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/migrate.h>
60 #include <linux/page-debug-flags.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
69 static DEFINE_MUTEX(pcp_batch_high_lock);
71 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
72 DEFINE_PER_CPU(int, numa_node);
73 EXPORT_PER_CPU_SYMBOL(numa_node);
76 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
78 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
79 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
80 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
81 * defined in <linux/topology.h>.
83 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
84 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
88 * Array of node states.
90 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
91 [N_POSSIBLE] = NODE_MASK_ALL,
92 [N_ONLINE] = { { [0] = 1UL } },
94 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
96 [N_HIGH_MEMORY] = { { [0] = 1UL } },
98 #ifdef CONFIG_MOVABLE_NODE
99 [N_MEMORY] = { { [0] = 1UL } },
101 [N_CPU] = { { [0] = 1UL } },
104 EXPORT_SYMBOL(node_states);
106 unsigned long totalram_pages __read_mostly;
107 unsigned long totalreserve_pages __read_mostly;
109 * When calculating the number of globally allowed dirty pages, there
110 * is a certain number of per-zone reserves that should not be
111 * considered dirtyable memory. This is the sum of those reserves
112 * over all existing zones that contribute dirtyable memory.
114 unsigned long dirty_balance_reserve __read_mostly;
116 int percpu_pagelist_fraction;
117 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
119 #ifdef CONFIG_PM_SLEEP
121 * The following functions are used by the suspend/hibernate code to temporarily
122 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
123 * while devices are suspended. To avoid races with the suspend/hibernate code,
124 * they should always be called with pm_mutex held (gfp_allowed_mask also should
125 * only be modified with pm_mutex held, unless the suspend/hibernate code is
126 * guaranteed not to run in parallel with that modification).
129 static gfp_t saved_gfp_mask;
131 void pm_restore_gfp_mask(void)
133 WARN_ON(!mutex_is_locked(&pm_mutex));
134 if (saved_gfp_mask) {
135 gfp_allowed_mask = saved_gfp_mask;
140 void pm_restrict_gfp_mask(void)
142 WARN_ON(!mutex_is_locked(&pm_mutex));
143 WARN_ON(saved_gfp_mask);
144 saved_gfp_mask = gfp_allowed_mask;
145 gfp_allowed_mask &= ~GFP_IOFS;
148 bool pm_suspended_storage(void)
150 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
154 #endif /* CONFIG_PM_SLEEP */
156 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
157 int pageblock_order __read_mostly;
160 static void __free_pages_ok(struct page *page, unsigned int order);
163 * results with 256, 32 in the lowmem_reserve sysctl:
164 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
165 * 1G machine -> (16M dma, 784M normal, 224M high)
166 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
167 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
168 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
170 * TBD: should special case ZONE_DMA32 machines here - in those we normally
171 * don't need any ZONE_NORMAL reservation
173 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
174 #ifdef CONFIG_ZONE_DMA
177 #ifdef CONFIG_ZONE_DMA32
180 #ifdef CONFIG_HIGHMEM
186 EXPORT_SYMBOL(totalram_pages);
188 static char * const zone_names[MAX_NR_ZONES] = {
189 #ifdef CONFIG_ZONE_DMA
192 #ifdef CONFIG_ZONE_DMA32
196 #ifdef CONFIG_HIGHMEM
202 int min_free_kbytes = 1024;
204 static unsigned long __meminitdata nr_kernel_pages;
205 static unsigned long __meminitdata nr_all_pages;
206 static unsigned long __meminitdata dma_reserve;
208 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
209 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
210 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
211 static unsigned long __initdata required_kernelcore;
212 static unsigned long __initdata required_movablecore;
213 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
215 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
217 EXPORT_SYMBOL(movable_zone);
218 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
221 int nr_node_ids __read_mostly = MAX_NUMNODES;
222 int nr_online_nodes __read_mostly = 1;
223 EXPORT_SYMBOL(nr_node_ids);
224 EXPORT_SYMBOL(nr_online_nodes);
227 int page_group_by_mobility_disabled __read_mostly;
229 void set_pageblock_migratetype(struct page *page, int migratetype)
232 if (unlikely(page_group_by_mobility_disabled))
233 migratetype = MIGRATE_UNMOVABLE;
235 set_pageblock_flags_group(page, (unsigned long)migratetype,
236 PB_migrate, PB_migrate_end);
239 bool oom_killer_disabled __read_mostly;
241 #ifdef CONFIG_DEBUG_VM
242 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
246 unsigned long pfn = page_to_pfn(page);
247 unsigned long sp, start_pfn;
250 seq = zone_span_seqbegin(zone);
251 start_pfn = zone->zone_start_pfn;
252 sp = zone->spanned_pages;
253 if (!zone_spans_pfn(zone, pfn))
255 } while (zone_span_seqretry(zone, seq));
258 pr_err("page %lu outside zone [ %lu - %lu ]\n",
259 pfn, start_pfn, start_pfn + sp);
264 static int page_is_consistent(struct zone *zone, struct page *page)
266 if (!pfn_valid_within(page_to_pfn(page)))
268 if (zone != page_zone(page))
274 * Temporary debugging check for pages not lying within a given zone.
276 static int bad_range(struct zone *zone, struct page *page)
278 if (page_outside_zone_boundaries(zone, page))
280 if (!page_is_consistent(zone, page))
286 static inline int bad_range(struct zone *zone, struct page *page)
292 static void bad_page(struct page *page)
294 static unsigned long resume;
295 static unsigned long nr_shown;
296 static unsigned long nr_unshown;
298 /* Don't complain about poisoned pages */
299 if (PageHWPoison(page)) {
300 page_mapcount_reset(page); /* remove PageBuddy */
305 * Allow a burst of 60 reports, then keep quiet for that minute;
306 * or allow a steady drip of one report per second.
308 if (nr_shown == 60) {
309 if (time_before(jiffies, resume)) {
315 "BUG: Bad page state: %lu messages suppressed\n",
322 resume = jiffies + 60 * HZ;
324 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
325 current->comm, page_to_pfn(page));
331 /* Leave bad fields for debug, except PageBuddy could make trouble */
332 page_mapcount_reset(page); /* remove PageBuddy */
333 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
337 * Higher-order pages are called "compound pages". They are structured thusly:
339 * The first PAGE_SIZE page is called the "head page".
341 * The remaining PAGE_SIZE pages are called "tail pages".
343 * All pages have PG_compound set. All tail pages have their ->first_page
344 * pointing at the head page.
346 * The first tail page's ->lru.next holds the address of the compound page's
347 * put_page() function. Its ->lru.prev holds the order of allocation.
348 * This usage means that zero-order pages may not be compound.
351 static void free_compound_page(struct page *page)
353 __free_pages_ok(page, compound_order(page));
356 void prep_compound_page(struct page *page, unsigned long order)
359 int nr_pages = 1 << order;
361 set_compound_page_dtor(page, free_compound_page);
362 set_compound_order(page, order);
364 for (i = 1; i < nr_pages; i++) {
365 struct page *p = page + i;
367 set_page_count(p, 0);
368 p->first_page = page;
372 /* update __split_huge_page_refcount if you change this function */
373 static int destroy_compound_page(struct page *page, unsigned long order)
376 int nr_pages = 1 << order;
379 if (unlikely(compound_order(page) != order)) {
384 __ClearPageHead(page);
386 for (i = 1; i < nr_pages; i++) {
387 struct page *p = page + i;
389 if (unlikely(!PageTail(p) || (p->first_page != page))) {
399 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
404 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
405 * and __GFP_HIGHMEM from hard or soft interrupt context.
407 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
408 for (i = 0; i < (1 << order); i++)
409 clear_highpage(page + i);
412 #ifdef CONFIG_DEBUG_PAGEALLOC
413 unsigned int _debug_guardpage_minorder;
415 static int __init debug_guardpage_minorder_setup(char *buf)
419 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
420 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
423 _debug_guardpage_minorder = res;
424 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
427 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
429 static inline void set_page_guard_flag(struct page *page)
431 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
434 static inline void clear_page_guard_flag(struct page *page)
436 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
439 static inline void set_page_guard_flag(struct page *page) { }
440 static inline void clear_page_guard_flag(struct page *page) { }
443 static inline void set_page_order(struct page *page, int order)
445 set_page_private(page, order);
446 __SetPageBuddy(page);
449 static inline void rmv_page_order(struct page *page)
451 __ClearPageBuddy(page);
452 set_page_private(page, 0);
456 * Locate the struct page for both the matching buddy in our
457 * pair (buddy1) and the combined O(n+1) page they form (page).
459 * 1) Any buddy B1 will have an order O twin B2 which satisfies
460 * the following equation:
462 * For example, if the starting buddy (buddy2) is #8 its order
464 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
466 * 2) Any buddy B will have an order O+1 parent P which
467 * satisfies the following equation:
470 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
472 static inline unsigned long
473 __find_buddy_index(unsigned long page_idx, unsigned int order)
475 return page_idx ^ (1 << order);
479 * This function checks whether a page is free && is the buddy
480 * we can do coalesce a page and its buddy if
481 * (a) the buddy is not in a hole &&
482 * (b) the buddy is in the buddy system &&
483 * (c) a page and its buddy have the same order &&
484 * (d) a page and its buddy are in the same zone.
486 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
487 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
489 * For recording page's order, we use page_private(page).
491 static inline int page_is_buddy(struct page *page, struct page *buddy,
494 if (!pfn_valid_within(page_to_pfn(buddy)))
497 if (page_zone_id(page) != page_zone_id(buddy))
500 if (page_is_guard(buddy) && page_order(buddy) == order) {
501 VM_BUG_ON(page_count(buddy) != 0);
505 if (PageBuddy(buddy) && page_order(buddy) == order) {
506 VM_BUG_ON(page_count(buddy) != 0);
513 * Freeing function for a buddy system allocator.
515 * The concept of a buddy system is to maintain direct-mapped table
516 * (containing bit values) for memory blocks of various "orders".
517 * The bottom level table contains the map for the smallest allocatable
518 * units of memory (here, pages), and each level above it describes
519 * pairs of units from the levels below, hence, "buddies".
520 * At a high level, all that happens here is marking the table entry
521 * at the bottom level available, and propagating the changes upward
522 * as necessary, plus some accounting needed to play nicely with other
523 * parts of the VM system.
524 * At each level, we keep a list of pages, which are heads of continuous
525 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
526 * order is recorded in page_private(page) field.
527 * So when we are allocating or freeing one, we can derive the state of the
528 * other. That is, if we allocate a small block, and both were
529 * free, the remainder of the region must be split into blocks.
530 * If a block is freed, and its buddy is also free, then this
531 * triggers coalescing into a block of larger size.
536 static inline void __free_one_page(struct page *page,
537 struct zone *zone, unsigned int order,
540 unsigned long page_idx;
541 unsigned long combined_idx;
542 unsigned long uninitialized_var(buddy_idx);
545 VM_BUG_ON(!zone_is_initialized(zone));
547 if (unlikely(PageCompound(page)))
548 if (unlikely(destroy_compound_page(page, order)))
551 VM_BUG_ON(migratetype == -1);
553 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
555 VM_BUG_ON(page_idx & ((1 << order) - 1));
556 VM_BUG_ON(bad_range(zone, page));
558 while (order < MAX_ORDER-1) {
559 buddy_idx = __find_buddy_index(page_idx, order);
560 buddy = page + (buddy_idx - page_idx);
561 if (!page_is_buddy(page, buddy, order))
564 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
565 * merge with it and move up one order.
567 if (page_is_guard(buddy)) {
568 clear_page_guard_flag(buddy);
569 set_page_private(page, 0);
570 __mod_zone_freepage_state(zone, 1 << order,
573 list_del(&buddy->lru);
574 zone->free_area[order].nr_free--;
575 rmv_page_order(buddy);
577 combined_idx = buddy_idx & page_idx;
578 page = page + (combined_idx - page_idx);
579 page_idx = combined_idx;
582 set_page_order(page, order);
585 * If this is not the largest possible page, check if the buddy
586 * of the next-highest order is free. If it is, it's possible
587 * that pages are being freed that will coalesce soon. In case,
588 * that is happening, add the free page to the tail of the list
589 * so it's less likely to be used soon and more likely to be merged
590 * as a higher order page
592 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
593 struct page *higher_page, *higher_buddy;
594 combined_idx = buddy_idx & page_idx;
595 higher_page = page + (combined_idx - page_idx);
596 buddy_idx = __find_buddy_index(combined_idx, order + 1);
597 higher_buddy = higher_page + (buddy_idx - combined_idx);
598 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
599 list_add_tail(&page->lru,
600 &zone->free_area[order].free_list[migratetype]);
605 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
607 zone->free_area[order].nr_free++;
610 static inline int free_pages_check(struct page *page)
612 if (unlikely(page_mapcount(page) |
613 (page->mapping != NULL) |
614 (atomic_read(&page->_count) != 0) |
615 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
616 (mem_cgroup_bad_page_check(page)))) {
620 page_nid_reset_last(page);
621 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
622 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
627 * Frees a number of pages from the PCP lists
628 * Assumes all pages on list are in same zone, and of same order.
629 * count is the number of pages to free.
631 * If the zone was previously in an "all pages pinned" state then look to
632 * see if this freeing clears that state.
634 * And clear the zone's pages_scanned counter, to hold off the "all pages are
635 * pinned" detection logic.
637 static void free_pcppages_bulk(struct zone *zone, int count,
638 struct per_cpu_pages *pcp)
644 spin_lock(&zone->lock);
645 zone->all_unreclaimable = 0;
646 zone->pages_scanned = 0;
650 struct list_head *list;
653 * Remove pages from lists in a round-robin fashion. A
654 * batch_free count is maintained that is incremented when an
655 * empty list is encountered. This is so more pages are freed
656 * off fuller lists instead of spinning excessively around empty
661 if (++migratetype == MIGRATE_PCPTYPES)
663 list = &pcp->lists[migratetype];
664 } while (list_empty(list));
666 /* This is the only non-empty list. Free them all. */
667 if (batch_free == MIGRATE_PCPTYPES)
668 batch_free = to_free;
671 int mt; /* migratetype of the to-be-freed page */
673 page = list_entry(list->prev, struct page, lru);
674 /* must delete as __free_one_page list manipulates */
675 list_del(&page->lru);
676 mt = get_freepage_migratetype(page);
677 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
678 __free_one_page(page, zone, 0, mt);
679 trace_mm_page_pcpu_drain(page, 0, mt);
680 if (likely(!is_migrate_isolate_page(page))) {
681 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
682 if (is_migrate_cma(mt))
683 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
685 } while (--to_free && --batch_free && !list_empty(list));
687 spin_unlock(&zone->lock);
690 static void free_one_page(struct zone *zone, struct page *page, int order,
693 spin_lock(&zone->lock);
694 zone->all_unreclaimable = 0;
695 zone->pages_scanned = 0;
697 __free_one_page(page, zone, order, migratetype);
698 if (unlikely(!is_migrate_isolate(migratetype)))
699 __mod_zone_freepage_state(zone, 1 << order, migratetype);
700 spin_unlock(&zone->lock);
703 static bool free_pages_prepare(struct page *page, unsigned int order)
708 trace_mm_page_free(page, order);
709 kmemcheck_free_shadow(page, order);
712 page->mapping = NULL;
713 for (i = 0; i < (1 << order); i++)
714 bad += free_pages_check(page + i);
718 if (!PageHighMem(page)) {
719 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
720 debug_check_no_obj_freed(page_address(page),
723 arch_free_page(page, order);
724 kernel_map_pages(page, 1 << order, 0);
729 static void __free_pages_ok(struct page *page, unsigned int order)
734 if (!free_pages_prepare(page, order))
737 local_irq_save(flags);
738 __count_vm_events(PGFREE, 1 << order);
739 migratetype = get_pageblock_migratetype(page);
740 set_freepage_migratetype(page, migratetype);
741 free_one_page(page_zone(page), page, order, migratetype);
742 local_irq_restore(flags);
746 * Read access to zone->managed_pages is safe because it's unsigned long,
747 * but we still need to serialize writers. Currently all callers of
748 * __free_pages_bootmem() except put_page_bootmem() should only be used
749 * at boot time. So for shorter boot time, we shift the burden to
750 * put_page_bootmem() to serialize writers.
752 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
754 unsigned int nr_pages = 1 << order;
758 for (loop = 0; loop < nr_pages; loop++) {
759 struct page *p = &page[loop];
761 if (loop + 1 < nr_pages)
763 __ClearPageReserved(p);
764 set_page_count(p, 0);
767 page_zone(page)->managed_pages += 1 << order;
768 set_page_refcounted(page);
769 __free_pages(page, order);
773 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
774 void __init init_cma_reserved_pageblock(struct page *page)
776 unsigned i = pageblock_nr_pages;
777 struct page *p = page;
780 __ClearPageReserved(p);
781 set_page_count(p, 0);
784 set_page_refcounted(page);
785 set_pageblock_migratetype(page, MIGRATE_CMA);
786 __free_pages(page, pageblock_order);
787 totalram_pages += pageblock_nr_pages;
788 #ifdef CONFIG_HIGHMEM
789 if (PageHighMem(page))
790 totalhigh_pages += pageblock_nr_pages;
796 * The order of subdivision here is critical for the IO subsystem.
797 * Please do not alter this order without good reasons and regression
798 * testing. Specifically, as large blocks of memory are subdivided,
799 * the order in which smaller blocks are delivered depends on the order
800 * they're subdivided in this function. This is the primary factor
801 * influencing the order in which pages are delivered to the IO
802 * subsystem according to empirical testing, and this is also justified
803 * by considering the behavior of a buddy system containing a single
804 * large block of memory acted on by a series of small allocations.
805 * This behavior is a critical factor in sglist merging's success.
809 static inline void expand(struct zone *zone, struct page *page,
810 int low, int high, struct free_area *area,
813 unsigned long size = 1 << high;
819 VM_BUG_ON(bad_range(zone, &page[size]));
821 #ifdef CONFIG_DEBUG_PAGEALLOC
822 if (high < debug_guardpage_minorder()) {
824 * Mark as guard pages (or page), that will allow to
825 * merge back to allocator when buddy will be freed.
