2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
69 DEFINE_PER_CPU(int, numa_node);
70 EXPORT_PER_CPU_SYMBOL(numa_node);
73 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
75 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
76 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
77 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
78 * defined in <linux/topology.h>.
80 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
81 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
85 * Array of node states.
87 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
88 [N_POSSIBLE] = NODE_MASK_ALL,
89 [N_ONLINE] = { { [0] = 1UL } },
91 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
93 [N_HIGH_MEMORY] = { { [0] = 1UL } },
95 #ifdef CONFIG_MOVABLE_NODE
96 [N_MEMORY] = { { [0] = 1UL } },
98 [N_CPU] = { { [0] = 1UL } },
101 EXPORT_SYMBOL(node_states);
103 unsigned long totalram_pages __read_mostly;
104 unsigned long totalreserve_pages __read_mostly;
106 * When calculating the number of globally allowed dirty pages, there
107 * is a certain number of per-zone reserves that should not be
108 * considered dirtyable memory. This is the sum of those reserves
109 * over all existing zones that contribute dirtyable memory.
111 unsigned long dirty_balance_reserve __read_mostly;
113 int percpu_pagelist_fraction;
114 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
116 #ifdef CONFIG_PM_SLEEP
118 * The following functions are used by the suspend/hibernate code to temporarily
119 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
120 * while devices are suspended. To avoid races with the suspend/hibernate code,
121 * they should always be called with pm_mutex held (gfp_allowed_mask also should
122 * only be modified with pm_mutex held, unless the suspend/hibernate code is
123 * guaranteed not to run in parallel with that modification).
126 static gfp_t saved_gfp_mask;
128 void pm_restore_gfp_mask(void)
130 WARN_ON(!mutex_is_locked(&pm_mutex));
131 if (saved_gfp_mask) {
132 gfp_allowed_mask = saved_gfp_mask;
137 void pm_restrict_gfp_mask(void)
139 WARN_ON(!mutex_is_locked(&pm_mutex));
140 WARN_ON(saved_gfp_mask);
141 saved_gfp_mask = gfp_allowed_mask;
142 gfp_allowed_mask &= ~GFP_IOFS;
145 bool pm_suspended_storage(void)
147 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
151 #endif /* CONFIG_PM_SLEEP */
153 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
154 int pageblock_order __read_mostly;
157 static void __free_pages_ok(struct page *page, unsigned int order);
160 * results with 256, 32 in the lowmem_reserve sysctl:
161 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
162 * 1G machine -> (16M dma, 784M normal, 224M high)
163 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
164 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
165 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
167 * TBD: should special case ZONE_DMA32 machines here - in those we normally
168 * don't need any ZONE_NORMAL reservation
170 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
171 #ifdef CONFIG_ZONE_DMA
174 #ifdef CONFIG_ZONE_DMA32
177 #ifdef CONFIG_HIGHMEM
183 EXPORT_SYMBOL(totalram_pages);
185 static char * const zone_names[MAX_NR_ZONES] = {
186 #ifdef CONFIG_ZONE_DMA
189 #ifdef CONFIG_ZONE_DMA32
193 #ifdef CONFIG_HIGHMEM
199 int min_free_kbytes = 1024;
200 int min_free_order_shift = 1;
202 static unsigned long __meminitdata nr_kernel_pages;
203 static unsigned long __meminitdata nr_all_pages;
204 static unsigned long __meminitdata dma_reserve;
206 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
207 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
208 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
209 static unsigned long __initdata required_kernelcore;
210 static unsigned long __initdata required_movablecore;
211 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
213 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
215 EXPORT_SYMBOL(movable_zone);
216 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
219 int nr_node_ids __read_mostly = MAX_NUMNODES;
220 int nr_online_nodes __read_mostly = 1;
221 EXPORT_SYMBOL(nr_node_ids);
222 EXPORT_SYMBOL(nr_online_nodes);
225 int page_group_by_mobility_disabled __read_mostly;
227 void set_pageblock_migratetype(struct page *page, int migratetype)
230 if (unlikely(page_group_by_mobility_disabled))
231 migratetype = MIGRATE_UNMOVABLE;
233 set_pageblock_flags_group(page, (unsigned long)migratetype,
234 PB_migrate, PB_migrate_end);
237 bool oom_killer_disabled __read_mostly;
239 #ifdef CONFIG_DEBUG_VM
240 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
244 unsigned long pfn = page_to_pfn(page);
245 unsigned long sp, start_pfn;
248 seq = zone_span_seqbegin(zone);
249 start_pfn = zone->zone_start_pfn;
250 sp = zone->spanned_pages;
251 if (!zone_spans_pfn(zone, pfn))
253 } while (zone_span_seqretry(zone, seq));
256 pr_err("page %lu outside zone [ %lu - %lu ]\n",
257 pfn, start_pfn, start_pfn + sp);
262 static int page_is_consistent(struct zone *zone, struct page *page)
264 if (!pfn_valid_within(page_to_pfn(page)))
266 if (zone != page_zone(page))
272 * Temporary debugging check for pages not lying within a given zone.
274 static int bad_range(struct zone *zone, struct page *page)
276 if (page_outside_zone_boundaries(zone, page))
278 if (!page_is_consistent(zone, page))
284 static inline int bad_range(struct zone *zone, struct page *page)
290 static void bad_page(struct page *page)
292 static unsigned long resume;
293 static unsigned long nr_shown;
294 static unsigned long nr_unshown;
296 /* Don't complain about poisoned pages */
297 if (PageHWPoison(page)) {
298 page_mapcount_reset(page); /* remove PageBuddy */
303 * Allow a burst of 60 reports, then keep quiet for that minute;
304 * or allow a steady drip of one report per second.
306 if (nr_shown == 60) {
307 if (time_before(jiffies, resume)) {
313 "BUG: Bad page state: %lu messages suppressed\n",
320 resume = jiffies + 60 * HZ;
322 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
323 current->comm, page_to_pfn(page));
329 /* Leave bad fields for debug, except PageBuddy could make trouble */
330 page_mapcount_reset(page); /* remove PageBuddy */
331 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
335 * Higher-order pages are called "compound pages". They are structured thusly:
337 * The first PAGE_SIZE page is called the "head page".
339 * The remaining PAGE_SIZE pages are called "tail pages".
341 * All pages have PG_compound set. All tail pages have their ->first_page
342 * pointing at the head page.
344 * The first tail page's ->lru.next holds the address of the compound page's
345 * put_page() function. Its ->lru.prev holds the order of allocation.
346 * This usage means that zero-order pages may not be compound.
349 static void free_compound_page(struct page *page)
351 __free_pages_ok(page, compound_order(page));
354 void prep_compound_page(struct page *page, unsigned long order)
357 int nr_pages = 1 << order;
359 set_compound_page_dtor(page, free_compound_page);
360 set_compound_order(page, order);
362 for (i = 1; i < nr_pages; i++) {
363 struct page *p = page + i;
365 set_page_count(p, 0);
366 p->first_page = page;
370 /* update __split_huge_page_refcount if you change this function */
371 static int destroy_compound_page(struct page *page, unsigned long order)
374 int nr_pages = 1 << order;
377 if (unlikely(compound_order(page) != order)) {
382 __ClearPageHead(page);
384 for (i = 1; i < nr_pages; i++) {
385 struct page *p = page + i;
387 if (unlikely(!PageTail(p) || (p->first_page != page))) {
397 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
402 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
403 * and __GFP_HIGHMEM from hard or soft interrupt context.
405 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
406 for (i = 0; i < (1 << order); i++)
407 clear_highpage(page + i);
410 #ifdef CONFIG_DEBUG_PAGEALLOC
411 unsigned int _debug_guardpage_minorder;
413 static int __init debug_guardpage_minorder_setup(char *buf)
417 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
418 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
421 _debug_guardpage_minorder = res;
422 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
425 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
427 static inline void set_page_guard_flag(struct page *page)
429 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
432 static inline void clear_page_guard_flag(struct page *page)
434 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
437 static inline void set_page_guard_flag(struct page *page) { }
438 static inline void clear_page_guard_flag(struct page *page) { }
441 static inline void set_page_order(struct page *page, int order)
443 set_page_private(page, order);
444 __SetPageBuddy(page);
447 static inline void rmv_page_order(struct page *page)
449 __ClearPageBuddy(page);
450 set_page_private(page, 0);
454 * Locate the struct page for both the matching buddy in our
455 * pair (buddy1) and the combined O(n+1) page they form (page).
457 * 1) Any buddy B1 will have an order O twin B2 which satisfies
458 * the following equation:
460 * For example, if the starting buddy (buddy2) is #8 its order
462 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
464 * 2) Any buddy B will have an order O+1 parent P which
465 * satisfies the following equation:
468 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
470 static inline unsigned long
471 __find_buddy_index(unsigned long page_idx, unsigned int order)
473 return page_idx ^ (1 << order);
477 * This function checks whether a page is free && is the buddy
478 * we can do coalesce a page and its buddy if
479 * (a) the buddy is not in a hole &&
480 * (b) the buddy is in the buddy system &&
481 * (c) a page and its buddy have the same order &&
482 * (d) a page and its buddy are in the same zone.
484 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
485 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
487 * For recording page's order, we use page_private(page).
489 static inline int page_is_buddy(struct page *page, struct page *buddy,
492 if (!pfn_valid_within(page_to_pfn(buddy)))
495 if (page_zone_id(page) != page_zone_id(buddy))
498 if (page_is_guard(buddy) && page_order(buddy) == order) {
499 VM_BUG_ON(page_count(buddy) != 0);
503 if (PageBuddy(buddy) && page_order(buddy) == order) {
504 VM_BUG_ON(page_count(buddy) != 0);
511 * Freeing function for a buddy system allocator.
513 * The concept of a buddy system is to maintain direct-mapped table
514 * (containing bit values) for memory blocks of various "orders".
515 * The bottom level table contains the map for the smallest allocatable
516 * units of memory (here, pages), and each level above it describes
517 * pairs of units from the levels below, hence, "buddies".
518 * At a high level, all that happens here is marking the table entry
519 * at the bottom level available, and propagating the changes upward
520 * as necessary, plus some accounting needed to play nicely with other
521 * parts of the VM system.
522 * At each level, we keep a list of pages, which are heads of continuous
523 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
524 * order is recorded in page_private(page) field.
525 * So when we are allocating or freeing one, we can derive the state of the
526 * other. That is, if we allocate a small block, and both were
527 * free, the remainder of the region must be split into blocks.
528 * If a block is freed, and its buddy is also free, then this
529 * triggers coalescing into a block of larger size.
534 static inline void __free_one_page(struct page *page,
535 struct zone *zone, unsigned int order,
538 unsigned long page_idx;
539 unsigned long combined_idx;
540 unsigned long uninitialized_var(buddy_idx);
543 VM_BUG_ON(!zone_is_initialized(zone));
545 if (unlikely(PageCompound(page)))
546 if (unlikely(destroy_compound_page(page, order)))
549 VM_BUG_ON(migratetype == -1);
551 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
553 VM_BUG_ON(page_idx & ((1 << order) - 1));
554 VM_BUG_ON(bad_range(zone, page));
556 while (order < MAX_ORDER-1) {
557 buddy_idx = __find_buddy_index(page_idx, order);
558 buddy = page + (buddy_idx - page_idx);
559 if (!page_is_buddy(page, buddy, order))
562 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
563 * merge with it and move up one order.
565 if (page_is_guard(buddy)) {
566 clear_page_guard_flag(buddy);
567 set_page_private(page, 0);
568 __mod_zone_freepage_state(zone, 1 << order,
571 list_del(&buddy->lru);
572 zone->free_area[order].nr_free--;
573 rmv_page_order(buddy);
575 combined_idx = buddy_idx & page_idx;
576 page = page + (combined_idx - page_idx);
577 page_idx = combined_idx;
580 set_page_order(page, order);
583 * If this is not the largest possible page, check if the buddy
584 * of the next-highest order is free. If it is, it's possible
585 * that pages are being freed that will coalesce soon. In case,
586 * that is happening, add the free page to the tail of the list
587 * so it's less likely to be used soon and more likely to be merged
588 * as a higher order page
590 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
591 struct page *higher_page, *higher_buddy;
592 combined_idx = buddy_idx & page_idx;
593 higher_page = page + (combined_idx - page_idx);
594 buddy_idx = __find_buddy_index(combined_idx, order + 1);
595 higher_buddy = higher_page + (buddy_idx - combined_idx);
596 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
597 list_add_tail(&page->lru,
598 &zone->free_area[order].free_list[migratetype]);
603 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
605 zone->free_area[order].nr_free++;
608 static inline int free_pages_check(struct page *page)
610 if (unlikely(page_mapcount(page) |
611 (page->mapping != NULL) |
612 (atomic_read(&page->_count) != 0) |
613 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
614 (mem_cgroup_bad_page_check(page)))) {
618 page_nid_reset_last(page);
619 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
620 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
625 * Frees a number of pages from the PCP lists
626 * Assumes all pages on list are in same zone, and of same order.
627 * count is the number of pages to free.
629 * If the zone was previously in an "all pages pinned" state then look to
630 * see if this freeing clears that state.
632 * And clear the zone's pages_scanned counter, to hold off the "all pages are
633 * pinned" detection logic.
635 static void free_pcppages_bulk(struct zone *zone, int count,
636 struct per_cpu_pages *pcp)
642 spin_lock(&zone->lock);
643 zone->all_unreclaimable = 0;
644 zone->pages_scanned = 0;
648 struct list_head *list;
651 * Remove pages from lists in a round-robin fashion. A
652 * batch_free count is maintained that is incremented when an
653 * empty list is encountered. This is so more pages are freed
654 * off fuller lists instead of spinning excessively around empty
659 if (++migratetype == MIGRATE_PCPTYPES)
661 list = &pcp->lists[migratetype];
662 } while (list_empty(list));
664 /* This is the only non-empty list. Free them all. */
665 if (batch_free == MIGRATE_PCPTYPES)
666 batch_free = to_free;
669 int mt; /* migratetype of the to-be-freed page */
671 page = list_entry(list->prev, struct page, lru);
672 /* must delete as __free_one_page list manipulates */
673 list_del(&page->lru);
674 mt = get_freepage_migratetype(page);
675 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
676 __free_one_page(page, zone, 0, mt);
677 trace_mm_page_pcpu_drain(page, 0, mt);
678 if (likely(!is_migrate_isolate_page(page))) {
679 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
680 if (is_migrate_cma(mt))
681 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
683 } while (--to_free && --batch_free && !list_empty(list));
685 spin_unlock(&zone->lock);
688 static void free_one_page(struct zone *zone, struct page *page, int order,
691 spin_lock(&zone->lock);
692 zone->all_unreclaimable = 0;
693 zone->pages_scanned = 0;
695 __free_one_page(page, zone, order, migratetype);
696 if (unlikely(!is_migrate_isolate(migratetype)))
697 __mod_zone_freepage_state(zone, 1 << order, migratetype);
698 spin_unlock(&zone->lock);
701 static bool free_pages_prepare(struct page *page, unsigned int order)
706 trace_mm_page_free(page, order);
707 kmemcheck_free_shadow(page, order);
710 page->mapping = NULL;
711 for (i = 0; i < (1 << order); i++)
712 bad += free_pages_check(page + i);
716 if (!PageHighMem(page)) {
717 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
718 debug_check_no_obj_freed(page_address(page),
721 arch_free_page(page, order);
722 kernel_map_pages(page, 1 << order, 0);
727 static void __free_pages_ok(struct page *page, unsigned int order)
732 if (!free_pages_prepare(page, order))
735 local_irq_save(flags);
736 __count_vm_events(PGFREE, 1 << order);
737 migratetype = get_pageblock_migratetype(page);
738 set_freepage_migratetype(page, migratetype);
739 free_one_page(page_zone(page), page, order, migratetype);
740 local_irq_restore(flags);
744 * Read access to zone->managed_pages is safe because it's unsigned long,
745 * but we still need to serialize writers. Currently all callers of
746 * __free_pages_bootmem() except put_page_bootmem() should only be used
747 * at boot time. So for shorter boot time, we shift the burden to
748 * put_page_bootmem() to serialize writers.
750 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
752 unsigned int nr_pages = 1 << order;
756 for (loop = 0; loop < nr_pages; loop++) {
757 struct page *p = &page[loop];
759 if (loop + 1 < nr_pages)
761 __ClearPageReserved(p);
762 set_page_count(p, 0);
765 page_zone(page)->managed_pages += 1 << order;
766 set_page_refcounted(page);
767 __free_pages(page, order);
771 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
772 void __init init_cma_reserved_pageblock(struct page *page)
774 unsigned i = pageblock_nr_pages;
775 struct page *p = page;
778 __ClearPageReserved(p);
779 set_page_count(p, 0);
782 set_page_refcounted(page);
783 set_pageblock_migratetype(page, MIGRATE_CMA);
784 __free_pages(page, pageblock_order);
785 totalram_pages += pageblock_nr_pages;
786 #ifdef CONFIG_HIGHMEM
787 if (PageHighMem(page))
788 totalhigh_pages += pageblock_nr_pages;
794 * The order of subdivision here is critical for the IO subsystem.
795 * Please do not alter this order without good reasons and regression
796 * testing. Specifically, as large blocks of memory are subdivided,
797 * the order in which smaller blocks are delivered depends on the order
798 * they're subdivided in this function. This is the primary factor
799 * influencing the order in which pages are delivered to the IO
800 * subsystem according to empirical testing, and this is also justified
801 * by considering the behavior of a buddy system containing a single
802 * large block of memory acted on by a series of small allocations.
803 * This behavior is a critical factor in sglist merging's success.
807 static inline void expand(struct zone *zone, struct page *page,
808 int low, int high, struct free_area *area,
811 unsigned long size = 1 << high;
817 VM_BUG_ON(bad_range(zone, &page[size]));
819 #ifdef CONFIG_DEBUG_PAGEALLOC
820 if (high < debug_guardpage_minorder()) {
822 * Mark as guard pages (or page), that will allow to
823 * merge back to allocator when buddy will be freed.
