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 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
69 DEFINE_PER_CPU(int, numa_node);
70 EXPORT_PER_CPU_SYMBOL(numa_node);
73 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
75 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
76 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
77 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
78 * defined in <linux/topology.h>.
80 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
81 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
85 * Array of node states.
87 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
88 [N_POSSIBLE] = NODE_MASK_ALL,
89 [N_ONLINE] = { { [0] = 1UL } },
91 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
93 [N_HIGH_MEMORY] = { { [0] = 1UL } },
95 #ifdef CONFIG_MOVABLE_NODE
96 [N_MEMORY] = { { [0] = 1UL } },
98 [N_CPU] = { { [0] = 1UL } },
101 EXPORT_SYMBOL(node_states);
103 unsigned long totalram_pages __read_mostly;
104 unsigned long totalreserve_pages __read_mostly;
106 * When calculating the number of globally allowed dirty pages, there
107 * is a certain number of per-zone reserves that should not be
108 * considered dirtyable memory. This is the sum of those reserves
109 * over all existing zones that contribute dirtyable memory.
111 unsigned long dirty_balance_reserve __read_mostly;
113 int percpu_pagelist_fraction;
114 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
116 #ifdef CONFIG_PM_SLEEP
118 * The following functions are used by the suspend/hibernate code to temporarily
119 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
120 * while devices are suspended. To avoid races with the suspend/hibernate code,
121 * they should always be called with pm_mutex held (gfp_allowed_mask also should
122 * only be modified with pm_mutex held, unless the suspend/hibernate code is
123 * guaranteed not to run in parallel with that modification).
126 static gfp_t saved_gfp_mask;
128 void pm_restore_gfp_mask(void)
130 WARN_ON(!mutex_is_locked(&pm_mutex));
131 if (saved_gfp_mask) {
132 gfp_allowed_mask = saved_gfp_mask;
137 void pm_restrict_gfp_mask(void)
139 WARN_ON(!mutex_is_locked(&pm_mutex));
140 WARN_ON(saved_gfp_mask);
141 saved_gfp_mask = gfp_allowed_mask;
142 gfp_allowed_mask &= ~GFP_IOFS;
145 bool pm_suspended_storage(void)
147 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
151 #endif /* CONFIG_PM_SLEEP */
153 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
154 int pageblock_order __read_mostly;
157 static void __free_pages_ok(struct page *page, unsigned int order);
160 * results with 256, 32 in the lowmem_reserve sysctl:
161 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
162 * 1G machine -> (16M dma, 784M normal, 224M high)
163 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
164 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
165 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
167 * TBD: should special case ZONE_DMA32 machines here - in those we normally
168 * don't need any ZONE_NORMAL reservation
170 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
171 #ifdef CONFIG_ZONE_DMA
174 #ifdef CONFIG_ZONE_DMA32
177 #ifdef CONFIG_HIGHMEM
183 EXPORT_SYMBOL(totalram_pages);
185 static char * const zone_names[MAX_NR_ZONES] = {
186 #ifdef CONFIG_ZONE_DMA
189 #ifdef CONFIG_ZONE_DMA32
193 #ifdef CONFIG_HIGHMEM
199 int min_free_kbytes = 1024;
201 static unsigned long __meminitdata nr_kernel_pages;
202 static unsigned long __meminitdata nr_all_pages;
203 static unsigned long __meminitdata dma_reserve;
205 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
206 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
207 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
208 static unsigned long __initdata required_kernelcore;
209 static unsigned long __initdata required_movablecore;
210 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
212 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
214 EXPORT_SYMBOL(movable_zone);
215 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
218 int nr_node_ids __read_mostly = MAX_NUMNODES;
219 int nr_online_nodes __read_mostly = 1;
220 EXPORT_SYMBOL(nr_node_ids);
221 EXPORT_SYMBOL(nr_online_nodes);
224 int page_group_by_mobility_disabled __read_mostly;
226 void set_pageblock_migratetype(struct page *page, int migratetype)
229 if (unlikely(page_group_by_mobility_disabled))
230 migratetype = MIGRATE_UNMOVABLE;
232 set_pageblock_flags_group(page, (unsigned long)migratetype,
233 PB_migrate, PB_migrate_end);
236 bool oom_killer_disabled __read_mostly;
238 #ifdef CONFIG_DEBUG_VM
239 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
243 unsigned long pfn = page_to_pfn(page);
244 unsigned long sp, start_pfn;
247 seq = zone_span_seqbegin(zone);
248 start_pfn = zone->zone_start_pfn;
249 sp = zone->spanned_pages;
250 if (!zone_spans_pfn(zone, pfn))
252 } while (zone_span_seqretry(zone, seq));
255 pr_err("page %lu outside zone [ %lu - %lu ]\n",
256 pfn, start_pfn, start_pfn + sp);
261 static int page_is_consistent(struct zone *zone, struct page *page)
263 if (!pfn_valid_within(page_to_pfn(page)))
265 if (zone != page_zone(page))
271 * Temporary debugging check for pages not lying within a given zone.
273 static int bad_range(struct zone *zone, struct page *page)
275 if (page_outside_zone_boundaries(zone, page))
277 if (!page_is_consistent(zone, page))
283 static inline int bad_range(struct zone *zone, struct page *page)
289 static void bad_page(struct page *page)
291 static unsigned long resume;
292 static unsigned long nr_shown;
293 static unsigned long nr_unshown;
295 /* Don't complain about poisoned pages */
296 if (PageHWPoison(page)) {
297 page_mapcount_reset(page); /* remove PageBuddy */
302 * Allow a burst of 60 reports, then keep quiet for that minute;
303 * or allow a steady drip of one report per second.
305 if (nr_shown == 60) {
306 if (time_before(jiffies, resume)) {
312 "BUG: Bad page state: %lu messages suppressed\n",
319 resume = jiffies + 60 * HZ;
321 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
322 current->comm, page_to_pfn(page));
328 /* Leave bad fields for debug, except PageBuddy could make trouble */
329 page_mapcount_reset(page); /* remove PageBuddy */
330 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
334 * Higher-order pages are called "compound pages". They are structured thusly:
336 * The first PAGE_SIZE page is called the "head page".
338 * The remaining PAGE_SIZE pages are called "tail pages".
340 * All pages have PG_compound set. All tail pages have their ->first_page
341 * pointing at the head page.
343 * The first tail page's ->lru.next holds the address of the compound page's
344 * put_page() function. Its ->lru.prev holds the order of allocation.
345 * This usage means that zero-order pages may not be compound.
348 static void free_compound_page(struct page *page)
350 __free_pages_ok(page, compound_order(page));
353 void prep_compound_page(struct page *page, unsigned long order)
356 int nr_pages = 1 << order;
358 set_compound_page_dtor(page, free_compound_page);
359 set_compound_order(page, order);
361 for (i = 1; i < nr_pages; i++) {
362 struct page *p = page + i;
364 set_page_count(p, 0);
365 p->first_page = page;
369 /* update __split_huge_page_refcount if you change this function */
370 static int destroy_compound_page(struct page *page, unsigned long order)
373 int nr_pages = 1 << order;
376 if (unlikely(compound_order(page) != order)) {
381 __ClearPageHead(page);
383 for (i = 1; i < nr_pages; i++) {
384 struct page *p = page + i;
386 if (unlikely(!PageTail(p) || (p->first_page != page))) {
396 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
401 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
402 * and __GFP_HIGHMEM from hard or soft interrupt context.
404 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
405 for (i = 0; i < (1 << order); i++)
406 clear_highpage(page + i);
409 #ifdef CONFIG_DEBUG_PAGEALLOC
410 unsigned int _debug_guardpage_minorder;
412 static int __init debug_guardpage_minorder_setup(char *buf)
416 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
417 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
420 _debug_guardpage_minorder = res;
421 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
424 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
426 static inline void set_page_guard_flag(struct page *page)
428 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
431 static inline void clear_page_guard_flag(struct page *page)
433 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
436 static inline void set_page_guard_flag(struct page *page) { }
437 static inline void clear_page_guard_flag(struct page *page) { }
440 static inline void set_page_order(struct page *page, int order)
442 set_page_private(page, order);
443 __SetPageBuddy(page);
446 static inline void rmv_page_order(struct page *page)
448 __ClearPageBuddy(page);
449 set_page_private(page, 0);
453 * Locate the struct page for both the matching buddy in our
454 * pair (buddy1) and the combined O(n+1) page they form (page).
456 * 1) Any buddy B1 will have an order O twin B2 which satisfies
457 * the following equation:
459 * For example, if the starting buddy (buddy2) is #8 its order
461 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
463 * 2) Any buddy B will have an order O+1 parent P which
464 * satisfies the following equation:
467 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
469 static inline unsigned long
470 __find_buddy_index(unsigned long page_idx, unsigned int order)
472 return page_idx ^ (1 << order);
476 * This function checks whether a page is free && is the buddy
477 * we can do coalesce a page and its buddy if
478 * (a) the buddy is not in a hole &&
479 * (b) the buddy is in the buddy system &&
480 * (c) a page and its buddy have the same order &&
481 * (d) a page and its buddy are in the same zone.
483 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
484 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
486 * For recording page's order, we use page_private(page).
488 static inline int page_is_buddy(struct page *page, struct page *buddy,
491 if (!pfn_valid_within(page_to_pfn(buddy)))
494 if (page_zone_id(page) != page_zone_id(buddy))
497 if (page_is_guard(buddy) && page_order(buddy) == order) {
498 VM_BUG_ON(page_count(buddy) != 0);
502 if (PageBuddy(buddy) && page_order(buddy) == order) {
503 VM_BUG_ON(page_count(buddy) != 0);
510 * Freeing function for a buddy system allocator.
512 * The concept of a buddy system is to maintain direct-mapped table
513 * (containing bit values) for memory blocks of various "orders".
514 * The bottom level table contains the map for the smallest allocatable
515 * units of memory (here, pages), and each level above it describes
516 * pairs of units from the levels below, hence, "buddies".
517 * At a high level, all that happens here is marking the table entry
518 * at the bottom level available, and propagating the changes upward
519 * as necessary, plus some accounting needed to play nicely with other
520 * parts of the VM system.
521 * At each level, we keep a list of pages, which are heads of continuous
522 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
523 * order is recorded in page_private(page) field.
524 * So when we are allocating or freeing one, we can derive the state of the
525 * other. That is, if we allocate a small block, and both were
526 * free, the remainder of the region must be split into blocks.
527 * If a block is freed, and its buddy is also free, then this
528 * triggers coalescing into a block of larger size.
533 static inline void __free_one_page(struct page *page,
534 struct zone *zone, unsigned int order,
537 unsigned long page_idx;
538 unsigned long combined_idx;
539 unsigned long uninitialized_var(buddy_idx);
542 VM_BUG_ON(!zone_is_initialized(zone));
544 if (unlikely(PageCompound(page)))
545 if (unlikely(destroy_compound_page(page, order)))
548 VM_BUG_ON(migratetype == -1);
550 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
552 VM_BUG_ON(page_idx & ((1 << order) - 1));
553 VM_BUG_ON(bad_range(zone, page));
555 while (order < MAX_ORDER-1) {
556 buddy_idx = __find_buddy_index(page_idx, order);
557 buddy = page + (buddy_idx - page_idx);
558 if (!page_is_buddy(page, buddy, order))
561 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
562 * merge with it and move up one order.
564 if (page_is_guard(buddy)) {
565 clear_page_guard_flag(buddy);
566 set_page_private(page, 0);
567 __mod_zone_freepage_state(zone, 1 << order,
570 list_del(&buddy->lru);
571 zone->free_area[order].nr_free--;
572 rmv_page_order(buddy);
574 combined_idx = buddy_idx & page_idx;
575 page = page + (combined_idx - page_idx);
576 page_idx = combined_idx;
579 set_page_order(page, order);
582 * If this is not the largest possible page, check if the buddy
583 * of the next-highest order is free. If it is, it's possible
584 * that pages are being freed that will coalesce soon. In case,
585 * that is happening, add the free page to the tail of the list
586 * so it's less likely to be used soon and more likely to be merged
587 * as a higher order page
589 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
590 struct page *higher_page, *higher_buddy;
591 combined_idx = buddy_idx & page_idx;
592 higher_page = page + (combined_idx - page_idx);
593 buddy_idx = __find_buddy_index(combined_idx, order + 1);
594 higher_buddy = higher_page + (buddy_idx - combined_idx);
595 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
596 list_add_tail(&page->lru,
597 &zone->free_area[order].free_list[migratetype]);
602 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
604 zone->free_area[order].nr_free++;
607 static inline int free_pages_check(struct page *page)
609 if (unlikely(page_mapcount(page) |
610 (page->mapping != NULL) |
611 (atomic_read(&page->_count) != 0) |
612 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
613 (mem_cgroup_bad_page_check(page)))) {
617 page_nid_reset_last(page);
618 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
619 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
624 * Frees a number of pages from the PCP lists
625 * Assumes all pages on list are in same zone, and of same order.
626 * count is the number of pages to free.
628 * If the zone was previously in an "all pages pinned" state then look to
629 * see if this freeing clears that state.
631 * And clear the zone's pages_scanned counter, to hold off the "all pages are
632 * pinned" detection logic.
634 static void free_pcppages_bulk(struct zone *zone, int count,
635 struct per_cpu_pages *pcp)
641 spin_lock(&zone->lock);
642 zone->all_unreclaimable = 0;
643 zone->pages_scanned = 0;
647 struct list_head *list;
650 * Remove pages from lists in a round-robin fashion. A
651 * batch_free count is maintained that is incremented when an
652 * empty list is encountered. This is so more pages are freed
653 * off fuller lists instead of spinning excessively around empty
658 if (++migratetype == MIGRATE_PCPTYPES)
660 list = &pcp->lists[migratetype];
661 } while (list_empty(list));
663 /* This is the only non-empty list. Free them all. */
664 if (batch_free == MIGRATE_PCPTYPES)
665 batch_free = to_free;
668 int mt; /* migratetype of the to-be-freed page */
670 page = list_entry(list->prev, struct page, lru);
671 /* must delete as __free_one_page list manipulates */
672 list_del(&page->lru);
673 mt = get_freepage_migratetype(page);
674 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
675 __free_one_page(page, zone, 0, mt);
676 trace_mm_page_pcpu_drain(page, 0, mt);
677 if (likely(!is_migrate_isolate_page(page))) {
678 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
679 if (is_migrate_cma(mt))
680 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
682 } while (--to_free && --batch_free && !list_empty(list));
684 spin_unlock(&zone->lock);
687 static void free_one_page(struct zone *zone, struct page *page, int order,
690 spin_lock(&zone->lock);
691 zone->all_unreclaimable = 0;
692 zone->pages_scanned = 0;
694 __free_one_page(page, zone, order, migratetype);
695 if (unlikely(!is_migrate_isolate(migratetype)))
696 __mod_zone_freepage_state(zone, 1 << order, migratetype);
697 spin_unlock(&zone->lock);
700 static bool free_pages_prepare(struct page *page, unsigned int order)
705 trace_mm_page_free(page, order);
706 kmemcheck_free_shadow(page, order);
709 page->mapping = NULL;
710 for (i = 0; i < (1 << order); i++)
711 bad += free_pages_check(page + i);
715 if (!PageHighMem(page)) {
716 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
717 debug_check_no_obj_freed(page_address(page),
720 arch_free_page(page, order);
721 kernel_map_pages(page, 1 << order, 0);
726 static void __free_pages_ok(struct page *page, unsigned int order)
731 if (!free_pages_prepare(page, order))
734 local_irq_save(flags);
735 __count_vm_events(PGFREE, 1 << order);
736 migratetype = get_pageblock_migratetype(page);
737 set_freepage_migratetype(page, migratetype);
738 free_one_page(page_zone(page), page, order, migratetype);
739 local_irq_restore(flags);
743 * Read access to zone->managed_pages is safe because it's unsigned long,
744 * but we still need to serialize writers. Currently all callers of
745 * __free_pages_bootmem() except put_page_bootmem() should only be used
746 * at boot time. So for shorter boot time, we shift the burden to
747 * put_page_bootmem() to serialize writers.
749 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
751 unsigned int nr_pages = 1 << order;
755 for (loop = 0; loop < nr_pages; loop++) {
756 struct page *p = &page[loop];
758 if (loop + 1 < nr_pages)
760 __ClearPageReserved(p);
761 set_page_count(p, 0);
764 page_zone(page)->managed_pages += 1 << order;
765 set_page_refcounted(page);
766 __free_pages(page, order);
770 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
771 void __init init_cma_reserved_pageblock(struct page *page)
773 unsigned i = pageblock_nr_pages;
774 struct page *p = page;
777 __ClearPageReserved(p);
778 set_page_count(p, 0);
781 set_page_refcounted(page);
782 set_pageblock_migratetype(page, MIGRATE_CMA);
783 __free_pages(page, pageblock_order);
784 totalram_pages += pageblock_nr_pages;
785 #ifdef CONFIG_HIGHMEM
786 if (PageHighMem(page))
787 totalhigh_pages += pageblock_nr_pages;
793 * The order of subdivision here is critical for the IO subsystem.
794 * Please do not alter this order without good reasons and regression
795 * testing. Specifically, as large blocks of memory are subdivided,
796 * the order in which smaller blocks are delivered depends on the order
797 * they're subdivided in this function. This is the primary factor
798 * influencing the order in which pages are delivered to the IO
799 * subsystem according to empirical testing, and this is also justified
800 * by considering the behavior of a buddy system containing a single
801 * large block of memory acted on by a series of small allocations.
802 * This behavior is a critical factor in sglist merging's success.
806 static inline void expand(struct zone *zone, struct page *page,
807 int low, int high, struct free_area *area,
810 unsigned long size = 1 << high;
816 VM_BUG_ON(bad_range(zone, &page[size]));
818 #ifdef CONFIG_DEBUG_PAGEALLOC
819 if (high < debug_guardpage_minorder()) {
821 * Mark as guard pages (or page), that will allow to
822 * merge back to allocator when buddy will be freed.
