2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
69 DEFINE_PER_CPU(int, numa_node);
70 EXPORT_PER_CPU_SYMBOL(numa_node);
73 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
75 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
76 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
77 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
78 * defined in <linux/topology.h>.
80 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
81 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
85 * Array of node states.
87 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
88 [N_POSSIBLE] = NODE_MASK_ALL,
89 [N_ONLINE] = { { [0] = 1UL } },
91 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
93 [N_HIGH_MEMORY] = { { [0] = 1UL } },
95 #ifdef CONFIG_MOVABLE_NODE
96 [N_MEMORY] = { { [0] = 1UL } },
98 [N_CPU] = { { [0] = 1UL } },
101 EXPORT_SYMBOL(node_states);
103 unsigned long totalram_pages __read_mostly;
104 unsigned long totalreserve_pages __read_mostly;
106 * When calculating the number of globally allowed dirty pages, there
107 * is a certain number of per-zone reserves that should not be
108 * considered dirtyable memory. This is the sum of those reserves
109 * over all existing zones that contribute dirtyable memory.
111 unsigned long dirty_balance_reserve __read_mostly;
113 int percpu_pagelist_fraction;
114 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
116 #ifdef CONFIG_PM_SLEEP
118 * The following functions are used by the suspend/hibernate code to temporarily
119 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
120 * while devices are suspended. To avoid races with the suspend/hibernate code,
121 * they should always be called with pm_mutex held (gfp_allowed_mask also should
122 * only be modified with pm_mutex held, unless the suspend/hibernate code is
123 * guaranteed not to run in parallel with that modification).
126 static gfp_t saved_gfp_mask;
128 void pm_restore_gfp_mask(void)
130 WARN_ON(!mutex_is_locked(&pm_mutex));
131 if (saved_gfp_mask) {
132 gfp_allowed_mask = saved_gfp_mask;
137 void pm_restrict_gfp_mask(void)
139 WARN_ON(!mutex_is_locked(&pm_mutex));
140 WARN_ON(saved_gfp_mask);
141 saved_gfp_mask = gfp_allowed_mask;
142 gfp_allowed_mask &= ~GFP_IOFS;
145 bool pm_suspended_storage(void)
147 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
151 #endif /* CONFIG_PM_SLEEP */
153 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
154 int pageblock_order __read_mostly;
157 static void __free_pages_ok(struct page *page, unsigned int order);
160 * results with 256, 32 in the lowmem_reserve sysctl:
161 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
162 * 1G machine -> (16M dma, 784M normal, 224M high)
163 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
164 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
165 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
167 * TBD: should special case ZONE_DMA32 machines here - in those we normally
168 * don't need any ZONE_NORMAL reservation
170 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
171 #ifdef CONFIG_ZONE_DMA
174 #ifdef CONFIG_ZONE_DMA32
177 #ifdef CONFIG_HIGHMEM
183 EXPORT_SYMBOL(totalram_pages);
185 static char * const zone_names[MAX_NR_ZONES] = {
186 #ifdef CONFIG_ZONE_DMA
189 #ifdef CONFIG_ZONE_DMA32
193 #ifdef CONFIG_HIGHMEM
200 * Try to keep at least this much lowmem free. Do not allow normal
201 * allocations below this point, only high priority ones. Automatically
202 * tuned according to the amount of memory in the system.
204 int min_free_kbytes = 1024;
205 int min_free_order_shift = 1;
208 * Extra memory for the system to try freeing. Used to temporarily
209 * free memory, to make space for new workloads. Anyone can allocate
210 * down to the min watermarks controlled by min_free_kbytes above.
212 int extra_free_kbytes = 0;
214 static unsigned long __meminitdata nr_kernel_pages;
215 static unsigned long __meminitdata nr_all_pages;
216 static unsigned long __meminitdata dma_reserve;
218 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
219 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
220 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
221 static unsigned long __initdata required_kernelcore;
222 static unsigned long __initdata required_movablecore;
223 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
225 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
227 EXPORT_SYMBOL(movable_zone);
228 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
231 int nr_node_ids __read_mostly = MAX_NUMNODES;
232 int nr_online_nodes __read_mostly = 1;
233 EXPORT_SYMBOL(nr_node_ids);
234 EXPORT_SYMBOL(nr_online_nodes);
237 int page_group_by_mobility_disabled __read_mostly;
239 void set_pageblock_migratetype(struct page *page, int migratetype)
242 if (unlikely(page_group_by_mobility_disabled))
243 migratetype = MIGRATE_UNMOVABLE;
245 set_pageblock_flags_group(page, (unsigned long)migratetype,
246 PB_migrate, PB_migrate_end);
249 bool oom_killer_disabled __read_mostly;
251 #ifdef CONFIG_DEBUG_VM
252 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
256 unsigned long pfn = page_to_pfn(page);
257 unsigned long sp, start_pfn;
260 seq = zone_span_seqbegin(zone);
261 start_pfn = zone->zone_start_pfn;
262 sp = zone->spanned_pages;
263 if (!zone_spans_pfn(zone, pfn))
265 } while (zone_span_seqretry(zone, seq));
268 pr_err("page %lu outside zone [ %lu - %lu ]\n",
269 pfn, start_pfn, start_pfn + sp);
274 static int page_is_consistent(struct zone *zone, struct page *page)
276 if (!pfn_valid_within(page_to_pfn(page)))
278 if (zone != page_zone(page))
284 * Temporary debugging check for pages not lying within a given zone.
286 static int bad_range(struct zone *zone, struct page *page)
288 if (page_outside_zone_boundaries(zone, page))
290 if (!page_is_consistent(zone, page))
296 static inline int bad_range(struct zone *zone, struct page *page)
302 static void bad_page(struct page *page)
304 static unsigned long resume;
305 static unsigned long nr_shown;
306 static unsigned long nr_unshown;
308 /* Don't complain about poisoned pages */
309 if (PageHWPoison(page)) {
310 page_mapcount_reset(page); /* remove PageBuddy */
315 * Allow a burst of 60 reports, then keep quiet for that minute;
316 * or allow a steady drip of one report per second.
318 if (nr_shown == 60) {
319 if (time_before(jiffies, resume)) {
325 "BUG: Bad page state: %lu messages suppressed\n",
332 resume = jiffies + 60 * HZ;
334 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
335 current->comm, page_to_pfn(page));
341 /* Leave bad fields for debug, except PageBuddy could make trouble */
342 page_mapcount_reset(page); /* remove PageBuddy */
343 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
347 * Higher-order pages are called "compound pages". They are structured thusly:
349 * The first PAGE_SIZE page is called the "head page".
351 * The remaining PAGE_SIZE pages are called "tail pages".
353 * All pages have PG_compound set. All tail pages have their ->first_page
354 * pointing at the head page.
356 * The first tail page's ->lru.next holds the address of the compound page's
357 * put_page() function. Its ->lru.prev holds the order of allocation.
358 * This usage means that zero-order pages may not be compound.
361 static void free_compound_page(struct page *page)
363 __free_pages_ok(page, compound_order(page));
366 void prep_compound_page(struct page *page, unsigned long order)
369 int nr_pages = 1 << order;
371 set_compound_page_dtor(page, free_compound_page);
372 set_compound_order(page, order);
374 for (i = 1; i < nr_pages; i++) {
375 struct page *p = page + i;
376 set_page_count(p, 0);
377 p->first_page = page;
378 /* Make sure p->first_page is always valid for PageTail() */
384 /* update __split_huge_page_refcount if you change this function */
385 static int destroy_compound_page(struct page *page, unsigned long order)
388 int nr_pages = 1 << order;
391 if (unlikely(compound_order(page) != order)) {
396 __ClearPageHead(page);
398 for (i = 1; i < nr_pages; i++) {
399 struct page *p = page + i;
401 if (unlikely(!PageTail(p) || (p->first_page != page))) {
411 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
416 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
417 * and __GFP_HIGHMEM from hard or soft interrupt context.
419 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
420 for (i = 0; i < (1 << order); i++)
421 clear_highpage(page + i);
424 #ifdef CONFIG_DEBUG_PAGEALLOC
425 unsigned int _debug_guardpage_minorder;
427 static int __init debug_guardpage_minorder_setup(char *buf)
431 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
432 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
435 _debug_guardpage_minorder = res;
436 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
439 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
441 static inline void set_page_guard_flag(struct page *page)
443 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
446 static inline void clear_page_guard_flag(struct page *page)
448 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
451 static inline void set_page_guard_flag(struct page *page) { }
452 static inline void clear_page_guard_flag(struct page *page) { }
455 static inline void set_page_order(struct page *page, int order)
457 set_page_private(page, order);
458 __SetPageBuddy(page);
461 static inline void rmv_page_order(struct page *page)
463 __ClearPageBuddy(page);
464 set_page_private(page, 0);
468 * Locate the struct page for both the matching buddy in our
469 * pair (buddy1) and the combined O(n+1) page they form (page).
471 * 1) Any buddy B1 will have an order O twin B2 which satisfies
472 * the following equation:
474 * For example, if the starting buddy (buddy2) is #8 its order
476 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
478 * 2) Any buddy B will have an order O+1 parent P which
479 * satisfies the following equation:
482 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
484 static inline unsigned long
485 __find_buddy_index(unsigned long page_idx, unsigned int order)
487 return page_idx ^ (1 << order);
491 * This function checks whether a page is free && is the buddy
492 * we can do coalesce a page and its buddy if
493 * (a) the buddy is not in a hole &&
494 * (b) the buddy is in the buddy system &&
495 * (c) a page and its buddy have the same order &&
496 * (d) a page and its buddy are in the same zone.
498 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
499 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
501 * For recording page's order, we use page_private(page).
503 static inline int page_is_buddy(struct page *page, struct page *buddy,
506 if (!pfn_valid_within(page_to_pfn(buddy)))
509 if (page_zone_id(page) != page_zone_id(buddy))
512 if (page_is_guard(buddy) && page_order(buddy) == order) {
513 VM_BUG_ON(page_count(buddy) != 0);
517 if (PageBuddy(buddy) && page_order(buddy) == order) {
518 VM_BUG_ON(page_count(buddy) != 0);
525 * Freeing function for a buddy system allocator.
527 * The concept of a buddy system is to maintain direct-mapped table
528 * (containing bit values) for memory blocks of various "orders".
529 * The bottom level table contains the map for the smallest allocatable
530 * units of memory (here, pages), and each level above it describes
531 * pairs of units from the levels below, hence, "buddies".
532 * At a high level, all that happens here is marking the table entry
533 * at the bottom level available, and propagating the changes upward
534 * as necessary, plus some accounting needed to play nicely with other
535 * parts of the VM system.
536 * At each level, we keep a list of pages, which are heads of continuous
537 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
538 * order is recorded in page_private(page) field.
539 * So when we are allocating or freeing one, we can derive the state of the
540 * other. That is, if we allocate a small block, and both were
541 * free, the remainder of the region must be split into blocks.
542 * If a block is freed, and its buddy is also free, then this
543 * triggers coalescing into a block of larger size.
548 static inline void __free_one_page(struct page *page,
549 struct zone *zone, unsigned int order,
552 unsigned long page_idx;
553 unsigned long combined_idx;
554 unsigned long uninitialized_var(buddy_idx);
557 VM_BUG_ON(!zone_is_initialized(zone));
559 if (unlikely(PageCompound(page)))
560 if (unlikely(destroy_compound_page(page, order)))
563 VM_BUG_ON(migratetype == -1);
565 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
567 VM_BUG_ON(page_idx & ((1 << order) - 1));
568 VM_BUG_ON(bad_range(zone, page));
570 while (order < MAX_ORDER-1) {
571 buddy_idx = __find_buddy_index(page_idx, order);
572 buddy = page + (buddy_idx - page_idx);
573 if (!page_is_buddy(page, buddy, order))
576 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
577 * merge with it and move up one order.
579 if (page_is_guard(buddy)) {
580 clear_page_guard_flag(buddy);
581 set_page_private(page, 0);
582 __mod_zone_freepage_state(zone, 1 << order,
585 list_del(&buddy->lru);
586 zone->free_area[order].nr_free--;
587 rmv_page_order(buddy);
589 combined_idx = buddy_idx & page_idx;
590 page = page + (combined_idx - page_idx);
591 page_idx = combined_idx;
594 set_page_order(page, order);
597 * If this is not the largest possible page, check if the buddy
598 * of the next-highest order is free. If it is, it's possible
599 * that pages are being freed that will coalesce soon. In case,
600 * that is happening, add the free page to the tail of the list
601 * so it's less likely to be used soon and more likely to be merged
602 * as a higher order page
604 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
605 struct page *higher_page, *higher_buddy;
606 combined_idx = buddy_idx & page_idx;
607 higher_page = page + (combined_idx - page_idx);
608 buddy_idx = __find_buddy_index(combined_idx, order + 1);
609 higher_buddy = higher_page + (buddy_idx - combined_idx);
610 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
611 list_add_tail(&page->lru,
612 &zone->free_area[order].free_list[migratetype]);
617 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
619 zone->free_area[order].nr_free++;
622 static inline int free_pages_check(struct page *page)
624 if (unlikely(page_mapcount(page) |
625 (page->mapping != NULL) |
626 (atomic_read(&page->_count) != 0) |
627 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
628 (mem_cgroup_bad_page_check(page)))) {
632 page_nid_reset_last(page);
633 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
634 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
639 * Frees a number of pages from the PCP lists
640 * Assumes all pages on list are in same zone, and of same order.
641 * count is the number of pages to free.
643 * If the zone was previously in an "all pages pinned" state then look to
644 * see if this freeing clears that state.
646 * And clear the zone's pages_scanned counter, to hold off the "all pages are
647 * pinned" detection logic.
649 static void free_pcppages_bulk(struct zone *zone, int count,
650 struct per_cpu_pages *pcp)
656 spin_lock(&zone->lock);
657 zone->all_unreclaimable = 0;
658 zone->pages_scanned = 0;
662 struct list_head *list;
665 * Remove pages from lists in a round-robin fashion. A
666 * batch_free count is maintained that is incremented when an
667 * empty list is encountered. This is so more pages are freed
668 * off fuller lists instead of spinning excessively around empty
673 if (++migratetype == MIGRATE_PCPTYPES)
675 list = &pcp->lists[migratetype];
676 } while (list_empty(list));
678 /* This is the only non-empty list. Free them all. */
679 if (batch_free == MIGRATE_PCPTYPES)
680 batch_free = to_free;
683 int mt; /* migratetype of the to-be-freed page */
685 page = list_entry(list->prev, struct page, lru);
686 /* must delete as __free_one_page list manipulates */
687 list_del(&page->lru);
688 mt = get_freepage_migratetype(page);
689 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
690 __free_one_page(page, zone, 0, mt);
691 trace_mm_page_pcpu_drain(page, 0, mt);
692 if (likely(!is_migrate_isolate_page(page))) {
693 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
694 if (is_migrate_cma(mt))
695 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
697 } while (--to_free && --batch_free && !list_empty(list));
699 spin_unlock(&zone->lock);
702 static void free_one_page(struct zone *zone, struct page *page, int order,
705 spin_lock(&zone->lock);
706 zone->all_unreclaimable = 0;
707 zone->pages_scanned = 0;
709 __free_one_page(page, zone, order, migratetype);
710 if (unlikely(!is_migrate_isolate(migratetype)))
711 __mod_zone_freepage_state(zone, 1 << order, migratetype);
712 spin_unlock(&zone->lock);
715 static bool free_pages_prepare(struct page *page, unsigned int order)
720 trace_mm_page_free(page, order);
721 kmemcheck_free_shadow(page, order);
724 page->mapping = NULL;
725 for (i = 0; i < (1 << order); i++)
726 bad += free_pages_check(page + i);
730 if (!PageHighMem(page)) {
731 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
732 debug_check_no_obj_freed(page_address(page),
735 arch_free_page(page, order);
736 kernel_map_pages(page, 1 << order, 0);
741 static void __free_pages_ok(struct page *page, unsigned int order)
746 if (!free_pages_prepare(page, order))
749 local_irq_save(flags);
750 __count_vm_events(PGFREE, 1 << order);
751 migratetype = get_pageblock_migratetype(page);
752 set_freepage_migratetype(page, migratetype);
753 free_one_page(page_zone(page), page, order, migratetype);
754 local_irq_restore(flags);
758 * Read access to zone->managed_pages is safe because it's unsigned long,
759 * but we still need to serialize writers. Currently all callers of
760 * __free_pages_bootmem() except put_page_bootmem() should only be used
761 * at boot time. So for shorter boot time, we shift the burden to
762 * put_page_bootmem() to serialize writers.
764 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
766 unsigned int nr_pages = 1 << order;
770 for (loop = 0; loop < nr_pages; loop++) {
771 struct page *p = &page[loop];
773 if (loop + 1 < nr_pages)
775 __ClearPageReserved(p);
776 set_page_count(p, 0);
779 page_zone(page)->managed_pages += 1 << order;
780 set_page_refcounted(page);
781 __free_pages(page, order);
785 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
786 void __init init_cma_reserved_pageblock(struct page *page)
788 unsigned i = pageblock_nr_pages;
789 struct page *p = page;
792 __ClearPageReserved(p);
793 set_page_count(p, 0);
796 set_page_refcounted(page);
797 set_pageblock_migratetype(page, MIGRATE_CMA);
798 __free_pages(page, pageblock_order);
799 totalram_pages += pageblock_nr_pages;
800 #ifdef CONFIG_HIGHMEM
801 if (PageHighMem(page))
802 totalhigh_pages += pageblock_nr_pages;
808 * The order of subdivision here is critical for the IO subsystem.
809 * Please do not alter this order without good reasons and regression
810 * testing. Specifically, as large blocks of memory are subdivided,
811 * the order in which smaller blocks are delivered depends on the order
812 * they're subdivided in this function. This is the primary factor
813 * influencing the order in which pages are delivered to the IO
814 * subsystem according to empirical testing, and this is also justified
815 * by considering the behavior of a buddy system containing a single
816 * large block of memory acted on by a series of small allocations.
817 * This behavior is a critical factor in sglist merging's success.
821 static inline void expand(struct zone *zone, struct page *page,
822 int low, int high, struct free_area *area,
825 unsigned long size = 1 << high;
831 VM_BUG_ON(bad_range(zone, &page[size]));
833 #ifdef CONFIG_DEBUG_PAGEALLOC
834 if (high < debug_guardpage_minorder()) {
836 * Mark as guard pages (or page), that will allow to
837 * merge back to allocator when buddy will be freed.
