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;
363 set_page_count(p, 0);
364 p->first_page = page;
365 /* Make sure p->first_page is always valid for PageTail() */
371 /* update __split_huge_page_refcount if you change this function */
372 static int destroy_compound_page(struct page *page, unsigned long order)
375 int nr_pages = 1 << order;
378 if (unlikely(compound_order(page) != order)) {
383 __ClearPageHead(page);
385 for (i = 1; i < nr_pages; i++) {
386 struct page *p = page + i;
388 if (unlikely(!PageTail(p) || (p->first_page != page))) {
398 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
403 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
404 * and __GFP_HIGHMEM from hard or soft interrupt context.
406 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
407 for (i = 0; i < (1 << order); i++)
408 clear_highpage(page + i);
411 #ifdef CONFIG_DEBUG_PAGEALLOC
412 unsigned int _debug_guardpage_minorder;
414 static int __init debug_guardpage_minorder_setup(char *buf)
418 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
419 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
422 _debug_guardpage_minorder = res;
423 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
426 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
428 static inline void set_page_guard_flag(struct page *page)
430 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
433 static inline void clear_page_guard_flag(struct page *page)
435 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
438 static inline void set_page_guard_flag(struct page *page) { }
439 static inline void clear_page_guard_flag(struct page *page) { }
442 static inline void set_page_order(struct page *page, int order)
444 set_page_private(page, order);
445 __SetPageBuddy(page);
448 static inline void rmv_page_order(struct page *page)
450 __ClearPageBuddy(page);
451 set_page_private(page, 0);
455 * Locate the struct page for both the matching buddy in our
456 * pair (buddy1) and the combined O(n+1) page they form (page).
458 * 1) Any buddy B1 will have an order O twin B2 which satisfies
459 * the following equation:
461 * For example, if the starting buddy (buddy2) is #8 its order
463 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
465 * 2) Any buddy B will have an order O+1 parent P which
466 * satisfies the following equation:
469 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
471 static inline unsigned long
472 __find_buddy_index(unsigned long page_idx, unsigned int order)
474 return page_idx ^ (1 << order);
478 * This function checks whether a page is free && is the buddy
479 * we can do coalesce a page and its buddy if
480 * (a) the buddy is not in a hole &&
481 * (b) the buddy is in the buddy system &&
482 * (c) a page and its buddy have the same order &&
483 * (d) a page and its buddy are in the same zone.
485 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
486 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
488 * For recording page's order, we use page_private(page).
490 static inline int page_is_buddy(struct page *page, struct page *buddy,
493 if (!pfn_valid_within(page_to_pfn(buddy)))
496 if (page_zone_id(page) != page_zone_id(buddy))
499 if (page_is_guard(buddy) && page_order(buddy) == order) {
500 VM_BUG_ON(page_count(buddy) != 0);
504 if (PageBuddy(buddy) && page_order(buddy) == order) {
505 VM_BUG_ON(page_count(buddy) != 0);
512 * Freeing function for a buddy system allocator.
514 * The concept of a buddy system is to maintain direct-mapped table
515 * (containing bit values) for memory blocks of various "orders".
516 * The bottom level table contains the map for the smallest allocatable
517 * units of memory (here, pages), and each level above it describes
518 * pairs of units from the levels below, hence, "buddies".
519 * At a high level, all that happens here is marking the table entry
520 * at the bottom level available, and propagating the changes upward
521 * as necessary, plus some accounting needed to play nicely with other
522 * parts of the VM system.
523 * At each level, we keep a list of pages, which are heads of continuous
524 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
525 * order is recorded in page_private(page) field.
526 * So when we are allocating or freeing one, we can derive the state of the
527 * other. That is, if we allocate a small block, and both were
528 * free, the remainder of the region must be split into blocks.
529 * If a block is freed, and its buddy is also free, then this
530 * triggers coalescing into a block of larger size.
535 static inline void __free_one_page(struct page *page,
536 struct zone *zone, unsigned int order,
539 unsigned long page_idx;
540 unsigned long combined_idx;
541 unsigned long uninitialized_var(buddy_idx);
544 VM_BUG_ON(!zone_is_initialized(zone));
546 if (unlikely(PageCompound(page)))
547 if (unlikely(destroy_compound_page(page, order)))
550 VM_BUG_ON(migratetype == -1);
552 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
554 VM_BUG_ON(page_idx & ((1 << order) - 1));
555 VM_BUG_ON(bad_range(zone, page));
557 while (order < MAX_ORDER-1) {
558 buddy_idx = __find_buddy_index(page_idx, order);
559 buddy = page + (buddy_idx - page_idx);
560 if (!page_is_buddy(page, buddy, order))
563 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
564 * merge with it and move up one order.
566 if (page_is_guard(buddy)) {
567 clear_page_guard_flag(buddy);
568 set_page_private(page, 0);
569 __mod_zone_freepage_state(zone, 1 << order,
572 list_del(&buddy->lru);
573 zone->free_area[order].nr_free--;
574 rmv_page_order(buddy);
576 combined_idx = buddy_idx & page_idx;
577 page = page + (combined_idx - page_idx);
578 page_idx = combined_idx;
581 set_page_order(page, order);
584 * If this is not the largest possible page, check if the buddy
585 * of the next-highest order is free. If it is, it's possible
586 * that pages are being freed that will coalesce soon. In case,
587 * that is happening, add the free page to the tail of the list
588 * so it's less likely to be used soon and more likely to be merged
589 * as a higher order page
591 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
592 struct page *higher_page, *higher_buddy;
593 combined_idx = buddy_idx & page_idx;
594 higher_page = page + (combined_idx - page_idx);
595 buddy_idx = __find_buddy_index(combined_idx, order + 1);
596 higher_buddy = higher_page + (buddy_idx - combined_idx);
597 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
598 list_add_tail(&page->lru,
599 &zone->free_area[order].free_list[migratetype]);
604 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
606 zone->free_area[order].nr_free++;
609 static inline int free_pages_check(struct page *page)
611 if (unlikely(page_mapcount(page) |
612 (page->mapping != NULL) |
613 (atomic_read(&page->_count) != 0) |
614 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
615 (mem_cgroup_bad_page_check(page)))) {
619 page_nid_reset_last(page);
620 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
621 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
626 * Frees a number of pages from the PCP lists
627 * Assumes all pages on list are in same zone, and of same order.
628 * count is the number of pages to free.
630 * If the zone was previously in an "all pages pinned" state then look to
631 * see if this freeing clears that state.
633 * And clear the zone's pages_scanned counter, to hold off the "all pages are
634 * pinned" detection logic.
636 static void free_pcppages_bulk(struct zone *zone, int count,
637 struct per_cpu_pages *pcp)
643 spin_lock(&zone->lock);
644 zone->all_unreclaimable = 0;
645 zone->pages_scanned = 0;
649 struct list_head *list;
652 * Remove pages from lists in a round-robin fashion. A
653 * batch_free count is maintained that is incremented when an
654 * empty list is encountered. This is so more pages are freed
655 * off fuller lists instead of spinning excessively around empty
660 if (++migratetype == MIGRATE_PCPTYPES)
662 list = &pcp->lists[migratetype];
663 } while (list_empty(list));
665 /* This is the only non-empty list. Free them all. */
666 if (batch_free == MIGRATE_PCPTYPES)
667 batch_free = to_free;
670 int mt; /* migratetype of the to-be-freed page */
672 page = list_entry(list->prev, struct page, lru);
673 /* must delete as __free_one_page list manipulates */
674 list_del(&page->lru);
675 mt = get_freepage_migratetype(page);
676 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
677 __free_one_page(page, zone, 0, mt);
678 trace_mm_page_pcpu_drain(page, 0, mt);
679 if (likely(!is_migrate_isolate_page(page))) {
680 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
681 if (is_migrate_cma(mt))
682 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
684 } while (--to_free && --batch_free && !list_empty(list));
686 spin_unlock(&zone->lock);
689 static void free_one_page(struct zone *zone, struct page *page, int order,
692 spin_lock(&zone->lock);
693 zone->all_unreclaimable = 0;
694 zone->pages_scanned = 0;
696 __free_one_page(page, zone, order, migratetype);
697 if (unlikely(!is_migrate_isolate(migratetype)))
698 __mod_zone_freepage_state(zone, 1 << order, migratetype);
699 spin_unlock(&zone->lock);
702 static bool free_pages_prepare(struct page *page, unsigned int order)
707 trace_mm_page_free(page, order);
708 kmemcheck_free_shadow(page, order);
711 page->mapping = NULL;
712 for (i = 0; i < (1 << order); i++)
713 bad += free_pages_check(page + i);
717 if (!PageHighMem(page)) {
718 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
719 debug_check_no_obj_freed(page_address(page),
722 arch_free_page(page, order);
723 kernel_map_pages(page, 1 << order, 0);
728 static void __free_pages_ok(struct page *page, unsigned int order)
733 if (!free_pages_prepare(page, order))
736 local_irq_save(flags);
737 __count_vm_events(PGFREE, 1 << order);
738 migratetype = get_pageblock_migratetype(page);
739 set_freepage_migratetype(page, migratetype);
740 free_one_page(page_zone(page), page, order, migratetype);
741 local_irq_restore(flags);
745 * Read access to zone->managed_pages is safe because it's unsigned long,
746 * but we still need to serialize writers. Currently all callers of
747 * __free_pages_bootmem() except put_page_bootmem() should only be used
748 * at boot time. So for shorter boot time, we shift the burden to
749 * put_page_bootmem() to serialize writers.
751 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
753 unsigned int nr_pages = 1 << order;
757 for (loop = 0; loop < nr_pages; loop++) {
758 struct page *p = &page[loop];
760 if (loop + 1 < nr_pages)
762 __ClearPageReserved(p);
763 set_page_count(p, 0);
766 page_zone(page)->managed_pages += 1 << order;
767 set_page_refcounted(page);
768 __free_pages(page, order);
772 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
773 void __init init_cma_reserved_pageblock(struct page *page)
775 unsigned i = pageblock_nr_pages;
776 struct page *p = page;
779 __ClearPageReserved(p);
780 set_page_count(p, 0);
783 set_page_refcounted(page);
784 set_pageblock_migratetype(page, MIGRATE_CMA);
785 __free_pages(page, pageblock_order);
786 totalram_pages += pageblock_nr_pages;
787 #ifdef CONFIG_HIGHMEM
788 if (PageHighMem(page))
789 totalhigh_pages += pageblock_nr_pages;
795 * The order of subdivision here is critical for the IO subsystem.
796 * Please do not alter this order without good reasons and regression
797 * testing. Specifically, as large blocks of memory are subdivided,
798 * the order in which smaller blocks are delivered depends on the order
799 * they're subdivided in this function. This is the primary factor
800 * influencing the order in which pages are delivered to the IO
801 * subsystem according to empirical testing, and this is also justified
802 * by considering the behavior of a buddy system containing a single
803 * large block of memory acted on by a series of small allocations.
804 * This behavior is a critical factor in sglist merging's success.
808 static inline void expand(struct zone *zone, struct page *page,
809 int low, int high, struct free_area *area,
812 unsigned long size = 1 << high;
818 VM_BUG_ON(bad_range(zone, &page[size]));
820 #ifdef CONFIG_DEBUG_PAGEALLOC
821 if (high < debug_guardpage_minorder()) {
823 * Mark as guard pages (or page), that will allow to
824 * merge back to allocator when buddy will be freed.
825 * Corresponding page table entries will not be touched,
826 * pages will stay not present in virtual address space
828 INIT_LIST_HEAD(&page[size].lru);
829 set_page_guard_flag(&page[size]);
830 set_page_private(&page[size], high);
831 /* Guard pages are not available for any usage */
832 __mod_zone_freepage_state(zone, -(1 << high),
837 list_add(&page[size].lru, &area->free_list[migratetype]);
839 set_page_order(&page[size], high);
844 * This page is about to be returned from the page allocator
846 static inline int check_new_page(struct page *page)
848 if (unlikely(page_mapcount(page) |
849 (page->mapping != NULL) |
850 (atomic_read(&page->_count) != 0) |
851 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
852 (mem_cgroup_bad_page_check(page)))) {
859 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
863 for (i = 0; i < (1 << order); i++) {
864 struct page *p = page + i;
865 if (unlikely(check_new_page(p)))
869 set_page_private(page, 0);
870 set_page_refcounted(page);
872 arch_alloc_page(page, order);
873 kernel_map_pages(page, 1 << order, 1);
875 if (gfp_flags & __GFP_ZERO)
876 prep_zero_page(page, order, gfp_flags);
878 if (order && (gfp_flags & __GFP_COMP))
879 prep_compound_page(page, order);
885 * Go through the free lists for the given migratetype and remove
886 * the smallest available page from the freelists
889 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
892 unsigned int current_order;
893 struct free_area * area;
896 /* Find a page of the appropriate size in the preferred list */
897 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
898 area = &(zone->free_area[current_order]);
899 if (list_empty(&area->free_list[migratetype]))
902 page = list_entry(area->free_list[migratetype].next,
904 list_del(&page->lru);
905 rmv_page_order(page);
907 expand(zone, page, order, current_order, area, migratetype);
916 * This array describes the order lists are fallen back to when
917 * the free lists for the desirable migrate type are depleted
919 static int fallbacks[MIGRATE_TYPES][4] = {
920 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
921 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
923 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
924 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
926 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
928 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
929 #ifdef CONFIG_MEMORY_ISOLATION
930 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
935 * Move the free pages in a range to the free lists of the requested type.
936 * Note that start_page and end_pages are not aligned on a pageblock
937 * boundary. If alignment is required, use move_freepages_block()
939 int move_freepages(struct zone *zone,
940 struct page *start_page, struct page *end_page,
947 #ifndef CONFIG_HOLES_IN_ZONE
949 * page_zone is not safe to call in this context when
950 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
951 * anyway as we check zone boundaries in move_freepages_block().
952 * Remove at a later date when no bug reports exist related to
953 * grouping pages by mobility
955 BUG_ON(page_zone(start_page) != page_zone(end_page));
958 for (page = start_page; page <= end_page;) {
959 /* Make sure we are not inadvertently changing nodes */
960 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
962 if (!pfn_valid_within(page_to_pfn(page))) {
967 if (!PageBuddy(page)) {
972 order = page_order(page);
973 list_move(&page->lru,
974 &zone->free_area[order].free_list[migratetype]);
975 set_freepage_migratetype(page, migratetype);
977 pages_moved += 1 << order;
983 int move_freepages_block(struct zone *zone, struct page *page,
986 unsigned long start_pfn, end_pfn;
987 struct page *start_page, *end_page;
989 start_pfn = page_to_pfn(page);
990 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
991 start_page = pfn_to_page(start_pfn);
992 end_page = start_page + pageblock_nr_pages - 1;
993 end_pfn = start_pfn + pageblock_nr_pages - 1;
995 /* Do not cross zone boundaries */
996 if (!zone_spans_pfn(zone, start_pfn))
998 if (!zone_spans_pfn(zone, end_pfn))
1001 return move_freepages(zone, start_page, end_page, migratetype);
1004 static void change_pageblock_range(struct page *pageblock_page,
1005 int start_order, int migratetype)
1007 int nr_pageblocks = 1 << (start_order - pageblock_order);
1009 while (nr_pageblocks--) {
1010 set_pageblock_migratetype(pageblock_page, migratetype);
1011 pageblock_page += pageblock_nr_pages;
1015 /* Remove an element from the buddy allocator from the fallback list */
1016 static inline struct page *
1017 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1019 struct free_area * area;
1024 /* Find the largest possible block of pages in the other list */
1025 for (current_order = MAX_ORDER-1; current_order >= order;
1028 migratetype = fallbacks[start_migratetype][i];
1030 /* MIGRATE_RESERVE handled later if necessary */
1031 if (migratetype == MIGRATE_RESERVE)
1034 area = &(zone->free_area[current_order]);
1035 if (list_empty(&area->free_list[migratetype]))
1038 page = list_entry(area->free_list[migratetype].next,
1043 * If breaking a large block of pages, move all free
1044 * pages to the preferred allocation list. If falling
1045 * back for a reclaimable kernel allocation, be more
1046 * aggressive about taking ownership of free pages
1048 * On the other hand, never change migration
1049 * type of MIGRATE_CMA pageblocks nor move CMA
1050 * pages on different free lists. We don't
1051 * want unmovable pages to be allocated from
1052 * MIGRATE_CMA areas.
1054 if (!is_migrate_cma(migratetype) &&
1055 (unlikely(current_order >= pageblock_order / 2) ||
1056 start_migratetype == MIGRATE_RECLAIMABLE ||
1057 page_group_by_mobility_disabled)) {
1059 pages = move_freepages_block(zone, page,
1062 /* Claim the whole block if over half of it is free */
1063 if (pages >= (1 << (pageblock_order-1)) ||
1064 page_group_by_mobility_disabled)
1065 set_pageblock_migratetype(page,
1068 migratetype = start_migratetype;
1071 /* Remove the page from the freelists */
1072 list_del(&page->lru);
1073 rmv_page_order(page);
1075 /* Take ownership for orders >= pageblock_order */
1076 if (current_order >= pageblock_order &&
1077 !is_migrate_cma(migratetype))
1078 change_pageblock_range(page, current_order,
1081 expand(zone, page, order, current_order, area,
1082 is_migrate_cma(migratetype)
1083 ? migratetype : start_migratetype);
1085 trace_mm_page_alloc_extfrag(page, order, current_order,
1086 start_migratetype, migratetype);
1096 * Do the hard work of removing an element from the buddy allocator.
1097 * Call me with the zone->lock already held.