824 * Corresponding page table entries will not be touched,
825 * pages will stay not present in virtual address space
827 INIT_LIST_HEAD(&page[size].lru);
828 set_page_guard_flag(&page[size]);
829 set_page_private(&page[size], high);
830 /* Guard pages are not available for any usage */
831 __mod_zone_freepage_state(zone, -(1 << high),
836 list_add(&page[size].lru, &area->free_list[migratetype]);
838 set_page_order(&page[size], high);
843 * This page is about to be returned from the page allocator
845 static inline int check_new_page(struct page *page)
847 if (unlikely(page_mapcount(page) |
848 (page->mapping != NULL) |
849 (atomic_read(&page->_count) != 0) |
850 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
851 (mem_cgroup_bad_page_check(page)))) {
858 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
862 for (i = 0; i < (1 << order); i++) {
863 struct page *p = page + i;
864 if (unlikely(check_new_page(p)))
868 set_page_private(page, 0);
869 set_page_refcounted(page);
871 arch_alloc_page(page, order);
872 kernel_map_pages(page, 1 << order, 1);
874 if (gfp_flags & __GFP_ZERO)
875 prep_zero_page(page, order, gfp_flags);
877 if (order && (gfp_flags & __GFP_COMP))
878 prep_compound_page(page, order);
884 * Go through the free lists for the given migratetype and remove
885 * the smallest available page from the freelists
888 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
891 unsigned int current_order;
892 struct free_area * area;
895 /* Find a page of the appropriate size in the preferred list */
896 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
897 area = &(zone->free_area[current_order]);
898 if (list_empty(&area->free_list[migratetype]))
901 page = list_entry(area->free_list[migratetype].next,
903 list_del(&page->lru);
904 rmv_page_order(page);
906 expand(zone, page, order, current_order, area, migratetype);
915 * This array describes the order lists are fallen back to when
916 * the free lists for the desirable migrate type are depleted
918 static int fallbacks[MIGRATE_TYPES][4] = {
919 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
920 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
922 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
923 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
925 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
927 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
928 #ifdef CONFIG_MEMORY_ISOLATION
929 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
934 * Move the free pages in a range to the free lists of the requested type.
935 * Note that start_page and end_pages are not aligned on a pageblock
936 * boundary. If alignment is required, use move_freepages_block()
938 int move_freepages(struct zone *zone,
939 struct page *start_page, struct page *end_page,
946 #ifndef CONFIG_HOLES_IN_ZONE
948 * page_zone is not safe to call in this context when
949 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
950 * anyway as we check zone boundaries in move_freepages_block().
951 * Remove at a later date when no bug reports exist related to
952 * grouping pages by mobility
954 BUG_ON(page_zone(start_page) != page_zone(end_page));
957 for (page = start_page; page <= end_page;) {
958 /* Make sure we are not inadvertently changing nodes */
959 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
961 if (!pfn_valid_within(page_to_pfn(page))) {
966 if (!PageBuddy(page)) {
971 order = page_order(page);
972 list_move(&page->lru,
973 &zone->free_area[order].free_list[migratetype]);
974 set_freepage_migratetype(page, migratetype);
976 pages_moved += 1 << order;
982 int move_freepages_block(struct zone *zone, struct page *page,
985 unsigned long start_pfn, end_pfn;
986 struct page *start_page, *end_page;
988 start_pfn = page_to_pfn(page);
989 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
990 start_page = pfn_to_page(start_pfn);
991 end_page = start_page + pageblock_nr_pages - 1;
992 end_pfn = start_pfn + pageblock_nr_pages - 1;
994 /* Do not cross zone boundaries */
995 if (!zone_spans_pfn(zone, start_pfn))
997 if (!zone_spans_pfn(zone, end_pfn))
1000 return move_freepages(zone, start_page, end_page, migratetype);
1003 static void change_pageblock_range(struct page *pageblock_page,
1004 int start_order, int migratetype)
1006 int nr_pageblocks = 1 << (start_order - pageblock_order);
1008 while (nr_pageblocks--) {
1009 set_pageblock_migratetype(pageblock_page, migratetype);
1010 pageblock_page += pageblock_nr_pages;
1014 /* Remove an element from the buddy allocator from the fallback list */
1015 static inline struct page *
1016 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1018 struct free_area * area;
1023 /* Find the largest possible block of pages in the other list */
1024 for (current_order = MAX_ORDER-1; current_order >= order;
1027 migratetype = fallbacks[start_migratetype][i];
1029 /* MIGRATE_RESERVE handled later if necessary */
1030 if (migratetype == MIGRATE_RESERVE)
1033 area = &(zone->free_area[current_order]);
1034 if (list_empty(&area->free_list[migratetype]))
1037 page = list_entry(area->free_list[migratetype].next,
1042 * If breaking a large block of pages, move all free
1043 * pages to the preferred allocation list. If falling
1044 * back for a reclaimable kernel allocation, be more
1045 * aggressive about taking ownership of free pages
1047 * On the other hand, never change migration
1048 * type of MIGRATE_CMA pageblocks nor move CMA
1049 * pages on different free lists. We don't
1050 * want unmovable pages to be allocated from
1051 * MIGRATE_CMA areas.
1053 if (!is_migrate_cma(migratetype) &&
1054 (unlikely(current_order >= pageblock_order / 2) ||
1055 start_migratetype == MIGRATE_RECLAIMABLE ||
1056 page_group_by_mobility_disabled)) {
1058 pages = move_freepages_block(zone, page,
1061 /* Claim the whole block if over half of it is free */
1062 if (pages >= (1 << (pageblock_order-1)) ||
1063 page_group_by_mobility_disabled)
1064 set_pageblock_migratetype(page,
1067 migratetype = start_migratetype;
1070 /* Remove the page from the freelists */
1071 list_del(&page->lru);
1072 rmv_page_order(page);
1074 /* Take ownership for orders >= pageblock_order */
1075 if (current_order >= pageblock_order &&
1076 !is_migrate_cma(migratetype))
1077 change_pageblock_range(page, current_order,
1080 expand(zone, page, order, current_order, area,
1081 is_migrate_cma(migratetype)
1082 ? migratetype : start_migratetype);
1084 trace_mm_page_alloc_extfrag(page, order, current_order,
1085 start_migratetype, migratetype);
1095 * Do the hard work of removing an element from the buddy allocator.
1096 * Call me with the zone->lock already held.
1098 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1104 page = __rmqueue_smallest(zone, order, migratetype);
1106 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1107 page = __rmqueue_fallback(zone, order, migratetype);
1110 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1111 * is used because __rmqueue_smallest is an inline function
1112 * and we want just one call site
1115 migratetype = MIGRATE_RESERVE;
1120 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1125 * Obtain a specified number of elements from the buddy allocator, all under
1126 * a single hold of the lock, for efficiency. Add them to the supplied list.
1127 * Returns the number of new pages which were placed at *list.
1129 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1130 unsigned long count, struct list_head *list,
1131 int migratetype, int cold)
1133 int mt = migratetype, i;
1135 spin_lock(&zone->lock);
1136 for (i = 0; i < count; ++i) {
1137 struct page *page = __rmqueue(zone, order, migratetype);
1138 if (unlikely(page == NULL))
1142 * Split buddy pages returned by expand() are received here
1143 * in physical page order. The page is added to the callers and
1144 * list and the list head then moves forward. From the callers
1145 * perspective, the linked list is ordered by page number in
1146 * some conditions. This is useful for IO devices that can
1147 * merge IO requests if the physical pages are ordered
1150 if (likely(cold == 0))
1151 list_add(&page->lru, list);
1153 list_add_tail(&page->lru, list);
1154 if (IS_ENABLED(CONFIG_CMA)) {
1155 mt = get_pageblock_migratetype(page);
1156 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1159 set_freepage_migratetype(page, mt);
1161 if (is_migrate_cma(mt))
1162 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1165 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1166 spin_unlock(&zone->lock);
1172 * Called from the vmstat counter updater to drain pagesets of this
1173 * currently executing processor on remote nodes after they have
1176 * Note that this function must be called with the thread pinned to
1177 * a single processor.
1179 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1181 unsigned long flags;
1184 local_irq_save(flags);
1185 if (pcp->count >= pcp->batch)
1186 to_drain = pcp->batch;
1188 to_drain = pcp->count;
1190 free_pcppages_bulk(zone, to_drain, pcp);
1191 pcp->count -= to_drain;
1193 local_irq_restore(flags);
1198 * Drain pages of the indicated processor.
1200 * The processor must either be the current processor and the
1201 * thread pinned to the current processor or a processor that
1204 static void drain_pages(unsigned int cpu)
1206 unsigned long flags;
1209 for_each_populated_zone(zone) {
1210 struct per_cpu_pageset *pset;
1211 struct per_cpu_pages *pcp;
1213 local_irq_save(flags);
1214 pset = per_cpu_ptr(zone->pageset, cpu);
1218 free_pcppages_bulk(zone, pcp->count, pcp);
1221 local_irq_restore(flags);
1226 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1228 void drain_local_pages(void *arg)
1230 drain_pages(smp_processor_id());
1234 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1236 * Note that this code is protected against sending an IPI to an offline
1237 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1238 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1239 * nothing keeps CPUs from showing up after we populated the cpumask and
1240 * before the call to on_each_cpu_mask().
1242 void drain_all_pages(void)
1245 struct per_cpu_pageset *pcp;
1249 * Allocate in the BSS so we wont require allocation in
1250 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1252 static cpumask_t cpus_with_pcps;
1255 * We don't care about racing with CPU hotplug event
1256 * as offline notification will cause the notified
1257 * cpu to drain that CPU pcps and on_each_cpu_mask
1258 * disables preemption as part of its processing
1260 for_each_online_cpu(cpu) {
1261 bool has_pcps = false;
1262 for_each_populated_zone(zone) {
1263 pcp = per_cpu_ptr(zone->pageset, cpu);
1264 if (pcp->pcp.count) {
1270 cpumask_set_cpu(cpu, &cpus_with_pcps);
1272 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1274 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1277 #ifdef CONFIG_HIBERNATION
1279 void mark_free_pages(struct zone *zone)
1281 unsigned long pfn, max_zone_pfn;
1282 unsigned long flags;
1284 struct list_head *curr;
1286 if (!zone->spanned_pages)
1289 spin_lock_irqsave(&zone->lock, flags);
1291 max_zone_pfn = zone_end_pfn(zone);
1292 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1293 if (pfn_valid(pfn)) {
1294 struct page *page = pfn_to_page(pfn);
1296 if (!swsusp_page_is_forbidden(page))
1297 swsusp_unset_page_free(page);
1300 for_each_migratetype_order(order, t) {
1301 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1304 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1305 for (i = 0; i < (1UL << order); i++)
1306 swsusp_set_page_free(pfn_to_page(pfn + i));
1309 spin_unlock_irqrestore(&zone->lock, flags);
1311 #endif /* CONFIG_PM */
1314 * Free a 0-order page
1315 * cold == 1 ? free a cold page : free a hot page
1317 void free_hot_cold_page(struct page *page, int cold)
1319 struct zone *zone = page_zone(page);
1320 struct per_cpu_pages *pcp;
1321 unsigned long flags;
1324 if (!free_pages_prepare(page, 0))
1327 migratetype = get_pageblock_migratetype(page);
1328 set_freepage_migratetype(page, migratetype);
1329 local_irq_save(flags);
1330 __count_vm_event(PGFREE);
1333 * We only track unmovable, reclaimable and movable on pcp lists.
1334 * Free ISOLATE pages back to the allocator because they are being
1335 * offlined but treat RESERVE as movable pages so we can get those
1336 * areas back if necessary. Otherwise, we may have to free
1337 * excessively into the page allocator
1339 if (migratetype >= MIGRATE_PCPTYPES) {
1340 if (unlikely(is_migrate_isolate(migratetype))) {
1341 free_one_page(zone, page, 0, migratetype);
1344 migratetype = MIGRATE_MOVABLE;
1347 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1349 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1351 list_add(&page->lru, &pcp->lists[migratetype]);
1353 if (pcp->count >= pcp->high) {
1354 free_pcppages_bulk(zone, pcp->batch, pcp);
1355 pcp->count -= pcp->batch;
1359 local_irq_restore(flags);
1363 * Free a list of 0-order pages
1365 void free_hot_cold_page_list(struct list_head *list, int cold)
1367 struct page *page, *next;
1369 list_for_each_entry_safe(page, next, list, lru) {
1370 trace_mm_page_free_batched(page, cold);
1371 free_hot_cold_page(page, cold);
1376 * split_page takes a non-compound higher-order page, and splits it into
1377 * n (1<<order) sub-pages: page[0..n]
1378 * Each sub-page must be freed individually.
1380 * Note: this is probably too low level an operation for use in drivers.
1381 * Please consult with lkml before using this in your driver.
1383 void split_page(struct page *page, unsigned int order)
1387 VM_BUG_ON(PageCompound(page));
1388 VM_BUG_ON(!page_count(page));
1390 #ifdef CONFIG_KMEMCHECK
1392 * Split shadow pages too, because free(page[0]) would
1393 * otherwise free the whole shadow.
1395 if (kmemcheck_page_is_tracked(page))
1396 split_page(virt_to_page(page[0].shadow), order);
1399 for (i = 1; i < (1 << order); i++)
1400 set_page_refcounted(page + i);
1402 EXPORT_SYMBOL_GPL(split_page);
1404 static int __isolate_free_page(struct page *page, unsigned int order)
1406 unsigned long watermark;
1410 BUG_ON(!PageBuddy(page));
1412 zone = page_zone(page);
1413 mt = get_pageblock_migratetype(page);
1415 if (!is_migrate_isolate(mt)) {
1416 /* Obey watermarks as if the page was being allocated */
1417 watermark = low_wmark_pages(zone) + (1 << order);
1418 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1421 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1424 /* Remove page from free list */
1425 list_del(&page->lru);
1426 zone->free_area[order].nr_free--;
1427 rmv_page_order(page);
1429 /* Set the pageblock if the isolated page is at least a pageblock */
1430 if (order >= pageblock_order - 1) {
1431 struct page *endpage = page + (1 << order) - 1;
1432 for (; page < endpage; page += pageblock_nr_pages) {
1433 int mt = get_pageblock_migratetype(page);
1434 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1435 set_pageblock_migratetype(page,
1440 return 1UL << order;
1444 * Similar to split_page except the page is already free. As this is only
1445 * being used for migration, the migratetype of the block also changes.
1446 * As this is called with interrupts disabled, the caller is responsible
1447 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1450 * Note: this is probably too low level an operation for use in drivers.
1451 * Please consult with lkml before using this in your driver.
1453 int split_free_page(struct page *page)
1458 order = page_order(page);
1460 nr_pages = __isolate_free_page(page, order);
1464 /* Split into individual pages */
1465 set_page_refcounted(page);
1466 split_page(page, order);
1471 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1472 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1476 struct page *buffered_rmqueue(struct zone *preferred_zone,
1477 struct zone *zone, int order, gfp_t gfp_flags,
1480 unsigned long flags;
1482 int cold = !!(gfp_flags & __GFP_COLD);
1485 if (likely(order == 0)) {
1486 struct per_cpu_pages *pcp;
1487 struct list_head *list;
1489 local_irq_save(flags);
1490 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1491 list = &pcp->lists[migratetype];
1492 if (list_empty(list)) {
1493 pcp->count += rmqueue_bulk(zone, 0,
1496 if (unlikely(list_empty(list)))
1501 page = list_entry(list->prev, struct page, lru);
1503 page = list_entry(list->next, struct page, lru);
1505 list_del(&page->lru);
1508 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1510 * __GFP_NOFAIL is not to be used in new code.
1512 * All __GFP_NOFAIL callers should be fixed so that they
1513 * properly detect and handle allocation failures.
1515 * We most definitely don't want callers attempting to
1516 * allocate greater than order-1 page units with
1519 WARN_ON_ONCE(order > 1);
1521 spin_lock_irqsave(&zone->lock, flags);
1522 page = __rmqueue(zone, order, migratetype);
1523 spin_unlock(&zone->lock);
1526 __mod_zone_freepage_state(zone, -(1 << order),
1527 get_pageblock_migratetype(page));
1530 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1531 zone_statistics(preferred_zone, zone, gfp_flags);
1532 local_irq_restore(flags);
1534 VM_BUG_ON(bad_range(zone, page));
1535 if (prep_new_page(page, order, gfp_flags))
1540 local_irq_restore(flags);
1544 #ifdef CONFIG_FAIL_PAGE_ALLOC
1547 struct fault_attr attr;
1549 u32 ignore_gfp_highmem;
1550 u32 ignore_gfp_wait;
1552 } fail_page_alloc = {
1553 .attr = FAULT_ATTR_INITIALIZER,
1554 .ignore_gfp_wait = 1,
1555 .ignore_gfp_highmem = 1,
1559 static int __init setup_fail_page_alloc(char *str)
1561 return setup_fault_attr(&fail_page_alloc.attr, str);
1563 __setup("fail_page_alloc=", setup_fail_page_alloc);
1565 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1567 if (order < fail_page_alloc.min_order)
1569 if (gfp_mask & __GFP_NOFAIL)
1571 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1573 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1576 return should_fail(&fail_page_alloc.attr, 1 << order);
1579 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1581 static int __init fail_page_alloc_debugfs(void)
1583 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1586 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1587 &fail_page_alloc.attr);
1589 return PTR_ERR(dir);
1591 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1592 &fail_page_alloc.ignore_gfp_wait))
1594 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1595 &fail_page_alloc.ignore_gfp_highmem))
1597 if (!debugfs_create_u32("min-order", mode, dir,
1598 &fail_page_alloc.min_order))
1603 debugfs_remove_recursive(dir);
1608 late_initcall(fail_page_alloc_debugfs);
1610 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1612 #else /* CONFIG_FAIL_PAGE_ALLOC */
1614 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1619 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1622 * Return true if free pages are above 'mark'. This takes into account the order
1623 * of the allocation.
1625 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1626 int classzone_idx, int alloc_flags, long free_pages)
1628 /* free_pages my go negative - that's OK */
1630 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1634 free_pages -= (1 << order) - 1;
1635 if (alloc_flags & ALLOC_HIGH)
1637 if (alloc_flags & ALLOC_HARDER)
1640 /* If allocation can't use CMA areas don't use free CMA pages */
1641 if (!(alloc_flags & ALLOC_CMA))
1642 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1645 if (free_pages - free_cma <= min + lowmem_reserve)
1647 for (o = 0; o < order; o++) {
1648 /* At the next order, this order's pages become unavailable */
1649 free_pages -= z->free_area[o].nr_free << o;
1651 /* Require fewer higher order pages to be free */
1652 min >>= min_free_order_shift;
1654 if (free_pages <= min)
1660 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1661 int classzone_idx, int alloc_flags)
1663 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1664 zone_page_state(z, NR_FREE_PAGES));
1667 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1668 int classzone_idx, int alloc_flags)
1670 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1672 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1673 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1675 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1681 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1682 * skip over zones that are not allowed by the cpuset, or that have
1683 * been recently (in last second) found to be nearly full. See further
1684 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1685 * that have to skip over a lot of full or unallowed zones.
