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/debugobjects.h>
52 #include <linux/kmemleak.h>
53 #include <linux/compaction.h>
54 #include <trace/events/kmem.h>
55 #include <linux/prefetch.h>
56 #include <linux/mm_inline.h>
57 #include <linux/migrate.h>
58 #include <linux/page-debug-flags.h>
59 #include <linux/hugetlb.h>
60 #include <linux/sched/rt.h>
62 #include <asm/sections.h>
63 #include <asm/tlbflush.h>
64 #include <asm/div64.h>
67 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
68 static DEFINE_MUTEX(pcp_batch_high_lock);
69 #define MIN_PERCPU_PAGELIST_FRACTION (8)
71 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
72 DEFINE_PER_CPU(int, numa_node);
73 EXPORT_PER_CPU_SYMBOL(numa_node);
76 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
78 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
79 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
80 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
81 * defined in <linux/topology.h>.
83 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
84 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
85 int _node_numa_mem_[MAX_NUMNODES];
89 * Array of node states.
91 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
92 [N_POSSIBLE] = NODE_MASK_ALL,
93 [N_ONLINE] = { { [0] = 1UL } },
95 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
97 [N_HIGH_MEMORY] = { { [0] = 1UL } },
99 #ifdef CONFIG_MOVABLE_NODE
100 [N_MEMORY] = { { [0] = 1UL } },
102 [N_CPU] = { { [0] = 1UL } },
105 EXPORT_SYMBOL(node_states);
107 /* Protect totalram_pages and zone->managed_pages */
108 static DEFINE_SPINLOCK(managed_page_count_lock);
110 unsigned long totalram_pages __read_mostly;
111 unsigned long totalreserve_pages __read_mostly;
113 * When calculating the number of globally allowed dirty pages, there
114 * is a certain number of per-zone reserves that should not be
115 * considered dirtyable memory. This is the sum of those reserves
116 * over all existing zones that contribute dirtyable memory.
118 unsigned long dirty_balance_reserve __read_mostly;
120 int percpu_pagelist_fraction;
121 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
123 #ifdef CONFIG_PM_SLEEP
125 * The following functions are used by the suspend/hibernate code to temporarily
126 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
127 * while devices are suspended. To avoid races with the suspend/hibernate code,
128 * they should always be called with pm_mutex held (gfp_allowed_mask also should
129 * only be modified with pm_mutex held, unless the suspend/hibernate code is
130 * guaranteed not to run in parallel with that modification).
133 static gfp_t saved_gfp_mask;
135 void pm_restore_gfp_mask(void)
137 WARN_ON(!mutex_is_locked(&pm_mutex));
138 if (saved_gfp_mask) {
139 gfp_allowed_mask = saved_gfp_mask;
144 void pm_restrict_gfp_mask(void)
146 WARN_ON(!mutex_is_locked(&pm_mutex));
147 WARN_ON(saved_gfp_mask);
148 saved_gfp_mask = gfp_allowed_mask;
149 gfp_allowed_mask &= ~GFP_IOFS;
152 bool pm_suspended_storage(void)
154 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
158 #endif /* CONFIG_PM_SLEEP */
160 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
161 int pageblock_order __read_mostly;
164 static void __free_pages_ok(struct page *page, unsigned int order);
167 * results with 256, 32 in the lowmem_reserve sysctl:
168 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
169 * 1G machine -> (16M dma, 784M normal, 224M high)
170 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
171 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
172 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
174 * TBD: should special case ZONE_DMA32 machines here - in those we normally
175 * don't need any ZONE_NORMAL reservation
177 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
178 #ifdef CONFIG_ZONE_DMA
181 #ifdef CONFIG_ZONE_DMA32
184 #ifdef CONFIG_HIGHMEM
190 EXPORT_SYMBOL(totalram_pages);
192 static char * const zone_names[MAX_NR_ZONES] = {
193 #ifdef CONFIG_ZONE_DMA
196 #ifdef CONFIG_ZONE_DMA32
200 #ifdef CONFIG_HIGHMEM
206 int min_free_kbytes = 1024;
207 int user_min_free_kbytes = -1;
209 static unsigned long __meminitdata nr_kernel_pages;
210 static unsigned long __meminitdata nr_all_pages;
211 static unsigned long __meminitdata dma_reserve;
213 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
214 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
215 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
216 static unsigned long __initdata required_kernelcore;
217 static unsigned long __initdata required_movablecore;
218 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
220 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
222 EXPORT_SYMBOL(movable_zone);
223 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
226 int nr_node_ids __read_mostly = MAX_NUMNODES;
227 int nr_online_nodes __read_mostly = 1;
228 EXPORT_SYMBOL(nr_node_ids);
229 EXPORT_SYMBOL(nr_online_nodes);
232 int page_group_by_mobility_disabled __read_mostly;
234 void set_pageblock_migratetype(struct page *page, int migratetype)
236 if (unlikely(page_group_by_mobility_disabled &&
237 migratetype < MIGRATE_PCPTYPES))
238 migratetype = MIGRATE_UNMOVABLE;
240 set_pageblock_flags_group(page, (unsigned long)migratetype,
241 PB_migrate, PB_migrate_end);
244 bool oom_killer_disabled __read_mostly;
246 #ifdef CONFIG_DEBUG_VM
247 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
251 unsigned long pfn = page_to_pfn(page);
252 unsigned long sp, start_pfn;
255 seq = zone_span_seqbegin(zone);
256 start_pfn = zone->zone_start_pfn;
257 sp = zone->spanned_pages;
258 if (!zone_spans_pfn(zone, pfn))
260 } while (zone_span_seqretry(zone, seq));
263 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
264 pfn, zone_to_nid(zone), zone->name,
265 start_pfn, start_pfn + sp);
270 static int page_is_consistent(struct zone *zone, struct page *page)
272 if (!pfn_valid_within(page_to_pfn(page)))
274 if (zone != page_zone(page))
280 * Temporary debugging check for pages not lying within a given zone.
282 static int bad_range(struct zone *zone, struct page *page)
284 if (page_outside_zone_boundaries(zone, page))
286 if (!page_is_consistent(zone, page))
292 static inline int bad_range(struct zone *zone, struct page *page)
298 static void bad_page(struct page *page, const char *reason,
299 unsigned long bad_flags)
301 static unsigned long resume;
302 static unsigned long nr_shown;
303 static unsigned long nr_unshown;
305 /* Don't complain about poisoned pages */
306 if (PageHWPoison(page)) {
307 page_mapcount_reset(page); /* remove PageBuddy */
312 * Allow a burst of 60 reports, then keep quiet for that minute;
313 * or allow a steady drip of one report per second.
315 if (nr_shown == 60) {
316 if (time_before(jiffies, resume)) {
322 "BUG: Bad page state: %lu messages suppressed\n",
329 resume = jiffies + 60 * HZ;
331 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
332 current->comm, page_to_pfn(page));
333 dump_page_badflags(page, reason, bad_flags);
338 /* Leave bad fields for debug, except PageBuddy could make trouble */
339 page_mapcount_reset(page); /* remove PageBuddy */
340 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
344 * Higher-order pages are called "compound pages". They are structured thusly:
346 * The first PAGE_SIZE page is called the "head page".
348 * The remaining PAGE_SIZE pages are called "tail pages".
350 * All pages have PG_compound set. All tail pages have their ->first_page
351 * pointing at the head page.
353 * The first tail page's ->lru.next holds the address of the compound page's
354 * put_page() function. Its ->lru.prev holds the order of allocation.
355 * This usage means that zero-order pages may not be compound.
358 static void free_compound_page(struct page *page)
360 __free_pages_ok(page, compound_order(page));
363 void prep_compound_page(struct page *page, unsigned long order)
366 int nr_pages = 1 << order;
368 set_compound_page_dtor(page, free_compound_page);
369 set_compound_order(page, order);
371 for (i = 1; i < nr_pages; i++) {
372 struct page *p = page + i;
373 set_page_count(p, 0);
374 p->first_page = page;
375 /* Make sure p->first_page is always valid for PageTail() */
381 /* update __split_huge_page_refcount if you change this function */
382 static int destroy_compound_page(struct page *page, unsigned long order)
385 int nr_pages = 1 << order;
388 if (unlikely(compound_order(page) != order)) {
389 bad_page(page, "wrong compound order", 0);
393 __ClearPageHead(page);
395 for (i = 1; i < nr_pages; i++) {
396 struct page *p = page + i;
398 if (unlikely(!PageTail(p))) {
399 bad_page(page, "PageTail not set", 0);
401 } else if (unlikely(p->first_page != page)) {
402 bad_page(page, "first_page not consistent", 0);
411 static inline void prep_zero_page(struct page *page, unsigned int order,
417 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
418 * and __GFP_HIGHMEM from hard or soft interrupt context.
420 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
421 for (i = 0; i < (1 << order); i++)
422 clear_highpage(page + i);
425 #ifdef CONFIG_DEBUG_PAGEALLOC
426 unsigned int _debug_guardpage_minorder;
428 static int __init debug_guardpage_minorder_setup(char *buf)
432 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
433 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
436 _debug_guardpage_minorder = res;
437 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
440 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
442 static inline void set_page_guard_flag(struct page *page)
444 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
447 static inline void clear_page_guard_flag(struct page *page)
449 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
452 static inline void set_page_guard_flag(struct page *page) { }
453 static inline void clear_page_guard_flag(struct page *page) { }
456 static inline void set_page_order(struct page *page, unsigned int order)
458 set_page_private(page, order);
459 __SetPageBuddy(page);
462 static inline void rmv_page_order(struct page *page)
464 __ClearPageBuddy(page);
465 set_page_private(page, 0);
469 * This function checks whether a page is free && is the buddy
470 * we can do coalesce a page and its buddy if
471 * (a) the buddy is not in a hole &&
472 * (b) the buddy is in the buddy system &&
473 * (c) a page and its buddy have the same order &&
474 * (d) a page and its buddy are in the same zone.
476 * For recording whether a page is in the buddy system, we set ->_mapcount
477 * PAGE_BUDDY_MAPCOUNT_VALUE.
478 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
479 * serialized by zone->lock.
481 * For recording page's order, we use page_private(page).
483 static inline int page_is_buddy(struct page *page, struct page *buddy,
486 if (!pfn_valid_within(page_to_pfn(buddy)))
489 if (page_is_guard(buddy) && page_order(buddy) == order) {
490 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
492 if (page_zone_id(page) != page_zone_id(buddy))
498 if (PageBuddy(buddy) && page_order(buddy) == order) {
499 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
502 * zone check is done late to avoid uselessly
503 * calculating zone/node ids for pages that could
506 if (page_zone_id(page) != page_zone_id(buddy))
515 * Freeing function for a buddy system allocator.
517 * The concept of a buddy system is to maintain direct-mapped table
518 * (containing bit values) for memory blocks of various "orders".
519 * The bottom level table contains the map for the smallest allocatable
520 * units of memory (here, pages), and each level above it describes
521 * pairs of units from the levels below, hence, "buddies".
522 * At a high level, all that happens here is marking the table entry
523 * at the bottom level available, and propagating the changes upward
524 * as necessary, plus some accounting needed to play nicely with other
525 * parts of the VM system.
526 * At each level, we keep a list of pages, which are heads of continuous
527 * free pages of length of (1 << order) and marked with _mapcount
528 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
530 * So when we are allocating or freeing one, we can derive the state of the
531 * other. That is, if we allocate a small block, and both were
532 * free, the remainder of the region must be split into blocks.
533 * If a block is freed, and its buddy is also free, then this
534 * triggers coalescing into a block of larger size.
539 static inline void __free_one_page(struct page *page,
541 struct zone *zone, unsigned int order,
544 unsigned long page_idx;
545 unsigned long combined_idx;
546 unsigned long uninitialized_var(buddy_idx);
548 int max_order = MAX_ORDER;
550 VM_BUG_ON(!zone_is_initialized(zone));
552 if (unlikely(PageCompound(page)))
553 if (unlikely(destroy_compound_page(page, order)))
556 VM_BUG_ON(migratetype == -1);
557 if (is_migrate_isolate(migratetype)) {
559 * We restrict max order of merging to prevent merge
560 * between freepages on isolate pageblock and normal
561 * pageblock. Without this, pageblock isolation
562 * could cause incorrect freepage accounting.
564 max_order = min(MAX_ORDER, pageblock_order + 1);
566 __mod_zone_freepage_state(zone, 1 << order, migratetype);
569 page_idx = pfn & ((1 << max_order) - 1);
571 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
572 VM_BUG_ON_PAGE(bad_range(zone, page), page);
574 while (order < max_order - 1) {
575 buddy_idx = __find_buddy_index(page_idx, order);
576 buddy = page + (buddy_idx - page_idx);
577 if (!page_is_buddy(page, buddy, order))
580 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
581 * merge with it and move up one order.
583 if (page_is_guard(buddy)) {
584 clear_page_guard_flag(buddy);
585 set_page_private(buddy, 0);
586 if (!is_migrate_isolate(migratetype)) {
587 __mod_zone_freepage_state(zone, 1 << order,
591 list_del(&buddy->lru);
592 zone->free_area[order].nr_free--;
593 rmv_page_order(buddy);
595 combined_idx = buddy_idx & page_idx;
596 page = page + (combined_idx - page_idx);
597 page_idx = combined_idx;
600 set_page_order(page, order);
603 * If this is not the largest possible page, check if the buddy
604 * of the next-highest order is free. If it is, it's possible
605 * that pages are being freed that will coalesce soon. In case,
606 * that is happening, add the free page to the tail of the list
607 * so it's less likely to be used soon and more likely to be merged
608 * as a higher order page
610 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
611 struct page *higher_page, *higher_buddy;
612 combined_idx = buddy_idx & page_idx;
613 higher_page = page + (combined_idx - page_idx);
614 buddy_idx = __find_buddy_index(combined_idx, order + 1);
615 higher_buddy = higher_page + (buddy_idx - combined_idx);
616 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
617 list_add_tail(&page->lru,
618 &zone->free_area[order].free_list[migratetype]);
623 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
625 zone->free_area[order].nr_free++;
628 static inline int free_pages_check(struct page *page)
630 const char *bad_reason = NULL;
631 unsigned long bad_flags = 0;
633 if (unlikely(page_mapcount(page)))
634 bad_reason = "nonzero mapcount";
635 if (unlikely(page->mapping != NULL))
636 bad_reason = "non-NULL mapping";
637 if (unlikely(atomic_read(&page->_count) != 0))
638 bad_reason = "nonzero _count";
639 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
640 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
641 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
643 if (unlikely(mem_cgroup_bad_page_check(page)))
644 bad_reason = "cgroup check failed";
645 if (unlikely(bad_reason)) {
646 bad_page(page, bad_reason, bad_flags);
649 page_cpupid_reset_last(page);
650 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
651 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
656 * Frees a number of pages from the PCP lists
657 * Assumes all pages on list are in same zone, and of same order.
658 * count is the number of pages to free.
660 * If the zone was previously in an "all pages pinned" state then look to
661 * see if this freeing clears that state.
663 * And clear the zone's pages_scanned counter, to hold off the "all pages are
664 * pinned" detection logic.
666 static void free_pcppages_bulk(struct zone *zone, int count,
667 struct per_cpu_pages *pcp)
672 unsigned long nr_scanned;
674 spin_lock(&zone->lock);
675 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
677 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
681 struct list_head *list;
684 * Remove pages from lists in a round-robin fashion. A
685 * batch_free count is maintained that is incremented when an
686 * empty list is encountered. This is so more pages are freed
687 * off fuller lists instead of spinning excessively around empty
692 if (++migratetype == MIGRATE_PCPTYPES)
694 list = &pcp->lists[migratetype];
695 } while (list_empty(list));
697 /* This is the only non-empty list. Free them all. */
698 if (batch_free == MIGRATE_PCPTYPES)
699 batch_free = to_free;
702 int mt; /* migratetype of the to-be-freed page */
704 page = list_entry(list->prev, struct page, lru);
705 /* must delete as __free_one_page list manipulates */
706 list_del(&page->lru);
707 mt = get_freepage_migratetype(page);
708 if (unlikely(has_isolate_pageblock(zone)))
709 mt = get_pageblock_migratetype(page);
711 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
712 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
713 trace_mm_page_pcpu_drain(page, 0, mt);
714 } while (--to_free && --batch_free && !list_empty(list));
716 spin_unlock(&zone->lock);
719 static void free_one_page(struct zone *zone,
720 struct page *page, unsigned long pfn,
724 unsigned long nr_scanned;
725 spin_lock(&zone->lock);
726 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
728 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
730 if (unlikely(has_isolate_pageblock(zone) ||
731 is_migrate_isolate(migratetype))) {
732 migratetype = get_pfnblock_migratetype(page, pfn);
734 __free_one_page(page, pfn, zone, order, migratetype);
735 spin_unlock(&zone->lock);
738 static bool free_pages_prepare(struct page *page, unsigned int order)
743 VM_BUG_ON_PAGE(PageTail(page), page);
744 VM_BUG_ON_PAGE(PageHead(page) && compound_order(page) != order, page);
746 trace_mm_page_free(page, order);
747 kmemcheck_free_shadow(page, order);
750 page->mapping = NULL;
751 for (i = 0; i < (1 << order); i++)
752 bad += free_pages_check(page + i);
756 if (!PageHighMem(page)) {
757 debug_check_no_locks_freed(page_address(page),
759 debug_check_no_obj_freed(page_address(page),
762 arch_free_page(page, order);
763 kernel_map_pages(page, 1 << order, 0);
768 static void __free_pages_ok(struct page *page, unsigned int order)
772 unsigned long pfn = page_to_pfn(page);
774 if (!free_pages_prepare(page, order))
777 migratetype = get_pfnblock_migratetype(page, pfn);
778 local_irq_save(flags);
779 __count_vm_events(PGFREE, 1 << order);
780 set_freepage_migratetype(page, migratetype);
781 free_one_page(page_zone(page), page, pfn, order, migratetype);
782 local_irq_restore(flags);
785 void __init __free_pages_bootmem(struct page *page, unsigned int order)
787 unsigned int nr_pages = 1 << order;
788 struct page *p = page;
792 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
794 __ClearPageReserved(p);
795 set_page_count(p, 0);
797 __ClearPageReserved(p);
798 set_page_count(p, 0);
800 page_zone(page)->managed_pages += nr_pages;
801 set_page_refcounted(page);
802 __free_pages(page, order);
806 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
807 void __init init_cma_reserved_pageblock(struct page *page)
809 unsigned i = pageblock_nr_pages;
810 struct page *p = page;
813 __ClearPageReserved(p);
814 set_page_count(p, 0);
817 set_pageblock_migratetype(page, MIGRATE_CMA);
819 if (pageblock_order >= MAX_ORDER) {
820 i = pageblock_nr_pages;
823 set_page_refcounted(p);
824 __free_pages(p, MAX_ORDER - 1);
825 p += MAX_ORDER_NR_PAGES;
826 } while (i -= MAX_ORDER_NR_PAGES);
828 set_page_refcounted(page);
829 __free_pages(page, pageblock_order);
832 adjust_managed_page_count(page, pageblock_nr_pages);
837 * The order of subdivision here is critical for the IO subsystem.
838 * Please do not alter this order without good reasons and regression
839 * testing. Specifically, as large blocks of memory are subdivided,
840 * the order in which smaller blocks are delivered depends on the order
841 * they're subdivided in this function. This is the primary factor
842 * influencing the order in which pages are delivered to the IO
843 * subsystem according to empirical testing, and this is also justified
844 * by considering the behavior of a buddy system containing a single
845 * large block of memory acted on by a series of small allocations.