1099 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1105 page = __rmqueue_smallest(zone, order, migratetype);
1107 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1108 page = __rmqueue_fallback(zone, order, migratetype);
1111 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1112 * is used because __rmqueue_smallest is an inline function
1113 * and we want just one call site
1116 migratetype = MIGRATE_RESERVE;
1121 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1126 * Obtain a specified number of elements from the buddy allocator, all under
1127 * a single hold of the lock, for efficiency. Add them to the supplied list.
1128 * Returns the number of new pages which were placed at *list.
1130 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1131 unsigned long count, struct list_head *list,
1132 int migratetype, int cold)
1134 int mt = migratetype, i;
1136 spin_lock(&zone->lock);
1137 for (i = 0; i < count; ++i) {
1138 struct page *page = __rmqueue(zone, order, migratetype);
1139 if (unlikely(page == NULL))
1143 * Split buddy pages returned by expand() are received here
1144 * in physical page order. The page is added to the callers and
1145 * list and the list head then moves forward. From the callers
1146 * perspective, the linked list is ordered by page number in
1147 * some conditions. This is useful for IO devices that can
1148 * merge IO requests if the physical pages are ordered
1151 if (likely(cold == 0))
1152 list_add(&page->lru, list);
1154 list_add_tail(&page->lru, list);
1155 if (IS_ENABLED(CONFIG_CMA)) {
1156 mt = get_pageblock_migratetype(page);
1157 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1160 set_freepage_migratetype(page, mt);
1162 if (is_migrate_cma(mt))
1163 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1166 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1167 spin_unlock(&zone->lock);
1173 * Called from the vmstat counter updater to drain pagesets of this
1174 * currently executing processor on remote nodes after they have
1177 * Note that this function must be called with the thread pinned to
1178 * a single processor.
1180 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1182 unsigned long flags;
1185 local_irq_save(flags);
1186 if (pcp->count >= pcp->batch)
1187 to_drain = pcp->batch;
1189 to_drain = pcp->count;
1191 free_pcppages_bulk(zone, to_drain, pcp);
1192 pcp->count -= to_drain;
1194 local_irq_restore(flags);
1199 * Drain pages of the indicated processor.
1201 * The processor must either be the current processor and the
1202 * thread pinned to the current processor or a processor that
1205 static void drain_pages(unsigned int cpu)
1207 unsigned long flags;
1210 for_each_populated_zone(zone) {
1211 struct per_cpu_pageset *pset;
1212 struct per_cpu_pages *pcp;
1214 local_irq_save(flags);
1215 pset = per_cpu_ptr(zone->pageset, cpu);
1219 free_pcppages_bulk(zone, pcp->count, pcp);
1222 local_irq_restore(flags);
1227 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1229 void drain_local_pages(void *arg)
1231 drain_pages(smp_processor_id());
1235 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1237 * Note that this code is protected against sending an IPI to an offline
1238 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1239 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1240 * nothing keeps CPUs from showing up after we populated the cpumask and
1241 * before the call to on_each_cpu_mask().
1243 void drain_all_pages(void)
1246 struct per_cpu_pageset *pcp;
1250 * Allocate in the BSS so we wont require allocation in
1251 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1253 static cpumask_t cpus_with_pcps;
1256 * We don't care about racing with CPU hotplug event
1257 * as offline notification will cause the notified
1258 * cpu to drain that CPU pcps and on_each_cpu_mask
1259 * disables preemption as part of its processing
1261 for_each_online_cpu(cpu) {
1262 bool has_pcps = false;
1263 for_each_populated_zone(zone) {
1264 pcp = per_cpu_ptr(zone->pageset, cpu);
1265 if (pcp->pcp.count) {
1271 cpumask_set_cpu(cpu, &cpus_with_pcps);
1273 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1275 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1278 #ifdef CONFIG_HIBERNATION
1280 void mark_free_pages(struct zone *zone)
1282 unsigned long pfn, max_zone_pfn;
1283 unsigned long flags;
1285 struct list_head *curr;
1287 if (!zone->spanned_pages)
1290 spin_lock_irqsave(&zone->lock, flags);
1292 max_zone_pfn = zone_end_pfn(zone);
1293 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1294 if (pfn_valid(pfn)) {
1295 struct page *page = pfn_to_page(pfn);
1297 if (!swsusp_page_is_forbidden(page))
1298 swsusp_unset_page_free(page);
1301 for_each_migratetype_order(order, t) {
1302 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1305 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1306 for (i = 0; i < (1UL << order); i++)
1307 swsusp_set_page_free(pfn_to_page(pfn + i));
1310 spin_unlock_irqrestore(&zone->lock, flags);
1312 #endif /* CONFIG_PM */
1315 * Free a 0-order page
1316 * cold == 1 ? free a cold page : free a hot page
1318 void free_hot_cold_page(struct page *page, int cold)
1320 struct zone *zone = page_zone(page);
1321 struct per_cpu_pages *pcp;
1322 unsigned long flags;
1325 if (!free_pages_prepare(page, 0))
1328 migratetype = get_pageblock_migratetype(page);
1329 set_freepage_migratetype(page, migratetype);
1330 local_irq_save(flags);
1331 __count_vm_event(PGFREE);
1334 * We only track unmovable, reclaimable and movable on pcp lists.
1335 * Free ISOLATE pages back to the allocator because they are being
1336 * offlined but treat RESERVE as movable pages so we can get those
1337 * areas back if necessary. Otherwise, we may have to free
1338 * excessively into the page allocator
1340 if (migratetype >= MIGRATE_PCPTYPES) {
1341 if (unlikely(is_migrate_isolate(migratetype))) {
1342 free_one_page(zone, page, 0, migratetype);
1345 migratetype = MIGRATE_MOVABLE;
1348 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1350 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1352 list_add(&page->lru, &pcp->lists[migratetype]);
1354 if (pcp->count >= pcp->high) {
1355 free_pcppages_bulk(zone, pcp->batch, pcp);
1356 pcp->count -= pcp->batch;
1360 local_irq_restore(flags);
1364 * Free a list of 0-order pages
1366 void free_hot_cold_page_list(struct list_head *list, int cold)
1368 struct page *page, *next;
1370 list_for_each_entry_safe(page, next, list, lru) {
1371 trace_mm_page_free_batched(page, cold);
1372 free_hot_cold_page(page, cold);
1377 * split_page takes a non-compound higher-order page, and splits it into
1378 * n (1<<order) sub-pages: page[0..n]
1379 * Each sub-page must be freed individually.
1381 * Note: this is probably too low level an operation for use in drivers.
1382 * Please consult with lkml before using this in your driver.
1384 void split_page(struct page *page, unsigned int order)
1388 VM_BUG_ON(PageCompound(page));
1389 VM_BUG_ON(!page_count(page));
1391 #ifdef CONFIG_KMEMCHECK
1393 * Split shadow pages too, because free(page[0]) would
1394 * otherwise free the whole shadow.
1396 if (kmemcheck_page_is_tracked(page))
1397 split_page(virt_to_page(page[0].shadow), order);
1400 for (i = 1; i < (1 << order); i++)
1401 set_page_refcounted(page + i);
1403 EXPORT_SYMBOL_GPL(split_page);
1405 static int __isolate_free_page(struct page *page, unsigned int order)
1407 unsigned long watermark;
1411 BUG_ON(!PageBuddy(page));
1413 zone = page_zone(page);
1414 mt = get_pageblock_migratetype(page);
1416 if (!is_migrate_isolate(mt)) {
1417 /* Obey watermarks as if the page was being allocated */
1418 watermark = low_wmark_pages(zone) + (1 << order);
1419 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1422 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1425 /* Remove page from free list */
1426 list_del(&page->lru);
1427 zone->free_area[order].nr_free--;
1428 rmv_page_order(page);
1430 /* Set the pageblock if the isolated page is at least a pageblock */
1431 if (order >= pageblock_order - 1) {
1432 struct page *endpage = page + (1 << order) - 1;
1433 for (; page < endpage; page += pageblock_nr_pages) {
1434 int mt = get_pageblock_migratetype(page);
1435 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1436 set_pageblock_migratetype(page,
1441 return 1UL << order;
1445 * Similar to split_page except the page is already free. As this is only
1446 * being used for migration, the migratetype of the block also changes.
1447 * As this is called with interrupts disabled, the caller is responsible
1448 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1451 * Note: this is probably too low level an operation for use in drivers.
1452 * Please consult with lkml before using this in your driver.
1454 int split_free_page(struct page *page)
1459 order = page_order(page);
1461 nr_pages = __isolate_free_page(page, order);
1465 /* Split into individual pages */
1466 set_page_refcounted(page);
1467 split_page(page, order);
1472 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1473 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1477 struct page *buffered_rmqueue(struct zone *preferred_zone,
1478 struct zone *zone, int order, gfp_t gfp_flags,
1481 unsigned long flags;
1483 int cold = !!(gfp_flags & __GFP_COLD);
1486 if (likely(order == 0)) {
1487 struct per_cpu_pages *pcp;
1488 struct list_head *list;
1490 local_irq_save(flags);
1491 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1492 list = &pcp->lists[migratetype];
1493 if (list_empty(list)) {
1494 pcp->count += rmqueue_bulk(zone, 0,
1497 if (unlikely(list_empty(list)))
1502 page = list_entry(list->prev, struct page, lru);
1504 page = list_entry(list->next, struct page, lru);
1506 list_del(&page->lru);
1509 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1511 * __GFP_NOFAIL is not to be used in new code.
1513 * All __GFP_NOFAIL callers should be fixed so that they
1514 * properly detect and handle allocation failures.
1516 * We most definitely don't want callers attempting to
1517 * allocate greater than order-1 page units with
1520 WARN_ON_ONCE(order > 1);
1522 spin_lock_irqsave(&zone->lock, flags);
1523 page = __rmqueue(zone, order, migratetype);
1524 spin_unlock(&zone->lock);
1527 __mod_zone_freepage_state(zone, -(1 << order),
1528 get_pageblock_migratetype(page));
1531 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1532 zone_statistics(preferred_zone, zone, gfp_flags);
1533 local_irq_restore(flags);
1535 VM_BUG_ON(bad_range(zone, page));
1536 if (prep_new_page(page, order, gfp_flags))
1541 local_irq_restore(flags);
1545 #ifdef CONFIG_FAIL_PAGE_ALLOC
1548 struct fault_attr attr;
1550 u32 ignore_gfp_highmem;
1551 u32 ignore_gfp_wait;
1553 } fail_page_alloc = {
1554 .attr = FAULT_ATTR_INITIALIZER,
1555 .ignore_gfp_wait = 1,
1556 .ignore_gfp_highmem = 1,
1560 static int __init setup_fail_page_alloc(char *str)
1562 return setup_fault_attr(&fail_page_alloc.attr, str);
1564 __setup("fail_page_alloc=", setup_fail_page_alloc);
1566 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1568 if (order < fail_page_alloc.min_order)
1570 if (gfp_mask & __GFP_NOFAIL)
1572 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1574 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1577 return should_fail(&fail_page_alloc.attr, 1 << order);
1580 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1582 static int __init fail_page_alloc_debugfs(void)
1584 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1587 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1588 &fail_page_alloc.attr);
1590 return PTR_ERR(dir);
1592 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1593 &fail_page_alloc.ignore_gfp_wait))
1595 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1596 &fail_page_alloc.ignore_gfp_highmem))
1598 if (!debugfs_create_u32("min-order", mode, dir,
1599 &fail_page_alloc.min_order))
1604 debugfs_remove_recursive(dir);
1609 late_initcall(fail_page_alloc_debugfs);
1611 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1613 #else /* CONFIG_FAIL_PAGE_ALLOC */
1615 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1620 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1623 * Return true if free pages are above 'mark'. This takes into account the order
1624 * of the allocation.
1626 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1627 int classzone_idx, int alloc_flags, long free_pages)
1629 /* free_pages my go negative - that's OK */
1631 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1635 free_pages -= (1 << order) - 1;
1636 if (alloc_flags & ALLOC_HIGH)
1638 if (alloc_flags & ALLOC_HARDER)
1641 /* If allocation can't use CMA areas don't use free CMA pages */
1642 if (!(alloc_flags & ALLOC_CMA))
1643 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1646 if (free_pages - free_cma <= min + lowmem_reserve)
1648 for (o = 0; o < order; o++) {
1649 /* At the next order, this order's pages become unavailable */
1650 free_pages -= z->free_area[o].nr_free << o;
1652 /* Require fewer higher order pages to be free */
1655 if (free_pages <= min)
1661 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1662 int classzone_idx, int alloc_flags)
1664 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1665 zone_page_state(z, NR_FREE_PAGES));
1668 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1669 int classzone_idx, int alloc_flags)
1671 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1673 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1674 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1676 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1682 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1683 * skip over zones that are not allowed by the cpuset, or that have
1684 * been recently (in last second) found to be nearly full. See further
1685 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1686 * that have to skip over a lot of full or unallowed zones.
1688 * If the zonelist cache is present in the passed in zonelist, then
1689 * returns a pointer to the allowed node mask (either the current
1690 * tasks mems_allowed, or node_states[N_MEMORY].)
1692 * If the zonelist cache is not available for this zonelist, does
1693 * nothing and returns NULL.
1695 * If the fullzones BITMAP in the zonelist cache is stale (more than
1696 * a second since last zap'd) then we zap it out (clear its bits.)
1698 * We hold off even calling zlc_setup, until after we've checked the
1699 * first zone in the zonelist, on the theory that most allocations will
1700 * be satisfied from that first zone, so best to examine that zone as
1701 * quickly as we can.
1703 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1705 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1706 nodemask_t *allowednodes; /* zonelist_cache approximation */
1708 zlc = zonelist->zlcache_ptr;
1712 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1713 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1714 zlc->last_full_zap = jiffies;
1717 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1718 &cpuset_current_mems_allowed :
1719 &node_states[N_MEMORY];
1720 return allowednodes;
1724 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1725 * if it is worth looking at further for free memory:
1726 * 1) Check that the zone isn't thought to be full (doesn't have its
1727 * bit set in the zonelist_cache fullzones BITMAP).
1728 * 2) Check that the zones node (obtained from the zonelist_cache
1729 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1730 * Return true (non-zero) if zone is worth looking at further, or
1731 * else return false (zero) if it is not.
1733 * This check -ignores- the distinction between various watermarks,
1734 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1735 * found to be full for any variation of these watermarks, it will
1736 * be considered full for up to one second by all requests, unless
1737 * we are so low on memory on all allowed nodes that we are forced
1738 * into the second scan of the zonelist.
1740 * In the second scan we ignore this zonelist cache and exactly
1741 * apply the watermarks to all zones, even it is slower to do so.
1742 * We are low on memory in the second scan, and should leave no stone
1743 * unturned looking for a free page.
1745 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1746 nodemask_t *allowednodes)
1748 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1749 int i; /* index of *z in zonelist zones */
1750 int n; /* node that zone *z is on */
1752 zlc = zonelist->zlcache_ptr;
1756 i = z - zonelist->_zonerefs;
1759 /* This zone is worth trying if it is allowed but not full */
1760 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1764 * Given 'z' scanning a zonelist, set the corresponding bit in
1765 * zlc->fullzones, so that subsequent attempts to allocate a page
1766 * from that zone don't waste time re-examining it.
1768 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1770 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1771 int i; /* index of *z in zonelist zones */
1773 zlc = zonelist->zlcache_ptr;
1777 i = z - zonelist->_zonerefs;
1779 set_bit(i, zlc->fullzones);
1783 * clear all zones full, called after direct reclaim makes progress so that
1784 * a zone that was recently full is not skipped over for up to a second
1786 static void zlc_clear_zones_full(struct zonelist *zonelist)
1788 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1790 zlc = zonelist->zlcache_ptr;
1794 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1797 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1799 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1802 static void __paginginit init_zone_allows_reclaim(int nid)
1806 for_each_online_node(i)
1807 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1808 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1810 zone_reclaim_mode = 1;
1813 #else /* CONFIG_NUMA */
1815 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1820 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1821 nodemask_t *allowednodes)
1826 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1830 static void zlc_clear_zones_full(struct zonelist *zonelist)
1834 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1839 static inline void init_zone_allows_reclaim(int nid)
1842 #endif /* CONFIG_NUMA */
1845 * get_page_from_freelist goes through the zonelist trying to allocate
1848 static struct page *
1849 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1850 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1851 struct zone *preferred_zone, int migratetype)
1854 struct page *page = NULL;
1857 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1858 int zlc_active = 0; /* set if using zonelist_cache */
1859 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1861 classzone_idx = zone_idx(preferred_zone);
1864 * Scan zonelist, looking for a zone with enough free.
1865 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1867 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1868 high_zoneidx, nodemask) {
1869 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1870 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1872 if ((alloc_flags & ALLOC_CPUSET) &&
1873 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1876 * When allocating a page cache page for writing, we
1877 * want to get it from a zone that is within its dirty
1878 * limit, such that no single zone holds more than its
1879 * proportional share of globally allowed dirty pages.
1880 * The dirty limits take into account the zone's
1881 * lowmem reserves and high watermark so that kswapd
1882 * should be able to balance it without having to
1883 * write pages from its LRU list.
1885 * This may look like it could increase pressure on
1886 * lower zones by failing allocations in higher zones
1887 * before they are full. But the pages that do spill
1888 * over are limited as the lower zones are protected
1889 * by this very same mechanism. It should not become
1890 * a practical burden to them.
1892 * XXX: For now, allow allocations to potentially
1893 * exceed the per-zone dirty limit in the slowpath
1894 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1895 * which is important when on a NUMA setup the allowed
1896 * zones are together not big enough to reach the
1897 * global limit. The proper fix for these situations
1898 * will require awareness of zones in the
1899 * dirty-throttling and the flusher threads.