1687 * If the zonelist cache is present in the passed in zonelist, then
1688 * returns a pointer to the allowed node mask (either the current
1689 * tasks mems_allowed, or node_states[N_MEMORY].)
1691 * If the zonelist cache is not available for this zonelist, does
1692 * nothing and returns NULL.
1694 * If the fullzones BITMAP in the zonelist cache is stale (more than
1695 * a second since last zap'd) then we zap it out (clear its bits.)
1697 * We hold off even calling zlc_setup, until after we've checked the
1698 * first zone in the zonelist, on the theory that most allocations will
1699 * be satisfied from that first zone, so best to examine that zone as
1700 * quickly as we can.
1702 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1704 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1705 nodemask_t *allowednodes; /* zonelist_cache approximation */
1707 zlc = zonelist->zlcache_ptr;
1711 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1712 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1713 zlc->last_full_zap = jiffies;
1716 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1717 &cpuset_current_mems_allowed :
1718 &node_states[N_MEMORY];
1719 return allowednodes;
1723 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1724 * if it is worth looking at further for free memory:
1725 * 1) Check that the zone isn't thought to be full (doesn't have its
1726 * bit set in the zonelist_cache fullzones BITMAP).
1727 * 2) Check that the zones node (obtained from the zonelist_cache
1728 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1729 * Return true (non-zero) if zone is worth looking at further, or
1730 * else return false (zero) if it is not.
1732 * This check -ignores- the distinction between various watermarks,
1733 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1734 * found to be full for any variation of these watermarks, it will
1735 * be considered full for up to one second by all requests, unless
1736 * we are so low on memory on all allowed nodes that we are forced
1737 * into the second scan of the zonelist.
1739 * In the second scan we ignore this zonelist cache and exactly
1740 * apply the watermarks to all zones, even it is slower to do so.
1741 * We are low on memory in the second scan, and should leave no stone
1742 * unturned looking for a free page.
1744 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1745 nodemask_t *allowednodes)
1747 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1748 int i; /* index of *z in zonelist zones */
1749 int n; /* node that zone *z is on */
1751 zlc = zonelist->zlcache_ptr;
1755 i = z - zonelist->_zonerefs;
1758 /* This zone is worth trying if it is allowed but not full */
1759 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1763 * Given 'z' scanning a zonelist, set the corresponding bit in
1764 * zlc->fullzones, so that subsequent attempts to allocate a page
1765 * from that zone don't waste time re-examining it.
1767 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1769 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1770 int i; /* index of *z in zonelist zones */
1772 zlc = zonelist->zlcache_ptr;
1776 i = z - zonelist->_zonerefs;
1778 set_bit(i, zlc->fullzones);
1782 * clear all zones full, called after direct reclaim makes progress so that
1783 * a zone that was recently full is not skipped over for up to a second
1785 static void zlc_clear_zones_full(struct zonelist *zonelist)
1787 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1789 zlc = zonelist->zlcache_ptr;
1793 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1796 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1798 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1801 static void __paginginit init_zone_allows_reclaim(int nid)
1805 for_each_online_node(i)
1806 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1807 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1809 zone_reclaim_mode = 1;
1812 #else /* CONFIG_NUMA */
1814 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1819 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1820 nodemask_t *allowednodes)
1825 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1829 static void zlc_clear_zones_full(struct zonelist *zonelist)
1833 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1838 static inline void init_zone_allows_reclaim(int nid)
1841 #endif /* CONFIG_NUMA */
1844 * get_page_from_freelist goes through the zonelist trying to allocate
1847 static struct page *
1848 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1849 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1850 struct zone *preferred_zone, int migratetype)
1853 struct page *page = NULL;
1856 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1857 int zlc_active = 0; /* set if using zonelist_cache */
1858 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1860 classzone_idx = zone_idx(preferred_zone);
1863 * Scan zonelist, looking for a zone with enough free.
1864 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1866 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1867 high_zoneidx, nodemask) {
1868 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1869 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1871 if ((alloc_flags & ALLOC_CPUSET) &&
1872 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1875 * When allocating a page cache page for writing, we
1876 * want to get it from a zone that is within its dirty
1877 * limit, such that no single zone holds more than its
1878 * proportional share of globally allowed dirty pages.
1879 * The dirty limits take into account the zone's
1880 * lowmem reserves and high watermark so that kswapd
1881 * should be able to balance it without having to
1882 * write pages from its LRU list.
1884 * This may look like it could increase pressure on
1885 * lower zones by failing allocations in higher zones
1886 * before they are full. But the pages that do spill
1887 * over are limited as the lower zones are protected
1888 * by this very same mechanism. It should not become
1889 * a practical burden to them.
1891 * XXX: For now, allow allocations to potentially
1892 * exceed the per-zone dirty limit in the slowpath
1893 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1894 * which is important when on a NUMA setup the allowed
1895 * zones are together not big enough to reach the
1896 * global limit. The proper fix for these situations
1897 * will require awareness of zones in the
1898 * dirty-throttling and the flusher threads.
1900 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1901 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1902 goto this_zone_full;
1904 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1905 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1909 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1910 if (zone_watermark_ok(zone, order, mark,
1911 classzone_idx, alloc_flags))
1914 if (IS_ENABLED(CONFIG_NUMA) &&
1915 !did_zlc_setup && nr_online_nodes > 1) {
1917 * we do zlc_setup if there are multiple nodes
1918 * and before considering the first zone allowed
1921 allowednodes = zlc_setup(zonelist, alloc_flags);
1926 if (zone_reclaim_mode == 0 ||
1927 !zone_allows_reclaim(preferred_zone, zone))
1928 goto this_zone_full;
1931 * As we may have just activated ZLC, check if the first
1932 * eligible zone has failed zone_reclaim recently.
1934 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1935 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1938 ret = zone_reclaim(zone, gfp_mask, order);
1940 case ZONE_RECLAIM_NOSCAN:
1943 case ZONE_RECLAIM_FULL:
1944 /* scanned but unreclaimable */
1947 /* did we reclaim enough */
1948 if (zone_watermark_ok(zone, order, mark,
1949 classzone_idx, alloc_flags))
1953 * Failed to reclaim enough to meet watermark.
1954 * Only mark the zone full if checking the min
1955 * watermark or if we failed to reclaim just
1956 * 1<<order pages or else the page allocator
1957 * fastpath will prematurely mark zones full
1958 * when the watermark is between the low and
1961 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
1962 ret == ZONE_RECLAIM_SOME)
1963 goto this_zone_full;
1970 page = buffered_rmqueue(preferred_zone, zone, order,
1971 gfp_mask, migratetype);
1975 if (IS_ENABLED(CONFIG_NUMA))
1976 zlc_mark_zone_full(zonelist, z);
1979 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1980 /* Disable zlc cache for second zonelist scan */
1987 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1988 * necessary to allocate the page. The expectation is
1989 * that the caller is taking steps that will free more
1990 * memory. The caller should avoid the page being used
1991 * for !PFMEMALLOC purposes.
1993 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1999 * Large machines with many possible nodes should not always dump per-node
2000 * meminfo in irq context.
2002 static inline bool should_suppress_show_mem(void)
2007 ret = in_interrupt();
2012 static DEFINE_RATELIMIT_STATE(nopage_rs,
2013 DEFAULT_RATELIMIT_INTERVAL,
2014 DEFAULT_RATELIMIT_BURST);
2016 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2018 unsigned int filter = SHOW_MEM_FILTER_NODES;
2020 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2021 debug_guardpage_minorder() > 0)
2025 * Walking all memory to count page types is very expensive and should
2026 * be inhibited in non-blockable contexts.
2028 if (!(gfp_mask & __GFP_WAIT))
2029 filter |= SHOW_MEM_FILTER_PAGE_COUNT;
2032 * This documents exceptions given to allocations in certain
2033 * contexts that are allowed to allocate outside current's set
2036 if (!(gfp_mask & __GFP_NOMEMALLOC))
2037 if (test_thread_flag(TIF_MEMDIE) ||
2038 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2039 filter &= ~SHOW_MEM_FILTER_NODES;
2040 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2041 filter &= ~SHOW_MEM_FILTER_NODES;
2044 struct va_format vaf;
2047 va_start(args, fmt);
2052 pr_warn("%pV", &vaf);
2057 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2058 current->comm, order, gfp_mask);
2061 if (!should_suppress_show_mem())
2066 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2067 unsigned long did_some_progress,
2068 unsigned long pages_reclaimed)
2070 /* Do not loop if specifically requested */
2071 if (gfp_mask & __GFP_NORETRY)
2074 /* Always retry if specifically requested */
2075 if (gfp_mask & __GFP_NOFAIL)
2079 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2080 * making forward progress without invoking OOM. Suspend also disables
2081 * storage devices so kswapd will not help. Bail if we are suspending.
2083 if (!did_some_progress && pm_suspended_storage())
2087 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2088 * means __GFP_NOFAIL, but that may not be true in other
2091 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2095 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2096 * specified, then we retry until we no longer reclaim any pages
2097 * (above), or we've reclaimed an order of pages at least as
2098 * large as the allocation's order. In both cases, if the
2099 * allocation still fails, we stop retrying.
2101 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2107 static inline struct page *
2108 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2109 struct zonelist *zonelist, enum zone_type high_zoneidx,
2110 nodemask_t *nodemask, struct zone *preferred_zone,
2115 /* Acquire the OOM killer lock for the zones in zonelist */
2116 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2117 schedule_timeout_uninterruptible(1);
2122 * Go through the zonelist yet one more time, keep very high watermark
2123 * here, this is only to catch a parallel oom killing, we must fail if
2124 * we're still under heavy pressure.
2126 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2127 order, zonelist, high_zoneidx,
2128 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2129 preferred_zone, migratetype);
2133 if (!(gfp_mask & __GFP_NOFAIL)) {
2134 /* The OOM killer will not help higher order allocs */
2135 if (order > PAGE_ALLOC_COSTLY_ORDER)
2137 /* The OOM killer does not needlessly kill tasks for lowmem */
2138 if (high_zoneidx < ZONE_NORMAL)
2141 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2142 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2143 * The caller should handle page allocation failure by itself if
2144 * it specifies __GFP_THISNODE.
2145 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2147 if (gfp_mask & __GFP_THISNODE)
2150 /* Exhausted what can be done so it's blamo time */
2151 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2154 clear_zonelist_oom(zonelist, gfp_mask);
2158 #ifdef CONFIG_COMPACTION
2159 /* Try memory compaction for high-order allocations before reclaim */
2160 static struct page *
2161 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2162 struct zonelist *zonelist, enum zone_type high_zoneidx,
2163 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2164 int migratetype, bool sync_migration,
2165 bool *contended_compaction, bool *deferred_compaction,
2166 unsigned long *did_some_progress)
2171 if (compaction_deferred(preferred_zone, order)) {
2172 *deferred_compaction = true;
2176 current->flags |= PF_MEMALLOC;
2177 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2178 nodemask, sync_migration,
2179 contended_compaction);
2180 current->flags &= ~PF_MEMALLOC;
2182 if (*did_some_progress != COMPACT_SKIPPED) {
2185 /* Page migration frees to the PCP lists but we want merging */
2186 drain_pages(get_cpu());
2189 page = get_page_from_freelist(gfp_mask, nodemask,
2190 order, zonelist, high_zoneidx,
2191 alloc_flags & ~ALLOC_NO_WATERMARKS,
2192 preferred_zone, migratetype);
2194 preferred_zone->compact_blockskip_flush = false;
2195 preferred_zone->compact_considered = 0;
2196 preferred_zone->compact_defer_shift = 0;
2197 if (order >= preferred_zone->compact_order_failed)
2198 preferred_zone->compact_order_failed = order + 1;
2199 count_vm_event(COMPACTSUCCESS);
2204 * It's bad if compaction run occurs and fails.
2205 * The most likely reason is that pages exist,
2206 * but not enough to satisfy watermarks.
2208 count_vm_event(COMPACTFAIL);
2211 * As async compaction considers a subset of pageblocks, only
2212 * defer if the failure was a sync compaction failure.
2215 defer_compaction(preferred_zone, order);
2223 static inline struct page *
2224 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2225 struct zonelist *zonelist, enum zone_type high_zoneidx,
2226 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2227 int migratetype, bool sync_migration,
2228 bool *contended_compaction, bool *deferred_compaction,
2229 unsigned long *did_some_progress)
2233 #endif /* CONFIG_COMPACTION */
2235 /* Perform direct synchronous page reclaim */
2237 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2238 nodemask_t *nodemask)
2240 struct reclaim_state reclaim_state;
2245 /* We now go into synchronous reclaim */
2246 cpuset_memory_pressure_bump();
2247 current->flags |= PF_MEMALLOC;
2248 lockdep_set_current_reclaim_state(gfp_mask);
2249 reclaim_state.reclaimed_slab = 0;
2250 current->reclaim_state = &reclaim_state;
2252 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2254 current->reclaim_state = NULL;
2255 lockdep_clear_current_reclaim_state();
2256 current->flags &= ~PF_MEMALLOC;
2263 /* The really slow allocator path where we enter direct reclaim */
2264 static inline struct page *
2265 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2266 struct zonelist *zonelist, enum zone_type high_zoneidx,
2267 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2268 int migratetype, unsigned long *did_some_progress)
2270 struct page *page = NULL;
2271 bool drained = false;
2273 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2275 if (unlikely(!(*did_some_progress)))
2278 /* After successful reclaim, reconsider all zones for allocation */
2279 if (IS_ENABLED(CONFIG_NUMA))
2280 zlc_clear_zones_full(zonelist);
2283 page = get_page_from_freelist(gfp_mask, nodemask, order,
2284 zonelist, high_zoneidx,
2285 alloc_flags & ~ALLOC_NO_WATERMARKS,
2286 preferred_zone, migratetype);
2289 * If an allocation failed after direct reclaim, it could be because
2290 * pages are pinned on the per-cpu lists. Drain them and try again
2292 if (!page && !drained) {
2302 * This is called in the allocator slow-path if the allocation request is of
2303 * sufficient urgency to ignore watermarks and take other desperate measures
2305 static inline struct page *
2306 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2307 struct zonelist *zonelist, enum zone_type high_zoneidx,
2308 nodemask_t *nodemask, struct zone *preferred_zone,
2314 page = get_page_from_freelist(gfp_mask, nodemask, order,
2315 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2316 preferred_zone, migratetype);
2318 if (!page && gfp_mask & __GFP_NOFAIL)
2319 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2320 } while (!page && (gfp_mask & __GFP_NOFAIL));
2326 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2327 enum zone_type high_zoneidx,
2328 enum zone_type classzone_idx)
2333 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2334 wakeup_kswapd(zone, order, classzone_idx);
2338 gfp_to_alloc_flags(gfp_t gfp_mask)
2340 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2341 const gfp_t wait = gfp_mask & __GFP_WAIT;
2343 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2344 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2347 * The caller may dip into page reserves a bit more if the caller
2348 * cannot run direct reclaim, or if the caller has realtime scheduling
2349 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2350 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2352 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2356 * Not worth trying to allocate harder for
2357 * __GFP_NOMEMALLOC even if it can't schedule.
2359 if (!(gfp_mask & __GFP_NOMEMALLOC))
2360 alloc_flags |= ALLOC_HARDER;
2362 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2363 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2365 alloc_flags &= ~ALLOC_CPUSET;
2366 } else if (unlikely(rt_task(current)) && !in_interrupt())
2367 alloc_flags |= ALLOC_HARDER;
2369 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2370 if (gfp_mask & __GFP_MEMALLOC)
2371 alloc_flags |= ALLOC_NO_WATERMARKS;
2372 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2373 alloc_flags |= ALLOC_NO_WATERMARKS;
2374 else if (!in_interrupt() &&
2375 ((current->flags & PF_MEMALLOC) ||
2376 unlikely(test_thread_flag(TIF_MEMDIE))))
2377 alloc_flags |= ALLOC_NO_WATERMARKS;
2380 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2381 alloc_flags |= ALLOC_CMA;
2386 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2388 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2391 static inline struct page *
2392 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2393 struct zonelist *zonelist, enum zone_type high_zoneidx,
2394 nodemask_t *nodemask, struct zone *preferred_zone,
2397 const gfp_t wait = gfp_mask & __GFP_WAIT;
2398 struct page *page = NULL;
2400 unsigned long pages_reclaimed = 0;
2401 unsigned long did_some_progress;
2402 bool sync_migration = false;
2403 bool deferred_compaction = false;
2404 bool contended_compaction = false;
2407 * In the slowpath, we sanity check order to avoid ever trying to
2408 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2409 * be using allocators in order of preference for an area that is
2412 if (order >= MAX_ORDER) {
2413 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2418 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2419 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2420 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2421 * using a larger set of nodes after it has established that the
2422 * allowed per node queues are empty and that nodes are
2425 if (IS_ENABLED(CONFIG_NUMA) &&
2426 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2430 if (!(gfp_mask & __GFP_NO_KSWAPD))
2431 wake_all_kswapd(order, zonelist, high_zoneidx,
2432 zone_idx(preferred_zone));
2435 * OK, we're below the kswapd watermark and have kicked background
2436 * reclaim. Now things get more complex, so set up alloc_flags according
2437 * to how we want to proceed.
2439 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2442 * Find the true preferred zone if the allocation is unconstrained by
2445 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2446 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2450 /* This is the last chance, in general, before the goto nopage. */
2451 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2452 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2453 preferred_zone, migratetype);
2457 /* Allocate without watermarks if the context allows */
2458 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2460 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2461 * the allocation is high priority and these type of
2462 * allocations are system rather than user orientated
2464 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2466 page = __alloc_pages_high_priority(gfp_mask, order,
2467 zonelist, high_zoneidx, nodemask,
2468 preferred_zone, migratetype);
2474 /* Atomic allocations - we can't balance anything */
2478 /* Avoid recursion of direct reclaim */
2479 if (current->flags & PF_MEMALLOC)
2482 /* Avoid allocations with no watermarks from looping endlessly */
2483 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2487 * Try direct compaction. The first pass is asynchronous. Subsequent
2488 * attempts after direct reclaim are synchronous
2490 page = __alloc_pages_direct_compact(gfp_mask, order,
2491 zonelist, high_zoneidx,
2493 alloc_flags, preferred_zone,
2494 migratetype, sync_migration,
2495 &contended_compaction,
2496 &deferred_compaction,
2497 &did_some_progress);
2500 sync_migration = true;
2503 * If compaction is deferred for high-order allocations, it is because
2504 * sync compaction recently failed. In this is the case and the caller
2505 * requested a movable allocation that does not heavily disrupt the
2506 * system then fail the allocation instead of entering direct reclaim.