838 * Corresponding page table entries will not be touched,
839 * pages will stay not present in virtual address space
841 INIT_LIST_HEAD(&page[size].lru);
842 set_page_guard_flag(&page[size]);
843 set_page_private(&page[size], high);
844 /* Guard pages are not available for any usage */
845 __mod_zone_freepage_state(zone, -(1 << high),
850 list_add(&page[size].lru, &area->free_list[migratetype]);
852 set_page_order(&page[size], high);
857 * This page is about to be returned from the page allocator
859 static inline int check_new_page(struct page *page)
861 if (unlikely(page_mapcount(page) |
862 (page->mapping != NULL) |
863 (atomic_read(&page->_count) != 0) |
864 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
865 (mem_cgroup_bad_page_check(page)))) {
872 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
876 for (i = 0; i < (1 << order); i++) {
877 struct page *p = page + i;
878 if (unlikely(check_new_page(p)))
882 set_page_private(page, 0);
883 set_page_refcounted(page);
885 arch_alloc_page(page, order);
886 kernel_map_pages(page, 1 << order, 1);
888 if (gfp_flags & __GFP_ZERO)
889 prep_zero_page(page, order, gfp_flags);
891 if (order && (gfp_flags & __GFP_COMP))
892 prep_compound_page(page, order);
898 * Go through the free lists for the given migratetype and remove
899 * the smallest available page from the freelists
902 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
905 unsigned int current_order;
906 struct free_area * area;
909 /* Find a page of the appropriate size in the preferred list */
910 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
911 area = &(zone->free_area[current_order]);
912 if (list_empty(&area->free_list[migratetype]))
915 page = list_entry(area->free_list[migratetype].next,
917 list_del(&page->lru);
918 rmv_page_order(page);
920 expand(zone, page, order, current_order, area, migratetype);
929 * This array describes the order lists are fallen back to when
930 * the free lists for the desirable migrate type are depleted
932 static int fallbacks[MIGRATE_TYPES][4] = {
933 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
934 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
936 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
937 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
939 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
941 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
942 #ifdef CONFIG_MEMORY_ISOLATION
943 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
948 * Move the free pages in a range to the free lists of the requested type.
949 * Note that start_page and end_pages are not aligned on a pageblock
950 * boundary. If alignment is required, use move_freepages_block()
952 int move_freepages(struct zone *zone,
953 struct page *start_page, struct page *end_page,
960 #ifndef CONFIG_HOLES_IN_ZONE
962 * page_zone is not safe to call in this context when
963 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
964 * anyway as we check zone boundaries in move_freepages_block().
965 * Remove at a later date when no bug reports exist related to
966 * grouping pages by mobility
968 BUG_ON(page_zone(start_page) != page_zone(end_page));
971 for (page = start_page; page <= end_page;) {
972 /* Make sure we are not inadvertently changing nodes */
973 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
975 if (!pfn_valid_within(page_to_pfn(page))) {
980 if (!PageBuddy(page)) {
985 order = page_order(page);
986 list_move(&page->lru,
987 &zone->free_area[order].free_list[migratetype]);
988 set_freepage_migratetype(page, migratetype);
990 pages_moved += 1 << order;
996 int move_freepages_block(struct zone *zone, struct page *page,
999 unsigned long start_pfn, end_pfn;
1000 struct page *start_page, *end_page;
1002 start_pfn = page_to_pfn(page);
1003 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1004 start_page = pfn_to_page(start_pfn);
1005 end_page = start_page + pageblock_nr_pages - 1;
1006 end_pfn = start_pfn + pageblock_nr_pages - 1;
1008 /* Do not cross zone boundaries */
1009 if (!zone_spans_pfn(zone, start_pfn))
1011 if (!zone_spans_pfn(zone, end_pfn))
1014 return move_freepages(zone, start_page, end_page, migratetype);
1017 static void change_pageblock_range(struct page *pageblock_page,
1018 int start_order, int migratetype)
1020 int nr_pageblocks = 1 << (start_order - pageblock_order);
1022 while (nr_pageblocks--) {
1023 set_pageblock_migratetype(pageblock_page, migratetype);
1024 pageblock_page += pageblock_nr_pages;
1028 /* Remove an element from the buddy allocator from the fallback list */
1029 static inline struct page *
1030 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1032 struct free_area * area;
1037 /* Find the largest possible block of pages in the other list */
1038 for (current_order = MAX_ORDER-1; current_order >= order;
1041 migratetype = fallbacks[start_migratetype][i];
1043 /* MIGRATE_RESERVE handled later if necessary */
1044 if (migratetype == MIGRATE_RESERVE)
1047 area = &(zone->free_area[current_order]);
1048 if (list_empty(&area->free_list[migratetype]))
1051 page = list_entry(area->free_list[migratetype].next,
1056 * If breaking a large block of pages, move all free
1057 * pages to the preferred allocation list. If falling
1058 * back for a reclaimable kernel allocation, be more
1059 * aggressive about taking ownership of free pages
1061 * On the other hand, never change migration
1062 * type of MIGRATE_CMA pageblocks nor move CMA
1063 * pages on different free lists. We don't
1064 * want unmovable pages to be allocated from
1065 * MIGRATE_CMA areas.
1067 if (!is_migrate_cma(migratetype) &&
1068 (unlikely(current_order >= pageblock_order / 2) ||
1069 start_migratetype == MIGRATE_RECLAIMABLE ||
1070 page_group_by_mobility_disabled)) {
1072 pages = move_freepages_block(zone, page,
1075 /* Claim the whole block if over half of it is free */
1076 if (pages >= (1 << (pageblock_order-1)) ||
1077 page_group_by_mobility_disabled)
1078 set_pageblock_migratetype(page,
1081 migratetype = start_migratetype;
1084 /* Remove the page from the freelists */
1085 list_del(&page->lru);
1086 rmv_page_order(page);
1088 /* Take ownership for orders >= pageblock_order */
1089 if (current_order >= pageblock_order &&
1090 !is_migrate_cma(migratetype))
1091 change_pageblock_range(page, current_order,
1094 expand(zone, page, order, current_order, area,
1095 is_migrate_cma(migratetype)
1096 ? migratetype : start_migratetype);
1098 trace_mm_page_alloc_extfrag(page, order, current_order,
1099 start_migratetype, migratetype);
1109 * Do the hard work of removing an element from the buddy allocator.
1110 * Call me with the zone->lock already held.
1112 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1118 page = __rmqueue_smallest(zone, order, migratetype);
1120 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1121 page = __rmqueue_fallback(zone, order, migratetype);
1124 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1125 * is used because __rmqueue_smallest is an inline function
1126 * and we want just one call site
1129 migratetype = MIGRATE_RESERVE;
1134 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1139 * Obtain a specified number of elements from the buddy allocator, all under
1140 * a single hold of the lock, for efficiency. Add them to the supplied list.
1141 * Returns the number of new pages which were placed at *list.
1143 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1144 unsigned long count, struct list_head *list,
1145 int migratetype, int cold)
1147 int mt = migratetype, i;
1149 spin_lock(&zone->lock);
1150 for (i = 0; i < count; ++i) {
1151 struct page *page = __rmqueue(zone, order, migratetype);
1152 if (unlikely(page == NULL))
1156 * Split buddy pages returned by expand() are received here
1157 * in physical page order. The page is added to the callers and
1158 * list and the list head then moves forward. From the callers
1159 * perspective, the linked list is ordered by page number in
1160 * some conditions. This is useful for IO devices that can
1161 * merge IO requests if the physical pages are ordered
1164 if (likely(cold == 0))
1165 list_add(&page->lru, list);
1167 list_add_tail(&page->lru, list);
1168 if (IS_ENABLED(CONFIG_CMA)) {
1169 mt = get_pageblock_migratetype(page);
1170 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1173 set_freepage_migratetype(page, mt);
1175 if (is_migrate_cma(mt))
1176 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1179 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1180 spin_unlock(&zone->lock);
1186 * Called from the vmstat counter updater to drain pagesets of this
1187 * currently executing processor on remote nodes after they have
1190 * Note that this function must be called with the thread pinned to
1191 * a single processor.
1193 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1195 unsigned long flags;
1198 local_irq_save(flags);
1199 if (pcp->count >= pcp->batch)
1200 to_drain = pcp->batch;
1202 to_drain = pcp->count;
1204 free_pcppages_bulk(zone, to_drain, pcp);
1205 pcp->count -= to_drain;
1207 local_irq_restore(flags);
1212 * Drain pages of the indicated processor.
1214 * The processor must either be the current processor and the
1215 * thread pinned to the current processor or a processor that
1218 static void drain_pages(unsigned int cpu)
1220 unsigned long flags;
1223 for_each_populated_zone(zone) {
1224 struct per_cpu_pageset *pset;
1225 struct per_cpu_pages *pcp;
1227 local_irq_save(flags);
1228 pset = per_cpu_ptr(zone->pageset, cpu);
1232 free_pcppages_bulk(zone, pcp->count, pcp);
1235 local_irq_restore(flags);
1240 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1242 void drain_local_pages(void *arg)
1244 drain_pages(smp_processor_id());
1248 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1250 * Note that this code is protected against sending an IPI to an offline
1251 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1252 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1253 * nothing keeps CPUs from showing up after we populated the cpumask and
1254 * before the call to on_each_cpu_mask().
1256 void drain_all_pages(void)
1259 struct per_cpu_pageset *pcp;
1263 * Allocate in the BSS so we wont require allocation in
1264 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1266 static cpumask_t cpus_with_pcps;
1269 * We don't care about racing with CPU hotplug event
1270 * as offline notification will cause the notified
1271 * cpu to drain that CPU pcps and on_each_cpu_mask
1272 * disables preemption as part of its processing
1274 for_each_online_cpu(cpu) {
1275 bool has_pcps = false;
1276 for_each_populated_zone(zone) {
1277 pcp = per_cpu_ptr(zone->pageset, cpu);
1278 if (pcp->pcp.count) {
1284 cpumask_set_cpu(cpu, &cpus_with_pcps);
1286 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1288 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1291 #ifdef CONFIG_HIBERNATION
1293 void mark_free_pages(struct zone *zone)
1295 unsigned long pfn, max_zone_pfn;
1296 unsigned long flags;
1298 struct list_head *curr;
1300 if (!zone->spanned_pages)
1303 spin_lock_irqsave(&zone->lock, flags);
1305 max_zone_pfn = zone_end_pfn(zone);
1306 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1307 if (pfn_valid(pfn)) {
1308 struct page *page = pfn_to_page(pfn);
1310 if (!swsusp_page_is_forbidden(page))
1311 swsusp_unset_page_free(page);
1314 for_each_migratetype_order(order, t) {
1315 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1318 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1319 for (i = 0; i < (1UL << order); i++)
1320 swsusp_set_page_free(pfn_to_page(pfn + i));
1323 spin_unlock_irqrestore(&zone->lock, flags);
1325 #endif /* CONFIG_PM */
1328 * Free a 0-order page
1329 * cold == 1 ? free a cold page : free a hot page
1331 void free_hot_cold_page(struct page *page, int cold)
1333 struct zone *zone = page_zone(page);
1334 struct per_cpu_pages *pcp;
1335 unsigned long flags;
1338 if (!free_pages_prepare(page, 0))
1341 migratetype = get_pageblock_migratetype(page);
1342 set_freepage_migratetype(page, migratetype);
1343 local_irq_save(flags);
1344 __count_vm_event(PGFREE);
1347 * We only track unmovable, reclaimable and movable on pcp lists.
1348 * Free ISOLATE pages back to the allocator because they are being
1349 * offlined but treat RESERVE as movable pages so we can get those
1350 * areas back if necessary. Otherwise, we may have to free
1351 * excessively into the page allocator
1353 if (migratetype >= MIGRATE_PCPTYPES) {
1354 if (unlikely(is_migrate_isolate(migratetype))) {
1355 free_one_page(zone, page, 0, migratetype);
1358 migratetype = MIGRATE_MOVABLE;
1361 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1363 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1365 list_add(&page->lru, &pcp->lists[migratetype]);
1367 if (pcp->count >= pcp->high) {
1368 free_pcppages_bulk(zone, pcp->batch, pcp);
1369 pcp->count -= pcp->batch;
1373 local_irq_restore(flags);
1377 * Free a list of 0-order pages
1379 void free_hot_cold_page_list(struct list_head *list, int cold)
1381 struct page *page, *next;
1383 list_for_each_entry_safe(page, next, list, lru) {
1384 trace_mm_page_free_batched(page, cold);
1385 free_hot_cold_page(page, cold);
1390 * split_page takes a non-compound higher-order page, and splits it into
1391 * n (1<<order) sub-pages: page[0..n]
1392 * Each sub-page must be freed individually.
1394 * Note: this is probably too low level an operation for use in drivers.
1395 * Please consult with lkml before using this in your driver.
1397 void split_page(struct page *page, unsigned int order)
1401 VM_BUG_ON(PageCompound(page));
1402 VM_BUG_ON(!page_count(page));
1404 #ifdef CONFIG_KMEMCHECK
1406 * Split shadow pages too, because free(page[0]) would
1407 * otherwise free the whole shadow.
1409 if (kmemcheck_page_is_tracked(page))
1410 split_page(virt_to_page(page[0].shadow), order);
1413 for (i = 1; i < (1 << order); i++)
1414 set_page_refcounted(page + i);
1416 EXPORT_SYMBOL_GPL(split_page);
1418 static int __isolate_free_page(struct page *page, unsigned int order)
1420 unsigned long watermark;
1424 BUG_ON(!PageBuddy(page));
1426 zone = page_zone(page);
1427 mt = get_pageblock_migratetype(page);
1429 if (!is_migrate_isolate(mt)) {
1430 /* Obey watermarks as if the page was being allocated */
1431 watermark = low_wmark_pages(zone) + (1 << order);
1432 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1435 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1438 /* Remove page from free list */
1439 list_del(&page->lru);
1440 zone->free_area[order].nr_free--;
1441 rmv_page_order(page);
1443 /* Set the pageblock if the isolated page is at least a pageblock */
1444 if (order >= pageblock_order - 1) {
1445 struct page *endpage = page + (1 << order) - 1;
1446 for (; page < endpage; page += pageblock_nr_pages) {
1447 int mt = get_pageblock_migratetype(page);
1448 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1449 set_pageblock_migratetype(page,
1454 return 1UL << order;
1458 * Similar to split_page except the page is already free. As this is only
1459 * being used for migration, the migratetype of the block also changes.
1460 * As this is called with interrupts disabled, the caller is responsible
1461 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1464 * Note: this is probably too low level an operation for use in drivers.
1465 * Please consult with lkml before using this in your driver.
1467 int split_free_page(struct page *page)
1472 order = page_order(page);
1474 nr_pages = __isolate_free_page(page, order);
1478 /* Split into individual pages */
1479 set_page_refcounted(page);
1480 split_page(page, order);
1485 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1486 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1490 struct page *buffered_rmqueue(struct zone *preferred_zone,
1491 struct zone *zone, int order, gfp_t gfp_flags,
1494 unsigned long flags;
1496 int cold = !!(gfp_flags & __GFP_COLD);
1499 if (likely(order == 0)) {
1500 struct per_cpu_pages *pcp;
1501 struct list_head *list;
1503 local_irq_save(flags);
1504 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1505 list = &pcp->lists[migratetype];
1506 if (list_empty(list)) {
1507 pcp->count += rmqueue_bulk(zone, 0,
1510 if (unlikely(list_empty(list)))
1515 page = list_entry(list->prev, struct page, lru);
1517 page = list_entry(list->next, struct page, lru);
1519 list_del(&page->lru);
1522 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1524 * __GFP_NOFAIL is not to be used in new code.
1526 * All __GFP_NOFAIL callers should be fixed so that they
1527 * properly detect and handle allocation failures.
1529 * We most definitely don't want callers attempting to
1530 * allocate greater than order-1 page units with
1533 WARN_ON_ONCE(order > 1);
1535 spin_lock_irqsave(&zone->lock, flags);
1536 page = __rmqueue(zone, order, migratetype);
1537 spin_unlock(&zone->lock);
1540 __mod_zone_freepage_state(zone, -(1 << order),
1541 get_pageblock_migratetype(page));
1544 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1545 zone_statistics(preferred_zone, zone, gfp_flags);
1546 local_irq_restore(flags);
1548 VM_BUG_ON(bad_range(zone, page));
1549 if (prep_new_page(page, order, gfp_flags))
1554 local_irq_restore(flags);
1558 #ifdef CONFIG_FAIL_PAGE_ALLOC
1561 struct fault_attr attr;
1563 u32 ignore_gfp_highmem;
1564 u32 ignore_gfp_wait;
1566 } fail_page_alloc = {
1567 .attr = FAULT_ATTR_INITIALIZER,
1568 .ignore_gfp_wait = 1,
1569 .ignore_gfp_highmem = 1,
1573 static int __init setup_fail_page_alloc(char *str)
1575 return setup_fault_attr(&fail_page_alloc.attr, str);
1577 __setup("fail_page_alloc=", setup_fail_page_alloc);
1579 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1581 if (order < fail_page_alloc.min_order)
1583 if (gfp_mask & __GFP_NOFAIL)
1585 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1587 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1590 return should_fail(&fail_page_alloc.attr, 1 << order);
1593 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1595 static int __init fail_page_alloc_debugfs(void)
1597 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1600 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1601 &fail_page_alloc.attr);
1603 return PTR_ERR(dir);
1605 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1606 &fail_page_alloc.ignore_gfp_wait))
1608 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1609 &fail_page_alloc.ignore_gfp_highmem))
1611 if (!debugfs_create_u32("min-order", mode, dir,
1612 &fail_page_alloc.min_order))
1617 debugfs_remove_recursive(dir);
1622 late_initcall(fail_page_alloc_debugfs);
1624 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1626 #else /* CONFIG_FAIL_PAGE_ALLOC */
1628 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1633 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1636 * Return true if free pages are above 'mark'. This takes into account the order
1637 * of the allocation.
1639 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1640 int classzone_idx, int alloc_flags, long free_pages)
1642 /* free_pages my go negative - that's OK */
1644 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1648 free_pages -= (1 << order) - 1;
1649 if (alloc_flags & ALLOC_HIGH)
1651 if (alloc_flags & ALLOC_HARDER)
1654 /* If allocation can't use CMA areas don't use free CMA pages */
1655 if (!(alloc_flags & ALLOC_CMA))
1656 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1659 if (free_pages - free_cma <= min + lowmem_reserve)
1661 for (o = 0; o < order; o++) {
1662 /* At the next order, this order's pages become unavailable */
1663 free_pages -= z->free_area[o].nr_free << o;
1665 /* Require fewer higher order pages to be free */
1666 min >>= min_free_order_shift;
1668 if (free_pages <= min)
1674 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1675 int classzone_idx, int alloc_flags)
1677 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1678 zone_page_state(z, NR_FREE_PAGES));
1681 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1682 int classzone_idx, int alloc_flags)
1684 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1686 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1687 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1689 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1695 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1696 * skip over zones that are not allowed by the cpuset, or that have
1697 * been recently (in last second) found to be nearly full. See further
1698 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1699 * that have to skip over a lot of full or unallowed zones.
