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 /* Protect totalram_pages and zone->managed_pages */
107 static DEFINE_SPINLOCK(managed_page_count_lock);
109 unsigned long totalram_pages __read_mostly;
110 unsigned long totalreserve_pages __read_mostly;
112 * When calculating the number of globally allowed dirty pages, there
113 * is a certain number of per-zone reserves that should not be
114 * considered dirtyable memory. This is the sum of those reserves
115 * over all existing zones that contribute dirtyable memory.
117 unsigned long dirty_balance_reserve __read_mostly;
119 int percpu_pagelist_fraction;
120 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
122 #ifdef CONFIG_PM_SLEEP
124 * The following functions are used by the suspend/hibernate code to temporarily
125 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
126 * while devices are suspended. To avoid races with the suspend/hibernate code,
127 * they should always be called with pm_mutex held (gfp_allowed_mask also should
128 * only be modified with pm_mutex held, unless the suspend/hibernate code is
129 * guaranteed not to run in parallel with that modification).
132 static gfp_t saved_gfp_mask;
134 void pm_restore_gfp_mask(void)
136 WARN_ON(!mutex_is_locked(&pm_mutex));
137 if (saved_gfp_mask) {
138 gfp_allowed_mask = saved_gfp_mask;
143 void pm_restrict_gfp_mask(void)
145 WARN_ON(!mutex_is_locked(&pm_mutex));
146 WARN_ON(saved_gfp_mask);
147 saved_gfp_mask = gfp_allowed_mask;
148 gfp_allowed_mask &= ~GFP_IOFS;
151 bool pm_suspended_storage(void)
153 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
157 #endif /* CONFIG_PM_SLEEP */
159 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
160 int pageblock_order __read_mostly;
163 static void __free_pages_ok(struct page *page, unsigned int order);
166 * results with 256, 32 in the lowmem_reserve sysctl:
167 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
168 * 1G machine -> (16M dma, 784M normal, 224M high)
169 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
170 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
171 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
173 * TBD: should special case ZONE_DMA32 machines here - in those we normally
174 * don't need any ZONE_NORMAL reservation
176 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
177 #ifdef CONFIG_ZONE_DMA
180 #ifdef CONFIG_ZONE_DMA32
183 #ifdef CONFIG_HIGHMEM
189 EXPORT_SYMBOL(totalram_pages);
191 static char * const zone_names[MAX_NR_ZONES] = {
192 #ifdef CONFIG_ZONE_DMA
195 #ifdef CONFIG_ZONE_DMA32
199 #ifdef CONFIG_HIGHMEM
205 int min_free_kbytes = 1024;
207 static unsigned long __meminitdata nr_kernel_pages;
208 static unsigned long __meminitdata nr_all_pages;
209 static unsigned long __meminitdata dma_reserve;
211 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
212 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
213 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
214 static unsigned long __initdata required_kernelcore;
215 static unsigned long __initdata required_movablecore;
216 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
218 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
220 EXPORT_SYMBOL(movable_zone);
221 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
224 int nr_node_ids __read_mostly = MAX_NUMNODES;
225 int nr_online_nodes __read_mostly = 1;
226 EXPORT_SYMBOL(nr_node_ids);
227 EXPORT_SYMBOL(nr_online_nodes);
230 int page_group_by_mobility_disabled __read_mostly;
232 void set_pageblock_migratetype(struct page *page, int migratetype)
235 if (unlikely(page_group_by_mobility_disabled))
236 migratetype = MIGRATE_UNMOVABLE;
238 set_pageblock_flags_group(page, (unsigned long)migratetype,
239 PB_migrate, PB_migrate_end);
242 bool oom_killer_disabled __read_mostly;
244 #ifdef CONFIG_DEBUG_VM
245 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
249 unsigned long pfn = page_to_pfn(page);
250 unsigned long sp, start_pfn;
253 seq = zone_span_seqbegin(zone);
254 start_pfn = zone->zone_start_pfn;
255 sp = zone->spanned_pages;
256 if (!zone_spans_pfn(zone, pfn))
258 } while (zone_span_seqretry(zone, seq));
261 pr_err("page %lu outside zone [ %lu - %lu ]\n",
262 pfn, start_pfn, start_pfn + sp);
267 static int page_is_consistent(struct zone *zone, struct page *page)
269 if (!pfn_valid_within(page_to_pfn(page)))
271 if (zone != page_zone(page))
277 * Temporary debugging check for pages not lying within a given zone.
279 static int bad_range(struct zone *zone, struct page *page)
281 if (page_outside_zone_boundaries(zone, page))
283 if (!page_is_consistent(zone, page))
289 static inline int bad_range(struct zone *zone, struct page *page)
295 static void bad_page(struct page *page)
297 static unsigned long resume;
298 static unsigned long nr_shown;
299 static unsigned long nr_unshown;
301 /* Don't complain about poisoned pages */
302 if (PageHWPoison(page)) {
303 page_mapcount_reset(page); /* remove PageBuddy */
308 * Allow a burst of 60 reports, then keep quiet for that minute;
309 * or allow a steady drip of one report per second.
311 if (nr_shown == 60) {
312 if (time_before(jiffies, resume)) {
318 "BUG: Bad page state: %lu messages suppressed\n",
325 resume = jiffies + 60 * HZ;
327 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
328 current->comm, page_to_pfn(page));
334 /* Leave bad fields for debug, except PageBuddy could make trouble */
335 page_mapcount_reset(page); /* remove PageBuddy */
336 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
340 * Higher-order pages are called "compound pages". They are structured thusly:
342 * The first PAGE_SIZE page is called the "head page".
344 * The remaining PAGE_SIZE pages are called "tail pages".
346 * All pages have PG_compound set. All tail pages have their ->first_page
347 * pointing at the head page.
349 * The first tail page's ->lru.next holds the address of the compound page's
350 * put_page() function. Its ->lru.prev holds the order of allocation.
351 * This usage means that zero-order pages may not be compound.
354 static void free_compound_page(struct page *page)
356 __free_pages_ok(page, compound_order(page));
359 void prep_compound_page(struct page *page, unsigned long order)
362 int nr_pages = 1 << order;
364 set_compound_page_dtor(page, free_compound_page);
365 set_compound_order(page, order);
367 for (i = 1; i < nr_pages; i++) {
368 struct page *p = page + i;
370 set_page_count(p, 0);
371 p->first_page = page;
375 /* update __split_huge_page_refcount if you change this function */
376 static int destroy_compound_page(struct page *page, unsigned long order)
379 int nr_pages = 1 << order;
382 if (unlikely(compound_order(page) != order)) {
387 __ClearPageHead(page);
389 for (i = 1; i < nr_pages; i++) {
390 struct page *p = page + i;
392 if (unlikely(!PageTail(p) || (p->first_page != page))) {
402 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
407 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
408 * and __GFP_HIGHMEM from hard or soft interrupt context.
410 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
411 for (i = 0; i < (1 << order); i++)
412 clear_highpage(page + i);
415 #ifdef CONFIG_DEBUG_PAGEALLOC
416 unsigned int _debug_guardpage_minorder;
418 static int __init debug_guardpage_minorder_setup(char *buf)
422 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
423 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
426 _debug_guardpage_minorder = res;
427 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
430 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
432 static inline void set_page_guard_flag(struct page *page)
434 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
437 static inline void clear_page_guard_flag(struct page *page)
439 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
442 static inline void set_page_guard_flag(struct page *page) { }
443 static inline void clear_page_guard_flag(struct page *page) { }
446 static inline void set_page_order(struct page *page, int order)
448 set_page_private(page, order);
449 __SetPageBuddy(page);
452 static inline void rmv_page_order(struct page *page)
454 __ClearPageBuddy(page);
455 set_page_private(page, 0);
459 * Locate the struct page for both the matching buddy in our
460 * pair (buddy1) and the combined O(n+1) page they form (page).
462 * 1) Any buddy B1 will have an order O twin B2 which satisfies
463 * the following equation:
465 * For example, if the starting buddy (buddy2) is #8 its order
467 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
469 * 2) Any buddy B will have an order O+1 parent P which
470 * satisfies the following equation:
473 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
475 static inline unsigned long
476 __find_buddy_index(unsigned long page_idx, unsigned int order)
478 return page_idx ^ (1 << order);
482 * This function checks whether a page is free && is the buddy
483 * we can do coalesce a page and its buddy if
484 * (a) the buddy is not in a hole &&
485 * (b) the buddy is in the buddy system &&
486 * (c) a page and its buddy have the same order &&
487 * (d) a page and its buddy are in the same zone.
489 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
490 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
492 * For recording page's order, we use page_private(page).
494 static inline int page_is_buddy(struct page *page, struct page *buddy,
497 if (!pfn_valid_within(page_to_pfn(buddy)))
500 if (page_zone_id(page) != page_zone_id(buddy))
503 if (page_is_guard(buddy) && page_order(buddy) == order) {
504 VM_BUG_ON(page_count(buddy) != 0);
508 if (PageBuddy(buddy) && page_order(buddy) == order) {
509 VM_BUG_ON(page_count(buddy) != 0);
516 * Freeing function for a buddy system allocator.
518 * The concept of a buddy system is to maintain direct-mapped table
519 * (containing bit values) for memory blocks of various "orders".
520 * The bottom level table contains the map for the smallest allocatable
521 * units of memory (here, pages), and each level above it describes
522 * pairs of units from the levels below, hence, "buddies".
523 * At a high level, all that happens here is marking the table entry
524 * at the bottom level available, and propagating the changes upward
525 * as necessary, plus some accounting needed to play nicely with other
526 * parts of the VM system.
527 * At each level, we keep a list of pages, which are heads of continuous
528 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
529 * order is recorded in page_private(page) field.
530 * So when we are allocating or freeing one, we can derive the state of the
531 * other. That is, if we allocate a small block, and both were
532 * free, the remainder of the region must be split into blocks.
533 * If a block is freed, and its buddy is also free, then this
534 * triggers coalescing into a block of larger size.
539 static inline void __free_one_page(struct page *page,
540 struct zone *zone, unsigned int order,
543 unsigned long page_idx;
544 unsigned long combined_idx;
545 unsigned long uninitialized_var(buddy_idx);
548 VM_BUG_ON(!zone_is_initialized(zone));
550 if (unlikely(PageCompound(page)))
551 if (unlikely(destroy_compound_page(page, order)))
554 VM_BUG_ON(migratetype == -1);
556 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
558 VM_BUG_ON(page_idx & ((1 << order) - 1));
559 VM_BUG_ON(bad_range(zone, page));
561 while (order < MAX_ORDER-1) {
562 buddy_idx = __find_buddy_index(page_idx, order);
563 buddy = page + (buddy_idx - page_idx);
564 if (!page_is_buddy(page, buddy, order))
567 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
568 * merge with it and move up one order.
570 if (page_is_guard(buddy)) {
571 clear_page_guard_flag(buddy);
572 set_page_private(page, 0);
573 __mod_zone_freepage_state(zone, 1 << order,
576 list_del(&buddy->lru);
577 zone->free_area[order].nr_free--;
578 rmv_page_order(buddy);
580 combined_idx = buddy_idx & page_idx;
581 page = page + (combined_idx - page_idx);
582 page_idx = combined_idx;
585 set_page_order(page, order);
588 * If this is not the largest possible page, check if the buddy
589 * of the next-highest order is free. If it is, it's possible
590 * that pages are being freed that will coalesce soon. In case,
591 * that is happening, add the free page to the tail of the list
592 * so it's less likely to be used soon and more likely to be merged
593 * as a higher order page
595 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
596 struct page *higher_page, *higher_buddy;
597 combined_idx = buddy_idx & page_idx;
598 higher_page = page + (combined_idx - page_idx);
599 buddy_idx = __find_buddy_index(combined_idx, order + 1);
600 higher_buddy = higher_page + (buddy_idx - combined_idx);
601 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
602 list_add_tail(&page->lru,
603 &zone->free_area[order].free_list[migratetype]);
608 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
610 zone->free_area[order].nr_free++;
613 static inline int free_pages_check(struct page *page)
615 if (unlikely(page_mapcount(page) |
616 (page->mapping != NULL) |
617 (atomic_read(&page->_count) != 0) |
618 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
619 (mem_cgroup_bad_page_check(page)))) {
623 page_nid_reset_last(page);
624 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
625 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
630 * Frees a number of pages from the PCP lists
631 * Assumes all pages on list are in same zone, and of same order.
632 * count is the number of pages to free.
634 * If the zone was previously in an "all pages pinned" state then look to
635 * see if this freeing clears that state.
637 * And clear the zone's pages_scanned counter, to hold off the "all pages are
638 * pinned" detection logic.
640 static void free_pcppages_bulk(struct zone *zone, int count,
641 struct per_cpu_pages *pcp)
647 spin_lock(&zone->lock);
648 zone->all_unreclaimable = 0;
649 zone->pages_scanned = 0;
653 struct list_head *list;
656 * Remove pages from lists in a round-robin fashion. A
657 * batch_free count is maintained that is incremented when an
658 * empty list is encountered. This is so more pages are freed
659 * off fuller lists instead of spinning excessively around empty
664 if (++migratetype == MIGRATE_PCPTYPES)
666 list = &pcp->lists[migratetype];
667 } while (list_empty(list));
669 /* This is the only non-empty list. Free them all. */
670 if (batch_free == MIGRATE_PCPTYPES)
671 batch_free = to_free;
674 int mt; /* migratetype of the to-be-freed page */
676 page = list_entry(list->prev, struct page, lru);
677 /* must delete as __free_one_page list manipulates */
678 list_del(&page->lru);
679 mt = get_freepage_migratetype(page);
680 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
681 __free_one_page(page, zone, 0, mt);
682 trace_mm_page_pcpu_drain(page, 0, mt);
683 if (likely(!is_migrate_isolate_page(page))) {
684 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
685 if (is_migrate_cma(mt))
686 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
688 } while (--to_free && --batch_free && !list_empty(list));
690 spin_unlock(&zone->lock);
693 static void free_one_page(struct zone *zone, struct page *page, int order,
696 spin_lock(&zone->lock);
697 zone->all_unreclaimable = 0;
698 zone->pages_scanned = 0;
700 __free_one_page(page, zone, order, migratetype);
701 if (unlikely(!is_migrate_isolate(migratetype)))
702 __mod_zone_freepage_state(zone, 1 << order, migratetype);
703 spin_unlock(&zone->lock);
706 static bool free_pages_prepare(struct page *page, unsigned int order)
711 trace_mm_page_free(page, order);
712 kmemcheck_free_shadow(page, order);
715 page->mapping = NULL;
716 for (i = 0; i < (1 << order); i++)
717 bad += free_pages_check(page + i);
721 if (!PageHighMem(page)) {
722 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
723 debug_check_no_obj_freed(page_address(page),
726 arch_free_page(page, order);
727 kernel_map_pages(page, 1 << order, 0);
732 static void __free_pages_ok(struct page *page, unsigned int order)
737 if (!free_pages_prepare(page, order))
740 local_irq_save(flags);
741 __count_vm_events(PGFREE, 1 << order);
742 migratetype = get_pageblock_migratetype(page);
743 set_freepage_migratetype(page, migratetype);
744 free_one_page(page_zone(page), page, order, migratetype);
745 local_irq_restore(flags);
749 * Read access to zone->managed_pages is safe because it's unsigned long,
750 * but we still need to serialize writers. Currently all callers of
751 * __free_pages_bootmem() except put_page_bootmem() should only be used
752 * at boot time. So for shorter boot time, we shift the burden to
753 * put_page_bootmem() to serialize writers.
755 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
757 unsigned int nr_pages = 1 << order;
761 for (loop = 0; loop < nr_pages; loop++) {
762 struct page *p = &page[loop];
764 if (loop + 1 < nr_pages)
766 __ClearPageReserved(p);
767 set_page_count(p, 0);
770 page_zone(page)->managed_pages += 1 << order;
771 set_page_refcounted(page);
772 __free_pages(page, order);
776 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
777 void __init init_cma_reserved_pageblock(struct page *page)
779 unsigned i = pageblock_nr_pages;
780 struct page *p = page;
783 __ClearPageReserved(p);
784 set_page_count(p, 0);
787 set_page_refcounted(page);
788 set_pageblock_migratetype(page, MIGRATE_CMA);
789 __free_pages(page, pageblock_order);
790 totalram_pages += pageblock_nr_pages;
791 #ifdef CONFIG_HIGHMEM
792 if (PageHighMem(page))
793 totalhigh_pages += pageblock_nr_pages;
799 * The order of subdivision here is critical for the IO subsystem.
800 * Please do not alter this order without good reasons and regression
801 * testing. Specifically, as large blocks of memory are subdivided,
802 * the order in which smaller blocks are delivered depends on the order
803 * they're subdivided in this function. This is the primary factor
804 * influencing the order in which pages are delivered to the IO
805 * subsystem according to empirical testing, and this is also justified
806 * by considering the behavior of a buddy system containing a single
807 * large block of memory acted on by a series of small allocations.
808 * This behavior is a critical factor in sglist merging's success.
812 static inline void expand(struct zone *zone, struct page *page,
813 int low, int high, struct free_area *area,
816 unsigned long size = 1 << high;
822 VM_BUG_ON(bad_range(zone, &page[size]));
824 #ifdef CONFIG_DEBUG_PAGEALLOC
825 if (high < debug_guardpage_minorder()) {
827 * Mark as guard pages (or page), that will allow to
828 * merge back to allocator when buddy will be freed.
829 * Corresponding page table entries will not be touched,
830 * pages will stay not present in virtual address space
832 INIT_LIST_HEAD(&page[size].lru);
833 set_page_guard_flag(&page[size]);
834 set_page_private(&page[size], high);
835 /* Guard pages are not available for any usage */
836 __mod_zone_freepage_state(zone, -(1 << high),
841 list_add(&page[size].lru, &area->free_list[migratetype]);
843 set_page_order(&page[size], high);
848 * This page is about to be returned from the page allocator
850 static inline int check_new_page(struct page *page)
852 if (unlikely(page_mapcount(page) |
853 (page->mapping != NULL) |
854 (atomic_read(&page->_count) != 0) |
855 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
856 (mem_cgroup_bad_page_check(page)))) {
863 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
867 for (i = 0; i < (1 << order); i++) {
868 struct page *p = page + i;
869 if (unlikely(check_new_page(p)))
873 set_page_private(page, 0);
874 set_page_refcounted(page);
876 arch_alloc_page(page, order);
877 kernel_map_pages(page, 1 << order, 1);
879 if (gfp_flags & __GFP_ZERO)
880 prep_zero_page(page, order, gfp_flags);
882 if (order && (gfp_flags & __GFP_COMP))
883 prep_compound_page(page, order);
889 * Go through the free lists for the given migratetype and remove
890 * the smallest available page from the freelists
893 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
896 unsigned int current_order;
897 struct free_area * area;
900 /* Find a page of the appropriate size in the preferred list */
901 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
902 area = &(zone->free_area[current_order]);
903 if (list_empty(&area->free_list[migratetype]))
906 page = list_entry(area->free_list[migratetype].next,
908 list_del(&page->lru);
909 rmv_page_order(page);
911 expand(zone, page, order, current_order, area, migratetype);
920 * This array describes the order lists are fallen back to when
921 * the free lists for the desirable migrate type are depleted
923 static int fallbacks[MIGRATE_TYPES][4] = {
924 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
925 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
927 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
928 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
930 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
932 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
933 #ifdef CONFIG_MEMORY_ISOLATION
934 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
939 * Move the free pages in a range to the free lists of the requested type.