823 * Corresponding page table entries will not be touched,
824 * pages will stay not present in virtual address space
826 INIT_LIST_HEAD(&page[size].lru);
827 set_page_guard_flag(&page[size]);
828 set_page_private(&page[size], high);
829 /* Guard pages are not available for any usage */
830 __mod_zone_freepage_state(zone, -(1 << high),
835 list_add(&page[size].lru, &area->free_list[migratetype]);
837 set_page_order(&page[size], high);
842 * This page is about to be returned from the page allocator
844 static inline int check_new_page(struct page *page)
846 if (unlikely(page_mapcount(page) |
847 (page->mapping != NULL) |
848 (atomic_read(&page->_count) != 0) |
849 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
850 (mem_cgroup_bad_page_check(page)))) {
857 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
861 for (i = 0; i < (1 << order); i++) {
862 struct page *p = page + i;
863 if (unlikely(check_new_page(p)))
867 set_page_private(page, 0);
868 set_page_refcounted(page);
870 arch_alloc_page(page, order);
871 kernel_map_pages(page, 1 << order, 1);
873 if (gfp_flags & __GFP_ZERO)
874 prep_zero_page(page, order, gfp_flags);
876 if (order && (gfp_flags & __GFP_COMP))
877 prep_compound_page(page, order);
883 * Go through the free lists for the given migratetype and remove
884 * the smallest available page from the freelists
887 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
890 unsigned int current_order;
891 struct free_area * area;
894 /* Find a page of the appropriate size in the preferred list */
895 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
896 area = &(zone->free_area[current_order]);
897 if (list_empty(&area->free_list[migratetype]))
900 page = list_entry(area->free_list[migratetype].next,
902 list_del(&page->lru);
903 rmv_page_order(page);
905 expand(zone, page, order, current_order, area, migratetype);
914 * This array describes the order lists are fallen back to when
915 * the free lists for the desirable migrate type are depleted
917 static int fallbacks[MIGRATE_TYPES][4] = {
918 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
919 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
921 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
922 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
924 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
926 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
927 #ifdef CONFIG_MEMORY_ISOLATION
928 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
933 * Move the free pages in a range to the free lists of the requested type.
934 * Note that start_page and end_pages are not aligned on a pageblock
935 * boundary. If alignment is required, use move_freepages_block()
937 int move_freepages(struct zone *zone,
938 struct page *start_page, struct page *end_page,
945 #ifndef CONFIG_HOLES_IN_ZONE
947 * page_zone is not safe to call in this context when
948 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
949 * anyway as we check zone boundaries in move_freepages_block().
950 * Remove at a later date when no bug reports exist related to
951 * grouping pages by mobility
953 BUG_ON(page_zone(start_page) != page_zone(end_page));
956 for (page = start_page; page <= end_page;) {
957 /* Make sure we are not inadvertently changing nodes */
958 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
960 if (!pfn_valid_within(page_to_pfn(page))) {
965 if (!PageBuddy(page)) {
970 order = page_order(page);
971 list_move(&page->lru,
972 &zone->free_area[order].free_list[migratetype]);
973 set_freepage_migratetype(page, migratetype);
975 pages_moved += 1 << order;
981 int move_freepages_block(struct zone *zone, struct page *page,
984 unsigned long start_pfn, end_pfn;
985 struct page *start_page, *end_page;
987 start_pfn = page_to_pfn(page);
988 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
989 start_page = pfn_to_page(start_pfn);
990 end_page = start_page + pageblock_nr_pages - 1;
991 end_pfn = start_pfn + pageblock_nr_pages - 1;
993 /* Do not cross zone boundaries */
994 if (!zone_spans_pfn(zone, start_pfn))
996 if (!zone_spans_pfn(zone, end_pfn))
999 return move_freepages(zone, start_page, end_page, migratetype);
1002 static void change_pageblock_range(struct page *pageblock_page,
1003 int start_order, int migratetype)
1005 int nr_pageblocks = 1 << (start_order - pageblock_order);
1007 while (nr_pageblocks--) {
1008 set_pageblock_migratetype(pageblock_page, migratetype);
1009 pageblock_page += pageblock_nr_pages;
1013 /* Remove an element from the buddy allocator from the fallback list */
1014 static inline struct page *
1015 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1017 struct free_area * area;
1022 /* Find the largest possible block of pages in the other list */
1023 for (current_order = MAX_ORDER-1; current_order >= order;
1026 migratetype = fallbacks[start_migratetype][i];
1028 /* MIGRATE_RESERVE handled later if necessary */
1029 if (migratetype == MIGRATE_RESERVE)
1032 area = &(zone->free_area[current_order]);
1033 if (list_empty(&area->free_list[migratetype]))
1036 page = list_entry(area->free_list[migratetype].next,
1041 * If breaking a large block of pages, move all free
1042 * pages to the preferred allocation list. If falling
1043 * back for a reclaimable kernel allocation, be more
1044 * aggressive about taking ownership of free pages
1046 * On the other hand, never change migration
1047 * type of MIGRATE_CMA pageblocks nor move CMA
1048 * pages on different free lists. We don't
1049 * want unmovable pages to be allocated from
1050 * MIGRATE_CMA areas.
1052 if (!is_migrate_cma(migratetype) &&
1053 (unlikely(current_order >= pageblock_order / 2) ||
1054 start_migratetype == MIGRATE_RECLAIMABLE ||
1055 page_group_by_mobility_disabled)) {
1057 pages = move_freepages_block(zone, page,
1060 /* Claim the whole block if over half of it is free */
1061 if (pages >= (1 << (pageblock_order-1)) ||
1062 page_group_by_mobility_disabled)
1063 set_pageblock_migratetype(page,
1066 migratetype = start_migratetype;
1069 /* Remove the page from the freelists */
1070 list_del(&page->lru);
1071 rmv_page_order(page);
1073 /* Take ownership for orders >= pageblock_order */
1074 if (current_order >= pageblock_order &&
1075 !is_migrate_cma(migratetype))
1076 change_pageblock_range(page, current_order,
1079 expand(zone, page, order, current_order, area,
1080 is_migrate_cma(migratetype)
1081 ? migratetype : start_migratetype);
1083 trace_mm_page_alloc_extfrag(page, order, current_order,
1084 start_migratetype, migratetype);
1094 * Do the hard work of removing an element from the buddy allocator.
1095 * Call me with the zone->lock already held.
1097 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1103 page = __rmqueue_smallest(zone, order, migratetype);
1105 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1106 page = __rmqueue_fallback(zone, order, migratetype);
1109 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1110 * is used because __rmqueue_smallest is an inline function
1111 * and we want just one call site
1114 migratetype = MIGRATE_RESERVE;
1119 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1124 * Obtain a specified number of elements from the buddy allocator, all under
1125 * a single hold of the lock, for efficiency. Add them to the supplied list.
1126 * Returns the number of new pages which were placed at *list.
1128 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1129 unsigned long count, struct list_head *list,
1130 int migratetype, int cold)
1132 int mt = migratetype, i;
1134 spin_lock(&zone->lock);
1135 for (i = 0; i < count; ++i) {
1136 struct page *page = __rmqueue(zone, order, migratetype);
1137 if (unlikely(page == NULL))
1141 * Split buddy pages returned by expand() are received here
1142 * in physical page order. The page is added to the callers and
1143 * list and the list head then moves forward. From the callers
1144 * perspective, the linked list is ordered by page number in
1145 * some conditions. This is useful for IO devices that can
1146 * merge IO requests if the physical pages are ordered
1149 if (likely(cold == 0))
1150 list_add(&page->lru, list);
1152 list_add_tail(&page->lru, list);
1153 if (IS_ENABLED(CONFIG_CMA)) {
1154 mt = get_pageblock_migratetype(page);
1155 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1158 set_freepage_migratetype(page, mt);
1160 if (is_migrate_cma(mt))
1161 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1164 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1165 spin_unlock(&zone->lock);
1171 * Called from the vmstat counter updater to drain pagesets of this
1172 * currently executing processor on remote nodes after they have
1175 * Note that this function must be called with the thread pinned to
1176 * a single processor.
1178 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1180 unsigned long flags;
1183 local_irq_save(flags);
1184 if (pcp->count >= pcp->batch)
1185 to_drain = pcp->batch;
1187 to_drain = pcp->count;
1189 free_pcppages_bulk(zone, to_drain, pcp);
1190 pcp->count -= to_drain;
1192 local_irq_restore(flags);
1197 * Drain pages of the indicated processor.
1199 * The processor must either be the current processor and the
1200 * thread pinned to the current processor or a processor that
1203 static void drain_pages(unsigned int cpu)
1205 unsigned long flags;
1208 for_each_populated_zone(zone) {
1209 struct per_cpu_pageset *pset;
1210 struct per_cpu_pages *pcp;
1212 local_irq_save(flags);
1213 pset = per_cpu_ptr(zone->pageset, cpu);
1217 free_pcppages_bulk(zone, pcp->count, pcp);
1220 local_irq_restore(flags);
1225 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1227 void drain_local_pages(void *arg)
1229 drain_pages(smp_processor_id());
1233 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1235 * Note that this code is protected against sending an IPI to an offline
1236 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1237 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1238 * nothing keeps CPUs from showing up after we populated the cpumask and
1239 * before the call to on_each_cpu_mask().
1241 void drain_all_pages(void)
1244 struct per_cpu_pageset *pcp;
1248 * Allocate in the BSS so we wont require allocation in
1249 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1251 static cpumask_t cpus_with_pcps;
1254 * We don't care about racing with CPU hotplug event
1255 * as offline notification will cause the notified
1256 * cpu to drain that CPU pcps and on_each_cpu_mask
1257 * disables preemption as part of its processing
1259 for_each_online_cpu(cpu) {
1260 bool has_pcps = false;
1261 for_each_populated_zone(zone) {
1262 pcp = per_cpu_ptr(zone->pageset, cpu);
1263 if (pcp->pcp.count) {
1269 cpumask_set_cpu(cpu, &cpus_with_pcps);
1271 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1273 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1276 #ifdef CONFIG_HIBERNATION
1278 void mark_free_pages(struct zone *zone)
1280 unsigned long pfn, max_zone_pfn;
1281 unsigned long flags;
1283 struct list_head *curr;
1285 if (!zone->spanned_pages)
1288 spin_lock_irqsave(&zone->lock, flags);
1290 max_zone_pfn = zone_end_pfn(zone);
1291 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1292 if (pfn_valid(pfn)) {
1293 struct page *page = pfn_to_page(pfn);
1295 if (!swsusp_page_is_forbidden(page))
1296 swsusp_unset_page_free(page);
1299 for_each_migratetype_order(order, t) {
1300 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1303 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1304 for (i = 0; i < (1UL << order); i++)
1305 swsusp_set_page_free(pfn_to_page(pfn + i));
1308 spin_unlock_irqrestore(&zone->lock, flags);
1310 #endif /* CONFIG_PM */
1313 * Free a 0-order page
1314 * cold == 1 ? free a cold page : free a hot page
1316 void free_hot_cold_page(struct page *page, int cold)
1318 struct zone *zone = page_zone(page);
1319 struct per_cpu_pages *pcp;
1320 unsigned long flags;
1323 if (!free_pages_prepare(page, 0))
1326 migratetype = get_pageblock_migratetype(page);
1327 set_freepage_migratetype(page, migratetype);
1328 local_irq_save(flags);
1329 __count_vm_event(PGFREE);
1332 * We only track unmovable, reclaimable and movable on pcp lists.
1333 * Free ISOLATE pages back to the allocator because they are being
1334 * offlined but treat RESERVE as movable pages so we can get those
1335 * areas back if necessary. Otherwise, we may have to free
1336 * excessively into the page allocator
1338 if (migratetype >= MIGRATE_PCPTYPES) {
1339 if (unlikely(is_migrate_isolate(migratetype))) {
1340 free_one_page(zone, page, 0, migratetype);
1343 migratetype = MIGRATE_MOVABLE;
1346 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1348 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1350 list_add(&page->lru, &pcp->lists[migratetype]);
1352 if (pcp->count >= pcp->high) {
1353 free_pcppages_bulk(zone, pcp->batch, pcp);
1354 pcp->count -= pcp->batch;
1358 local_irq_restore(flags);
1362 * Free a list of 0-order pages
1364 void free_hot_cold_page_list(struct list_head *list, int cold)
1366 struct page *page, *next;
1368 list_for_each_entry_safe(page, next, list, lru) {
1369 trace_mm_page_free_batched(page, cold);
1370 free_hot_cold_page(page, cold);
1375 * split_page takes a non-compound higher-order page, and splits it into
1376 * n (1<<order) sub-pages: page[0..n]
1377 * Each sub-page must be freed individually.
1379 * Note: this is probably too low level an operation for use in drivers.
1380 * Please consult with lkml before using this in your driver.
1382 void split_page(struct page *page, unsigned int order)
1386 VM_BUG_ON(PageCompound(page));
1387 VM_BUG_ON(!page_count(page));
1389 #ifdef CONFIG_KMEMCHECK
1391 * Split shadow pages too, because free(page[0]) would
1392 * otherwise free the whole shadow.
1394 if (kmemcheck_page_is_tracked(page))
1395 split_page(virt_to_page(page[0].shadow), order);
1398 for (i = 1; i < (1 << order); i++)
1399 set_page_refcounted(page + i);
1401 EXPORT_SYMBOL_GPL(split_page);
1403 static int __isolate_free_page(struct page *page, unsigned int order)
1405 unsigned long watermark;
1409 BUG_ON(!PageBuddy(page));
1411 zone = page_zone(page);
1412 mt = get_pageblock_migratetype(page);
1414 if (!is_migrate_isolate(mt)) {
1415 /* Obey watermarks as if the page was being allocated */
1416 watermark = low_wmark_pages(zone) + (1 << order);
1417 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1420 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1423 /* Remove page from free list */
1424 list_del(&page->lru);
1425 zone->free_area[order].nr_free--;
1426 rmv_page_order(page);
1428 /* Set the pageblock if the isolated page is at least a pageblock */
1429 if (order >= pageblock_order - 1) {
1430 struct page *endpage = page + (1 << order) - 1;
1431 for (; page < endpage; page += pageblock_nr_pages) {
1432 int mt = get_pageblock_migratetype(page);
1433 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1434 set_pageblock_migratetype(page,
1439 return 1UL << order;
1443 * Similar to split_page except the page is already free. As this is only
1444 * being used for migration, the migratetype of the block also changes.
1445 * As this is called with interrupts disabled, the caller is responsible
1446 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1449 * Note: this is probably too low level an operation for use in drivers.
1450 * Please consult with lkml before using this in your driver.
1452 int split_free_page(struct page *page)
1457 order = page_order(page);
1459 nr_pages = __isolate_free_page(page, order);
1463 /* Split into individual pages */
1464 set_page_refcounted(page);
1465 split_page(page, order);
1470 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1471 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1475 struct page *buffered_rmqueue(struct zone *preferred_zone,
1476 struct zone *zone, int order, gfp_t gfp_flags,
1479 unsigned long flags;
1481 int cold = !!(gfp_flags & __GFP_COLD);
1484 if (likely(order == 0)) {
1485 struct per_cpu_pages *pcp;
1486 struct list_head *list;
1488 local_irq_save(flags);
1489 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1490 list = &pcp->lists[migratetype];
1491 if (list_empty(list)) {
1492 pcp->count += rmqueue_bulk(zone, 0,
1495 if (unlikely(list_empty(list)))
1500 page = list_entry(list->prev, struct page, lru);
1502 page = list_entry(list->next, struct page, lru);
1504 list_del(&page->lru);
1507 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1509 * __GFP_NOFAIL is not to be used in new code.
1511 * All __GFP_NOFAIL callers should be fixed so that they
1512 * properly detect and handle allocation failures.
1514 * We most definitely don't want callers attempting to
1515 * allocate greater than order-1 page units with
1518 WARN_ON_ONCE(order > 1);
1520 spin_lock_irqsave(&zone->lock, flags);
1521 page = __rmqueue(zone, order, migratetype);
1522 spin_unlock(&zone->lock);
1525 __mod_zone_freepage_state(zone, -(1 << order),
1526 get_pageblock_migratetype(page));
1529 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1530 zone_statistics(preferred_zone, zone, gfp_flags);
1531 local_irq_restore(flags);
1533 VM_BUG_ON(bad_range(zone, page));
1534 if (prep_new_page(page, order, gfp_flags))
1539 local_irq_restore(flags);
1543 #ifdef CONFIG_FAIL_PAGE_ALLOC
1546 struct fault_attr attr;
1548 u32 ignore_gfp_highmem;
1549 u32 ignore_gfp_wait;
1551 } fail_page_alloc = {
1552 .attr = FAULT_ATTR_INITIALIZER,
1553 .ignore_gfp_wait = 1,
1554 .ignore_gfp_highmem = 1,
1558 static int __init setup_fail_page_alloc(char *str)
1560 return setup_fault_attr(&fail_page_alloc.attr, str);
1562 __setup("fail_page_alloc=", setup_fail_page_alloc);
1564 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1566 if (order < fail_page_alloc.min_order)
1568 if (gfp_mask & __GFP_NOFAIL)
1570 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1572 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1575 return should_fail(&fail_page_alloc.attr, 1 << order);
1578 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1580 static int __init fail_page_alloc_debugfs(void)
1582 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1585 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1586 &fail_page_alloc.attr);
1588 return PTR_ERR(dir);
1590 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1591 &fail_page_alloc.ignore_gfp_wait))
1593 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1594 &fail_page_alloc.ignore_gfp_highmem))
1596 if (!debugfs_create_u32("min-order", mode, dir,
1597 &fail_page_alloc.min_order))
1602 debugfs_remove_recursive(dir);
1607 late_initcall(fail_page_alloc_debugfs);
1609 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1611 #else /* CONFIG_FAIL_PAGE_ALLOC */
1613 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1618 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1621 * Return true if free pages are above 'mark'. This takes into account the order
1622 * of the allocation.
1624 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1625 int classzone_idx, int alloc_flags, long free_pages)
1627 /* free_pages my go negative - that's OK */
1629 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1632 free_pages -= (1 << order) - 1;
1633 if (alloc_flags & ALLOC_HIGH)
1635 if (alloc_flags & ALLOC_HARDER)
1638 /* If allocation can't use CMA areas don't use free CMA pages */
1639 if (!(alloc_flags & ALLOC_CMA))
1640 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1642 if (free_pages <= min + lowmem_reserve)
1644 for (o = 0; o < order; o++) {
1645 /* At the next order, this order's pages become unavailable */
1646 free_pages -= z->free_area[o].nr_free << o;
1648 /* Require fewer higher order pages to be free */
1651 if (free_pages <= min)
1657 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1658 int classzone_idx, int alloc_flags)
1660 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1661 zone_page_state(z, NR_FREE_PAGES));
1664 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1665 int classzone_idx, int alloc_flags)
1667 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1669 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1670 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1672 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1678 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1679 * skip over zones that are not allowed by the cpuset, or that have
1680 * been recently (in last second) found to be nearly full. See further
1681 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1682 * that have to skip over a lot of full or unallowed zones.
