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>
62 #include <asm/tlbflush.h>
63 #include <asm/div64.h>
66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67 DEFINE_PER_CPU(int, numa_node);
68 EXPORT_PER_CPU_SYMBOL(numa_node);
71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
78 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
83 * Array of node states.
85 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86 [N_POSSIBLE] = NODE_MASK_ALL,
87 [N_ONLINE] = { { [0] = 1UL } },
89 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
91 [N_HIGH_MEMORY] = { { [0] = 1UL } },
93 #ifdef CONFIG_MOVABLE_NODE
94 [N_MEMORY] = { { [0] = 1UL } },
96 [N_CPU] = { { [0] = 1UL } },
99 EXPORT_SYMBOL(node_states);
101 unsigned long totalram_pages __read_mostly;
102 unsigned long totalreserve_pages __read_mostly;
104 * When calculating the number of globally allowed dirty pages, there
105 * is a certain number of per-zone reserves that should not be
106 * considered dirtyable memory. This is the sum of those reserves
107 * over all existing zones that contribute dirtyable memory.
109 unsigned long dirty_balance_reserve __read_mostly;
111 int percpu_pagelist_fraction;
112 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
114 #ifdef CONFIG_PM_SLEEP
116 * The following functions are used by the suspend/hibernate code to temporarily
117 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
118 * while devices are suspended. To avoid races with the suspend/hibernate code,
119 * they should always be called with pm_mutex held (gfp_allowed_mask also should
120 * only be modified with pm_mutex held, unless the suspend/hibernate code is
121 * guaranteed not to run in parallel with that modification).
124 static gfp_t saved_gfp_mask;
126 void pm_restore_gfp_mask(void)
128 WARN_ON(!mutex_is_locked(&pm_mutex));
129 if (saved_gfp_mask) {
130 gfp_allowed_mask = saved_gfp_mask;
135 void pm_restrict_gfp_mask(void)
137 WARN_ON(!mutex_is_locked(&pm_mutex));
138 WARN_ON(saved_gfp_mask);
139 saved_gfp_mask = gfp_allowed_mask;
140 gfp_allowed_mask &= ~GFP_IOFS;
143 bool pm_suspended_storage(void)
145 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
149 #endif /* CONFIG_PM_SLEEP */
151 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
152 int pageblock_order __read_mostly;
155 static void __free_pages_ok(struct page *page, unsigned int order);
158 * results with 256, 32 in the lowmem_reserve sysctl:
159 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
160 * 1G machine -> (16M dma, 784M normal, 224M high)
161 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
162 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
163 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
165 * TBD: should special case ZONE_DMA32 machines here - in those we normally
166 * don't need any ZONE_NORMAL reservation
168 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
169 #ifdef CONFIG_ZONE_DMA
172 #ifdef CONFIG_ZONE_DMA32
175 #ifdef CONFIG_HIGHMEM
181 EXPORT_SYMBOL(totalram_pages);
183 static char * const zone_names[MAX_NR_ZONES] = {
184 #ifdef CONFIG_ZONE_DMA
187 #ifdef CONFIG_ZONE_DMA32
191 #ifdef CONFIG_HIGHMEM
197 int min_free_kbytes = 1024;
199 static unsigned long __meminitdata nr_kernel_pages;
200 static unsigned long __meminitdata nr_all_pages;
201 static unsigned long __meminitdata dma_reserve;
203 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
204 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
205 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
206 static unsigned long __initdata required_kernelcore;
207 static unsigned long __initdata required_movablecore;
208 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
210 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
212 EXPORT_SYMBOL(movable_zone);
213 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
216 int nr_node_ids __read_mostly = MAX_NUMNODES;
217 int nr_online_nodes __read_mostly = 1;
218 EXPORT_SYMBOL(nr_node_ids);
219 EXPORT_SYMBOL(nr_online_nodes);
222 int page_group_by_mobility_disabled __read_mostly;
226 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
227 * Instead, use {un}set_pageblock_isolate.
229 void set_pageblock_migratetype(struct page *page, int migratetype)
232 if (unlikely(page_group_by_mobility_disabled))
233 migratetype = MIGRATE_UNMOVABLE;
235 set_pageblock_flags_group(page, (unsigned long)migratetype,
236 PB_migrate, PB_migrate_end);
239 bool oom_killer_disabled __read_mostly;
241 #ifdef CONFIG_DEBUG_VM
242 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
246 unsigned long pfn = page_to_pfn(page);
249 seq = zone_span_seqbegin(zone);
250 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
252 else if (pfn < zone->zone_start_pfn)
254 } while (zone_span_seqretry(zone, seq));
259 static int page_is_consistent(struct zone *zone, struct page *page)
261 if (!pfn_valid_within(page_to_pfn(page)))
263 if (zone != page_zone(page))
269 * Temporary debugging check for pages not lying within a given zone.
271 static int bad_range(struct zone *zone, struct page *page)
273 if (page_outside_zone_boundaries(zone, page))
275 if (!page_is_consistent(zone, page))
281 static inline int bad_range(struct zone *zone, struct page *page)
287 static void bad_page(struct page *page)
289 static unsigned long resume;
290 static unsigned long nr_shown;
291 static unsigned long nr_unshown;
293 /* Don't complain about poisoned pages */
294 if (PageHWPoison(page)) {
295 reset_page_mapcount(page); /* remove PageBuddy */
300 * Allow a burst of 60 reports, then keep quiet for that minute;
301 * or allow a steady drip of one report per second.
303 if (nr_shown == 60) {
304 if (time_before(jiffies, resume)) {
310 "BUG: Bad page state: %lu messages suppressed\n",
317 resume = jiffies + 60 * HZ;
319 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
320 current->comm, page_to_pfn(page));
326 /* Leave bad fields for debug, except PageBuddy could make trouble */
327 reset_page_mapcount(page); /* remove PageBuddy */
328 add_taint(TAINT_BAD_PAGE);
332 * Higher-order pages are called "compound pages". They are structured thusly:
334 * The first PAGE_SIZE page is called the "head page".
336 * The remaining PAGE_SIZE pages are called "tail pages".
338 * All pages have PG_compound set. All tail pages have their ->first_page
339 * pointing at the head page.
341 * The first tail page's ->lru.next holds the address of the compound page's
342 * put_page() function. Its ->lru.prev holds the order of allocation.
343 * This usage means that zero-order pages may not be compound.
346 static void free_compound_page(struct page *page)
348 __free_pages_ok(page, compound_order(page));
351 void prep_compound_page(struct page *page, unsigned long order)
354 int nr_pages = 1 << order;
356 set_compound_page_dtor(page, free_compound_page);
357 set_compound_order(page, order);
359 for (i = 1; i < nr_pages; i++) {
360 struct page *p = page + i;
362 set_page_count(p, 0);
363 p->first_page = page;
367 /* update __split_huge_page_refcount if you change this function */
368 static int destroy_compound_page(struct page *page, unsigned long order)
371 int nr_pages = 1 << order;
374 if (unlikely(compound_order(page) != order) ||
375 unlikely(!PageHead(page))) {
380 __ClearPageHead(page);
382 for (i = 1; i < nr_pages; i++) {
383 struct page *p = page + i;
385 if (unlikely(!PageTail(p) || (p->first_page != page))) {
395 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
400 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
401 * and __GFP_HIGHMEM from hard or soft interrupt context.
403 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
404 for (i = 0; i < (1 << order); i++)
405 clear_highpage(page + i);
408 #ifdef CONFIG_DEBUG_PAGEALLOC
409 unsigned int _debug_guardpage_minorder;
411 static int __init debug_guardpage_minorder_setup(char *buf)
415 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
416 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
419 _debug_guardpage_minorder = res;
420 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
423 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
425 static inline void set_page_guard_flag(struct page *page)
427 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
430 static inline void clear_page_guard_flag(struct page *page)
432 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
435 static inline void set_page_guard_flag(struct page *page) { }
436 static inline void clear_page_guard_flag(struct page *page) { }
439 static inline void set_page_order(struct page *page, int order)
441 set_page_private(page, order);
442 __SetPageBuddy(page);
445 static inline void rmv_page_order(struct page *page)
447 __ClearPageBuddy(page);
448 set_page_private(page, 0);
452 * Locate the struct page for both the matching buddy in our
453 * pair (buddy1) and the combined O(n+1) page they form (page).
455 * 1) Any buddy B1 will have an order O twin B2 which satisfies
456 * the following equation:
458 * For example, if the starting buddy (buddy2) is #8 its order
460 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
462 * 2) Any buddy B will have an order O+1 parent P which
463 * satisfies the following equation:
466 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
468 static inline unsigned long
469 __find_buddy_index(unsigned long page_idx, unsigned int order)
471 return page_idx ^ (1 << order);
475 * This function checks whether a page is free && is the buddy
476 * we can do coalesce a page and its buddy if
477 * (a) the buddy is not in a hole &&
478 * (b) the buddy is in the buddy system &&
479 * (c) a page and its buddy have the same order &&
480 * (d) a page and its buddy are in the same zone.
482 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
483 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
485 * For recording page's order, we use page_private(page).
487 static inline int page_is_buddy(struct page *page, struct page *buddy,
490 if (!pfn_valid_within(page_to_pfn(buddy)))
493 if (page_zone_id(page) != page_zone_id(buddy))
496 if (page_is_guard(buddy) && page_order(buddy) == order) {
497 VM_BUG_ON(page_count(buddy) != 0);
501 if (PageBuddy(buddy) && page_order(buddy) == order) {
502 VM_BUG_ON(page_count(buddy) != 0);
509 * Freeing function for a buddy system allocator.
511 * The concept of a buddy system is to maintain direct-mapped table
512 * (containing bit values) for memory blocks of various "orders".
513 * The bottom level table contains the map for the smallest allocatable
514 * units of memory (here, pages), and each level above it describes
515 * pairs of units from the levels below, hence, "buddies".
516 * At a high level, all that happens here is marking the table entry
517 * at the bottom level available, and propagating the changes upward
518 * as necessary, plus some accounting needed to play nicely with other
519 * parts of the VM system.
520 * At each level, we keep a list of pages, which are heads of continuous
521 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
522 * order is recorded in page_private(page) field.
523 * So when we are allocating or freeing one, we can derive the state of the
524 * other. That is, if we allocate a small block, and both were
525 * free, the remainder of the region must be split into blocks.
526 * If a block is freed, and its buddy is also free, then this
527 * triggers coalescing into a block of larger size.
532 static inline void __free_one_page(struct page *page,
533 struct zone *zone, unsigned int order,
536 unsigned long page_idx;
537 unsigned long combined_idx;
538 unsigned long uninitialized_var(buddy_idx);
541 if (unlikely(PageCompound(page)))
542 if (unlikely(destroy_compound_page(page, order)))
545 VM_BUG_ON(migratetype == -1);
547 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
549 VM_BUG_ON(page_idx & ((1 << order) - 1));
550 VM_BUG_ON(bad_range(zone, page));
552 while (order < MAX_ORDER-1) {
553 buddy_idx = __find_buddy_index(page_idx, order);
554 buddy = page + (buddy_idx - page_idx);
555 if (!page_is_buddy(page, buddy, order))
558 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
559 * merge with it and move up one order.
561 if (page_is_guard(buddy)) {
562 clear_page_guard_flag(buddy);
563 set_page_private(page, 0);
564 __mod_zone_freepage_state(zone, 1 << order,
567 list_del(&buddy->lru);
568 zone->free_area[order].nr_free--;
569 rmv_page_order(buddy);
571 combined_idx = buddy_idx & page_idx;
572 page = page + (combined_idx - page_idx);
573 page_idx = combined_idx;
576 set_page_order(page, order);
579 * If this is not the largest possible page, check if the buddy
580 * of the next-highest order is free. If it is, it's possible
581 * that pages are being freed that will coalesce soon. In case,
582 * that is happening, add the free page to the tail of the list
583 * so it's less likely to be used soon and more likely to be merged
584 * as a higher order page
586 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
587 struct page *higher_page, *higher_buddy;
588 combined_idx = buddy_idx & page_idx;
589 higher_page = page + (combined_idx - page_idx);
590 buddy_idx = __find_buddy_index(combined_idx, order + 1);
591 higher_buddy = higher_page + (buddy_idx - combined_idx);
592 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
593 list_add_tail(&page->lru,
594 &zone->free_area[order].free_list[migratetype]);
599 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
601 zone->free_area[order].nr_free++;
604 static inline int free_pages_check(struct page *page)
606 if (unlikely(page_mapcount(page) |
607 (page->mapping != NULL) |
608 (atomic_read(&page->_count) != 0) |
609 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
610 (mem_cgroup_bad_page_check(page)))) {
614 reset_page_last_nid(page);
615 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
616 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
621 * Frees a number of pages from the PCP lists
622 * Assumes all pages on list are in same zone, and of same order.
623 * count is the number of pages to free.
625 * If the zone was previously in an "all pages pinned" state then look to
626 * see if this freeing clears that state.
628 * And clear the zone's pages_scanned counter, to hold off the "all pages are
629 * pinned" detection logic.
631 static void free_pcppages_bulk(struct zone *zone, int count,
632 struct per_cpu_pages *pcp)
638 spin_lock(&zone->lock);
639 zone->all_unreclaimable = 0;
640 zone->pages_scanned = 0;
644 struct list_head *list;
647 * Remove pages from lists in a round-robin fashion. A
648 * batch_free count is maintained that is incremented when an
649 * empty list is encountered. This is so more pages are freed
650 * off fuller lists instead of spinning excessively around empty
655 if (++migratetype == MIGRATE_PCPTYPES)
657 list = &pcp->lists[migratetype];
658 } while (list_empty(list));
660 /* This is the only non-empty list. Free them all. */
661 if (batch_free == MIGRATE_PCPTYPES)
662 batch_free = to_free;
665 int mt; /* migratetype of the to-be-freed page */
667 page = list_entry(list->prev, struct page, lru);
668 /* must delete as __free_one_page list manipulates */
669 list_del(&page->lru);
670 mt = get_freepage_migratetype(page);
671 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
672 __free_one_page(page, zone, 0, mt);
673 trace_mm_page_pcpu_drain(page, 0, mt);
674 if (likely(get_pageblock_migratetype(page) != MIGRATE_ISOLATE)) {
675 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
676 if (is_migrate_cma(mt))
677 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
679 } while (--to_free && --batch_free && !list_empty(list));
681 spin_unlock(&zone->lock);
684 static void free_one_page(struct zone *zone, struct page *page, int order,
687 spin_lock(&zone->lock);
688 zone->all_unreclaimable = 0;
689 zone->pages_scanned = 0;
691 __free_one_page(page, zone, order, migratetype);
692 if (unlikely(migratetype != MIGRATE_ISOLATE))
693 __mod_zone_freepage_state(zone, 1 << order, migratetype);
694 spin_unlock(&zone->lock);
697 static bool free_pages_prepare(struct page *page, unsigned int order)
702 trace_mm_page_free(page, order);
703 kmemcheck_free_shadow(page, order);
706 page->mapping = NULL;
707 for (i = 0; i < (1 << order); i++)
708 bad += free_pages_check(page + i);
712 if (!PageHighMem(page)) {
713 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
714 debug_check_no_obj_freed(page_address(page),
717 arch_free_page(page, order);
718 kernel_map_pages(page, 1 << order, 0);
723 static void __free_pages_ok(struct page *page, unsigned int order)
728 if (!free_pages_prepare(page, order))
731 local_irq_save(flags);
732 __count_vm_events(PGFREE, 1 << order);
733 migratetype = get_pageblock_migratetype(page);
734 set_freepage_migratetype(page, migratetype);
735 free_one_page(page_zone(page), page, order, migratetype);
736 local_irq_restore(flags);
740 * Read access to zone->managed_pages is safe because it's unsigned long,
741 * but we still need to serialize writers. Currently all callers of
742 * __free_pages_bootmem() except put_page_bootmem() should only be used
743 * at boot time. So for shorter boot time, we shift the burden to
744 * put_page_bootmem() to serialize writers.
746 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
748 unsigned int nr_pages = 1 << order;
752 for (loop = 0; loop < nr_pages; loop++) {
753 struct page *p = &page[loop];
755 if (loop + 1 < nr_pages)
757 __ClearPageReserved(p);
758 set_page_count(p, 0);
761 page_zone(page)->managed_pages += 1 << order;
762 set_page_refcounted(page);
763 __free_pages(page, order);
767 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
768 void __init init_cma_reserved_pageblock(struct page *page)
770 unsigned i = pageblock_nr_pages;
771 struct page *p = page;
774 __ClearPageReserved(p);
775 set_page_count(p, 0);
778 set_page_refcounted(page);
779 set_pageblock_migratetype(page, MIGRATE_CMA);
780 __free_pages(page, pageblock_order);
781 totalram_pages += pageblock_nr_pages;
786 * The order of subdivision here is critical for the IO subsystem.
787 * Please do not alter this order without good reasons and regression
788 * testing. Specifically, as large blocks of memory are subdivided,
789 * the order in which smaller blocks are delivered depends on the order
790 * they're subdivided in this function. This is the primary factor
791 * influencing the order in which pages are delivered to the IO
792 * subsystem according to empirical testing, and this is also justified
793 * by considering the behavior of a buddy system containing a single
794 * large block of memory acted on by a series of small allocations.
795 * This behavior is a critical factor in sglist merging's success.
799 static inline void expand(struct zone *zone, struct page *page,
800 int low, int high, struct free_area *area,
803 unsigned long size = 1 << high;
809 VM_BUG_ON(bad_range(zone, &page[size]));
811 #ifdef CONFIG_DEBUG_PAGEALLOC
812 if (high < debug_guardpage_minorder()) {
814 * Mark as guard pages (or page), that will allow to
815 * merge back to allocator when buddy will be freed.