940 * Note that start_page and end_pages are not aligned on a pageblock
941 * boundary. If alignment is required, use move_freepages_block()
943 int move_freepages(struct zone *zone,
944 struct page *start_page, struct page *end_page,
951 #ifndef CONFIG_HOLES_IN_ZONE
953 * page_zone is not safe to call in this context when
954 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
955 * anyway as we check zone boundaries in move_freepages_block().
956 * Remove at a later date when no bug reports exist related to
957 * grouping pages by mobility
959 BUG_ON(page_zone(start_page) != page_zone(end_page));
962 for (page = start_page; page <= end_page;) {
963 /* Make sure we are not inadvertently changing nodes */
964 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
966 if (!pfn_valid_within(page_to_pfn(page))) {
971 if (!PageBuddy(page)) {
976 order = page_order(page);
977 list_move(&page->lru,
978 &zone->free_area[order].free_list[migratetype]);
979 set_freepage_migratetype(page, migratetype);
981 pages_moved += 1 << order;
987 int move_freepages_block(struct zone *zone, struct page *page,
990 unsigned long start_pfn, end_pfn;
991 struct page *start_page, *end_page;
993 start_pfn = page_to_pfn(page);
994 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
995 start_page = pfn_to_page(start_pfn);
996 end_page = start_page + pageblock_nr_pages - 1;
997 end_pfn = start_pfn + pageblock_nr_pages - 1;
999 /* Do not cross zone boundaries */
1000 if (!zone_spans_pfn(zone, start_pfn))
1002 if (!zone_spans_pfn(zone, end_pfn))
1005 return move_freepages(zone, start_page, end_page, migratetype);
1008 static void change_pageblock_range(struct page *pageblock_page,
1009 int start_order, int migratetype)
1011 int nr_pageblocks = 1 << (start_order - pageblock_order);
1013 while (nr_pageblocks--) {
1014 set_pageblock_migratetype(pageblock_page, migratetype);
1015 pageblock_page += pageblock_nr_pages;
1019 /* Remove an element from the buddy allocator from the fallback list */
1020 static inline struct page *
1021 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1023 struct free_area * area;
1028 /* Find the largest possible block of pages in the other list */
1029 for (current_order = MAX_ORDER-1; current_order >= order;
1032 migratetype = fallbacks[start_migratetype][i];
1034 /* MIGRATE_RESERVE handled later if necessary */
1035 if (migratetype == MIGRATE_RESERVE)
1038 area = &(zone->free_area[current_order]);
1039 if (list_empty(&area->free_list[migratetype]))
1042 page = list_entry(area->free_list[migratetype].next,
1047 * If breaking a large block of pages, move all free
1048 * pages to the preferred allocation list. If falling
1049 * back for a reclaimable kernel allocation, be more
1050 * aggressive about taking ownership of free pages
1052 * On the other hand, never change migration
1053 * type of MIGRATE_CMA pageblocks nor move CMA
1054 * pages on different free lists. We don't
1055 * want unmovable pages to be allocated from
1056 * MIGRATE_CMA areas.
1058 if (!is_migrate_cma(migratetype) &&
1059 (unlikely(current_order >= pageblock_order / 2) ||
1060 start_migratetype == MIGRATE_RECLAIMABLE ||
1061 page_group_by_mobility_disabled)) {
1063 pages = move_freepages_block(zone, page,
1066 /* Claim the whole block if over half of it is free */
1067 if (pages >= (1 << (pageblock_order-1)) ||
1068 page_group_by_mobility_disabled)
1069 set_pageblock_migratetype(page,
1072 migratetype = start_migratetype;
1075 /* Remove the page from the freelists */
1076 list_del(&page->lru);
1077 rmv_page_order(page);
1079 /* Take ownership for orders >= pageblock_order */
1080 if (current_order >= pageblock_order &&
1081 !is_migrate_cma(migratetype))
1082 change_pageblock_range(page, current_order,
1085 expand(zone, page, order, current_order, area,
1086 is_migrate_cma(migratetype)
1087 ? migratetype : start_migratetype);
1089 trace_mm_page_alloc_extfrag(page, order, current_order,
1090 start_migratetype, migratetype);
1100 * Do the hard work of removing an element from the buddy allocator.
1101 * Call me with the zone->lock already held.
1103 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1109 page = __rmqueue_smallest(zone, order, migratetype);
1111 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1112 page = __rmqueue_fallback(zone, order, migratetype);
1115 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1116 * is used because __rmqueue_smallest is an inline function
1117 * and we want just one call site
1120 migratetype = MIGRATE_RESERVE;
1125 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1130 * Obtain a specified number of elements from the buddy allocator, all under
1131 * a single hold of the lock, for efficiency. Add them to the supplied list.
1132 * Returns the number of new pages which were placed at *list.
1134 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1135 unsigned long count, struct list_head *list,
1136 int migratetype, int cold)
1138 int mt = migratetype, i;
1140 spin_lock(&zone->lock);
1141 for (i = 0; i < count; ++i) {
1142 struct page *page = __rmqueue(zone, order, migratetype);
1143 if (unlikely(page == NULL))
1147 * Split buddy pages returned by expand() are received here
1148 * in physical page order. The page is added to the callers and
1149 * list and the list head then moves forward. From the callers
1150 * perspective, the linked list is ordered by page number in
1151 * some conditions. This is useful for IO devices that can
1152 * merge IO requests if the physical pages are ordered
1155 if (likely(cold == 0))
1156 list_add(&page->lru, list);
1158 list_add_tail(&page->lru, list);
1159 if (IS_ENABLED(CONFIG_CMA)) {
1160 mt = get_pageblock_migratetype(page);
1161 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1164 set_freepage_migratetype(page, mt);
1166 if (is_migrate_cma(mt))
1167 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1170 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1171 spin_unlock(&zone->lock);
1177 * Called from the vmstat counter updater to drain pagesets of this
1178 * currently executing processor on remote nodes after they have
1181 * Note that this function must be called with the thread pinned to
1182 * a single processor.
1184 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1186 unsigned long flags;
1188 unsigned long batch;
1190 local_irq_save(flags);
1191 batch = ACCESS_ONCE(pcp->batch);
1192 if (pcp->count >= batch)
1195 to_drain = pcp->count;
1197 free_pcppages_bulk(zone, to_drain, pcp);
1198 pcp->count -= to_drain;
1200 local_irq_restore(flags);
1205 * Drain pages of the indicated processor.
1207 * The processor must either be the current processor and the
1208 * thread pinned to the current processor or a processor that
1211 static void drain_pages(unsigned int cpu)
1213 unsigned long flags;
1216 for_each_populated_zone(zone) {
1217 struct per_cpu_pageset *pset;
1218 struct per_cpu_pages *pcp;
1220 local_irq_save(flags);
1221 pset = per_cpu_ptr(zone->pageset, cpu);
1225 free_pcppages_bulk(zone, pcp->count, pcp);
1228 local_irq_restore(flags);
1233 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1235 void drain_local_pages(void *arg)
1237 drain_pages(smp_processor_id());
1241 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1243 * Note that this code is protected against sending an IPI to an offline
1244 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1245 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1246 * nothing keeps CPUs from showing up after we populated the cpumask and
1247 * before the call to on_each_cpu_mask().
1249 void drain_all_pages(void)
1252 struct per_cpu_pageset *pcp;
1256 * Allocate in the BSS so we wont require allocation in
1257 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1259 static cpumask_t cpus_with_pcps;
1262 * We don't care about racing with CPU hotplug event
1263 * as offline notification will cause the notified
1264 * cpu to drain that CPU pcps and on_each_cpu_mask
1265 * disables preemption as part of its processing
1267 for_each_online_cpu(cpu) {
1268 bool has_pcps = false;
1269 for_each_populated_zone(zone) {
1270 pcp = per_cpu_ptr(zone->pageset, cpu);
1271 if (pcp->pcp.count) {
1277 cpumask_set_cpu(cpu, &cpus_with_pcps);
1279 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1281 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1284 #ifdef CONFIG_HIBERNATION
1286 void mark_free_pages(struct zone *zone)
1288 unsigned long pfn, max_zone_pfn;
1289 unsigned long flags;
1291 struct list_head *curr;
1293 if (!zone->spanned_pages)
1296 spin_lock_irqsave(&zone->lock, flags);
1298 max_zone_pfn = zone_end_pfn(zone);
1299 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1300 if (pfn_valid(pfn)) {
1301 struct page *page = pfn_to_page(pfn);
1303 if (!swsusp_page_is_forbidden(page))
1304 swsusp_unset_page_free(page);
1307 for_each_migratetype_order(order, t) {
1308 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1311 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1312 for (i = 0; i < (1UL << order); i++)
1313 swsusp_set_page_free(pfn_to_page(pfn + i));
1316 spin_unlock_irqrestore(&zone->lock, flags);
1318 #endif /* CONFIG_PM */
1321 * Free a 0-order page
1322 * cold == 1 ? free a cold page : free a hot page
1324 void free_hot_cold_page(struct page *page, int cold)
1326 struct zone *zone = page_zone(page);
1327 struct per_cpu_pages *pcp;
1328 unsigned long flags;
1331 if (!free_pages_prepare(page, 0))
1334 migratetype = get_pageblock_migratetype(page);
1335 set_freepage_migratetype(page, migratetype);
1336 local_irq_save(flags);
1337 __count_vm_event(PGFREE);
1340 * We only track unmovable, reclaimable and movable on pcp lists.
1341 * Free ISOLATE pages back to the allocator because they are being
1342 * offlined but treat RESERVE as movable pages so we can get those
1343 * areas back if necessary. Otherwise, we may have to free
1344 * excessively into the page allocator
1346 if (migratetype >= MIGRATE_PCPTYPES) {
1347 if (unlikely(is_migrate_isolate(migratetype))) {
1348 free_one_page(zone, page, 0, migratetype);
1351 migratetype = MIGRATE_MOVABLE;
1354 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1356 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1358 list_add(&page->lru, &pcp->lists[migratetype]);
1360 if (pcp->count >= pcp->high) {
1361 unsigned long batch = ACCESS_ONCE(pcp->batch);
1362 free_pcppages_bulk(zone, batch, pcp);
1363 pcp->count -= batch;
1367 local_irq_restore(flags);
1371 * Free a list of 0-order pages
1373 void free_hot_cold_page_list(struct list_head *list, int cold)
1375 struct page *page, *next;
1377 list_for_each_entry_safe(page, next, list, lru) {
1378 trace_mm_page_free_batched(page, cold);
1379 free_hot_cold_page(page, cold);
1384 * split_page takes a non-compound higher-order page, and splits it into
1385 * n (1<<order) sub-pages: page[0..n]
1386 * Each sub-page must be freed individually.
1388 * Note: this is probably too low level an operation for use in drivers.
1389 * Please consult with lkml before using this in your driver.
1391 void split_page(struct page *page, unsigned int order)
1395 VM_BUG_ON(PageCompound(page));
1396 VM_BUG_ON(!page_count(page));
1398 #ifdef CONFIG_KMEMCHECK
1400 * Split shadow pages too, because free(page[0]) would
1401 * otherwise free the whole shadow.
1403 if (kmemcheck_page_is_tracked(page))
1404 split_page(virt_to_page(page[0].shadow), order);
1407 for (i = 1; i < (1 << order); i++)
1408 set_page_refcounted(page + i);
1410 EXPORT_SYMBOL_GPL(split_page);
1412 static int __isolate_free_page(struct page *page, unsigned int order)
1414 unsigned long watermark;
1418 BUG_ON(!PageBuddy(page));
1420 zone = page_zone(page);
1421 mt = get_pageblock_migratetype(page);
1423 if (!is_migrate_isolate(mt)) {
1424 /* Obey watermarks as if the page was being allocated */
1425 watermark = low_wmark_pages(zone) + (1 << order);
1426 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1429 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1432 /* Remove page from free list */
1433 list_del(&page->lru);
1434 zone->free_area[order].nr_free--;
1435 rmv_page_order(page);
1437 /* Set the pageblock if the isolated page is at least a pageblock */
1438 if (order >= pageblock_order - 1) {
1439 struct page *endpage = page + (1 << order) - 1;
1440 for (; page < endpage; page += pageblock_nr_pages) {
1441 int mt = get_pageblock_migratetype(page);
1442 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1443 set_pageblock_migratetype(page,
1448 return 1UL << order;
1452 * Similar to split_page except the page is already free. As this is only
1453 * being used for migration, the migratetype of the block also changes.
1454 * As this is called with interrupts disabled, the caller is responsible
1455 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1458 * Note: this is probably too low level an operation for use in drivers.
1459 * Please consult with lkml before using this in your driver.
1461 int split_free_page(struct page *page)
1466 order = page_order(page);
1468 nr_pages = __isolate_free_page(page, order);
1472 /* Split into individual pages */
1473 set_page_refcounted(page);
1474 split_page(page, order);
1479 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1480 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1484 struct page *buffered_rmqueue(struct zone *preferred_zone,
1485 struct zone *zone, int order, gfp_t gfp_flags,
1488 unsigned long flags;
1490 int cold = !!(gfp_flags & __GFP_COLD);
1493 if (likely(order == 0)) {
1494 struct per_cpu_pages *pcp;
1495 struct list_head *list;
1497 local_irq_save(flags);
1498 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1499 list = &pcp->lists[migratetype];
1500 if (list_empty(list)) {
1501 pcp->count += rmqueue_bulk(zone, 0,
1504 if (unlikely(list_empty(list)))
1509 page = list_entry(list->prev, struct page, lru);
1511 page = list_entry(list->next, struct page, lru);
1513 list_del(&page->lru);
1516 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1518 * __GFP_NOFAIL is not to be used in new code.
1520 * All __GFP_NOFAIL callers should be fixed so that they
1521 * properly detect and handle allocation failures.
1523 * We most definitely don't want callers attempting to
1524 * allocate greater than order-1 page units with
1527 WARN_ON_ONCE(order > 1);
1529 spin_lock_irqsave(&zone->lock, flags);
1530 page = __rmqueue(zone, order, migratetype);
1531 spin_unlock(&zone->lock);
1534 __mod_zone_freepage_state(zone, -(1 << order),
1535 get_pageblock_migratetype(page));
1538 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1539 zone_statistics(preferred_zone, zone, gfp_flags);
1540 local_irq_restore(flags);
1542 VM_BUG_ON(bad_range(zone, page));
1543 if (prep_new_page(page, order, gfp_flags))
1548 local_irq_restore(flags);
1552 #ifdef CONFIG_FAIL_PAGE_ALLOC
1555 struct fault_attr attr;
1557 u32 ignore_gfp_highmem;
1558 u32 ignore_gfp_wait;
1560 } fail_page_alloc = {
1561 .attr = FAULT_ATTR_INITIALIZER,
1562 .ignore_gfp_wait = 1,
1563 .ignore_gfp_highmem = 1,
1567 static int __init setup_fail_page_alloc(char *str)
1569 return setup_fault_attr(&fail_page_alloc.attr, str);
1571 __setup("fail_page_alloc=", setup_fail_page_alloc);
1573 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1575 if (order < fail_page_alloc.min_order)
1577 if (gfp_mask & __GFP_NOFAIL)
1579 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1581 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1584 return should_fail(&fail_page_alloc.attr, 1 << order);
1587 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1589 static int __init fail_page_alloc_debugfs(void)
1591 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1594 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1595 &fail_page_alloc.attr);
1597 return PTR_ERR(dir);
1599 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1600 &fail_page_alloc.ignore_gfp_wait))
1602 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1603 &fail_page_alloc.ignore_gfp_highmem))
1605 if (!debugfs_create_u32("min-order", mode, dir,
1606 &fail_page_alloc.min_order))
1611 debugfs_remove_recursive(dir);
1616 late_initcall(fail_page_alloc_debugfs);
1618 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1620 #else /* CONFIG_FAIL_PAGE_ALLOC */
1622 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1627 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1630 * Return true if free pages are above 'mark'. This takes into account the order
1631 * of the allocation.
1633 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1634 int classzone_idx, int alloc_flags, long free_pages)
1636 /* free_pages my go negative - that's OK */
1638 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1642 free_pages -= (1 << order) - 1;
1643 if (alloc_flags & ALLOC_HIGH)
1645 if (alloc_flags & ALLOC_HARDER)
1648 /* If allocation can't use CMA areas don't use free CMA pages */
1649 if (!(alloc_flags & ALLOC_CMA))
1650 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1653 if (free_pages - free_cma <= min + lowmem_reserve)
1655 for (o = 0; o < order; o++) {
1656 /* At the next order, this order's pages become unavailable */
1657 free_pages -= z->free_area[o].nr_free << o;
1659 /* Require fewer higher order pages to be free */
1662 if (free_pages <= min)
1668 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1669 int classzone_idx, int alloc_flags)
1671 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1672 zone_page_state(z, NR_FREE_PAGES));
1675 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1676 int classzone_idx, int alloc_flags)
1678 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1680 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1681 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1683 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1689 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1690 * skip over zones that are not allowed by the cpuset, or that have
1691 * been recently (in last second) found to be nearly full. See further
1692 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1693 * that have to skip over a lot of full or unallowed zones.
1695 * If the zonelist cache is present in the passed in zonelist, then
1696 * returns a pointer to the allowed node mask (either the current
1697 * tasks mems_allowed, or node_states[N_MEMORY].)
1699 * If the zonelist cache is not available for this zonelist, does
1700 * nothing and returns NULL.
1702 * If the fullzones BITMAP in the zonelist cache is stale (more than
1703 * a second since last zap'd) then we zap it out (clear its bits.)