826 * Corresponding page table entries will not be touched,
827 * pages will stay not present in virtual address space
829 INIT_LIST_HEAD(&page[size].lru);
830 set_page_guard_flag(&page[size]);
831 set_page_private(&page[size], high);
832 /* Guard pages are not available for any usage */
833 __mod_zone_freepage_state(zone, -(1 << high),
838 list_add(&page[size].lru, &area->free_list[migratetype]);
840 set_page_order(&page[size], high);
845 * This page is about to be returned from the page allocator
847 static inline int check_new_page(struct page *page)
849 if (unlikely(page_mapcount(page) |
850 (page->mapping != NULL) |
851 (atomic_read(&page->_count) != 0) |
852 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
853 (mem_cgroup_bad_page_check(page)))) {
860 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
864 for (i = 0; i < (1 << order); i++) {
865 struct page *p = page + i;
866 if (unlikely(check_new_page(p)))
870 set_page_private(page, 0);
871 set_page_refcounted(page);
873 arch_alloc_page(page, order);
874 kernel_map_pages(page, 1 << order, 1);
876 if (gfp_flags & __GFP_ZERO)
877 prep_zero_page(page, order, gfp_flags);
879 if (order && (gfp_flags & __GFP_COMP))
880 prep_compound_page(page, order);
886 * Go through the free lists for the given migratetype and remove
887 * the smallest available page from the freelists
890 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
893 unsigned int current_order;
894 struct free_area * area;
897 /* Find a page of the appropriate size in the preferred list */
898 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
899 area = &(zone->free_area[current_order]);
900 if (list_empty(&area->free_list[migratetype]))
903 page = list_entry(area->free_list[migratetype].next,
905 list_del(&page->lru);
906 rmv_page_order(page);
908 expand(zone, page, order, current_order, area, migratetype);
917 * This array describes the order lists are fallen back to when
918 * the free lists for the desirable migrate type are depleted
920 static int fallbacks[MIGRATE_TYPES][4] = {
921 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
922 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
924 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
925 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
927 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
929 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
930 #ifdef CONFIG_MEMORY_ISOLATION
931 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
936 * Move the free pages in a range to the free lists of the requested type.
937 * Note that start_page and end_pages are not aligned on a pageblock
938 * boundary. If alignment is required, use move_freepages_block()
940 int move_freepages(struct zone *zone,
941 struct page *start_page, struct page *end_page,
948 #ifndef CONFIG_HOLES_IN_ZONE
950 * page_zone is not safe to call in this context when
951 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
952 * anyway as we check zone boundaries in move_freepages_block().
953 * Remove at a later date when no bug reports exist related to
954 * grouping pages by mobility
956 BUG_ON(page_zone(start_page) != page_zone(end_page));
959 for (page = start_page; page <= end_page;) {
960 /* Make sure we are not inadvertently changing nodes */
961 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
963 if (!pfn_valid_within(page_to_pfn(page))) {
968 if (!PageBuddy(page)) {
973 order = page_order(page);
974 list_move(&page->lru,
975 &zone->free_area[order].free_list[migratetype]);
976 set_freepage_migratetype(page, migratetype);
978 pages_moved += 1 << order;
984 int move_freepages_block(struct zone *zone, struct page *page,
987 unsigned long start_pfn, end_pfn;
988 struct page *start_page, *end_page;
990 start_pfn = page_to_pfn(page);
991 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
992 start_page = pfn_to_page(start_pfn);
993 end_page = start_page + pageblock_nr_pages - 1;
994 end_pfn = start_pfn + pageblock_nr_pages - 1;
996 /* Do not cross zone boundaries */
997 if (!zone_spans_pfn(zone, start_pfn))
999 if (!zone_spans_pfn(zone, end_pfn))
1002 return move_freepages(zone, start_page, end_page, migratetype);
1005 static void change_pageblock_range(struct page *pageblock_page,
1006 int start_order, int migratetype)
1008 int nr_pageblocks = 1 << (start_order - pageblock_order);
1010 while (nr_pageblocks--) {
1011 set_pageblock_migratetype(pageblock_page, migratetype);
1012 pageblock_page += pageblock_nr_pages;
1016 /* Remove an element from the buddy allocator from the fallback list */
1017 static inline struct page *
1018 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1020 struct free_area * area;
1025 /* Find the largest possible block of pages in the other list */
1026 for (current_order = MAX_ORDER-1; current_order >= order;
1029 migratetype = fallbacks[start_migratetype][i];
1031 /* MIGRATE_RESERVE handled later if necessary */
1032 if (migratetype == MIGRATE_RESERVE)
1035 area = &(zone->free_area[current_order]);
1036 if (list_empty(&area->free_list[migratetype]))
1039 page = list_entry(area->free_list[migratetype].next,
1044 * If breaking a large block of pages, move all free
1045 * pages to the preferred allocation list. If falling
1046 * back for a reclaimable kernel allocation, be more
1047 * aggressive about taking ownership of free pages
1049 * On the other hand, never change migration
1050 * type of MIGRATE_CMA pageblocks nor move CMA
1051 * pages on different free lists. We don't
1052 * want unmovable pages to be allocated from
1053 * MIGRATE_CMA areas.
1055 if (!is_migrate_cma(migratetype) &&
1056 (unlikely(current_order >= pageblock_order / 2) ||
1057 start_migratetype == MIGRATE_RECLAIMABLE ||
1058 page_group_by_mobility_disabled)) {
1060 pages = move_freepages_block(zone, page,
1063 /* Claim the whole block if over half of it is free */
1064 if (pages >= (1 << (pageblock_order-1)) ||
1065 page_group_by_mobility_disabled)
1066 set_pageblock_migratetype(page,
1069 migratetype = start_migratetype;
1072 /* Remove the page from the freelists */
1073 list_del(&page->lru);
1074 rmv_page_order(page);
1076 /* Take ownership for orders >= pageblock_order */
1077 if (current_order >= pageblock_order &&
1078 !is_migrate_cma(migratetype))
1079 change_pageblock_range(page, current_order,
1082 expand(zone, page, order, current_order, area,
1083 is_migrate_cma(migratetype)
1084 ? migratetype : start_migratetype);
1086 trace_mm_page_alloc_extfrag(page, order, current_order,
1087 start_migratetype, migratetype);
1097 * Do the hard work of removing an element from the buddy allocator.
1098 * Call me with the zone->lock already held.
1100 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1106 page = __rmqueue_smallest(zone, order, migratetype);
1108 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1109 page = __rmqueue_fallback(zone, order, migratetype);
1112 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1113 * is used because __rmqueue_smallest is an inline function
1114 * and we want just one call site
1117 migratetype = MIGRATE_RESERVE;
1122 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1127 * Obtain a specified number of elements from the buddy allocator, all under
1128 * a single hold of the lock, for efficiency. Add them to the supplied list.
1129 * Returns the number of new pages which were placed at *list.
1131 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1132 unsigned long count, struct list_head *list,
1133 int migratetype, int cold)
1135 int mt = migratetype, i;
1137 spin_lock(&zone->lock);
1138 for (i = 0; i < count; ++i) {
1139 struct page *page = __rmqueue(zone, order, migratetype);
1140 if (unlikely(page == NULL))
1144 * Split buddy pages returned by expand() are received here
1145 * in physical page order. The page is added to the callers and
1146 * list and the list head then moves forward. From the callers
1147 * perspective, the linked list is ordered by page number in
1148 * some conditions. This is useful for IO devices that can
1149 * merge IO requests if the physical pages are ordered
1152 if (likely(cold == 0))
1153 list_add(&page->lru, list);
1155 list_add_tail(&page->lru, list);
1156 if (IS_ENABLED(CONFIG_CMA)) {
1157 mt = get_pageblock_migratetype(page);
1158 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1161 set_freepage_migratetype(page, mt);
1163 if (is_migrate_cma(mt))
1164 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1167 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1168 spin_unlock(&zone->lock);
1174 * Called from the vmstat counter updater to drain pagesets of this
1175 * currently executing processor on remote nodes after they have
1178 * Note that this function must be called with the thread pinned to
1179 * a single processor.
1181 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1183 unsigned long flags;
1186 local_irq_save(flags);
1187 if (pcp->count >= pcp->batch)
1188 to_drain = pcp->batch;
1190 to_drain = pcp->count;
1192 free_pcppages_bulk(zone, to_drain, pcp);
1193 pcp->count -= to_drain;
1195 local_irq_restore(flags);
1200 * Drain pages of the indicated processor.
1202 * The processor must either be the current processor and the
1203 * thread pinned to the current processor or a processor that
1206 static void drain_pages(unsigned int cpu)
1208 unsigned long flags;
1211 for_each_populated_zone(zone) {
1212 struct per_cpu_pageset *pset;
1213 struct per_cpu_pages *pcp;
1215 local_irq_save(flags);
1216 pset = per_cpu_ptr(zone->pageset, cpu);
1220 free_pcppages_bulk(zone, pcp->count, pcp);
1223 local_irq_restore(flags);
1228 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1230 void drain_local_pages(void *arg)
1232 drain_pages(smp_processor_id());
1236 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1238 * Note that this code is protected against sending an IPI to an offline
1239 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1240 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1241 * nothing keeps CPUs from showing up after we populated the cpumask and
1242 * before the call to on_each_cpu_mask().
1244 void drain_all_pages(void)
1247 struct per_cpu_pageset *pcp;
1251 * Allocate in the BSS so we wont require allocation in
1252 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1254 static cpumask_t cpus_with_pcps;
1257 * We don't care about racing with CPU hotplug event
1258 * as offline notification will cause the notified
1259 * cpu to drain that CPU pcps and on_each_cpu_mask
1260 * disables preemption as part of its processing
1262 for_each_online_cpu(cpu) {
1263 bool has_pcps = false;
1264 for_each_populated_zone(zone) {
1265 pcp = per_cpu_ptr(zone->pageset, cpu);
1266 if (pcp->pcp.count) {
1272 cpumask_set_cpu(cpu, &cpus_with_pcps);
1274 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1276 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1279 #ifdef CONFIG_HIBERNATION
1281 void mark_free_pages(struct zone *zone)
1283 unsigned long pfn, max_zone_pfn;
1284 unsigned long flags;
1286 struct list_head *curr;
1288 if (!zone->spanned_pages)
1291 spin_lock_irqsave(&zone->lock, flags);
1293 max_zone_pfn = zone_end_pfn(zone);
1294 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1295 if (pfn_valid(pfn)) {
1296 struct page *page = pfn_to_page(pfn);
1298 if (!swsusp_page_is_forbidden(page))
1299 swsusp_unset_page_free(page);
1302 for_each_migratetype_order(order, t) {
1303 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1306 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1307 for (i = 0; i < (1UL << order); i++)
1308 swsusp_set_page_free(pfn_to_page(pfn + i));
1311 spin_unlock_irqrestore(&zone->lock, flags);
1313 #endif /* CONFIG_PM */
1316 * Free a 0-order page
1317 * cold == 1 ? free a cold page : free a hot page
1319 void free_hot_cold_page(struct page *page, int cold)
1321 struct zone *zone = page_zone(page);
1322 struct per_cpu_pages *pcp;
1323 unsigned long flags;
1326 if (!free_pages_prepare(page, 0))
1329 migratetype = get_pageblock_migratetype(page);
1330 set_freepage_migratetype(page, migratetype);
1331 local_irq_save(flags);
1332 __count_vm_event(PGFREE);
1335 * We only track unmovable, reclaimable and movable on pcp lists.
1336 * Free ISOLATE pages back to the allocator because they are being
1337 * offlined but treat RESERVE as movable pages so we can get those
1338 * areas back if necessary. Otherwise, we may have to free
1339 * excessively into the page allocator
1341 if (migratetype >= MIGRATE_PCPTYPES) {
1342 if (unlikely(is_migrate_isolate(migratetype))) {
1343 free_one_page(zone, page, 0, migratetype);
1346 migratetype = MIGRATE_MOVABLE;
1349 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1351 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1353 list_add(&page->lru, &pcp->lists[migratetype]);
1355 if (pcp->count >= pcp->high) {
1356 free_pcppages_bulk(zone, pcp->batch, pcp);
1357 pcp->count -= pcp->batch;
1361 local_irq_restore(flags);
1365 * Free a list of 0-order pages
1367 void free_hot_cold_page_list(struct list_head *list, int cold)
1369 struct page *page, *next;
1371 list_for_each_entry_safe(page, next, list, lru) {
1372 trace_mm_page_free_batched(page, cold);
1373 free_hot_cold_page(page, cold);
1378 * split_page takes a non-compound higher-order page, and splits it into
1379 * n (1<<order) sub-pages: page[0..n]
1380 * Each sub-page must be freed individually.
1382 * Note: this is probably too low level an operation for use in drivers.
1383 * Please consult with lkml before using this in your driver.
1385 void split_page(struct page *page, unsigned int order)
1389 VM_BUG_ON(PageCompound(page));
1390 VM_BUG_ON(!page_count(page));
1392 #ifdef CONFIG_KMEMCHECK
1394 * Split shadow pages too, because free(page[0]) would
1395 * otherwise free the whole shadow.
1397 if (kmemcheck_page_is_tracked(page))
1398 split_page(virt_to_page(page[0].shadow), order);
1401 for (i = 1; i < (1 << order); i++)
1402 set_page_refcounted(page + i);
1404 EXPORT_SYMBOL_GPL(split_page);
1406 static int __isolate_free_page(struct page *page, unsigned int order)
1408 unsigned long watermark;
1412 BUG_ON(!PageBuddy(page));
1414 zone = page_zone(page);
1415 mt = get_pageblock_migratetype(page);
1417 if (!is_migrate_isolate(mt)) {
1418 /* Obey watermarks as if the page was being allocated */
1419 watermark = low_wmark_pages(zone) + (1 << order);
1420 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1423 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1426 /* Remove page from free list */
1427 list_del(&page->lru);
1428 zone->free_area[order].nr_free--;
1429 rmv_page_order(page);
1431 /* Set the pageblock if the isolated page is at least a pageblock */
1432 if (order >= pageblock_order - 1) {
1433 struct page *endpage = page + (1 << order) - 1;
1434 for (; page < endpage; page += pageblock_nr_pages) {
1435 int mt = get_pageblock_migratetype(page);
1436 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1437 set_pageblock_migratetype(page,
1442 return 1UL << order;
1446 * Similar to split_page except the page is already free. As this is only
1447 * being used for migration, the migratetype of the block also changes.
1448 * As this is called with interrupts disabled, the caller is responsible
1449 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1452 * Note: this is probably too low level an operation for use in drivers.
1453 * Please consult with lkml before using this in your driver.
1455 int split_free_page(struct page *page)
1460 order = page_order(page);
1462 nr_pages = __isolate_free_page(page, order);
1466 /* Split into individual pages */
1467 set_page_refcounted(page);
1468 split_page(page, order);
1473 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1474 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1478 struct page *buffered_rmqueue(struct zone *preferred_zone,
1479 struct zone *zone, int order, gfp_t gfp_flags,
1482 unsigned long flags;
1484 int cold = !!(gfp_flags & __GFP_COLD);
1487 if (likely(order == 0)) {
1488 struct per_cpu_pages *pcp;
1489 struct list_head *list;
1491 local_irq_save(flags);
1492 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1493 list = &pcp->lists[migratetype];
1494 if (list_empty(list)) {
1495 pcp->count += rmqueue_bulk(zone, 0,
1498 if (unlikely(list_empty(list)))
1503 page = list_entry(list->prev, struct page, lru);
1505 page = list_entry(list->next, struct page, lru);
1507 list_del(&page->lru);
1510 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1512 * __GFP_NOFAIL is not to be used in new code.
1514 * All __GFP_NOFAIL callers should be fixed so that they
1515 * properly detect and handle allocation failures.
1517 * We most definitely don't want callers attempting to
1518 * allocate greater than order-1 page units with
1521 WARN_ON_ONCE(order > 1);
1523 spin_lock_irqsave(&zone->lock, flags);
1524 page = __rmqueue(zone, order, migratetype);
1525 spin_unlock(&zone->lock);
1528 __mod_zone_freepage_state(zone, -(1 << order),
1529 get_pageblock_migratetype(page));
1532 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1533 zone_statistics(preferred_zone, zone, gfp_flags);
1534 local_irq_restore(flags);
1536 VM_BUG_ON(bad_range(zone, page));
1537 if (prep_new_page(page, order, gfp_flags))
1542 local_irq_restore(flags);
1546 #ifdef CONFIG_FAIL_PAGE_ALLOC
1549 struct fault_attr attr;
1551 u32 ignore_gfp_highmem;
1552 u32 ignore_gfp_wait;
1554 } fail_page_alloc = {
1555 .attr = FAULT_ATTR_INITIALIZER,
1556 .ignore_gfp_wait = 1,
1557 .ignore_gfp_highmem = 1,
1561 static int __init setup_fail_page_alloc(char *str)
1563 return setup_fault_attr(&fail_page_alloc.attr, str);
1565 __setup("fail_page_alloc=", setup_fail_page_alloc);
1567 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1569 if (order < fail_page_alloc.min_order)
1571 if (gfp_mask & __GFP_NOFAIL)
1573 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1575 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1578 return should_fail(&fail_page_alloc.attr, 1 << order);
1581 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1583 static int __init fail_page_alloc_debugfs(void)
1585 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1588 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1589 &fail_page_alloc.attr);
1591 return PTR_ERR(dir);
1593 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1594 &fail_page_alloc.ignore_gfp_wait))
1596 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1597 &fail_page_alloc.ignore_gfp_highmem))
1599 if (!debugfs_create_u32("min-order", mode, dir,
1600 &fail_page_alloc.min_order))
1605 debugfs_remove_recursive(dir);
1610 late_initcall(fail_page_alloc_debugfs);
1612 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1614 #else /* CONFIG_FAIL_PAGE_ALLOC */
1616 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1621 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1624 * Return true if free pages are above 'mark'. This takes into account the order
1625 * of the allocation.
1627 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1628 int classzone_idx, int alloc_flags, long free_pages)
1630 /* free_pages my go negative - that's OK */
1632 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1636 free_pages -= (1 << order) - 1;
1637 if (alloc_flags & ALLOC_HIGH)
1639 if (alloc_flags & ALLOC_HARDER)
1642 /* If allocation can't use CMA areas don't use free CMA pages */
1643 if (!(alloc_flags & ALLOC_CMA))
1644 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1647 if (free_pages - free_cma <= min + lowmem_reserve)
1649 for (o = 0; o < order; o++) {
1650 /* At the next order, this order's pages become unavailable */
1651 free_pages -= z->free_area[o].nr_free << o;
1653 /* Require fewer higher order pages to be free */
1656 if (free_pages <= min)
1662 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1663 int classzone_idx, int alloc_flags)
1665 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1666 zone_page_state(z, NR_FREE_PAGES));
1669 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1670 int classzone_idx, int alloc_flags)
1672 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1674 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1675 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1677 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1683 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1684 * skip over zones that are not allowed by the cpuset, or that have
1685 * been recently (in last second) found to be nearly full. See further
1686 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1687 * that have to skip over a lot of full or unallowed zones.