816 * Corresponding page table entries will not be touched,
817 * pages will stay not present in virtual address space
819 INIT_LIST_HEAD(&page[size].lru);
820 set_page_guard_flag(&page[size]);
821 set_page_private(&page[size], high);
822 /* Guard pages are not available for any usage */
823 __mod_zone_freepage_state(zone, -(1 << high),
828 list_add(&page[size].lru, &area->free_list[migratetype]);
830 set_page_order(&page[size], high);
835 * This page is about to be returned from the page allocator
837 static inline int check_new_page(struct page *page)
839 if (unlikely(page_mapcount(page) |
840 (page->mapping != NULL) |
841 (atomic_read(&page->_count) != 0) |
842 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
843 (mem_cgroup_bad_page_check(page)))) {
850 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
854 for (i = 0; i < (1 << order); i++) {
855 struct page *p = page + i;
856 if (unlikely(check_new_page(p)))
860 set_page_private(page, 0);
861 set_page_refcounted(page);
863 arch_alloc_page(page, order);
864 kernel_map_pages(page, 1 << order, 1);
866 if (gfp_flags & __GFP_ZERO)
867 prep_zero_page(page, order, gfp_flags);
869 if (order && (gfp_flags & __GFP_COMP))
870 prep_compound_page(page, order);
876 * Go through the free lists for the given migratetype and remove
877 * the smallest available page from the freelists
880 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
883 unsigned int current_order;
884 struct free_area * area;
887 /* Find a page of the appropriate size in the preferred list */
888 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
889 area = &(zone->free_area[current_order]);
890 if (list_empty(&area->free_list[migratetype]))
893 page = list_entry(area->free_list[migratetype].next,
895 list_del(&page->lru);
896 rmv_page_order(page);
898 expand(zone, page, order, current_order, area, migratetype);
907 * This array describes the order lists are fallen back to when
908 * the free lists for the desirable migrate type are depleted
910 static int fallbacks[MIGRATE_TYPES][4] = {
911 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
912 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
914 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
915 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
917 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
919 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
920 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
924 * Move the free pages in a range to the free lists of the requested type.
925 * Note that start_page and end_pages are not aligned on a pageblock
926 * boundary. If alignment is required, use move_freepages_block()
928 int move_freepages(struct zone *zone,
929 struct page *start_page, struct page *end_page,
936 #ifndef CONFIG_HOLES_IN_ZONE
938 * page_zone is not safe to call in this context when
939 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
940 * anyway as we check zone boundaries in move_freepages_block().
941 * Remove at a later date when no bug reports exist related to
942 * grouping pages by mobility
944 BUG_ON(page_zone(start_page) != page_zone(end_page));
947 for (page = start_page; page <= end_page;) {
948 /* Make sure we are not inadvertently changing nodes */
949 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
951 if (!pfn_valid_within(page_to_pfn(page))) {
956 if (!PageBuddy(page)) {
961 order = page_order(page);
962 list_move(&page->lru,
963 &zone->free_area[order].free_list[migratetype]);
964 set_freepage_migratetype(page, migratetype);
966 pages_moved += 1 << order;
972 int move_freepages_block(struct zone *zone, struct page *page,
975 unsigned long start_pfn, end_pfn;
976 struct page *start_page, *end_page;
978 start_pfn = page_to_pfn(page);
979 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
980 start_page = pfn_to_page(start_pfn);
981 end_page = start_page + pageblock_nr_pages - 1;
982 end_pfn = start_pfn + pageblock_nr_pages - 1;
984 /* Do not cross zone boundaries */
985 if (start_pfn < zone->zone_start_pfn)
987 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
990 return move_freepages(zone, start_page, end_page, migratetype);
993 static void change_pageblock_range(struct page *pageblock_page,
994 int start_order, int migratetype)
996 int nr_pageblocks = 1 << (start_order - pageblock_order);
998 while (nr_pageblocks--) {
999 set_pageblock_migratetype(pageblock_page, migratetype);
1000 pageblock_page += pageblock_nr_pages;
1004 /* Remove an element from the buddy allocator from the fallback list */
1005 static inline struct page *
1006 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1008 struct free_area * area;
1013 /* Find the largest possible block of pages in the other list */
1014 for (current_order = MAX_ORDER-1; current_order >= order;
1017 migratetype = fallbacks[start_migratetype][i];
1019 /* MIGRATE_RESERVE handled later if necessary */
1020 if (migratetype == MIGRATE_RESERVE)
1023 area = &(zone->free_area[current_order]);
1024 if (list_empty(&area->free_list[migratetype]))
1027 page = list_entry(area->free_list[migratetype].next,
1032 * If breaking a large block of pages, move all free
1033 * pages to the preferred allocation list. If falling
1034 * back for a reclaimable kernel allocation, be more
1035 * aggressive about taking ownership of free pages
1037 * On the other hand, never change migration
1038 * type of MIGRATE_CMA pageblocks nor move CMA
1039 * pages on different free lists. We don't
1040 * want unmovable pages to be allocated from
1041 * MIGRATE_CMA areas.
1043 if (!is_migrate_cma(migratetype) &&
1044 (unlikely(current_order >= pageblock_order / 2) ||
1045 start_migratetype == MIGRATE_RECLAIMABLE ||
1046 page_group_by_mobility_disabled)) {
1048 pages = move_freepages_block(zone, page,
1051 /* Claim the whole block if over half of it is free */
1052 if (pages >= (1 << (pageblock_order-1)) ||
1053 page_group_by_mobility_disabled)
1054 set_pageblock_migratetype(page,
1057 migratetype = start_migratetype;
1060 /* Remove the page from the freelists */
1061 list_del(&page->lru);
1062 rmv_page_order(page);
1064 /* Take ownership for orders >= pageblock_order */
1065 if (current_order >= pageblock_order &&
1066 !is_migrate_cma(migratetype))
1067 change_pageblock_range(page, current_order,
1070 expand(zone, page, order, current_order, area,
1071 is_migrate_cma(migratetype)
1072 ? migratetype : start_migratetype);
1074 trace_mm_page_alloc_extfrag(page, order, current_order,
1075 start_migratetype, migratetype);
1085 * Do the hard work of removing an element from the buddy allocator.
1086 * Call me with the zone->lock already held.
1088 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1094 page = __rmqueue_smallest(zone, order, migratetype);
1096 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1097 page = __rmqueue_fallback(zone, order, migratetype);
1100 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1101 * is used because __rmqueue_smallest is an inline function
1102 * and we want just one call site
1105 migratetype = MIGRATE_RESERVE;
1110 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1115 * Obtain a specified number of elements from the buddy allocator, all under
1116 * a single hold of the lock, for efficiency. Add them to the supplied list.
1117 * Returns the number of new pages which were placed at *list.
1119 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1120 unsigned long count, struct list_head *list,
1121 int migratetype, int cold)
1123 int mt = migratetype, i;
1125 spin_lock(&zone->lock);
1126 for (i = 0; i < count; ++i) {
1127 struct page *page = __rmqueue(zone, order, migratetype);
1128 if (unlikely(page == NULL))
1132 * Split buddy pages returned by expand() are received here
1133 * in physical page order. The page is added to the callers and
1134 * list and the list head then moves forward. From the callers
1135 * perspective, the linked list is ordered by page number in
1136 * some conditions. This is useful for IO devices that can
1137 * merge IO requests if the physical pages are ordered
1140 if (likely(cold == 0))
1141 list_add(&page->lru, list);
1143 list_add_tail(&page->lru, list);
1144 if (IS_ENABLED(CONFIG_CMA)) {
1145 mt = get_pageblock_migratetype(page);
1146 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1149 set_freepage_migratetype(page, mt);
1151 if (is_migrate_cma(mt))
1152 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1155 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1156 spin_unlock(&zone->lock);
1162 * Called from the vmstat counter updater to drain pagesets of this
1163 * currently executing processor on remote nodes after they have
1166 * Note that this function must be called with the thread pinned to
1167 * a single processor.
1169 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1171 unsigned long flags;
1174 local_irq_save(flags);
1175 if (pcp->count >= pcp->batch)
1176 to_drain = pcp->batch;
1178 to_drain = pcp->count;
1180 free_pcppages_bulk(zone, to_drain, pcp);
1181 pcp->count -= to_drain;
1183 local_irq_restore(flags);
1188 * Drain pages of the indicated processor.
1190 * The processor must either be the current processor and the
1191 * thread pinned to the current processor or a processor that
1194 static void drain_pages(unsigned int cpu)
1196 unsigned long flags;
1199 for_each_populated_zone(zone) {
1200 struct per_cpu_pageset *pset;
1201 struct per_cpu_pages *pcp;
1203 local_irq_save(flags);
1204 pset = per_cpu_ptr(zone->pageset, cpu);
1208 free_pcppages_bulk(zone, pcp->count, pcp);
1211 local_irq_restore(flags);
1216 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1218 void drain_local_pages(void *arg)
1220 drain_pages(smp_processor_id());
1224 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1226 * Note that this code is protected against sending an IPI to an offline
1227 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1228 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1229 * nothing keeps CPUs from showing up after we populated the cpumask and
1230 * before the call to on_each_cpu_mask().
1232 void drain_all_pages(void)
1235 struct per_cpu_pageset *pcp;
1239 * Allocate in the BSS so we wont require allocation in
1240 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1242 static cpumask_t cpus_with_pcps;
1245 * We don't care about racing with CPU hotplug event
1246 * as offline notification will cause the notified
1247 * cpu to drain that CPU pcps and on_each_cpu_mask
1248 * disables preemption as part of its processing
1250 for_each_online_cpu(cpu) {
1251 bool has_pcps = false;
1252 for_each_populated_zone(zone) {
1253 pcp = per_cpu_ptr(zone->pageset, cpu);
1254 if (pcp->pcp.count) {
1260 cpumask_set_cpu(cpu, &cpus_with_pcps);
1262 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1264 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1267 #ifdef CONFIG_HIBERNATION
1269 void mark_free_pages(struct zone *zone)
1271 unsigned long pfn, max_zone_pfn;
1272 unsigned long flags;
1274 struct list_head *curr;
1276 if (!zone->spanned_pages)
1279 spin_lock_irqsave(&zone->lock, flags);
1281 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1282 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1283 if (pfn_valid(pfn)) {
1284 struct page *page = pfn_to_page(pfn);
1286 if (!swsusp_page_is_forbidden(page))
1287 swsusp_unset_page_free(page);
1290 for_each_migratetype_order(order, t) {
1291 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1294 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1295 for (i = 0; i < (1UL << order); i++)
1296 swsusp_set_page_free(pfn_to_page(pfn + i));
1299 spin_unlock_irqrestore(&zone->lock, flags);
1301 #endif /* CONFIG_PM */
1304 * Free a 0-order page
1305 * cold == 1 ? free a cold page : free a hot page
1307 void free_hot_cold_page(struct page *page, int cold)
1309 struct zone *zone = page_zone(page);
1310 struct per_cpu_pages *pcp;
1311 unsigned long flags;
1314 if (!free_pages_prepare(page, 0))
1317 migratetype = get_pageblock_migratetype(page);
1318 set_freepage_migratetype(page, migratetype);
1319 local_irq_save(flags);
1320 __count_vm_event(PGFREE);
1323 * We only track unmovable, reclaimable and movable on pcp lists.
1324 * Free ISOLATE pages back to the allocator because they are being
1325 * offlined but treat RESERVE as movable pages so we can get those
1326 * areas back if necessary. Otherwise, we may have to free
1327 * excessively into the page allocator
1329 if (migratetype >= MIGRATE_PCPTYPES) {
1330 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1331 free_one_page(zone, page, 0, migratetype);
1334 migratetype = MIGRATE_MOVABLE;
1337 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1339 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1341 list_add(&page->lru, &pcp->lists[migratetype]);
1343 if (pcp->count >= pcp->high) {
1344 free_pcppages_bulk(zone, pcp->batch, pcp);
1345 pcp->count -= pcp->batch;
1349 local_irq_restore(flags);
1353 * Free a list of 0-order pages
1355 void free_hot_cold_page_list(struct list_head *list, int cold)
1357 struct page *page, *next;
1359 list_for_each_entry_safe(page, next, list, lru) {
1360 trace_mm_page_free_batched(page, cold);
1361 free_hot_cold_page(page, cold);
1366 * split_page takes a non-compound higher-order page, and splits it into
1367 * n (1<<order) sub-pages: page[0..n]
1368 * Each sub-page must be freed individually.
1370 * Note: this is probably too low level an operation for use in drivers.
1371 * Please consult with lkml before using this in your driver.
1373 void split_page(struct page *page, unsigned int order)
1377 VM_BUG_ON(PageCompound(page));
1378 VM_BUG_ON(!page_count(page));
1380 #ifdef CONFIG_KMEMCHECK
1382 * Split shadow pages too, because free(page[0]) would
1383 * otherwise free the whole shadow.
1385 if (kmemcheck_page_is_tracked(page))
1386 split_page(virt_to_page(page[0].shadow), order);
1389 for (i = 1; i < (1 << order); i++)
1390 set_page_refcounted(page + i);
1394 * Similar to the split_page family of functions except that the page
1395 * required at the given order and being isolated now to prevent races
1396 * with parallel allocators
1398 int capture_free_page(struct page *page, int alloc_order, int migratetype)
1401 unsigned long watermark;
1405 BUG_ON(!PageBuddy(page));
1407 zone = page_zone(page);
1408 order = page_order(page);
1409 mt = get_pageblock_migratetype(page);
1411 if (mt != MIGRATE_ISOLATE) {
1412 /* Obey watermarks as if the page was being allocated */
1413 watermark = low_wmark_pages(zone) + (1 << order);
1414 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1417 __mod_zone_freepage_state(zone, -(1UL << alloc_order), mt);
1420 /* Remove page from free list */
1421 list_del(&page->lru);
1422 zone->free_area[order].nr_free--;
1423 rmv_page_order(page);
1425 if (alloc_order != order)
1426 expand(zone, page, alloc_order, order,
1427 &zone->free_area[order], migratetype);
1429 /* Set the pageblock if the captured page is at least a pageblock */
1430 if (order >= pageblock_order - 1) {
1431 struct page *endpage = page + (1 << order) - 1;
1432 for (; page < endpage; page += pageblock_nr_pages) {
1433 int mt = get_pageblock_migratetype(page);
1434 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1435 set_pageblock_migratetype(page,
1440 return 1UL << alloc_order;
1444 * Similar to split_page except the page is already free. As this is only
1445 * being used for migration, the migratetype of the block also changes.
1446 * As this is called with interrupts disabled, the caller is responsible
1447 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1450 * Note: this is probably too low level an operation for use in drivers.
1451 * Please consult with lkml before using this in your driver.
1453 int split_free_page(struct page *page)
1458 BUG_ON(!PageBuddy(page));
1459 order = page_order(page);
1461 nr_pages = capture_free_page(page, order, 0);
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];
1634 free_pages -= (1 << order) - 1;
1635 if (alloc_flags & ALLOC_HIGH)
1637 if (alloc_flags & ALLOC_HARDER)
1640 /* If allocation can't use CMA areas don't use free CMA pages */
1641 if (!(alloc_flags & ALLOC_CMA))
1642 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1644 if (free_pages <= min + lowmem_reserve)
1646 for (o = 0; o < order; o++) {
1647 /* At the next order, this order's pages become unavailable */
1648 free_pages -= z->free_area[o].nr_free << o;
1650 /* Require fewer higher order pages to be free */
1653 if (free_pages <= min)
1659 #ifdef CONFIG_MEMORY_ISOLATION
1660 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1662 if (unlikely(zone->nr_pageblock_isolate))
1663 return zone->nr_pageblock_isolate * pageblock_nr_pages;
1667 static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1673 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1674 int classzone_idx, int alloc_flags)
1676 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1677 zone_page_state(z, NR_FREE_PAGES));
1680 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1681 int classzone_idx, int alloc_flags)
1683 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1685 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1686 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1689 * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1690 * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1691 * sleep although it could do so. But this is more desirable for memory
1692 * hotplug than sleeping which can cause a livelock in the direct
1695 free_pages -= nr_zone_isolate_freepages(z);
1696 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1702 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1703 * skip over zones that are not allowed by the cpuset, or that have
1704 * been recently (in last second) found to be nearly full. See further
1705 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1706 * that have to skip over a lot of full or unallowed zones.
1708 * If the zonelist cache is present in the passed in zonelist, then
1709 * returns a pointer to the allowed node mask (either the current
1710 * tasks mems_allowed, or node_states[N_MEMORY].)
1712 * If the zonelist cache is not available for this zonelist, does
1713 * nothing and returns NULL.
1715 * If the fullzones BITMAP in the zonelist cache is stale (more than
1716 * a second since last zap'd) then we zap it out (clear its bits.)
1718 * We hold off even calling zlc_setup, until after we've checked the
1719 * first zone in the zonelist, on the theory that most allocations will
1720 * be satisfied from that first zone, so best to examine that zone as
1721 * quickly as we can.
1723 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1725 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1726 nodemask_t *allowednodes; /* zonelist_cache approximation */
1728 zlc = zonelist->zlcache_ptr;
1732 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1733 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1734 zlc->last_full_zap = jiffies;
1737 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1738 &cpuset_current_mems_allowed :
1739 &node_states[N_MEMORY];
1740 return allowednodes;
1744 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1745 * if it is worth looking at further for free memory:
1746 * 1) Check that the zone isn't thought to be full (doesn't have its
1747 * bit set in the zonelist_cache fullzones BITMAP).
1748 * 2) Check that the zones node (obtained from the zonelist_cache
1749 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1750 * Return true (non-zero) if zone is worth looking at further, or
1751 * else return false (zero) if it is not.