1705 * We hold off even calling zlc_setup, until after we've checked the
1706 * first zone in the zonelist, on the theory that most allocations will
1707 * be satisfied from that first zone, so best to examine that zone as
1708 * quickly as we can.
1710 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1712 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1713 nodemask_t *allowednodes; /* zonelist_cache approximation */
1715 zlc = zonelist->zlcache_ptr;
1719 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1720 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1721 zlc->last_full_zap = jiffies;
1724 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1725 &cpuset_current_mems_allowed :
1726 &node_states[N_MEMORY];
1727 return allowednodes;
1731 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1732 * if it is worth looking at further for free memory:
1733 * 1) Check that the zone isn't thought to be full (doesn't have its
1734 * bit set in the zonelist_cache fullzones BITMAP).
1735 * 2) Check that the zones node (obtained from the zonelist_cache
1736 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1737 * Return true (non-zero) if zone is worth looking at further, or
1738 * else return false (zero) if it is not.
1740 * This check -ignores- the distinction between various watermarks,
1741 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1742 * found to be full for any variation of these watermarks, it will
1743 * be considered full for up to one second by all requests, unless
1744 * we are so low on memory on all allowed nodes that we are forced
1745 * into the second scan of the zonelist.
1747 * In the second scan we ignore this zonelist cache and exactly
1748 * apply the watermarks to all zones, even it is slower to do so.
1749 * We are low on memory in the second scan, and should leave no stone
1750 * unturned looking for a free page.
1752 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1753 nodemask_t *allowednodes)
1755 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1756 int i; /* index of *z in zonelist zones */
1757 int n; /* node that zone *z is on */
1759 zlc = zonelist->zlcache_ptr;
1763 i = z - zonelist->_zonerefs;
1766 /* This zone is worth trying if it is allowed but not full */
1767 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1771 * Given 'z' scanning a zonelist, set the corresponding bit in
1772 * zlc->fullzones, so that subsequent attempts to allocate a page
1773 * from that zone don't waste time re-examining it.
1775 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1777 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1778 int i; /* index of *z in zonelist zones */
1780 zlc = zonelist->zlcache_ptr;
1784 i = z - zonelist->_zonerefs;
1786 set_bit(i, zlc->fullzones);
1790 * clear all zones full, called after direct reclaim makes progress so that
1791 * a zone that was recently full is not skipped over for up to a second
1793 static void zlc_clear_zones_full(struct zonelist *zonelist)
1795 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1797 zlc = zonelist->zlcache_ptr;
1801 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1804 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1806 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1809 static void __paginginit init_zone_allows_reclaim(int nid)
1813 for_each_online_node(i)
1814 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1815 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1817 zone_reclaim_mode = 1;
1820 #else /* CONFIG_NUMA */
1822 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1827 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1828 nodemask_t *allowednodes)
1833 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1837 static void zlc_clear_zones_full(struct zonelist *zonelist)
1841 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1846 static inline void init_zone_allows_reclaim(int nid)
1849 #endif /* CONFIG_NUMA */
1852 * get_page_from_freelist goes through the zonelist trying to allocate
1855 static struct page *
1856 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1857 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1858 struct zone *preferred_zone, int migratetype)
1861 struct page *page = NULL;
1864 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1865 int zlc_active = 0; /* set if using zonelist_cache */
1866 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1868 classzone_idx = zone_idx(preferred_zone);
1871 * Scan zonelist, looking for a zone with enough free.
1872 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1874 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1875 high_zoneidx, nodemask) {
1876 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1877 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1879 if ((alloc_flags & ALLOC_CPUSET) &&
1880 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1883 * When allocating a page cache page for writing, we
1884 * want to get it from a zone that is within its dirty
1885 * limit, such that no single zone holds more than its
1886 * proportional share of globally allowed dirty pages.
1887 * The dirty limits take into account the zone's
1888 * lowmem reserves and high watermark so that kswapd
1889 * should be able to balance it without having to
1890 * write pages from its LRU list.
1892 * This may look like it could increase pressure on
1893 * lower zones by failing allocations in higher zones
1894 * before they are full. But the pages that do spill
1895 * over are limited as the lower zones are protected
1896 * by this very same mechanism. It should not become
1897 * a practical burden to them.
1899 * XXX: For now, allow allocations to potentially
1900 * exceed the per-zone dirty limit in the slowpath
1901 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1902 * which is important when on a NUMA setup the allowed
1903 * zones are together not big enough to reach the
1904 * global limit. The proper fix for these situations
1905 * will require awareness of zones in the
1906 * dirty-throttling and the flusher threads.
1908 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1909 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1910 goto this_zone_full;
1912 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1913 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1917 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1918 if (zone_watermark_ok(zone, order, mark,
1919 classzone_idx, alloc_flags))
1922 if (IS_ENABLED(CONFIG_NUMA) &&
1923 !did_zlc_setup && nr_online_nodes > 1) {
1925 * we do zlc_setup if there are multiple nodes
1926 * and before considering the first zone allowed
1929 allowednodes = zlc_setup(zonelist, alloc_flags);
1934 if (zone_reclaim_mode == 0 ||
1935 !zone_allows_reclaim(preferred_zone, zone))
1936 goto this_zone_full;
1939 * As we may have just activated ZLC, check if the first
1940 * eligible zone has failed zone_reclaim recently.
1942 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1943 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1946 ret = zone_reclaim(zone, gfp_mask, order);
1948 case ZONE_RECLAIM_NOSCAN:
1951 case ZONE_RECLAIM_FULL:
1952 /* scanned but unreclaimable */
1955 /* did we reclaim enough */
1956 if (zone_watermark_ok(zone, order, mark,
1957 classzone_idx, alloc_flags))
1961 * Failed to reclaim enough to meet watermark.
1962 * Only mark the zone full if checking the min
1963 * watermark or if we failed to reclaim just
1964 * 1<<order pages or else the page allocator
1965 * fastpath will prematurely mark zones full
1966 * when the watermark is between the low and
1969 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
1970 ret == ZONE_RECLAIM_SOME)
1971 goto this_zone_full;
1978 page = buffered_rmqueue(preferred_zone, zone, order,
1979 gfp_mask, migratetype);
1983 if (IS_ENABLED(CONFIG_NUMA))
1984 zlc_mark_zone_full(zonelist, z);
1987 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1988 /* Disable zlc cache for second zonelist scan */
1995 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1996 * necessary to allocate the page. The expectation is
1997 * that the caller is taking steps that will free more
1998 * memory. The caller should avoid the page being used
1999 * for !PFMEMALLOC purposes.
2001 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2007 * Large machines with many possible nodes should not always dump per-node
2008 * meminfo in irq context.
2010 static inline bool should_suppress_show_mem(void)
2015 ret = in_interrupt();
2020 static DEFINE_RATELIMIT_STATE(nopage_rs,
2021 DEFAULT_RATELIMIT_INTERVAL,
2022 DEFAULT_RATELIMIT_BURST);
2024 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2026 unsigned int filter = SHOW_MEM_FILTER_NODES;
2028 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2029 debug_guardpage_minorder() > 0)
2033 * Walking all memory to count page types is very expensive and should
2034 * be inhibited in non-blockable contexts.
2036 if (!(gfp_mask & __GFP_WAIT))
2037 filter |= SHOW_MEM_FILTER_PAGE_COUNT;
2040 * This documents exceptions given to allocations in certain
2041 * contexts that are allowed to allocate outside current's set
2044 if (!(gfp_mask & __GFP_NOMEMALLOC))
2045 if (test_thread_flag(TIF_MEMDIE) ||
2046 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2047 filter &= ~SHOW_MEM_FILTER_NODES;
2048 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2049 filter &= ~SHOW_MEM_FILTER_NODES;
2052 struct va_format vaf;
2055 va_start(args, fmt);
2060 pr_warn("%pV", &vaf);
2065 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2066 current->comm, order, gfp_mask);
2069 if (!should_suppress_show_mem())
2074 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2075 unsigned long did_some_progress,
2076 unsigned long pages_reclaimed)
2078 /* Do not loop if specifically requested */
2079 if (gfp_mask & __GFP_NORETRY)
2082 /* Always retry if specifically requested */
2083 if (gfp_mask & __GFP_NOFAIL)
2087 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2088 * making forward progress without invoking OOM. Suspend also disables
2089 * storage devices so kswapd will not help. Bail if we are suspending.
2091 if (!did_some_progress && pm_suspended_storage())
2095 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2096 * means __GFP_NOFAIL, but that may not be true in other
2099 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2103 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2104 * specified, then we retry until we no longer reclaim any pages
2105 * (above), or we've reclaimed an order of pages at least as
2106 * large as the allocation's order. In both cases, if the
2107 * allocation still fails, we stop retrying.
2109 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2115 static inline struct page *
2116 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2117 struct zonelist *zonelist, enum zone_type high_zoneidx,
2118 nodemask_t *nodemask, struct zone *preferred_zone,
2123 /* Acquire the OOM killer lock for the zones in zonelist */
2124 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2125 schedule_timeout_uninterruptible(1);
2130 * Go through the zonelist yet one more time, keep very high watermark
2131 * here, this is only to catch a parallel oom killing, we must fail if
2132 * we're still under heavy pressure.
2134 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2135 order, zonelist, high_zoneidx,
2136 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2137 preferred_zone, migratetype);
2141 if (!(gfp_mask & __GFP_NOFAIL)) {
2142 /* The OOM killer will not help higher order allocs */
2143 if (order > PAGE_ALLOC_COSTLY_ORDER)
2145 /* The OOM killer does not needlessly kill tasks for lowmem */
2146 if (high_zoneidx < ZONE_NORMAL)
2149 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2150 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2151 * The caller should handle page allocation failure by itself if
2152 * it specifies __GFP_THISNODE.
2153 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2155 if (gfp_mask & __GFP_THISNODE)
2158 /* Exhausted what can be done so it's blamo time */
2159 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2162 clear_zonelist_oom(zonelist, gfp_mask);
2166 #ifdef CONFIG_COMPACTION
2167 /* Try memory compaction for high-order allocations before reclaim */
2168 static struct page *
2169 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2170 struct zonelist *zonelist, enum zone_type high_zoneidx,
2171 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2172 int migratetype, bool sync_migration,
2173 bool *contended_compaction, bool *deferred_compaction,
2174 unsigned long *did_some_progress)
2179 if (compaction_deferred(preferred_zone, order)) {
2180 *deferred_compaction = true;
2184 current->flags |= PF_MEMALLOC;
2185 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2186 nodemask, sync_migration,
2187 contended_compaction);
2188 current->flags &= ~PF_MEMALLOC;
2190 if (*did_some_progress != COMPACT_SKIPPED) {
2193 /* Page migration frees to the PCP lists but we want merging */
2194 drain_pages(get_cpu());
2197 page = get_page_from_freelist(gfp_mask, nodemask,
2198 order, zonelist, high_zoneidx,
2199 alloc_flags & ~ALLOC_NO_WATERMARKS,
2200 preferred_zone, migratetype);
2202 preferred_zone->compact_blockskip_flush = false;
2203 preferred_zone->compact_considered = 0;
2204 preferred_zone->compact_defer_shift = 0;
2205 if (order >= preferred_zone->compact_order_failed)
2206 preferred_zone->compact_order_failed = order + 1;
2207 count_vm_event(COMPACTSUCCESS);
2212 * It's bad if compaction run occurs and fails.
2213 * The most likely reason is that pages exist,
2214 * but not enough to satisfy watermarks.
2216 count_vm_event(COMPACTFAIL);
2219 * As async compaction considers a subset of pageblocks, only
2220 * defer if the failure was a sync compaction failure.
2223 defer_compaction(preferred_zone, order);
2231 static inline struct page *
2232 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2233 struct zonelist *zonelist, enum zone_type high_zoneidx,
2234 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2235 int migratetype, bool sync_migration,
2236 bool *contended_compaction, bool *deferred_compaction,
2237 unsigned long *did_some_progress)
2241 #endif /* CONFIG_COMPACTION */
2243 /* Perform direct synchronous page reclaim */
2245 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2246 nodemask_t *nodemask)
2248 struct reclaim_state reclaim_state;
2253 /* We now go into synchronous reclaim */
2254 cpuset_memory_pressure_bump();
2255 current->flags |= PF_MEMALLOC;
2256 lockdep_set_current_reclaim_state(gfp_mask);
2257 reclaim_state.reclaimed_slab = 0;
2258 current->reclaim_state = &reclaim_state;
2260 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2262 current->reclaim_state = NULL;
2263 lockdep_clear_current_reclaim_state();
2264 current->flags &= ~PF_MEMALLOC;
2271 /* The really slow allocator path where we enter direct reclaim */
2272 static inline struct page *
2273 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2274 struct zonelist *zonelist, enum zone_type high_zoneidx,
2275 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2276 int migratetype, unsigned long *did_some_progress)
2278 struct page *page = NULL;
2279 bool drained = false;
2281 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2283 if (unlikely(!(*did_some_progress)))
2286 /* After successful reclaim, reconsider all zones for allocation */
2287 if (IS_ENABLED(CONFIG_NUMA))
2288 zlc_clear_zones_full(zonelist);
2291 page = get_page_from_freelist(gfp_mask, nodemask, order,
2292 zonelist, high_zoneidx,
2293 alloc_flags & ~ALLOC_NO_WATERMARKS,
2294 preferred_zone, migratetype);
2297 * If an allocation failed after direct reclaim, it could be because
2298 * pages are pinned on the per-cpu lists. Drain them and try again
2300 if (!page && !drained) {
2310 * This is called in the allocator slow-path if the allocation request is of
2311 * sufficient urgency to ignore watermarks and take other desperate measures
2313 static inline struct page *
2314 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2315 struct zonelist *zonelist, enum zone_type high_zoneidx,
2316 nodemask_t *nodemask, struct zone *preferred_zone,
2322 page = get_page_from_freelist(gfp_mask, nodemask, order,
2323 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2324 preferred_zone, migratetype);
2326 if (!page && gfp_mask & __GFP_NOFAIL)
2327 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2328 } while (!page && (gfp_mask & __GFP_NOFAIL));
2334 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2335 enum zone_type high_zoneidx,
2336 enum zone_type classzone_idx)
2341 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2342 wakeup_kswapd(zone, order, classzone_idx);
2346 gfp_to_alloc_flags(gfp_t gfp_mask)
2348 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2349 const gfp_t wait = gfp_mask & __GFP_WAIT;
2351 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2352 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2355 * The caller may dip into page reserves a bit more if the caller
2356 * cannot run direct reclaim, or if the caller has realtime scheduling
2357 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2358 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2360 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2364 * Not worth trying to allocate harder for
2365 * __GFP_NOMEMALLOC even if it can't schedule.
2367 if (!(gfp_mask & __GFP_NOMEMALLOC))
2368 alloc_flags |= ALLOC_HARDER;
2370 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2371 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2373 alloc_flags &= ~ALLOC_CPUSET;
2374 } else if (unlikely(rt_task(current)) && !in_interrupt())
2375 alloc_flags |= ALLOC_HARDER;
2377 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2378 if (gfp_mask & __GFP_MEMALLOC)
2379 alloc_flags |= ALLOC_NO_WATERMARKS;
2380 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2381 alloc_flags |= ALLOC_NO_WATERMARKS;
2382 else if (!in_interrupt() &&
2383 ((current->flags & PF_MEMALLOC) ||
2384 unlikely(test_thread_flag(TIF_MEMDIE))))
2385 alloc_flags |= ALLOC_NO_WATERMARKS;
2388 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2389 alloc_flags |= ALLOC_CMA;
2394 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2396 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2399 static inline struct page *
2400 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2401 struct zonelist *zonelist, enum zone_type high_zoneidx,
2402 nodemask_t *nodemask, struct zone *preferred_zone,
2405 const gfp_t wait = gfp_mask & __GFP_WAIT;
2406 struct page *page = NULL;
2408 unsigned long pages_reclaimed = 0;
2409 unsigned long did_some_progress;
2410 bool sync_migration = false;
2411 bool deferred_compaction = false;
2412 bool contended_compaction = false;
2415 * In the slowpath, we sanity check order to avoid ever trying to
2416 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2417 * be using allocators in order of preference for an area that is
2420 if (order >= MAX_ORDER) {
2421 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2426 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2427 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2428 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2429 * using a larger set of nodes after it has established that the
2430 * allowed per node queues are empty and that nodes are
2433 if (IS_ENABLED(CONFIG_NUMA) &&
2434 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2438 if (!(gfp_mask & __GFP_NO_KSWAPD))
2439 wake_all_kswapd(order, zonelist, high_zoneidx,
2440 zone_idx(preferred_zone));
2443 * OK, we're below the kswapd watermark and have kicked background
2444 * reclaim. Now things get more complex, so set up alloc_flags according
2445 * to how we want to proceed.
2447 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2450 * Find the true preferred zone if the allocation is unconstrained by
2453 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2454 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2458 /* This is the last chance, in general, before the goto nopage. */
2459 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2460 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2461 preferred_zone, migratetype);
2465 /* Allocate without watermarks if the context allows */
2466 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2468 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2469 * the allocation is high priority and these type of
2470 * allocations are system rather than user orientated
2472 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2474 page = __alloc_pages_high_priority(gfp_mask, order,
2475 zonelist, high_zoneidx, nodemask,
2476 preferred_zone, migratetype);
2482 /* Atomic allocations - we can't balance anything */
2486 /* Avoid recursion of direct reclaim */
2487 if (current->flags & PF_MEMALLOC)
2490 /* Avoid allocations with no watermarks from looping endlessly */
2491 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2495 * Try direct compaction. The first pass is asynchronous. Subsequent
2496 * attempts after direct reclaim are synchronous
2498 page = __alloc_pages_direct_compact(gfp_mask, order,
2499 zonelist, high_zoneidx,
2501 alloc_flags, preferred_zone,
2502 migratetype, sync_migration,
2503 &contended_compaction,
2504 &deferred_compaction,
2505 &did_some_progress);
2508 sync_migration = true;
2511 * If compaction is deferred for high-order allocations, it is because
2512 * sync compaction recently failed. In this is the case and the caller
2513 * requested a movable allocation that does not heavily disrupt the
2514 * system then fail the allocation instead of entering direct reclaim.
