2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74 DEFINE_PER_CPU(int, numa_node);
75 EXPORT_PER_CPU_SYMBOL(numa_node);
78 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
85 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
86 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
90 * Array of node states.
92 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
93 [N_POSSIBLE] = NODE_MASK_ALL,
94 [N_ONLINE] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY] = { { [0] = 1UL } },
103 [N_CPU] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock);
111 unsigned long totalram_pages __read_mostly;
112 unsigned long totalreserve_pages __read_mostly;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly;
121 int percpu_pagelist_fraction;
122 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex));
139 if (saved_gfp_mask) {
140 gfp_allowed_mask = saved_gfp_mask;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex));
148 WARN_ON(saved_gfp_mask);
149 saved_gfp_mask = gfp_allowed_mask;
150 gfp_allowed_mask &= ~GFP_IOFS;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly;
165 static void __free_pages_ok(struct page *page, unsigned int order);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages);
193 static char * const zone_names[MAX_NR_ZONES] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes = 1024;
208 int user_min_free_kbytes = -1;
210 static unsigned long __meminitdata nr_kernel_pages;
211 static unsigned long __meminitdata nr_all_pages;
212 static unsigned long __meminitdata dma_reserve;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
217 static unsigned long __initdata required_kernelcore;
218 static unsigned long __initdata required_movablecore;
219 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly = MAX_NUMNODES;
228 int nr_online_nodes __read_mostly = 1;
229 EXPORT_SYMBOL(nr_node_ids);
230 EXPORT_SYMBOL(nr_online_nodes);
233 int page_group_by_mobility_disabled __read_mostly;
235 void set_pageblock_migratetype(struct page *page, int migratetype)
237 if (unlikely(page_group_by_mobility_disabled &&
238 migratetype < MIGRATE_PCPTYPES))
239 migratetype = MIGRATE_UNMOVABLE;
241 set_pageblock_flags_group(page, (unsigned long)migratetype,
242 PB_migrate, PB_migrate_end);
245 bool oom_killer_disabled __read_mostly;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
252 unsigned long pfn = page_to_pfn(page);
253 unsigned long sp, start_pfn;
256 seq = zone_span_seqbegin(zone);
257 start_pfn = zone->zone_start_pfn;
258 sp = zone->spanned_pages;
259 if (!zone_spans_pfn(zone, pfn))
261 } while (zone_span_seqretry(zone, seq));
264 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
265 pfn, zone_to_nid(zone), zone->name,
266 start_pfn, start_pfn + sp);
271 static int page_is_consistent(struct zone *zone, struct page *page)
273 if (!pfn_valid_within(page_to_pfn(page)))
275 if (zone != page_zone(page))
281 * Temporary debugging check for pages not lying within a given zone.
283 static int bad_range(struct zone *zone, struct page *page)
285 if (page_outside_zone_boundaries(zone, page))
287 if (!page_is_consistent(zone, page))
293 static inline int bad_range(struct zone *zone, struct page *page)
299 static void bad_page(struct page *page, const char *reason,
300 unsigned long bad_flags)
302 static unsigned long resume;
303 static unsigned long nr_shown;
304 static unsigned long nr_unshown;
306 /* Don't complain about poisoned pages */
307 if (PageHWPoison(page)) {
308 page_mapcount_reset(page); /* remove PageBuddy */
313 * Allow a burst of 60 reports, then keep quiet for that minute;
314 * or allow a steady drip of one report per second.
316 if (nr_shown == 60) {
317 if (time_before(jiffies, resume)) {
323 "BUG: Bad page state: %lu messages suppressed\n",
330 resume = jiffies + 60 * HZ;
332 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
333 current->comm, page_to_pfn(page));
334 dump_page_badflags(page, reason, bad_flags);
339 /* Leave bad fields for debug, except PageBuddy could make trouble */
340 page_mapcount_reset(page); /* remove PageBuddy */
341 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
345 * Higher-order pages are called "compound pages". They are structured thusly:
347 * The first PAGE_SIZE page is called the "head page".
349 * The remaining PAGE_SIZE pages are called "tail pages".
351 * All pages have PG_compound set. All tail pages have their ->first_page
352 * pointing at the head page.
354 * The first tail page's ->lru.next holds the address of the compound page's
355 * put_page() function. Its ->lru.prev holds the order of allocation.
356 * This usage means that zero-order pages may not be compound.
359 static void free_compound_page(struct page *page)
361 __free_pages_ok(page, compound_order(page));
364 void prep_compound_page(struct page *page, unsigned long order)
367 int nr_pages = 1 << order;
369 set_compound_page_dtor(page, free_compound_page);
370 set_compound_order(page, order);
372 for (i = 1; i < nr_pages; i++) {
373 struct page *p = page + i;
374 set_page_count(p, 0);
375 p->first_page = page;
376 /* Make sure p->first_page is always valid for PageTail() */
382 /* update __split_huge_page_refcount if you change this function */
383 static int destroy_compound_page(struct page *page, unsigned long order)
386 int nr_pages = 1 << order;
389 if (unlikely(compound_order(page) != order)) {
390 bad_page(page, "wrong compound order", 0);
394 __ClearPageHead(page);
396 for (i = 1; i < nr_pages; i++) {
397 struct page *p = page + i;
399 if (unlikely(!PageTail(p))) {
400 bad_page(page, "PageTail not set", 0);
402 } else if (unlikely(p->first_page != page)) {
403 bad_page(page, "first_page not consistent", 0);
412 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
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, 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 * Locate the struct page for both the matching buddy in our
470 * pair (buddy1) and the combined O(n+1) page they form (page).
472 * 1) Any buddy B1 will have an order O twin B2 which satisfies
473 * the following equation:
475 * For example, if the starting buddy (buddy2) is #8 its order
477 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
479 * 2) Any buddy B will have an order O+1 parent P which
480 * satisfies the following equation:
483 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
485 static inline unsigned long
486 __find_buddy_index(unsigned long page_idx, unsigned int order)
488 return page_idx ^ (1 << order);
492 * This function checks whether a page is free && is the buddy
493 * we can do coalesce a page and its buddy if
494 * (a) the buddy is not in a hole &&
495 * (b) the buddy is in the buddy system &&
496 * (c) a page and its buddy have the same order &&
497 * (d) a page and its buddy are in the same zone.
499 * For recording whether a page is in the buddy system, we set ->_mapcount
500 * PAGE_BUDDY_MAPCOUNT_VALUE.
501 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
502 * serialized by zone->lock.
504 * For recording page's order, we use page_private(page).
506 static inline int page_is_buddy(struct page *page, struct page *buddy,
509 if (!pfn_valid_within(page_to_pfn(buddy)))
512 if (page_zone_id(page) != page_zone_id(buddy))
515 if (page_is_guard(buddy) && page_order(buddy) == order) {
516 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
520 if (PageBuddy(buddy) && page_order(buddy) == order) {
521 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
528 * Freeing function for a buddy system allocator.
530 * The concept of a buddy system is to maintain direct-mapped table
531 * (containing bit values) for memory blocks of various "orders".
532 * The bottom level table contains the map for the smallest allocatable
533 * units of memory (here, pages), and each level above it describes
534 * pairs of units from the levels below, hence, "buddies".
535 * At a high level, all that happens here is marking the table entry
536 * at the bottom level available, and propagating the changes upward
537 * as necessary, plus some accounting needed to play nicely with other
538 * parts of the VM system.
539 * At each level, we keep a list of pages, which are heads of continuous
540 * free pages of length of (1 << order) and marked with _mapcount
541 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
543 * So when we are allocating or freeing one, we can derive the state of the
544 * other. That is, if we allocate a small block, and both were
545 * free, the remainder of the region must be split into blocks.
546 * If a block is freed, and its buddy is also free, then this
547 * triggers coalescing into a block of larger size.
552 static inline void __free_one_page(struct page *page,
553 struct zone *zone, unsigned int order,
556 unsigned long page_idx;
557 unsigned long combined_idx;
558 unsigned long uninitialized_var(buddy_idx);
561 VM_BUG_ON(!zone_is_initialized(zone));
563 if (unlikely(PageCompound(page)))
564 if (unlikely(destroy_compound_page(page, order)))
567 VM_BUG_ON(migratetype == -1);
569 page_idx = page_to_pfn(page) & ((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(page, 0);
586 __mod_zone_freepage_state(zone, 1 << order,
589 list_del(&buddy->lru);
590 zone->free_area[order].nr_free--;
591 rmv_page_order(buddy);
593 combined_idx = buddy_idx & page_idx;
594 page = page + (combined_idx - page_idx);
595 page_idx = combined_idx;
598 set_page_order(page, order);
601 * If this is not the largest possible page, check if the buddy
602 * of the next-highest order is free. If it is, it's possible
603 * that pages are being freed that will coalesce soon. In case,
604 * that is happening, add the free page to the tail of the list
605 * so it's less likely to be used soon and more likely to be merged
606 * as a higher order page
608 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
609 struct page *higher_page, *higher_buddy;
610 combined_idx = buddy_idx & page_idx;
611 higher_page = page + (combined_idx - page_idx);
612 buddy_idx = __find_buddy_index(combined_idx, order + 1);
613 higher_buddy = higher_page + (buddy_idx - combined_idx);
614 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
615 list_add_tail(&page->lru,
616 &zone->free_area[order].free_list[migratetype]);
621 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
623 zone->free_area[order].nr_free++;
626 static inline int free_pages_check(struct page *page)
628 const char *bad_reason = NULL;
629 unsigned long bad_flags = 0;
631 if (unlikely(page_mapcount(page)))
632 bad_reason = "nonzero mapcount";
633 if (unlikely(page->mapping != NULL))
634 bad_reason = "non-NULL mapping";
635 if (unlikely(atomic_read(&page->_count) != 0))
636 bad_reason = "nonzero _count";
637 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
638 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
639 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
641 if (unlikely(mem_cgroup_bad_page_check(page)))
642 bad_reason = "cgroup check failed";
643 if (unlikely(bad_reason)) {
644 bad_page(page, bad_reason, bad_flags);
647 page_cpupid_reset_last(page);
648 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
649 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
654 * Frees a number of pages from the PCP lists
655 * Assumes all pages on list are in same zone, and of same order.
656 * count is the number of pages to free.
658 * If the zone was previously in an "all pages pinned" state then look to
659 * see if this freeing clears that state.
661 * And clear the zone's pages_scanned counter, to hold off the "all pages are
662 * pinned" detection logic.
664 static void free_pcppages_bulk(struct zone *zone, int count,
665 struct per_cpu_pages *pcp)
671 spin_lock(&zone->lock);
672 zone->pages_scanned = 0;
676 struct list_head *list;
679 * Remove pages from lists in a round-robin fashion. A
680 * batch_free count is maintained that is incremented when an
681 * empty list is encountered. This is so more pages are freed
682 * off fuller lists instead of spinning excessively around empty
687 if (++migratetype == MIGRATE_PCPTYPES)
689 list = &pcp->lists[migratetype];
690 } while (list_empty(list));
692 /* This is the only non-empty list. Free them all. */
693 if (batch_free == MIGRATE_PCPTYPES)
694 batch_free = to_free;
697 int mt; /* migratetype of the to-be-freed page */
699 page = list_entry(list->prev, struct page, lru);
700 /* must delete as __free_one_page list manipulates */
701 list_del(&page->lru);
702 mt = get_freepage_migratetype(page);
703 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
704 __free_one_page(page, zone, 0, mt);
705 trace_mm_page_pcpu_drain(page, 0, mt);
706 if (likely(!is_migrate_isolate_page(page))) {
707 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
708 if (is_migrate_cma(mt))
709 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
711 } while (--to_free && --batch_free && !list_empty(list));
713 spin_unlock(&zone->lock);
716 static void free_one_page(struct zone *zone, struct page *page, int order,
719 spin_lock(&zone->lock);
720 zone->pages_scanned = 0;
722 __free_one_page(page, zone, order, migratetype);
723 if (unlikely(!is_migrate_isolate(migratetype)))
724 __mod_zone_freepage_state(zone, 1 << order, migratetype);
725 spin_unlock(&zone->lock);
728 static bool free_pages_prepare(struct page *page, unsigned int order)
733 trace_mm_page_free(page, order);
734 kmemcheck_free_shadow(page, order);
737 page->mapping = NULL;
738 for (i = 0; i < (1 << order); i++)
739 bad += free_pages_check(page + i);
743 if (!PageHighMem(page)) {
744 debug_check_no_locks_freed(page_address(page),
746 debug_check_no_obj_freed(page_address(page),
749 arch_free_page(page, order);
750 kernel_map_pages(page, 1 << order, 0);
755 static void __free_pages_ok(struct page *page, unsigned int order)
760 if (!free_pages_prepare(page, order))
763 local_irq_save(flags);
764 __count_vm_events(PGFREE, 1 << order);
765 migratetype = get_pageblock_migratetype(page);
766 set_freepage_migratetype(page, migratetype);
767 free_one_page(page_zone(page), page, order, migratetype);
768 local_irq_restore(flags);
771 void __init __free_pages_bootmem(struct page *page, unsigned int order)
773 unsigned int nr_pages = 1 << order;
774 struct page *p = page;
778 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
780 __ClearPageReserved(p);
781 set_page_count(p, 0);
783 __ClearPageReserved(p);
784 set_page_count(p, 0);
786 page_zone(page)->managed_pages += nr_pages;
787 set_page_refcounted(page);
788 __free_pages(page, order);
792 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
793 void __init init_cma_reserved_pageblock(struct page *page)
795 unsigned i = pageblock_nr_pages;
796 struct page *p = page;
799 __ClearPageReserved(p);
800 set_page_count(p, 0);
803 set_page_refcounted(page);
804 set_pageblock_migratetype(page, MIGRATE_CMA);
805 __free_pages(page, pageblock_order);
806 adjust_managed_page_count(page, pageblock_nr_pages);
811 * The order of subdivision here is critical for the IO subsystem.
812 * Please do not alter this order without good reasons and regression
813 * testing. Specifically, as large blocks of memory are subdivided,
814 * the order in which smaller blocks are delivered depends on the order
815 * they're subdivided in this function. This is the primary factor
816 * influencing the order in which pages are delivered to the IO
817 * subsystem according to empirical testing, and this is also justified
818 * by considering the behavior of a buddy system containing a single
819 * large block of memory acted on by a series of small allocations.
820 * This behavior is a critical factor in sglist merging's success.
824 static inline void expand(struct zone *zone, struct page *page,
825 int low, int high, struct free_area *area,
828 unsigned long size = 1 << high;
834 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
836 #ifdef CONFIG_DEBUG_PAGEALLOC
837 if (high < debug_guardpage_minorder()) {
839 * Mark as guard pages (or page), that will allow to
840 * merge back to allocator when buddy will be freed.
841 * Corresponding page table entries will not be touched,
842 * pages will stay not present in virtual address space
844 INIT_LIST_HEAD(&page[size].lru);
845 set_page_guard_flag(&page[size]);
846 set_page_private(&page[size], high);
847 /* Guard pages are not available for any usage */
848 __mod_zone_freepage_state(zone, -(1 << high),
853 list_add(&page[size].lru, &area->free_list[migratetype]);
855 set_page_order(&page[size], high);
860 * This page is about to be returned from the page allocator
862 static inline int check_new_page(struct page *page)
864 const char *bad_reason = NULL;
865 unsigned long bad_flags = 0;
867 if (unlikely(page_mapcount(page)))
868 bad_reason = "nonzero mapcount";
869 if (unlikely(page->mapping != NULL))
870 bad_reason = "non-NULL mapping";
871 if (unlikely(atomic_read(&page->_count) != 0))
872 bad_reason = "nonzero _count";
873 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
874 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
875 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
877 if (unlikely(mem_cgroup_bad_page_check(page)))
878 bad_reason = "cgroup check failed";
879 if (unlikely(bad_reason)) {
880 bad_page(page, bad_reason, bad_flags);
886 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
890 for (i = 0; i < (1 << order); i++) {
891 struct page *p = page + i;
892 if (unlikely(check_new_page(p)))
896 set_page_private(page, 0);
897 set_page_refcounted(page);
899 arch_alloc_page(page, order);
900 kernel_map_pages(page, 1 << order, 1);
902 if (gfp_flags & __GFP_ZERO)
903 prep_zero_page(page, order, gfp_flags);
905 if (order && (gfp_flags & __GFP_COMP))
906 prep_compound_page(page, order);
912 * Go through the free lists for the given migratetype and remove
913 * the smallest available page from the freelists
916 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
919 unsigned int current_order;
920 struct free_area *area;
923 /* Find a page of the appropriate size in the preferred list */
924 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
925 area = &(zone->free_area[current_order]);
926 if (list_empty(&area->free_list[migratetype]))
929 page = list_entry(area->free_list[migratetype].next,
931 list_del(&page->lru);
932 rmv_page_order(page);
934 expand(zone, page, order, current_order, area, migratetype);
935 set_freepage_migratetype(page, migratetype);
944 * This array describes the order lists are fallen back to when
945 * the free lists for the desirable migrate type are depleted
947 static int fallbacks[MIGRATE_TYPES][4] = {
948 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
949 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
951 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
952 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
954 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
956 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
957 #ifdef CONFIG_MEMORY_ISOLATION
958 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
963 * Move the free pages in a range to the free lists of the requested type.
964 * Note that start_page and end_pages are not aligned on a pageblock
965 * boundary. If alignment is required, use move_freepages_block()
967 int move_freepages(struct zone *zone,
968 struct page *start_page, struct page *end_page,
975 #ifndef CONFIG_HOLES_IN_ZONE
977 * page_zone is not safe to call in this context when
978 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
979 * anyway as we check zone boundaries in move_freepages_block().
980 * Remove at a later date when no bug reports exist related to
981 * grouping pages by mobility
983 BUG_ON(page_zone(start_page) != page_zone(end_page));
986 for (page = start_page; page <= end_page;) {
987 /* Make sure we are not inadvertently changing nodes */
988 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
990 if (!pfn_valid_within(page_to_pfn(page))) {
995 if (!PageBuddy(page)) {
1000 order = page_order(page);
1001 list_move(&page->lru,
1002 &zone->free_area[order].free_list[migratetype]);
1003 set_freepage_migratetype(page, migratetype);
1005 pages_moved += 1 << order;
1011 int move_freepages_block(struct zone *zone, struct page *page,
1014 unsigned long start_pfn, end_pfn;
1015 struct page *start_page, *end_page;
1017 start_pfn = page_to_pfn(page);
1018 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1019 start_page = pfn_to_page(start_pfn);
1020 end_page = start_page + pageblock_nr_pages - 1;
1021 end_pfn = start_pfn + pageblock_nr_pages - 1;
1023 /* Do not cross zone boundaries */
1024 if (!zone_spans_pfn(zone, start_pfn))
1026 if (!zone_spans_pfn(zone, end_pfn))
1029 return move_freepages(zone, start_page, end_page, migratetype);
1032 static void change_pageblock_range(struct page *pageblock_page,
1033 int start_order, int migratetype)
1035 int nr_pageblocks = 1 << (start_order - pageblock_order);
1037 while (nr_pageblocks--) {
1038 set_pageblock_migratetype(pageblock_page, migratetype);
1039 pageblock_page += pageblock_nr_pages;
1044 * If breaking a large block of pages, move all free pages to the preferred
1045 * allocation list. If falling back for a reclaimable kernel allocation, be
1046 * more aggressive about taking ownership of free pages.