1684 * If the zonelist cache is present in the passed in zonelist, then
1685 * returns a pointer to the allowed node mask (either the current
1686 * tasks mems_allowed, or node_states[N_MEMORY].)
1688 * If the zonelist cache is not available for this zonelist, does
1689 * nothing and returns NULL.
1691 * If the fullzones BITMAP in the zonelist cache is stale (more than
1692 * a second since last zap'd) then we zap it out (clear its bits.)
1694 * We hold off even calling zlc_setup, until after we've checked the
1695 * first zone in the zonelist, on the theory that most allocations will
1696 * be satisfied from that first zone, so best to examine that zone as
1697 * quickly as we can.
1699 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1701 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1702 nodemask_t *allowednodes; /* zonelist_cache approximation */
1704 zlc = zonelist->zlcache_ptr;
1708 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1709 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1710 zlc->last_full_zap = jiffies;
1713 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1714 &cpuset_current_mems_allowed :
1715 &node_states[N_MEMORY];
1716 return allowednodes;
1720 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1721 * if it is worth looking at further for free memory:
1722 * 1) Check that the zone isn't thought to be full (doesn't have its
1723 * bit set in the zonelist_cache fullzones BITMAP).
1724 * 2) Check that the zones node (obtained from the zonelist_cache
1725 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1726 * Return true (non-zero) if zone is worth looking at further, or
1727 * else return false (zero) if it is not.
1729 * This check -ignores- the distinction between various watermarks,
1730 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1731 * found to be full for any variation of these watermarks, it will
1732 * be considered full for up to one second by all requests, unless
1733 * we are so low on memory on all allowed nodes that we are forced
1734 * into the second scan of the zonelist.
1736 * In the second scan we ignore this zonelist cache and exactly
1737 * apply the watermarks to all zones, even it is slower to do so.
1738 * We are low on memory in the second scan, and should leave no stone
1739 * unturned looking for a free page.
1741 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1742 nodemask_t *allowednodes)
1744 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1745 int i; /* index of *z in zonelist zones */
1746 int n; /* node that zone *z is on */
1748 zlc = zonelist->zlcache_ptr;
1752 i = z - zonelist->_zonerefs;
1755 /* This zone is worth trying if it is allowed but not full */
1756 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1760 * Given 'z' scanning a zonelist, set the corresponding bit in
1761 * zlc->fullzones, so that subsequent attempts to allocate a page
1762 * from that zone don't waste time re-examining it.
1764 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1766 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1767 int i; /* index of *z in zonelist zones */
1769 zlc = zonelist->zlcache_ptr;
1773 i = z - zonelist->_zonerefs;
1775 set_bit(i, zlc->fullzones);
1779 * clear all zones full, called after direct reclaim makes progress so that
1780 * a zone that was recently full is not skipped over for up to a second
1782 static void zlc_clear_zones_full(struct zonelist *zonelist)
1784 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1786 zlc = zonelist->zlcache_ptr;
1790 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1793 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1795 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1798 static void __paginginit init_zone_allows_reclaim(int nid)
1802 for_each_online_node(i)
1803 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1804 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1806 zone_reclaim_mode = 1;
1809 #else /* CONFIG_NUMA */
1811 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1816 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1817 nodemask_t *allowednodes)
1822 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1826 static void zlc_clear_zones_full(struct zonelist *zonelist)
1830 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1835 static inline void init_zone_allows_reclaim(int nid)
1838 #endif /* CONFIG_NUMA */
1841 * get_page_from_freelist goes through the zonelist trying to allocate
1844 static struct page *
1845 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1846 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1847 struct zone *preferred_zone, int migratetype)
1850 struct page *page = NULL;
1853 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1854 int zlc_active = 0; /* set if using zonelist_cache */
1855 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1857 classzone_idx = zone_idx(preferred_zone);
1860 * Scan zonelist, looking for a zone with enough free.
1861 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1863 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1864 high_zoneidx, nodemask) {
1865 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1866 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1868 if ((alloc_flags & ALLOC_CPUSET) &&
1869 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1872 * When allocating a page cache page for writing, we
1873 * want to get it from a zone that is within its dirty
1874 * limit, such that no single zone holds more than its
1875 * proportional share of globally allowed dirty pages.
1876 * The dirty limits take into account the zone's
1877 * lowmem reserves and high watermark so that kswapd
1878 * should be able to balance it without having to
1879 * write pages from its LRU list.
1881 * This may look like it could increase pressure on
1882 * lower zones by failing allocations in higher zones
1883 * before they are full. But the pages that do spill
1884 * over are limited as the lower zones are protected
1885 * by this very same mechanism. It should not become
1886 * a practical burden to them.
1888 * XXX: For now, allow allocations to potentially
1889 * exceed the per-zone dirty limit in the slowpath
1890 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1891 * which is important when on a NUMA setup the allowed
1892 * zones are together not big enough to reach the
1893 * global limit. The proper fix for these situations
1894 * will require awareness of zones in the
1895 * dirty-throttling and the flusher threads.
1897 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1898 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1899 goto this_zone_full;
1901 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1902 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1906 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1907 if (zone_watermark_ok(zone, order, mark,
1908 classzone_idx, alloc_flags))
1911 if (IS_ENABLED(CONFIG_NUMA) &&
1912 !did_zlc_setup && nr_online_nodes > 1) {
1914 * we do zlc_setup if there are multiple nodes
1915 * and before considering the first zone allowed
1918 allowednodes = zlc_setup(zonelist, alloc_flags);
1923 if (zone_reclaim_mode == 0 ||
1924 !zone_allows_reclaim(preferred_zone, zone))
1925 goto this_zone_full;
1928 * As we may have just activated ZLC, check if the first
1929 * eligible zone has failed zone_reclaim recently.
1931 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1932 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1935 ret = zone_reclaim(zone, gfp_mask, order);
1937 case ZONE_RECLAIM_NOSCAN:
1940 case ZONE_RECLAIM_FULL:
1941 /* scanned but unreclaimable */
1944 /* did we reclaim enough */
1945 if (zone_watermark_ok(zone, order, mark,
1946 classzone_idx, alloc_flags))
1950 * Failed to reclaim enough to meet watermark.
1951 * Only mark the zone full if checking the min
1952 * watermark or if we failed to reclaim just
1953 * 1<<order pages or else the page allocator
1954 * fastpath will prematurely mark zones full
1955 * when the watermark is between the low and
1958 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
1959 ret == ZONE_RECLAIM_SOME)
1960 goto this_zone_full;
1967 page = buffered_rmqueue(preferred_zone, zone, order,
1968 gfp_mask, migratetype);
1972 if (IS_ENABLED(CONFIG_NUMA))
1973 zlc_mark_zone_full(zonelist, z);
1976 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1977 /* Disable zlc cache for second zonelist scan */
1984 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1985 * necessary to allocate the page. The expectation is
1986 * that the caller is taking steps that will free more
1987 * memory. The caller should avoid the page being used
1988 * for !PFMEMALLOC purposes.
1990 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1996 * Large machines with many possible nodes should not always dump per-node
1997 * meminfo in irq context.
1999 static inline bool should_suppress_show_mem(void)
2004 ret = in_interrupt();
2009 static DEFINE_RATELIMIT_STATE(nopage_rs,
2010 DEFAULT_RATELIMIT_INTERVAL,
2011 DEFAULT_RATELIMIT_BURST);
2013 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2015 unsigned int filter = SHOW_MEM_FILTER_NODES;
2017 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2018 debug_guardpage_minorder() > 0)
2022 * Walking all memory to count page types is very expensive and should
2023 * be inhibited in non-blockable contexts.
2025 if (!(gfp_mask & __GFP_WAIT))
2026 filter |= SHOW_MEM_FILTER_PAGE_COUNT;
2029 * This documents exceptions given to allocations in certain
2030 * contexts that are allowed to allocate outside current's set
2033 if (!(gfp_mask & __GFP_NOMEMALLOC))
2034 if (test_thread_flag(TIF_MEMDIE) ||
2035 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2036 filter &= ~SHOW_MEM_FILTER_NODES;
2037 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2038 filter &= ~SHOW_MEM_FILTER_NODES;
2041 struct va_format vaf;
2044 va_start(args, fmt);
2049 pr_warn("%pV", &vaf);
2054 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2055 current->comm, order, gfp_mask);
2058 if (!should_suppress_show_mem())
2063 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2064 unsigned long did_some_progress,
2065 unsigned long pages_reclaimed)
2067 /* Do not loop if specifically requested */
2068 if (gfp_mask & __GFP_NORETRY)
2071 /* Always retry if specifically requested */
2072 if (gfp_mask & __GFP_NOFAIL)
2076 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2077 * making forward progress without invoking OOM. Suspend also disables
2078 * storage devices so kswapd will not help. Bail if we are suspending.
2080 if (!did_some_progress && pm_suspended_storage())
2084 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2085 * means __GFP_NOFAIL, but that may not be true in other
2088 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2092 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2093 * specified, then we retry until we no longer reclaim any pages
2094 * (above), or we've reclaimed an order of pages at least as
2095 * large as the allocation's order. In both cases, if the
2096 * allocation still fails, we stop retrying.
2098 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2104 static inline struct page *
2105 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2106 struct zonelist *zonelist, enum zone_type high_zoneidx,
2107 nodemask_t *nodemask, struct zone *preferred_zone,
2112 /* Acquire the OOM killer lock for the zones in zonelist */
2113 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2114 schedule_timeout_uninterruptible(1);
2119 * Go through the zonelist yet one more time, keep very high watermark
2120 * here, this is only to catch a parallel oom killing, we must fail if
2121 * we're still under heavy pressure.
2123 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2124 order, zonelist, high_zoneidx,
2125 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2126 preferred_zone, migratetype);
2130 if (!(gfp_mask & __GFP_NOFAIL)) {
2131 /* The OOM killer will not help higher order allocs */
2132 if (order > PAGE_ALLOC_COSTLY_ORDER)
2134 /* The OOM killer does not needlessly kill tasks for lowmem */
2135 if (high_zoneidx < ZONE_NORMAL)
2138 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2139 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2140 * The caller should handle page allocation failure by itself if
2141 * it specifies __GFP_THISNODE.
2142 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2144 if (gfp_mask & __GFP_THISNODE)
2147 /* Exhausted what can be done so it's blamo time */
2148 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2151 clear_zonelist_oom(zonelist, gfp_mask);
2155 #ifdef CONFIG_COMPACTION
2156 /* Try memory compaction for high-order allocations before reclaim */
2157 static struct page *
2158 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2159 struct zonelist *zonelist, enum zone_type high_zoneidx,
2160 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2161 int migratetype, bool sync_migration,
2162 bool *contended_compaction, bool *deferred_compaction,
2163 unsigned long *did_some_progress)
2168 if (compaction_deferred(preferred_zone, order)) {
2169 *deferred_compaction = true;
2173 current->flags |= PF_MEMALLOC;
2174 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2175 nodemask, sync_migration,
2176 contended_compaction);
2177 current->flags &= ~PF_MEMALLOC;
2179 if (*did_some_progress != COMPACT_SKIPPED) {
2182 /* Page migration frees to the PCP lists but we want merging */
2183 drain_pages(get_cpu());
2186 page = get_page_from_freelist(gfp_mask, nodemask,
2187 order, zonelist, high_zoneidx,
2188 alloc_flags & ~ALLOC_NO_WATERMARKS,
2189 preferred_zone, migratetype);
2191 preferred_zone->compact_blockskip_flush = false;
2192 preferred_zone->compact_considered = 0;
2193 preferred_zone->compact_defer_shift = 0;
2194 if (order >= preferred_zone->compact_order_failed)
2195 preferred_zone->compact_order_failed = order + 1;
2196 count_vm_event(COMPACTSUCCESS);
2201 * It's bad if compaction run occurs and fails.
2202 * The most likely reason is that pages exist,
2203 * but not enough to satisfy watermarks.
2205 count_vm_event(COMPACTFAIL);
2208 * As async compaction considers a subset of pageblocks, only
2209 * defer if the failure was a sync compaction failure.
2212 defer_compaction(preferred_zone, order);
2220 static inline struct page *
2221 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2222 struct zonelist *zonelist, enum zone_type high_zoneidx,
2223 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2224 int migratetype, bool sync_migration,
2225 bool *contended_compaction, bool *deferred_compaction,
2226 unsigned long *did_some_progress)
2230 #endif /* CONFIG_COMPACTION */
2232 /* Perform direct synchronous page reclaim */
2234 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2235 nodemask_t *nodemask)
2237 struct reclaim_state reclaim_state;
2242 /* We now go into synchronous reclaim */
2243 cpuset_memory_pressure_bump();
2244 current->flags |= PF_MEMALLOC;
2245 lockdep_set_current_reclaim_state(gfp_mask);
2246 reclaim_state.reclaimed_slab = 0;
2247 current->reclaim_state = &reclaim_state;
2249 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2251 current->reclaim_state = NULL;
2252 lockdep_clear_current_reclaim_state();
2253 current->flags &= ~PF_MEMALLOC;
2260 /* The really slow allocator path where we enter direct reclaim */
2261 static inline struct page *
2262 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2263 struct zonelist *zonelist, enum zone_type high_zoneidx,
2264 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2265 int migratetype, unsigned long *did_some_progress)
2267 struct page *page = NULL;
2268 bool drained = false;
2270 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2272 if (unlikely(!(*did_some_progress)))
2275 /* After successful reclaim, reconsider all zones for allocation */
2276 if (IS_ENABLED(CONFIG_NUMA))
2277 zlc_clear_zones_full(zonelist);
2280 page = get_page_from_freelist(gfp_mask, nodemask, order,
2281 zonelist, high_zoneidx,
2282 alloc_flags & ~ALLOC_NO_WATERMARKS,
2283 preferred_zone, migratetype);
2286 * If an allocation failed after direct reclaim, it could be because
2287 * pages are pinned on the per-cpu lists. Drain them and try again
2289 if (!page && !drained) {
2299 * This is called in the allocator slow-path if the allocation request is of
2300 * sufficient urgency to ignore watermarks and take other desperate measures
2302 static inline struct page *
2303 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2304 struct zonelist *zonelist, enum zone_type high_zoneidx,
2305 nodemask_t *nodemask, struct zone *preferred_zone,
2311 page = get_page_from_freelist(gfp_mask, nodemask, order,
2312 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2313 preferred_zone, migratetype);
2315 if (!page && gfp_mask & __GFP_NOFAIL)
2316 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2317 } while (!page && (gfp_mask & __GFP_NOFAIL));
2323 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2324 enum zone_type high_zoneidx,
2325 enum zone_type classzone_idx)
2330 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2331 wakeup_kswapd(zone, order, classzone_idx);
2335 gfp_to_alloc_flags(gfp_t gfp_mask)
2337 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2338 const gfp_t wait = gfp_mask & __GFP_WAIT;
2340 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2341 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2344 * The caller may dip into page reserves a bit more if the caller
2345 * cannot run direct reclaim, or if the caller has realtime scheduling
2346 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2347 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2349 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2353 * Not worth trying to allocate harder for
2354 * __GFP_NOMEMALLOC even if it can't schedule.
2356 if (!(gfp_mask & __GFP_NOMEMALLOC))
2357 alloc_flags |= ALLOC_HARDER;
2359 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2360 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2362 alloc_flags &= ~ALLOC_CPUSET;
2363 } else if (unlikely(rt_task(current)) && !in_interrupt())
2364 alloc_flags |= ALLOC_HARDER;
2366 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2367 if (gfp_mask & __GFP_MEMALLOC)
2368 alloc_flags |= ALLOC_NO_WATERMARKS;
2369 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2370 alloc_flags |= ALLOC_NO_WATERMARKS;
2371 else if (!in_interrupt() &&
2372 ((current->flags & PF_MEMALLOC) ||
2373 unlikely(test_thread_flag(TIF_MEMDIE))))
2374 alloc_flags |= ALLOC_NO_WATERMARKS;
2377 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2378 alloc_flags |= ALLOC_CMA;
2383 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2385 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2388 static inline struct page *
2389 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2390 struct zonelist *zonelist, enum zone_type high_zoneidx,
2391 nodemask_t *nodemask, struct zone *preferred_zone,
2394 const gfp_t wait = gfp_mask & __GFP_WAIT;
2395 struct page *page = NULL;
2397 unsigned long pages_reclaimed = 0;
2398 unsigned long did_some_progress;
2399 bool sync_migration = false;
2400 bool deferred_compaction = false;
2401 bool contended_compaction = false;
2404 * In the slowpath, we sanity check order to avoid ever trying to
2405 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2406 * be using allocators in order of preference for an area that is
2409 if (order >= MAX_ORDER) {
2410 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2415 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2416 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2417 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2418 * using a larger set of nodes after it has established that the
2419 * allowed per node queues are empty and that nodes are
2422 if (IS_ENABLED(CONFIG_NUMA) &&
2423 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2427 if (!(gfp_mask & __GFP_NO_KSWAPD))
2428 wake_all_kswapd(order, zonelist, high_zoneidx,
2429 zone_idx(preferred_zone));
2432 * OK, we're below the kswapd watermark and have kicked background
2433 * reclaim. Now things get more complex, so set up alloc_flags according
2434 * to how we want to proceed.
2436 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2439 * Find the true preferred zone if the allocation is unconstrained by
2442 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2443 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2447 /* This is the last chance, in general, before the goto nopage. */
2448 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2449 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2450 preferred_zone, migratetype);
2454 /* Allocate without watermarks if the context allows */
2455 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2457 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2458 * the allocation is high priority and these type of
2459 * allocations are system rather than user orientated
2461 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2463 page = __alloc_pages_high_priority(gfp_mask, order,
2464 zonelist, high_zoneidx, nodemask,
2465 preferred_zone, migratetype);
2471 /* Atomic allocations - we can't balance anything */
2475 /* Avoid recursion of direct reclaim */
2476 if (current->flags & PF_MEMALLOC)
2479 /* Avoid allocations with no watermarks from looping endlessly */
2480 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2484 * Try direct compaction. The first pass is asynchronous. Subsequent
2485 * attempts after direct reclaim are synchronous
2487 page = __alloc_pages_direct_compact(gfp_mask, order,
2488 zonelist, high_zoneidx,
2490 alloc_flags, preferred_zone,
2491 migratetype, sync_migration,
2492 &contended_compaction,
2493 &deferred_compaction,
2494 &did_some_progress);
2497 sync_migration = true;
2500 * If compaction is deferred for high-order allocations, it is because
2501 * sync compaction recently failed. In this is the case and the caller
2502 * requested a movable allocation that does not heavily disrupt the
2503 * system then fail the allocation instead of entering direct reclaim.