2508 if ((deferred_compaction || contended_compaction) &&
2509 (gfp_mask & __GFP_NO_KSWAPD))
2512 /* Try direct reclaim and then allocating */
2513 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2514 zonelist, high_zoneidx,
2516 alloc_flags, preferred_zone,
2517 migratetype, &did_some_progress);
2522 * If we failed to make any progress reclaiming, then we are
2523 * running out of options and have to consider going OOM
2525 if (!did_some_progress) {
2526 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2527 if (oom_killer_disabled)
2529 /* Coredumps can quickly deplete all memory reserves */
2530 if ((current->flags & PF_DUMPCORE) &&
2531 !(gfp_mask & __GFP_NOFAIL))
2533 page = __alloc_pages_may_oom(gfp_mask, order,
2534 zonelist, high_zoneidx,
2535 nodemask, preferred_zone,
2540 if (!(gfp_mask & __GFP_NOFAIL)) {
2542 * The oom killer is not called for high-order
2543 * allocations that may fail, so if no progress
2544 * is being made, there are no other options and
2545 * retrying is unlikely to help.
2547 if (order > PAGE_ALLOC_COSTLY_ORDER)
2550 * The oom killer is not called for lowmem
2551 * allocations to prevent needlessly killing
2554 if (high_zoneidx < ZONE_NORMAL)
2562 /* Check if we should retry the allocation */
2563 pages_reclaimed += did_some_progress;
2564 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2566 /* Wait for some write requests to complete then retry */
2567 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2571 * High-order allocations do not necessarily loop after
2572 * direct reclaim and reclaim/compaction depends on compaction
2573 * being called after reclaim so call directly if necessary
2575 page = __alloc_pages_direct_compact(gfp_mask, order,
2576 zonelist, high_zoneidx,
2578 alloc_flags, preferred_zone,
2579 migratetype, sync_migration,
2580 &contended_compaction,
2581 &deferred_compaction,
2582 &did_some_progress);
2588 warn_alloc_failed(gfp_mask, order, NULL);
2591 if (kmemcheck_enabled)
2592 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2598 * This is the 'heart' of the zoned buddy allocator.
2601 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2602 struct zonelist *zonelist, nodemask_t *nodemask)
2604 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2605 struct zone *preferred_zone;
2606 struct page *page = NULL;
2607 int migratetype = allocflags_to_migratetype(gfp_mask);
2608 unsigned int cpuset_mems_cookie;
2609 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2610 struct mem_cgroup *memcg = NULL;
2612 gfp_mask &= gfp_allowed_mask;
2614 lockdep_trace_alloc(gfp_mask);
2616 might_sleep_if(gfp_mask & __GFP_WAIT);
2618 if (should_fail_alloc_page(gfp_mask, order))
2622 * Check the zones suitable for the gfp_mask contain at least one
2623 * valid zone. It's possible to have an empty zonelist as a result
2624 * of GFP_THISNODE and a memoryless node
2626 if (unlikely(!zonelist->_zonerefs->zone))
2630 * Will only have any effect when __GFP_KMEMCG is set. This is
2631 * verified in the (always inline) callee
2633 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2637 cpuset_mems_cookie = get_mems_allowed();
2639 /* The preferred zone is used for statistics later */
2640 first_zones_zonelist(zonelist, high_zoneidx,
2641 nodemask ? : &cpuset_current_mems_allowed,
2643 if (!preferred_zone)
2647 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2648 alloc_flags |= ALLOC_CMA;
2650 /* First allocation attempt */
2651 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2652 zonelist, high_zoneidx, alloc_flags,
2653 preferred_zone, migratetype);
2654 if (unlikely(!page)) {
2656 * Runtime PM, block IO and its error handling path
2657 * can deadlock because I/O on the device might not
2660 gfp_mask = memalloc_noio_flags(gfp_mask);
2661 page = __alloc_pages_slowpath(gfp_mask, order,
2662 zonelist, high_zoneidx, nodemask,
2663 preferred_zone, migratetype);
2666 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2670 * When updating a task's mems_allowed, it is possible to race with
2671 * parallel threads in such a way that an allocation can fail while
2672 * the mask is being updated. If a page allocation is about to fail,
2673 * check if the cpuset changed during allocation and if so, retry.
2675 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2678 memcg_kmem_commit_charge(page, memcg, order);
2682 EXPORT_SYMBOL(__alloc_pages_nodemask);
2685 * Common helper functions.
2687 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2692 * __get_free_pages() returns a 32-bit address, which cannot represent
2695 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2697 page = alloc_pages(gfp_mask, order);
2700 return (unsigned long) page_address(page);
2702 EXPORT_SYMBOL(__get_free_pages);
2704 unsigned long get_zeroed_page(gfp_t gfp_mask)
2706 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2708 EXPORT_SYMBOL(get_zeroed_page);
2710 void __free_pages(struct page *page, unsigned int order)
2712 if (put_page_testzero(page)) {
2714 free_hot_cold_page(page, 0);
2716 __free_pages_ok(page, order);
2720 EXPORT_SYMBOL(__free_pages);
2722 void free_pages(unsigned long addr, unsigned int order)
2725 VM_BUG_ON(!virt_addr_valid((void *)addr));
2726 __free_pages(virt_to_page((void *)addr), order);
2730 EXPORT_SYMBOL(free_pages);
2733 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2734 * pages allocated with __GFP_KMEMCG.
2736 * Those pages are accounted to a particular memcg, embedded in the
2737 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2738 * for that information only to find out that it is NULL for users who have no
2739 * interest in that whatsoever, we provide these functions.
2741 * The caller knows better which flags it relies on.
2743 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2745 memcg_kmem_uncharge_pages(page, order);
2746 __free_pages(page, order);
2749 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2752 VM_BUG_ON(!virt_addr_valid((void *)addr));
2753 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2757 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2760 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2761 unsigned long used = addr + PAGE_ALIGN(size);
2763 split_page(virt_to_page((void *)addr), order);
2764 while (used < alloc_end) {
2769 return (void *)addr;
2773 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2774 * @size: the number of bytes to allocate
2775 * @gfp_mask: GFP flags for the allocation
2777 * This function is similar to alloc_pages(), except that it allocates the
2778 * minimum number of pages to satisfy the request. alloc_pages() can only
2779 * allocate memory in power-of-two pages.
2781 * This function is also limited by MAX_ORDER.
2783 * Memory allocated by this function must be released by free_pages_exact().
2785 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2787 unsigned int order = get_order(size);
2790 addr = __get_free_pages(gfp_mask, order);
2791 return make_alloc_exact(addr, order, size);
2793 EXPORT_SYMBOL(alloc_pages_exact);
2796 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2798 * @nid: the preferred node ID where memory should be allocated
2799 * @size: the number of bytes to allocate
2800 * @gfp_mask: GFP flags for the allocation
2802 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2804 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2807 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2809 unsigned order = get_order(size);
2810 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2813 return make_alloc_exact((unsigned long)page_address(p), order, size);
2815 EXPORT_SYMBOL(alloc_pages_exact_nid);
2818 * free_pages_exact - release memory allocated via alloc_pages_exact()
2819 * @virt: the value returned by alloc_pages_exact.
2820 * @size: size of allocation, same value as passed to alloc_pages_exact().
2822 * Release the memory allocated by a previous call to alloc_pages_exact.
2824 void free_pages_exact(void *virt, size_t size)
2826 unsigned long addr = (unsigned long)virt;
2827 unsigned long end = addr + PAGE_ALIGN(size);
2829 while (addr < end) {
2834 EXPORT_SYMBOL(free_pages_exact);
2837 * nr_free_zone_pages - count number of pages beyond high watermark
2838 * @offset: The zone index of the highest zone
2840 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2841 * high watermark within all zones at or below a given zone index. For each
2842 * zone, the number of pages is calculated as:
2843 * present_pages - high_pages
2845 static unsigned long nr_free_zone_pages(int offset)
2850 /* Just pick one node, since fallback list is circular */
2851 unsigned long sum = 0;
2853 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2855 for_each_zone_zonelist(zone, z, zonelist, offset) {
2856 unsigned long size = zone->managed_pages;
2857 unsigned long high = high_wmark_pages(zone);
2866 * nr_free_buffer_pages - count number of pages beyond high watermark
2868 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2869 * watermark within ZONE_DMA and ZONE_NORMAL.
2871 unsigned long nr_free_buffer_pages(void)
2873 return nr_free_zone_pages(gfp_zone(GFP_USER));
2875 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2878 * nr_free_pagecache_pages - count number of pages beyond high watermark
2880 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2881 * high watermark within all zones.
2883 unsigned long nr_free_pagecache_pages(void)
2885 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2888 static inline void show_node(struct zone *zone)
2890 if (IS_ENABLED(CONFIG_NUMA))
2891 printk("Node %d ", zone_to_nid(zone));
2894 void si_meminfo(struct sysinfo *val)
2896 val->totalram = totalram_pages;
2898 val->freeram = global_page_state(NR_FREE_PAGES);
2899 val->bufferram = nr_blockdev_pages();
2900 val->totalhigh = totalhigh_pages;
2901 val->freehigh = nr_free_highpages();
2902 val->mem_unit = PAGE_SIZE;
2905 EXPORT_SYMBOL(si_meminfo);
2908 void si_meminfo_node(struct sysinfo *val, int nid)
2910 pg_data_t *pgdat = NODE_DATA(nid);
2912 val->totalram = pgdat->node_present_pages;
2913 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2914 #ifdef CONFIG_HIGHMEM
2915 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
2916 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2922 val->mem_unit = PAGE_SIZE;
2927 * Determine whether the node should be displayed or not, depending on whether
2928 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2930 bool skip_free_areas_node(unsigned int flags, int nid)
2933 unsigned int cpuset_mems_cookie;
2935 if (!(flags & SHOW_MEM_FILTER_NODES))
2939 cpuset_mems_cookie = get_mems_allowed();
2940 ret = !node_isset(nid, cpuset_current_mems_allowed);
2941 } while (!put_mems_allowed(cpuset_mems_cookie));
2946 #define K(x) ((x) << (PAGE_SHIFT-10))
2948 static void show_migration_types(unsigned char type)
2950 static const char types[MIGRATE_TYPES] = {
2951 [MIGRATE_UNMOVABLE] = 'U',
2952 [MIGRATE_RECLAIMABLE] = 'E',
2953 [MIGRATE_MOVABLE] = 'M',
2954 [MIGRATE_RESERVE] = 'R',
2956 [MIGRATE_CMA] = 'C',
2958 #ifdef CONFIG_MEMORY_ISOLATION
2959 [MIGRATE_ISOLATE] = 'I',
2962 char tmp[MIGRATE_TYPES + 1];
2966 for (i = 0; i < MIGRATE_TYPES; i++) {
2967 if (type & (1 << i))
2972 printk("(%s) ", tmp);
2976 * Show free area list (used inside shift_scroll-lock stuff)
2977 * We also calculate the percentage fragmentation. We do this by counting the
2978 * memory on each free list with the exception of the first item on the list.
2979 * Suppresses nodes that are not allowed by current's cpuset if
2980 * SHOW_MEM_FILTER_NODES is passed.
2982 void show_free_areas(unsigned int filter)
2987 for_each_populated_zone(zone) {
2988 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2991 printk("%s per-cpu:\n", zone->name);
2993 for_each_online_cpu(cpu) {
2994 struct per_cpu_pageset *pageset;
2996 pageset = per_cpu_ptr(zone->pageset, cpu);
2998 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2999 cpu, pageset->pcp.high,
3000 pageset->pcp.batch, pageset->pcp.count);
3004 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3005 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3007 " dirty:%lu writeback:%lu unstable:%lu\n"
3008 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3009 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3011 global_page_state(NR_ACTIVE_ANON),
3012 global_page_state(NR_INACTIVE_ANON),
3013 global_page_state(NR_ISOLATED_ANON),
3014 global_page_state(NR_ACTIVE_FILE),
3015 global_page_state(NR_INACTIVE_FILE),
3016 global_page_state(NR_ISOLATED_FILE),
3017 global_page_state(NR_UNEVICTABLE),
3018 global_page_state(NR_FILE_DIRTY),
3019 global_page_state(NR_WRITEBACK),
3020 global_page_state(NR_UNSTABLE_NFS),
3021 global_page_state(NR_FREE_PAGES),
3022 global_page_state(NR_SLAB_RECLAIMABLE),
3023 global_page_state(NR_SLAB_UNRECLAIMABLE),
3024 global_page_state(NR_FILE_MAPPED),
3025 global_page_state(NR_SHMEM),
3026 global_page_state(NR_PAGETABLE),
3027 global_page_state(NR_BOUNCE),
3028 global_page_state(NR_FREE_CMA_PAGES));
3030 for_each_populated_zone(zone) {
3033 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3041 " active_anon:%lukB"
3042 " inactive_anon:%lukB"
3043 " active_file:%lukB"
3044 " inactive_file:%lukB"
3045 " unevictable:%lukB"
3046 " isolated(anon):%lukB"
3047 " isolated(file):%lukB"
3055 " slab_reclaimable:%lukB"
3056 " slab_unreclaimable:%lukB"
3057 " kernel_stack:%lukB"
3062 " writeback_tmp:%lukB"
3063 " pages_scanned:%lu"
3064 " all_unreclaimable? %s"
3067 K(zone_page_state(zone, NR_FREE_PAGES)),
3068 K(min_wmark_pages(zone)),
3069 K(low_wmark_pages(zone)),
3070 K(high_wmark_pages(zone)),
3071 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3072 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3073 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3074 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3075 K(zone_page_state(zone, NR_UNEVICTABLE)),
3076 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3077 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3078 K(zone->present_pages),
3079 K(zone->managed_pages),
3080 K(zone_page_state(zone, NR_MLOCK)),
3081 K(zone_page_state(zone, NR_FILE_DIRTY)),
3082 K(zone_page_state(zone, NR_WRITEBACK)),
3083 K(zone_page_state(zone, NR_FILE_MAPPED)),
3084 K(zone_page_state(zone, NR_SHMEM)),
3085 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3086 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3087 zone_page_state(zone, NR_KERNEL_STACK) *
3089 K(zone_page_state(zone, NR_PAGETABLE)),
3090 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3091 K(zone_page_state(zone, NR_BOUNCE)),
3092 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3093 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3094 zone->pages_scanned,
3095 (zone->all_unreclaimable ? "yes" : "no")
3097 printk("lowmem_reserve[]:");
3098 for (i = 0; i < MAX_NR_ZONES; i++)
3099 printk(" %lu", zone->lowmem_reserve[i]);
3103 for_each_populated_zone(zone) {
3104 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3105 unsigned char types[MAX_ORDER];
3107 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3110 printk("%s: ", zone->name);
3112 spin_lock_irqsave(&zone->lock, flags);
3113 for (order = 0; order < MAX_ORDER; order++) {
3114 struct free_area *area = &zone->free_area[order];
3117 nr[order] = area->nr_free;
3118 total += nr[order] << order;
3121 for (type = 0; type < MIGRATE_TYPES; type++) {
3122 if (!list_empty(&area->free_list[type]))
3123 types[order] |= 1 << type;
3126 spin_unlock_irqrestore(&zone->lock, flags);
3127 for (order = 0; order < MAX_ORDER; order++) {
3128 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3130 show_migration_types(types[order]);
3132 printk("= %lukB\n", K(total));
3135 hugetlb_show_meminfo();
3137 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3139 show_swap_cache_info();
3142 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3144 zoneref->zone = zone;
3145 zoneref->zone_idx = zone_idx(zone);
3149 * Builds allocation fallback zone lists.
3151 * Add all populated zones of a node to the zonelist.
3153 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3154 int nr_zones, enum zone_type zone_type)
3158 BUG_ON(zone_type >= MAX_NR_ZONES);
3163 zone = pgdat->node_zones + zone_type;
3164 if (populated_zone(zone)) {
3165 zoneref_set_zone(zone,
3166 &zonelist->_zonerefs[nr_zones++]);
3167 check_highest_zone(zone_type);
3170 } while (zone_type);
3177 * 0 = automatic detection of better ordering.
3178 * 1 = order by ([node] distance, -zonetype)
3179 * 2 = order by (-zonetype, [node] distance)
3181 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3182 * the same zonelist. So only NUMA can configure this param.
3184 #define ZONELIST_ORDER_DEFAULT 0
3185 #define ZONELIST_ORDER_NODE 1
3186 #define ZONELIST_ORDER_ZONE 2
3188 /* zonelist order in the kernel.
3189 * set_zonelist_order() will set this to NODE or ZONE.
3191 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3192 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3196 /* The value user specified ....changed by config */
3197 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3198 /* string for sysctl */
3199 #define NUMA_ZONELIST_ORDER_LEN 16
3200 char numa_zonelist_order[16] = "default";
3203 * interface for configure zonelist ordering.
3204 * command line option "numa_zonelist_order"
3205 * = "[dD]efault - default, automatic configuration.
3206 * = "[nN]ode - order by node locality, then by zone within node
3207 * = "[zZ]one - order by zone, then by locality within zone
3210 static int __parse_numa_zonelist_order(char *s)
3212 if (*s == 'd' || *s == 'D') {
3213 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3214 } else if (*s == 'n' || *s == 'N') {
3215 user_zonelist_order = ZONELIST_ORDER_NODE;
3216 } else if (*s == 'z' || *s == 'Z') {
3217 user_zonelist_order = ZONELIST_ORDER_ZONE;
3220 "Ignoring invalid numa_zonelist_order value: "
3227 static __init int setup_numa_zonelist_order(char *s)
3234 ret = __parse_numa_zonelist_order(s);
3236 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3240 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3243 * sysctl handler for numa_zonelist_order
3245 int numa_zonelist_order_handler(ctl_table *table, int write,
3246 void __user *buffer, size_t *length,
3249 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3251 static DEFINE_MUTEX(zl_order_mutex);
3253 mutex_lock(&zl_order_mutex);
3255 strcpy(saved_string, (char*)table->data);
3256 ret = proc_dostring(table, write, buffer, length, ppos);
3260 int oldval = user_zonelist_order;
3261 if (__parse_numa_zonelist_order((char*)table->data)) {
3263 * bogus value. restore saved string
3265 strncpy((char*)table->data, saved_string,
3266 NUMA_ZONELIST_ORDER_LEN);
3267 user_zonelist_order = oldval;
3268 } else if (oldval != user_zonelist_order) {
3269 mutex_lock(&zonelists_mutex);
3270 build_all_zonelists(NULL, NULL);
3271 mutex_unlock(&zonelists_mutex);
3275 mutex_unlock(&zl_order_mutex);
3280 #define MAX_NODE_LOAD (nr_online_nodes)
3281 static int node_load[MAX_NUMNODES];
3284 * find_next_best_node - find the next node that should appear in a given node's fallback list
3285 * @node: node whose fallback list we're appending
3286 * @used_node_mask: nodemask_t of already used nodes
3288 * We use a number of factors to determine which is the next node that should
3289 * appear on a given node's fallback list. The node should not have appeared
3290 * already in @node's fallback list, and it should be the next closest node
3291 * according to the distance array (which contains arbitrary distance values
3292 * from each node to each node in the system), and should also prefer nodes
3293 * with no CPUs, since presumably they'll have very little allocation pressure
3294 * on them otherwise.