1901 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1902 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1903 goto this_zone_full;
1905 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1906 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1910 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1911 if (zone_watermark_ok(zone, order, mark,
1912 classzone_idx, alloc_flags))
1915 if (IS_ENABLED(CONFIG_NUMA) &&
1916 !did_zlc_setup && nr_online_nodes > 1) {
1918 * we do zlc_setup if there are multiple nodes
1919 * and before considering the first zone allowed
1922 allowednodes = zlc_setup(zonelist, alloc_flags);
1927 if (zone_reclaim_mode == 0 ||
1928 !zone_allows_reclaim(preferred_zone, zone))
1929 goto this_zone_full;
1932 * As we may have just activated ZLC, check if the first
1933 * eligible zone has failed zone_reclaim recently.
1935 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1936 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1939 ret = zone_reclaim(zone, gfp_mask, order);
1941 case ZONE_RECLAIM_NOSCAN:
1944 case ZONE_RECLAIM_FULL:
1945 /* scanned but unreclaimable */
1948 /* did we reclaim enough */
1949 if (zone_watermark_ok(zone, order, mark,
1950 classzone_idx, alloc_flags))
1954 * Failed to reclaim enough to meet watermark.
1955 * Only mark the zone full if checking the min
1956 * watermark or if we failed to reclaim just
1957 * 1<<order pages or else the page allocator
1958 * fastpath will prematurely mark zones full
1959 * when the watermark is between the low and
1962 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
1963 ret == ZONE_RECLAIM_SOME)
1964 goto this_zone_full;
1971 page = buffered_rmqueue(preferred_zone, zone, order,
1972 gfp_mask, migratetype);
1976 if (IS_ENABLED(CONFIG_NUMA))
1977 zlc_mark_zone_full(zonelist, z);
1980 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1981 /* Disable zlc cache for second zonelist scan */
1988 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1989 * necessary to allocate the page. The expectation is
1990 * that the caller is taking steps that will free more
1991 * memory. The caller should avoid the page being used
1992 * for !PFMEMALLOC purposes.
1994 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2000 * Large machines with many possible nodes should not always dump per-node
2001 * meminfo in irq context.
2003 static inline bool should_suppress_show_mem(void)
2008 ret = in_interrupt();
2013 static DEFINE_RATELIMIT_STATE(nopage_rs,
2014 DEFAULT_RATELIMIT_INTERVAL,
2015 DEFAULT_RATELIMIT_BURST);
2017 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2019 unsigned int filter = SHOW_MEM_FILTER_NODES;
2021 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2022 debug_guardpage_minorder() > 0)
2026 * Walking all memory to count page types is very expensive and should
2027 * be inhibited in non-blockable contexts.
2029 if (!(gfp_mask & __GFP_WAIT))
2030 filter |= SHOW_MEM_FILTER_PAGE_COUNT;
2033 * This documents exceptions given to allocations in certain
2034 * contexts that are allowed to allocate outside current's set
2037 if (!(gfp_mask & __GFP_NOMEMALLOC))
2038 if (test_thread_flag(TIF_MEMDIE) ||
2039 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2040 filter &= ~SHOW_MEM_FILTER_NODES;
2041 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2042 filter &= ~SHOW_MEM_FILTER_NODES;
2045 struct va_format vaf;
2048 va_start(args, fmt);
2053 pr_warn("%pV", &vaf);
2058 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2059 current->comm, order, gfp_mask);
2062 if (!should_suppress_show_mem())
2067 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2068 unsigned long did_some_progress,
2069 unsigned long pages_reclaimed)
2071 /* Do not loop if specifically requested */
2072 if (gfp_mask & __GFP_NORETRY)
2075 /* Always retry if specifically requested */
2076 if (gfp_mask & __GFP_NOFAIL)
2080 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2081 * making forward progress without invoking OOM. Suspend also disables
2082 * storage devices so kswapd will not help. Bail if we are suspending.
2084 if (!did_some_progress && pm_suspended_storage())
2088 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2089 * means __GFP_NOFAIL, but that may not be true in other
2092 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2096 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2097 * specified, then we retry until we no longer reclaim any pages
2098 * (above), or we've reclaimed an order of pages at least as
2099 * large as the allocation's order. In both cases, if the
2100 * allocation still fails, we stop retrying.
2102 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2108 static inline struct page *
2109 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2110 struct zonelist *zonelist, enum zone_type high_zoneidx,
2111 nodemask_t *nodemask, struct zone *preferred_zone,
2116 /* Acquire the OOM killer lock for the zones in zonelist */
2117 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2118 schedule_timeout_uninterruptible(1);
2123 * Go through the zonelist yet one more time, keep very high watermark
2124 * here, this is only to catch a parallel oom killing, we must fail if
2125 * we're still under heavy pressure.
2127 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2128 order, zonelist, high_zoneidx,
2129 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2130 preferred_zone, migratetype);
2134 if (!(gfp_mask & __GFP_NOFAIL)) {
2135 /* The OOM killer will not help higher order allocs */
2136 if (order > PAGE_ALLOC_COSTLY_ORDER)
2138 /* The OOM killer does not needlessly kill tasks for lowmem */
2139 if (high_zoneidx < ZONE_NORMAL)
2142 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2143 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2144 * The caller should handle page allocation failure by itself if
2145 * it specifies __GFP_THISNODE.
2146 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2148 if (gfp_mask & __GFP_THISNODE)
2151 /* Exhausted what can be done so it's blamo time */
2152 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2155 clear_zonelist_oom(zonelist, gfp_mask);
2159 #ifdef CONFIG_COMPACTION
2160 /* Try memory compaction for high-order allocations before reclaim */
2161 static struct page *
2162 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2163 struct zonelist *zonelist, enum zone_type high_zoneidx,
2164 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2165 int migratetype, bool sync_migration,
2166 bool *contended_compaction, bool *deferred_compaction,
2167 unsigned long *did_some_progress)
2172 if (compaction_deferred(preferred_zone, order)) {
2173 *deferred_compaction = true;
2177 current->flags |= PF_MEMALLOC;
2178 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2179 nodemask, sync_migration,
2180 contended_compaction);
2181 current->flags &= ~PF_MEMALLOC;
2183 if (*did_some_progress != COMPACT_SKIPPED) {
2186 /* Page migration frees to the PCP lists but we want merging */
2187 drain_pages(get_cpu());
2190 page = get_page_from_freelist(gfp_mask, nodemask,
2191 order, zonelist, high_zoneidx,
2192 alloc_flags & ~ALLOC_NO_WATERMARKS,
2193 preferred_zone, migratetype);
2195 preferred_zone->compact_blockskip_flush = false;
2196 preferred_zone->compact_considered = 0;
2197 preferred_zone->compact_defer_shift = 0;
2198 if (order >= preferred_zone->compact_order_failed)
2199 preferred_zone->compact_order_failed = order + 1;
2200 count_vm_event(COMPACTSUCCESS);
2205 * It's bad if compaction run occurs and fails.
2206 * The most likely reason is that pages exist,
2207 * but not enough to satisfy watermarks.
2209 count_vm_event(COMPACTFAIL);
2212 * As async compaction considers a subset of pageblocks, only
2213 * defer if the failure was a sync compaction failure.
2216 defer_compaction(preferred_zone, order);
2224 static inline struct page *
2225 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2226 struct zonelist *zonelist, enum zone_type high_zoneidx,
2227 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2228 int migratetype, bool sync_migration,
2229 bool *contended_compaction, bool *deferred_compaction,
2230 unsigned long *did_some_progress)
2234 #endif /* CONFIG_COMPACTION */
2236 /* Perform direct synchronous page reclaim */
2238 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2239 nodemask_t *nodemask)
2241 struct reclaim_state reclaim_state;
2246 /* We now go into synchronous reclaim */
2247 cpuset_memory_pressure_bump();
2248 current->flags |= PF_MEMALLOC;
2249 lockdep_set_current_reclaim_state(gfp_mask);
2250 reclaim_state.reclaimed_slab = 0;
2251 current->reclaim_state = &reclaim_state;
2253 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2255 current->reclaim_state = NULL;
2256 lockdep_clear_current_reclaim_state();
2257 current->flags &= ~PF_MEMALLOC;
2264 /* The really slow allocator path where we enter direct reclaim */
2265 static inline struct page *
2266 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2267 struct zonelist *zonelist, enum zone_type high_zoneidx,
2268 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2269 int migratetype, unsigned long *did_some_progress)
2271 struct page *page = NULL;
2272 bool drained = false;
2274 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2276 if (unlikely(!(*did_some_progress)))
2279 /* After successful reclaim, reconsider all zones for allocation */
2280 if (IS_ENABLED(CONFIG_NUMA))
2281 zlc_clear_zones_full(zonelist);
2284 page = get_page_from_freelist(gfp_mask, nodemask, order,
2285 zonelist, high_zoneidx,
2286 alloc_flags & ~ALLOC_NO_WATERMARKS,
2287 preferred_zone, migratetype);
2290 * If an allocation failed after direct reclaim, it could be because
2291 * pages are pinned on the per-cpu lists. Drain them and try again
2293 if (!page && !drained) {
2303 * This is called in the allocator slow-path if the allocation request is of
2304 * sufficient urgency to ignore watermarks and take other desperate measures
2306 static inline struct page *
2307 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2308 struct zonelist *zonelist, enum zone_type high_zoneidx,
2309 nodemask_t *nodemask, struct zone *preferred_zone,
2315 page = get_page_from_freelist(gfp_mask, nodemask, order,
2316 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2317 preferred_zone, migratetype);
2319 if (!page && gfp_mask & __GFP_NOFAIL)
2320 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2321 } while (!page && (gfp_mask & __GFP_NOFAIL));
2327 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2328 enum zone_type high_zoneidx,
2329 enum zone_type classzone_idx)
2334 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2335 wakeup_kswapd(zone, order, classzone_idx);
2339 gfp_to_alloc_flags(gfp_t gfp_mask)
2341 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2342 const gfp_t wait = gfp_mask & __GFP_WAIT;
2344 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2345 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2348 * The caller may dip into page reserves a bit more if the caller
2349 * cannot run direct reclaim, or if the caller has realtime scheduling
2350 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2351 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2353 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2357 * Not worth trying to allocate harder for
2358 * __GFP_NOMEMALLOC even if it can't schedule.
2360 if (!(gfp_mask & __GFP_NOMEMALLOC))
2361 alloc_flags |= ALLOC_HARDER;
2363 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2364 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2366 alloc_flags &= ~ALLOC_CPUSET;
2367 } else if (unlikely(rt_task(current)) && !in_interrupt())
2368 alloc_flags |= ALLOC_HARDER;
2370 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2371 if (gfp_mask & __GFP_MEMALLOC)
2372 alloc_flags |= ALLOC_NO_WATERMARKS;
2373 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2374 alloc_flags |= ALLOC_NO_WATERMARKS;
2375 else if (!in_interrupt() &&
2376 ((current->flags & PF_MEMALLOC) ||
2377 unlikely(test_thread_flag(TIF_MEMDIE))))
2378 alloc_flags |= ALLOC_NO_WATERMARKS;
2381 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2382 alloc_flags |= ALLOC_CMA;
2387 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2389 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2392 static inline struct page *
2393 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2394 struct zonelist *zonelist, enum zone_type high_zoneidx,
2395 nodemask_t *nodemask, struct zone *preferred_zone,
2398 const gfp_t wait = gfp_mask & __GFP_WAIT;
2399 struct page *page = NULL;
2401 unsigned long pages_reclaimed = 0;
2402 unsigned long did_some_progress;
2403 bool sync_migration = false;
2404 bool deferred_compaction = false;
2405 bool contended_compaction = false;
2408 * In the slowpath, we sanity check order to avoid ever trying to
2409 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2410 * be using allocators in order of preference for an area that is
2413 if (order >= MAX_ORDER) {
2414 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2419 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2420 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2421 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2422 * using a larger set of nodes after it has established that the
2423 * allowed per node queues are empty and that nodes are
2426 if (IS_ENABLED(CONFIG_NUMA) &&
2427 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2431 if (!(gfp_mask & __GFP_NO_KSWAPD))
2432 wake_all_kswapd(order, zonelist, high_zoneidx,
2433 zone_idx(preferred_zone));
2436 * OK, we're below the kswapd watermark and have kicked background
2437 * reclaim. Now things get more complex, so set up alloc_flags according
2438 * to how we want to proceed.
2440 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2443 * Find the true preferred zone if the allocation is unconstrained by
2446 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2447 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2451 /* This is the last chance, in general, before the goto nopage. */
2452 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2453 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2454 preferred_zone, migratetype);
2458 /* Allocate without watermarks if the context allows */
2459 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2461 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2462 * the allocation is high priority and these type of
2463 * allocations are system rather than user orientated
2465 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2467 page = __alloc_pages_high_priority(gfp_mask, order,
2468 zonelist, high_zoneidx, nodemask,
2469 preferred_zone, migratetype);
2475 /* Atomic allocations - we can't balance anything */
2479 /* Avoid recursion of direct reclaim */
2480 if (current->flags & PF_MEMALLOC)
2483 /* Avoid allocations with no watermarks from looping endlessly */
2484 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2488 * Try direct compaction. The first pass is asynchronous. Subsequent
2489 * attempts after direct reclaim are synchronous
2491 page = __alloc_pages_direct_compact(gfp_mask, order,
2492 zonelist, high_zoneidx,
2494 alloc_flags, preferred_zone,
2495 migratetype, sync_migration,
2496 &contended_compaction,
2497 &deferred_compaction,
2498 &did_some_progress);
2501 sync_migration = true;
2504 * If compaction is deferred for high-order allocations, it is because
2505 * sync compaction recently failed. In this is the case and the caller
2506 * requested a movable allocation that does not heavily disrupt the
2507 * system then fail the allocation instead of entering direct reclaim.
2509 if ((deferred_compaction || contended_compaction) &&
2510 (gfp_mask & __GFP_NO_KSWAPD))
2513 /* Try direct reclaim and then allocating */
2514 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2515 zonelist, high_zoneidx,
2517 alloc_flags, preferred_zone,
2518 migratetype, &did_some_progress);
2523 * If we failed to make any progress reclaiming, then we are
2524 * running out of options and have to consider going OOM
2526 if (!did_some_progress) {
2527 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2528 if (oom_killer_disabled)
2530 /* Coredumps can quickly deplete all memory reserves */
2531 if ((current->flags & PF_DUMPCORE) &&
2532 !(gfp_mask & __GFP_NOFAIL))
2534 page = __alloc_pages_may_oom(gfp_mask, order,
2535 zonelist, high_zoneidx,
2536 nodemask, preferred_zone,
2541 if (!(gfp_mask & __GFP_NOFAIL)) {
2543 * The oom killer is not called for high-order
2544 * allocations that may fail, so if no progress
2545 * is being made, there are no other options and
2546 * retrying is unlikely to help.
2548 if (order > PAGE_ALLOC_COSTLY_ORDER)
2551 * The oom killer is not called for lowmem
2552 * allocations to prevent needlessly killing
2555 if (high_zoneidx < ZONE_NORMAL)
2563 /* Check if we should retry the allocation */
2564 pages_reclaimed += did_some_progress;
2565 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2567 /* Wait for some write requests to complete then retry */
2568 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2572 * High-order allocations do not necessarily loop after
2573 * direct reclaim and reclaim/compaction depends on compaction
2574 * being called after reclaim so call directly if necessary
2576 page = __alloc_pages_direct_compact(gfp_mask, order,
2577 zonelist, high_zoneidx,
2579 alloc_flags, preferred_zone,
2580 migratetype, sync_migration,
2581 &contended_compaction,
2582 &deferred_compaction,
2583 &did_some_progress);
2589 warn_alloc_failed(gfp_mask, order, NULL);
2592 if (kmemcheck_enabled)
2593 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2599 * This is the 'heart' of the zoned buddy allocator.
2602 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2603 struct zonelist *zonelist, nodemask_t *nodemask)
2605 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2606 struct zone *preferred_zone;
2607 struct page *page = NULL;
2608 int migratetype = allocflags_to_migratetype(gfp_mask);
2609 unsigned int cpuset_mems_cookie;
2610 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2611 struct mem_cgroup *memcg = NULL;
2613 gfp_mask &= gfp_allowed_mask;
2615 lockdep_trace_alloc(gfp_mask);
2617 might_sleep_if(gfp_mask & __GFP_WAIT);
2619 if (should_fail_alloc_page(gfp_mask, order))
2623 * Check the zones suitable for the gfp_mask contain at least one
2624 * valid zone. It's possible to have an empty zonelist as a result
2625 * of GFP_THISNODE and a memoryless node
2627 if (unlikely(!zonelist->_zonerefs->zone))
2631 * Will only have any effect when __GFP_KMEMCG is set. This is
2632 * verified in the (always inline) callee
2634 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2638 cpuset_mems_cookie = get_mems_allowed();
2640 /* The preferred zone is used for statistics later */
2641 first_zones_zonelist(zonelist, high_zoneidx,
2642 nodemask ? : &cpuset_current_mems_allowed,
2644 if (!preferred_zone)
2648 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2649 alloc_flags |= ALLOC_CMA;
2651 /* First allocation attempt */
2652 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2653 zonelist, high_zoneidx, alloc_flags,
2654 preferred_zone, migratetype);
2655 if (unlikely(!page)) {
2657 * Runtime PM, block IO and its error handling path
2658 * can deadlock because I/O on the device might not
2661 gfp_mask = memalloc_noio_flags(gfp_mask);
2662 page = __alloc_pages_slowpath(gfp_mask, order,
2663 zonelist, high_zoneidx, nodemask,
2664 preferred_zone, migratetype);
2667 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2671 * When updating a task's mems_allowed, it is possible to race with
2672 * parallel threads in such a way that an allocation can fail while
2673 * the mask is being updated. If a page allocation is about to fail,
2674 * check if the cpuset changed during allocation and if so, retry.