2505 if ((deferred_compaction || contended_compaction) &&
2506 (gfp_mask & __GFP_NO_KSWAPD))
2509 /* Try direct reclaim and then allocating */
2510 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2511 zonelist, high_zoneidx,
2513 alloc_flags, preferred_zone,
2514 migratetype, &did_some_progress);
2519 * If we failed to make any progress reclaiming, then we are
2520 * running out of options and have to consider going OOM
2522 if (!did_some_progress) {
2523 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2524 if (oom_killer_disabled)
2526 /* Coredumps can quickly deplete all memory reserves */
2527 if ((current->flags & PF_DUMPCORE) &&
2528 !(gfp_mask & __GFP_NOFAIL))
2530 page = __alloc_pages_may_oom(gfp_mask, order,
2531 zonelist, high_zoneidx,
2532 nodemask, preferred_zone,
2537 if (!(gfp_mask & __GFP_NOFAIL)) {
2539 * The oom killer is not called for high-order
2540 * allocations that may fail, so if no progress
2541 * is being made, there are no other options and
2542 * retrying is unlikely to help.
2544 if (order > PAGE_ALLOC_COSTLY_ORDER)
2547 * The oom killer is not called for lowmem
2548 * allocations to prevent needlessly killing
2551 if (high_zoneidx < ZONE_NORMAL)
2559 /* Check if we should retry the allocation */
2560 pages_reclaimed += did_some_progress;
2561 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2563 /* Wait for some write requests to complete then retry */
2564 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2568 * High-order allocations do not necessarily loop after
2569 * direct reclaim and reclaim/compaction depends on compaction
2570 * being called after reclaim so call directly if necessary
2572 page = __alloc_pages_direct_compact(gfp_mask, order,
2573 zonelist, high_zoneidx,
2575 alloc_flags, preferred_zone,
2576 migratetype, sync_migration,
2577 &contended_compaction,
2578 &deferred_compaction,
2579 &did_some_progress);
2585 warn_alloc_failed(gfp_mask, order, NULL);
2588 if (kmemcheck_enabled)
2589 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2595 * This is the 'heart' of the zoned buddy allocator.
2598 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2599 struct zonelist *zonelist, nodemask_t *nodemask)
2601 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2602 struct zone *preferred_zone;
2603 struct page *page = NULL;
2604 int migratetype = allocflags_to_migratetype(gfp_mask);
2605 unsigned int cpuset_mems_cookie;
2606 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2607 struct mem_cgroup *memcg = NULL;
2609 gfp_mask &= gfp_allowed_mask;
2611 lockdep_trace_alloc(gfp_mask);
2613 might_sleep_if(gfp_mask & __GFP_WAIT);
2615 if (should_fail_alloc_page(gfp_mask, order))
2619 * Check the zones suitable for the gfp_mask contain at least one
2620 * valid zone. It's possible to have an empty zonelist as a result
2621 * of GFP_THISNODE and a memoryless node
2623 if (unlikely(!zonelist->_zonerefs->zone))
2627 * Will only have any effect when __GFP_KMEMCG is set. This is
2628 * verified in the (always inline) callee
2630 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2634 cpuset_mems_cookie = get_mems_allowed();
2636 /* The preferred zone is used for statistics later */
2637 first_zones_zonelist(zonelist, high_zoneidx,
2638 nodemask ? : &cpuset_current_mems_allowed,
2640 if (!preferred_zone)
2644 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2645 alloc_flags |= ALLOC_CMA;
2647 /* First allocation attempt */
2648 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2649 zonelist, high_zoneidx, alloc_flags,
2650 preferred_zone, migratetype);
2651 if (unlikely(!page)) {
2653 * Runtime PM, block IO and its error handling path
2654 * can deadlock because I/O on the device might not
2657 gfp_mask = memalloc_noio_flags(gfp_mask);
2658 page = __alloc_pages_slowpath(gfp_mask, order,
2659 zonelist, high_zoneidx, nodemask,
2660 preferred_zone, migratetype);
2663 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2667 * When updating a task's mems_allowed, it is possible to race with
2668 * parallel threads in such a way that an allocation can fail while
2669 * the mask is being updated. If a page allocation is about to fail,
2670 * check if the cpuset changed during allocation and if so, retry.
2672 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2675 memcg_kmem_commit_charge(page, memcg, order);
2679 EXPORT_SYMBOL(__alloc_pages_nodemask);
2682 * Common helper functions.
2684 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2689 * __get_free_pages() returns a 32-bit address, which cannot represent
2692 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2694 page = alloc_pages(gfp_mask, order);
2697 return (unsigned long) page_address(page);
2699 EXPORT_SYMBOL(__get_free_pages);
2701 unsigned long get_zeroed_page(gfp_t gfp_mask)
2703 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2705 EXPORT_SYMBOL(get_zeroed_page);
2707 void __free_pages(struct page *page, unsigned int order)
2709 if (put_page_testzero(page)) {
2711 free_hot_cold_page(page, 0);
2713 __free_pages_ok(page, order);
2717 EXPORT_SYMBOL(__free_pages);
2719 void free_pages(unsigned long addr, unsigned int order)
2722 VM_BUG_ON(!virt_addr_valid((void *)addr));
2723 __free_pages(virt_to_page((void *)addr), order);
2727 EXPORT_SYMBOL(free_pages);
2730 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2731 * pages allocated with __GFP_KMEMCG.
2733 * Those pages are accounted to a particular memcg, embedded in the
2734 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2735 * for that information only to find out that it is NULL for users who have no
2736 * interest in that whatsoever, we provide these functions.
2738 * The caller knows better which flags it relies on.
2740 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2742 memcg_kmem_uncharge_pages(page, order);
2743 __free_pages(page, order);
2746 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2749 VM_BUG_ON(!virt_addr_valid((void *)addr));
2750 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2754 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2757 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2758 unsigned long used = addr + PAGE_ALIGN(size);
2760 split_page(virt_to_page((void *)addr), order);
2761 while (used < alloc_end) {
2766 return (void *)addr;
2770 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2771 * @size: the number of bytes to allocate
2772 * @gfp_mask: GFP flags for the allocation
2774 * This function is similar to alloc_pages(), except that it allocates the
2775 * minimum number of pages to satisfy the request. alloc_pages() can only
2776 * allocate memory in power-of-two pages.
2778 * This function is also limited by MAX_ORDER.
2780 * Memory allocated by this function must be released by free_pages_exact().
2782 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2784 unsigned int order = get_order(size);
2787 addr = __get_free_pages(gfp_mask, order);
2788 return make_alloc_exact(addr, order, size);
2790 EXPORT_SYMBOL(alloc_pages_exact);
2793 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2795 * @nid: the preferred node ID where memory should be allocated
2796 * @size: the number of bytes to allocate
2797 * @gfp_mask: GFP flags for the allocation
2799 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2801 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2804 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2806 unsigned order = get_order(size);
2807 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2810 return make_alloc_exact((unsigned long)page_address(p), order, size);
2812 EXPORT_SYMBOL(alloc_pages_exact_nid);
2815 * free_pages_exact - release memory allocated via alloc_pages_exact()
2816 * @virt: the value returned by alloc_pages_exact.
2817 * @size: size of allocation, same value as passed to alloc_pages_exact().
2819 * Release the memory allocated by a previous call to alloc_pages_exact.
2821 void free_pages_exact(void *virt, size_t size)
2823 unsigned long addr = (unsigned long)virt;
2824 unsigned long end = addr + PAGE_ALIGN(size);
2826 while (addr < end) {
2831 EXPORT_SYMBOL(free_pages_exact);
2834 * nr_free_zone_pages - count number of pages beyond high watermark
2835 * @offset: The zone index of the highest zone
2837 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2838 * high watermark within all zones at or below a given zone index. For each
2839 * zone, the number of pages is calculated as:
2840 * present_pages - high_pages
2842 static unsigned long nr_free_zone_pages(int offset)
2847 /* Just pick one node, since fallback list is circular */
2848 unsigned long sum = 0;
2850 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2852 for_each_zone_zonelist(zone, z, zonelist, offset) {
2853 unsigned long size = zone->managed_pages;
2854 unsigned long high = high_wmark_pages(zone);
2863 * nr_free_buffer_pages - count number of pages beyond high watermark
2865 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2866 * watermark within ZONE_DMA and ZONE_NORMAL.
2868 unsigned long nr_free_buffer_pages(void)
2870 return nr_free_zone_pages(gfp_zone(GFP_USER));
2872 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2875 * nr_free_pagecache_pages - count number of pages beyond high watermark
2877 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2878 * high watermark within all zones.
2880 unsigned long nr_free_pagecache_pages(void)
2882 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2885 static inline void show_node(struct zone *zone)
2887 if (IS_ENABLED(CONFIG_NUMA))
2888 printk("Node %d ", zone_to_nid(zone));
2891 void si_meminfo(struct sysinfo *val)
2893 val->totalram = totalram_pages;
2895 val->freeram = global_page_state(NR_FREE_PAGES);
2896 val->bufferram = nr_blockdev_pages();
2897 val->totalhigh = totalhigh_pages;
2898 val->freehigh = nr_free_highpages();
2899 val->mem_unit = PAGE_SIZE;
2902 EXPORT_SYMBOL(si_meminfo);
2905 void si_meminfo_node(struct sysinfo *val, int nid)
2907 pg_data_t *pgdat = NODE_DATA(nid);
2909 val->totalram = pgdat->node_present_pages;
2910 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2911 #ifdef CONFIG_HIGHMEM
2912 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
2913 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2919 val->mem_unit = PAGE_SIZE;
2924 * Determine whether the node should be displayed or not, depending on whether
2925 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2927 bool skip_free_areas_node(unsigned int flags, int nid)
2930 unsigned int cpuset_mems_cookie;
2932 if (!(flags & SHOW_MEM_FILTER_NODES))
2936 cpuset_mems_cookie = get_mems_allowed();
2937 ret = !node_isset(nid, cpuset_current_mems_allowed);
2938 } while (!put_mems_allowed(cpuset_mems_cookie));
2943 #define K(x) ((x) << (PAGE_SHIFT-10))
2945 static void show_migration_types(unsigned char type)
2947 static const char types[MIGRATE_TYPES] = {
2948 [MIGRATE_UNMOVABLE] = 'U',
2949 [MIGRATE_RECLAIMABLE] = 'E',
2950 [MIGRATE_MOVABLE] = 'M',
2951 [MIGRATE_RESERVE] = 'R',
2953 [MIGRATE_CMA] = 'C',
2955 #ifdef CONFIG_MEMORY_ISOLATION
2956 [MIGRATE_ISOLATE] = 'I',
2959 char tmp[MIGRATE_TYPES + 1];
2963 for (i = 0; i < MIGRATE_TYPES; i++) {
2964 if (type & (1 << i))
2969 printk("(%s) ", tmp);
2973 * Show free area list (used inside shift_scroll-lock stuff)
2974 * We also calculate the percentage fragmentation. We do this by counting the
2975 * memory on each free list with the exception of the first item on the list.
2976 * Suppresses nodes that are not allowed by current's cpuset if
2977 * SHOW_MEM_FILTER_NODES is passed.
2979 void show_free_areas(unsigned int filter)
2984 for_each_populated_zone(zone) {
2985 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2988 printk("%s per-cpu:\n", zone->name);
2990 for_each_online_cpu(cpu) {
2991 struct per_cpu_pageset *pageset;
2993 pageset = per_cpu_ptr(zone->pageset, cpu);
2995 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2996 cpu, pageset->pcp.high,
2997 pageset->pcp.batch, pageset->pcp.count);
3001 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3002 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3004 " dirty:%lu writeback:%lu unstable:%lu\n"
3005 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3006 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3008 global_page_state(NR_ACTIVE_ANON),
3009 global_page_state(NR_INACTIVE_ANON),
3010 global_page_state(NR_ISOLATED_ANON),
3011 global_page_state(NR_ACTIVE_FILE),
3012 global_page_state(NR_INACTIVE_FILE),
3013 global_page_state(NR_ISOLATED_FILE),
3014 global_page_state(NR_UNEVICTABLE),
3015 global_page_state(NR_FILE_DIRTY),
3016 global_page_state(NR_WRITEBACK),
3017 global_page_state(NR_UNSTABLE_NFS),
3018 global_page_state(NR_FREE_PAGES),
3019 global_page_state(NR_SLAB_RECLAIMABLE),
3020 global_page_state(NR_SLAB_UNRECLAIMABLE),
3021 global_page_state(NR_FILE_MAPPED),
3022 global_page_state(NR_SHMEM),
3023 global_page_state(NR_PAGETABLE),
3024 global_page_state(NR_BOUNCE),
3025 global_page_state(NR_FREE_CMA_PAGES));
3027 for_each_populated_zone(zone) {
3030 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3038 " active_anon:%lukB"
3039 " inactive_anon:%lukB"
3040 " active_file:%lukB"
3041 " inactive_file:%lukB"
3042 " unevictable:%lukB"
3043 " isolated(anon):%lukB"
3044 " isolated(file):%lukB"
3052 " slab_reclaimable:%lukB"
3053 " slab_unreclaimable:%lukB"
3054 " kernel_stack:%lukB"
3059 " writeback_tmp:%lukB"
3060 " pages_scanned:%lu"
3061 " all_unreclaimable? %s"
3064 K(zone_page_state(zone, NR_FREE_PAGES)),
3065 K(min_wmark_pages(zone)),
3066 K(low_wmark_pages(zone)),
3067 K(high_wmark_pages(zone)),
3068 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3069 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3070 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3071 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3072 K(zone_page_state(zone, NR_UNEVICTABLE)),
3073 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3074 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3075 K(zone->present_pages),
3076 K(zone->managed_pages),
3077 K(zone_page_state(zone, NR_MLOCK)),
3078 K(zone_page_state(zone, NR_FILE_DIRTY)),
3079 K(zone_page_state(zone, NR_WRITEBACK)),
3080 K(zone_page_state(zone, NR_FILE_MAPPED)),
3081 K(zone_page_state(zone, NR_SHMEM)),
3082 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3083 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3084 zone_page_state(zone, NR_KERNEL_STACK) *
3086 K(zone_page_state(zone, NR_PAGETABLE)),
3087 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3088 K(zone_page_state(zone, NR_BOUNCE)),
3089 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3090 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3091 zone->pages_scanned,
3092 (zone->all_unreclaimable ? "yes" : "no")
3094 printk("lowmem_reserve[]:");
3095 for (i = 0; i < MAX_NR_ZONES; i++)
3096 printk(" %lu", zone->lowmem_reserve[i]);
3100 for_each_populated_zone(zone) {
3101 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3102 unsigned char types[MAX_ORDER];
3104 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3107 printk("%s: ", zone->name);
3109 spin_lock_irqsave(&zone->lock, flags);
3110 for (order = 0; order < MAX_ORDER; order++) {
3111 struct free_area *area = &zone->free_area[order];
3114 nr[order] = area->nr_free;
3115 total += nr[order] << order;
3118 for (type = 0; type < MIGRATE_TYPES; type++) {
3119 if (!list_empty(&area->free_list[type]))
3120 types[order] |= 1 << type;
3123 spin_unlock_irqrestore(&zone->lock, flags);
3124 for (order = 0; order < MAX_ORDER; order++) {
3125 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3127 show_migration_types(types[order]);
3129 printk("= %lukB\n", K(total));
3132 hugetlb_show_meminfo();
3134 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3136 show_swap_cache_info();
3139 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3141 zoneref->zone = zone;
3142 zoneref->zone_idx = zone_idx(zone);
3146 * Builds allocation fallback zone lists.
3148 * Add all populated zones of a node to the zonelist.
3150 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3151 int nr_zones, enum zone_type zone_type)
3155 BUG_ON(zone_type >= MAX_NR_ZONES);
3160 zone = pgdat->node_zones + zone_type;
3161 if (populated_zone(zone)) {
3162 zoneref_set_zone(zone,
3163 &zonelist->_zonerefs[nr_zones++]);
3164 check_highest_zone(zone_type);
3167 } while (zone_type);
3174 * 0 = automatic detection of better ordering.
3175 * 1 = order by ([node] distance, -zonetype)
3176 * 2 = order by (-zonetype, [node] distance)
3178 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3179 * the same zonelist. So only NUMA can configure this param.
3181 #define ZONELIST_ORDER_DEFAULT 0
3182 #define ZONELIST_ORDER_NODE 1
3183 #define ZONELIST_ORDER_ZONE 2
3185 /* zonelist order in the kernel.
3186 * set_zonelist_order() will set this to NODE or ZONE.
3188 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3189 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3193 /* The value user specified ....changed by config */
3194 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3195 /* string for sysctl */
3196 #define NUMA_ZONELIST_ORDER_LEN 16
3197 char numa_zonelist_order[16] = "default";
3200 * interface for configure zonelist ordering.
3201 * command line option "numa_zonelist_order"
3202 * = "[dD]efault - default, automatic configuration.