2516 if ((deferred_compaction || contended_compaction) &&
2517 (gfp_mask & __GFP_NO_KSWAPD))
2520 /* Try direct reclaim and then allocating */
2521 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2522 zonelist, high_zoneidx,
2524 alloc_flags, preferred_zone,
2525 migratetype, &did_some_progress);
2530 * If we failed to make any progress reclaiming, then we are
2531 * running out of options and have to consider going OOM
2533 if (!did_some_progress) {
2534 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2535 if (oom_killer_disabled)
2537 /* Coredumps can quickly deplete all memory reserves */
2538 if ((current->flags & PF_DUMPCORE) &&
2539 !(gfp_mask & __GFP_NOFAIL))
2541 page = __alloc_pages_may_oom(gfp_mask, order,
2542 zonelist, high_zoneidx,
2543 nodemask, preferred_zone,
2548 if (!(gfp_mask & __GFP_NOFAIL)) {
2550 * The oom killer is not called for high-order
2551 * allocations that may fail, so if no progress
2552 * is being made, there are no other options and
2553 * retrying is unlikely to help.
2555 if (order > PAGE_ALLOC_COSTLY_ORDER)
2558 * The oom killer is not called for lowmem
2559 * allocations to prevent needlessly killing
2562 if (high_zoneidx < ZONE_NORMAL)
2570 /* Check if we should retry the allocation */
2571 pages_reclaimed += did_some_progress;
2572 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2574 /* Wait for some write requests to complete then retry */
2575 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2579 * High-order allocations do not necessarily loop after
2580 * direct reclaim and reclaim/compaction depends on compaction
2581 * being called after reclaim so call directly if necessary
2583 page = __alloc_pages_direct_compact(gfp_mask, order,
2584 zonelist, high_zoneidx,
2586 alloc_flags, preferred_zone,
2587 migratetype, sync_migration,
2588 &contended_compaction,
2589 &deferred_compaction,
2590 &did_some_progress);
2596 warn_alloc_failed(gfp_mask, order, NULL);
2599 if (kmemcheck_enabled)
2600 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2606 * This is the 'heart' of the zoned buddy allocator.
2609 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2610 struct zonelist *zonelist, nodemask_t *nodemask)
2612 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2613 struct zone *preferred_zone;
2614 struct page *page = NULL;
2615 int migratetype = allocflags_to_migratetype(gfp_mask);
2616 unsigned int cpuset_mems_cookie;
2617 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2618 struct mem_cgroup *memcg = NULL;
2620 gfp_mask &= gfp_allowed_mask;
2622 lockdep_trace_alloc(gfp_mask);
2624 might_sleep_if(gfp_mask & __GFP_WAIT);
2626 if (should_fail_alloc_page(gfp_mask, order))
2630 * Check the zones suitable for the gfp_mask contain at least one
2631 * valid zone. It's possible to have an empty zonelist as a result
2632 * of GFP_THISNODE and a memoryless node
2634 if (unlikely(!zonelist->_zonerefs->zone))
2638 * Will only have any effect when __GFP_KMEMCG is set. This is
2639 * verified in the (always inline) callee
2641 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2645 cpuset_mems_cookie = get_mems_allowed();
2647 /* The preferred zone is used for statistics later */
2648 first_zones_zonelist(zonelist, high_zoneidx,
2649 nodemask ? : &cpuset_current_mems_allowed,
2651 if (!preferred_zone)
2655 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2656 alloc_flags |= ALLOC_CMA;
2658 /* First allocation attempt */
2659 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2660 zonelist, high_zoneidx, alloc_flags,
2661 preferred_zone, migratetype);
2662 if (unlikely(!page)) {
2664 * Runtime PM, block IO and its error handling path
2665 * can deadlock because I/O on the device might not
2668 gfp_mask = memalloc_noio_flags(gfp_mask);
2669 page = __alloc_pages_slowpath(gfp_mask, order,
2670 zonelist, high_zoneidx, nodemask,
2671 preferred_zone, migratetype);
2674 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2678 * When updating a task's mems_allowed, it is possible to race with
2679 * parallel threads in such a way that an allocation can fail while
2680 * the mask is being updated. If a page allocation is about to fail,
2681 * check if the cpuset changed during allocation and if so, retry.
2683 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2686 memcg_kmem_commit_charge(page, memcg, order);
2690 EXPORT_SYMBOL(__alloc_pages_nodemask);
2693 * Common helper functions.
2695 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2700 * __get_free_pages() returns a 32-bit address, which cannot represent
2703 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2705 page = alloc_pages(gfp_mask, order);
2708 return (unsigned long) page_address(page);
2710 EXPORT_SYMBOL(__get_free_pages);
2712 unsigned long get_zeroed_page(gfp_t gfp_mask)
2714 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2716 EXPORT_SYMBOL(get_zeroed_page);
2718 void __free_pages(struct page *page, unsigned int order)
2720 if (put_page_testzero(page)) {
2722 free_hot_cold_page(page, 0);
2724 __free_pages_ok(page, order);
2728 EXPORT_SYMBOL(__free_pages);
2730 void free_pages(unsigned long addr, unsigned int order)
2733 VM_BUG_ON(!virt_addr_valid((void *)addr));
2734 __free_pages(virt_to_page((void *)addr), order);
2738 EXPORT_SYMBOL(free_pages);
2741 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2742 * pages allocated with __GFP_KMEMCG.
2744 * Those pages are accounted to a particular memcg, embedded in the
2745 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2746 * for that information only to find out that it is NULL for users who have no
2747 * interest in that whatsoever, we provide these functions.
2749 * The caller knows better which flags it relies on.
2751 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2753 memcg_kmem_uncharge_pages(page, order);
2754 __free_pages(page, order);
2757 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2760 VM_BUG_ON(!virt_addr_valid((void *)addr));
2761 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2765 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2768 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2769 unsigned long used = addr + PAGE_ALIGN(size);
2771 split_page(virt_to_page((void *)addr), order);
2772 while (used < alloc_end) {
2777 return (void *)addr;
2781 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2782 * @size: the number of bytes to allocate
2783 * @gfp_mask: GFP flags for the allocation
2785 * This function is similar to alloc_pages(), except that it allocates the
2786 * minimum number of pages to satisfy the request. alloc_pages() can only
2787 * allocate memory in power-of-two pages.
2789 * This function is also limited by MAX_ORDER.
2791 * Memory allocated by this function must be released by free_pages_exact().
2793 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2795 unsigned int order = get_order(size);
2798 addr = __get_free_pages(gfp_mask, order);
2799 return make_alloc_exact(addr, order, size);
2801 EXPORT_SYMBOL(alloc_pages_exact);
2804 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2806 * @nid: the preferred node ID where memory should be allocated
2807 * @size: the number of bytes to allocate
2808 * @gfp_mask: GFP flags for the allocation
2810 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2812 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2815 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2817 unsigned order = get_order(size);
2818 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2821 return make_alloc_exact((unsigned long)page_address(p), order, size);
2823 EXPORT_SYMBOL(alloc_pages_exact_nid);
2826 * free_pages_exact - release memory allocated via alloc_pages_exact()
2827 * @virt: the value returned by alloc_pages_exact.
2828 * @size: size of allocation, same value as passed to alloc_pages_exact().
2830 * Release the memory allocated by a previous call to alloc_pages_exact.
2832 void free_pages_exact(void *virt, size_t size)
2834 unsigned long addr = (unsigned long)virt;
2835 unsigned long end = addr + PAGE_ALIGN(size);
2837 while (addr < end) {
2842 EXPORT_SYMBOL(free_pages_exact);
2845 * nr_free_zone_pages - count number of pages beyond high watermark
2846 * @offset: The zone index of the highest zone
2848 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2849 * high watermark within all zones at or below a given zone index. For each
2850 * zone, the number of pages is calculated as:
2851 * managed_pages - high_pages
2853 static unsigned long nr_free_zone_pages(int offset)
2858 /* Just pick one node, since fallback list is circular */
2859 unsigned long sum = 0;
2861 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2863 for_each_zone_zonelist(zone, z, zonelist, offset) {
2864 unsigned long size = zone->managed_pages;
2865 unsigned long high = high_wmark_pages(zone);
2874 * nr_free_buffer_pages - count number of pages beyond high watermark
2876 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2877 * watermark within ZONE_DMA and ZONE_NORMAL.
2879 unsigned long nr_free_buffer_pages(void)
2881 return nr_free_zone_pages(gfp_zone(GFP_USER));
2883 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2886 * nr_free_pagecache_pages - count number of pages beyond high watermark
2888 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2889 * high watermark within all zones.
2891 unsigned long nr_free_pagecache_pages(void)
2893 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2896 static inline void show_node(struct zone *zone)
2898 if (IS_ENABLED(CONFIG_NUMA))
2899 printk("Node %d ", zone_to_nid(zone));
2902 void si_meminfo(struct sysinfo *val)
2904 val->totalram = totalram_pages;
2906 val->freeram = global_page_state(NR_FREE_PAGES);
2907 val->bufferram = nr_blockdev_pages();
2908 val->totalhigh = totalhigh_pages;
2909 val->freehigh = nr_free_highpages();
2910 val->mem_unit = PAGE_SIZE;
2913 EXPORT_SYMBOL(si_meminfo);
2916 void si_meminfo_node(struct sysinfo *val, int nid)
2918 pg_data_t *pgdat = NODE_DATA(nid);
2920 val->totalram = pgdat->node_present_pages;
2921 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2922 #ifdef CONFIG_HIGHMEM
2923 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
2924 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2930 val->mem_unit = PAGE_SIZE;
2935 * Determine whether the node should be displayed or not, depending on whether
2936 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2938 bool skip_free_areas_node(unsigned int flags, int nid)
2941 unsigned int cpuset_mems_cookie;
2943 if (!(flags & SHOW_MEM_FILTER_NODES))
2947 cpuset_mems_cookie = get_mems_allowed();
2948 ret = !node_isset(nid, cpuset_current_mems_allowed);
2949 } while (!put_mems_allowed(cpuset_mems_cookie));
2954 #define K(x) ((x) << (PAGE_SHIFT-10))
2956 static void show_migration_types(unsigned char type)
2958 static const char types[MIGRATE_TYPES] = {
2959 [MIGRATE_UNMOVABLE] = 'U',
2960 [MIGRATE_RECLAIMABLE] = 'E',
2961 [MIGRATE_MOVABLE] = 'M',
2962 [MIGRATE_RESERVE] = 'R',
2964 [MIGRATE_CMA] = 'C',
2966 #ifdef CONFIG_MEMORY_ISOLATION
2967 [MIGRATE_ISOLATE] = 'I',
2970 char tmp[MIGRATE_TYPES + 1];
2974 for (i = 0; i < MIGRATE_TYPES; i++) {
2975 if (type & (1 << i))
2980 printk("(%s) ", tmp);
2984 * Show free area list (used inside shift_scroll-lock stuff)
2985 * We also calculate the percentage fragmentation. We do this by counting the
2986 * memory on each free list with the exception of the first item on the list.
2987 * Suppresses nodes that are not allowed by current's cpuset if
2988 * SHOW_MEM_FILTER_NODES is passed.
2990 void show_free_areas(unsigned int filter)
2995 for_each_populated_zone(zone) {
2996 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2999 printk("%s per-cpu:\n", zone->name);
3001 for_each_online_cpu(cpu) {
3002 struct per_cpu_pageset *pageset;
3004 pageset = per_cpu_ptr(zone->pageset, cpu);
3006 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3007 cpu, pageset->pcp.high,
3008 pageset->pcp.batch, pageset->pcp.count);
3012 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3013 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3015 " dirty:%lu writeback:%lu unstable:%lu\n"
3016 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3017 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3019 global_page_state(NR_ACTIVE_ANON),
3020 global_page_state(NR_INACTIVE_ANON),
3021 global_page_state(NR_ISOLATED_ANON),
3022 global_page_state(NR_ACTIVE_FILE),
3023 global_page_state(NR_INACTIVE_FILE),
3024 global_page_state(NR_ISOLATED_FILE),
3025 global_page_state(NR_UNEVICTABLE),
3026 global_page_state(NR_FILE_DIRTY),
3027 global_page_state(NR_WRITEBACK),
3028 global_page_state(NR_UNSTABLE_NFS),
3029 global_page_state(NR_FREE_PAGES),
3030 global_page_state(NR_SLAB_RECLAIMABLE),
3031 global_page_state(NR_SLAB_UNRECLAIMABLE),
3032 global_page_state(NR_FILE_MAPPED),
3033 global_page_state(NR_SHMEM),
3034 global_page_state(NR_PAGETABLE),
3035 global_page_state(NR_BOUNCE),
3036 global_page_state(NR_FREE_CMA_PAGES));
3038 for_each_populated_zone(zone) {
3041 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3049 " active_anon:%lukB"
3050 " inactive_anon:%lukB"
3051 " active_file:%lukB"
3052 " inactive_file:%lukB"
3053 " unevictable:%lukB"
3054 " isolated(anon):%lukB"
3055 " isolated(file):%lukB"
3063 " slab_reclaimable:%lukB"
3064 " slab_unreclaimable:%lukB"
3065 " kernel_stack:%lukB"
3070 " writeback_tmp:%lukB"
3071 " pages_scanned:%lu"
3072 " all_unreclaimable? %s"
3075 K(zone_page_state(zone, NR_FREE_PAGES)),
3076 K(min_wmark_pages(zone)),
3077 K(low_wmark_pages(zone)),
3078 K(high_wmark_pages(zone)),
3079 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3080 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3081 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3082 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3083 K(zone_page_state(zone, NR_UNEVICTABLE)),
3084 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3085 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3086 K(zone->present_pages),
3087 K(zone->managed_pages),
3088 K(zone_page_state(zone, NR_MLOCK)),
3089 K(zone_page_state(zone, NR_FILE_DIRTY)),
3090 K(zone_page_state(zone, NR_WRITEBACK)),
3091 K(zone_page_state(zone, NR_FILE_MAPPED)),
3092 K(zone_page_state(zone, NR_SHMEM)),
3093 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3094 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3095 zone_page_state(zone, NR_KERNEL_STACK) *
3097 K(zone_page_state(zone, NR_PAGETABLE)),
3098 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3099 K(zone_page_state(zone, NR_BOUNCE)),
3100 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3101 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3102 zone->pages_scanned,
3103 (zone->all_unreclaimable ? "yes" : "no")
3105 printk("lowmem_reserve[]:");
3106 for (i = 0; i < MAX_NR_ZONES; i++)
3107 printk(" %lu", zone->lowmem_reserve[i]);
3111 for_each_populated_zone(zone) {
3112 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3113 unsigned char types[MAX_ORDER];
3115 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3118 printk("%s: ", zone->name);
3120 spin_lock_irqsave(&zone->lock, flags);
3121 for (order = 0; order < MAX_ORDER; order++) {
3122 struct free_area *area = &zone->free_area[order];
3125 nr[order] = area->nr_free;
3126 total += nr[order] << order;
3129 for (type = 0; type < MIGRATE_TYPES; type++) {
3130 if (!list_empty(&area->free_list[type]))
3131 types[order] |= 1 << type;
3134 spin_unlock_irqrestore(&zone->lock, flags);
3135 for (order = 0; order < MAX_ORDER; order++) {
3136 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3138 show_migration_types(types[order]);
3140 printk("= %lukB\n", K(total));
3143 hugetlb_show_meminfo();
3145 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3147 show_swap_cache_info();
3150 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3152 zoneref->zone = zone;
3153 zoneref->zone_idx = zone_idx(zone);
3157 * Builds allocation fallback zone lists.
3159 * Add all populated zones of a node to the zonelist.
3161 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3162 int nr_zones, enum zone_type zone_type)
3166 BUG_ON(zone_type >= MAX_NR_ZONES);
3171 zone = pgdat->node_zones + zone_type;
3172 if (populated_zone(zone)) {
3173 zoneref_set_zone(zone,
3174 &zonelist->_zonerefs[nr_zones++]);
3175 check_highest_zone(zone_type);
3178 } while (zone_type);
3185 * 0 = automatic detection of better ordering.
3186 * 1 = order by ([node] distance, -zonetype)
3187 * 2 = order by (-zonetype, [node] distance)
3189 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3190 * the same zonelist. So only NUMA can configure this param.
3192 #define ZONELIST_ORDER_DEFAULT 0
3193 #define ZONELIST_ORDER_NODE 1
3194 #define ZONELIST_ORDER_ZONE 2
3196 /* zonelist order in the kernel.
3197 * set_zonelist_order() will set this to NODE or ZONE.
3199 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3200 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3204 /* The value user specified ....changed by config */
3205 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3206 /* string for sysctl */
3207 #define NUMA_ZONELIST_ORDER_LEN 16
3208 char numa_zonelist_order[16] = "default";
3211 * interface for configure zonelist ordering.
3212 * command line option "numa_zonelist_order"
3213 * = "[dD]efault - default, automatic configuration.
3214 * = "[nN]ode - order by node locality, then by zone within node
3215 * = "[zZ]one - order by zone, then by locality within zone
3218 static int __parse_numa_zonelist_order(char *s)
3220 if (*s == 'd' || *s == 'D') {
3221 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3222 } else if (*s == 'n' || *s == 'N') {
3223 user_zonelist_order = ZONELIST_ORDER_NODE;
3224 } else if (*s == 'z' || *s == 'Z') {
3225 user_zonelist_order = ZONELIST_ORDER_ZONE;
3228 "Ignoring invalid numa_zonelist_order value: "
3235 static __init int setup_numa_zonelist_order(char *s)
3242 ret = __parse_numa_zonelist_order(s);
3244 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3248 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3251 * sysctl handler for numa_zonelist_order
3253 int numa_zonelist_order_handler(ctl_table *table, int write,
3254 void __user *buffer, size_t *length,
3257 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3259 static DEFINE_MUTEX(zl_order_mutex);
3261 mutex_lock(&zl_order_mutex);
3263 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3267 strcpy(saved_string, (char *)table->data);
3269 ret = proc_dostring(table, write, buffer, length, ppos);
3273 int oldval = user_zonelist_order;
3275 ret = __parse_numa_zonelist_order((char *)table->data);
3278 * bogus value. restore saved string
3280 strncpy((char *)table->data, saved_string,
3281 NUMA_ZONELIST_ORDER_LEN);
3282 user_zonelist_order = oldval;
3283 } else if (oldval != user_zonelist_order) {
3284 mutex_lock(&zonelists_mutex);
3285 build_all_zonelists(NULL, NULL);
3286 mutex_unlock(&zonelists_mutex);
3290 mutex_unlock(&zl_order_mutex);
3295 #define MAX_NODE_LOAD (nr_online_nodes)
3296 static int node_load[MAX_NUMNODES];
3299 * find_next_best_node - find the next node that should appear in a given node's fallback list
3300 * @node: node whose fallback list we're appending
3301 * @used_node_mask: nodemask_t of already used nodes
3303 * We use a number of factors to determine which is the next node that should
3304 * appear on a given node's fallback list. The node should not have appeared
3305 * already in @node's fallback list, and it should be the next closest node
3306 * according to the distance array (which contains arbitrary distance values
3307 * from each node to each node in the system), and should also prefer nodes
3308 * with no CPUs, since presumably they'll have very little allocation pressure
3309 * on them otherwise.