846 * This behavior is a critical factor in sglist merging's success.
850 static inline void expand(struct zone *zone, struct page *page,
851 int low, int high, struct free_area *area,
854 unsigned long size = 1 << high;
860 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
862 #ifdef CONFIG_DEBUG_PAGEALLOC
863 if (high < debug_guardpage_minorder()) {
865 * Mark as guard pages (or page), that will allow to
866 * merge back to allocator when buddy will be freed.
867 * Corresponding page table entries will not be touched,
868 * pages will stay not present in virtual address space
870 INIT_LIST_HEAD(&page[size].lru);
871 set_page_guard_flag(&page[size]);
872 set_page_private(&page[size], high);
873 /* Guard pages are not available for any usage */
874 __mod_zone_freepage_state(zone, -(1 << high),
879 list_add(&page[size].lru, &area->free_list[migratetype]);
881 set_page_order(&page[size], high);
886 * This page is about to be returned from the page allocator
888 static inline int check_new_page(struct page *page)
890 const char *bad_reason = NULL;
891 unsigned long bad_flags = 0;
893 if (unlikely(page_mapcount(page)))
894 bad_reason = "nonzero mapcount";
895 if (unlikely(page->mapping != NULL))
896 bad_reason = "non-NULL mapping";
897 if (unlikely(atomic_read(&page->_count) != 0))
898 bad_reason = "nonzero _count";
899 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
900 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
901 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
903 if (unlikely(mem_cgroup_bad_page_check(page)))
904 bad_reason = "cgroup check failed";
905 if (unlikely(bad_reason)) {
906 bad_page(page, bad_reason, bad_flags);
912 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags)
916 for (i = 0; i < (1 << order); i++) {
917 struct page *p = page + i;
918 if (unlikely(check_new_page(p)))
922 set_page_private(page, 0);
923 set_page_refcounted(page);
925 arch_alloc_page(page, order);
926 kernel_map_pages(page, 1 << order, 1);
928 if (gfp_flags & __GFP_ZERO)
929 prep_zero_page(page, order, gfp_flags);
931 if (order && (gfp_flags & __GFP_COMP))
932 prep_compound_page(page, order);
938 * Go through the free lists for the given migratetype and remove
939 * the smallest available page from the freelists
942 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
945 unsigned int current_order;
946 struct free_area *area;
949 /* Find a page of the appropriate size in the preferred list */
950 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
951 area = &(zone->free_area[current_order]);
952 if (list_empty(&area->free_list[migratetype]))
955 page = list_entry(area->free_list[migratetype].next,
957 list_del(&page->lru);
958 rmv_page_order(page);
960 expand(zone, page, order, current_order, area, migratetype);
961 set_freepage_migratetype(page, migratetype);
970 * This array describes the order lists are fallen back to when
971 * the free lists for the desirable migrate type are depleted
973 static int fallbacks[MIGRATE_TYPES][4] = {
974 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
975 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
977 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
978 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
980 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
982 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
983 #ifdef CONFIG_MEMORY_ISOLATION
984 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
989 * Move the free pages in a range to the free lists of the requested type.
990 * Note that start_page and end_pages are not aligned on a pageblock
991 * boundary. If alignment is required, use move_freepages_block()
993 int move_freepages(struct zone *zone,
994 struct page *start_page, struct page *end_page,
1001 #ifndef CONFIG_HOLES_IN_ZONE
1003 * page_zone is not safe to call in this context when
1004 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1005 * anyway as we check zone boundaries in move_freepages_block().
1006 * Remove at a later date when no bug reports exist related to
1007 * grouping pages by mobility
1009 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1012 for (page = start_page; page <= end_page;) {
1013 /* Make sure we are not inadvertently changing nodes */
1014 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1016 if (!pfn_valid_within(page_to_pfn(page))) {
1021 if (!PageBuddy(page)) {
1026 order = page_order(page);
1027 list_move(&page->lru,
1028 &zone->free_area[order].free_list[migratetype]);
1029 set_freepage_migratetype(page, migratetype);
1031 pages_moved += 1 << order;
1037 int move_freepages_block(struct zone *zone, struct page *page,
1040 unsigned long start_pfn, end_pfn;
1041 struct page *start_page, *end_page;
1043 start_pfn = page_to_pfn(page);
1044 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1045 start_page = pfn_to_page(start_pfn);
1046 end_page = start_page + pageblock_nr_pages - 1;
1047 end_pfn = start_pfn + pageblock_nr_pages - 1;
1049 /* Do not cross zone boundaries */
1050 if (!zone_spans_pfn(zone, start_pfn))
1052 if (!zone_spans_pfn(zone, end_pfn))
1055 return move_freepages(zone, start_page, end_page, migratetype);
1058 static void change_pageblock_range(struct page *pageblock_page,
1059 int start_order, int migratetype)
1061 int nr_pageblocks = 1 << (start_order - pageblock_order);
1063 while (nr_pageblocks--) {
1064 set_pageblock_migratetype(pageblock_page, migratetype);
1065 pageblock_page += pageblock_nr_pages;
1070 * If breaking a large block of pages, move all free pages to the preferred
1071 * allocation list. If falling back for a reclaimable kernel allocation, be
1072 * more aggressive about taking ownership of free pages.
1074 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1075 * nor move CMA pages to different free lists. We don't want unmovable pages
1076 * to be allocated from MIGRATE_CMA areas.
1078 * Returns the new migratetype of the pageblock (or the same old migratetype
1079 * if it was unchanged).
1081 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1082 int start_type, int fallback_type)
1084 int current_order = page_order(page);
1087 * When borrowing from MIGRATE_CMA, we need to release the excess
1088 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1089 * is set to CMA so it is returned to the correct freelist in case
1090 * the page ends up being not actually allocated from the pcp lists.
1092 if (is_migrate_cma(fallback_type))
1093 return fallback_type;
1095 /* Take ownership for orders >= pageblock_order */
1096 if (current_order >= pageblock_order) {
1097 change_pageblock_range(page, current_order, start_type);
1101 if (current_order >= pageblock_order / 2 ||
1102 start_type == MIGRATE_RECLAIMABLE ||
1103 page_group_by_mobility_disabled) {
1106 pages = move_freepages_block(zone, page, start_type);
1108 /* Claim the whole block if over half of it is free */
1109 if (pages >= (1 << (pageblock_order-1)) ||
1110 page_group_by_mobility_disabled) {
1112 set_pageblock_migratetype(page, start_type);
1118 return fallback_type;
1121 /* Remove an element from the buddy allocator from the fallback list */
1122 static inline struct page *
1123 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1125 struct free_area *area;
1126 unsigned int current_order;
1128 int migratetype, new_type, i;
1130 /* Find the largest possible block of pages in the other list */
1131 for (current_order = MAX_ORDER-1;
1132 current_order >= order && current_order <= MAX_ORDER-1;
1135 migratetype = fallbacks[start_migratetype][i];
1137 /* MIGRATE_RESERVE handled later if necessary */
1138 if (migratetype == MIGRATE_RESERVE)
1141 area = &(zone->free_area[current_order]);
1142 if (list_empty(&area->free_list[migratetype]))
1145 page = list_entry(area->free_list[migratetype].next,
1149 new_type = try_to_steal_freepages(zone, page,
1153 /* Remove the page from the freelists */
1154 list_del(&page->lru);
1155 rmv_page_order(page);
1157 expand(zone, page, order, current_order, area,
1159 /* The freepage_migratetype may differ from pageblock's
1160 * migratetype depending on the decisions in
1161 * try_to_steal_freepages. This is OK as long as it does
1162 * not differ for MIGRATE_CMA type.
1164 set_freepage_migratetype(page, new_type);
1166 trace_mm_page_alloc_extfrag(page, order, current_order,
1167 start_migratetype, migratetype, new_type);
1177 * Do the hard work of removing an element from the buddy allocator.
1178 * Call me with the zone->lock already held.
1180 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1186 page = __rmqueue_smallest(zone, order, migratetype);
1188 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1189 page = __rmqueue_fallback(zone, order, migratetype);
1192 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1193 * is used because __rmqueue_smallest is an inline function
1194 * and we want just one call site
1197 migratetype = MIGRATE_RESERVE;
1202 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1207 * Obtain a specified number of elements from the buddy allocator, all under
1208 * a single hold of the lock, for efficiency. Add them to the supplied list.
1209 * Returns the number of new pages which were placed at *list.
1211 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1212 unsigned long count, struct list_head *list,
1213 int migratetype, bool cold)
1217 spin_lock(&zone->lock);
1218 for (i = 0; i < count; ++i) {
1219 struct page *page = __rmqueue(zone, order, migratetype);
1220 if (unlikely(page == NULL))
1224 * Split buddy pages returned by expand() are received here
1225 * in physical page order. The page is added to the callers and
1226 * list and the list head then moves forward. From the callers
1227 * perspective, the linked list is ordered by page number in
1228 * some conditions. This is useful for IO devices that can
1229 * merge IO requests if the physical pages are ordered
1233 list_add(&page->lru, list);
1235 list_add_tail(&page->lru, list);
1237 if (is_migrate_cma(get_freepage_migratetype(page)))
1238 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1241 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1242 spin_unlock(&zone->lock);
1248 * Called from the vmstat counter updater to drain pagesets of this
1249 * currently executing processor on remote nodes after they have
1252 * Note that this function must be called with the thread pinned to
1253 * a single processor.
1255 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1257 unsigned long flags;
1258 int to_drain, batch;
1260 local_irq_save(flags);
1261 batch = ACCESS_ONCE(pcp->batch);
1262 to_drain = min(pcp->count, batch);
1264 free_pcppages_bulk(zone, to_drain, pcp);
1265 pcp->count -= to_drain;
1267 local_irq_restore(flags);
1272 * Drain pcplists of the indicated processor and zone.
1274 * The processor must either be the current processor and the
1275 * thread pinned to the current processor or a processor that
1278 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1280 unsigned long flags;
1281 struct per_cpu_pageset *pset;
1282 struct per_cpu_pages *pcp;
1284 local_irq_save(flags);
1285 pset = per_cpu_ptr(zone->pageset, cpu);
1289 free_pcppages_bulk(zone, pcp->count, pcp);
1292 local_irq_restore(flags);
1296 * Drain pcplists of all zones on the indicated processor.
1298 * The processor must either be the current processor and the
1299 * thread pinned to the current processor or a processor that
1302 static void drain_pages(unsigned int cpu)
1306 for_each_populated_zone(zone) {
1307 drain_pages_zone(cpu, zone);
1312 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1314 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1315 * the single zone's pages.
1317 void drain_local_pages(struct zone *zone)
1319 int cpu = smp_processor_id();
1322 drain_pages_zone(cpu, zone);
1328 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1330 * When zone parameter is non-NULL, spill just the single zone's pages.
1332 * Note that this code is protected against sending an IPI to an offline
1333 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1334 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1335 * nothing keeps CPUs from showing up after we populated the cpumask and
1336 * before the call to on_each_cpu_mask().
1338 void drain_all_pages(struct zone *zone)
1343 * Allocate in the BSS so we wont require allocation in
1344 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1346 static cpumask_t cpus_with_pcps;
1349 * We don't care about racing with CPU hotplug event
1350 * as offline notification will cause the notified
1351 * cpu to drain that CPU pcps and on_each_cpu_mask
1352 * disables preemption as part of its processing
1354 for_each_online_cpu(cpu) {
1355 struct per_cpu_pageset *pcp;
1357 bool has_pcps = false;
1360 pcp = per_cpu_ptr(zone->pageset, cpu);
1364 for_each_populated_zone(z) {
1365 pcp = per_cpu_ptr(z->pageset, cpu);
1366 if (pcp->pcp.count) {
1374 cpumask_set_cpu(cpu, &cpus_with_pcps);
1376 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1378 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1382 #ifdef CONFIG_HIBERNATION
1384 void mark_free_pages(struct zone *zone)
1386 unsigned long pfn, max_zone_pfn;
1387 unsigned long flags;
1388 unsigned int order, t;
1389 struct list_head *curr;
1391 if (zone_is_empty(zone))
1394 spin_lock_irqsave(&zone->lock, flags);
1396 max_zone_pfn = zone_end_pfn(zone);
1397 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1398 if (pfn_valid(pfn)) {
1399 struct page *page = pfn_to_page(pfn);
1401 if (!swsusp_page_is_forbidden(page))
1402 swsusp_unset_page_free(page);
1405 for_each_migratetype_order(order, t) {
1406 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1409 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1410 for (i = 0; i < (1UL << order); i++)
1411 swsusp_set_page_free(pfn_to_page(pfn + i));
1414 spin_unlock_irqrestore(&zone->lock, flags);
1416 #endif /* CONFIG_PM */
1419 * Free a 0-order page
1420 * cold == true ? free a cold page : free a hot page
1422 void free_hot_cold_page(struct page *page, bool cold)
1424 struct zone *zone = page_zone(page);
1425 struct per_cpu_pages *pcp;
1426 unsigned long flags;
1427 unsigned long pfn = page_to_pfn(page);
1430 if (!free_pages_prepare(page, 0))
1433 migratetype = get_pfnblock_migratetype(page, pfn);
1434 set_freepage_migratetype(page, migratetype);
1435 local_irq_save(flags);
1436 __count_vm_event(PGFREE);
1439 * We only track unmovable, reclaimable and movable on pcp lists.
1440 * Free ISOLATE pages back to the allocator because they are being
1441 * offlined but treat RESERVE as movable pages so we can get those
1442 * areas back if necessary. Otherwise, we may have to free
1443 * excessively into the page allocator
1445 if (migratetype >= MIGRATE_PCPTYPES) {
1446 if (unlikely(is_migrate_isolate(migratetype))) {
1447 free_one_page(zone, page, pfn, 0, migratetype);
1450 migratetype = MIGRATE_MOVABLE;
1453 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1455 list_add(&page->lru, &pcp->lists[migratetype]);
1457 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1459 if (pcp->count >= pcp->high) {
1460 unsigned long batch = ACCESS_ONCE(pcp->batch);
1461 free_pcppages_bulk(zone, batch, pcp);
1462 pcp->count -= batch;
1466 local_irq_restore(flags);
1470 * Free a list of 0-order pages
1472 void free_hot_cold_page_list(struct list_head *list, bool cold)
1474 struct page *page, *next;
1476 list_for_each_entry_safe(page, next, list, lru) {
1477 trace_mm_page_free_batched(page, cold);
1478 free_hot_cold_page(page, cold);
1483 * split_page takes a non-compound higher-order page, and splits it into
1484 * n (1<<order) sub-pages: page[0..n]
1485 * Each sub-page must be freed individually.
1487 * Note: this is probably too low level an operation for use in drivers.
1488 * Please consult with lkml before using this in your driver.
1490 void split_page(struct page *page, unsigned int order)
1494 VM_BUG_ON_PAGE(PageCompound(page), page);
1495 VM_BUG_ON_PAGE(!page_count(page), page);
1497 #ifdef CONFIG_KMEMCHECK
1499 * Split shadow pages too, because free(page[0]) would
1500 * otherwise free the whole shadow.
1502 if (kmemcheck_page_is_tracked(page))
1503 split_page(virt_to_page(page[0].shadow), order);
1506 for (i = 1; i < (1 << order); i++)
1507 set_page_refcounted(page + i);
1509 EXPORT_SYMBOL_GPL(split_page);
1511 int __isolate_free_page(struct page *page, unsigned int order)
1513 unsigned long watermark;
1517 BUG_ON(!PageBuddy(page));
1519 zone = page_zone(page);
1520 mt = get_pageblock_migratetype(page);
1522 if (!is_migrate_isolate(mt)) {
1523 /* Obey watermarks as if the page was being allocated */
1524 watermark = low_wmark_pages(zone) + (1 << order);
1525 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1528 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1531 /* Remove page from free list */
1532 list_del(&page->lru);
1533 zone->free_area[order].nr_free--;
1534 rmv_page_order(page);
1536 /* Set the pageblock if the isolated page is at least a pageblock */
1537 if (order >= pageblock_order - 1) {
1538 struct page *endpage = page + (1 << order) - 1;
1539 for (; page < endpage; page += pageblock_nr_pages) {
1540 int mt = get_pageblock_migratetype(page);
1541 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1542 set_pageblock_migratetype(page,
1547 return 1UL << order;
1551 * Similar to split_page except the page is already free. As this is only
1552 * being used for migration, the migratetype of the block also changes.
1553 * As this is called with interrupts disabled, the caller is responsible
1554 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1557 * Note: this is probably too low level an operation for use in drivers.
1558 * Please consult with lkml before using this in your driver.
1560 int split_free_page(struct page *page)
1565 order = page_order(page);
1567 nr_pages = __isolate_free_page(page, order);
1571 /* Split into individual pages */
1572 set_page_refcounted(page);
1573 split_page(page, order);
1578 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1579 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1583 struct page *buffered_rmqueue(struct zone *preferred_zone,
1584 struct zone *zone, unsigned int order,
1585 gfp_t gfp_flags, int migratetype)
1587 unsigned long flags;
1589 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1592 if (likely(order == 0)) {
1593 struct per_cpu_pages *pcp;
1594 struct list_head *list;
1596 local_irq_save(flags);
1597 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1598 list = &pcp->lists[migratetype];
1599 if (list_empty(list)) {
1600 pcp->count += rmqueue_bulk(zone, 0,
1603 if (unlikely(list_empty(list)))
1608 page = list_entry(list->prev, struct page, lru);
1610 page = list_entry(list->next, struct page, lru);
1612 list_del(&page->lru);
1615 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1617 * __GFP_NOFAIL is not to be used in new code.
1619 * All __GFP_NOFAIL callers should be fixed so that they
1620 * properly detect and handle allocation failures.
1622 * We most definitely don't want callers attempting to
1623 * allocate greater than order-1 page units with
1626 WARN_ON_ONCE(order > 1);
1628 spin_lock_irqsave(&zone->lock, flags);
1629 page = __rmqueue(zone, order, migratetype);
1630 spin_unlock(&zone->lock);
1633 __mod_zone_freepage_state(zone, -(1 << order),
1634 get_freepage_migratetype(page));
1637 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1638 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1639 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1640 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1642 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1643 zone_statistics(preferred_zone, zone, gfp_flags);
1644 local_irq_restore(flags);
1646 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1647 if (prep_new_page(page, order, gfp_flags))
1652 local_irq_restore(flags);
1656 #ifdef CONFIG_FAIL_PAGE_ALLOC
1659 struct fault_attr attr;
1661 u32 ignore_gfp_highmem;
1662 u32 ignore_gfp_wait;
1664 } fail_page_alloc = {
1665 .attr = FAULT_ATTR_INITIALIZER,
1666 .ignore_gfp_wait = 1,
1667 .ignore_gfp_highmem = 1,
1671 static int __init setup_fail_page_alloc(char *str)
1673 return setup_fault_attr(&fail_page_alloc.attr, str);
1675 __setup("fail_page_alloc=", setup_fail_page_alloc);
1677 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1679 if (order < fail_page_alloc.min_order)
1681 if (gfp_mask & __GFP_NOFAIL)
1683 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1685 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1688 return should_fail(&fail_page_alloc.attr, 1 << order);
1691 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1693 static int __init fail_page_alloc_debugfs(void)
1695 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1698 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1699 &fail_page_alloc.attr);
1701 return PTR_ERR(dir);
1703 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1704 &fail_page_alloc.ignore_gfp_wait))
1706 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1707 &fail_page_alloc.ignore_gfp_highmem))
1709 if (!debugfs_create_u32("min-order", mode, dir,
1710 &fail_page_alloc.min_order))
1715 debugfs_remove_recursive(dir);
1720 late_initcall(fail_page_alloc_debugfs);
1722 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1724 #else /* CONFIG_FAIL_PAGE_ALLOC */
1726 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1731 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1734 * Return true if free pages are above 'mark'. This takes into account the order
1735 * of the allocation.