1753 * This check -ignores- the distinction between various watermarks,
1754 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1755 * found to be full for any variation of these watermarks, it will
1756 * be considered full for up to one second by all requests, unless
1757 * we are so low on memory on all allowed nodes that we are forced
1758 * into the second scan of the zonelist.
1760 * In the second scan we ignore this zonelist cache and exactly
1761 * apply the watermarks to all zones, even it is slower to do so.
1762 * We are low on memory in the second scan, and should leave no stone
1763 * unturned looking for a free page.
1765 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1766 nodemask_t *allowednodes)
1768 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1769 int i; /* index of *z in zonelist zones */
1770 int n; /* node that zone *z is on */
1772 zlc = zonelist->zlcache_ptr;
1776 i = z - zonelist->_zonerefs;
1779 /* This zone is worth trying if it is allowed but not full */
1780 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1784 * Given 'z' scanning a zonelist, set the corresponding bit in
1785 * zlc->fullzones, so that subsequent attempts to allocate a page
1786 * from that zone don't waste time re-examining it.
1788 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1790 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1791 int i; /* index of *z in zonelist zones */
1793 zlc = zonelist->zlcache_ptr;
1797 i = z - zonelist->_zonerefs;
1799 set_bit(i, zlc->fullzones);
1803 * clear all zones full, called after direct reclaim makes progress so that
1804 * a zone that was recently full is not skipped over for up to a second
1806 static void zlc_clear_zones_full(struct zonelist *zonelist)
1808 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1810 zlc = zonelist->zlcache_ptr;
1814 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1817 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1819 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1822 static void __paginginit init_zone_allows_reclaim(int nid)
1826 for_each_online_node(i)
1827 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1828 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1830 zone_reclaim_mode = 1;
1833 #else /* CONFIG_NUMA */
1835 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1840 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1841 nodemask_t *allowednodes)
1846 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1850 static void zlc_clear_zones_full(struct zonelist *zonelist)
1854 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1859 static inline void init_zone_allows_reclaim(int nid)
1862 #endif /* CONFIG_NUMA */
1865 * get_page_from_freelist goes through the zonelist trying to allocate
1868 static struct page *
1869 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1870 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1871 struct zone *preferred_zone, int migratetype)
1874 struct page *page = NULL;
1877 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1878 int zlc_active = 0; /* set if using zonelist_cache */
1879 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1881 classzone_idx = zone_idx(preferred_zone);
1884 * Scan zonelist, looking for a zone with enough free.
1885 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1887 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1888 high_zoneidx, nodemask) {
1889 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1890 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1892 if ((alloc_flags & ALLOC_CPUSET) &&
1893 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1896 * When allocating a page cache page for writing, we
1897 * want to get it from a zone that is within its dirty
1898 * limit, such that no single zone holds more than its
1899 * proportional share of globally allowed dirty pages.
1900 * The dirty limits take into account the zone's
1901 * lowmem reserves and high watermark so that kswapd
1902 * should be able to balance it without having to
1903 * write pages from its LRU list.
1905 * This may look like it could increase pressure on
1906 * lower zones by failing allocations in higher zones
1907 * before they are full. But the pages that do spill
1908 * over are limited as the lower zones are protected
1909 * by this very same mechanism. It should not become
1910 * a practical burden to them.
1912 * XXX: For now, allow allocations to potentially
1913 * exceed the per-zone dirty limit in the slowpath
1914 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1915 * which is important when on a NUMA setup the allowed
1916 * zones are together not big enough to reach the
1917 * global limit. The proper fix for these situations
1918 * will require awareness of zones in the
1919 * dirty-throttling and the flusher threads.
1921 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1922 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1923 goto this_zone_full;
1925 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1926 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1930 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1931 if (zone_watermark_ok(zone, order, mark,
1932 classzone_idx, alloc_flags))
1935 if (IS_ENABLED(CONFIG_NUMA) &&
1936 !did_zlc_setup && nr_online_nodes > 1) {
1938 * we do zlc_setup if there are multiple nodes
1939 * and before considering the first zone allowed
1942 allowednodes = zlc_setup(zonelist, alloc_flags);
1947 if (zone_reclaim_mode == 0 ||
1948 !zone_allows_reclaim(preferred_zone, zone))
1949 goto this_zone_full;
1952 * As we may have just activated ZLC, check if the first
1953 * eligible zone has failed zone_reclaim recently.
1955 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1956 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1959 ret = zone_reclaim(zone, gfp_mask, order);
1961 case ZONE_RECLAIM_NOSCAN:
1964 case ZONE_RECLAIM_FULL:
1965 /* scanned but unreclaimable */
1968 /* did we reclaim enough */
1969 if (!zone_watermark_ok(zone, order, mark,
1970 classzone_idx, alloc_flags))
1971 goto this_zone_full;
1976 page = buffered_rmqueue(preferred_zone, zone, order,
1977 gfp_mask, migratetype);
1981 if (IS_ENABLED(CONFIG_NUMA))
1982 zlc_mark_zone_full(zonelist, z);
1985 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1986 /* Disable zlc cache for second zonelist scan */
1993 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1994 * necessary to allocate the page. The expectation is
1995 * that the caller is taking steps that will free more
1996 * memory. The caller should avoid the page being used
1997 * for !PFMEMALLOC purposes.
1999 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2005 * Large machines with many possible nodes should not always dump per-node
2006 * meminfo in irq context.
2008 static inline bool should_suppress_show_mem(void)
2013 ret = in_interrupt();
2018 static DEFINE_RATELIMIT_STATE(nopage_rs,
2019 DEFAULT_RATELIMIT_INTERVAL,
2020 DEFAULT_RATELIMIT_BURST);
2022 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2024 unsigned int filter = SHOW_MEM_FILTER_NODES;
2026 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2027 debug_guardpage_minorder() > 0)
2031 * This documents exceptions given to allocations in certain
2032 * contexts that are allowed to allocate outside current's set
2035 if (!(gfp_mask & __GFP_NOMEMALLOC))
2036 if (test_thread_flag(TIF_MEMDIE) ||
2037 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2038 filter &= ~SHOW_MEM_FILTER_NODES;
2039 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2040 filter &= ~SHOW_MEM_FILTER_NODES;
2043 struct va_format vaf;
2046 va_start(args, fmt);
2051 pr_warn("%pV", &vaf);
2056 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2057 current->comm, order, gfp_mask);
2060 if (!should_suppress_show_mem())
2065 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2066 unsigned long did_some_progress,
2067 unsigned long pages_reclaimed)
2069 /* Do not loop if specifically requested */
2070 if (gfp_mask & __GFP_NORETRY)
2073 /* Always retry if specifically requested */
2074 if (gfp_mask & __GFP_NOFAIL)
2078 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2079 * making forward progress without invoking OOM. Suspend also disables
2080 * storage devices so kswapd will not help. Bail if we are suspending.
2082 if (!did_some_progress && pm_suspended_storage())
2086 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2087 * means __GFP_NOFAIL, but that may not be true in other
2090 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2094 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2095 * specified, then we retry until we no longer reclaim any pages
2096 * (above), or we've reclaimed an order of pages at least as
2097 * large as the allocation's order. In both cases, if the
2098 * allocation still fails, we stop retrying.
2100 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2106 static inline struct page *
2107 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2108 struct zonelist *zonelist, enum zone_type high_zoneidx,
2109 nodemask_t *nodemask, struct zone *preferred_zone,
2114 /* Acquire the OOM killer lock for the zones in zonelist */
2115 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2116 schedule_timeout_uninterruptible(1);
2121 * Go through the zonelist yet one more time, keep very high watermark
2122 * here, this is only to catch a parallel oom killing, we must fail if
2123 * we're still under heavy pressure.
2125 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2126 order, zonelist, high_zoneidx,
2127 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2128 preferred_zone, migratetype);
2132 if (!(gfp_mask & __GFP_NOFAIL)) {
2133 /* The OOM killer will not help higher order allocs */
2134 if (order > PAGE_ALLOC_COSTLY_ORDER)
2136 /* The OOM killer does not needlessly kill tasks for lowmem */
2137 if (high_zoneidx < ZONE_NORMAL)
2140 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2141 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2142 * The caller should handle page allocation failure by itself if
2143 * it specifies __GFP_THISNODE.
2144 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2146 if (gfp_mask & __GFP_THISNODE)
2149 /* Exhausted what can be done so it's blamo time */
2150 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2153 clear_zonelist_oom(zonelist, gfp_mask);
2157 #ifdef CONFIG_COMPACTION
2158 /* Try memory compaction for high-order allocations before reclaim */
2159 static struct page *
2160 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2161 struct zonelist *zonelist, enum zone_type high_zoneidx,
2162 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2163 int migratetype, bool sync_migration,
2164 bool *contended_compaction, bool *deferred_compaction,
2165 unsigned long *did_some_progress)
2167 struct page *page = NULL;
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, &page);
2181 current->flags &= ~PF_MEMALLOC;
2183 /* If compaction captured a page, prep and use it */
2185 prep_new_page(page, order, gfp_mask);
2189 if (*did_some_progress != COMPACT_SKIPPED) {
2190 /* Page migration frees to the PCP lists but we want merging */
2191 drain_pages(get_cpu());
2194 page = get_page_from_freelist(gfp_mask, nodemask,
2195 order, zonelist, high_zoneidx,
2196 alloc_flags & ~ALLOC_NO_WATERMARKS,
2197 preferred_zone, migratetype);
2200 preferred_zone->compact_blockskip_flush = false;
2201 preferred_zone->compact_considered = 0;
2202 preferred_zone->compact_defer_shift = 0;
2203 if (order >= preferred_zone->compact_order_failed)
2204 preferred_zone->compact_order_failed = order + 1;
2205 count_vm_event(COMPACTSUCCESS);
2210 * It's bad if compaction run occurs and fails.
2211 * The most likely reason is that pages exist,
2212 * but not enough to satisfy watermarks.
2214 count_vm_event(COMPACTFAIL);
2217 * As async compaction considers a subset of pageblocks, only
2218 * defer if the failure was a sync compaction failure.
2221 defer_compaction(preferred_zone, order);
2229 static inline struct page *
2230 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2231 struct zonelist *zonelist, enum zone_type high_zoneidx,
2232 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2233 int migratetype, bool sync_migration,
2234 bool *contended_compaction, bool *deferred_compaction,
2235 unsigned long *did_some_progress)
2239 #endif /* CONFIG_COMPACTION */
2241 /* Perform direct synchronous page reclaim */
2243 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2244 nodemask_t *nodemask)
2246 struct reclaim_state reclaim_state;
2251 /* We now go into synchronous reclaim */
2252 cpuset_memory_pressure_bump();
2253 current->flags |= PF_MEMALLOC;
2254 lockdep_set_current_reclaim_state(gfp_mask);
2255 reclaim_state.reclaimed_slab = 0;
2256 current->reclaim_state = &reclaim_state;
2258 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2260 current->reclaim_state = NULL;
2261 lockdep_clear_current_reclaim_state();
2262 current->flags &= ~PF_MEMALLOC;
2269 /* The really slow allocator path where we enter direct reclaim */
2270 static inline struct page *
2271 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2272 struct zonelist *zonelist, enum zone_type high_zoneidx,
2273 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2274 int migratetype, unsigned long *did_some_progress)
2276 struct page *page = NULL;
2277 bool drained = false;
2279 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2281 if (unlikely(!(*did_some_progress)))
2284 /* After successful reclaim, reconsider all zones for allocation */
2285 if (IS_ENABLED(CONFIG_NUMA))
2286 zlc_clear_zones_full(zonelist);
2289 page = get_page_from_freelist(gfp_mask, nodemask, order,
2290 zonelist, high_zoneidx,
2291 alloc_flags & ~ALLOC_NO_WATERMARKS,
2292 preferred_zone, migratetype);
2295 * If an allocation failed after direct reclaim, it could be because
2296 * pages are pinned on the per-cpu lists. Drain them and try again
2298 if (!page && !drained) {
2308 * This is called in the allocator slow-path if the allocation request is of
2309 * sufficient urgency to ignore watermarks and take other desperate measures
2311 static inline struct page *
2312 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2313 struct zonelist *zonelist, enum zone_type high_zoneidx,
2314 nodemask_t *nodemask, struct zone *preferred_zone,
2320 page = get_page_from_freelist(gfp_mask, nodemask, order,
2321 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2322 preferred_zone, migratetype);
2324 if (!page && gfp_mask & __GFP_NOFAIL)
2325 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2326 } while (!page && (gfp_mask & __GFP_NOFAIL));
2332 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2333 enum zone_type high_zoneidx,
2334 enum zone_type classzone_idx)
2339 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2340 wakeup_kswapd(zone, order, classzone_idx);
2344 gfp_to_alloc_flags(gfp_t gfp_mask)
2346 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2347 const gfp_t wait = gfp_mask & __GFP_WAIT;
2349 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2350 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2353 * The caller may dip into page reserves a bit more if the caller
2354 * cannot run direct reclaim, or if the caller has realtime scheduling
2355 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2356 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2358 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2362 * Not worth trying to allocate harder for
2363 * __GFP_NOMEMALLOC even if it can't schedule.
2365 if (!(gfp_mask & __GFP_NOMEMALLOC))
2366 alloc_flags |= ALLOC_HARDER;
2368 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2369 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2371 alloc_flags &= ~ALLOC_CPUSET;
2372 } else if (unlikely(rt_task(current)) && !in_interrupt())
2373 alloc_flags |= ALLOC_HARDER;
2375 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2376 if (gfp_mask & __GFP_MEMALLOC)
2377 alloc_flags |= ALLOC_NO_WATERMARKS;
2378 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2379 alloc_flags |= ALLOC_NO_WATERMARKS;
2380 else if (!in_interrupt() &&
2381 ((current->flags & PF_MEMALLOC) ||
2382 unlikely(test_thread_flag(TIF_MEMDIE))))
2383 alloc_flags |= ALLOC_NO_WATERMARKS;
2386 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2387 alloc_flags |= ALLOC_CMA;
2392 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2394 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2397 static inline struct page *
2398 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2399 struct zonelist *zonelist, enum zone_type high_zoneidx,
2400 nodemask_t *nodemask, struct zone *preferred_zone,
2403 const gfp_t wait = gfp_mask & __GFP_WAIT;
2404 struct page *page = NULL;
2406 unsigned long pages_reclaimed = 0;
2407 unsigned long did_some_progress;
2408 bool sync_migration = false;
2409 bool deferred_compaction = false;
2410 bool contended_compaction = false;
2413 * In the slowpath, we sanity check order to avoid ever trying to
2414 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2415 * be using allocators in order of preference for an area that is
2418 if (order >= MAX_ORDER) {
2419 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2424 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2425 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2426 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2427 * using a larger set of nodes after it has established that the
2428 * allowed per node queues are empty and that nodes are
2431 if (IS_ENABLED(CONFIG_NUMA) &&
2432 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2436 if (!(gfp_mask & __GFP_NO_KSWAPD))
2437 wake_all_kswapd(order, zonelist, high_zoneidx,
2438 zone_idx(preferred_zone));
2441 * OK, we're below the kswapd watermark and have kicked background
2442 * reclaim. Now things get more complex, so set up alloc_flags according
2443 * to how we want to proceed.
2445 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2448 * Find the true preferred zone if the allocation is unconstrained by
2451 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2452 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2456 /* This is the last chance, in general, before the goto nopage. */
2457 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2458 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2459 preferred_zone, migratetype);
2463 /* Allocate without watermarks if the context allows */
2464 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2466 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2467 * the allocation is high priority and these type of
2468 * allocations are system rather than user orientated
2470 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2472 page = __alloc_pages_high_priority(gfp_mask, order,
2473 zonelist, high_zoneidx, nodemask,
2474 preferred_zone, migratetype);
2480 /* Atomic allocations - we can't balance anything */
2484 /* Avoid recursion of direct reclaim */
2485 if (current->flags & PF_MEMALLOC)
2488 /* Avoid allocations with no watermarks from looping endlessly */
2489 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2493 * Try direct compaction. The first pass is asynchronous. Subsequent
2494 * attempts after direct reclaim are synchronous
2496 page = __alloc_pages_direct_compact(gfp_mask, order,
2497 zonelist, high_zoneidx,
2499 alloc_flags, preferred_zone,
2500 migratetype, sync_migration,
2501 &contended_compaction,
2502 &deferred_compaction,
2503 &did_some_progress);
2506 sync_migration = true;
2509 * If compaction is deferred for high-order allocations, it is because
2510 * sync compaction recently failed. In this is the case and the caller
2511 * requested a movable allocation that does not heavily disrupt the
2512 * system then fail the allocation instead of entering direct reclaim.
2514 if ((deferred_compaction || contended_compaction) &&
2515 (gfp_mask & __GFP_NO_KSWAPD))
2518 /* Try direct reclaim and then allocating */
2519 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2520 zonelist, high_zoneidx,
2522 alloc_flags, preferred_zone,
2523 migratetype, &did_some_progress);
2528 * If we failed to make any progress reclaiming, then we are
2529 * running out of options and have to consider going OOM
2531 if (!did_some_progress) {
2532 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2533 if (oom_killer_disabled)
2535 /* Coredumps can quickly deplete all memory reserves */
2536 if ((current->flags & PF_DUMPCORE) &&
2537 !(gfp_mask & __GFP_NOFAIL))
2539 page = __alloc_pages_may_oom(gfp_mask, order,
2540 zonelist, high_zoneidx,
2541 nodemask, preferred_zone,
2546 if (!(gfp_mask & __GFP_NOFAIL)) {
2548 * The oom killer is not called for high-order
2549 * allocations that may fail, so if no progress
2550 * is being made, there are no other options and
2551 * retrying is unlikely to help.