1701 * If the zonelist cache is present in the passed in zonelist, then
1702 * returns a pointer to the allowed node mask (either the current
1703 * tasks mems_allowed, or node_states[N_MEMORY].)
1705 * If the zonelist cache is not available for this zonelist, does
1706 * nothing and returns NULL.
1708 * If the fullzones BITMAP in the zonelist cache is stale (more than
1709 * a second since last zap'd) then we zap it out (clear its bits.)
1711 * We hold off even calling zlc_setup, until after we've checked the
1712 * first zone in the zonelist, on the theory that most allocations will
1713 * be satisfied from that first zone, so best to examine that zone as
1714 * quickly as we can.
1716 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1718 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1719 nodemask_t *allowednodes; /* zonelist_cache approximation */
1721 zlc = zonelist->zlcache_ptr;
1725 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1726 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1727 zlc->last_full_zap = jiffies;
1730 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1731 &cpuset_current_mems_allowed :
1732 &node_states[N_MEMORY];
1733 return allowednodes;
1737 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1738 * if it is worth looking at further for free memory:
1739 * 1) Check that the zone isn't thought to be full (doesn't have its
1740 * bit set in the zonelist_cache fullzones BITMAP).
1741 * 2) Check that the zones node (obtained from the zonelist_cache
1742 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1743 * Return true (non-zero) if zone is worth looking at further, or
1744 * else return false (zero) if it is not.
1746 * This check -ignores- the distinction between various watermarks,
1747 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1748 * found to be full for any variation of these watermarks, it will
1749 * be considered full for up to one second by all requests, unless
1750 * we are so low on memory on all allowed nodes that we are forced
1751 * into the second scan of the zonelist.
1753 * In the second scan we ignore this zonelist cache and exactly
1754 * apply the watermarks to all zones, even it is slower to do so.
1755 * We are low on memory in the second scan, and should leave no stone
1756 * unturned looking for a free page.
1758 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1759 nodemask_t *allowednodes)
1761 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1762 int i; /* index of *z in zonelist zones */
1763 int n; /* node that zone *z is on */
1765 zlc = zonelist->zlcache_ptr;
1769 i = z - zonelist->_zonerefs;
1772 /* This zone is worth trying if it is allowed but not full */
1773 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1777 * Given 'z' scanning a zonelist, set the corresponding bit in
1778 * zlc->fullzones, so that subsequent attempts to allocate a page
1779 * from that zone don't waste time re-examining it.
1781 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1783 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1784 int i; /* index of *z in zonelist zones */
1786 zlc = zonelist->zlcache_ptr;
1790 i = z - zonelist->_zonerefs;
1792 set_bit(i, zlc->fullzones);
1796 * clear all zones full, called after direct reclaim makes progress so that
1797 * a zone that was recently full is not skipped over for up to a second
1799 static void zlc_clear_zones_full(struct zonelist *zonelist)
1801 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1803 zlc = zonelist->zlcache_ptr;
1807 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1810 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1812 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1815 static void __paginginit init_zone_allows_reclaim(int nid)
1819 for_each_online_node(i)
1820 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1821 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1823 zone_reclaim_mode = 1;
1826 #else /* CONFIG_NUMA */
1828 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1833 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1834 nodemask_t *allowednodes)
1839 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1843 static void zlc_clear_zones_full(struct zonelist *zonelist)
1847 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1852 static inline void init_zone_allows_reclaim(int nid)
1855 #endif /* CONFIG_NUMA */
1858 * get_page_from_freelist goes through the zonelist trying to allocate
1861 static struct page *
1862 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1863 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1864 struct zone *preferred_zone, int migratetype)
1867 struct page *page = NULL;
1870 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1871 int zlc_active = 0; /* set if using zonelist_cache */
1872 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1874 classzone_idx = zone_idx(preferred_zone);
1877 * Scan zonelist, looking for a zone with enough free.
1878 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1880 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1881 high_zoneidx, nodemask) {
1882 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1883 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1885 if ((alloc_flags & ALLOC_CPUSET) &&
1886 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1889 * When allocating a page cache page for writing, we
1890 * want to get it from a zone that is within its dirty
1891 * limit, such that no single zone holds more than its
1892 * proportional share of globally allowed dirty pages.
1893 * The dirty limits take into account the zone's
1894 * lowmem reserves and high watermark so that kswapd
1895 * should be able to balance it without having to
1896 * write pages from its LRU list.
1898 * This may look like it could increase pressure on
1899 * lower zones by failing allocations in higher zones
1900 * before they are full. But the pages that do spill
1901 * over are limited as the lower zones are protected
1902 * by this very same mechanism. It should not become
1903 * a practical burden to them.
1905 * XXX: For now, allow allocations to potentially
1906 * exceed the per-zone dirty limit in the slowpath
1907 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1908 * which is important when on a NUMA setup the allowed
1909 * zones are together not big enough to reach the
1910 * global limit. The proper fix for these situations
1911 * will require awareness of zones in the
1912 * dirty-throttling and the flusher threads.
1914 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1915 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1916 goto this_zone_full;
1918 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1919 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1923 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1924 if (zone_watermark_ok(zone, order, mark,
1925 classzone_idx, alloc_flags))
1928 if (IS_ENABLED(CONFIG_NUMA) &&
1929 !did_zlc_setup && nr_online_nodes > 1) {
1931 * we do zlc_setup if there are multiple nodes
1932 * and before considering the first zone allowed
1935 allowednodes = zlc_setup(zonelist, alloc_flags);
1940 if (zone_reclaim_mode == 0 ||
1941 !zone_allows_reclaim(preferred_zone, zone))
1942 goto this_zone_full;
1945 * As we may have just activated ZLC, check if the first
1946 * eligible zone has failed zone_reclaim recently.
1948 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1949 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1952 ret = zone_reclaim(zone, gfp_mask, order);
1954 case ZONE_RECLAIM_NOSCAN:
1957 case ZONE_RECLAIM_FULL:
1958 /* scanned but unreclaimable */
1961 /* did we reclaim enough */
1962 if (zone_watermark_ok(zone, order, mark,
1963 classzone_idx, alloc_flags))
1967 * Failed to reclaim enough to meet watermark.
1968 * Only mark the zone full if checking the min
1969 * watermark or if we failed to reclaim just
1970 * 1<<order pages or else the page allocator
1971 * fastpath will prematurely mark zones full
1972 * when the watermark is between the low and
1975 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
1976 ret == ZONE_RECLAIM_SOME)
1977 goto this_zone_full;
1984 page = buffered_rmqueue(preferred_zone, zone, order,
1985 gfp_mask, migratetype);
1989 if (IS_ENABLED(CONFIG_NUMA))
1990 zlc_mark_zone_full(zonelist, z);
1993 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1994 /* Disable zlc cache for second zonelist scan */
2001 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2002 * necessary to allocate the page. The expectation is
2003 * that the caller is taking steps that will free more
2004 * memory. The caller should avoid the page being used
2005 * for !PFMEMALLOC purposes.
2007 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2013 * Large machines with many possible nodes should not always dump per-node
2014 * meminfo in irq context.
2016 static inline bool should_suppress_show_mem(void)
2021 ret = in_interrupt();
2026 static DEFINE_RATELIMIT_STATE(nopage_rs,
2027 DEFAULT_RATELIMIT_INTERVAL,
2028 DEFAULT_RATELIMIT_BURST);
2030 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2032 unsigned int filter = SHOW_MEM_FILTER_NODES;
2034 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2035 debug_guardpage_minorder() > 0)
2039 * Walking all memory to count page types is very expensive and should
2040 * be inhibited in non-blockable contexts.
2042 if (!(gfp_mask & __GFP_WAIT))
2043 filter |= SHOW_MEM_FILTER_PAGE_COUNT;
2046 * This documents exceptions given to allocations in certain
2047 * contexts that are allowed to allocate outside current's set
2050 if (!(gfp_mask & __GFP_NOMEMALLOC))
2051 if (test_thread_flag(TIF_MEMDIE) ||
2052 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2053 filter &= ~SHOW_MEM_FILTER_NODES;
2054 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2055 filter &= ~SHOW_MEM_FILTER_NODES;
2058 struct va_format vaf;
2061 va_start(args, fmt);
2066 pr_warn("%pV", &vaf);
2071 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2072 current->comm, order, gfp_mask);
2075 if (!should_suppress_show_mem())
2080 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2081 unsigned long did_some_progress,
2082 unsigned long pages_reclaimed)
2084 /* Do not loop if specifically requested */
2085 if (gfp_mask & __GFP_NORETRY)
2088 /* Always retry if specifically requested */
2089 if (gfp_mask & __GFP_NOFAIL)
2093 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2094 * making forward progress without invoking OOM. Suspend also disables
2095 * storage devices so kswapd will not help. Bail if we are suspending.
2097 if (!did_some_progress && pm_suspended_storage())
2101 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2102 * means __GFP_NOFAIL, but that may not be true in other
2105 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2109 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2110 * specified, then we retry until we no longer reclaim any pages
2111 * (above), or we've reclaimed an order of pages at least as
2112 * large as the allocation's order. In both cases, if the
2113 * allocation still fails, we stop retrying.
2115 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2121 static inline struct page *
2122 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2123 struct zonelist *zonelist, enum zone_type high_zoneidx,
2124 nodemask_t *nodemask, struct zone *preferred_zone,
2129 /* Acquire the OOM killer lock for the zones in zonelist */
2130 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2131 schedule_timeout_uninterruptible(1);
2136 * PM-freezer should be notified that there might be an OOM killer on
2137 * its way to kill and wake somebody up. This is too early and we might
2138 * end up not killing anything but false positives are acceptable.
2139 * See freeze_processes.
2144 * Go through the zonelist yet one more time, keep very high watermark
2145 * here, this is only to catch a parallel oom killing, we must fail if
2146 * we're still under heavy pressure.
2148 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2149 order, zonelist, high_zoneidx,
2150 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2151 preferred_zone, migratetype);
2155 if (!(gfp_mask & __GFP_NOFAIL)) {
2156 /* The OOM killer will not help higher order allocs */
2157 if (order > PAGE_ALLOC_COSTLY_ORDER)
2159 /* The OOM killer does not needlessly kill tasks for lowmem */
2160 if (high_zoneidx < ZONE_NORMAL)
2163 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2164 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2165 * The caller should handle page allocation failure by itself if
2166 * it specifies __GFP_THISNODE.
2167 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2169 if (gfp_mask & __GFP_THISNODE)
2172 /* Exhausted what can be done so it's blamo time */
2173 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2176 clear_zonelist_oom(zonelist, gfp_mask);
2180 #ifdef CONFIG_COMPACTION
2181 /* Try memory compaction for high-order allocations before reclaim */
2182 static struct page *
2183 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2184 struct zonelist *zonelist, enum zone_type high_zoneidx,
2185 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2186 int migratetype, bool sync_migration,
2187 bool *contended_compaction, bool *deferred_compaction,
2188 unsigned long *did_some_progress)
2193 if (compaction_deferred(preferred_zone, order)) {
2194 *deferred_compaction = true;
2198 current->flags |= PF_MEMALLOC;
2199 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2200 nodemask, sync_migration,
2201 contended_compaction);
2202 current->flags &= ~PF_MEMALLOC;
2204 if (*did_some_progress != COMPACT_SKIPPED) {
2207 /* Page migration frees to the PCP lists but we want merging */
2208 drain_pages(get_cpu());
2211 page = get_page_from_freelist(gfp_mask, nodemask,
2212 order, zonelist, high_zoneidx,
2213 alloc_flags & ~ALLOC_NO_WATERMARKS,
2214 preferred_zone, migratetype);
2216 preferred_zone->compact_blockskip_flush = false;
2217 preferred_zone->compact_considered = 0;
2218 preferred_zone->compact_defer_shift = 0;
2219 if (order >= preferred_zone->compact_order_failed)
2220 preferred_zone->compact_order_failed = order + 1;
2221 count_vm_event(COMPACTSUCCESS);
2226 * It's bad if compaction run occurs and fails.
2227 * The most likely reason is that pages exist,
2228 * but not enough to satisfy watermarks.
2230 count_vm_event(COMPACTFAIL);
2233 * As async compaction considers a subset of pageblocks, only
2234 * defer if the failure was a sync compaction failure.
2237 defer_compaction(preferred_zone, order);
2245 static inline struct page *
2246 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2247 struct zonelist *zonelist, enum zone_type high_zoneidx,
2248 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2249 int migratetype, bool sync_migration,
2250 bool *contended_compaction, bool *deferred_compaction,
2251 unsigned long *did_some_progress)
2255 #endif /* CONFIG_COMPACTION */
2257 /* Perform direct synchronous page reclaim */
2259 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2260 nodemask_t *nodemask)
2262 struct reclaim_state reclaim_state;
2267 /* We now go into synchronous reclaim */
2268 cpuset_memory_pressure_bump();
2269 current->flags |= PF_MEMALLOC;
2270 lockdep_set_current_reclaim_state(gfp_mask);
2271 reclaim_state.reclaimed_slab = 0;
2272 current->reclaim_state = &reclaim_state;
2274 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2276 current->reclaim_state = NULL;
2277 lockdep_clear_current_reclaim_state();
2278 current->flags &= ~PF_MEMALLOC;
2285 /* The really slow allocator path where we enter direct reclaim */
2286 static inline struct page *
2287 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2288 struct zonelist *zonelist, enum zone_type high_zoneidx,
2289 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2290 int migratetype, unsigned long *did_some_progress)
2292 struct page *page = NULL;
2293 bool drained = false;
2295 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2297 if (unlikely(!(*did_some_progress)))
2300 /* After successful reclaim, reconsider all zones for allocation */
2301 if (IS_ENABLED(CONFIG_NUMA))
2302 zlc_clear_zones_full(zonelist);
2305 page = get_page_from_freelist(gfp_mask, nodemask, order,
2306 zonelist, high_zoneidx,
2307 alloc_flags & ~ALLOC_NO_WATERMARKS,
2308 preferred_zone, migratetype);
2311 * If an allocation failed after direct reclaim, it could be because
2312 * pages are pinned on the per-cpu lists. Drain them and try again
2314 if (!page && !drained) {
2324 * This is called in the allocator slow-path if the allocation request is of
2325 * sufficient urgency to ignore watermarks and take other desperate measures
2327 static inline struct page *
2328 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2329 struct zonelist *zonelist, enum zone_type high_zoneidx,
2330 nodemask_t *nodemask, struct zone *preferred_zone,
2336 page = get_page_from_freelist(gfp_mask, nodemask, order,
2337 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2338 preferred_zone, migratetype);
2340 if (!page && gfp_mask & __GFP_NOFAIL)
2341 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2342 } while (!page && (gfp_mask & __GFP_NOFAIL));
2348 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2349 enum zone_type high_zoneidx,
2350 enum zone_type classzone_idx)
2355 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2356 wakeup_kswapd(zone, order, classzone_idx);
2360 gfp_to_alloc_flags(gfp_t gfp_mask)
2362 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2363 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2365 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2366 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2369 * The caller may dip into page reserves a bit more if the caller
2370 * cannot run direct reclaim, or if the caller has realtime scheduling
2371 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2372 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2374 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2378 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2379 * if it can't schedule.
2381 if (!(gfp_mask & __GFP_NOMEMALLOC))
2382 alloc_flags |= ALLOC_HARDER;
2384 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2385 * comment for __cpuset_node_allowed_softwall().
2387 alloc_flags &= ~ALLOC_CPUSET;
2388 } else if (unlikely(rt_task(current)) && !in_interrupt())
2389 alloc_flags |= ALLOC_HARDER;
2391 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2392 if (gfp_mask & __GFP_MEMALLOC)
2393 alloc_flags |= ALLOC_NO_WATERMARKS;
2394 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2395 alloc_flags |= ALLOC_NO_WATERMARKS;
2396 else if (!in_interrupt() &&
2397 ((current->flags & PF_MEMALLOC) ||
2398 unlikely(test_thread_flag(TIF_MEMDIE))))
2399 alloc_flags |= ALLOC_NO_WATERMARKS;
2402 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2403 alloc_flags |= ALLOC_CMA;
2408 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2410 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2413 static inline struct page *
2414 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2415 struct zonelist *zonelist, enum zone_type high_zoneidx,
2416 nodemask_t *nodemask, struct zone *preferred_zone,
2419 const gfp_t wait = gfp_mask & __GFP_WAIT;
2420 struct page *page = NULL;
2422 unsigned long pages_reclaimed = 0;
2423 unsigned long did_some_progress;
2424 bool sync_migration = false;
2425 bool deferred_compaction = false;
2426 bool contended_compaction = false;
2429 * In the slowpath, we sanity check order to avoid ever trying to
2430 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2431 * be using allocators in order of preference for an area that is
2434 if (order >= MAX_ORDER) {
2435 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2440 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2441 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2442 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2443 * using a larger set of nodes after it has established that the
2444 * allowed per node queues are empty and that nodes are
2447 if (IS_ENABLED(CONFIG_NUMA) &&
2448 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2452 if (!(gfp_mask & __GFP_NO_KSWAPD))
2453 wake_all_kswapd(order, zonelist, high_zoneidx,
2454 zone_idx(preferred_zone));
2457 * OK, we're below the kswapd watermark and have kicked background
2458 * reclaim. Now things get more complex, so set up alloc_flags according
2459 * to how we want to proceed.
2461 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2464 * Find the true preferred zone if the allocation is unconstrained by
2467 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2468 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2472 /* This is the last chance, in general, before the goto nopage. */
2473 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2474 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2475 preferred_zone, migratetype);
2479 /* Allocate without watermarks if the context allows */
2480 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2482 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2483 * the allocation is high priority and these type of
2484 * allocations are system rather than user orientated
2486 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2488 page = __alloc_pages_high_priority(gfp_mask, order,
2489 zonelist, high_zoneidx, nodemask,
2490 preferred_zone, migratetype);
2496 /* Atomic allocations - we can't balance anything */
2500 /* Avoid recursion of direct reclaim */
2501 if (current->flags & PF_MEMALLOC)
2504 /* Avoid allocations with no watermarks from looping endlessly */
2505 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2509 * Try direct compaction. The first pass is asynchronous. Subsequent
2510 * attempts after direct reclaim are synchronous
2512 page = __alloc_pages_direct_compact(gfp_mask, order,
2513 zonelist, high_zoneidx,
2515 alloc_flags, preferred_zone,
2516 migratetype, sync_migration,
2517 &contended_compaction,
2518 &deferred_compaction,
2519 &did_some_progress);
2522 sync_migration = true;
2525 * If compaction is deferred for high-order allocations, it is because
2526 * sync compaction recently failed. In this is the case and the caller
2527 * requested a movable allocation that does not heavily disrupt the
2528 * system then fail the allocation instead of entering direct reclaim.