1737 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1738 unsigned long mark, int classzone_idx, int alloc_flags,
1741 /* free_pages may go negative - that's OK */
1746 free_pages -= (1 << order) - 1;
1747 if (alloc_flags & ALLOC_HIGH)
1749 if (alloc_flags & ALLOC_HARDER)
1752 /* If allocation can't use CMA areas don't use free CMA pages */
1753 if (!(alloc_flags & ALLOC_CMA))
1754 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1757 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1759 for (o = 0; o < order; o++) {
1760 /* At the next order, this order's pages become unavailable */
1761 free_pages -= z->free_area[o].nr_free << o;
1763 /* Require fewer higher order pages to be free */
1766 if (free_pages <= min)
1772 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1773 int classzone_idx, int alloc_flags)
1775 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1776 zone_page_state(z, NR_FREE_PAGES));
1779 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1780 unsigned long mark, int classzone_idx, int alloc_flags)
1782 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1784 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1785 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1787 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1793 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1794 * skip over zones that are not allowed by the cpuset, or that have
1795 * been recently (in last second) found to be nearly full. See further
1796 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1797 * that have to skip over a lot of full or unallowed zones.
1799 * If the zonelist cache is present in the passed zonelist, then
1800 * returns a pointer to the allowed node mask (either the current
1801 * tasks mems_allowed, or node_states[N_MEMORY].)
1803 * If the zonelist cache is not available for this zonelist, does
1804 * nothing and returns NULL.
1806 * If the fullzones BITMAP in the zonelist cache is stale (more than
1807 * a second since last zap'd) then we zap it out (clear its bits.)
1809 * We hold off even calling zlc_setup, until after we've checked the
1810 * first zone in the zonelist, on the theory that most allocations will
1811 * be satisfied from that first zone, so best to examine that zone as
1812 * quickly as we can.
1814 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1816 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1817 nodemask_t *allowednodes; /* zonelist_cache approximation */
1819 zlc = zonelist->zlcache_ptr;
1823 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1824 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1825 zlc->last_full_zap = jiffies;
1828 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1829 &cpuset_current_mems_allowed :
1830 &node_states[N_MEMORY];
1831 return allowednodes;
1835 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1836 * if it is worth looking at further for free memory:
1837 * 1) Check that the zone isn't thought to be full (doesn't have its
1838 * bit set in the zonelist_cache fullzones BITMAP).
1839 * 2) Check that the zones node (obtained from the zonelist_cache
1840 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1841 * Return true (non-zero) if zone is worth looking at further, or
1842 * else return false (zero) if it is not.
1844 * This check -ignores- the distinction between various watermarks,
1845 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1846 * found to be full for any variation of these watermarks, it will
1847 * be considered full for up to one second by all requests, unless
1848 * we are so low on memory on all allowed nodes that we are forced
1849 * into the second scan of the zonelist.
1851 * In the second scan we ignore this zonelist cache and exactly
1852 * apply the watermarks to all zones, even it is slower to do so.
1853 * We are low on memory in the second scan, and should leave no stone
1854 * unturned looking for a free page.
1856 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1857 nodemask_t *allowednodes)
1859 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1860 int i; /* index of *z in zonelist zones */
1861 int n; /* node that zone *z is on */
1863 zlc = zonelist->zlcache_ptr;
1867 i = z - zonelist->_zonerefs;
1870 /* This zone is worth trying if it is allowed but not full */
1871 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1875 * Given 'z' scanning a zonelist, set the corresponding bit in
1876 * zlc->fullzones, so that subsequent attempts to allocate a page
1877 * from that zone don't waste time re-examining it.
1879 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1881 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1882 int i; /* index of *z in zonelist zones */
1884 zlc = zonelist->zlcache_ptr;
1888 i = z - zonelist->_zonerefs;
1890 set_bit(i, zlc->fullzones);
1894 * clear all zones full, called after direct reclaim makes progress so that
1895 * a zone that was recently full is not skipped over for up to a second
1897 static void zlc_clear_zones_full(struct zonelist *zonelist)
1899 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1901 zlc = zonelist->zlcache_ptr;
1905 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1908 static bool zone_local(struct zone *local_zone, struct zone *zone)
1910 return local_zone->node == zone->node;
1913 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1915 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1919 #else /* CONFIG_NUMA */
1921 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1926 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1927 nodemask_t *allowednodes)
1932 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1936 static void zlc_clear_zones_full(struct zonelist *zonelist)
1940 static bool zone_local(struct zone *local_zone, struct zone *zone)
1945 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1950 #endif /* CONFIG_NUMA */
1952 static void reset_alloc_batches(struct zone *preferred_zone)
1954 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
1957 mod_zone_page_state(zone, NR_ALLOC_BATCH,
1958 high_wmark_pages(zone) - low_wmark_pages(zone) -
1959 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
1960 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1961 } while (zone++ != preferred_zone);
1965 * get_page_from_freelist goes through the zonelist trying to allocate
1968 static struct page *
1969 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1970 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1971 struct zone *preferred_zone, int classzone_idx, int migratetype)
1974 struct page *page = NULL;
1976 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1977 int zlc_active = 0; /* set if using zonelist_cache */
1978 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1979 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
1980 (gfp_mask & __GFP_WRITE);
1981 int nr_fair_skipped = 0;
1982 bool zonelist_rescan;
1985 zonelist_rescan = false;
1988 * Scan zonelist, looking for a zone with enough free.
1989 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1991 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1992 high_zoneidx, nodemask) {
1995 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1996 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1998 if (cpusets_enabled() &&
1999 (alloc_flags & ALLOC_CPUSET) &&
2000 !cpuset_zone_allowed_softwall(zone, gfp_mask))
2003 * Distribute pages in proportion to the individual
2004 * zone size to ensure fair page aging. The zone a
2005 * page was allocated in should have no effect on the
2006 * time the page has in memory before being reclaimed.
2008 if (alloc_flags & ALLOC_FAIR) {
2009 if (!zone_local(preferred_zone, zone))
2011 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2017 * When allocating a page cache page for writing, we
2018 * want to get it from a zone that is within its dirty
2019 * limit, such that no single zone holds more than its
2020 * proportional share of globally allowed dirty pages.
2021 * The dirty limits take into account the zone's
2022 * lowmem reserves and high watermark so that kswapd
2023 * should be able to balance it without having to
2024 * write pages from its LRU list.
2026 * This may look like it could increase pressure on
2027 * lower zones by failing allocations in higher zones
2028 * before they are full. But the pages that do spill
2029 * over are limited as the lower zones are protected
2030 * by this very same mechanism. It should not become
2031 * a practical burden to them.
2033 * XXX: For now, allow allocations to potentially
2034 * exceed the per-zone dirty limit in the slowpath
2035 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2036 * which is important when on a NUMA setup the allowed
2037 * zones are together not big enough to reach the
2038 * global limit. The proper fix for these situations
2039 * will require awareness of zones in the
2040 * dirty-throttling and the flusher threads.
2042 if (consider_zone_dirty && !zone_dirty_ok(zone))
2045 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2046 if (!zone_watermark_ok(zone, order, mark,
2047 classzone_idx, alloc_flags)) {
2050 /* Checked here to keep the fast path fast */
2051 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2052 if (alloc_flags & ALLOC_NO_WATERMARKS)
2055 if (IS_ENABLED(CONFIG_NUMA) &&
2056 !did_zlc_setup && nr_online_nodes > 1) {
2058 * we do zlc_setup if there are multiple nodes
2059 * and before considering the first zone allowed
2062 allowednodes = zlc_setup(zonelist, alloc_flags);
2067 if (zone_reclaim_mode == 0 ||
2068 !zone_allows_reclaim(preferred_zone, zone))
2069 goto this_zone_full;
2072 * As we may have just activated ZLC, check if the first
2073 * eligible zone has failed zone_reclaim recently.
2075 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2076 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2079 ret = zone_reclaim(zone, gfp_mask, order);
2081 case ZONE_RECLAIM_NOSCAN:
2084 case ZONE_RECLAIM_FULL:
2085 /* scanned but unreclaimable */
2088 /* did we reclaim enough */
2089 if (zone_watermark_ok(zone, order, mark,
2090 classzone_idx, alloc_flags))
2094 * Failed to reclaim enough to meet watermark.
2095 * Only mark the zone full if checking the min
2096 * watermark or if we failed to reclaim just
2097 * 1<<order pages or else the page allocator
2098 * fastpath will prematurely mark zones full
2099 * when the watermark is between the low and
2102 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2103 ret == ZONE_RECLAIM_SOME)
2104 goto this_zone_full;
2111 page = buffered_rmqueue(preferred_zone, zone, order,
2112 gfp_mask, migratetype);
2116 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2117 zlc_mark_zone_full(zonelist, z);
2122 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2123 * necessary to allocate the page. The expectation is
2124 * that the caller is taking steps that will free more
2125 * memory. The caller should avoid the page being used
2126 * for !PFMEMALLOC purposes.
2128 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2133 * The first pass makes sure allocations are spread fairly within the
2134 * local node. However, the local node might have free pages left
2135 * after the fairness batches are exhausted, and remote zones haven't
2136 * even been considered yet. Try once more without fairness, and
2137 * include remote zones now, before entering the slowpath and waking
2138 * kswapd: prefer spilling to a remote zone over swapping locally.
2140 if (alloc_flags & ALLOC_FAIR) {
2141 alloc_flags &= ~ALLOC_FAIR;
2142 if (nr_fair_skipped) {
2143 zonelist_rescan = true;
2144 reset_alloc_batches(preferred_zone);
2146 if (nr_online_nodes > 1)
2147 zonelist_rescan = true;
2150 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2151 /* Disable zlc cache for second zonelist scan */
2153 zonelist_rescan = true;
2156 if (zonelist_rescan)
2163 * Large machines with many possible nodes should not always dump per-node
2164 * meminfo in irq context.
2166 static inline bool should_suppress_show_mem(void)
2171 ret = in_interrupt();
2176 static DEFINE_RATELIMIT_STATE(nopage_rs,
2177 DEFAULT_RATELIMIT_INTERVAL,
2178 DEFAULT_RATELIMIT_BURST);
2180 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2182 unsigned int filter = SHOW_MEM_FILTER_NODES;
2184 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2185 debug_guardpage_minorder() > 0)
2189 * This documents exceptions given to allocations in certain
2190 * contexts that are allowed to allocate outside current's set
2193 if (!(gfp_mask & __GFP_NOMEMALLOC))
2194 if (test_thread_flag(TIF_MEMDIE) ||
2195 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2196 filter &= ~SHOW_MEM_FILTER_NODES;
2197 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2198 filter &= ~SHOW_MEM_FILTER_NODES;
2201 struct va_format vaf;
2204 va_start(args, fmt);
2209 pr_warn("%pV", &vaf);
2214 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2215 current->comm, order, gfp_mask);
2218 if (!should_suppress_show_mem())
2223 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2224 unsigned long did_some_progress,
2225 unsigned long pages_reclaimed)
2227 /* Do not loop if specifically requested */
2228 if (gfp_mask & __GFP_NORETRY)
2231 /* Always retry if specifically requested */
2232 if (gfp_mask & __GFP_NOFAIL)
2236 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2237 * making forward progress without invoking OOM. Suspend also disables
2238 * storage devices so kswapd will not help. Bail if we are suspending.
2240 if (!did_some_progress && pm_suspended_storage())
2244 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2245 * means __GFP_NOFAIL, but that may not be true in other
2248 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2252 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2253 * specified, then we retry until we no longer reclaim any pages
2254 * (above), or we've reclaimed an order of pages at least as
2255 * large as the allocation's order. In both cases, if the
2256 * allocation still fails, we stop retrying.
2258 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2264 static inline struct page *
2265 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2266 struct zonelist *zonelist, enum zone_type high_zoneidx,
2267 nodemask_t *nodemask, struct zone *preferred_zone,
2268 int classzone_idx, int migratetype)
2272 /* Acquire the per-zone oom lock for each zone */
2273 if (!oom_zonelist_trylock(zonelist, gfp_mask)) {
2274 schedule_timeout_uninterruptible(1);
2279 * PM-freezer should be notified that there might be an OOM killer on
2280 * its way to kill and wake somebody up. This is too early and we might
2281 * end up not killing anything but false positives are acceptable.
2282 * See freeze_processes.
2287 * Go through the zonelist yet one more time, keep very high watermark
2288 * here, this is only to catch a parallel oom killing, we must fail if
2289 * we're still under heavy pressure.
2291 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2292 order, zonelist, high_zoneidx,
2293 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2294 preferred_zone, classzone_idx, migratetype);
2298 if (!(gfp_mask & __GFP_NOFAIL)) {
2299 /* The OOM killer will not help higher order allocs */
2300 if (order > PAGE_ALLOC_COSTLY_ORDER)
2302 /* The OOM killer does not needlessly kill tasks for lowmem */
2303 if (high_zoneidx < ZONE_NORMAL)
2306 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2307 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2308 * The caller should handle page allocation failure by itself if
2309 * it specifies __GFP_THISNODE.
2310 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2312 if (gfp_mask & __GFP_THISNODE)
2315 /* Exhausted what can be done so it's blamo time */
2316 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2319 oom_zonelist_unlock(zonelist, gfp_mask);
2323 #ifdef CONFIG_COMPACTION
2324 /* Try memory compaction for high-order allocations before reclaim */
2325 static struct page *
2326 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2327 struct zonelist *zonelist, enum zone_type high_zoneidx,
2328 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2329 int classzone_idx, int migratetype, enum migrate_mode mode,
2330 int *contended_compaction, bool *deferred_compaction)
2332 unsigned long compact_result;
2338 current->flags |= PF_MEMALLOC;
2339 compact_result = try_to_compact_pages(zonelist, order, gfp_mask,
2341 contended_compaction,
2342 alloc_flags, classzone_idx);
2343 current->flags &= ~PF_MEMALLOC;
2345 switch (compact_result) {
2346 case COMPACT_DEFERRED:
2347 *deferred_compaction = true;
2349 case COMPACT_SKIPPED:
2356 * At least in one zone compaction wasn't deferred or skipped, so let's
2357 * count a compaction stall
2359 count_vm_event(COMPACTSTALL);
2361 page = get_page_from_freelist(gfp_mask, nodemask,
2362 order, zonelist, high_zoneidx,
2363 alloc_flags & ~ALLOC_NO_WATERMARKS,
2364 preferred_zone, classzone_idx, migratetype);
2367 struct zone *zone = page_zone(page);
2369 zone->compact_blockskip_flush = false;
2370 compaction_defer_reset(zone, order, true);
2371 count_vm_event(COMPACTSUCCESS);
2376 * It's bad if compaction run occurs and fails. The most likely reason
2377 * is that pages exist, but not enough to satisfy watermarks.
2379 count_vm_event(COMPACTFAIL);
2386 static inline struct page *
2387 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2388 struct zonelist *zonelist, enum zone_type high_zoneidx,
2389 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2390 int classzone_idx, int migratetype, enum migrate_mode mode,
2391 int *contended_compaction, bool *deferred_compaction)
2395 #endif /* CONFIG_COMPACTION */
2397 /* Perform direct synchronous page reclaim */
2399 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2400 nodemask_t *nodemask)
2402 struct reclaim_state reclaim_state;
2407 /* We now go into synchronous reclaim */
2408 cpuset_memory_pressure_bump();
2409 current->flags |= PF_MEMALLOC;
2410 lockdep_set_current_reclaim_state(gfp_mask);
2411 reclaim_state.reclaimed_slab = 0;
2412 current->reclaim_state = &reclaim_state;
2414 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2416 current->reclaim_state = NULL;
2417 lockdep_clear_current_reclaim_state();
2418 current->flags &= ~PF_MEMALLOC;
2425 /* The really slow allocator path where we enter direct reclaim */
2426 static inline struct page *
2427 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2428 struct zonelist *zonelist, enum zone_type high_zoneidx,
2429 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2430 int classzone_idx, int migratetype, unsigned long *did_some_progress)
2432 struct page *page = NULL;
2433 bool drained = false;
2435 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2437 if (unlikely(!(*did_some_progress)))
2440 /* After successful reclaim, reconsider all zones for allocation */
2441 if (IS_ENABLED(CONFIG_NUMA))
2442 zlc_clear_zones_full(zonelist);
2445 page = get_page_from_freelist(gfp_mask, nodemask, order,
2446 zonelist, high_zoneidx,
2447 alloc_flags & ~ALLOC_NO_WATERMARKS,
2448 preferred_zone, classzone_idx,
2452 * If an allocation failed after direct reclaim, it could be because
2453 * pages are pinned on the per-cpu lists. Drain them and try again
2455 if (!page && !drained) {
2456 drain_all_pages(NULL);
2465 * This is called in the allocator slow-path if the allocation request is of
2466 * sufficient urgency to ignore watermarks and take other desperate measures
2468 static inline struct page *
2469 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2470 struct zonelist *zonelist, enum zone_type high_zoneidx,
2471 nodemask_t *nodemask, struct zone *preferred_zone,
2472 int classzone_idx, int migratetype)
2477 page = get_page_from_freelist(gfp_mask, nodemask, order,
2478 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2479 preferred_zone, classzone_idx, migratetype);
2481 if (!page && gfp_mask & __GFP_NOFAIL)
2482 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2483 } while (!page && (gfp_mask & __GFP_NOFAIL));
2488 static void wake_all_kswapds(unsigned int order,
2489 struct zonelist *zonelist,
2490 enum zone_type high_zoneidx,
2491 struct zone *preferred_zone,
2492 nodemask_t *nodemask)
2497 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2498 high_zoneidx, nodemask)
2499 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2503 gfp_to_alloc_flags(gfp_t gfp_mask)
2505 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2506 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2508 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2509 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2512 * The caller may dip into page reserves a bit more if the caller
2513 * cannot run direct reclaim, or if the caller has realtime scheduling
2514 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2515 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2517 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2521 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2522 * if it can't schedule.
2524 if (!(gfp_mask & __GFP_NOMEMALLOC))
2525 alloc_flags |= ALLOC_HARDER;
2527 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2528 * comment for __cpuset_node_allowed_softwall().