3310 * It returns -1 if no node is found.
3312 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3315 int min_val = INT_MAX;
3316 int best_node = NUMA_NO_NODE;
3317 const struct cpumask *tmp = cpumask_of_node(0);
3319 /* Use the local node if we haven't already */
3320 if (!node_isset(node, *used_node_mask)) {
3321 node_set(node, *used_node_mask);
3325 for_each_node_state(n, N_MEMORY) {
3327 /* Don't want a node to appear more than once */
3328 if (node_isset(n, *used_node_mask))
3331 /* Use the distance array to find the distance */
3332 val = node_distance(node, n);
3334 /* Penalize nodes under us ("prefer the next node") */
3337 /* Give preference to headless and unused nodes */
3338 tmp = cpumask_of_node(n);
3339 if (!cpumask_empty(tmp))
3340 val += PENALTY_FOR_NODE_WITH_CPUS;
3342 /* Slight preference for less loaded node */
3343 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3344 val += node_load[n];
3346 if (val < min_val) {
3353 node_set(best_node, *used_node_mask);
3360 * Build zonelists ordered by node and zones within node.
3361 * This results in maximum locality--normal zone overflows into local
3362 * DMA zone, if any--but risks exhausting DMA zone.
3364 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3367 struct zonelist *zonelist;
3369 zonelist = &pgdat->node_zonelists[0];
3370 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3372 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3374 zonelist->_zonerefs[j].zone = NULL;
3375 zonelist->_zonerefs[j].zone_idx = 0;
3379 * Build gfp_thisnode zonelists
3381 static void build_thisnode_zonelists(pg_data_t *pgdat)
3384 struct zonelist *zonelist;
3386 zonelist = &pgdat->node_zonelists[1];
3387 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3388 zonelist->_zonerefs[j].zone = NULL;
3389 zonelist->_zonerefs[j].zone_idx = 0;
3393 * Build zonelists ordered by zone and nodes within zones.
3394 * This results in conserving DMA zone[s] until all Normal memory is
3395 * exhausted, but results in overflowing to remote node while memory
3396 * may still exist in local DMA zone.
3398 static int node_order[MAX_NUMNODES];
3400 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3403 int zone_type; /* needs to be signed */
3405 struct zonelist *zonelist;
3407 zonelist = &pgdat->node_zonelists[0];
3409 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3410 for (j = 0; j < nr_nodes; j++) {
3411 node = node_order[j];
3412 z = &NODE_DATA(node)->node_zones[zone_type];
3413 if (populated_zone(z)) {
3415 &zonelist->_zonerefs[pos++]);
3416 check_highest_zone(zone_type);
3420 zonelist->_zonerefs[pos].zone = NULL;
3421 zonelist->_zonerefs[pos].zone_idx = 0;
3424 static int default_zonelist_order(void)
3427 unsigned long low_kmem_size,total_size;
3431 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3432 * If they are really small and used heavily, the system can fall
3433 * into OOM very easily.
3434 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3436 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3439 for_each_online_node(nid) {
3440 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3441 z = &NODE_DATA(nid)->node_zones[zone_type];
3442 if (populated_zone(z)) {
3443 if (zone_type < ZONE_NORMAL)
3444 low_kmem_size += z->managed_pages;
3445 total_size += z->managed_pages;
3446 } else if (zone_type == ZONE_NORMAL) {
3448 * If any node has only lowmem, then node order
3449 * is preferred to allow kernel allocations
3450 * locally; otherwise, they can easily infringe
3451 * on other nodes when there is an abundance of
3452 * lowmem available to allocate from.
3454 return ZONELIST_ORDER_NODE;
3458 if (!low_kmem_size || /* there are no DMA area. */
3459 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3460 return ZONELIST_ORDER_NODE;
3462 * look into each node's config.
3463 * If there is a node whose DMA/DMA32 memory is very big area on
3464 * local memory, NODE_ORDER may be suitable.
3466 average_size = total_size /
3467 (nodes_weight(node_states[N_MEMORY]) + 1);
3468 for_each_online_node(nid) {
3471 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3472 z = &NODE_DATA(nid)->node_zones[zone_type];
3473 if (populated_zone(z)) {
3474 if (zone_type < ZONE_NORMAL)
3475 low_kmem_size += z->present_pages;
3476 total_size += z->present_pages;
3479 if (low_kmem_size &&
3480 total_size > average_size && /* ignore small node */
3481 low_kmem_size > total_size * 70/100)
3482 return ZONELIST_ORDER_NODE;
3484 return ZONELIST_ORDER_ZONE;
3487 static void set_zonelist_order(void)
3489 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3490 current_zonelist_order = default_zonelist_order();
3492 current_zonelist_order = user_zonelist_order;
3495 static void build_zonelists(pg_data_t *pgdat)
3499 nodemask_t used_mask;
3500 int local_node, prev_node;
3501 struct zonelist *zonelist;
3502 int order = current_zonelist_order;
3504 /* initialize zonelists */
3505 for (i = 0; i < MAX_ZONELISTS; i++) {
3506 zonelist = pgdat->node_zonelists + i;
3507 zonelist->_zonerefs[0].zone = NULL;
3508 zonelist->_zonerefs[0].zone_idx = 0;
3511 /* NUMA-aware ordering of nodes */
3512 local_node = pgdat->node_id;
3513 load = nr_online_nodes;
3514 prev_node = local_node;
3515 nodes_clear(used_mask);
3517 memset(node_order, 0, sizeof(node_order));
3520 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3522 * We don't want to pressure a particular node.
3523 * So adding penalty to the first node in same
3524 * distance group to make it round-robin.
3526 if (node_distance(local_node, node) !=
3527 node_distance(local_node, prev_node))
3528 node_load[node] = load;
3532 if (order == ZONELIST_ORDER_NODE)
3533 build_zonelists_in_node_order(pgdat, node);
3535 node_order[j++] = node; /* remember order */
3538 if (order == ZONELIST_ORDER_ZONE) {
3539 /* calculate node order -- i.e., DMA last! */
3540 build_zonelists_in_zone_order(pgdat, j);
3543 build_thisnode_zonelists(pgdat);
3546 /* Construct the zonelist performance cache - see further mmzone.h */
3547 static void build_zonelist_cache(pg_data_t *pgdat)
3549 struct zonelist *zonelist;
3550 struct zonelist_cache *zlc;
3553 zonelist = &pgdat->node_zonelists[0];
3554 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3555 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3556 for (z = zonelist->_zonerefs; z->zone; z++)
3557 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3560 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3562 * Return node id of node used for "local" allocations.
3563 * I.e., first node id of first zone in arg node's generic zonelist.
3564 * Used for initializing percpu 'numa_mem', which is used primarily
3565 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3567 int local_memory_node(int node)
3571 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3572 gfp_zone(GFP_KERNEL),
3579 #else /* CONFIG_NUMA */
3581 static void set_zonelist_order(void)
3583 current_zonelist_order = ZONELIST_ORDER_ZONE;
3586 static void build_zonelists(pg_data_t *pgdat)
3588 int node, local_node;
3590 struct zonelist *zonelist;
3592 local_node = pgdat->node_id;
3594 zonelist = &pgdat->node_zonelists[0];
3595 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3598 * Now we build the zonelist so that it contains the zones
3599 * of all the other nodes.
3600 * We don't want to pressure a particular node, so when
3601 * building the zones for node N, we make sure that the
3602 * zones coming right after the local ones are those from
3603 * node N+1 (modulo N)
3605 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3606 if (!node_online(node))
3608 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3611 for (node = 0; node < local_node; node++) {
3612 if (!node_online(node))
3614 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3618 zonelist->_zonerefs[j].zone = NULL;
3619 zonelist->_zonerefs[j].zone_idx = 0;
3622 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3623 static void build_zonelist_cache(pg_data_t *pgdat)
3625 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3628 #endif /* CONFIG_NUMA */
3631 * Boot pageset table. One per cpu which is going to be used for all
3632 * zones and all nodes. The parameters will be set in such a way
3633 * that an item put on a list will immediately be handed over to
3634 * the buddy list. This is safe since pageset manipulation is done
3635 * with interrupts disabled.
3637 * The boot_pagesets must be kept even after bootup is complete for
3638 * unused processors and/or zones. They do play a role for bootstrapping
3639 * hotplugged processors.
3641 * zoneinfo_show() and maybe other functions do
3642 * not check if the processor is online before following the pageset pointer.
3643 * Other parts of the kernel may not check if the zone is available.
3645 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3646 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3647 static void setup_zone_pageset(struct zone *zone);
3650 * Global mutex to protect against size modification of zonelists
3651 * as well as to serialize pageset setup for the new populated zone.
3653 DEFINE_MUTEX(zonelists_mutex);
3655 /* return values int ....just for stop_machine() */
3656 static int __build_all_zonelists(void *data)
3660 pg_data_t *self = data;
3663 memset(node_load, 0, sizeof(node_load));
3666 if (self && !node_online(self->node_id)) {
3667 build_zonelists(self);
3668 build_zonelist_cache(self);
3671 for_each_online_node(nid) {
3672 pg_data_t *pgdat = NODE_DATA(nid);
3674 build_zonelists(pgdat);
3675 build_zonelist_cache(pgdat);
3679 * Initialize the boot_pagesets that are going to be used
3680 * for bootstrapping processors. The real pagesets for
3681 * each zone will be allocated later when the per cpu
3682 * allocator is available.
3684 * boot_pagesets are used also for bootstrapping offline
3685 * cpus if the system is already booted because the pagesets
3686 * are needed to initialize allocators on a specific cpu too.
3687 * F.e. the percpu allocator needs the page allocator which
3688 * needs the percpu allocator in order to allocate its pagesets
3689 * (a chicken-egg dilemma).
3691 for_each_possible_cpu(cpu) {
3692 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3694 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3696 * We now know the "local memory node" for each node--
3697 * i.e., the node of the first zone in the generic zonelist.
3698 * Set up numa_mem percpu variable for on-line cpus. During
3699 * boot, only the boot cpu should be on-line; we'll init the
3700 * secondary cpus' numa_mem as they come on-line. During
3701 * node/memory hotplug, we'll fixup all on-line cpus.
3703 if (cpu_online(cpu))
3704 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3712 * Called with zonelists_mutex held always
3713 * unless system_state == SYSTEM_BOOTING.
3715 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3717 set_zonelist_order();
3719 if (system_state == SYSTEM_BOOTING) {
3720 __build_all_zonelists(NULL);
3721 mminit_verify_zonelist();
3722 cpuset_init_current_mems_allowed();
3724 #ifdef CONFIG_MEMORY_HOTPLUG
3726 setup_zone_pageset(zone);
3728 /* we have to stop all cpus to guarantee there is no user
3730 stop_machine(__build_all_zonelists, pgdat, NULL);
3731 /* cpuset refresh routine should be here */
3733 vm_total_pages = nr_free_pagecache_pages();
3735 * Disable grouping by mobility if the number of pages in the
3736 * system is too low to allow the mechanism to work. It would be
3737 * more accurate, but expensive to check per-zone. This check is
3738 * made on memory-hotadd so a system can start with mobility
3739 * disabled and enable it later
3741 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3742 page_group_by_mobility_disabled = 1;
3744 page_group_by_mobility_disabled = 0;
3746 printk("Built %i zonelists in %s order, mobility grouping %s. "
3747 "Total pages: %ld\n",
3749 zonelist_order_name[current_zonelist_order],
3750 page_group_by_mobility_disabled ? "off" : "on",
3753 printk("Policy zone: %s\n", zone_names[policy_zone]);
3758 * Helper functions to size the waitqueue hash table.
3759 * Essentially these want to choose hash table sizes sufficiently
3760 * large so that collisions trying to wait on pages are rare.
3761 * But in fact, the number of active page waitqueues on typical
3762 * systems is ridiculously low, less than 200. So this is even
3763 * conservative, even though it seems large.
3765 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3766 * waitqueues, i.e. the size of the waitq table given the number of pages.
3768 #define PAGES_PER_WAITQUEUE 256
3770 #ifndef CONFIG_MEMORY_HOTPLUG
3771 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3773 unsigned long size = 1;
3775 pages /= PAGES_PER_WAITQUEUE;
3777 while (size < pages)
3781 * Once we have dozens or even hundreds of threads sleeping
3782 * on IO we've got bigger problems than wait queue collision.
3783 * Limit the size of the wait table to a reasonable size.
3785 size = min(size, 4096UL);
3787 return max(size, 4UL);
3791 * A zone's size might be changed by hot-add, so it is not possible to determine
3792 * a suitable size for its wait_table. So we use the maximum size now.
3794 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3796 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3797 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3798 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3800 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3801 * or more by the traditional way. (See above). It equals:
3803 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3804 * ia64(16K page size) : = ( 8G + 4M)byte.
3805 * powerpc (64K page size) : = (32G +16M)byte.
3807 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3814 * This is an integer logarithm so that shifts can be used later
3815 * to extract the more random high bits from the multiplicative
3816 * hash function before the remainder is taken.
3818 static inline unsigned long wait_table_bits(unsigned long size)
3823 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3826 * Check if a pageblock contains reserved pages
3828 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3832 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3833 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3840 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3841 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3842 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3843 * higher will lead to a bigger reserve which will get freed as contiguous
3844 * blocks as reclaim kicks in
3846 static void setup_zone_migrate_reserve(struct zone *zone)
3848 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3850 unsigned long block_migratetype;
3854 * Get the start pfn, end pfn and the number of blocks to reserve
3855 * We have to be careful to be aligned to pageblock_nr_pages to
3856 * make sure that we always check pfn_valid for the first page in
3859 start_pfn = zone->zone_start_pfn;
3860 end_pfn = zone_end_pfn(zone);
3861 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3862 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3866 * Reserve blocks are generally in place to help high-order atomic
3867 * allocations that are short-lived. A min_free_kbytes value that
3868 * would result in more than 2 reserve blocks for atomic allocations
3869 * is assumed to be in place to help anti-fragmentation for the
3870 * future allocation of hugepages at runtime.
3872 reserve = min(2, reserve);
3874 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3875 if (!pfn_valid(pfn))
3877 page = pfn_to_page(pfn);
3879 /* Watch out for overlapping nodes */
3880 if (page_to_nid(page) != zone_to_nid(zone))
3883 block_migratetype = get_pageblock_migratetype(page);
3885 /* Only test what is necessary when the reserves are not met */
3888 * Blocks with reserved pages will never free, skip
3891 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3892 if (pageblock_is_reserved(pfn, block_end_pfn))
3895 /* If this block is reserved, account for it */
3896 if (block_migratetype == MIGRATE_RESERVE) {
3901 /* Suitable for reserving if this block is movable */
3902 if (block_migratetype == MIGRATE_MOVABLE) {
3903 set_pageblock_migratetype(page,
3905 move_freepages_block(zone, page,
3913 * If the reserve is met and this is a previous reserved block,
3916 if (block_migratetype == MIGRATE_RESERVE) {
3917 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3918 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3924 * Initially all pages are reserved - free ones are freed
3925 * up by free_all_bootmem() once the early boot process is
3926 * done. Non-atomic initialization, single-pass.
3928 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3929 unsigned long start_pfn, enum memmap_context context)
3932 unsigned long end_pfn = start_pfn + size;
3936 if (highest_memmap_pfn < end_pfn - 1)
3937 highest_memmap_pfn = end_pfn - 1;
3939 z = &NODE_DATA(nid)->node_zones[zone];
3940 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3942 * There can be holes in boot-time mem_map[]s
3943 * handed to this function. They do not
3944 * exist on hotplugged memory.
3946 if (context == MEMMAP_EARLY) {
3947 if (!early_pfn_valid(pfn))
3949 if (!early_pfn_in_nid(pfn, nid))
3952 page = pfn_to_page(pfn);
3953 set_page_links(page, zone, nid, pfn);
3954 mminit_verify_page_links(page, zone, nid, pfn);
3955 init_page_count(page);
3956 page_mapcount_reset(page);
3957 page_nid_reset_last(page);
3958 SetPageReserved(page);
3960 * Mark the block movable so that blocks are reserved for
3961 * movable at startup. This will force kernel allocations
3962 * to reserve their blocks rather than leaking throughout
3963 * the address space during boot when many long-lived
3964 * kernel allocations are made. Later some blocks near
3965 * the start are marked MIGRATE_RESERVE by
3966 * setup_zone_migrate_reserve()
3968 * bitmap is created for zone's valid pfn range. but memmap
3969 * can be created for invalid pages (for alignment)
3970 * check here not to call set_pageblock_migratetype() against
3973 if ((z->zone_start_pfn <= pfn)
3974 && (pfn < zone_end_pfn(z))
3975 && !(pfn & (pageblock_nr_pages - 1)))
3976 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3978 INIT_LIST_HEAD(&page->lru);
3979 #ifdef WANT_PAGE_VIRTUAL
3980 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3981 if (!is_highmem_idx(zone))
3982 set_page_address(page, __va(pfn << PAGE_SHIFT));
3987 static void __meminit zone_init_free_lists(struct zone *zone)
3990 for_each_migratetype_order(order, t) {
3991 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3992 zone->free_area[order].nr_free = 0;
3996 #ifndef __HAVE_ARCH_MEMMAP_INIT
3997 #define memmap_init(size, nid, zone, start_pfn) \
3998 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4001 static int __meminit zone_batchsize(struct zone *zone)
4007 * The per-cpu-pages pools are set to around 1000th of the
4008 * size of the zone. But no more than 1/2 of a meg.
4010 * OK, so we don't know how big the cache is. So guess.
4012 batch = zone->managed_pages / 1024;
4013 if (batch * PAGE_SIZE > 512 * 1024)
4014 batch = (512 * 1024) / PAGE_SIZE;
4015 batch /= 4; /* We effectively *= 4 below */
4020 * Clamp the batch to a 2^n - 1 value. Having a power
4021 * of 2 value was found to be more likely to have
4022 * suboptimal cache aliasing properties in some cases.
4024 * For example if 2 tasks are alternately allocating
4025 * batches of pages, one task can end up with a lot
4026 * of pages of one half of the possible page colors
4027 * and the other with pages of the other colors.
