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);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node);
76 EXPORT_PER_CPU_SYMBOL(numa_node);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
91 * Array of node states.
93 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
94 [N_POSSIBLE] = NODE_MASK_ALL,
95 [N_ONLINE] = { { [0] = 1UL } },
97 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
99 [N_HIGH_MEMORY] = { { [0] = 1UL } },
101 #ifdef CONFIG_MOVABLE_NODE
102 [N_MEMORY] = { { [0] = 1UL } },
104 [N_CPU] = { { [0] = 1UL } },
107 EXPORT_SYMBOL(node_states);
109 /* Protect totalram_pages and zone->managed_pages */
110 static DEFINE_SPINLOCK(managed_page_count_lock);
112 unsigned long totalram_pages __read_mostly;
113 unsigned long totalreserve_pages __read_mostly;
115 * When calculating the number of globally allowed dirty pages, there
116 * is a certain number of per-zone reserves that should not be
117 * considered dirtyable memory. This is the sum of those reserves
118 * over all existing zones that contribute dirtyable memory.
120 unsigned long dirty_balance_reserve __read_mostly;
122 int percpu_pagelist_fraction;
123 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
125 #ifdef CONFIG_PM_SLEEP
127 * The following functions are used by the suspend/hibernate code to temporarily
128 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
129 * while devices are suspended. To avoid races with the suspend/hibernate code,
130 * they should always be called with pm_mutex held (gfp_allowed_mask also should
131 * only be modified with pm_mutex held, unless the suspend/hibernate code is
132 * guaranteed not to run in parallel with that modification).
135 static gfp_t saved_gfp_mask;
137 void pm_restore_gfp_mask(void)
139 WARN_ON(!mutex_is_locked(&pm_mutex));
140 if (saved_gfp_mask) {
141 gfp_allowed_mask = saved_gfp_mask;
146 void pm_restrict_gfp_mask(void)
148 WARN_ON(!mutex_is_locked(&pm_mutex));
149 WARN_ON(saved_gfp_mask);
150 saved_gfp_mask = gfp_allowed_mask;
151 gfp_allowed_mask &= ~GFP_IOFS;
154 bool pm_suspended_storage(void)
156 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
160 #endif /* CONFIG_PM_SLEEP */
162 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
163 int pageblock_order __read_mostly;
166 static void __free_pages_ok(struct page *page, unsigned int order);
169 * results with 256, 32 in the lowmem_reserve sysctl:
170 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
171 * 1G machine -> (16M dma, 784M normal, 224M high)
172 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
173 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
174 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
176 * TBD: should special case ZONE_DMA32 machines here - in those we normally
177 * don't need any ZONE_NORMAL reservation
179 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
180 #ifdef CONFIG_ZONE_DMA
183 #ifdef CONFIG_ZONE_DMA32
186 #ifdef CONFIG_HIGHMEM
192 EXPORT_SYMBOL(totalram_pages);
194 static char * const zone_names[MAX_NR_ZONES] = {
195 #ifdef CONFIG_ZONE_DMA
198 #ifdef CONFIG_ZONE_DMA32
202 #ifdef CONFIG_HIGHMEM
208 int min_free_kbytes = 1024;
209 int user_min_free_kbytes = -1;
211 static unsigned long __meminitdata nr_kernel_pages;
212 static unsigned long __meminitdata nr_all_pages;
213 static unsigned long __meminitdata dma_reserve;
215 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
216 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
217 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
218 static unsigned long __initdata required_kernelcore;
219 static unsigned long __initdata required_movablecore;
220 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
222 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
224 EXPORT_SYMBOL(movable_zone);
225 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
228 int nr_node_ids __read_mostly = MAX_NUMNODES;
229 int nr_online_nodes __read_mostly = 1;
230 EXPORT_SYMBOL(nr_node_ids);
231 EXPORT_SYMBOL(nr_online_nodes);
234 int page_group_by_mobility_disabled __read_mostly;
236 void set_pageblock_migratetype(struct page *page, int migratetype)
238 if (unlikely(page_group_by_mobility_disabled &&
239 migratetype < MIGRATE_PCPTYPES))
240 migratetype = MIGRATE_UNMOVABLE;
242 set_pageblock_flags_group(page, (unsigned long)migratetype,
243 PB_migrate, PB_migrate_end);
246 bool oom_killer_disabled __read_mostly;
248 #ifdef CONFIG_DEBUG_VM
249 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
253 unsigned long pfn = page_to_pfn(page);
254 unsigned long sp, start_pfn;
257 seq = zone_span_seqbegin(zone);
258 start_pfn = zone->zone_start_pfn;
259 sp = zone->spanned_pages;
260 if (!zone_spans_pfn(zone, pfn))
262 } while (zone_span_seqretry(zone, seq));
265 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
266 pfn, zone_to_nid(zone), zone->name,
267 start_pfn, start_pfn + sp);
272 static int page_is_consistent(struct zone *zone, struct page *page)
274 if (!pfn_valid_within(page_to_pfn(page)))
276 if (zone != page_zone(page))
282 * Temporary debugging check for pages not lying within a given zone.
284 static int bad_range(struct zone *zone, struct page *page)
286 if (page_outside_zone_boundaries(zone, page))
288 if (!page_is_consistent(zone, page))
294 static inline int bad_range(struct zone *zone, struct page *page)
300 static void bad_page(struct page *page, const char *reason,
301 unsigned long bad_flags)
303 static unsigned long resume;
304 static unsigned long nr_shown;
305 static unsigned long nr_unshown;
307 /* Don't complain about poisoned pages */
308 if (PageHWPoison(page)) {
309 page_mapcount_reset(page); /* remove PageBuddy */
314 * Allow a burst of 60 reports, then keep quiet for that minute;
315 * or allow a steady drip of one report per second.