2553 if (order > PAGE_ALLOC_COSTLY_ORDER)
2556 * The oom killer is not called for lowmem
2557 * allocations to prevent needlessly killing
2560 if (high_zoneidx < ZONE_NORMAL)
2568 /* Check if we should retry the allocation */
2569 pages_reclaimed += did_some_progress;
2570 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2572 /* Wait for some write requests to complete then retry */
2573 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2577 * High-order allocations do not necessarily loop after
2578 * direct reclaim and reclaim/compaction depends on compaction
2579 * being called after reclaim so call directly if necessary
2581 page = __alloc_pages_direct_compact(gfp_mask, order,
2582 zonelist, high_zoneidx,
2584 alloc_flags, preferred_zone,
2585 migratetype, sync_migration,
2586 &contended_compaction,
2587 &deferred_compaction,
2588 &did_some_progress);
2594 warn_alloc_failed(gfp_mask, order, NULL);
2597 if (kmemcheck_enabled)
2598 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2604 * This is the 'heart' of the zoned buddy allocator.
2607 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2608 struct zonelist *zonelist, nodemask_t *nodemask)
2610 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2611 struct zone *preferred_zone;
2612 struct page *page = NULL;
2613 int migratetype = allocflags_to_migratetype(gfp_mask);
2614 unsigned int cpuset_mems_cookie;
2615 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2617 gfp_mask &= gfp_allowed_mask;
2619 lockdep_trace_alloc(gfp_mask);
2621 might_sleep_if(gfp_mask & __GFP_WAIT);
2623 if (should_fail_alloc_page(gfp_mask, order))
2627 * Check the zones suitable for the gfp_mask contain at least one
2628 * valid zone. It's possible to have an empty zonelist as a result
2629 * of GFP_THISNODE and a memoryless node
2631 if (unlikely(!zonelist->_zonerefs->zone))
2635 cpuset_mems_cookie = get_mems_allowed();
2637 /* The preferred zone is used for statistics later */
2638 first_zones_zonelist(zonelist, high_zoneidx,
2639 nodemask ? : &cpuset_current_mems_allowed,
2641 if (!preferred_zone)
2645 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2646 alloc_flags |= ALLOC_CMA;
2648 /* First allocation attempt */
2649 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2650 zonelist, high_zoneidx, alloc_flags,
2651 preferred_zone, migratetype);
2652 if (unlikely(!page))
2653 page = __alloc_pages_slowpath(gfp_mask, order,
2654 zonelist, high_zoneidx, nodemask,
2655 preferred_zone, migratetype);
2657 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2661 * When updating a task's mems_allowed, it is possible to race with
2662 * parallel threads in such a way that an allocation can fail while
2663 * the mask is being updated. If a page allocation is about to fail,
2664 * check if the cpuset changed during allocation and if so, retry.
2666 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2671 EXPORT_SYMBOL(__alloc_pages_nodemask);
2674 * Common helper functions.
2676 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2681 * __get_free_pages() returns a 32-bit address, which cannot represent
2684 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2686 page = alloc_pages(gfp_mask, order);
2689 return (unsigned long) page_address(page);
2691 EXPORT_SYMBOL(__get_free_pages);
2693 unsigned long get_zeroed_page(gfp_t gfp_mask)
2695 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2697 EXPORT_SYMBOL(get_zeroed_page);
2699 void __free_pages(struct page *page, unsigned int order)
2701 if (put_page_testzero(page)) {
2703 free_hot_cold_page(page, 0);
2705 __free_pages_ok(page, order);
2709 EXPORT_SYMBOL(__free_pages);
2711 void free_pages(unsigned long addr, unsigned int order)
2714 VM_BUG_ON(!virt_addr_valid((void *)addr));
2715 __free_pages(virt_to_page((void *)addr), order);
2719 EXPORT_SYMBOL(free_pages);
2721 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2724 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2725 unsigned long used = addr + PAGE_ALIGN(size);
2727 split_page(virt_to_page((void *)addr), order);
2728 while (used < alloc_end) {
2733 return (void *)addr;
2737 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2738 * @size: the number of bytes to allocate
2739 * @gfp_mask: GFP flags for the allocation
2741 * This function is similar to alloc_pages(), except that it allocates the
2742 * minimum number of pages to satisfy the request. alloc_pages() can only
2743 * allocate memory in power-of-two pages.
2745 * This function is also limited by MAX_ORDER.
2747 * Memory allocated by this function must be released by free_pages_exact().
2749 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2751 unsigned int order = get_order(size);
2754 addr = __get_free_pages(gfp_mask, order);
2755 return make_alloc_exact(addr, order, size);
2757 EXPORT_SYMBOL(alloc_pages_exact);
2760 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2762 * @nid: the preferred node ID where memory should be allocated
2763 * @size: the number of bytes to allocate
2764 * @gfp_mask: GFP flags for the allocation
2766 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2768 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2771 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2773 unsigned order = get_order(size);
2774 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2777 return make_alloc_exact((unsigned long)page_address(p), order, size);
2779 EXPORT_SYMBOL(alloc_pages_exact_nid);
2782 * free_pages_exact - release memory allocated via alloc_pages_exact()
2783 * @virt: the value returned by alloc_pages_exact.
2784 * @size: size of allocation, same value as passed to alloc_pages_exact().
2786 * Release the memory allocated by a previous call to alloc_pages_exact.
2788 void free_pages_exact(void *virt, size_t size)
2790 unsigned long addr = (unsigned long)virt;
2791 unsigned long end = addr + PAGE_ALIGN(size);
2793 while (addr < end) {
2798 EXPORT_SYMBOL(free_pages_exact);
2800 static unsigned int nr_free_zone_pages(int offset)
2805 /* Just pick one node, since fallback list is circular */
2806 unsigned int sum = 0;
2808 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2810 for_each_zone_zonelist(zone, z, zonelist, offset) {
2811 unsigned long size = zone->present_pages;
2812 unsigned long high = high_wmark_pages(zone);
2821 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2823 unsigned int nr_free_buffer_pages(void)
2825 return nr_free_zone_pages(gfp_zone(GFP_USER));
2827 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2830 * Amount of free RAM allocatable within all zones
2832 unsigned int nr_free_pagecache_pages(void)
2834 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2837 static inline void show_node(struct zone *zone)
2839 if (IS_ENABLED(CONFIG_NUMA))
2840 printk("Node %d ", zone_to_nid(zone));
2843 void si_meminfo(struct sysinfo *val)
2845 val->totalram = totalram_pages;
2847 val->freeram = global_page_state(NR_FREE_PAGES);
2848 val->bufferram = nr_blockdev_pages();
2849 val->totalhigh = totalhigh_pages;
2850 val->freehigh = nr_free_highpages();
2851 val->mem_unit = PAGE_SIZE;
2854 EXPORT_SYMBOL(si_meminfo);
2857 void si_meminfo_node(struct sysinfo *val, int nid)
2859 pg_data_t *pgdat = NODE_DATA(nid);
2861 val->totalram = pgdat->node_present_pages;
2862 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2863 #ifdef CONFIG_HIGHMEM
2864 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2865 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2871 val->mem_unit = PAGE_SIZE;
2876 * Determine whether the node should be displayed or not, depending on whether
2877 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2879 bool skip_free_areas_node(unsigned int flags, int nid)
2882 unsigned int cpuset_mems_cookie;
2884 if (!(flags & SHOW_MEM_FILTER_NODES))
2888 cpuset_mems_cookie = get_mems_allowed();
2889 ret = !node_isset(nid, cpuset_current_mems_allowed);
2890 } while (!put_mems_allowed(cpuset_mems_cookie));
2895 #define K(x) ((x) << (PAGE_SHIFT-10))
2897 static void show_migration_types(unsigned char type)
2899 static const char types[MIGRATE_TYPES] = {
2900 [MIGRATE_UNMOVABLE] = 'U',
2901 [MIGRATE_RECLAIMABLE] = 'E',
2902 [MIGRATE_MOVABLE] = 'M',
2903 [MIGRATE_RESERVE] = 'R',
2905 [MIGRATE_CMA] = 'C',
2907 [MIGRATE_ISOLATE] = 'I',
2909 char tmp[MIGRATE_TYPES + 1];
2913 for (i = 0; i < MIGRATE_TYPES; i++) {
2914 if (type & (1 << i))
2919 printk("(%s) ", tmp);
2923 * Show free area list (used inside shift_scroll-lock stuff)
2924 * We also calculate the percentage fragmentation. We do this by counting the
2925 * memory on each free list with the exception of the first item on the list.
2926 * Suppresses nodes that are not allowed by current's cpuset if
2927 * SHOW_MEM_FILTER_NODES is passed.
2929 void show_free_areas(unsigned int filter)
2934 for_each_populated_zone(zone) {
2935 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2938 printk("%s per-cpu:\n", zone->name);
2940 for_each_online_cpu(cpu) {
2941 struct per_cpu_pageset *pageset;
2943 pageset = per_cpu_ptr(zone->pageset, cpu);
2945 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2946 cpu, pageset->pcp.high,
2947 pageset->pcp.batch, pageset->pcp.count);
2951 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2952 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2954 " dirty:%lu writeback:%lu unstable:%lu\n"
2955 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2956 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2958 global_page_state(NR_ACTIVE_ANON),
2959 global_page_state(NR_INACTIVE_ANON),
2960 global_page_state(NR_ISOLATED_ANON),
2961 global_page_state(NR_ACTIVE_FILE),
2962 global_page_state(NR_INACTIVE_FILE),
2963 global_page_state(NR_ISOLATED_FILE),
2964 global_page_state(NR_UNEVICTABLE),
2965 global_page_state(NR_FILE_DIRTY),
2966 global_page_state(NR_WRITEBACK),
2967 global_page_state(NR_UNSTABLE_NFS),
2968 global_page_state(NR_FREE_PAGES),
2969 global_page_state(NR_SLAB_RECLAIMABLE),
2970 global_page_state(NR_SLAB_UNRECLAIMABLE),
2971 global_page_state(NR_FILE_MAPPED),
2972 global_page_state(NR_SHMEM),
2973 global_page_state(NR_PAGETABLE),
2974 global_page_state(NR_BOUNCE),
2975 global_page_state(NR_FREE_CMA_PAGES));
2977 for_each_populated_zone(zone) {
2980 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2988 " active_anon:%lukB"
2989 " inactive_anon:%lukB"
2990 " active_file:%lukB"
2991 " inactive_file:%lukB"
2992 " unevictable:%lukB"
2993 " isolated(anon):%lukB"
2994 " isolated(file):%lukB"
3002 " slab_reclaimable:%lukB"
3003 " slab_unreclaimable:%lukB"
3004 " kernel_stack:%lukB"
3009 " writeback_tmp:%lukB"
3010 " pages_scanned:%lu"
3011 " all_unreclaimable? %s"
3014 K(zone_page_state(zone, NR_FREE_PAGES)),
3015 K(min_wmark_pages(zone)),
3016 K(low_wmark_pages(zone)),
3017 K(high_wmark_pages(zone)),
3018 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3019 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3020 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3021 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3022 K(zone_page_state(zone, NR_UNEVICTABLE)),
3023 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3024 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3025 K(zone->present_pages),
3026 K(zone->managed_pages),
3027 K(zone_page_state(zone, NR_MLOCK)),
3028 K(zone_page_state(zone, NR_FILE_DIRTY)),
3029 K(zone_page_state(zone, NR_WRITEBACK)),
3030 K(zone_page_state(zone, NR_FILE_MAPPED)),
3031 K(zone_page_state(zone, NR_SHMEM)),
3032 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3033 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3034 zone_page_state(zone, NR_KERNEL_STACK) *
3036 K(zone_page_state(zone, NR_PAGETABLE)),
3037 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3038 K(zone_page_state(zone, NR_BOUNCE)),
3039 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3040 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3041 zone->pages_scanned,
3042 (zone->all_unreclaimable ? "yes" : "no")
3044 printk("lowmem_reserve[]:");
3045 for (i = 0; i < MAX_NR_ZONES; i++)
3046 printk(" %lu", zone->lowmem_reserve[i]);
3050 for_each_populated_zone(zone) {
3051 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3052 unsigned char types[MAX_ORDER];
3054 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3057 printk("%s: ", zone->name);
3059 spin_lock_irqsave(&zone->lock, flags);
3060 for (order = 0; order < MAX_ORDER; order++) {
3061 struct free_area *area = &zone->free_area[order];
3064 nr[order] = area->nr_free;
3065 total += nr[order] << order;
3068 for (type = 0; type < MIGRATE_TYPES; type++) {
3069 if (!list_empty(&area->free_list[type]))
3070 types[order] |= 1 << type;
3073 spin_unlock_irqrestore(&zone->lock, flags);
3074 for (order = 0; order < MAX_ORDER; order++) {
3075 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3077 show_migration_types(types[order]);
3079 printk("= %lukB\n", K(total));
3082 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3084 show_swap_cache_info();
3087 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3089 zoneref->zone = zone;
3090 zoneref->zone_idx = zone_idx(zone);
3094 * Builds allocation fallback zone lists.
3096 * Add all populated zones of a node to the zonelist.
3098 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3099 int nr_zones, enum zone_type zone_type)
3103 BUG_ON(zone_type >= MAX_NR_ZONES);
3108 zone = pgdat->node_zones + zone_type;
3109 if (populated_zone(zone)) {
3110 zoneref_set_zone(zone,
3111 &zonelist->_zonerefs[nr_zones++]);
3112 check_highest_zone(zone_type);
3115 } while (zone_type);
3122 * 0 = automatic detection of better ordering.
3123 * 1 = order by ([node] distance, -zonetype)
3124 * 2 = order by (-zonetype, [node] distance)
3126 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3127 * the same zonelist. So only NUMA can configure this param.
3129 #define ZONELIST_ORDER_DEFAULT 0
3130 #define ZONELIST_ORDER_NODE 1
3131 #define ZONELIST_ORDER_ZONE 2
3133 /* zonelist order in the kernel.
3134 * set_zonelist_order() will set this to NODE or ZONE.
3136 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3137 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3141 /* The value user specified ....changed by config */
3142 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3143 /* string for sysctl */
3144 #define NUMA_ZONELIST_ORDER_LEN 16
3145 char numa_zonelist_order[16] = "default";
3148 * interface for configure zonelist ordering.
3149 * command line option "numa_zonelist_order"
3150 * = "[dD]efault - default, automatic configuration.
3151 * = "[nN]ode - order by node locality, then by zone within node
3152 * = "[zZ]one - order by zone, then by locality within zone
3155 static int __parse_numa_zonelist_order(char *s)
3157 if (*s == 'd' || *s == 'D') {
3158 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3159 } else if (*s == 'n' || *s == 'N') {
3160 user_zonelist_order = ZONELIST_ORDER_NODE;
3161 } else if (*s == 'z' || *s == 'Z') {
3162 user_zonelist_order = ZONELIST_ORDER_ZONE;
3165 "Ignoring invalid numa_zonelist_order value: "
3172 static __init int setup_numa_zonelist_order(char *s)
3179 ret = __parse_numa_zonelist_order(s);
3181 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3185 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3188 * sysctl handler for numa_zonelist_order
3190 int numa_zonelist_order_handler(ctl_table *table, int write,
3191 void __user *buffer, size_t *length,
3194 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3196 static DEFINE_MUTEX(zl_order_mutex);
3198 mutex_lock(&zl_order_mutex);
3200 strcpy(saved_string, (char*)table->data);
3201 ret = proc_dostring(table, write, buffer, length, ppos);
3205 int oldval = user_zonelist_order;
3206 if (__parse_numa_zonelist_order((char*)table->data)) {
3208 * bogus value. restore saved string
3210 strncpy((char*)table->data, saved_string,
3211 NUMA_ZONELIST_ORDER_LEN);
3212 user_zonelist_order = oldval;
3213 } else if (oldval != user_zonelist_order) {
3214 mutex_lock(&zonelists_mutex);
3215 build_all_zonelists(NULL, NULL);
3216 mutex_unlock(&zonelists_mutex);
3220 mutex_unlock(&zl_order_mutex);
3225 #define MAX_NODE_LOAD (nr_online_nodes)
3226 static int node_load[MAX_NUMNODES];
3229 * find_next_best_node - find the next node that should appear in a given node's fallback list
3230 * @node: node whose fallback list we're appending
3231 * @used_node_mask: nodemask_t of already used nodes
3233 * We use a number of factors to determine which is the next node that should
3234 * appear on a given node's fallback list. The node should not have appeared
3235 * already in @node's fallback list, and it should be the next closest node
3236 * according to the distance array (which contains arbitrary distance values
3237 * from each node to each node in the system), and should also prefer nodes
3238 * with no CPUs, since presumably they'll have very little allocation pressure
3239 * on them otherwise.
3240 * It returns -1 if no node is found.
3242 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3245 int min_val = INT_MAX;
3247 const struct cpumask *tmp = cpumask_of_node(0);
3249 /* Use the local node if we haven't already */
3250 if (!node_isset(node, *used_node_mask)) {
3251 node_set(node, *used_node_mask);
3255 for_each_node_state(n, N_MEMORY) {
3257 /* Don't want a node to appear more than once */
3258 if (node_isset(n, *used_node_mask))
3261 /* Use the distance array to find the distance */
3262 val = node_distance(node, n);
3264 /* Penalize nodes under us ("prefer the next node") */
3267 /* Give preference to headless and unused nodes */
3268 tmp = cpumask_of_node(n);
3269 if (!cpumask_empty(tmp))
3270 val += PENALTY_FOR_NODE_WITH_CPUS;
3272 /* Slight preference for less loaded node */
3273 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3274 val += node_load[n];
3276 if (val < min_val) {
3283 node_set(best_node, *used_node_mask);
3290 * Build zonelists ordered by node and zones within node.
3291 * This results in maximum locality--normal zone overflows into local
3292 * DMA zone, if any--but risks exhausting DMA zone.