1048 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1049 * nor move CMA pages to different free lists. We don't want unmovable pages
1050 * to be allocated from MIGRATE_CMA areas.
1052 * Returns the new migratetype of the pageblock (or the same old migratetype
1053 * if it was unchanged).
1055 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1056 int start_type, int fallback_type)
1058 int current_order = page_order(page);
1061 * When borrowing from MIGRATE_CMA, we need to release the excess
1062 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1063 * is set to CMA so it is returned to the correct freelist in case
1064 * the page ends up being not actually allocated from the pcp lists.
1066 if (is_migrate_cma(fallback_type))
1067 return fallback_type;
1069 /* Take ownership for orders >= pageblock_order */
1070 if (current_order >= pageblock_order) {
1071 change_pageblock_range(page, current_order, start_type);
1075 if (current_order >= pageblock_order / 2 ||
1076 start_type == MIGRATE_RECLAIMABLE ||
1077 page_group_by_mobility_disabled) {
1080 pages = move_freepages_block(zone, page, start_type);
1082 /* Claim the whole block if over half of it is free */
1083 if (pages >= (1 << (pageblock_order-1)) ||
1084 page_group_by_mobility_disabled) {
1086 set_pageblock_migratetype(page, start_type);
1092 return fallback_type;
1095 /* Remove an element from the buddy allocator from the fallback list */
1096 static inline struct page *
1097 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1099 struct free_area *area;
1102 int migratetype, new_type, i;
1104 /* Find the largest possible block of pages in the other list */
1105 for (current_order = MAX_ORDER-1; current_order >= order;
1108 migratetype = fallbacks[start_migratetype][i];
1110 /* MIGRATE_RESERVE handled later if necessary */
1111 if (migratetype == MIGRATE_RESERVE)
1114 area = &(zone->free_area[current_order]);
1115 if (list_empty(&area->free_list[migratetype]))
1118 page = list_entry(area->free_list[migratetype].next,
1122 new_type = try_to_steal_freepages(zone, page,
1126 /* Remove the page from the freelists */
1127 list_del(&page->lru);
1128 rmv_page_order(page);
1130 expand(zone, page, order, current_order, area,
1132 /* The freepage_migratetype may differ from pageblock's
1133 * migratetype depending on the decisions in
1134 * try_to_steal_freepages. This is OK as long as it does
1135 * not differ for MIGRATE_CMA type.
1137 set_freepage_migratetype(page, new_type);
1139 trace_mm_page_alloc_extfrag(page, order, current_order,
1140 start_migratetype, migratetype, new_type);
1150 * Do the hard work of removing an element from the buddy allocator.
1151 * Call me with the zone->lock already held.
1153 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1159 page = __rmqueue_smallest(zone, order, migratetype);
1161 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1162 page = __rmqueue_fallback(zone, order, migratetype);
1165 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1166 * is used because __rmqueue_smallest is an inline function
1167 * and we want just one call site
1170 migratetype = MIGRATE_RESERVE;
1175 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1180 * Obtain a specified number of elements from the buddy allocator, all under
1181 * a single hold of the lock, for efficiency. Add them to the supplied list.
1182 * Returns the number of new pages which were placed at *list.
1184 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1185 unsigned long count, struct list_head *list,
1186 int migratetype, int cold)
1190 spin_lock(&zone->lock);
1191 for (i = 0; i < count; ++i) {
1192 struct page *page = __rmqueue(zone, order, migratetype);
1193 if (unlikely(page == NULL))
1197 * Split buddy pages returned by expand() are received here
1198 * in physical page order. The page is added to the callers and
1199 * list and the list head then moves forward. From the callers
1200 * perspective, the linked list is ordered by page number in
1201 * some conditions. This is useful for IO devices that can
1202 * merge IO requests if the physical pages are ordered
1205 if (likely(cold == 0))
1206 list_add(&page->lru, list);
1208 list_add_tail(&page->lru, list);
1210 if (is_migrate_cma(get_freepage_migratetype(page)))
1211 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1214 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1215 spin_unlock(&zone->lock);
1221 * Called from the vmstat counter updater to drain pagesets of this
1222 * currently executing processor on remote nodes after they have
1225 * Note that this function must be called with the thread pinned to
1226 * a single processor.
1228 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1230 unsigned long flags;
1232 unsigned long batch;
1234 local_irq_save(flags);
1235 batch = ACCESS_ONCE(pcp->batch);
1236 if (pcp->count >= batch)
1239 to_drain = pcp->count;
1241 free_pcppages_bulk(zone, to_drain, pcp);
1242 pcp->count -= to_drain;
1244 local_irq_restore(flags);
1249 * Drain pages of the indicated processor.
1251 * The processor must either be the current processor and the
1252 * thread pinned to the current processor or a processor that
1255 static void drain_pages(unsigned int cpu)
1257 unsigned long flags;
1260 for_each_populated_zone(zone) {
1261 struct per_cpu_pageset *pset;
1262 struct per_cpu_pages *pcp;
1264 local_irq_save(flags);
1265 pset = per_cpu_ptr(zone->pageset, cpu);
1269 free_pcppages_bulk(zone, pcp->count, pcp);
1272 local_irq_restore(flags);
1277 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1279 void drain_local_pages(void *arg)
1281 drain_pages(smp_processor_id());
1285 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1287 * Note that this code is protected against sending an IPI to an offline
1288 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1289 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1290 * nothing keeps CPUs from showing up after we populated the cpumask and
1291 * before the call to on_each_cpu_mask().
1293 void drain_all_pages(void)
1296 struct per_cpu_pageset *pcp;
1300 * Allocate in the BSS so we wont require allocation in
1301 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1303 static cpumask_t cpus_with_pcps;
1306 * We don't care about racing with CPU hotplug event
1307 * as offline notification will cause the notified
1308 * cpu to drain that CPU pcps and on_each_cpu_mask
1309 * disables preemption as part of its processing
1311 for_each_online_cpu(cpu) {
1312 bool has_pcps = false;
1313 for_each_populated_zone(zone) {
1314 pcp = per_cpu_ptr(zone->pageset, cpu);
1315 if (pcp->pcp.count) {
1321 cpumask_set_cpu(cpu, &cpus_with_pcps);
1323 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1325 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1328 #ifdef CONFIG_HIBERNATION
1330 void mark_free_pages(struct zone *zone)
1332 unsigned long pfn, max_zone_pfn;
1333 unsigned long flags;
1335 struct list_head *curr;
1337 if (zone_is_empty(zone))
1340 spin_lock_irqsave(&zone->lock, flags);
1342 max_zone_pfn = zone_end_pfn(zone);
1343 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1344 if (pfn_valid(pfn)) {
1345 struct page *page = pfn_to_page(pfn);
1347 if (!swsusp_page_is_forbidden(page))
1348 swsusp_unset_page_free(page);
1351 for_each_migratetype_order(order, t) {
1352 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1355 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1356 for (i = 0; i < (1UL << order); i++)
1357 swsusp_set_page_free(pfn_to_page(pfn + i));
1360 spin_unlock_irqrestore(&zone->lock, flags);
1362 #endif /* CONFIG_PM */
1365 * Free a 0-order page
1366 * cold == 1 ? free a cold page : free a hot page
1368 void free_hot_cold_page(struct page *page, int cold)
1370 struct zone *zone = page_zone(page);
1371 struct per_cpu_pages *pcp;
1372 unsigned long flags;
1375 if (!free_pages_prepare(page, 0))
1378 migratetype = get_pageblock_migratetype(page);
1379 set_freepage_migratetype(page, migratetype);
1380 local_irq_save(flags);
1381 __count_vm_event(PGFREE);
1384 * We only track unmovable, reclaimable and movable on pcp lists.
1385 * Free ISOLATE pages back to the allocator because they are being
1386 * offlined but treat RESERVE as movable pages so we can get those
1387 * areas back if necessary. Otherwise, we may have to free
1388 * excessively into the page allocator
1390 if (migratetype >= MIGRATE_PCPTYPES) {
1391 if (unlikely(is_migrate_isolate(migratetype))) {
1392 free_one_page(zone, page, 0, migratetype);
1395 migratetype = MIGRATE_MOVABLE;
1398 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1400 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1402 list_add(&page->lru, &pcp->lists[migratetype]);
1404 if (pcp->count >= pcp->high) {
1405 unsigned long batch = ACCESS_ONCE(pcp->batch);
1406 free_pcppages_bulk(zone, batch, pcp);
1407 pcp->count -= batch;
1411 local_irq_restore(flags);
1415 * Free a list of 0-order pages
1417 void free_hot_cold_page_list(struct list_head *list, int cold)
1419 struct page *page, *next;
1421 list_for_each_entry_safe(page, next, list, lru) {
1422 trace_mm_page_free_batched(page, cold);
1423 free_hot_cold_page(page, cold);
1428 * split_page takes a non-compound higher-order page, and splits it into
1429 * n (1<<order) sub-pages: page[0..n]
1430 * Each sub-page must be freed individually.
1432 * Note: this is probably too low level an operation for use in drivers.
1433 * Please consult with lkml before using this in your driver.
1435 void split_page(struct page *page, unsigned int order)
1439 VM_BUG_ON_PAGE(PageCompound(page), page);
1440 VM_BUG_ON_PAGE(!page_count(page), page);
1442 #ifdef CONFIG_KMEMCHECK
1444 * Split shadow pages too, because free(page[0]) would
1445 * otherwise free the whole shadow.
1447 if (kmemcheck_page_is_tracked(page))
1448 split_page(virt_to_page(page[0].shadow), order);
1451 for (i = 1; i < (1 << order); i++)
1452 set_page_refcounted(page + i);
1454 EXPORT_SYMBOL_GPL(split_page);
1456 static int __isolate_free_page(struct page *page, unsigned int order)
1458 unsigned long watermark;
1462 BUG_ON(!PageBuddy(page));
1464 zone = page_zone(page);
1465 mt = get_pageblock_migratetype(page);
1467 if (!is_migrate_isolate(mt)) {
1468 /* Obey watermarks as if the page was being allocated */
1469 watermark = low_wmark_pages(zone) + (1 << order);
1470 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1473 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1476 /* Remove page from free list */
1477 list_del(&page->lru);
1478 zone->free_area[order].nr_free--;
1479 rmv_page_order(page);
1481 /* Set the pageblock if the isolated page is at least a pageblock */
1482 if (order >= pageblock_order - 1) {
1483 struct page *endpage = page + (1 << order) - 1;
1484 for (; page < endpage; page += pageblock_nr_pages) {
1485 int mt = get_pageblock_migratetype(page);
1486 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1487 set_pageblock_migratetype(page,
1492 return 1UL << order;
1496 * Similar to split_page except the page is already free. As this is only
1497 * being used for migration, the migratetype of the block also changes.
1498 * As this is called with interrupts disabled, the caller is responsible
1499 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1502 * Note: this is probably too low level an operation for use in drivers.
1503 * Please consult with lkml before using this in your driver.
1505 int split_free_page(struct page *page)
1510 order = page_order(page);
1512 nr_pages = __isolate_free_page(page, order);
1516 /* Split into individual pages */
1517 set_page_refcounted(page);
1518 split_page(page, order);
1523 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1524 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1528 struct page *buffered_rmqueue(struct zone *preferred_zone,
1529 struct zone *zone, int order, gfp_t gfp_flags,
1532 unsigned long flags;
1534 int cold = !!(gfp_flags & __GFP_COLD);
1537 if (likely(order == 0)) {
1538 struct per_cpu_pages *pcp;
1539 struct list_head *list;
1541 local_irq_save(flags);
1542 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1543 list = &pcp->lists[migratetype];
1544 if (list_empty(list)) {
1545 pcp->count += rmqueue_bulk(zone, 0,
1548 if (unlikely(list_empty(list)))
1553 page = list_entry(list->prev, struct page, lru);
1555 page = list_entry(list->next, struct page, lru);
1557 list_del(&page->lru);
1560 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1562 * __GFP_NOFAIL is not to be used in new code.
1564 * All __GFP_NOFAIL callers should be fixed so that they
1565 * properly detect and handle allocation failures.
1567 * We most definitely don't want callers attempting to
1568 * allocate greater than order-1 page units with
1571 WARN_ON_ONCE(order > 1);
1573 spin_lock_irqsave(&zone->lock, flags);
1574 page = __rmqueue(zone, order, migratetype);
1575 spin_unlock(&zone->lock);
1578 __mod_zone_freepage_state(zone, -(1 << order),
1579 get_freepage_migratetype(page));
1582 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1584 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1585 zone_statistics(preferred_zone, zone, gfp_flags);
1586 local_irq_restore(flags);
1588 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1589 if (prep_new_page(page, order, gfp_flags))
1594 local_irq_restore(flags);
1598 #ifdef CONFIG_FAIL_PAGE_ALLOC
1601 struct fault_attr attr;
1603 u32 ignore_gfp_highmem;
1604 u32 ignore_gfp_wait;
1606 } fail_page_alloc = {
1607 .attr = FAULT_ATTR_INITIALIZER,
1608 .ignore_gfp_wait = 1,
1609 .ignore_gfp_highmem = 1,
1613 static int __init setup_fail_page_alloc(char *str)
1615 return setup_fault_attr(&fail_page_alloc.attr, str);
1617 __setup("fail_page_alloc=", setup_fail_page_alloc);
1619 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1621 if (order < fail_page_alloc.min_order)
1623 if (gfp_mask & __GFP_NOFAIL)
1625 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1627 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1630 return should_fail(&fail_page_alloc.attr, 1 << order);
1633 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1635 static int __init fail_page_alloc_debugfs(void)
1637 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1640 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1641 &fail_page_alloc.attr);
1643 return PTR_ERR(dir);
1645 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1646 &fail_page_alloc.ignore_gfp_wait))
1648 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1649 &fail_page_alloc.ignore_gfp_highmem))
1651 if (!debugfs_create_u32("min-order", mode, dir,
1652 &fail_page_alloc.min_order))
1657 debugfs_remove_recursive(dir);
1662 late_initcall(fail_page_alloc_debugfs);
1664 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1666 #else /* CONFIG_FAIL_PAGE_ALLOC */
1668 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1673 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1676 * Return true if free pages are above 'mark'. This takes into account the order
1677 * of the allocation.
1679 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1680 int classzone_idx, int alloc_flags, long free_pages)
1682 /* free_pages my go negative - that's OK */
1684 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1688 free_pages -= (1 << order) - 1;
1689 if (alloc_flags & ALLOC_HIGH)
1691 if (alloc_flags & ALLOC_HARDER)
1694 /* If allocation can't use CMA areas don't use free CMA pages */
1695 if (!(alloc_flags & ALLOC_CMA))
1696 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1699 if (free_pages - free_cma <= min + lowmem_reserve)
1701 for (o = 0; o < order; o++) {
1702 /* At the next order, this order's pages become unavailable */
1703 free_pages -= z->free_area[o].nr_free << o;
1705 /* Require fewer higher order pages to be free */
1708 if (free_pages <= min)
1714 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1715 int classzone_idx, int alloc_flags)
1717 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1718 zone_page_state(z, NR_FREE_PAGES));
1721 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1722 int classzone_idx, int alloc_flags)
1724 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1726 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1727 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1729 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1735 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1736 * skip over zones that are not allowed by the cpuset, or that have
1737 * been recently (in last second) found to be nearly full. See further
1738 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1739 * that have to skip over a lot of full or unallowed zones.
1741 * If the zonelist cache is present in the passed zonelist, then
1742 * returns a pointer to the allowed node mask (either the current
1743 * tasks mems_allowed, or node_states[N_MEMORY].)
1745 * If the zonelist cache is not available for this zonelist, does
1746 * nothing and returns NULL.
1748 * If the fullzones BITMAP in the zonelist cache is stale (more than
1749 * a second since last zap'd) then we zap it out (clear its bits.)
1751 * We hold off even calling zlc_setup, until after we've checked the
1752 * first zone in the zonelist, on the theory that most allocations will
1753 * be satisfied from that first zone, so best to examine that zone as
1754 * quickly as we can.
1756 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1758 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1759 nodemask_t *allowednodes; /* zonelist_cache approximation */
1761 zlc = zonelist->zlcache_ptr;
1765 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1766 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1767 zlc->last_full_zap = jiffies;
1770 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1771 &cpuset_current_mems_allowed :
1772 &node_states[N_MEMORY];
1773 return allowednodes;
1777 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1778 * if it is worth looking at further for free memory:
1779 * 1) Check that the zone isn't thought to be full (doesn't have its
1780 * bit set in the zonelist_cache fullzones BITMAP).
1781 * 2) Check that the zones node (obtained from the zonelist_cache
1782 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1783 * Return true (non-zero) if zone is worth looking at further, or
1784 * else return false (zero) if it is not.
1786 * This check -ignores- the distinction between various watermarks,
1787 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1788 * found to be full for any variation of these watermarks, it will
1789 * be considered full for up to one second by all requests, unless
1790 * we are so low on memory on all allowed nodes that we are forced
1791 * into the second scan of the zonelist.
1793 * In the second scan we ignore this zonelist cache and exactly
1794 * apply the watermarks to all zones, even it is slower to do so.
1795 * We are low on memory in the second scan, and should leave no stone
1796 * unturned looking for a free page.
1798 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1799 nodemask_t *allowednodes)
1801 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1802 int i; /* index of *z in zonelist zones */
1803 int n; /* node that zone *z is on */
1805 zlc = zonelist->zlcache_ptr;
1809 i = z - zonelist->_zonerefs;
1812 /* This zone is worth trying if it is allowed but not full */
1813 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1817 * Given 'z' scanning a zonelist, set the corresponding bit in
1818 * zlc->fullzones, so that subsequent attempts to allocate a page
1819 * from that zone don't waste time re-examining it.
1821 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1823 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1824 int i; /* index of *z in zonelist zones */
1826 zlc = zonelist->zlcache_ptr;
1830 i = z - zonelist->_zonerefs;
1832 set_bit(i, zlc->fullzones);
1836 * clear all zones full, called after direct reclaim makes progress so that
1837 * a zone that was recently full is not skipped over for up to a second
1839 static void zlc_clear_zones_full(struct zonelist *zonelist)
1841 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1843 zlc = zonelist->zlcache_ptr;
1847 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1850 static bool zone_local(struct zone *local_zone, struct zone *zone)
1852 return local_zone->node == zone->node;
1855 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1857 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1861 #else /* CONFIG_NUMA */
1863 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1868 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1869 nodemask_t *allowednodes)
1874 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1878 static void zlc_clear_zones_full(struct zonelist *zonelist)
1882 static bool zone_local(struct zone *local_zone, struct zone *zone)
1887 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1892 #endif /* CONFIG_NUMA */
1895 * get_page_from_freelist goes through the zonelist trying to allocate
1898 static struct page *
1899 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1900 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1901 struct zone *preferred_zone, int migratetype)
1904 struct page *page = NULL;
1907 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1908 int zlc_active = 0; /* set if using zonelist_cache */
1909 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1911 classzone_idx = zone_idx(preferred_zone);
1914 * Scan zonelist, looking for a zone with enough free.