317 if (nr_shown == 60) {
318 if (time_before(jiffies, resume)) {
324 "BUG: Bad page state: %lu messages suppressed\n",
331 resume = jiffies + 60 * HZ;
333 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
334 current->comm, page_to_pfn(page));
335 dump_page_badflags(page, reason, bad_flags);
340 /* Leave bad fields for debug, except PageBuddy could make trouble */
341 page_mapcount_reset(page); /* remove PageBuddy */
342 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
346 * Higher-order pages are called "compound pages". They are structured thusly:
348 * The first PAGE_SIZE page is called the "head page".
350 * The remaining PAGE_SIZE pages are called "tail pages".
352 * All pages have PG_compound set. All tail pages have their ->first_page
353 * pointing at the head page.
355 * The first tail page's ->lru.next holds the address of the compound page's
356 * put_page() function. Its ->lru.prev holds the order of allocation.
357 * This usage means that zero-order pages may not be compound.
360 static void free_compound_page(struct page *page)
362 __free_pages_ok(page, compound_order(page));
365 void prep_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
370 set_compound_page_dtor(page, free_compound_page);
371 set_compound_order(page, order);
373 for (i = 1; i < nr_pages; i++) {
374 struct page *p = page + i;
375 set_page_count(p, 0);
376 p->first_page = page;
377 /* Make sure p->first_page is always valid for PageTail() */
383 /* update __split_huge_page_refcount if you change this function */
384 static int destroy_compound_page(struct page *page, unsigned long order)
387 int nr_pages = 1 << order;
390 if (unlikely(compound_order(page) != order)) {
391 bad_page(page, "wrong compound order", 0);
395 __ClearPageHead(page);
397 for (i = 1; i < nr_pages; i++) {
398 struct page *p = page + i;
400 if (unlikely(!PageTail(p))) {
401 bad_page(page, "PageTail not set", 0);
403 } else if (unlikely(p->first_page != page)) {
404 bad_page(page, "first_page not consistent", 0);
413 static inline void prep_zero_page(struct page *page, unsigned int order,
419 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
420 * and __GFP_HIGHMEM from hard or soft interrupt context.
422 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
423 for (i = 0; i < (1 << order); i++)
424 clear_highpage(page + i);
427 #ifdef CONFIG_DEBUG_PAGEALLOC
428 unsigned int _debug_guardpage_minorder;
430 static int __init debug_guardpage_minorder_setup(char *buf)
434 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
435 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
438 _debug_guardpage_minorder = res;
439 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
442 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
444 static inline void set_page_guard_flag(struct page *page)
446 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
449 static inline void clear_page_guard_flag(struct page *page)
451 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
454 static inline void set_page_guard_flag(struct page *page) { }
455 static inline void clear_page_guard_flag(struct page *page) { }
458 static inline void set_page_order(struct page *page, unsigned int order)
460 set_page_private(page, order);
461 __SetPageBuddy(page);
464 static inline void rmv_page_order(struct page *page)
466 __ClearPageBuddy(page);
467 set_page_private(page, 0);
471 * Locate the struct page for both the matching buddy in our
472 * pair (buddy1) and the combined O(n+1) page they form (page).
474 * 1) Any buddy B1 will have an order O twin B2 which satisfies
475 * the following equation:
477 * For example, if the starting buddy (buddy2) is #8 its order
479 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
481 * 2) Any buddy B will have an order O+1 parent P which
482 * satisfies the following equation:
485 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
487 static inline unsigned long
488 __find_buddy_index(unsigned long page_idx, unsigned int order)
490 return page_idx ^ (1 << order);
494 * This function checks whether a page is free && is the buddy
495 * we can do coalesce a page and its buddy if
496 * (a) the buddy is not in a hole &&
497 * (b) the buddy is in the buddy system &&
498 * (c) a page and its buddy have the same order &&
499 * (d) a page and its buddy are in the same zone.
501 * For recording whether a page is in the buddy system, we set ->_mapcount
502 * PAGE_BUDDY_MAPCOUNT_VALUE.
503 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
504 * serialized by zone->lock.
506 * For recording page's order, we use page_private(page).
508 static inline int page_is_buddy(struct page *page, struct page *buddy,
511 if (!pfn_valid_within(page_to_pfn(buddy)))
514 if (page_is_guard(buddy) && page_order(buddy) == order) {
515 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
517 if (page_zone_id(page) != page_zone_id(buddy))
523 if (PageBuddy(buddy) && page_order(buddy) == order) {
524 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
527 * zone check is done late to avoid uselessly
528 * calculating zone/node ids for pages that could
531 if (page_zone_id(page) != page_zone_id(buddy))
540 * Freeing function for a buddy system allocator.
542 * The concept of a buddy system is to maintain direct-mapped table
543 * (containing bit values) for memory blocks of various "orders".
544 * The bottom level table contains the map for the smallest allocatable
545 * units of memory (here, pages), and each level above it describes
546 * pairs of units from the levels below, hence, "buddies".
547 * At a high level, all that happens here is marking the table entry
548 * at the bottom level available, and propagating the changes upward
549 * as necessary, plus some accounting needed to play nicely with other
550 * parts of the VM system.
551 * At each level, we keep a list of pages, which are heads of continuous
552 * free pages of length of (1 << order) and marked with _mapcount
553 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
555 * So when we are allocating or freeing one, we can derive the state of the
556 * other. That is, if we allocate a small block, and both were
557 * free, the remainder of the region must be split into blocks.
558 * If a block is freed, and its buddy is also free, then this
559 * triggers coalescing into a block of larger size.