2530 alloc_flags &= ~ALLOC_CPUSET;
2531 } else if (unlikely(rt_task(current)) && !in_interrupt())
2532 alloc_flags |= ALLOC_HARDER;
2534 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2535 if (gfp_mask & __GFP_MEMALLOC)
2536 alloc_flags |= ALLOC_NO_WATERMARKS;
2537 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2538 alloc_flags |= ALLOC_NO_WATERMARKS;
2539 else if (!in_interrupt() &&
2540 ((current->flags & PF_MEMALLOC) ||
2541 unlikely(test_thread_flag(TIF_MEMDIE))))
2542 alloc_flags |= ALLOC_NO_WATERMARKS;
2545 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2546 alloc_flags |= ALLOC_CMA;
2551 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2553 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2556 static inline struct page *
2557 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2558 struct zonelist *zonelist, enum zone_type high_zoneidx,
2559 nodemask_t *nodemask, struct zone *preferred_zone,
2560 int classzone_idx, int migratetype)
2562 const gfp_t wait = gfp_mask & __GFP_WAIT;
2563 struct page *page = NULL;
2565 unsigned long pages_reclaimed = 0;
2566 unsigned long did_some_progress;
2567 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2568 bool deferred_compaction = false;
2569 int contended_compaction = COMPACT_CONTENDED_NONE;
2572 * In the slowpath, we sanity check order to avoid ever trying to
2573 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2574 * be using allocators in order of preference for an area that is
2577 if (order >= MAX_ORDER) {
2578 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2583 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2584 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2585 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2586 * using a larger set of nodes after it has established that the
2587 * allowed per node queues are empty and that nodes are
2590 if (IS_ENABLED(CONFIG_NUMA) &&
2591 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2595 if (!(gfp_mask & __GFP_NO_KSWAPD))
2596 wake_all_kswapds(order, zonelist, high_zoneidx,
2597 preferred_zone, nodemask);
2600 * OK, we're below the kswapd watermark and have kicked background
2601 * reclaim. Now things get more complex, so set up alloc_flags according
2602 * to how we want to proceed.
2604 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2607 * Find the true preferred zone if the allocation is unconstrained by
2610 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) {
2611 struct zoneref *preferred_zoneref;
2612 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2613 NULL, &preferred_zone);
2614 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2618 /* This is the last chance, in general, before the goto nopage. */
2619 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2620 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2621 preferred_zone, classzone_idx, migratetype);
2625 /* Allocate without watermarks if the context allows */
2626 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2628 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2629 * the allocation is high priority and these type of
2630 * allocations are system rather than user orientated
2632 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2634 page = __alloc_pages_high_priority(gfp_mask, order,
2635 zonelist, high_zoneidx, nodemask,
2636 preferred_zone, classzone_idx, migratetype);
2642 /* Atomic allocations - we can't balance anything */
2645 * All existing users of the deprecated __GFP_NOFAIL are
2646 * blockable, so warn of any new users that actually allow this
2647 * type of allocation to fail.
2649 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2653 /* Avoid recursion of direct reclaim */
2654 if (current->flags & PF_MEMALLOC)
2657 /* Avoid allocations with no watermarks from looping endlessly */
2658 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2662 * Try direct compaction. The first pass is asynchronous. Subsequent
2663 * attempts after direct reclaim are synchronous
2665 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2666 high_zoneidx, nodemask, alloc_flags,
2668 classzone_idx, migratetype,
2669 migration_mode, &contended_compaction,
2670 &deferred_compaction);
2674 /* Checks for THP-specific high-order allocations */
2675 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2677 * If compaction is deferred for high-order allocations, it is
2678 * because sync compaction recently failed. If this is the case
2679 * and the caller requested a THP allocation, we do not want
2680 * to heavily disrupt the system, so we fail the allocation
2681 * instead of entering direct reclaim.
2683 if (deferred_compaction)
2687 * In all zones where compaction was attempted (and not
2688 * deferred or skipped), lock contention has been detected.
2689 * For THP allocation we do not want to disrupt the others
2690 * so we fallback to base pages instead.
2692 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2696 * If compaction was aborted due to need_resched(), we do not
2697 * want to further increase allocation latency, unless it is
2698 * khugepaged trying to collapse.
2700 if (contended_compaction == COMPACT_CONTENDED_SCHED
2701 && !(current->flags & PF_KTHREAD))
2706 * It can become very expensive to allocate transparent hugepages at
2707 * fault, so use asynchronous memory compaction for THP unless it is
2708 * khugepaged trying to collapse.
2710 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2711 (current->flags & PF_KTHREAD))
2712 migration_mode = MIGRATE_SYNC_LIGHT;
2714 /* Try direct reclaim and then allocating */
2715 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2716 zonelist, high_zoneidx,
2718 alloc_flags, preferred_zone,
2719 classzone_idx, migratetype,
2720 &did_some_progress);
2725 * If we failed to make any progress reclaiming, then we are
2726 * running out of options and have to consider going OOM
2728 if (!did_some_progress) {
2729 if (oom_gfp_allowed(gfp_mask)) {
2730 if (oom_killer_disabled)
2732 /* Coredumps can quickly deplete all memory reserves */
2733 if ((current->flags & PF_DUMPCORE) &&
2734 !(gfp_mask & __GFP_NOFAIL))
2736 page = __alloc_pages_may_oom(gfp_mask, order,
2737 zonelist, high_zoneidx,
2738 nodemask, preferred_zone,
2739 classzone_idx, migratetype);
2743 if (!(gfp_mask & __GFP_NOFAIL)) {
2745 * The oom killer is not called for high-order
2746 * allocations that may fail, so if no progress
2747 * is being made, there are no other options and
2748 * retrying is unlikely to help.
2750 if (order > PAGE_ALLOC_COSTLY_ORDER)
2753 * The oom killer is not called for lowmem
2754 * allocations to prevent needlessly killing
2757 if (high_zoneidx < ZONE_NORMAL)
2765 /* Check if we should retry the allocation */
2766 pages_reclaimed += did_some_progress;
2767 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2769 /* Wait for some write requests to complete then retry */
2770 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2774 * High-order allocations do not necessarily loop after
2775 * direct reclaim and reclaim/compaction depends on compaction
2776 * being called after reclaim so call directly if necessary
2778 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2779 high_zoneidx, nodemask, alloc_flags,
2781 classzone_idx, migratetype,
2782 migration_mode, &contended_compaction,
2783 &deferred_compaction);
2789 warn_alloc_failed(gfp_mask, order, NULL);
2792 if (kmemcheck_enabled)
2793 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2799 * This is the 'heart' of the zoned buddy allocator.
2802 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2803 struct zonelist *zonelist, nodemask_t *nodemask)
2805 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2806 struct zone *preferred_zone;
2807 struct zoneref *preferred_zoneref;
2808 struct page *page = NULL;
2809 int migratetype = gfpflags_to_migratetype(gfp_mask);
2810 unsigned int cpuset_mems_cookie;
2811 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2814 gfp_mask &= gfp_allowed_mask;
2816 lockdep_trace_alloc(gfp_mask);
2818 might_sleep_if(gfp_mask & __GFP_WAIT);
2820 if (should_fail_alloc_page(gfp_mask, order))
2824 * Check the zones suitable for the gfp_mask contain at least one
2825 * valid zone. It's possible to have an empty zonelist as a result
2826 * of GFP_THISNODE and a memoryless node
2828 if (unlikely(!zonelist->_zonerefs->zone))
2831 if (IS_ENABLED(CONFIG_CMA) && migratetype == MIGRATE_MOVABLE)
2832 alloc_flags |= ALLOC_CMA;
2835 cpuset_mems_cookie = read_mems_allowed_begin();
2837 /* The preferred zone is used for statistics later */
2838 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2839 nodemask ? : &cpuset_current_mems_allowed,
2841 if (!preferred_zone)
2843 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2845 /* First allocation attempt */
2846 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2847 zonelist, high_zoneidx, alloc_flags,
2848 preferred_zone, classzone_idx, migratetype);
2849 if (unlikely(!page)) {
2851 * Runtime PM, block IO and its error handling path
2852 * can deadlock because I/O on the device might not
2855 gfp_mask = memalloc_noio_flags(gfp_mask);
2856 page = __alloc_pages_slowpath(gfp_mask, order,
2857 zonelist, high_zoneidx, nodemask,
2858 preferred_zone, classzone_idx, migratetype);
2861 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2865 * When updating a task's mems_allowed, it is possible to race with
2866 * parallel threads in such a way that an allocation can fail while
2867 * the mask is being updated. If a page allocation is about to fail,
2868 * check if the cpuset changed during allocation and if so, retry.
2870 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2875 EXPORT_SYMBOL(__alloc_pages_nodemask);
2878 * Common helper functions.
2880 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2885 * __get_free_pages() returns a 32-bit address, which cannot represent
2888 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2890 page = alloc_pages(gfp_mask, order);
2893 return (unsigned long) page_address(page);
2895 EXPORT_SYMBOL(__get_free_pages);
2897 unsigned long get_zeroed_page(gfp_t gfp_mask)
2899 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2901 EXPORT_SYMBOL(get_zeroed_page);
2903 void __free_pages(struct page *page, unsigned int order)
2905 if (put_page_testzero(page)) {
2907 free_hot_cold_page(page, false);
2909 __free_pages_ok(page, order);
2913 EXPORT_SYMBOL(__free_pages);
2915 void free_pages(unsigned long addr, unsigned int order)
2918 VM_BUG_ON(!virt_addr_valid((void *)addr));
2919 __free_pages(virt_to_page((void *)addr), order);
2923 EXPORT_SYMBOL(free_pages);
2926 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2927 * of the current memory cgroup.
2929 * It should be used when the caller would like to use kmalloc, but since the
2930 * allocation is large, it has to fall back to the page allocator.
2932 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2935 struct mem_cgroup *memcg = NULL;
2937 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2939 page = alloc_pages(gfp_mask, order);
2940 memcg_kmem_commit_charge(page, memcg, order);
2944 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2947 struct mem_cgroup *memcg = NULL;
2949 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2951 page = alloc_pages_node(nid, gfp_mask, order);
2952 memcg_kmem_commit_charge(page, memcg, order);
2957 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2960 void __free_kmem_pages(struct page *page, unsigned int order)
2962 memcg_kmem_uncharge_pages(page, order);
2963 __free_pages(page, order);
2966 void free_kmem_pages(unsigned long addr, unsigned int order)
2969 VM_BUG_ON(!virt_addr_valid((void *)addr));
2970 __free_kmem_pages(virt_to_page((void *)addr), order);
2974 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2977 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2978 unsigned long used = addr + PAGE_ALIGN(size);
2980 split_page(virt_to_page((void *)addr), order);
2981 while (used < alloc_end) {
2986 return (void *)addr;
2990 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2991 * @size: the number of bytes to allocate
2992 * @gfp_mask: GFP flags for the allocation
2994 * This function is similar to alloc_pages(), except that it allocates the
2995 * minimum number of pages to satisfy the request. alloc_pages() can only
2996 * allocate memory in power-of-two pages.
2998 * This function is also limited by MAX_ORDER.
3000 * Memory allocated by this function must be released by free_pages_exact().
3002 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3004 unsigned int order = get_order(size);
3007 addr = __get_free_pages(gfp_mask, order);
3008 return make_alloc_exact(addr, order, size);
3010 EXPORT_SYMBOL(alloc_pages_exact);
3013 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3015 * @nid: the preferred node ID where memory should be allocated
3016 * @size: the number of bytes to allocate
3017 * @gfp_mask: GFP flags for the allocation
3019 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3021 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3024 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3026 unsigned order = get_order(size);
3027 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3030 return make_alloc_exact((unsigned long)page_address(p), order, size);
3034 * free_pages_exact - release memory allocated via alloc_pages_exact()
3035 * @virt: the value returned by alloc_pages_exact.
3036 * @size: size of allocation, same value as passed to alloc_pages_exact().
3038 * Release the memory allocated by a previous call to alloc_pages_exact.
3040 void free_pages_exact(void *virt, size_t size)
3042 unsigned long addr = (unsigned long)virt;
3043 unsigned long end = addr + PAGE_ALIGN(size);
3045 while (addr < end) {
3050 EXPORT_SYMBOL(free_pages_exact);
3053 * nr_free_zone_pages - count number of pages beyond high watermark
3054 * @offset: The zone index of the highest zone
3056 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3057 * high watermark within all zones at or below a given zone index. For each
3058 * zone, the number of pages is calculated as:
3059 * managed_pages - high_pages
3061 static unsigned long nr_free_zone_pages(int offset)
3066 /* Just pick one node, since fallback list is circular */
3067 unsigned long sum = 0;
3069 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3071 for_each_zone_zonelist(zone, z, zonelist, offset) {
3072 unsigned long size = zone->managed_pages;
3073 unsigned long high = high_wmark_pages(zone);
3082 * nr_free_buffer_pages - count number of pages beyond high watermark
3084 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3085 * watermark within ZONE_DMA and ZONE_NORMAL.
3087 unsigned long nr_free_buffer_pages(void)
3089 return nr_free_zone_pages(gfp_zone(GFP_USER));
3091 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3094 * nr_free_pagecache_pages - count number of pages beyond high watermark
3096 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3097 * high watermark within all zones.
3099 unsigned long nr_free_pagecache_pages(void)
3101 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3104 static inline void show_node(struct zone *zone)
3106 if (IS_ENABLED(CONFIG_NUMA))
3107 printk("Node %d ", zone_to_nid(zone));
3110 void si_meminfo(struct sysinfo *val)
3112 val->totalram = totalram_pages;
3113 val->sharedram = global_page_state(NR_SHMEM);
3114 val->freeram = global_page_state(NR_FREE_PAGES);
3115 val->bufferram = nr_blockdev_pages();
3116 val->totalhigh = totalhigh_pages;
3117 val->freehigh = nr_free_highpages();
3118 val->mem_unit = PAGE_SIZE;
3121 EXPORT_SYMBOL(si_meminfo);
3124 void si_meminfo_node(struct sysinfo *val, int nid)
3126 int zone_type; /* needs to be signed */
3127 unsigned long managed_pages = 0;
3128 pg_data_t *pgdat = NODE_DATA(nid);
3130 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3131 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3132 val->totalram = managed_pages;
3133 val->sharedram = node_page_state(nid, NR_SHMEM);
3134 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3135 #ifdef CONFIG_HIGHMEM
3136 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3137 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3143 val->mem_unit = PAGE_SIZE;
3148 * Determine whether the node should be displayed or not, depending on whether
3149 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3151 bool skip_free_areas_node(unsigned int flags, int nid)
3154 unsigned int cpuset_mems_cookie;
3156 if (!(flags & SHOW_MEM_FILTER_NODES))
3160 cpuset_mems_cookie = read_mems_allowed_begin();
3161 ret = !node_isset(nid, cpuset_current_mems_allowed);
3162 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3167 #define K(x) ((x) << (PAGE_SHIFT-10))
3169 static void show_migration_types(unsigned char type)
3171 static const char types[MIGRATE_TYPES] = {
3172 [MIGRATE_UNMOVABLE] = 'U',
3173 [MIGRATE_RECLAIMABLE] = 'E',
3174 [MIGRATE_MOVABLE] = 'M',
3175 [MIGRATE_RESERVE] = 'R',
3177 [MIGRATE_CMA] = 'C',
3179 #ifdef CONFIG_MEMORY_ISOLATION
3180 [MIGRATE_ISOLATE] = 'I',
3183 char tmp[MIGRATE_TYPES + 1];
3187 for (i = 0; i < MIGRATE_TYPES; i++) {
3188 if (type & (1 << i))
3193 printk("(%s) ", tmp);
3197 * Show free area list (used inside shift_scroll-lock stuff)
3198 * We also calculate the percentage fragmentation. We do this by counting the
3199 * memory on each free list with the exception of the first item on the list.
3200 * Suppresses nodes that are not allowed by current's cpuset if
3201 * SHOW_MEM_FILTER_NODES is passed.
3203 void show_free_areas(unsigned int filter)
3208 for_each_populated_zone(zone) {
3209 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3212 printk("%s per-cpu:\n", zone->name);
3214 for_each_online_cpu(cpu) {
3215 struct per_cpu_pageset *pageset;
3217 pageset = per_cpu_ptr(zone->pageset, cpu);
3219 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3220 cpu, pageset->pcp.high,
3221 pageset->pcp.batch, pageset->pcp.count);
3225 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3226 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3228 " dirty:%lu writeback:%lu unstable:%lu\n"
3229 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3230 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3232 global_page_state(NR_ACTIVE_ANON),
3233 global_page_state(NR_INACTIVE_ANON),
3234 global_page_state(NR_ISOLATED_ANON),
3235 global_page_state(NR_ACTIVE_FILE),
3236 global_page_state(NR_INACTIVE_FILE),
3237 global_page_state(NR_ISOLATED_FILE),
3238 global_page_state(NR_UNEVICTABLE),
3239 global_page_state(NR_FILE_DIRTY),
3240 global_page_state(NR_WRITEBACK),
3241 global_page_state(NR_UNSTABLE_NFS),
3242 global_page_state(NR_FREE_PAGES),
3243 global_page_state(NR_SLAB_RECLAIMABLE),
3244 global_page_state(NR_SLAB_UNRECLAIMABLE),
3245 global_page_state(NR_FILE_MAPPED),
3246 global_page_state(NR_SHMEM),
3247 global_page_state(NR_PAGETABLE),
3248 global_page_state(NR_BOUNCE),
3249 global_page_state(NR_FREE_CMA_PAGES));
3251 for_each_populated_zone(zone) {
3254 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3262 " active_anon:%lukB"
3263 " inactive_anon:%lukB"
3264 " active_file:%lukB"
3265 " inactive_file:%lukB"
3266 " unevictable:%lukB"
3267 " isolated(anon):%lukB"
3268 " isolated(file):%lukB"
3276 " slab_reclaimable:%lukB"
3277 " slab_unreclaimable:%lukB"
3278 " kernel_stack:%lukB"
3283 " writeback_tmp:%lukB"
3284 " pages_scanned:%lu"
3285 " all_unreclaimable? %s"
3288 K(zone_page_state(zone, NR_FREE_PAGES)),
3289 K(min_wmark_pages(zone)),
3290 K(low_wmark_pages(zone)),
3291 K(high_wmark_pages(zone)),
3292 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3293 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3294 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3295 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3296 K(zone_page_state(zone, NR_UNEVICTABLE)),
3297 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3298 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3299 K(zone->present_pages),
3300 K(zone->managed_pages),
3301 K(zone_page_state(zone, NR_MLOCK)),
3302 K(zone_page_state(zone, NR_FILE_DIRTY)),
3303 K(zone_page_state(zone, NR_WRITEBACK)),
3304 K(zone_page_state(zone, NR_FILE_MAPPED)),
3305 K(zone_page_state(zone, NR_SHMEM)),
3306 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3307 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3308 zone_page_state(zone, NR_KERNEL_STACK) *
3310 K(zone_page_state(zone, NR_PAGETABLE)),
3311 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3312 K(zone_page_state(zone, NR_BOUNCE)),
3313 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3314 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3315 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3316 (!zone_reclaimable(zone) ? "yes" : "no")
3318 printk("lowmem_reserve[]:");
3319 for (i = 0; i < MAX_NR_ZONES; i++)
3320 printk(" %ld", zone->lowmem_reserve[i]);
3324 for_each_populated_zone(zone) {
3325 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3326 unsigned char types[MAX_ORDER];
3328 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3331 printk("%s: ", zone->name);
3333 spin_lock_irqsave(&zone->lock, flags);
3334 for (order = 0; order < MAX_ORDER; order++) {
3335 struct free_area *area = &zone->free_area[order];
3338 nr[order] = area->nr_free;
3339 total += nr[order] << order;
3342 for (type = 0; type < MIGRATE_TYPES; type++) {
3343 if (!list_empty(&area->free_list[type]))
3344 types[order] |= 1 << type;
3347 spin_unlock_irqrestore(&zone->lock, flags);
3348 for (order = 0; order < MAX_ORDER; order++) {
3349 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3351 show_migration_types(types[order]);
3353 printk("= %lukB\n", K(total));
3356 hugetlb_show_meminfo();
3358 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3360 show_swap_cache_info();
3363 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3365 zoneref->zone = zone;
3366 zoneref->zone_idx = zone_idx(zone);
3370 * Builds allocation fallback zone lists.
3372 * Add all populated zones of a node to the zonelist.
