273 while (pfn < end || !list_empty(&cc->migratepages)) {
6274 if (fatal_signal_pending(current)) {
6279 if (list_empty(&cc->migratepages)) {
6280 cc->nr_migratepages = 0;
6281 pfn = isolate_migratepages_range(cc->zone, cc,
6288 } else if (++tries == 5) {
6289 ret = ret < 0 ? ret : -EBUSY;
6293 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6295 cc->nr_migratepages -= nr_reclaimed;
6297 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6298 NULL, 0, cc->mode, MR_CMA);
6301 putback_movable_pages(&cc->migratepages);
6308 * alloc_contig_range() -- tries to allocate given range of pages
6309 * @start: start PFN to allocate
6310 * @end: one-past-the-last PFN to allocate
6311 * @migratetype: migratetype of the underlaying pageblocks (either
6312 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6313 * in range must have the same migratetype and it must
6314 * be either of the two.
6316 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6317 * aligned, however it's the caller's responsibility to guarantee that
6318 * we are the only thread that changes migrate type of pageblocks the
6321 * The PFN range must belong to a single zone.
6323 * Returns zero on success or negative error code. On success all
6324 * pages which PFN is in [start, end) are allocated for the caller and
6325 * need to be freed with free_contig_range().
6327 int alloc_contig_range(unsigned long start, unsigned long end,
6328 unsigned migratetype)
6330 unsigned long outer_start, outer_end;
6333 struct compact_control cc = {
6334 .nr_migratepages = 0,
6336 .zone = page_zone(pfn_to_page(start)),
6337 .mode = MIGRATE_SYNC,
6338 .ignore_skip_hint = true,
6340 INIT_LIST_HEAD(&cc.migratepages);
6343 * What we do here is we mark all pageblocks in range as
6344 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6345 * have different sizes, and due to the way page allocator
6346 * work, we align the range to biggest of the two pages so
6347 * that page allocator won't try to merge buddies from
6348 * different pageblocks and change MIGRATE_ISOLATE to some
6349 * other migration type.
6351 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6352 * migrate the pages from an unaligned range (ie. pages that
6353 * we are interested in). This will put all the pages in
6354 * range back to page allocator as MIGRATE_ISOLATE.
6356 * When this is done, we take the pages in range from page
6357 * allocator removing them from the buddy system. This way
6358 * page allocator will never consider using them.
6360 * This lets us mark the pageblocks back as
6361 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6362 * aligned range but not in the unaligned, original range are
6363 * put back to page allocator so that buddy can use them.
6366 ret = start_isolate_page_range(pfn_max_align_down(start),
6367 pfn_max_align_up(end), migratetype,
6372 ret = __alloc_contig_migrate_range(&cc, start, end);
6377 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6378 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6379 * more, all pages in [start, end) are free in page allocator.
6380 * What we are going to do is to allocate all pages from
6381 * [start, end) (that is remove them from page allocator).
6383 * The only problem is that pages at the beginning and at the
6384 * end of interesting range may be not aligned with pages that
6385 * page allocator holds, ie. they can be part of higher order
6386 * pages. Because of this, we reserve the bigger range and
6387 * once this is done free the pages we are not interested in.
6389 * We don't have to hold zone->lock here because the pages are
6390 * isolated thus they won't get removed from buddy.
6393 lru_add_drain_all();
6397 outer_start = start;
6398 while (!PageBuddy(pfn_to_page(outer_start))) {
6399 if (++order >= MAX_ORDER) {
6403 outer_start &= ~0UL << order;
6406 /* Make sure the range is really isolated. */
6407 if (test_pages_isolated(outer_start, end, false)) {
6408 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6415 /* Grab isolated pages from freelists. */
6416 outer_end = isolate_freepages_range(&cc, outer_start, end);
6422 /* Free head and tail (if any) */
6423 if (start != outer_start)
6424 free_contig_range(outer_start, start - outer_start);
6425 if (end != outer_end)
6426 free_contig_range(end, outer_end - end);
6429 undo_isolate_page_range(pfn_max_align_down(start),
6430 pfn_max_align_up(end), migratetype);
6434 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6436 unsigned int count = 0;
6438 for (; nr_pages--; pfn++) {
6439 struct page *page = pfn_to_page(pfn);
6441 count += page_count(page) != 1;
6444 WARN(count != 0, "%d pages are still in use!\n", count);
6448 #ifdef CONFIG_MEMORY_HOTPLUG
6450 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6451 * page high values need to be recalulated.
6453 void __meminit zone_pcp_update(struct zone *zone)
6456 mutex_lock(&pcp_batch_high_lock);
6457 for_each_possible_cpu(cpu)
6458 pageset_set_high_and_batch(zone,
6459 per_cpu_ptr(zone->pageset, cpu));
6460 mutex_unlock(&pcp_batch_high_lock);
6464 void zone_pcp_reset(struct zone *zone)
6466 unsigned long flags;
6468 struct per_cpu_pageset *pset;
6470 /* avoid races with drain_pages() */
6471 local_irq_save(flags);
6472 if (zone->pageset != &boot_pageset) {
6473 for_each_online_cpu(cpu) {
6474 pset = per_cpu_ptr(zone->pageset, cpu);
6475 drain_zonestat(zone, pset);
6477 free_percpu(zone->pageset);
6478 zone->pageset = &boot_pageset;
6480 local_irq_restore(flags);
6483 #ifdef CONFIG_MEMORY_HOTREMOVE
6485 * All pages in the range must be isolated before calling this.