1689 * If the zonelist cache is present in the passed in zonelist, then
1690 * returns a pointer to the allowed node mask (either the current
1691 * tasks mems_allowed, or node_states[N_MEMORY].)
1693 * If the zonelist cache is not available for this zonelist, does
1694 * nothing and returns NULL.
1696 * If the fullzones BITMAP in the zonelist cache is stale (more than
1697 * a second since last zap'd) then we zap it out (clear its bits.)
1699 * We hold off even calling zlc_setup, until after we've checked the
1700 * first zone in the zonelist, on the theory that most allocations will
1701 * be satisfied from that first zone, so best to examine that zone as
1702 * quickly as we can.
1704 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1706 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1707 nodemask_t *allowednodes; /* zonelist_cache approximation */
1709 zlc = zonelist->zlcache_ptr;
1713 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1714 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1715 zlc->last_full_zap = jiffies;
1718 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1719 &cpuset_current_mems_allowed :
1720 &node_states[N_MEMORY];
1721 return allowednodes;
1725 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1726 * if it is worth looking at further for free memory:
1727 * 1) Check that the zone isn't thought to be full (doesn't have its
1728 * bit set in the zonelist_cache fullzones BITMAP).
1729 * 2) Check that the zones node (obtained from the zonelist_cache
1730 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1731 * Return true (non-zero) if zone is worth looking at further, or
1732 * else return false (zero) if it is not.
1734 * This check -ignores- the distinction between various watermarks,
1735 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1736 * found to be full for any variation of these watermarks, it will
1737 * be considered full for up to one second by all requests, unless
1738 * we are so low on memory on all allowed nodes that we are forced
1739 * into the second scan of the zonelist.
1741 * In the second scan we ignore this zonelist cache and exactly
1742 * apply the watermarks to all zones, even it is slower to do so.
1743 * We are low on memory in the second scan, and should leave no stone
1744 * unturned looking for a free page.
1746 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1747 nodemask_t *allowednodes)
1749 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1750 int i; /* index of *z in zonelist zones */
1751 int n; /* node that zone *z is on */
1753 zlc = zonelist->zlcache_ptr;
1757 i = z - zonelist->_zonerefs;
1760 /* This zone is worth trying if it is allowed but not full */
1761 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1765 * Given 'z' scanning a zonelist, set the corresponding bit in
1766 * zlc->fullzones, so that subsequent attempts to allocate a page
1767 * from that zone don't waste time re-examining it.
1769 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1771 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1772 int i; /* index of *z in zonelist zones */
1774 zlc = zonelist->zlcache_ptr;
1778 i = z - zonelist->_zonerefs;
1780 set_bit(i, zlc->fullzones);
1784 * clear all zones full, called after direct reclaim makes progress so that
1785 * a zone that was recently full is not skipped over for up to a second
1787 static void zlc_clear_zones_full(struct zonelist *zonelist)
1789 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1791 zlc = zonelist->zlcache_ptr;
1795 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1798 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1800 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1803 static void __paginginit init_zone_allows_reclaim(int nid)
1807 for_each_online_node(i)
1808 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1809 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1811 zone_reclaim_mode = 1;
1814 #else /* CONFIG_NUMA */
1816 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1821 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1822 nodemask_t *allowednodes)
1827 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1831 static void zlc_clear_zones_full(struct zonelist *zonelist)
1835 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1840 static inline void init_zone_allows_reclaim(int nid)
1843 #endif /* CONFIG_NUMA */
1846 * get_page_from_freelist goes through the zonelist trying to allocate
1849 static struct page *
1850 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1851 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1852 struct zone *preferred_zone, int migratetype)
1855 struct page *page = NULL;
1858 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1859 int zlc_active = 0; /* set if using zonelist_cache */
1860 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1862 classzone_idx = zone_idx(preferred_zone);
1865 * Scan zonelist, looking for a zone with enough free.
1866 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1868 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1869 high_zoneidx, nodemask) {
1870 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1871 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1873 if ((alloc_flags & ALLOC_CPUSET) &&
1874 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1877 * When allocating a page cache page for writing, we
1878 * want to get it from a zone that is within its dirty
1879 * limit, such that no single zone holds more than its
1880 * proportional share of globally allowed dirty pages.
1881 * The dirty limits take into account the zone's
1882 * lowmem reserves and high watermark so that kswapd
1883 * should be able to balance it without having to
1884 * write pages from its LRU list.
1886 * This may look like it could increase pressure on
1887 * lower zones by failing allocations in higher zones
1888 * before they are full. But the pages that do spill
1889 * over are limited as the lower zones are protected
1890 * by this very same mechanism. It should not become
1891 * a practical burden to them.
1893 * XXX: For now, allow allocations to potentially
1894 * exceed the per-zone dirty limit in the slowpath
1895 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1896 * which is important when on a NUMA setup the allowed
1897 * zones are together not big enough to reach the
1898 * global limit. The proper fix for these situations
1899 * will require awareness of zones in the
1900 * dirty-throttling and the flusher threads.
1902 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1903 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1904 goto this_zone_full;
1906 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1907 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1911 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1912 if (zone_watermark_ok(zone, order, mark,
1913 classzone_idx, alloc_flags))
1916 if (IS_ENABLED(CONFIG_NUMA) &&
1917 !did_zlc_setup && nr_online_nodes > 1) {
1919 * we do zlc_setup if there are multiple nodes
1920 * and before considering the first zone allowed
1923 allowednodes = zlc_setup(zonelist, alloc_flags);
1928 if (zone_reclaim_mode == 0 ||
1929 !zone_allows_reclaim(preferred_zone, zone))
1930 goto this_zone_full;
1933 * As we may have just activated ZLC, check if the first
1934 * eligible zone has failed zone_reclaim recently.
1936 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1937 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1940 ret = zone_reclaim(zone, gfp_mask, order);
1942 case ZONE_RECLAIM_NOSCAN:
1945 case ZONE_RECLAIM_FULL:
1946 /* scanned but unreclaimable */
1949 /* did we reclaim enough */
1950 if (zone_watermark_ok(zone, order, mark,
1951 classzone_idx, alloc_flags))
1955 * Failed to reclaim enough to meet watermark.
1956 * Only mark the zone full if checking the min
1957 * watermark or if we failed to reclaim just
1958 * 1<<order pages or else the page allocator
1959 * fastpath will prematurely mark zones full
1960 * when the watermark is between the low and
1963 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
1964 ret == ZONE_RECLAIM_SOME)
1965 goto this_zone_full;
1972 page = buffered_rmqueue(preferred_zone, zone, order,
1973 gfp_mask, migratetype);
1977 if (IS_ENABLED(CONFIG_NUMA))
1978 zlc_mark_zone_full(zonelist, z);
1981 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1982 /* Disable zlc cache for second zonelist scan */
1989 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1990 * necessary to allocate the page. The expectation is
1991 * that the caller is taking steps that will free more
1992 * memory. The caller should avoid the page being used
1993 * for !PFMEMALLOC purposes.
1995 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2001 * Large machines with many possible nodes should not always dump per-node
2002 * meminfo in irq context.
2004 static inline bool should_suppress_show_mem(void)
2009 ret = in_interrupt();
2014 static DEFINE_RATELIMIT_STATE(nopage_rs,
2015 DEFAULT_RATELIMIT_INTERVAL,
2016 DEFAULT_RATELIMIT_BURST);
2018 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2020 unsigned int filter = SHOW_MEM_FILTER_NODES;
2022 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2023 debug_guardpage_minorder() > 0)
2027 * Walking all memory to count page types is very expensive and should
2028 * be inhibited in non-blockable contexts.
2030 if (!(gfp_mask & __GFP_WAIT))
2031 filter |= SHOW_MEM_FILTER_PAGE_COUNT;
2034 * This documents exceptions given to allocations in certain
2035 * contexts that are allowed to allocate outside current's set
2038 if (!(gfp_mask & __GFP_NOMEMALLOC))
2039 if (test_thread_flag(TIF_MEMDIE) ||
2040 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2041 filter &= ~SHOW_MEM_FILTER_NODES;
2042 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2043 filter &= ~SHOW_MEM_FILTER_NODES;
2046 struct va_format vaf;
2049 va_start(args, fmt);
2054 pr_warn("%pV", &vaf);
2059 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2060 current->comm, order, gfp_mask);
2063 if (!should_suppress_show_mem())
2068 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2069 unsigned long did_some_progress,
2070 unsigned long pages_reclaimed)
2072 /* Do not loop if specifically requested */
2073 if (gfp_mask & __GFP_NORETRY)
2076 /* Always retry if specifically requested */
2077 if (gfp_mask & __GFP_NOFAIL)
2081 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2082 * making forward progress without invoking OOM. Suspend also disables
2083 * storage devices so kswapd will not help. Bail if we are suspending.
2085 if (!did_some_progress && pm_suspended_storage())
2089 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2090 * means __GFP_NOFAIL, but that may not be true in other
2093 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2097 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2098 * specified, then we retry until we no longer reclaim any pages
2099 * (above), or we've reclaimed an order of pages at least as
2100 * large as the allocation's order. In both cases, if the
2101 * allocation still fails, we stop retrying.
2103 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2109 static inline struct page *
2110 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2111 struct zonelist *zonelist, enum zone_type high_zoneidx,
2112 nodemask_t *nodemask, struct zone *preferred_zone,
2117 /* Acquire the OOM killer lock for the zones in zonelist */
2118 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2119 schedule_timeout_uninterruptible(1);
2124 * Go through the zonelist yet one more time, keep very high watermark
2125 * here, this is only to catch a parallel oom killing, we must fail if
2126 * we're still under heavy pressure.
2128 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2129 order, zonelist, high_zoneidx,
2130 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2131 preferred_zone, migratetype);
2135 if (!(gfp_mask & __GFP_NOFAIL)) {
2136 /* The OOM killer will not help higher order allocs */
2137 if (order > PAGE_ALLOC_COSTLY_ORDER)
2139 /* The OOM killer does not needlessly kill tasks for lowmem */
2140 if (high_zoneidx < ZONE_NORMAL)
2143 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2144 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2145 * The caller should handle page allocation failure by itself if
2146 * it specifies __GFP_THISNODE.
2147 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2149 if (gfp_mask & __GFP_THISNODE)
2152 /* Exhausted what can be done so it's blamo time */
2153 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2156 clear_zonelist_oom(zonelist, gfp_mask);
2160 #ifdef CONFIG_COMPACTION
2161 /* Try memory compaction for high-order allocations before reclaim */
2162 static struct page *
2163 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2164 struct zonelist *zonelist, enum zone_type high_zoneidx,
2165 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2166 int migratetype, bool sync_migration,
2167 bool *contended_compaction, bool *deferred_compaction,
2168 unsigned long *did_some_progress)
2173 if (compaction_deferred(preferred_zone, order)) {
2174 *deferred_compaction = true;
2178 current->flags |= PF_MEMALLOC;
2179 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2180 nodemask, sync_migration,
2181 contended_compaction);
2182 current->flags &= ~PF_MEMALLOC;
2184 if (*did_some_progress != COMPACT_SKIPPED) {
2187 /* Page migration frees to the PCP lists but we want merging */
2188 drain_pages(get_cpu());
2191 page = get_page_from_freelist(gfp_mask, nodemask,
2192 order, zonelist, high_zoneidx,
2193 alloc_flags & ~ALLOC_NO_WATERMARKS,
2194 preferred_zone, migratetype);
2196 preferred_zone->compact_blockskip_flush = false;
2197 preferred_zone->compact_considered = 0;
2198 preferred_zone->compact_defer_shift = 0;
2199 if (order >= preferred_zone->compact_order_failed)
2200 preferred_zone->compact_order_failed = order + 1;
2201 count_vm_event(COMPACTSUCCESS);
2206 * It's bad if compaction run occurs and fails.
2207 * The most likely reason is that pages exist,
2208 * but not enough to satisfy watermarks.
2210 count_vm_event(COMPACTFAIL);
2213 * As async compaction considers a subset of pageblocks, only
2214 * defer if the failure was a sync compaction failure.
2217 defer_compaction(preferred_zone, order);
2225 static inline struct page *
2226 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2227 struct zonelist *zonelist, enum zone_type high_zoneidx,
2228 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2229 int migratetype, bool sync_migration,
2230 bool *contended_compaction, bool *deferred_compaction,
2231 unsigned long *did_some_progress)
2235 #endif /* CONFIG_COMPACTION */
2237 /* Perform direct synchronous page reclaim */
2239 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2240 nodemask_t *nodemask)
2242 struct reclaim_state reclaim_state;
2247 /* We now go into synchronous reclaim */
2248 cpuset_memory_pressure_bump();
2249 current->flags |= PF_MEMALLOC;
2250 lockdep_set_current_reclaim_state(gfp_mask);
2251 reclaim_state.reclaimed_slab = 0;
2252 current->reclaim_state = &reclaim_state;
2254 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2256 current->reclaim_state = NULL;
2257 lockdep_clear_current_reclaim_state();
2258 current->flags &= ~PF_MEMALLOC;
2265 /* The really slow allocator path where we enter direct reclaim */
2266 static inline struct page *
2267 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2268 struct zonelist *zonelist, enum zone_type high_zoneidx,
2269 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2270 int migratetype, unsigned long *did_some_progress)
2272 struct page *page = NULL;
2273 bool drained = false;
2275 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2277 if (unlikely(!(*did_some_progress)))
2280 /* After successful reclaim, reconsider all zones for allocation */
2281 if (IS_ENABLED(CONFIG_NUMA))
2282 zlc_clear_zones_full(zonelist);
2285 page = get_page_from_freelist(gfp_mask, nodemask, order,
2286 zonelist, high_zoneidx,
2287 alloc_flags & ~ALLOC_NO_WATERMARKS,
2288 preferred_zone, migratetype);
2291 * If an allocation failed after direct reclaim, it could be because
2292 * pages are pinned on the per-cpu lists. Drain them and try again
2294 if (!page && !drained) {
2304 * This is called in the allocator slow-path if the allocation request is of
2305 * sufficient urgency to ignore watermarks and take other desperate measures
2307 static inline struct page *
2308 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2309 struct zonelist *zonelist, enum zone_type high_zoneidx,
2310 nodemask_t *nodemask, struct zone *preferred_zone,
2316 page = get_page_from_freelist(gfp_mask, nodemask, order,
2317 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2318 preferred_zone, migratetype);
2320 if (!page && gfp_mask & __GFP_NOFAIL)
2321 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2322 } while (!page && (gfp_mask & __GFP_NOFAIL));
2328 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2329 enum zone_type high_zoneidx,
2330 enum zone_type classzone_idx)
2335 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2336 wakeup_kswapd(zone, order, classzone_idx);
2340 gfp_to_alloc_flags(gfp_t gfp_mask)
2342 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2343 const gfp_t wait = gfp_mask & __GFP_WAIT;
2345 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2346 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2349 * The caller may dip into page reserves a bit more if the caller
2350 * cannot run direct reclaim, or if the caller has realtime scheduling
2351 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2352 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2354 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2358 * Not worth trying to allocate harder for
2359 * __GFP_NOMEMALLOC even if it can't schedule.
2361 if (!(gfp_mask & __GFP_NOMEMALLOC))
2362 alloc_flags |= ALLOC_HARDER;
2364 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2365 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2367 alloc_flags &= ~ALLOC_CPUSET;
2368 } else if (unlikely(rt_task(current)) && !in_interrupt())
2369 alloc_flags |= ALLOC_HARDER;
2371 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2372 if (gfp_mask & __GFP_MEMALLOC)
2373 alloc_flags |= ALLOC_NO_WATERMARKS;
2374 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2375 alloc_flags |= ALLOC_NO_WATERMARKS;
2376 else if (!in_interrupt() &&
2377 ((current->flags & PF_MEMALLOC) ||
2378 unlikely(test_thread_flag(TIF_MEMDIE))))
2379 alloc_flags |= ALLOC_NO_WATERMARKS;
2382 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2383 alloc_flags |= ALLOC_CMA;
2388 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2390 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2393 static inline struct page *
2394 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2395 struct zonelist *zonelist, enum zone_type high_zoneidx,
2396 nodemask_t *nodemask, struct zone *preferred_zone,
2399 const gfp_t wait = gfp_mask & __GFP_WAIT;
2400 struct page *page = NULL;
2402 unsigned long pages_reclaimed = 0;
2403 unsigned long did_some_progress;
2404 bool sync_migration = false;
2405 bool deferred_compaction = false;
2406 bool contended_compaction = false;
2409 * In the slowpath, we sanity check order to avoid ever trying to
2410 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2411 * be using allocators in order of preference for an area that is
2414 if (order >= MAX_ORDER) {
2415 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2420 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2421 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2422 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2423 * using a larger set of nodes after it has established that the
2424 * allowed per node queues are empty and that nodes are
2427 if (IS_ENABLED(CONFIG_NUMA) &&
2428 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2432 if (!(gfp_mask & __GFP_NO_KSWAPD))
2433 wake_all_kswapd(order, zonelist, high_zoneidx,
2434 zone_idx(preferred_zone));
2437 * OK, we're below the kswapd watermark and have kicked background
2438 * reclaim. Now things get more complex, so set up alloc_flags according
2439 * to how we want to proceed.
2441 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2444 * Find the true preferred zone if the allocation is unconstrained by
2447 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2448 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2452 /* This is the last chance, in general, before the goto nopage. */
2453 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2454 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2455 preferred_zone, migratetype);
2459 /* Allocate without watermarks if the context allows */
2460 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2462 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2463 * the allocation is high priority and these type of
2464 * allocations are system rather than user orientated
2466 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2468 page = __alloc_pages_high_priority(gfp_mask, order,
2469 zonelist, high_zoneidx, nodemask,
2470 preferred_zone, migratetype);
2476 /* Atomic allocations - we can't balance anything */
2480 /* Avoid recursion of direct reclaim */
2481 if (current->flags & PF_MEMALLOC)
2484 /* Avoid allocations with no watermarks from looping endlessly */
2485 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2489 * Try direct compaction. The first pass is asynchronous. Subsequent
2490 * attempts after direct reclaim are synchronous
2492 page = __alloc_pages_direct_compact(gfp_mask, order,
2493 zonelist, high_zoneidx,
2495 alloc_flags, preferred_zone,
2496 migratetype, sync_migration,
2497 &contended_compaction,
2498 &deferred_compaction,
2499 &did_some_progress);
2502 sync_migration = true;
2505 * If compaction is deferred for high-order allocations, it is because
2506 * sync compaction recently failed. In this is the case and the caller
2507 * requested a movable allocation that does not heavily disrupt the
2508 * system then fail the allocation instead of entering direct reclaim.