1915 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1917 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1918 high_zoneidx, nodemask) {
1921 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1922 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1924 if ((alloc_flags & ALLOC_CPUSET) &&
1925 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1927 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1928 if (unlikely(alloc_flags & ALLOC_NO_WATERMARKS))
1931 * Distribute pages in proportion to the individual
1932 * zone size to ensure fair page aging. The zone a
1933 * page was allocated in should have no effect on the
1934 * time the page has in memory before being reclaimed.
1936 if (alloc_flags & ALLOC_FAIR) {
1937 if (!zone_local(preferred_zone, zone))
1939 if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1943 * When allocating a page cache page for writing, we
1944 * want to get it from a zone that is within its dirty
1945 * limit, such that no single zone holds more than its
1946 * proportional share of globally allowed dirty pages.
1947 * The dirty limits take into account the zone's
1948 * lowmem reserves and high watermark so that kswapd
1949 * should be able to balance it without having to
1950 * write pages from its LRU list.
1952 * This may look like it could increase pressure on
1953 * lower zones by failing allocations in higher zones
1954 * before they are full. But the pages that do spill
1955 * over are limited as the lower zones are protected
1956 * by this very same mechanism. It should not become
1957 * a practical burden to them.
1959 * XXX: For now, allow allocations to potentially
1960 * exceed the per-zone dirty limit in the slowpath
1961 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1962 * which is important when on a NUMA setup the allowed
1963 * zones are together not big enough to reach the
1964 * global limit. The proper fix for these situations
1965 * will require awareness of zones in the
1966 * dirty-throttling and the flusher threads.
1968 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1969 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1970 goto this_zone_full;
1972 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1973 if (!zone_watermark_ok(zone, order, mark,
1974 classzone_idx, alloc_flags)) {
1977 if (IS_ENABLED(CONFIG_NUMA) &&
1978 !did_zlc_setup && nr_online_nodes > 1) {
1980 * we do zlc_setup if there are multiple nodes
1981 * and before considering the first zone allowed
1984 allowednodes = zlc_setup(zonelist, alloc_flags);
1989 if (zone_reclaim_mode == 0 ||
1990 !zone_allows_reclaim(preferred_zone, zone))
1991 goto this_zone_full;
1994 * As we may have just activated ZLC, check if the first
1995 * eligible zone has failed zone_reclaim recently.
1997 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1998 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2001 ret = zone_reclaim(zone, gfp_mask, order);
2003 case ZONE_RECLAIM_NOSCAN:
2006 case ZONE_RECLAIM_FULL:
2007 /* scanned but unreclaimable */
2010 /* did we reclaim enough */
2011 if (zone_watermark_ok(zone, order, mark,
2012 classzone_idx, alloc_flags))
2016 * Failed to reclaim enough to meet watermark.
2017 * Only mark the zone full if checking the min
2018 * watermark or if we failed to reclaim just
2019 * 1<<order pages or else the page allocator
2020 * fastpath will prematurely mark zones full
2021 * when the watermark is between the low and
2024 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2025 ret == ZONE_RECLAIM_SOME)
2026 goto this_zone_full;
2033 page = buffered_rmqueue(preferred_zone, zone, order,
2034 gfp_mask, migratetype);
2038 if (IS_ENABLED(CONFIG_NUMA))
2039 zlc_mark_zone_full(zonelist, z);
2042 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2043 /* Disable zlc cache for second zonelist scan */
2050 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2051 * necessary to allocate the page. The expectation is
2052 * that the caller is taking steps that will free more
2053 * memory. The caller should avoid the page being used
2054 * for !PFMEMALLOC purposes.
2056 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2062 * Large machines with many possible nodes should not always dump per-node
2063 * meminfo in irq context.
2065 static inline bool should_suppress_show_mem(void)
2070 ret = in_interrupt();
2075 static DEFINE_RATELIMIT_STATE(nopage_rs,
2076 DEFAULT_RATELIMIT_INTERVAL,
2077 DEFAULT_RATELIMIT_BURST);
2079 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2081 unsigned int filter = SHOW_MEM_FILTER_NODES;
2083 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2084 debug_guardpage_minorder() > 0)
2088 * This documents exceptions given to allocations in certain
2089 * contexts that are allowed to allocate outside current's set
2092 if (!(gfp_mask & __GFP_NOMEMALLOC))
2093 if (test_thread_flag(TIF_MEMDIE) ||
2094 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2095 filter &= ~SHOW_MEM_FILTER_NODES;
2096 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2097 filter &= ~SHOW_MEM_FILTER_NODES;
2100 struct va_format vaf;
2103 va_start(args, fmt);
2108 pr_warn("%pV", &vaf);
2113 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2114 current->comm, order, gfp_mask);
2117 if (!should_suppress_show_mem())
2122 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2123 unsigned long did_some_progress,
2124 unsigned long pages_reclaimed)
2126 /* Do not loop if specifically requested */
2127 if (gfp_mask & __GFP_NORETRY)
2130 /* Always retry if specifically requested */
2131 if (gfp_mask & __GFP_NOFAIL)
2135 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2136 * making forward progress without invoking OOM. Suspend also disables
2137 * storage devices so kswapd will not help. Bail if we are suspending.
2139 if (!did_some_progress && pm_suspended_storage())
2143 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2144 * means __GFP_NOFAIL, but that may not be true in other
2147 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2151 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2152 * specified, then we retry until we no longer reclaim any pages
2153 * (above), or we've reclaimed an order of pages at least as
2154 * large as the allocation's order. In both cases, if the
2155 * allocation still fails, we stop retrying.
2157 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2163 static inline struct page *
2164 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2165 struct zonelist *zonelist, enum zone_type high_zoneidx,
2166 nodemask_t *nodemask, struct zone *preferred_zone,
2171 /* Acquire the OOM killer lock for the zones in zonelist */
2172 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2173 schedule_timeout_uninterruptible(1);
2178 * Go through the zonelist yet one more time, keep very high watermark
2179 * here, this is only to catch a parallel oom killing, we must fail if
2180 * we're still under heavy pressure.
2182 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2183 order, zonelist, high_zoneidx,
2184 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2185 preferred_zone, migratetype);
2189 if (!(gfp_mask & __GFP_NOFAIL)) {
2190 /* The OOM killer will not help higher order allocs */
2191 if (order > PAGE_ALLOC_COSTLY_ORDER)
2193 /* The OOM killer does not needlessly kill tasks for lowmem */
2194 if (high_zoneidx < ZONE_NORMAL)
2197 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2198 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2199 * The caller should handle page allocation failure by itself if
2200 * it specifies __GFP_THISNODE.
2201 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2203 if (gfp_mask & __GFP_THISNODE)
2206 /* Exhausted what can be done so it's blamo time */
2207 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2210 clear_zonelist_oom(zonelist, gfp_mask);
2214 #ifdef CONFIG_COMPACTION
2215 /* Try memory compaction for high-order allocations before reclaim */
2216 static struct page *
2217 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2218 struct zonelist *zonelist, enum zone_type high_zoneidx,
2219 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2220 int migratetype, enum migrate_mode mode,
2221 bool *contended_compaction, bool *deferred_compaction,
2222 unsigned long *did_some_progress)
2227 if (compaction_deferred(preferred_zone, order)) {
2228 *deferred_compaction = true;
2232 current->flags |= PF_MEMALLOC;
2233 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2235 contended_compaction);
2236 current->flags &= ~PF_MEMALLOC;
2238 if (*did_some_progress != COMPACT_SKIPPED) {
2241 /* Page migration frees to the PCP lists but we want merging */
2242 drain_pages(get_cpu());
2245 page = get_page_from_freelist(gfp_mask, nodemask,
2246 order, zonelist, high_zoneidx,
2247 alloc_flags & ~ALLOC_NO_WATERMARKS,
2248 preferred_zone, migratetype);
2250 preferred_zone->compact_blockskip_flush = false;
2251 compaction_defer_reset(preferred_zone, order, true);
2252 count_vm_event(COMPACTSUCCESS);
2257 * It's bad if compaction run occurs and fails.
2258 * The most likely reason is that pages exist,
2259 * but not enough to satisfy watermarks.
2261 count_vm_event(COMPACTFAIL);
2264 * As async compaction considers a subset of pageblocks, only
2265 * defer if the failure was a sync compaction failure.
2267 if (mode != MIGRATE_ASYNC)
2268 defer_compaction(preferred_zone, order);
2276 static inline struct page *
2277 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2278 struct zonelist *zonelist, enum zone_type high_zoneidx,
2279 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2280 int migratetype, enum migrate_mode mode, bool *contended_compaction,
2281 bool *deferred_compaction, unsigned long *did_some_progress)
2285 #endif /* CONFIG_COMPACTION */
2287 /* Perform direct synchronous page reclaim */
2289 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2290 nodemask_t *nodemask)
2292 struct reclaim_state reclaim_state;
2297 /* We now go into synchronous reclaim */
2298 cpuset_memory_pressure_bump();
2299 current->flags |= PF_MEMALLOC;
2300 lockdep_set_current_reclaim_state(gfp_mask);
2301 reclaim_state.reclaimed_slab = 0;
2302 current->reclaim_state = &reclaim_state;
2304 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2306 current->reclaim_state = NULL;
2307 lockdep_clear_current_reclaim_state();
2308 current->flags &= ~PF_MEMALLOC;
2315 /* The really slow allocator path where we enter direct reclaim */
2316 static inline struct page *
2317 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2318 struct zonelist *zonelist, enum zone_type high_zoneidx,
2319 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2320 int migratetype, unsigned long *did_some_progress)
2322 struct page *page = NULL;
2323 bool drained = false;
2325 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2327 if (unlikely(!(*did_some_progress)))
2330 /* After successful reclaim, reconsider all zones for allocation */
2331 if (IS_ENABLED(CONFIG_NUMA))
2332 zlc_clear_zones_full(zonelist);
2335 page = get_page_from_freelist(gfp_mask, nodemask, order,
2336 zonelist, high_zoneidx,
2337 alloc_flags & ~ALLOC_NO_WATERMARKS,
2338 preferred_zone, migratetype);
2341 * If an allocation failed after direct reclaim, it could be because
2342 * pages are pinned on the per-cpu lists. Drain them and try again
2344 if (!page && !drained) {
2354 * This is called in the allocator slow-path if the allocation request is of
2355 * sufficient urgency to ignore watermarks and take other desperate measures
2357 static inline struct page *
2358 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2359 struct zonelist *zonelist, enum zone_type high_zoneidx,
2360 nodemask_t *nodemask, struct zone *preferred_zone,
2366 page = get_page_from_freelist(gfp_mask, nodemask, order,
2367 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2368 preferred_zone, migratetype);
2370 if (!page && gfp_mask & __GFP_NOFAIL)
2371 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2372 } while (!page && (gfp_mask & __GFP_NOFAIL));
2377 static void reset_alloc_batches(struct zonelist *zonelist,
2378 enum zone_type high_zoneidx,
2379 struct zone *preferred_zone)
2384 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2386 * Only reset the batches of zones that were actually
2387 * considered in the fairness pass, we don't want to
2388 * trash fairness information for zones that are not
2389 * actually part of this zonelist's round-robin cycle.
2391 if (!zone_local(preferred_zone, zone))
2393 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2394 high_wmark_pages(zone) - low_wmark_pages(zone) -
2395 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2399 static void wake_all_kswapds(unsigned int order,
2400 struct zonelist *zonelist,
2401 enum zone_type high_zoneidx,
2402 struct zone *preferred_zone)
2407 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2408 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2412 gfp_to_alloc_flags(gfp_t gfp_mask)
2414 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2415 const gfp_t wait = gfp_mask & __GFP_WAIT;
2417 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2418 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2421 * The caller may dip into page reserves a bit more if the caller
2422 * cannot run direct reclaim, or if the caller has realtime scheduling
2423 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2424 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2426 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2430 * Not worth trying to allocate harder for
2431 * __GFP_NOMEMALLOC even if it can't schedule.
2433 if (!(gfp_mask & __GFP_NOMEMALLOC))
2434 alloc_flags |= ALLOC_HARDER;
2436 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2437 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2439 alloc_flags &= ~ALLOC_CPUSET;
2440 } else if (unlikely(rt_task(current)) && !in_interrupt())
2441 alloc_flags |= ALLOC_HARDER;
2443 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2444 if (gfp_mask & __GFP_MEMALLOC)
2445 alloc_flags |= ALLOC_NO_WATERMARKS;
2446 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2447 alloc_flags |= ALLOC_NO_WATERMARKS;
2448 else if (!in_interrupt() &&
2449 ((current->flags & PF_MEMALLOC) ||
2450 unlikely(test_thread_flag(TIF_MEMDIE))))
2451 alloc_flags |= ALLOC_NO_WATERMARKS;
2454 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2455 alloc_flags |= ALLOC_CMA;
2460 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2462 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2465 static inline struct page *
2466 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2467 struct zonelist *zonelist, enum zone_type high_zoneidx,
2468 nodemask_t *nodemask, struct zone *preferred_zone,
2471 const gfp_t wait = gfp_mask & __GFP_WAIT;
2472 struct page *page = NULL;
2474 unsigned long pages_reclaimed = 0;
2475 unsigned long did_some_progress;
2476 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2477 bool deferred_compaction = false;
2478 bool contended_compaction = false;
2481 * In the slowpath, we sanity check order to avoid ever trying to
2482 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2483 * be using allocators in order of preference for an area that is
2486 if (order >= MAX_ORDER) {
2487 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2492 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2493 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2494 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2495 * using a larger set of nodes after it has established that the
2496 * allowed per node queues are empty and that nodes are
2499 if (IS_ENABLED(CONFIG_NUMA) &&
2500 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2504 if (!(gfp_mask & __GFP_NO_KSWAPD))
2505 wake_all_kswapds(order, zonelist, high_zoneidx, preferred_zone);
2508 * OK, we're below the kswapd watermark and have kicked background
2509 * reclaim. Now things get more complex, so set up alloc_flags according
2510 * to how we want to proceed.
2512 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2515 * Find the true preferred zone if the allocation is unconstrained by
2518 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2519 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2523 /* This is the last chance, in general, before the goto nopage. */
2524 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2525 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2526 preferred_zone, migratetype);
2530 /* Allocate without watermarks if the context allows */
2531 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2533 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2534 * the allocation is high priority and these type of
2535 * allocations are system rather than user orientated
2537 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2539 page = __alloc_pages_high_priority(gfp_mask, order,
2540 zonelist, high_zoneidx, nodemask,
2541 preferred_zone, migratetype);
2547 /* Atomic allocations - we can't balance anything */
2550 * All existing users of the deprecated __GFP_NOFAIL are
2551 * blockable, so warn of any new users that actually allow this
2552 * type of allocation to fail.
2554 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2558 /* Avoid recursion of direct reclaim */
2559 if (current->flags & PF_MEMALLOC)
2562 /* Avoid allocations with no watermarks from looping endlessly */
2563 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2567 * Try direct compaction. The first pass is asynchronous. Subsequent
2568 * attempts after direct reclaim are synchronous
2570 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2571 high_zoneidx, nodemask, alloc_flags,
2572 preferred_zone, migratetype,
2573 migration_mode, &contended_compaction,
2574 &deferred_compaction,
2575 &did_some_progress);
2580 * It can become very expensive to allocate transparent hugepages at
2581 * fault, so use asynchronous memory compaction for THP unless it is
2582 * khugepaged trying to collapse.
2584 if (!(gfp_mask & __GFP_NO_KSWAPD) || (current->flags & PF_KTHREAD))
2585 migration_mode = MIGRATE_SYNC_LIGHT;
2588 * If compaction is deferred for high-order allocations, it is because
2589 * sync compaction recently failed. In this is the case and the caller
2590 * requested a movable allocation that does not heavily disrupt the
2591 * system then fail the allocation instead of entering direct reclaim.
2593 if ((deferred_compaction || contended_compaction) &&
2594 (gfp_mask & __GFP_NO_KSWAPD))
2597 /* Try direct reclaim and then allocating */
2598 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2599 zonelist, high_zoneidx,
2601 alloc_flags, preferred_zone,
2602 migratetype, &did_some_progress);
2607 * If we failed to make any progress reclaiming, then we are
2608 * running out of options and have to consider going OOM
2610 if (!did_some_progress) {
2611 if (oom_gfp_allowed(gfp_mask)) {
2612 if (oom_killer_disabled)
2614 /* Coredumps can quickly deplete all memory reserves */
2615 if ((current->flags & PF_DUMPCORE) &&
2616 !(gfp_mask & __GFP_NOFAIL))
2618 page = __alloc_pages_may_oom(gfp_mask, order,
2619 zonelist, high_zoneidx,
2620 nodemask, preferred_zone,
2625 if (!(gfp_mask & __GFP_NOFAIL)) {
2627 * The oom killer is not called for high-order
2628 * allocations that may fail, so if no progress
2629 * is being made, there are no other options and
2630 * retrying is unlikely to help.
2632 if (order > PAGE_ALLOC_COSTLY_ORDER)
2635 * The oom killer is not called for lowmem
2636 * allocations to prevent needlessly killing
2639 if (high_zoneidx < ZONE_NORMAL)
2647 /* Check if we should retry the allocation */
2648 pages_reclaimed += did_some_progress;
2649 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2651 /* Wait for some write requests to complete then retry */
2652 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2656 * High-order allocations do not necessarily loop after
2657 * direct reclaim and reclaim/compaction depends on compaction
2658 * being called after reclaim so call directly if necessary
2660 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2661 high_zoneidx, nodemask, alloc_flags,
2662 preferred_zone, migratetype,
2663 migration_mode, &contended_compaction,
2664 &deferred_compaction,
2665 &did_some_progress);
2671 warn_alloc_failed(gfp_mask, order, NULL);
2674 if (kmemcheck_enabled)
2675 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2681 * This is the 'heart' of the zoned buddy allocator.