2530 if ((deferred_compaction || contended_compaction) &&
2531 (gfp_mask & __GFP_NO_KSWAPD))
2534 /* Try direct reclaim and then allocating */
2535 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2536 zonelist, high_zoneidx,
2538 alloc_flags, preferred_zone,
2539 migratetype, &did_some_progress);
2544 * If we failed to make any progress reclaiming, then we are
2545 * running out of options and have to consider going OOM
2547 if (!did_some_progress) {
2548 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2549 if (oom_killer_disabled)
2551 /* Coredumps can quickly deplete all memory reserves */
2552 if ((current->flags & PF_DUMPCORE) &&
2553 !(gfp_mask & __GFP_NOFAIL))
2555 page = __alloc_pages_may_oom(gfp_mask, order,
2556 zonelist, high_zoneidx,
2557 nodemask, preferred_zone,
2562 if (!(gfp_mask & __GFP_NOFAIL)) {
2564 * The oom killer is not called for high-order
2565 * allocations that may fail, so if no progress
2566 * is being made, there are no other options and
2567 * retrying is unlikely to help.
2569 if (order > PAGE_ALLOC_COSTLY_ORDER)
2572 * The oom killer is not called for lowmem
2573 * allocations to prevent needlessly killing
2576 if (high_zoneidx < ZONE_NORMAL)
2584 /* Check if we should retry the allocation */
2585 pages_reclaimed += did_some_progress;
2586 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2588 /* Wait for some write requests to complete then retry */
2589 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2593 * High-order allocations do not necessarily loop after
2594 * direct reclaim and reclaim/compaction depends on compaction
2595 * being called after reclaim so call directly if necessary
2597 page = __alloc_pages_direct_compact(gfp_mask, order,
2598 zonelist, high_zoneidx,
2600 alloc_flags, preferred_zone,
2601 migratetype, sync_migration,
2602 &contended_compaction,
2603 &deferred_compaction,
2604 &did_some_progress);
2610 warn_alloc_failed(gfp_mask, order, NULL);
2613 if (kmemcheck_enabled)
2614 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2620 * This is the 'heart' of the zoned buddy allocator.
2623 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2624 struct zonelist *zonelist, nodemask_t *nodemask)
2626 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2627 struct zone *preferred_zone;
2628 struct page *page = NULL;
2629 int migratetype = allocflags_to_migratetype(gfp_mask);
2630 unsigned int cpuset_mems_cookie;
2631 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2632 struct mem_cgroup *memcg = NULL;
2634 gfp_mask &= gfp_allowed_mask;
2636 lockdep_trace_alloc(gfp_mask);
2638 might_sleep_if(gfp_mask & __GFP_WAIT);
2640 if (should_fail_alloc_page(gfp_mask, order))
2644 * Check the zones suitable for the gfp_mask contain at least one
2645 * valid zone. It's possible to have an empty zonelist as a result
2646 * of GFP_THISNODE and a memoryless node
2648 if (unlikely(!zonelist->_zonerefs->zone))
2652 * Will only have any effect when __GFP_KMEMCG is set. This is
2653 * verified in the (always inline) callee
2655 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2659 cpuset_mems_cookie = get_mems_allowed();
2661 /* The preferred zone is used for statistics later */
2662 first_zones_zonelist(zonelist, high_zoneidx,
2663 nodemask ? : &cpuset_current_mems_allowed,
2665 if (!preferred_zone)
2669 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2670 alloc_flags |= ALLOC_CMA;
2672 /* First allocation attempt */
2673 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2674 zonelist, high_zoneidx, alloc_flags,
2675 preferred_zone, migratetype);
2676 if (unlikely(!page)) {
2678 * Runtime PM, block IO and its error handling path
2679 * can deadlock because I/O on the device might not
2682 gfp_mask = memalloc_noio_flags(gfp_mask);
2683 page = __alloc_pages_slowpath(gfp_mask, order,
2684 zonelist, high_zoneidx, nodemask,
2685 preferred_zone, migratetype);
2688 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2692 * When updating a task's mems_allowed, it is possible to race with
2693 * parallel threads in such a way that an allocation can fail while
2694 * the mask is being updated. If a page allocation is about to fail,
2695 * check if the cpuset changed during allocation and if so, retry.
2697 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2700 memcg_kmem_commit_charge(page, memcg, order);
2704 EXPORT_SYMBOL(__alloc_pages_nodemask);
2707 * Common helper functions.
2709 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2714 * __get_free_pages() returns a 32-bit address, which cannot represent
2717 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2719 page = alloc_pages(gfp_mask, order);
2722 return (unsigned long) page_address(page);
2724 EXPORT_SYMBOL(__get_free_pages);
2726 unsigned long get_zeroed_page(gfp_t gfp_mask)
2728 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2730 EXPORT_SYMBOL(get_zeroed_page);
2732 void __free_pages(struct page *page, unsigned int order)
2734 if (put_page_testzero(page)) {
2736 free_hot_cold_page(page, 0);
2738 __free_pages_ok(page, order);
2742 EXPORT_SYMBOL(__free_pages);
2744 void free_pages(unsigned long addr, unsigned int order)
2747 VM_BUG_ON(!virt_addr_valid((void *)addr));
2748 __free_pages(virt_to_page((void *)addr), order);
2752 EXPORT_SYMBOL(free_pages);
2755 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2756 * pages allocated with __GFP_KMEMCG.
2758 * Those pages are accounted to a particular memcg, embedded in the
2759 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2760 * for that information only to find out that it is NULL for users who have no
2761 * interest in that whatsoever, we provide these functions.
2763 * The caller knows better which flags it relies on.
2765 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2767 memcg_kmem_uncharge_pages(page, order);
2768 __free_pages(page, order);
2771 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2774 VM_BUG_ON(!virt_addr_valid((void *)addr));
2775 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2779 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2782 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2783 unsigned long used = addr + PAGE_ALIGN(size);
2785 split_page(virt_to_page((void *)addr), order);
2786 while (used < alloc_end) {
2791 return (void *)addr;
2795 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2796 * @size: the number of bytes to allocate
2797 * @gfp_mask: GFP flags for the allocation
2799 * This function is similar to alloc_pages(), except that it allocates the
2800 * minimum number of pages to satisfy the request. alloc_pages() can only
2801 * allocate memory in power-of-two pages.
2803 * This function is also limited by MAX_ORDER.
2805 * Memory allocated by this function must be released by free_pages_exact().
2807 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2809 unsigned int order = get_order(size);
2812 addr = __get_free_pages(gfp_mask, order);
2813 return make_alloc_exact(addr, order, size);
2815 EXPORT_SYMBOL(alloc_pages_exact);
2818 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2820 * @nid: the preferred node ID where memory should be allocated
2821 * @size: the number of bytes to allocate
2822 * @gfp_mask: GFP flags for the allocation
2824 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2826 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2829 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2831 unsigned order = get_order(size);
2832 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2835 return make_alloc_exact((unsigned long)page_address(p), order, size);
2837 EXPORT_SYMBOL(alloc_pages_exact_nid);
2840 * free_pages_exact - release memory allocated via alloc_pages_exact()
2841 * @virt: the value returned by alloc_pages_exact.
2842 * @size: size of allocation, same value as passed to alloc_pages_exact().
2844 * Release the memory allocated by a previous call to alloc_pages_exact.
2846 void free_pages_exact(void *virt, size_t size)
2848 unsigned long addr = (unsigned long)virt;
2849 unsigned long end = addr + PAGE_ALIGN(size);
2851 while (addr < end) {
2856 EXPORT_SYMBOL(free_pages_exact);
2859 * nr_free_zone_pages - count number of pages beyond high watermark
2860 * @offset: The zone index of the highest zone
2862 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2863 * high watermark within all zones at or below a given zone index. For each
2864 * zone, the number of pages is calculated as:
2865 * present_pages - high_pages
2867 static unsigned long nr_free_zone_pages(int offset)
2872 /* Just pick one node, since fallback list is circular */
2873 unsigned long sum = 0;
2875 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2877 for_each_zone_zonelist(zone, z, zonelist, offset) {
2878 unsigned long size = zone->managed_pages;
2879 unsigned long high = high_wmark_pages(zone);
2888 * nr_free_buffer_pages - count number of pages beyond high watermark
2890 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2891 * watermark within ZONE_DMA and ZONE_NORMAL.
2893 unsigned long nr_free_buffer_pages(void)
2895 return nr_free_zone_pages(gfp_zone(GFP_USER));
2897 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2900 * nr_free_pagecache_pages - count number of pages beyond high watermark
2902 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2903 * high watermark within all zones.
2905 unsigned long nr_free_pagecache_pages(void)
2907 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2910 static inline void show_node(struct zone *zone)
2912 if (IS_ENABLED(CONFIG_NUMA))
2913 printk("Node %d ", zone_to_nid(zone));
2916 void si_meminfo(struct sysinfo *val)
2918 val->totalram = totalram_pages;
2920 val->freeram = global_page_state(NR_FREE_PAGES);
2921 val->bufferram = nr_blockdev_pages();
2922 val->totalhigh = totalhigh_pages;
2923 val->freehigh = nr_free_highpages();
2924 val->mem_unit = PAGE_SIZE;
2927 EXPORT_SYMBOL(si_meminfo);
2930 void si_meminfo_node(struct sysinfo *val, int nid)
2932 pg_data_t *pgdat = NODE_DATA(nid);
2934 val->totalram = pgdat->node_present_pages;
2935 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2936 #ifdef CONFIG_HIGHMEM
2937 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
2938 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2944 val->mem_unit = PAGE_SIZE;
2949 * Determine whether the node should be displayed or not, depending on whether
2950 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2952 bool skip_free_areas_node(unsigned int flags, int nid)
2955 unsigned int cpuset_mems_cookie;
2957 if (!(flags & SHOW_MEM_FILTER_NODES))
2961 cpuset_mems_cookie = get_mems_allowed();
2962 ret = !node_isset(nid, cpuset_current_mems_allowed);
2963 } while (!put_mems_allowed(cpuset_mems_cookie));
2968 #define K(x) ((x) << (PAGE_SHIFT-10))
2970 static void show_migration_types(unsigned char type)
2972 static const char types[MIGRATE_TYPES] = {
2973 [MIGRATE_UNMOVABLE] = 'U',
2974 [MIGRATE_RECLAIMABLE] = 'E',
2975 [MIGRATE_MOVABLE] = 'M',
2976 [MIGRATE_RESERVE] = 'R',
2978 [MIGRATE_CMA] = 'C',
2980 #ifdef CONFIG_MEMORY_ISOLATION
2981 [MIGRATE_ISOLATE] = 'I',
2984 char tmp[MIGRATE_TYPES + 1];
2988 for (i = 0; i < MIGRATE_TYPES; i++) {
2989 if (type & (1 << i))
2994 printk("(%s) ", tmp);
2998 * Show free area list (used inside shift_scroll-lock stuff)
2999 * We also calculate the percentage fragmentation. We do this by counting the
3000 * memory on each free list with the exception of the first item on the list.
3001 * Suppresses nodes that are not allowed by current's cpuset if
3002 * SHOW_MEM_FILTER_NODES is passed.
3004 void show_free_areas(unsigned int filter)
3009 for_each_populated_zone(zone) {
3010 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3013 printk("%s per-cpu:\n", zone->name);
3015 for_each_online_cpu(cpu) {
3016 struct per_cpu_pageset *pageset;
3018 pageset = per_cpu_ptr(zone->pageset, cpu);
3020 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3021 cpu, pageset->pcp.high,
3022 pageset->pcp.batch, pageset->pcp.count);
3026 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3027 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3029 " dirty:%lu writeback:%lu unstable:%lu\n"
3030 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3031 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3033 global_page_state(NR_ACTIVE_ANON),
3034 global_page_state(NR_INACTIVE_ANON),
3035 global_page_state(NR_ISOLATED_ANON),
3036 global_page_state(NR_ACTIVE_FILE),
3037 global_page_state(NR_INACTIVE_FILE),
3038 global_page_state(NR_ISOLATED_FILE),
3039 global_page_state(NR_UNEVICTABLE),
3040 global_page_state(NR_FILE_DIRTY),
3041 global_page_state(NR_WRITEBACK),
3042 global_page_state(NR_UNSTABLE_NFS),
3043 global_page_state(NR_FREE_PAGES),
3044 global_page_state(NR_SLAB_RECLAIMABLE),
3045 global_page_state(NR_SLAB_UNRECLAIMABLE),
3046 global_page_state(NR_FILE_MAPPED),
3047 global_page_state(NR_SHMEM),
3048 global_page_state(NR_PAGETABLE),
3049 global_page_state(NR_BOUNCE),
3050 global_page_state(NR_FREE_CMA_PAGES));
3052 for_each_populated_zone(zone) {
3055 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3063 " active_anon:%lukB"
3064 " inactive_anon:%lukB"
3065 " active_file:%lukB"
3066 " inactive_file:%lukB"
3067 " unevictable:%lukB"
3068 " isolated(anon):%lukB"
3069 " isolated(file):%lukB"
3077 " slab_reclaimable:%lukB"
3078 " slab_unreclaimable:%lukB"
3079 " kernel_stack:%lukB"
3084 " writeback_tmp:%lukB"
3085 " pages_scanned:%lu"
3086 " all_unreclaimable? %s"
3089 K(zone_page_state(zone, NR_FREE_PAGES)),
3090 K(min_wmark_pages(zone)),
3091 K(low_wmark_pages(zone)),
3092 K(high_wmark_pages(zone)),
3093 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3094 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3095 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3096 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3097 K(zone_page_state(zone, NR_UNEVICTABLE)),
3098 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3099 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3100 K(zone->present_pages),
3101 K(zone->managed_pages),
3102 K(zone_page_state(zone, NR_MLOCK)),
3103 K(zone_page_state(zone, NR_FILE_DIRTY)),
3104 K(zone_page_state(zone, NR_WRITEBACK)),
3105 K(zone_page_state(zone, NR_FILE_MAPPED)),
3106 K(zone_page_state(zone, NR_SHMEM)),
3107 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3108 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3109 zone_page_state(zone, NR_KERNEL_STACK) *
3111 K(zone_page_state(zone, NR_PAGETABLE)),
3112 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3113 K(zone_page_state(zone, NR_BOUNCE)),
3114 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3115 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3116 zone->pages_scanned,
3117 (zone->all_unreclaimable ? "yes" : "no")
3119 printk("lowmem_reserve[]:");
3120 for (i = 0; i < MAX_NR_ZONES; i++)
3121 printk(" %lu", zone->lowmem_reserve[i]);
3125 for_each_populated_zone(zone) {
3126 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3127 unsigned char types[MAX_ORDER];
3129 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3132 printk("%s: ", zone->name);
3134 spin_lock_irqsave(&zone->lock, flags);
3135 for (order = 0; order < MAX_ORDER; order++) {
3136 struct free_area *area = &zone->free_area[order];
3139 nr[order] = area->nr_free;
3140 total += nr[order] << order;
3143 for (type = 0; type < MIGRATE_TYPES; type++) {
3144 if (!list_empty(&area->free_list[type]))
3145 types[order] |= 1 << type;
3148 spin_unlock_irqrestore(&zone->lock, flags);
3149 for (order = 0; order < MAX_ORDER; order++) {
3150 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3152 show_migration_types(types[order]);
3154 printk("= %lukB\n", K(total));
3157 hugetlb_show_meminfo();
3159 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3161 show_swap_cache_info();
3164 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3166 zoneref->zone = zone;
3167 zoneref->zone_idx = zone_idx(zone);
3171 * Builds allocation fallback zone lists.
3173 * Add all populated zones of a node to the zonelist.
3175 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3176 int nr_zones, enum zone_type zone_type)
3180 BUG_ON(zone_type >= MAX_NR_ZONES);
3185 zone = pgdat->node_zones + zone_type;
3186 if (populated_zone(zone)) {
3187 zoneref_set_zone(zone,
3188 &zonelist->_zonerefs[nr_zones++]);
3189 check_highest_zone(zone_type);
3192 } while (zone_type);
3199 * 0 = automatic detection of better ordering.
3200 * 1 = order by ([node] distance, -zonetype)
3201 * 2 = order by (-zonetype, [node] distance)
3203 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3204 * the same zonelist. So only NUMA can configure this param.
3206 #define ZONELIST_ORDER_DEFAULT 0
3207 #define ZONELIST_ORDER_NODE 1
3208 #define ZONELIST_ORDER_ZONE 2
3210 /* zonelist order in the kernel.
3211 * set_zonelist_order() will set this to NODE or ZONE.
3213 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3214 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3218 /* The value user specified ....changed by config */
3219 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3220 /* string for sysctl */
3221 #define NUMA_ZONELIST_ORDER_LEN 16
3222 char numa_zonelist_order[16] = "default";
3225 * interface for configure zonelist ordering.
3226 * command line option "numa_zonelist_order"
3227 * = "[dD]efault - default, automatic configuration.