3295 * It returns -1 if no node is found.
3297 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3300 int min_val = INT_MAX;
3301 int best_node = NUMA_NO_NODE;
3302 const struct cpumask *tmp = cpumask_of_node(0);
3304 /* Use the local node if we haven't already */
3305 if (!node_isset(node, *used_node_mask)) {
3306 node_set(node, *used_node_mask);
3310 for_each_node_state(n, N_MEMORY) {
3312 /* Don't want a node to appear more than once */
3313 if (node_isset(n, *used_node_mask))
3316 /* Use the distance array to find the distance */
3317 val = node_distance(node, n);
3319 /* Penalize nodes under us ("prefer the next node") */
3322 /* Give preference to headless and unused nodes */
3323 tmp = cpumask_of_node(n);
3324 if (!cpumask_empty(tmp))
3325 val += PENALTY_FOR_NODE_WITH_CPUS;
3327 /* Slight preference for less loaded node */
3328 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3329 val += node_load[n];
3331 if (val < min_val) {
3338 node_set(best_node, *used_node_mask);
3345 * Build zonelists ordered by node and zones within node.
3346 * This results in maximum locality--normal zone overflows into local
3347 * DMA zone, if any--but risks exhausting DMA zone.
3349 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3352 struct zonelist *zonelist;
3354 zonelist = &pgdat->node_zonelists[0];
3355 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3357 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3359 zonelist->_zonerefs[j].zone = NULL;
3360 zonelist->_zonerefs[j].zone_idx = 0;
3364 * Build gfp_thisnode zonelists
3366 static void build_thisnode_zonelists(pg_data_t *pgdat)
3369 struct zonelist *zonelist;
3371 zonelist = &pgdat->node_zonelists[1];
3372 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3373 zonelist->_zonerefs[j].zone = NULL;
3374 zonelist->_zonerefs[j].zone_idx = 0;
3378 * Build zonelists ordered by zone and nodes within zones.
3379 * This results in conserving DMA zone[s] until all Normal memory is
3380 * exhausted, but results in overflowing to remote node while memory
3381 * may still exist in local DMA zone.
3383 static int node_order[MAX_NUMNODES];
3385 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3388 int zone_type; /* needs to be signed */
3390 struct zonelist *zonelist;
3392 zonelist = &pgdat->node_zonelists[0];
3394 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3395 for (j = 0; j < nr_nodes; j++) {
3396 node = node_order[j];
3397 z = &NODE_DATA(node)->node_zones[zone_type];
3398 if (populated_zone(z)) {
3400 &zonelist->_zonerefs[pos++]);
3401 check_highest_zone(zone_type);
3405 zonelist->_zonerefs[pos].zone = NULL;
3406 zonelist->_zonerefs[pos].zone_idx = 0;
3409 static int default_zonelist_order(void)
3412 unsigned long low_kmem_size,total_size;
3416 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3417 * If they are really small and used heavily, the system can fall
3418 * into OOM very easily.
3419 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3421 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3424 for_each_online_node(nid) {
3425 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3426 z = &NODE_DATA(nid)->node_zones[zone_type];
3427 if (populated_zone(z)) {
3428 if (zone_type < ZONE_NORMAL)
3429 low_kmem_size += z->present_pages;
3430 total_size += z->present_pages;
3431 } else if (zone_type == ZONE_NORMAL) {
3433 * If any node has only lowmem, then node order
3434 * is preferred to allow kernel allocations
3435 * locally; otherwise, they can easily infringe
3436 * on other nodes when there is an abundance of
3437 * lowmem available to allocate from.
3439 return ZONELIST_ORDER_NODE;
3443 if (!low_kmem_size || /* there are no DMA area. */
3444 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3445 return ZONELIST_ORDER_NODE;
3447 * look into each node's config.
3448 * If there is a node whose DMA/DMA32 memory is very big area on
3449 * local memory, NODE_ORDER may be suitable.
3451 average_size = total_size /
3452 (nodes_weight(node_states[N_MEMORY]) + 1);
3453 for_each_online_node(nid) {
3456 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3457 z = &NODE_DATA(nid)->node_zones[zone_type];
3458 if (populated_zone(z)) {
3459 if (zone_type < ZONE_NORMAL)
3460 low_kmem_size += z->present_pages;
3461 total_size += z->present_pages;
3464 if (low_kmem_size &&
3465 total_size > average_size && /* ignore small node */
3466 low_kmem_size > total_size * 70/100)
3467 return ZONELIST_ORDER_NODE;
3469 return ZONELIST_ORDER_ZONE;
3472 static void set_zonelist_order(void)
3474 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3475 current_zonelist_order = default_zonelist_order();
3477 current_zonelist_order = user_zonelist_order;
3480 static void build_zonelists(pg_data_t *pgdat)
3484 nodemask_t used_mask;
3485 int local_node, prev_node;
3486 struct zonelist *zonelist;
3487 int order = current_zonelist_order;
3489 /* initialize zonelists */
3490 for (i = 0; i < MAX_ZONELISTS; i++) {
3491 zonelist = pgdat->node_zonelists + i;
3492 zonelist->_zonerefs[0].zone = NULL;
3493 zonelist->_zonerefs[0].zone_idx = 0;
3496 /* NUMA-aware ordering of nodes */
3497 local_node = pgdat->node_id;
3498 load = nr_online_nodes;
3499 prev_node = local_node;
3500 nodes_clear(used_mask);
3502 memset(node_order, 0, sizeof(node_order));
3505 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3507 * We don't want to pressure a particular node.
3508 * So adding penalty to the first node in same
3509 * distance group to make it round-robin.
3511 if (node_distance(local_node, node) !=
3512 node_distance(local_node, prev_node))
3513 node_load[node] = load;
3517 if (order == ZONELIST_ORDER_NODE)
3518 build_zonelists_in_node_order(pgdat, node);
3520 node_order[j++] = node; /* remember order */
3523 if (order == ZONELIST_ORDER_ZONE) {
3524 /* calculate node order -- i.e., DMA last! */
3525 build_zonelists_in_zone_order(pgdat, j);
3528 build_thisnode_zonelists(pgdat);
3531 /* Construct the zonelist performance cache - see further mmzone.h */
3532 static void build_zonelist_cache(pg_data_t *pgdat)
3534 struct zonelist *zonelist;
3535 struct zonelist_cache *zlc;
3538 zonelist = &pgdat->node_zonelists[0];
3539 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3540 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3541 for (z = zonelist->_zonerefs; z->zone; z++)
3542 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3545 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3547 * Return node id of node used for "local" allocations.
3548 * I.e., first node id of first zone in arg node's generic zonelist.
3549 * Used for initializing percpu 'numa_mem', which is used primarily
3550 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3552 int local_memory_node(int node)
3556 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3557 gfp_zone(GFP_KERNEL),
3564 #else /* CONFIG_NUMA */
3566 static void set_zonelist_order(void)
3568 current_zonelist_order = ZONELIST_ORDER_ZONE;
3571 static void build_zonelists(pg_data_t *pgdat)
3573 int node, local_node;
3575 struct zonelist *zonelist;
3577 local_node = pgdat->node_id;
3579 zonelist = &pgdat->node_zonelists[0];
3580 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3583 * Now we build the zonelist so that it contains the zones
3584 * of all the other nodes.
3585 * We don't want to pressure a particular node, so when
3586 * building the zones for node N, we make sure that the
3587 * zones coming right after the local ones are those from
3588 * node N+1 (modulo N)
3590 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3591 if (!node_online(node))
3593 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3596 for (node = 0; node < local_node; node++) {
3597 if (!node_online(node))
3599 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3603 zonelist->_zonerefs[j].zone = NULL;
3604 zonelist->_zonerefs[j].zone_idx = 0;
3607 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3608 static void build_zonelist_cache(pg_data_t *pgdat)
3610 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3613 #endif /* CONFIG_NUMA */
3616 * Boot pageset table. One per cpu which is going to be used for all
3617 * zones and all nodes. The parameters will be set in such a way
3618 * that an item put on a list will immediately be handed over to
3619 * the buddy list. This is safe since pageset manipulation is done
3620 * with interrupts disabled.
3622 * The boot_pagesets must be kept even after bootup is complete for
3623 * unused processors and/or zones. They do play a role for bootstrapping
3624 * hotplugged processors.
3626 * zoneinfo_show() and maybe other functions do
3627 * not check if the processor is online before following the pageset pointer.
3628 * Other parts of the kernel may not check if the zone is available.
3630 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3631 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3632 static void setup_zone_pageset(struct zone *zone);
3635 * Global mutex to protect against size modification of zonelists
3636 * as well as to serialize pageset setup for the new populated zone.
3638 DEFINE_MUTEX(zonelists_mutex);
3640 /* return values int ....just for stop_machine() */
3641 static int __build_all_zonelists(void *data)
3645 pg_data_t *self = data;
3648 memset(node_load, 0, sizeof(node_load));
3651 if (self && !node_online(self->node_id)) {
3652 build_zonelists(self);
3653 build_zonelist_cache(self);
3656 for_each_online_node(nid) {
3657 pg_data_t *pgdat = NODE_DATA(nid);
3659 build_zonelists(pgdat);
3660 build_zonelist_cache(pgdat);
3664 * Initialize the boot_pagesets that are going to be used
3665 * for bootstrapping processors. The real pagesets for
3666 * each zone will be allocated later when the per cpu
3667 * allocator is available.
3669 * boot_pagesets are used also for bootstrapping offline
3670 * cpus if the system is already booted because the pagesets
3671 * are needed to initialize allocators on a specific cpu too.
3672 * F.e. the percpu allocator needs the page allocator which
3673 * needs the percpu allocator in order to allocate its pagesets
3674 * (a chicken-egg dilemma).
3676 for_each_possible_cpu(cpu) {
3677 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3679 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3681 * We now know the "local memory node" for each node--
3682 * i.e., the node of the first zone in the generic zonelist.
3683 * Set up numa_mem percpu variable for on-line cpus. During
3684 * boot, only the boot cpu should be on-line; we'll init the
3685 * secondary cpus' numa_mem as they come on-line. During
3686 * node/memory hotplug, we'll fixup all on-line cpus.
3688 if (cpu_online(cpu))
3689 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3697 * Called with zonelists_mutex held always
3698 * unless system_state == SYSTEM_BOOTING.
3700 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3702 set_zonelist_order();
3704 if (system_state == SYSTEM_BOOTING) {
3705 __build_all_zonelists(NULL);
3706 mminit_verify_zonelist();
3707 cpuset_init_current_mems_allowed();
3709 /* we have to stop all cpus to guarantee there is no user
3711 #ifdef CONFIG_MEMORY_HOTPLUG
3713 setup_zone_pageset(zone);
3715 stop_machine(__build_all_zonelists, pgdat, NULL);
3716 /* cpuset refresh routine should be here */
3718 vm_total_pages = nr_free_pagecache_pages();
3720 * Disable grouping by mobility if the number of pages in the
3721 * system is too low to allow the mechanism to work. It would be
3722 * more accurate, but expensive to check per-zone. This check is
3723 * made on memory-hotadd so a system can start with mobility
3724 * disabled and enable it later
3726 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3727 page_group_by_mobility_disabled = 1;
3729 page_group_by_mobility_disabled = 0;
3731 printk("Built %i zonelists in %s order, mobility grouping %s. "
3732 "Total pages: %ld\n",
3734 zonelist_order_name[current_zonelist_order],
3735 page_group_by_mobility_disabled ? "off" : "on",
3738 printk("Policy zone: %s\n", zone_names[policy_zone]);
3743 * Helper functions to size the waitqueue hash table.
3744 * Essentially these want to choose hash table sizes sufficiently
3745 * large so that collisions trying to wait on pages are rare.
3746 * But in fact, the number of active page waitqueues on typical
3747 * systems is ridiculously low, less than 200. So this is even
3748 * conservative, even though it seems large.
3750 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3751 * waitqueues, i.e. the size of the waitq table given the number of pages.
3753 #define PAGES_PER_WAITQUEUE 256
3755 #ifndef CONFIG_MEMORY_HOTPLUG
3756 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3758 unsigned long size = 1;
3760 pages /= PAGES_PER_WAITQUEUE;
3762 while (size < pages)
3766 * Once we have dozens or even hundreds of threads sleeping
3767 * on IO we've got bigger problems than wait queue collision.
3768 * Limit the size of the wait table to a reasonable size.
3770 size = min(size, 4096UL);
3772 return max(size, 4UL);
3776 * A zone's size might be changed by hot-add, so it is not possible to determine
3777 * a suitable size for its wait_table. So we use the maximum size now.
3779 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3781 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3782 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3783 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3785 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3786 * or more by the traditional way. (See above). It equals:
3788 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3789 * ia64(16K page size) : = ( 8G + 4M)byte.
3790 * powerpc (64K page size) : = (32G +16M)byte.
3792 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3799 * This is an integer logarithm so that shifts can be used later
3800 * to extract the more random high bits from the multiplicative
3801 * hash function before the remainder is taken.
3803 static inline unsigned long wait_table_bits(unsigned long size)
3808 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3811 * Check if a pageblock contains reserved pages
3813 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3817 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3818 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3825 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3826 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3827 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3828 * higher will lead to a bigger reserve which will get freed as contiguous
3829 * blocks as reclaim kicks in
3831 static void setup_zone_migrate_reserve(struct zone *zone)
3833 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3835 unsigned long block_migratetype;
3839 * Get the start pfn, end pfn and the number of blocks to reserve
3840 * We have to be careful to be aligned to pageblock_nr_pages to
3841 * make sure that we always check pfn_valid for the first page in
3844 start_pfn = zone->zone_start_pfn;
3845 end_pfn = zone_end_pfn(zone);
3846 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3847 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3851 * Reserve blocks are generally in place to help high-order atomic
3852 * allocations that are short-lived. A min_free_kbytes value that
3853 * would result in more than 2 reserve blocks for atomic allocations
3854 * is assumed to be in place to help anti-fragmentation for the
3855 * future allocation of hugepages at runtime.
3857 reserve = min(2, reserve);
3859 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3860 if (!pfn_valid(pfn))
3862 page = pfn_to_page(pfn);
3864 /* Watch out for overlapping nodes */
3865 if (page_to_nid(page) != zone_to_nid(zone))
3868 block_migratetype = get_pageblock_migratetype(page);
3870 /* Only test what is necessary when the reserves are not met */
3873 * Blocks with reserved pages will never free, skip
3876 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3877 if (pageblock_is_reserved(pfn, block_end_pfn))
3880 /* If this block is reserved, account for it */
3881 if (block_migratetype == MIGRATE_RESERVE) {
3886 /* Suitable for reserving if this block is movable */
3887 if (block_migratetype == MIGRATE_MOVABLE) {
3888 set_pageblock_migratetype(page,
3890 move_freepages_block(zone, page,
3898 * If the reserve is met and this is a previous reserved block,
3901 if (block_migratetype == MIGRATE_RESERVE) {
3902 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3903 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3909 * Initially all pages are reserved - free ones are freed
3910 * up by free_all_bootmem() once the early boot process is
3911 * done. Non-atomic initialization, single-pass.
3913 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3914 unsigned long start_pfn, enum memmap_context context)
3917 unsigned long end_pfn = start_pfn + size;
3921 if (highest_memmap_pfn < end_pfn - 1)
3922 highest_memmap_pfn = end_pfn - 1;
3924 z = &NODE_DATA(nid)->node_zones[zone];
3925 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3927 * There can be holes in boot-time mem_map[]s
3928 * handed to this function. They do not
3929 * exist on hotplugged memory.
3931 if (context == MEMMAP_EARLY) {
3932 if (!early_pfn_valid(pfn))
3934 if (!early_pfn_in_nid(pfn, nid))
3937 page = pfn_to_page(pfn);
3938 set_page_links(page, zone, nid, pfn);
3939 mminit_verify_page_links(page, zone, nid, pfn);
3940 init_page_count(page);
3941 page_mapcount_reset(page);
3942 page_nid_reset_last(page);
3943 SetPageReserved(page);
3945 * Mark the block movable so that blocks are reserved for
3946 * movable at startup. This will force kernel allocations
3947 * to reserve their blocks rather than leaking throughout
3948 * the address space during boot when many long-lived
3949 * kernel allocations are made. Later some blocks near
3950 * the start are marked MIGRATE_RESERVE by
3951 * setup_zone_migrate_reserve()
3953 * bitmap is created for zone's valid pfn range. but memmap
3954 * can be created for invalid pages (for alignment)
3955 * check here not to call set_pageblock_migratetype() against
3958 if ((z->zone_start_pfn <= pfn)
3959 && (pfn < zone_end_pfn(z))
3960 && !(pfn & (pageblock_nr_pages - 1)))
3961 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3963 INIT_LIST_HEAD(&page->lru);
3964 #ifdef WANT_PAGE_VIRTUAL
3965 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3966 if (!is_highmem_idx(zone))
3967 set_page_address(page, __va(pfn << PAGE_SHIFT));
3972 static void __meminit zone_init_free_lists(struct zone *zone)
3975 for_each_migratetype_order(order, t) {
3976 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3977 zone->free_area[order].nr_free = 0;
3981 #ifndef __HAVE_ARCH_MEMMAP_INIT
3982 #define memmap_init(size, nid, zone, start_pfn) \
3983 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3986 static int __meminit zone_batchsize(struct zone *zone)
3992 * The per-cpu-pages pools are set to around 1000th of the
3993 * size of the zone. But no more than 1/2 of a meg.