2684 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2685 struct zonelist *zonelist, nodemask_t *nodemask)
2687 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2688 struct zone *preferred_zone;
2689 struct page *page = NULL;
2690 int migratetype = allocflags_to_migratetype(gfp_mask);
2691 unsigned int cpuset_mems_cookie;
2692 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2694 gfp_mask &= gfp_allowed_mask;
2696 lockdep_trace_alloc(gfp_mask);
2698 might_sleep_if(gfp_mask & __GFP_WAIT);
2700 if (should_fail_alloc_page(gfp_mask, order))
2704 * Check the zones suitable for the gfp_mask contain at least one
2705 * valid zone. It's possible to have an empty zonelist as a result
2706 * of GFP_THISNODE and a memoryless node
2708 if (unlikely(!zonelist->_zonerefs->zone))
2712 cpuset_mems_cookie = read_mems_allowed_begin();
2714 /* The preferred zone is used for statistics later */
2715 first_zones_zonelist(zonelist, high_zoneidx,
2716 nodemask ? : &cpuset_current_mems_allowed,
2718 if (!preferred_zone)
2722 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2723 alloc_flags |= ALLOC_CMA;
2726 /* First allocation attempt */
2727 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2728 zonelist, high_zoneidx, alloc_flags,
2729 preferred_zone, migratetype);
2730 if (unlikely(!page)) {
2732 * The first pass makes sure allocations are spread
2733 * fairly within the local node. However, the local
2734 * node might have free pages left after the fairness
2735 * batches are exhausted, and remote zones haven't
2736 * even been considered yet. Try once more without
2737 * fairness, and include remote zones now, before
2738 * entering the slowpath and waking kswapd: prefer
2739 * spilling to a remote zone over swapping locally.
2741 if (alloc_flags & ALLOC_FAIR) {
2742 reset_alloc_batches(zonelist, high_zoneidx,
2744 alloc_flags &= ~ALLOC_FAIR;
2748 * Runtime PM, block IO and its error handling path
2749 * can deadlock because I/O on the device might not
2752 gfp_mask = memalloc_noio_flags(gfp_mask);
2753 page = __alloc_pages_slowpath(gfp_mask, order,
2754 zonelist, high_zoneidx, nodemask,
2755 preferred_zone, migratetype);
2758 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2762 * When updating a task's mems_allowed, it is possible to race with
2763 * parallel threads in such a way that an allocation can fail while
2764 * the mask is being updated. If a page allocation is about to fail,
2765 * check if the cpuset changed during allocation and if so, retry.
2767 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2772 EXPORT_SYMBOL(__alloc_pages_nodemask);
2775 * Common helper functions.
2777 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2782 * __get_free_pages() returns a 32-bit address, which cannot represent
2785 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2787 page = alloc_pages(gfp_mask, order);
2790 return (unsigned long) page_address(page);
2792 EXPORT_SYMBOL(__get_free_pages);
2794 unsigned long get_zeroed_page(gfp_t gfp_mask)
2796 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2798 EXPORT_SYMBOL(get_zeroed_page);
2800 void __free_pages(struct page *page, unsigned int order)
2802 if (put_page_testzero(page)) {
2804 free_hot_cold_page(page, 0);
2806 __free_pages_ok(page, order);
2810 EXPORT_SYMBOL(__free_pages);
2812 void free_pages(unsigned long addr, unsigned int order)
2815 VM_BUG_ON(!virt_addr_valid((void *)addr));
2816 __free_pages(virt_to_page((void *)addr), order);
2820 EXPORT_SYMBOL(free_pages);
2823 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2824 * of the current memory cgroup.
2826 * It should be used when the caller would like to use kmalloc, but since the
2827 * allocation is large, it has to fall back to the page allocator.
2829 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2832 struct mem_cgroup *memcg = NULL;
2834 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2836 page = alloc_pages(gfp_mask, order);
2837 memcg_kmem_commit_charge(page, memcg, order);
2841 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2844 struct mem_cgroup *memcg = NULL;
2846 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2848 page = alloc_pages_node(nid, gfp_mask, order);
2849 memcg_kmem_commit_charge(page, memcg, order);
2854 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2857 void __free_kmem_pages(struct page *page, unsigned int order)
2859 memcg_kmem_uncharge_pages(page, order);
2860 __free_pages(page, order);
2863 void free_kmem_pages(unsigned long addr, unsigned int order)
2866 VM_BUG_ON(!virt_addr_valid((void *)addr));
2867 __free_kmem_pages(virt_to_page((void *)addr), order);
2871 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2874 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2875 unsigned long used = addr + PAGE_ALIGN(size);
2877 split_page(virt_to_page((void *)addr), order);
2878 while (used < alloc_end) {
2883 return (void *)addr;
2887 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2888 * @size: the number of bytes to allocate
2889 * @gfp_mask: GFP flags for the allocation
2891 * This function is similar to alloc_pages(), except that it allocates the
2892 * minimum number of pages to satisfy the request. alloc_pages() can only
2893 * allocate memory in power-of-two pages.
2895 * This function is also limited by MAX_ORDER.
2897 * Memory allocated by this function must be released by free_pages_exact().
2899 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2901 unsigned int order = get_order(size);
2904 addr = __get_free_pages(gfp_mask, order);
2905 return make_alloc_exact(addr, order, size);
2907 EXPORT_SYMBOL(alloc_pages_exact);
2910 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2912 * @nid: the preferred node ID where memory should be allocated
2913 * @size: the number of bytes to allocate
2914 * @gfp_mask: GFP flags for the allocation
2916 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2918 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2921 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2923 unsigned order = get_order(size);
2924 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2927 return make_alloc_exact((unsigned long)page_address(p), order, size);
2929 EXPORT_SYMBOL(alloc_pages_exact_nid);
2932 * free_pages_exact - release memory allocated via alloc_pages_exact()
2933 * @virt: the value returned by alloc_pages_exact.
2934 * @size: size of allocation, same value as passed to alloc_pages_exact().
2936 * Release the memory allocated by a previous call to alloc_pages_exact.
2938 void free_pages_exact(void *virt, size_t size)
2940 unsigned long addr = (unsigned long)virt;
2941 unsigned long end = addr + PAGE_ALIGN(size);
2943 while (addr < end) {
2948 EXPORT_SYMBOL(free_pages_exact);
2951 * nr_free_zone_pages - count number of pages beyond high watermark
2952 * @offset: The zone index of the highest zone
2954 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2955 * high watermark within all zones at or below a given zone index. For each
2956 * zone, the number of pages is calculated as:
2957 * managed_pages - high_pages
2959 static unsigned long nr_free_zone_pages(int offset)
2964 /* Just pick one node, since fallback list is circular */
2965 unsigned long sum = 0;
2967 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2969 for_each_zone_zonelist(zone, z, zonelist, offset) {
2970 unsigned long size = zone->managed_pages;
2971 unsigned long high = high_wmark_pages(zone);
2980 * nr_free_buffer_pages - count number of pages beyond high watermark
2982 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2983 * watermark within ZONE_DMA and ZONE_NORMAL.
2985 unsigned long nr_free_buffer_pages(void)
2987 return nr_free_zone_pages(gfp_zone(GFP_USER));
2989 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2992 * nr_free_pagecache_pages - count number of pages beyond high watermark
2994 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2995 * high watermark within all zones.
2997 unsigned long nr_free_pagecache_pages(void)
2999 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3002 static inline void show_node(struct zone *zone)
3004 if (IS_ENABLED(CONFIG_NUMA))
3005 printk("Node %d ", zone_to_nid(zone));
3008 void si_meminfo(struct sysinfo *val)
3010 val->totalram = totalram_pages;
3012 val->freeram = global_page_state(NR_FREE_PAGES);
3013 val->bufferram = nr_blockdev_pages();
3014 val->totalhigh = totalhigh_pages;
3015 val->freehigh = nr_free_highpages();
3016 val->mem_unit = PAGE_SIZE;
3019 EXPORT_SYMBOL(si_meminfo);
3022 void si_meminfo_node(struct sysinfo *val, int nid)
3024 int zone_type; /* needs to be signed */
3025 unsigned long managed_pages = 0;
3026 pg_data_t *pgdat = NODE_DATA(nid);
3028 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3029 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3030 val->totalram = managed_pages;
3031 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3032 #ifdef CONFIG_HIGHMEM
3033 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3034 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3040 val->mem_unit = PAGE_SIZE;
3045 * Determine whether the node should be displayed or not, depending on whether
3046 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3048 bool skip_free_areas_node(unsigned int flags, int nid)
3051 unsigned int cpuset_mems_cookie;
3053 if (!(flags & SHOW_MEM_FILTER_NODES))
3057 cpuset_mems_cookie = read_mems_allowed_begin();
3058 ret = !node_isset(nid, cpuset_current_mems_allowed);
3059 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3064 #define K(x) ((x) << (PAGE_SHIFT-10))
3066 static void show_migration_types(unsigned char type)
3068 static const char types[MIGRATE_TYPES] = {
3069 [MIGRATE_UNMOVABLE] = 'U',
3070 [MIGRATE_RECLAIMABLE] = 'E',
3071 [MIGRATE_MOVABLE] = 'M',
3072 [MIGRATE_RESERVE] = 'R',
3074 [MIGRATE_CMA] = 'C',
3076 #ifdef CONFIG_MEMORY_ISOLATION
3077 [MIGRATE_ISOLATE] = 'I',
3080 char tmp[MIGRATE_TYPES + 1];
3084 for (i = 0; i < MIGRATE_TYPES; i++) {
3085 if (type & (1 << i))
3090 printk("(%s) ", tmp);
3094 * Show free area list (used inside shift_scroll-lock stuff)
3095 * We also calculate the percentage fragmentation. We do this by counting the
3096 * memory on each free list with the exception of the first item on the list.
3097 * Suppresses nodes that are not allowed by current's cpuset if
3098 * SHOW_MEM_FILTER_NODES is passed.
3100 void show_free_areas(unsigned int filter)
3105 for_each_populated_zone(zone) {
3106 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3109 printk("%s per-cpu:\n", zone->name);
3111 for_each_online_cpu(cpu) {
3112 struct per_cpu_pageset *pageset;
3114 pageset = per_cpu_ptr(zone->pageset, cpu);
3116 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3117 cpu, pageset->pcp.high,
3118 pageset->pcp.batch, pageset->pcp.count);
3122 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3123 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3125 " dirty:%lu writeback:%lu unstable:%lu\n"
3126 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3127 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3129 global_page_state(NR_ACTIVE_ANON),
3130 global_page_state(NR_INACTIVE_ANON),
3131 global_page_state(NR_ISOLATED_ANON),
3132 global_page_state(NR_ACTIVE_FILE),
3133 global_page_state(NR_INACTIVE_FILE),
3134 global_page_state(NR_ISOLATED_FILE),
3135 global_page_state(NR_UNEVICTABLE),
3136 global_page_state(NR_FILE_DIRTY),
3137 global_page_state(NR_WRITEBACK),
3138 global_page_state(NR_UNSTABLE_NFS),
3139 global_page_state(NR_FREE_PAGES),
3140 global_page_state(NR_SLAB_RECLAIMABLE),
3141 global_page_state(NR_SLAB_UNRECLAIMABLE),
3142 global_page_state(NR_FILE_MAPPED),
3143 global_page_state(NR_SHMEM),
3144 global_page_state(NR_PAGETABLE),
3145 global_page_state(NR_BOUNCE),
3146 global_page_state(NR_FREE_CMA_PAGES));
3148 for_each_populated_zone(zone) {
3151 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3159 " active_anon:%lukB"
3160 " inactive_anon:%lukB"
3161 " active_file:%lukB"
3162 " inactive_file:%lukB"
3163 " unevictable:%lukB"
3164 " isolated(anon):%lukB"
3165 " isolated(file):%lukB"
3173 " slab_reclaimable:%lukB"
3174 " slab_unreclaimable:%lukB"
3175 " kernel_stack:%lukB"
3180 " writeback_tmp:%lukB"
3181 " pages_scanned:%lu"
3182 " all_unreclaimable? %s"
3185 K(zone_page_state(zone, NR_FREE_PAGES)),
3186 K(min_wmark_pages(zone)),
3187 K(low_wmark_pages(zone)),
3188 K(high_wmark_pages(zone)),
3189 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3190 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3191 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3192 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3193 K(zone_page_state(zone, NR_UNEVICTABLE)),
3194 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3195 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3196 K(zone->present_pages),
3197 K(zone->managed_pages),
3198 K(zone_page_state(zone, NR_MLOCK)),
3199 K(zone_page_state(zone, NR_FILE_DIRTY)),
3200 K(zone_page_state(zone, NR_WRITEBACK)),
3201 K(zone_page_state(zone, NR_FILE_MAPPED)),
3202 K(zone_page_state(zone, NR_SHMEM)),
3203 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3204 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3205 zone_page_state(zone, NR_KERNEL_STACK) *
3207 K(zone_page_state(zone, NR_PAGETABLE)),
3208 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3209 K(zone_page_state(zone, NR_BOUNCE)),
3210 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3211 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3212 zone->pages_scanned,
3213 (!zone_reclaimable(zone) ? "yes" : "no")
3215 printk("lowmem_reserve[]:");
3216 for (i = 0; i < MAX_NR_ZONES; i++)
3217 printk(" %lu", zone->lowmem_reserve[i]);
3221 for_each_populated_zone(zone) {
3222 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3223 unsigned char types[MAX_ORDER];
3225 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3228 printk("%s: ", zone->name);
3230 spin_lock_irqsave(&zone->lock, flags);
3231 for (order = 0; order < MAX_ORDER; order++) {
3232 struct free_area *area = &zone->free_area[order];
3235 nr[order] = area->nr_free;
3236 total += nr[order] << order;
3239 for (type = 0; type < MIGRATE_TYPES; type++) {
3240 if (!list_empty(&area->free_list[type]))
3241 types[order] |= 1 << type;
3244 spin_unlock_irqrestore(&zone->lock, flags);
3245 for (order = 0; order < MAX_ORDER; order++) {
3246 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3248 show_migration_types(types[order]);
3250 printk("= %lukB\n", K(total));
3253 hugetlb_show_meminfo();
3255 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3257 show_swap_cache_info();
3260 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3262 zoneref->zone = zone;
3263 zoneref->zone_idx = zone_idx(zone);
3267 * Builds allocation fallback zone lists.
3269 * Add all populated zones of a node to the zonelist.
3271 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3275 enum zone_type zone_type = MAX_NR_ZONES;
3279 zone = pgdat->node_zones + zone_type;
3280 if (populated_zone(zone)) {
3281 zoneref_set_zone(zone,
3282 &zonelist->_zonerefs[nr_zones++]);
3283 check_highest_zone(zone_type);
3285 } while (zone_type);
3293 * 0 = automatic detection of better ordering.
3294 * 1 = order by ([node] distance, -zonetype)
3295 * 2 = order by (-zonetype, [node] distance)
3297 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3298 * the same zonelist. So only NUMA can configure this param.
3300 #define ZONELIST_ORDER_DEFAULT 0
3301 #define ZONELIST_ORDER_NODE 1
3302 #define ZONELIST_ORDER_ZONE 2
3304 /* zonelist order in the kernel.
3305 * set_zonelist_order() will set this to NODE or ZONE.
3307 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3308 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3312 /* The value user specified ....changed by config */
3313 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3314 /* string for sysctl */
3315 #define NUMA_ZONELIST_ORDER_LEN 16
3316 char numa_zonelist_order[16] = "default";
3319 * interface for configure zonelist ordering.
3320 * command line option "numa_zonelist_order"
3321 * = "[dD]efault - default, automatic configuration.
3322 * = "[nN]ode - order by node locality, then by zone within node
3323 * = "[zZ]one - order by zone, then by locality within zone
3326 static int __parse_numa_zonelist_order(char *s)
3328 if (*s == 'd' || *s == 'D') {
3329 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3330 } else if (*s == 'n' || *s == 'N') {
3331 user_zonelist_order = ZONELIST_ORDER_NODE;
3332 } else if (*s == 'z' || *s == 'Z') {
3333 user_zonelist_order = ZONELIST_ORDER_ZONE;
3336 "Ignoring invalid numa_zonelist_order value: "
3343 static __init int setup_numa_zonelist_order(char *s)
3350 ret = __parse_numa_zonelist_order(s);
3352 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3356 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3359 * sysctl handler for numa_zonelist_order
3361 int numa_zonelist_order_handler(ctl_table *table, int write,
3362 void __user *buffer, size_t *length,
3365 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3367 static DEFINE_MUTEX(zl_order_mutex);
3369 mutex_lock(&zl_order_mutex);
3371 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3375 strcpy(saved_string, (char *)table->data);
3377 ret = proc_dostring(table, write, buffer, length, ppos);
3381 int oldval = user_zonelist_order;
3383 ret = __parse_numa_zonelist_order((char *)table->data);
3386 * bogus value. restore saved string
3388 strncpy((char *)table->data, saved_string,
3389 NUMA_ZONELIST_ORDER_LEN);
3390 user_zonelist_order = oldval;
3391 } else if (oldval != user_zonelist_order) {
3392 mutex_lock(&zonelists_mutex);
3393 build_all_zonelists(NULL, NULL);
3394 mutex_unlock(&zonelists_mutex);
3398 mutex_unlock(&zl_order_mutex);
3403 #define MAX_NODE_LOAD (nr_online_nodes)
3404 static int node_load[MAX_NUMNODES];
3407 * find_next_best_node - find the next node that should appear in a given node's fallback list
3408 * @node: node whose fallback list we're appending
3409 * @used_node_mask: nodemask_t of already used nodes
3411 * We use a number of factors to determine which is the next node that should
3412 * appear on a given node's fallback list. The node should not have appeared
3413 * already in @node's fallback list, and it should be the next closest node
3414 * according to the distance array (which contains arbitrary distance values
3415 * from each node to each node in the system), and should also prefer nodes
3416 * with no CPUs, since presumably they'll have very little allocation pressure
3417 * on them otherwise.
3418 * It returns -1 if no node is found.
3420 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3423 int min_val = INT_MAX;
3424 int best_node = NUMA_NO_NODE;
3425 const struct cpumask *tmp = cpumask_of_node(0);
3427 /* Use the local node if we haven't already */
3428 if (!node_isset(node, *used_node_mask)) {
3429 node_set(node, *used_node_mask);
3433 for_each_node_state(n, N_MEMORY) {
3435 /* Don't want a node to appear more than once */
3436 if (node_isset(n, *used_node_mask))
3439 /* Use the distance array to find the distance */
3440 val = node_distance(node, n);
3442 /* Penalize nodes under us ("prefer the next node") */
3445 /* Give preference to headless and unused nodes */
3446 tmp = cpumask_of_node(n);
3447 if (!cpumask_empty(tmp))
3448 val += PENALTY_FOR_NODE_WITH_CPUS;
3450 /* Slight preference for less loaded node */
3451 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3452 val += node_load[n];
3454 if (val < min_val) {
3461 node_set(best_node, *used_node_mask);
3468 * Build zonelists ordered by node and zones within node.
3469 * This results in maximum locality--normal zone overflows into local
3470 * DMA zone, if any--but risks exhausting DMA zone.
3472 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3475 struct zonelist *zonelist;
3477 zonelist = &pgdat->node_zonelists[0];
3478 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3480 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3481 zonelist->_zonerefs[j].zone = NULL;
3482 zonelist->_zonerefs[j].zone_idx = 0;
3486 * Build gfp_thisnode zonelists
3488 static void build_thisnode_zonelists(pg_data_t *pgdat)
3491 struct zonelist *zonelist;
3493 zonelist = &pgdat->node_zonelists[1];
3494 j = build_zonelists_node(pgdat, zonelist, 0);
3495 zonelist->_zonerefs[j].zone = NULL;
3496 zonelist->_zonerefs[j].zone_idx = 0;
3500 * Build zonelists ordered by zone and nodes within zones.