2676 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2679 memcg_kmem_commit_charge(page, memcg, order);
2683 EXPORT_SYMBOL(__alloc_pages_nodemask);
2686 * Common helper functions.
2688 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2693 * __get_free_pages() returns a 32-bit address, which cannot represent
2696 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2698 page = alloc_pages(gfp_mask, order);
2701 return (unsigned long) page_address(page);
2703 EXPORT_SYMBOL(__get_free_pages);
2705 unsigned long get_zeroed_page(gfp_t gfp_mask)
2707 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2709 EXPORT_SYMBOL(get_zeroed_page);
2711 void __free_pages(struct page *page, unsigned int order)
2713 if (put_page_testzero(page)) {
2715 free_hot_cold_page(page, 0);
2717 __free_pages_ok(page, order);
2721 EXPORT_SYMBOL(__free_pages);
2723 void free_pages(unsigned long addr, unsigned int order)
2726 VM_BUG_ON(!virt_addr_valid((void *)addr));
2727 __free_pages(virt_to_page((void *)addr), order);
2731 EXPORT_SYMBOL(free_pages);
2734 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2735 * pages allocated with __GFP_KMEMCG.
2737 * Those pages are accounted to a particular memcg, embedded in the
2738 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2739 * for that information only to find out that it is NULL for users who have no
2740 * interest in that whatsoever, we provide these functions.
2742 * The caller knows better which flags it relies on.
2744 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2746 memcg_kmem_uncharge_pages(page, order);
2747 __free_pages(page, order);
2750 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2753 VM_BUG_ON(!virt_addr_valid((void *)addr));
2754 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2758 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2761 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2762 unsigned long used = addr + PAGE_ALIGN(size);
2764 split_page(virt_to_page((void *)addr), order);
2765 while (used < alloc_end) {
2770 return (void *)addr;
2774 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2775 * @size: the number of bytes to allocate
2776 * @gfp_mask: GFP flags for the allocation
2778 * This function is similar to alloc_pages(), except that it allocates the
2779 * minimum number of pages to satisfy the request. alloc_pages() can only
2780 * allocate memory in power-of-two pages.
2782 * This function is also limited by MAX_ORDER.
2784 * Memory allocated by this function must be released by free_pages_exact().
2786 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2788 unsigned int order = get_order(size);
2791 addr = __get_free_pages(gfp_mask, order);
2792 return make_alloc_exact(addr, order, size);
2794 EXPORT_SYMBOL(alloc_pages_exact);
2797 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2799 * @nid: the preferred node ID where memory should be allocated
2800 * @size: the number of bytes to allocate
2801 * @gfp_mask: GFP flags for the allocation
2803 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2805 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2808 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2810 unsigned order = get_order(size);
2811 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2814 return make_alloc_exact((unsigned long)page_address(p), order, size);
2816 EXPORT_SYMBOL(alloc_pages_exact_nid);
2819 * free_pages_exact - release memory allocated via alloc_pages_exact()
2820 * @virt: the value returned by alloc_pages_exact.
2821 * @size: size of allocation, same value as passed to alloc_pages_exact().
2823 * Release the memory allocated by a previous call to alloc_pages_exact.
2825 void free_pages_exact(void *virt, size_t size)
2827 unsigned long addr = (unsigned long)virt;
2828 unsigned long end = addr + PAGE_ALIGN(size);
2830 while (addr < end) {
2835 EXPORT_SYMBOL(free_pages_exact);
2838 * nr_free_zone_pages - count number of pages beyond high watermark
2839 * @offset: The zone index of the highest zone
2841 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2842 * high watermark within all zones at or below a given zone index. For each
2843 * zone, the number of pages is calculated as:
2844 * present_pages - high_pages
2846 static unsigned long nr_free_zone_pages(int offset)
2851 /* Just pick one node, since fallback list is circular */
2852 unsigned long sum = 0;
2854 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2856 for_each_zone_zonelist(zone, z, zonelist, offset) {
2857 unsigned long size = zone->managed_pages;
2858 unsigned long high = high_wmark_pages(zone);
2867 * nr_free_buffer_pages - count number of pages beyond high watermark
2869 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2870 * watermark within ZONE_DMA and ZONE_NORMAL.
2872 unsigned long nr_free_buffer_pages(void)
2874 return nr_free_zone_pages(gfp_zone(GFP_USER));
2876 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2879 * nr_free_pagecache_pages - count number of pages beyond high watermark
2881 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2882 * high watermark within all zones.
2884 unsigned long nr_free_pagecache_pages(void)
2886 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2889 static inline void show_node(struct zone *zone)
2891 if (IS_ENABLED(CONFIG_NUMA))
2892 printk("Node %d ", zone_to_nid(zone));
2895 void si_meminfo(struct sysinfo *val)
2897 val->totalram = totalram_pages;
2899 val->freeram = global_page_state(NR_FREE_PAGES);
2900 val->bufferram = nr_blockdev_pages();
2901 val->totalhigh = totalhigh_pages;
2902 val->freehigh = nr_free_highpages();
2903 val->mem_unit = PAGE_SIZE;
2906 EXPORT_SYMBOL(si_meminfo);
2909 void si_meminfo_node(struct sysinfo *val, int nid)
2911 pg_data_t *pgdat = NODE_DATA(nid);
2913 val->totalram = pgdat->node_present_pages;
2914 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2915 #ifdef CONFIG_HIGHMEM
2916 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
2917 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2923 val->mem_unit = PAGE_SIZE;
2928 * Determine whether the node should be displayed or not, depending on whether
2929 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2931 bool skip_free_areas_node(unsigned int flags, int nid)
2934 unsigned int cpuset_mems_cookie;
2936 if (!(flags & SHOW_MEM_FILTER_NODES))
2940 cpuset_mems_cookie = get_mems_allowed();
2941 ret = !node_isset(nid, cpuset_current_mems_allowed);
2942 } while (!put_mems_allowed(cpuset_mems_cookie));
2947 #define K(x) ((x) << (PAGE_SHIFT-10))
2949 static void show_migration_types(unsigned char type)
2951 static const char types[MIGRATE_TYPES] = {
2952 [MIGRATE_UNMOVABLE] = 'U',
2953 [MIGRATE_RECLAIMABLE] = 'E',
2954 [MIGRATE_MOVABLE] = 'M',
2955 [MIGRATE_RESERVE] = 'R',
2957 [MIGRATE_CMA] = 'C',
2959 #ifdef CONFIG_MEMORY_ISOLATION
2960 [MIGRATE_ISOLATE] = 'I',
2963 char tmp[MIGRATE_TYPES + 1];
2967 for (i = 0; i < MIGRATE_TYPES; i++) {
2968 if (type & (1 << i))
2973 printk("(%s) ", tmp);
2977 * Show free area list (used inside shift_scroll-lock stuff)
2978 * We also calculate the percentage fragmentation. We do this by counting the
2979 * memory on each free list with the exception of the first item on the list.
2980 * Suppresses nodes that are not allowed by current's cpuset if
2981 * SHOW_MEM_FILTER_NODES is passed.
2983 void show_free_areas(unsigned int filter)
2988 for_each_populated_zone(zone) {
2989 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2992 printk("%s per-cpu:\n", zone->name);
2994 for_each_online_cpu(cpu) {
2995 struct per_cpu_pageset *pageset;
2997 pageset = per_cpu_ptr(zone->pageset, cpu);
2999 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3000 cpu, pageset->pcp.high,
3001 pageset->pcp.batch, pageset->pcp.count);
3005 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3006 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3008 " dirty:%lu writeback:%lu unstable:%lu\n"
3009 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3010 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3012 global_page_state(NR_ACTIVE_ANON),
3013 global_page_state(NR_INACTIVE_ANON),
3014 global_page_state(NR_ISOLATED_ANON),
3015 global_page_state(NR_ACTIVE_FILE),
3016 global_page_state(NR_INACTIVE_FILE),
3017 global_page_state(NR_ISOLATED_FILE),
3018 global_page_state(NR_UNEVICTABLE),
3019 global_page_state(NR_FILE_DIRTY),
3020 global_page_state(NR_WRITEBACK),
3021 global_page_state(NR_UNSTABLE_NFS),
3022 global_page_state(NR_FREE_PAGES),
3023 global_page_state(NR_SLAB_RECLAIMABLE),
3024 global_page_state(NR_SLAB_UNRECLAIMABLE),
3025 global_page_state(NR_FILE_MAPPED),
3026 global_page_state(NR_SHMEM),
3027 global_page_state(NR_PAGETABLE),
3028 global_page_state(NR_BOUNCE),
3029 global_page_state(NR_FREE_CMA_PAGES));
3031 for_each_populated_zone(zone) {
3034 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3042 " active_anon:%lukB"
3043 " inactive_anon:%lukB"
3044 " active_file:%lukB"
3045 " inactive_file:%lukB"
3046 " unevictable:%lukB"
3047 " isolated(anon):%lukB"
3048 " isolated(file):%lukB"
3056 " slab_reclaimable:%lukB"
3057 " slab_unreclaimable:%lukB"
3058 " kernel_stack:%lukB"
3063 " writeback_tmp:%lukB"
3064 " pages_scanned:%lu"
3065 " all_unreclaimable? %s"
3068 K(zone_page_state(zone, NR_FREE_PAGES)),
3069 K(min_wmark_pages(zone)),
3070 K(low_wmark_pages(zone)),
3071 K(high_wmark_pages(zone)),
3072 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3073 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3074 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3075 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3076 K(zone_page_state(zone, NR_UNEVICTABLE)),
3077 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3078 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3079 K(zone->present_pages),
3080 K(zone->managed_pages),
3081 K(zone_page_state(zone, NR_MLOCK)),
3082 K(zone_page_state(zone, NR_FILE_DIRTY)),
3083 K(zone_page_state(zone, NR_WRITEBACK)),
3084 K(zone_page_state(zone, NR_FILE_MAPPED)),
3085 K(zone_page_state(zone, NR_SHMEM)),
3086 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3087 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3088 zone_page_state(zone, NR_KERNEL_STACK) *
3090 K(zone_page_state(zone, NR_PAGETABLE)),
3091 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3092 K(zone_page_state(zone, NR_BOUNCE)),
3093 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3094 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3095 zone->pages_scanned,
3096 (zone->all_unreclaimable ? "yes" : "no")
3098 printk("lowmem_reserve[]:");
3099 for (i = 0; i < MAX_NR_ZONES; i++)
3100 printk(" %lu", zone->lowmem_reserve[i]);
3104 for_each_populated_zone(zone) {
3105 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3106 unsigned char types[MAX_ORDER];
3108 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3111 printk("%s: ", zone->name);
3113 spin_lock_irqsave(&zone->lock, flags);
3114 for (order = 0; order < MAX_ORDER; order++) {
3115 struct free_area *area = &zone->free_area[order];
3118 nr[order] = area->nr_free;
3119 total += nr[order] << order;
3122 for (type = 0; type < MIGRATE_TYPES; type++) {
3123 if (!list_empty(&area->free_list[type]))
3124 types[order] |= 1 << type;
3127 spin_unlock_irqrestore(&zone->lock, flags);
3128 for (order = 0; order < MAX_ORDER; order++) {
3129 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3131 show_migration_types(types[order]);
3133 printk("= %lukB\n", K(total));
3136 hugetlb_show_meminfo();
3138 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3140 show_swap_cache_info();
3143 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3145 zoneref->zone = zone;
3146 zoneref->zone_idx = zone_idx(zone);
3150 * Builds allocation fallback zone lists.
3152 * Add all populated zones of a node to the zonelist.
3154 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3155 int nr_zones, enum zone_type zone_type)
3159 BUG_ON(zone_type >= MAX_NR_ZONES);
3164 zone = pgdat->node_zones + zone_type;
3165 if (populated_zone(zone)) {
3166 zoneref_set_zone(zone,
3167 &zonelist->_zonerefs[nr_zones++]);
3168 check_highest_zone(zone_type);
3171 } while (zone_type);
3178 * 0 = automatic detection of better ordering.
3179 * 1 = order by ([node] distance, -zonetype)
3180 * 2 = order by (-zonetype, [node] distance)
3182 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3183 * the same zonelist. So only NUMA can configure this param.
3185 #define ZONELIST_ORDER_DEFAULT 0
3186 #define ZONELIST_ORDER_NODE 1
3187 #define ZONELIST_ORDER_ZONE 2
3189 /* zonelist order in the kernel.
3190 * set_zonelist_order() will set this to NODE or ZONE.
3192 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3193 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3197 /* The value user specified ....changed by config */
3198 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3199 /* string for sysctl */
3200 #define NUMA_ZONELIST_ORDER_LEN 16
3201 char numa_zonelist_order[16] = "default";
3204 * interface for configure zonelist ordering.
3205 * command line option "numa_zonelist_order"
3206 * = "[dD]efault - default, automatic configuration.
3207 * = "[nN]ode - order by node locality, then by zone within node
3208 * = "[zZ]one - order by zone, then by locality within zone
3211 static int __parse_numa_zonelist_order(char *s)
3213 if (*s == 'd' || *s == 'D') {
3214 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3215 } else if (*s == 'n' || *s == 'N') {
3216 user_zonelist_order = ZONELIST_ORDER_NODE;
3217 } else if (*s == 'z' || *s == 'Z') {
3218 user_zonelist_order = ZONELIST_ORDER_ZONE;
3221 "Ignoring invalid numa_zonelist_order value: "
3228 static __init int setup_numa_zonelist_order(char *s)
3235 ret = __parse_numa_zonelist_order(s);
3237 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3241 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3244 * sysctl handler for numa_zonelist_order
3246 int numa_zonelist_order_handler(ctl_table *table, int write,
3247 void __user *buffer, size_t *length,
3250 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3252 static DEFINE_MUTEX(zl_order_mutex);
3254 mutex_lock(&zl_order_mutex);
3256 strcpy(saved_string, (char*)table->data);
3257 ret = proc_dostring(table, write, buffer, length, ppos);
3261 int oldval = user_zonelist_order;
3262 if (__parse_numa_zonelist_order((char*)table->data)) {
3264 * bogus value. restore saved string
3266 strncpy((char*)table->data, saved_string,
3267 NUMA_ZONELIST_ORDER_LEN);
3268 user_zonelist_order = oldval;
3269 } else if (oldval != user_zonelist_order) {
3270 mutex_lock(&zonelists_mutex);
3271 build_all_zonelists(NULL, NULL);
3272 mutex_unlock(&zonelists_mutex);
3276 mutex_unlock(&zl_order_mutex);
3281 #define MAX_NODE_LOAD (nr_online_nodes)
3282 static int node_load[MAX_NUMNODES];
3285 * find_next_best_node - find the next node that should appear in a given node's fallback list
3286 * @node: node whose fallback list we're appending
3287 * @used_node_mask: nodemask_t of already used nodes
3289 * We use a number of factors to determine which is the next node that should
3290 * appear on a given node's fallback list. The node should not have appeared
3291 * already in @node's fallback list, and it should be the next closest node
3292 * according to the distance array (which contains arbitrary distance values
3293 * from each node to each node in the system), and should also prefer nodes
3294 * with no CPUs, since presumably they'll have very little allocation pressure
3295 * on them otherwise.
3296 * It returns -1 if no node is found.
3298 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3301 int min_val = INT_MAX;
3302 int best_node = NUMA_NO_NODE;
3303 const struct cpumask *tmp = cpumask_of_node(0);
3305 /* Use the local node if we haven't already */
3306 if (!node_isset(node, *used_node_mask)) {
3307 node_set(node, *used_node_mask);
3311 for_each_node_state(n, N_MEMORY) {
3313 /* Don't want a node to appear more than once */
3314 if (node_isset(n, *used_node_mask))
3317 /* Use the distance array to find the distance */
3318 val = node_distance(node, n);
3320 /* Penalize nodes under us ("prefer the next node") */
3323 /* Give preference to headless and unused nodes */
3324 tmp = cpumask_of_node(n);
3325 if (!cpumask_empty(tmp))
3326 val += PENALTY_FOR_NODE_WITH_CPUS;
3328 /* Slight preference for less loaded node */
3329 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3330 val += node_load[n];
3332 if (val < min_val) {
3339 node_set(best_node, *used_node_mask);
3346 * Build zonelists ordered by node and zones within node.
3347 * This results in maximum locality--normal zone overflows into local
3348 * DMA zone, if any--but risks exhausting DMA zone.
3350 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3353 struct zonelist *zonelist;
3355 zonelist = &pgdat->node_zonelists[0];
3356 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3358 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3360 zonelist->_zonerefs[j].zone = NULL;
3361 zonelist->_zonerefs[j].zone_idx = 0;
3365 * Build gfp_thisnode zonelists
3367 static void build_thisnode_zonelists(pg_data_t *pgdat)
3370 struct zonelist *zonelist;
3372 zonelist = &pgdat->node_zonelists[1];
3373 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3374 zonelist->_zonerefs[j].zone = NULL;
3375 zonelist->_zonerefs[j].zone_idx = 0;
3379 * Build zonelists ordered by zone and nodes within zones.
3380 * This results in conserving DMA zone[s] until all Normal memory is
3381 * exhausted, but results in overflowing to remote node while memory
3382 * may still exist in local DMA zone.