3374 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3378 enum zone_type zone_type = MAX_NR_ZONES;
3382 zone = pgdat->node_zones + zone_type;
3383 if (populated_zone(zone)) {
3384 zoneref_set_zone(zone,
3385 &zonelist->_zonerefs[nr_zones++]);
3386 check_highest_zone(zone_type);
3388 } while (zone_type);
3396 * 0 = automatic detection of better ordering.
3397 * 1 = order by ([node] distance, -zonetype)
3398 * 2 = order by (-zonetype, [node] distance)
3400 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3401 * the same zonelist. So only NUMA can configure this param.
3403 #define ZONELIST_ORDER_DEFAULT 0
3404 #define ZONELIST_ORDER_NODE 1
3405 #define ZONELIST_ORDER_ZONE 2
3407 /* zonelist order in the kernel.
3408 * set_zonelist_order() will set this to NODE or ZONE.
3410 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3411 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3415 /* The value user specified ....changed by config */
3416 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3417 /* string for sysctl */
3418 #define NUMA_ZONELIST_ORDER_LEN 16
3419 char numa_zonelist_order[16] = "default";
3422 * interface for configure zonelist ordering.
3423 * command line option "numa_zonelist_order"
3424 * = "[dD]efault - default, automatic configuration.
3425 * = "[nN]ode - order by node locality, then by zone within node
3426 * = "[zZ]one - order by zone, then by locality within zone
3429 static int __parse_numa_zonelist_order(char *s)
3431 if (*s == 'd' || *s == 'D') {
3432 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3433 } else if (*s == 'n' || *s == 'N') {
3434 user_zonelist_order = ZONELIST_ORDER_NODE;
3435 } else if (*s == 'z' || *s == 'Z') {
3436 user_zonelist_order = ZONELIST_ORDER_ZONE;
3439 "Ignoring invalid numa_zonelist_order value: "
3446 static __init int setup_numa_zonelist_order(char *s)
3453 ret = __parse_numa_zonelist_order(s);
3455 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3459 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3462 * sysctl handler for numa_zonelist_order
3464 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3465 void __user *buffer, size_t *length,
3468 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3470 static DEFINE_MUTEX(zl_order_mutex);
3472 mutex_lock(&zl_order_mutex);
3474 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3478 strcpy(saved_string, (char *)table->data);
3480 ret = proc_dostring(table, write, buffer, length, ppos);
3484 int oldval = user_zonelist_order;
3486 ret = __parse_numa_zonelist_order((char *)table->data);
3489 * bogus value. restore saved string
3491 strncpy((char *)table->data, saved_string,
3492 NUMA_ZONELIST_ORDER_LEN);
3493 user_zonelist_order = oldval;
3494 } else if (oldval != user_zonelist_order) {
3495 mutex_lock(&zonelists_mutex);
3496 build_all_zonelists(NULL, NULL);
3497 mutex_unlock(&zonelists_mutex);
3501 mutex_unlock(&zl_order_mutex);
3506 #define MAX_NODE_LOAD (nr_online_nodes)
3507 static int node_load[MAX_NUMNODES];
3510 * find_next_best_node - find the next node that should appear in a given node's fallback list
3511 * @node: node whose fallback list we're appending
3512 * @used_node_mask: nodemask_t of already used nodes
3514 * We use a number of factors to determine which is the next node that should
3515 * appear on a given node's fallback list. The node should not have appeared
3516 * already in @node's fallback list, and it should be the next closest node
3517 * according to the distance array (which contains arbitrary distance values
3518 * from each node to each node in the system), and should also prefer nodes
3519 * with no CPUs, since presumably they'll have very little allocation pressure
3520 * on them otherwise.
3521 * It returns -1 if no node is found.
3523 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3526 int min_val = INT_MAX;
3527 int best_node = NUMA_NO_NODE;
3528 const struct cpumask *tmp = cpumask_of_node(0);
3530 /* Use the local node if we haven't already */
3531 if (!node_isset(node, *used_node_mask)) {
3532 node_set(node, *used_node_mask);
3536 for_each_node_state(n, N_MEMORY) {
3538 /* Don't want a node to appear more than once */
3539 if (node_isset(n, *used_node_mask))
3542 /* Use the distance array to find the distance */
3543 val = node_distance(node, n);
3545 /* Penalize nodes under us ("prefer the next node") */
3548 /* Give preference to headless and unused nodes */
3549 tmp = cpumask_of_node(n);
3550 if (!cpumask_empty(tmp))
3551 val += PENALTY_FOR_NODE_WITH_CPUS;
3553 /* Slight preference for less loaded node */
3554 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3555 val += node_load[n];
3557 if (val < min_val) {
3564 node_set(best_node, *used_node_mask);
3571 * Build zonelists ordered by node and zones within node.
3572 * This results in maximum locality--normal zone overflows into local
3573 * DMA zone, if any--but risks exhausting DMA zone.
3575 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3578 struct zonelist *zonelist;
3580 zonelist = &pgdat->node_zonelists[0];
3581 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3583 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3584 zonelist->_zonerefs[j].zone = NULL;
3585 zonelist->_zonerefs[j].zone_idx = 0;
3589 * Build gfp_thisnode zonelists
3591 static void build_thisnode_zonelists(pg_data_t *pgdat)
3594 struct zonelist *zonelist;
3596 zonelist = &pgdat->node_zonelists[1];
3597 j = build_zonelists_node(pgdat, zonelist, 0);
3598 zonelist->_zonerefs[j].zone = NULL;
3599 zonelist->_zonerefs[j].zone_idx = 0;
3603 * Build zonelists ordered by zone and nodes within zones.
3604 * This results in conserving DMA zone[s] until all Normal memory is
3605 * exhausted, but results in overflowing to remote node while memory
3606 * may still exist in local DMA zone.
3608 static int node_order[MAX_NUMNODES];
3610 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3613 int zone_type; /* needs to be signed */
3615 struct zonelist *zonelist;
3617 zonelist = &pgdat->node_zonelists[0];
3619 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3620 for (j = 0; j < nr_nodes; j++) {
3621 node = node_order[j];
3622 z = &NODE_DATA(node)->node_zones[zone_type];
3623 if (populated_zone(z)) {
3625 &zonelist->_zonerefs[pos++]);
3626 check_highest_zone(zone_type);
3630 zonelist->_zonerefs[pos].zone = NULL;
3631 zonelist->_zonerefs[pos].zone_idx = 0;
3634 #if defined(CONFIG_64BIT)
3636 * Devices that require DMA32/DMA are relatively rare and do not justify a
3637 * penalty to every machine in case the specialised case applies. Default
3638 * to Node-ordering on 64-bit NUMA machines
3640 static int default_zonelist_order(void)
3642 return ZONELIST_ORDER_NODE;
3646 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3647 * by the kernel. If processes running on node 0 deplete the low memory zone
3648 * then reclaim will occur more frequency increasing stalls and potentially
3649 * be easier to OOM if a large percentage of the zone is under writeback or
3650 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3651 * Hence, default to zone ordering on 32-bit.
3653 static int default_zonelist_order(void)
3655 return ZONELIST_ORDER_ZONE;
3657 #endif /* CONFIG_64BIT */
3659 static void set_zonelist_order(void)
3661 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3662 current_zonelist_order = default_zonelist_order();
3664 current_zonelist_order = user_zonelist_order;
3667 static void build_zonelists(pg_data_t *pgdat)
3671 nodemask_t used_mask;
3672 int local_node, prev_node;
3673 struct zonelist *zonelist;
3674 int order = current_zonelist_order;
3676 /* initialize zonelists */
3677 for (i = 0; i < MAX_ZONELISTS; i++) {
3678 zonelist = pgdat->node_zonelists + i;
3679 zonelist->_zonerefs[0].zone = NULL;
3680 zonelist->_zonerefs[0].zone_idx = 0;
3683 /* NUMA-aware ordering of nodes */
3684 local_node = pgdat->node_id;
3685 load = nr_online_nodes;
3686 prev_node = local_node;
3687 nodes_clear(used_mask);
3689 memset(node_order, 0, sizeof(node_order));
3692 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3694 * We don't want to pressure a particular node.
3695 * So adding penalty to the first node in same
3696 * distance group to make it round-robin.
3698 if (node_distance(local_node, node) !=
3699 node_distance(local_node, prev_node))
3700 node_load[node] = load;
3704 if (order == ZONELIST_ORDER_NODE)
3705 build_zonelists_in_node_order(pgdat, node);
3707 node_order[j++] = node; /* remember order */
3710 if (order == ZONELIST_ORDER_ZONE) {
3711 /* calculate node order -- i.e., DMA last! */
3712 build_zonelists_in_zone_order(pgdat, j);
3715 build_thisnode_zonelists(pgdat);
3718 /* Construct the zonelist performance cache - see further mmzone.h */
3719 static void build_zonelist_cache(pg_data_t *pgdat)
3721 struct zonelist *zonelist;
3722 struct zonelist_cache *zlc;
3725 zonelist = &pgdat->node_zonelists[0];
3726 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3727 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3728 for (z = zonelist->_zonerefs; z->zone; z++)
3729 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3732 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3734 * Return node id of node used for "local" allocations.
3735 * I.e., first node id of first zone in arg node's generic zonelist.
3736 * Used for initializing percpu 'numa_mem', which is used primarily
3737 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3739 int local_memory_node(int node)
3743 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3744 gfp_zone(GFP_KERNEL),
3751 #else /* CONFIG_NUMA */
3753 static void set_zonelist_order(void)
3755 current_zonelist_order = ZONELIST_ORDER_ZONE;
3758 static void build_zonelists(pg_data_t *pgdat)
3760 int node, local_node;
3762 struct zonelist *zonelist;
3764 local_node = pgdat->node_id;
3766 zonelist = &pgdat->node_zonelists[0];
3767 j = build_zonelists_node(pgdat, zonelist, 0);
3770 * Now we build the zonelist so that it contains the zones
3771 * of all the other nodes.
3772 * We don't want to pressure a particular node, so when
3773 * building the zones for node N, we make sure that the
3774 * zones coming right after the local ones are those from
3775 * node N+1 (modulo N)
3777 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3778 if (!node_online(node))
3780 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3782 for (node = 0; node < local_node; node++) {
3783 if (!node_online(node))
3785 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3788 zonelist->_zonerefs[j].zone = NULL;
3789 zonelist->_zonerefs[j].zone_idx = 0;
3792 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3793 static void build_zonelist_cache(pg_data_t *pgdat)
3795 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3798 #endif /* CONFIG_NUMA */
3801 * Boot pageset table. One per cpu which is going to be used for all
3802 * zones and all nodes. The parameters will be set in such a way
3803 * that an item put on a list will immediately be handed over to
3804 * the buddy list. This is safe since pageset manipulation is done
3805 * with interrupts disabled.
3807 * The boot_pagesets must be kept even after bootup is complete for
3808 * unused processors and/or zones. They do play a role for bootstrapping
3809 * hotplugged processors.
3811 * zoneinfo_show() and maybe other functions do
3812 * not check if the processor is online before following the pageset pointer.
3813 * Other parts of the kernel may not check if the zone is available.
3815 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3816 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3817 static void setup_zone_pageset(struct zone *zone);
3820 * Global mutex to protect against size modification of zonelists
3821 * as well as to serialize pageset setup for the new populated zone.
3823 DEFINE_MUTEX(zonelists_mutex);
3825 /* return values int ....just for stop_machine() */
3826 static int __build_all_zonelists(void *data)
3830 pg_data_t *self = data;
3833 memset(node_load, 0, sizeof(node_load));
3836 if (self && !node_online(self->node_id)) {
3837 build_zonelists(self);
3838 build_zonelist_cache(self);
3841 for_each_online_node(nid) {
3842 pg_data_t *pgdat = NODE_DATA(nid);
3844 build_zonelists(pgdat);
3845 build_zonelist_cache(pgdat);
3849 * Initialize the boot_pagesets that are going to be used
3850 * for bootstrapping processors. The real pagesets for
3851 * each zone will be allocated later when the per cpu
3852 * allocator is available.
3854 * boot_pagesets are used also for bootstrapping offline
3855 * cpus if the system is already booted because the pagesets
3856 * are needed to initialize allocators on a specific cpu too.
3857 * F.e. the percpu allocator needs the page allocator which
3858 * needs the percpu allocator in order to allocate its pagesets
3859 * (a chicken-egg dilemma).
3861 for_each_possible_cpu(cpu) {
3862 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3864 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3866 * We now know the "local memory node" for each node--
3867 * i.e., the node of the first zone in the generic zonelist.
3868 * Set up numa_mem percpu variable for on-line cpus. During
3869 * boot, only the boot cpu should be on-line; we'll init the
3870 * secondary cpus' numa_mem as they come on-line. During
3871 * node/memory hotplug, we'll fixup all on-line cpus.
3873 if (cpu_online(cpu))
3874 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3882 * Called with zonelists_mutex held always
3883 * unless system_state == SYSTEM_BOOTING.
3885 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3887 set_zonelist_order();
3889 if (system_state == SYSTEM_BOOTING) {
3890 __build_all_zonelists(NULL);
3891 mminit_verify_zonelist();
3892 cpuset_init_current_mems_allowed();
3894 #ifdef CONFIG_MEMORY_HOTPLUG
3896 setup_zone_pageset(zone);
3898 /* we have to stop all cpus to guarantee there is no user
3900 stop_machine(__build_all_zonelists, pgdat, NULL);
3901 /* cpuset refresh routine should be here */
3903 vm_total_pages = nr_free_pagecache_pages();
3905 * Disable grouping by mobility if the number of pages in the
3906 * system is too low to allow the mechanism to work. It would be
3907 * more accurate, but expensive to check per-zone. This check is
3908 * made on memory-hotadd so a system can start with mobility
3909 * disabled and enable it later
3911 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3912 page_group_by_mobility_disabled = 1;
3914 page_group_by_mobility_disabled = 0;
3916 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3917 "Total pages: %ld\n",
3919 zonelist_order_name[current_zonelist_order],
3920 page_group_by_mobility_disabled ? "off" : "on",
3923 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
3928 * Helper functions to size the waitqueue hash table.
3929 * Essentially these want to choose hash table sizes sufficiently
3930 * large so that collisions trying to wait on pages are rare.
3931 * But in fact, the number of active page waitqueues on typical
3932 * systems is ridiculously low, less than 200. So this is even
3933 * conservative, even though it seems large.
3935 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3936 * waitqueues, i.e. the size of the waitq table given the number of pages.
3938 #define PAGES_PER_WAITQUEUE 256
3940 #ifndef CONFIG_MEMORY_HOTPLUG
3941 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3943 unsigned long size = 1;
3945 pages /= PAGES_PER_WAITQUEUE;
3947 while (size < pages)
3951 * Once we have dozens or even hundreds of threads sleeping
3952 * on IO we've got bigger problems than wait queue collision.
3953 * Limit the size of the wait table to a reasonable size.
3955 size = min(size, 4096UL);
3957 return max(size, 4UL);
3961 * A zone's size might be changed by hot-add, so it is not possible to determine
3962 * a suitable size for its wait_table. So we use the maximum size now.
3964 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3966 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3967 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3968 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3970 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3971 * or more by the traditional way. (See above). It equals:
3973 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3974 * ia64(16K page size) : = ( 8G + 4M)byte.
3975 * powerpc (64K page size) : = (32G +16M)byte.
3977 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3984 * This is an integer logarithm so that shifts can be used later
3985 * to extract the more random high bits from the multiplicative
3986 * hash function before the remainder is taken.
3988 static inline unsigned long wait_table_bits(unsigned long size)
3994 * Check if a pageblock contains reserved pages
3996 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4000 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4001 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4008 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4009 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4010 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4011 * higher will lead to a bigger reserve which will get freed as contiguous
4012 * blocks as reclaim kicks in
4014 static void setup_zone_migrate_reserve(struct zone *zone)
4016 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4018 unsigned long block_migratetype;
4023 * Get the start pfn, end pfn and the number of blocks to reserve
4024 * We have to be careful to be aligned to pageblock_nr_pages to
4025 * make sure that we always check pfn_valid for the first page in
4028 start_pfn = zone->zone_start_pfn;
4029 end_pfn = zone_end_pfn(zone);
4030 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4031 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4035 * Reserve blocks are generally in place to help high-order atomic
4036 * allocations that are short-lived. A min_free_kbytes value that
4037 * would result in more than 2 reserve blocks for atomic allocations
4038 * is assumed to be in place to help anti-fragmentation for the
4039 * future allocation of hugepages at runtime.
4041 reserve = min(2, reserve);
4042 old_reserve = zone->nr_migrate_reserve_block;
4044 /* When memory hot-add, we almost always need to do nothing */
4045 if (reserve == old_reserve)
4047 zone->nr_migrate_reserve_block = reserve;
4049 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4050 if (!pfn_valid(pfn))
4052 page = pfn_to_page(pfn);
4054 /* Watch out for overlapping nodes */
4055 if (page_to_nid(page) != zone_to_nid(zone))
4058 block_migratetype = get_pageblock_migratetype(page);
4060 /* Only test what is necessary when the reserves are not met */
4063 * Blocks with reserved pages will never free, skip
4066 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4067 if (pageblock_is_reserved(pfn, block_end_pfn))
4070 /* If this block is reserved, account for it */
4071 if (block_migratetype == MIGRATE_RESERVE) {
4076 /* Suitable for reserving if this block is movable */
4077 if (block_migratetype == MIGRATE_MOVABLE) {
4078 set_pageblock_migratetype(page,
4080 move_freepages_block(zone, page,
4085 } else if (!old_reserve) {
4087 * At boot time we don't need to scan the whole zone
4088 * for turning off MIGRATE_RESERVE.
4094 * If the reserve is met and this is a previous reserved block,
4097 if (block_migratetype == MIGRATE_RESERVE) {
4098 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4099 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4105 * Initially all pages are reserved - free ones are freed
4106 * up by free_all_bootmem() once the early boot process is
4107 * done. Non-atomic initialization, single-pass.
4109 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4110 unsigned long start_pfn, enum memmap_context context)
4113 unsigned long end_pfn = start_pfn + size;
4117 if (highest_memmap_pfn < end_pfn - 1)
4118 highest_memmap_pfn = end_pfn - 1;
4120 z = &NODE_DATA(nid)->node_zones[zone];
4121 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4123 * There can be holes in boot-time mem_map[]s
4124 * handed to this function. They do not
4125 * exist on hotplugged memory.
4127 if (context == MEMMAP_EARLY) {
4128 if (!early_pfn_valid(pfn))
4130 if (!early_pfn_in_nid(pfn, nid))
4133 page = pfn_to_page(pfn);
4134 set_page_links(page, zone, nid, pfn);
4135 mminit_verify_page_links(page, zone, nid, pfn);
4136 init_page_count(page);
4137 page_mapcount_reset(page);
4138 page_cpupid_reset_last(page);
4139 SetPageReserved(page);
4141 * Mark the block movable so that blocks are reserved for
4142 * movable at startup. This will force kernel allocations
4143 * to reserve their blocks rather than leaking throughout
4144 * the address space during boot when many long-lived
4145 * kernel allocations are made. Later some blocks near
4146 * the start are marked MIGRATE_RESERVE by
4147 * setup_zone_migrate_reserve()
4149 * bitmap is created for zone's valid pfn range. but memmap
4150 * can be created for invalid pages (for alignment)
4151 * check here not to call set_pageblock_migratetype() against
4154 if ((z->zone_start_pfn <= pfn)
4155 && (pfn < zone_end_pfn(z))
4156 && !(pfn & (pageblock_nr_pages - 1)))
4157 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4159 INIT_LIST_HEAD(&page->lru);
4160 #ifdef WANT_PAGE_VIRTUAL
4161 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4162 if (!is_highmem_idx(zone))
4163 set_page_address(page, __va(pfn << PAGE_SHIFT));
4168 static void __meminit zone_init_free_lists(struct zone *zone)
4170 unsigned int order, t;
4171 for_each_migratetype_order(order, t) {
4172 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4173 zone->free_area[order].nr_free = 0;
4177 #ifndef __HAVE_ARCH_MEMMAP_INIT
4178 #define memmap_init(size, nid, zone, start_pfn) \
4179 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4182 static int zone_batchsize(struct zone *zone)
4188 * The per-cpu-pages pools are set to around 1000th of the
4189 * size of the zone. But no more than 1/2 of a meg.