3228 * = "[nN]ode - order by node locality, then by zone within node
3229 * = "[zZ]one - order by zone, then by locality within zone
3232 static int __parse_numa_zonelist_order(char *s)
3234 if (*s == 'd' || *s == 'D') {
3235 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3236 } else if (*s == 'n' || *s == 'N') {
3237 user_zonelist_order = ZONELIST_ORDER_NODE;
3238 } else if (*s == 'z' || *s == 'Z') {
3239 user_zonelist_order = ZONELIST_ORDER_ZONE;
3242 "Ignoring invalid numa_zonelist_order value: "
3249 static __init int setup_numa_zonelist_order(char *s)
3256 ret = __parse_numa_zonelist_order(s);
3258 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3262 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3265 * sysctl handler for numa_zonelist_order
3267 int numa_zonelist_order_handler(ctl_table *table, int write,
3268 void __user *buffer, size_t *length,
3271 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3273 static DEFINE_MUTEX(zl_order_mutex);
3275 mutex_lock(&zl_order_mutex);
3277 strcpy(saved_string, (char*)table->data);
3278 ret = proc_dostring(table, write, buffer, length, ppos);
3282 int oldval = user_zonelist_order;
3283 if (__parse_numa_zonelist_order((char*)table->data)) {
3285 * bogus value. restore saved string
3287 strncpy((char*)table->data, saved_string,
3288 NUMA_ZONELIST_ORDER_LEN);
3289 user_zonelist_order = oldval;
3290 } else if (oldval != user_zonelist_order) {
3291 mutex_lock(&zonelists_mutex);
3292 build_all_zonelists(NULL, NULL);
3293 mutex_unlock(&zonelists_mutex);
3297 mutex_unlock(&zl_order_mutex);
3302 #define MAX_NODE_LOAD (nr_online_nodes)
3303 static int node_load[MAX_NUMNODES];
3306 * find_next_best_node - find the next node that should appear in a given node's fallback list
3307 * @node: node whose fallback list we're appending
3308 * @used_node_mask: nodemask_t of already used nodes
3310 * We use a number of factors to determine which is the next node that should
3311 * appear on a given node's fallback list. The node should not have appeared
3312 * already in @node's fallback list, and it should be the next closest node
3313 * according to the distance array (which contains arbitrary distance values
3314 * from each node to each node in the system), and should also prefer nodes
3315 * with no CPUs, since presumably they'll have very little allocation pressure
3316 * on them otherwise.
3317 * It returns -1 if no node is found.
3319 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3322 int min_val = INT_MAX;
3323 int best_node = NUMA_NO_NODE;
3324 const struct cpumask *tmp = cpumask_of_node(0);
3326 /* Use the local node if we haven't already */
3327 if (!node_isset(node, *used_node_mask)) {
3328 node_set(node, *used_node_mask);
3332 for_each_node_state(n, N_MEMORY) {
3334 /* Don't want a node to appear more than once */
3335 if (node_isset(n, *used_node_mask))
3338 /* Use the distance array to find the distance */
3339 val = node_distance(node, n);
3341 /* Penalize nodes under us ("prefer the next node") */
3344 /* Give preference to headless and unused nodes */
3345 tmp = cpumask_of_node(n);
3346 if (!cpumask_empty(tmp))
3347 val += PENALTY_FOR_NODE_WITH_CPUS;
3349 /* Slight preference for less loaded node */
3350 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3351 val += node_load[n];
3353 if (val < min_val) {
3360 node_set(best_node, *used_node_mask);
3367 * Build zonelists ordered by node and zones within node.
3368 * This results in maximum locality--normal zone overflows into local
3369 * DMA zone, if any--but risks exhausting DMA zone.
3371 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3374 struct zonelist *zonelist;
3376 zonelist = &pgdat->node_zonelists[0];
3377 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3379 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3381 zonelist->_zonerefs[j].zone = NULL;
3382 zonelist->_zonerefs[j].zone_idx = 0;
3386 * Build gfp_thisnode zonelists
3388 static void build_thisnode_zonelists(pg_data_t *pgdat)
3391 struct zonelist *zonelist;
3393 zonelist = &pgdat->node_zonelists[1];
3394 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3395 zonelist->_zonerefs[j].zone = NULL;
3396 zonelist->_zonerefs[j].zone_idx = 0;
3400 * Build zonelists ordered by zone and nodes within zones.
3401 * This results in conserving DMA zone[s] until all Normal memory is
3402 * exhausted, but results in overflowing to remote node while memory
3403 * may still exist in local DMA zone.
3405 static int node_order[MAX_NUMNODES];
3407 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3410 int zone_type; /* needs to be signed */
3412 struct zonelist *zonelist;
3414 zonelist = &pgdat->node_zonelists[0];
3416 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3417 for (j = 0; j < nr_nodes; j++) {
3418 node = node_order[j];
3419 z = &NODE_DATA(node)->node_zones[zone_type];
3420 if (populated_zone(z)) {
3422 &zonelist->_zonerefs[pos++]);
3423 check_highest_zone(zone_type);
3427 zonelist->_zonerefs[pos].zone = NULL;
3428 zonelist->_zonerefs[pos].zone_idx = 0;
3431 static int default_zonelist_order(void)
3434 unsigned long low_kmem_size,total_size;
3438 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3439 * If they are really small and used heavily, the system can fall
3440 * into OOM very easily.
3441 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3443 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3446 for_each_online_node(nid) {
3447 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3448 z = &NODE_DATA(nid)->node_zones[zone_type];
3449 if (populated_zone(z)) {
3450 if (zone_type < ZONE_NORMAL)
3451 low_kmem_size += z->present_pages;
3452 total_size += z->present_pages;
3453 } else if (zone_type == ZONE_NORMAL) {
3455 * If any node has only lowmem, then node order
3456 * is preferred to allow kernel allocations
3457 * locally; otherwise, they can easily infringe
3458 * on other nodes when there is an abundance of
3459 * lowmem available to allocate from.
3461 return ZONELIST_ORDER_NODE;
3465 if (!low_kmem_size || /* there are no DMA area. */
3466 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3467 return ZONELIST_ORDER_NODE;
3469 * look into each node's config.
3470 * If there is a node whose DMA/DMA32 memory is very big area on
3471 * local memory, NODE_ORDER may be suitable.
3473 average_size = total_size /
3474 (nodes_weight(node_states[N_MEMORY]) + 1);
3475 for_each_online_node(nid) {
3478 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3479 z = &NODE_DATA(nid)->node_zones[zone_type];
3480 if (populated_zone(z)) {
3481 if (zone_type < ZONE_NORMAL)
3482 low_kmem_size += z->present_pages;
3483 total_size += z->present_pages;
3486 if (low_kmem_size &&
3487 total_size > average_size && /* ignore small node */
3488 low_kmem_size > total_size * 70/100)
3489 return ZONELIST_ORDER_NODE;
3491 return ZONELIST_ORDER_ZONE;
3494 static void set_zonelist_order(void)
3496 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3497 current_zonelist_order = default_zonelist_order();
3499 current_zonelist_order = user_zonelist_order;
3502 static void build_zonelists(pg_data_t *pgdat)
3506 nodemask_t used_mask;
3507 int local_node, prev_node;
3508 struct zonelist *zonelist;
3509 int order = current_zonelist_order;
3511 /* initialize zonelists */
3512 for (i = 0; i < MAX_ZONELISTS; i++) {
3513 zonelist = pgdat->node_zonelists + i;
3514 zonelist->_zonerefs[0].zone = NULL;
3515 zonelist->_zonerefs[0].zone_idx = 0;
3518 /* NUMA-aware ordering of nodes */
3519 local_node = pgdat->node_id;
3520 load = nr_online_nodes;
3521 prev_node = local_node;
3522 nodes_clear(used_mask);
3524 memset(node_order, 0, sizeof(node_order));
3527 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3529 * We don't want to pressure a particular node.
3530 * So adding penalty to the first node in same
3531 * distance group to make it round-robin.
3533 if (node_distance(local_node, node) !=
3534 node_distance(local_node, prev_node))
3535 node_load[node] = load;
3539 if (order == ZONELIST_ORDER_NODE)
3540 build_zonelists_in_node_order(pgdat, node);
3542 node_order[j++] = node; /* remember order */
3545 if (order == ZONELIST_ORDER_ZONE) {
3546 /* calculate node order -- i.e., DMA last! */
3547 build_zonelists_in_zone_order(pgdat, j);
3550 build_thisnode_zonelists(pgdat);
3553 /* Construct the zonelist performance cache - see further mmzone.h */
3554 static void build_zonelist_cache(pg_data_t *pgdat)
3556 struct zonelist *zonelist;
3557 struct zonelist_cache *zlc;
3560 zonelist = &pgdat->node_zonelists[0];
3561 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3562 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3563 for (z = zonelist->_zonerefs; z->zone; z++)
3564 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3567 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3569 * Return node id of node used for "local" allocations.
3570 * I.e., first node id of first zone in arg node's generic zonelist.
3571 * Used for initializing percpu 'numa_mem', which is used primarily
3572 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3574 int local_memory_node(int node)
3578 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3579 gfp_zone(GFP_KERNEL),
3586 #else /* CONFIG_NUMA */
3588 static void set_zonelist_order(void)
3590 current_zonelist_order = ZONELIST_ORDER_ZONE;
3593 static void build_zonelists(pg_data_t *pgdat)
3595 int node, local_node;
3597 struct zonelist *zonelist;
3599 local_node = pgdat->node_id;
3601 zonelist = &pgdat->node_zonelists[0];
3602 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3605 * Now we build the zonelist so that it contains the zones
3606 * of all the other nodes.
3607 * We don't want to pressure a particular node, so when
3608 * building the zones for node N, we make sure that the
3609 * zones coming right after the local ones are those from
3610 * node N+1 (modulo N)
3612 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3613 if (!node_online(node))
3615 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3618 for (node = 0; node < local_node; node++) {
3619 if (!node_online(node))
3621 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3625 zonelist->_zonerefs[j].zone = NULL;
3626 zonelist->_zonerefs[j].zone_idx = 0;
3629 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3630 static void build_zonelist_cache(pg_data_t *pgdat)
3632 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3635 #endif /* CONFIG_NUMA */
3638 * Boot pageset table. One per cpu which is going to be used for all
3639 * zones and all nodes. The parameters will be set in such a way
3640 * that an item put on a list will immediately be handed over to
3641 * the buddy list. This is safe since pageset manipulation is done
3642 * with interrupts disabled.
3644 * The boot_pagesets must be kept even after bootup is complete for
3645 * unused processors and/or zones. They do play a role for bootstrapping
3646 * hotplugged processors.
3648 * zoneinfo_show() and maybe other functions do
3649 * not check if the processor is online before following the pageset pointer.
3650 * Other parts of the kernel may not check if the zone is available.
3652 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3653 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3654 static void setup_zone_pageset(struct zone *zone);
3657 * Global mutex to protect against size modification of zonelists
3658 * as well as to serialize pageset setup for the new populated zone.
3660 DEFINE_MUTEX(zonelists_mutex);
3662 /* return values int ....just for stop_machine() */
3663 static int __build_all_zonelists(void *data)
3667 pg_data_t *self = data;
3670 memset(node_load, 0, sizeof(node_load));
3673 if (self && !node_online(self->node_id)) {
3674 build_zonelists(self);
3675 build_zonelist_cache(self);
3678 for_each_online_node(nid) {
3679 pg_data_t *pgdat = NODE_DATA(nid);
3681 build_zonelists(pgdat);
3682 build_zonelist_cache(pgdat);
3686 * Initialize the boot_pagesets that are going to be used
3687 * for bootstrapping processors. The real pagesets for
3688 * each zone will be allocated later when the per cpu
3689 * allocator is available.
3691 * boot_pagesets are used also for bootstrapping offline
3692 * cpus if the system is already booted because the pagesets
3693 * are needed to initialize allocators on a specific cpu too.
3694 * F.e. the percpu allocator needs the page allocator which
3695 * needs the percpu allocator in order to allocate its pagesets
3696 * (a chicken-egg dilemma).
3698 for_each_possible_cpu(cpu) {
3699 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3701 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3703 * We now know the "local memory node" for each node--
3704 * i.e., the node of the first zone in the generic zonelist.
3705 * Set up numa_mem percpu variable for on-line cpus. During
3706 * boot, only the boot cpu should be on-line; we'll init the
3707 * secondary cpus' numa_mem as they come on-line. During
3708 * node/memory hotplug, we'll fixup all on-line cpus.
3710 if (cpu_online(cpu))
3711 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3719 * Called with zonelists_mutex held always
3720 * unless system_state == SYSTEM_BOOTING.
3722 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3724 set_zonelist_order();
3726 if (system_state == SYSTEM_BOOTING) {
3727 __build_all_zonelists(NULL);
3728 mminit_verify_zonelist();
3729 cpuset_init_current_mems_allowed();
3731 /* we have to stop all cpus to guarantee there is no user
3733 #ifdef CONFIG_MEMORY_HOTPLUG
3735 setup_zone_pageset(zone);
3737 stop_machine(__build_all_zonelists, pgdat, NULL);
3738 /* cpuset refresh routine should be here */
3740 vm_total_pages = nr_free_pagecache_pages();
3742 * Disable grouping by mobility if the number of pages in the
3743 * system is too low to allow the mechanism to work. It would be
3744 * more accurate, but expensive to check per-zone. This check is
3745 * made on memory-hotadd so a system can start with mobility
3746 * disabled and enable it later
3748 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3749 page_group_by_mobility_disabled = 1;
3751 page_group_by_mobility_disabled = 0;
3753 printk("Built %i zonelists in %s order, mobility grouping %s. "
3754 "Total pages: %ld\n",
3756 zonelist_order_name[current_zonelist_order],
3757 page_group_by_mobility_disabled ? "off" : "on",
3760 printk("Policy zone: %s\n", zone_names[policy_zone]);
3765 * Helper functions to size the waitqueue hash table.
3766 * Essentially these want to choose hash table sizes sufficiently
3767 * large so that collisions trying to wait on pages are rare.
3768 * But in fact, the number of active page waitqueues on typical
3769 * systems is ridiculously low, less than 200. So this is even
3770 * conservative, even though it seems large.
3772 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3773 * waitqueues, i.e. the size of the waitq table given the number of pages.
3775 #define PAGES_PER_WAITQUEUE 256
3777 #ifndef CONFIG_MEMORY_HOTPLUG
3778 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3780 unsigned long size = 1;
3782 pages /= PAGES_PER_WAITQUEUE;
3784 while (size < pages)
3788 * Once we have dozens or even hundreds of threads sleeping
3789 * on IO we've got bigger problems than wait queue collision.
3790 * Limit the size of the wait table to a reasonable size.
3792 size = min(size, 4096UL);
3794 return max(size, 4UL);
3798 * A zone's size might be changed by hot-add, so it is not possible to determine
3799 * a suitable size for its wait_table. So we use the maximum size now.
3801 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3803 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3804 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3805 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3807 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3808 * or more by the traditional way. (See above). It equals:
3810 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3811 * ia64(16K page size) : = ( 8G + 4M)byte.
3812 * powerpc (64K page size) : = (32G +16M)byte.
3814 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3821 * This is an integer logarithm so that shifts can be used later
3822 * to extract the more random high bits from the multiplicative
3823 * hash function before the remainder is taken.
3825 static inline unsigned long wait_table_bits(unsigned long size)
3830 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3833 * Check if a pageblock contains reserved pages
3835 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3839 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3840 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3847 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3848 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3849 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3850 * higher will lead to a bigger reserve which will get freed as contiguous
3851 * blocks as reclaim kicks in
3853 static void setup_zone_migrate_reserve(struct zone *zone)
3855 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3857 unsigned long block_migratetype;
3861 * Get the start pfn, end pfn and the number of blocks to reserve
3862 * We have to be careful to be aligned to pageblock_nr_pages to
3863 * make sure that we always check pfn_valid for the first page in
3866 start_pfn = zone->zone_start_pfn;
3867 end_pfn = zone_end_pfn(zone);
3868 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3869 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3873 * Reserve blocks are generally in place to help high-order atomic
3874 * allocations that are short-lived. A min_free_kbytes value that
3875 * would result in more than 2 reserve blocks for atomic allocations
3876 * is assumed to be in place to help anti-fragmentation for the
3877 * future allocation of hugepages at runtime.
3879 reserve = min(2, reserve);
3881 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3882 if (!pfn_valid(pfn))
3884 page = pfn_to_page(pfn);
3886 /* Watch out for overlapping nodes */
3887 if (page_to_nid(page) != zone_to_nid(zone))
3890 block_migratetype = get_pageblock_migratetype(page);
3892 /* Only test what is necessary when the reserves are not met */
3895 * Blocks with reserved pages will never free, skip
3898 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3899 if (pageblock_is_reserved(pfn, block_end_pfn))
3902 /* If this block is reserved, account for it */
3903 if (block_migratetype == MIGRATE_RESERVE) {
3908 /* Suitable for reserving if this block is movable */
3909 if (block_migratetype == MIGRATE_MOVABLE) {
3910 set_pageblock_migratetype(page,
3912 move_freepages_block(zone, page,
3920 * If the reserve is met and this is a previous reserved block,
3923 if (block_migratetype == MIGRATE_RESERVE) {
3924 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3925 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3931 * Initially all pages are reserved - free ones are freed
3932 * up by free_all_bootmem() once the early boot process is
3933 * done. Non-atomic initialization, single-pass.
3935 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3936 unsigned long start_pfn, enum memmap_context context)
3939 unsigned long end_pfn = start_pfn + size;
3943 if (highest_memmap_pfn < end_pfn - 1)
3944 highest_memmap_pfn = end_pfn - 1;
3946 z = &NODE_DATA(nid)->node_zones[zone];
3947 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3949 * There can be holes in boot-time mem_map[]s
3950 * handed to this function. They do not
3951 * exist on hotplugged memory.
3953 if (context == MEMMAP_EARLY) {
3954 if (!early_pfn_valid(pfn))
3956 if (!early_pfn_in_nid(pfn, nid))
3959 page = pfn_to_page(pfn);
3960 set_page_links(page, zone, nid, pfn);
3961 mminit_verify_page_links(page, zone, nid, pfn);
3962 init_page_count(page);
3963 page_mapcount_reset(page);
3964 page_nid_reset_last(page);
3965 SetPageReserved(page);
3967 * Mark the block movable so that blocks are reserved for
3968 * movable at startup. This will force kernel allocations
3969 * to reserve their blocks rather than leaking throughout
3970 * the address space during boot when many long-lived
3971 * kernel allocations are made. Later some blocks near
3972 * the start are marked MIGRATE_RESERVE by
3973 * setup_zone_migrate_reserve()
3975 * bitmap is created for zone's valid pfn range. but memmap
3976 * can be created for invalid pages (for alignment)
3977 * check here not to call set_pageblock_migratetype() against
3980 if ((z->zone_start_pfn <= pfn)
3981 && (pfn < zone_end_pfn(z))
3982 && !(pfn & (pageblock_nr_pages - 1)))
3983 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3985 INIT_LIST_HEAD(&page->lru);
3986 #ifdef WANT_PAGE_VIRTUAL
3987 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3988 if (!is_highmem_idx(zone))
3989 set_page_address(page, __va(pfn << PAGE_SHIFT));
3994 static void __meminit zone_init_free_lists(struct zone *zone)
3997 for_each_migratetype_order(order, t) {
3998 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3999 zone->free_area[order].nr_free = 0;
4003 #ifndef __HAVE_ARCH_MEMMAP_INIT
4004 #define memmap_init(size, nid, zone, start_pfn) \
4005 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4008 static int __meminit zone_batchsize(struct zone *zone)
4014 * The per-cpu-pages pools are set to around 1000th of the
4015 * size of the zone. But no more than 1/2 of a meg.