3384 static int node_order[MAX_NUMNODES];
3386 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3389 int zone_type; /* needs to be signed */
3391 struct zonelist *zonelist;
3393 zonelist = &pgdat->node_zonelists[0];
3395 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3396 for (j = 0; j < nr_nodes; j++) {
3397 node = node_order[j];
3398 z = &NODE_DATA(node)->node_zones[zone_type];
3399 if (populated_zone(z)) {
3401 &zonelist->_zonerefs[pos++]);
3402 check_highest_zone(zone_type);
3406 zonelist->_zonerefs[pos].zone = NULL;
3407 zonelist->_zonerefs[pos].zone_idx = 0;
3410 static int default_zonelist_order(void)
3413 unsigned long low_kmem_size,total_size;
3417 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3418 * If they are really small and used heavily, the system can fall
3419 * into OOM very easily.
3420 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3422 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3425 for_each_online_node(nid) {
3426 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3427 z = &NODE_DATA(nid)->node_zones[zone_type];
3428 if (populated_zone(z)) {
3429 if (zone_type < ZONE_NORMAL)
3430 low_kmem_size += z->present_pages;
3431 total_size += z->present_pages;
3432 } else if (zone_type == ZONE_NORMAL) {
3434 * If any node has only lowmem, then node order
3435 * is preferred to allow kernel allocations
3436 * locally; otherwise, they can easily infringe
3437 * on other nodes when there is an abundance of
3438 * lowmem available to allocate from.
3440 return ZONELIST_ORDER_NODE;
3444 if (!low_kmem_size || /* there are no DMA area. */
3445 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3446 return ZONELIST_ORDER_NODE;
3448 * look into each node's config.
3449 * If there is a node whose DMA/DMA32 memory is very big area on
3450 * local memory, NODE_ORDER may be suitable.
3452 average_size = total_size /
3453 (nodes_weight(node_states[N_MEMORY]) + 1);
3454 for_each_online_node(nid) {
3457 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3458 z = &NODE_DATA(nid)->node_zones[zone_type];
3459 if (populated_zone(z)) {
3460 if (zone_type < ZONE_NORMAL)
3461 low_kmem_size += z->present_pages;
3462 total_size += z->present_pages;
3465 if (low_kmem_size &&
3466 total_size > average_size && /* ignore small node */
3467 low_kmem_size > total_size * 70/100)
3468 return ZONELIST_ORDER_NODE;
3470 return ZONELIST_ORDER_ZONE;
3473 static void set_zonelist_order(void)
3475 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3476 current_zonelist_order = default_zonelist_order();
3478 current_zonelist_order = user_zonelist_order;
3481 static void build_zonelists(pg_data_t *pgdat)
3485 nodemask_t used_mask;
3486 int local_node, prev_node;
3487 struct zonelist *zonelist;
3488 int order = current_zonelist_order;
3490 /* initialize zonelists */
3491 for (i = 0; i < MAX_ZONELISTS; i++) {
3492 zonelist = pgdat->node_zonelists + i;
3493 zonelist->_zonerefs[0].zone = NULL;
3494 zonelist->_zonerefs[0].zone_idx = 0;
3497 /* NUMA-aware ordering of nodes */
3498 local_node = pgdat->node_id;
3499 load = nr_online_nodes;
3500 prev_node = local_node;
3501 nodes_clear(used_mask);
3503 memset(node_order, 0, sizeof(node_order));
3506 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3508 * We don't want to pressure a particular node.
3509 * So adding penalty to the first node in same
3510 * distance group to make it round-robin.
3512 if (node_distance(local_node, node) !=
3513 node_distance(local_node, prev_node))
3514 node_load[node] = load;
3518 if (order == ZONELIST_ORDER_NODE)
3519 build_zonelists_in_node_order(pgdat, node);
3521 node_order[j++] = node; /* remember order */
3524 if (order == ZONELIST_ORDER_ZONE) {
3525 /* calculate node order -- i.e., DMA last! */
3526 build_zonelists_in_zone_order(pgdat, j);
3529 build_thisnode_zonelists(pgdat);
3532 /* Construct the zonelist performance cache - see further mmzone.h */
3533 static void build_zonelist_cache(pg_data_t *pgdat)
3535 struct zonelist *zonelist;
3536 struct zonelist_cache *zlc;
3539 zonelist = &pgdat->node_zonelists[0];
3540 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3541 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3542 for (z = zonelist->_zonerefs; z->zone; z++)
3543 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3546 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3548 * Return node id of node used for "local" allocations.
3549 * I.e., first node id of first zone in arg node's generic zonelist.
3550 * Used for initializing percpu 'numa_mem', which is used primarily
3551 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3553 int local_memory_node(int node)
3557 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3558 gfp_zone(GFP_KERNEL),
3565 #else /* CONFIG_NUMA */
3567 static void set_zonelist_order(void)
3569 current_zonelist_order = ZONELIST_ORDER_ZONE;
3572 static void build_zonelists(pg_data_t *pgdat)
3574 int node, local_node;
3576 struct zonelist *zonelist;
3578 local_node = pgdat->node_id;
3580 zonelist = &pgdat->node_zonelists[0];
3581 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3584 * Now we build the zonelist so that it contains the zones
3585 * of all the other nodes.
3586 * We don't want to pressure a particular node, so when
3587 * building the zones for node N, we make sure that the
3588 * zones coming right after the local ones are those from
3589 * node N+1 (modulo N)
3591 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3592 if (!node_online(node))
3594 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3597 for (node = 0; node < local_node; node++) {
3598 if (!node_online(node))
3600 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3604 zonelist->_zonerefs[j].zone = NULL;
3605 zonelist->_zonerefs[j].zone_idx = 0;
3608 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3609 static void build_zonelist_cache(pg_data_t *pgdat)
3611 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3614 #endif /* CONFIG_NUMA */
3617 * Boot pageset table. One per cpu which is going to be used for all
3618 * zones and all nodes. The parameters will be set in such a way
3619 * that an item put on a list will immediately be handed over to
3620 * the buddy list. This is safe since pageset manipulation is done
3621 * with interrupts disabled.
3623 * The boot_pagesets must be kept even after bootup is complete for
3624 * unused processors and/or zones. They do play a role for bootstrapping
3625 * hotplugged processors.
3627 * zoneinfo_show() and maybe other functions do
3628 * not check if the processor is online before following the pageset pointer.
3629 * Other parts of the kernel may not check if the zone is available.
3631 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3632 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3633 static void setup_zone_pageset(struct zone *zone);
3636 * Global mutex to protect against size modification of zonelists
3637 * as well as to serialize pageset setup for the new populated zone.
3639 DEFINE_MUTEX(zonelists_mutex);
3641 /* return values int ....just for stop_machine() */
3642 static int __build_all_zonelists(void *data)
3646 pg_data_t *self = data;
3649 memset(node_load, 0, sizeof(node_load));
3652 if (self && !node_online(self->node_id)) {
3653 build_zonelists(self);
3654 build_zonelist_cache(self);
3657 for_each_online_node(nid) {
3658 pg_data_t *pgdat = NODE_DATA(nid);
3660 build_zonelists(pgdat);
3661 build_zonelist_cache(pgdat);
3665 * Initialize the boot_pagesets that are going to be used
3666 * for bootstrapping processors. The real pagesets for
3667 * each zone will be allocated later when the per cpu
3668 * allocator is available.
3670 * boot_pagesets are used also for bootstrapping offline
3671 * cpus if the system is already booted because the pagesets
3672 * are needed to initialize allocators on a specific cpu too.
3673 * F.e. the percpu allocator needs the page allocator which
3674 * needs the percpu allocator in order to allocate its pagesets
3675 * (a chicken-egg dilemma).
3677 for_each_possible_cpu(cpu) {
3678 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3680 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3682 * We now know the "local memory node" for each node--
3683 * i.e., the node of the first zone in the generic zonelist.
3684 * Set up numa_mem percpu variable for on-line cpus. During
3685 * boot, only the boot cpu should be on-line; we'll init the
3686 * secondary cpus' numa_mem as they come on-line. During
3687 * node/memory hotplug, we'll fixup all on-line cpus.
3689 if (cpu_online(cpu))
3690 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3698 * Called with zonelists_mutex held always
3699 * unless system_state == SYSTEM_BOOTING.
3701 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3703 set_zonelist_order();
3705 if (system_state == SYSTEM_BOOTING) {
3706 __build_all_zonelists(NULL);
3707 mminit_verify_zonelist();
3708 cpuset_init_current_mems_allowed();
3710 /* we have to stop all cpus to guarantee there is no user
3712 #ifdef CONFIG_MEMORY_HOTPLUG
3714 setup_zone_pageset(zone);
3716 stop_machine(__build_all_zonelists, pgdat, NULL);
3717 /* cpuset refresh routine should be here */
3719 vm_total_pages = nr_free_pagecache_pages();
3721 * Disable grouping by mobility if the number of pages in the
3722 * system is too low to allow the mechanism to work. It would be
3723 * more accurate, but expensive to check per-zone. This check is
3724 * made on memory-hotadd so a system can start with mobility
3725 * disabled and enable it later
3727 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3728 page_group_by_mobility_disabled = 1;
3730 page_group_by_mobility_disabled = 0;
3732 printk("Built %i zonelists in %s order, mobility grouping %s. "
3733 "Total pages: %ld\n",
3735 zonelist_order_name[current_zonelist_order],
3736 page_group_by_mobility_disabled ? "off" : "on",
3739 printk("Policy zone: %s\n", zone_names[policy_zone]);
3744 * Helper functions to size the waitqueue hash table.
3745 * Essentially these want to choose hash table sizes sufficiently
3746 * large so that collisions trying to wait on pages are rare.
3747 * But in fact, the number of active page waitqueues on typical
3748 * systems is ridiculously low, less than 200. So this is even
3749 * conservative, even though it seems large.
3751 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3752 * waitqueues, i.e. the size of the waitq table given the number of pages.
3754 #define PAGES_PER_WAITQUEUE 256
3756 #ifndef CONFIG_MEMORY_HOTPLUG
3757 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3759 unsigned long size = 1;
3761 pages /= PAGES_PER_WAITQUEUE;
3763 while (size < pages)
3767 * Once we have dozens or even hundreds of threads sleeping
3768 * on IO we've got bigger problems than wait queue collision.
3769 * Limit the size of the wait table to a reasonable size.
3771 size = min(size, 4096UL);
3773 return max(size, 4UL);
3777 * A zone's size might be changed by hot-add, so it is not possible to determine
3778 * a suitable size for its wait_table. So we use the maximum size now.
3780 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3782 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3783 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3784 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3786 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3787 * or more by the traditional way. (See above). It equals:
3789 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3790 * ia64(16K page size) : = ( 8G + 4M)byte.
3791 * powerpc (64K page size) : = (32G +16M)byte.
3793 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3800 * This is an integer logarithm so that shifts can be used later
3801 * to extract the more random high bits from the multiplicative
3802 * hash function before the remainder is taken.
3804 static inline unsigned long wait_table_bits(unsigned long size)
3809 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3812 * Check if a pageblock contains reserved pages
3814 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3818 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3819 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3826 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3827 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3828 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3829 * higher will lead to a bigger reserve which will get freed as contiguous
3830 * blocks as reclaim kicks in
3832 static void setup_zone_migrate_reserve(struct zone *zone)
3834 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3836 unsigned long block_migratetype;
3840 * Get the start pfn, end pfn and the number of blocks to reserve
3841 * We have to be careful to be aligned to pageblock_nr_pages to
3842 * make sure that we always check pfn_valid for the first page in
3845 start_pfn = zone->zone_start_pfn;
3846 end_pfn = zone_end_pfn(zone);
3847 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3848 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3852 * Reserve blocks are generally in place to help high-order atomic
3853 * allocations that are short-lived. A min_free_kbytes value that
3854 * would result in more than 2 reserve blocks for atomic allocations
3855 * is assumed to be in place to help anti-fragmentation for the
3856 * future allocation of hugepages at runtime.
3858 reserve = min(2, reserve);
3860 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3861 if (!pfn_valid(pfn))
3863 page = pfn_to_page(pfn);
3865 /* Watch out for overlapping nodes */
3866 if (page_to_nid(page) != zone_to_nid(zone))
3869 block_migratetype = get_pageblock_migratetype(page);
3871 /* Only test what is necessary when the reserves are not met */
3874 * Blocks with reserved pages will never free, skip
3877 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3878 if (pageblock_is_reserved(pfn, block_end_pfn))
3881 /* If this block is reserved, account for it */
3882 if (block_migratetype == MIGRATE_RESERVE) {
3887 /* Suitable for reserving if this block is movable */
3888 if (block_migratetype == MIGRATE_MOVABLE) {
3889 set_pageblock_migratetype(page,
3891 move_freepages_block(zone, page,
3899 * If the reserve is met and this is a previous reserved block,
3902 if (block_migratetype == MIGRATE_RESERVE) {
3903 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3904 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3910 * Initially all pages are reserved - free ones are freed
3911 * up by free_all_bootmem() once the early boot process is
3912 * done. Non-atomic initialization, single-pass.
3914 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3915 unsigned long start_pfn, enum memmap_context context)
3918 unsigned long end_pfn = start_pfn + size;
3922 if (highest_memmap_pfn < end_pfn - 1)
3923 highest_memmap_pfn = end_pfn - 1;
3925 z = &NODE_DATA(nid)->node_zones[zone];
3926 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3928 * There can be holes in boot-time mem_map[]s
3929 * handed to this function. They do not
3930 * exist on hotplugged memory.
3932 if (context == MEMMAP_EARLY) {
3933 if (!early_pfn_valid(pfn))
3935 if (!early_pfn_in_nid(pfn, nid))
3938 page = pfn_to_page(pfn);
3939 set_page_links(page, zone, nid, pfn);
3940 mminit_verify_page_links(page, zone, nid, pfn);
3941 init_page_count(page);
3942 page_mapcount_reset(page);
3943 page_nid_reset_last(page);
3944 SetPageReserved(page);
3946 * Mark the block movable so that blocks are reserved for
3947 * movable at startup. This will force kernel allocations
3948 * to reserve their blocks rather than leaking throughout
3949 * the address space during boot when many long-lived
3950 * kernel allocations are made. Later some blocks near
3951 * the start are marked MIGRATE_RESERVE by
3952 * setup_zone_migrate_reserve()
3954 * bitmap is created for zone's valid pfn range. but memmap
3955 * can be created for invalid pages (for alignment)
3956 * check here not to call set_pageblock_migratetype() against
3959 if ((z->zone_start_pfn <= pfn)
3960 && (pfn < zone_end_pfn(z))
3961 && !(pfn & (pageblock_nr_pages - 1)))
3962 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3964 INIT_LIST_HEAD(&page->lru);
3965 #ifdef WANT_PAGE_VIRTUAL
3966 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3967 if (!is_highmem_idx(zone))
3968 set_page_address(page, __va(pfn << PAGE_SHIFT));
3973 static void __meminit zone_init_free_lists(struct zone *zone)
3976 for_each_migratetype_order(order, t) {
3977 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3978 zone->free_area[order].nr_free = 0;
3982 #ifndef __HAVE_ARCH_MEMMAP_INIT
3983 #define memmap_init(size, nid, zone, start_pfn) \
3984 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3987 static int __meminit zone_batchsize(struct zone *zone)
3993 * The per-cpu-pages pools are set to around 1000th of the
3994 * size of the zone. But no more than 1/2 of a meg.
3996 * OK, so we don't know how big the cache is. So guess.
3998 batch = zone->managed_pages / 1024;
3999 if (batch * PAGE_SIZE > 512 * 1024)
4000 batch = (512 * 1024) / PAGE_SIZE;
4001 batch /= 4; /* We effectively *= 4 below */
4006 * Clamp the batch to a 2^n - 1 value. Having a power
4007 * of 2 value was found to be more likely to have
4008 * suboptimal cache aliasing properties in some cases.
4010 * For example if 2 tasks are alternately allocating
4011 * batches of pages, one task can end up with a lot
4012 * of pages of one half of the possible page colors
4013 * and the other with pages of the other colors.
4015 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4020 /* The deferral and batching of frees should be suppressed under NOMMU
4023 * The problem is that NOMMU needs to be able to allocate large chunks
4024 * of contiguous memory as there's no hardware page translation to
4025 * assemble apparent contiguous memory from discontiguous pages.
4027 * Queueing large contiguous runs of pages for batching, however,
4028 * causes the pages to actually be freed in smaller chunks. As there
4029 * can be a significant delay between the individual batches being
4030 * recycled, this leads to the once large chunks of space being
4031 * fragmented and becoming unavailable for high-order allocations.
4037 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4039 struct per_cpu_pages *pcp;
4042 memset(p, 0, sizeof(*p));
4046 pcp->high = 6 * batch;
4047 pcp->batch = max(1UL, 1 * batch);
4048 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4049 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4053 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4054 * to the value high for the pageset p.
4057 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
4060 struct per_cpu_pages *pcp;
4064 pcp->batch = max(1UL, high/4);
4065 if ((high/4) > (PAGE_SHIFT * 8))
4066 pcp->batch = PAGE_SHIFT * 8;
4069 static void __meminit setup_zone_pageset(struct zone *zone)
4073 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4075 for_each_possible_cpu(cpu) {
4076 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4078 setup_pageset(pcp, zone_batchsize(zone));
4080 if (percpu_pagelist_fraction)
4081 setup_pagelist_highmark(pcp,
4082 (zone->managed_pages /
4083 percpu_pagelist_fraction));
4088 * Allocate per cpu pagesets and initialize them.
4089 * Before this call only boot pagesets were available.