2510 if ((deferred_compaction || contended_compaction) &&
2511 (gfp_mask & __GFP_NO_KSWAPD))
2514 /* Try direct reclaim and then allocating */
2515 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2516 zonelist, high_zoneidx,
2518 alloc_flags, preferred_zone,
2519 migratetype, &did_some_progress);
2524 * If we failed to make any progress reclaiming, then we are
2525 * running out of options and have to consider going OOM
2527 if (!did_some_progress) {
2528 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2529 if (oom_killer_disabled)
2531 /* Coredumps can quickly deplete all memory reserves */
2532 if ((current->flags & PF_DUMPCORE) &&
2533 !(gfp_mask & __GFP_NOFAIL))
2535 page = __alloc_pages_may_oom(gfp_mask, order,
2536 zonelist, high_zoneidx,
2537 nodemask, preferred_zone,
2542 if (!(gfp_mask & __GFP_NOFAIL)) {
2544 * The oom killer is not called for high-order
2545 * allocations that may fail, so if no progress
2546 * is being made, there are no other options and
2547 * retrying is unlikely to help.
2549 if (order > PAGE_ALLOC_COSTLY_ORDER)
2552 * The oom killer is not called for lowmem
2553 * allocations to prevent needlessly killing
2556 if (high_zoneidx < ZONE_NORMAL)
2564 /* Check if we should retry the allocation */
2565 pages_reclaimed += did_some_progress;
2566 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2568 /* Wait for some write requests to complete then retry */
2569 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2573 * High-order allocations do not necessarily loop after
2574 * direct reclaim and reclaim/compaction depends on compaction
2575 * being called after reclaim so call directly if necessary
2577 page = __alloc_pages_direct_compact(gfp_mask, order,
2578 zonelist, high_zoneidx,
2580 alloc_flags, preferred_zone,
2581 migratetype, sync_migration,
2582 &contended_compaction,
2583 &deferred_compaction,
2584 &did_some_progress);
2590 warn_alloc_failed(gfp_mask, order, NULL);
2593 if (kmemcheck_enabled)
2594 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2600 * This is the 'heart' of the zoned buddy allocator.
2603 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2604 struct zonelist *zonelist, nodemask_t *nodemask)
2606 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2607 struct zone *preferred_zone;
2608 struct page *page = NULL;
2609 int migratetype = allocflags_to_migratetype(gfp_mask);
2610 unsigned int cpuset_mems_cookie;
2611 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2612 struct mem_cgroup *memcg = NULL;
2614 gfp_mask &= gfp_allowed_mask;
2616 lockdep_trace_alloc(gfp_mask);
2618 might_sleep_if(gfp_mask & __GFP_WAIT);
2620 if (should_fail_alloc_page(gfp_mask, order))
2624 * Check the zones suitable for the gfp_mask contain at least one
2625 * valid zone. It's possible to have an empty zonelist as a result
2626 * of GFP_THISNODE and a memoryless node
2628 if (unlikely(!zonelist->_zonerefs->zone))
2632 * Will only have any effect when __GFP_KMEMCG is set. This is
2633 * verified in the (always inline) callee
2635 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2639 cpuset_mems_cookie = get_mems_allowed();
2641 /* The preferred zone is used for statistics later */
2642 first_zones_zonelist(zonelist, high_zoneidx,
2643 nodemask ? : &cpuset_current_mems_allowed,
2645 if (!preferred_zone)
2649 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2650 alloc_flags |= ALLOC_CMA;
2652 /* First allocation attempt */
2653 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2654 zonelist, high_zoneidx, alloc_flags,
2655 preferred_zone, migratetype);
2656 if (unlikely(!page)) {
2658 * Runtime PM, block IO and its error handling path
2659 * can deadlock because I/O on the device might not
2662 gfp_mask = memalloc_noio_flags(gfp_mask);
2663 page = __alloc_pages_slowpath(gfp_mask, order,
2664 zonelist, high_zoneidx, nodemask,
2665 preferred_zone, migratetype);
2668 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2672 * When updating a task's mems_allowed, it is possible to race with
2673 * parallel threads in such a way that an allocation can fail while
2674 * the mask is being updated. If a page allocation is about to fail,
2675 * check if the cpuset changed during allocation and if so, retry.
2677 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2680 memcg_kmem_commit_charge(page, memcg, order);
2684 EXPORT_SYMBOL(__alloc_pages_nodemask);
2687 * Common helper functions.
2689 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2694 * __get_free_pages() returns a 32-bit address, which cannot represent
2697 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2699 page = alloc_pages(gfp_mask, order);
2702 return (unsigned long) page_address(page);
2704 EXPORT_SYMBOL(__get_free_pages);
2706 unsigned long get_zeroed_page(gfp_t gfp_mask)
2708 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2710 EXPORT_SYMBOL(get_zeroed_page);
2712 void __free_pages(struct page *page, unsigned int order)
2714 if (put_page_testzero(page)) {
2716 free_hot_cold_page(page, 0);
2718 __free_pages_ok(page, order);
2722 EXPORT_SYMBOL(__free_pages);
2724 void free_pages(unsigned long addr, unsigned int order)
2727 VM_BUG_ON(!virt_addr_valid((void *)addr));
2728 __free_pages(virt_to_page((void *)addr), order);
2732 EXPORT_SYMBOL(free_pages);
2735 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2736 * pages allocated with __GFP_KMEMCG.
2738 * Those pages are accounted to a particular memcg, embedded in the
2739 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2740 * for that information only to find out that it is NULL for users who have no
2741 * interest in that whatsoever, we provide these functions.
2743 * The caller knows better which flags it relies on.
2745 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2747 memcg_kmem_uncharge_pages(page, order);
2748 __free_pages(page, order);
2751 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2754 VM_BUG_ON(!virt_addr_valid((void *)addr));
2755 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2759 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2762 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2763 unsigned long used = addr + PAGE_ALIGN(size);
2765 split_page(virt_to_page((void *)addr), order);
2766 while (used < alloc_end) {
2771 return (void *)addr;
2775 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2776 * @size: the number of bytes to allocate
2777 * @gfp_mask: GFP flags for the allocation
2779 * This function is similar to alloc_pages(), except that it allocates the
2780 * minimum number of pages to satisfy the request. alloc_pages() can only
2781 * allocate memory in power-of-two pages.
2783 * This function is also limited by MAX_ORDER.
2785 * Memory allocated by this function must be released by free_pages_exact().
2787 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2789 unsigned int order = get_order(size);
2792 addr = __get_free_pages(gfp_mask, order);
2793 return make_alloc_exact(addr, order, size);
2795 EXPORT_SYMBOL(alloc_pages_exact);
2798 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2800 * @nid: the preferred node ID where memory should be allocated
2801 * @size: the number of bytes to allocate
2802 * @gfp_mask: GFP flags for the allocation
2804 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2806 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2809 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2811 unsigned order = get_order(size);
2812 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2815 return make_alloc_exact((unsigned long)page_address(p), order, size);
2817 EXPORT_SYMBOL(alloc_pages_exact_nid);
2820 * free_pages_exact - release memory allocated via alloc_pages_exact()
2821 * @virt: the value returned by alloc_pages_exact.
2822 * @size: size of allocation, same value as passed to alloc_pages_exact().
2824 * Release the memory allocated by a previous call to alloc_pages_exact.
2826 void free_pages_exact(void *virt, size_t size)
2828 unsigned long addr = (unsigned long)virt;
2829 unsigned long end = addr + PAGE_ALIGN(size);
2831 while (addr < end) {
2836 EXPORT_SYMBOL(free_pages_exact);
2839 * nr_free_zone_pages - count number of pages beyond high watermark
2840 * @offset: The zone index of the highest zone
2842 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2843 * high watermark within all zones at or below a given zone index. For each
2844 * zone, the number of pages is calculated as:
2845 * present_pages - high_pages
2847 static unsigned long nr_free_zone_pages(int offset)
2852 /* Just pick one node, since fallback list is circular */
2853 unsigned long sum = 0;
2855 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2857 for_each_zone_zonelist(zone, z, zonelist, offset) {
2858 unsigned long size = zone->managed_pages;
2859 unsigned long high = high_wmark_pages(zone);
2868 * nr_free_buffer_pages - count number of pages beyond high watermark
2870 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2871 * watermark within ZONE_DMA and ZONE_NORMAL.
2873 unsigned long nr_free_buffer_pages(void)
2875 return nr_free_zone_pages(gfp_zone(GFP_USER));
2877 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2880 * nr_free_pagecache_pages - count number of pages beyond high watermark
2882 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2883 * high watermark within all zones.
2885 unsigned long nr_free_pagecache_pages(void)
2887 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2890 static inline void show_node(struct zone *zone)
2892 if (IS_ENABLED(CONFIG_NUMA))
2893 printk("Node %d ", zone_to_nid(zone));
2896 void si_meminfo(struct sysinfo *val)
2898 val->totalram = totalram_pages;
2900 val->freeram = global_page_state(NR_FREE_PAGES);
2901 val->bufferram = nr_blockdev_pages();
2902 val->totalhigh = totalhigh_pages;
2903 val->freehigh = nr_free_highpages();
2904 val->mem_unit = PAGE_SIZE;
2907 EXPORT_SYMBOL(si_meminfo);
2910 void si_meminfo_node(struct sysinfo *val, int nid)
2912 pg_data_t *pgdat = NODE_DATA(nid);
2914 val->totalram = pgdat->node_present_pages;
2915 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2916 #ifdef CONFIG_HIGHMEM
2917 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
2918 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2924 val->mem_unit = PAGE_SIZE;
2929 * Determine whether the node should be displayed or not, depending on whether
2930 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2932 bool skip_free_areas_node(unsigned int flags, int nid)
2935 unsigned int cpuset_mems_cookie;
2937 if (!(flags & SHOW_MEM_FILTER_NODES))
2941 cpuset_mems_cookie = get_mems_allowed();
2942 ret = !node_isset(nid, cpuset_current_mems_allowed);
2943 } while (!put_mems_allowed(cpuset_mems_cookie));
2948 #define K(x) ((x) << (PAGE_SHIFT-10))
2950 static void show_migration_types(unsigned char type)
2952 static const char types[MIGRATE_TYPES] = {
2953 [MIGRATE_UNMOVABLE] = 'U',
2954 [MIGRATE_RECLAIMABLE] = 'E',
2955 [MIGRATE_MOVABLE] = 'M',
2956 [MIGRATE_RESERVE] = 'R',
2958 [MIGRATE_CMA] = 'C',
2960 #ifdef CONFIG_MEMORY_ISOLATION
2961 [MIGRATE_ISOLATE] = 'I',
2964 char tmp[MIGRATE_TYPES + 1];
2968 for (i = 0; i < MIGRATE_TYPES; i++) {
2969 if (type & (1 << i))
2974 printk("(%s) ", tmp);
2978 * Show free area list (used inside shift_scroll-lock stuff)
2979 * We also calculate the percentage fragmentation. We do this by counting the
2980 * memory on each free list with the exception of the first item on the list.
2981 * Suppresses nodes that are not allowed by current's cpuset if
2982 * SHOW_MEM_FILTER_NODES is passed.
2984 void show_free_areas(unsigned int filter)
2989 for_each_populated_zone(zone) {
2990 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2993 printk("%s per-cpu:\n", zone->name);
2995 for_each_online_cpu(cpu) {
2996 struct per_cpu_pageset *pageset;
2998 pageset = per_cpu_ptr(zone->pageset, cpu);
3000 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3001 cpu, pageset->pcp.high,
3002 pageset->pcp.batch, pageset->pcp.count);
3006 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3007 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3009 " dirty:%lu writeback:%lu unstable:%lu\n"
3010 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3011 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3013 global_page_state(NR_ACTIVE_ANON),
3014 global_page_state(NR_INACTIVE_ANON),
3015 global_page_state(NR_ISOLATED_ANON),
3016 global_page_state(NR_ACTIVE_FILE),
3017 global_page_state(NR_INACTIVE_FILE),
3018 global_page_state(NR_ISOLATED_FILE),
3019 global_page_state(NR_UNEVICTABLE),
3020 global_page_state(NR_FILE_DIRTY),
3021 global_page_state(NR_WRITEBACK),
3022 global_page_state(NR_UNSTABLE_NFS),
3023 global_page_state(NR_FREE_PAGES),
3024 global_page_state(NR_SLAB_RECLAIMABLE),
3025 global_page_state(NR_SLAB_UNRECLAIMABLE),
3026 global_page_state(NR_FILE_MAPPED),
3027 global_page_state(NR_SHMEM),
3028 global_page_state(NR_PAGETABLE),
3029 global_page_state(NR_BOUNCE),
3030 global_page_state(NR_FREE_CMA_PAGES));
3032 for_each_populated_zone(zone) {
3035 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3043 " active_anon:%lukB"
3044 " inactive_anon:%lukB"
3045 " active_file:%lukB"
3046 " inactive_file:%lukB"
3047 " unevictable:%lukB"
3048 " isolated(anon):%lukB"
3049 " isolated(file):%lukB"
3057 " slab_reclaimable:%lukB"
3058 " slab_unreclaimable:%lukB"
3059 " kernel_stack:%lukB"
3064 " writeback_tmp:%lukB"
3065 " pages_scanned:%lu"
3066 " all_unreclaimable? %s"
3069 K(zone_page_state(zone, NR_FREE_PAGES)),
3070 K(min_wmark_pages(zone)),
3071 K(low_wmark_pages(zone)),
3072 K(high_wmark_pages(zone)),
3073 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3074 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3075 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3076 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3077 K(zone_page_state(zone, NR_UNEVICTABLE)),
3078 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3079 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3080 K(zone->present_pages),
3081 K(zone->managed_pages),
3082 K(zone_page_state(zone, NR_MLOCK)),
3083 K(zone_page_state(zone, NR_FILE_DIRTY)),
3084 K(zone_page_state(zone, NR_WRITEBACK)),
3085 K(zone_page_state(zone, NR_FILE_MAPPED)),
3086 K(zone_page_state(zone, NR_SHMEM)),
3087 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3088 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3089 zone_page_state(zone, NR_KERNEL_STACK) *
3091 K(zone_page_state(zone, NR_PAGETABLE)),
3092 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3093 K(zone_page_state(zone, NR_BOUNCE)),
3094 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3095 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3096 zone->pages_scanned,
3097 (zone->all_unreclaimable ? "yes" : "no")
3099 printk("lowmem_reserve[]:");
3100 for (i = 0; i < MAX_NR_ZONES; i++)
3101 printk(" %lu", zone->lowmem_reserve[i]);
3105 for_each_populated_zone(zone) {
3106 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3107 unsigned char types[MAX_ORDER];
3109 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3112 printk("%s: ", zone->name);
3114 spin_lock_irqsave(&zone->lock, flags);
3115 for (order = 0; order < MAX_ORDER; order++) {
3116 struct free_area *area = &zone->free_area[order];
3119 nr[order] = area->nr_free;
3120 total += nr[order] << order;
3123 for (type = 0; type < MIGRATE_TYPES; type++) {
3124 if (!list_empty(&area->free_list[type]))
3125 types[order] |= 1 << type;
3128 spin_unlock_irqrestore(&zone->lock, flags);
3129 for (order = 0; order < MAX_ORDER; order++) {
3130 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3132 show_migration_types(types[order]);
3134 printk("= %lukB\n", K(total));
3137 hugetlb_show_meminfo();
3139 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3141 show_swap_cache_info();
3144 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3146 zoneref->zone = zone;
3147 zoneref->zone_idx = zone_idx(zone);
3151 * Builds allocation fallback zone lists.
3153 * Add all populated zones of a node to the zonelist.
3155 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3156 int nr_zones, enum zone_type zone_type)
3160 BUG_ON(zone_type >= MAX_NR_ZONES);
3165 zone = pgdat->node_zones + zone_type;
3166 if (populated_zone(zone)) {
3167 zoneref_set_zone(zone,
3168 &zonelist->_zonerefs[nr_zones++]);
3169 check_highest_zone(zone_type);
3172 } while (zone_type);
3179 * 0 = automatic detection of better ordering.
3180 * 1 = order by ([node] distance, -zonetype)
3181 * 2 = order by (-zonetype, [node] distance)
3183 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3184 * the same zonelist. So only NUMA can configure this param.
3186 #define ZONELIST_ORDER_DEFAULT 0
3187 #define ZONELIST_ORDER_NODE 1
3188 #define ZONELIST_ORDER_ZONE 2
3190 /* zonelist order in the kernel.
3191 * set_zonelist_order() will set this to NODE or ZONE.
3193 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3194 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3198 /* The value user specified ....changed by config */
3199 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3200 /* string for sysctl */
3201 #define NUMA_ZONELIST_ORDER_LEN 16
3202 char numa_zonelist_order[16] = "default";
3205 * interface for configure zonelist ordering.
3206 * command line option "numa_zonelist_order"
3207 * = "[dD]efault - default, automatic configuration.
3208 * = "[nN]ode - order by node locality, then by zone within node
3209 * = "[zZ]one - order by zone, then by locality within zone
3212 static int __parse_numa_zonelist_order(char *s)
3214 if (*s == 'd' || *s == 'D') {
3215 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3216 } else if (*s == 'n' || *s == 'N') {
3217 user_zonelist_order = ZONELIST_ORDER_NODE;
3218 } else if (*s == 'z' || *s == 'Z') {
3219 user_zonelist_order = ZONELIST_ORDER_ZONE;
3222 "Ignoring invalid numa_zonelist_order value: "
3229 static __init int setup_numa_zonelist_order(char *s)
3236 ret = __parse_numa_zonelist_order(s);
3238 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3242 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3245 * sysctl handler for numa_zonelist_order
3247 int numa_zonelist_order_handler(ctl_table *table, int write,
3248 void __user *buffer, size_t *length,
3251 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3253 static DEFINE_MUTEX(zl_order_mutex);
3255 mutex_lock(&zl_order_mutex);
3257 strcpy(saved_string, (char*)table->data);
3258 ret = proc_dostring(table, write, buffer, length, ppos);
3262 int oldval = user_zonelist_order;
3263 if (__parse_numa_zonelist_order((char*)table->data)) {
3265 * bogus value. restore saved string
3267 strncpy((char*)table->data, saved_string,
3268 NUMA_ZONELIST_ORDER_LEN);
3269 user_zonelist_order = oldval;
3270 } else if (oldval != user_zonelist_order) {
3271 mutex_lock(&zonelists_mutex);
3272 build_all_zonelists(NULL, NULL);
3273 mutex_unlock(&zonelists_mutex);
3277 mutex_unlock(&zl_order_mutex);
3282 #define MAX_NODE_LOAD (nr_online_nodes)
3283 static int node_load[MAX_NUMNODES];
3286 * find_next_best_node - find the next node that should appear in a given node's fallback list
3287 * @node: node whose fallback list we're appending
3288 * @used_node_mask: nodemask_t of already used nodes
3290 * We use a number of factors to determine which is the next node that should
3291 * appear on a given node's fallback list. The node should not have appeared
3292 * already in @node's fallback list, and it should be the next closest node
3293 * according to the distance array (which contains arbitrary distance values
3294 * from each node to each node in the system), and should also prefer nodes
3295 * with no CPUs, since presumably they'll have very little allocation pressure
3296 * on them otherwise.