4017 * OK, so we don't know how big the cache is. So guess.
4019 batch = zone->managed_pages / 1024;
4020 if (batch * PAGE_SIZE > 512 * 1024)
4021 batch = (512 * 1024) / PAGE_SIZE;
4022 batch /= 4; /* We effectively *= 4 below */
4027 * Clamp the batch to a 2^n - 1 value. Having a power
4028 * of 2 value was found to be more likely to have
4029 * suboptimal cache aliasing properties in some cases.
4031 * For example if 2 tasks are alternately allocating
4032 * batches of pages, one task can end up with a lot
4033 * of pages of one half of the possible page colors
4034 * and the other with pages of the other colors.
4036 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4041 /* The deferral and batching of frees should be suppressed under NOMMU
4044 * The problem is that NOMMU needs to be able to allocate large chunks
4045 * of contiguous memory as there's no hardware page translation to
4046 * assemble apparent contiguous memory from discontiguous pages.
4048 * Queueing large contiguous runs of pages for batching, however,
4049 * causes the pages to actually be freed in smaller chunks. As there
4050 * can be a significant delay between the individual batches being
4051 * recycled, this leads to the once large chunks of space being
4052 * fragmented and becoming unavailable for high-order allocations.
4058 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4060 struct per_cpu_pages *pcp;
4063 memset(p, 0, sizeof(*p));
4067 pcp->high = 6 * batch;
4068 pcp->batch = max(1UL, 1 * batch);
4069 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4070 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4074 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4075 * to the value high for the pageset p.
4078 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
4081 struct per_cpu_pages *pcp;
4085 pcp->batch = max(1UL, high/4);
4086 if ((high/4) > (PAGE_SHIFT * 8))
4087 pcp->batch = PAGE_SHIFT * 8;
4090 static void __meminit setup_zone_pageset(struct zone *zone)
4094 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4096 for_each_possible_cpu(cpu) {
4097 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4099 setup_pageset(pcp, zone_batchsize(zone));
4101 if (percpu_pagelist_fraction)
4102 setup_pagelist_highmark(pcp,
4103 (zone->managed_pages /
4104 percpu_pagelist_fraction));
4109 * Allocate per cpu pagesets and initialize them.
4110 * Before this call only boot pagesets were available.
4112 void __init setup_per_cpu_pageset(void)
4116 for_each_populated_zone(zone)
4117 setup_zone_pageset(zone);
4120 static noinline __init_refok
4121 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4124 struct pglist_data *pgdat = zone->zone_pgdat;
4128 * The per-page waitqueue mechanism uses hashed waitqueues
4131 zone->wait_table_hash_nr_entries =
4132 wait_table_hash_nr_entries(zone_size_pages);
4133 zone->wait_table_bits =
4134 wait_table_bits(zone->wait_table_hash_nr_entries);
4135 alloc_size = zone->wait_table_hash_nr_entries
4136 * sizeof(wait_queue_head_t);
4138 if (!slab_is_available()) {
4139 zone->wait_table = (wait_queue_head_t *)
4140 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4143 * This case means that a zone whose size was 0 gets new memory
4144 * via memory hot-add.
4145 * But it may be the case that a new node was hot-added. In
4146 * this case vmalloc() will not be able to use this new node's
4147 * memory - this wait_table must be initialized to use this new
4148 * node itself as well.
4149 * To use this new node's memory, further consideration will be
4152 zone->wait_table = vmalloc(alloc_size);
4154 if (!zone->wait_table)
4157 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4158 init_waitqueue_head(zone->wait_table + i);
4163 static __meminit void zone_pcp_init(struct zone *zone)
4166 * per cpu subsystem is not up at this point. The following code
4167 * relies on the ability of the linker to provide the
4168 * offset of a (static) per cpu variable into the per cpu area.
4170 zone->pageset = &boot_pageset;
4172 if (zone->present_pages)
4173 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4174 zone->name, zone->present_pages,
4175 zone_batchsize(zone));
4178 int __meminit init_currently_empty_zone(struct zone *zone,
4179 unsigned long zone_start_pfn,
4181 enum memmap_context context)
4183 struct pglist_data *pgdat = zone->zone_pgdat;
4185 ret = zone_wait_table_init(zone, size);
4188 pgdat->nr_zones = zone_idx(zone) + 1;
4190 zone->zone_start_pfn = zone_start_pfn;
4192 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4193 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4195 (unsigned long)zone_idx(zone),
4196 zone_start_pfn, (zone_start_pfn + size));
4198 zone_init_free_lists(zone);
4203 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4204 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4206 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4207 * Architectures may implement their own version but if add_active_range()
4208 * was used and there are no special requirements, this is a convenient
4211 int __meminit __early_pfn_to_nid(unsigned long pfn)
4213 unsigned long start_pfn, end_pfn;
4216 * NOTE: The following SMP-unsafe globals are only used early in boot
4217 * when the kernel is running single-threaded.
4219 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4220 static int __meminitdata last_nid;
4222 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4225 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4226 if (start_pfn <= pfn && pfn < end_pfn) {
4227 last_start_pfn = start_pfn;
4228 last_end_pfn = end_pfn;
4232 /* This is a memory hole */
4235 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4237 int __meminit early_pfn_to_nid(unsigned long pfn)
4241 nid = __early_pfn_to_nid(pfn);
4244 /* just returns 0 */
4248 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4249 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4253 nid = __early_pfn_to_nid(pfn);
4254 if (nid >= 0 && nid != node)
4261 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4262 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4263 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4265 * If an architecture guarantees that all ranges registered with
4266 * add_active_ranges() contain no holes and may be freed, this
4267 * this function may be used instead of calling free_bootmem() manually.
4269 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4271 unsigned long start_pfn, end_pfn;
4274 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4275 start_pfn = min(start_pfn, max_low_pfn);
4276 end_pfn = min(end_pfn, max_low_pfn);
4278 if (start_pfn < end_pfn)
4279 free_bootmem_node(NODE_DATA(this_nid),
4280 PFN_PHYS(start_pfn),
4281 (end_pfn - start_pfn) << PAGE_SHIFT);
4286 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4287 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4289 * If an architecture guarantees that all ranges registered with
4290 * add_active_ranges() contain no holes and may be freed, this
4291 * function may be used instead of calling memory_present() manually.
4293 void __init sparse_memory_present_with_active_regions(int nid)
4295 unsigned long start_pfn, end_pfn;
4298 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4299 memory_present(this_nid, start_pfn, end_pfn);
4303 * get_pfn_range_for_nid - Return the start and end page frames for a node
4304 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4305 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4306 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4308 * It returns the start and end page frame of a node based on information
4309 * provided by an arch calling add_active_range(). If called for a node
4310 * with no available memory, a warning is printed and the start and end
4313 void __meminit get_pfn_range_for_nid(unsigned int nid,
4314 unsigned long *start_pfn, unsigned long *end_pfn)
4316 unsigned long this_start_pfn, this_end_pfn;
4322 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4323 *start_pfn = min(*start_pfn, this_start_pfn);
4324 *end_pfn = max(*end_pfn, this_end_pfn);
4327 if (*start_pfn == -1UL)
4332 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4333 * assumption is made that zones within a node are ordered in monotonic
4334 * increasing memory addresses so that the "highest" populated zone is used
4336 static void __init find_usable_zone_for_movable(void)
4339 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4340 if (zone_index == ZONE_MOVABLE)
4343 if (arch_zone_highest_possible_pfn[zone_index] >
4344 arch_zone_lowest_possible_pfn[zone_index])
4348 VM_BUG_ON(zone_index == -1);
4349 movable_zone = zone_index;
4353 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4354 * because it is sized independent of architecture. Unlike the other zones,
4355 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4356 * in each node depending on the size of each node and how evenly kernelcore
4357 * is distributed. This helper function adjusts the zone ranges
4358 * provided by the architecture for a given node by using the end of the
4359 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4360 * zones within a node are in order of monotonic increases memory addresses
4362 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4363 unsigned long zone_type,
4364 unsigned long node_start_pfn,
4365 unsigned long node_end_pfn,
4366 unsigned long *zone_start_pfn,
4367 unsigned long *zone_end_pfn)
4369 /* Only adjust if ZONE_MOVABLE is on this node */
4370 if (zone_movable_pfn[nid]) {
4371 /* Size ZONE_MOVABLE */
4372 if (zone_type == ZONE_MOVABLE) {
4373 *zone_start_pfn = zone_movable_pfn[nid];
4374 *zone_end_pfn = min(node_end_pfn,
4375 arch_zone_highest_possible_pfn[movable_zone]);
4377 /* Adjust for ZONE_MOVABLE starting within this range */
4378 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4379 *zone_end_pfn > zone_movable_pfn[nid]) {
4380 *zone_end_pfn = zone_movable_pfn[nid];
4382 /* Check if this whole range is within ZONE_MOVABLE */
4383 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4384 *zone_start_pfn = *zone_end_pfn;
4389 * Return the number of pages a zone spans in a node, including holes
4390 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4392 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4393 unsigned long zone_type,
4394 unsigned long *ignored)
4396 unsigned long node_start_pfn, node_end_pfn;
4397 unsigned long zone_start_pfn, zone_end_pfn;
4399 /* Get the start and end of the node and zone */
4400 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4401 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4402 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4403 adjust_zone_range_for_zone_movable(nid, zone_type,
4404 node_start_pfn, node_end_pfn,
4405 &zone_start_pfn, &zone_end_pfn);
4407 /* Check that this node has pages within the zone's required range */
4408 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4411 /* Move the zone boundaries inside the node if necessary */
4412 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4413 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4415 /* Return the spanned pages */
4416 return zone_end_pfn - zone_start_pfn;
4420 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4421 * then all holes in the requested range will be accounted for.
4423 unsigned long __meminit __absent_pages_in_range(int nid,
4424 unsigned long range_start_pfn,
4425 unsigned long range_end_pfn)
4427 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4428 unsigned long start_pfn, end_pfn;
4431 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4432 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4433 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4434 nr_absent -= end_pfn - start_pfn;
4440 * absent_pages_in_range - Return number of page frames in holes within a range
4441 * @start_pfn: The start PFN to start searching for holes
4442 * @end_pfn: The end PFN to stop searching for holes
4444 * It returns the number of pages frames in memory holes within a range.
4446 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4447 unsigned long end_pfn)
4449 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4452 /* Return the number of page frames in holes in a zone on a node */
4453 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4454 unsigned long zone_type,
4455 unsigned long *ignored)
4457 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4458 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4459 unsigned long node_start_pfn, node_end_pfn;
4460 unsigned long zone_start_pfn, zone_end_pfn;
4462 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4463 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4464 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4466 adjust_zone_range_for_zone_movable(nid, zone_type,
4467 node_start_pfn, node_end_pfn,
4468 &zone_start_pfn, &zone_end_pfn);
4469 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4472 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4473 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4474 unsigned long zone_type,
4475 unsigned long *zones_size)
4477 return zones_size[zone_type];
4480 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4481 unsigned long zone_type,
4482 unsigned long *zholes_size)
4487 return zholes_size[zone_type];
4490 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4492 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4493 unsigned long *zones_size, unsigned long *zholes_size)
4495 unsigned long realtotalpages, totalpages = 0;
4498 for (i = 0; i < MAX_NR_ZONES; i++)
4499 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4501 pgdat->node_spanned_pages = totalpages;
4503 realtotalpages = totalpages;
4504 for (i = 0; i < MAX_NR_ZONES; i++)
4506 zone_absent_pages_in_node(pgdat->node_id, i,
4508 pgdat->node_present_pages = realtotalpages;
4509 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4513 #ifndef CONFIG_SPARSEMEM
4515 * Calculate the size of the zone->blockflags rounded to an unsigned long
4516 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4517 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4518 * round what is now in bits to nearest long in bits, then return it in
4521 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4523 unsigned long usemapsize;
4525 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4526 usemapsize = roundup(zonesize, pageblock_nr_pages);
4527 usemapsize = usemapsize >> pageblock_order;
4528 usemapsize *= NR_PAGEBLOCK_BITS;
4529 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4531 return usemapsize / 8;
4534 static void __init setup_usemap(struct pglist_data *pgdat,
4536 unsigned long zone_start_pfn,
4537 unsigned long zonesize)
4539 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4540 zone->pageblock_flags = NULL;
4542 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4546 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4547 unsigned long zone_start_pfn, unsigned long zonesize) {}
4548 #endif /* CONFIG_SPARSEMEM */
4550 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4552 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4553 void __init set_pageblock_order(void)
4557 /* Check that pageblock_nr_pages has not already been setup */
4558 if (pageblock_order)
4561 if (HPAGE_SHIFT > PAGE_SHIFT)
4562 order = HUGETLB_PAGE_ORDER;
4564 order = MAX_ORDER - 1;
4567 * Assume the largest contiguous order of interest is a huge page.
4568 * This value may be variable depending on boot parameters on IA64 and
4571 pageblock_order = order;
4573 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4576 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4577 * is unused as pageblock_order is set at compile-time. See
4578 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4581 void __init set_pageblock_order(void)
4585 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4587 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4588 unsigned long present_pages)
4590 unsigned long pages = spanned_pages;
4593 * Provide a more accurate estimation if there are holes within
4594 * the zone and SPARSEMEM is in use. If there are holes within the
4595 * zone, each populated memory region may cost us one or two extra
4596 * memmap pages due to alignment because memmap pages for each
4597 * populated regions may not naturally algined on page boundary.
4598 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4600 if (spanned_pages > present_pages + (present_pages >> 4) &&
4601 IS_ENABLED(CONFIG_SPARSEMEM))
4602 pages = present_pages;
4604 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4608 * Set up the zone data structures:
4609 * - mark all pages reserved
4610 * - mark all memory queues empty
4611 * - clear the memory bitmaps
4613 * NOTE: pgdat should get zeroed by caller.
4615 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4616 unsigned long *zones_size, unsigned long *zholes_size)
4619 int nid = pgdat->node_id;
4620 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4623 pgdat_resize_init(pgdat);
4624 #ifdef CONFIG_NUMA_BALANCING
4625 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4626 pgdat->numabalancing_migrate_nr_pages = 0;
4627 pgdat->numabalancing_migrate_next_window = jiffies;
4629 init_waitqueue_head(&pgdat->kswapd_wait);
4630 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4631 pgdat_page_cgroup_init(pgdat);
4633 for (j = 0; j < MAX_NR_ZONES; j++) {
4634 struct zone *zone = pgdat->node_zones + j;
4635 unsigned long size, realsize, freesize, memmap_pages;
4637 size = zone_spanned_pages_in_node(nid, j, zones_size);
4638 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4642 * Adjust freesize so that it accounts for how much memory
4643 * is used by this zone for memmap. This affects the watermark
4644 * and per-cpu initialisations
4646 memmap_pages = calc_memmap_size(size, realsize);
4647 if (freesize >= memmap_pages) {
4648 freesize -= memmap_pages;
4651 " %s zone: %lu pages used for memmap\n",
4652 zone_names[j], memmap_pages);
4655 " %s zone: %lu pages exceeds freesize %lu\n",
4656 zone_names[j], memmap_pages, freesize);
4658 /* Account for reserved pages */
4659 if (j == 0 && freesize > dma_reserve) {
4660 freesize -= dma_reserve;
4661 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4662 zone_names[0], dma_reserve);
4665 if (!is_highmem_idx(j))
4666 nr_kernel_pages += freesize;
4667 /* Charge for highmem memmap if there are enough kernel pages */
4668 else if (nr_kernel_pages > memmap_pages * 2)
4669 nr_kernel_pages -= memmap_pages;
4670 nr_all_pages += freesize;
4672 zone->spanned_pages = size;
4673 zone->present_pages = realsize;
4675 * Set an approximate value for lowmem here, it will be adjusted
4676 * when the bootmem allocator frees pages into the buddy system.
4677 * And all highmem pages will be managed by the buddy system.
4679 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4682 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4684 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4686 zone->name = zone_names[j];
4687 spin_lock_init(&zone->lock);
4688 spin_lock_init(&zone->lru_lock);
4689 zone_seqlock_init(zone);
4690 zone->zone_pgdat = pgdat;
4692 zone_pcp_init(zone);
4693 lruvec_init(&zone->lruvec);
4697 set_pageblock_order();
4698 setup_usemap(pgdat, zone, zone_start_pfn, size);
4699 ret = init_currently_empty_zone(zone, zone_start_pfn,
4700 size, MEMMAP_EARLY);
4702 memmap_init(size, nid, j, zone_start_pfn);
4703 zone_start_pfn += size;
4707 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4709 /* Skip empty nodes */
4710 if (!pgdat->node_spanned_pages)
4713 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4714 /* ia64 gets its own node_mem_map, before this, without bootmem */
4715 if (!pgdat->node_mem_map) {
4716 unsigned long size, start, end;
4720 * The zone's endpoints aren't required to be MAX_ORDER
4721 * aligned but the node_mem_map endpoints must be in order
4722 * for the buddy allocator to function correctly.
4724 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4725 end = pgdat_end_pfn(pgdat);
4726 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4727 size = (end - start) * sizeof(struct page);
4728 map = alloc_remap(pgdat->node_id, size);
4730 map = alloc_bootmem_node_nopanic(pgdat, size);
4731 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4733 #ifndef CONFIG_NEED_MULTIPLE_NODES
4735 * With no DISCONTIG, the global mem_map is just set as node 0's
4737 if (pgdat == NODE_DATA(0)) {
4738 mem_map = NODE_DATA(0)->node_mem_map;
4739 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4740 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4741 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4742 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4745 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4748 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4749 unsigned long node_start_pfn, unsigned long *zholes_size)
4751 pg_data_t *pgdat = NODE_DATA(nid);
4753 /* pg_data_t should be reset to zero when it's allocated */
4754 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4756 pgdat->node_id = nid;
4757 pgdat->node_start_pfn = node_start_pfn;
4758 init_zone_allows_reclaim(nid);
4759 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4761 alloc_node_mem_map(pgdat);
4762 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4763 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4764 nid, (unsigned long)pgdat,
4765 (unsigned long)pgdat->node_mem_map);
4768 free_area_init_core(pgdat, zones_size, zholes_size);
4771 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4773 #if MAX_NUMNODES > 1
4775 * Figure out the number of possible node ids.
4777 void __init setup_nr_node_ids(void)
4780 unsigned int highest = 0;
4782 for_each_node_mask(node, node_possible_map)
4784 nr_node_ids = highest + 1;
4789 * node_map_pfn_alignment - determine the maximum internode alignment
4791 * This function should be called after node map is populated and sorted.