4191 * OK, so we don't know how big the cache is. So guess.
4193 batch = zone->managed_pages / 1024;
4194 if (batch * PAGE_SIZE > 512 * 1024)
4195 batch = (512 * 1024) / PAGE_SIZE;
4196 batch /= 4; /* We effectively *= 4 below */
4201 * Clamp the batch to a 2^n - 1 value. Having a power
4202 * of 2 value was found to be more likely to have
4203 * suboptimal cache aliasing properties in some cases.
4205 * For example if 2 tasks are alternately allocating
4206 * batches of pages, one task can end up with a lot
4207 * of pages of one half of the possible page colors
4208 * and the other with pages of the other colors.
4210 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4215 /* The deferral and batching of frees should be suppressed under NOMMU
4218 * The problem is that NOMMU needs to be able to allocate large chunks
4219 * of contiguous memory as there's no hardware page translation to
4220 * assemble apparent contiguous memory from discontiguous pages.
4222 * Queueing large contiguous runs of pages for batching, however,
4223 * causes the pages to actually be freed in smaller chunks. As there
4224 * can be a significant delay between the individual batches being
4225 * recycled, this leads to the once large chunks of space being
4226 * fragmented and becoming unavailable for high-order allocations.
4233 * pcp->high and pcp->batch values are related and dependent on one another:
4234 * ->batch must never be higher then ->high.
4235 * The following function updates them in a safe manner without read side
4238 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4239 * those fields changing asynchronously (acording the the above rule).
4241 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4242 * outside of boot time (or some other assurance that no concurrent updaters
4245 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4246 unsigned long batch)
4248 /* start with a fail safe value for batch */
4252 /* Update high, then batch, in order */
4259 /* a companion to pageset_set_high() */
4260 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4262 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4265 static void pageset_init(struct per_cpu_pageset *p)
4267 struct per_cpu_pages *pcp;
4270 memset(p, 0, sizeof(*p));
4274 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4275 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4278 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4281 pageset_set_batch(p, batch);
4285 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4286 * to the value high for the pageset p.
4288 static void pageset_set_high(struct per_cpu_pageset *p,
4291 unsigned long batch = max(1UL, high / 4);
4292 if ((high / 4) > (PAGE_SHIFT * 8))
4293 batch = PAGE_SHIFT * 8;
4295 pageset_update(&p->pcp, high, batch);
4298 static void pageset_set_high_and_batch(struct zone *zone,
4299 struct per_cpu_pageset *pcp)
4301 if (percpu_pagelist_fraction)
4302 pageset_set_high(pcp,
4303 (zone->managed_pages /
4304 percpu_pagelist_fraction));
4306 pageset_set_batch(pcp, zone_batchsize(zone));
4309 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4311 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4314 pageset_set_high_and_batch(zone, pcp);
4317 static void __meminit setup_zone_pageset(struct zone *zone)
4320 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4321 for_each_possible_cpu(cpu)
4322 zone_pageset_init(zone, cpu);
4326 * Allocate per cpu pagesets and initialize them.
4327 * Before this call only boot pagesets were available.
4329 void __init setup_per_cpu_pageset(void)
4333 for_each_populated_zone(zone)
4334 setup_zone_pageset(zone);
4337 static noinline __init_refok
4338 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4344 * The per-page waitqueue mechanism uses hashed waitqueues
4347 zone->wait_table_hash_nr_entries =
4348 wait_table_hash_nr_entries(zone_size_pages);
4349 zone->wait_table_bits =
4350 wait_table_bits(zone->wait_table_hash_nr_entries);
4351 alloc_size = zone->wait_table_hash_nr_entries
4352 * sizeof(wait_queue_head_t);
4354 if (!slab_is_available()) {
4355 zone->wait_table = (wait_queue_head_t *)
4356 memblock_virt_alloc_node_nopanic(
4357 alloc_size, zone->zone_pgdat->node_id);
4360 * This case means that a zone whose size was 0 gets new memory
4361 * via memory hot-add.
4362 * But it may be the case that a new node was hot-added. In
4363 * this case vmalloc() will not be able to use this new node's
4364 * memory - this wait_table must be initialized to use this new
4365 * node itself as well.
4366 * To use this new node's memory, further consideration will be
4369 zone->wait_table = vmalloc(alloc_size);
4371 if (!zone->wait_table)
4374 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4375 init_waitqueue_head(zone->wait_table + i);
4380 static __meminit void zone_pcp_init(struct zone *zone)
4383 * per cpu subsystem is not up at this point. The following code
4384 * relies on the ability of the linker to provide the
4385 * offset of a (static) per cpu variable into the per cpu area.
4387 zone->pageset = &boot_pageset;
4389 if (populated_zone(zone))
4390 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4391 zone->name, zone->present_pages,
4392 zone_batchsize(zone));
4395 int __meminit init_currently_empty_zone(struct zone *zone,
4396 unsigned long zone_start_pfn,
4398 enum memmap_context context)
4400 struct pglist_data *pgdat = zone->zone_pgdat;
4402 ret = zone_wait_table_init(zone, size);
4405 pgdat->nr_zones = zone_idx(zone) + 1;
4407 zone->zone_start_pfn = zone_start_pfn;
4409 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4410 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4412 (unsigned long)zone_idx(zone),
4413 zone_start_pfn, (zone_start_pfn + size));
4415 zone_init_free_lists(zone);
4420 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4421 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4423 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4425 int __meminit __early_pfn_to_nid(unsigned long pfn)
4427 unsigned long start_pfn, end_pfn;
4430 * NOTE: The following SMP-unsafe globals are only used early in boot
4431 * when the kernel is running single-threaded.
4433 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4434 static int __meminitdata last_nid;
4436 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4439 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4441 last_start_pfn = start_pfn;
4442 last_end_pfn = end_pfn;
4448 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4450 int __meminit early_pfn_to_nid(unsigned long pfn)
4454 nid = __early_pfn_to_nid(pfn);
4457 /* just returns 0 */
4461 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4462 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4466 nid = __early_pfn_to_nid(pfn);
4467 if (nid >= 0 && nid != node)
4474 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4475 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4476 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4478 * If an architecture guarantees that all ranges registered contain no holes
4479 * and may be freed, this this function may be used instead of calling
4480 * memblock_free_early_nid() manually.
4482 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4484 unsigned long start_pfn, end_pfn;
4487 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4488 start_pfn = min(start_pfn, max_low_pfn);
4489 end_pfn = min(end_pfn, max_low_pfn);
4491 if (start_pfn < end_pfn)
4492 memblock_free_early_nid(PFN_PHYS(start_pfn),
4493 (end_pfn - start_pfn) << PAGE_SHIFT,
4499 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4500 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4502 * If an architecture guarantees that all ranges registered contain no holes and may
4503 * be freed, this function may be used instead of calling memory_present() manually.
4505 void __init sparse_memory_present_with_active_regions(int nid)
4507 unsigned long start_pfn, end_pfn;
4510 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4511 memory_present(this_nid, start_pfn, end_pfn);
4515 * get_pfn_range_for_nid - Return the start and end page frames for a node
4516 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4517 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4518 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4520 * It returns the start and end page frame of a node based on information
4521 * provided by memblock_set_node(). If called for a node
4522 * with no available memory, a warning is printed and the start and end
4525 void __meminit get_pfn_range_for_nid(unsigned int nid,
4526 unsigned long *start_pfn, unsigned long *end_pfn)
4528 unsigned long this_start_pfn, this_end_pfn;
4534 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4535 *start_pfn = min(*start_pfn, this_start_pfn);
4536 *end_pfn = max(*end_pfn, this_end_pfn);
4539 if (*start_pfn == -1UL)
4544 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4545 * assumption is made that zones within a node are ordered in monotonic
4546 * increasing memory addresses so that the "highest" populated zone is used
4548 static void __init find_usable_zone_for_movable(void)
4551 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4552 if (zone_index == ZONE_MOVABLE)
4555 if (arch_zone_highest_possible_pfn[zone_index] >
4556 arch_zone_lowest_possible_pfn[zone_index])
4560 VM_BUG_ON(zone_index == -1);
4561 movable_zone = zone_index;
4565 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4566 * because it is sized independent of architecture. Unlike the other zones,
4567 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4568 * in each node depending on the size of each node and how evenly kernelcore
4569 * is distributed. This helper function adjusts the zone ranges
4570 * provided by the architecture for a given node by using the end of the
4571 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4572 * zones within a node are in order of monotonic increases memory addresses
4574 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4575 unsigned long zone_type,
4576 unsigned long node_start_pfn,
4577 unsigned long node_end_pfn,
4578 unsigned long *zone_start_pfn,
4579 unsigned long *zone_end_pfn)
4581 /* Only adjust if ZONE_MOVABLE is on this node */
4582 if (zone_movable_pfn[nid]) {
4583 /* Size ZONE_MOVABLE */
4584 if (zone_type == ZONE_MOVABLE) {
4585 *zone_start_pfn = zone_movable_pfn[nid];
4586 *zone_end_pfn = min(node_end_pfn,
4587 arch_zone_highest_possible_pfn[movable_zone]);
4589 /* Adjust for ZONE_MOVABLE starting within this range */
4590 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4591 *zone_end_pfn > zone_movable_pfn[nid]) {
4592 *zone_end_pfn = zone_movable_pfn[nid];
4594 /* Check if this whole range is within ZONE_MOVABLE */
4595 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4596 *zone_start_pfn = *zone_end_pfn;
4601 * Return the number of pages a zone spans in a node, including holes
4602 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4604 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4605 unsigned long zone_type,
4606 unsigned long node_start_pfn,
4607 unsigned long node_end_pfn,
4608 unsigned long *ignored)
4610 unsigned long zone_start_pfn, zone_end_pfn;
4612 /* Get the start and end of the zone */
4613 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4614 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4615 adjust_zone_range_for_zone_movable(nid, zone_type,
4616 node_start_pfn, node_end_pfn,
4617 &zone_start_pfn, &zone_end_pfn);
4619 /* Check that this node has pages within the zone's required range */
4620 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4623 /* Move the zone boundaries inside the node if necessary */
4624 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4625 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4627 /* Return the spanned pages */
4628 return zone_end_pfn - zone_start_pfn;
4632 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4633 * then all holes in the requested range will be accounted for.
4635 unsigned long __meminit __absent_pages_in_range(int nid,
4636 unsigned long range_start_pfn,
4637 unsigned long range_end_pfn)
4639 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4640 unsigned long start_pfn, end_pfn;
4643 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4644 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4645 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4646 nr_absent -= end_pfn - start_pfn;
4652 * absent_pages_in_range - Return number of page frames in holes within a range
4653 * @start_pfn: The start PFN to start searching for holes
4654 * @end_pfn: The end PFN to stop searching for holes
4656 * It returns the number of pages frames in memory holes within a range.
4658 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4659 unsigned long end_pfn)
4661 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4664 /* Return the number of page frames in holes in a zone on a node */
4665 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4666 unsigned long zone_type,
4667 unsigned long node_start_pfn,
4668 unsigned long node_end_pfn,
4669 unsigned long *ignored)
4671 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4672 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4673 unsigned long zone_start_pfn, zone_end_pfn;
4675 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4676 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4678 adjust_zone_range_for_zone_movable(nid, zone_type,
4679 node_start_pfn, node_end_pfn,
4680 &zone_start_pfn, &zone_end_pfn);
4681 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4684 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4685 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4686 unsigned long zone_type,
4687 unsigned long node_start_pfn,
4688 unsigned long node_end_pfn,
4689 unsigned long *zones_size)
4691 return zones_size[zone_type];
4694 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4695 unsigned long zone_type,
4696 unsigned long node_start_pfn,
4697 unsigned long node_end_pfn,
4698 unsigned long *zholes_size)
4703 return zholes_size[zone_type];
4706 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4708 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4709 unsigned long node_start_pfn,
4710 unsigned long node_end_pfn,
4711 unsigned long *zones_size,
4712 unsigned long *zholes_size)
4714 unsigned long realtotalpages, totalpages = 0;
4717 for (i = 0; i < MAX_NR_ZONES; i++)
4718 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4722 pgdat->node_spanned_pages = totalpages;
4724 realtotalpages = totalpages;
4725 for (i = 0; i < MAX_NR_ZONES; i++)
4727 zone_absent_pages_in_node(pgdat->node_id, i,
4728 node_start_pfn, node_end_pfn,
4730 pgdat->node_present_pages = realtotalpages;
4731 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4735 #ifndef CONFIG_SPARSEMEM
4737 * Calculate the size of the zone->blockflags rounded to an unsigned long
4738 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4739 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4740 * round what is now in bits to nearest long in bits, then return it in
4743 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4745 unsigned long usemapsize;
4747 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4748 usemapsize = roundup(zonesize, pageblock_nr_pages);
4749 usemapsize = usemapsize >> pageblock_order;
4750 usemapsize *= NR_PAGEBLOCK_BITS;
4751 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4753 return usemapsize / 8;
4756 static void __init setup_usemap(struct pglist_data *pgdat,
4758 unsigned long zone_start_pfn,
4759 unsigned long zonesize)
4761 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4762 zone->pageblock_flags = NULL;
4764 zone->pageblock_flags =
4765 memblock_virt_alloc_node_nopanic(usemapsize,
4769 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4770 unsigned long zone_start_pfn, unsigned long zonesize) {}
4771 #endif /* CONFIG_SPARSEMEM */
4773 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4775 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4776 void __paginginit set_pageblock_order(void)
4780 /* Check that pageblock_nr_pages has not already been setup */
4781 if (pageblock_order)
4784 if (HPAGE_SHIFT > PAGE_SHIFT)
4785 order = HUGETLB_PAGE_ORDER;
4787 order = MAX_ORDER - 1;
4790 * Assume the largest contiguous order of interest is a huge page.
4791 * This value may be variable depending on boot parameters on IA64 and
4794 pageblock_order = order;
4796 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4799 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4800 * is unused as pageblock_order is set at compile-time. See
4801 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4804 void __paginginit set_pageblock_order(void)
4808 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4810 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4811 unsigned long present_pages)
4813 unsigned long pages = spanned_pages;
4816 * Provide a more accurate estimation if there are holes within
4817 * the zone and SPARSEMEM is in use. If there are holes within the
4818 * zone, each populated memory region may cost us one or two extra
4819 * memmap pages due to alignment because memmap pages for each
4820 * populated regions may not naturally algined on page boundary.
4821 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4823 if (spanned_pages > present_pages + (present_pages >> 4) &&
4824 IS_ENABLED(CONFIG_SPARSEMEM))
4825 pages = present_pages;
4827 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4831 * Set up the zone data structures:
4832 * - mark all pages reserved
4833 * - mark all memory queues empty
4834 * - clear the memory bitmaps
4836 * NOTE: pgdat should get zeroed by caller.
4838 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4839 unsigned long node_start_pfn, unsigned long node_end_pfn,
4840 unsigned long *zones_size, unsigned long *zholes_size)
4843 int nid = pgdat->node_id;
4844 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4847 pgdat_resize_init(pgdat);
4848 #ifdef CONFIG_NUMA_BALANCING
4849 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4850 pgdat->numabalancing_migrate_nr_pages = 0;
4851 pgdat->numabalancing_migrate_next_window = jiffies;
4853 init_waitqueue_head(&pgdat->kswapd_wait);
4854 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4856 for (j = 0; j < MAX_NR_ZONES; j++) {
4857 struct zone *zone = pgdat->node_zones + j;
4858 unsigned long size, realsize, freesize, memmap_pages;
4860 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4861 node_end_pfn, zones_size);
4862 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4868 * Adjust freesize so that it accounts for how much memory
4869 * is used by this zone for memmap. This affects the watermark
4870 * and per-cpu initialisations
4872 memmap_pages = calc_memmap_size(size, realsize);
4873 if (freesize >= memmap_pages) {
4874 freesize -= memmap_pages;
4877 " %s zone: %lu pages used for memmap\n",
4878 zone_names[j], memmap_pages);
4881 " %s zone: %lu pages exceeds freesize %lu\n",
4882 zone_names[j], memmap_pages, freesize);
4884 /* Account for reserved pages */
4885 if (j == 0 && freesize > dma_reserve) {
4886 freesize -= dma_reserve;
4887 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4888 zone_names[0], dma_reserve);
4891 if (!is_highmem_idx(j))
4892 nr_kernel_pages += freesize;
4893 /* Charge for highmem memmap if there are enough kernel pages */
4894 else if (nr_kernel_pages > memmap_pages * 2)
4895 nr_kernel_pages -= memmap_pages;
4896 nr_all_pages += freesize;
4898 zone->spanned_pages = size;
4899 zone->present_pages = realsize;
4901 * Set an approximate value for lowmem here, it will be adjusted
4902 * when the bootmem allocator frees pages into the buddy system.
4903 * And all highmem pages will be managed by the buddy system.
4905 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4908 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4910 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4912 zone->name = zone_names[j];
4913 spin_lock_init(&zone->lock);
4914 spin_lock_init(&zone->lru_lock);
4915 zone_seqlock_init(zone);
4916 zone->zone_pgdat = pgdat;
4917 zone_pcp_init(zone);
4919 /* For bootup, initialized properly in watermark setup */
4920 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4922 lruvec_init(&zone->lruvec);
4926 set_pageblock_order();
4927 setup_usemap(pgdat, zone, zone_start_pfn, size);
4928 ret = init_currently_empty_zone(zone, zone_start_pfn,
4929 size, MEMMAP_EARLY);
4931 memmap_init(size, nid, j, zone_start_pfn);
4932 zone_start_pfn += size;
4936 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4938 /* Skip empty nodes */
4939 if (!pgdat->node_spanned_pages)
4942 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4943 /* ia64 gets its own node_mem_map, before this, without bootmem */
4944 if (!pgdat->node_mem_map) {
4945 unsigned long size, start, end;
4949 * The zone's endpoints aren't required to be MAX_ORDER
4950 * aligned but the node_mem_map endpoints must be in order
4951 * for the buddy allocator to function correctly.
4953 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4954 end = pgdat_end_pfn(pgdat);
4955 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4956 size = (end - start) * sizeof(struct page);
4957 map = alloc_remap(pgdat->node_id, size);
4959 map = memblock_virt_alloc_node_nopanic(size,
4961 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4963 #ifndef CONFIG_NEED_MULTIPLE_NODES
4965 * With no DISCONTIG, the global mem_map is just set as node 0's
4967 if (pgdat == NODE_DATA(0)) {
4968 mem_map = NODE_DATA(0)->node_mem_map;
4969 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4970 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4971 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4972 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4975 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4978 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4979 unsigned long node_start_pfn, unsigned long *zholes_size)
4981 pg_data_t *pgdat = NODE_DATA(nid);
4982 unsigned long start_pfn = 0;
4983 unsigned long end_pfn = 0;
4985 /* pg_data_t should be reset to zero when it's allocated */
4986 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4988 pgdat->node_id = nid;
4989 pgdat->node_start_pfn = node_start_pfn;
4990 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4991 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4992 printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid,
4993 (u64) start_pfn << PAGE_SHIFT, (u64) (end_pfn << PAGE_SHIFT) - 1);
4995 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4996 zones_size, zholes_size);
4998 alloc_node_mem_map(pgdat);
4999 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5000 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5001 nid, (unsigned long)pgdat,
5002 (unsigned long)pgdat->node_mem_map);
5005 free_area_init_core(pgdat, start_pfn, end_pfn,
5006 zones_size, zholes_size);
5009 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5011 #if MAX_NUMNODES > 1
5013 * Figure out the number of possible node ids.