3297 * It returns -1 if no node is found.
3299 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3302 int min_val = INT_MAX;
3303 int best_node = NUMA_NO_NODE;
3304 const struct cpumask *tmp = cpumask_of_node(0);
3306 /* Use the local node if we haven't already */
3307 if (!node_isset(node, *used_node_mask)) {
3308 node_set(node, *used_node_mask);
3312 for_each_node_state(n, N_MEMORY) {
3314 /* Don't want a node to appear more than once */
3315 if (node_isset(n, *used_node_mask))
3318 /* Use the distance array to find the distance */
3319 val = node_distance(node, n);
3321 /* Penalize nodes under us ("prefer the next node") */
3324 /* Give preference to headless and unused nodes */
3325 tmp = cpumask_of_node(n);
3326 if (!cpumask_empty(tmp))
3327 val += PENALTY_FOR_NODE_WITH_CPUS;
3329 /* Slight preference for less loaded node */
3330 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3331 val += node_load[n];
3333 if (val < min_val) {
3340 node_set(best_node, *used_node_mask);
3347 * Build zonelists ordered by node and zones within node.
3348 * This results in maximum locality--normal zone overflows into local
3349 * DMA zone, if any--but risks exhausting DMA zone.
3351 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3354 struct zonelist *zonelist;
3356 zonelist = &pgdat->node_zonelists[0];
3357 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3359 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3361 zonelist->_zonerefs[j].zone = NULL;
3362 zonelist->_zonerefs[j].zone_idx = 0;
3366 * Build gfp_thisnode zonelists
3368 static void build_thisnode_zonelists(pg_data_t *pgdat)
3371 struct zonelist *zonelist;
3373 zonelist = &pgdat->node_zonelists[1];
3374 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3375 zonelist->_zonerefs[j].zone = NULL;
3376 zonelist->_zonerefs[j].zone_idx = 0;
3380 * Build zonelists ordered by zone and nodes within zones.
3381 * This results in conserving DMA zone[s] until all Normal memory is
3382 * exhausted, but results in overflowing to remote node while memory
3383 * may still exist in local DMA zone.
3385 static int node_order[MAX_NUMNODES];
3387 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3390 int zone_type; /* needs to be signed */
3392 struct zonelist *zonelist;
3394 zonelist = &pgdat->node_zonelists[0];
3396 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3397 for (j = 0; j < nr_nodes; j++) {
3398 node = node_order[j];
3399 z = &NODE_DATA(node)->node_zones[zone_type];
3400 if (populated_zone(z)) {
3402 &zonelist->_zonerefs[pos++]);
3403 check_highest_zone(zone_type);
3407 zonelist->_zonerefs[pos].zone = NULL;
3408 zonelist->_zonerefs[pos].zone_idx = 0;
3411 static int default_zonelist_order(void)
3414 unsigned long low_kmem_size,total_size;
3418 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3419 * If they are really small and used heavily, the system can fall
3420 * into OOM very easily.
3421 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3423 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3426 for_each_online_node(nid) {
3427 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3428 z = &NODE_DATA(nid)->node_zones[zone_type];
3429 if (populated_zone(z)) {
3430 if (zone_type < ZONE_NORMAL)
3431 low_kmem_size += z->present_pages;
3432 total_size += z->present_pages;
3433 } else if (zone_type == ZONE_NORMAL) {
3435 * If any node has only lowmem, then node order
3436 * is preferred to allow kernel allocations
3437 * locally; otherwise, they can easily infringe
3438 * on other nodes when there is an abundance of
3439 * lowmem available to allocate from.
3441 return ZONELIST_ORDER_NODE;
3445 if (!low_kmem_size || /* there are no DMA area. */
3446 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3447 return ZONELIST_ORDER_NODE;
3449 * look into each node's config.
3450 * If there is a node whose DMA/DMA32 memory is very big area on
3451 * local memory, NODE_ORDER may be suitable.
3453 average_size = total_size /
3454 (nodes_weight(node_states[N_MEMORY]) + 1);
3455 for_each_online_node(nid) {
3458 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3459 z = &NODE_DATA(nid)->node_zones[zone_type];
3460 if (populated_zone(z)) {
3461 if (zone_type < ZONE_NORMAL)
3462 low_kmem_size += z->present_pages;
3463 total_size += z->present_pages;
3466 if (low_kmem_size &&
3467 total_size > average_size && /* ignore small node */
3468 low_kmem_size > total_size * 70/100)
3469 return ZONELIST_ORDER_NODE;
3471 return ZONELIST_ORDER_ZONE;
3474 static void set_zonelist_order(void)
3476 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3477 current_zonelist_order = default_zonelist_order();
3479 current_zonelist_order = user_zonelist_order;
3482 static void build_zonelists(pg_data_t *pgdat)
3486 nodemask_t used_mask;
3487 int local_node, prev_node;
3488 struct zonelist *zonelist;
3489 int order = current_zonelist_order;
3491 /* initialize zonelists */
3492 for (i = 0; i < MAX_ZONELISTS; i++) {
3493 zonelist = pgdat->node_zonelists + i;
3494 zonelist->_zonerefs[0].zone = NULL;
3495 zonelist->_zonerefs[0].zone_idx = 0;
3498 /* NUMA-aware ordering of nodes */
3499 local_node = pgdat->node_id;
3500 load = nr_online_nodes;
3501 prev_node = local_node;
3502 nodes_clear(used_mask);
3504 memset(node_order, 0, sizeof(node_order));
3507 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3509 * We don't want to pressure a particular node.
3510 * So adding penalty to the first node in same
3511 * distance group to make it round-robin.
3513 if (node_distance(local_node, node) !=
3514 node_distance(local_node, prev_node))
3515 node_load[node] = load;
3519 if (order == ZONELIST_ORDER_NODE)
3520 build_zonelists_in_node_order(pgdat, node);
3522 node_order[j++] = node; /* remember order */
3525 if (order == ZONELIST_ORDER_ZONE) {
3526 /* calculate node order -- i.e., DMA last! */
3527 build_zonelists_in_zone_order(pgdat, j);
3530 build_thisnode_zonelists(pgdat);
3533 /* Construct the zonelist performance cache - see further mmzone.h */
3534 static void build_zonelist_cache(pg_data_t *pgdat)
3536 struct zonelist *zonelist;
3537 struct zonelist_cache *zlc;
3540 zonelist = &pgdat->node_zonelists[0];
3541 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3542 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3543 for (z = zonelist->_zonerefs; z->zone; z++)
3544 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3547 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3549 * Return node id of node used for "local" allocations.
3550 * I.e., first node id of first zone in arg node's generic zonelist.
3551 * Used for initializing percpu 'numa_mem', which is used primarily
3552 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3554 int local_memory_node(int node)
3558 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3559 gfp_zone(GFP_KERNEL),
3566 #else /* CONFIG_NUMA */
3568 static void set_zonelist_order(void)
3570 current_zonelist_order = ZONELIST_ORDER_ZONE;
3573 static void build_zonelists(pg_data_t *pgdat)
3575 int node, local_node;
3577 struct zonelist *zonelist;
3579 local_node = pgdat->node_id;
3581 zonelist = &pgdat->node_zonelists[0];
3582 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3585 * Now we build the zonelist so that it contains the zones
3586 * of all the other nodes.
3587 * We don't want to pressure a particular node, so when
3588 * building the zones for node N, we make sure that the
3589 * zones coming right after the local ones are those from
3590 * node N+1 (modulo N)
3592 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3593 if (!node_online(node))
3595 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3598 for (node = 0; node < local_node; node++) {
3599 if (!node_online(node))
3601 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3605 zonelist->_zonerefs[j].zone = NULL;
3606 zonelist->_zonerefs[j].zone_idx = 0;
3609 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3610 static void build_zonelist_cache(pg_data_t *pgdat)
3612 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3615 #endif /* CONFIG_NUMA */
3618 * Boot pageset table. One per cpu which is going to be used for all
3619 * zones and all nodes. The parameters will be set in such a way
3620 * that an item put on a list will immediately be handed over to
3621 * the buddy list. This is safe since pageset manipulation is done
3622 * with interrupts disabled.
3624 * The boot_pagesets must be kept even after bootup is complete for
3625 * unused processors and/or zones. They do play a role for bootstrapping
3626 * hotplugged processors.
3628 * zoneinfo_show() and maybe other functions do
3629 * not check if the processor is online before following the pageset pointer.
3630 * Other parts of the kernel may not check if the zone is available.
3632 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3633 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3634 static void setup_zone_pageset(struct zone *zone);
3637 * Global mutex to protect against size modification of zonelists
3638 * as well as to serialize pageset setup for the new populated zone.
3640 DEFINE_MUTEX(zonelists_mutex);
3642 /* return values int ....just for stop_machine() */
3643 static int __build_all_zonelists(void *data)
3647 pg_data_t *self = data;
3650 memset(node_load, 0, sizeof(node_load));
3653 if (self && !node_online(self->node_id)) {
3654 build_zonelists(self);
3655 build_zonelist_cache(self);
3658 for_each_online_node(nid) {
3659 pg_data_t *pgdat = NODE_DATA(nid);
3661 build_zonelists(pgdat);
3662 build_zonelist_cache(pgdat);
3666 * Initialize the boot_pagesets that are going to be used
3667 * for bootstrapping processors. The real pagesets for
3668 * each zone will be allocated later when the per cpu
3669 * allocator is available.
3671 * boot_pagesets are used also for bootstrapping offline
3672 * cpus if the system is already booted because the pagesets
3673 * are needed to initialize allocators on a specific cpu too.
3674 * F.e. the percpu allocator needs the page allocator which
3675 * needs the percpu allocator in order to allocate its pagesets
3676 * (a chicken-egg dilemma).
3678 for_each_possible_cpu(cpu) {
3679 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3681 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3683 * We now know the "local memory node" for each node--
3684 * i.e., the node of the first zone in the generic zonelist.
3685 * Set up numa_mem percpu variable for on-line cpus. During
3686 * boot, only the boot cpu should be on-line; we'll init the
3687 * secondary cpus' numa_mem as they come on-line. During
3688 * node/memory hotplug, we'll fixup all on-line cpus.
3690 if (cpu_online(cpu))
3691 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3699 * Called with zonelists_mutex held always
3700 * unless system_state == SYSTEM_BOOTING.
3702 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3704 set_zonelist_order();
3706 if (system_state == SYSTEM_BOOTING) {
3707 __build_all_zonelists(NULL);
3708 mminit_verify_zonelist();
3709 cpuset_init_current_mems_allowed();
3711 /* we have to stop all cpus to guarantee there is no user
3713 #ifdef CONFIG_MEMORY_HOTPLUG
3715 setup_zone_pageset(zone);
3717 stop_machine(__build_all_zonelists, pgdat, NULL);
3718 /* cpuset refresh routine should be here */
3720 vm_total_pages = nr_free_pagecache_pages();
3722 * Disable grouping by mobility if the number of pages in the
3723 * system is too low to allow the mechanism to work. It would be
3724 * more accurate, but expensive to check per-zone. This check is
3725 * made on memory-hotadd so a system can start with mobility
3726 * disabled and enable it later
3728 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3729 page_group_by_mobility_disabled = 1;
3731 page_group_by_mobility_disabled = 0;
3733 printk("Built %i zonelists in %s order, mobility grouping %s. "
3734 "Total pages: %ld\n",
3736 zonelist_order_name[current_zonelist_order],
3737 page_group_by_mobility_disabled ? "off" : "on",
3740 printk("Policy zone: %s\n", zone_names[policy_zone]);
3745 * Helper functions to size the waitqueue hash table.
3746 * Essentially these want to choose hash table sizes sufficiently
3747 * large so that collisions trying to wait on pages are rare.
3748 * But in fact, the number of active page waitqueues on typical
3749 * systems is ridiculously low, less than 200. So this is even
3750 * conservative, even though it seems large.
3752 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3753 * waitqueues, i.e. the size of the waitq table given the number of pages.
3755 #define PAGES_PER_WAITQUEUE 256
3757 #ifndef CONFIG_MEMORY_HOTPLUG
3758 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3760 unsigned long size = 1;
3762 pages /= PAGES_PER_WAITQUEUE;
3764 while (size < pages)
3768 * Once we have dozens or even hundreds of threads sleeping
3769 * on IO we've got bigger problems than wait queue collision.
3770 * Limit the size of the wait table to a reasonable size.
3772 size = min(size, 4096UL);
3774 return max(size, 4UL);
3778 * A zone's size might be changed by hot-add, so it is not possible to determine
3779 * a suitable size for its wait_table. So we use the maximum size now.
3781 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3783 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3784 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3785 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3787 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3788 * or more by the traditional way. (See above). It equals:
3790 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3791 * ia64(16K page size) : = ( 8G + 4M)byte.
3792 * powerpc (64K page size) : = (32G +16M)byte.
3794 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3801 * This is an integer logarithm so that shifts can be used later
3802 * to extract the more random high bits from the multiplicative
3803 * hash function before the remainder is taken.
3805 static inline unsigned long wait_table_bits(unsigned long size)
3810 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3813 * Check if a pageblock contains reserved pages
3815 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3819 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3820 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3827 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3828 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3829 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3830 * higher will lead to a bigger reserve which will get freed as contiguous
3831 * blocks as reclaim kicks in
3833 static void setup_zone_migrate_reserve(struct zone *zone)
3835 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3837 unsigned long block_migratetype;
3841 * Get the start pfn, end pfn and the number of blocks to reserve
3842 * We have to be careful to be aligned to pageblock_nr_pages to
3843 * make sure that we always check pfn_valid for the first page in
3846 start_pfn = zone->zone_start_pfn;
3847 end_pfn = zone_end_pfn(zone);
3848 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3849 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3853 * Reserve blocks are generally in place to help high-order atomic
3854 * allocations that are short-lived. A min_free_kbytes value that
3855 * would result in more than 2 reserve blocks for atomic allocations
3856 * is assumed to be in place to help anti-fragmentation for the
3857 * future allocation of hugepages at runtime.
3859 reserve = min(2, reserve);
3861 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3862 if (!pfn_valid(pfn))
3864 page = pfn_to_page(pfn);
3866 /* Watch out for overlapping nodes */
3867 if (page_to_nid(page) != zone_to_nid(zone))
3870 block_migratetype = get_pageblock_migratetype(page);
3872 /* Only test what is necessary when the reserves are not met */
3875 * Blocks with reserved pages will never free, skip
3878 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3879 if (pageblock_is_reserved(pfn, block_end_pfn))
3882 /* If this block is reserved, account for it */
3883 if (block_migratetype == MIGRATE_RESERVE) {
3888 /* Suitable for reserving if this block is movable */
3889 if (block_migratetype == MIGRATE_MOVABLE) {
3890 set_pageblock_migratetype(page,
3892 move_freepages_block(zone, page,
3900 * If the reserve is met and this is a previous reserved block,
3903 if (block_migratetype == MIGRATE_RESERVE) {
3904 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3905 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3911 * Initially all pages are reserved - free ones are freed
3912 * up by free_all_bootmem() once the early boot process is
3913 * done. Non-atomic initialization, single-pass.
3915 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3916 unsigned long start_pfn, enum memmap_context context)
3919 unsigned long end_pfn = start_pfn + size;
3923 if (highest_memmap_pfn < end_pfn - 1)
3924 highest_memmap_pfn = end_pfn - 1;
3926 z = &NODE_DATA(nid)->node_zones[zone];
3927 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3929 * There can be holes in boot-time mem_map[]s
3930 * handed to this function. They do not
3931 * exist on hotplugged memory.
3933 if (context == MEMMAP_EARLY) {
3934 if (!early_pfn_valid(pfn))
3936 if (!early_pfn_in_nid(pfn, nid))
3939 page = pfn_to_page(pfn);
3940 set_page_links(page, zone, nid, pfn);
3941 mminit_verify_page_links(page, zone, nid, pfn);
3942 init_page_count(page);
3943 page_mapcount_reset(page);
3944 page_nid_reset_last(page);
3945 SetPageReserved(page);
3947 * Mark the block movable so that blocks are reserved for
3948 * movable at startup. This will force kernel allocations
3949 * to reserve their blocks rather than leaking throughout
3950 * the address space during boot when many long-lived
3951 * kernel allocations are made. Later some blocks near
3952 * the start are marked MIGRATE_RESERVE by
3953 * setup_zone_migrate_reserve()
3955 * bitmap is created for zone's valid pfn range. but memmap
3956 * can be created for invalid pages (for alignment)
3957 * check here not to call set_pageblock_migratetype() against
3960 if ((z->zone_start_pfn <= pfn)
3961 && (pfn < zone_end_pfn(z))
3962 && !(pfn & (pageblock_nr_pages - 1)))
3963 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3965 INIT_LIST_HEAD(&page->lru);
3966 #ifdef WANT_PAGE_VIRTUAL
3967 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3968 if (!is_highmem_idx(zone))
3969 set_page_address(page, __va(pfn << PAGE_SHIFT));
3974 static void __meminit zone_init_free_lists(struct zone *zone)
3977 for_each_migratetype_order(order, t) {
3978 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3979 zone->free_area[order].nr_free = 0;
3983 #ifndef __HAVE_ARCH_MEMMAP_INIT
3984 #define memmap_init(size, nid, zone, start_pfn) \
3985 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3988 static int __meminit zone_batchsize(struct zone *zone)
3994 * The per-cpu-pages pools are set to around 1000th of the
3995 * size of the zone. But no more than 1/2 of a meg.
3997 * OK, so we don't know how big the cache is. So guess.
3999 batch = zone->managed_pages / 1024;
4000 if (batch * PAGE_SIZE > 512 * 1024)
4001 batch = (512 * 1024) / PAGE_SIZE;
4002 batch /= 4; /* We effectively *= 4 below */
4007 * Clamp the batch to a 2^n - 1 value. Having a power
4008 * of 2 value was found to be more likely to have
4009 * suboptimal cache aliasing properties in some cases.