564 static inline void __free_one_page(struct page *page,
566 struct zone *zone, unsigned int order,
569 unsigned long page_idx;
570 unsigned long combined_idx;
571 unsigned long uninitialized_var(buddy_idx);
574 VM_BUG_ON(!zone_is_initialized(zone));
576 if (unlikely(PageCompound(page)))
577 if (unlikely(destroy_compound_page(page, order)))
580 VM_BUG_ON(migratetype == -1);
582 page_idx = pfn & ((1 << MAX_ORDER) - 1);
584 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
585 VM_BUG_ON_PAGE(bad_range(zone, page), page);
587 while (order < MAX_ORDER-1) {
588 buddy_idx = __find_buddy_index(page_idx, order);
589 buddy = page + (buddy_idx - page_idx);
590 if (!page_is_buddy(page, buddy, order))
593 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
594 * merge with it and move up one order.
596 if (page_is_guard(buddy)) {
597 clear_page_guard_flag(buddy);
598 set_page_private(page, 0);
599 __mod_zone_freepage_state(zone, 1 << order,
602 list_del(&buddy->lru);
603 zone->free_area[order].nr_free--;
604 rmv_page_order(buddy);
606 combined_idx = buddy_idx & page_idx;
607 page = page + (combined_idx - page_idx);
608 page_idx = combined_idx;
611 set_page_order(page, order);
614 * If this is not the largest possible page, check if the buddy
615 * of the next-highest order is free. If it is, it's possible
616 * that pages are being freed that will coalesce soon. In case,
617 * that is happening, add the free page to the tail of the list
618 * so it's less likely to be used soon and more likely to be merged
619 * as a higher order page
621 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
622 struct page *higher_page, *higher_buddy;
623 combined_idx = buddy_idx & page_idx;
624 higher_page = page + (combined_idx - page_idx);
625 buddy_idx = __find_buddy_index(combined_idx, order + 1);
626 higher_buddy = higher_page + (buddy_idx - combined_idx);
627 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
628 list_add_tail(&page->lru,
629 &zone->free_area[order].free_list[migratetype]);
634 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
636 zone->free_area[order].nr_free++;
639 static inline int free_pages_check(struct page *page)
641 const char *bad_reason = NULL;
642 unsigned long bad_flags = 0;
644 if (unlikely(page_mapcount(page)))
645 bad_reason = "nonzero mapcount";
646 if (unlikely(page->mapping != NULL))
647 bad_reason = "non-NULL mapping";
648 if (unlikely(atomic_read(&page->_count) != 0))
649 bad_reason = "nonzero _count";
650 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
651 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
652 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
654 if (unlikely(mem_cgroup_bad_page_check(page)))
655 bad_reason = "cgroup check failed";
656 if (unlikely(bad_reason)) {
657 bad_page(page, bad_reason, bad_flags);
660 page_cpupid_reset_last(page);
661 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
662 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
667 * Frees a number of pages from the PCP lists
668 * Assumes all pages on list are in same zone, and of same order.
669 * count is the number of pages to free.
671 * If the zone was previously in an "all pages pinned" state then look to
672 * see if this freeing clears that state.
674 * And clear the zone's pages_scanned counter, to hold off the "all pages are
675 * pinned" detection logic.
677 static void free_pcppages_bulk(struct zone *zone, int count,
678 struct per_cpu_pages *pcp)
684 spin_lock(&zone->lock);
685 zone->pages_scanned = 0;
689 struct list_head *list;
692 * Remove pages from lists in a round-robin fashion. A
693 * batch_free count is maintained that is incremented when an
694 * empty list is encountered. This is so more pages are freed
695 * off fuller lists instead of spinning excessively around empty
700 if (++migratetype == MIGRATE_PCPTYPES)
702 list = &pcp->lists[migratetype];
703 } while (list_empty(list));
705 /* This is the only non-empty list. Free them all. */
706 if (batch_free == MIGRATE_PCPTYPES)
707 batch_free = to_free;
710 int mt; /* migratetype of the to-be-freed page */
712 page = list_entry(list->prev, struct page, lru);
713 /* must delete as __free_one_page list manipulates */
714 list_del(&page->lru);
715 mt = get_freepage_migratetype(page);
716 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
717 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
718 trace_mm_page_pcpu_drain(page, 0, mt);
719 if (likely(!is_migrate_isolate_page(page))) {
720 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
721 if (is_migrate_cma(mt))
722 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
724 } while (--to_free && --batch_free && !list_empty(list));
726 spin_unlock(&zone->lock);
729 static void free_one_page(struct zone *zone,
730 struct page *page, unsigned long pfn,
734 spin_lock(&zone->lock);
735 zone->pages_scanned = 0;
737 __free_one_page(page, pfn, zone, order, migratetype);
738 if (unlikely(!is_migrate_isolate(migratetype)))
739 __mod_zone_freepage_state(zone, 1 << order, migratetype);
740 spin_unlock(&zone->lock);
743 static bool free_pages_prepare(struct page *page, unsigned int order)
748 trace_mm_page_free(page, order);
749 kmemcheck_free_shadow(page, order);
752 page->mapping = NULL;
753 for (i = 0; i < (1 << order); i++)
754 bad += free_pages_check(page + i);
758 if (!PageHighMem(page)) {
759 debug_check_no_locks_freed(page_address(page),
761 debug_check_no_obj_freed(page_address(page),
764 arch_free_page(page, order);
765 kernel_map_pages(page, 1 << order, 0);
770 static void __free_pages_ok(struct page *page, unsigned int order)
774 unsigned long pfn = page_to_pfn(page);
776 if (!free_pages_prepare(page, order))
779 migratetype = get_pfnblock_migratetype(page, pfn);
780 local_irq_save(flags);
781 __count_vm_events(PGFREE, 1 << order);
782 set_freepage_migratetype(page, migratetype);
783 free_one_page(page_zone(page), page, pfn, order, migratetype);
784 local_irq_restore(flags);
787 void __init __free_pages_bootmem(struct page *page, unsigned int order)
789 unsigned int nr_pages = 1 << order;
790 struct page *p = page;
794 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
796 __ClearPageReserved(p);
797 set_page_count(p, 0);
799 __ClearPageReserved(p);
800 set_page_count(p, 0);
802 page_zone(page)->managed_pages += nr_pages;
803 set_page_refcounted(page);
804 __free_pages(page, order);
808 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
809 void __init init_cma_reserved_pageblock(struct page *page)
811 unsigned i = pageblock_nr_pages;
812 struct page *p = page;
815 __ClearPageReserved(p);
816 set_page_count(p, 0);
819 set_pageblock_migratetype(page, MIGRATE_CMA);
821 if (pageblock_order >= MAX_ORDER) {
822 i = pageblock_nr_pages;
825 set_page_refcounted(p);
826 __free_pages(p, MAX_ORDER - 1);
827 p += MAX_ORDER_NR_PAGES;
828 } while (i -= MAX_ORDER_NR_PAGES);
830 set_page_refcounted(page);
831 __free_pages(page, pageblock_order);
834 adjust_managed_page_count(page, pageblock_nr_pages);
839 * The order of subdivision here is critical for the IO subsystem.