3294 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3297 struct zonelist *zonelist;
3299 zonelist = &pgdat->node_zonelists[0];
3300 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3302 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3304 zonelist->_zonerefs[j].zone = NULL;
3305 zonelist->_zonerefs[j].zone_idx = 0;
3309 * Build gfp_thisnode zonelists
3311 static void build_thisnode_zonelists(pg_data_t *pgdat)
3314 struct zonelist *zonelist;
3316 zonelist = &pgdat->node_zonelists[1];
3317 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3318 zonelist->_zonerefs[j].zone = NULL;
3319 zonelist->_zonerefs[j].zone_idx = 0;
3323 * Build zonelists ordered by zone and nodes within zones.
3324 * This results in conserving DMA zone[s] until all Normal memory is
3325 * exhausted, but results in overflowing to remote node while memory
3326 * may still exist in local DMA zone.
3328 static int node_order[MAX_NUMNODES];
3330 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3333 int zone_type; /* needs to be signed */
3335 struct zonelist *zonelist;
3337 zonelist = &pgdat->node_zonelists[0];
3339 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3340 for (j = 0; j < nr_nodes; j++) {
3341 node = node_order[j];
3342 z = &NODE_DATA(node)->node_zones[zone_type];
3343 if (populated_zone(z)) {
3345 &zonelist->_zonerefs[pos++]);
3346 check_highest_zone(zone_type);
3350 zonelist->_zonerefs[pos].zone = NULL;
3351 zonelist->_zonerefs[pos].zone_idx = 0;
3354 static int default_zonelist_order(void)
3357 unsigned long low_kmem_size,total_size;
3361 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3362 * If they are really small and used heavily, the system can fall
3363 * into OOM very easily.
3364 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3366 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3369 for_each_online_node(nid) {
3370 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3371 z = &NODE_DATA(nid)->node_zones[zone_type];
3372 if (populated_zone(z)) {
3373 if (zone_type < ZONE_NORMAL)
3374 low_kmem_size += z->present_pages;
3375 total_size += z->present_pages;
3376 } else if (zone_type == ZONE_NORMAL) {
3378 * If any node has only lowmem, then node order
3379 * is preferred to allow kernel allocations
3380 * locally; otherwise, they can easily infringe
3381 * on other nodes when there is an abundance of
3382 * lowmem available to allocate from.
3384 return ZONELIST_ORDER_NODE;
3388 if (!low_kmem_size || /* there are no DMA area. */
3389 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3390 return ZONELIST_ORDER_NODE;
3392 * look into each node's config.
3393 * If there is a node whose DMA/DMA32 memory is very big area on
3394 * local memory, NODE_ORDER may be suitable.
3396 average_size = total_size /
3397 (nodes_weight(node_states[N_MEMORY]) + 1);
3398 for_each_online_node(nid) {
3401 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3402 z = &NODE_DATA(nid)->node_zones[zone_type];
3403 if (populated_zone(z)) {
3404 if (zone_type < ZONE_NORMAL)
3405 low_kmem_size += z->present_pages;
3406 total_size += z->present_pages;
3409 if (low_kmem_size &&
3410 total_size > average_size && /* ignore small node */
3411 low_kmem_size > total_size * 70/100)
3412 return ZONELIST_ORDER_NODE;
3414 return ZONELIST_ORDER_ZONE;
3417 static void set_zonelist_order(void)
3419 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3420 current_zonelist_order = default_zonelist_order();
3422 current_zonelist_order = user_zonelist_order;
3425 static void build_zonelists(pg_data_t *pgdat)
3429 nodemask_t used_mask;
3430 int local_node, prev_node;
3431 struct zonelist *zonelist;
3432 int order = current_zonelist_order;
3434 /* initialize zonelists */
3435 for (i = 0; i < MAX_ZONELISTS; i++) {
3436 zonelist = pgdat->node_zonelists + i;
3437 zonelist->_zonerefs[0].zone = NULL;
3438 zonelist->_zonerefs[0].zone_idx = 0;
3441 /* NUMA-aware ordering of nodes */
3442 local_node = pgdat->node_id;
3443 load = nr_online_nodes;
3444 prev_node = local_node;
3445 nodes_clear(used_mask);
3447 memset(node_order, 0, sizeof(node_order));
3450 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3452 * We don't want to pressure a particular node.
3453 * So adding penalty to the first node in same
3454 * distance group to make it round-robin.
3456 if (node_distance(local_node, node) !=
3457 node_distance(local_node, prev_node))
3458 node_load[node] = load;
3462 if (order == ZONELIST_ORDER_NODE)
3463 build_zonelists_in_node_order(pgdat, node);
3465 node_order[j++] = node; /* remember order */
3468 if (order == ZONELIST_ORDER_ZONE) {
3469 /* calculate node order -- i.e., DMA last! */
3470 build_zonelists_in_zone_order(pgdat, j);
3473 build_thisnode_zonelists(pgdat);
3476 /* Construct the zonelist performance cache - see further mmzone.h */
3477 static void build_zonelist_cache(pg_data_t *pgdat)
3479 struct zonelist *zonelist;
3480 struct zonelist_cache *zlc;
3483 zonelist = &pgdat->node_zonelists[0];
3484 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3485 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3486 for (z = zonelist->_zonerefs; z->zone; z++)
3487 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3490 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3492 * Return node id of node used for "local" allocations.
3493 * I.e., first node id of first zone in arg node's generic zonelist.
3494 * Used for initializing percpu 'numa_mem', which is used primarily
3495 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3497 int local_memory_node(int node)
3501 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3502 gfp_zone(GFP_KERNEL),
3509 #else /* CONFIG_NUMA */
3511 static void set_zonelist_order(void)
3513 current_zonelist_order = ZONELIST_ORDER_ZONE;
3516 static void build_zonelists(pg_data_t *pgdat)
3518 int node, local_node;
3520 struct zonelist *zonelist;
3522 local_node = pgdat->node_id;
3524 zonelist = &pgdat->node_zonelists[0];
3525 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3528 * Now we build the zonelist so that it contains the zones
3529 * of all the other nodes.
3530 * We don't want to pressure a particular node, so when
3531 * building the zones for node N, we make sure that the
3532 * zones coming right after the local ones are those from
3533 * node N+1 (modulo N)
3535 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3536 if (!node_online(node))
3538 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3541 for (node = 0; node < local_node; node++) {
3542 if (!node_online(node))
3544 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3548 zonelist->_zonerefs[j].zone = NULL;
3549 zonelist->_zonerefs[j].zone_idx = 0;
3552 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3553 static void build_zonelist_cache(pg_data_t *pgdat)
3555 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3558 #endif /* CONFIG_NUMA */
3561 * Boot pageset table. One per cpu which is going to be used for all
3562 * zones and all nodes. The parameters will be set in such a way
3563 * that an item put on a list will immediately be handed over to
3564 * the buddy list. This is safe since pageset manipulation is done
3565 * with interrupts disabled.
3567 * The boot_pagesets must be kept even after bootup is complete for
3568 * unused processors and/or zones. They do play a role for bootstrapping
3569 * hotplugged processors.
3571 * zoneinfo_show() and maybe other functions do
3572 * not check if the processor is online before following the pageset pointer.
3573 * Other parts of the kernel may not check if the zone is available.
3575 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3576 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3577 static void setup_zone_pageset(struct zone *zone);
3580 * Global mutex to protect against size modification of zonelists
3581 * as well as to serialize pageset setup for the new populated zone.
3583 DEFINE_MUTEX(zonelists_mutex);
3585 /* return values int ....just for stop_machine() */
3586 static int __build_all_zonelists(void *data)
3590 pg_data_t *self = data;
3593 memset(node_load, 0, sizeof(node_load));
3596 if (self && !node_online(self->node_id)) {
3597 build_zonelists(self);
3598 build_zonelist_cache(self);
3601 for_each_online_node(nid) {
3602 pg_data_t *pgdat = NODE_DATA(nid);
3604 build_zonelists(pgdat);
3605 build_zonelist_cache(pgdat);
3609 * Initialize the boot_pagesets that are going to be used
3610 * for bootstrapping processors. The real pagesets for
3611 * each zone will be allocated later when the per cpu
3612 * allocator is available.
3614 * boot_pagesets are used also for bootstrapping offline
3615 * cpus if the system is already booted because the pagesets
3616 * are needed to initialize allocators on a specific cpu too.
3617 * F.e. the percpu allocator needs the page allocator which
3618 * needs the percpu allocator in order to allocate its pagesets
3619 * (a chicken-egg dilemma).
3621 for_each_possible_cpu(cpu) {
3622 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3624 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3626 * We now know the "local memory node" for each node--
3627 * i.e., the node of the first zone in the generic zonelist.
3628 * Set up numa_mem percpu variable for on-line cpus. During
3629 * boot, only the boot cpu should be on-line; we'll init the
3630 * secondary cpus' numa_mem as they come on-line. During
3631 * node/memory hotplug, we'll fixup all on-line cpus.
3633 if (cpu_online(cpu))
3634 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3642 * Called with zonelists_mutex held always
3643 * unless system_state == SYSTEM_BOOTING.
3645 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3647 set_zonelist_order();
3649 if (system_state == SYSTEM_BOOTING) {
3650 __build_all_zonelists(NULL);
3651 mminit_verify_zonelist();
3652 cpuset_init_current_mems_allowed();
3654 /* we have to stop all cpus to guarantee there is no user
3656 #ifdef CONFIG_MEMORY_HOTPLUG
3658 setup_zone_pageset(zone);
3660 stop_machine(__build_all_zonelists, pgdat, NULL);
3661 /* cpuset refresh routine should be here */
3663 vm_total_pages = nr_free_pagecache_pages();
3665 * Disable grouping by mobility if the number of pages in the
3666 * system is too low to allow the mechanism to work. It would be
3667 * more accurate, but expensive to check per-zone. This check is
3668 * made on memory-hotadd so a system can start with mobility
3669 * disabled and enable it later
3671 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3672 page_group_by_mobility_disabled = 1;
3674 page_group_by_mobility_disabled = 0;
3676 printk("Built %i zonelists in %s order, mobility grouping %s. "
3677 "Total pages: %ld\n",
3679 zonelist_order_name[current_zonelist_order],
3680 page_group_by_mobility_disabled ? "off" : "on",
3683 printk("Policy zone: %s\n", zone_names[policy_zone]);
3688 * Helper functions to size the waitqueue hash table.
3689 * Essentially these want to choose hash table sizes sufficiently
3690 * large so that collisions trying to wait on pages are rare.
3691 * But in fact, the number of active page waitqueues on typical
3692 * systems is ridiculously low, less than 200. So this is even
3693 * conservative, even though it seems large.
3695 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3696 * waitqueues, i.e. the size of the waitq table given the number of pages.
3698 #define PAGES_PER_WAITQUEUE 256
3700 #ifndef CONFIG_MEMORY_HOTPLUG
3701 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3703 unsigned long size = 1;
3705 pages /= PAGES_PER_WAITQUEUE;
3707 while (size < pages)
3711 * Once we have dozens or even hundreds of threads sleeping
3712 * on IO we've got bigger problems than wait queue collision.
3713 * Limit the size of the wait table to a reasonable size.
3715 size = min(size, 4096UL);
3717 return max(size, 4UL);
3721 * A zone's size might be changed by hot-add, so it is not possible to determine
3722 * a suitable size for its wait_table. So we use the maximum size now.
3724 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3726 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3727 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3728 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3730 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3731 * or more by the traditional way. (See above). It equals:
3733 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3734 * ia64(16K page size) : = ( 8G + 4M)byte.
3735 * powerpc (64K page size) : = (32G +16M)byte.
3737 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3744 * This is an integer logarithm so that shifts can be used later
3745 * to extract the more random high bits from the multiplicative
3746 * hash function before the remainder is taken.
3748 static inline unsigned long wait_table_bits(unsigned long size)
3753 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3756 * Check if a pageblock contains reserved pages
3758 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3762 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3763 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3770 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3771 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3772 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3773 * higher will lead to a bigger reserve which will get freed as contiguous
3774 * blocks as reclaim kicks in
3776 static void setup_zone_migrate_reserve(struct zone *zone)
3778 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3780 unsigned long block_migratetype;
3784 * Get the start pfn, end pfn and the number of blocks to reserve
3785 * We have to be careful to be aligned to pageblock_nr_pages to
3786 * make sure that we always check pfn_valid for the first page in
3789 start_pfn = zone->zone_start_pfn;
3790 end_pfn = start_pfn + zone->spanned_pages;
3791 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3792 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3796 * Reserve blocks are generally in place to help high-order atomic
3797 * allocations that are short-lived. A min_free_kbytes value that
3798 * would result in more than 2 reserve blocks for atomic allocations
3799 * is assumed to be in place to help anti-fragmentation for the
3800 * future allocation of hugepages at runtime.
3802 reserve = min(2, reserve);
3804 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3805 if (!pfn_valid(pfn))
3807 page = pfn_to_page(pfn);
3809 /* Watch out for overlapping nodes */
3810 if (page_to_nid(page) != zone_to_nid(zone))
3813 block_migratetype = get_pageblock_migratetype(page);
3815 /* Only test what is necessary when the reserves are not met */
3818 * Blocks with reserved pages will never free, skip
3821 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3822 if (pageblock_is_reserved(pfn, block_end_pfn))
3825 /* If this block is reserved, account for it */
3826 if (block_migratetype == MIGRATE_RESERVE) {
3831 /* Suitable for reserving if this block is movable */
3832 if (block_migratetype == MIGRATE_MOVABLE) {
3833 set_pageblock_migratetype(page,
3835 move_freepages_block(zone, page,
3843 * If the reserve is met and this is a previous reserved block,
3846 if (block_migratetype == MIGRATE_RESERVE) {
3847 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3848 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3854 * Initially all pages are reserved - free ones are freed
3855 * up by free_all_bootmem() once the early boot process is
3856 * done. Non-atomic initialization, single-pass.
3858 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3859 unsigned long start_pfn, enum memmap_context context)
3862 unsigned long end_pfn = start_pfn + size;
3866 if (highest_memmap_pfn < end_pfn - 1)
3867 highest_memmap_pfn = end_pfn - 1;
3869 z = &NODE_DATA(nid)->node_zones[zone];
3870 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3872 * There can be holes in boot-time mem_map[]s
3873 * handed to this function. They do not
3874 * exist on hotplugged memory.
3876 if (context == MEMMAP_EARLY) {
3877 if (!early_pfn_valid(pfn))
3879 if (!early_pfn_in_nid(pfn, nid))
3882 page = pfn_to_page(pfn);
3883 set_page_links(page, zone, nid, pfn);
3884 mminit_verify_page_links(page, zone, nid, pfn);
3885 init_page_count(page);
3886 reset_page_mapcount(page);
3887 reset_page_last_nid(page);
3888 SetPageReserved(page);
3890 * Mark the block movable so that blocks are reserved for
3891 * movable at startup. This will force kernel allocations
3892 * to reserve their blocks rather than leaking throughout
3893 * the address space during boot when many long-lived
3894 * kernel allocations are made. Later some blocks near
3895 * the start are marked MIGRATE_RESERVE by
3896 * setup_zone_migrate_reserve()
3898 * bitmap is created for zone's valid pfn range. but memmap
3899 * can be created for invalid pages (for alignment)
3900 * check here not to call set_pageblock_migratetype() against
3903 if ((z->zone_start_pfn <= pfn)
3904 && (pfn < z->zone_start_pfn + z->spanned_pages)
3905 && !(pfn & (pageblock_nr_pages - 1)))
3906 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3908 INIT_LIST_HEAD(&page->lru);
3909 #ifdef WANT_PAGE_VIRTUAL
3910 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3911 if (!is_highmem_idx(zone))
3912 set_page_address(page, __va(pfn << PAGE_SHIFT));
3917 static void __meminit zone_init_free_lists(struct zone *zone)
3920 for_each_migratetype_order(order, t) {
3921 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3922 zone->free_area[order].nr_free = 0;
3926 #ifndef __HAVE_ARCH_MEMMAP_INIT
3927 #define memmap_init(size, nid, zone, start_pfn) \
3928 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3931 static int __meminit zone_batchsize(struct zone *zone)
3937 * The per-cpu-pages pools are set to around 1000th of the
3938 * size of the zone. But no more than 1/2 of a meg.
3940 * OK, so we don't know how big the cache is. So guess.
3942 batch = zone->present_pages / 1024;
3943 if (batch * PAGE_SIZE > 512 * 1024)
3944 batch = (512 * 1024) / PAGE_SIZE;
3945 batch /= 4; /* We effectively *= 4 below */
3950 * Clamp the batch to a 2^n - 1 value. Having a power
3951 * of 2 value was found to be more likely to have
3952 * suboptimal cache aliasing properties in some cases.
3954 * For example if 2 tasks are alternately allocating
3955 * batches of pages, one task can end up with a lot
3956 * of pages of one half of the possible page colors
3957 * and the other with pages of the other colors.
3959 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3964 /* The deferral and batching of frees should be suppressed under NOMMU
3967 * The problem is that NOMMU needs to be able to allocate large chunks
3968 * of contiguous memory as there's no hardware page translation to
3969 * assemble apparent contiguous memory from discontiguous pages.
3971 * Queueing large contiguous runs of pages for batching, however,
3972 * causes the pages to actually be freed in smaller chunks. As there
3973 * can be a significant delay between the individual batches being
3974 * recycled, this leads to the once large chunks of space being
3975 * fragmented and becoming unavailable for high-order allocations.
3981 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3983 struct per_cpu_pages *pcp;
3986 memset(p, 0, sizeof(*p));
3990 pcp->high = 6 * batch;
3991 pcp->batch = max(1UL, 1 * batch);
3992 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3993 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3997 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3998 * to the value high for the pageset p.