4029 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4034 /* The deferral and batching of frees should be suppressed under NOMMU
4037 * The problem is that NOMMU needs to be able to allocate large chunks
4038 * of contiguous memory as there's no hardware page translation to
4039 * assemble apparent contiguous memory from discontiguous pages.
4041 * Queueing large contiguous runs of pages for batching, however,
4042 * causes the pages to actually be freed in smaller chunks. As there
4043 * can be a significant delay between the individual batches being
4044 * recycled, this leads to the once large chunks of space being
4045 * fragmented and becoming unavailable for high-order allocations.
4052 * pcp->high and pcp->batch values are related and dependent on one another:
4053 * ->batch must never be higher then ->high.
4054 * The following function updates them in a safe manner without read side
4057 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4058 * those fields changing asynchronously (acording the the above rule).
4060 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4061 * outside of boot time (or some other assurance that no concurrent updaters
4064 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4065 unsigned long batch)
4067 /* start with a fail safe value for batch */
4071 /* Update high, then batch, in order */
4078 /* a companion to pageset_set_high() */
4079 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4081 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4084 static void pageset_init(struct per_cpu_pageset *p)
4086 struct per_cpu_pages *pcp;
4089 memset(p, 0, sizeof(*p));
4093 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4094 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4097 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4100 pageset_set_batch(p, batch);
4104 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4105 * to the value high for the pageset p.
4107 static void pageset_set_high(struct per_cpu_pageset *p,
4110 unsigned long batch = max(1UL, high / 4);
4111 if ((high / 4) > (PAGE_SHIFT * 8))
4112 batch = PAGE_SHIFT * 8;
4114 pageset_update(&p->pcp, high, batch);
4117 static void __meminit pageset_set_high_and_batch(struct zone *zone,
4118 struct per_cpu_pageset *pcp)
4120 if (percpu_pagelist_fraction)
4121 pageset_set_high(pcp,
4122 (zone->managed_pages /
4123 percpu_pagelist_fraction));
4125 pageset_set_batch(pcp, zone_batchsize(zone));
4128 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4130 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4133 pageset_set_high_and_batch(zone, pcp);
4136 static void __meminit setup_zone_pageset(struct zone *zone)
4139 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4140 for_each_possible_cpu(cpu)
4141 zone_pageset_init(zone, cpu);
4145 * Allocate per cpu pagesets and initialize them.
4146 * Before this call only boot pagesets were available.
4148 void __init setup_per_cpu_pageset(void)
4152 for_each_populated_zone(zone)
4153 setup_zone_pageset(zone);
4156 static noinline __init_refok
4157 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4160 struct pglist_data *pgdat = zone->zone_pgdat;
4164 * The per-page waitqueue mechanism uses hashed waitqueues
4167 zone->wait_table_hash_nr_entries =
4168 wait_table_hash_nr_entries(zone_size_pages);
4169 zone->wait_table_bits =
4170 wait_table_bits(zone->wait_table_hash_nr_entries);
4171 alloc_size = zone->wait_table_hash_nr_entries
4172 * sizeof(wait_queue_head_t);
4174 if (!slab_is_available()) {
4175 zone->wait_table = (wait_queue_head_t *)
4176 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4179 * This case means that a zone whose size was 0 gets new memory
4180 * via memory hot-add.
4181 * But it may be the case that a new node was hot-added. In
4182 * this case vmalloc() will not be able to use this new node's
4183 * memory - this wait_table must be initialized to use this new
4184 * node itself as well.
4185 * To use this new node's memory, further consideration will be
4188 zone->wait_table = vmalloc(alloc_size);
4190 if (!zone->wait_table)
4193 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4194 init_waitqueue_head(zone->wait_table + i);
4199 static __meminit void zone_pcp_init(struct zone *zone)
4202 * per cpu subsystem is not up at this point. The following code
4203 * relies on the ability of the linker to provide the
4204 * offset of a (static) per cpu variable into the per cpu area.
4206 zone->pageset = &boot_pageset;
4208 if (zone->present_pages)
4209 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4210 zone->name, zone->present_pages,
4211 zone_batchsize(zone));
4214 int __meminit init_currently_empty_zone(struct zone *zone,
4215 unsigned long zone_start_pfn,
4217 enum memmap_context context)
4219 struct pglist_data *pgdat = zone->zone_pgdat;
4221 ret = zone_wait_table_init(zone, size);
4224 pgdat->nr_zones = zone_idx(zone) + 1;
4226 zone->zone_start_pfn = zone_start_pfn;
4228 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4229 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4231 (unsigned long)zone_idx(zone),
4232 zone_start_pfn, (zone_start_pfn + size));
4234 zone_init_free_lists(zone);
4239 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4240 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4242 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4243 * Architectures may implement their own version but if add_active_range()
4244 * was used and there are no special requirements, this is a convenient
4247 int __meminit __early_pfn_to_nid(unsigned long pfn)
4249 unsigned long start_pfn, end_pfn;
4252 * NOTE: The following SMP-unsafe globals are only used early in boot
4253 * when the kernel is running single-threaded.
4255 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4256 static int __meminitdata last_nid;
4258 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4261 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4262 if (start_pfn <= pfn && pfn < end_pfn) {
4263 last_start_pfn = start_pfn;
4264 last_end_pfn = end_pfn;
4268 /* This is a memory hole */
4271 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4273 int __meminit early_pfn_to_nid(unsigned long pfn)
4277 nid = __early_pfn_to_nid(pfn);
4280 /* just returns 0 */
4284 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4285 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4289 nid = __early_pfn_to_nid(pfn);
4290 if (nid >= 0 && nid != node)
4297 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4298 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4299 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4301 * If an architecture guarantees that all ranges registered with
4302 * add_active_ranges() contain no holes and may be freed, this
4303 * this function may be used instead of calling free_bootmem() manually.
4305 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4307 unsigned long start_pfn, end_pfn;
4310 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4311 start_pfn = min(start_pfn, max_low_pfn);
4312 end_pfn = min(end_pfn, max_low_pfn);
4314 if (start_pfn < end_pfn)
4315 free_bootmem_node(NODE_DATA(this_nid),
4316 PFN_PHYS(start_pfn),
4317 (end_pfn - start_pfn) << PAGE_SHIFT);
4322 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4323 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4325 * If an architecture guarantees that all ranges registered with
4326 * add_active_ranges() contain no holes and may be freed, this
4327 * function may be used instead of calling memory_present() manually.
4329 void __init sparse_memory_present_with_active_regions(int nid)
4331 unsigned long start_pfn, end_pfn;
4334 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4335 memory_present(this_nid, start_pfn, end_pfn);
4339 * get_pfn_range_for_nid - Return the start and end page frames for a node
4340 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4341 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4342 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4344 * It returns the start and end page frame of a node based on information
4345 * provided by an arch calling add_active_range(). If called for a node
4346 * with no available memory, a warning is printed and the start and end
4349 void __meminit get_pfn_range_for_nid(unsigned int nid,
4350 unsigned long *start_pfn, unsigned long *end_pfn)
4352 unsigned long this_start_pfn, this_end_pfn;
4358 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4359 *start_pfn = min(*start_pfn, this_start_pfn);
4360 *end_pfn = max(*end_pfn, this_end_pfn);
4363 if (*start_pfn == -1UL)
4368 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4369 * assumption is made that zones within a node are ordered in monotonic
4370 * increasing memory addresses so that the "highest" populated zone is used
4372 static void __init find_usable_zone_for_movable(void)
4375 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4376 if (zone_index == ZONE_MOVABLE)
4379 if (arch_zone_highest_possible_pfn[zone_index] >
4380 arch_zone_lowest_possible_pfn[zone_index])
4384 VM_BUG_ON(zone_index == -1);
4385 movable_zone = zone_index;
4389 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4390 * because it is sized independent of architecture. Unlike the other zones,
4391 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4392 * in each node depending on the size of each node and how evenly kernelcore
4393 * is distributed. This helper function adjusts the zone ranges
4394 * provided by the architecture for a given node by using the end of the
4395 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4396 * zones within a node are in order of monotonic increases memory addresses
4398 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4399 unsigned long zone_type,
4400 unsigned long node_start_pfn,
4401 unsigned long node_end_pfn,
4402 unsigned long *zone_start_pfn,
4403 unsigned long *zone_end_pfn)
4405 /* Only adjust if ZONE_MOVABLE is on this node */
4406 if (zone_movable_pfn[nid]) {
4407 /* Size ZONE_MOVABLE */
4408 if (zone_type == ZONE_MOVABLE) {
4409 *zone_start_pfn = zone_movable_pfn[nid];
4410 *zone_end_pfn = min(node_end_pfn,
4411 arch_zone_highest_possible_pfn[movable_zone]);
4413 /* Adjust for ZONE_MOVABLE starting within this range */
4414 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4415 *zone_end_pfn > zone_movable_pfn[nid]) {
4416 *zone_end_pfn = zone_movable_pfn[nid];
4418 /* Check if this whole range is within ZONE_MOVABLE */
4419 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4420 *zone_start_pfn = *zone_end_pfn;
4425 * Return the number of pages a zone spans in a node, including holes
4426 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4428 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4429 unsigned long zone_type,
4430 unsigned long *ignored)
4432 unsigned long node_start_pfn, node_end_pfn;
4433 unsigned long zone_start_pfn, zone_end_pfn;
4435 /* Get the start and end of the node and zone */
4436 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4437 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4438 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4439 adjust_zone_range_for_zone_movable(nid, zone_type,
4440 node_start_pfn, node_end_pfn,
4441 &zone_start_pfn, &zone_end_pfn);
4443 /* Check that this node has pages within the zone's required range */
4444 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4447 /* Move the zone boundaries inside the node if necessary */
4448 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4449 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4451 /* Return the spanned pages */
4452 return zone_end_pfn - zone_start_pfn;
4456 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4457 * then all holes in the requested range will be accounted for.
4459 unsigned long __meminit __absent_pages_in_range(int nid,
4460 unsigned long range_start_pfn,
4461 unsigned long range_end_pfn)
4463 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4464 unsigned long start_pfn, end_pfn;
4467 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4468 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4469 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4470 nr_absent -= end_pfn - start_pfn;
4476 * absent_pages_in_range - Return number of page frames in holes within a range
4477 * @start_pfn: The start PFN to start searching for holes
4478 * @end_pfn: The end PFN to stop searching for holes
4480 * It returns the number of pages frames in memory holes within a range.
4482 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4483 unsigned long end_pfn)
4485 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4488 /* Return the number of page frames in holes in a zone on a node */
4489 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4490 unsigned long zone_type,
4491 unsigned long *ignored)
4493 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4494 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4495 unsigned long node_start_pfn, node_end_pfn;
4496 unsigned long zone_start_pfn, zone_end_pfn;
4498 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4499 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4500 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4502 adjust_zone_range_for_zone_movable(nid, zone_type,
4503 node_start_pfn, node_end_pfn,
4504 &zone_start_pfn, &zone_end_pfn);
4505 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4508 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4509 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4510 unsigned long zone_type,
4511 unsigned long *zones_size)
4513 return zones_size[zone_type];
4516 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4517 unsigned long zone_type,
4518 unsigned long *zholes_size)
4523 return zholes_size[zone_type];
4526 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4528 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4529 unsigned long *zones_size, unsigned long *zholes_size)
4531 unsigned long realtotalpages, totalpages = 0;
4534 for (i = 0; i < MAX_NR_ZONES; i++)
4535 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4537 pgdat->node_spanned_pages = totalpages;
4539 realtotalpages = totalpages;
4540 for (i = 0; i < MAX_NR_ZONES; i++)
4542 zone_absent_pages_in_node(pgdat->node_id, i,
4544 pgdat->node_present_pages = realtotalpages;
4545 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4549 #ifndef CONFIG_SPARSEMEM
4551 * Calculate the size of the zone->blockflags rounded to an unsigned long
4552 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4553 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4554 * round what is now in bits to nearest long in bits, then return it in
4557 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4559 unsigned long usemapsize;
4561 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4562 usemapsize = roundup(zonesize, pageblock_nr_pages);
4563 usemapsize = usemapsize >> pageblock_order;
4564 usemapsize *= NR_PAGEBLOCK_BITS;
4565 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4567 return usemapsize / 8;
4570 static void __init setup_usemap(struct pglist_data *pgdat,
4572 unsigned long zone_start_pfn,
4573 unsigned long zonesize)
4575 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4576 zone->pageblock_flags = NULL;
4578 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4582 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4583 unsigned long zone_start_pfn, unsigned long zonesize) {}
4584 #endif /* CONFIG_SPARSEMEM */
4586 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4588 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4589 void __init set_pageblock_order(void)
4593 /* Check that pageblock_nr_pages has not already been setup */
4594 if (pageblock_order)
4597 if (HPAGE_SHIFT > PAGE_SHIFT)
4598 order = HUGETLB_PAGE_ORDER;
4600 order = MAX_ORDER - 1;
4603 * Assume the largest contiguous order of interest is a huge page.
4604 * This value may be variable depending on boot parameters on IA64 and
4607 pageblock_order = order;
4609 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4612 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4613 * is unused as pageblock_order is set at compile-time. See
4614 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4617 void __init set_pageblock_order(void)
4621 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4623 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4624 unsigned long present_pages)
4626 unsigned long pages = spanned_pages;
4629 * Provide a more accurate estimation if there are holes within
4630 * the zone and SPARSEMEM is in use. If there are holes within the
4631 * zone, each populated memory region may cost us one or two extra
4632 * memmap pages due to alignment because memmap pages for each
4633 * populated regions may not naturally algined on page boundary.
4634 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4636 if (spanned_pages > present_pages + (present_pages >> 4) &&
4637 IS_ENABLED(CONFIG_SPARSEMEM))
4638 pages = present_pages;
4640 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4644 * Set up the zone data structures:
4645 * - mark all pages reserved
4646 * - mark all memory queues empty
4647 * - clear the memory bitmaps
4649 * NOTE: pgdat should get zeroed by caller.
4651 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4652 unsigned long *zones_size, unsigned long *zholes_size)
4655 int nid = pgdat->node_id;
4656 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4659 pgdat_resize_init(pgdat);
4660 #ifdef CONFIG_NUMA_BALANCING
4661 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4662 pgdat->numabalancing_migrate_nr_pages = 0;
4663 pgdat->numabalancing_migrate_next_window = jiffies;
4665 init_waitqueue_head(&pgdat->kswapd_wait);
4666 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4667 pgdat_page_cgroup_init(pgdat);
4669 for (j = 0; j < MAX_NR_ZONES; j++) {
4670 struct zone *zone = pgdat->node_zones + j;
4671 unsigned long size, realsize, freesize, memmap_pages;
4673 size = zone_spanned_pages_in_node(nid, j, zones_size);
4674 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4678 * Adjust freesize so that it accounts for how much memory
4679 * is used by this zone for memmap. This affects the watermark
4680 * and per-cpu initialisations
4682 memmap_pages = calc_memmap_size(size, realsize);
4683 if (freesize >= memmap_pages) {
4684 freesize -= memmap_pages;
4687 " %s zone: %lu pages used for memmap\n",
4688 zone_names[j], memmap_pages);
4691 " %s zone: %lu pages exceeds freesize %lu\n",
4692 zone_names[j], memmap_pages, freesize);
4694 /* Account for reserved pages */
4695 if (j == 0 && freesize > dma_reserve) {
4696 freesize -= dma_reserve;
4697 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4698 zone_names[0], dma_reserve);
4701 if (!is_highmem_idx(j))
4702 nr_kernel_pages += freesize;
4703 /* Charge for highmem memmap if there are enough kernel pages */
4704 else if (nr_kernel_pages > memmap_pages * 2)
4705 nr_kernel_pages -= memmap_pages;
4706 nr_all_pages += freesize;
4708 zone->spanned_pages = size;
4709 zone->present_pages = realsize;
4711 * Set an approximate value for lowmem here, it will be adjusted
4712 * when the bootmem allocator frees pages into the buddy system.
4713 * And all highmem pages will be managed by the buddy system.
4715 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4718 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4720 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4722 zone->name = zone_names[j];
4723 spin_lock_init(&zone->lock);
4724 spin_lock_init(&zone->lru_lock);
4725 zone_seqlock_init(zone);
4726 zone->zone_pgdat = pgdat;
4728 zone_pcp_init(zone);
4729 lruvec_init(&zone->lruvec);
4733 set_pageblock_order();
4734 setup_usemap(pgdat, zone, zone_start_pfn, size);
4735 ret = init_currently_empty_zone(zone, zone_start_pfn,
4736 size, MEMMAP_EARLY);
4738 memmap_init(size, nid, j, zone_start_pfn);
4739 zone_start_pfn += size;
4743 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4745 /* Skip empty nodes */
4746 if (!pgdat->node_spanned_pages)
4749 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4750 /* ia64 gets its own node_mem_map, before this, without bootmem */
4751 if (!pgdat->node_mem_map) {
4752 unsigned long size, start, end;
4756 * The zone's endpoints aren't required to be MAX_ORDER
4757 * aligned but the node_mem_map endpoints must be in order
4758 * for the buddy allocator to function correctly.
4760 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4761 end = pgdat_end_pfn(pgdat);
4762 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4763 size = (end - start) * sizeof(struct page);
4764 map = alloc_remap(pgdat->node_id, size);
4766 map = alloc_bootmem_node_nopanic(pgdat, size);
4767 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4769 #ifndef CONFIG_NEED_MULTIPLE_NODES
4771 * With no DISCONTIG, the global mem_map is just set as node 0's
4773 if (pgdat == NODE_DATA(0)) {
4774 mem_map = NODE_DATA(0)->node_mem_map;
4775 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4776 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4777 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4778 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4781 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4784 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4785 unsigned long node_start_pfn, unsigned long *zholes_size)
4787 pg_data_t *pgdat = NODE_DATA(nid);
4789 /* pg_data_t should be reset to zero when it's allocated */
4790 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4792 pgdat->node_id = nid;
4793 pgdat->node_start_pfn = node_start_pfn;
4794 init_zone_allows_reclaim(nid);
4795 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4797 alloc_node_mem_map(pgdat);
4798 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4799 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4800 nid, (unsigned long)pgdat,
4801 (unsigned long)pgdat->node_mem_map);
4804 free_area_init_core(pgdat, zones_size, zholes_size);
4807 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4809 #if MAX_NUMNODES > 1
4811 * Figure out the number of possible node ids.