4091 void __init setup_per_cpu_pageset(void)
4095 for_each_populated_zone(zone)
4096 setup_zone_pageset(zone);
4099 static noinline __init_refok
4100 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4103 struct pglist_data *pgdat = zone->zone_pgdat;
4107 * The per-page waitqueue mechanism uses hashed waitqueues
4110 zone->wait_table_hash_nr_entries =
4111 wait_table_hash_nr_entries(zone_size_pages);
4112 zone->wait_table_bits =
4113 wait_table_bits(zone->wait_table_hash_nr_entries);
4114 alloc_size = zone->wait_table_hash_nr_entries
4115 * sizeof(wait_queue_head_t);
4117 if (!slab_is_available()) {
4118 zone->wait_table = (wait_queue_head_t *)
4119 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4122 * This case means that a zone whose size was 0 gets new memory
4123 * via memory hot-add.
4124 * But it may be the case that a new node was hot-added. In
4125 * this case vmalloc() will not be able to use this new node's
4126 * memory - this wait_table must be initialized to use this new
4127 * node itself as well.
4128 * To use this new node's memory, further consideration will be
4131 zone->wait_table = vmalloc(alloc_size);
4133 if (!zone->wait_table)
4136 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4137 init_waitqueue_head(zone->wait_table + i);
4142 static __meminit void zone_pcp_init(struct zone *zone)
4145 * per cpu subsystem is not up at this point. The following code
4146 * relies on the ability of the linker to provide the
4147 * offset of a (static) per cpu variable into the per cpu area.
4149 zone->pageset = &boot_pageset;
4151 if (zone->present_pages)
4152 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4153 zone->name, zone->present_pages,
4154 zone_batchsize(zone));
4157 int __meminit init_currently_empty_zone(struct zone *zone,
4158 unsigned long zone_start_pfn,
4160 enum memmap_context context)
4162 struct pglist_data *pgdat = zone->zone_pgdat;
4164 ret = zone_wait_table_init(zone, size);
4167 pgdat->nr_zones = zone_idx(zone) + 1;
4169 zone->zone_start_pfn = zone_start_pfn;
4171 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4172 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4174 (unsigned long)zone_idx(zone),
4175 zone_start_pfn, (zone_start_pfn + size));
4177 zone_init_free_lists(zone);
4182 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4183 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4185 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4186 * Architectures may implement their own version but if add_active_range()
4187 * was used and there are no special requirements, this is a convenient
4190 int __meminit __early_pfn_to_nid(unsigned long pfn)
4192 unsigned long start_pfn, end_pfn;
4195 * NOTE: The following SMP-unsafe globals are only used early in boot
4196 * when the kernel is running single-threaded.
4198 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4199 static int __meminitdata last_nid;
4201 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4204 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4205 if (start_pfn <= pfn && pfn < end_pfn) {
4206 last_start_pfn = start_pfn;
4207 last_end_pfn = end_pfn;
4211 /* This is a memory hole */
4214 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4216 int __meminit early_pfn_to_nid(unsigned long pfn)
4220 nid = __early_pfn_to_nid(pfn);
4223 /* just returns 0 */
4227 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4228 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4232 nid = __early_pfn_to_nid(pfn);
4233 if (nid >= 0 && nid != node)
4240 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4241 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4242 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4244 * If an architecture guarantees that all ranges registered with
4245 * add_active_ranges() contain no holes and may be freed, this
4246 * this function may be used instead of calling free_bootmem() manually.
4248 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4250 unsigned long start_pfn, end_pfn;
4253 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4254 start_pfn = min(start_pfn, max_low_pfn);
4255 end_pfn = min(end_pfn, max_low_pfn);
4257 if (start_pfn < end_pfn)
4258 free_bootmem_node(NODE_DATA(this_nid),
4259 PFN_PHYS(start_pfn),
4260 (end_pfn - start_pfn) << PAGE_SHIFT);
4265 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4266 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4268 * If an architecture guarantees that all ranges registered with
4269 * add_active_ranges() contain no holes and may be freed, this
4270 * function may be used instead of calling memory_present() manually.
4272 void __init sparse_memory_present_with_active_regions(int nid)
4274 unsigned long start_pfn, end_pfn;
4277 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4278 memory_present(this_nid, start_pfn, end_pfn);
4282 * get_pfn_range_for_nid - Return the start and end page frames for a node
4283 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4284 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4285 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4287 * It returns the start and end page frame of a node based on information
4288 * provided by an arch calling add_active_range(). If called for a node
4289 * with no available memory, a warning is printed and the start and end
4292 void __meminit get_pfn_range_for_nid(unsigned int nid,
4293 unsigned long *start_pfn, unsigned long *end_pfn)
4295 unsigned long this_start_pfn, this_end_pfn;
4301 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4302 *start_pfn = min(*start_pfn, this_start_pfn);
4303 *end_pfn = max(*end_pfn, this_end_pfn);
4306 if (*start_pfn == -1UL)
4311 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4312 * assumption is made that zones within a node are ordered in monotonic
4313 * increasing memory addresses so that the "highest" populated zone is used
4315 static void __init find_usable_zone_for_movable(void)
4318 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4319 if (zone_index == ZONE_MOVABLE)
4322 if (arch_zone_highest_possible_pfn[zone_index] >
4323 arch_zone_lowest_possible_pfn[zone_index])
4327 VM_BUG_ON(zone_index == -1);
4328 movable_zone = zone_index;
4332 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4333 * because it is sized independent of architecture. Unlike the other zones,
4334 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4335 * in each node depending on the size of each node and how evenly kernelcore
4336 * is distributed. This helper function adjusts the zone ranges
4337 * provided by the architecture for a given node by using the end of the
4338 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4339 * zones within a node are in order of monotonic increases memory addresses
4341 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4342 unsigned long zone_type,
4343 unsigned long node_start_pfn,
4344 unsigned long node_end_pfn,
4345 unsigned long *zone_start_pfn,
4346 unsigned long *zone_end_pfn)
4348 /* Only adjust if ZONE_MOVABLE is on this node */
4349 if (zone_movable_pfn[nid]) {
4350 /* Size ZONE_MOVABLE */
4351 if (zone_type == ZONE_MOVABLE) {
4352 *zone_start_pfn = zone_movable_pfn[nid];
4353 *zone_end_pfn = min(node_end_pfn,
4354 arch_zone_highest_possible_pfn[movable_zone]);
4356 /* Adjust for ZONE_MOVABLE starting within this range */
4357 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4358 *zone_end_pfn > zone_movable_pfn[nid]) {
4359 *zone_end_pfn = zone_movable_pfn[nid];
4361 /* Check if this whole range is within ZONE_MOVABLE */
4362 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4363 *zone_start_pfn = *zone_end_pfn;
4368 * Return the number of pages a zone spans in a node, including holes
4369 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4371 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4372 unsigned long zone_type,
4373 unsigned long *ignored)
4375 unsigned long node_start_pfn, node_end_pfn;
4376 unsigned long zone_start_pfn, zone_end_pfn;
4378 /* Get the start and end of the node and zone */
4379 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4380 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4381 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4382 adjust_zone_range_for_zone_movable(nid, zone_type,
4383 node_start_pfn, node_end_pfn,
4384 &zone_start_pfn, &zone_end_pfn);
4386 /* Check that this node has pages within the zone's required range */
4387 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4390 /* Move the zone boundaries inside the node if necessary */
4391 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4392 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4394 /* Return the spanned pages */
4395 return zone_end_pfn - zone_start_pfn;
4399 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4400 * then all holes in the requested range will be accounted for.
4402 unsigned long __meminit __absent_pages_in_range(int nid,
4403 unsigned long range_start_pfn,
4404 unsigned long range_end_pfn)
4406 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4407 unsigned long start_pfn, end_pfn;
4410 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4411 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4412 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4413 nr_absent -= end_pfn - start_pfn;
4419 * absent_pages_in_range - Return number of page frames in holes within a range
4420 * @start_pfn: The start PFN to start searching for holes
4421 * @end_pfn: The end PFN to stop searching for holes
4423 * It returns the number of pages frames in memory holes within a range.
4425 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4426 unsigned long end_pfn)
4428 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4431 /* Return the number of page frames in holes in a zone on a node */
4432 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4433 unsigned long zone_type,
4434 unsigned long *ignored)
4436 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4437 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4438 unsigned long node_start_pfn, node_end_pfn;
4439 unsigned long zone_start_pfn, zone_end_pfn;
4441 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4442 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4443 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4445 adjust_zone_range_for_zone_movable(nid, zone_type,
4446 node_start_pfn, node_end_pfn,
4447 &zone_start_pfn, &zone_end_pfn);
4448 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4451 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4452 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4453 unsigned long zone_type,
4454 unsigned long *zones_size)
4456 return zones_size[zone_type];
4459 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4460 unsigned long zone_type,
4461 unsigned long *zholes_size)
4466 return zholes_size[zone_type];
4469 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4471 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4472 unsigned long *zones_size, unsigned long *zholes_size)
4474 unsigned long realtotalpages, totalpages = 0;
4477 for (i = 0; i < MAX_NR_ZONES; i++)
4478 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4480 pgdat->node_spanned_pages = totalpages;
4482 realtotalpages = totalpages;
4483 for (i = 0; i < MAX_NR_ZONES; i++)
4485 zone_absent_pages_in_node(pgdat->node_id, i,
4487 pgdat->node_present_pages = realtotalpages;
4488 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4492 #ifndef CONFIG_SPARSEMEM
4494 * Calculate the size of the zone->blockflags rounded to an unsigned long
4495 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4496 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4497 * round what is now in bits to nearest long in bits, then return it in
4500 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4502 unsigned long usemapsize;
4504 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4505 usemapsize = roundup(zonesize, pageblock_nr_pages);
4506 usemapsize = usemapsize >> pageblock_order;
4507 usemapsize *= NR_PAGEBLOCK_BITS;
4508 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4510 return usemapsize / 8;
4513 static void __init setup_usemap(struct pglist_data *pgdat,
4515 unsigned long zone_start_pfn,
4516 unsigned long zonesize)
4518 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4519 zone->pageblock_flags = NULL;
4521 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4525 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4526 unsigned long zone_start_pfn, unsigned long zonesize) {}
4527 #endif /* CONFIG_SPARSEMEM */
4529 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4531 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4532 void __init set_pageblock_order(void)
4536 /* Check that pageblock_nr_pages has not already been setup */
4537 if (pageblock_order)
4540 if (HPAGE_SHIFT > PAGE_SHIFT)
4541 order = HUGETLB_PAGE_ORDER;
4543 order = MAX_ORDER - 1;
4546 * Assume the largest contiguous order of interest is a huge page.
4547 * This value may be variable depending on boot parameters on IA64 and
4550 pageblock_order = order;
4552 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4555 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4556 * is unused as pageblock_order is set at compile-time. See
4557 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4560 void __init set_pageblock_order(void)
4564 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4566 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4567 unsigned long present_pages)
4569 unsigned long pages = spanned_pages;
4572 * Provide a more accurate estimation if there are holes within
4573 * the zone and SPARSEMEM is in use. If there are holes within the
4574 * zone, each populated memory region may cost us one or two extra
4575 * memmap pages due to alignment because memmap pages for each
4576 * populated regions may not naturally algined on page boundary.
4577 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4579 if (spanned_pages > present_pages + (present_pages >> 4) &&
4580 IS_ENABLED(CONFIG_SPARSEMEM))
4581 pages = present_pages;
4583 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4587 * Set up the zone data structures:
4588 * - mark all pages reserved
4589 * - mark all memory queues empty
4590 * - clear the memory bitmaps
4592 * NOTE: pgdat should get zeroed by caller.
4594 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4595 unsigned long *zones_size, unsigned long *zholes_size)
4598 int nid = pgdat->node_id;
4599 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4602 pgdat_resize_init(pgdat);
4603 #ifdef CONFIG_NUMA_BALANCING
4604 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4605 pgdat->numabalancing_migrate_nr_pages = 0;
4606 pgdat->numabalancing_migrate_next_window = jiffies;
4608 init_waitqueue_head(&pgdat->kswapd_wait);
4609 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4610 pgdat_page_cgroup_init(pgdat);
4612 for (j = 0; j < MAX_NR_ZONES; j++) {
4613 struct zone *zone = pgdat->node_zones + j;
4614 unsigned long size, realsize, freesize, memmap_pages;
4616 size = zone_spanned_pages_in_node(nid, j, zones_size);
4617 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4621 * Adjust freesize so that it accounts for how much memory
4622 * is used by this zone for memmap. This affects the watermark
4623 * and per-cpu initialisations
4625 memmap_pages = calc_memmap_size(size, realsize);
4626 if (freesize >= memmap_pages) {
4627 freesize -= memmap_pages;
4630 " %s zone: %lu pages used for memmap\n",
4631 zone_names[j], memmap_pages);
4634 " %s zone: %lu pages exceeds freesize %lu\n",
4635 zone_names[j], memmap_pages, freesize);
4637 /* Account for reserved pages */
4638 if (j == 0 && freesize > dma_reserve) {
4639 freesize -= dma_reserve;
4640 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4641 zone_names[0], dma_reserve);
4644 if (!is_highmem_idx(j))
4645 nr_kernel_pages += freesize;
4646 /* Charge for highmem memmap if there are enough kernel pages */
4647 else if (nr_kernel_pages > memmap_pages * 2)
4648 nr_kernel_pages -= memmap_pages;
4649 nr_all_pages += freesize;
4651 zone->spanned_pages = size;
4652 zone->present_pages = realsize;
4654 * Set an approximate value for lowmem here, it will be adjusted
4655 * when the bootmem allocator frees pages into the buddy system.
4656 * And all highmem pages will be managed by the buddy system.
4658 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4661 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4663 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4665 zone->name = zone_names[j];
4666 spin_lock_init(&zone->lock);
4667 spin_lock_init(&zone->lru_lock);
4668 zone_seqlock_init(zone);
4669 zone->zone_pgdat = pgdat;
4671 zone_pcp_init(zone);
4672 lruvec_init(&zone->lruvec);
4676 set_pageblock_order();
4677 setup_usemap(pgdat, zone, zone_start_pfn, size);
4678 ret = init_currently_empty_zone(zone, zone_start_pfn,
4679 size, MEMMAP_EARLY);
4681 memmap_init(size, nid, j, zone_start_pfn);
4682 zone_start_pfn += size;
4686 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4688 /* Skip empty nodes */
4689 if (!pgdat->node_spanned_pages)
4692 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4693 /* ia64 gets its own node_mem_map, before this, without bootmem */
4694 if (!pgdat->node_mem_map) {
4695 unsigned long size, start, end;
4699 * The zone's endpoints aren't required to be MAX_ORDER
4700 * aligned but the node_mem_map endpoints must be in order
4701 * for the buddy allocator to function correctly.
4703 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4704 end = pgdat_end_pfn(pgdat);
4705 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4706 size = (end - start) * sizeof(struct page);
4707 map = alloc_remap(pgdat->node_id, size);
4709 map = alloc_bootmem_node_nopanic(pgdat, size);
4710 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4712 #ifndef CONFIG_NEED_MULTIPLE_NODES
4714 * With no DISCONTIG, the global mem_map is just set as node 0's
4716 if (pgdat == NODE_DATA(0)) {
4717 mem_map = NODE_DATA(0)->node_mem_map;
4718 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4719 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4720 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4721 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4724 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4727 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4728 unsigned long node_start_pfn, unsigned long *zholes_size)
4730 pg_data_t *pgdat = NODE_DATA(nid);
4732 /* pg_data_t should be reset to zero when it's allocated */
4733 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4735 pgdat->node_id = nid;
4736 pgdat->node_start_pfn = node_start_pfn;
4737 init_zone_allows_reclaim(nid);
4738 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4740 alloc_node_mem_map(pgdat);
4741 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4742 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4743 nid, (unsigned long)pgdat,
4744 (unsigned long)pgdat->node_mem_map);
4747 free_area_init_core(pgdat, zones_size, zholes_size);
4750 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4752 #if MAX_NUMNODES > 1
4754 * Figure out the number of possible node ids.
4756 void __init setup_nr_node_ids(void)
4759 unsigned int highest = 0;
4761 for_each_node_mask(node, node_possible_map)
4763 nr_node_ids = highest + 1;
4768 * node_map_pfn_alignment - determine the maximum internode alignment
4770 * This function should be called after node map is populated and sorted.
4771 * It calculates the maximum power of two alignment which can distinguish
4774 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4775 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4776 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4777 * shifted, 1GiB is enough and this function will indicate so.
4779 * This is used to test whether pfn -> nid mapping of the chosen memory
4780 * model has fine enough granularity to avoid incorrect mapping for the
4781 * populated node map.
4783 * Returns the determined alignment in pfn's. 0 if there is no alignment
4784 * requirement (single node).
4786 unsigned long __init node_map_pfn_alignment(void)
4788 unsigned long accl_mask = 0, last_end = 0;
4789 unsigned long start, end, mask;
4793 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4794 if (!start || last_nid < 0 || last_nid == nid) {
4801 * Start with a mask granular enough to pin-point to the
4802 * start pfn and tick off bits one-by-one until it becomes
4803 * too coarse to separate the current node from the last.
4805 mask = ~((1 << __ffs(start)) - 1);
4806 while (mask && last_end <= (start & (mask << 1)))
4809 /* accumulate all internode masks */
4813 /* convert mask to number of pages */
4814 return ~accl_mask + 1;
4817 /* Find the lowest pfn for a node */
4818 static unsigned long __init find_min_pfn_for_node(int nid)
4820 unsigned long min_pfn = ULONG_MAX;
4821 unsigned long start_pfn;
4824 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4825 min_pfn = min(min_pfn, start_pfn);
4827 if (min_pfn == ULONG_MAX) {
4829 "Could not find start_pfn for node %d\n", nid);
4837 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4839 * It returns the minimum PFN based on information provided via
4840 * add_active_range().
4842 unsigned long __init find_min_pfn_with_active_regions(void)
4844 return find_min_pfn_for_node(MAX_NUMNODES);
4848 * early_calculate_totalpages()
4849 * Sum pages in active regions for movable zone.