3501 * This results in conserving DMA zone[s] until all Normal memory is
3502 * exhausted, but results in overflowing to remote node while memory
3503 * may still exist in local DMA zone.
3505 static int node_order[MAX_NUMNODES];
3507 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3510 int zone_type; /* needs to be signed */
3512 struct zonelist *zonelist;
3514 zonelist = &pgdat->node_zonelists[0];
3516 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3517 for (j = 0; j < nr_nodes; j++) {
3518 node = node_order[j];
3519 z = &NODE_DATA(node)->node_zones[zone_type];
3520 if (populated_zone(z)) {
3522 &zonelist->_zonerefs[pos++]);
3523 check_highest_zone(zone_type);
3527 zonelist->_zonerefs[pos].zone = NULL;
3528 zonelist->_zonerefs[pos].zone_idx = 0;
3531 static int default_zonelist_order(void)
3534 unsigned long low_kmem_size, total_size;
3538 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3539 * If they are really small and used heavily, the system can fall
3540 * into OOM very easily.
3541 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3543 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3546 for_each_online_node(nid) {
3547 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3548 z = &NODE_DATA(nid)->node_zones[zone_type];
3549 if (populated_zone(z)) {
3550 if (zone_type < ZONE_NORMAL)
3551 low_kmem_size += z->managed_pages;
3552 total_size += z->managed_pages;
3553 } else if (zone_type == ZONE_NORMAL) {
3555 * If any node has only lowmem, then node order
3556 * is preferred to allow kernel allocations
3557 * locally; otherwise, they can easily infringe
3558 * on other nodes when there is an abundance of
3559 * lowmem available to allocate from.
3561 return ZONELIST_ORDER_NODE;
3565 if (!low_kmem_size || /* there are no DMA area. */
3566 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3567 return ZONELIST_ORDER_NODE;
3569 * look into each node's config.
3570 * If there is a node whose DMA/DMA32 memory is very big area on
3571 * local memory, NODE_ORDER may be suitable.
3573 average_size = total_size /
3574 (nodes_weight(node_states[N_MEMORY]) + 1);
3575 for_each_online_node(nid) {
3578 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3579 z = &NODE_DATA(nid)->node_zones[zone_type];
3580 if (populated_zone(z)) {
3581 if (zone_type < ZONE_NORMAL)
3582 low_kmem_size += z->present_pages;
3583 total_size += z->present_pages;
3586 if (low_kmem_size &&
3587 total_size > average_size && /* ignore small node */
3588 low_kmem_size > total_size * 70/100)
3589 return ZONELIST_ORDER_NODE;
3591 return ZONELIST_ORDER_ZONE;
3594 static void set_zonelist_order(void)
3596 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3597 current_zonelist_order = default_zonelist_order();
3599 current_zonelist_order = user_zonelist_order;
3602 static void build_zonelists(pg_data_t *pgdat)
3606 nodemask_t used_mask;
3607 int local_node, prev_node;
3608 struct zonelist *zonelist;
3609 int order = current_zonelist_order;
3611 /* initialize zonelists */
3612 for (i = 0; i < MAX_ZONELISTS; i++) {
3613 zonelist = pgdat->node_zonelists + i;
3614 zonelist->_zonerefs[0].zone = NULL;
3615 zonelist->_zonerefs[0].zone_idx = 0;
3618 /* NUMA-aware ordering of nodes */
3619 local_node = pgdat->node_id;
3620 load = nr_online_nodes;
3621 prev_node = local_node;
3622 nodes_clear(used_mask);
3624 memset(node_order, 0, sizeof(node_order));
3627 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3629 * We don't want to pressure a particular node.
3630 * So adding penalty to the first node in same
3631 * distance group to make it round-robin.
3633 if (node_distance(local_node, node) !=
3634 node_distance(local_node, prev_node))
3635 node_load[node] = load;
3639 if (order == ZONELIST_ORDER_NODE)
3640 build_zonelists_in_node_order(pgdat, node);
3642 node_order[j++] = node; /* remember order */
3645 if (order == ZONELIST_ORDER_ZONE) {
3646 /* calculate node order -- i.e., DMA last! */
3647 build_zonelists_in_zone_order(pgdat, j);
3650 build_thisnode_zonelists(pgdat);
3653 /* Construct the zonelist performance cache - see further mmzone.h */
3654 static void build_zonelist_cache(pg_data_t *pgdat)
3656 struct zonelist *zonelist;
3657 struct zonelist_cache *zlc;
3660 zonelist = &pgdat->node_zonelists[0];
3661 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3662 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3663 for (z = zonelist->_zonerefs; z->zone; z++)
3664 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3667 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3669 * Return node id of node used for "local" allocations.
3670 * I.e., first node id of first zone in arg node's generic zonelist.
3671 * Used for initializing percpu 'numa_mem', which is used primarily
3672 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3674 int local_memory_node(int node)
3678 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3679 gfp_zone(GFP_KERNEL),
3686 #else /* CONFIG_NUMA */
3688 static void set_zonelist_order(void)
3690 current_zonelist_order = ZONELIST_ORDER_ZONE;
3693 static void build_zonelists(pg_data_t *pgdat)
3695 int node, local_node;
3697 struct zonelist *zonelist;
3699 local_node = pgdat->node_id;
3701 zonelist = &pgdat->node_zonelists[0];
3702 j = build_zonelists_node(pgdat, zonelist, 0);
3705 * Now we build the zonelist so that it contains the zones
3706 * of all the other nodes.
3707 * We don't want to pressure a particular node, so when
3708 * building the zones for node N, we make sure that the
3709 * zones coming right after the local ones are those from
3710 * node N+1 (modulo N)
3712 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3713 if (!node_online(node))
3715 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3717 for (node = 0; node < local_node; node++) {
3718 if (!node_online(node))
3720 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3723 zonelist->_zonerefs[j].zone = NULL;
3724 zonelist->_zonerefs[j].zone_idx = 0;
3727 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3728 static void build_zonelist_cache(pg_data_t *pgdat)
3730 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3733 #endif /* CONFIG_NUMA */
3736 * Boot pageset table. One per cpu which is going to be used for all
3737 * zones and all nodes. The parameters will be set in such a way
3738 * that an item put on a list will immediately be handed over to
3739 * the buddy list. This is safe since pageset manipulation is done
3740 * with interrupts disabled.
3742 * The boot_pagesets must be kept even after bootup is complete for
3743 * unused processors and/or zones. They do play a role for bootstrapping
3744 * hotplugged processors.
3746 * zoneinfo_show() and maybe other functions do
3747 * not check if the processor is online before following the pageset pointer.
3748 * Other parts of the kernel may not check if the zone is available.
3750 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3751 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3752 static void setup_zone_pageset(struct zone *zone);
3755 * Global mutex to protect against size modification of zonelists
3756 * as well as to serialize pageset setup for the new populated zone.
3758 DEFINE_MUTEX(zonelists_mutex);
3760 /* return values int ....just for stop_machine() */
3761 static int __build_all_zonelists(void *data)
3765 pg_data_t *self = data;
3768 memset(node_load, 0, sizeof(node_load));
3771 if (self && !node_online(self->node_id)) {
3772 build_zonelists(self);
3773 build_zonelist_cache(self);
3776 for_each_online_node(nid) {
3777 pg_data_t *pgdat = NODE_DATA(nid);
3779 build_zonelists(pgdat);
3780 build_zonelist_cache(pgdat);
3784 * Initialize the boot_pagesets that are going to be used
3785 * for bootstrapping processors. The real pagesets for
3786 * each zone will be allocated later when the per cpu
3787 * allocator is available.
3789 * boot_pagesets are used also for bootstrapping offline
3790 * cpus if the system is already booted because the pagesets
3791 * are needed to initialize allocators on a specific cpu too.
3792 * F.e. the percpu allocator needs the page allocator which
3793 * needs the percpu allocator in order to allocate its pagesets
3794 * (a chicken-egg dilemma).
3796 for_each_possible_cpu(cpu) {
3797 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3799 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3801 * We now know the "local memory node" for each node--
3802 * i.e., the node of the first zone in the generic zonelist.
3803 * Set up numa_mem percpu variable for on-line cpus. During
3804 * boot, only the boot cpu should be on-line; we'll init the
3805 * secondary cpus' numa_mem as they come on-line. During
3806 * node/memory hotplug, we'll fixup all on-line cpus.
3808 if (cpu_online(cpu))
3809 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3817 * Called with zonelists_mutex held always
3818 * unless system_state == SYSTEM_BOOTING.
3820 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3822 set_zonelist_order();
3824 if (system_state == SYSTEM_BOOTING) {
3825 __build_all_zonelists(NULL);
3826 mminit_verify_zonelist();
3827 cpuset_init_current_mems_allowed();
3829 #ifdef CONFIG_MEMORY_HOTPLUG
3831 setup_zone_pageset(zone);
3833 /* we have to stop all cpus to guarantee there is no user
3835 stop_machine(__build_all_zonelists, pgdat, NULL);
3836 /* cpuset refresh routine should be here */
3838 vm_total_pages = nr_free_pagecache_pages();
3840 * Disable grouping by mobility if the number of pages in the
3841 * system is too low to allow the mechanism to work. It would be
3842 * more accurate, but expensive to check per-zone. This check is
3843 * made on memory-hotadd so a system can start with mobility
3844 * disabled and enable it later
3846 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3847 page_group_by_mobility_disabled = 1;
3849 page_group_by_mobility_disabled = 0;
3851 printk("Built %i zonelists in %s order, mobility grouping %s. "
3852 "Total pages: %ld\n",
3854 zonelist_order_name[current_zonelist_order],
3855 page_group_by_mobility_disabled ? "off" : "on",
3858 printk("Policy zone: %s\n", zone_names[policy_zone]);
3863 * Helper functions to size the waitqueue hash table.
3864 * Essentially these want to choose hash table sizes sufficiently
3865 * large so that collisions trying to wait on pages are rare.
3866 * But in fact, the number of active page waitqueues on typical
3867 * systems is ridiculously low, less than 200. So this is even
3868 * conservative, even though it seems large.
3870 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3871 * waitqueues, i.e. the size of the waitq table given the number of pages.
3873 #define PAGES_PER_WAITQUEUE 256
3875 #ifndef CONFIG_MEMORY_HOTPLUG
3876 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3878 unsigned long size = 1;
3880 pages /= PAGES_PER_WAITQUEUE;
3882 while (size < pages)
3886 * Once we have dozens or even hundreds of threads sleeping
3887 * on IO we've got bigger problems than wait queue collision.
3888 * Limit the size of the wait table to a reasonable size.
3890 size = min(size, 4096UL);
3892 return max(size, 4UL);
3896 * A zone's size might be changed by hot-add, so it is not possible to determine
3897 * a suitable size for its wait_table. So we use the maximum size now.
3899 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3901 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3902 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3903 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3905 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3906 * or more by the traditional way. (See above). It equals:
3908 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3909 * ia64(16K page size) : = ( 8G + 4M)byte.
3910 * powerpc (64K page size) : = (32G +16M)byte.
3912 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3919 * This is an integer logarithm so that shifts can be used later
3920 * to extract the more random high bits from the multiplicative
3921 * hash function before the remainder is taken.
3923 static inline unsigned long wait_table_bits(unsigned long size)
3929 * Check if a pageblock contains reserved pages
3931 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3935 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3936 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3943 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3944 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3945 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3946 * higher will lead to a bigger reserve which will get freed as contiguous
3947 * blocks as reclaim kicks in
3949 static void setup_zone_migrate_reserve(struct zone *zone)
3951 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3953 unsigned long block_migratetype;
3958 * Get the start pfn, end pfn and the number of blocks to reserve
3959 * We have to be careful to be aligned to pageblock_nr_pages to
3960 * make sure that we always check pfn_valid for the first page in
3963 start_pfn = zone->zone_start_pfn;
3964 end_pfn = zone_end_pfn(zone);
3965 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3966 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3970 * Reserve blocks are generally in place to help high-order atomic
3971 * allocations that are short-lived. A min_free_kbytes value that
3972 * would result in more than 2 reserve blocks for atomic allocations
3973 * is assumed to be in place to help anti-fragmentation for the
3974 * future allocation of hugepages at runtime.
3976 reserve = min(2, reserve);
3977 old_reserve = zone->nr_migrate_reserve_block;
3979 /* When memory hot-add, we almost always need to do nothing */
3980 if (reserve == old_reserve)
3982 zone->nr_migrate_reserve_block = reserve;
3984 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3985 if (!pfn_valid(pfn))
3987 page = pfn_to_page(pfn);
3989 /* Watch out for overlapping nodes */
3990 if (page_to_nid(page) != zone_to_nid(zone))
3993 block_migratetype = get_pageblock_migratetype(page);
3995 /* Only test what is necessary when the reserves are not met */
3998 * Blocks with reserved pages will never free, skip
4001 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4002 if (pageblock_is_reserved(pfn, block_end_pfn))
4005 /* If this block is reserved, account for it */
4006 if (block_migratetype == MIGRATE_RESERVE) {
4011 /* Suitable for reserving if this block is movable */
4012 if (block_migratetype == MIGRATE_MOVABLE) {
4013 set_pageblock_migratetype(page,
4015 move_freepages_block(zone, page,
4020 } else if (!old_reserve) {
4022 * At boot time we don't need to scan the whole zone
4023 * for turning off MIGRATE_RESERVE.
4029 * If the reserve is met and this is a previous reserved block,
4032 if (block_migratetype == MIGRATE_RESERVE) {
4033 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4034 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4040 * Initially all pages are reserved - free ones are freed
4041 * up by free_all_bootmem() once the early boot process is
4042 * done. Non-atomic initialization, single-pass.
4044 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4045 unsigned long start_pfn, enum memmap_context context)
4048 unsigned long end_pfn = start_pfn + size;
4052 if (highest_memmap_pfn < end_pfn - 1)
4053 highest_memmap_pfn = end_pfn - 1;
4055 z = &NODE_DATA(nid)->node_zones[zone];
4056 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4058 * There can be holes in boot-time mem_map[]s
4059 * handed to this function. They do not
4060 * exist on hotplugged memory.
4062 if (context == MEMMAP_EARLY) {
4063 if (!early_pfn_valid(pfn))
4065 if (!early_pfn_in_nid(pfn, nid))
4068 page = pfn_to_page(pfn);
4069 set_page_links(page, zone, nid, pfn);
4070 mminit_verify_page_links(page, zone, nid, pfn);
4071 init_page_count(page);
4072 page_mapcount_reset(page);
4073 page_cpupid_reset_last(page);
4074 SetPageReserved(page);
4076 * Mark the block movable so that blocks are reserved for
4077 * movable at startup. This will force kernel allocations
4078 * to reserve their blocks rather than leaking throughout
4079 * the address space during boot when many long-lived
4080 * kernel allocations are made. Later some blocks near
4081 * the start are marked MIGRATE_RESERVE by
4082 * setup_zone_migrate_reserve()
4084 * bitmap is created for zone's valid pfn range. but memmap
4085 * can be created for invalid pages (for alignment)
4086 * check here not to call set_pageblock_migratetype() against
4089 if ((z->zone_start_pfn <= pfn)
4090 && (pfn < zone_end_pfn(z))
4091 && !(pfn & (pageblock_nr_pages - 1)))
4092 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4094 INIT_LIST_HEAD(&page->lru);
4095 #ifdef WANT_PAGE_VIRTUAL
4096 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4097 if (!is_highmem_idx(zone))
4098 set_page_address(page, __va(pfn << PAGE_SHIFT));
4103 static void __meminit zone_init_free_lists(struct zone *zone)
4106 for_each_migratetype_order(order, t) {
4107 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4108 zone->free_area[order].nr_free = 0;
4112 #ifndef __HAVE_ARCH_MEMMAP_INIT
4113 #define memmap_init(size, nid, zone, start_pfn) \
4114 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4117 static int __meminit zone_batchsize(struct zone *zone)
4123 * The per-cpu-pages pools are set to around 1000th of the
4124 * size of the zone. But no more than 1/2 of a meg.
4126 * OK, so we don't know how big the cache is. So guess.
4128 batch = zone->managed_pages / 1024;
4129 if (batch * PAGE_SIZE > 512 * 1024)
4130 batch = (512 * 1024) / PAGE_SIZE;
4131 batch /= 4; /* We effectively *= 4 below */
4136 * Clamp the batch to a 2^n - 1 value. Having a power
4137 * of 2 value was found to be more likely to have
4138 * suboptimal cache aliasing properties in some cases.
4140 * For example if 2 tasks are alternately allocating
4141 * batches of pages, one task can end up with a lot
4142 * of pages of one half of the possible page colors
4143 * and the other with pages of the other colors.
4145 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4150 /* The deferral and batching of frees should be suppressed under NOMMU
4153 * The problem is that NOMMU needs to be able to allocate large chunks
4154 * of contiguous memory as there's no hardware page translation to
4155 * assemble apparent contiguous memory from discontiguous pages.
4157 * Queueing large contiguous runs of pages for batching, however,
4158 * causes the pages to actually be freed in smaller chunks. As there
4159 * can be a significant delay between the individual batches being
4160 * recycled, this leads to the once large chunks of space being
4161 * fragmented and becoming unavailable for high-order allocations.
4168 * pcp->high and pcp->batch values are related and dependent on one another:
4169 * ->batch must never be higher then ->high.
4170 * The following function updates them in a safe manner without read side
4173 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4174 * those fields changing asynchronously (acording the the above rule).
4176 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4177 * outside of boot time (or some other assurance that no concurrent updaters
4180 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4181 unsigned long batch)
4183 /* start with a fail safe value for batch */
4187 /* Update high, then batch, in order */
4194 /* a companion to pageset_set_high() */
4195 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4197 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4200 static void pageset_init(struct per_cpu_pageset *p)
4202 struct per_cpu_pages *pcp;
4205 memset(p, 0, sizeof(*p));
4209 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4210 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4213 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4216 pageset_set_batch(p, batch);
4220 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4221 * to the value high for the pageset p.
4223 static void pageset_set_high(struct per_cpu_pageset *p,
4226 unsigned long batch = max(1UL, high / 4);
4227 if ((high / 4) > (PAGE_SHIFT * 8))
4228 batch = PAGE_SHIFT * 8;
4230 pageset_update(&p->pcp, high, batch);
4233 static void __meminit pageset_set_high_and_batch(struct zone *zone,
4234 struct per_cpu_pageset *pcp)
4236 if (percpu_pagelist_fraction)
4237 pageset_set_high(pcp,
4238 (zone->managed_pages /
4239 percpu_pagelist_fraction));
4241 pageset_set_batch(pcp, zone_batchsize(zone));
4244 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4246 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4249 pageset_set_high_and_batch(zone, pcp);
4252 static void __meminit setup_zone_pageset(struct zone *zone)
4255 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4256 for_each_possible_cpu(cpu)
4257 zone_pageset_init(zone, cpu);
4261 * Allocate per cpu pagesets and initialize them.