4792 * It calculates the maximum power of two alignment which can distinguish
4795 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4796 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4797 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4798 * shifted, 1GiB is enough and this function will indicate so.
4800 * This is used to test whether pfn -> nid mapping of the chosen memory
4801 * model has fine enough granularity to avoid incorrect mapping for the
4802 * populated node map.
4804 * Returns the determined alignment in pfn's. 0 if there is no alignment
4805 * requirement (single node).
4807 unsigned long __init node_map_pfn_alignment(void)
4809 unsigned long accl_mask = 0, last_end = 0;
4810 unsigned long start, end, mask;
4814 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4815 if (!start || last_nid < 0 || last_nid == nid) {
4822 * Start with a mask granular enough to pin-point to the
4823 * start pfn and tick off bits one-by-one until it becomes
4824 * too coarse to separate the current node from the last.
4826 mask = ~((1 << __ffs(start)) - 1);
4827 while (mask && last_end <= (start & (mask << 1)))
4830 /* accumulate all internode masks */
4834 /* convert mask to number of pages */
4835 return ~accl_mask + 1;
4838 /* Find the lowest pfn for a node */
4839 static unsigned long __init find_min_pfn_for_node(int nid)
4841 unsigned long min_pfn = ULONG_MAX;
4842 unsigned long start_pfn;
4845 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4846 min_pfn = min(min_pfn, start_pfn);
4848 if (min_pfn == ULONG_MAX) {
4850 "Could not find start_pfn for node %d\n", nid);
4858 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4860 * It returns the minimum PFN based on information provided via
4861 * add_active_range().
4863 unsigned long __init find_min_pfn_with_active_regions(void)
4865 return find_min_pfn_for_node(MAX_NUMNODES);
4869 * early_calculate_totalpages()
4870 * Sum pages in active regions for movable zone.
4871 * Populate N_MEMORY for calculating usable_nodes.
4873 static unsigned long __init early_calculate_totalpages(void)
4875 unsigned long totalpages = 0;
4876 unsigned long start_pfn, end_pfn;
4879 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4880 unsigned long pages = end_pfn - start_pfn;
4882 totalpages += pages;
4884 node_set_state(nid, N_MEMORY);
4890 * Find the PFN the Movable zone begins in each node. Kernel memory
4891 * is spread evenly between nodes as long as the nodes have enough
4892 * memory. When they don't, some nodes will have more kernelcore than
4895 static void __init find_zone_movable_pfns_for_nodes(void)
4898 unsigned long usable_startpfn;
4899 unsigned long kernelcore_node, kernelcore_remaining;
4900 /* save the state before borrow the nodemask */
4901 nodemask_t saved_node_state = node_states[N_MEMORY];
4902 unsigned long totalpages = early_calculate_totalpages();
4903 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4906 * If movablecore was specified, calculate what size of
4907 * kernelcore that corresponds so that memory usable for
4908 * any allocation type is evenly spread. If both kernelcore
4909 * and movablecore are specified, then the value of kernelcore
4910 * will be used for required_kernelcore if it's greater than
4911 * what movablecore would have allowed.
4913 if (required_movablecore) {
4914 unsigned long corepages;
4917 * Round-up so that ZONE_MOVABLE is at least as large as what
4918 * was requested by the user
4920 required_movablecore =
4921 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4922 corepages = totalpages - required_movablecore;
4924 required_kernelcore = max(required_kernelcore, corepages);
4927 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4928 if (!required_kernelcore)
4931 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4932 find_usable_zone_for_movable();
4933 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4936 /* Spread kernelcore memory as evenly as possible throughout nodes */
4937 kernelcore_node = required_kernelcore / usable_nodes;
4938 for_each_node_state(nid, N_MEMORY) {
4939 unsigned long start_pfn, end_pfn;
4942 * Recalculate kernelcore_node if the division per node
4943 * now exceeds what is necessary to satisfy the requested
4944 * amount of memory for the kernel
4946 if (required_kernelcore < kernelcore_node)
4947 kernelcore_node = required_kernelcore / usable_nodes;
4950 * As the map is walked, we track how much memory is usable
4951 * by the kernel using kernelcore_remaining. When it is
4952 * 0, the rest of the node is usable by ZONE_MOVABLE
4954 kernelcore_remaining = kernelcore_node;
4956 /* Go through each range of PFNs within this node */
4957 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4958 unsigned long size_pages;
4960 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4961 if (start_pfn >= end_pfn)
4964 /* Account for what is only usable for kernelcore */
4965 if (start_pfn < usable_startpfn) {
4966 unsigned long kernel_pages;
4967 kernel_pages = min(end_pfn, usable_startpfn)
4970 kernelcore_remaining -= min(kernel_pages,
4971 kernelcore_remaining);
4972 required_kernelcore -= min(kernel_pages,
4973 required_kernelcore);
4975 /* Continue if range is now fully accounted */
4976 if (end_pfn <= usable_startpfn) {
4979 * Push zone_movable_pfn to the end so
4980 * that if we have to rebalance
4981 * kernelcore across nodes, we will
4982 * not double account here
4984 zone_movable_pfn[nid] = end_pfn;
4987 start_pfn = usable_startpfn;
4991 * The usable PFN range for ZONE_MOVABLE is from
4992 * start_pfn->end_pfn. Calculate size_pages as the
4993 * number of pages used as kernelcore
4995 size_pages = end_pfn - start_pfn;
4996 if (size_pages > kernelcore_remaining)
4997 size_pages = kernelcore_remaining;
4998 zone_movable_pfn[nid] = start_pfn + size_pages;
5001 * Some kernelcore has been met, update counts and
5002 * break if the kernelcore for this node has been
5005 required_kernelcore -= min(required_kernelcore,
5007 kernelcore_remaining -= size_pages;
5008 if (!kernelcore_remaining)
5014 * If there is still required_kernelcore, we do another pass with one
5015 * less node in the count. This will push zone_movable_pfn[nid] further
5016 * along on the nodes that still have memory until kernelcore is
5020 if (usable_nodes && required_kernelcore > usable_nodes)
5023 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5024 for (nid = 0; nid < MAX_NUMNODES; nid++)
5025 zone_movable_pfn[nid] =
5026 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5029 /* restore the node_state */
5030 node_states[N_MEMORY] = saved_node_state;
5033 /* Any regular or high memory on that node ? */
5034 static void check_for_memory(pg_data_t *pgdat, int nid)
5036 enum zone_type zone_type;
5038 if (N_MEMORY == N_NORMAL_MEMORY)
5041 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5042 struct zone *zone = &pgdat->node_zones[zone_type];
5043 if (zone->present_pages) {
5044 node_set_state(nid, N_HIGH_MEMORY);
5045 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5046 zone_type <= ZONE_NORMAL)
5047 node_set_state(nid, N_NORMAL_MEMORY);
5054 * free_area_init_nodes - Initialise all pg_data_t and zone data
5055 * @max_zone_pfn: an array of max PFNs for each zone
5057 * This will call free_area_init_node() for each active node in the system.
5058 * Using the page ranges provided by add_active_range(), the size of each
5059 * zone in each node and their holes is calculated. If the maximum PFN
5060 * between two adjacent zones match, it is assumed that the zone is empty.
5061 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5062 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5063 * starts where the previous one ended. For example, ZONE_DMA32 starts
5064 * at arch_max_dma_pfn.
5066 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5068 unsigned long start_pfn, end_pfn;
5071 /* Record where the zone boundaries are */
5072 memset(arch_zone_lowest_possible_pfn, 0,
5073 sizeof(arch_zone_lowest_possible_pfn));
5074 memset(arch_zone_highest_possible_pfn, 0,
5075 sizeof(arch_zone_highest_possible_pfn));
5076 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5077 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5078 for (i = 1; i < MAX_NR_ZONES; i++) {
5079 if (i == ZONE_MOVABLE)
5081 arch_zone_lowest_possible_pfn[i] =
5082 arch_zone_highest_possible_pfn[i-1];
5083 arch_zone_highest_possible_pfn[i] =
5084 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5086 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5087 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5089 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5090 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5091 find_zone_movable_pfns_for_nodes();
5093 /* Print out the zone ranges */
5094 printk("Zone ranges:\n");
5095 for (i = 0; i < MAX_NR_ZONES; i++) {
5096 if (i == ZONE_MOVABLE)
5098 printk(KERN_CONT " %-8s ", zone_names[i]);
5099 if (arch_zone_lowest_possible_pfn[i] ==
5100 arch_zone_highest_possible_pfn[i])
5101 printk(KERN_CONT "empty\n");
5103 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5104 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5105 (arch_zone_highest_possible_pfn[i]
5106 << PAGE_SHIFT) - 1);
5109 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5110 printk("Movable zone start for each node\n");
5111 for (i = 0; i < MAX_NUMNODES; i++) {
5112 if (zone_movable_pfn[i])
5113 printk(" Node %d: %#010lx\n", i,
5114 zone_movable_pfn[i] << PAGE_SHIFT);
5117 /* Print out the early node map */
5118 printk("Early memory node ranges\n");
5119 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5120 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5121 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5123 /* Initialise every node */
5124 mminit_verify_pageflags_layout();
5125 setup_nr_node_ids();
5126 for_each_online_node(nid) {
5127 pg_data_t *pgdat = NODE_DATA(nid);
5128 free_area_init_node(nid, NULL,
5129 find_min_pfn_for_node(nid), NULL);
5131 /* Any memory on that node */
5132 if (pgdat->node_present_pages)
5133 node_set_state(nid, N_MEMORY);
5134 check_for_memory(pgdat, nid);
5138 static int __init cmdline_parse_core(char *p, unsigned long *core)
5140 unsigned long long coremem;
5144 coremem = memparse(p, &p);
5145 *core = coremem >> PAGE_SHIFT;
5147 /* Paranoid check that UL is enough for the coremem value */
5148 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5154 * kernelcore=size sets the amount of memory for use for allocations that
5155 * cannot be reclaimed or migrated.
5157 static int __init cmdline_parse_kernelcore(char *p)
5159 return cmdline_parse_core(p, &required_kernelcore);
5163 * movablecore=size sets the amount of memory for use for allocations that
5164 * can be reclaimed or migrated.
5166 static int __init cmdline_parse_movablecore(char *p)
5168 return cmdline_parse_core(p, &required_movablecore);
5171 early_param("kernelcore", cmdline_parse_kernelcore);
5172 early_param("movablecore", cmdline_parse_movablecore);
5174 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5176 unsigned long free_reserved_area(unsigned long start, unsigned long end,
5177 int poison, char *s)
5179 unsigned long pages, pos;
5181 pos = start = PAGE_ALIGN(start);
5183 for (pages = 0; pos < end; pos += PAGE_SIZE, pages++) {
5185 memset((void *)pos, poison, PAGE_SIZE);
5186 free_reserved_page(virt_to_page((void *)pos));
5190 pr_info("Freeing %s memory: %ldK (%lx - %lx)\n",
5191 s, pages << (PAGE_SHIFT - 10), start, end);
5196 #ifdef CONFIG_HIGHMEM
5197 void free_highmem_page(struct page *page)
5199 __free_reserved_page(page);
5206 * set_dma_reserve - set the specified number of pages reserved in the first zone
5207 * @new_dma_reserve: The number of pages to mark reserved
5209 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5210 * In the DMA zone, a significant percentage may be consumed by kernel image
5211 * and other unfreeable allocations which can skew the watermarks badly. This
5212 * function may optionally be used to account for unfreeable pages in the
5213 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5214 * smaller per-cpu batchsize.
5216 void __init set_dma_reserve(unsigned long new_dma_reserve)
5218 dma_reserve = new_dma_reserve;
5221 void __init free_area_init(unsigned long *zones_size)
5223 free_area_init_node(0, zones_size,
5224 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5227 static int page_alloc_cpu_notify(struct notifier_block *self,
5228 unsigned long action, void *hcpu)
5230 int cpu = (unsigned long)hcpu;
5232 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5233 lru_add_drain_cpu(cpu);
5237 * Spill the event counters of the dead processor
5238 * into the current processors event counters.
5239 * This artificially elevates the count of the current
5242 vm_events_fold_cpu(cpu);
5245 * Zero the differential counters of the dead processor
5246 * so that the vm statistics are consistent.
5248 * This is only okay since the processor is dead and cannot
5249 * race with what we are doing.
5251 refresh_cpu_vm_stats(cpu);
5256 void __init page_alloc_init(void)
5258 hotcpu_notifier(page_alloc_cpu_notify, 0);
5262 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5263 * or min_free_kbytes changes.
5265 static void calculate_totalreserve_pages(void)
5267 struct pglist_data *pgdat;
5268 unsigned long reserve_pages = 0;
5269 enum zone_type i, j;
5271 for_each_online_pgdat(pgdat) {
5272 for (i = 0; i < MAX_NR_ZONES; i++) {
5273 struct zone *zone = pgdat->node_zones + i;
5274 unsigned long max = 0;
5276 /* Find valid and maximum lowmem_reserve in the zone */
5277 for (j = i; j < MAX_NR_ZONES; j++) {
5278 if (zone->lowmem_reserve[j] > max)
5279 max = zone->lowmem_reserve[j];
5282 /* we treat the high watermark as reserved pages. */
5283 max += high_wmark_pages(zone);
5285 if (max > zone->managed_pages)
5286 max = zone->managed_pages;
5287 reserve_pages += max;
5289 * Lowmem reserves are not available to
5290 * GFP_HIGHUSER page cache allocations and
5291 * kswapd tries to balance zones to their high
5292 * watermark. As a result, neither should be
5293 * regarded as dirtyable memory, to prevent a
5294 * situation where reclaim has to clean pages
5295 * in order to balance the zones.
5297 zone->dirty_balance_reserve = max;
5300 dirty_balance_reserve = reserve_pages;
5301 totalreserve_pages = reserve_pages;
5305 * setup_per_zone_lowmem_reserve - called whenever
5306 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5307 * has a correct pages reserved value, so an adequate number of
5308 * pages are left in the zone after a successful __alloc_pages().
5310 static void setup_per_zone_lowmem_reserve(void)
5312 struct pglist_data *pgdat;
5313 enum zone_type j, idx;
5315 for_each_online_pgdat(pgdat) {
5316 for (j = 0; j < MAX_NR_ZONES; j++) {
5317 struct zone *zone = pgdat->node_zones + j;
5318 unsigned long managed_pages = zone->managed_pages;
5320 zone->lowmem_reserve[j] = 0;
5324 struct zone *lower_zone;
5328 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5329 sysctl_lowmem_reserve_ratio[idx] = 1;
5331 lower_zone = pgdat->node_zones + idx;
5332 lower_zone->lowmem_reserve[j] = managed_pages /
5333 sysctl_lowmem_reserve_ratio[idx];
5334 managed_pages += lower_zone->managed_pages;
5339 /* update totalreserve_pages */
5340 calculate_totalreserve_pages();
5343 static void __setup_per_zone_wmarks(void)
5345 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5346 unsigned long pages_low = extra_free_kbytes >> (PAGE_SHIFT - 10);
5347 unsigned long lowmem_pages = 0;
5349 unsigned long flags;
5351 /* Calculate total number of !ZONE_HIGHMEM pages */
5352 for_each_zone(zone) {
5353 if (!is_highmem(zone))
5354 lowmem_pages += zone->managed_pages;
5357 for_each_zone(zone) {
5360 spin_lock_irqsave(&zone->lock, flags);
5361 min = (u64)pages_min * zone->managed_pages;
5362 do_div(min, lowmem_pages);
5363 low = (u64)pages_low * zone->managed_pages;
5364 do_div(low, vm_total_pages);
5366 if (is_highmem(zone)) {
5368 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5369 * need highmem pages, so cap pages_min to a small
5372 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5373 * deltas controls asynch page reclaim, and so should
5374 * not be capped for highmem.
5376 unsigned long min_pages;
5378 min_pages = zone->managed_pages / 1024;
5379 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5380 zone->watermark[WMARK_MIN] = min_pages;
5383 * If it's a lowmem zone, reserve a number of pages
5384 * proportionate to the zone's size.
5386 zone->watermark[WMARK_MIN] = min;
5389 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) +
5391 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) +
5394 setup_zone_migrate_reserve(zone);
5395 spin_unlock_irqrestore(&zone->lock, flags);
5398 /* update totalreserve_pages */
5399 calculate_totalreserve_pages();
5403 * setup_per_zone_wmarks - called when min_free_kbytes changes
5404 * or when memory is hot-{added|removed}
5406 * Ensures that the watermark[min,low,high] values for each zone are set
5407 * correctly with respect to min_free_kbytes.
5409 void setup_per_zone_wmarks(void)
5411 mutex_lock(&zonelists_mutex);
5412 __setup_per_zone_wmarks();
5413 mutex_unlock(&zonelists_mutex);
5417 * The inactive anon list should be small enough that the VM never has to
5418 * do too much work, but large enough that each inactive page has a chance
5419 * to be referenced again before it is swapped out.
5421 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5422 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5423 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5424 * the anonymous pages are kept on the inactive list.
5427 * memory ratio inactive anon
5428 * -------------------------------------
5437 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5439 unsigned int gb, ratio;
5441 /* Zone size in gigabytes */
5442 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5444 ratio = int_sqrt(10 * gb);
5448 zone->inactive_ratio = ratio;
5451 static void __meminit setup_per_zone_inactive_ratio(void)
5456 calculate_zone_inactive_ratio(zone);
5460 * Initialise min_free_kbytes.
5462 * For small machines we want it small (128k min). For large machines
5463 * we want it large (64MB max). But it is not linear, because network
5464 * bandwidth does not increase linearly with machine size. We use
5466 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5467 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5483 int __meminit init_per_zone_wmark_min(void)
5485 unsigned long lowmem_kbytes;
5487 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5489 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5490 if (min_free_kbytes < 128)
5491 min_free_kbytes = 128;
5492 if (min_free_kbytes > 65536)
5493 min_free_kbytes = 65536;
5494 setup_per_zone_wmarks();
5495 refresh_zone_stat_thresholds();
5496 setup_per_zone_lowmem_reserve();
5497 setup_per_zone_inactive_ratio();
5500 module_init(init_per_zone_wmark_min)
5503 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5504 * that we can call two helper functions whenever min_free_kbytes
5505 * or extra_free_kbytes changes.
5507 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5508 void __user *buffer, size_t *length, loff_t *ppos)
5510 proc_dointvec(table, write, buffer, length, ppos);
5512 setup_per_zone_wmarks();
5517 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5518 void __user *buffer, size_t *length, loff_t *ppos)
5523 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5528 zone->min_unmapped_pages = (zone->managed_pages *
5529 sysctl_min_unmapped_ratio) / 100;
5533 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5534 void __user *buffer, size_t *length, loff_t *ppos)
5539 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5544 zone->min_slab_pages = (zone->managed_pages *
5545 sysctl_min_slab_ratio) / 100;
5551 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5552 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5553 * whenever sysctl_lowmem_reserve_ratio changes.