4850 * Populate N_MEMORY for calculating usable_nodes.
4852 static unsigned long __init early_calculate_totalpages(void)
4854 unsigned long totalpages = 0;
4855 unsigned long start_pfn, end_pfn;
4858 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4859 unsigned long pages = end_pfn - start_pfn;
4861 totalpages += pages;
4863 node_set_state(nid, N_MEMORY);
4869 * Find the PFN the Movable zone begins in each node. Kernel memory
4870 * is spread evenly between nodes as long as the nodes have enough
4871 * memory. When they don't, some nodes will have more kernelcore than
4874 static void __init find_zone_movable_pfns_for_nodes(void)
4877 unsigned long usable_startpfn;
4878 unsigned long kernelcore_node, kernelcore_remaining;
4879 /* save the state before borrow the nodemask */
4880 nodemask_t saved_node_state = node_states[N_MEMORY];
4881 unsigned long totalpages = early_calculate_totalpages();
4882 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4885 * If movablecore was specified, calculate what size of
4886 * kernelcore that corresponds so that memory usable for
4887 * any allocation type is evenly spread. If both kernelcore
4888 * and movablecore are specified, then the value of kernelcore
4889 * will be used for required_kernelcore if it's greater than
4890 * what movablecore would have allowed.
4892 if (required_movablecore) {
4893 unsigned long corepages;
4896 * Round-up so that ZONE_MOVABLE is at least as large as what
4897 * was requested by the user
4899 required_movablecore =
4900 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4901 corepages = totalpages - required_movablecore;
4903 required_kernelcore = max(required_kernelcore, corepages);
4906 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4907 if (!required_kernelcore)
4910 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4911 find_usable_zone_for_movable();
4912 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4915 /* Spread kernelcore memory as evenly as possible throughout nodes */
4916 kernelcore_node = required_kernelcore / usable_nodes;
4917 for_each_node_state(nid, N_MEMORY) {
4918 unsigned long start_pfn, end_pfn;
4921 * Recalculate kernelcore_node if the division per node
4922 * now exceeds what is necessary to satisfy the requested
4923 * amount of memory for the kernel
4925 if (required_kernelcore < kernelcore_node)
4926 kernelcore_node = required_kernelcore / usable_nodes;
4929 * As the map is walked, we track how much memory is usable
4930 * by the kernel using kernelcore_remaining. When it is
4931 * 0, the rest of the node is usable by ZONE_MOVABLE
4933 kernelcore_remaining = kernelcore_node;
4935 /* Go through each range of PFNs within this node */
4936 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4937 unsigned long size_pages;
4939 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4940 if (start_pfn >= end_pfn)
4943 /* Account for what is only usable for kernelcore */
4944 if (start_pfn < usable_startpfn) {
4945 unsigned long kernel_pages;
4946 kernel_pages = min(end_pfn, usable_startpfn)
4949 kernelcore_remaining -= min(kernel_pages,
4950 kernelcore_remaining);
4951 required_kernelcore -= min(kernel_pages,
4952 required_kernelcore);
4954 /* Continue if range is now fully accounted */
4955 if (end_pfn <= usable_startpfn) {
4958 * Push zone_movable_pfn to the end so
4959 * that if we have to rebalance
4960 * kernelcore across nodes, we will
4961 * not double account here
4963 zone_movable_pfn[nid] = end_pfn;
4966 start_pfn = usable_startpfn;
4970 * The usable PFN range for ZONE_MOVABLE is from
4971 * start_pfn->end_pfn. Calculate size_pages as the
4972 * number of pages used as kernelcore
4974 size_pages = end_pfn - start_pfn;
4975 if (size_pages > kernelcore_remaining)
4976 size_pages = kernelcore_remaining;
4977 zone_movable_pfn[nid] = start_pfn + size_pages;
4980 * Some kernelcore has been met, update counts and
4981 * break if the kernelcore for this node has been
4984 required_kernelcore -= min(required_kernelcore,
4986 kernelcore_remaining -= size_pages;
4987 if (!kernelcore_remaining)
4993 * If there is still required_kernelcore, we do another pass with one
4994 * less node in the count. This will push zone_movable_pfn[nid] further
4995 * along on the nodes that still have memory until kernelcore is
4999 if (usable_nodes && required_kernelcore > usable_nodes)
5002 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5003 for (nid = 0; nid < MAX_NUMNODES; nid++)
5004 zone_movable_pfn[nid] =
5005 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5008 /* restore the node_state */
5009 node_states[N_MEMORY] = saved_node_state;
5012 /* Any regular or high memory on that node ? */
5013 static void check_for_memory(pg_data_t *pgdat, int nid)
5015 enum zone_type zone_type;
5017 if (N_MEMORY == N_NORMAL_MEMORY)
5020 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5021 struct zone *zone = &pgdat->node_zones[zone_type];
5022 if (zone->present_pages) {
5023 node_set_state(nid, N_HIGH_MEMORY);
5024 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5025 zone_type <= ZONE_NORMAL)
5026 node_set_state(nid, N_NORMAL_MEMORY);
5033 * free_area_init_nodes - Initialise all pg_data_t and zone data
5034 * @max_zone_pfn: an array of max PFNs for each zone
5036 * This will call free_area_init_node() for each active node in the system.
5037 * Using the page ranges provided by add_active_range(), the size of each
5038 * zone in each node and their holes is calculated. If the maximum PFN
5039 * between two adjacent zones match, it is assumed that the zone is empty.
5040 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5041 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5042 * starts where the previous one ended. For example, ZONE_DMA32 starts
5043 * at arch_max_dma_pfn.
5045 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5047 unsigned long start_pfn, end_pfn;
5050 /* Record where the zone boundaries are */
5051 memset(arch_zone_lowest_possible_pfn, 0,
5052 sizeof(arch_zone_lowest_possible_pfn));
5053 memset(arch_zone_highest_possible_pfn, 0,
5054 sizeof(arch_zone_highest_possible_pfn));
5055 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5056 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5057 for (i = 1; i < MAX_NR_ZONES; i++) {
5058 if (i == ZONE_MOVABLE)
5060 arch_zone_lowest_possible_pfn[i] =
5061 arch_zone_highest_possible_pfn[i-1];
5062 arch_zone_highest_possible_pfn[i] =
5063 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5065 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5066 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5068 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5069 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5070 find_zone_movable_pfns_for_nodes();
5072 /* Print out the zone ranges */
5073 printk("Zone ranges:\n");
5074 for (i = 0; i < MAX_NR_ZONES; i++) {
5075 if (i == ZONE_MOVABLE)
5077 printk(KERN_CONT " %-8s ", zone_names[i]);
5078 if (arch_zone_lowest_possible_pfn[i] ==
5079 arch_zone_highest_possible_pfn[i])
5080 printk(KERN_CONT "empty\n");
5082 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5083 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5084 (arch_zone_highest_possible_pfn[i]
5085 << PAGE_SHIFT) - 1);
5088 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5089 printk("Movable zone start for each node\n");
5090 for (i = 0; i < MAX_NUMNODES; i++) {
5091 if (zone_movable_pfn[i])
5092 printk(" Node %d: %#010lx\n", i,
5093 zone_movable_pfn[i] << PAGE_SHIFT);
5096 /* Print out the early node map */
5097 printk("Early memory node ranges\n");
5098 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5099 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5100 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5102 /* Initialise every node */
5103 mminit_verify_pageflags_layout();
5104 setup_nr_node_ids();
5105 for_each_online_node(nid) {
5106 pg_data_t *pgdat = NODE_DATA(nid);
5107 free_area_init_node(nid, NULL,
5108 find_min_pfn_for_node(nid), NULL);
5110 /* Any memory on that node */
5111 if (pgdat->node_present_pages)
5112 node_set_state(nid, N_MEMORY);
5113 check_for_memory(pgdat, nid);
5117 static int __init cmdline_parse_core(char *p, unsigned long *core)
5119 unsigned long long coremem;
5123 coremem = memparse(p, &p);
5124 *core = coremem >> PAGE_SHIFT;
5126 /* Paranoid check that UL is enough for the coremem value */
5127 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5133 * kernelcore=size sets the amount of memory for use for allocations that
5134 * cannot be reclaimed or migrated.
5136 static int __init cmdline_parse_kernelcore(char *p)
5138 return cmdline_parse_core(p, &required_kernelcore);
5142 * movablecore=size sets the amount of memory for use for allocations that
5143 * can be reclaimed or migrated.
5145 static int __init cmdline_parse_movablecore(char *p)
5147 return cmdline_parse_core(p, &required_movablecore);
5150 early_param("kernelcore", cmdline_parse_kernelcore);
5151 early_param("movablecore", cmdline_parse_movablecore);
5153 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5155 unsigned long free_reserved_area(unsigned long start, unsigned long end,
5156 int poison, char *s)
5158 unsigned long pages, pos;
5160 pos = start = PAGE_ALIGN(start);
5162 for (pages = 0; pos < end; pos += PAGE_SIZE, pages++) {
5164 memset((void *)pos, poison, PAGE_SIZE);
5165 free_reserved_page(virt_to_page((void *)pos));
5169 pr_info("Freeing %s memory: %ldK (%lx - %lx)\n",
5170 s, pages << (PAGE_SHIFT - 10), start, end);
5175 #ifdef CONFIG_HIGHMEM
5176 void free_highmem_page(struct page *page)
5178 __free_reserved_page(page);
5185 * set_dma_reserve - set the specified number of pages reserved in the first zone
5186 * @new_dma_reserve: The number of pages to mark reserved
5188 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5189 * In the DMA zone, a significant percentage may be consumed by kernel image
5190 * and other unfreeable allocations which can skew the watermarks badly. This
5191 * function may optionally be used to account for unfreeable pages in the
5192 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5193 * smaller per-cpu batchsize.
5195 void __init set_dma_reserve(unsigned long new_dma_reserve)
5197 dma_reserve = new_dma_reserve;
5200 void __init free_area_init(unsigned long *zones_size)
5202 free_area_init_node(0, zones_size,
5203 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5206 static int page_alloc_cpu_notify(struct notifier_block *self,
5207 unsigned long action, void *hcpu)
5209 int cpu = (unsigned long)hcpu;
5211 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5212 lru_add_drain_cpu(cpu);
5216 * Spill the event counters of the dead processor
5217 * into the current processors event counters.
5218 * This artificially elevates the count of the current
5221 vm_events_fold_cpu(cpu);
5224 * Zero the differential counters of the dead processor
5225 * so that the vm statistics are consistent.
5227 * This is only okay since the processor is dead and cannot
5228 * race with what we are doing.
5230 refresh_cpu_vm_stats(cpu);
5235 void __init page_alloc_init(void)
5237 hotcpu_notifier(page_alloc_cpu_notify, 0);
5241 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5242 * or min_free_kbytes changes.
5244 static void calculate_totalreserve_pages(void)
5246 struct pglist_data *pgdat;
5247 unsigned long reserve_pages = 0;
5248 enum zone_type i, j;
5250 for_each_online_pgdat(pgdat) {
5251 for (i = 0; i < MAX_NR_ZONES; i++) {
5252 struct zone *zone = pgdat->node_zones + i;
5253 unsigned long max = 0;
5255 /* Find valid and maximum lowmem_reserve in the zone */
5256 for (j = i; j < MAX_NR_ZONES; j++) {
5257 if (zone->lowmem_reserve[j] > max)
5258 max = zone->lowmem_reserve[j];
5261 /* we treat the high watermark as reserved pages. */
5262 max += high_wmark_pages(zone);
5264 if (max > zone->managed_pages)
5265 max = zone->managed_pages;
5266 reserve_pages += max;
5268 * Lowmem reserves are not available to
5269 * GFP_HIGHUSER page cache allocations and
5270 * kswapd tries to balance zones to their high
5271 * watermark. As a result, neither should be
5272 * regarded as dirtyable memory, to prevent a
5273 * situation where reclaim has to clean pages
5274 * in order to balance the zones.
5276 zone->dirty_balance_reserve = max;
5279 dirty_balance_reserve = reserve_pages;
5280 totalreserve_pages = reserve_pages;
5284 * setup_per_zone_lowmem_reserve - called whenever
5285 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5286 * has a correct pages reserved value, so an adequate number of
5287 * pages are left in the zone after a successful __alloc_pages().
5289 static void setup_per_zone_lowmem_reserve(void)
5291 struct pglist_data *pgdat;
5292 enum zone_type j, idx;
5294 for_each_online_pgdat(pgdat) {
5295 for (j = 0; j < MAX_NR_ZONES; j++) {
5296 struct zone *zone = pgdat->node_zones + j;
5297 unsigned long managed_pages = zone->managed_pages;
5299 zone->lowmem_reserve[j] = 0;
5303 struct zone *lower_zone;
5307 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5308 sysctl_lowmem_reserve_ratio[idx] = 1;
5310 lower_zone = pgdat->node_zones + idx;
5311 lower_zone->lowmem_reserve[j] = managed_pages /
5312 sysctl_lowmem_reserve_ratio[idx];
5313 managed_pages += lower_zone->managed_pages;
5318 /* update totalreserve_pages */
5319 calculate_totalreserve_pages();
5322 static void __setup_per_zone_wmarks(void)
5324 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5325 unsigned long lowmem_pages = 0;
5327 unsigned long flags;
5329 /* Calculate total number of !ZONE_HIGHMEM pages */
5330 for_each_zone(zone) {
5331 if (!is_highmem(zone))
5332 lowmem_pages += zone->managed_pages;
5335 for_each_zone(zone) {
5338 spin_lock_irqsave(&zone->lock, flags);
5339 tmp = (u64)pages_min * zone->managed_pages;
5340 do_div(tmp, lowmem_pages);
5341 if (is_highmem(zone)) {
5343 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5344 * need highmem pages, so cap pages_min to a small
5347 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5348 * deltas controls asynch page reclaim, and so should
5349 * not be capped for highmem.
5351 unsigned long min_pages;
5353 min_pages = zone->managed_pages / 1024;
5354 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5355 zone->watermark[WMARK_MIN] = min_pages;
5358 * If it's a lowmem zone, reserve a number of pages
5359 * proportionate to the zone's size.
5361 zone->watermark[WMARK_MIN] = tmp;
5364 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5365 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5367 setup_zone_migrate_reserve(zone);
5368 spin_unlock_irqrestore(&zone->lock, flags);
5371 /* update totalreserve_pages */
5372 calculate_totalreserve_pages();
5376 * setup_per_zone_wmarks - called when min_free_kbytes changes
5377 * or when memory is hot-{added|removed}
5379 * Ensures that the watermark[min,low,high] values for each zone are set
5380 * correctly with respect to min_free_kbytes.
5382 void setup_per_zone_wmarks(void)
5384 mutex_lock(&zonelists_mutex);
5385 __setup_per_zone_wmarks();
5386 mutex_unlock(&zonelists_mutex);
5390 * The inactive anon list should be small enough that the VM never has to
5391 * do too much work, but large enough that each inactive page has a chance
5392 * to be referenced again before it is swapped out.
5394 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5395 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5396 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5397 * the anonymous pages are kept on the inactive list.
5400 * memory ratio inactive anon
5401 * -------------------------------------
5410 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5412 unsigned int gb, ratio;
5414 /* Zone size in gigabytes */
5415 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5417 ratio = int_sqrt(10 * gb);
5421 zone->inactive_ratio = ratio;
5424 static void __meminit setup_per_zone_inactive_ratio(void)
5429 calculate_zone_inactive_ratio(zone);
5433 * Initialise min_free_kbytes.
5435 * For small machines we want it small (128k min). For large machines
5436 * we want it large (64MB max). But it is not linear, because network
5437 * bandwidth does not increase linearly with machine size. We use
5439 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5440 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5456 int __meminit init_per_zone_wmark_min(void)
5458 unsigned long lowmem_kbytes;
5460 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5462 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5463 if (min_free_kbytes < 128)
5464 min_free_kbytes = 128;
5465 if (min_free_kbytes > 65536)
5466 min_free_kbytes = 65536;
5467 setup_per_zone_wmarks();
5468 refresh_zone_stat_thresholds();
5469 setup_per_zone_lowmem_reserve();
5470 setup_per_zone_inactive_ratio();
5473 module_init(init_per_zone_wmark_min)
5476 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5477 * that we can call two helper functions whenever min_free_kbytes
5480 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5481 void __user *buffer, size_t *length, loff_t *ppos)
5483 proc_dointvec(table, write, buffer, length, ppos);
5485 setup_per_zone_wmarks();
5490 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5491 void __user *buffer, size_t *length, loff_t *ppos)
5496 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5501 zone->min_unmapped_pages = (zone->managed_pages *
5502 sysctl_min_unmapped_ratio) / 100;
5506 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5507 void __user *buffer, size_t *length, loff_t *ppos)
5512 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5517 zone->min_slab_pages = (zone->managed_pages *
5518 sysctl_min_slab_ratio) / 100;
5524 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5525 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5526 * whenever sysctl_lowmem_reserve_ratio changes.
5528 * The reserve ratio obviously has absolutely no relation with the
5529 * minimum watermarks. The lowmem reserve ratio can only make sense
5530 * if in function of the boot time zone sizes.
5532 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5533 void __user *buffer, size_t *length, loff_t *ppos)
5535 proc_dointvec_minmax(table, write, buffer, length, ppos);
5536 setup_per_zone_lowmem_reserve();
5541 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5542 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5543 * can have before it gets flushed back to buddy allocator.