3203 * = "[nN]ode - order by node locality, then by zone within node
3204 * = "[zZ]one - order by zone, then by locality within zone
3207 static int __parse_numa_zonelist_order(char *s)
3209 if (*s == 'd' || *s == 'D') {
3210 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3211 } else if (*s == 'n' || *s == 'N') {
3212 user_zonelist_order = ZONELIST_ORDER_NODE;
3213 } else if (*s == 'z' || *s == 'Z') {
3214 user_zonelist_order = ZONELIST_ORDER_ZONE;
3217 "Ignoring invalid numa_zonelist_order value: "
3224 static __init int setup_numa_zonelist_order(char *s)
3231 ret = __parse_numa_zonelist_order(s);
3233 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3237 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3240 * sysctl handler for numa_zonelist_order
3242 int numa_zonelist_order_handler(ctl_table *table, int write,
3243 void __user *buffer, size_t *length,
3246 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3248 static DEFINE_MUTEX(zl_order_mutex);
3250 mutex_lock(&zl_order_mutex);
3252 strcpy(saved_string, (char*)table->data);
3253 ret = proc_dostring(table, write, buffer, length, ppos);
3257 int oldval = user_zonelist_order;
3258 if (__parse_numa_zonelist_order((char*)table->data)) {
3260 * bogus value. restore saved string
3262 strncpy((char*)table->data, saved_string,
3263 NUMA_ZONELIST_ORDER_LEN);
3264 user_zonelist_order = oldval;
3265 } else if (oldval != user_zonelist_order) {
3266 mutex_lock(&zonelists_mutex);
3267 build_all_zonelists(NULL, NULL);
3268 mutex_unlock(&zonelists_mutex);
3272 mutex_unlock(&zl_order_mutex);
3277 #define MAX_NODE_LOAD (nr_online_nodes)
3278 static int node_load[MAX_NUMNODES];
3281 * find_next_best_node - find the next node that should appear in a given node's fallback list
3282 * @node: node whose fallback list we're appending
3283 * @used_node_mask: nodemask_t of already used nodes
3285 * We use a number of factors to determine which is the next node that should
3286 * appear on a given node's fallback list. The node should not have appeared
3287 * already in @node's fallback list, and it should be the next closest node
3288 * according to the distance array (which contains arbitrary distance values
3289 * from each node to each node in the system), and should also prefer nodes
3290 * with no CPUs, since presumably they'll have very little allocation pressure
3291 * on them otherwise.
3292 * It returns -1 if no node is found.
3294 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3297 int min_val = INT_MAX;
3298 int best_node = NUMA_NO_NODE;
3299 const struct cpumask *tmp = cpumask_of_node(0);
3301 /* Use the local node if we haven't already */
3302 if (!node_isset(node, *used_node_mask)) {
3303 node_set(node, *used_node_mask);
3307 for_each_node_state(n, N_MEMORY) {
3309 /* Don't want a node to appear more than once */
3310 if (node_isset(n, *used_node_mask))
3313 /* Use the distance array to find the distance */
3314 val = node_distance(node, n);
3316 /* Penalize nodes under us ("prefer the next node") */
3319 /* Give preference to headless and unused nodes */
3320 tmp = cpumask_of_node(n);
3321 if (!cpumask_empty(tmp))
3322 val += PENALTY_FOR_NODE_WITH_CPUS;
3324 /* Slight preference for less loaded node */
3325 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3326 val += node_load[n];
3328 if (val < min_val) {
3335 node_set(best_node, *used_node_mask);
3342 * Build zonelists ordered by node and zones within node.
3343 * This results in maximum locality--normal zone overflows into local
3344 * DMA zone, if any--but risks exhausting DMA zone.
3346 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3349 struct zonelist *zonelist;
3351 zonelist = &pgdat->node_zonelists[0];
3352 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3354 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3356 zonelist->_zonerefs[j].zone = NULL;
3357 zonelist->_zonerefs[j].zone_idx = 0;
3361 * Build gfp_thisnode zonelists
3363 static void build_thisnode_zonelists(pg_data_t *pgdat)
3366 struct zonelist *zonelist;
3368 zonelist = &pgdat->node_zonelists[1];
3369 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3370 zonelist->_zonerefs[j].zone = NULL;
3371 zonelist->_zonerefs[j].zone_idx = 0;
3375 * Build zonelists ordered by zone and nodes within zones.
3376 * This results in conserving DMA zone[s] until all Normal memory is
3377 * exhausted, but results in overflowing to remote node while memory
3378 * may still exist in local DMA zone.
3380 static int node_order[MAX_NUMNODES];
3382 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3385 int zone_type; /* needs to be signed */
3387 struct zonelist *zonelist;
3389 zonelist = &pgdat->node_zonelists[0];
3391 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3392 for (j = 0; j < nr_nodes; j++) {
3393 node = node_order[j];
3394 z = &NODE_DATA(node)->node_zones[zone_type];
3395 if (populated_zone(z)) {
3397 &zonelist->_zonerefs[pos++]);
3398 check_highest_zone(zone_type);
3402 zonelist->_zonerefs[pos].zone = NULL;
3403 zonelist->_zonerefs[pos].zone_idx = 0;
3406 static int default_zonelist_order(void)
3409 unsigned long low_kmem_size,total_size;
3413 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3414 * If they are really small and used heavily, the system can fall
3415 * into OOM very easily.
3416 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3418 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3421 for_each_online_node(nid) {
3422 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3423 z = &NODE_DATA(nid)->node_zones[zone_type];
3424 if (populated_zone(z)) {
3425 if (zone_type < ZONE_NORMAL)
3426 low_kmem_size += z->present_pages;
3427 total_size += z->present_pages;
3428 } else if (zone_type == ZONE_NORMAL) {
3430 * If any node has only lowmem, then node order
3431 * is preferred to allow kernel allocations
3432 * locally; otherwise, they can easily infringe
3433 * on other nodes when there is an abundance of
3434 * lowmem available to allocate from.
3436 return ZONELIST_ORDER_NODE;
3440 if (!low_kmem_size || /* there are no DMA area. */
3441 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3442 return ZONELIST_ORDER_NODE;
3444 * look into each node's config.
3445 * If there is a node whose DMA/DMA32 memory is very big area on
3446 * local memory, NODE_ORDER may be suitable.
3448 average_size = total_size /
3449 (nodes_weight(node_states[N_MEMORY]) + 1);
3450 for_each_online_node(nid) {
3453 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3454 z = &NODE_DATA(nid)->node_zones[zone_type];
3455 if (populated_zone(z)) {
3456 if (zone_type < ZONE_NORMAL)
3457 low_kmem_size += z->present_pages;
3458 total_size += z->present_pages;
3461 if (low_kmem_size &&
3462 total_size > average_size && /* ignore small node */
3463 low_kmem_size > total_size * 70/100)
3464 return ZONELIST_ORDER_NODE;
3466 return ZONELIST_ORDER_ZONE;
3469 static void set_zonelist_order(void)
3471 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3472 current_zonelist_order = default_zonelist_order();
3474 current_zonelist_order = user_zonelist_order;
3477 static void build_zonelists(pg_data_t *pgdat)
3481 nodemask_t used_mask;
3482 int local_node, prev_node;
3483 struct zonelist *zonelist;
3484 int order = current_zonelist_order;
3486 /* initialize zonelists */
3487 for (i = 0; i < MAX_ZONELISTS; i++) {
3488 zonelist = pgdat->node_zonelists + i;
3489 zonelist->_zonerefs[0].zone = NULL;
3490 zonelist->_zonerefs[0].zone_idx = 0;
3493 /* NUMA-aware ordering of nodes */
3494 local_node = pgdat->node_id;
3495 load = nr_online_nodes;
3496 prev_node = local_node;
3497 nodes_clear(used_mask);
3499 memset(node_order, 0, sizeof(node_order));
3502 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3504 * We don't want to pressure a particular node.
3505 * So adding penalty to the first node in same
3506 * distance group to make it round-robin.
3508 if (node_distance(local_node, node) !=
3509 node_distance(local_node, prev_node))
3510 node_load[node] = load;
3514 if (order == ZONELIST_ORDER_NODE)
3515 build_zonelists_in_node_order(pgdat, node);
3517 node_order[j++] = node; /* remember order */
3520 if (order == ZONELIST_ORDER_ZONE) {
3521 /* calculate node order -- i.e., DMA last! */
3522 build_zonelists_in_zone_order(pgdat, j);
3525 build_thisnode_zonelists(pgdat);
3528 /* Construct the zonelist performance cache - see further mmzone.h */
3529 static void build_zonelist_cache(pg_data_t *pgdat)
3531 struct zonelist *zonelist;
3532 struct zonelist_cache *zlc;
3535 zonelist = &pgdat->node_zonelists[0];
3536 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3537 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3538 for (z = zonelist->_zonerefs; z->zone; z++)
3539 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3542 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3544 * Return node id of node used for "local" allocations.
3545 * I.e., first node id of first zone in arg node's generic zonelist.
3546 * Used for initializing percpu 'numa_mem', which is used primarily
3547 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3549 int local_memory_node(int node)
3553 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3554 gfp_zone(GFP_KERNEL),
3561 #else /* CONFIG_NUMA */
3563 static void set_zonelist_order(void)
3565 current_zonelist_order = ZONELIST_ORDER_ZONE;
3568 static void build_zonelists(pg_data_t *pgdat)
3570 int node, local_node;
3572 struct zonelist *zonelist;
3574 local_node = pgdat->node_id;
3576 zonelist = &pgdat->node_zonelists[0];
3577 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3580 * Now we build the zonelist so that it contains the zones
3581 * of all the other nodes.
3582 * We don't want to pressure a particular node, so when
3583 * building the zones for node N, we make sure that the
3584 * zones coming right after the local ones are those from
3585 * node N+1 (modulo N)
3587 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3588 if (!node_online(node))
3590 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3593 for (node = 0; node < local_node; node++) {
3594 if (!node_online(node))
3596 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3600 zonelist->_zonerefs[j].zone = NULL;
3601 zonelist->_zonerefs[j].zone_idx = 0;
3604 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3605 static void build_zonelist_cache(pg_data_t *pgdat)
3607 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3610 #endif /* CONFIG_NUMA */
3613 * Boot pageset table. One per cpu which is going to be used for all
3614 * zones and all nodes. The parameters will be set in such a way
3615 * that an item put on a list will immediately be handed over to
3616 * the buddy list. This is safe since pageset manipulation is done
3617 * with interrupts disabled.
3619 * The boot_pagesets must be kept even after bootup is complete for
3620 * unused processors and/or zones. They do play a role for bootstrapping
3621 * hotplugged processors.
3623 * zoneinfo_show() and maybe other functions do
3624 * not check if the processor is online before following the pageset pointer.
3625 * Other parts of the kernel may not check if the zone is available.
3627 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3628 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3629 static void setup_zone_pageset(struct zone *zone);
3632 * Global mutex to protect against size modification of zonelists
3633 * as well as to serialize pageset setup for the new populated zone.
3635 DEFINE_MUTEX(zonelists_mutex);
3637 /* return values int ....just for stop_machine() */
3638 static int __build_all_zonelists(void *data)
3642 pg_data_t *self = data;
3645 memset(node_load, 0, sizeof(node_load));
3648 if (self && !node_online(self->node_id)) {
3649 build_zonelists(self);
3650 build_zonelist_cache(self);
3653 for_each_online_node(nid) {
3654 pg_data_t *pgdat = NODE_DATA(nid);
3656 build_zonelists(pgdat);
3657 build_zonelist_cache(pgdat);
3661 * Initialize the boot_pagesets that are going to be used
3662 * for bootstrapping processors. The real pagesets for
3663 * each zone will be allocated later when the per cpu
3664 * allocator is available.
3666 * boot_pagesets are used also for bootstrapping offline
3667 * cpus if the system is already booted because the pagesets
3668 * are needed to initialize allocators on a specific cpu too.
3669 * F.e. the percpu allocator needs the page allocator which
3670 * needs the percpu allocator in order to allocate its pagesets
3671 * (a chicken-egg dilemma).
3673 for_each_possible_cpu(cpu) {
3674 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3676 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3678 * We now know the "local memory node" for each node--
3679 * i.e., the node of the first zone in the generic zonelist.
3680 * Set up numa_mem percpu variable for on-line cpus. During
3681 * boot, only the boot cpu should be on-line; we'll init the
3682 * secondary cpus' numa_mem as they come on-line. During
3683 * node/memory hotplug, we'll fixup all on-line cpus.
3685 if (cpu_online(cpu))
3686 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3694 * Called with zonelists_mutex held always
3695 * unless system_state == SYSTEM_BOOTING.
3697 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3699 set_zonelist_order();
3701 if (system_state == SYSTEM_BOOTING) {
3702 __build_all_zonelists(NULL);
3703 mminit_verify_zonelist();
3704 cpuset_init_current_mems_allowed();
3706 /* we have to stop all cpus to guarantee there is no user
3708 #ifdef CONFIG_MEMORY_HOTPLUG
3710 setup_zone_pageset(zone);
3712 stop_machine(__build_all_zonelists, pgdat, NULL);
3713 /* cpuset refresh routine should be here */
3715 vm_total_pages = nr_free_pagecache_pages();
3717 * Disable grouping by mobility if the number of pages in the
3718 * system is too low to allow the mechanism to work. It would be
3719 * more accurate, but expensive to check per-zone. This check is
3720 * made on memory-hotadd so a system can start with mobility
3721 * disabled and enable it later
3723 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3724 page_group_by_mobility_disabled = 1;
3726 page_group_by_mobility_disabled = 0;
3728 printk("Built %i zonelists in %s order, mobility grouping %s. "
3729 "Total pages: %ld\n",
3731 zonelist_order_name[current_zonelist_order],
3732 page_group_by_mobility_disabled ? "off" : "on",
3735 printk("Policy zone: %s\n", zone_names[policy_zone]);
3740 * Helper functions to size the waitqueue hash table.
3741 * Essentially these want to choose hash table sizes sufficiently
3742 * large so that collisions trying to wait on pages are rare.
3743 * But in fact, the number of active page waitqueues on typical
3744 * systems is ridiculously low, less than 200. So this is even
3745 * conservative, even though it seems large.
3747 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3748 * waitqueues, i.e. the size of the waitq table given the number of pages.
3750 #define PAGES_PER_WAITQUEUE 256
3752 #ifndef CONFIG_MEMORY_HOTPLUG
3753 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3755 unsigned long size = 1;
3757 pages /= PAGES_PER_WAITQUEUE;
3759 while (size < pages)
3763 * Once we have dozens or even hundreds of threads sleeping
3764 * on IO we've got bigger problems than wait queue collision.
3765 * Limit the size of the wait table to a reasonable size.
3767 size = min(size, 4096UL);
3769 return max(size, 4UL);
3773 * A zone's size might be changed by hot-add, so it is not possible to determine
3774 * a suitable size for its wait_table. So we use the maximum size now.
3776 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3778 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3779 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3780 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3782 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3783 * or more by the traditional way. (See above). It equals:
3785 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3786 * ia64(16K page size) : = ( 8G + 4M)byte.
3787 * powerpc (64K page size) : = (32G +16M)byte.
3789 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3796 * This is an integer logarithm so that shifts can be used later
3797 * to extract the more random high bits from the multiplicative
3798 * hash function before the remainder is taken.
3800 static inline unsigned long wait_table_bits(unsigned long size)
3805 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3808 * Check if a pageblock contains reserved pages
3810 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3814 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3815 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3822 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3823 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3824 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3825 * higher will lead to a bigger reserve which will get freed as contiguous
3826 * blocks as reclaim kicks in
3828 static void setup_zone_migrate_reserve(struct zone *zone)
3830 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3832 unsigned long block_migratetype;
3836 * Get the start pfn, end pfn and the number of blocks to reserve
3837 * We have to be careful to be aligned to pageblock_nr_pages to
3838 * make sure that we always check pfn_valid for the first page in
3841 start_pfn = zone->zone_start_pfn;
3842 end_pfn = zone_end_pfn(zone);
3843 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3844 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3848 * Reserve blocks are generally in place to help high-order atomic
3849 * allocations that are short-lived. A min_free_kbytes value that
3850 * would result in more than 2 reserve blocks for atomic allocations
3851 * is assumed to be in place to help anti-fragmentation for the
3852 * future allocation of hugepages at runtime.
3854 reserve = min(2, reserve);
3856 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3857 if (!pfn_valid(pfn))
3859 page = pfn_to_page(pfn);
3861 /* Watch out for overlapping nodes */
3862 if (page_to_nid(page) != zone_to_nid(zone))
3865 block_migratetype = get_pageblock_migratetype(page);
3867 /* Only test what is necessary when the reserves are not met */
3870 * Blocks with reserved pages will never free, skip
3873 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3874 if (pageblock_is_reserved(pfn, block_end_pfn))
3877 /* If this block is reserved, account for it */
3878 if (block_migratetype == MIGRATE_RESERVE) {
3883 /* Suitable for reserving if this block is movable */
3884 if (block_migratetype == MIGRATE_MOVABLE) {
3885 set_pageblock_migratetype(page,
3887 move_freepages_block(zone, page,
3895 * If the reserve is met and this is a previous reserved block,
3898 if (block_migratetype == MIGRATE_RESERVE) {
3899 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3900 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3906 * Initially all pages are reserved - free ones are freed
3907 * up by free_all_bootmem() once the early boot process is
3908 * done. Non-atomic initialization, single-pass.
3910 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3911 unsigned long start_pfn, enum memmap_context context)
3914 unsigned long end_pfn = start_pfn + size;
3918 if (highest_memmap_pfn < end_pfn - 1)
3919 highest_memmap_pfn = end_pfn - 1;
3921 z = &NODE_DATA(nid)->node_zones[zone];
3922 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3924 * There can be holes in boot-time mem_map[]s
3925 * handed to this function. They do not
3926 * exist on hotplugged memory.
3928 if (context == MEMMAP_EARLY) {
3929 if (!early_pfn_valid(pfn))
3931 if (!early_pfn_in_nid(pfn, nid))
3934 page = pfn_to_page(pfn);
3935 set_page_links(page, zone, nid, pfn);
3936 mminit_verify_page_links(page, zone, nid, pfn);
3937 init_page_count(page);
3938 page_mapcount_reset(page);
3939 page_nid_reset_last(page);
3940 SetPageReserved(page);
3942 * Mark the block movable so that blocks are reserved for
3943 * movable at startup. This will force kernel allocations
3944 * to reserve their blocks rather than leaking throughout
3945 * the address space during boot when many long-lived
3946 * kernel allocations are made. Later some blocks near
3947 * the start are marked MIGRATE_RESERVE by
3948 * setup_zone_migrate_reserve()
3950 * bitmap is created for zone's valid pfn range. but memmap
3951 * can be created for invalid pages (for alignment)
3952 * check here not to call set_pageblock_migratetype() against
3955 if ((z->zone_start_pfn <= pfn)
3956 && (pfn < zone_end_pfn(z))
3957 && !(pfn & (pageblock_nr_pages - 1)))
3958 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3960 INIT_LIST_HEAD(&page->lru);
3961 #ifdef WANT_PAGE_VIRTUAL
3962 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3963 if (!is_highmem_idx(zone))
3964 set_page_address(page, __va(pfn << PAGE_SHIFT));
3969 static void __meminit zone_init_free_lists(struct zone *zone)
3972 for_each_migratetype_order(order, t) {
3973 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3974 zone->free_area[order].nr_free = 0;
3978 #ifndef __HAVE_ARCH_MEMMAP_INIT
3979 #define memmap_init(size, nid, zone, start_pfn) \
3980 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3983 static int __meminit zone_batchsize(struct zone *zone)
3989 * The per-cpu-pages pools are set to around 1000th of the
3990 * size of the zone. But no more than 1/2 of a meg.