4001 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
4004 struct per_cpu_pages *pcp;
4008 pcp->batch = max(1UL, high/4);
4009 if ((high/4) > (PAGE_SHIFT * 8))
4010 pcp->batch = PAGE_SHIFT * 8;
4013 static void __meminit setup_zone_pageset(struct zone *zone)
4017 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4019 for_each_possible_cpu(cpu) {
4020 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4022 setup_pageset(pcp, zone_batchsize(zone));
4024 if (percpu_pagelist_fraction)
4025 setup_pagelist_highmark(pcp,
4026 (zone->present_pages /
4027 percpu_pagelist_fraction));
4032 * Allocate per cpu pagesets and initialize them.
4033 * Before this call only boot pagesets were available.
4035 void __init setup_per_cpu_pageset(void)
4039 for_each_populated_zone(zone)
4040 setup_zone_pageset(zone);
4043 static noinline __init_refok
4044 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4047 struct pglist_data *pgdat = zone->zone_pgdat;
4051 * The per-page waitqueue mechanism uses hashed waitqueues
4054 zone->wait_table_hash_nr_entries =
4055 wait_table_hash_nr_entries(zone_size_pages);
4056 zone->wait_table_bits =
4057 wait_table_bits(zone->wait_table_hash_nr_entries);
4058 alloc_size = zone->wait_table_hash_nr_entries
4059 * sizeof(wait_queue_head_t);
4061 if (!slab_is_available()) {
4062 zone->wait_table = (wait_queue_head_t *)
4063 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4066 * This case means that a zone whose size was 0 gets new memory
4067 * via memory hot-add.
4068 * But it may be the case that a new node was hot-added. In
4069 * this case vmalloc() will not be able to use this new node's
4070 * memory - this wait_table must be initialized to use this new
4071 * node itself as well.
4072 * To use this new node's memory, further consideration will be
4075 zone->wait_table = vmalloc(alloc_size);
4077 if (!zone->wait_table)
4080 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4081 init_waitqueue_head(zone->wait_table + i);
4086 static __meminit void zone_pcp_init(struct zone *zone)
4089 * per cpu subsystem is not up at this point. The following code
4090 * relies on the ability of the linker to provide the
4091 * offset of a (static) per cpu variable into the per cpu area.
4093 zone->pageset = &boot_pageset;
4095 if (zone->present_pages)
4096 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4097 zone->name, zone->present_pages,
4098 zone_batchsize(zone));
4101 int __meminit init_currently_empty_zone(struct zone *zone,
4102 unsigned long zone_start_pfn,
4104 enum memmap_context context)
4106 struct pglist_data *pgdat = zone->zone_pgdat;
4108 ret = zone_wait_table_init(zone, size);
4111 pgdat->nr_zones = zone_idx(zone) + 1;
4113 zone->zone_start_pfn = zone_start_pfn;
4115 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4116 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4118 (unsigned long)zone_idx(zone),
4119 zone_start_pfn, (zone_start_pfn + size));
4121 zone_init_free_lists(zone);
4126 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4127 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4129 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4130 * Architectures may implement their own version but if add_active_range()
4131 * was used and there are no special requirements, this is a convenient
4134 int __meminit __early_pfn_to_nid(unsigned long pfn)
4136 unsigned long start_pfn, end_pfn;
4139 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4140 if (start_pfn <= pfn && pfn < end_pfn)
4142 /* This is a memory hole */
4145 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4147 int __meminit early_pfn_to_nid(unsigned long pfn)
4151 nid = __early_pfn_to_nid(pfn);
4154 /* just returns 0 */
4158 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4159 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4163 nid = __early_pfn_to_nid(pfn);
4164 if (nid >= 0 && nid != node)
4171 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4172 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4173 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4175 * If an architecture guarantees that all ranges registered with
4176 * add_active_ranges() contain no holes and may be freed, this
4177 * this function may be used instead of calling free_bootmem() manually.
4179 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4181 unsigned long start_pfn, end_pfn;
4184 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4185 start_pfn = min(start_pfn, max_low_pfn);
4186 end_pfn = min(end_pfn, max_low_pfn);
4188 if (start_pfn < end_pfn)
4189 free_bootmem_node(NODE_DATA(this_nid),
4190 PFN_PHYS(start_pfn),
4191 (end_pfn - start_pfn) << PAGE_SHIFT);
4196 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4197 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4199 * If an architecture guarantees that all ranges registered with
4200 * add_active_ranges() contain no holes and may be freed, this
4201 * function may be used instead of calling memory_present() manually.
4203 void __init sparse_memory_present_with_active_regions(int nid)
4205 unsigned long start_pfn, end_pfn;
4208 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4209 memory_present(this_nid, start_pfn, end_pfn);
4213 * get_pfn_range_for_nid - Return the start and end page frames for a node
4214 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4215 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4216 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4218 * It returns the start and end page frame of a node based on information
4219 * provided by an arch calling add_active_range(). If called for a node
4220 * with no available memory, a warning is printed and the start and end
4223 void __meminit get_pfn_range_for_nid(unsigned int nid,
4224 unsigned long *start_pfn, unsigned long *end_pfn)
4226 unsigned long this_start_pfn, this_end_pfn;
4232 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4233 *start_pfn = min(*start_pfn, this_start_pfn);
4234 *end_pfn = max(*end_pfn, this_end_pfn);
4237 if (*start_pfn == -1UL)
4242 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4243 * assumption is made that zones within a node are ordered in monotonic
4244 * increasing memory addresses so that the "highest" populated zone is used
4246 static void __init find_usable_zone_for_movable(void)
4249 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4250 if (zone_index == ZONE_MOVABLE)
4253 if (arch_zone_highest_possible_pfn[zone_index] >
4254 arch_zone_lowest_possible_pfn[zone_index])
4258 VM_BUG_ON(zone_index == -1);
4259 movable_zone = zone_index;
4263 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4264 * because it is sized independent of architecture. Unlike the other zones,
4265 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4266 * in each node depending on the size of each node and how evenly kernelcore
4267 * is distributed. This helper function adjusts the zone ranges
4268 * provided by the architecture for a given node by using the end of the
4269 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4270 * zones within a node are in order of monotonic increases memory addresses
4272 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4273 unsigned long zone_type,
4274 unsigned long node_start_pfn,
4275 unsigned long node_end_pfn,
4276 unsigned long *zone_start_pfn,
4277 unsigned long *zone_end_pfn)
4279 /* Only adjust if ZONE_MOVABLE is on this node */
4280 if (zone_movable_pfn[nid]) {
4281 /* Size ZONE_MOVABLE */
4282 if (zone_type == ZONE_MOVABLE) {
4283 *zone_start_pfn = zone_movable_pfn[nid];
4284 *zone_end_pfn = min(node_end_pfn,
4285 arch_zone_highest_possible_pfn[movable_zone]);
4287 /* Adjust for ZONE_MOVABLE starting within this range */
4288 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4289 *zone_end_pfn > zone_movable_pfn[nid]) {
4290 *zone_end_pfn = zone_movable_pfn[nid];
4292 /* Check if this whole range is within ZONE_MOVABLE */
4293 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4294 *zone_start_pfn = *zone_end_pfn;
4299 * Return the number of pages a zone spans in a node, including holes
4300 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4302 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4303 unsigned long zone_type,
4304 unsigned long *ignored)
4306 unsigned long node_start_pfn, node_end_pfn;
4307 unsigned long zone_start_pfn, zone_end_pfn;
4309 /* Get the start and end of the node and zone */
4310 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4311 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4312 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4313 adjust_zone_range_for_zone_movable(nid, zone_type,
4314 node_start_pfn, node_end_pfn,
4315 &zone_start_pfn, &zone_end_pfn);
4317 /* Check that this node has pages within the zone's required range */
4318 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4321 /* Move the zone boundaries inside the node if necessary */
4322 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4323 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4325 /* Return the spanned pages */
4326 return zone_end_pfn - zone_start_pfn;
4330 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4331 * then all holes in the requested range will be accounted for.
4333 unsigned long __meminit __absent_pages_in_range(int nid,
4334 unsigned long range_start_pfn,
4335 unsigned long range_end_pfn)
4337 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4338 unsigned long start_pfn, end_pfn;
4341 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4342 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4343 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4344 nr_absent -= end_pfn - start_pfn;
4350 * absent_pages_in_range - Return number of page frames in holes within a range
4351 * @start_pfn: The start PFN to start searching for holes
4352 * @end_pfn: The end PFN to stop searching for holes
4354 * It returns the number of pages frames in memory holes within a range.
4356 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4357 unsigned long end_pfn)
4359 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4362 /* Return the number of page frames in holes in a zone on a node */
4363 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4364 unsigned long zone_type,
4365 unsigned long *ignored)
4367 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4368 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4369 unsigned long node_start_pfn, node_end_pfn;
4370 unsigned long zone_start_pfn, zone_end_pfn;
4372 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4373 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4374 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4376 adjust_zone_range_for_zone_movable(nid, zone_type,
4377 node_start_pfn, node_end_pfn,
4378 &zone_start_pfn, &zone_end_pfn);
4379 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4382 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4383 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4384 unsigned long zone_type,
4385 unsigned long *zones_size)
4387 return zones_size[zone_type];
4390 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4391 unsigned long zone_type,
4392 unsigned long *zholes_size)
4397 return zholes_size[zone_type];
4400 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4402 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4403 unsigned long *zones_size, unsigned long *zholes_size)
4405 unsigned long realtotalpages, totalpages = 0;
4408 for (i = 0; i < MAX_NR_ZONES; i++)
4409 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4411 pgdat->node_spanned_pages = totalpages;
4413 realtotalpages = totalpages;
4414 for (i = 0; i < MAX_NR_ZONES; i++)
4416 zone_absent_pages_in_node(pgdat->node_id, i,
4418 pgdat->node_present_pages = realtotalpages;
4419 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4423 #ifndef CONFIG_SPARSEMEM
4425 * Calculate the size of the zone->blockflags rounded to an unsigned long
4426 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4427 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4428 * round what is now in bits to nearest long in bits, then return it in
4431 static unsigned long __init usemap_size(unsigned long zonesize)
4433 unsigned long usemapsize;
4435 usemapsize = roundup(zonesize, pageblock_nr_pages);
4436 usemapsize = usemapsize >> pageblock_order;
4437 usemapsize *= NR_PAGEBLOCK_BITS;
4438 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4440 return usemapsize / 8;
4443 static void __init setup_usemap(struct pglist_data *pgdat,
4444 struct zone *zone, unsigned long zonesize)
4446 unsigned long usemapsize = usemap_size(zonesize);
4447 zone->pageblock_flags = NULL;
4449 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4453 static inline void setup_usemap(struct pglist_data *pgdat,
4454 struct zone *zone, unsigned long zonesize) {}
4455 #endif /* CONFIG_SPARSEMEM */
4457 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4459 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4460 void __init set_pageblock_order(void)
4464 /* Check that pageblock_nr_pages has not already been setup */
4465 if (pageblock_order)
4468 if (HPAGE_SHIFT > PAGE_SHIFT)
4469 order = HUGETLB_PAGE_ORDER;
4471 order = MAX_ORDER - 1;
4474 * Assume the largest contiguous order of interest is a huge page.
4475 * This value may be variable depending on boot parameters on IA64 and
4478 pageblock_order = order;
4480 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4483 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4484 * is unused as pageblock_order is set at compile-time. See
4485 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4488 void __init set_pageblock_order(void)
4492 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4494 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4495 unsigned long present_pages)
4497 unsigned long pages = spanned_pages;
4500 * Provide a more accurate estimation if there are holes within
4501 * the zone and SPARSEMEM is in use. If there are holes within the
4502 * zone, each populated memory region may cost us one or two extra
4503 * memmap pages due to alignment because memmap pages for each
4504 * populated regions may not naturally algined on page boundary.
4505 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4507 if (spanned_pages > present_pages + (present_pages >> 4) &&
4508 IS_ENABLED(CONFIG_SPARSEMEM))
4509 pages = present_pages;
4511 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4515 * Set up the zone data structures:
4516 * - mark all pages reserved
4517 * - mark all memory queues empty
4518 * - clear the memory bitmaps
4520 * NOTE: pgdat should get zeroed by caller.
4522 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4523 unsigned long *zones_size, unsigned long *zholes_size)
4526 int nid = pgdat->node_id;
4527 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4530 pgdat_resize_init(pgdat);
4531 #ifdef CONFIG_NUMA_BALANCING
4532 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4533 pgdat->numabalancing_migrate_nr_pages = 0;
4534 pgdat->numabalancing_migrate_next_window = jiffies;
4536 init_waitqueue_head(&pgdat->kswapd_wait);
4537 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4538 pgdat_page_cgroup_init(pgdat);
4540 for (j = 0; j < MAX_NR_ZONES; j++) {
4541 struct zone *zone = pgdat->node_zones + j;
4542 unsigned long size, realsize, freesize, memmap_pages;
4544 size = zone_spanned_pages_in_node(nid, j, zones_size);
4545 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4549 * Adjust freesize so that it accounts for how much memory
4550 * is used by this zone for memmap. This affects the watermark
4551 * and per-cpu initialisations
4553 memmap_pages = calc_memmap_size(size, realsize);
4554 if (freesize >= memmap_pages) {
4555 freesize -= memmap_pages;
4558 " %s zone: %lu pages used for memmap\n",
4559 zone_names[j], memmap_pages);
4562 " %s zone: %lu pages exceeds freesize %lu\n",
4563 zone_names[j], memmap_pages, freesize);
4565 /* Account for reserved pages */
4566 if (j == 0 && freesize > dma_reserve) {
4567 freesize -= dma_reserve;
4568 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4569 zone_names[0], dma_reserve);
4572 if (!is_highmem_idx(j))
4573 nr_kernel_pages += freesize;
4574 /* Charge for highmem memmap if there are enough kernel pages */
4575 else if (nr_kernel_pages > memmap_pages * 2)
4576 nr_kernel_pages -= memmap_pages;
4577 nr_all_pages += freesize;
4579 zone->spanned_pages = size;
4580 zone->present_pages = freesize;
4582 * Set an approximate value for lowmem here, it will be adjusted
4583 * when the bootmem allocator frees pages into the buddy system.
4584 * And all highmem pages will be managed by the buddy system.
4586 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4589 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4591 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4593 zone->name = zone_names[j];
4594 spin_lock_init(&zone->lock);
4595 spin_lock_init(&zone->lru_lock);
4596 zone_seqlock_init(zone);
4597 zone->zone_pgdat = pgdat;
4599 zone_pcp_init(zone);
4600 lruvec_init(&zone->lruvec);
4604 set_pageblock_order();
4605 setup_usemap(pgdat, zone, size);
4606 ret = init_currently_empty_zone(zone, zone_start_pfn,
4607 size, MEMMAP_EARLY);
4609 memmap_init(size, nid, j, zone_start_pfn);
4610 zone_start_pfn += size;
4614 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4616 /* Skip empty nodes */
4617 if (!pgdat->node_spanned_pages)
4620 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4621 /* ia64 gets its own node_mem_map, before this, without bootmem */
4622 if (!pgdat->node_mem_map) {
4623 unsigned long size, start, end;
4627 * The zone's endpoints aren't required to be MAX_ORDER
4628 * aligned but the node_mem_map endpoints must be in order
4629 * for the buddy allocator to function correctly.
4631 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4632 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4633 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4634 size = (end - start) * sizeof(struct page);
4635 map = alloc_remap(pgdat->node_id, size);
4637 map = alloc_bootmem_node_nopanic(pgdat, size);
4638 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4640 #ifndef CONFIG_NEED_MULTIPLE_NODES
4642 * With no DISCONTIG, the global mem_map is just set as node 0's
4644 if (pgdat == NODE_DATA(0)) {
4645 mem_map = NODE_DATA(0)->node_mem_map;
4646 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4647 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4648 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4649 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4652 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4655 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4656 unsigned long node_start_pfn, unsigned long *zholes_size)
4658 pg_data_t *pgdat = NODE_DATA(nid);
4660 /* pg_data_t should be reset to zero when it's allocated */
4661 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4663 pgdat->node_id = nid;
4664 pgdat->node_start_pfn = node_start_pfn;
4665 init_zone_allows_reclaim(nid);
4666 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4668 alloc_node_mem_map(pgdat);
4669 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4670 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4671 nid, (unsigned long)pgdat,
4672 (unsigned long)pgdat->node_mem_map);
4675 free_area_init_core(pgdat, zones_size, zholes_size);
4678 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4680 #if MAX_NUMNODES > 1
4682 * Figure out the number of possible node ids.
4684 static void __init setup_nr_node_ids(void)
4687 unsigned int highest = 0;
4689 for_each_node_mask(node, node_possible_map)
4691 nr_node_ids = highest + 1;
4694 static inline void setup_nr_node_ids(void)
4700 * node_map_pfn_alignment - determine the maximum internode alignment
4702 * This function should be called after node map is populated and sorted.
4703 * It calculates the maximum power of two alignment which can distinguish
4706 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4707 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4708 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4709 * shifted, 1GiB is enough and this function will indicate so.
4711 * This is used to test whether pfn -> nid mapping of the chosen memory
4712 * model has fine enough granularity to avoid incorrect mapping for the
4713 * populated node map.
4715 * Returns the determined alignment in pfn's. 0 if there is no alignment
4716 * requirement (single node).
4718 unsigned long __init node_map_pfn_alignment(void)
4720 unsigned long accl_mask = 0, last_end = 0;
4721 unsigned long start, end, mask;
4725 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4726 if (!start || last_nid < 0 || last_nid == nid) {
4733 * Start with a mask granular enough to pin-point to the
4734 * start pfn and tick off bits one-by-one until it becomes
4735 * too coarse to separate the current node from the last.