4011 * For example if 2 tasks are alternately allocating
4012 * batches of pages, one task can end up with a lot
4013 * of pages of one half of the possible page colors
4014 * and the other with pages of the other colors.
4016 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4021 /* The deferral and batching of frees should be suppressed under NOMMU
4024 * The problem is that NOMMU needs to be able to allocate large chunks
4025 * of contiguous memory as there's no hardware page translation to
4026 * assemble apparent contiguous memory from discontiguous pages.
4028 * Queueing large contiguous runs of pages for batching, however,
4029 * causes the pages to actually be freed in smaller chunks. As there
4030 * can be a significant delay between the individual batches being
4031 * recycled, this leads to the once large chunks of space being
4032 * fragmented and becoming unavailable for high-order allocations.
4039 * pcp->high and pcp->batch values are related and dependent on one another:
4040 * ->batch must never be higher then ->high.
4041 * The following function updates them in a safe manner without read side
4044 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4045 * those fields changing asynchronously (acording the the above rule).
4047 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4048 * outside of boot time (or some other assurance that no concurrent updaters
4051 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4052 unsigned long batch)
4054 /* start with a fail safe value for batch */
4058 /* Update high, then batch, in order */
4065 /* a companion to setup_pagelist_highmark() */
4066 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4068 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4071 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4073 struct per_cpu_pages *pcp;
4076 memset(p, 0, sizeof(*p));
4080 pageset_set_batch(p, batch);
4081 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4082 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4086 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4087 * to the value high for the pageset p.
4089 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
4092 unsigned long batch = max(1UL, high / 4);
4093 if ((high / 4) > (PAGE_SHIFT * 8))
4094 batch = PAGE_SHIFT * 8;
4096 pageset_update(&p->pcp, high, batch);
4099 static void __meminit setup_zone_pageset(struct zone *zone)
4103 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4105 for_each_possible_cpu(cpu) {
4106 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4108 setup_pageset(pcp, zone_batchsize(zone));
4110 if (percpu_pagelist_fraction)
4111 setup_pagelist_highmark(pcp,
4112 (zone->managed_pages /
4113 percpu_pagelist_fraction));
4118 * Allocate per cpu pagesets and initialize them.
4119 * Before this call only boot pagesets were available.
4121 void __init setup_per_cpu_pageset(void)
4125 for_each_populated_zone(zone)
4126 setup_zone_pageset(zone);
4129 static noinline __init_refok
4130 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4133 struct pglist_data *pgdat = zone->zone_pgdat;
4137 * The per-page waitqueue mechanism uses hashed waitqueues
4140 zone->wait_table_hash_nr_entries =
4141 wait_table_hash_nr_entries(zone_size_pages);
4142 zone->wait_table_bits =
4143 wait_table_bits(zone->wait_table_hash_nr_entries);
4144 alloc_size = zone->wait_table_hash_nr_entries
4145 * sizeof(wait_queue_head_t);
4147 if (!slab_is_available()) {
4148 zone->wait_table = (wait_queue_head_t *)
4149 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4152 * This case means that a zone whose size was 0 gets new memory
4153 * via memory hot-add.
4154 * But it may be the case that a new node was hot-added. In
4155 * this case vmalloc() will not be able to use this new node's
4156 * memory - this wait_table must be initialized to use this new
4157 * node itself as well.
4158 * To use this new node's memory, further consideration will be
4161 zone->wait_table = vmalloc(alloc_size);
4163 if (!zone->wait_table)
4166 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4167 init_waitqueue_head(zone->wait_table + i);
4172 static __meminit void zone_pcp_init(struct zone *zone)
4175 * per cpu subsystem is not up at this point. The following code
4176 * relies on the ability of the linker to provide the
4177 * offset of a (static) per cpu variable into the per cpu area.
4179 zone->pageset = &boot_pageset;
4181 if (zone->present_pages)
4182 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4183 zone->name, zone->present_pages,
4184 zone_batchsize(zone));
4187 int __meminit init_currently_empty_zone(struct zone *zone,
4188 unsigned long zone_start_pfn,
4190 enum memmap_context context)
4192 struct pglist_data *pgdat = zone->zone_pgdat;
4194 ret = zone_wait_table_init(zone, size);
4197 pgdat->nr_zones = zone_idx(zone) + 1;
4199 zone->zone_start_pfn = zone_start_pfn;
4201 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4202 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4204 (unsigned long)zone_idx(zone),
4205 zone_start_pfn, (zone_start_pfn + size));
4207 zone_init_free_lists(zone);
4212 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4213 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4215 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4216 * Architectures may implement their own version but if add_active_range()
4217 * was used and there are no special requirements, this is a convenient
4220 int __meminit __early_pfn_to_nid(unsigned long pfn)
4222 unsigned long start_pfn, end_pfn;
4225 * NOTE: The following SMP-unsafe globals are only used early in boot
4226 * when the kernel is running single-threaded.
4228 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4229 static int __meminitdata last_nid;
4231 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4234 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4235 if (start_pfn <= pfn && pfn < end_pfn) {
4236 last_start_pfn = start_pfn;
4237 last_end_pfn = end_pfn;
4241 /* This is a memory hole */
4244 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4246 int __meminit early_pfn_to_nid(unsigned long pfn)
4250 nid = __early_pfn_to_nid(pfn);
4253 /* just returns 0 */
4257 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4258 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4262 nid = __early_pfn_to_nid(pfn);
4263 if (nid >= 0 && nid != node)
4270 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4271 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4272 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4274 * If an architecture guarantees that all ranges registered with
4275 * add_active_ranges() contain no holes and may be freed, this
4276 * this function may be used instead of calling free_bootmem() manually.
4278 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4280 unsigned long start_pfn, end_pfn;
4283 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4284 start_pfn = min(start_pfn, max_low_pfn);
4285 end_pfn = min(end_pfn, max_low_pfn);
4287 if (start_pfn < end_pfn)
4288 free_bootmem_node(NODE_DATA(this_nid),
4289 PFN_PHYS(start_pfn),
4290 (end_pfn - start_pfn) << PAGE_SHIFT);
4295 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4296 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4298 * If an architecture guarantees that all ranges registered with
4299 * add_active_ranges() contain no holes and may be freed, this
4300 * function may be used instead of calling memory_present() manually.
4302 void __init sparse_memory_present_with_active_regions(int nid)
4304 unsigned long start_pfn, end_pfn;
4307 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4308 memory_present(this_nid, start_pfn, end_pfn);
4312 * get_pfn_range_for_nid - Return the start and end page frames for a node
4313 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4314 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4315 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4317 * It returns the start and end page frame of a node based on information
4318 * provided by an arch calling add_active_range(). If called for a node
4319 * with no available memory, a warning is printed and the start and end
4322 void __meminit get_pfn_range_for_nid(unsigned int nid,
4323 unsigned long *start_pfn, unsigned long *end_pfn)
4325 unsigned long this_start_pfn, this_end_pfn;
4331 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4332 *start_pfn = min(*start_pfn, this_start_pfn);
4333 *end_pfn = max(*end_pfn, this_end_pfn);
4336 if (*start_pfn == -1UL)
4341 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4342 * assumption is made that zones within a node are ordered in monotonic
4343 * increasing memory addresses so that the "highest" populated zone is used
4345 static void __init find_usable_zone_for_movable(void)
4348 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4349 if (zone_index == ZONE_MOVABLE)
4352 if (arch_zone_highest_possible_pfn[zone_index] >
4353 arch_zone_lowest_possible_pfn[zone_index])
4357 VM_BUG_ON(zone_index == -1);
4358 movable_zone = zone_index;
4362 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4363 * because it is sized independent of architecture. Unlike the other zones,
4364 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4365 * in each node depending on the size of each node and how evenly kernelcore
4366 * is distributed. This helper function adjusts the zone ranges
4367 * provided by the architecture for a given node by using the end of the
4368 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4369 * zones within a node are in order of monotonic increases memory addresses
4371 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4372 unsigned long zone_type,
4373 unsigned long node_start_pfn,
4374 unsigned long node_end_pfn,
4375 unsigned long *zone_start_pfn,
4376 unsigned long *zone_end_pfn)
4378 /* Only adjust if ZONE_MOVABLE is on this node */
4379 if (zone_movable_pfn[nid]) {
4380 /* Size ZONE_MOVABLE */
4381 if (zone_type == ZONE_MOVABLE) {
4382 *zone_start_pfn = zone_movable_pfn[nid];
4383 *zone_end_pfn = min(node_end_pfn,
4384 arch_zone_highest_possible_pfn[movable_zone]);
4386 /* Adjust for ZONE_MOVABLE starting within this range */
4387 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4388 *zone_end_pfn > zone_movable_pfn[nid]) {
4389 *zone_end_pfn = zone_movable_pfn[nid];
4391 /* Check if this whole range is within ZONE_MOVABLE */
4392 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4393 *zone_start_pfn = *zone_end_pfn;
4398 * Return the number of pages a zone spans in a node, including holes
4399 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4401 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4402 unsigned long zone_type,
4403 unsigned long *ignored)
4405 unsigned long node_start_pfn, node_end_pfn;
4406 unsigned long zone_start_pfn, zone_end_pfn;
4408 /* Get the start and end of the node and zone */
4409 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4410 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4411 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4412 adjust_zone_range_for_zone_movable(nid, zone_type,
4413 node_start_pfn, node_end_pfn,
4414 &zone_start_pfn, &zone_end_pfn);
4416 /* Check that this node has pages within the zone's required range */
4417 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4420 /* Move the zone boundaries inside the node if necessary */
4421 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4422 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4424 /* Return the spanned pages */
4425 return zone_end_pfn - zone_start_pfn;
4429 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4430 * then all holes in the requested range will be accounted for.
4432 unsigned long __meminit __absent_pages_in_range(int nid,
4433 unsigned long range_start_pfn,
4434 unsigned long range_end_pfn)
4436 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4437 unsigned long start_pfn, end_pfn;
4440 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4441 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4442 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4443 nr_absent -= end_pfn - start_pfn;
4449 * absent_pages_in_range - Return number of page frames in holes within a range
4450 * @start_pfn: The start PFN to start searching for holes
4451 * @end_pfn: The end PFN to stop searching for holes
4453 * It returns the number of pages frames in memory holes within a range.
4455 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4456 unsigned long end_pfn)
4458 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4461 /* Return the number of page frames in holes in a zone on a node */
4462 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4463 unsigned long zone_type,
4464 unsigned long *ignored)
4466 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4467 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4468 unsigned long node_start_pfn, node_end_pfn;
4469 unsigned long zone_start_pfn, zone_end_pfn;
4471 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4472 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4473 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4475 adjust_zone_range_for_zone_movable(nid, zone_type,
4476 node_start_pfn, node_end_pfn,
4477 &zone_start_pfn, &zone_end_pfn);
4478 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4481 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4482 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4483 unsigned long zone_type,
4484 unsigned long *zones_size)
4486 return zones_size[zone_type];
4489 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4490 unsigned long zone_type,
4491 unsigned long *zholes_size)
4496 return zholes_size[zone_type];
4499 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4501 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4502 unsigned long *zones_size, unsigned long *zholes_size)
4504 unsigned long realtotalpages, totalpages = 0;
4507 for (i = 0; i < MAX_NR_ZONES; i++)
4508 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4510 pgdat->node_spanned_pages = totalpages;
4512 realtotalpages = totalpages;
4513 for (i = 0; i < MAX_NR_ZONES; i++)
4515 zone_absent_pages_in_node(pgdat->node_id, i,
4517 pgdat->node_present_pages = realtotalpages;
4518 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4522 #ifndef CONFIG_SPARSEMEM
4524 * Calculate the size of the zone->blockflags rounded to an unsigned long
4525 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4526 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4527 * round what is now in bits to nearest long in bits, then return it in
4530 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4532 unsigned long usemapsize;
4534 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4535 usemapsize = roundup(zonesize, pageblock_nr_pages);
4536 usemapsize = usemapsize >> pageblock_order;
4537 usemapsize *= NR_PAGEBLOCK_BITS;
4538 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4540 return usemapsize / 8;
4543 static void __init setup_usemap(struct pglist_data *pgdat,
4545 unsigned long zone_start_pfn,
4546 unsigned long zonesize)
4548 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4549 zone->pageblock_flags = NULL;
4551 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4555 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4556 unsigned long zone_start_pfn, unsigned long zonesize) {}
4557 #endif /* CONFIG_SPARSEMEM */
4559 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4561 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4562 void __init set_pageblock_order(void)
4566 /* Check that pageblock_nr_pages has not already been setup */
4567 if (pageblock_order)
4570 if (HPAGE_SHIFT > PAGE_SHIFT)
4571 order = HUGETLB_PAGE_ORDER;
4573 order = MAX_ORDER - 1;
4576 * Assume the largest contiguous order of interest is a huge page.
4577 * This value may be variable depending on boot parameters on IA64 and
4580 pageblock_order = order;
4582 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4585 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4586 * is unused as pageblock_order is set at compile-time. See
4587 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4590 void __init set_pageblock_order(void)
4594 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4596 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4597 unsigned long present_pages)
4599 unsigned long pages = spanned_pages;
4602 * Provide a more accurate estimation if there are holes within
4603 * the zone and SPARSEMEM is in use. If there are holes within the
4604 * zone, each populated memory region may cost us one or two extra
4605 * memmap pages due to alignment because memmap pages for each
4606 * populated regions may not naturally algined on page boundary.
4607 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4609 if (spanned_pages > present_pages + (present_pages >> 4) &&
4610 IS_ENABLED(CONFIG_SPARSEMEM))
4611 pages = present_pages;
4613 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4617 * Set up the zone data structures:
4618 * - mark all pages reserved
4619 * - mark all memory queues empty
4620 * - clear the memory bitmaps
4622 * NOTE: pgdat should get zeroed by caller.
4624 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4625 unsigned long *zones_size, unsigned long *zholes_size)
4628 int nid = pgdat->node_id;
4629 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4632 pgdat_resize_init(pgdat);
4633 #ifdef CONFIG_NUMA_BALANCING
4634 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4635 pgdat->numabalancing_migrate_nr_pages = 0;
4636 pgdat->numabalancing_migrate_next_window = jiffies;
4638 init_waitqueue_head(&pgdat->kswapd_wait);
4639 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4640 pgdat_page_cgroup_init(pgdat);
4642 for (j = 0; j < MAX_NR_ZONES; j++) {
4643 struct zone *zone = pgdat->node_zones + j;
4644 unsigned long size, realsize, freesize, memmap_pages;
4646 size = zone_spanned_pages_in_node(nid, j, zones_size);
4647 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4651 * Adjust freesize so that it accounts for how much memory
4652 * is used by this zone for memmap. This affects the watermark
4653 * and per-cpu initialisations
4655 memmap_pages = calc_memmap_size(size, realsize);
4656 if (freesize >= memmap_pages) {
4657 freesize -= memmap_pages;
4660 " %s zone: %lu pages used for memmap\n",
4661 zone_names[j], memmap_pages);
4664 " %s zone: %lu pages exceeds freesize %lu\n",
4665 zone_names[j], memmap_pages, freesize);
4667 /* Account for reserved pages */
4668 if (j == 0 && freesize > dma_reserve) {
4669 freesize -= dma_reserve;
4670 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4671 zone_names[0], dma_reserve);
4674 if (!is_highmem_idx(j))
4675 nr_kernel_pages += freesize;
4676 /* Charge for highmem memmap if there are enough kernel pages */
4677 else if (nr_kernel_pages > memmap_pages * 2)
4678 nr_kernel_pages -= memmap_pages;
4679 nr_all_pages += freesize;
4681 zone->spanned_pages = size;
4682 zone->present_pages = realsize;
4684 * Set an approximate value for lowmem here, it will be adjusted
4685 * when the bootmem allocator frees pages into the buddy system.
4686 * And all highmem pages will be managed by the buddy system.
4688 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4691 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4693 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4695 zone->name = zone_names[j];
4696 spin_lock_init(&zone->lock);
4697 spin_lock_init(&zone->lru_lock);
4698 zone_seqlock_init(zone);
4699 zone->zone_pgdat = pgdat;
4701 zone_pcp_init(zone);
4702 lruvec_init(&zone->lruvec);
4706 set_pageblock_order();
4707 setup_usemap(pgdat, zone, zone_start_pfn, size);
4708 ret = init_currently_empty_zone(zone, zone_start_pfn,
4709 size, MEMMAP_EARLY);
4711 memmap_init(size, nid, j, zone_start_pfn);
4712 zone_start_pfn += size;
4716 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4718 /* Skip empty nodes */
4719 if (!pgdat->node_spanned_pages)
4722 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4723 /* ia64 gets its own node_mem_map, before this, without bootmem */
4724 if (!pgdat->node_mem_map) {
4725 unsigned long size, start, end;
4729 * The zone's endpoints aren't required to be MAX_ORDER
4730 * aligned but the node_mem_map endpoints must be in order
4731 * for the buddy allocator to function correctly.
4733 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4734 end = pgdat_end_pfn(pgdat);
4735 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4736 size = (end - start) * sizeof(struct page);
4737 map = alloc_remap(pgdat->node_id, size);
4739 map = alloc_bootmem_node_nopanic(pgdat, size);
4740 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4742 #ifndef CONFIG_NEED_MULTIPLE_NODES
4744 * With no DISCONTIG, the global mem_map is just set as node 0's
4746 if (pgdat == NODE_DATA(0)) {
4747 mem_map = NODE_DATA(0)->node_mem_map;
4748 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4749 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4750 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4751 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4754 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4757 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4758 unsigned long node_start_pfn, unsigned long *zholes_size)
4760 pg_data_t *pgdat = NODE_DATA(nid);
4762 /* pg_data_t should be reset to zero when it's allocated */
4763 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4765 pgdat->node_id = nid;
4766 pgdat->node_start_pfn = node_start_pfn;
4767 init_zone_allows_reclaim(nid);
4768 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4770 alloc_node_mem_map(pgdat);
4771 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4772 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4773 nid, (unsigned long)pgdat,
4774 (unsigned long)pgdat->node_mem_map);
4777 free_area_init_core(pgdat, zones_size, zholes_size);
4780 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4782 #if MAX_NUMNODES > 1
4784 * Figure out the number of possible node ids.
4786 void __init setup_nr_node_ids(void)
4789 unsigned int highest = 0;
4791 for_each_node_mask(node, node_possible_map)
4793 nr_node_ids = highest + 1;
4798 * node_map_pfn_alignment - determine the maximum internode alignment
4800 * This function should be called after node map is populated and sorted.