3992 * OK, so we don't know how big the cache is. So guess.
3994 batch = zone->managed_pages / 1024;
3995 if (batch * PAGE_SIZE > 512 * 1024)
3996 batch = (512 * 1024) / PAGE_SIZE;
3997 batch /= 4; /* We effectively *= 4 below */
4002 * Clamp the batch to a 2^n - 1 value. Having a power
4003 * of 2 value was found to be more likely to have
4004 * suboptimal cache aliasing properties in some cases.
4006 * For example if 2 tasks are alternately allocating
4007 * batches of pages, one task can end up with a lot
4008 * of pages of one half of the possible page colors
4009 * and the other with pages of the other colors.
4011 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4016 /* The deferral and batching of frees should be suppressed under NOMMU
4019 * The problem is that NOMMU needs to be able to allocate large chunks
4020 * of contiguous memory as there's no hardware page translation to
4021 * assemble apparent contiguous memory from discontiguous pages.
4023 * Queueing large contiguous runs of pages for batching, however,
4024 * causes the pages to actually be freed in smaller chunks. As there
4025 * can be a significant delay between the individual batches being
4026 * recycled, this leads to the once large chunks of space being
4027 * fragmented and becoming unavailable for high-order allocations.
4033 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4035 struct per_cpu_pages *pcp;
4038 memset(p, 0, sizeof(*p));
4042 pcp->high = 6 * batch;
4043 pcp->batch = max(1UL, 1 * batch);
4044 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4045 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4049 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4050 * to the value high for the pageset p.
4053 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
4056 struct per_cpu_pages *pcp;
4060 pcp->batch = max(1UL, high/4);
4061 if ((high/4) > (PAGE_SHIFT * 8))
4062 pcp->batch = PAGE_SHIFT * 8;
4065 static void __meminit setup_zone_pageset(struct zone *zone)
4069 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4071 for_each_possible_cpu(cpu) {
4072 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4074 setup_pageset(pcp, zone_batchsize(zone));
4076 if (percpu_pagelist_fraction)
4077 setup_pagelist_highmark(pcp,
4078 (zone->managed_pages /
4079 percpu_pagelist_fraction));
4084 * Allocate per cpu pagesets and initialize them.
4085 * Before this call only boot pagesets were available.
4087 void __init setup_per_cpu_pageset(void)
4091 for_each_populated_zone(zone)
4092 setup_zone_pageset(zone);
4095 static noinline __init_refok
4096 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4099 struct pglist_data *pgdat = zone->zone_pgdat;
4103 * The per-page waitqueue mechanism uses hashed waitqueues
4106 zone->wait_table_hash_nr_entries =
4107 wait_table_hash_nr_entries(zone_size_pages);
4108 zone->wait_table_bits =
4109 wait_table_bits(zone->wait_table_hash_nr_entries);
4110 alloc_size = zone->wait_table_hash_nr_entries
4111 * sizeof(wait_queue_head_t);
4113 if (!slab_is_available()) {
4114 zone->wait_table = (wait_queue_head_t *)
4115 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4118 * This case means that a zone whose size was 0 gets new memory
4119 * via memory hot-add.
4120 * But it may be the case that a new node was hot-added. In
4121 * this case vmalloc() will not be able to use this new node's
4122 * memory - this wait_table must be initialized to use this new
4123 * node itself as well.
4124 * To use this new node's memory, further consideration will be
4127 zone->wait_table = vmalloc(alloc_size);
4129 if (!zone->wait_table)
4132 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4133 init_waitqueue_head(zone->wait_table + i);
4138 static __meminit void zone_pcp_init(struct zone *zone)
4141 * per cpu subsystem is not up at this point. The following code
4142 * relies on the ability of the linker to provide the
4143 * offset of a (static) per cpu variable into the per cpu area.
4145 zone->pageset = &boot_pageset;
4147 if (zone->present_pages)
4148 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4149 zone->name, zone->present_pages,
4150 zone_batchsize(zone));
4153 int __meminit init_currently_empty_zone(struct zone *zone,
4154 unsigned long zone_start_pfn,
4156 enum memmap_context context)
4158 struct pglist_data *pgdat = zone->zone_pgdat;
4160 ret = zone_wait_table_init(zone, size);
4163 pgdat->nr_zones = zone_idx(zone) + 1;
4165 zone->zone_start_pfn = zone_start_pfn;
4167 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4168 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4170 (unsigned long)zone_idx(zone),
4171 zone_start_pfn, (zone_start_pfn + size));
4173 zone_init_free_lists(zone);
4178 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4179 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4181 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4182 * Architectures may implement their own version but if add_active_range()
4183 * was used and there are no special requirements, this is a convenient
4186 int __meminit __early_pfn_to_nid(unsigned long pfn)
4188 unsigned long start_pfn, end_pfn;
4191 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4192 if (start_pfn <= pfn && pfn < end_pfn)
4194 /* This is a memory hole */
4197 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4199 int __meminit early_pfn_to_nid(unsigned long pfn)
4203 nid = __early_pfn_to_nid(pfn);
4206 /* just returns 0 */
4210 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4211 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4215 nid = __early_pfn_to_nid(pfn);
4216 if (nid >= 0 && nid != node)
4223 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4224 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4225 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4227 * If an architecture guarantees that all ranges registered with
4228 * add_active_ranges() contain no holes and may be freed, this
4229 * this function may be used instead of calling free_bootmem() manually.
4231 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4233 unsigned long start_pfn, end_pfn;
4236 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4237 start_pfn = min(start_pfn, max_low_pfn);
4238 end_pfn = min(end_pfn, max_low_pfn);
4240 if (start_pfn < end_pfn)
4241 free_bootmem_node(NODE_DATA(this_nid),
4242 PFN_PHYS(start_pfn),
4243 (end_pfn - start_pfn) << PAGE_SHIFT);
4248 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4249 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4251 * If an architecture guarantees that all ranges registered with
4252 * add_active_ranges() contain no holes and may be freed, this
4253 * function may be used instead of calling memory_present() manually.
4255 void __init sparse_memory_present_with_active_regions(int nid)
4257 unsigned long start_pfn, end_pfn;
4260 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4261 memory_present(this_nid, start_pfn, end_pfn);
4265 * get_pfn_range_for_nid - Return the start and end page frames for a node
4266 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4267 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4268 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4270 * It returns the start and end page frame of a node based on information
4271 * provided by an arch calling add_active_range(). If called for a node
4272 * with no available memory, a warning is printed and the start and end
4275 void __meminit get_pfn_range_for_nid(unsigned int nid,
4276 unsigned long *start_pfn, unsigned long *end_pfn)
4278 unsigned long this_start_pfn, this_end_pfn;
4284 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4285 *start_pfn = min(*start_pfn, this_start_pfn);
4286 *end_pfn = max(*end_pfn, this_end_pfn);
4289 if (*start_pfn == -1UL)
4294 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4295 * assumption is made that zones within a node are ordered in monotonic
4296 * increasing memory addresses so that the "highest" populated zone is used
4298 static void __init find_usable_zone_for_movable(void)
4301 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4302 if (zone_index == ZONE_MOVABLE)
4305 if (arch_zone_highest_possible_pfn[zone_index] >
4306 arch_zone_lowest_possible_pfn[zone_index])
4310 VM_BUG_ON(zone_index == -1);
4311 movable_zone = zone_index;
4315 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4316 * because it is sized independent of architecture. Unlike the other zones,
4317 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4318 * in each node depending on the size of each node and how evenly kernelcore
4319 * is distributed. This helper function adjusts the zone ranges
4320 * provided by the architecture for a given node by using the end of the
4321 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4322 * zones within a node are in order of monotonic increases memory addresses
4324 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4325 unsigned long zone_type,
4326 unsigned long node_start_pfn,
4327 unsigned long node_end_pfn,
4328 unsigned long *zone_start_pfn,
4329 unsigned long *zone_end_pfn)
4331 /* Only adjust if ZONE_MOVABLE is on this node */
4332 if (zone_movable_pfn[nid]) {
4333 /* Size ZONE_MOVABLE */
4334 if (zone_type == ZONE_MOVABLE) {
4335 *zone_start_pfn = zone_movable_pfn[nid];
4336 *zone_end_pfn = min(node_end_pfn,
4337 arch_zone_highest_possible_pfn[movable_zone]);
4339 /* Adjust for ZONE_MOVABLE starting within this range */
4340 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4341 *zone_end_pfn > zone_movable_pfn[nid]) {
4342 *zone_end_pfn = zone_movable_pfn[nid];
4344 /* Check if this whole range is within ZONE_MOVABLE */
4345 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4346 *zone_start_pfn = *zone_end_pfn;
4351 * Return the number of pages a zone spans in a node, including holes
4352 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4354 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4355 unsigned long zone_type,
4356 unsigned long *ignored)
4358 unsigned long node_start_pfn, node_end_pfn;
4359 unsigned long zone_start_pfn, zone_end_pfn;
4361 /* Get the start and end of the node and zone */
4362 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4363 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4364 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4365 adjust_zone_range_for_zone_movable(nid, zone_type,
4366 node_start_pfn, node_end_pfn,
4367 &zone_start_pfn, &zone_end_pfn);
4369 /* Check that this node has pages within the zone's required range */
4370 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4373 /* Move the zone boundaries inside the node if necessary */
4374 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4375 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4377 /* Return the spanned pages */
4378 return zone_end_pfn - zone_start_pfn;
4382 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4383 * then all holes in the requested range will be accounted for.
4385 unsigned long __meminit __absent_pages_in_range(int nid,
4386 unsigned long range_start_pfn,
4387 unsigned long range_end_pfn)
4389 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4390 unsigned long start_pfn, end_pfn;
4393 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4394 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4395 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4396 nr_absent -= end_pfn - start_pfn;
4402 * absent_pages_in_range - Return number of page frames in holes within a range
4403 * @start_pfn: The start PFN to start searching for holes
4404 * @end_pfn: The end PFN to stop searching for holes
4406 * It returns the number of pages frames in memory holes within a range.
4408 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4409 unsigned long end_pfn)
4411 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4414 /* Return the number of page frames in holes in a zone on a node */
4415 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4416 unsigned long zone_type,
4417 unsigned long *ignored)
4419 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4420 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4421 unsigned long node_start_pfn, node_end_pfn;
4422 unsigned long zone_start_pfn, zone_end_pfn;
4424 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4425 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4426 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4428 adjust_zone_range_for_zone_movable(nid, zone_type,
4429 node_start_pfn, node_end_pfn,
4430 &zone_start_pfn, &zone_end_pfn);
4431 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4434 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4435 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4436 unsigned long zone_type,
4437 unsigned long *zones_size)
4439 return zones_size[zone_type];
4442 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4443 unsigned long zone_type,
4444 unsigned long *zholes_size)
4449 return zholes_size[zone_type];
4452 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4454 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4455 unsigned long *zones_size, unsigned long *zholes_size)
4457 unsigned long realtotalpages, totalpages = 0;
4460 for (i = 0; i < MAX_NR_ZONES; i++)
4461 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4463 pgdat->node_spanned_pages = totalpages;
4465 realtotalpages = totalpages;
4466 for (i = 0; i < MAX_NR_ZONES; i++)
4468 zone_absent_pages_in_node(pgdat->node_id, i,
4470 pgdat->node_present_pages = realtotalpages;
4471 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4475 #ifndef CONFIG_SPARSEMEM
4477 * Calculate the size of the zone->blockflags rounded to an unsigned long
4478 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4479 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4480 * round what is now in bits to nearest long in bits, then return it in
4483 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4485 unsigned long usemapsize;
4487 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4488 usemapsize = roundup(zonesize, pageblock_nr_pages);
4489 usemapsize = usemapsize >> pageblock_order;
4490 usemapsize *= NR_PAGEBLOCK_BITS;
4491 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4493 return usemapsize / 8;
4496 static void __init setup_usemap(struct pglist_data *pgdat,
4498 unsigned long zone_start_pfn,
4499 unsigned long zonesize)
4501 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4502 zone->pageblock_flags = NULL;
4504 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4508 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4509 unsigned long zone_start_pfn, unsigned long zonesize) {}
4510 #endif /* CONFIG_SPARSEMEM */
4512 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4514 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4515 void __init set_pageblock_order(void)
4519 /* Check that pageblock_nr_pages has not already been setup */
4520 if (pageblock_order)
4523 if (HPAGE_SHIFT > PAGE_SHIFT)
4524 order = HUGETLB_PAGE_ORDER;
4526 order = MAX_ORDER - 1;
4529 * Assume the largest contiguous order of interest is a huge page.
4530 * This value may be variable depending on boot parameters on IA64 and
4533 pageblock_order = order;
4535 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4538 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4539 * is unused as pageblock_order is set at compile-time. See
4540 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4543 void __init set_pageblock_order(void)
4547 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4549 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4550 unsigned long present_pages)
4552 unsigned long pages = spanned_pages;
4555 * Provide a more accurate estimation if there are holes within
4556 * the zone and SPARSEMEM is in use. If there are holes within the
4557 * zone, each populated memory region may cost us one or two extra
4558 * memmap pages due to alignment because memmap pages for each
4559 * populated regions may not naturally algined on page boundary.
4560 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4562 if (spanned_pages > present_pages + (present_pages >> 4) &&
4563 IS_ENABLED(CONFIG_SPARSEMEM))
4564 pages = present_pages;
4566 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4570 * Set up the zone data structures:
4571 * - mark all pages reserved
4572 * - mark all memory queues empty
4573 * - clear the memory bitmaps
4575 * NOTE: pgdat should get zeroed by caller.
4577 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4578 unsigned long *zones_size, unsigned long *zholes_size)
4581 int nid = pgdat->node_id;
4582 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4585 pgdat_resize_init(pgdat);
4586 #ifdef CONFIG_NUMA_BALANCING
4587 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4588 pgdat->numabalancing_migrate_nr_pages = 0;
4589 pgdat->numabalancing_migrate_next_window = jiffies;
4591 init_waitqueue_head(&pgdat->kswapd_wait);
4592 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4593 pgdat_page_cgroup_init(pgdat);
4595 for (j = 0; j < MAX_NR_ZONES; j++) {
4596 struct zone *zone = pgdat->node_zones + j;
4597 unsigned long size, realsize, freesize, memmap_pages;
4599 size = zone_spanned_pages_in_node(nid, j, zones_size);
4600 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4604 * Adjust freesize so that it accounts for how much memory
4605 * is used by this zone for memmap. This affects the watermark
4606 * and per-cpu initialisations
4608 memmap_pages = calc_memmap_size(size, realsize);
4609 if (freesize >= memmap_pages) {
4610 freesize -= memmap_pages;
4613 " %s zone: %lu pages used for memmap\n",
4614 zone_names[j], memmap_pages);
4617 " %s zone: %lu pages exceeds freesize %lu\n",
4618 zone_names[j], memmap_pages, freesize);
4620 /* Account for reserved pages */
4621 if (j == 0 && freesize > dma_reserve) {
4622 freesize -= dma_reserve;
4623 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4624 zone_names[0], dma_reserve);
4627 if (!is_highmem_idx(j))
4628 nr_kernel_pages += freesize;
4629 /* Charge for highmem memmap if there are enough kernel pages */
4630 else if (nr_kernel_pages > memmap_pages * 2)
4631 nr_kernel_pages -= memmap_pages;
4632 nr_all_pages += freesize;
4634 zone->spanned_pages = size;
4635 zone->present_pages = realsize;
4637 * Set an approximate value for lowmem here, it will be adjusted
4638 * when the bootmem allocator frees pages into the buddy system.
4639 * And all highmem pages will be managed by the buddy system.
4641 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4644 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4646 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4648 zone->name = zone_names[j];
4649 spin_lock_init(&zone->lock);
4650 spin_lock_init(&zone->lru_lock);
4651 zone_seqlock_init(zone);
4652 zone->zone_pgdat = pgdat;
4654 zone_pcp_init(zone);
4655 lruvec_init(&zone->lruvec);
4659 set_pageblock_order();
4660 setup_usemap(pgdat, zone, zone_start_pfn, size);
4661 ret = init_currently_empty_zone(zone, zone_start_pfn,
4662 size, MEMMAP_EARLY);
4664 memmap_init(size, nid, j, zone_start_pfn);
4665 zone_start_pfn += size;
4669 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4671 /* Skip empty nodes */
4672 if (!pgdat->node_spanned_pages)
4675 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4676 /* ia64 gets its own node_mem_map, before this, without bootmem */
4677 if (!pgdat->node_mem_map) {
4678 unsigned long size, start, end;
4682 * The zone's endpoints aren't required to be MAX_ORDER
4683 * aligned but the node_mem_map endpoints must be in order
4684 * for the buddy allocator to function correctly.
4686 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4687 end = pgdat_end_pfn(pgdat);
4688 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4689 size = (end - start) * sizeof(struct page);
4690 map = alloc_remap(pgdat->node_id, size);
4692 map = alloc_bootmem_node_nopanic(pgdat, size);
4693 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4695 #ifndef CONFIG_NEED_MULTIPLE_NODES
4697 * With no DISCONTIG, the global mem_map is just set as node 0's
4699 if (pgdat == NODE_DATA(0)) {
4700 mem_map = NODE_DATA(0)->node_mem_map;
4701 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4702 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4703 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4704 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4707 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4710 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4711 unsigned long node_start_pfn, unsigned long *zholes_size)
4713 pg_data_t *pgdat = NODE_DATA(nid);
4715 /* pg_data_t should be reset to zero when it's allocated */
4716 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4718 pgdat->node_id = nid;
4719 pgdat->node_start_pfn = node_start_pfn;
4720 init_zone_allows_reclaim(nid);
4721 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4723 alloc_node_mem_map(pgdat);
4724 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4725 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4726 nid, (unsigned long)pgdat,
4727 (unsigned long)pgdat->node_mem_map);
4730 free_area_init_core(pgdat, zones_size, zholes_size);
4733 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4735 #if MAX_NUMNODES > 1
4737 * Figure out the number of possible node ids.
4739 static void __init setup_nr_node_ids(void)
4742 unsigned int highest = 0;
4744 for_each_node_mask(node, node_possible_map)
4746 nr_node_ids = highest + 1;
4749 static inline void setup_nr_node_ids(void)
4755 * node_map_pfn_alignment - determine the maximum internode alignment
4757 * This function should be called after node map is populated and sorted.