4737 mask = ~((1 << __ffs(start)) - 1);
4738 while (mask && last_end <= (start & (mask << 1)))
4741 /* accumulate all internode masks */
4745 /* convert mask to number of pages */
4746 return ~accl_mask + 1;
4749 /* Find the lowest pfn for a node */
4750 static unsigned long __init find_min_pfn_for_node(int nid)
4752 unsigned long min_pfn = ULONG_MAX;
4753 unsigned long start_pfn;
4756 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4757 min_pfn = min(min_pfn, start_pfn);
4759 if (min_pfn == ULONG_MAX) {
4761 "Could not find start_pfn for node %d\n", nid);
4769 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4771 * It returns the minimum PFN based on information provided via
4772 * add_active_range().
4774 unsigned long __init find_min_pfn_with_active_regions(void)
4776 return find_min_pfn_for_node(MAX_NUMNODES);
4780 * early_calculate_totalpages()
4781 * Sum pages in active regions for movable zone.
4782 * Populate N_MEMORY for calculating usable_nodes.
4784 static unsigned long __init early_calculate_totalpages(void)
4786 unsigned long totalpages = 0;
4787 unsigned long start_pfn, end_pfn;
4790 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4791 unsigned long pages = end_pfn - start_pfn;
4793 totalpages += pages;
4795 node_set_state(nid, N_MEMORY);
4801 * Find the PFN the Movable zone begins in each node. Kernel memory
4802 * is spread evenly between nodes as long as the nodes have enough
4803 * memory. When they don't, some nodes will have more kernelcore than
4806 static void __init find_zone_movable_pfns_for_nodes(void)
4809 unsigned long usable_startpfn;
4810 unsigned long kernelcore_node, kernelcore_remaining;
4811 /* save the state before borrow the nodemask */
4812 nodemask_t saved_node_state = node_states[N_MEMORY];
4813 unsigned long totalpages = early_calculate_totalpages();
4814 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4817 * If movablecore was specified, calculate what size of
4818 * kernelcore that corresponds so that memory usable for
4819 * any allocation type is evenly spread. If both kernelcore
4820 * and movablecore are specified, then the value of kernelcore
4821 * will be used for required_kernelcore if it's greater than
4822 * what movablecore would have allowed.
4824 if (required_movablecore) {
4825 unsigned long corepages;
4828 * Round-up so that ZONE_MOVABLE is at least as large as what
4829 * was requested by the user
4831 required_movablecore =
4832 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4833 corepages = totalpages - required_movablecore;
4835 required_kernelcore = max(required_kernelcore, corepages);
4838 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4839 if (!required_kernelcore)
4842 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4843 find_usable_zone_for_movable();
4844 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4847 /* Spread kernelcore memory as evenly as possible throughout nodes */
4848 kernelcore_node = required_kernelcore / usable_nodes;
4849 for_each_node_state(nid, N_MEMORY) {
4850 unsigned long start_pfn, end_pfn;
4853 * Recalculate kernelcore_node if the division per node
4854 * now exceeds what is necessary to satisfy the requested
4855 * amount of memory for the kernel
4857 if (required_kernelcore < kernelcore_node)
4858 kernelcore_node = required_kernelcore / usable_nodes;
4861 * As the map is walked, we track how much memory is usable
4862 * by the kernel using kernelcore_remaining. When it is
4863 * 0, the rest of the node is usable by ZONE_MOVABLE
4865 kernelcore_remaining = kernelcore_node;
4867 /* Go through each range of PFNs within this node */
4868 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4869 unsigned long size_pages;
4871 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4872 if (start_pfn >= end_pfn)
4875 /* Account for what is only usable for kernelcore */
4876 if (start_pfn < usable_startpfn) {
4877 unsigned long kernel_pages;
4878 kernel_pages = min(end_pfn, usable_startpfn)
4881 kernelcore_remaining -= min(kernel_pages,
4882 kernelcore_remaining);
4883 required_kernelcore -= min(kernel_pages,
4884 required_kernelcore);
4886 /* Continue if range is now fully accounted */
4887 if (end_pfn <= usable_startpfn) {
4890 * Push zone_movable_pfn to the end so
4891 * that if we have to rebalance
4892 * kernelcore across nodes, we will
4893 * not double account here
4895 zone_movable_pfn[nid] = end_pfn;
4898 start_pfn = usable_startpfn;
4902 * The usable PFN range for ZONE_MOVABLE is from
4903 * start_pfn->end_pfn. Calculate size_pages as the
4904 * number of pages used as kernelcore
4906 size_pages = end_pfn - start_pfn;
4907 if (size_pages > kernelcore_remaining)
4908 size_pages = kernelcore_remaining;
4909 zone_movable_pfn[nid] = start_pfn + size_pages;
4912 * Some kernelcore has been met, update counts and
4913 * break if the kernelcore for this node has been
4916 required_kernelcore -= min(required_kernelcore,
4918 kernelcore_remaining -= size_pages;
4919 if (!kernelcore_remaining)
4925 * If there is still required_kernelcore, we do another pass with one
4926 * less node in the count. This will push zone_movable_pfn[nid] further
4927 * along on the nodes that still have memory until kernelcore is
4931 if (usable_nodes && required_kernelcore > usable_nodes)
4934 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4935 for (nid = 0; nid < MAX_NUMNODES; nid++)
4936 zone_movable_pfn[nid] =
4937 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4940 /* restore the node_state */
4941 node_states[N_MEMORY] = saved_node_state;
4944 /* Any regular or high memory on that node ? */
4945 static void check_for_memory(pg_data_t *pgdat, int nid)
4947 enum zone_type zone_type;
4949 if (N_MEMORY == N_NORMAL_MEMORY)
4952 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
4953 struct zone *zone = &pgdat->node_zones[zone_type];
4954 if (zone->present_pages) {
4955 node_set_state(nid, N_HIGH_MEMORY);
4956 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
4957 zone_type <= ZONE_NORMAL)
4958 node_set_state(nid, N_NORMAL_MEMORY);
4965 * free_area_init_nodes - Initialise all pg_data_t and zone data
4966 * @max_zone_pfn: an array of max PFNs for each zone
4968 * This will call free_area_init_node() for each active node in the system.
4969 * Using the page ranges provided by add_active_range(), the size of each
4970 * zone in each node and their holes is calculated. If the maximum PFN
4971 * between two adjacent zones match, it is assumed that the zone is empty.
4972 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4973 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4974 * starts where the previous one ended. For example, ZONE_DMA32 starts
4975 * at arch_max_dma_pfn.
4977 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4979 unsigned long start_pfn, end_pfn;
4982 /* Record where the zone boundaries are */
4983 memset(arch_zone_lowest_possible_pfn, 0,
4984 sizeof(arch_zone_lowest_possible_pfn));
4985 memset(arch_zone_highest_possible_pfn, 0,
4986 sizeof(arch_zone_highest_possible_pfn));
4987 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4988 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4989 for (i = 1; i < MAX_NR_ZONES; i++) {
4990 if (i == ZONE_MOVABLE)
4992 arch_zone_lowest_possible_pfn[i] =
4993 arch_zone_highest_possible_pfn[i-1];
4994 arch_zone_highest_possible_pfn[i] =
4995 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4997 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4998 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5000 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5001 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5002 find_zone_movable_pfns_for_nodes();
5004 /* Print out the zone ranges */
5005 printk("Zone ranges:\n");
5006 for (i = 0; i < MAX_NR_ZONES; i++) {
5007 if (i == ZONE_MOVABLE)
5009 printk(KERN_CONT " %-8s ", zone_names[i]);
5010 if (arch_zone_lowest_possible_pfn[i] ==
5011 arch_zone_highest_possible_pfn[i])
5012 printk(KERN_CONT "empty\n");
5014 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5015 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5016 (arch_zone_highest_possible_pfn[i]
5017 << PAGE_SHIFT) - 1);
5020 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5021 printk("Movable zone start for each node\n");
5022 for (i = 0; i < MAX_NUMNODES; i++) {
5023 if (zone_movable_pfn[i])
5024 printk(" Node %d: %#010lx\n", i,
5025 zone_movable_pfn[i] << PAGE_SHIFT);
5028 /* Print out the early node map */
5029 printk("Early memory node ranges\n");
5030 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5031 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5032 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5034 /* Initialise every node */
5035 mminit_verify_pageflags_layout();
5036 setup_nr_node_ids();
5037 for_each_online_node(nid) {
5038 pg_data_t *pgdat = NODE_DATA(nid);
5039 free_area_init_node(nid, NULL,
5040 find_min_pfn_for_node(nid), NULL);
5042 /* Any memory on that node */
5043 if (pgdat->node_present_pages)
5044 node_set_state(nid, N_MEMORY);
5045 check_for_memory(pgdat, nid);
5049 static int __init cmdline_parse_core(char *p, unsigned long *core)
5051 unsigned long long coremem;
5055 coremem = memparse(p, &p);
5056 *core = coremem >> PAGE_SHIFT;
5058 /* Paranoid check that UL is enough for the coremem value */
5059 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5065 * kernelcore=size sets the amount of memory for use for allocations that
5066 * cannot be reclaimed or migrated.
5068 static int __init cmdline_parse_kernelcore(char *p)
5070 return cmdline_parse_core(p, &required_kernelcore);
5074 * movablecore=size sets the amount of memory for use for allocations that
5075 * can be reclaimed or migrated.
5077 static int __init cmdline_parse_movablecore(char *p)
5079 return cmdline_parse_core(p, &required_movablecore);
5082 early_param("kernelcore", cmdline_parse_kernelcore);
5083 early_param("movablecore", cmdline_parse_movablecore);
5085 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5088 * set_dma_reserve - set the specified number of pages reserved in the first zone
5089 * @new_dma_reserve: The number of pages to mark reserved
5091 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5092 * In the DMA zone, a significant percentage may be consumed by kernel image
5093 * and other unfreeable allocations which can skew the watermarks badly. This
5094 * function may optionally be used to account for unfreeable pages in the
5095 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5096 * smaller per-cpu batchsize.
5098 void __init set_dma_reserve(unsigned long new_dma_reserve)
5100 dma_reserve = new_dma_reserve;
5103 void __init free_area_init(unsigned long *zones_size)
5105 free_area_init_node(0, zones_size,
5106 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5109 static int page_alloc_cpu_notify(struct notifier_block *self,
5110 unsigned long action, void *hcpu)
5112 int cpu = (unsigned long)hcpu;
5114 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5115 lru_add_drain_cpu(cpu);
5119 * Spill the event counters of the dead processor
5120 * into the current processors event counters.
5121 * This artificially elevates the count of the current
5124 vm_events_fold_cpu(cpu);
5127 * Zero the differential counters of the dead processor
5128 * so that the vm statistics are consistent.
5130 * This is only okay since the processor is dead and cannot
5131 * race with what we are doing.
5133 refresh_cpu_vm_stats(cpu);
5138 void __init page_alloc_init(void)
5140 hotcpu_notifier(page_alloc_cpu_notify, 0);
5144 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5145 * or min_free_kbytes changes.
5147 static void calculate_totalreserve_pages(void)
5149 struct pglist_data *pgdat;
5150 unsigned long reserve_pages = 0;
5151 enum zone_type i, j;
5153 for_each_online_pgdat(pgdat) {
5154 for (i = 0; i < MAX_NR_ZONES; i++) {
5155 struct zone *zone = pgdat->node_zones + i;
5156 unsigned long max = 0;
5158 /* Find valid and maximum lowmem_reserve in the zone */
5159 for (j = i; j < MAX_NR_ZONES; j++) {
5160 if (zone->lowmem_reserve[j] > max)
5161 max = zone->lowmem_reserve[j];
5164 /* we treat the high watermark as reserved pages. */
5165 max += high_wmark_pages(zone);
5167 if (max > zone->present_pages)
5168 max = zone->present_pages;
5169 reserve_pages += max;
5171 * Lowmem reserves are not available to
5172 * GFP_HIGHUSER page cache allocations and
5173 * kswapd tries to balance zones to their high
5174 * watermark. As a result, neither should be
5175 * regarded as dirtyable memory, to prevent a
5176 * situation where reclaim has to clean pages
5177 * in order to balance the zones.
5179 zone->dirty_balance_reserve = max;
5182 dirty_balance_reserve = reserve_pages;
5183 totalreserve_pages = reserve_pages;
5187 * setup_per_zone_lowmem_reserve - called whenever
5188 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5189 * has a correct pages reserved value, so an adequate number of
5190 * pages are left in the zone after a successful __alloc_pages().
5192 static void setup_per_zone_lowmem_reserve(void)
5194 struct pglist_data *pgdat;
5195 enum zone_type j, idx;
5197 for_each_online_pgdat(pgdat) {
5198 for (j = 0; j < MAX_NR_ZONES; j++) {
5199 struct zone *zone = pgdat->node_zones + j;
5200 unsigned long present_pages = zone->present_pages;
5202 zone->lowmem_reserve[j] = 0;
5206 struct zone *lower_zone;
5210 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5211 sysctl_lowmem_reserve_ratio[idx] = 1;
5213 lower_zone = pgdat->node_zones + idx;
5214 lower_zone->lowmem_reserve[j] = present_pages /
5215 sysctl_lowmem_reserve_ratio[idx];
5216 present_pages += lower_zone->present_pages;
5221 /* update totalreserve_pages */
5222 calculate_totalreserve_pages();
5225 static void __setup_per_zone_wmarks(void)
5227 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5228 unsigned long lowmem_pages = 0;
5230 unsigned long flags;
5232 /* Calculate total number of !ZONE_HIGHMEM pages */
5233 for_each_zone(zone) {
5234 if (!is_highmem(zone))
5235 lowmem_pages += zone->present_pages;
5238 for_each_zone(zone) {
5241 spin_lock_irqsave(&zone->lock, flags);
5242 tmp = (u64)pages_min * zone->present_pages;
5243 do_div(tmp, lowmem_pages);
5244 if (is_highmem(zone)) {
5246 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5247 * need highmem pages, so cap pages_min to a small
5250 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5251 * deltas controls asynch page reclaim, and so should
5252 * not be capped for highmem.
5256 min_pages = zone->present_pages / 1024;
5257 if (min_pages < SWAP_CLUSTER_MAX)
5258 min_pages = SWAP_CLUSTER_MAX;
5259 if (min_pages > 128)
5261 zone->watermark[WMARK_MIN] = min_pages;
5264 * If it's a lowmem zone, reserve a number of pages
5265 * proportionate to the zone's size.
5267 zone->watermark[WMARK_MIN] = tmp;
5270 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5271 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5273 setup_zone_migrate_reserve(zone);
5274 spin_unlock_irqrestore(&zone->lock, flags);
5277 /* update totalreserve_pages */
5278 calculate_totalreserve_pages();
5282 * setup_per_zone_wmarks - called when min_free_kbytes changes
5283 * or when memory is hot-{added|removed}
5285 * Ensures that the watermark[min,low,high] values for each zone are set
5286 * correctly with respect to min_free_kbytes.
5288 void setup_per_zone_wmarks(void)
5290 mutex_lock(&zonelists_mutex);
5291 __setup_per_zone_wmarks();
5292 mutex_unlock(&zonelists_mutex);
5296 * The inactive anon list should be small enough that the VM never has to
5297 * do too much work, but large enough that each inactive page has a chance
5298 * to be referenced again before it is swapped out.
5300 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5301 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5302 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5303 * the anonymous pages are kept on the inactive list.
5306 * memory ratio inactive anon
5307 * -------------------------------------
5316 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5318 unsigned int gb, ratio;
5320 /* Zone size in gigabytes */
5321 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5323 ratio = int_sqrt(10 * gb);
5327 zone->inactive_ratio = ratio;
5330 static void __meminit setup_per_zone_inactive_ratio(void)
5335 calculate_zone_inactive_ratio(zone);
5339 * Initialise min_free_kbytes.
5341 * For small machines we want it small (128k min). For large machines
5342 * we want it large (64MB max). But it is not linear, because network
5343 * bandwidth does not increase linearly with machine size. We use
5345 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5346 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5362 int __meminit init_per_zone_wmark_min(void)
5364 unsigned long lowmem_kbytes;
5366 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5368 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5369 if (min_free_kbytes < 128)
5370 min_free_kbytes = 128;
5371 if (min_free_kbytes > 65536)
5372 min_free_kbytes = 65536;
5373 setup_per_zone_wmarks();
5374 refresh_zone_stat_thresholds();
5375 setup_per_zone_lowmem_reserve();
5376 setup_per_zone_inactive_ratio();
5379 module_init(init_per_zone_wmark_min)
5382 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5383 * that we can call two helper functions whenever min_free_kbytes
5386 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5387 void __user *buffer, size_t *length, loff_t *ppos)
5389 proc_dointvec(table, write, buffer, length, ppos);
5391 setup_per_zone_wmarks();
5396 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5397 void __user *buffer, size_t *length, loff_t *ppos)
5402 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5407 zone->min_unmapped_pages = (zone->present_pages *
5408 sysctl_min_unmapped_ratio) / 100;
5412 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5413 void __user *buffer, size_t *length, loff_t *ppos)
5418 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5423 zone->min_slab_pages = (zone->present_pages *
5424 sysctl_min_slab_ratio) / 100;
5430 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5431 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5432 * whenever sysctl_lowmem_reserve_ratio changes.
5434 * The reserve ratio obviously has absolutely no relation with the
5435 * minimum watermarks. The lowmem reserve ratio can only make sense
5436 * if in function of the boot time zone sizes.
5438 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5439 void __user *buffer, size_t *length, loff_t *ppos)
5441 proc_dointvec_minmax(table, write, buffer, length, ppos);
5442 setup_per_zone_lowmem_reserve();
5447 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5448 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5449 * can have before it gets flushed back to buddy allocator.