840 * Please do not alter this order without good reasons and regression
841 * testing. Specifically, as large blocks of memory are subdivided,
842 * the order in which smaller blocks are delivered depends on the order
843 * they're subdivided in this function. This is the primary factor
844 * influencing the order in which pages are delivered to the IO
845 * subsystem according to empirical testing, and this is also justified
846 * by considering the behavior of a buddy system containing a single
847 * large block of memory acted on by a series of small allocations.
848 * This behavior is a critical factor in sglist merging's success.
852 static inline void expand(struct zone *zone, struct page *page,
853 int low, int high, struct free_area *area,
856 unsigned long size = 1 << high;
862 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
864 #ifdef CONFIG_DEBUG_PAGEALLOC
865 if (high < debug_guardpage_minorder()) {
867 * Mark as guard pages (or page), that will allow to
868 * merge back to allocator when buddy will be freed.
869 * Corresponding page table entries will not be touched,
870 * pages will stay not present in virtual address space
872 INIT_LIST_HEAD(&page[size].lru);
873 set_page_guard_flag(&page[size]);
874 set_page_private(&page[size], high);
875 /* Guard pages are not available for any usage */
876 __mod_zone_freepage_state(zone, -(1 << high),
881 list_add(&page[size].lru, &area->free_list[migratetype]);
883 set_page_order(&page[size], high);
888 * This page is about to be returned from the page allocator
890 static inline int check_new_page(struct page *page)
892 const char *bad_reason = NULL;
893 unsigned long bad_flags = 0;
895 if (unlikely(page_mapcount(page)))
896 bad_reason = "nonzero mapcount";
897 if (unlikely(page->mapping != NULL))
898 bad_reason = "non-NULL mapping";
899 if (unlikely(atomic_read(&page->_count) != 0))
900 bad_reason = "nonzero _count";
901 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
902 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
903 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
905 if (unlikely(mem_cgroup_bad_page_check(page)))
906 bad_reason = "cgroup check failed";
907 if (unlikely(bad_reason)) {
908 bad_page(page, bad_reason, bad_flags);
914 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags)
918 for (i = 0; i < (1 << order); i++) {
919 struct page *p = page + i;
920 if (unlikely(check_new_page(p)))
924 set_page_private(page, 0);
925 set_page_refcounted(page);
927 arch_alloc_page(page, order);
928 kernel_map_pages(page, 1 << order, 1);
930 if (gfp_flags & __GFP_ZERO)
931 prep_zero_page(page, order, gfp_flags);
933 if (order && (gfp_flags & __GFP_COMP))
934 prep_compound_page(page, order);
940 * Go through the free lists for the given migratetype and remove
941 * the smallest available page from the freelists
944 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
947 unsigned int current_order;
948 struct free_area *area;
951 /* Find a page of the appropriate size in the preferred list */
952 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
953 area = &(zone->free_area[current_order]);
954 if (list_empty(&area->free_list[migratetype]))
957 page = list_entry(area->free_list[migratetype].next,
959 list_del(&page->lru);
960 rmv_page_order(page);
962 expand(zone, page, order, current_order, area, migratetype);
963 set_freepage_migratetype(page, migratetype);
972 * This array describes the order lists are fallen back to when
973 * the free lists for the desirable migrate type are depleted
975 static int fallbacks[MIGRATE_TYPES][4] = {
976 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
977 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
979 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
980 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
982 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
984 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
985 #ifdef CONFIG_MEMORY_ISOLATION
986 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
991 * Move the free pages in a range to the free lists of the requested type.
992 * Note that start_page and end_pages are not aligned on a pageblock
993 * boundary. If alignment is required, use move_freepages_block()
995 int move_freepages(struct zone *zone,
996 struct page *start_page, struct page *end_page,
1000 unsigned long order;
1001 int pages_moved = 0;
1003 #ifndef CONFIG_HOLES_IN_ZONE
1005 * page_zone is not safe to call in this context when
1006 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1007 * anyway as we check zone boundaries in move_freepages_block().