4758 * It calculates the maximum power of two alignment which can distinguish
4761 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4762 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4763 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4764 * shifted, 1GiB is enough and this function will indicate so.
4766 * This is used to test whether pfn -> nid mapping of the chosen memory
4767 * model has fine enough granularity to avoid incorrect mapping for the
4768 * populated node map.
4770 * Returns the determined alignment in pfn's. 0 if there is no alignment
4771 * requirement (single node).
4773 unsigned long __init node_map_pfn_alignment(void)
4775 unsigned long accl_mask = 0, last_end = 0;
4776 unsigned long start, end, mask;
4780 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4781 if (!start || last_nid < 0 || last_nid == nid) {
4788 * Start with a mask granular enough to pin-point to the
4789 * start pfn and tick off bits one-by-one until it becomes
4790 * too coarse to separate the current node from the last.
4792 mask = ~((1 << __ffs(start)) - 1);
4793 while (mask && last_end <= (start & (mask << 1)))
4796 /* accumulate all internode masks */
4800 /* convert mask to number of pages */
4801 return ~accl_mask + 1;
4804 /* Find the lowest pfn for a node */
4805 static unsigned long __init find_min_pfn_for_node(int nid)
4807 unsigned long min_pfn = ULONG_MAX;
4808 unsigned long start_pfn;
4811 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4812 min_pfn = min(min_pfn, start_pfn);
4814 if (min_pfn == ULONG_MAX) {
4816 "Could not find start_pfn for node %d\n", nid);
4824 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4826 * It returns the minimum PFN based on information provided via
4827 * add_active_range().
4829 unsigned long __init find_min_pfn_with_active_regions(void)
4831 return find_min_pfn_for_node(MAX_NUMNODES);
4835 * early_calculate_totalpages()
4836 * Sum pages in active regions for movable zone.
4837 * Populate N_MEMORY for calculating usable_nodes.
4839 static unsigned long __init early_calculate_totalpages(void)
4841 unsigned long totalpages = 0;
4842 unsigned long start_pfn, end_pfn;
4845 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4846 unsigned long pages = end_pfn - start_pfn;
4848 totalpages += pages;
4850 node_set_state(nid, N_MEMORY);
4856 * Find the PFN the Movable zone begins in each node. Kernel memory
4857 * is spread evenly between nodes as long as the nodes have enough
4858 * memory. When they don't, some nodes will have more kernelcore than
4861 static void __init find_zone_movable_pfns_for_nodes(void)
4864 unsigned long usable_startpfn;
4865 unsigned long kernelcore_node, kernelcore_remaining;
4866 /* save the state before borrow the nodemask */
4867 nodemask_t saved_node_state = node_states[N_MEMORY];
4868 unsigned long totalpages = early_calculate_totalpages();
4869 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4872 * If movablecore was specified, calculate what size of
4873 * kernelcore that corresponds so that memory usable for
4874 * any allocation type is evenly spread. If both kernelcore
4875 * and movablecore are specified, then the value of kernelcore
4876 * will be used for required_kernelcore if it's greater than
4877 * what movablecore would have allowed.
4879 if (required_movablecore) {
4880 unsigned long corepages;
4883 * Round-up so that ZONE_MOVABLE is at least as large as what
4884 * was requested by the user
4886 required_movablecore =
4887 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4888 corepages = totalpages - required_movablecore;
4890 required_kernelcore = max(required_kernelcore, corepages);
4893 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4894 if (!required_kernelcore)
4897 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4898 find_usable_zone_for_movable();
4899 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4902 /* Spread kernelcore memory as evenly as possible throughout nodes */
4903 kernelcore_node = required_kernelcore / usable_nodes;
4904 for_each_node_state(nid, N_MEMORY) {
4905 unsigned long start_pfn, end_pfn;
4908 * Recalculate kernelcore_node if the division per node
4909 * now exceeds what is necessary to satisfy the requested
4910 * amount of memory for the kernel
4912 if (required_kernelcore < kernelcore_node)
4913 kernelcore_node = required_kernelcore / usable_nodes;
4916 * As the map is walked, we track how much memory is usable
4917 * by the kernel using kernelcore_remaining. When it is
4918 * 0, the rest of the node is usable by ZONE_MOVABLE
4920 kernelcore_remaining = kernelcore_node;
4922 /* Go through each range of PFNs within this node */
4923 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4924 unsigned long size_pages;
4926 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4927 if (start_pfn >= end_pfn)
4930 /* Account for what is only usable for kernelcore */
4931 if (start_pfn < usable_startpfn) {
4932 unsigned long kernel_pages;
4933 kernel_pages = min(end_pfn, usable_startpfn)
4936 kernelcore_remaining -= min(kernel_pages,
4937 kernelcore_remaining);
4938 required_kernelcore -= min(kernel_pages,
4939 required_kernelcore);
4941 /* Continue if range is now fully accounted */
4942 if (end_pfn <= usable_startpfn) {
4945 * Push zone_movable_pfn to the end so
4946 * that if we have to rebalance
4947 * kernelcore across nodes, we will
4948 * not double account here
4950 zone_movable_pfn[nid] = end_pfn;
4953 start_pfn = usable_startpfn;
4957 * The usable PFN range for ZONE_MOVABLE is from
4958 * start_pfn->end_pfn. Calculate size_pages as the
4959 * number of pages used as kernelcore
4961 size_pages = end_pfn - start_pfn;
4962 if (size_pages > kernelcore_remaining)
4963 size_pages = kernelcore_remaining;
4964 zone_movable_pfn[nid] = start_pfn + size_pages;
4967 * Some kernelcore has been met, update counts and
4968 * break if the kernelcore for this node has been
4971 required_kernelcore -= min(required_kernelcore,
4973 kernelcore_remaining -= size_pages;
4974 if (!kernelcore_remaining)
4980 * If there is still required_kernelcore, we do another pass with one
4981 * less node in the count. This will push zone_movable_pfn[nid] further
4982 * along on the nodes that still have memory until kernelcore is
4986 if (usable_nodes && required_kernelcore > usable_nodes)
4989 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4990 for (nid = 0; nid < MAX_NUMNODES; nid++)
4991 zone_movable_pfn[nid] =
4992 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4995 /* restore the node_state */
4996 node_states[N_MEMORY] = saved_node_state;
4999 /* Any regular or high memory on that node ? */
5000 static void check_for_memory(pg_data_t *pgdat, int nid)
5002 enum zone_type zone_type;
5004 if (N_MEMORY == N_NORMAL_MEMORY)
5007 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5008 struct zone *zone = &pgdat->node_zones[zone_type];
5009 if (zone->present_pages) {
5010 node_set_state(nid, N_HIGH_MEMORY);
5011 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5012 zone_type <= ZONE_NORMAL)
5013 node_set_state(nid, N_NORMAL_MEMORY);
5020 * free_area_init_nodes - Initialise all pg_data_t and zone data
5021 * @max_zone_pfn: an array of max PFNs for each zone
5023 * This will call free_area_init_node() for each active node in the system.
5024 * Using the page ranges provided by add_active_range(), the size of each
5025 * zone in each node and their holes is calculated. If the maximum PFN
5026 * between two adjacent zones match, it is assumed that the zone is empty.
5027 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5028 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5029 * starts where the previous one ended. For example, ZONE_DMA32 starts
5030 * at arch_max_dma_pfn.
5032 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5034 unsigned long start_pfn, end_pfn;
5037 /* Record where the zone boundaries are */
5038 memset(arch_zone_lowest_possible_pfn, 0,
5039 sizeof(arch_zone_lowest_possible_pfn));
5040 memset(arch_zone_highest_possible_pfn, 0,
5041 sizeof(arch_zone_highest_possible_pfn));
5042 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5043 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5044 for (i = 1; i < MAX_NR_ZONES; i++) {
5045 if (i == ZONE_MOVABLE)
5047 arch_zone_lowest_possible_pfn[i] =
5048 arch_zone_highest_possible_pfn[i-1];
5049 arch_zone_highest_possible_pfn[i] =
5050 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5052 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5053 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5055 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5056 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5057 find_zone_movable_pfns_for_nodes();
5059 /* Print out the zone ranges */
5060 printk("Zone ranges:\n");
5061 for (i = 0; i < MAX_NR_ZONES; i++) {
5062 if (i == ZONE_MOVABLE)
5064 printk(KERN_CONT " %-8s ", zone_names[i]);
5065 if (arch_zone_lowest_possible_pfn[i] ==
5066 arch_zone_highest_possible_pfn[i])
5067 printk(KERN_CONT "empty\n");
5069 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5070 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5071 (arch_zone_highest_possible_pfn[i]
5072 << PAGE_SHIFT) - 1);
5075 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5076 printk("Movable zone start for each node\n");
5077 for (i = 0; i < MAX_NUMNODES; i++) {
5078 if (zone_movable_pfn[i])
5079 printk(" Node %d: %#010lx\n", i,
5080 zone_movable_pfn[i] << PAGE_SHIFT);
5083 /* Print out the early node map */
5084 printk("Early memory node ranges\n");
5085 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5086 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5087 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5089 /* Initialise every node */
5090 mminit_verify_pageflags_layout();
5091 setup_nr_node_ids();
5092 for_each_online_node(nid) {
5093 pg_data_t *pgdat = NODE_DATA(nid);
5094 free_area_init_node(nid, NULL,
5095 find_min_pfn_for_node(nid), NULL);
5097 /* Any memory on that node */
5098 if (pgdat->node_present_pages)
5099 node_set_state(nid, N_MEMORY);
5100 check_for_memory(pgdat, nid);
5104 static int __init cmdline_parse_core(char *p, unsigned long *core)
5106 unsigned long long coremem;
5110 coremem = memparse(p, &p);
5111 *core = coremem >> PAGE_SHIFT;
5113 /* Paranoid check that UL is enough for the coremem value */
5114 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5120 * kernelcore=size sets the amount of memory for use for allocations that
5121 * cannot be reclaimed or migrated.
5123 static int __init cmdline_parse_kernelcore(char *p)
5125 return cmdline_parse_core(p, &required_kernelcore);
5129 * movablecore=size sets the amount of memory for use for allocations that
5130 * can be reclaimed or migrated.
5132 static int __init cmdline_parse_movablecore(char *p)
5134 return cmdline_parse_core(p, &required_movablecore);
5137 early_param("kernelcore", cmdline_parse_kernelcore);
5138 early_param("movablecore", cmdline_parse_movablecore);
5140 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5142 unsigned long free_reserved_area(unsigned long start, unsigned long end,
5143 int poison, char *s)
5145 unsigned long pages, pos;
5147 pos = start = PAGE_ALIGN(start);
5149 for (pages = 0; pos < end; pos += PAGE_SIZE, pages++) {
5151 memset((void *)pos, poison, PAGE_SIZE);
5152 free_reserved_page(virt_to_page(pos));
5156 pr_info("Freeing %s memory: %ldK (%lx - %lx)\n",
5157 s, pages << (PAGE_SHIFT - 10), start, end);
5162 #ifdef CONFIG_HIGHMEM
5163 void free_highmem_page(struct page *page)
5165 __free_reserved_page(page);
5172 * set_dma_reserve - set the specified number of pages reserved in the first zone
5173 * @new_dma_reserve: The number of pages to mark reserved
5175 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5176 * In the DMA zone, a significant percentage may be consumed by kernel image
5177 * and other unfreeable allocations which can skew the watermarks badly. This
5178 * function may optionally be used to account for unfreeable pages in the
5179 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5180 * smaller per-cpu batchsize.
5182 void __init set_dma_reserve(unsigned long new_dma_reserve)
5184 dma_reserve = new_dma_reserve;
5187 void __init free_area_init(unsigned long *zones_size)
5189 free_area_init_node(0, zones_size,
5190 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5193 static int page_alloc_cpu_notify(struct notifier_block *self,
5194 unsigned long action, void *hcpu)
5196 int cpu = (unsigned long)hcpu;
5198 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5199 lru_add_drain_cpu(cpu);
5203 * Spill the event counters of the dead processor
5204 * into the current processors event counters.
5205 * This artificially elevates the count of the current
5208 vm_events_fold_cpu(cpu);
5211 * Zero the differential counters of the dead processor
5212 * so that the vm statistics are consistent.
5214 * This is only okay since the processor is dead and cannot
5215 * race with what we are doing.
5217 refresh_cpu_vm_stats(cpu);
5222 void __init page_alloc_init(void)
5224 hotcpu_notifier(page_alloc_cpu_notify, 0);
5228 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5229 * or min_free_kbytes changes.
5231 static void calculate_totalreserve_pages(void)
5233 struct pglist_data *pgdat;
5234 unsigned long reserve_pages = 0;
5235 enum zone_type i, j;
5237 for_each_online_pgdat(pgdat) {
5238 for (i = 0; i < MAX_NR_ZONES; i++) {
5239 struct zone *zone = pgdat->node_zones + i;
5240 unsigned long max = 0;
5242 /* Find valid and maximum lowmem_reserve in the zone */
5243 for (j = i; j < MAX_NR_ZONES; j++) {
5244 if (zone->lowmem_reserve[j] > max)
5245 max = zone->lowmem_reserve[j];
5248 /* we treat the high watermark as reserved pages. */
5249 max += high_wmark_pages(zone);
5251 if (max > zone->managed_pages)
5252 max = zone->managed_pages;
5253 reserve_pages += max;
5255 * Lowmem reserves are not available to
5256 * GFP_HIGHUSER page cache allocations and
5257 * kswapd tries to balance zones to their high
5258 * watermark. As a result, neither should be
5259 * regarded as dirtyable memory, to prevent a
5260 * situation where reclaim has to clean pages
5261 * in order to balance the zones.
5263 zone->dirty_balance_reserve = max;
5266 dirty_balance_reserve = reserve_pages;
5267 totalreserve_pages = reserve_pages;
5271 * setup_per_zone_lowmem_reserve - called whenever
5272 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5273 * has a correct pages reserved value, so an adequate number of
5274 * pages are left in the zone after a successful __alloc_pages().
5276 static void setup_per_zone_lowmem_reserve(void)
5278 struct pglist_data *pgdat;
5279 enum zone_type j, idx;
5281 for_each_online_pgdat(pgdat) {
5282 for (j = 0; j < MAX_NR_ZONES; j++) {
5283 struct zone *zone = pgdat->node_zones + j;
5284 unsigned long managed_pages = zone->managed_pages;
5286 zone->lowmem_reserve[j] = 0;
5290 struct zone *lower_zone;
5294 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5295 sysctl_lowmem_reserve_ratio[idx] = 1;
5297 lower_zone = pgdat->node_zones + idx;
5298 lower_zone->lowmem_reserve[j] = managed_pages /
5299 sysctl_lowmem_reserve_ratio[idx];
5300 managed_pages += lower_zone->managed_pages;
5305 /* update totalreserve_pages */
5306 calculate_totalreserve_pages();
5309 static void __setup_per_zone_wmarks(void)
5311 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5312 unsigned long lowmem_pages = 0;
5314 unsigned long flags;
5316 /* Calculate total number of !ZONE_HIGHMEM pages */
5317 for_each_zone(zone) {
5318 if (!is_highmem(zone))
5319 lowmem_pages += zone->managed_pages;
5322 for_each_zone(zone) {
5325 spin_lock_irqsave(&zone->lock, flags);
5326 tmp = (u64)pages_min * zone->managed_pages;
5327 do_div(tmp, lowmem_pages);
5328 if (is_highmem(zone)) {
5330 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5331 * need highmem pages, so cap pages_min to a small
5334 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5335 * deltas controls asynch page reclaim, and so should
5336 * not be capped for highmem.
5338 unsigned long min_pages;
5340 min_pages = zone->managed_pages / 1024;
5341 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5342 zone->watermark[WMARK_MIN] = min_pages;
5345 * If it's a lowmem zone, reserve a number of pages
5346 * proportionate to the zone's size.
5348 zone->watermark[WMARK_MIN] = tmp;
5351 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5352 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5354 setup_zone_migrate_reserve(zone);
5355 spin_unlock_irqrestore(&zone->lock, flags);
5358 /* update totalreserve_pages */
5359 calculate_totalreserve_pages();
5363 * setup_per_zone_wmarks - called when min_free_kbytes changes
5364 * or when memory is hot-{added|removed}
5366 * Ensures that the watermark[min,low,high] values for each zone are set
5367 * correctly with respect to min_free_kbytes.
5369 void setup_per_zone_wmarks(void)
5371 mutex_lock(&zonelists_mutex);
5372 __setup_per_zone_wmarks();
5373 mutex_unlock(&zonelists_mutex);
5377 * The inactive anon list should be small enough that the VM never has to
5378 * do too much work, but large enough that each inactive page has a chance
5379 * to be referenced again before it is swapped out.
5381 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5382 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5383 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5384 * the anonymous pages are kept on the inactive list.
5387 * memory ratio inactive anon
5388 * -------------------------------------
5397 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5399 unsigned int gb, ratio;
5401 /* Zone size in gigabytes */
5402 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5404 ratio = int_sqrt(10 * gb);
5408 zone->inactive_ratio = ratio;
5411 static void __meminit setup_per_zone_inactive_ratio(void)
5416 calculate_zone_inactive_ratio(zone);
5420 * Initialise min_free_kbytes.
5422 * For small machines we want it small (128k min). For large machines
5423 * we want it large (64MB max). But it is not linear, because network
5424 * bandwidth does not increase linearly with machine size. We use
5426 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5427 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5443 int __meminit init_per_zone_wmark_min(void)
5445 unsigned long lowmem_kbytes;
5447 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5449 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5450 if (min_free_kbytes < 128)
5451 min_free_kbytes = 128;
5452 if (min_free_kbytes > 65536)
5453 min_free_kbytes = 65536;
5454 setup_per_zone_wmarks();
5455 refresh_zone_stat_thresholds();
5456 setup_per_zone_lowmem_reserve();
5457 setup_per_zone_inactive_ratio();
5460 module_init(init_per_zone_wmark_min)
5463 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5464 * that we can call two helper functions whenever min_free_kbytes
5467 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5468 void __user *buffer, size_t *length, loff_t *ppos)
5470 proc_dointvec(table, write, buffer, length, ppos);
5472 setup_per_zone_wmarks();
5477 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5478 void __user *buffer, size_t *length, loff_t *ppos)
5483 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5488 zone->min_unmapped_pages = (zone->managed_pages *
5489 sysctl_min_unmapped_ratio) / 100;
5493 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5494 void __user *buffer, size_t *length, loff_t *ppos)
5499 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5504 zone->min_slab_pages = (zone->managed_pages *
5505 sysctl_min_slab_ratio) / 100;
5511 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5512 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5513 * whenever sysctl_lowmem_reserve_ratio changes.