5015 void __init setup_nr_node_ids(void)
5018 unsigned int highest = 0;
5020 for_each_node_mask(node, node_possible_map)
5022 nr_node_ids = highest + 1;
5027 * node_map_pfn_alignment - determine the maximum internode alignment
5029 * This function should be called after node map is populated and sorted.
5030 * It calculates the maximum power of two alignment which can distinguish
5033 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5034 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5035 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5036 * shifted, 1GiB is enough and this function will indicate so.
5038 * This is used to test whether pfn -> nid mapping of the chosen memory
5039 * model has fine enough granularity to avoid incorrect mapping for the
5040 * populated node map.
5042 * Returns the determined alignment in pfn's. 0 if there is no alignment
5043 * requirement (single node).
5045 unsigned long __init node_map_pfn_alignment(void)
5047 unsigned long accl_mask = 0, last_end = 0;
5048 unsigned long start, end, mask;
5052 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5053 if (!start || last_nid < 0 || last_nid == nid) {
5060 * Start with a mask granular enough to pin-point to the
5061 * start pfn and tick off bits one-by-one until it becomes
5062 * too coarse to separate the current node from the last.
5064 mask = ~((1 << __ffs(start)) - 1);
5065 while (mask && last_end <= (start & (mask << 1)))
5068 /* accumulate all internode masks */
5072 /* convert mask to number of pages */
5073 return ~accl_mask + 1;
5076 /* Find the lowest pfn for a node */
5077 static unsigned long __init find_min_pfn_for_node(int nid)
5079 unsigned long min_pfn = ULONG_MAX;
5080 unsigned long start_pfn;
5083 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5084 min_pfn = min(min_pfn, start_pfn);
5086 if (min_pfn == ULONG_MAX) {
5088 "Could not find start_pfn for node %d\n", nid);
5096 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5098 * It returns the minimum PFN based on information provided via
5099 * memblock_set_node().
5101 unsigned long __init find_min_pfn_with_active_regions(void)
5103 return find_min_pfn_for_node(MAX_NUMNODES);
5107 * early_calculate_totalpages()
5108 * Sum pages in active regions for movable zone.
5109 * Populate N_MEMORY for calculating usable_nodes.
5111 static unsigned long __init early_calculate_totalpages(void)
5113 unsigned long totalpages = 0;
5114 unsigned long start_pfn, end_pfn;
5117 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5118 unsigned long pages = end_pfn - start_pfn;
5120 totalpages += pages;
5122 node_set_state(nid, N_MEMORY);
5128 * Find the PFN the Movable zone begins in each node. Kernel memory
5129 * is spread evenly between nodes as long as the nodes have enough
5130 * memory. When they don't, some nodes will have more kernelcore than
5133 static void __init find_zone_movable_pfns_for_nodes(void)
5136 unsigned long usable_startpfn;
5137 unsigned long kernelcore_node, kernelcore_remaining;
5138 /* save the state before borrow the nodemask */
5139 nodemask_t saved_node_state = node_states[N_MEMORY];
5140 unsigned long totalpages = early_calculate_totalpages();
5141 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5142 struct memblock_region *r;
5144 /* Need to find movable_zone earlier when movable_node is specified. */
5145 find_usable_zone_for_movable();
5148 * If movable_node is specified, ignore kernelcore and movablecore
5151 if (movable_node_is_enabled()) {
5152 for_each_memblock(memory, r) {
5153 if (!memblock_is_hotpluggable(r))
5158 usable_startpfn = PFN_DOWN(r->base);
5159 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5160 min(usable_startpfn, zone_movable_pfn[nid]) :
5168 * If movablecore=nn[KMG] was specified, calculate what size of
5169 * kernelcore that corresponds so that memory usable for
5170 * any allocation type is evenly spread. If both kernelcore
5171 * and movablecore are specified, then the value of kernelcore
5172 * will be used for required_kernelcore if it's greater than
5173 * what movablecore would have allowed.
5175 if (required_movablecore) {
5176 unsigned long corepages;
5179 * Round-up so that ZONE_MOVABLE is at least as large as what
5180 * was requested by the user
5182 required_movablecore =
5183 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5184 corepages = totalpages - required_movablecore;
5186 required_kernelcore = max(required_kernelcore, corepages);
5189 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5190 if (!required_kernelcore)
5193 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5194 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5197 /* Spread kernelcore memory as evenly as possible throughout nodes */
5198 kernelcore_node = required_kernelcore / usable_nodes;
5199 for_each_node_state(nid, N_MEMORY) {
5200 unsigned long start_pfn, end_pfn;
5203 * Recalculate kernelcore_node if the division per node
5204 * now exceeds what is necessary to satisfy the requested
5205 * amount of memory for the kernel
5207 if (required_kernelcore < kernelcore_node)
5208 kernelcore_node = required_kernelcore / usable_nodes;
5211 * As the map is walked, we track how much memory is usable
5212 * by the kernel using kernelcore_remaining. When it is
5213 * 0, the rest of the node is usable by ZONE_MOVABLE
5215 kernelcore_remaining = kernelcore_node;
5217 /* Go through each range of PFNs within this node */
5218 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5219 unsigned long size_pages;
5221 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5222 if (start_pfn >= end_pfn)
5225 /* Account for what is only usable for kernelcore */
5226 if (start_pfn < usable_startpfn) {
5227 unsigned long kernel_pages;
5228 kernel_pages = min(end_pfn, usable_startpfn)
5231 kernelcore_remaining -= min(kernel_pages,
5232 kernelcore_remaining);
5233 required_kernelcore -= min(kernel_pages,
5234 required_kernelcore);
5236 /* Continue if range is now fully accounted */
5237 if (end_pfn <= usable_startpfn) {
5240 * Push zone_movable_pfn to the end so
5241 * that if we have to rebalance
5242 * kernelcore across nodes, we will
5243 * not double account here
5245 zone_movable_pfn[nid] = end_pfn;
5248 start_pfn = usable_startpfn;
5252 * The usable PFN range for ZONE_MOVABLE is from
5253 * start_pfn->end_pfn. Calculate size_pages as the
5254 * number of pages used as kernelcore
5256 size_pages = end_pfn - start_pfn;
5257 if (size_pages > kernelcore_remaining)
5258 size_pages = kernelcore_remaining;
5259 zone_movable_pfn[nid] = start_pfn + size_pages;
5262 * Some kernelcore has been met, update counts and
5263 * break if the kernelcore for this node has been
5266 required_kernelcore -= min(required_kernelcore,
5268 kernelcore_remaining -= size_pages;
5269 if (!kernelcore_remaining)
5275 * If there is still required_kernelcore, we do another pass with one
5276 * less node in the count. This will push zone_movable_pfn[nid] further
5277 * along on the nodes that still have memory until kernelcore is
5281 if (usable_nodes && required_kernelcore > usable_nodes)
5285 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5286 for (nid = 0; nid < MAX_NUMNODES; nid++)
5287 zone_movable_pfn[nid] =
5288 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5291 /* restore the node_state */
5292 node_states[N_MEMORY] = saved_node_state;
5295 /* Any regular or high memory on that node ? */
5296 static void check_for_memory(pg_data_t *pgdat, int nid)
5298 enum zone_type zone_type;
5300 if (N_MEMORY == N_NORMAL_MEMORY)
5303 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5304 struct zone *zone = &pgdat->node_zones[zone_type];
5305 if (populated_zone(zone)) {
5306 node_set_state(nid, N_HIGH_MEMORY);
5307 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5308 zone_type <= ZONE_NORMAL)
5309 node_set_state(nid, N_NORMAL_MEMORY);
5316 * free_area_init_nodes - Initialise all pg_data_t and zone data
5317 * @max_zone_pfn: an array of max PFNs for each zone
5319 * This will call free_area_init_node() for each active node in the system.
5320 * Using the page ranges provided by memblock_set_node(), the size of each
5321 * zone in each node and their holes is calculated. If the maximum PFN
5322 * between two adjacent zones match, it is assumed that the zone is empty.
5323 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5324 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5325 * starts where the previous one ended. For example, ZONE_DMA32 starts
5326 * at arch_max_dma_pfn.
5328 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5330 unsigned long start_pfn, end_pfn;
5333 /* Record where the zone boundaries are */
5334 memset(arch_zone_lowest_possible_pfn, 0,
5335 sizeof(arch_zone_lowest_possible_pfn));
5336 memset(arch_zone_highest_possible_pfn, 0,
5337 sizeof(arch_zone_highest_possible_pfn));
5338 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5339 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5340 for (i = 1; i < MAX_NR_ZONES; i++) {
5341 if (i == ZONE_MOVABLE)
5343 arch_zone_lowest_possible_pfn[i] =
5344 arch_zone_highest_possible_pfn[i-1];
5345 arch_zone_highest_possible_pfn[i] =
5346 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5348 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5349 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5351 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5352 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5353 find_zone_movable_pfns_for_nodes();
5355 /* Print out the zone ranges */
5356 pr_info("Zone ranges:\n");
5357 for (i = 0; i < MAX_NR_ZONES; i++) {
5358 if (i == ZONE_MOVABLE)
5360 pr_info(" %-8s ", zone_names[i]);
5361 if (arch_zone_lowest_possible_pfn[i] ==
5362 arch_zone_highest_possible_pfn[i])
5365 pr_cont("[mem %0#10lx-%0#10lx]\n",
5366 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5367 (arch_zone_highest_possible_pfn[i]
5368 << PAGE_SHIFT) - 1);
5371 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5372 pr_info("Movable zone start for each node\n");
5373 for (i = 0; i < MAX_NUMNODES; i++) {
5374 if (zone_movable_pfn[i])
5375 pr_info(" Node %d: %#010lx\n", i,
5376 zone_movable_pfn[i] << PAGE_SHIFT);
5379 /* Print out the early node map */
5380 pr_info("Early memory node ranges\n");
5381 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5382 pr_info(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5383 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5385 /* Initialise every node */
5386 mminit_verify_pageflags_layout();
5387 setup_nr_node_ids();
5388 for_each_online_node(nid) {
5389 pg_data_t *pgdat = NODE_DATA(nid);
5390 free_area_init_node(nid, NULL,
5391 find_min_pfn_for_node(nid), NULL);
5393 /* Any memory on that node */
5394 if (pgdat->node_present_pages)
5395 node_set_state(nid, N_MEMORY);
5396 check_for_memory(pgdat, nid);
5400 static int __init cmdline_parse_core(char *p, unsigned long *core)
5402 unsigned long long coremem;
5406 coremem = memparse(p, &p);
5407 *core = coremem >> PAGE_SHIFT;
5409 /* Paranoid check that UL is enough for the coremem value */
5410 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5416 * kernelcore=size sets the amount of memory for use for allocations that
5417 * cannot be reclaimed or migrated.
5419 static int __init cmdline_parse_kernelcore(char *p)
5421 return cmdline_parse_core(p, &required_kernelcore);
5425 * movablecore=size sets the amount of memory for use for allocations that
5426 * can be reclaimed or migrated.
5428 static int __init cmdline_parse_movablecore(char *p)
5430 return cmdline_parse_core(p, &required_movablecore);
5433 early_param("kernelcore", cmdline_parse_kernelcore);
5434 early_param("movablecore", cmdline_parse_movablecore);
5436 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5438 void adjust_managed_page_count(struct page *page, long count)
5440 spin_lock(&managed_page_count_lock);
5441 page_zone(page)->managed_pages += count;
5442 totalram_pages += count;
5443 #ifdef CONFIG_HIGHMEM
5444 if (PageHighMem(page))
5445 totalhigh_pages += count;
5447 spin_unlock(&managed_page_count_lock);
5449 EXPORT_SYMBOL(adjust_managed_page_count);
5451 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5454 unsigned long pages = 0;
5456 start = (void *)PAGE_ALIGN((unsigned long)start);
5457 end = (void *)((unsigned long)end & PAGE_MASK);
5458 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5459 if ((unsigned int)poison <= 0xFF)
5460 memset(pos, poison, PAGE_SIZE);
5461 free_reserved_page(virt_to_page(pos));
5465 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5466 s, pages << (PAGE_SHIFT - 10), start, end);
5470 EXPORT_SYMBOL(free_reserved_area);
5472 #ifdef CONFIG_HIGHMEM
5473 void free_highmem_page(struct page *page)
5475 __free_reserved_page(page);
5477 page_zone(page)->managed_pages++;
5483 void __init mem_init_print_info(const char *str)
5485 unsigned long physpages, codesize, datasize, rosize, bss_size;
5486 unsigned long init_code_size, init_data_size;
5488 physpages = get_num_physpages();
5489 codesize = _etext - _stext;
5490 datasize = _edata - _sdata;
5491 rosize = __end_rodata - __start_rodata;
5492 bss_size = __bss_stop - __bss_start;
5493 init_data_size = __init_end - __init_begin;
5494 init_code_size = _einittext - _sinittext;
5497 * Detect special cases and adjust section sizes accordingly:
5498 * 1) .init.* may be embedded into .data sections
5499 * 2) .init.text.* may be out of [__init_begin, __init_end],
5500 * please refer to arch/tile/kernel/vmlinux.lds.S.
5501 * 3) .rodata.* may be embedded into .text or .data sections.
5503 #define adj_init_size(start, end, size, pos, adj) \
5505 if (start <= pos && pos < end && size > adj) \
5509 adj_init_size(__init_begin, __init_end, init_data_size,
5510 _sinittext, init_code_size);
5511 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5512 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5513 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5514 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5516 #undef adj_init_size
5518 pr_info("Memory: %luK/%luK available "
5519 "(%luK kernel code, %luK rwdata, %luK rodata, "
5520 "%luK init, %luK bss, %luK reserved"
5521 #ifdef CONFIG_HIGHMEM
5525 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5526 codesize >> 10, datasize >> 10, rosize >> 10,
5527 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5528 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5529 #ifdef CONFIG_HIGHMEM
5530 totalhigh_pages << (PAGE_SHIFT-10),
5532 str ? ", " : "", str ? str : "");
5536 * set_dma_reserve - set the specified number of pages reserved in the first zone
5537 * @new_dma_reserve: The number of pages to mark reserved
5539 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5540 * In the DMA zone, a significant percentage may be consumed by kernel image
5541 * and other unfreeable allocations which can skew the watermarks badly. This
5542 * function may optionally be used to account for unfreeable pages in the
5543 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5544 * smaller per-cpu batchsize.
5546 void __init set_dma_reserve(unsigned long new_dma_reserve)
5548 dma_reserve = new_dma_reserve;
5551 void __init free_area_init(unsigned long *zones_size)
5553 free_area_init_node(0, zones_size,
5554 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5557 static int page_alloc_cpu_notify(struct notifier_block *self,
5558 unsigned long action, void *hcpu)
5560 int cpu = (unsigned long)hcpu;
5562 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5563 lru_add_drain_cpu(cpu);
5567 * Spill the event counters of the dead processor
5568 * into the current processors event counters.
5569 * This artificially elevates the count of the current
5572 vm_events_fold_cpu(cpu);
5575 * Zero the differential counters of the dead processor
5576 * so that the vm statistics are consistent.
5578 * This is only okay since the processor is dead and cannot
5579 * race with what we are doing.
5581 cpu_vm_stats_fold(cpu);
5586 void __init page_alloc_init(void)
5588 hotcpu_notifier(page_alloc_cpu_notify, 0);
5592 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5593 * or min_free_kbytes changes.
5595 static void calculate_totalreserve_pages(void)
5597 struct pglist_data *pgdat;
5598 unsigned long reserve_pages = 0;
5599 enum zone_type i, j;
5601 for_each_online_pgdat(pgdat) {
5602 for (i = 0; i < MAX_NR_ZONES; i++) {
5603 struct zone *zone = pgdat->node_zones + i;
5606 /* Find valid and maximum lowmem_reserve in the zone */
5607 for (j = i; j < MAX_NR_ZONES; j++) {
5608 if (zone->lowmem_reserve[j] > max)
5609 max = zone->lowmem_reserve[j];
5612 /* we treat the high watermark as reserved pages. */
5613 max += high_wmark_pages(zone);
5615 if (max > zone->managed_pages)
5616 max = zone->managed_pages;
5617 reserve_pages += max;
5619 * Lowmem reserves are not available to
5620 * GFP_HIGHUSER page cache allocations and
5621 * kswapd tries to balance zones to their high
5622 * watermark. As a result, neither should be
5623 * regarded as dirtyable memory, to prevent a
5624 * situation where reclaim has to clean pages
5625 * in order to balance the zones.
5627 zone->dirty_balance_reserve = max;
5630 dirty_balance_reserve = reserve_pages;
5631 totalreserve_pages = reserve_pages;
5635 * setup_per_zone_lowmem_reserve - called whenever
5636 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5637 * has a correct pages reserved value, so an adequate number of
5638 * pages are left in the zone after a successful __alloc_pages().
5640 static void setup_per_zone_lowmem_reserve(void)
5642 struct pglist_data *pgdat;
5643 enum zone_type j, idx;
5645 for_each_online_pgdat(pgdat) {
5646 for (j = 0; j < MAX_NR_ZONES; j++) {
5647 struct zone *zone = pgdat->node_zones + j;
5648 unsigned long managed_pages = zone->managed_pages;
5650 zone->lowmem_reserve[j] = 0;
5654 struct zone *lower_zone;
5658 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5659 sysctl_lowmem_reserve_ratio[idx] = 1;
5661 lower_zone = pgdat->node_zones + idx;
5662 lower_zone->lowmem_reserve[j] = managed_pages /
5663 sysctl_lowmem_reserve_ratio[idx];
5664 managed_pages += lower_zone->managed_pages;
5669 /* update totalreserve_pages */
5670 calculate_totalreserve_pages();
5673 static void __setup_per_zone_wmarks(void)
5675 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5676 unsigned long lowmem_pages = 0;
5678 unsigned long flags;
5680 /* Calculate total number of !ZONE_HIGHMEM pages */
5681 for_each_zone(zone) {
5682 if (!is_highmem(zone))
5683 lowmem_pages += zone->managed_pages;
5686 for_each_zone(zone) {
5689 spin_lock_irqsave(&zone->lock, flags);
5690 tmp = (u64)pages_min * zone->managed_pages;
5691 do_div(tmp, lowmem_pages);
5692 if (is_highmem(zone)) {
5694 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5695 * need highmem pages, so cap pages_min to a small
5698 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5699 * deltas controls asynch page reclaim, and so should
5700 * not be capped for highmem.
5702 unsigned long min_pages;
5704 min_pages = zone->managed_pages / 1024;
5705 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5706 zone->watermark[WMARK_MIN] = min_pages;
5709 * If it's a lowmem zone, reserve a number of pages
5710 * proportionate to the zone's size.