1008 * Remove at a later date when no bug reports exist related to
1009 * grouping pages by mobility
1011 BUG_ON(page_zone(start_page) != page_zone(end_page));
1014 for (page = start_page; page <= end_page;) {
1015 /* Make sure we are not inadvertently changing nodes */
1016 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1018 if (!pfn_valid_within(page_to_pfn(page))) {
1023 if (!PageBuddy(page)) {
1028 order = page_order(page);
1029 list_move(&page->lru,
1030 &zone->free_area[order].free_list[migratetype]);
1031 set_freepage_migratetype(page, migratetype);
1033 pages_moved += 1 << order;
1039 int move_freepages_block(struct zone *zone, struct page *page,
1042 unsigned long start_pfn, end_pfn;
1043 struct page *start_page, *end_page;
1045 start_pfn = page_to_pfn(page);
1046 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1047 start_page = pfn_to_page(start_pfn);
1048 end_page = start_page + pageblock_nr_pages - 1;
1049 end_pfn = start_pfn + pageblock_nr_pages - 1;
1051 /* Do not cross zone boundaries */
1052 if (!zone_spans_pfn(zone, start_pfn))
1054 if (!zone_spans_pfn(zone, end_pfn))
1057 return move_freepages(zone, start_page, end_page, migratetype);
1060 static void change_pageblock_range(struct page *pageblock_page,
1061 int start_order, int migratetype)
1063 int nr_pageblocks = 1 << (start_order - pageblock_order);
1065 while (nr_pageblocks--) {
1066 set_pageblock_migratetype(pageblock_page, migratetype);
1067 pageblock_page += pageblock_nr_pages;
1072 * If breaking a large block of pages, move all free pages to the preferred
1073 * allocation list. If falling back for a reclaimable kernel allocation, be
1074 * more aggressive about taking ownership of free pages.
1076 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1077 * nor move CMA pages to different free lists. We don't want unmovable pages
1078 * to be allocated from MIGRATE_CMA areas.
1080 * Returns the new migratetype of the pageblock (or the same old migratetype
1081 * if it was unchanged).
1083 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1084 int start_type, int fallback_type)
1086 int current_order = page_order(page);
1089 * When borrowing from MIGRATE_CMA, we need to release the excess
1090 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1091 * is set to CMA so it is returned to the correct freelist in case
1092 * the page ends up being not actually allocated from the pcp lists.
1094 if (is_migrate_cma(fallback_type))
1095 return fallback_type;
1097 /* Take ownership for orders >= pageblock_order */
1098 if (current_order >= pageblock_order) {
1099 change_pageblock_range(page, current_order, start_type);
1103 if (current_order >= pageblock_order / 2 ||
1104 start_type == MIGRATE_RECLAIMABLE ||
1105 page_group_by_mobility_disabled) {
1108 pages = move_freepages_block(zone, page, start_type);
1110 /* Claim the whole block if over half of it is free */
1111 if (pages >= (1 << (pageblock_order-1)) ||
1112 page_group_by_mobility_disabled) {
1114 set_pageblock_migratetype(page, start_type);
1120 return fallback_type;
1123 /* Remove an element from the buddy allocator from the fallback list */
1124 static inline struct page *
1125 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1127 struct free_area *area;
1128 unsigned int current_order;
1130 int migratetype, new_type, i;
1132 /* Find the largest possible block of pages in the other list */
1133 for (current_order = MAX_ORDER-1;
1134 current_order >= order && current_order <= MAX_ORDER-1;
1137 migratetype = fallbacks[start_migratetype][i];
1139 /* MIGRATE_RESERVE handled later if necessary */
1140 if (migratetype == MIGRATE_RESERVE)
1143 area = &(zone->free_area[current_order]);
1144 if (list_empty(&area->free_list[migratetype]))
1147 page = list_entry(area->free_list[migratetype].next,
1151 new_type = try_to_steal_freepages(zone, page,
1155 /* Remove the page from the freelists */
1156 list_del(&page->lru);
1157 rmv_page_order(page);
1159 expand(zone, page, order, current_order, area,
1161 /* The freepage_migratetype may differ from pageblock's
1162 * migratetype depending on the decisions in
1163 * try_to_steal_freepages. This is OK as long as it does
1164 * not differ for MIGRATE_CMA type.
1166 set_freepage_migratetype(page, new_type);
1168 trace_mm_page_alloc_extfrag(page, order, current_order,
1169 start_migratetype, migratetype, new_type);
1179 * Do the hard work of removing an element from the buddy allocator.
1180 * Call me with the zone->lock already held.
1182 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1188 page = __rmqueue_smallest(zone, order, migratetype);
1190 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1191 page = __rmqueue_fallback(zone, order, migratetype);
1194 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1195 * is used because __rmqueue_smallest is an inline function
1196 * and we want just one call site
1199 migratetype = MIGRATE_RESERVE;
1204 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1209 * Obtain a specified number of elements from the buddy allocator, all under
1210 * a single hold of the lock, for efficiency. Add them to the supplied list.
1211 * Returns the number of new pages which were placed at *list.
1213 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1214 unsigned long count, struct list_head *list,
1215 int migratetype, bool cold)
1219 spin_lock(&zone->lock);
1220 for (i = 0; i < count; ++i) {
1221 struct page *page = __rmqueue(zone, order, migratetype);
1222 if (unlikely(page == NULL))
1226 * Split buddy pages returned by expand() are received here
1227 * in physical page order. The page is added to the callers and
1228 * list and the list head then moves forward. From the callers
1229 * perspective, the linked list is ordered by page number in
1230 * some conditions. This is useful for IO devices that can
1231 * merge IO requests if the physical pages are ordered
1235 list_add(&page->lru, list);
1237 list_add_tail(&page->lru, list);
1239 if (is_migrate_cma(get_freepage_migratetype(page)))
1240 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1243 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1244 spin_unlock(&zone->lock);
1250 * Called from the vmstat counter updater to drain pagesets of this
1251 * currently executing processor on remote nodes after they have
1254 * Note that this function must be called with the thread pinned to
1255 * a single processor.
1257 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1259 unsigned long flags;
1260 int to_drain, batch;
1262 local_irq_save(flags);
1263 batch = ACCESS_ONCE(pcp->batch);
1264 to_drain = min(pcp->count, batch);
1266 free_pcppages_bulk(zone, to_drain, pcp);
1267 pcp->count -= to_drain;
1269 local_irq_restore(flags);
1274 * Drain pages of the indicated processor.