5555 * The reserve ratio obviously has absolutely no relation with the
5556 * minimum watermarks. The lowmem reserve ratio can only make sense
5557 * if in function of the boot time zone sizes.
5559 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5560 void __user *buffer, size_t *length, loff_t *ppos)
5562 proc_dointvec_minmax(table, write, buffer, length, ppos);
5563 setup_per_zone_lowmem_reserve();
5568 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5569 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5570 * can have before it gets flushed back to buddy allocator.
5573 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5574 void __user *buffer, size_t *length, loff_t *ppos)
5580 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5581 if (!write || (ret < 0))
5583 for_each_populated_zone(zone) {
5584 for_each_possible_cpu(cpu) {
5586 high = zone->managed_pages / percpu_pagelist_fraction;
5587 setup_pagelist_highmark(
5588 per_cpu_ptr(zone->pageset, cpu), high);
5594 int hashdist = HASHDIST_DEFAULT;
5597 static int __init set_hashdist(char *str)
5601 hashdist = simple_strtoul(str, &str, 0);
5604 __setup("hashdist=", set_hashdist);
5608 * allocate a large system hash table from bootmem
5609 * - it is assumed that the hash table must contain an exact power-of-2
5610 * quantity of entries
5611 * - limit is the number of hash buckets, not the total allocation size
5613 void *__init alloc_large_system_hash(const char *tablename,
5614 unsigned long bucketsize,
5615 unsigned long numentries,
5618 unsigned int *_hash_shift,
5619 unsigned int *_hash_mask,
5620 unsigned long low_limit,
5621 unsigned long high_limit)
5623 unsigned long long max = high_limit;
5624 unsigned long log2qty, size;
5627 /* allow the kernel cmdline to have a say */
5629 /* round applicable memory size up to nearest megabyte */
5630 numentries = nr_kernel_pages;
5631 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5632 numentries >>= 20 - PAGE_SHIFT;
5633 numentries <<= 20 - PAGE_SHIFT;
5635 /* limit to 1 bucket per 2^scale bytes of low memory */
5636 if (scale > PAGE_SHIFT)
5637 numentries >>= (scale - PAGE_SHIFT);
5639 numentries <<= (PAGE_SHIFT - scale);
5641 /* Make sure we've got at least a 0-order allocation.. */
5642 if (unlikely(flags & HASH_SMALL)) {
5643 /* Makes no sense without HASH_EARLY */
5644 WARN_ON(!(flags & HASH_EARLY));
5645 if (!(numentries >> *_hash_shift)) {
5646 numentries = 1UL << *_hash_shift;
5647 BUG_ON(!numentries);
5649 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5650 numentries = PAGE_SIZE / bucketsize;
5652 numentries = roundup_pow_of_two(numentries);
5654 /* limit allocation size to 1/16 total memory by default */
5656 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5657 do_div(max, bucketsize);
5659 max = min(max, 0x80000000ULL);
5661 if (numentries < low_limit)
5662 numentries = low_limit;
5663 if (numentries > max)
5666 log2qty = ilog2(numentries);
5669 size = bucketsize << log2qty;
5670 if (flags & HASH_EARLY)
5671 table = alloc_bootmem_nopanic(size);
5673 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5676 * If bucketsize is not a power-of-two, we may free
5677 * some pages at the end of hash table which
5678 * alloc_pages_exact() automatically does
5680 if (get_order(size) < MAX_ORDER) {
5681 table = alloc_pages_exact(size, GFP_ATOMIC);
5682 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5685 } while (!table && size > PAGE_SIZE && --log2qty);
5688 panic("Failed to allocate %s hash table\n", tablename);
5690 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5693 ilog2(size) - PAGE_SHIFT,
5697 *_hash_shift = log2qty;
5699 *_hash_mask = (1 << log2qty) - 1;
5704 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5705 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5708 #ifdef CONFIG_SPARSEMEM
5709 return __pfn_to_section(pfn)->pageblock_flags;
5711 return zone->pageblock_flags;
5712 #endif /* CONFIG_SPARSEMEM */
5715 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5717 #ifdef CONFIG_SPARSEMEM
5718 pfn &= (PAGES_PER_SECTION-1);
5719 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5721 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5722 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5723 #endif /* CONFIG_SPARSEMEM */
5727 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5728 * @page: The page within the block of interest
5729 * @start_bitidx: The first bit of interest to retrieve
5730 * @end_bitidx: The last bit of interest
5731 * returns pageblock_bits flags
5733 unsigned long get_pageblock_flags_group(struct page *page,
5734 int start_bitidx, int end_bitidx)
5737 unsigned long *bitmap;
5738 unsigned long pfn, bitidx;
5739 unsigned long flags = 0;
5740 unsigned long value = 1;
5742 zone = page_zone(page);
5743 pfn = page_to_pfn(page);
5744 bitmap = get_pageblock_bitmap(zone, pfn);
5745 bitidx = pfn_to_bitidx(zone, pfn);
5747 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5748 if (test_bit(bitidx + start_bitidx, bitmap))
5755 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5756 * @page: The page within the block of interest
5757 * @start_bitidx: The first bit of interest
5758 * @end_bitidx: The last bit of interest
5759 * @flags: The flags to set
5761 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5762 int start_bitidx, int end_bitidx)
5765 unsigned long *bitmap;
5766 unsigned long pfn, bitidx;
5767 unsigned long value = 1;
5769 zone = page_zone(page);
5770 pfn = page_to_pfn(page);
5771 bitmap = get_pageblock_bitmap(zone, pfn);
5772 bitidx = pfn_to_bitidx(zone, pfn);
5773 VM_BUG_ON(!zone_spans_pfn(zone, pfn));
5775 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5777 __set_bit(bitidx + start_bitidx, bitmap);
5779 __clear_bit(bitidx + start_bitidx, bitmap);
5783 * This function checks whether pageblock includes unmovable pages or not.
5784 * If @count is not zero, it is okay to include less @count unmovable pages
5786 * PageLRU check wihtout isolation or lru_lock could race so that
5787 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5788 * expect this function should be exact.
5790 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5791 bool skip_hwpoisoned_pages)
5793 unsigned long pfn, iter, found;
5797 * For avoiding noise data, lru_add_drain_all() should be called
5798 * If ZONE_MOVABLE, the zone never contains unmovable pages
5800 if (zone_idx(zone) == ZONE_MOVABLE)
5802 mt = get_pageblock_migratetype(page);
5803 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5806 pfn = page_to_pfn(page);
5807 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5808 unsigned long check = pfn + iter;
5810 if (!pfn_valid_within(check))
5813 page = pfn_to_page(check);
5815 * We can't use page_count without pin a page
5816 * because another CPU can free compound page.
5817 * This check already skips compound tails of THP
5818 * because their page->_count is zero at all time.
5820 if (!atomic_read(&page->_count)) {
5821 if (PageBuddy(page))
5822 iter += (1 << page_order(page)) - 1;
5827 * The HWPoisoned page may be not in buddy system, and
5828 * page_count() is not 0.
5830 if (skip_hwpoisoned_pages && PageHWPoison(page))
5836 * If there are RECLAIMABLE pages, we need to check it.
5837 * But now, memory offline itself doesn't call shrink_slab()
5838 * and it still to be fixed.
5841 * If the page is not RAM, page_count()should be 0.
5842 * we don't need more check. This is an _used_ not-movable page.
5844 * The problematic thing here is PG_reserved pages. PG_reserved
5845 * is set to both of a memory hole page and a _used_ kernel
5854 bool is_pageblock_removable_nolock(struct page *page)
5860 * We have to be careful here because we are iterating over memory
5861 * sections which are not zone aware so we might end up outside of
5862 * the zone but still within the section.
5863 * We have to take care about the node as well. If the node is offline
5864 * its NODE_DATA will be NULL - see page_zone.
5866 if (!node_online(page_to_nid(page)))
5869 zone = page_zone(page);
5870 pfn = page_to_pfn(page);
5871 if (!zone_spans_pfn(zone, pfn))
5874 return !has_unmovable_pages(zone, page, 0, true);
5879 static unsigned long pfn_max_align_down(unsigned long pfn)
5881 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5882 pageblock_nr_pages) - 1);
5885 static unsigned long pfn_max_align_up(unsigned long pfn)
5887 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5888 pageblock_nr_pages));
5891 /* [start, end) must belong to a single zone. */
5892 static int __alloc_contig_migrate_range(struct compact_control *cc,
5893 unsigned long start, unsigned long end)
5895 /* This function is based on compact_zone() from compaction.c. */
5896 unsigned long nr_reclaimed;
5897 unsigned long pfn = start;
5898 unsigned int tries = 0;
5903 while (pfn < end || !list_empty(&cc->migratepages)) {
5904 if (fatal_signal_pending(current)) {
5909 if (list_empty(&cc->migratepages)) {
5910 cc->nr_migratepages = 0;
5911 pfn = isolate_migratepages_range(cc->zone, cc,
5918 } else if (++tries == 5) {
5919 ret = ret < 0 ? ret : -EBUSY;
5923 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5925 cc->nr_migratepages -= nr_reclaimed;
5927 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
5928 0, MIGRATE_SYNC, MR_CMA);
5931 putback_movable_pages(&cc->migratepages);
5938 * alloc_contig_range() -- tries to allocate given range of pages
5939 * @start: start PFN to allocate
5940 * @end: one-past-the-last PFN to allocate
5941 * @migratetype: migratetype of the underlaying pageblocks (either
5942 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5943 * in range must have the same migratetype and it must
5944 * be either of the two.
5946 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5947 * aligned, however it's the caller's responsibility to guarantee that
5948 * we are the only thread that changes migrate type of pageblocks the
5951 * The PFN range must belong to a single zone.
5953 * Returns zero on success or negative error code. On success all
5954 * pages which PFN is in [start, end) are allocated for the caller and
5955 * need to be freed with free_contig_range().
5957 int alloc_contig_range(unsigned long start, unsigned long end,
5958 unsigned migratetype)
5960 unsigned long outer_start, outer_end;
5963 struct compact_control cc = {
5964 .nr_migratepages = 0,
5966 .zone = page_zone(pfn_to_page(start)),
5968 .ignore_skip_hint = true,
5970 INIT_LIST_HEAD(&cc.migratepages);
5973 * What we do here is we mark all pageblocks in range as
5974 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5975 * have different sizes, and due to the way page allocator
5976 * work, we align the range to biggest of the two pages so
5977 * that page allocator won't try to merge buddies from
5978 * different pageblocks and change MIGRATE_ISOLATE to some
5979 * other migration type.
5981 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5982 * migrate the pages from an unaligned range (ie. pages that
5983 * we are interested in). This will put all the pages in
5984 * range back to page allocator as MIGRATE_ISOLATE.
5986 * When this is done, we take the pages in range from page
5987 * allocator removing them from the buddy system. This way
5988 * page allocator will never consider using them.
5990 * This lets us mark the pageblocks back as
5991 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5992 * aligned range but not in the unaligned, original range are
5993 * put back to page allocator so that buddy can use them.
5996 ret = start_isolate_page_range(pfn_max_align_down(start),
5997 pfn_max_align_up(end), migratetype,
6002 ret = __alloc_contig_migrate_range(&cc, start, end);
6007 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6008 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6009 * more, all pages in [start, end) are free in page allocator.
6010 * What we are going to do is to allocate all pages from
6011 * [start, end) (that is remove them from page allocator).
6013 * The only problem is that pages at the beginning and at the
6014 * end of interesting range may be not aligned with pages that
6015 * page allocator holds, ie. they can be part of higher order
6016 * pages. Because of this, we reserve the bigger range and
6017 * once this is done free the pages we are not interested in.
6019 * We don't have to hold zone->lock here because the pages are
6020 * isolated thus they won't get removed from buddy.
6023 lru_add_drain_all();
6027 outer_start = start;
6028 while (!PageBuddy(pfn_to_page(outer_start))) {
6029 if (++order >= MAX_ORDER) {
6033 outer_start &= ~0UL << order;
6036 /* Make sure the range is really isolated. */
6037 if (test_pages_isolated(outer_start, end, false)) {
6038 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6045 /* Grab isolated pages from freelists. */
6046 outer_end = isolate_freepages_range(&cc, outer_start, end);
6052 /* Free head and tail (if any) */
6053 if (start != outer_start)
6054 free_contig_range(outer_start, start - outer_start);
6055 if (end != outer_end)
6056 free_contig_range(end, outer_end - end);
6059 undo_isolate_page_range(pfn_max_align_down(start),
6060 pfn_max_align_up(end), migratetype);
6064 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6066 unsigned int count = 0;
6068 for (; nr_pages--; pfn++) {
6069 struct page *page = pfn_to_page(pfn);
6071 count += page_count(page) != 1;
6074 WARN(count != 0, "%d pages are still in use!\n", count);
6078 #ifdef CONFIG_MEMORY_HOTPLUG
6079 static int __meminit __zone_pcp_update(void *data)
6081 struct zone *zone = data;
6083 unsigned long batch = zone_batchsize(zone), flags;
6085 for_each_possible_cpu(cpu) {
6086 struct per_cpu_pageset *pset;
6087 struct per_cpu_pages *pcp;
6089 pset = per_cpu_ptr(zone->pageset, cpu);
6092 local_irq_save(flags);
6094 free_pcppages_bulk(zone, pcp->count, pcp);
6095 drain_zonestat(zone, pset);
6096 setup_pageset(pset, batch);
6097 local_irq_restore(flags);
6102 void __meminit zone_pcp_update(struct zone *zone)
6104 stop_machine(__zone_pcp_update, zone, NULL);
6108 void zone_pcp_reset(struct zone *zone)
6110 unsigned long flags;
6112 struct per_cpu_pageset *pset;
6114 /* avoid races with drain_pages() */
6115 local_irq_save(flags);
6116 if (zone->pageset != &boot_pageset) {
6117 for_each_online_cpu(cpu) {
6118 pset = per_cpu_ptr(zone->pageset, cpu);
6119 drain_zonestat(zone, pset);
6121 free_percpu(zone->pageset);
6122 zone->pageset = &boot_pageset;
6124 local_irq_restore(flags);
6127 #ifdef CONFIG_MEMORY_HOTREMOVE
6129 * All pages in the range must be isolated before calling this.
6132 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6138 unsigned long flags;
6139 /* find the first valid pfn */
6140 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6145 zone = page_zone(pfn_to_page(pfn));
6146 spin_lock_irqsave(&zone->lock, flags);
6148 while (pfn < end_pfn) {
6149 if (!pfn_valid(pfn)) {
6153 page = pfn_to_page(pfn);
6155 * The HWPoisoned page may be not in buddy system, and
6156 * page_count() is not 0.
6158 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6160 SetPageReserved(page);
6164 BUG_ON(page_count(page));
6165 BUG_ON(!PageBuddy(page));
6166 order = page_order(page);
6167 #ifdef CONFIG_DEBUG_VM
6168 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6169 pfn, 1 << order, end_pfn);
6171 list_del(&page->lru);
6172 rmv_page_order(page);
6173 zone->free_area[order].nr_free--;
6174 #ifdef CONFIG_HIGHMEM
6175 if (PageHighMem(page))
6176 totalhigh_pages -= 1 << order;
6178 for (i = 0; i < (1 << order); i++)
6179 SetPageReserved((page+i));
6180 pfn += (1 << order);
6182 spin_unlock_irqrestore(&zone->lock, flags);
6186 #ifdef CONFIG_MEMORY_FAILURE
6187 bool is_free_buddy_page(struct page *page)
6189 struct zone *zone = page_zone(page);
6190 unsigned long pfn = page_to_pfn(page);
6191 unsigned long flags;
6194 spin_lock_irqsave(&zone->lock, flags);
6195 for (order = 0; order < MAX_ORDER; order++) {
6196 struct page *page_head = page - (pfn & ((1 << order) - 1));
6198 if (PageBuddy(page_head) && page_order(page_head) >= order)
6201 spin_unlock_irqrestore(&zone->lock, flags);
6203 return order < MAX_ORDER;
6207 static const struct trace_print_flags pageflag_names[] = {
6208 {1UL << PG_locked, "locked" },
6209 {1UL << PG_error, "error" },
6210 {1UL << PG_referenced, "referenced" },
6211 {1UL << PG_uptodate, "uptodate" },
6212 {1UL << PG_dirty, "dirty" },
6213 {1UL << PG_lru, "lru" },
6214 {1UL << PG_active, "active" },
6215 {1UL << PG_slab, "slab" },
6216 {1UL << PG_owner_priv_1, "owner_priv_1" },
6217 {1UL << PG_arch_1, "arch_1" },
6218 {1UL << PG_reserved, "reserved" },
6219 {1UL << PG_private, "private" },
6220 {1UL << PG_private_2, "private_2" },
6221 {1UL << PG_writeback, "writeback" },
6222 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6223 {1UL << PG_head, "head" },
6224 {1UL << PG_tail, "tail" },
6226 {1UL << PG_compound, "compound" },
6228 {1UL << PG_swapcache, "swapcache" },
6229 {1UL << PG_mappedtodisk, "mappedtodisk" },
6230 {1UL << PG_reclaim, "reclaim" },
6231 {1UL << PG_swapbacked, "swapbacked" },
6232 {1UL << PG_unevictable, "unevictable" },
6234 {1UL << PG_mlocked, "mlocked" },
6236 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6237 {1UL << PG_uncached, "uncached" },
6239 #ifdef CONFIG_MEMORY_FAILURE
6240 {1UL << PG_hwpoison, "hwpoison" },
6242 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6243 {1UL << PG_compound_lock, "compound_lock" },
6247 static void dump_page_flags(unsigned long flags)
6249 const char *delim = "";
6253 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6255 printk(KERN_ALERT "page flags: %#lx(", flags);
6257 /* remove zone id */
6258 flags &= (1UL << NR_PAGEFLAGS) - 1;
6260 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6262 mask = pageflag_names[i].mask;
6263 if ((flags & mask) != mask)
6267 printk("%s%s", delim, pageflag_names[i].name);
6271 /* check for left over flags */
6273 printk("%s%#lx", delim, flags);
6278 void dump_page(struct page *page)
6281 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6282 page, atomic_read(&page->_count), page_mapcount(page),
6283 page->mapping, page->index);
6284 dump_page_flags(page->flags);
6285 mem_cgroup_print_bad_page(page);