4801 * It calculates the maximum power of two alignment which can distinguish
4804 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4805 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4806 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4807 * shifted, 1GiB is enough and this function will indicate so.
4809 * This is used to test whether pfn -> nid mapping of the chosen memory
4810 * model has fine enough granularity to avoid incorrect mapping for the
4811 * populated node map.
4813 * Returns the determined alignment in pfn's. 0 if there is no alignment
4814 * requirement (single node).
4816 unsigned long __init node_map_pfn_alignment(void)
4818 unsigned long accl_mask = 0, last_end = 0;
4819 unsigned long start, end, mask;
4823 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4824 if (!start || last_nid < 0 || last_nid == nid) {
4831 * Start with a mask granular enough to pin-point to the
4832 * start pfn and tick off bits one-by-one until it becomes
4833 * too coarse to separate the current node from the last.
4835 mask = ~((1 << __ffs(start)) - 1);
4836 while (mask && last_end <= (start & (mask << 1)))
4839 /* accumulate all internode masks */
4843 /* convert mask to number of pages */
4844 return ~accl_mask + 1;
4847 /* Find the lowest pfn for a node */
4848 static unsigned long __init find_min_pfn_for_node(int nid)
4850 unsigned long min_pfn = ULONG_MAX;
4851 unsigned long start_pfn;
4854 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4855 min_pfn = min(min_pfn, start_pfn);
4857 if (min_pfn == ULONG_MAX) {
4859 "Could not find start_pfn for node %d\n", nid);
4867 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4869 * It returns the minimum PFN based on information provided via
4870 * add_active_range().
4872 unsigned long __init find_min_pfn_with_active_regions(void)
4874 return find_min_pfn_for_node(MAX_NUMNODES);
4878 * early_calculate_totalpages()
4879 * Sum pages in active regions for movable zone.
4880 * Populate N_MEMORY for calculating usable_nodes.
4882 static unsigned long __init early_calculate_totalpages(void)
4884 unsigned long totalpages = 0;
4885 unsigned long start_pfn, end_pfn;
4888 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4889 unsigned long pages = end_pfn - start_pfn;
4891 totalpages += pages;
4893 node_set_state(nid, N_MEMORY);
4899 * Find the PFN the Movable zone begins in each node. Kernel memory
4900 * is spread evenly between nodes as long as the nodes have enough
4901 * memory. When they don't, some nodes will have more kernelcore than
4904 static void __init find_zone_movable_pfns_for_nodes(void)
4907 unsigned long usable_startpfn;
4908 unsigned long kernelcore_node, kernelcore_remaining;
4909 /* save the state before borrow the nodemask */
4910 nodemask_t saved_node_state = node_states[N_MEMORY];
4911 unsigned long totalpages = early_calculate_totalpages();
4912 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4915 * If movablecore was specified, calculate what size of
4916 * kernelcore that corresponds so that memory usable for
4917 * any allocation type is evenly spread. If both kernelcore
4918 * and movablecore are specified, then the value of kernelcore
4919 * will be used for required_kernelcore if it's greater than
4920 * what movablecore would have allowed.
4922 if (required_movablecore) {
4923 unsigned long corepages;
4926 * Round-up so that ZONE_MOVABLE is at least as large as what
4927 * was requested by the user
4929 required_movablecore =
4930 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4931 corepages = totalpages - required_movablecore;
4933 required_kernelcore = max(required_kernelcore, corepages);
4936 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4937 if (!required_kernelcore)
4940 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4941 find_usable_zone_for_movable();
4942 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4945 /* Spread kernelcore memory as evenly as possible throughout nodes */
4946 kernelcore_node = required_kernelcore / usable_nodes;
4947 for_each_node_state(nid, N_MEMORY) {
4948 unsigned long start_pfn, end_pfn;
4951 * Recalculate kernelcore_node if the division per node
4952 * now exceeds what is necessary to satisfy the requested
4953 * amount of memory for the kernel
4955 if (required_kernelcore < kernelcore_node)
4956 kernelcore_node = required_kernelcore / usable_nodes;
4959 * As the map is walked, we track how much memory is usable
4960 * by the kernel using kernelcore_remaining. When it is
4961 * 0, the rest of the node is usable by ZONE_MOVABLE
4963 kernelcore_remaining = kernelcore_node;
4965 /* Go through each range of PFNs within this node */
4966 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4967 unsigned long size_pages;
4969 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4970 if (start_pfn >= end_pfn)
4973 /* Account for what is only usable for kernelcore */
4974 if (start_pfn < usable_startpfn) {
4975 unsigned long kernel_pages;
4976 kernel_pages = min(end_pfn, usable_startpfn)
4979 kernelcore_remaining -= min(kernel_pages,
4980 kernelcore_remaining);
4981 required_kernelcore -= min(kernel_pages,
4982 required_kernelcore);
4984 /* Continue if range is now fully accounted */
4985 if (end_pfn <= usable_startpfn) {
4988 * Push zone_movable_pfn to the end so
4989 * that if we have to rebalance
4990 * kernelcore across nodes, we will
4991 * not double account here
4993 zone_movable_pfn[nid] = end_pfn;
4996 start_pfn = usable_startpfn;
5000 * The usable PFN range for ZONE_MOVABLE is from
5001 * start_pfn->end_pfn. Calculate size_pages as the
5002 * number of pages used as kernelcore
5004 size_pages = end_pfn - start_pfn;
5005 if (size_pages > kernelcore_remaining)
5006 size_pages = kernelcore_remaining;
5007 zone_movable_pfn[nid] = start_pfn + size_pages;
5010 * Some kernelcore has been met, update counts and
5011 * break if the kernelcore for this node has been
5014 required_kernelcore -= min(required_kernelcore,
5016 kernelcore_remaining -= size_pages;
5017 if (!kernelcore_remaining)
5023 * If there is still required_kernelcore, we do another pass with one
5024 * less node in the count. This will push zone_movable_pfn[nid] further
5025 * along on the nodes that still have memory until kernelcore is
5029 if (usable_nodes && required_kernelcore > usable_nodes)
5032 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5033 for (nid = 0; nid < MAX_NUMNODES; nid++)
5034 zone_movable_pfn[nid] =
5035 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5038 /* restore the node_state */
5039 node_states[N_MEMORY] = saved_node_state;
5042 /* Any regular or high memory on that node ? */
5043 static void check_for_memory(pg_data_t *pgdat, int nid)
5045 enum zone_type zone_type;
5047 if (N_MEMORY == N_NORMAL_MEMORY)
5050 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5051 struct zone *zone = &pgdat->node_zones[zone_type];
5052 if (zone->present_pages) {
5053 node_set_state(nid, N_HIGH_MEMORY);
5054 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5055 zone_type <= ZONE_NORMAL)
5056 node_set_state(nid, N_NORMAL_MEMORY);
5063 * free_area_init_nodes - Initialise all pg_data_t and zone data
5064 * @max_zone_pfn: an array of max PFNs for each zone
5066 * This will call free_area_init_node() for each active node in the system.
5067 * Using the page ranges provided by add_active_range(), the size of each
5068 * zone in each node and their holes is calculated. If the maximum PFN
5069 * between two adjacent zones match, it is assumed that the zone is empty.
5070 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5071 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5072 * starts where the previous one ended. For example, ZONE_DMA32 starts
5073 * at arch_max_dma_pfn.
5075 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5077 unsigned long start_pfn, end_pfn;
5080 /* Record where the zone boundaries are */
5081 memset(arch_zone_lowest_possible_pfn, 0,
5082 sizeof(arch_zone_lowest_possible_pfn));
5083 memset(arch_zone_highest_possible_pfn, 0,
5084 sizeof(arch_zone_highest_possible_pfn));
5085 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5086 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5087 for (i = 1; i < MAX_NR_ZONES; i++) {
5088 if (i == ZONE_MOVABLE)
5090 arch_zone_lowest_possible_pfn[i] =
5091 arch_zone_highest_possible_pfn[i-1];
5092 arch_zone_highest_possible_pfn[i] =
5093 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5095 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5096 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5098 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5099 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5100 find_zone_movable_pfns_for_nodes();
5102 /* Print out the zone ranges */
5103 printk("Zone ranges:\n");
5104 for (i = 0; i < MAX_NR_ZONES; i++) {
5105 if (i == ZONE_MOVABLE)
5107 printk(KERN_CONT " %-8s ", zone_names[i]);
5108 if (arch_zone_lowest_possible_pfn[i] ==
5109 arch_zone_highest_possible_pfn[i])
5110 printk(KERN_CONT "empty\n");
5112 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5113 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5114 (arch_zone_highest_possible_pfn[i]
5115 << PAGE_SHIFT) - 1);
5118 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5119 printk("Movable zone start for each node\n");
5120 for (i = 0; i < MAX_NUMNODES; i++) {
5121 if (zone_movable_pfn[i])
5122 printk(" Node %d: %#010lx\n", i,
5123 zone_movable_pfn[i] << PAGE_SHIFT);
5126 /* Print out the early node map */
5127 printk("Early memory node ranges\n");
5128 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5129 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5130 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5132 /* Initialise every node */
5133 mminit_verify_pageflags_layout();
5134 setup_nr_node_ids();
5135 for_each_online_node(nid) {
5136 pg_data_t *pgdat = NODE_DATA(nid);
5137 free_area_init_node(nid, NULL,
5138 find_min_pfn_for_node(nid), NULL);
5140 /* Any memory on that node */
5141 if (pgdat->node_present_pages)
5142 node_set_state(nid, N_MEMORY);
5143 check_for_memory(pgdat, nid);
5147 static int __init cmdline_parse_core(char *p, unsigned long *core)
5149 unsigned long long coremem;
5153 coremem = memparse(p, &p);
5154 *core = coremem >> PAGE_SHIFT;
5156 /* Paranoid check that UL is enough for the coremem value */
5157 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5163 * kernelcore=size sets the amount of memory for use for allocations that
5164 * cannot be reclaimed or migrated.
5166 static int __init cmdline_parse_kernelcore(char *p)
5168 return cmdline_parse_core(p, &required_kernelcore);
5172 * movablecore=size sets the amount of memory for use for allocations that
5173 * can be reclaimed or migrated.
5175 static int __init cmdline_parse_movablecore(char *p)
5177 return cmdline_parse_core(p, &required_movablecore);
5180 early_param("kernelcore", cmdline_parse_kernelcore);
5181 early_param("movablecore", cmdline_parse_movablecore);
5183 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5185 unsigned long free_reserved_area(unsigned long start, unsigned long end,
5186 int poison, char *s)
5188 unsigned long pages, pos;
5190 pos = start = PAGE_ALIGN(start);
5192 for (pages = 0; pos < end; pos += PAGE_SIZE, pages++) {
5194 memset((void *)pos, poison, PAGE_SIZE);
5195 free_reserved_page(virt_to_page((void *)pos));
5199 pr_info("Freeing %s memory: %ldK (%lx - %lx)\n",
5200 s, pages << (PAGE_SHIFT - 10), start, end);
5205 #ifdef CONFIG_HIGHMEM
5206 void free_highmem_page(struct page *page)
5208 __free_reserved_page(page);
5215 * set_dma_reserve - set the specified number of pages reserved in the first zone
5216 * @new_dma_reserve: The number of pages to mark reserved
5218 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5219 * In the DMA zone, a significant percentage may be consumed by kernel image
5220 * and other unfreeable allocations which can skew the watermarks badly. This
5221 * function may optionally be used to account for unfreeable pages in the
5222 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5223 * smaller per-cpu batchsize.
5225 void __init set_dma_reserve(unsigned long new_dma_reserve)
5227 dma_reserve = new_dma_reserve;
5230 void __init free_area_init(unsigned long *zones_size)
5232 free_area_init_node(0, zones_size,
5233 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5236 static int page_alloc_cpu_notify(struct notifier_block *self,
5237 unsigned long action, void *hcpu)
5239 int cpu = (unsigned long)hcpu;
5241 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5242 lru_add_drain_cpu(cpu);
5246 * Spill the event counters of the dead processor
5247 * into the current processors event counters.
5248 * This artificially elevates the count of the current
5251 vm_events_fold_cpu(cpu);
5254 * Zero the differential counters of the dead processor
5255 * so that the vm statistics are consistent.
5257 * This is only okay since the processor is dead and cannot
5258 * race with what we are doing.
5260 refresh_cpu_vm_stats(cpu);
5265 void __init page_alloc_init(void)
5267 hotcpu_notifier(page_alloc_cpu_notify, 0);
5271 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5272 * or min_free_kbytes changes.
5274 static void calculate_totalreserve_pages(void)
5276 struct pglist_data *pgdat;
5277 unsigned long reserve_pages = 0;
5278 enum zone_type i, j;
5280 for_each_online_pgdat(pgdat) {
5281 for (i = 0; i < MAX_NR_ZONES; i++) {
5282 struct zone *zone = pgdat->node_zones + i;
5283 unsigned long max = 0;
5285 /* Find valid and maximum lowmem_reserve in the zone */
5286 for (j = i; j < MAX_NR_ZONES; j++) {
5287 if (zone->lowmem_reserve[j] > max)
5288 max = zone->lowmem_reserve[j];
5291 /* we treat the high watermark as reserved pages. */
5292 max += high_wmark_pages(zone);
5294 if (max > zone->managed_pages)
5295 max = zone->managed_pages;
5296 reserve_pages += max;
5298 * Lowmem reserves are not available to
5299 * GFP_HIGHUSER page cache allocations and
5300 * kswapd tries to balance zones to their high
5301 * watermark. As a result, neither should be
5302 * regarded as dirtyable memory, to prevent a
5303 * situation where reclaim has to clean pages
5304 * in order to balance the zones.
5306 zone->dirty_balance_reserve = max;
5309 dirty_balance_reserve = reserve_pages;
5310 totalreserve_pages = reserve_pages;
5314 * setup_per_zone_lowmem_reserve - called whenever
5315 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5316 * has a correct pages reserved value, so an adequate number of
5317 * pages are left in the zone after a successful __alloc_pages().
5319 static void setup_per_zone_lowmem_reserve(void)
5321 struct pglist_data *pgdat;
5322 enum zone_type j, idx;
5324 for_each_online_pgdat(pgdat) {
5325 for (j = 0; j < MAX_NR_ZONES; j++) {
5326 struct zone *zone = pgdat->node_zones + j;
5327 unsigned long managed_pages = zone->managed_pages;
5329 zone->lowmem_reserve[j] = 0;
5333 struct zone *lower_zone;
5337 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5338 sysctl_lowmem_reserve_ratio[idx] = 1;
5340 lower_zone = pgdat->node_zones + idx;
5341 lower_zone->lowmem_reserve[j] = managed_pages /
5342 sysctl_lowmem_reserve_ratio[idx];
5343 managed_pages += lower_zone->managed_pages;
5348 /* update totalreserve_pages */
5349 calculate_totalreserve_pages();
5352 static void __setup_per_zone_wmarks(void)
5354 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5355 unsigned long lowmem_pages = 0;
5357 unsigned long flags;
5359 /* Calculate total number of !ZONE_HIGHMEM pages */
5360 for_each_zone(zone) {
5361 if (!is_highmem(zone))
5362 lowmem_pages += zone->managed_pages;
5365 for_each_zone(zone) {
5368 spin_lock_irqsave(&zone->lock, flags);
5369 tmp = (u64)pages_min * zone->managed_pages;
5370 do_div(tmp, lowmem_pages);
5371 if (is_highmem(zone)) {
5373 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5374 * need highmem pages, so cap pages_min to a small
5377 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5378 * deltas controls asynch page reclaim, and so should
5379 * not be capped for highmem.
5381 unsigned long min_pages;
5383 min_pages = zone->managed_pages / 1024;
5384 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5385 zone->watermark[WMARK_MIN] = min_pages;
5388 * If it's a lowmem zone, reserve a number of pages
5389 * proportionate to the zone's size.
5391 zone->watermark[WMARK_MIN] = tmp;
5394 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5395 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5397 setup_zone_migrate_reserve(zone);
5398 spin_unlock_irqrestore(&zone->lock, flags);
5401 /* update totalreserve_pages */
5402 calculate_totalreserve_pages();
5406 * setup_per_zone_wmarks - called when min_free_kbytes changes
5407 * or when memory is hot-{added|removed}
5409 * Ensures that the watermark[min,low,high] values for each zone are set
5410 * correctly with respect to min_free_kbytes.
5412 void setup_per_zone_wmarks(void)
5414 mutex_lock(&zonelists_mutex);
5415 __setup_per_zone_wmarks();
5416 mutex_unlock(&zonelists_mutex);
5420 * The inactive anon list should be small enough that the VM never has to
5421 * do too much work, but large enough that each inactive page has a chance
5422 * to be referenced again before it is swapped out.
5424 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5425 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5426 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5427 * the anonymous pages are kept on the inactive list.
5430 * memory ratio inactive anon
5431 * -------------------------------------
5440 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5442 unsigned int gb, ratio;
5444 /* Zone size in gigabytes */
5445 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5447 ratio = int_sqrt(10 * gb);
5451 zone->inactive_ratio = ratio;
5454 static void __meminit setup_per_zone_inactive_ratio(void)
5459 calculate_zone_inactive_ratio(zone);
5463 * Initialise min_free_kbytes.
5465 * For small machines we want it small (128k min). For large machines
5466 * we want it large (64MB max). But it is not linear, because network
5467 * bandwidth does not increase linearly with machine size. We use
5469 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5470 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5486 int __meminit init_per_zone_wmark_min(void)
5488 unsigned long lowmem_kbytes;
5490 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5492 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5493 if (min_free_kbytes < 128)
5494 min_free_kbytes = 128;
5495 if (min_free_kbytes > 65536)
5496 min_free_kbytes = 65536;
5497 setup_per_zone_wmarks();
5498 refresh_zone_stat_thresholds();
5499 setup_per_zone_lowmem_reserve();
5500 setup_per_zone_inactive_ratio();
5503 module_init(init_per_zone_wmark_min)
5506 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5507 * that we can call two helper functions whenever min_free_kbytes
5510 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5511 void __user *buffer, size_t *length, loff_t *ppos)
5513 proc_dointvec(table, write, buffer, length, ppos);
5515 setup_per_zone_wmarks();
5520 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5521 void __user *buffer, size_t *length, loff_t *ppos)
5526 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5531 zone->min_unmapped_pages = (zone->managed_pages *
5532 sysctl_min_unmapped_ratio) / 100;
5536 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5537 void __user *buffer, size_t *length, loff_t *ppos)
5542 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5547 zone->min_slab_pages = (zone->managed_pages *
5548 sysctl_min_slab_ratio) / 100;
5554 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5555 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5556 * whenever sysctl_lowmem_reserve_ratio changes.