3995 * OK, so we don't know how big the cache is. So guess.
3997 batch = zone->managed_pages / 1024;
3998 if (batch * PAGE_SIZE > 512 * 1024)
3999 batch = (512 * 1024) / PAGE_SIZE;
4000 batch /= 4; /* We effectively *= 4 below */
4005 * Clamp the batch to a 2^n - 1 value. Having a power
4006 * of 2 value was found to be more likely to have
4007 * suboptimal cache aliasing properties in some cases.
4009 * For example if 2 tasks are alternately allocating
4010 * batches of pages, one task can end up with a lot
4011 * of pages of one half of the possible page colors
4012 * and the other with pages of the other colors.
4014 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4019 /* The deferral and batching of frees should be suppressed under NOMMU
4022 * The problem is that NOMMU needs to be able to allocate large chunks
4023 * of contiguous memory as there's no hardware page translation to
4024 * assemble apparent contiguous memory from discontiguous pages.
4026 * Queueing large contiguous runs of pages for batching, however,
4027 * causes the pages to actually be freed in smaller chunks. As there
4028 * can be a significant delay between the individual batches being
4029 * recycled, this leads to the once large chunks of space being
4030 * fragmented and becoming unavailable for high-order allocations.
4036 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4038 struct per_cpu_pages *pcp;
4041 memset(p, 0, sizeof(*p));
4045 pcp->high = 6 * batch;
4046 pcp->batch = max(1UL, 1 * batch);
4047 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4048 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4052 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4053 * to the value high for the pageset p.
4056 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
4059 struct per_cpu_pages *pcp;
4063 pcp->batch = max(1UL, high/4);
4064 if ((high/4) > (PAGE_SHIFT * 8))
4065 pcp->batch = PAGE_SHIFT * 8;
4068 static void __meminit setup_zone_pageset(struct zone *zone)
4072 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4074 for_each_possible_cpu(cpu) {
4075 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4077 setup_pageset(pcp, zone_batchsize(zone));
4079 if (percpu_pagelist_fraction)
4080 setup_pagelist_highmark(pcp,
4081 (zone->managed_pages /
4082 percpu_pagelist_fraction));
4087 * Allocate per cpu pagesets and initialize them.
4088 * Before this call only boot pagesets were available.
4090 void __init setup_per_cpu_pageset(void)
4094 for_each_populated_zone(zone)
4095 setup_zone_pageset(zone);
4098 static noinline __init_refok
4099 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4102 struct pglist_data *pgdat = zone->zone_pgdat;
4106 * The per-page waitqueue mechanism uses hashed waitqueues
4109 zone->wait_table_hash_nr_entries =
4110 wait_table_hash_nr_entries(zone_size_pages);
4111 zone->wait_table_bits =
4112 wait_table_bits(zone->wait_table_hash_nr_entries);
4113 alloc_size = zone->wait_table_hash_nr_entries
4114 * sizeof(wait_queue_head_t);
4116 if (!slab_is_available()) {
4117 zone->wait_table = (wait_queue_head_t *)
4118 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4121 * This case means that a zone whose size was 0 gets new memory
4122 * via memory hot-add.
4123 * But it may be the case that a new node was hot-added. In
4124 * this case vmalloc() will not be able to use this new node's
4125 * memory - this wait_table must be initialized to use this new
4126 * node itself as well.
4127 * To use this new node's memory, further consideration will be
4130 zone->wait_table = vmalloc(alloc_size);
4132 if (!zone->wait_table)
4135 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4136 init_waitqueue_head(zone->wait_table + i);
4141 static __meminit void zone_pcp_init(struct zone *zone)
4144 * per cpu subsystem is not up at this point. The following code
4145 * relies on the ability of the linker to provide the
4146 * offset of a (static) per cpu variable into the per cpu area.
4148 zone->pageset = &boot_pageset;
4150 if (zone->present_pages)
4151 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4152 zone->name, zone->present_pages,
4153 zone_batchsize(zone));
4156 int __meminit init_currently_empty_zone(struct zone *zone,
4157 unsigned long zone_start_pfn,
4159 enum memmap_context context)
4161 struct pglist_data *pgdat = zone->zone_pgdat;
4163 ret = zone_wait_table_init(zone, size);
4166 pgdat->nr_zones = zone_idx(zone) + 1;
4168 zone->zone_start_pfn = zone_start_pfn;
4170 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4171 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4173 (unsigned long)zone_idx(zone),
4174 zone_start_pfn, (zone_start_pfn + size));
4176 zone_init_free_lists(zone);
4181 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4182 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4184 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4185 * Architectures may implement their own version but if add_active_range()
4186 * was used and there are no special requirements, this is a convenient
4189 int __meminit __early_pfn_to_nid(unsigned long pfn)
4191 unsigned long start_pfn, end_pfn;
4194 * NOTE: The following SMP-unsafe globals are only used early in boot
4195 * when the kernel is running single-threaded.
4197 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4198 static int __meminitdata last_nid;
4200 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4203 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4204 if (start_pfn <= pfn && pfn < end_pfn) {
4205 last_start_pfn = start_pfn;
4206 last_end_pfn = end_pfn;
4210 /* This is a memory hole */
4213 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4215 int __meminit early_pfn_to_nid(unsigned long pfn)
4219 nid = __early_pfn_to_nid(pfn);
4222 /* just returns 0 */
4226 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4227 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4231 nid = __early_pfn_to_nid(pfn);
4232 if (nid >= 0 && nid != node)
4239 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4240 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4241 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4243 * If an architecture guarantees that all ranges registered with
4244 * add_active_ranges() contain no holes and may be freed, this
4245 * this function may be used instead of calling free_bootmem() manually.
4247 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4249 unsigned long start_pfn, end_pfn;
4252 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4253 start_pfn = min(start_pfn, max_low_pfn);
4254 end_pfn = min(end_pfn, max_low_pfn);
4256 if (start_pfn < end_pfn)
4257 free_bootmem_node(NODE_DATA(this_nid),
4258 PFN_PHYS(start_pfn),
4259 (end_pfn - start_pfn) << PAGE_SHIFT);
4264 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4265 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4267 * If an architecture guarantees that all ranges registered with
4268 * add_active_ranges() contain no holes and may be freed, this
4269 * function may be used instead of calling memory_present() manually.
4271 void __init sparse_memory_present_with_active_regions(int nid)
4273 unsigned long start_pfn, end_pfn;
4276 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4277 memory_present(this_nid, start_pfn, end_pfn);
4281 * get_pfn_range_for_nid - Return the start and end page frames for a node
4282 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4283 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4284 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4286 * It returns the start and end page frame of a node based on information
4287 * provided by an arch calling add_active_range(). If called for a node
4288 * with no available memory, a warning is printed and the start and end
4291 void __meminit get_pfn_range_for_nid(unsigned int nid,
4292 unsigned long *start_pfn, unsigned long *end_pfn)
4294 unsigned long this_start_pfn, this_end_pfn;
4300 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4301 *start_pfn = min(*start_pfn, this_start_pfn);
4302 *end_pfn = max(*end_pfn, this_end_pfn);
4305 if (*start_pfn == -1UL)
4310 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4311 * assumption is made that zones within a node are ordered in monotonic
4312 * increasing memory addresses so that the "highest" populated zone is used
4314 static void __init find_usable_zone_for_movable(void)
4317 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4318 if (zone_index == ZONE_MOVABLE)
4321 if (arch_zone_highest_possible_pfn[zone_index] >
4322 arch_zone_lowest_possible_pfn[zone_index])
4326 VM_BUG_ON(zone_index == -1);
4327 movable_zone = zone_index;
4331 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4332 * because it is sized independent of architecture. Unlike the other zones,
4333 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4334 * in each node depending on the size of each node and how evenly kernelcore
4335 * is distributed. This helper function adjusts the zone ranges
4336 * provided by the architecture for a given node by using the end of the
4337 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4338 * zones within a node are in order of monotonic increases memory addresses
4340 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4341 unsigned long zone_type,
4342 unsigned long node_start_pfn,
4343 unsigned long node_end_pfn,
4344 unsigned long *zone_start_pfn,
4345 unsigned long *zone_end_pfn)
4347 /* Only adjust if ZONE_MOVABLE is on this node */
4348 if (zone_movable_pfn[nid]) {
4349 /* Size ZONE_MOVABLE */
4350 if (zone_type == ZONE_MOVABLE) {
4351 *zone_start_pfn = zone_movable_pfn[nid];
4352 *zone_end_pfn = min(node_end_pfn,
4353 arch_zone_highest_possible_pfn[movable_zone]);
4355 /* Adjust for ZONE_MOVABLE starting within this range */
4356 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4357 *zone_end_pfn > zone_movable_pfn[nid]) {
4358 *zone_end_pfn = zone_movable_pfn[nid];
4360 /* Check if this whole range is within ZONE_MOVABLE */
4361 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4362 *zone_start_pfn = *zone_end_pfn;
4367 * Return the number of pages a zone spans in a node, including holes
4368 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4370 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4371 unsigned long zone_type,
4372 unsigned long *ignored)
4374 unsigned long node_start_pfn, node_end_pfn;
4375 unsigned long zone_start_pfn, zone_end_pfn;
4377 /* Get the start and end of the node and zone */
4378 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4379 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4380 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4381 adjust_zone_range_for_zone_movable(nid, zone_type,
4382 node_start_pfn, node_end_pfn,
4383 &zone_start_pfn, &zone_end_pfn);
4385 /* Check that this node has pages within the zone's required range */
4386 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4389 /* Move the zone boundaries inside the node if necessary */
4390 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4391 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4393 /* Return the spanned pages */
4394 return zone_end_pfn - zone_start_pfn;
4398 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4399 * then all holes in the requested range will be accounted for.
4401 unsigned long __meminit __absent_pages_in_range(int nid,
4402 unsigned long range_start_pfn,
4403 unsigned long range_end_pfn)
4405 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4406 unsigned long start_pfn, end_pfn;
4409 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4410 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4411 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4412 nr_absent -= end_pfn - start_pfn;
4418 * absent_pages_in_range - Return number of page frames in holes within a range
4419 * @start_pfn: The start PFN to start searching for holes
4420 * @end_pfn: The end PFN to stop searching for holes
4422 * It returns the number of pages frames in memory holes within a range.
4424 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4425 unsigned long end_pfn)
4427 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4430 /* Return the number of page frames in holes in a zone on a node */
4431 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4432 unsigned long zone_type,
4433 unsigned long *ignored)
4435 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4436 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4437 unsigned long node_start_pfn, node_end_pfn;
4438 unsigned long zone_start_pfn, zone_end_pfn;
4440 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4441 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4442 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4444 adjust_zone_range_for_zone_movable(nid, zone_type,
4445 node_start_pfn, node_end_pfn,
4446 &zone_start_pfn, &zone_end_pfn);
4447 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4450 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4451 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4452 unsigned long zone_type,
4453 unsigned long *zones_size)
4455 return zones_size[zone_type];
4458 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4459 unsigned long zone_type,
4460 unsigned long *zholes_size)
4465 return zholes_size[zone_type];
4468 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4470 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4471 unsigned long *zones_size, unsigned long *zholes_size)
4473 unsigned long realtotalpages, totalpages = 0;
4476 for (i = 0; i < MAX_NR_ZONES; i++)
4477 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4479 pgdat->node_spanned_pages = totalpages;
4481 realtotalpages = totalpages;
4482 for (i = 0; i < MAX_NR_ZONES; i++)
4484 zone_absent_pages_in_node(pgdat->node_id, i,
4486 pgdat->node_present_pages = realtotalpages;
4487 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4491 #ifndef CONFIG_SPARSEMEM
4493 * Calculate the size of the zone->blockflags rounded to an unsigned long
4494 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4495 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4496 * round what is now in bits to nearest long in bits, then return it in
4499 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4501 unsigned long usemapsize;
4503 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4504 usemapsize = roundup(zonesize, pageblock_nr_pages);
4505 usemapsize = usemapsize >> pageblock_order;
4506 usemapsize *= NR_PAGEBLOCK_BITS;
4507 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4509 return usemapsize / 8;
4512 static void __init setup_usemap(struct pglist_data *pgdat,
4514 unsigned long zone_start_pfn,
4515 unsigned long zonesize)
4517 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4518 zone->pageblock_flags = NULL;
4520 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4524 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4525 unsigned long zone_start_pfn, unsigned long zonesize) {}
4526 #endif /* CONFIG_SPARSEMEM */
4528 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4530 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4531 void __init set_pageblock_order(void)
4535 /* Check that pageblock_nr_pages has not already been setup */
4536 if (pageblock_order)
4539 if (HPAGE_SHIFT > PAGE_SHIFT)
4540 order = HUGETLB_PAGE_ORDER;
4542 order = MAX_ORDER - 1;
4545 * Assume the largest contiguous order of interest is a huge page.
4546 * This value may be variable depending on boot parameters on IA64 and
4549 pageblock_order = order;
4551 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4554 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4555 * is unused as pageblock_order is set at compile-time. See
4556 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4559 void __init set_pageblock_order(void)
4563 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4565 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4566 unsigned long present_pages)
4568 unsigned long pages = spanned_pages;
4571 * Provide a more accurate estimation if there are holes within
4572 * the zone and SPARSEMEM is in use. If there are holes within the
4573 * zone, each populated memory region may cost us one or two extra
4574 * memmap pages due to alignment because memmap pages for each
4575 * populated regions may not naturally algined on page boundary.
4576 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4578 if (spanned_pages > present_pages + (present_pages >> 4) &&
4579 IS_ENABLED(CONFIG_SPARSEMEM))
4580 pages = present_pages;
4582 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4586 * Set up the zone data structures:
4587 * - mark all pages reserved
4588 * - mark all memory queues empty
4589 * - clear the memory bitmaps
4591 * NOTE: pgdat should get zeroed by caller.
4593 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4594 unsigned long *zones_size, unsigned long *zholes_size)
4597 int nid = pgdat->node_id;
4598 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4601 pgdat_resize_init(pgdat);
4602 #ifdef CONFIG_NUMA_BALANCING
4603 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4604 pgdat->numabalancing_migrate_nr_pages = 0;
4605 pgdat->numabalancing_migrate_next_window = jiffies;
4607 init_waitqueue_head(&pgdat->kswapd_wait);
4608 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4609 pgdat_page_cgroup_init(pgdat);
4611 for (j = 0; j < MAX_NR_ZONES; j++) {
4612 struct zone *zone = pgdat->node_zones + j;
4613 unsigned long size, realsize, freesize, memmap_pages;
4615 size = zone_spanned_pages_in_node(nid, j, zones_size);
4616 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4620 * Adjust freesize so that it accounts for how much memory
4621 * is used by this zone for memmap. This affects the watermark
4622 * and per-cpu initialisations
4624 memmap_pages = calc_memmap_size(size, realsize);
4625 if (freesize >= memmap_pages) {
4626 freesize -= memmap_pages;
4629 " %s zone: %lu pages used for memmap\n",
4630 zone_names[j], memmap_pages);
4633 " %s zone: %lu pages exceeds freesize %lu\n",
4634 zone_names[j], memmap_pages, freesize);
4636 /* Account for reserved pages */
4637 if (j == 0 && freesize > dma_reserve) {
4638 freesize -= dma_reserve;
4639 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4640 zone_names[0], dma_reserve);
4643 if (!is_highmem_idx(j))
4644 nr_kernel_pages += freesize;
4645 /* Charge for highmem memmap if there are enough kernel pages */
4646 else if (nr_kernel_pages > memmap_pages * 2)
4647 nr_kernel_pages -= memmap_pages;
4648 nr_all_pages += freesize;
4650 zone->spanned_pages = size;
4651 zone->present_pages = realsize;
4653 * Set an approximate value for lowmem here, it will be adjusted
4654 * when the bootmem allocator frees pages into the buddy system.
4655 * And all highmem pages will be managed by the buddy system.
4657 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4660 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4662 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4664 zone->name = zone_names[j];
4665 spin_lock_init(&zone->lock);
4666 spin_lock_init(&zone->lru_lock);
4667 zone_seqlock_init(zone);
4668 zone->zone_pgdat = pgdat;
4670 zone_pcp_init(zone);
4671 lruvec_init(&zone->lruvec);
4675 set_pageblock_order();
4676 setup_usemap(pgdat, zone, zone_start_pfn, size);
4677 ret = init_currently_empty_zone(zone, zone_start_pfn,
4678 size, MEMMAP_EARLY);
4680 memmap_init(size, nid, j, zone_start_pfn);
4681 zone_start_pfn += size;
4685 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4687 /* Skip empty nodes */
4688 if (!pgdat->node_spanned_pages)
4691 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4692 /* ia64 gets its own node_mem_map, before this, without bootmem */
4693 if (!pgdat->node_mem_map) {
4694 unsigned long size, start, end;
4698 * The zone's endpoints aren't required to be MAX_ORDER
4699 * aligned but the node_mem_map endpoints must be in order
4700 * for the buddy allocator to function correctly.
4702 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4703 end = pgdat_end_pfn(pgdat);
4704 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4705 size = (end - start) * sizeof(struct page);
4706 map = alloc_remap(pgdat->node_id, size);
4708 map = alloc_bootmem_node_nopanic(pgdat, size);
4709 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4711 #ifndef CONFIG_NEED_MULTIPLE_NODES
4713 * With no DISCONTIG, the global mem_map is just set as node 0's
4715 if (pgdat == NODE_DATA(0)) {
4716 mem_map = NODE_DATA(0)->node_mem_map;
4717 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4718 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4719 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4720 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4723 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4726 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4727 unsigned long node_start_pfn, unsigned long *zholes_size)
4729 pg_data_t *pgdat = NODE_DATA(nid);
4731 /* pg_data_t should be reset to zero when it's allocated */
4732 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4734 pgdat->node_id = nid;
4735 pgdat->node_start_pfn = node_start_pfn;
4736 init_zone_allows_reclaim(nid);
4737 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4739 alloc_node_mem_map(pgdat);
4740 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4741 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4742 nid, (unsigned long)pgdat,
4743 (unsigned long)pgdat->node_mem_map);
4746 free_area_init_core(pgdat, zones_size, zholes_size);
4749 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4751 #if MAX_NUMNODES > 1
4753 * Figure out the number of possible node ids.
4755 void __init setup_nr_node_ids(void)
4758 unsigned int highest = 0;
4760 for_each_node_mask(node, node_possible_map)
4762 nr_node_ids = highest + 1;
4767 * node_map_pfn_alignment - determine the maximum internode alignment
4769 * This function should be called after node map is populated and sorted.