1276 * The processor must either be the current processor and the
1277 * thread pinned to the current processor or a processor that
1280 static void drain_pages(unsigned int cpu)
1282 unsigned long flags;
1285 for_each_populated_zone(zone) {
1286 struct per_cpu_pageset *pset;
1287 struct per_cpu_pages *pcp;
1289 local_irq_save(flags);
1290 pset = per_cpu_ptr(zone->pageset, cpu);
1294 free_pcppages_bulk(zone, pcp->count, pcp);
1297 local_irq_restore(flags);
1302 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1304 void drain_local_pages(void *arg)
1306 drain_pages(smp_processor_id());
1310 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1312 * Note that this code is protected against sending an IPI to an offline
1313 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1314 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1315 * nothing keeps CPUs from showing up after we populated the cpumask and
1316 * before the call to on_each_cpu_mask().
1318 void drain_all_pages(void)
1321 struct per_cpu_pageset *pcp;
1325 * Allocate in the BSS so we wont require allocation in
1326 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1328 static cpumask_t cpus_with_pcps;
1331 * We don't care about racing with CPU hotplug event
1332 * as offline notification will cause the notified
1333 * cpu to drain that CPU pcps and on_each_cpu_mask
1334 * disables preemption as part of its processing
1336 for_each_online_cpu(cpu) {
1337 bool has_pcps = false;
1338 for_each_populated_zone(zone) {
1339 pcp = per_cpu_ptr(zone->pageset, cpu);
1340 if (pcp->pcp.count) {
1346 cpumask_set_cpu(cpu, &cpus_with_pcps);
1348 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1350 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1353 #ifdef CONFIG_HIBERNATION
1355 void mark_free_pages(struct zone *zone)
1357 unsigned long pfn, max_zone_pfn;
1358 unsigned long flags;
1359 unsigned int order, t;
1360 struct list_head *curr;
1362 if (zone_is_empty(zone))
1365 spin_lock_irqsave(&zone->lock, flags);
1367 max_zone_pfn = zone_end_pfn(zone);
1368 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1369 if (pfn_valid(pfn)) {
1370 struct page *page = pfn_to_page(pfn);
1372 if (!swsusp_page_is_forbidden(page))
1373 swsusp_unset_page_free(page);
1376 for_each_migratetype_order(order, t) {
1377 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1380 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1381 for (i = 0; i < (1UL << order); i++)
1382 swsusp_set_page_free(pfn_to_page(pfn + i));
1385 spin_unlock_irqrestore(&zone->lock, flags);
1387 #endif /* CONFIG_PM */
1390 * Free a 0-order page
1391 * cold == true ? free a cold page : free a hot page
1393 void free_hot_cold_page(struct page *page, bool cold)
1395 struct zone *zone = page_zone(page);
1396 struct per_cpu_pages *pcp;
1397 unsigned long flags;
1398 unsigned long pfn = page_to_pfn(page);
1401 if (!free_pages_prepare(page, 0))
1404 migratetype = get_pfnblock_migratetype(page, pfn);
1405 set_freepage_migratetype(page, migratetype);
1406 local_irq_save(flags);
1407 __count_vm_event(PGFREE);
1410 * We only track unmovable, reclaimable and movable on pcp lists.
1411 * Free ISOLATE pages back to the allocator because they are being
1412 * offlined but treat RESERVE as movable pages so we can get those
1413 * areas back if necessary. Otherwise, we may have to free
1414 * excessively into the page allocator
1416 if (migratetype >= MIGRATE_PCPTYPES) {
1417 if (unlikely(is_migrate_isolate(migratetype))) {
1418 free_one_page(zone, page, pfn, 0, migratetype);
1421 migratetype = MIGRATE_MOVABLE;
1424 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1426 list_add(&page->lru, &pcp->lists[migratetype]);
1428 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1430 if (pcp->count >= pcp->high) {
1431 unsigned long batch = ACCESS_ONCE(pcp->batch);
1432 free_pcppages_bulk(zone, batch, pcp);
1433 pcp->count -= batch;
1437 local_irq_restore(flags);
1441 * Free a list of 0-order pages
1443 void free_hot_cold_page_list(struct list_head *list, bool cold)
1445 struct page *page, *next;
1447 list_for_each_entry_safe(page, next, list, lru) {
1448 trace_mm_page_free_batched(page, cold);
1449 free_hot_cold_page(page, cold);
1454 * split_page takes a non-compound higher-order page, and splits it into
1455 * n (1<<order) sub-pages: page[0..n]
1456 * Each sub-page must be freed individually.
1458 * Note: this is probably too low level an operation for use in drivers.
1459 * Please consult with lkml before using this in your driver.
1461 void split_page(struct page *page, unsigned int order)
1465 VM_BUG_ON_PAGE(PageCompound(page), page);
1466 VM_BUG_ON_PAGE(!page_count(page), page);
1468 #ifdef CONFIG_KMEMCHECK
1470 * Split shadow pages too, because free(page[0]) would
1471 * otherwise free the whole shadow.
1473 if (kmemcheck_page_is_tracked(page))
1474 split_page(virt_to_page(page[0].shadow), order);
1477 for (i = 1; i < (1 << order); i++)
1478 set_page_refcounted(page + i);
1480 EXPORT_SYMBOL_GPL(split_page);
1482 static int __isolate_free_page(struct page *page, unsigned int order)
1484 unsigned long watermark;
1488 BUG_ON(!PageBuddy(page));
1490 zone = page_zone(page);
1491 mt = get_pageblock_migratetype(page);
1493 if (!is_migrate_isolate(mt)) {
1494 /* Obey watermarks as if the page was being allocated */
1495 watermark = low_wmark_pages(zone) + (1 << order);
1496 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1499 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1502 /* Remove page from free list */
1503 list_del(&page->lru);
1504 zone->free_area[order].nr_free--;
1505 rmv_page_order(page);
1507 /* Set the pageblock if the isolated page is at least a pageblock */
1508 if (order >= pageblock_order - 1) {
1509 struct page *endpage = page + (1 << order) - 1;
1510 for (; page < endpage; page += pageblock_nr_pages) {
1511 int mt = get_pageblock_migratetype(page);
1512 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1513 set_pageblock_migratetype(page,
1518 return 1UL << order;
1522 * Similar to split_page except the page is already free. As this is only
1523 * being used for migration, the migratetype of the block also changes.