5452 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5453 void __user *buffer, size_t *length, loff_t *ppos)
5459 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5460 if (!write || (ret < 0))
5462 for_each_populated_zone(zone) {
5463 for_each_possible_cpu(cpu) {
5465 high = zone->present_pages / percpu_pagelist_fraction;
5466 setup_pagelist_highmark(
5467 per_cpu_ptr(zone->pageset, cpu), high);
5473 int hashdist = HASHDIST_DEFAULT;
5476 static int __init set_hashdist(char *str)
5480 hashdist = simple_strtoul(str, &str, 0);
5483 __setup("hashdist=", set_hashdist);
5487 * allocate a large system hash table from bootmem
5488 * - it is assumed that the hash table must contain an exact power-of-2
5489 * quantity of entries
5490 * - limit is the number of hash buckets, not the total allocation size
5492 void *__init alloc_large_system_hash(const char *tablename,
5493 unsigned long bucketsize,
5494 unsigned long numentries,
5497 unsigned int *_hash_shift,
5498 unsigned int *_hash_mask,
5499 unsigned long low_limit,
5500 unsigned long high_limit)
5502 unsigned long long max = high_limit;
5503 unsigned long log2qty, size;
5506 /* allow the kernel cmdline to have a say */
5508 /* round applicable memory size up to nearest megabyte */
5509 numentries = nr_kernel_pages;
5510 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5511 numentries >>= 20 - PAGE_SHIFT;
5512 numentries <<= 20 - PAGE_SHIFT;
5514 /* limit to 1 bucket per 2^scale bytes of low memory */
5515 if (scale > PAGE_SHIFT)
5516 numentries >>= (scale - PAGE_SHIFT);
5518 numentries <<= (PAGE_SHIFT - scale);
5520 /* Make sure we've got at least a 0-order allocation.. */
5521 if (unlikely(flags & HASH_SMALL)) {
5522 /* Makes no sense without HASH_EARLY */
5523 WARN_ON(!(flags & HASH_EARLY));
5524 if (!(numentries >> *_hash_shift)) {
5525 numentries = 1UL << *_hash_shift;
5526 BUG_ON(!numentries);
5528 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5529 numentries = PAGE_SIZE / bucketsize;
5531 numentries = roundup_pow_of_two(numentries);
5533 /* limit allocation size to 1/16 total memory by default */
5535 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5536 do_div(max, bucketsize);
5538 max = min(max, 0x80000000ULL);
5540 if (numentries < low_limit)
5541 numentries = low_limit;
5542 if (numentries > max)
5545 log2qty = ilog2(numentries);
5548 size = bucketsize << log2qty;
5549 if (flags & HASH_EARLY)
5550 table = alloc_bootmem_nopanic(size);
5552 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5555 * If bucketsize is not a power-of-two, we may free
5556 * some pages at the end of hash table which
5557 * alloc_pages_exact() automatically does
5559 if (get_order(size) < MAX_ORDER) {
5560 table = alloc_pages_exact(size, GFP_ATOMIC);
5561 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5564 } while (!table && size > PAGE_SIZE && --log2qty);
5567 panic("Failed to allocate %s hash table\n", tablename);
5569 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5572 ilog2(size) - PAGE_SHIFT,
5576 *_hash_shift = log2qty;
5578 *_hash_mask = (1 << log2qty) - 1;
5583 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5584 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5587 #ifdef CONFIG_SPARSEMEM
5588 return __pfn_to_section(pfn)->pageblock_flags;
5590 return zone->pageblock_flags;
5591 #endif /* CONFIG_SPARSEMEM */
5594 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5596 #ifdef CONFIG_SPARSEMEM
5597 pfn &= (PAGES_PER_SECTION-1);
5598 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5600 pfn = pfn - zone->zone_start_pfn;
5601 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5602 #endif /* CONFIG_SPARSEMEM */
5606 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5607 * @page: The page within the block of interest
5608 * @start_bitidx: The first bit of interest to retrieve
5609 * @end_bitidx: The last bit of interest
5610 * returns pageblock_bits flags
5612 unsigned long get_pageblock_flags_group(struct page *page,
5613 int start_bitidx, int end_bitidx)
5616 unsigned long *bitmap;
5617 unsigned long pfn, bitidx;
5618 unsigned long flags = 0;
5619 unsigned long value = 1;
5621 zone = page_zone(page);
5622 pfn = page_to_pfn(page);
5623 bitmap = get_pageblock_bitmap(zone, pfn);
5624 bitidx = pfn_to_bitidx(zone, pfn);
5626 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5627 if (test_bit(bitidx + start_bitidx, bitmap))
5634 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5635 * @page: The page within the block of interest
5636 * @start_bitidx: The first bit of interest
5637 * @end_bitidx: The last bit of interest
5638 * @flags: The flags to set
5640 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5641 int start_bitidx, int end_bitidx)
5644 unsigned long *bitmap;
5645 unsigned long pfn, bitidx;
5646 unsigned long value = 1;
5648 zone = page_zone(page);
5649 pfn = page_to_pfn(page);
5650 bitmap = get_pageblock_bitmap(zone, pfn);
5651 bitidx = pfn_to_bitidx(zone, pfn);
5652 VM_BUG_ON(pfn < zone->zone_start_pfn);
5653 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5655 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5657 __set_bit(bitidx + start_bitidx, bitmap);
5659 __clear_bit(bitidx + start_bitidx, bitmap);
5663 * This function checks whether pageblock includes unmovable pages or not.
5664 * If @count is not zero, it is okay to include less @count unmovable pages
5666 * PageLRU check wihtout isolation or lru_lock could race so that
5667 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5668 * expect this function should be exact.
5670 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5671 bool skip_hwpoisoned_pages)
5673 unsigned long pfn, iter, found;
5677 * For avoiding noise data, lru_add_drain_all() should be called
5678 * If ZONE_MOVABLE, the zone never contains unmovable pages
5680 if (zone_idx(zone) == ZONE_MOVABLE)
5682 mt = get_pageblock_migratetype(page);
5683 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5686 pfn = page_to_pfn(page);
5687 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5688 unsigned long check = pfn + iter;
5690 if (!pfn_valid_within(check))
5693 page = pfn_to_page(check);
5695 * We can't use page_count without pin a page
5696 * because another CPU can free compound page.
5697 * This check already skips compound tails of THP
5698 * because their page->_count is zero at all time.
5700 if (!atomic_read(&page->_count)) {
5701 if (PageBuddy(page))
5702 iter += (1 << page_order(page)) - 1;
5707 * The HWPoisoned page may be not in buddy system, and
5708 * page_count() is not 0.
5710 if (skip_hwpoisoned_pages && PageHWPoison(page))
5716 * If there are RECLAIMABLE pages, we need to check it.
5717 * But now, memory offline itself doesn't call shrink_slab()
5718 * and it still to be fixed.
5721 * If the page is not RAM, page_count()should be 0.
5722 * we don't need more check. This is an _used_ not-movable page.
5724 * The problematic thing here is PG_reserved pages. PG_reserved
5725 * is set to both of a memory hole page and a _used_ kernel
5734 bool is_pageblock_removable_nolock(struct page *page)
5740 * We have to be careful here because we are iterating over memory
5741 * sections which are not zone aware so we might end up outside of
5742 * the zone but still within the section.
5743 * We have to take care about the node as well. If the node is offline
5744 * its NODE_DATA will be NULL - see page_zone.
5746 if (!node_online(page_to_nid(page)))
5749 zone = page_zone(page);
5750 pfn = page_to_pfn(page);
5751 if (zone->zone_start_pfn > pfn ||
5752 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5755 return !has_unmovable_pages(zone, page, 0, true);
5760 static unsigned long pfn_max_align_down(unsigned long pfn)
5762 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5763 pageblock_nr_pages) - 1);
5766 static unsigned long pfn_max_align_up(unsigned long pfn)
5768 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5769 pageblock_nr_pages));
5772 /* [start, end) must belong to a single zone. */
5773 static int __alloc_contig_migrate_range(struct compact_control *cc,
5774 unsigned long start, unsigned long end)
5776 /* This function is based on compact_zone() from compaction.c. */
5777 unsigned long nr_reclaimed;
5778 unsigned long pfn = start;
5779 unsigned int tries = 0;
5784 while (pfn < end || !list_empty(&cc->migratepages)) {
5785 if (fatal_signal_pending(current)) {
5790 if (list_empty(&cc->migratepages)) {
5791 cc->nr_migratepages = 0;
5792 pfn = isolate_migratepages_range(cc->zone, cc,
5799 } else if (++tries == 5) {
5800 ret = ret < 0 ? ret : -EBUSY;
5804 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5806 cc->nr_migratepages -= nr_reclaimed;
5808 ret = migrate_pages(&cc->migratepages,
5809 alloc_migrate_target,
5810 0, false, MIGRATE_SYNC,
5814 putback_movable_pages(&cc->migratepages);
5815 return ret > 0 ? 0 : ret;
5819 * alloc_contig_range() -- tries to allocate given range of pages
5820 * @start: start PFN to allocate
5821 * @end: one-past-the-last PFN to allocate
5822 * @migratetype: migratetype of the underlaying pageblocks (either
5823 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5824 * in range must have the same migratetype and it must
5825 * be either of the two.
5827 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5828 * aligned, however it's the caller's responsibility to guarantee that
5829 * we are the only thread that changes migrate type of pageblocks the
5832 * The PFN range must belong to a single zone.
5834 * Returns zero on success or negative error code. On success all
5835 * pages which PFN is in [start, end) are allocated for the caller and
5836 * need to be freed with free_contig_range().
5838 int alloc_contig_range(unsigned long start, unsigned long end,
5839 unsigned migratetype)
5841 unsigned long outer_start, outer_end;
5844 struct compact_control cc = {
5845 .nr_migratepages = 0,
5847 .zone = page_zone(pfn_to_page(start)),
5849 .ignore_skip_hint = true,
5851 INIT_LIST_HEAD(&cc.migratepages);
5854 * What we do here is we mark all pageblocks in range as
5855 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5856 * have different sizes, and due to the way page allocator
5857 * work, we align the range to biggest of the two pages so
5858 * that page allocator won't try to merge buddies from
5859 * different pageblocks and change MIGRATE_ISOLATE to some
5860 * other migration type.
5862 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5863 * migrate the pages from an unaligned range (ie. pages that
5864 * we are interested in). This will put all the pages in
5865 * range back to page allocator as MIGRATE_ISOLATE.
5867 * When this is done, we take the pages in range from page
5868 * allocator removing them from the buddy system. This way
5869 * page allocator will never consider using them.
5871 * This lets us mark the pageblocks back as
5872 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5873 * aligned range but not in the unaligned, original range are
5874 * put back to page allocator so that buddy can use them.
5877 ret = start_isolate_page_range(pfn_max_align_down(start),
5878 pfn_max_align_up(end), migratetype,
5883 ret = __alloc_contig_migrate_range(&cc, start, end);
5888 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5889 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5890 * more, all pages in [start, end) are free in page allocator.
5891 * What we are going to do is to allocate all pages from
5892 * [start, end) (that is remove them from page allocator).
5894 * The only problem is that pages at the beginning and at the
5895 * end of interesting range may be not aligned with pages that
5896 * page allocator holds, ie. they can be part of higher order
5897 * pages. Because of this, we reserve the bigger range and
5898 * once this is done free the pages we are not interested in.
5900 * We don't have to hold zone->lock here because the pages are
5901 * isolated thus they won't get removed from buddy.
5904 lru_add_drain_all();
5908 outer_start = start;
5909 while (!PageBuddy(pfn_to_page(outer_start))) {
5910 if (++order >= MAX_ORDER) {
5914 outer_start &= ~0UL << order;
5917 /* Make sure the range is really isolated. */
5918 if (test_pages_isolated(outer_start, end, false)) {
5919 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5926 /* Grab isolated pages from freelists. */
5927 outer_end = isolate_freepages_range(&cc, outer_start, end);
5933 /* Free head and tail (if any) */
5934 if (start != outer_start)
5935 free_contig_range(outer_start, start - outer_start);
5936 if (end != outer_end)
5937 free_contig_range(end, outer_end - end);
5940 undo_isolate_page_range(pfn_max_align_down(start),
5941 pfn_max_align_up(end), migratetype);
5945 void free_contig_range(unsigned long pfn, unsigned nr_pages)
5947 for (; nr_pages--; ++pfn)
5948 __free_page(pfn_to_page(pfn));
5952 #ifdef CONFIG_MEMORY_HOTPLUG
5953 static int __meminit __zone_pcp_update(void *data)
5955 struct zone *zone = data;
5957 unsigned long batch = zone_batchsize(zone), flags;
5959 for_each_possible_cpu(cpu) {
5960 struct per_cpu_pageset *pset;
5961 struct per_cpu_pages *pcp;
5963 pset = per_cpu_ptr(zone->pageset, cpu);
5966 local_irq_save(flags);
5968 free_pcppages_bulk(zone, pcp->count, pcp);
5969 drain_zonestat(zone, pset);
5970 setup_pageset(pset, batch);
5971 local_irq_restore(flags);
5976 void __meminit zone_pcp_update(struct zone *zone)
5978 stop_machine(__zone_pcp_update, zone, NULL);
5982 void zone_pcp_reset(struct zone *zone)
5984 unsigned long flags;
5986 struct per_cpu_pageset *pset;
5988 /* avoid races with drain_pages() */
5989 local_irq_save(flags);
5990 if (zone->pageset != &boot_pageset) {
5991 for_each_online_cpu(cpu) {
5992 pset = per_cpu_ptr(zone->pageset, cpu);
5993 drain_zonestat(zone, pset);
5995 free_percpu(zone->pageset);
5996 zone->pageset = &boot_pageset;
5998 local_irq_restore(flags);
6001 #ifdef CONFIG_MEMORY_HOTREMOVE
6003 * All pages in the range must be isolated before calling this.
6006 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6012 unsigned long flags;
6013 /* find the first valid pfn */
6014 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6019 zone = page_zone(pfn_to_page(pfn));
6020 spin_lock_irqsave(&zone->lock, flags);
6022 while (pfn < end_pfn) {
6023 if (!pfn_valid(pfn)) {
6027 page = pfn_to_page(pfn);
6029 * The HWPoisoned page may be not in buddy system, and
6030 * page_count() is not 0.
6032 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6034 SetPageReserved(page);
6038 BUG_ON(page_count(page));
6039 BUG_ON(!PageBuddy(page));
6040 order = page_order(page);
6041 #ifdef CONFIG_DEBUG_VM
6042 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6043 pfn, 1 << order, end_pfn);
6045 list_del(&page->lru);
6046 rmv_page_order(page);
6047 zone->free_area[order].nr_free--;
6048 for (i = 0; i < (1 << order); i++)
6049 SetPageReserved((page+i));
6050 pfn += (1 << order);
6052 spin_unlock_irqrestore(&zone->lock, flags);
6056 #ifdef CONFIG_MEMORY_FAILURE
6057 bool is_free_buddy_page(struct page *page)
6059 struct zone *zone = page_zone(page);
6060 unsigned long pfn = page_to_pfn(page);
6061 unsigned long flags;
6064 spin_lock_irqsave(&zone->lock, flags);
6065 for (order = 0; order < MAX_ORDER; order++) {
6066 struct page *page_head = page - (pfn & ((1 << order) - 1));
6068 if (PageBuddy(page_head) && page_order(page_head) >= order)
6071 spin_unlock_irqrestore(&zone->lock, flags);
6073 return order < MAX_ORDER;
6077 static const struct trace_print_flags pageflag_names[] = {
6078 {1UL << PG_locked, "locked" },
6079 {1UL << PG_error, "error" },
6080 {1UL << PG_referenced, "referenced" },
6081 {1UL << PG_uptodate, "uptodate" },
6082 {1UL << PG_dirty, "dirty" },
6083 {1UL << PG_lru, "lru" },
6084 {1UL << PG_active, "active" },
6085 {1UL << PG_slab, "slab" },
6086 {1UL << PG_owner_priv_1, "owner_priv_1" },
6087 {1UL << PG_arch_1, "arch_1" },
6088 {1UL << PG_reserved, "reserved" },
6089 {1UL << PG_private, "private" },
6090 {1UL << PG_private_2, "private_2" },
6091 {1UL << PG_writeback, "writeback" },
6092 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6093 {1UL << PG_head, "head" },
6094 {1UL << PG_tail, "tail" },
6096 {1UL << PG_compound, "compound" },
6098 {1UL << PG_swapcache, "swapcache" },
6099 {1UL << PG_mappedtodisk, "mappedtodisk" },
6100 {1UL << PG_reclaim, "reclaim" },
6101 {1UL << PG_swapbacked, "swapbacked" },
6102 {1UL << PG_unevictable, "unevictable" },
6104 {1UL << PG_mlocked, "mlocked" },
6106 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6107 {1UL << PG_uncached, "uncached" },
6109 #ifdef CONFIG_MEMORY_FAILURE
6110 {1UL << PG_hwpoison, "hwpoison" },
6112 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6113 {1UL << PG_compound_lock, "compound_lock" },
6117 static void dump_page_flags(unsigned long flags)
6119 const char *delim = "";
6123 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6125 printk(KERN_ALERT "page flags: %#lx(", flags);
6127 /* remove zone id */
6128 flags &= (1UL << NR_PAGEFLAGS) - 1;
6130 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6132 mask = pageflag_names[i].mask;
6133 if ((flags & mask) != mask)
6137 printk("%s%s", delim, pageflag_names[i].name);
6141 /* check for left over flags */
6143 printk("%s%#lx", delim, flags);
6148 void dump_page(struct page *page)
6151 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6152 page, atomic_read(&page->_count), page_mapcount(page),
6153 page->mapping, page->index);
6154 dump_page_flags(page->flags);
6155 mem_cgroup_print_bad_page(page);