5558 * The reserve ratio obviously has absolutely no relation with the
5559 * minimum watermarks. The lowmem reserve ratio can only make sense
5560 * if in function of the boot time zone sizes.
5562 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5563 void __user *buffer, size_t *length, loff_t *ppos)
5565 proc_dointvec_minmax(table, write, buffer, length, ppos);
5566 setup_per_zone_lowmem_reserve();
5571 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5572 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5573 * can have before it gets flushed back to buddy allocator.
5576 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5577 void __user *buffer, size_t *length, loff_t *ppos)
5583 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5584 if (!write || (ret < 0))
5587 mutex_lock(&pcp_batch_high_lock);
5588 for_each_populated_zone(zone) {
5589 for_each_possible_cpu(cpu) {
5591 high = zone->managed_pages / percpu_pagelist_fraction;
5592 setup_pagelist_highmark(
5593 per_cpu_ptr(zone->pageset, cpu), high);
5596 mutex_unlock(&pcp_batch_high_lock);
5600 int hashdist = HASHDIST_DEFAULT;
5603 static int __init set_hashdist(char *str)
5607 hashdist = simple_strtoul(str, &str, 0);
5610 __setup("hashdist=", set_hashdist);
5614 * allocate a large system hash table from bootmem
5615 * - it is assumed that the hash table must contain an exact power-of-2
5616 * quantity of entries
5617 * - limit is the number of hash buckets, not the total allocation size
5619 void *__init alloc_large_system_hash(const char *tablename,
5620 unsigned long bucketsize,
5621 unsigned long numentries,
5624 unsigned int *_hash_shift,
5625 unsigned int *_hash_mask,
5626 unsigned long low_limit,
5627 unsigned long high_limit)
5629 unsigned long long max = high_limit;
5630 unsigned long log2qty, size;
5633 /* allow the kernel cmdline to have a say */
5635 /* round applicable memory size up to nearest megabyte */
5636 numentries = nr_kernel_pages;
5637 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5638 numentries >>= 20 - PAGE_SHIFT;
5639 numentries <<= 20 - PAGE_SHIFT;
5641 /* limit to 1 bucket per 2^scale bytes of low memory */
5642 if (scale > PAGE_SHIFT)
5643 numentries >>= (scale - PAGE_SHIFT);
5645 numentries <<= (PAGE_SHIFT - scale);
5647 /* Make sure we've got at least a 0-order allocation.. */
5648 if (unlikely(flags & HASH_SMALL)) {
5649 /* Makes no sense without HASH_EARLY */
5650 WARN_ON(!(flags & HASH_EARLY));
5651 if (!(numentries >> *_hash_shift)) {
5652 numentries = 1UL << *_hash_shift;
5653 BUG_ON(!numentries);
5655 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5656 numentries = PAGE_SIZE / bucketsize;
5658 numentries = roundup_pow_of_two(numentries);
5660 /* limit allocation size to 1/16 total memory by default */
5662 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5663 do_div(max, bucketsize);
5665 max = min(max, 0x80000000ULL);
5667 if (numentries < low_limit)
5668 numentries = low_limit;
5669 if (numentries > max)
5672 log2qty = ilog2(numentries);
5675 size = bucketsize << log2qty;
5676 if (flags & HASH_EARLY)
5677 table = alloc_bootmem_nopanic(size);
5679 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5682 * If bucketsize is not a power-of-two, we may free
5683 * some pages at the end of hash table which
5684 * alloc_pages_exact() automatically does
5686 if (get_order(size) < MAX_ORDER) {
5687 table = alloc_pages_exact(size, GFP_ATOMIC);
5688 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5691 } while (!table && size > PAGE_SIZE && --log2qty);
5694 panic("Failed to allocate %s hash table\n", tablename);
5696 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5699 ilog2(size) - PAGE_SHIFT,
5703 *_hash_shift = log2qty;
5705 *_hash_mask = (1 << log2qty) - 1;
5710 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5711 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5714 #ifdef CONFIG_SPARSEMEM
5715 return __pfn_to_section(pfn)->pageblock_flags;
5717 return zone->pageblock_flags;
5718 #endif /* CONFIG_SPARSEMEM */
5721 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5723 #ifdef CONFIG_SPARSEMEM
5724 pfn &= (PAGES_PER_SECTION-1);
5725 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5727 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5728 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5729 #endif /* CONFIG_SPARSEMEM */
5733 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5734 * @page: The page within the block of interest
5735 * @start_bitidx: The first bit of interest to retrieve
5736 * @end_bitidx: The last bit of interest
5737 * returns pageblock_bits flags
5739 unsigned long get_pageblock_flags_group(struct page *page,
5740 int start_bitidx, int end_bitidx)
5743 unsigned long *bitmap;
5744 unsigned long pfn, bitidx;
5745 unsigned long flags = 0;
5746 unsigned long value = 1;
5748 zone = page_zone(page);
5749 pfn = page_to_pfn(page);
5750 bitmap = get_pageblock_bitmap(zone, pfn);
5751 bitidx = pfn_to_bitidx(zone, pfn);
5753 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5754 if (test_bit(bitidx + start_bitidx, bitmap))
5761 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5762 * @page: The page within the block of interest
5763 * @start_bitidx: The first bit of interest
5764 * @end_bitidx: The last bit of interest
5765 * @flags: The flags to set
5767 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5768 int start_bitidx, int end_bitidx)
5771 unsigned long *bitmap;
5772 unsigned long pfn, bitidx;
5773 unsigned long value = 1;
5775 zone = page_zone(page);
5776 pfn = page_to_pfn(page);
5777 bitmap = get_pageblock_bitmap(zone, pfn);
5778 bitidx = pfn_to_bitidx(zone, pfn);
5779 VM_BUG_ON(!zone_spans_pfn(zone, pfn));
5781 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5783 __set_bit(bitidx + start_bitidx, bitmap);
5785 __clear_bit(bitidx + start_bitidx, bitmap);
5789 * This function checks whether pageblock includes unmovable pages or not.
5790 * If @count is not zero, it is okay to include less @count unmovable pages
5792 * PageLRU check wihtout isolation or lru_lock could race so that
5793 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5794 * expect this function should be exact.
5796 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5797 bool skip_hwpoisoned_pages)
5799 unsigned long pfn, iter, found;
5803 * For avoiding noise data, lru_add_drain_all() should be called
5804 * If ZONE_MOVABLE, the zone never contains unmovable pages
5806 if (zone_idx(zone) == ZONE_MOVABLE)
5808 mt = get_pageblock_migratetype(page);
5809 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5812 pfn = page_to_pfn(page);
5813 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5814 unsigned long check = pfn + iter;
5816 if (!pfn_valid_within(check))
5819 page = pfn_to_page(check);
5821 * We can't use page_count without pin a page
5822 * because another CPU can free compound page.
5823 * This check already skips compound tails of THP
5824 * because their page->_count is zero at all time.
5826 if (!atomic_read(&page->_count)) {
5827 if (PageBuddy(page))
5828 iter += (1 << page_order(page)) - 1;
5833 * The HWPoisoned page may be not in buddy system, and
5834 * page_count() is not 0.
5836 if (skip_hwpoisoned_pages && PageHWPoison(page))
5842 * If there are RECLAIMABLE pages, we need to check it.
5843 * But now, memory offline itself doesn't call shrink_slab()
5844 * and it still to be fixed.
5847 * If the page is not RAM, page_count()should be 0.
5848 * we don't need more check. This is an _used_ not-movable page.
5850 * The problematic thing here is PG_reserved pages. PG_reserved
5851 * is set to both of a memory hole page and a _used_ kernel
5860 bool is_pageblock_removable_nolock(struct page *page)
5866 * We have to be careful here because we are iterating over memory
5867 * sections which are not zone aware so we might end up outside of
5868 * the zone but still within the section.
5869 * We have to take care about the node as well. If the node is offline
5870 * its NODE_DATA will be NULL - see page_zone.
5872 if (!node_online(page_to_nid(page)))
5875 zone = page_zone(page);
5876 pfn = page_to_pfn(page);
5877 if (!zone_spans_pfn(zone, pfn))
5880 return !has_unmovable_pages(zone, page, 0, true);
5885 static unsigned long pfn_max_align_down(unsigned long pfn)
5887 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5888 pageblock_nr_pages) - 1);
5891 static unsigned long pfn_max_align_up(unsigned long pfn)
5893 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5894 pageblock_nr_pages));
5897 /* [start, end) must belong to a single zone. */
5898 static int __alloc_contig_migrate_range(struct compact_control *cc,
5899 unsigned long start, unsigned long end)
5901 /* This function is based on compact_zone() from compaction.c. */
5902 unsigned long nr_reclaimed;
5903 unsigned long pfn = start;
5904 unsigned int tries = 0;
5909 while (pfn < end || !list_empty(&cc->migratepages)) {
5910 if (fatal_signal_pending(current)) {
5915 if (list_empty(&cc->migratepages)) {
5916 cc->nr_migratepages = 0;
5917 pfn = isolate_migratepages_range(cc->zone, cc,
5924 } else if (++tries == 5) {
5925 ret = ret < 0 ? ret : -EBUSY;
5929 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5931 cc->nr_migratepages -= nr_reclaimed;
5933 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
5934 0, MIGRATE_SYNC, MR_CMA);
5937 putback_movable_pages(&cc->migratepages);
5944 * alloc_contig_range() -- tries to allocate given range of pages
5945 * @start: start PFN to allocate
5946 * @end: one-past-the-last PFN to allocate
5947 * @migratetype: migratetype of the underlaying pageblocks (either
5948 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5949 * in range must have the same migratetype and it must
5950 * be either of the two.
5952 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5953 * aligned, however it's the caller's responsibility to guarantee that
5954 * we are the only thread that changes migrate type of pageblocks the
5957 * The PFN range must belong to a single zone.
5959 * Returns zero on success or negative error code. On success all
5960 * pages which PFN is in [start, end) are allocated for the caller and
5961 * need to be freed with free_contig_range().
5963 int alloc_contig_range(unsigned long start, unsigned long end,
5964 unsigned migratetype)
5966 unsigned long outer_start, outer_end;
5969 struct compact_control cc = {
5970 .nr_migratepages = 0,
5972 .zone = page_zone(pfn_to_page(start)),
5974 .ignore_skip_hint = true,
5976 INIT_LIST_HEAD(&cc.migratepages);
5979 * What we do here is we mark all pageblocks in range as
5980 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5981 * have different sizes, and due to the way page allocator
5982 * work, we align the range to biggest of the two pages so
5983 * that page allocator won't try to merge buddies from
5984 * different pageblocks and change MIGRATE_ISOLATE to some
5985 * other migration type.
5987 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5988 * migrate the pages from an unaligned range (ie. pages that
5989 * we are interested in). This will put all the pages in
5990 * range back to page allocator as MIGRATE_ISOLATE.
5992 * When this is done, we take the pages in range from page
5993 * allocator removing them from the buddy system. This way
5994 * page allocator will never consider using them.
5996 * This lets us mark the pageblocks back as
5997 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5998 * aligned range but not in the unaligned, original range are
5999 * put back to page allocator so that buddy can use them.
6002 ret = start_isolate_page_range(pfn_max_align_down(start),
6003 pfn_max_align_up(end), migratetype,
6008 ret = __alloc_contig_migrate_range(&cc, start, end);
6013 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6014 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6015 * more, all pages in [start, end) are free in page allocator.
6016 * What we are going to do is to allocate all pages from
6017 * [start, end) (that is remove them from page allocator).
6019 * The only problem is that pages at the beginning and at the
6020 * end of interesting range may be not aligned with pages that
6021 * page allocator holds, ie. they can be part of higher order
6022 * pages. Because of this, we reserve the bigger range and
6023 * once this is done free the pages we are not interested in.
6025 * We don't have to hold zone->lock here because the pages are
6026 * isolated thus they won't get removed from buddy.
6029 lru_add_drain_all();
6033 outer_start = start;
6034 while (!PageBuddy(pfn_to_page(outer_start))) {
6035 if (++order >= MAX_ORDER) {
6039 outer_start &= ~0UL << order;
6042 /* Make sure the range is really isolated. */
6043 if (test_pages_isolated(outer_start, end, false)) {
6044 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6051 /* Grab isolated pages from freelists. */
6052 outer_end = isolate_freepages_range(&cc, outer_start, end);
6058 /* Free head and tail (if any) */
6059 if (start != outer_start)
6060 free_contig_range(outer_start, start - outer_start);
6061 if (end != outer_end)
6062 free_contig_range(end, outer_end - end);
6065 undo_isolate_page_range(pfn_max_align_down(start),
6066 pfn_max_align_up(end), migratetype);
6070 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6072 unsigned int count = 0;
6074 for (; nr_pages--; pfn++) {
6075 struct page *page = pfn_to_page(pfn);
6077 count += page_count(page) != 1;
6080 WARN(count != 0, "%d pages are still in use!\n", count);
6084 #ifdef CONFIG_MEMORY_HOTPLUG
6085 static int __meminit __zone_pcp_update(void *data)
6087 struct zone *zone = data;
6089 unsigned long batch = zone_batchsize(zone), flags;
6091 for_each_possible_cpu(cpu) {
6092 struct per_cpu_pageset *pset;
6093 struct per_cpu_pages *pcp;
6095 pset = per_cpu_ptr(zone->pageset, cpu);
6098 local_irq_save(flags);
6100 free_pcppages_bulk(zone, pcp->count, pcp);
6101 drain_zonestat(zone, pset);
6102 setup_pageset(pset, batch);
6103 local_irq_restore(flags);
6108 void __meminit zone_pcp_update(struct zone *zone)
6110 mutex_lock(&pcp_batch_high_lock);
6111 stop_machine(__zone_pcp_update, zone, NULL);
6112 mutex_unlock(&pcp_batch_high_lock);
6116 void zone_pcp_reset(struct zone *zone)
6118 unsigned long flags;
6120 struct per_cpu_pageset *pset;
6122 /* avoid races with drain_pages() */
6123 local_irq_save(flags);
6124 if (zone->pageset != &boot_pageset) {
6125 for_each_online_cpu(cpu) {
6126 pset = per_cpu_ptr(zone->pageset, cpu);
6127 drain_zonestat(zone, pset);
6129 free_percpu(zone->pageset);
6130 zone->pageset = &boot_pageset;
6132 local_irq_restore(flags);
6135 #ifdef CONFIG_MEMORY_HOTREMOVE
6137 * All pages in the range must be isolated before calling this.
6140 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6146 unsigned long flags;
6147 /* find the first valid pfn */
6148 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6153 zone = page_zone(pfn_to_page(pfn));
6154 spin_lock_irqsave(&zone->lock, flags);
6156 while (pfn < end_pfn) {
6157 if (!pfn_valid(pfn)) {
6161 page = pfn_to_page(pfn);
6163 * The HWPoisoned page may be not in buddy system, and
6164 * page_count() is not 0.
6166 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6168 SetPageReserved(page);
6172 BUG_ON(page_count(page));
6173 BUG_ON(!PageBuddy(page));
6174 order = page_order(page);
6175 #ifdef CONFIG_DEBUG_VM
6176 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6177 pfn, 1 << order, end_pfn);
6179 list_del(&page->lru);
6180 rmv_page_order(page);
6181 zone->free_area[order].nr_free--;
6182 for (i = 0; i < (1 << order); i++)
6183 SetPageReserved((page+i));
6184 pfn += (1 << order);
6186 spin_unlock_irqrestore(&zone->lock, flags);
6190 #ifdef CONFIG_MEMORY_FAILURE
6191 bool is_free_buddy_page(struct page *page)
6193 struct zone *zone = page_zone(page);
6194 unsigned long pfn = page_to_pfn(page);
6195 unsigned long flags;
6198 spin_lock_irqsave(&zone->lock, flags);
6199 for (order = 0; order < MAX_ORDER; order++) {
6200 struct page *page_head = page - (pfn & ((1 << order) - 1));
6202 if (PageBuddy(page_head) && page_order(page_head) >= order)
6205 spin_unlock_irqrestore(&zone->lock, flags);
6207 return order < MAX_ORDER;
6211 static const struct trace_print_flags pageflag_names[] = {
6212 {1UL << PG_locked, "locked" },
6213 {1UL << PG_error, "error" },
6214 {1UL << PG_referenced, "referenced" },
6215 {1UL << PG_uptodate, "uptodate" },
6216 {1UL << PG_dirty, "dirty" },
6217 {1UL << PG_lru, "lru" },
6218 {1UL << PG_active, "active" },
6219 {1UL << PG_slab, "slab" },
6220 {1UL << PG_owner_priv_1, "owner_priv_1" },
6221 {1UL << PG_arch_1, "arch_1" },
6222 {1UL << PG_reserved, "reserved" },
6223 {1UL << PG_private, "private" },
6224 {1UL << PG_private_2, "private_2" },
6225 {1UL << PG_writeback, "writeback" },
6226 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6227 {1UL << PG_head, "head" },
6228 {1UL << PG_tail, "tail" },
6230 {1UL << PG_compound, "compound" },
6232 {1UL << PG_swapcache, "swapcache" },
6233 {1UL << PG_mappedtodisk, "mappedtodisk" },
6234 {1UL << PG_reclaim, "reclaim" },
6235 {1UL << PG_swapbacked, "swapbacked" },
6236 {1UL << PG_unevictable, "unevictable" },
6238 {1UL << PG_mlocked, "mlocked" },
6240 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6241 {1UL << PG_uncached, "uncached" },
6243 #ifdef CONFIG_MEMORY_FAILURE
6244 {1UL << PG_hwpoison, "hwpoison" },
6246 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6247 {1UL << PG_compound_lock, "compound_lock" },
6251 static void dump_page_flags(unsigned long flags)
6253 const char *delim = "";
6257 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6259 printk(KERN_ALERT "page flags: %#lx(", flags);
6261 /* remove zone id */
6262 flags &= (1UL << NR_PAGEFLAGS) - 1;
6264 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6266 mask = pageflag_names[i].mask;
6267 if ((flags & mask) != mask)
6271 printk("%s%s", delim, pageflag_names[i].name);
6275 /* check for left over flags */
6277 printk("%s%#lx", delim, flags);
6282 void dump_page(struct page *page)
6285 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6286 page, atomic_read(&page->_count), page_mapcount(page),
6287 page->mapping, page->index);
6288 dump_page_flags(page->flags);
6289 mem_cgroup_print_bad_page(page);