4770 * It calculates the maximum power of two alignment which can distinguish
4773 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4774 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4775 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4776 * shifted, 1GiB is enough and this function will indicate so.
4778 * This is used to test whether pfn -> nid mapping of the chosen memory
4779 * model has fine enough granularity to avoid incorrect mapping for the
4780 * populated node map.
4782 * Returns the determined alignment in pfn's. 0 if there is no alignment
4783 * requirement (single node).
4785 unsigned long __init node_map_pfn_alignment(void)
4787 unsigned long accl_mask = 0, last_end = 0;
4788 unsigned long start, end, mask;
4792 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4793 if (!start || last_nid < 0 || last_nid == nid) {
4800 * Start with a mask granular enough to pin-point to the
4801 * start pfn and tick off bits one-by-one until it becomes
4802 * too coarse to separate the current node from the last.
4804 mask = ~((1 << __ffs(start)) - 1);
4805 while (mask && last_end <= (start & (mask << 1)))
4808 /* accumulate all internode masks */
4812 /* convert mask to number of pages */
4813 return ~accl_mask + 1;
4816 /* Find the lowest pfn for a node */
4817 static unsigned long __init find_min_pfn_for_node(int nid)
4819 unsigned long min_pfn = ULONG_MAX;
4820 unsigned long start_pfn;
4823 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4824 min_pfn = min(min_pfn, start_pfn);
4826 if (min_pfn == ULONG_MAX) {
4828 "Could not find start_pfn for node %d\n", nid);
4836 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4838 * It returns the minimum PFN based on information provided via
4839 * add_active_range().
4841 unsigned long __init find_min_pfn_with_active_regions(void)
4843 return find_min_pfn_for_node(MAX_NUMNODES);
4847 * early_calculate_totalpages()
4848 * Sum pages in active regions for movable zone.
4849 * Populate N_MEMORY for calculating usable_nodes.
4851 static unsigned long __init early_calculate_totalpages(void)
4853 unsigned long totalpages = 0;
4854 unsigned long start_pfn, end_pfn;
4857 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4858 unsigned long pages = end_pfn - start_pfn;
4860 totalpages += pages;
4862 node_set_state(nid, N_MEMORY);
4868 * Find the PFN the Movable zone begins in each node. Kernel memory
4869 * is spread evenly between nodes as long as the nodes have enough
4870 * memory. When they don't, some nodes will have more kernelcore than
4873 static void __init find_zone_movable_pfns_for_nodes(void)
4876 unsigned long usable_startpfn;
4877 unsigned long kernelcore_node, kernelcore_remaining;
4878 /* save the state before borrow the nodemask */
4879 nodemask_t saved_node_state = node_states[N_MEMORY];
4880 unsigned long totalpages = early_calculate_totalpages();
4881 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4884 * If movablecore was specified, calculate what size of
4885 * kernelcore that corresponds so that memory usable for
4886 * any allocation type is evenly spread. If both kernelcore
4887 * and movablecore are specified, then the value of kernelcore
4888 * will be used for required_kernelcore if it's greater than
4889 * what movablecore would have allowed.
4891 if (required_movablecore) {
4892 unsigned long corepages;
4895 * Round-up so that ZONE_MOVABLE is at least as large as what
4896 * was requested by the user
4898 required_movablecore =
4899 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4900 corepages = totalpages - required_movablecore;
4902 required_kernelcore = max(required_kernelcore, corepages);
4905 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4906 if (!required_kernelcore)
4909 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4910 find_usable_zone_for_movable();
4911 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4914 /* Spread kernelcore memory as evenly as possible throughout nodes */
4915 kernelcore_node = required_kernelcore / usable_nodes;
4916 for_each_node_state(nid, N_MEMORY) {
4917 unsigned long start_pfn, end_pfn;
4920 * Recalculate kernelcore_node if the division per node
4921 * now exceeds what is necessary to satisfy the requested
4922 * amount of memory for the kernel
4924 if (required_kernelcore < kernelcore_node)
4925 kernelcore_node = required_kernelcore / usable_nodes;
4928 * As the map is walked, we track how much memory is usable
4929 * by the kernel using kernelcore_remaining. When it is
4930 * 0, the rest of the node is usable by ZONE_MOVABLE
4932 kernelcore_remaining = kernelcore_node;
4934 /* Go through each range of PFNs within this node */
4935 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4936 unsigned long size_pages;
4938 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4939 if (start_pfn >= end_pfn)
4942 /* Account for what is only usable for kernelcore */
4943 if (start_pfn < usable_startpfn) {
4944 unsigned long kernel_pages;
4945 kernel_pages = min(end_pfn, usable_startpfn)
4948 kernelcore_remaining -= min(kernel_pages,
4949 kernelcore_remaining);
4950 required_kernelcore -= min(kernel_pages,
4951 required_kernelcore);
4953 /* Continue if range is now fully accounted */
4954 if (end_pfn <= usable_startpfn) {
4957 * Push zone_movable_pfn to the end so
4958 * that if we have to rebalance
4959 * kernelcore across nodes, we will
4960 * not double account here
4962 zone_movable_pfn[nid] = end_pfn;
4965 start_pfn = usable_startpfn;
4969 * The usable PFN range for ZONE_MOVABLE is from
4970 * start_pfn->end_pfn. Calculate size_pages as the
4971 * number of pages used as kernelcore
4973 size_pages = end_pfn - start_pfn;
4974 if (size_pages > kernelcore_remaining)
4975 size_pages = kernelcore_remaining;
4976 zone_movable_pfn[nid] = start_pfn + size_pages;
4979 * Some kernelcore has been met, update counts and
4980 * break if the kernelcore for this node has been
4983 required_kernelcore -= min(required_kernelcore,
4985 kernelcore_remaining -= size_pages;
4986 if (!kernelcore_remaining)
4992 * If there is still required_kernelcore, we do another pass with one
4993 * less node in the count. This will push zone_movable_pfn[nid] further
4994 * along on the nodes that still have memory until kernelcore is
4998 if (usable_nodes && required_kernelcore > usable_nodes)
5001 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5002 for (nid = 0; nid < MAX_NUMNODES; nid++)
5003 zone_movable_pfn[nid] =
5004 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5007 /* restore the node_state */
5008 node_states[N_MEMORY] = saved_node_state;
5011 /* Any regular or high memory on that node ? */
5012 static void check_for_memory(pg_data_t *pgdat, int nid)
5014 enum zone_type zone_type;
5016 if (N_MEMORY == N_NORMAL_MEMORY)
5019 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5020 struct zone *zone = &pgdat->node_zones[zone_type];
5021 if (zone->present_pages) {
5022 node_set_state(nid, N_HIGH_MEMORY);
5023 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5024 zone_type <= ZONE_NORMAL)
5025 node_set_state(nid, N_NORMAL_MEMORY);
5032 * free_area_init_nodes - Initialise all pg_data_t and zone data
5033 * @max_zone_pfn: an array of max PFNs for each zone
5035 * This will call free_area_init_node() for each active node in the system.
5036 * Using the page ranges provided by add_active_range(), the size of each
5037 * zone in each node and their holes is calculated. If the maximum PFN
5038 * between two adjacent zones match, it is assumed that the zone is empty.
5039 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5040 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5041 * starts where the previous one ended. For example, ZONE_DMA32 starts
5042 * at arch_max_dma_pfn.
5044 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5046 unsigned long start_pfn, end_pfn;
5049 /* Record where the zone boundaries are */
5050 memset(arch_zone_lowest_possible_pfn, 0,
5051 sizeof(arch_zone_lowest_possible_pfn));
5052 memset(arch_zone_highest_possible_pfn, 0,
5053 sizeof(arch_zone_highest_possible_pfn));
5054 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5055 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5056 for (i = 1; i < MAX_NR_ZONES; i++) {
5057 if (i == ZONE_MOVABLE)
5059 arch_zone_lowest_possible_pfn[i] =
5060 arch_zone_highest_possible_pfn[i-1];
5061 arch_zone_highest_possible_pfn[i] =
5062 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5064 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5065 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5067 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5068 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5069 find_zone_movable_pfns_for_nodes();
5071 /* Print out the zone ranges */
5072 printk("Zone ranges:\n");
5073 for (i = 0; i < MAX_NR_ZONES; i++) {
5074 if (i == ZONE_MOVABLE)
5076 printk(KERN_CONT " %-8s ", zone_names[i]);
5077 if (arch_zone_lowest_possible_pfn[i] ==
5078 arch_zone_highest_possible_pfn[i])
5079 printk(KERN_CONT "empty\n");
5081 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5082 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5083 (arch_zone_highest_possible_pfn[i]
5084 << PAGE_SHIFT) - 1);
5087 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5088 printk("Movable zone start for each node\n");
5089 for (i = 0; i < MAX_NUMNODES; i++) {
5090 if (zone_movable_pfn[i])
5091 printk(" Node %d: %#010lx\n", i,
5092 zone_movable_pfn[i] << PAGE_SHIFT);
5095 /* Print out the early node map */
5096 printk("Early memory node ranges\n");
5097 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5098 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5099 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5101 /* Initialise every node */
5102 mminit_verify_pageflags_layout();
5103 setup_nr_node_ids();
5104 for_each_online_node(nid) {
5105 pg_data_t *pgdat = NODE_DATA(nid);
5106 free_area_init_node(nid, NULL,
5107 find_min_pfn_for_node(nid), NULL);
5109 /* Any memory on that node */
5110 if (pgdat->node_present_pages)
5111 node_set_state(nid, N_MEMORY);
5112 check_for_memory(pgdat, nid);
5116 static int __init cmdline_parse_core(char *p, unsigned long *core)
5118 unsigned long long coremem;
5122 coremem = memparse(p, &p);
5123 *core = coremem >> PAGE_SHIFT;
5125 /* Paranoid check that UL is enough for the coremem value */
5126 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5132 * kernelcore=size sets the amount of memory for use for allocations that
5133 * cannot be reclaimed or migrated.
5135 static int __init cmdline_parse_kernelcore(char *p)
5137 return cmdline_parse_core(p, &required_kernelcore);
5141 * movablecore=size sets the amount of memory for use for allocations that
5142 * can be reclaimed or migrated.
5144 static int __init cmdline_parse_movablecore(char *p)
5146 return cmdline_parse_core(p, &required_movablecore);
5149 early_param("kernelcore", cmdline_parse_kernelcore);
5150 early_param("movablecore", cmdline_parse_movablecore);
5152 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5154 unsigned long free_reserved_area(unsigned long start, unsigned long end,
5155 int poison, char *s)
5157 unsigned long pages, pos;
5159 pos = start = PAGE_ALIGN(start);
5161 for (pages = 0; pos < end; pos += PAGE_SIZE, pages++) {
5163 memset((void *)pos, poison, PAGE_SIZE);
5164 free_reserved_page(virt_to_page((void *)pos));
5168 pr_info("Freeing %s memory: %ldK (%lx - %lx)\n",
5169 s, pages << (PAGE_SHIFT - 10), start, end);
5174 #ifdef CONFIG_HIGHMEM
5175 void free_highmem_page(struct page *page)
5177 __free_reserved_page(page);
5184 * set_dma_reserve - set the specified number of pages reserved in the first zone
5185 * @new_dma_reserve: The number of pages to mark reserved
5187 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5188 * In the DMA zone, a significant percentage may be consumed by kernel image
5189 * and other unfreeable allocations which can skew the watermarks badly. This
5190 * function may optionally be used to account for unfreeable pages in the
5191 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5192 * smaller per-cpu batchsize.
5194 void __init set_dma_reserve(unsigned long new_dma_reserve)
5196 dma_reserve = new_dma_reserve;
5199 void __init free_area_init(unsigned long *zones_size)
5201 free_area_init_node(0, zones_size,
5202 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5205 static int page_alloc_cpu_notify(struct notifier_block *self,
5206 unsigned long action, void *hcpu)
5208 int cpu = (unsigned long)hcpu;
5210 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5211 lru_add_drain_cpu(cpu);
5215 * Spill the event counters of the dead processor
5216 * into the current processors event counters.
5217 * This artificially elevates the count of the current
5220 vm_events_fold_cpu(cpu);
5223 * Zero the differential counters of the dead processor
5224 * so that the vm statistics are consistent.
5226 * This is only okay since the processor is dead and cannot
5227 * race with what we are doing.
5229 refresh_cpu_vm_stats(cpu);
5234 void __init page_alloc_init(void)
5236 hotcpu_notifier(page_alloc_cpu_notify, 0);
5240 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5241 * or min_free_kbytes changes.
5243 static void calculate_totalreserve_pages(void)
5245 struct pglist_data *pgdat;
5246 unsigned long reserve_pages = 0;
5247 enum zone_type i, j;
5249 for_each_online_pgdat(pgdat) {
5250 for (i = 0; i < MAX_NR_ZONES; i++) {
5251 struct zone *zone = pgdat->node_zones + i;
5252 unsigned long max = 0;
5254 /* Find valid and maximum lowmem_reserve in the zone */
5255 for (j = i; j < MAX_NR_ZONES; j++) {
5256 if (zone->lowmem_reserve[j] > max)
5257 max = zone->lowmem_reserve[j];
5260 /* we treat the high watermark as reserved pages. */
5261 max += high_wmark_pages(zone);
5263 if (max > zone->managed_pages)
5264 max = zone->managed_pages;
5265 reserve_pages += max;
5267 * Lowmem reserves are not available to
5268 * GFP_HIGHUSER page cache allocations and
5269 * kswapd tries to balance zones to their high
5270 * watermark. As a result, neither should be
5271 * regarded as dirtyable memory, to prevent a
5272 * situation where reclaim has to clean pages
5273 * in order to balance the zones.
5275 zone->dirty_balance_reserve = max;
5278 dirty_balance_reserve = reserve_pages;
5279 totalreserve_pages = reserve_pages;
5283 * setup_per_zone_lowmem_reserve - called whenever
5284 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5285 * has a correct pages reserved value, so an adequate number of
5286 * pages are left in the zone after a successful __alloc_pages().
5288 static void setup_per_zone_lowmem_reserve(void)
5290 struct pglist_data *pgdat;
5291 enum zone_type j, idx;
5293 for_each_online_pgdat(pgdat) {
5294 for (j = 0; j < MAX_NR_ZONES; j++) {
5295 struct zone *zone = pgdat->node_zones + j;
5296 unsigned long managed_pages = zone->managed_pages;
5298 zone->lowmem_reserve[j] = 0;
5302 struct zone *lower_zone;
5306 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5307 sysctl_lowmem_reserve_ratio[idx] = 1;
5309 lower_zone = pgdat->node_zones + idx;
5310 lower_zone->lowmem_reserve[j] = managed_pages /
5311 sysctl_lowmem_reserve_ratio[idx];
5312 managed_pages += lower_zone->managed_pages;
5317 /* update totalreserve_pages */
5318 calculate_totalreserve_pages();
5321 static void __setup_per_zone_wmarks(void)
5323 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5324 unsigned long lowmem_pages = 0;
5326 unsigned long flags;
5328 /* Calculate total number of !ZONE_HIGHMEM pages */
5329 for_each_zone(zone) {
5330 if (!is_highmem(zone))
5331 lowmem_pages += zone->managed_pages;
5334 for_each_zone(zone) {
5337 spin_lock_irqsave(&zone->lock, flags);
5338 tmp = (u64)pages_min * zone->managed_pages;
5339 do_div(tmp, lowmem_pages);
5340 if (is_highmem(zone)) {
5342 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5343 * need highmem pages, so cap pages_min to a small
5346 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5347 * deltas controls asynch page reclaim, and so should
5348 * not be capped for highmem.
5350 unsigned long min_pages;
5352 min_pages = zone->managed_pages / 1024;
5353 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5354 zone->watermark[WMARK_MIN] = min_pages;
5357 * If it's a lowmem zone, reserve a number of pages
5358 * proportionate to the zone's size.
5360 zone->watermark[WMARK_MIN] = tmp;
5363 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5364 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5366 setup_zone_migrate_reserve(zone);
5367 spin_unlock_irqrestore(&zone->lock, flags);
5370 /* update totalreserve_pages */
5371 calculate_totalreserve_pages();
5375 * setup_per_zone_wmarks - called when min_free_kbytes changes
5376 * or when memory is hot-{added|removed}
5378 * Ensures that the watermark[min,low,high] values for each zone are set
5379 * correctly with respect to min_free_kbytes.
5381 void setup_per_zone_wmarks(void)
5383 mutex_lock(&zonelists_mutex);
5384 __setup_per_zone_wmarks();
5385 mutex_unlock(&zonelists_mutex);
5389 * The inactive anon list should be small enough that the VM never has to
5390 * do too much work, but large enough that each inactive page has a chance
5391 * to be referenced again before it is swapped out.
5393 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5394 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5395 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5396 * the anonymous pages are kept on the inactive list.
5399 * memory ratio inactive anon
5400 * -------------------------------------
5409 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5411 unsigned int gb, ratio;
5413 /* Zone size in gigabytes */
5414 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5416 ratio = int_sqrt(10 * gb);
5420 zone->inactive_ratio = ratio;
5423 static void __meminit setup_per_zone_inactive_ratio(void)
5428 calculate_zone_inactive_ratio(zone);
5432 * Initialise min_free_kbytes.
5434 * For small machines we want it small (128k min). For large machines
5435 * we want it large (64MB max). But it is not linear, because network
5436 * bandwidth does not increase linearly with machine size. We use
5438 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5439 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5455 int __meminit init_per_zone_wmark_min(void)
5457 unsigned long lowmem_kbytes;
5459 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5461 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5462 if (min_free_kbytes < 128)
5463 min_free_kbytes = 128;
5464 if (min_free_kbytes > 65536)
5465 min_free_kbytes = 65536;
5466 setup_per_zone_wmarks();
5467 refresh_zone_stat_thresholds();
5468 setup_per_zone_lowmem_reserve();
5469 setup_per_zone_inactive_ratio();
5472 module_init(init_per_zone_wmark_min)
5475 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5476 * that we can call two helper functions whenever min_free_kbytes
5479 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5480 void __user *buffer, size_t *length, loff_t *ppos)
5482 proc_dointvec(table, write, buffer, length, ppos);
5484 setup_per_zone_wmarks();
5489 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5490 void __user *buffer, size_t *length, loff_t *ppos)
5495 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5500 zone->min_unmapped_pages = (zone->managed_pages *
5501 sysctl_min_unmapped_ratio) / 100;
5505 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5506 void __user *buffer, size_t *length, loff_t *ppos)
5511 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5516 zone->min_slab_pages = (zone->managed_pages *
5517 sysctl_min_slab_ratio) / 100;
5523 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5524 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5525 * whenever sysctl_lowmem_reserve_ratio changes.