4262 * Before this call only boot pagesets were available.
4264 void __init setup_per_cpu_pageset(void)
4268 for_each_populated_zone(zone)
4269 setup_zone_pageset(zone);
4272 static noinline __init_refok
4273 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4279 * The per-page waitqueue mechanism uses hashed waitqueues
4282 zone->wait_table_hash_nr_entries =
4283 wait_table_hash_nr_entries(zone_size_pages);
4284 zone->wait_table_bits =
4285 wait_table_bits(zone->wait_table_hash_nr_entries);
4286 alloc_size = zone->wait_table_hash_nr_entries
4287 * sizeof(wait_queue_head_t);
4289 if (!slab_is_available()) {
4290 zone->wait_table = (wait_queue_head_t *)
4291 memblock_virt_alloc_node_nopanic(
4292 alloc_size, zone->zone_pgdat->node_id);
4295 * This case means that a zone whose size was 0 gets new memory
4296 * via memory hot-add.
4297 * But it may be the case that a new node was hot-added. In
4298 * this case vmalloc() will not be able to use this new node's
4299 * memory - this wait_table must be initialized to use this new
4300 * node itself as well.
4301 * To use this new node's memory, further consideration will be
4304 zone->wait_table = vmalloc(alloc_size);
4306 if (!zone->wait_table)
4309 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4310 init_waitqueue_head(zone->wait_table + i);
4315 static __meminit void zone_pcp_init(struct zone *zone)
4318 * per cpu subsystem is not up at this point. The following code
4319 * relies on the ability of the linker to provide the
4320 * offset of a (static) per cpu variable into the per cpu area.
4322 zone->pageset = &boot_pageset;
4324 if (populated_zone(zone))
4325 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4326 zone->name, zone->present_pages,
4327 zone_batchsize(zone));
4330 int __meminit init_currently_empty_zone(struct zone *zone,
4331 unsigned long zone_start_pfn,
4333 enum memmap_context context)
4335 struct pglist_data *pgdat = zone->zone_pgdat;
4337 ret = zone_wait_table_init(zone, size);
4340 pgdat->nr_zones = zone_idx(zone) + 1;
4342 zone->zone_start_pfn = zone_start_pfn;
4344 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4345 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4347 (unsigned long)zone_idx(zone),
4348 zone_start_pfn, (zone_start_pfn + size));
4350 zone_init_free_lists(zone);
4355 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4356 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4358 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4359 * Architectures may implement their own version but if add_active_range()
4360 * was used and there are no special requirements, this is a convenient
4363 int __meminit __early_pfn_to_nid(unsigned long pfn)
4365 unsigned long start_pfn, end_pfn;
4368 * NOTE: The following SMP-unsafe globals are only used early in boot
4369 * when the kernel is running single-threaded.
4371 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4372 static int __meminitdata last_nid;
4374 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4377 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4379 last_start_pfn = start_pfn;
4380 last_end_pfn = end_pfn;
4386 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4388 int __meminit early_pfn_to_nid(unsigned long pfn)
4392 nid = __early_pfn_to_nid(pfn);
4395 /* just returns 0 */
4399 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4400 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4404 nid = __early_pfn_to_nid(pfn);
4405 if (nid >= 0 && nid != node)
4412 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4413 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4414 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4416 * If an architecture guarantees that all ranges registered with
4417 * add_active_ranges() contain no holes and may be freed, this
4418 * this function may be used instead of calling memblock_free_early_nid()
4421 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4423 unsigned long start_pfn, end_pfn;
4426 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4427 start_pfn = min(start_pfn, max_low_pfn);
4428 end_pfn = min(end_pfn, max_low_pfn);
4430 if (start_pfn < end_pfn)
4431 memblock_free_early_nid(PFN_PHYS(start_pfn),
4432 (end_pfn - start_pfn) << PAGE_SHIFT,
4438 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4439 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4441 * If an architecture guarantees that all ranges registered with
4442 * add_active_ranges() contain no holes and may be freed, this
4443 * function may be used instead of calling memory_present() manually.
4445 void __init sparse_memory_present_with_active_regions(int nid)
4447 unsigned long start_pfn, end_pfn;
4450 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4451 memory_present(this_nid, start_pfn, end_pfn);
4455 * get_pfn_range_for_nid - Return the start and end page frames for a node
4456 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4457 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4458 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4460 * It returns the start and end page frame of a node based on information
4461 * provided by an arch calling add_active_range(). If called for a node
4462 * with no available memory, a warning is printed and the start and end
4465 void __meminit get_pfn_range_for_nid(unsigned int nid,
4466 unsigned long *start_pfn, unsigned long *end_pfn)
4468 unsigned long this_start_pfn, this_end_pfn;
4474 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4475 *start_pfn = min(*start_pfn, this_start_pfn);
4476 *end_pfn = max(*end_pfn, this_end_pfn);
4479 if (*start_pfn == -1UL)
4484 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4485 * assumption is made that zones within a node are ordered in monotonic
4486 * increasing memory addresses so that the "highest" populated zone is used
4488 static void __init find_usable_zone_for_movable(void)
4491 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4492 if (zone_index == ZONE_MOVABLE)
4495 if (arch_zone_highest_possible_pfn[zone_index] >
4496 arch_zone_lowest_possible_pfn[zone_index])
4500 VM_BUG_ON(zone_index == -1);
4501 movable_zone = zone_index;
4505 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4506 * because it is sized independent of architecture. Unlike the other zones,
4507 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4508 * in each node depending on the size of each node and how evenly kernelcore
4509 * is distributed. This helper function adjusts the zone ranges
4510 * provided by the architecture for a given node by using the end of the
4511 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4512 * zones within a node are in order of monotonic increases memory addresses
4514 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4515 unsigned long zone_type,
4516 unsigned long node_start_pfn,
4517 unsigned long node_end_pfn,
4518 unsigned long *zone_start_pfn,
4519 unsigned long *zone_end_pfn)
4521 /* Only adjust if ZONE_MOVABLE is on this node */
4522 if (zone_movable_pfn[nid]) {
4523 /* Size ZONE_MOVABLE */
4524 if (zone_type == ZONE_MOVABLE) {
4525 *zone_start_pfn = zone_movable_pfn[nid];
4526 *zone_end_pfn = min(node_end_pfn,
4527 arch_zone_highest_possible_pfn[movable_zone]);
4529 /* Adjust for ZONE_MOVABLE starting within this range */
4530 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4531 *zone_end_pfn > zone_movable_pfn[nid]) {
4532 *zone_end_pfn = zone_movable_pfn[nid];
4534 /* Check if this whole range is within ZONE_MOVABLE */
4535 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4536 *zone_start_pfn = *zone_end_pfn;
4541 * Return the number of pages a zone spans in a node, including holes
4542 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4544 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4545 unsigned long zone_type,
4546 unsigned long node_start_pfn,
4547 unsigned long node_end_pfn,
4548 unsigned long *ignored)
4550 unsigned long zone_start_pfn, zone_end_pfn;
4552 /* Get the start and end of the zone */
4553 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4554 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4555 adjust_zone_range_for_zone_movable(nid, zone_type,
4556 node_start_pfn, node_end_pfn,
4557 &zone_start_pfn, &zone_end_pfn);
4559 /* Check that this node has pages within the zone's required range */
4560 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4563 /* Move the zone boundaries inside the node if necessary */
4564 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4565 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4567 /* Return the spanned pages */
4568 return zone_end_pfn - zone_start_pfn;
4572 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4573 * then all holes in the requested range will be accounted for.
4575 unsigned long __meminit __absent_pages_in_range(int nid,
4576 unsigned long range_start_pfn,
4577 unsigned long range_end_pfn)
4579 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4580 unsigned long start_pfn, end_pfn;
4583 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4584 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4585 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4586 nr_absent -= end_pfn - start_pfn;
4592 * absent_pages_in_range - Return number of page frames in holes within a range
4593 * @start_pfn: The start PFN to start searching for holes
4594 * @end_pfn: The end PFN to stop searching for holes
4596 * It returns the number of pages frames in memory holes within a range.
4598 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4599 unsigned long end_pfn)
4601 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4604 /* Return the number of page frames in holes in a zone on a node */
4605 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4606 unsigned long zone_type,
4607 unsigned long node_start_pfn,
4608 unsigned long node_end_pfn,
4609 unsigned long *ignored)
4611 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4612 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4613 unsigned long zone_start_pfn, zone_end_pfn;
4615 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4616 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4618 adjust_zone_range_for_zone_movable(nid, zone_type,
4619 node_start_pfn, node_end_pfn,
4620 &zone_start_pfn, &zone_end_pfn);
4621 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4624 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4625 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4626 unsigned long zone_type,
4627 unsigned long node_start_pfn,
4628 unsigned long node_end_pfn,
4629 unsigned long *zones_size)
4631 return zones_size[zone_type];
4634 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4635 unsigned long zone_type,
4636 unsigned long node_start_pfn,
4637 unsigned long node_end_pfn,
4638 unsigned long *zholes_size)
4643 return zholes_size[zone_type];
4646 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4648 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4649 unsigned long node_start_pfn,
4650 unsigned long node_end_pfn,
4651 unsigned long *zones_size,
4652 unsigned long *zholes_size)
4654 unsigned long realtotalpages, totalpages = 0;
4657 for (i = 0; i < MAX_NR_ZONES; i++)
4658 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4662 pgdat->node_spanned_pages = totalpages;
4664 realtotalpages = totalpages;
4665 for (i = 0; i < MAX_NR_ZONES; i++)
4667 zone_absent_pages_in_node(pgdat->node_id, i,
4668 node_start_pfn, node_end_pfn,
4670 pgdat->node_present_pages = realtotalpages;
4671 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4675 #ifndef CONFIG_SPARSEMEM
4677 * Calculate the size of the zone->blockflags rounded to an unsigned long
4678 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4679 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4680 * round what is now in bits to nearest long in bits, then return it in
4683 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4685 unsigned long usemapsize;
4687 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4688 usemapsize = roundup(zonesize, pageblock_nr_pages);
4689 usemapsize = usemapsize >> pageblock_order;
4690 usemapsize *= NR_PAGEBLOCK_BITS;
4691 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4693 return usemapsize / 8;
4696 static void __init setup_usemap(struct pglist_data *pgdat,
4698 unsigned long zone_start_pfn,
4699 unsigned long zonesize)
4701 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4702 zone->pageblock_flags = NULL;
4704 zone->pageblock_flags =
4705 memblock_virt_alloc_node_nopanic(usemapsize,
4709 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4710 unsigned long zone_start_pfn, unsigned long zonesize) {}
4711 #endif /* CONFIG_SPARSEMEM */
4713 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4715 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4716 void __paginginit set_pageblock_order(void)
4720 /* Check that pageblock_nr_pages has not already been setup */
4721 if (pageblock_order)
4724 if (HPAGE_SHIFT > PAGE_SHIFT)
4725 order = HUGETLB_PAGE_ORDER;
4727 order = MAX_ORDER - 1;
4730 * Assume the largest contiguous order of interest is a huge page.
4731 * This value may be variable depending on boot parameters on IA64 and
4734 pageblock_order = order;
4736 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4739 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4740 * is unused as pageblock_order is set at compile-time. See
4741 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4744 void __paginginit set_pageblock_order(void)
4748 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4750 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4751 unsigned long present_pages)
4753 unsigned long pages = spanned_pages;
4756 * Provide a more accurate estimation if there are holes within
4757 * the zone and SPARSEMEM is in use. If there are holes within the
4758 * zone, each populated memory region may cost us one or two extra
4759 * memmap pages due to alignment because memmap pages for each
4760 * populated regions may not naturally algined on page boundary.
4761 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4763 if (spanned_pages > present_pages + (present_pages >> 4) &&
4764 IS_ENABLED(CONFIG_SPARSEMEM))
4765 pages = present_pages;
4767 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4771 * Set up the zone data structures:
4772 * - mark all pages reserved
4773 * - mark all memory queues empty
4774 * - clear the memory bitmaps
4776 * NOTE: pgdat should get zeroed by caller.
4778 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4779 unsigned long node_start_pfn, unsigned long node_end_pfn,
4780 unsigned long *zones_size, unsigned long *zholes_size)
4783 int nid = pgdat->node_id;
4784 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4787 pgdat_resize_init(pgdat);
4788 #ifdef CONFIG_NUMA_BALANCING
4789 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4790 pgdat->numabalancing_migrate_nr_pages = 0;
4791 pgdat->numabalancing_migrate_next_window = jiffies;
4793 init_waitqueue_head(&pgdat->kswapd_wait);
4794 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4795 pgdat_page_cgroup_init(pgdat);
4797 for (j = 0; j < MAX_NR_ZONES; j++) {
4798 struct zone *zone = pgdat->node_zones + j;
4799 unsigned long size, realsize, freesize, memmap_pages;
4801 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4802 node_end_pfn, zones_size);
4803 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4809 * Adjust freesize so that it accounts for how much memory
4810 * is used by this zone for memmap. This affects the watermark
4811 * and per-cpu initialisations
4813 memmap_pages = calc_memmap_size(size, realsize);
4814 if (freesize >= memmap_pages) {
4815 freesize -= memmap_pages;
4818 " %s zone: %lu pages used for memmap\n",
4819 zone_names[j], memmap_pages);
4822 " %s zone: %lu pages exceeds freesize %lu\n",
4823 zone_names[j], memmap_pages, freesize);
4825 /* Account for reserved pages */
4826 if (j == 0 && freesize > dma_reserve) {
4827 freesize -= dma_reserve;
4828 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4829 zone_names[0], dma_reserve);
4832 if (!is_highmem_idx(j))
4833 nr_kernel_pages += freesize;
4834 /* Charge for highmem memmap if there are enough kernel pages */
4835 else if (nr_kernel_pages > memmap_pages * 2)
4836 nr_kernel_pages -= memmap_pages;
4837 nr_all_pages += freesize;
4839 zone->spanned_pages = size;
4840 zone->present_pages = realsize;
4842 * Set an approximate value for lowmem here, it will be adjusted
4843 * when the bootmem allocator frees pages into the buddy system.
4844 * And all highmem pages will be managed by the buddy system.
4846 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4849 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4851 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4853 zone->name = zone_names[j];
4854 spin_lock_init(&zone->lock);
4855 spin_lock_init(&zone->lru_lock);
4856 zone_seqlock_init(zone);
4857 zone->zone_pgdat = pgdat;
4858 zone_pcp_init(zone);
4860 /* For bootup, initialized properly in watermark setup */
4861 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4863 lruvec_init(&zone->lruvec);
4867 set_pageblock_order();
4868 setup_usemap(pgdat, zone, zone_start_pfn, size);
4869 ret = init_currently_empty_zone(zone, zone_start_pfn,
4870 size, MEMMAP_EARLY);
4872 memmap_init(size, nid, j, zone_start_pfn);
4873 zone_start_pfn += size;
4877 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4879 /* Skip empty nodes */
4880 if (!pgdat->node_spanned_pages)
4883 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4884 /* ia64 gets its own node_mem_map, before this, without bootmem */
4885 if (!pgdat->node_mem_map) {
4886 unsigned long size, start, end;
4890 * The zone's endpoints aren't required to be MAX_ORDER
4891 * aligned but the node_mem_map endpoints must be in order
4892 * for the buddy allocator to function correctly.
4894 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4895 end = pgdat_end_pfn(pgdat);
4896 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4897 size = (end - start) * sizeof(struct page);
4898 map = alloc_remap(pgdat->node_id, size);
4900 map = memblock_virt_alloc_node_nopanic(size,
4902 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4904 #ifndef CONFIG_NEED_MULTIPLE_NODES
4906 * With no DISCONTIG, the global mem_map is just set as node 0's
4908 if (pgdat == NODE_DATA(0)) {
4909 mem_map = NODE_DATA(0)->node_mem_map;
4910 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4911 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4912 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4913 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4916 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4919 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4920 unsigned long node_start_pfn, unsigned long *zholes_size)
4922 pg_data_t *pgdat = NODE_DATA(nid);
4923 unsigned long start_pfn = 0;
4924 unsigned long end_pfn = 0;
4926 /* pg_data_t should be reset to zero when it's allocated */
4927 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4929 pgdat->node_id = nid;
4930 pgdat->node_start_pfn = node_start_pfn;
4931 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4932 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4934 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4935 zones_size, zholes_size);
4937 alloc_node_mem_map(pgdat);
4938 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4939 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4940 nid, (unsigned long)pgdat,
4941 (unsigned long)pgdat->node_mem_map);
4944 free_area_init_core(pgdat, start_pfn, end_pfn,
4945 zones_size, zholes_size);
4948 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4950 #if MAX_NUMNODES > 1
4952 * Figure out the number of possible node ids.
4954 void __init setup_nr_node_ids(void)
4957 unsigned int highest = 0;
4959 for_each_node_mask(node, node_possible_map)
4961 nr_node_ids = highest + 1;
4966 * node_map_pfn_alignment - determine the maximum internode alignment
4968 * This function should be called after node map is populated and sorted.
4969 * It calculates the maximum power of two alignment which can distinguish
4972 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4973 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4974 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4975 * shifted, 1GiB is enough and this function will indicate so.
4977 * This is used to test whether pfn -> nid mapping of the chosen memory
4978 * model has fine enough granularity to avoid incorrect mapping for the
4979 * populated node map.
4981 * Returns the determined alignment in pfn's. 0 if there is no alignment
4982 * requirement (single node).
4984 unsigned long __init node_map_pfn_alignment(void)
4986 unsigned long accl_mask = 0, last_end = 0;
4987 unsigned long start, end, mask;
4991 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4992 if (!start || last_nid < 0 || last_nid == nid) {
4999 * Start with a mask granular enough to pin-point to the
5000 * start pfn and tick off bits one-by-one until it becomes
5001 * too coarse to separate the current node from the last.
5003 mask = ~((1 << __ffs(start)) - 1);
5004 while (mask && last_end <= (start & (mask << 1)))
5007 /* accumulate all internode masks */
5011 /* convert mask to number of pages */
5012 return ~accl_mask + 1;
5015 /* Find the lowest pfn for a node */
5016 static unsigned long __init find_min_pfn_for_node(int nid)
5018 unsigned long min_pfn = ULONG_MAX;
5019 unsigned long start_pfn;
5022 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5023 min_pfn = min(min_pfn, start_pfn);
5025 if (min_pfn == ULONG_MAX) {
5027 "Could not find start_pfn for node %d\n", nid);
5035 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5037 * It returns the minimum PFN based on information provided via
5038 * add_active_range().
5040 unsigned long __init find_min_pfn_with_active_regions(void)
5042 return find_min_pfn_for_node(MAX_NUMNODES);
5046 * early_calculate_totalpages()
5047 * Sum pages in active regions for movable zone.
5048 * Populate N_MEMORY for calculating usable_nodes.