5515 * The reserve ratio obviously has absolutely no relation with the
5516 * minimum watermarks. The lowmem reserve ratio can only make sense
5517 * if in function of the boot time zone sizes.
5519 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5520 void __user *buffer, size_t *length, loff_t *ppos)
5522 proc_dointvec_minmax(table, write, buffer, length, ppos);
5523 setup_per_zone_lowmem_reserve();
5528 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5529 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5530 * can have before it gets flushed back to buddy allocator.
5533 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5534 void __user *buffer, size_t *length, loff_t *ppos)
5540 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5541 if (!write || (ret < 0))
5543 for_each_populated_zone(zone) {
5544 for_each_possible_cpu(cpu) {
5546 high = zone->managed_pages / percpu_pagelist_fraction;
5547 setup_pagelist_highmark(
5548 per_cpu_ptr(zone->pageset, cpu), high);
5554 int hashdist = HASHDIST_DEFAULT;
5557 static int __init set_hashdist(char *str)
5561 hashdist = simple_strtoul(str, &str, 0);
5564 __setup("hashdist=", set_hashdist);
5568 * allocate a large system hash table from bootmem
5569 * - it is assumed that the hash table must contain an exact power-of-2
5570 * quantity of entries
5571 * - limit is the number of hash buckets, not the total allocation size
5573 void *__init alloc_large_system_hash(const char *tablename,
5574 unsigned long bucketsize,
5575 unsigned long numentries,
5578 unsigned int *_hash_shift,
5579 unsigned int *_hash_mask,
5580 unsigned long low_limit,
5581 unsigned long high_limit)
5583 unsigned long long max = high_limit;
5584 unsigned long log2qty, size;
5587 /* allow the kernel cmdline to have a say */
5589 /* round applicable memory size up to nearest megabyte */
5590 numentries = nr_kernel_pages;
5591 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5592 numentries >>= 20 - PAGE_SHIFT;
5593 numentries <<= 20 - PAGE_SHIFT;
5595 /* limit to 1 bucket per 2^scale bytes of low memory */
5596 if (scale > PAGE_SHIFT)
5597 numentries >>= (scale - PAGE_SHIFT);
5599 numentries <<= (PAGE_SHIFT - scale);
5601 /* Make sure we've got at least a 0-order allocation.. */
5602 if (unlikely(flags & HASH_SMALL)) {
5603 /* Makes no sense without HASH_EARLY */
5604 WARN_ON(!(flags & HASH_EARLY));
5605 if (!(numentries >> *_hash_shift)) {
5606 numentries = 1UL << *_hash_shift;
5607 BUG_ON(!numentries);
5609 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5610 numentries = PAGE_SIZE / bucketsize;
5612 numentries = roundup_pow_of_two(numentries);
5614 /* limit allocation size to 1/16 total memory by default */
5616 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5617 do_div(max, bucketsize);
5619 max = min(max, 0x80000000ULL);
5621 if (numentries < low_limit)
5622 numentries = low_limit;
5623 if (numentries > max)
5626 log2qty = ilog2(numentries);
5629 size = bucketsize << log2qty;
5630 if (flags & HASH_EARLY)
5631 table = alloc_bootmem_nopanic(size);
5633 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5636 * If bucketsize is not a power-of-two, we may free
5637 * some pages at the end of hash table which
5638 * alloc_pages_exact() automatically does
5640 if (get_order(size) < MAX_ORDER) {
5641 table = alloc_pages_exact(size, GFP_ATOMIC);
5642 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5645 } while (!table && size > PAGE_SIZE && --log2qty);
5648 panic("Failed to allocate %s hash table\n", tablename);
5650 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5653 ilog2(size) - PAGE_SHIFT,
5657 *_hash_shift = log2qty;
5659 *_hash_mask = (1 << log2qty) - 1;
5664 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5665 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5668 #ifdef CONFIG_SPARSEMEM
5669 return __pfn_to_section(pfn)->pageblock_flags;
5671 return zone->pageblock_flags;
5672 #endif /* CONFIG_SPARSEMEM */
5675 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5677 #ifdef CONFIG_SPARSEMEM
5678 pfn &= (PAGES_PER_SECTION-1);
5679 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5681 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5682 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5683 #endif /* CONFIG_SPARSEMEM */
5687 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5688 * @page: The page within the block of interest
5689 * @start_bitidx: The first bit of interest to retrieve
5690 * @end_bitidx: The last bit of interest
5691 * returns pageblock_bits flags
5693 unsigned long get_pageblock_flags_group(struct page *page,
5694 int start_bitidx, int end_bitidx)
5697 unsigned long *bitmap;
5698 unsigned long pfn, bitidx;
5699 unsigned long flags = 0;
5700 unsigned long value = 1;
5702 zone = page_zone(page);
5703 pfn = page_to_pfn(page);
5704 bitmap = get_pageblock_bitmap(zone, pfn);
5705 bitidx = pfn_to_bitidx(zone, pfn);
5707 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5708 if (test_bit(bitidx + start_bitidx, bitmap))
5715 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5716 * @page: The page within the block of interest
5717 * @start_bitidx: The first bit of interest
5718 * @end_bitidx: The last bit of interest
5719 * @flags: The flags to set
5721 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5722 int start_bitidx, int end_bitidx)
5725 unsigned long *bitmap;
5726 unsigned long pfn, bitidx;
5727 unsigned long value = 1;
5729 zone = page_zone(page);
5730 pfn = page_to_pfn(page);
5731 bitmap = get_pageblock_bitmap(zone, pfn);
5732 bitidx = pfn_to_bitidx(zone, pfn);
5733 VM_BUG_ON(!zone_spans_pfn(zone, pfn));
5735 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5737 __set_bit(bitidx + start_bitidx, bitmap);
5739 __clear_bit(bitidx + start_bitidx, bitmap);
5743 * This function checks whether pageblock includes unmovable pages or not.
5744 * If @count is not zero, it is okay to include less @count unmovable pages
5746 * PageLRU check wihtout isolation or lru_lock could race so that
5747 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5748 * expect this function should be exact.
5750 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5751 bool skip_hwpoisoned_pages)
5753 unsigned long pfn, iter, found;
5757 * For avoiding noise data, lru_add_drain_all() should be called
5758 * If ZONE_MOVABLE, the zone never contains unmovable pages
5760 if (zone_idx(zone) == ZONE_MOVABLE)
5762 mt = get_pageblock_migratetype(page);
5763 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5766 pfn = page_to_pfn(page);
5767 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5768 unsigned long check = pfn + iter;
5770 if (!pfn_valid_within(check))
5773 page = pfn_to_page(check);
5775 * We can't use page_count without pin a page
5776 * because another CPU can free compound page.
5777 * This check already skips compound tails of THP
5778 * because their page->_count is zero at all time.
5780 if (!atomic_read(&page->_count)) {
5781 if (PageBuddy(page))
5782 iter += (1 << page_order(page)) - 1;
5787 * The HWPoisoned page may be not in buddy system, and
5788 * page_count() is not 0.
5790 if (skip_hwpoisoned_pages && PageHWPoison(page))
5796 * If there are RECLAIMABLE pages, we need to check it.
5797 * But now, memory offline itself doesn't call shrink_slab()
5798 * and it still to be fixed.
5801 * If the page is not RAM, page_count()should be 0.
5802 * we don't need more check. This is an _used_ not-movable page.
5804 * The problematic thing here is PG_reserved pages. PG_reserved
5805 * is set to both of a memory hole page and a _used_ kernel
5814 bool is_pageblock_removable_nolock(struct page *page)
5820 * We have to be careful here because we are iterating over memory
5821 * sections which are not zone aware so we might end up outside of
5822 * the zone but still within the section.
5823 * We have to take care about the node as well. If the node is offline
5824 * its NODE_DATA will be NULL - see page_zone.
5826 if (!node_online(page_to_nid(page)))
5829 zone = page_zone(page);
5830 pfn = page_to_pfn(page);
5831 if (!zone_spans_pfn(zone, pfn))
5834 return !has_unmovable_pages(zone, page, 0, true);
5839 static unsigned long pfn_max_align_down(unsigned long pfn)
5841 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5842 pageblock_nr_pages) - 1);
5845 static unsigned long pfn_max_align_up(unsigned long pfn)
5847 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5848 pageblock_nr_pages));
5851 /* [start, end) must belong to a single zone. */
5852 static int __alloc_contig_migrate_range(struct compact_control *cc,
5853 unsigned long start, unsigned long end)
5855 /* This function is based on compact_zone() from compaction.c. */
5856 unsigned long nr_reclaimed;
5857 unsigned long pfn = start;
5858 unsigned int tries = 0;
5863 while (pfn < end || !list_empty(&cc->migratepages)) {
5864 if (fatal_signal_pending(current)) {
5869 if (list_empty(&cc->migratepages)) {
5870 cc->nr_migratepages = 0;
5871 pfn = isolate_migratepages_range(cc->zone, cc,
5878 } else if (++tries == 5) {
5879 ret = ret < 0 ? ret : -EBUSY;
5883 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5885 cc->nr_migratepages -= nr_reclaimed;
5887 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
5888 0, MIGRATE_SYNC, MR_CMA);
5891 putback_movable_pages(&cc->migratepages);
5898 * alloc_contig_range() -- tries to allocate given range of pages
5899 * @start: start PFN to allocate
5900 * @end: one-past-the-last PFN to allocate
5901 * @migratetype: migratetype of the underlaying pageblocks (either
5902 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5903 * in range must have the same migratetype and it must
5904 * be either of the two.
5906 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5907 * aligned, however it's the caller's responsibility to guarantee that
5908 * we are the only thread that changes migrate type of pageblocks the
5911 * The PFN range must belong to a single zone.
5913 * Returns zero on success or negative error code. On success all
5914 * pages which PFN is in [start, end) are allocated for the caller and
5915 * need to be freed with free_contig_range().
5917 int alloc_contig_range(unsigned long start, unsigned long end,
5918 unsigned migratetype)
5920 unsigned long outer_start, outer_end;
5923 struct compact_control cc = {
5924 .nr_migratepages = 0,
5926 .zone = page_zone(pfn_to_page(start)),
5928 .ignore_skip_hint = true,
5930 INIT_LIST_HEAD(&cc.migratepages);
5933 * What we do here is we mark all pageblocks in range as
5934 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5935 * have different sizes, and due to the way page allocator
5936 * work, we align the range to biggest of the two pages so
5937 * that page allocator won't try to merge buddies from
5938 * different pageblocks and change MIGRATE_ISOLATE to some
5939 * other migration type.
5941 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5942 * migrate the pages from an unaligned range (ie. pages that
5943 * we are interested in). This will put all the pages in
5944 * range back to page allocator as MIGRATE_ISOLATE.
5946 * When this is done, we take the pages in range from page
5947 * allocator removing them from the buddy system. This way
5948 * page allocator will never consider using them.
5950 * This lets us mark the pageblocks back as
5951 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5952 * aligned range but not in the unaligned, original range are
5953 * put back to page allocator so that buddy can use them.
5956 ret = start_isolate_page_range(pfn_max_align_down(start),
5957 pfn_max_align_up(end), migratetype,
5962 ret = __alloc_contig_migrate_range(&cc, start, end);
5967 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5968 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5969 * more, all pages in [start, end) are free in page allocator.
5970 * What we are going to do is to allocate all pages from
5971 * [start, end) (that is remove them from page allocator).
5973 * The only problem is that pages at the beginning and at the
5974 * end of interesting range may be not aligned with pages that
5975 * page allocator holds, ie. they can be part of higher order
5976 * pages. Because of this, we reserve the bigger range and
5977 * once this is done free the pages we are not interested in.
5979 * We don't have to hold zone->lock here because the pages are
5980 * isolated thus they won't get removed from buddy.
5983 lru_add_drain_all();
5987 outer_start = start;
5988 while (!PageBuddy(pfn_to_page(outer_start))) {
5989 if (++order >= MAX_ORDER) {
5993 outer_start &= ~0UL << order;
5996 /* Make sure the range is really isolated. */
5997 if (test_pages_isolated(outer_start, end, false)) {
5998 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6005 /* Grab isolated pages from freelists. */
6006 outer_end = isolate_freepages_range(&cc, outer_start, end);
6012 /* Free head and tail (if any) */
6013 if (start != outer_start)
6014 free_contig_range(outer_start, start - outer_start);
6015 if (end != outer_end)
6016 free_contig_range(end, outer_end - end);
6019 undo_isolate_page_range(pfn_max_align_down(start),
6020 pfn_max_align_up(end), migratetype);
6024 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6026 unsigned int count = 0;
6028 for (; nr_pages--; pfn++) {
6029 struct page *page = pfn_to_page(pfn);
6031 count += page_count(page) != 1;
6034 WARN(count != 0, "%d pages are still in use!\n", count);
6038 #ifdef CONFIG_MEMORY_HOTPLUG
6039 static int __meminit __zone_pcp_update(void *data)
6041 struct zone *zone = data;
6043 unsigned long batch = zone_batchsize(zone), flags;
6045 for_each_possible_cpu(cpu) {
6046 struct per_cpu_pageset *pset;
6047 struct per_cpu_pages *pcp;
6049 pset = per_cpu_ptr(zone->pageset, cpu);
6052 local_irq_save(flags);
6054 free_pcppages_bulk(zone, pcp->count, pcp);
6055 drain_zonestat(zone, pset);
6056 setup_pageset(pset, batch);
6057 local_irq_restore(flags);
6062 void __meminit zone_pcp_update(struct zone *zone)
6064 stop_machine(__zone_pcp_update, zone, NULL);
6068 void zone_pcp_reset(struct zone *zone)
6070 unsigned long flags;
6072 struct per_cpu_pageset *pset;
6074 /* avoid races with drain_pages() */
6075 local_irq_save(flags);
6076 if (zone->pageset != &boot_pageset) {
6077 for_each_online_cpu(cpu) {
6078 pset = per_cpu_ptr(zone->pageset, cpu);
6079 drain_zonestat(zone, pset);
6081 free_percpu(zone->pageset);
6082 zone->pageset = &boot_pageset;
6084 local_irq_restore(flags);
6087 #ifdef CONFIG_MEMORY_HOTREMOVE
6089 * All pages in the range must be isolated before calling this.
6092 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6098 unsigned long flags;
6099 /* find the first valid pfn */
6100 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6105 zone = page_zone(pfn_to_page(pfn));
6106 spin_lock_irqsave(&zone->lock, flags);
6108 while (pfn < end_pfn) {
6109 if (!pfn_valid(pfn)) {
6113 page = pfn_to_page(pfn);
6115 * The HWPoisoned page may be not in buddy system, and
6116 * page_count() is not 0.
6118 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6120 SetPageReserved(page);
6124 BUG_ON(page_count(page));
6125 BUG_ON(!PageBuddy(page));
6126 order = page_order(page);
6127 #ifdef CONFIG_DEBUG_VM
6128 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6129 pfn, 1 << order, end_pfn);
6131 list_del(&page->lru);
6132 rmv_page_order(page);
6133 zone->free_area[order].nr_free--;
6134 for (i = 0; i < (1 << order); i++)
6135 SetPageReserved((page+i));
6136 pfn += (1 << order);
6138 spin_unlock_irqrestore(&zone->lock, flags);
6142 #ifdef CONFIG_MEMORY_FAILURE
6143 bool is_free_buddy_page(struct page *page)
6145 struct zone *zone = page_zone(page);
6146 unsigned long pfn = page_to_pfn(page);
6147 unsigned long flags;
6150 spin_lock_irqsave(&zone->lock, flags);
6151 for (order = 0; order < MAX_ORDER; order++) {
6152 struct page *page_head = page - (pfn & ((1 << order) - 1));
6154 if (PageBuddy(page_head) && page_order(page_head) >= order)
6157 spin_unlock_irqrestore(&zone->lock, flags);
6159 return order < MAX_ORDER;
6163 static const struct trace_print_flags pageflag_names[] = {
6164 {1UL << PG_locked, "locked" },
6165 {1UL << PG_error, "error" },
6166 {1UL << PG_referenced, "referenced" },
6167 {1UL << PG_uptodate, "uptodate" },
6168 {1UL << PG_dirty, "dirty" },
6169 {1UL << PG_lru, "lru" },
6170 {1UL << PG_active, "active" },
6171 {1UL << PG_slab, "slab" },
6172 {1UL << PG_owner_priv_1, "owner_priv_1" },
6173 {1UL << PG_arch_1, "arch_1" },
6174 {1UL << PG_reserved, "reserved" },
6175 {1UL << PG_private, "private" },
6176 {1UL << PG_private_2, "private_2" },
6177 {1UL << PG_writeback, "writeback" },
6178 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6179 {1UL << PG_head, "head" },
6180 {1UL << PG_tail, "tail" },
6182 {1UL << PG_compound, "compound" },
6184 {1UL << PG_swapcache, "swapcache" },
6185 {1UL << PG_mappedtodisk, "mappedtodisk" },
6186 {1UL << PG_reclaim, "reclaim" },
6187 {1UL << PG_swapbacked, "swapbacked" },
6188 {1UL << PG_unevictable, "unevictable" },
6190 {1UL << PG_mlocked, "mlocked" },
6192 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6193 {1UL << PG_uncached, "uncached" },
6195 #ifdef CONFIG_MEMORY_FAILURE
6196 {1UL << PG_hwpoison, "hwpoison" },
6198 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6199 {1UL << PG_compound_lock, "compound_lock" },
6203 static void dump_page_flags(unsigned long flags)
6205 const char *delim = "";
6209 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6211 printk(KERN_ALERT "page flags: %#lx(", flags);
6213 /* remove zone id */
6214 flags &= (1UL << NR_PAGEFLAGS) - 1;
6216 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6218 mask = pageflag_names[i].mask;
6219 if ((flags & mask) != mask)
6223 printk("%s%s", delim, pageflag_names[i].name);
6227 /* check for left over flags */
6229 printk("%s%#lx", delim, flags);
6234 void dump_page(struct page *page)
6237 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6238 page, atomic_read(&page->_count), page_mapcount(page),
6239 page->mapping, page->index);
6240 dump_page_flags(page->flags);
6241 mem_cgroup_print_bad_page(page);