1524 * As this is called with interrupts disabled, the caller is responsible
1525 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1528 * Note: this is probably too low level an operation for use in drivers.
1529 * Please consult with lkml before using this in your driver.
1531 int split_free_page(struct page *page)
1536 order = page_order(page);
1538 nr_pages = __isolate_free_page(page, order);
1542 /* Split into individual pages */
1543 set_page_refcounted(page);
1544 split_page(page, order);
1549 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1550 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1554 struct page *buffered_rmqueue(struct zone *preferred_zone,
1555 struct zone *zone, unsigned int order,
1556 gfp_t gfp_flags, int migratetype)
1558 unsigned long flags;
1560 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1563 if (likely(order == 0)) {
1564 struct per_cpu_pages *pcp;
1565 struct list_head *list;
1567 local_irq_save(flags);
1568 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1569 list = &pcp->lists[migratetype];
1570 if (list_empty(list)) {
1571 pcp->count += rmqueue_bulk(zone, 0,
1574 if (unlikely(list_empty(list)))
1579 page = list_entry(list->prev, struct page, lru);
1581 page = list_entry(list->next, struct page, lru);
1583 list_del(&page->lru);
1586 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1588 * __GFP_NOFAIL is not to be used in new code.
1590 * All __GFP_NOFAIL callers should be fixed so that they
1591 * properly detect and handle allocation failures.
1593 * We most definitely don't want callers attempting to
1594 * allocate greater than order-1 page units with
1597 WARN_ON_ONCE(order > 1);
1599 spin_lock_irqsave(&zone->lock, flags);
1600 page = __rmqueue(zone, order, migratetype);
1601 spin_unlock(&zone->lock);
1604 __mod_zone_freepage_state(zone, -(1 << order),
1605 get_freepage_migratetype(page));
1608 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1610 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1611 zone_statistics(preferred_zone, zone, gfp_flags);
1612 local_irq_restore(flags);
1614 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1615 if (prep_new_page(page, order, gfp_flags))
1620 local_irq_restore(flags);
1624 #ifdef CONFIG_FAIL_PAGE_ALLOC
1627 struct fault_attr attr;
1629 u32 ignore_gfp_highmem;
1630 u32 ignore_gfp_wait;
1632 } fail_page_alloc = {
1633 .attr = FAULT_ATTR_INITIALIZER,
1634 .ignore_gfp_wait = 1,
1635 .ignore_gfp_highmem = 1,
1639 static int __init setup_fail_page_alloc(char *str)
1641 return setup_fault_attr(&fail_page_alloc.attr, str);
1643 __setup("fail_page_alloc=", setup_fail_page_alloc);
1645 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1647 if (order < fail_page_alloc.min_order)
1649 if (gfp_mask & __GFP_NOFAIL)
1651 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1653 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1656 return should_fail(&fail_page_alloc.attr, 1 << order);
1659 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1661 static int __init fail_page_alloc_debugfs(void)
1663 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1666 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1667 &fail_page_alloc.attr);
1669 return PTR_ERR(dir);
1671 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1672 &fail_page_alloc.ignore_gfp_wait))
1674 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1675 &fail_page_alloc.ignore_gfp_highmem))
1677 if (!debugfs_create_u32("min-order", mode, dir,
1678 &fail_page_alloc.min_order))
1683 debugfs_remove_recursive(dir);
1688 late_initcall(fail_page_alloc_debugfs);
1690 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1692 #else /* CONFIG_FAIL_PAGE_ALLOC */
1694 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1699 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1702 * Return true if free pages are above 'mark'. This takes into account the order
1703 * of the allocation.
1705 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1706 unsigned long mark, int classzone_idx, int alloc_flags,
1709 /* free_pages my go negative - that's OK */
1711 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1715 free_pages -= (1 << order) - 1;
1716 if (alloc_flags & ALLOC_HIGH)
1718 if (alloc_flags & ALLOC_HARDER)
1721 /* If allocation can't use CMA areas don't use free CMA pages */
1722 if (!(alloc_flags & ALLOC_CMA))
1723 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1726 if (free_pages - free_cma <= min + lowmem_reserve)
1728 for (o = 0; o < order; o++) {
1729 /* At the next order, this order's pages become unavailable */
1730 free_pages -= z->free_area[o].nr_free << o;
1732 /* Require fewer higher order pages to be free */
1735 if (free_pages <= min)
1741 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1742 int classzone_idx, int alloc_flags)
1744 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1745 zone_page_state(z, NR_FREE_PAGES));
1748 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1749 unsigned long mark, int classzone_idx, int alloc_flags)
1751 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1753 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1754 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1756 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1762 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1763 * skip over zones that are not allowed by the cpuset, or that have
1764 * been recently (in last second) found to be nearly full. See further
1765 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1766 * that have to skip over a lot of full or unallowed zones.
1768 * If the zonelist cache is present in the passed zonelist, then
1769 * returns a pointer to the allowed node mask (either the current
1770 * tasks mems_allowed, or node_states[N_MEMORY].)
1772 * If the zonelist cache is not available for this zonelist, does
1773 * nothing and returns NULL.
1775 * If the fullzones BITMAP in the zonelist cache is stale (more than
1776 * a second since last zap'd) then we zap it out (clear its bits.)