5050 static unsigned long __init early_calculate_totalpages(void)
5052 unsigned long totalpages = 0;
5053 unsigned long start_pfn, end_pfn;
5056 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5057 unsigned long pages = end_pfn - start_pfn;
5059 totalpages += pages;
5061 node_set_state(nid, N_MEMORY);
5067 * Find the PFN the Movable zone begins in each node. Kernel memory
5068 * is spread evenly between nodes as long as the nodes have enough
5069 * memory. When they don't, some nodes will have more kernelcore than
5072 static void __init find_zone_movable_pfns_for_nodes(void)
5075 unsigned long usable_startpfn;
5076 unsigned long kernelcore_node, kernelcore_remaining;
5077 /* save the state before borrow the nodemask */
5078 nodemask_t saved_node_state = node_states[N_MEMORY];
5079 unsigned long totalpages = early_calculate_totalpages();
5080 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5081 struct memblock_region *r;
5083 /* Need to find movable_zone earlier when movable_node is specified. */
5084 find_usable_zone_for_movable();
5087 * If movable_node is specified, ignore kernelcore and movablecore
5090 if (movable_node_is_enabled()) {
5091 for_each_memblock(memory, r) {
5092 if (!memblock_is_hotpluggable(r))
5097 usable_startpfn = PFN_DOWN(r->base);
5098 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5099 min(usable_startpfn, zone_movable_pfn[nid]) :
5107 * If movablecore=nn[KMG] was specified, calculate what size of
5108 * kernelcore that corresponds so that memory usable for
5109 * any allocation type is evenly spread. If both kernelcore
5110 * and movablecore are specified, then the value of kernelcore
5111 * will be used for required_kernelcore if it's greater than
5112 * what movablecore would have allowed.
5114 if (required_movablecore) {
5115 unsigned long corepages;
5118 * Round-up so that ZONE_MOVABLE is at least as large as what
5119 * was requested by the user
5121 required_movablecore =
5122 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5123 corepages = totalpages - required_movablecore;
5125 required_kernelcore = max(required_kernelcore, corepages);
5128 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5129 if (!required_kernelcore)
5132 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5133 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5136 /* Spread kernelcore memory as evenly as possible throughout nodes */
5137 kernelcore_node = required_kernelcore / usable_nodes;
5138 for_each_node_state(nid, N_MEMORY) {
5139 unsigned long start_pfn, end_pfn;
5142 * Recalculate kernelcore_node if the division per node
5143 * now exceeds what is necessary to satisfy the requested
5144 * amount of memory for the kernel
5146 if (required_kernelcore < kernelcore_node)
5147 kernelcore_node = required_kernelcore / usable_nodes;
5150 * As the map is walked, we track how much memory is usable
5151 * by the kernel using kernelcore_remaining. When it is
5152 * 0, the rest of the node is usable by ZONE_MOVABLE
5154 kernelcore_remaining = kernelcore_node;
5156 /* Go through each range of PFNs within this node */
5157 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5158 unsigned long size_pages;
5160 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5161 if (start_pfn >= end_pfn)
5164 /* Account for what is only usable for kernelcore */
5165 if (start_pfn < usable_startpfn) {
5166 unsigned long kernel_pages;
5167 kernel_pages = min(end_pfn, usable_startpfn)
5170 kernelcore_remaining -= min(kernel_pages,
5171 kernelcore_remaining);
5172 required_kernelcore -= min(kernel_pages,
5173 required_kernelcore);
5175 /* Continue if range is now fully accounted */
5176 if (end_pfn <= usable_startpfn) {
5179 * Push zone_movable_pfn to the end so
5180 * that if we have to rebalance
5181 * kernelcore across nodes, we will
5182 * not double account here
5184 zone_movable_pfn[nid] = end_pfn;
5187 start_pfn = usable_startpfn;
5191 * The usable PFN range for ZONE_MOVABLE is from
5192 * start_pfn->end_pfn. Calculate size_pages as the
5193 * number of pages used as kernelcore
5195 size_pages = end_pfn - start_pfn;
5196 if (size_pages > kernelcore_remaining)
5197 size_pages = kernelcore_remaining;
5198 zone_movable_pfn[nid] = start_pfn + size_pages;
5201 * Some kernelcore has been met, update counts and
5202 * break if the kernelcore for this node has been
5205 required_kernelcore -= min(required_kernelcore,
5207 kernelcore_remaining -= size_pages;
5208 if (!kernelcore_remaining)
5214 * If there is still required_kernelcore, we do another pass with one
5215 * less node in the count. This will push zone_movable_pfn[nid] further
5216 * along on the nodes that still have memory until kernelcore is
5220 if (usable_nodes && required_kernelcore > usable_nodes)
5224 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5225 for (nid = 0; nid < MAX_NUMNODES; nid++)
5226 zone_movable_pfn[nid] =
5227 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5230 /* restore the node_state */
5231 node_states[N_MEMORY] = saved_node_state;
5234 /* Any regular or high memory on that node ? */
5235 static void check_for_memory(pg_data_t *pgdat, int nid)
5237 enum zone_type zone_type;
5239 if (N_MEMORY == N_NORMAL_MEMORY)
5242 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5243 struct zone *zone = &pgdat->node_zones[zone_type];
5244 if (populated_zone(zone)) {
5245 node_set_state(nid, N_HIGH_MEMORY);
5246 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5247 zone_type <= ZONE_NORMAL)
5248 node_set_state(nid, N_NORMAL_MEMORY);
5255 * free_area_init_nodes - Initialise all pg_data_t and zone data
5256 * @max_zone_pfn: an array of max PFNs for each zone
5258 * This will call free_area_init_node() for each active node in the system.
5259 * Using the page ranges provided by add_active_range(), the size of each
5260 * zone in each node and their holes is calculated. If the maximum PFN
5261 * between two adjacent zones match, it is assumed that the zone is empty.
5262 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5263 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5264 * starts where the previous one ended. For example, ZONE_DMA32 starts
5265 * at arch_max_dma_pfn.
5267 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5269 unsigned long start_pfn, end_pfn;
5272 /* Record where the zone boundaries are */
5273 memset(arch_zone_lowest_possible_pfn, 0,
5274 sizeof(arch_zone_lowest_possible_pfn));
5275 memset(arch_zone_highest_possible_pfn, 0,
5276 sizeof(arch_zone_highest_possible_pfn));
5277 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5278 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5279 for (i = 1; i < MAX_NR_ZONES; i++) {
5280 if (i == ZONE_MOVABLE)
5282 arch_zone_lowest_possible_pfn[i] =
5283 arch_zone_highest_possible_pfn[i-1];
5284 arch_zone_highest_possible_pfn[i] =
5285 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5287 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5288 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5290 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5291 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5292 find_zone_movable_pfns_for_nodes();
5294 /* Print out the zone ranges */
5295 printk("Zone ranges:\n");
5296 for (i = 0; i < MAX_NR_ZONES; i++) {
5297 if (i == ZONE_MOVABLE)
5299 printk(KERN_CONT " %-8s ", zone_names[i]);
5300 if (arch_zone_lowest_possible_pfn[i] ==
5301 arch_zone_highest_possible_pfn[i])
5302 printk(KERN_CONT "empty\n");
5304 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5305 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5306 (arch_zone_highest_possible_pfn[i]
5307 << PAGE_SHIFT) - 1);
5310 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5311 printk("Movable zone start for each node\n");
5312 for (i = 0; i < MAX_NUMNODES; i++) {
5313 if (zone_movable_pfn[i])
5314 printk(" Node %d: %#010lx\n", i,
5315 zone_movable_pfn[i] << PAGE_SHIFT);
5318 /* Print out the early node map */
5319 printk("Early memory node ranges\n");
5320 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5321 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5322 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5324 /* Initialise every node */
5325 mminit_verify_pageflags_layout();
5326 setup_nr_node_ids();
5327 for_each_online_node(nid) {
5328 pg_data_t *pgdat = NODE_DATA(nid);
5329 free_area_init_node(nid, NULL,
5330 find_min_pfn_for_node(nid), NULL);
5332 /* Any memory on that node */
5333 if (pgdat->node_present_pages)
5334 node_set_state(nid, N_MEMORY);
5335 check_for_memory(pgdat, nid);
5339 static int __init cmdline_parse_core(char *p, unsigned long *core)
5341 unsigned long long coremem;
5345 coremem = memparse(p, &p);
5346 *core = coremem >> PAGE_SHIFT;
5348 /* Paranoid check that UL is enough for the coremem value */
5349 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5355 * kernelcore=size sets the amount of memory for use for allocations that
5356 * cannot be reclaimed or migrated.
5358 static int __init cmdline_parse_kernelcore(char *p)
5360 return cmdline_parse_core(p, &required_kernelcore);
5364 * movablecore=size sets the amount of memory for use for allocations that
5365 * can be reclaimed or migrated.
5367 static int __init cmdline_parse_movablecore(char *p)
5369 return cmdline_parse_core(p, &required_movablecore);
5372 early_param("kernelcore", cmdline_parse_kernelcore);
5373 early_param("movablecore", cmdline_parse_movablecore);
5375 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5377 void adjust_managed_page_count(struct page *page, long count)
5379 spin_lock(&managed_page_count_lock);
5380 page_zone(page)->managed_pages += count;
5381 totalram_pages += count;
5382 #ifdef CONFIG_HIGHMEM
5383 if (PageHighMem(page))
5384 totalhigh_pages += count;
5386 spin_unlock(&managed_page_count_lock);
5388 EXPORT_SYMBOL(adjust_managed_page_count);
5390 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5393 unsigned long pages = 0;
5395 start = (void *)PAGE_ALIGN((unsigned long)start);
5396 end = (void *)((unsigned long)end & PAGE_MASK);
5397 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5398 if ((unsigned int)poison <= 0xFF)
5399 memset(pos, poison, PAGE_SIZE);
5400 free_reserved_page(virt_to_page(pos));
5404 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5405 s, pages << (PAGE_SHIFT - 10), start, end);
5409 EXPORT_SYMBOL(free_reserved_area);
5411 #ifdef CONFIG_HIGHMEM
5412 void free_highmem_page(struct page *page)
5414 __free_reserved_page(page);
5416 page_zone(page)->managed_pages++;
5422 void __init mem_init_print_info(const char *str)
5424 unsigned long physpages, codesize, datasize, rosize, bss_size;
5425 unsigned long init_code_size, init_data_size;
5427 physpages = get_num_physpages();
5428 codesize = _etext - _stext;
5429 datasize = _edata - _sdata;
5430 rosize = __end_rodata - __start_rodata;
5431 bss_size = __bss_stop - __bss_start;
5432 init_data_size = __init_end - __init_begin;
5433 init_code_size = _einittext - _sinittext;
5436 * Detect special cases and adjust section sizes accordingly:
5437 * 1) .init.* may be embedded into .data sections
5438 * 2) .init.text.* may be out of [__init_begin, __init_end],
5439 * please refer to arch/tile/kernel/vmlinux.lds.S.
5440 * 3) .rodata.* may be embedded into .text or .data sections.
5442 #define adj_init_size(start, end, size, pos, adj) \
5444 if (start <= pos && pos < end && size > adj) \
5448 adj_init_size(__init_begin, __init_end, init_data_size,
5449 _sinittext, init_code_size);
5450 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5451 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5452 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5453 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5455 #undef adj_init_size
5457 printk("Memory: %luK/%luK available "
5458 "(%luK kernel code, %luK rwdata, %luK rodata, "
5459 "%luK init, %luK bss, %luK reserved"
5460 #ifdef CONFIG_HIGHMEM
5464 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5465 codesize >> 10, datasize >> 10, rosize >> 10,
5466 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5467 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5468 #ifdef CONFIG_HIGHMEM
5469 totalhigh_pages << (PAGE_SHIFT-10),
5471 str ? ", " : "", str ? str : "");
5475 * set_dma_reserve - set the specified number of pages reserved in the first zone
5476 * @new_dma_reserve: The number of pages to mark reserved
5478 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5479 * In the DMA zone, a significant percentage may be consumed by kernel image
5480 * and other unfreeable allocations which can skew the watermarks badly. This
5481 * function may optionally be used to account for unfreeable pages in the
5482 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5483 * smaller per-cpu batchsize.
5485 void __init set_dma_reserve(unsigned long new_dma_reserve)
5487 dma_reserve = new_dma_reserve;
5490 void __init free_area_init(unsigned long *zones_size)
5492 free_area_init_node(0, zones_size,
5493 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5496 static int page_alloc_cpu_notify(struct notifier_block *self,
5497 unsigned long action, void *hcpu)
5499 int cpu = (unsigned long)hcpu;
5501 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5502 lru_add_drain_cpu(cpu);
5506 * Spill the event counters of the dead processor
5507 * into the current processors event counters.
5508 * This artificially elevates the count of the current
5511 vm_events_fold_cpu(cpu);
5514 * Zero the differential counters of the dead processor
5515 * so that the vm statistics are consistent.
5517 * This is only okay since the processor is dead and cannot
5518 * race with what we are doing.
5520 cpu_vm_stats_fold(cpu);
5525 void __init page_alloc_init(void)
5527 hotcpu_notifier(page_alloc_cpu_notify, 0);
5531 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5532 * or min_free_kbytes changes.
5534 static void calculate_totalreserve_pages(void)
5536 struct pglist_data *pgdat;
5537 unsigned long reserve_pages = 0;
5538 enum zone_type i, j;
5540 for_each_online_pgdat(pgdat) {
5541 for (i = 0; i < MAX_NR_ZONES; i++) {
5542 struct zone *zone = pgdat->node_zones + i;
5543 unsigned long max = 0;
5545 /* Find valid and maximum lowmem_reserve in the zone */
5546 for (j = i; j < MAX_NR_ZONES; j++) {
5547 if (zone->lowmem_reserve[j] > max)
5548 max = zone->lowmem_reserve[j];
5551 /* we treat the high watermark as reserved pages. */
5552 max += high_wmark_pages(zone);
5554 if (max > zone->managed_pages)
5555 max = zone->managed_pages;
5556 reserve_pages += max;
5558 * Lowmem reserves are not available to
5559 * GFP_HIGHUSER page cache allocations and
5560 * kswapd tries to balance zones to their high
5561 * watermark. As a result, neither should be
5562 * regarded as dirtyable memory, to prevent a
5563 * situation where reclaim has to clean pages
5564 * in order to balance the zones.
5566 zone->dirty_balance_reserve = max;
5569 dirty_balance_reserve = reserve_pages;
5570 totalreserve_pages = reserve_pages;
5574 * setup_per_zone_lowmem_reserve - called whenever
5575 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5576 * has a correct pages reserved value, so an adequate number of
5577 * pages are left in the zone after a successful __alloc_pages().
5579 static void setup_per_zone_lowmem_reserve(void)
5581 struct pglist_data *pgdat;
5582 enum zone_type j, idx;
5584 for_each_online_pgdat(pgdat) {
5585 for (j = 0; j < MAX_NR_ZONES; j++) {
5586 struct zone *zone = pgdat->node_zones + j;
5587 unsigned long managed_pages = zone->managed_pages;
5589 zone->lowmem_reserve[j] = 0;
5593 struct zone *lower_zone;
5597 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5598 sysctl_lowmem_reserve_ratio[idx] = 1;
5600 lower_zone = pgdat->node_zones + idx;
5601 lower_zone->lowmem_reserve[j] = managed_pages /
5602 sysctl_lowmem_reserve_ratio[idx];
5603 managed_pages += lower_zone->managed_pages;
5608 /* update totalreserve_pages */
5609 calculate_totalreserve_pages();
5612 static void __setup_per_zone_wmarks(void)
5614 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5615 unsigned long lowmem_pages = 0;
5617 unsigned long flags;
5619 /* Calculate total number of !ZONE_HIGHMEM pages */
5620 for_each_zone(zone) {
5621 if (!is_highmem(zone))
5622 lowmem_pages += zone->managed_pages;
5625 for_each_zone(zone) {
5628 spin_lock_irqsave(&zone->lock, flags);
5629 tmp = (u64)pages_min * zone->managed_pages;
5630 do_div(tmp, lowmem_pages);
5631 if (is_highmem(zone)) {
5633 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5634 * need highmem pages, so cap pages_min to a small
5637 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5638 * deltas controls asynch page reclaim, and so should
5639 * not be capped for highmem.
5641 unsigned long min_pages;
5643 min_pages = zone->managed_pages / 1024;
5644 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5645 zone->watermark[WMARK_MIN] = min_pages;
5648 * If it's a lowmem zone, reserve a number of pages
5649 * proportionate to the zone's size.
5651 zone->watermark[WMARK_MIN] = tmp;
5654 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5655 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5657 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5658 high_wmark_pages(zone) -
5659 low_wmark_pages(zone) -
5660 zone_page_state(zone, NR_ALLOC_BATCH));
5662 setup_zone_migrate_reserve(zone);
5663 spin_unlock_irqrestore(&zone->lock, flags);
5666 /* update totalreserve_pages */
5667 calculate_totalreserve_pages();
5671 * setup_per_zone_wmarks - called when min_free_kbytes changes
5672 * or when memory is hot-{added|removed}
5674 * Ensures that the watermark[min,low,high] values for each zone are set
5675 * correctly with respect to min_free_kbytes.
5677 void setup_per_zone_wmarks(void)
5679 mutex_lock(&zonelists_mutex);
5680 __setup_per_zone_wmarks();
5681 mutex_unlock(&zonelists_mutex);
5685 * The inactive anon list should be small enough that the VM never has to
5686 * do too much work, but large enough that each inactive page has a chance
5687 * to be referenced again before it is swapped out.
5689 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5690 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5691 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5692 * the anonymous pages are kept on the inactive list.
5695 * memory ratio inactive anon
5696 * -------------------------------------
5705 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5707 unsigned int gb, ratio;
5709 /* Zone size in gigabytes */
5710 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5712 ratio = int_sqrt(10 * gb);
5716 zone->inactive_ratio = ratio;
5719 static void __meminit setup_per_zone_inactive_ratio(void)
5724 calculate_zone_inactive_ratio(zone);
5728 * Initialise min_free_kbytes.
5730 * For small machines we want it small (128k min). For large machines
5731 * we want it large (64MB max). But it is not linear, because network
5732 * bandwidth does not increase linearly with machine size. We use
5734 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5735 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5751 int __meminit init_per_zone_wmark_min(void)
5753 unsigned long lowmem_kbytes;
5754 int new_min_free_kbytes;
5756 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5757 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5759 if (new_min_free_kbytes > user_min_free_kbytes) {
5760 min_free_kbytes = new_min_free_kbytes;
5761 if (min_free_kbytes < 128)
5762 min_free_kbytes = 128;
5763 if (min_free_kbytes > 65536)
5764 min_free_kbytes = 65536;
5766 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5767 new_min_free_kbytes, user_min_free_kbytes);
5769 setup_per_zone_wmarks();
5770 refresh_zone_stat_thresholds();
5771 setup_per_zone_lowmem_reserve();
5772 setup_per_zone_inactive_ratio();
5775 module_init(init_per_zone_wmark_min)
5778 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5779 * that we can call two helper functions whenever min_free_kbytes
5782 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5783 void __user *buffer, size_t *length, loff_t *ppos)
5787 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5792 user_min_free_kbytes = min_free_kbytes;
5793 setup_per_zone_wmarks();
5799 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5800 void __user *buffer, size_t *length, loff_t *ppos)
5805 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5810 zone->min_unmapped_pages = (zone->managed_pages *
5811 sysctl_min_unmapped_ratio) / 100;
5815 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5816 void __user *buffer, size_t *length, loff_t *ppos)
5821 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5826 zone->min_slab_pages = (zone->managed_pages *
5827 sysctl_min_slab_ratio) / 100;
5833 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5834 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5835 * whenever sysctl_lowmem_reserve_ratio changes.