5546 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5547 void __user *buffer, size_t *length, loff_t *ppos)
5553 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5554 if (!write || (ret < 0))
5556 for_each_populated_zone(zone) {
5557 for_each_possible_cpu(cpu) {
5559 high = zone->managed_pages / percpu_pagelist_fraction;
5560 setup_pagelist_highmark(
5561 per_cpu_ptr(zone->pageset, cpu), high);
5567 int hashdist = HASHDIST_DEFAULT;
5570 static int __init set_hashdist(char *str)
5574 hashdist = simple_strtoul(str, &str, 0);
5577 __setup("hashdist=", set_hashdist);
5581 * allocate a large system hash table from bootmem
5582 * - it is assumed that the hash table must contain an exact power-of-2
5583 * quantity of entries
5584 * - limit is the number of hash buckets, not the total allocation size
5586 void *__init alloc_large_system_hash(const char *tablename,
5587 unsigned long bucketsize,
5588 unsigned long numentries,
5591 unsigned int *_hash_shift,
5592 unsigned int *_hash_mask,
5593 unsigned long low_limit,
5594 unsigned long high_limit)
5596 unsigned long long max = high_limit;
5597 unsigned long log2qty, size;
5600 /* allow the kernel cmdline to have a say */
5602 /* round applicable memory size up to nearest megabyte */
5603 numentries = nr_kernel_pages;
5604 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5605 numentries >>= 20 - PAGE_SHIFT;
5606 numentries <<= 20 - PAGE_SHIFT;
5608 /* limit to 1 bucket per 2^scale bytes of low memory */
5609 if (scale > PAGE_SHIFT)
5610 numentries >>= (scale - PAGE_SHIFT);
5612 numentries <<= (PAGE_SHIFT - scale);
5614 /* Make sure we've got at least a 0-order allocation.. */
5615 if (unlikely(flags & HASH_SMALL)) {
5616 /* Makes no sense without HASH_EARLY */
5617 WARN_ON(!(flags & HASH_EARLY));
5618 if (!(numentries >> *_hash_shift)) {
5619 numentries = 1UL << *_hash_shift;
5620 BUG_ON(!numentries);
5622 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5623 numentries = PAGE_SIZE / bucketsize;
5625 numentries = roundup_pow_of_two(numentries);
5627 /* limit allocation size to 1/16 total memory by default */
5629 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5630 do_div(max, bucketsize);
5632 max = min(max, 0x80000000ULL);
5634 if (numentries < low_limit)
5635 numentries = low_limit;
5636 if (numentries > max)
5639 log2qty = ilog2(numentries);
5642 size = bucketsize << log2qty;
5643 if (flags & HASH_EARLY)
5644 table = alloc_bootmem_nopanic(size);
5646 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5649 * If bucketsize is not a power-of-two, we may free
5650 * some pages at the end of hash table which
5651 * alloc_pages_exact() automatically does
5653 if (get_order(size) < MAX_ORDER) {
5654 table = alloc_pages_exact(size, GFP_ATOMIC);
5655 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5658 } while (!table && size > PAGE_SIZE && --log2qty);
5661 panic("Failed to allocate %s hash table\n", tablename);
5663 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5666 ilog2(size) - PAGE_SHIFT,
5670 *_hash_shift = log2qty;
5672 *_hash_mask = (1 << log2qty) - 1;
5677 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5678 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5681 #ifdef CONFIG_SPARSEMEM
5682 return __pfn_to_section(pfn)->pageblock_flags;
5684 return zone->pageblock_flags;
5685 #endif /* CONFIG_SPARSEMEM */
5688 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5690 #ifdef CONFIG_SPARSEMEM
5691 pfn &= (PAGES_PER_SECTION-1);
5692 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5694 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5695 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5696 #endif /* CONFIG_SPARSEMEM */
5700 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5701 * @page: The page within the block of interest
5702 * @start_bitidx: The first bit of interest to retrieve
5703 * @end_bitidx: The last bit of interest
5704 * returns pageblock_bits flags
5706 unsigned long get_pageblock_flags_group(struct page *page,
5707 int start_bitidx, int end_bitidx)
5710 unsigned long *bitmap;
5711 unsigned long pfn, bitidx;
5712 unsigned long flags = 0;
5713 unsigned long value = 1;
5715 zone = page_zone(page);
5716 pfn = page_to_pfn(page);
5717 bitmap = get_pageblock_bitmap(zone, pfn);
5718 bitidx = pfn_to_bitidx(zone, pfn);
5720 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5721 if (test_bit(bitidx + start_bitidx, bitmap))
5728 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5729 * @page: The page within the block of interest
5730 * @start_bitidx: The first bit of interest
5731 * @end_bitidx: The last bit of interest
5732 * @flags: The flags to set
5734 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5735 int start_bitidx, int end_bitidx)
5738 unsigned long *bitmap;
5739 unsigned long pfn, bitidx;
5740 unsigned long value = 1;
5742 zone = page_zone(page);
5743 pfn = page_to_pfn(page);
5744 bitmap = get_pageblock_bitmap(zone, pfn);
5745 bitidx = pfn_to_bitidx(zone, pfn);
5746 VM_BUG_ON(!zone_spans_pfn(zone, pfn));
5748 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5750 __set_bit(bitidx + start_bitidx, bitmap);
5752 __clear_bit(bitidx + start_bitidx, bitmap);
5756 * This function checks whether pageblock includes unmovable pages or not.
5757 * If @count is not zero, it is okay to include less @count unmovable pages
5759 * PageLRU check wihtout isolation or lru_lock could race so that
5760 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5761 * expect this function should be exact.
5763 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5764 bool skip_hwpoisoned_pages)
5766 unsigned long pfn, iter, found;
5770 * For avoiding noise data, lru_add_drain_all() should be called
5771 * If ZONE_MOVABLE, the zone never contains unmovable pages
5773 if (zone_idx(zone) == ZONE_MOVABLE)
5775 mt = get_pageblock_migratetype(page);
5776 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5779 pfn = page_to_pfn(page);
5780 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5781 unsigned long check = pfn + iter;
5783 if (!pfn_valid_within(check))
5786 page = pfn_to_page(check);
5788 * We can't use page_count without pin a page
5789 * because another CPU can free compound page.
5790 * This check already skips compound tails of THP
5791 * because their page->_count is zero at all time.
5793 if (!atomic_read(&page->_count)) {
5794 if (PageBuddy(page))
5795 iter += (1 << page_order(page)) - 1;
5800 * The HWPoisoned page may be not in buddy system, and
5801 * page_count() is not 0.
5803 if (skip_hwpoisoned_pages && PageHWPoison(page))
5809 * If there are RECLAIMABLE pages, we need to check it.
5810 * But now, memory offline itself doesn't call shrink_slab()
5811 * and it still to be fixed.
5814 * If the page is not RAM, page_count()should be 0.
5815 * we don't need more check. This is an _used_ not-movable page.
5817 * The problematic thing here is PG_reserved pages. PG_reserved
5818 * is set to both of a memory hole page and a _used_ kernel
5827 bool is_pageblock_removable_nolock(struct page *page)
5833 * We have to be careful here because we are iterating over memory
5834 * sections which are not zone aware so we might end up outside of
5835 * the zone but still within the section.
5836 * We have to take care about the node as well. If the node is offline
5837 * its NODE_DATA will be NULL - see page_zone.
5839 if (!node_online(page_to_nid(page)))
5842 zone = page_zone(page);
5843 pfn = page_to_pfn(page);
5844 if (!zone_spans_pfn(zone, pfn))
5847 return !has_unmovable_pages(zone, page, 0, true);
5852 static unsigned long pfn_max_align_down(unsigned long pfn)
5854 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5855 pageblock_nr_pages) - 1);
5858 static unsigned long pfn_max_align_up(unsigned long pfn)
5860 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5861 pageblock_nr_pages));
5864 /* [start, end) must belong to a single zone. */
5865 static int __alloc_contig_migrate_range(struct compact_control *cc,
5866 unsigned long start, unsigned long end)
5868 /* This function is based on compact_zone() from compaction.c. */
5869 unsigned long nr_reclaimed;
5870 unsigned long pfn = start;
5871 unsigned int tries = 0;
5876 while (pfn < end || !list_empty(&cc->migratepages)) {
5877 if (fatal_signal_pending(current)) {
5882 if (list_empty(&cc->migratepages)) {
5883 cc->nr_migratepages = 0;
5884 pfn = isolate_migratepages_range(cc->zone, cc,
5891 } else if (++tries == 5) {
5892 ret = ret < 0 ? ret : -EBUSY;
5896 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5898 cc->nr_migratepages -= nr_reclaimed;
5900 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
5901 0, MIGRATE_SYNC, MR_CMA);
5904 putback_movable_pages(&cc->migratepages);
5911 * alloc_contig_range() -- tries to allocate given range of pages
5912 * @start: start PFN to allocate
5913 * @end: one-past-the-last PFN to allocate
5914 * @migratetype: migratetype of the underlaying pageblocks (either
5915 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5916 * in range must have the same migratetype and it must
5917 * be either of the two.
5919 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5920 * aligned, however it's the caller's responsibility to guarantee that
5921 * we are the only thread that changes migrate type of pageblocks the
5924 * The PFN range must belong to a single zone.
5926 * Returns zero on success or negative error code. On success all
5927 * pages which PFN is in [start, end) are allocated for the caller and
5928 * need to be freed with free_contig_range().
5930 int alloc_contig_range(unsigned long start, unsigned long end,
5931 unsigned migratetype)
5933 unsigned long outer_start, outer_end;
5936 struct compact_control cc = {
5937 .nr_migratepages = 0,
5939 .zone = page_zone(pfn_to_page(start)),
5941 .ignore_skip_hint = true,
5943 INIT_LIST_HEAD(&cc.migratepages);
5946 * What we do here is we mark all pageblocks in range as
5947 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5948 * have different sizes, and due to the way page allocator
5949 * work, we align the range to biggest of the two pages so
5950 * that page allocator won't try to merge buddies from
5951 * different pageblocks and change MIGRATE_ISOLATE to some
5952 * other migration type.
5954 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5955 * migrate the pages from an unaligned range (ie. pages that
5956 * we are interested in). This will put all the pages in
5957 * range back to page allocator as MIGRATE_ISOLATE.
5959 * When this is done, we take the pages in range from page
5960 * allocator removing them from the buddy system. This way
5961 * page allocator will never consider using them.
5963 * This lets us mark the pageblocks back as
5964 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5965 * aligned range but not in the unaligned, original range are
5966 * put back to page allocator so that buddy can use them.
5969 ret = start_isolate_page_range(pfn_max_align_down(start),
5970 pfn_max_align_up(end), migratetype,
5975 ret = __alloc_contig_migrate_range(&cc, start, end);
5980 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5981 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5982 * more, all pages in [start, end) are free in page allocator.
5983 * What we are going to do is to allocate all pages from
5984 * [start, end) (that is remove them from page allocator).
5986 * The only problem is that pages at the beginning and at the
5987 * end of interesting range may be not aligned with pages that
5988 * page allocator holds, ie. they can be part of higher order
5989 * pages. Because of this, we reserve the bigger range and
5990 * once this is done free the pages we are not interested in.
5992 * We don't have to hold zone->lock here because the pages are
5993 * isolated thus they won't get removed from buddy.
5996 lru_add_drain_all();
6000 outer_start = start;
6001 while (!PageBuddy(pfn_to_page(outer_start))) {
6002 if (++order >= MAX_ORDER) {
6006 outer_start &= ~0UL << order;
6009 /* Make sure the range is really isolated. */
6010 if (test_pages_isolated(outer_start, end, false)) {
6011 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6018 /* Grab isolated pages from freelists. */
6019 outer_end = isolate_freepages_range(&cc, outer_start, end);
6025 /* Free head and tail (if any) */
6026 if (start != outer_start)
6027 free_contig_range(outer_start, start - outer_start);
6028 if (end != outer_end)
6029 free_contig_range(end, outer_end - end);
6032 undo_isolate_page_range(pfn_max_align_down(start),
6033 pfn_max_align_up(end), migratetype);
6037 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6039 unsigned int count = 0;
6041 for (; nr_pages--; pfn++) {
6042 struct page *page = pfn_to_page(pfn);
6044 count += page_count(page) != 1;
6047 WARN(count != 0, "%d pages are still in use!\n", count);
6051 #ifdef CONFIG_MEMORY_HOTPLUG
6052 static int __meminit __zone_pcp_update(void *data)
6054 struct zone *zone = data;
6056 unsigned long batch = zone_batchsize(zone), flags;
6058 for_each_possible_cpu(cpu) {
6059 struct per_cpu_pageset *pset;
6060 struct per_cpu_pages *pcp;
6062 pset = per_cpu_ptr(zone->pageset, cpu);
6065 local_irq_save(flags);
6067 free_pcppages_bulk(zone, pcp->count, pcp);
6068 drain_zonestat(zone, pset);
6069 setup_pageset(pset, batch);
6070 local_irq_restore(flags);
6075 void __meminit zone_pcp_update(struct zone *zone)
6077 stop_machine(__zone_pcp_update, zone, NULL);
6081 void zone_pcp_reset(struct zone *zone)
6083 unsigned long flags;
6085 struct per_cpu_pageset *pset;
6087 /* avoid races with drain_pages() */
6088 local_irq_save(flags);
6089 if (zone->pageset != &boot_pageset) {
6090 for_each_online_cpu(cpu) {
6091 pset = per_cpu_ptr(zone->pageset, cpu);
6092 drain_zonestat(zone, pset);
6094 free_percpu(zone->pageset);
6095 zone->pageset = &boot_pageset;
6097 local_irq_restore(flags);
6100 #ifdef CONFIG_MEMORY_HOTREMOVE
6102 * All pages in the range must be isolated before calling this.
6105 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6111 unsigned long flags;
6112 /* find the first valid pfn */
6113 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6118 zone = page_zone(pfn_to_page(pfn));
6119 spin_lock_irqsave(&zone->lock, flags);
6121 while (pfn < end_pfn) {
6122 if (!pfn_valid(pfn)) {
6126 page = pfn_to_page(pfn);
6128 * The HWPoisoned page may be not in buddy system, and
6129 * page_count() is not 0.
6131 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6133 SetPageReserved(page);
6137 BUG_ON(page_count(page));
6138 BUG_ON(!PageBuddy(page));
6139 order = page_order(page);
6140 #ifdef CONFIG_DEBUG_VM
6141 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6142 pfn, 1 << order, end_pfn);
6144 list_del(&page->lru);
6145 rmv_page_order(page);
6146 zone->free_area[order].nr_free--;
6147 #ifdef CONFIG_HIGHMEM
6148 if (PageHighMem(page))
6149 totalhigh_pages -= 1 << order;
6151 for (i = 0; i < (1 << order); i++)
6152 SetPageReserved((page+i));
6153 pfn += (1 << order);
6155 spin_unlock_irqrestore(&zone->lock, flags);
6159 #ifdef CONFIG_MEMORY_FAILURE
6160 bool is_free_buddy_page(struct page *page)
6162 struct zone *zone = page_zone(page);
6163 unsigned long pfn = page_to_pfn(page);
6164 unsigned long flags;
6167 spin_lock_irqsave(&zone->lock, flags);
6168 for (order = 0; order < MAX_ORDER; order++) {
6169 struct page *page_head = page - (pfn & ((1 << order) - 1));
6171 if (PageBuddy(page_head) && page_order(page_head) >= order)
6174 spin_unlock_irqrestore(&zone->lock, flags);
6176 return order < MAX_ORDER;
6180 static const struct trace_print_flags pageflag_names[] = {
6181 {1UL << PG_locked, "locked" },
6182 {1UL << PG_error, "error" },
6183 {1UL << PG_referenced, "referenced" },
6184 {1UL << PG_uptodate, "uptodate" },
6185 {1UL << PG_dirty, "dirty" },
6186 {1UL << PG_lru, "lru" },
6187 {1UL << PG_active, "active" },
6188 {1UL << PG_slab, "slab" },
6189 {1UL << PG_owner_priv_1, "owner_priv_1" },
6190 {1UL << PG_arch_1, "arch_1" },
6191 {1UL << PG_reserved, "reserved" },
6192 {1UL << PG_private, "private" },
6193 {1UL << PG_private_2, "private_2" },
6194 {1UL << PG_writeback, "writeback" },
6195 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6196 {1UL << PG_head, "head" },
6197 {1UL << PG_tail, "tail" },
6199 {1UL << PG_compound, "compound" },
6201 {1UL << PG_swapcache, "swapcache" },
6202 {1UL << PG_mappedtodisk, "mappedtodisk" },
6203 {1UL << PG_reclaim, "reclaim" },
6204 {1UL << PG_swapbacked, "swapbacked" },
6205 {1UL << PG_unevictable, "unevictable" },
6207 {1UL << PG_mlocked, "mlocked" },
6209 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6210 {1UL << PG_uncached, "uncached" },
6212 #ifdef CONFIG_MEMORY_FAILURE
6213 {1UL << PG_hwpoison, "hwpoison" },
6215 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6216 {1UL << PG_compound_lock, "compound_lock" },
6220 static void dump_page_flags(unsigned long flags)
6222 const char *delim = "";
6226 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6228 printk(KERN_ALERT "page flags: %#lx(", flags);
6230 /* remove zone id */
6231 flags &= (1UL << NR_PAGEFLAGS) - 1;
6233 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6235 mask = pageflag_names[i].mask;
6236 if ((flags & mask) != mask)
6240 printk("%s%s", delim, pageflag_names[i].name);
6244 /* check for left over flags */
6246 printk("%s%#lx", delim, flags);
6251 void dump_page(struct page *page)
6254 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6255 page, atomic_read(&page->_count), page_mapcount(page),
6256 page->mapping, page->index);
6257 dump_page_flags(page->flags);
6258 mem_cgroup_print_bad_page(page);