4813 void __init setup_nr_node_ids(void)
4816 unsigned int highest = 0;
4818 for_each_node_mask(node, node_possible_map)
4820 nr_node_ids = highest + 1;
4825 * node_map_pfn_alignment - determine the maximum internode alignment
4827 * This function should be called after node map is populated and sorted.
4828 * It calculates the maximum power of two alignment which can distinguish
4831 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4832 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4833 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4834 * shifted, 1GiB is enough and this function will indicate so.
4836 * This is used to test whether pfn -> nid mapping of the chosen memory
4837 * model has fine enough granularity to avoid incorrect mapping for the
4838 * populated node map.
4840 * Returns the determined alignment in pfn's. 0 if there is no alignment
4841 * requirement (single node).
4843 unsigned long __init node_map_pfn_alignment(void)
4845 unsigned long accl_mask = 0, last_end = 0;
4846 unsigned long start, end, mask;
4850 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4851 if (!start || last_nid < 0 || last_nid == nid) {
4858 * Start with a mask granular enough to pin-point to the
4859 * start pfn and tick off bits one-by-one until it becomes
4860 * too coarse to separate the current node from the last.
4862 mask = ~((1 << __ffs(start)) - 1);
4863 while (mask && last_end <= (start & (mask << 1)))
4866 /* accumulate all internode masks */
4870 /* convert mask to number of pages */
4871 return ~accl_mask + 1;
4874 /* Find the lowest pfn for a node */
4875 static unsigned long __init find_min_pfn_for_node(int nid)
4877 unsigned long min_pfn = ULONG_MAX;
4878 unsigned long start_pfn;
4881 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4882 min_pfn = min(min_pfn, start_pfn);
4884 if (min_pfn == ULONG_MAX) {
4886 "Could not find start_pfn for node %d\n", nid);
4894 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4896 * It returns the minimum PFN based on information provided via
4897 * add_active_range().
4899 unsigned long __init find_min_pfn_with_active_regions(void)
4901 return find_min_pfn_for_node(MAX_NUMNODES);
4905 * early_calculate_totalpages()
4906 * Sum pages in active regions for movable zone.
4907 * Populate N_MEMORY for calculating usable_nodes.
4909 static unsigned long __init early_calculate_totalpages(void)
4911 unsigned long totalpages = 0;
4912 unsigned long start_pfn, end_pfn;
4915 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4916 unsigned long pages = end_pfn - start_pfn;
4918 totalpages += pages;
4920 node_set_state(nid, N_MEMORY);
4926 * Find the PFN the Movable zone begins in each node. Kernel memory
4927 * is spread evenly between nodes as long as the nodes have enough
4928 * memory. When they don't, some nodes will have more kernelcore than
4931 static void __init find_zone_movable_pfns_for_nodes(void)
4934 unsigned long usable_startpfn;
4935 unsigned long kernelcore_node, kernelcore_remaining;
4936 /* save the state before borrow the nodemask */
4937 nodemask_t saved_node_state = node_states[N_MEMORY];
4938 unsigned long totalpages = early_calculate_totalpages();
4939 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4942 * If movablecore was specified, calculate what size of
4943 * kernelcore that corresponds so that memory usable for
4944 * any allocation type is evenly spread. If both kernelcore
4945 * and movablecore are specified, then the value of kernelcore
4946 * will be used for required_kernelcore if it's greater than
4947 * what movablecore would have allowed.
4949 if (required_movablecore) {
4950 unsigned long corepages;
4953 * Round-up so that ZONE_MOVABLE is at least as large as what
4954 * was requested by the user
4956 required_movablecore =
4957 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4958 corepages = totalpages - required_movablecore;
4960 required_kernelcore = max(required_kernelcore, corepages);
4963 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4964 if (!required_kernelcore)
4967 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4968 find_usable_zone_for_movable();
4969 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4972 /* Spread kernelcore memory as evenly as possible throughout nodes */
4973 kernelcore_node = required_kernelcore / usable_nodes;
4974 for_each_node_state(nid, N_MEMORY) {
4975 unsigned long start_pfn, end_pfn;
4978 * Recalculate kernelcore_node if the division per node
4979 * now exceeds what is necessary to satisfy the requested
4980 * amount of memory for the kernel
4982 if (required_kernelcore < kernelcore_node)
4983 kernelcore_node = required_kernelcore / usable_nodes;
4986 * As the map is walked, we track how much memory is usable
4987 * by the kernel using kernelcore_remaining. When it is
4988 * 0, the rest of the node is usable by ZONE_MOVABLE
4990 kernelcore_remaining = kernelcore_node;
4992 /* Go through each range of PFNs within this node */
4993 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4994 unsigned long size_pages;
4996 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4997 if (start_pfn >= end_pfn)
5000 /* Account for what is only usable for kernelcore */
5001 if (start_pfn < usable_startpfn) {
5002 unsigned long kernel_pages;
5003 kernel_pages = min(end_pfn, usable_startpfn)
5006 kernelcore_remaining -= min(kernel_pages,
5007 kernelcore_remaining);
5008 required_kernelcore -= min(kernel_pages,
5009 required_kernelcore);
5011 /* Continue if range is now fully accounted */
5012 if (end_pfn <= usable_startpfn) {
5015 * Push zone_movable_pfn to the end so
5016 * that if we have to rebalance
5017 * kernelcore across nodes, we will
5018 * not double account here
5020 zone_movable_pfn[nid] = end_pfn;
5023 start_pfn = usable_startpfn;
5027 * The usable PFN range for ZONE_MOVABLE is from
5028 * start_pfn->end_pfn. Calculate size_pages as the
5029 * number of pages used as kernelcore
5031 size_pages = end_pfn - start_pfn;
5032 if (size_pages > kernelcore_remaining)
5033 size_pages = kernelcore_remaining;
5034 zone_movable_pfn[nid] = start_pfn + size_pages;
5037 * Some kernelcore has been met, update counts and
5038 * break if the kernelcore for this node has been
5041 required_kernelcore -= min(required_kernelcore,
5043 kernelcore_remaining -= size_pages;
5044 if (!kernelcore_remaining)
5050 * If there is still required_kernelcore, we do another pass with one
5051 * less node in the count. This will push zone_movable_pfn[nid] further
5052 * along on the nodes that still have memory until kernelcore is
5056 if (usable_nodes && required_kernelcore > usable_nodes)
5059 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5060 for (nid = 0; nid < MAX_NUMNODES; nid++)
5061 zone_movable_pfn[nid] =
5062 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5065 /* restore the node_state */
5066 node_states[N_MEMORY] = saved_node_state;
5069 /* Any regular or high memory on that node ? */
5070 static void check_for_memory(pg_data_t *pgdat, int nid)
5072 enum zone_type zone_type;
5074 if (N_MEMORY == N_NORMAL_MEMORY)
5077 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5078 struct zone *zone = &pgdat->node_zones[zone_type];
5079 if (zone->present_pages) {
5080 node_set_state(nid, N_HIGH_MEMORY);
5081 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5082 zone_type <= ZONE_NORMAL)
5083 node_set_state(nid, N_NORMAL_MEMORY);
5090 * free_area_init_nodes - Initialise all pg_data_t and zone data
5091 * @max_zone_pfn: an array of max PFNs for each zone
5093 * This will call free_area_init_node() for each active node in the system.
5094 * Using the page ranges provided by add_active_range(), the size of each
5095 * zone in each node and their holes is calculated. If the maximum PFN
5096 * between two adjacent zones match, it is assumed that the zone is empty.
5097 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5098 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5099 * starts where the previous one ended. For example, ZONE_DMA32 starts
5100 * at arch_max_dma_pfn.
5102 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5104 unsigned long start_pfn, end_pfn;
5107 /* Record where the zone boundaries are */
5108 memset(arch_zone_lowest_possible_pfn, 0,
5109 sizeof(arch_zone_lowest_possible_pfn));
5110 memset(arch_zone_highest_possible_pfn, 0,
5111 sizeof(arch_zone_highest_possible_pfn));
5112 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5113 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5114 for (i = 1; i < MAX_NR_ZONES; i++) {
5115 if (i == ZONE_MOVABLE)
5117 arch_zone_lowest_possible_pfn[i] =
5118 arch_zone_highest_possible_pfn[i-1];
5119 arch_zone_highest_possible_pfn[i] =
5120 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5122 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5123 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5125 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5126 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5127 find_zone_movable_pfns_for_nodes();
5129 /* Print out the zone ranges */
5130 printk("Zone ranges:\n");
5131 for (i = 0; i < MAX_NR_ZONES; i++) {
5132 if (i == ZONE_MOVABLE)
5134 printk(KERN_CONT " %-8s ", zone_names[i]);
5135 if (arch_zone_lowest_possible_pfn[i] ==
5136 arch_zone_highest_possible_pfn[i])
5137 printk(KERN_CONT "empty\n");
5139 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5140 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5141 (arch_zone_highest_possible_pfn[i]
5142 << PAGE_SHIFT) - 1);
5145 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5146 printk("Movable zone start for each node\n");
5147 for (i = 0; i < MAX_NUMNODES; i++) {
5148 if (zone_movable_pfn[i])
5149 printk(" Node %d: %#010lx\n", i,
5150 zone_movable_pfn[i] << PAGE_SHIFT);
5153 /* Print out the early node map */
5154 printk("Early memory node ranges\n");
5155 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5156 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5157 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5159 /* Initialise every node */
5160 mminit_verify_pageflags_layout();
5161 setup_nr_node_ids();
5162 for_each_online_node(nid) {
5163 pg_data_t *pgdat = NODE_DATA(nid);
5164 free_area_init_node(nid, NULL,
5165 find_min_pfn_for_node(nid), NULL);
5167 /* Any memory on that node */
5168 if (pgdat->node_present_pages)
5169 node_set_state(nid, N_MEMORY);
5170 check_for_memory(pgdat, nid);
5174 static int __init cmdline_parse_core(char *p, unsigned long *core)
5176 unsigned long long coremem;
5180 coremem = memparse(p, &p);
5181 *core = coremem >> PAGE_SHIFT;
5183 /* Paranoid check that UL is enough for the coremem value */
5184 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5190 * kernelcore=size sets the amount of memory for use for allocations that
5191 * cannot be reclaimed or migrated.
5193 static int __init cmdline_parse_kernelcore(char *p)
5195 return cmdline_parse_core(p, &required_kernelcore);
5199 * movablecore=size sets the amount of memory for use for allocations that
5200 * can be reclaimed or migrated.
5202 static int __init cmdline_parse_movablecore(char *p)
5204 return cmdline_parse_core(p, &required_movablecore);
5207 early_param("kernelcore", cmdline_parse_kernelcore);
5208 early_param("movablecore", cmdline_parse_movablecore);
5210 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5212 void adjust_managed_page_count(struct page *page, long count)
5214 spin_lock(&managed_page_count_lock);
5215 page_zone(page)->managed_pages += count;
5216 totalram_pages += count;
5217 spin_unlock(&managed_page_count_lock);
5220 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5223 unsigned long pages = 0;
5225 start = (void *)PAGE_ALIGN((unsigned long)start);
5226 end = (void *)((unsigned long)end & PAGE_MASK);
5227 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5228 if ((unsigned int)poison <= 0xFF)
5229 memset(pos, poison, PAGE_SIZE);
5230 free_reserved_page(virt_to_page(pos));
5234 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5235 s, pages << (PAGE_SHIFT - 10), start, end);
5239 EXPORT_SYMBOL(free_reserved_area);
5241 #ifdef CONFIG_HIGHMEM
5242 void free_highmem_page(struct page *page)
5244 __free_reserved_page(page);
5246 page_zone(page)->managed_pages++;
5252 * set_dma_reserve - set the specified number of pages reserved in the first zone
5253 * @new_dma_reserve: The number of pages to mark reserved
5255 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5256 * In the DMA zone, a significant percentage may be consumed by kernel image
5257 * and other unfreeable allocations which can skew the watermarks badly. This
5258 * function may optionally be used to account for unfreeable pages in the
5259 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5260 * smaller per-cpu batchsize.
5262 void __init set_dma_reserve(unsigned long new_dma_reserve)
5264 dma_reserve = new_dma_reserve;
5267 void __init free_area_init(unsigned long *zones_size)
5269 free_area_init_node(0, zones_size,
5270 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5273 static int page_alloc_cpu_notify(struct notifier_block *self,
5274 unsigned long action, void *hcpu)
5276 int cpu = (unsigned long)hcpu;
5278 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5279 lru_add_drain_cpu(cpu);
5283 * Spill the event counters of the dead processor
5284 * into the current processors event counters.
5285 * This artificially elevates the count of the current
5288 vm_events_fold_cpu(cpu);
5291 * Zero the differential counters of the dead processor
5292 * so that the vm statistics are consistent.
5294 * This is only okay since the processor is dead and cannot
5295 * race with what we are doing.
5297 refresh_cpu_vm_stats(cpu);
5302 void __init page_alloc_init(void)
5304 hotcpu_notifier(page_alloc_cpu_notify, 0);
5308 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5309 * or min_free_kbytes changes.
5311 static void calculate_totalreserve_pages(void)
5313 struct pglist_data *pgdat;
5314 unsigned long reserve_pages = 0;
5315 enum zone_type i, j;
5317 for_each_online_pgdat(pgdat) {
5318 for (i = 0; i < MAX_NR_ZONES; i++) {
5319 struct zone *zone = pgdat->node_zones + i;
5320 unsigned long max = 0;
5322 /* Find valid and maximum lowmem_reserve in the zone */
5323 for (j = i; j < MAX_NR_ZONES; j++) {
5324 if (zone->lowmem_reserve[j] > max)
5325 max = zone->lowmem_reserve[j];
5328 /* we treat the high watermark as reserved pages. */
5329 max += high_wmark_pages(zone);
5331 if (max > zone->managed_pages)
5332 max = zone->managed_pages;
5333 reserve_pages += max;
5335 * Lowmem reserves are not available to
5336 * GFP_HIGHUSER page cache allocations and
5337 * kswapd tries to balance zones to their high
5338 * watermark. As a result, neither should be
5339 * regarded as dirtyable memory, to prevent a
5340 * situation where reclaim has to clean pages
5341 * in order to balance the zones.
5343 zone->dirty_balance_reserve = max;
5346 dirty_balance_reserve = reserve_pages;
5347 totalreserve_pages = reserve_pages;
5351 * setup_per_zone_lowmem_reserve - called whenever
5352 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5353 * has a correct pages reserved value, so an adequate number of
5354 * pages are left in the zone after a successful __alloc_pages().
5356 static void setup_per_zone_lowmem_reserve(void)
5358 struct pglist_data *pgdat;
5359 enum zone_type j, idx;
5361 for_each_online_pgdat(pgdat) {
5362 for (j = 0; j < MAX_NR_ZONES; j++) {
5363 struct zone *zone = pgdat->node_zones + j;
5364 unsigned long managed_pages = zone->managed_pages;
5366 zone->lowmem_reserve[j] = 0;
5370 struct zone *lower_zone;
5374 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5375 sysctl_lowmem_reserve_ratio[idx] = 1;
5377 lower_zone = pgdat->node_zones + idx;
5378 lower_zone->lowmem_reserve[j] = managed_pages /
5379 sysctl_lowmem_reserve_ratio[idx];
5380 managed_pages += lower_zone->managed_pages;
5385 /* update totalreserve_pages */
5386 calculate_totalreserve_pages();
5389 static void __setup_per_zone_wmarks(void)
5391 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5392 unsigned long lowmem_pages = 0;
5394 unsigned long flags;
5396 /* Calculate total number of !ZONE_HIGHMEM pages */
5397 for_each_zone(zone) {
5398 if (!is_highmem(zone))
5399 lowmem_pages += zone->managed_pages;
5402 for_each_zone(zone) {
5405 spin_lock_irqsave(&zone->lock, flags);
5406 tmp = (u64)pages_min * zone->managed_pages;
5407 do_div(tmp, lowmem_pages);
5408 if (is_highmem(zone)) {
5410 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5411 * need highmem pages, so cap pages_min to a small
5414 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5415 * deltas controls asynch page reclaim, and so should
5416 * not be capped for highmem.
5418 unsigned long min_pages;
5420 min_pages = zone->managed_pages / 1024;
5421 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5422 zone->watermark[WMARK_MIN] = min_pages;
5425 * If it's a lowmem zone, reserve a number of pages
5426 * proportionate to the zone's size.
5428 zone->watermark[WMARK_MIN] = tmp;
5431 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5432 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5434 setup_zone_migrate_reserve(zone);
5435 spin_unlock_irqrestore(&zone->lock, flags);
5438 /* update totalreserve_pages */
5439 calculate_totalreserve_pages();
5443 * setup_per_zone_wmarks - called when min_free_kbytes changes
5444 * or when memory is hot-{added|removed}
5446 * Ensures that the watermark[min,low,high] values for each zone are set
5447 * correctly with respect to min_free_kbytes.
5449 void setup_per_zone_wmarks(void)
5451 mutex_lock(&zonelists_mutex);
5452 __setup_per_zone_wmarks();
5453 mutex_unlock(&zonelists_mutex);
5457 * The inactive anon list should be small enough that the VM never has to
5458 * do too much work, but large enough that each inactive page has a chance
5459 * to be referenced again before it is swapped out.
5461 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5462 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5463 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5464 * the anonymous pages are kept on the inactive list.
5467 * memory ratio inactive anon
5468 * -------------------------------------
5477 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5479 unsigned int gb, ratio;
5481 /* Zone size in gigabytes */
5482 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5484 ratio = int_sqrt(10 * gb);
5488 zone->inactive_ratio = ratio;
5491 static void __meminit setup_per_zone_inactive_ratio(void)
5496 calculate_zone_inactive_ratio(zone);
5500 * Initialise min_free_kbytes.
5502 * For small machines we want it small (128k min). For large machines
5503 * we want it large (64MB max). But it is not linear, because network
5504 * bandwidth does not increase linearly with machine size. We use
5506 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5507 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5523 int __meminit init_per_zone_wmark_min(void)
5525 unsigned long lowmem_kbytes;
5527 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5529 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5530 if (min_free_kbytes < 128)
5531 min_free_kbytes = 128;
5532 if (min_free_kbytes > 65536)
5533 min_free_kbytes = 65536;
5534 setup_per_zone_wmarks();
5535 refresh_zone_stat_thresholds();
5536 setup_per_zone_lowmem_reserve();
5537 setup_per_zone_inactive_ratio();
5540 module_init(init_per_zone_wmark_min)
5543 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5544 * that we can call two helper functions whenever min_free_kbytes
5547 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5548 void __user *buffer, size_t *length, loff_t *ppos)
5550 proc_dointvec(table, write, buffer, length, ppos);
5552 setup_per_zone_wmarks();
5557 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5558 void __user *buffer, size_t *length, loff_t *ppos)
5563 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5568 zone->min_unmapped_pages = (zone->managed_pages *
5569 sysctl_min_unmapped_ratio) / 100;
5573 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5574 void __user *buffer, size_t *length, loff_t *ppos)
5579 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5584 zone->min_slab_pages = (zone->managed_pages *
5585 sysctl_min_slab_ratio) / 100;
5591 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5592 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5593 * whenever sysctl_lowmem_reserve_ratio changes.