6488 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6492 unsigned int order, i;
6494 unsigned long flags;
6495 /* find the first valid pfn */
6496 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6501 zone = page_zone(pfn_to_page(pfn));
6502 spin_lock_irqsave(&zone->lock, flags);
6504 while (pfn < end_pfn) {
6505 if (!pfn_valid(pfn)) {
6509 page = pfn_to_page(pfn);
6511 * The HWPoisoned page may be not in buddy system, and
6512 * page_count() is not 0.
6514 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6516 SetPageReserved(page);
6520 BUG_ON(page_count(page));
6521 BUG_ON(!PageBuddy(page));
6522 order = page_order(page);
6523 #ifdef CONFIG_DEBUG_VM
6524 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6525 pfn, 1 << order, end_pfn);
6527 list_del(&page->lru);
6528 rmv_page_order(page);
6529 zone->free_area[order].nr_free--;
6530 for (i = 0; i < (1 << order); i++)
6531 SetPageReserved((page+i));
6532 pfn += (1 << order);
6534 spin_unlock_irqrestore(&zone->lock, flags);
6538 #ifdef CONFIG_MEMORY_FAILURE
6539 bool is_free_buddy_page(struct page *page)
6541 struct zone *zone = page_zone(page);
6542 unsigned long pfn = page_to_pfn(page);
6543 unsigned long flags;
6546 spin_lock_irqsave(&zone->lock, flags);
6547 for (order = 0; order < MAX_ORDER; order++) {
6548 struct page *page_head = page - (pfn & ((1 << order) - 1));
6550 if (PageBuddy(page_head) && page_order(page_head) >= order)
6553 spin_unlock_irqrestore(&zone->lock, flags);
6555 return order < MAX_ORDER;
6559 static const struct trace_print_flags pageflag_names[] = {
6560 {1UL << PG_locked, "locked" },
6561 {1UL << PG_error, "error" },
6562 {1UL << PG_referenced, "referenced" },
6563 {1UL << PG_uptodate, "uptodate" },
6564 {1UL << PG_dirty, "dirty" },
6565 {1UL << PG_lru, "lru" },
6566 {1UL << PG_active, "active" },
6567 {1UL << PG_slab, "slab" },
6568 {1UL << PG_owner_priv_1, "owner_priv_1" },
6569 {1UL << PG_arch_1, "arch_1" },
6570 {1UL << PG_reserved, "reserved" },
6571 {1UL << PG_private, "private" },
6572 {1UL << PG_private_2, "private_2" },
6573 {1UL << PG_writeback, "writeback" },
6574 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6575 {1UL << PG_head, "head" },
6576 {1UL << PG_tail, "tail" },
6578 {1UL << PG_compound, "compound" },
6580 {1UL << PG_swapcache, "swapcache" },
6581 {1UL << PG_mappedtodisk, "mappedtodisk" },
6582 {1UL << PG_reclaim, "reclaim" },
6583 {1UL << PG_swapbacked, "swapbacked" },
6584 {1UL << PG_unevictable, "unevictable" },
6586 {1UL << PG_mlocked, "mlocked" },
6588 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6589 {1UL << PG_uncached, "uncached" },
6591 #ifdef CONFIG_MEMORY_FAILURE
6592 {1UL << PG_hwpoison, "hwpoison" },
6594 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6595 {1UL << PG_compound_lock, "compound_lock" },
6599 static void dump_page_flags(unsigned long flags)
6601 const char *delim = "";
6605 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6607 printk(KERN_ALERT "page flags: %#lx(", flags);
6609 /* remove zone id */
6610 flags &= (1UL << NR_PAGEFLAGS) - 1;
6612 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6614 mask = pageflag_names[i].mask;
6615 if ((flags & mask) != mask)
6619 printk("%s%s", delim, pageflag_names[i].name);
6623 /* check for left over flags */
6625 printk("%s%#lx", delim, flags);
6630 void dump_page_badflags(struct page *page, const char *reason,
6631 unsigned long badflags)
6634 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6635 page, atomic_read(&page->_count), page_mapcount(page),
6636 page->mapping, page->index);
6637 dump_page_flags(page->flags);
6639 pr_alert("page dumped because: %s\n", reason);
6640 if (page->flags & badflags) {
6641 pr_alert("bad because of flags:\n");
6642 dump_page_flags(page->flags & badflags);
6644 mem_cgroup_print_bad_page(page);
6647 void dump_page(struct page *page, const char *reason)
6649 dump_page_badflags(page, reason, 0);
6651 EXPORT_SYMBOL(dump_page);