5712 zone->watermark[WMARK_MIN] = tmp;
5715 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5716 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5718 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5719 high_wmark_pages(zone) - low_wmark_pages(zone) -
5720 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5722 setup_zone_migrate_reserve(zone);
5723 spin_unlock_irqrestore(&zone->lock, flags);
5726 /* update totalreserve_pages */
5727 calculate_totalreserve_pages();
5731 * setup_per_zone_wmarks - called when min_free_kbytes changes
5732 * or when memory is hot-{added|removed}
5734 * Ensures that the watermark[min,low,high] values for each zone are set
5735 * correctly with respect to min_free_kbytes.
5737 void setup_per_zone_wmarks(void)
5739 mutex_lock(&zonelists_mutex);
5740 __setup_per_zone_wmarks();
5741 mutex_unlock(&zonelists_mutex);
5745 * The inactive anon list should be small enough that the VM never has to
5746 * do too much work, but large enough that each inactive page has a chance
5747 * to be referenced again before it is swapped out.
5749 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5750 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5751 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5752 * the anonymous pages are kept on the inactive list.
5755 * memory ratio inactive anon
5756 * -------------------------------------
5765 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5767 unsigned int gb, ratio;
5769 /* Zone size in gigabytes */
5770 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5772 ratio = int_sqrt(10 * gb);
5776 zone->inactive_ratio = ratio;
5779 static void __meminit setup_per_zone_inactive_ratio(void)
5784 calculate_zone_inactive_ratio(zone);
5788 * Initialise min_free_kbytes.
5790 * For small machines we want it small (128k min). For large machines
5791 * we want it large (64MB max). But it is not linear, because network
5792 * bandwidth does not increase linearly with machine size. We use
5794 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5795 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5811 int __meminit init_per_zone_wmark_min(void)
5813 unsigned long lowmem_kbytes;
5814 int new_min_free_kbytes;
5816 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5817 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5819 if (new_min_free_kbytes > user_min_free_kbytes) {
5820 min_free_kbytes = new_min_free_kbytes;
5821 if (min_free_kbytes < 128)
5822 min_free_kbytes = 128;
5823 if (min_free_kbytes > 65536)
5824 min_free_kbytes = 65536;
5826 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5827 new_min_free_kbytes, user_min_free_kbytes);
5829 setup_per_zone_wmarks();
5830 refresh_zone_stat_thresholds();
5831 setup_per_zone_lowmem_reserve();
5832 setup_per_zone_inactive_ratio();
5835 module_init(init_per_zone_wmark_min)
5838 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5839 * that we can call two helper functions whenever min_free_kbytes
5842 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5843 void __user *buffer, size_t *length, loff_t *ppos)
5847 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5852 user_min_free_kbytes = min_free_kbytes;
5853 setup_per_zone_wmarks();
5859 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5860 void __user *buffer, size_t *length, loff_t *ppos)
5865 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5870 zone->min_unmapped_pages = (zone->managed_pages *
5871 sysctl_min_unmapped_ratio) / 100;
5875 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5876 void __user *buffer, size_t *length, loff_t *ppos)
5881 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5886 zone->min_slab_pages = (zone->managed_pages *
5887 sysctl_min_slab_ratio) / 100;
5893 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5894 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5895 * whenever sysctl_lowmem_reserve_ratio changes.
5897 * The reserve ratio obviously has absolutely no relation with the
5898 * minimum watermarks. The lowmem reserve ratio can only make sense
5899 * if in function of the boot time zone sizes.
5901 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5902 void __user *buffer, size_t *length, loff_t *ppos)
5904 proc_dointvec_minmax(table, write, buffer, length, ppos);
5905 setup_per_zone_lowmem_reserve();
5910 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5911 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5912 * pagelist can have before it gets flushed back to buddy allocator.
5914 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5915 void __user *buffer, size_t *length, loff_t *ppos)
5918 int old_percpu_pagelist_fraction;
5921 mutex_lock(&pcp_batch_high_lock);
5922 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5924 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5925 if (!write || ret < 0)
5928 /* Sanity checking to avoid pcp imbalance */
5929 if (percpu_pagelist_fraction &&
5930 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
5931 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
5937 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
5940 for_each_populated_zone(zone) {
5943 for_each_possible_cpu(cpu)
5944 pageset_set_high_and_batch(zone,
5945 per_cpu_ptr(zone->pageset, cpu));
5948 mutex_unlock(&pcp_batch_high_lock);
5952 int hashdist = HASHDIST_DEFAULT;
5955 static int __init set_hashdist(char *str)
5959 hashdist = simple_strtoul(str, &str, 0);
5962 __setup("hashdist=", set_hashdist);
5966 * allocate a large system hash table from bootmem
5967 * - it is assumed that the hash table must contain an exact power-of-2
5968 * quantity of entries
5969 * - limit is the number of hash buckets, not the total allocation size
5971 void *__init alloc_large_system_hash(const char *tablename,
5972 unsigned long bucketsize,
5973 unsigned long numentries,
5976 unsigned int *_hash_shift,
5977 unsigned int *_hash_mask,
5978 unsigned long low_limit,
5979 unsigned long high_limit)
5981 unsigned long long max = high_limit;
5982 unsigned long log2qty, size;
5985 /* allow the kernel cmdline to have a say */
5987 /* round applicable memory size up to nearest megabyte */
5988 numentries = nr_kernel_pages;
5990 /* It isn't necessary when PAGE_SIZE >= 1MB */
5991 if (PAGE_SHIFT < 20)
5992 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5994 /* limit to 1 bucket per 2^scale bytes of low memory */
5995 if (scale > PAGE_SHIFT)
5996 numentries >>= (scale - PAGE_SHIFT);
5998 numentries <<= (PAGE_SHIFT - scale);
6000 /* Make sure we've got at least a 0-order allocation.. */
6001 if (unlikely(flags & HASH_SMALL)) {
6002 /* Makes no sense without HASH_EARLY */
6003 WARN_ON(!(flags & HASH_EARLY));
6004 if (!(numentries >> *_hash_shift)) {
6005 numentries = 1UL << *_hash_shift;
6006 BUG_ON(!numentries);
6008 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6009 numentries = PAGE_SIZE / bucketsize;
6011 numentries = roundup_pow_of_two(numentries);
6013 /* limit allocation size to 1/16 total memory by default */
6015 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6016 do_div(max, bucketsize);
6018 max = min(max, 0x80000000ULL);
6020 if (numentries < low_limit)
6021 numentries = low_limit;
6022 if (numentries > max)
6025 log2qty = ilog2(numentries);
6028 size = bucketsize << log2qty;
6029 if (flags & HASH_EARLY)
6030 table = memblock_virt_alloc_nopanic(size, 0);
6032 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6035 * If bucketsize is not a power-of-two, we may free
6036 * some pages at the end of hash table which
6037 * alloc_pages_exact() automatically does
6039 if (get_order(size) < MAX_ORDER) {
6040 table = alloc_pages_exact(size, GFP_ATOMIC);
6041 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6044 } while (!table && size > PAGE_SIZE && --log2qty);
6047 panic("Failed to allocate %s hash table\n", tablename);
6049 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6052 ilog2(size) - PAGE_SHIFT,
6056 *_hash_shift = log2qty;
6058 *_hash_mask = (1 << log2qty) - 1;
6063 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6064 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6067 #ifdef CONFIG_SPARSEMEM
6068 return __pfn_to_section(pfn)->pageblock_flags;
6070 return zone->pageblock_flags;
6071 #endif /* CONFIG_SPARSEMEM */
6074 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6076 #ifdef CONFIG_SPARSEMEM
6077 pfn &= (PAGES_PER_SECTION-1);
6078 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6080 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6081 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6082 #endif /* CONFIG_SPARSEMEM */
6086 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6087 * @page: The page within the block of interest
6088 * @pfn: The target page frame number
6089 * @end_bitidx: The last bit of interest to retrieve
6090 * @mask: mask of bits that the caller is interested in
6092 * Return: pageblock_bits flags
6094 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6095 unsigned long end_bitidx,
6099 unsigned long *bitmap;
6100 unsigned long bitidx, word_bitidx;
6103 zone = page_zone(page);
6104 bitmap = get_pageblock_bitmap(zone, pfn);
6105 bitidx = pfn_to_bitidx(zone, pfn);
6106 word_bitidx = bitidx / BITS_PER_LONG;
6107 bitidx &= (BITS_PER_LONG-1);
6109 word = bitmap[word_bitidx];
6110 bitidx += end_bitidx;
6111 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6115 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6116 * @page: The page within the block of interest
6117 * @flags: The flags to set
6118 * @pfn: The target page frame number
6119 * @end_bitidx: The last bit of interest
6120 * @mask: mask of bits that the caller is interested in
6122 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6124 unsigned long end_bitidx,
6128 unsigned long *bitmap;
6129 unsigned long bitidx, word_bitidx;
6130 unsigned long old_word, word;
6132 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6134 zone = page_zone(page);
6135 bitmap = get_pageblock_bitmap(zone, pfn);
6136 bitidx = pfn_to_bitidx(zone, pfn);
6137 word_bitidx = bitidx / BITS_PER_LONG;
6138 bitidx &= (BITS_PER_LONG-1);
6140 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6142 bitidx += end_bitidx;
6143 mask <<= (BITS_PER_LONG - bitidx - 1);
6144 flags <<= (BITS_PER_LONG - bitidx - 1);
6146 word = ACCESS_ONCE(bitmap[word_bitidx]);
6148 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6149 if (word == old_word)
6156 * This function checks whether pageblock includes unmovable pages or not.
6157 * If @count is not zero, it is okay to include less @count unmovable pages
6159 * PageLRU check without isolation or lru_lock could race so that
6160 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6161 * expect this function should be exact.
6163 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6164 bool skip_hwpoisoned_pages)
6166 unsigned long pfn, iter, found;
6170 * For avoiding noise data, lru_add_drain_all() should be called
6171 * If ZONE_MOVABLE, the zone never contains unmovable pages
6173 if (zone_idx(zone) == ZONE_MOVABLE)
6175 mt = get_pageblock_migratetype(page);
6176 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6179 pfn = page_to_pfn(page);
6180 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6181 unsigned long check = pfn + iter;
6183 if (!pfn_valid_within(check))
6186 page = pfn_to_page(check);
6189 * Hugepages are not in LRU lists, but they're movable.
6190 * We need not scan over tail pages bacause we don't
6191 * handle each tail page individually in migration.
6193 if (PageHuge(page)) {
6194 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6199 * We can't use page_count without pin a page
6200 * because another CPU can free compound page.
6201 * This check already skips compound tails of THP
6202 * because their page->_count is zero at all time.
6204 if (!atomic_read(&page->_count)) {
6205 if (PageBuddy(page))
6206 iter += (1 << page_order(page)) - 1;
6211 * The HWPoisoned page may be not in buddy system, and
6212 * page_count() is not 0.
6214 if (skip_hwpoisoned_pages && PageHWPoison(page))
6220 * If there are RECLAIMABLE pages, we need to check it.
6221 * But now, memory offline itself doesn't call shrink_slab()
6222 * and it still to be fixed.
6225 * If the page is not RAM, page_count()should be 0.
6226 * we don't need more check. This is an _used_ not-movable page.
6228 * The problematic thing here is PG_reserved pages. PG_reserved
6229 * is set to both of a memory hole page and a _used_ kernel
6238 bool is_pageblock_removable_nolock(struct page *page)
6244 * We have to be careful here because we are iterating over memory
6245 * sections which are not zone aware so we might end up outside of
6246 * the zone but still within the section.
6247 * We have to take care about the node as well. If the node is offline
6248 * its NODE_DATA will be NULL - see page_zone.
6250 if (!node_online(page_to_nid(page)))
6253 zone = page_zone(page);
6254 pfn = page_to_pfn(page);
6255 if (!zone_spans_pfn(zone, pfn))
6258 return !has_unmovable_pages(zone, page, 0, true);
6263 static unsigned long pfn_max_align_down(unsigned long pfn)
6265 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6266 pageblock_nr_pages) - 1);
6269 static unsigned long pfn_max_align_up(unsigned long pfn)
6271 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6272 pageblock_nr_pages));
6275 /* [start, end) must belong to a single zone. */
6276 static int __alloc_contig_migrate_range(struct compact_control *cc,
6277 unsigned long start, unsigned long end)
6279 /* This function is based on compact_zone() from compaction.c. */
6280 unsigned long nr_reclaimed;
6281 unsigned long pfn = start;
6282 unsigned int tries = 0;
6287 while (pfn < end || !list_empty(&cc->migratepages)) {
6288 if (fatal_signal_pending(current)) {
6293 if (list_empty(&cc->migratepages)) {
6294 cc->nr_migratepages = 0;
6295 pfn = isolate_migratepages_range(cc, pfn, end);
6301 } else if (++tries == 5) {
6302 ret = ret < 0 ? ret : -EBUSY;
6306 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6308 cc->nr_migratepages -= nr_reclaimed;
6310 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6311 NULL, 0, cc->mode, MR_CMA);
6314 putback_movable_pages(&cc->migratepages);
6321 * alloc_contig_range() -- tries to allocate given range of pages
6322 * @start: start PFN to allocate
6323 * @end: one-past-the-last PFN to allocate
6324 * @migratetype: migratetype of the underlaying pageblocks (either
6325 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6326 * in range must have the same migratetype and it must
6327 * be either of the two.
6329 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6330 * aligned, however it's the caller's responsibility to guarantee that
6331 * we are the only thread that changes migrate type of pageblocks the
6334 * The PFN range must belong to a single zone.
6336 * Returns zero on success or negative error code. On success all
6337 * pages which PFN is in [start, end) are allocated for the caller and
6338 * need to be freed with free_contig_range().
6340 int alloc_contig_range(unsigned long start, unsigned long end,
6341 unsigned migratetype)
6343 unsigned long outer_start, outer_end;
6346 struct compact_control cc = {
6347 .nr_migratepages = 0,
6349 .zone = page_zone(pfn_to_page(start)),
6350 .mode = MIGRATE_SYNC,
6351 .ignore_skip_hint = true,
6353 INIT_LIST_HEAD(&cc.migratepages);
6356 * What we do here is we mark all pageblocks in range as
6357 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6358 * have different sizes, and due to the way page allocator
6359 * work, we align the range to biggest of the two pages so
6360 * that page allocator won't try to merge buddies from
6361 * different pageblocks and change MIGRATE_ISOLATE to some
6362 * other migration type.
6364 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6365 * migrate the pages from an unaligned range (ie. pages that
6366 * we are interested in). This will put all the pages in
6367 * range back to page allocator as MIGRATE_ISOLATE.
6369 * When this is done, we take the pages in range from page
6370 * allocator removing them from the buddy system. This way
6371 * page allocator will never consider using them.
6373 * This lets us mark the pageblocks back as
6374 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6375 * aligned range but not in the unaligned, original range are
6376 * put back to page allocator so that buddy can use them.
6379 ret = start_isolate_page_range(pfn_max_align_down(start),
6380 pfn_max_align_up(end), migratetype,
6385 ret = __alloc_contig_migrate_range(&cc, start, end);
6390 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6391 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6392 * more, all pages in [start, end) are free in page allocator.
6393 * What we are going to do is to allocate all pages from
6394 * [start, end) (that is remove them from page allocator).
6396 * The only problem is that pages at the beginning and at the
6397 * end of interesting range may be not aligned with pages that
6398 * page allocator holds, ie. they can be part of higher order
6399 * pages. Because of this, we reserve the bigger range and
6400 * once this is done free the pages we are not interested in.
6402 * We don't have to hold zone->lock here because the pages are
6403 * isolated thus they won't get removed from buddy.
6406 lru_add_drain_all();
6407 drain_all_pages(cc.zone);
6410 outer_start = start;
6411 while (!PageBuddy(pfn_to_page(outer_start))) {
6412 if (++order >= MAX_ORDER) {
6416 outer_start &= ~0UL << order;
6419 /* Make sure the range is really isolated. */
6420 if (test_pages_isolated(outer_start, end, false)) {
6421 pr_info("%s: [%lx, %lx) PFNs busy\n",
6422 __func__, outer_start, end);
6427 /* Grab isolated pages from freelists. */
6428 outer_end = isolate_freepages_range(&cc, outer_start, end);
6434 /* Free head and tail (if any) */
6435 if (start != outer_start)
6436 free_contig_range(outer_start, start - outer_start);
6437 if (end != outer_end)
6438 free_contig_range(end, outer_end - end);
6441 undo_isolate_page_range(pfn_max_align_down(start),
6442 pfn_max_align_up(end), migratetype);
6446 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6448 unsigned int count = 0;
6450 for (; nr_pages--; pfn++) {
6451 struct page *page = pfn_to_page(pfn);
6453 count += page_count(page) != 1;
6456 WARN(count != 0, "%d pages are still in use!\n", count);
6460 #ifdef CONFIG_MEMORY_HOTPLUG
6462 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6463 * page high values need to be recalulated.
6465 void __meminit zone_pcp_update(struct zone *zone)
6468 mutex_lock(&pcp_batch_high_lock);
6469 for_each_possible_cpu(cpu)
6470 pageset_set_high_and_batch(zone,
6471 per_cpu_ptr(zone->pageset, cpu));
6472 mutex_unlock(&pcp_batch_high_lock);
6476 void zone_pcp_reset(struct zone *zone)
6478 unsigned long flags;
6480 struct per_cpu_pageset *pset;
6482 /* avoid races with drain_pages() */
6483 local_irq_save(flags);
6484 if (zone->pageset != &boot_pageset) {
6485 for_each_online_cpu(cpu) {
6486 pset = per_cpu_ptr(zone->pageset, cpu);
6487 drain_zonestat(zone, pset);
6489 free_percpu(zone->pageset);
6490 zone->pageset = &boot_pageset;
6492 local_irq_restore(flags);
6495 #ifdef CONFIG_MEMORY_HOTREMOVE
6497 * All pages in the range must be isolated before calling this.
6500 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6504 unsigned int order, i;
6506 unsigned long flags;
6507 /* find the first valid pfn */
6508 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6513 zone = page_zone(pfn_to_page(pfn));
6514 spin_lock_irqsave(&zone->lock, flags);
6516 while (pfn < end_pfn) {
6517 if (!pfn_valid(pfn)) {
6521 page = pfn_to_page(pfn);
6523 * The HWPoisoned page may be not in buddy system, and
6524 * page_count() is not 0.
6526 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6528 SetPageReserved(page);
6532 BUG_ON(page_count(page));
6533 BUG_ON(!PageBuddy(page));
6534 order = page_order(page);
6535 #ifdef CONFIG_DEBUG_VM
6536 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6537 pfn, 1 << order, end_pfn);
6539 list_del(&page->lru);
6540 rmv_page_order(page);
6541 zone->free_area[order].nr_free--;
6542 for (i = 0; i < (1 << order); i++)
6543 SetPageReserved((page+i));
6544 pfn += (1 << order);
6546 spin_unlock_irqrestore(&zone->lock, flags);
6550 #ifdef CONFIG_MEMORY_FAILURE
6551 bool is_free_buddy_page(struct page *page)
6553 struct zone *zone = page_zone(page);
6554 unsigned long pfn = page_to_pfn(page);
6555 unsigned long flags;
6558 spin_lock_irqsave(&zone->lock, flags);
6559 for (order = 0; order < MAX_ORDER; order++) {
6560 struct page *page_head = page - (pfn & ((1 << order) - 1));
6562 if (PageBuddy(page_head) && page_order(page_head) >= order)
6565 spin_unlock_irqrestore(&zone->lock, flags);
6567 return order < MAX_ORDER;