5595 * The reserve ratio obviously has absolutely no relation with the
5596 * minimum watermarks. The lowmem reserve ratio can only make sense
5597 * if in function of the boot time zone sizes.
5599 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5600 void __user *buffer, size_t *length, loff_t *ppos)
5602 proc_dointvec_minmax(table, write, buffer, length, ppos);
5603 setup_per_zone_lowmem_reserve();
5608 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5609 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5610 * can have before it gets flushed back to buddy allocator.
5612 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5613 void __user *buffer, size_t *length, loff_t *ppos)
5619 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5620 if (!write || (ret < 0))
5623 mutex_lock(&pcp_batch_high_lock);
5624 for_each_populated_zone(zone) {
5626 high = zone->managed_pages / percpu_pagelist_fraction;
5627 for_each_possible_cpu(cpu)
5628 pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5631 mutex_unlock(&pcp_batch_high_lock);
5635 int hashdist = HASHDIST_DEFAULT;
5638 static int __init set_hashdist(char *str)
5642 hashdist = simple_strtoul(str, &str, 0);
5645 __setup("hashdist=", set_hashdist);
5649 * allocate a large system hash table from bootmem
5650 * - it is assumed that the hash table must contain an exact power-of-2
5651 * quantity of entries
5652 * - limit is the number of hash buckets, not the total allocation size
5654 void *__init alloc_large_system_hash(const char *tablename,
5655 unsigned long bucketsize,
5656 unsigned long numentries,
5659 unsigned int *_hash_shift,
5660 unsigned int *_hash_mask,
5661 unsigned long low_limit,
5662 unsigned long high_limit)
5664 unsigned long long max = high_limit;
5665 unsigned long log2qty, size;
5668 /* allow the kernel cmdline to have a say */
5670 /* round applicable memory size up to nearest megabyte */
5671 numentries = nr_kernel_pages;
5672 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5673 numentries >>= 20 - PAGE_SHIFT;
5674 numentries <<= 20 - PAGE_SHIFT;
5676 /* limit to 1 bucket per 2^scale bytes of low memory */
5677 if (scale > PAGE_SHIFT)
5678 numentries >>= (scale - PAGE_SHIFT);
5680 numentries <<= (PAGE_SHIFT - scale);
5682 /* Make sure we've got at least a 0-order allocation.. */
5683 if (unlikely(flags & HASH_SMALL)) {
5684 /* Makes no sense without HASH_EARLY */
5685 WARN_ON(!(flags & HASH_EARLY));
5686 if (!(numentries >> *_hash_shift)) {
5687 numentries = 1UL << *_hash_shift;
5688 BUG_ON(!numentries);
5690 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5691 numentries = PAGE_SIZE / bucketsize;
5693 numentries = roundup_pow_of_two(numentries);
5695 /* limit allocation size to 1/16 total memory by default */
5697 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5698 do_div(max, bucketsize);
5700 max = min(max, 0x80000000ULL);
5702 if (numentries < low_limit)
5703 numentries = low_limit;
5704 if (numentries > max)
5707 log2qty = ilog2(numentries);
5710 size = bucketsize << log2qty;
5711 if (flags & HASH_EARLY)
5712 table = alloc_bootmem_nopanic(size);
5714 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5717 * If bucketsize is not a power-of-two, we may free
5718 * some pages at the end of hash table which
5719 * alloc_pages_exact() automatically does
5721 if (get_order(size) < MAX_ORDER) {
5722 table = alloc_pages_exact(size, GFP_ATOMIC);
5723 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5726 } while (!table && size > PAGE_SIZE && --log2qty);
5729 panic("Failed to allocate %s hash table\n", tablename);
5731 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5734 ilog2(size) - PAGE_SHIFT,
5738 *_hash_shift = log2qty;
5740 *_hash_mask = (1 << log2qty) - 1;
5745 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5746 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5749 #ifdef CONFIG_SPARSEMEM
5750 return __pfn_to_section(pfn)->pageblock_flags;
5752 return zone->pageblock_flags;
5753 #endif /* CONFIG_SPARSEMEM */
5756 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5758 #ifdef CONFIG_SPARSEMEM
5759 pfn &= (PAGES_PER_SECTION-1);
5760 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5762 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5763 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5764 #endif /* CONFIG_SPARSEMEM */
5768 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5769 * @page: The page within the block of interest
5770 * @start_bitidx: The first bit of interest to retrieve
5771 * @end_bitidx: The last bit of interest
5772 * returns pageblock_bits flags
5774 unsigned long get_pageblock_flags_group(struct page *page,
5775 int start_bitidx, int end_bitidx)
5778 unsigned long *bitmap;
5779 unsigned long pfn, bitidx;
5780 unsigned long flags = 0;
5781 unsigned long value = 1;
5783 zone = page_zone(page);
5784 pfn = page_to_pfn(page);
5785 bitmap = get_pageblock_bitmap(zone, pfn);
5786 bitidx = pfn_to_bitidx(zone, pfn);
5788 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5789 if (test_bit(bitidx + start_bitidx, bitmap))
5796 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5797 * @page: The page within the block of interest
5798 * @start_bitidx: The first bit of interest
5799 * @end_bitidx: The last bit of interest
5800 * @flags: The flags to set
5802 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5803 int start_bitidx, int end_bitidx)
5806 unsigned long *bitmap;
5807 unsigned long pfn, bitidx;
5808 unsigned long value = 1;
5810 zone = page_zone(page);
5811 pfn = page_to_pfn(page);
5812 bitmap = get_pageblock_bitmap(zone, pfn);
5813 bitidx = pfn_to_bitidx(zone, pfn);
5814 VM_BUG_ON(!zone_spans_pfn(zone, pfn));
5816 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5818 __set_bit(bitidx + start_bitidx, bitmap);
5820 __clear_bit(bitidx + start_bitidx, bitmap);
5824 * This function checks whether pageblock includes unmovable pages or not.
5825 * If @count is not zero, it is okay to include less @count unmovable pages
5827 * PageLRU check wihtout isolation or lru_lock could race so that
5828 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5829 * expect this function should be exact.
5831 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5832 bool skip_hwpoisoned_pages)
5834 unsigned long pfn, iter, found;
5838 * For avoiding noise data, lru_add_drain_all() should be called
5839 * If ZONE_MOVABLE, the zone never contains unmovable pages
5841 if (zone_idx(zone) == ZONE_MOVABLE)
5843 mt = get_pageblock_migratetype(page);
5844 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5847 pfn = page_to_pfn(page);
5848 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5849 unsigned long check = pfn + iter;
5851 if (!pfn_valid_within(check))
5854 page = pfn_to_page(check);
5856 * We can't use page_count without pin a page
5857 * because another CPU can free compound page.
5858 * This check already skips compound tails of THP
5859 * because their page->_count is zero at all time.
5861 if (!atomic_read(&page->_count)) {
5862 if (PageBuddy(page))
5863 iter += (1 << page_order(page)) - 1;
5868 * The HWPoisoned page may be not in buddy system, and
5869 * page_count() is not 0.
5871 if (skip_hwpoisoned_pages && PageHWPoison(page))
5877 * If there are RECLAIMABLE pages, we need to check it.
5878 * But now, memory offline itself doesn't call shrink_slab()
5879 * and it still to be fixed.
5882 * If the page is not RAM, page_count()should be 0.
5883 * we don't need more check. This is an _used_ not-movable page.
5885 * The problematic thing here is PG_reserved pages. PG_reserved
5886 * is set to both of a memory hole page and a _used_ kernel
5895 bool is_pageblock_removable_nolock(struct page *page)
5901 * We have to be careful here because we are iterating over memory
5902 * sections which are not zone aware so we might end up outside of
5903 * the zone but still within the section.
5904 * We have to take care about the node as well. If the node is offline
5905 * its NODE_DATA will be NULL - see page_zone.
5907 if (!node_online(page_to_nid(page)))
5910 zone = page_zone(page);
5911 pfn = page_to_pfn(page);
5912 if (!zone_spans_pfn(zone, pfn))
5915 return !has_unmovable_pages(zone, page, 0, true);
5920 static unsigned long pfn_max_align_down(unsigned long pfn)
5922 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5923 pageblock_nr_pages) - 1);
5926 static unsigned long pfn_max_align_up(unsigned long pfn)
5928 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5929 pageblock_nr_pages));
5932 /* [start, end) must belong to a single zone. */
5933 static int __alloc_contig_migrate_range(struct compact_control *cc,
5934 unsigned long start, unsigned long end)
5936 /* This function is based on compact_zone() from compaction.c. */
5937 unsigned long nr_reclaimed;
5938 unsigned long pfn = start;
5939 unsigned int tries = 0;
5944 while (pfn < end || !list_empty(&cc->migratepages)) {
5945 if (fatal_signal_pending(current)) {
5950 if (list_empty(&cc->migratepages)) {
5951 cc->nr_migratepages = 0;
5952 pfn = isolate_migratepages_range(cc->zone, cc,
5959 } else if (++tries == 5) {
5960 ret = ret < 0 ? ret : -EBUSY;
5964 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5966 cc->nr_migratepages -= nr_reclaimed;
5968 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
5969 0, MIGRATE_SYNC, MR_CMA);
5972 putback_movable_pages(&cc->migratepages);
5979 * alloc_contig_range() -- tries to allocate given range of pages
5980 * @start: start PFN to allocate
5981 * @end: one-past-the-last PFN to allocate
5982 * @migratetype: migratetype of the underlaying pageblocks (either
5983 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5984 * in range must have the same migratetype and it must
5985 * be either of the two.
5987 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5988 * aligned, however it's the caller's responsibility to guarantee that
5989 * we are the only thread that changes migrate type of pageblocks the
5992 * The PFN range must belong to a single zone.
5994 * Returns zero on success or negative error code. On success all
5995 * pages which PFN is in [start, end) are allocated for the caller and
5996 * need to be freed with free_contig_range().
5998 int alloc_contig_range(unsigned long start, unsigned long end,
5999 unsigned migratetype)
6001 unsigned long outer_start, outer_end;
6004 struct compact_control cc = {
6005 .nr_migratepages = 0,
6007 .zone = page_zone(pfn_to_page(start)),
6009 .ignore_skip_hint = true,
6011 INIT_LIST_HEAD(&cc.migratepages);
6014 * What we do here is we mark all pageblocks in range as
6015 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6016 * have different sizes, and due to the way page allocator
6017 * work, we align the range to biggest of the two pages so
6018 * that page allocator won't try to merge buddies from
6019 * different pageblocks and change MIGRATE_ISOLATE to some
6020 * other migration type.
6022 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6023 * migrate the pages from an unaligned range (ie. pages that
6024 * we are interested in). This will put all the pages in
6025 * range back to page allocator as MIGRATE_ISOLATE.
6027 * When this is done, we take the pages in range from page
6028 * allocator removing them from the buddy system. This way
6029 * page allocator will never consider using them.
6031 * This lets us mark the pageblocks back as
6032 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6033 * aligned range but not in the unaligned, original range are
6034 * put back to page allocator so that buddy can use them.
6037 ret = start_isolate_page_range(pfn_max_align_down(start),
6038 pfn_max_align_up(end), migratetype,
6043 ret = __alloc_contig_migrate_range(&cc, start, end);
6048 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6049 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6050 * more, all pages in [start, end) are free in page allocator.
6051 * What we are going to do is to allocate all pages from
6052 * [start, end) (that is remove them from page allocator).
6054 * The only problem is that pages at the beginning and at the
6055 * end of interesting range may be not aligned with pages that
6056 * page allocator holds, ie. they can be part of higher order
6057 * pages. Because of this, we reserve the bigger range and
6058 * once this is done free the pages we are not interested in.
6060 * We don't have to hold zone->lock here because the pages are
6061 * isolated thus they won't get removed from buddy.
6064 lru_add_drain_all();
6068 outer_start = start;
6069 while (!PageBuddy(pfn_to_page(outer_start))) {
6070 if (++order >= MAX_ORDER) {
6074 outer_start &= ~0UL << order;
6077 /* Make sure the range is really isolated. */
6078 if (test_pages_isolated(outer_start, end, false)) {
6079 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6086 /* Grab isolated pages from freelists. */
6087 outer_end = isolate_freepages_range(&cc, outer_start, end);
6093 /* Free head and tail (if any) */
6094 if (start != outer_start)
6095 free_contig_range(outer_start, start - outer_start);
6096 if (end != outer_end)
6097 free_contig_range(end, outer_end - end);
6100 undo_isolate_page_range(pfn_max_align_down(start),
6101 pfn_max_align_up(end), migratetype);
6105 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6107 unsigned int count = 0;
6109 for (; nr_pages--; pfn++) {
6110 struct page *page = pfn_to_page(pfn);
6112 count += page_count(page) != 1;
6115 WARN(count != 0, "%d pages are still in use!\n", count);
6119 #ifdef CONFIG_MEMORY_HOTPLUG
6121 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6122 * page high values need to be recalulated.
6124 void __meminit zone_pcp_update(struct zone *zone)
6127 mutex_lock(&pcp_batch_high_lock);
6128 for_each_possible_cpu(cpu)
6129 pageset_set_high_and_batch(zone,
6130 per_cpu_ptr(zone->pageset, cpu));
6131 mutex_unlock(&pcp_batch_high_lock);
6135 void zone_pcp_reset(struct zone *zone)
6137 unsigned long flags;
6139 struct per_cpu_pageset *pset;
6141 /* avoid races with drain_pages() */
6142 local_irq_save(flags);
6143 if (zone->pageset != &boot_pageset) {
6144 for_each_online_cpu(cpu) {
6145 pset = per_cpu_ptr(zone->pageset, cpu);
6146 drain_zonestat(zone, pset);
6148 free_percpu(zone->pageset);
6149 zone->pageset = &boot_pageset;
6151 local_irq_restore(flags);
6154 #ifdef CONFIG_MEMORY_HOTREMOVE
6156 * All pages in the range must be isolated before calling this.
6159 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6165 unsigned long flags;
6166 /* find the first valid pfn */
6167 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6172 zone = page_zone(pfn_to_page(pfn));
6173 spin_lock_irqsave(&zone->lock, flags);
6175 while (pfn < end_pfn) {
6176 if (!pfn_valid(pfn)) {
6180 page = pfn_to_page(pfn);
6182 * The HWPoisoned page may be not in buddy system, and
6183 * page_count() is not 0.
6185 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6187 SetPageReserved(page);
6191 BUG_ON(page_count(page));
6192 BUG_ON(!PageBuddy(page));
6193 order = page_order(page);
6194 #ifdef CONFIG_DEBUG_VM
6195 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6196 pfn, 1 << order, end_pfn);
6198 list_del(&page->lru);
6199 rmv_page_order(page);
6200 zone->free_area[order].nr_free--;
6201 #ifdef CONFIG_HIGHMEM
6202 if (PageHighMem(page))
6203 totalhigh_pages -= 1 << order;
6205 for (i = 0; i < (1 << order); i++)
6206 SetPageReserved((page+i));
6207 pfn += (1 << order);
6209 spin_unlock_irqrestore(&zone->lock, flags);
6213 #ifdef CONFIG_MEMORY_FAILURE
6214 bool is_free_buddy_page(struct page *page)
6216 struct zone *zone = page_zone(page);
6217 unsigned long pfn = page_to_pfn(page);
6218 unsigned long flags;
6221 spin_lock_irqsave(&zone->lock, flags);
6222 for (order = 0; order < MAX_ORDER; order++) {
6223 struct page *page_head = page - (pfn & ((1 << order) - 1));
6225 if (PageBuddy(page_head) && page_order(page_head) >= order)
6228 spin_unlock_irqrestore(&zone->lock, flags);
6230 return order < MAX_ORDER;
6234 static const struct trace_print_flags pageflag_names[] = {
6235 {1UL << PG_locked, "locked" },
6236 {1UL << PG_error, "error" },
6237 {1UL << PG_referenced, "referenced" },
6238 {1UL << PG_uptodate, "uptodate" },
6239 {1UL << PG_dirty, "dirty" },
6240 {1UL << PG_lru, "lru" },
6241 {1UL << PG_active, "active" },
6242 {1UL << PG_slab, "slab" },
6243 {1UL << PG_owner_priv_1, "owner_priv_1" },
6244 {1UL << PG_arch_1, "arch_1" },
6245 {1UL << PG_reserved, "reserved" },
6246 {1UL << PG_private, "private" },
6247 {1UL << PG_private_2, "private_2" },
6248 {1UL << PG_writeback, "writeback" },
6249 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6250 {1UL << PG_head, "head" },
6251 {1UL << PG_tail, "tail" },
6253 {1UL << PG_compound, "compound" },
6255 {1UL << PG_swapcache, "swapcache" },
6256 {1UL << PG_mappedtodisk, "mappedtodisk" },
6257 {1UL << PG_reclaim, "reclaim" },
6258 {1UL << PG_swapbacked, "swapbacked" },
6259 {1UL << PG_unevictable, "unevictable" },
6261 {1UL << PG_mlocked, "mlocked" },
6263 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6264 {1UL << PG_uncached, "uncached" },
6266 #ifdef CONFIG_MEMORY_FAILURE
6267 {1UL << PG_hwpoison, "hwpoison" },
6269 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6270 {1UL << PG_compound_lock, "compound_lock" },
6274 static void dump_page_flags(unsigned long flags)
6276 const char *delim = "";
6280 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6282 printk(KERN_ALERT "page flags: %#lx(", flags);
6284 /* remove zone id */
6285 flags &= (1UL << NR_PAGEFLAGS) - 1;
6287 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6289 mask = pageflag_names[i].mask;
6290 if ((flags & mask) != mask)
6294 printk("%s%s", delim, pageflag_names[i].name);
6298 /* check for left over flags */
6300 printk("%s%#lx", delim, flags);
6305 void dump_page(struct page *page)
6308 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6309 page, atomic_read(&page->_count), page_mapcount(page),
6310 page->mapping, page->index);
6311 dump_page_flags(page->flags);
6312 mem_cgroup_print_bad_page(page);