5837 * The reserve ratio obviously has absolutely no relation with the
5838 * minimum watermarks. The lowmem reserve ratio can only make sense
5839 * if in function of the boot time zone sizes.
5841 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5842 void __user *buffer, size_t *length, loff_t *ppos)
5844 proc_dointvec_minmax(table, write, buffer, length, ppos);
5845 setup_per_zone_lowmem_reserve();
5850 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5851 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5852 * pagelist can have before it gets flushed back to buddy allocator.
5854 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5855 void __user *buffer, size_t *length, loff_t *ppos)
5861 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5862 if (!write || (ret < 0))
5865 mutex_lock(&pcp_batch_high_lock);
5866 for_each_populated_zone(zone) {
5868 high = zone->managed_pages / percpu_pagelist_fraction;
5869 for_each_possible_cpu(cpu)
5870 pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5873 mutex_unlock(&pcp_batch_high_lock);
5877 int hashdist = HASHDIST_DEFAULT;
5880 static int __init set_hashdist(char *str)
5884 hashdist = simple_strtoul(str, &str, 0);
5887 __setup("hashdist=", set_hashdist);
5891 * allocate a large system hash table from bootmem
5892 * - it is assumed that the hash table must contain an exact power-of-2
5893 * quantity of entries
5894 * - limit is the number of hash buckets, not the total allocation size
5896 void *__init alloc_large_system_hash(const char *tablename,
5897 unsigned long bucketsize,
5898 unsigned long numentries,
5901 unsigned int *_hash_shift,
5902 unsigned int *_hash_mask,
5903 unsigned long low_limit,
5904 unsigned long high_limit)
5906 unsigned long long max = high_limit;
5907 unsigned long log2qty, size;
5910 /* allow the kernel cmdline to have a say */
5912 /* round applicable memory size up to nearest megabyte */
5913 numentries = nr_kernel_pages;
5915 /* It isn't necessary when PAGE_SIZE >= 1MB */
5916 if (PAGE_SHIFT < 20)
5917 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5919 /* limit to 1 bucket per 2^scale bytes of low memory */
5920 if (scale > PAGE_SHIFT)
5921 numentries >>= (scale - PAGE_SHIFT);
5923 numentries <<= (PAGE_SHIFT - scale);
5925 /* Make sure we've got at least a 0-order allocation.. */
5926 if (unlikely(flags & HASH_SMALL)) {
5927 /* Makes no sense without HASH_EARLY */
5928 WARN_ON(!(flags & HASH_EARLY));
5929 if (!(numentries >> *_hash_shift)) {
5930 numentries = 1UL << *_hash_shift;
5931 BUG_ON(!numentries);
5933 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5934 numentries = PAGE_SIZE / bucketsize;
5936 numentries = roundup_pow_of_two(numentries);
5938 /* limit allocation size to 1/16 total memory by default */
5940 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5941 do_div(max, bucketsize);
5943 max = min(max, 0x80000000ULL);
5945 if (numentries < low_limit)
5946 numentries = low_limit;
5947 if (numentries > max)
5950 log2qty = ilog2(numentries);
5953 size = bucketsize << log2qty;
5954 if (flags & HASH_EARLY)
5955 table = memblock_virt_alloc_nopanic(size, 0);
5957 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5960 * If bucketsize is not a power-of-two, we may free
5961 * some pages at the end of hash table which
5962 * alloc_pages_exact() automatically does
5964 if (get_order(size) < MAX_ORDER) {
5965 table = alloc_pages_exact(size, GFP_ATOMIC);
5966 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5969 } while (!table && size > PAGE_SIZE && --log2qty);
5972 panic("Failed to allocate %s hash table\n", tablename);
5974 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5977 ilog2(size) - PAGE_SHIFT,
5981 *_hash_shift = log2qty;
5983 *_hash_mask = (1 << log2qty) - 1;
5988 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5989 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5992 #ifdef CONFIG_SPARSEMEM
5993 return __pfn_to_section(pfn)->pageblock_flags;
5995 return zone->pageblock_flags;
5996 #endif /* CONFIG_SPARSEMEM */
5999 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6001 #ifdef CONFIG_SPARSEMEM
6002 pfn &= (PAGES_PER_SECTION-1);
6003 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6005 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6006 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6007 #endif /* CONFIG_SPARSEMEM */
6011 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
6012 * @page: The page within the block of interest
6013 * @start_bitidx: The first bit of interest to retrieve
6014 * @end_bitidx: The last bit of interest
6015 * returns pageblock_bits flags
6017 unsigned long get_pageblock_flags_group(struct page *page,
6018 int start_bitidx, int end_bitidx)
6021 unsigned long *bitmap;
6022 unsigned long pfn, bitidx;
6023 unsigned long flags = 0;
6024 unsigned long value = 1;
6026 zone = page_zone(page);
6027 pfn = page_to_pfn(page);
6028 bitmap = get_pageblock_bitmap(zone, pfn);
6029 bitidx = pfn_to_bitidx(zone, pfn);
6031 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6032 if (test_bit(bitidx + start_bitidx, bitmap))
6039 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
6040 * @page: The page within the block of interest
6041 * @start_bitidx: The first bit of interest
6042 * @end_bitidx: The last bit of interest
6043 * @flags: The flags to set
6045 void set_pageblock_flags_group(struct page *page, unsigned long flags,
6046 int start_bitidx, int end_bitidx)
6049 unsigned long *bitmap;
6050 unsigned long pfn, bitidx;
6051 unsigned long value = 1;
6053 zone = page_zone(page);
6054 pfn = page_to_pfn(page);
6055 bitmap = get_pageblock_bitmap(zone, pfn);
6056 bitidx = pfn_to_bitidx(zone, pfn);
6057 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6059 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6061 __set_bit(bitidx + start_bitidx, bitmap);
6063 __clear_bit(bitidx + start_bitidx, bitmap);
6067 * This function checks whether pageblock includes unmovable pages or not.
6068 * If @count is not zero, it is okay to include less @count unmovable pages
6070 * PageLRU check without isolation or lru_lock could race so that
6071 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6072 * expect this function should be exact.
6074 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6075 bool skip_hwpoisoned_pages)
6077 unsigned long pfn, iter, found;
6081 * For avoiding noise data, lru_add_drain_all() should be called
6082 * If ZONE_MOVABLE, the zone never contains unmovable pages
6084 if (zone_idx(zone) == ZONE_MOVABLE)
6086 mt = get_pageblock_migratetype(page);
6087 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6090 pfn = page_to_pfn(page);
6091 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6092 unsigned long check = pfn + iter;
6094 if (!pfn_valid_within(check))
6097 page = pfn_to_page(check);
6100 * Hugepages are not in LRU lists, but they're movable.
6101 * We need not scan over tail pages bacause we don't
6102 * handle each tail page individually in migration.
6104 if (PageHuge(page)) {
6105 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6110 * We can't use page_count without pin a page
6111 * because another CPU can free compound page.
6112 * This check already skips compound tails of THP
6113 * because their page->_count is zero at all time.
6115 if (!atomic_read(&page->_count)) {
6116 if (PageBuddy(page))
6117 iter += (1 << page_order(page)) - 1;
6122 * The HWPoisoned page may be not in buddy system, and
6123 * page_count() is not 0.
6125 if (skip_hwpoisoned_pages && PageHWPoison(page))
6131 * If there are RECLAIMABLE pages, we need to check it.
6132 * But now, memory offline itself doesn't call shrink_slab()
6133 * and it still to be fixed.
6136 * If the page is not RAM, page_count()should be 0.
6137 * we don't need more check. This is an _used_ not-movable page.
6139 * The problematic thing here is PG_reserved pages. PG_reserved
6140 * is set to both of a memory hole page and a _used_ kernel
6149 bool is_pageblock_removable_nolock(struct page *page)
6155 * We have to be careful here because we are iterating over memory
6156 * sections which are not zone aware so we might end up outside of
6157 * the zone but still within the section.
6158 * We have to take care about the node as well. If the node is offline
6159 * its NODE_DATA will be NULL - see page_zone.
6161 if (!node_online(page_to_nid(page)))
6164 zone = page_zone(page);
6165 pfn = page_to_pfn(page);
6166 if (!zone_spans_pfn(zone, pfn))
6169 return !has_unmovable_pages(zone, page, 0, true);
6174 static unsigned long pfn_max_align_down(unsigned long pfn)
6176 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6177 pageblock_nr_pages) - 1);
6180 static unsigned long pfn_max_align_up(unsigned long pfn)
6182 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6183 pageblock_nr_pages));
6186 /* [start, end) must belong to a single zone. */
6187 static int __alloc_contig_migrate_range(struct compact_control *cc,
6188 unsigned long start, unsigned long end)
6190 /* This function is based on compact_zone() from compaction.c. */
6191 unsigned long nr_reclaimed;
6192 unsigned long pfn = start;
6193 unsigned int tries = 0;
6198 while (pfn < end || !list_empty(&cc->migratepages)) {
6199 if (fatal_signal_pending(current)) {
6204 if (list_empty(&cc->migratepages)) {
6205 cc->nr_migratepages = 0;
6206 pfn = isolate_migratepages_range(cc->zone, cc,
6213 } else if (++tries == 5) {
6214 ret = ret < 0 ? ret : -EBUSY;
6218 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6220 cc->nr_migratepages -= nr_reclaimed;
6222 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6223 NULL, 0, cc->mode, MR_CMA);
6226 putback_movable_pages(&cc->migratepages);
6233 * alloc_contig_range() -- tries to allocate given range of pages
6234 * @start: start PFN to allocate
6235 * @end: one-past-the-last PFN to allocate
6236 * @migratetype: migratetype of the underlaying pageblocks (either
6237 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6238 * in range must have the same migratetype and it must
6239 * be either of the two.
6241 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6242 * aligned, however it's the caller's responsibility to guarantee that
6243 * we are the only thread that changes migrate type of pageblocks the
6246 * The PFN range must belong to a single zone.
6248 * Returns zero on success or negative error code. On success all
6249 * pages which PFN is in [start, end) are allocated for the caller and
6250 * need to be freed with free_contig_range().
6252 int alloc_contig_range(unsigned long start, unsigned long end,
6253 unsigned migratetype)
6255 unsigned long outer_start, outer_end;
6258 struct compact_control cc = {
6259 .nr_migratepages = 0,
6261 .zone = page_zone(pfn_to_page(start)),
6262 .mode = MIGRATE_SYNC,
6263 .ignore_skip_hint = true,
6265 INIT_LIST_HEAD(&cc.migratepages);
6268 * What we do here is we mark all pageblocks in range as
6269 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6270 * have different sizes, and due to the way page allocator
6271 * work, we align the range to biggest of the two pages so
6272 * that page allocator won't try to merge buddies from
6273 * different pageblocks and change MIGRATE_ISOLATE to some
6274 * other migration type.
6276 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6277 * migrate the pages from an unaligned range (ie. pages that
6278 * we are interested in). This will put all the pages in
6279 * range back to page allocator as MIGRATE_ISOLATE.
6281 * When this is done, we take the pages in range from page
6282 * allocator removing them from the buddy system. This way
6283 * page allocator will never consider using them.
6285 * This lets us mark the pageblocks back as
6286 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6287 * aligned range but not in the unaligned, original range are
6288 * put back to page allocator so that buddy can use them.
6291 ret = start_isolate_page_range(pfn_max_align_down(start),
6292 pfn_max_align_up(end), migratetype,
6297 ret = __alloc_contig_migrate_range(&cc, start, end);
6302 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6303 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6304 * more, all pages in [start, end) are free in page allocator.
6305 * What we are going to do is to allocate all pages from
6306 * [start, end) (that is remove them from page allocator).
6308 * The only problem is that pages at the beginning and at the
6309 * end of interesting range may be not aligned with pages that
6310 * page allocator holds, ie. they can be part of higher order
6311 * pages. Because of this, we reserve the bigger range and
6312 * once this is done free the pages we are not interested in.
6314 * We don't have to hold zone->lock here because the pages are
6315 * isolated thus they won't get removed from buddy.
6318 lru_add_drain_all();
6322 outer_start = start;
6323 while (!PageBuddy(pfn_to_page(outer_start))) {
6324 if (++order >= MAX_ORDER) {
6328 outer_start &= ~0UL << order;
6331 /* Make sure the range is really isolated. */
6332 if (test_pages_isolated(outer_start, end, false)) {
6333 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6340 /* Grab isolated pages from freelists. */
6341 outer_end = isolate_freepages_range(&cc, outer_start, end);
6347 /* Free head and tail (if any) */
6348 if (start != outer_start)
6349 free_contig_range(outer_start, start - outer_start);
6350 if (end != outer_end)
6351 free_contig_range(end, outer_end - end);
6354 undo_isolate_page_range(pfn_max_align_down(start),
6355 pfn_max_align_up(end), migratetype);
6359 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6361 unsigned int count = 0;
6363 for (; nr_pages--; pfn++) {
6364 struct page *page = pfn_to_page(pfn);
6366 count += page_count(page) != 1;
6369 WARN(count != 0, "%d pages are still in use!\n", count);
6373 #ifdef CONFIG_MEMORY_HOTPLUG
6375 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6376 * page high values need to be recalulated.
6378 void __meminit zone_pcp_update(struct zone *zone)
6381 mutex_lock(&pcp_batch_high_lock);
6382 for_each_possible_cpu(cpu)
6383 pageset_set_high_and_batch(zone,
6384 per_cpu_ptr(zone->pageset, cpu));
6385 mutex_unlock(&pcp_batch_high_lock);
6389 void zone_pcp_reset(struct zone *zone)
6391 unsigned long flags;
6393 struct per_cpu_pageset *pset;
6395 /* avoid races with drain_pages() */
6396 local_irq_save(flags);
6397 if (zone->pageset != &boot_pageset) {
6398 for_each_online_cpu(cpu) {
6399 pset = per_cpu_ptr(zone->pageset, cpu);
6400 drain_zonestat(zone, pset);
6402 free_percpu(zone->pageset);
6403 zone->pageset = &boot_pageset;
6405 local_irq_restore(flags);
6408 #ifdef CONFIG_MEMORY_HOTREMOVE
6410 * All pages in the range must be isolated before calling this.
6413 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6419 unsigned long flags;
6420 /* find the first valid pfn */
6421 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6426 zone = page_zone(pfn_to_page(pfn));
6427 spin_lock_irqsave(&zone->lock, flags);
6429 while (pfn < end_pfn) {
6430 if (!pfn_valid(pfn)) {
6434 page = pfn_to_page(pfn);
6436 * The HWPoisoned page may be not in buddy system, and
6437 * page_count() is not 0.
6439 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6441 SetPageReserved(page);
6445 BUG_ON(page_count(page));
6446 BUG_ON(!PageBuddy(page));
6447 order = page_order(page);
6448 #ifdef CONFIG_DEBUG_VM
6449 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6450 pfn, 1 << order, end_pfn);
6452 list_del(&page->lru);
6453 rmv_page_order(page);
6454 zone->free_area[order].nr_free--;
6455 for (i = 0; i < (1 << order); i++)
6456 SetPageReserved((page+i));
6457 pfn += (1 << order);
6459 spin_unlock_irqrestore(&zone->lock, flags);
6463 #ifdef CONFIG_MEMORY_FAILURE
6464 bool is_free_buddy_page(struct page *page)
6466 struct zone *zone = page_zone(page);
6467 unsigned long pfn = page_to_pfn(page);
6468 unsigned long flags;
6471 spin_lock_irqsave(&zone->lock, flags);
6472 for (order = 0; order < MAX_ORDER; order++) {
6473 struct page *page_head = page - (pfn & ((1 << order) - 1));
6475 if (PageBuddy(page_head) && page_order(page_head) >= order)
6478 spin_unlock_irqrestore(&zone->lock, flags);
6480 return order < MAX_ORDER;
6484 static const struct trace_print_flags pageflag_names[] = {
6485 {1UL << PG_locked, "locked" },
6486 {1UL << PG_error, "error" },
6487 {1UL << PG_referenced, "referenced" },
6488 {1UL << PG_uptodate, "uptodate" },
6489 {1UL << PG_dirty, "dirty" },
6490 {1UL << PG_lru, "lru" },
6491 {1UL << PG_active, "active" },
6492 {1UL << PG_slab, "slab" },
6493 {1UL << PG_owner_priv_1, "owner_priv_1" },
6494 {1UL << PG_arch_1, "arch_1" },
6495 {1UL << PG_reserved, "reserved" },
6496 {1UL << PG_private, "private" },
6497 {1UL << PG_private_2, "private_2" },
6498 {1UL << PG_writeback, "writeback" },
6499 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6500 {1UL << PG_head, "head" },
6501 {1UL << PG_tail, "tail" },
6503 {1UL << PG_compound, "compound" },
6505 {1UL << PG_swapcache, "swapcache" },
6506 {1UL << PG_mappedtodisk, "mappedtodisk" },
6507 {1UL << PG_reclaim, "reclaim" },
6508 {1UL << PG_swapbacked, "swapbacked" },
6509 {1UL << PG_unevictable, "unevictable" },
6511 {1UL << PG_mlocked, "mlocked" },
6513 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6514 {1UL << PG_uncached, "uncached" },
6516 #ifdef CONFIG_MEMORY_FAILURE
6517 {1UL << PG_hwpoison, "hwpoison" },
6519 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6520 {1UL << PG_compound_lock, "compound_lock" },
6524 static void dump_page_flags(unsigned long flags)
6526 const char *delim = "";
6530 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6532 printk(KERN_ALERT "page flags: %#lx(", flags);
6534 /* remove zone id */
6535 flags &= (1UL << NR_PAGEFLAGS) - 1;
6537 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6539 mask = pageflag_names[i].mask;
6540 if ((flags & mask) != mask)
6544 printk("%s%s", delim, pageflag_names[i].name);
6548 /* check for left over flags */
6550 printk("%s%#lx", delim, flags);
6555 void dump_page_badflags(struct page *page, const char *reason,
6556 unsigned long badflags)
6559 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6560 page, atomic_read(&page->_count), page_mapcount(page),
6561 page->mapping, page->index);
6562 dump_page_flags(page->flags);
6564 pr_alert("page dumped because: %s\n", reason);
6565 if (page->flags & badflags) {
6566 pr_alert("bad because of flags:\n");
6567 dump_page_flags(page->flags & badflags);
6569 mem_cgroup_print_bad_page(page);
6572 void dump_page(struct page *page, const char *reason)
6574 dump_page_badflags(page, reason, 0);
6576 EXPORT_SYMBOL(dump_page);