5527 * The reserve ratio obviously has absolutely no relation with the
5528 * minimum watermarks. The lowmem reserve ratio can only make sense
5529 * if in function of the boot time zone sizes.
5531 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5532 void __user *buffer, size_t *length, loff_t *ppos)
5534 proc_dointvec_minmax(table, write, buffer, length, ppos);
5535 setup_per_zone_lowmem_reserve();
5540 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5541 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5542 * can have before it gets flushed back to buddy allocator.
5545 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5546 void __user *buffer, size_t *length, loff_t *ppos)
5552 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5553 if (!write || (ret < 0))
5555 for_each_populated_zone(zone) {
5556 for_each_possible_cpu(cpu) {
5558 high = zone->managed_pages / percpu_pagelist_fraction;
5559 setup_pagelist_highmark(
5560 per_cpu_ptr(zone->pageset, cpu), high);
5566 int hashdist = HASHDIST_DEFAULT;
5569 static int __init set_hashdist(char *str)
5573 hashdist = simple_strtoul(str, &str, 0);
5576 __setup("hashdist=", set_hashdist);
5580 * allocate a large system hash table from bootmem
5581 * - it is assumed that the hash table must contain an exact power-of-2
5582 * quantity of entries
5583 * - limit is the number of hash buckets, not the total allocation size
5585 void *__init alloc_large_system_hash(const char *tablename,
5586 unsigned long bucketsize,
5587 unsigned long numentries,
5590 unsigned int *_hash_shift,
5591 unsigned int *_hash_mask,
5592 unsigned long low_limit,
5593 unsigned long high_limit)
5595 unsigned long long max = high_limit;
5596 unsigned long log2qty, size;
5599 /* allow the kernel cmdline to have a say */
5601 /* round applicable memory size up to nearest megabyte */
5602 numentries = nr_kernel_pages;
5603 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5604 numentries >>= 20 - PAGE_SHIFT;
5605 numentries <<= 20 - PAGE_SHIFT;
5607 /* limit to 1 bucket per 2^scale bytes of low memory */
5608 if (scale > PAGE_SHIFT)
5609 numentries >>= (scale - PAGE_SHIFT);
5611 numentries <<= (PAGE_SHIFT - scale);
5613 /* Make sure we've got at least a 0-order allocation.. */
5614 if (unlikely(flags & HASH_SMALL)) {
5615 /* Makes no sense without HASH_EARLY */
5616 WARN_ON(!(flags & HASH_EARLY));
5617 if (!(numentries >> *_hash_shift)) {
5618 numentries = 1UL << *_hash_shift;
5619 BUG_ON(!numentries);
5621 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5622 numentries = PAGE_SIZE / bucketsize;
5624 numentries = roundup_pow_of_two(numentries);
5626 /* limit allocation size to 1/16 total memory by default */
5628 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5629 do_div(max, bucketsize);
5631 max = min(max, 0x80000000ULL);
5633 if (numentries < low_limit)
5634 numentries = low_limit;
5635 if (numentries > max)
5638 log2qty = ilog2(numentries);
5641 size = bucketsize << log2qty;
5642 if (flags & HASH_EARLY)
5643 table = alloc_bootmem_nopanic(size);
5645 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5648 * If bucketsize is not a power-of-two, we may free
5649 * some pages at the end of hash table which
5650 * alloc_pages_exact() automatically does
5652 if (get_order(size) < MAX_ORDER) {
5653 table = alloc_pages_exact(size, GFP_ATOMIC);
5654 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5657 } while (!table && size > PAGE_SIZE && --log2qty);
5660 panic("Failed to allocate %s hash table\n", tablename);
5662 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5665 ilog2(size) - PAGE_SHIFT,
5669 *_hash_shift = log2qty;
5671 *_hash_mask = (1 << log2qty) - 1;
5676 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5677 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5680 #ifdef CONFIG_SPARSEMEM
5681 return __pfn_to_section(pfn)->pageblock_flags;
5683 return zone->pageblock_flags;
5684 #endif /* CONFIG_SPARSEMEM */
5687 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5689 #ifdef CONFIG_SPARSEMEM
5690 pfn &= (PAGES_PER_SECTION-1);
5691 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5693 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5694 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5695 #endif /* CONFIG_SPARSEMEM */
5699 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5700 * @page: The page within the block of interest
5701 * @start_bitidx: The first bit of interest to retrieve
5702 * @end_bitidx: The last bit of interest
5703 * returns pageblock_bits flags
5705 unsigned long get_pageblock_flags_group(struct page *page,
5706 int start_bitidx, int end_bitidx)
5709 unsigned long *bitmap;
5710 unsigned long pfn, bitidx;
5711 unsigned long flags = 0;
5712 unsigned long value = 1;
5714 zone = page_zone(page);
5715 pfn = page_to_pfn(page);
5716 bitmap = get_pageblock_bitmap(zone, pfn);
5717 bitidx = pfn_to_bitidx(zone, pfn);
5719 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5720 if (test_bit(bitidx + start_bitidx, bitmap))
5727 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5728 * @page: The page within the block of interest
5729 * @start_bitidx: The first bit of interest
5730 * @end_bitidx: The last bit of interest
5731 * @flags: The flags to set
5733 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5734 int start_bitidx, int end_bitidx)
5737 unsigned long *bitmap;
5738 unsigned long pfn, bitidx;
5739 unsigned long value = 1;
5741 zone = page_zone(page);
5742 pfn = page_to_pfn(page);
5743 bitmap = get_pageblock_bitmap(zone, pfn);
5744 bitidx = pfn_to_bitidx(zone, pfn);
5745 VM_BUG_ON(!zone_spans_pfn(zone, pfn));
5747 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5749 __set_bit(bitidx + start_bitidx, bitmap);
5751 __clear_bit(bitidx + start_bitidx, bitmap);
5755 * This function checks whether pageblock includes unmovable pages or not.
5756 * If @count is not zero, it is okay to include less @count unmovable pages
5758 * PageLRU check wihtout isolation or lru_lock could race so that
5759 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5760 * expect this function should be exact.
5762 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5763 bool skip_hwpoisoned_pages)
5765 unsigned long pfn, iter, found;
5769 * For avoiding noise data, lru_add_drain_all() should be called
5770 * If ZONE_MOVABLE, the zone never contains unmovable pages
5772 if (zone_idx(zone) == ZONE_MOVABLE)
5774 mt = get_pageblock_migratetype(page);
5775 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5778 pfn = page_to_pfn(page);
5779 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5780 unsigned long check = pfn + iter;
5782 if (!pfn_valid_within(check))
5785 page = pfn_to_page(check);
5787 * We can't use page_count without pin a page
5788 * because another CPU can free compound page.
5789 * This check already skips compound tails of THP
5790 * because their page->_count is zero at all time.
5792 if (!atomic_read(&page->_count)) {
5793 if (PageBuddy(page))
5794 iter += (1 << page_order(page)) - 1;
5799 * The HWPoisoned page may be not in buddy system, and
5800 * page_count() is not 0.
5802 if (skip_hwpoisoned_pages && PageHWPoison(page))
5808 * If there are RECLAIMABLE pages, we need to check it.
5809 * But now, memory offline itself doesn't call shrink_slab()
5810 * and it still to be fixed.
5813 * If the page is not RAM, page_count()should be 0.
5814 * we don't need more check. This is an _used_ not-movable page.
5816 * The problematic thing here is PG_reserved pages. PG_reserved
5817 * is set to both of a memory hole page and a _used_ kernel
5826 bool is_pageblock_removable_nolock(struct page *page)
5832 * We have to be careful here because we are iterating over memory
5833 * sections which are not zone aware so we might end up outside of
5834 * the zone but still within the section.
5835 * We have to take care about the node as well. If the node is offline
5836 * its NODE_DATA will be NULL - see page_zone.
5838 if (!node_online(page_to_nid(page)))
5841 zone = page_zone(page);
5842 pfn = page_to_pfn(page);
5843 if (!zone_spans_pfn(zone, pfn))
5846 return !has_unmovable_pages(zone, page, 0, true);
5851 static unsigned long pfn_max_align_down(unsigned long pfn)
5853 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5854 pageblock_nr_pages) - 1);
5857 static unsigned long pfn_max_align_up(unsigned long pfn)
5859 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5860 pageblock_nr_pages));
5863 /* [start, end) must belong to a single zone. */
5864 static int __alloc_contig_migrate_range(struct compact_control *cc,
5865 unsigned long start, unsigned long end)
5867 /* This function is based on compact_zone() from compaction.c. */
5868 unsigned long nr_reclaimed;
5869 unsigned long pfn = start;
5870 unsigned int tries = 0;
5875 while (pfn < end || !list_empty(&cc->migratepages)) {
5876 if (fatal_signal_pending(current)) {
5881 if (list_empty(&cc->migratepages)) {
5882 cc->nr_migratepages = 0;
5883 pfn = isolate_migratepages_range(cc->zone, cc,
5890 } else if (++tries == 5) {
5891 ret = ret < 0 ? ret : -EBUSY;
5895 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5897 cc->nr_migratepages -= nr_reclaimed;
5899 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
5900 0, MIGRATE_SYNC, MR_CMA);
5903 putback_movable_pages(&cc->migratepages);
5910 * alloc_contig_range() -- tries to allocate given range of pages
5911 * @start: start PFN to allocate
5912 * @end: one-past-the-last PFN to allocate
5913 * @migratetype: migratetype of the underlaying pageblocks (either
5914 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5915 * in range must have the same migratetype and it must
5916 * be either of the two.
5918 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5919 * aligned, however it's the caller's responsibility to guarantee that
5920 * we are the only thread that changes migrate type of pageblocks the
5923 * The PFN range must belong to a single zone.
5925 * Returns zero on success or negative error code. On success all
5926 * pages which PFN is in [start, end) are allocated for the caller and
5927 * need to be freed with free_contig_range().
5929 int alloc_contig_range(unsigned long start, unsigned long end,
5930 unsigned migratetype)
5932 unsigned long outer_start, outer_end;
5935 struct compact_control cc = {
5936 .nr_migratepages = 0,
5938 .zone = page_zone(pfn_to_page(start)),
5940 .ignore_skip_hint = true,
5942 INIT_LIST_HEAD(&cc.migratepages);
5945 * What we do here is we mark all pageblocks in range as
5946 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5947 * have different sizes, and due to the way page allocator
5948 * work, we align the range to biggest of the two pages so
5949 * that page allocator won't try to merge buddies from
5950 * different pageblocks and change MIGRATE_ISOLATE to some
5951 * other migration type.
5953 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5954 * migrate the pages from an unaligned range (ie. pages that
5955 * we are interested in). This will put all the pages in
5956 * range back to page allocator as MIGRATE_ISOLATE.
5958 * When this is done, we take the pages in range from page
5959 * allocator removing them from the buddy system. This way
5960 * page allocator will never consider using them.
5962 * This lets us mark the pageblocks back as
5963 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5964 * aligned range but not in the unaligned, original range are
5965 * put back to page allocator so that buddy can use them.
5968 ret = start_isolate_page_range(pfn_max_align_down(start),
5969 pfn_max_align_up(end), migratetype,
5974 ret = __alloc_contig_migrate_range(&cc, start, end);
5979 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5980 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5981 * more, all pages in [start, end) are free in page allocator.
5982 * What we are going to do is to allocate all pages from
5983 * [start, end) (that is remove them from page allocator).
5985 * The only problem is that pages at the beginning and at the
5986 * end of interesting range may be not aligned with pages that
5987 * page allocator holds, ie. they can be part of higher order
5988 * pages. Because of this, we reserve the bigger range and
5989 * once this is done free the pages we are not interested in.
5991 * We don't have to hold zone->lock here because the pages are
5992 * isolated thus they won't get removed from buddy.
5995 lru_add_drain_all();
5999 outer_start = start;
6000 while (!PageBuddy(pfn_to_page(outer_start))) {
6001 if (++order >= MAX_ORDER) {
6005 outer_start &= ~0UL << order;
6008 /* Make sure the range is really isolated. */
6009 if (test_pages_isolated(outer_start, end, false)) {
6010 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6017 /* Grab isolated pages from freelists. */
6018 outer_end = isolate_freepages_range(&cc, outer_start, end);
6024 /* Free head and tail (if any) */
6025 if (start != outer_start)
6026 free_contig_range(outer_start, start - outer_start);
6027 if (end != outer_end)
6028 free_contig_range(end, outer_end - end);
6031 undo_isolate_page_range(pfn_max_align_down(start),
6032 pfn_max_align_up(end), migratetype);
6036 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6038 unsigned int count = 0;
6040 for (; nr_pages--; pfn++) {
6041 struct page *page = pfn_to_page(pfn);
6043 count += page_count(page) != 1;
6046 WARN(count != 0, "%d pages are still in use!\n", count);
6050 #ifdef CONFIG_MEMORY_HOTPLUG
6051 static int __meminit __zone_pcp_update(void *data)
6053 struct zone *zone = data;
6055 unsigned long batch = zone_batchsize(zone), flags;
6057 for_each_possible_cpu(cpu) {
6058 struct per_cpu_pageset *pset;
6059 struct per_cpu_pages *pcp;
6061 pset = per_cpu_ptr(zone->pageset, cpu);
6064 local_irq_save(flags);
6066 free_pcppages_bulk(zone, pcp->count, pcp);
6067 drain_zonestat(zone, pset);
6068 setup_pageset(pset, batch);
6069 local_irq_restore(flags);
6074 void __meminit zone_pcp_update(struct zone *zone)
6076 stop_machine(__zone_pcp_update, zone, NULL);
6080 void zone_pcp_reset(struct zone *zone)
6082 unsigned long flags;
6084 struct per_cpu_pageset *pset;
6086 /* avoid races with drain_pages() */
6087 local_irq_save(flags);
6088 if (zone->pageset != &boot_pageset) {
6089 for_each_online_cpu(cpu) {
6090 pset = per_cpu_ptr(zone->pageset, cpu);
6091 drain_zonestat(zone, pset);
6093 free_percpu(zone->pageset);
6094 zone->pageset = &boot_pageset;
6096 local_irq_restore(flags);
6099 #ifdef CONFIG_MEMORY_HOTREMOVE
6101 * All pages in the range must be isolated before calling this.
6104 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6110 unsigned long flags;
6111 /* find the first valid pfn */
6112 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6117 zone = page_zone(pfn_to_page(pfn));
6118 spin_lock_irqsave(&zone->lock, flags);
6120 while (pfn < end_pfn) {
6121 if (!pfn_valid(pfn)) {
6125 page = pfn_to_page(pfn);
6127 * The HWPoisoned page may be not in buddy system, and
6128 * page_count() is not 0.
6130 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6132 SetPageReserved(page);
6136 BUG_ON(page_count(page));
6137 BUG_ON(!PageBuddy(page));
6138 order = page_order(page);
6139 #ifdef CONFIG_DEBUG_VM
6140 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6141 pfn, 1 << order, end_pfn);
6143 list_del(&page->lru);
6144 rmv_page_order(page);
6145 zone->free_area[order].nr_free--;
6146 #ifdef CONFIG_HIGHMEM
6147 if (PageHighMem(page))
6148 totalhigh_pages -= 1 << order;
6150 for (i = 0; i < (1 << order); i++)
6151 SetPageReserved((page+i));
6152 pfn += (1 << order);
6154 spin_unlock_irqrestore(&zone->lock, flags);
6158 #ifdef CONFIG_MEMORY_FAILURE
6159 bool is_free_buddy_page(struct page *page)
6161 struct zone *zone = page_zone(page);
6162 unsigned long pfn = page_to_pfn(page);
6163 unsigned long flags;
6166 spin_lock_irqsave(&zone->lock, flags);
6167 for (order = 0; order < MAX_ORDER; order++) {
6168 struct page *page_head = page - (pfn & ((1 << order) - 1));
6170 if (PageBuddy(page_head) && page_order(page_head) >= order)
6173 spin_unlock_irqrestore(&zone->lock, flags);
6175 return order < MAX_ORDER;
6179 static const struct trace_print_flags pageflag_names[] = {
6180 {1UL << PG_locked, "locked" },
6181 {1UL << PG_error, "error" },
6182 {1UL << PG_referenced, "referenced" },
6183 {1UL << PG_uptodate, "uptodate" },
6184 {1UL << PG_dirty, "dirty" },
6185 {1UL << PG_lru, "lru" },
6186 {1UL << PG_active, "active" },
6187 {1UL << PG_slab, "slab" },
6188 {1UL << PG_owner_priv_1, "owner_priv_1" },
6189 {1UL << PG_arch_1, "arch_1" },
6190 {1UL << PG_reserved, "reserved" },
6191 {1UL << PG_private, "private" },
6192 {1UL << PG_private_2, "private_2" },
6193 {1UL << PG_writeback, "writeback" },
6194 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6195 {1UL << PG_head, "head" },
6196 {1UL << PG_tail, "tail" },
6198 {1UL << PG_compound, "compound" },
6200 {1UL << PG_swapcache, "swapcache" },
6201 {1UL << PG_mappedtodisk, "mappedtodisk" },
6202 {1UL << PG_reclaim, "reclaim" },
6203 {1UL << PG_swapbacked, "swapbacked" },
6204 {1UL << PG_unevictable, "unevictable" },
6206 {1UL << PG_mlocked, "mlocked" },
6208 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6209 {1UL << PG_uncached, "uncached" },
6211 #ifdef CONFIG_MEMORY_FAILURE
6212 {1UL << PG_hwpoison, "hwpoison" },
6214 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6215 {1UL << PG_compound_lock, "compound_lock" },
6219 static void dump_page_flags(unsigned long flags)
6221 const char *delim = "";
6225 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6227 printk(KERN_ALERT "page flags: %#lx(", flags);
6229 /* remove zone id */
6230 flags &= (1UL << NR_PAGEFLAGS) - 1;
6232 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6234 mask = pageflag_names[i].mask;
6235 if ((flags & mask) != mask)
6239 printk("%s%s", delim, pageflag_names[i].name);
6243 /* check for left over flags */
6245 printk("%s%#lx", delim, flags);
6250 void dump_page(struct page *page)
6253 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6254 page, atomic_read(&page->_count), page_mapcount(page),
6255 page->mapping, page->index);
6256 dump_page_flags(page->flags);
6257 mem_cgroup_print_bad_page(page);