1778 * We hold off even calling zlc_setup, until after we've checked the
1779 * first zone in the zonelist, on the theory that most allocations will
1780 * be satisfied from that first zone, so best to examine that zone as
1781 * quickly as we can.
1783 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1785 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1786 nodemask_t *allowednodes; /* zonelist_cache approximation */
1788 zlc = zonelist->zlcache_ptr;
1792 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1793 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1794 zlc->last_full_zap = jiffies;
1797 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1798 &cpuset_current_mems_allowed :
1799 &node_states[N_MEMORY];
1800 return allowednodes;
1804 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1805 * if it is worth looking at further for free memory:
1806 * 1) Check that the zone isn't thought to be full (doesn't have its
1807 * bit set in the zonelist_cache fullzones BITMAP).
1808 * 2) Check that the zones node (obtained from the zonelist_cache
1809 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1810 * Return true (non-zero) if zone is worth looking at further, or
1811 * else return false (zero) if it is not.
1813 * This check -ignores- the distinction between various watermarks,
1814 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1815 * found to be full for any variation of these watermarks, it will
1816 * be considered full for up to one second by all requests, unless
1817 * we are so low on memory on all allowed nodes that we are forced
1818 * into the second scan of the zonelist.
1820 * In the second scan we ignore this zonelist cache and exactly
1821 * apply the watermarks to all zones, even it is slower to do so.
1822 * We are low on memory in the second scan, and should leave no stone
1823 * unturned looking for a free page.
1825 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1826 nodemask_t *allowednodes)
1828 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1829 int i; /* index of *z in zonelist zones */
1830 int n; /* node that zone *z is on */
1832 zlc = zonelist->zlcache_ptr;
1836 i = z - zonelist->_zonerefs;
1839 /* This zone is worth trying if it is allowed but not full */
1840 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1844 * Given 'z' scanning a zonelist, set the corresponding bit in
1845 * zlc->fullzones, so that subsequent attempts to allocate a page
1846 * from that zone don't waste time re-examining it.
1848 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1850 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1851 int i; /* index of *z in zonelist zones */
1853 zlc = zonelist->zlcache_ptr;
1857 i = z - zonelist->_zonerefs;
1859 set_bit(i, zlc->fullzones);
1863 * clear all zones full, called after direct reclaim makes progress so that
1864 * a zone that was recently full is not skipped over for up to a second
1866 static void zlc_clear_zones_full(struct zonelist *zonelist)
1868 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1870 zlc = zonelist->zlcache_ptr;
1874 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1877 static bool zone_local(struct zone *local_zone, struct zone *zone)
1879 return local_zone->node == zone->node;
1882 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1884 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
1888 #else /* CONFIG_NUMA */
1890 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1895 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1896 nodemask_t *allowednodes)
1901 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1905 static void zlc_clear_zones_full(struct zonelist *zonelist)
1909 static bool zone_local(struct zone *local_zone, struct zone *zone)
1914 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1919 #endif /* CONFIG_NUMA */
1922 * get_page_from_freelist goes through the zonelist trying to allocate
1925 static struct page *
1926 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1927 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1928 struct zone *preferred_zone, int classzone_idx, int migratetype)
1931 struct page *page = NULL;
1933 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1934 int zlc_active = 0; /* set if using zonelist_cache */
1935 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1936 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
1937 (gfp_mask & __GFP_WRITE);
1941 * Scan zonelist, looking for a zone with enough free.
1942 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1944 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1945 high_zoneidx, nodemask) {
1948 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1949 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1951 if (cpusets_enabled() &&
1952 (alloc_flags & ALLOC_CPUSET) &&
1953 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1956 * Distribute pages in proportion to the individual
1957 * zone size to ensure fair page aging. The zone a
1958 * page was allocated in should have no effect on the
1959 * time the page has in memory before being reclaimed.
1961 if (alloc_flags & ALLOC_FAIR) {
1962 if (!zone_local(preferred_zone, zone))
1964 if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1968 * When allocating a page cache page for writing, we
1969 * want to get it from a zone that is within its dirty
1970 * limit, such that no single zone holds more than its
1971 * proportional share of globally allowed dirty pages.
1972 * The dirty limits take into account the zone's
1973 * lowmem reserves and high watermark so that kswapd
1974 * should be able to balance it without having to
1975 * write pages from its LRU list.
1977 * This may look like it could increase pressure on
1978 * lower zones by failing allocations in higher zones
1979 * before they are full. But the pages that do spill
1980 * over are limited as the lower zones are protected
1981 * by this very same mechanism. It should not become
1982 * a practical burden to them.
1984 * XXX: For now, allow allocations to potentially
1985 * exceed the per-zone dirty limit in the slowpath
1986 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1987 * which is important when on a NUMA setup the allowed
1988 * zones are together not big enough to reach the
1989 * global limit. The proper fix for these situations
1990 * will require awareness of zones in the
1991 * dirty-throttling and the flusher threads.
1993 if (consider_zone_dirty && !zone_dirty_ok(zone))
1996 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1997 if (!zone_watermark_ok(zone, order, mark,
1998 classzone_idx, alloc_flags)) {
2001 /* Checked here to keep the fast path fast */
2002 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2003 if (alloc_flags & ALLOC_NO_WATERMARKS)
2006 if (IS_ENABLED(CONFIG_NUMA) &&
2007 !did_zlc_setup && nr_online_nodes > 1) {
2009 * we do zlc_setup if there are multiple nodes
2010 * and before considering the first zone allowed
2013 allowednodes = zlc_setup(zonelist, alloc_flags);
2018 if (zone_reclaim_mode == 0 ||
2019 !zone_allows_reclaim(preferred_zone, zone))
2020 goto this_zone_full;
2023 * As we may have just activated ZLC, check if the first
2024 * eligible zone has failed zone_reclaim recently.
2026 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2027 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2030 ret = zone_reclaim(zone, gfp_mask, order);
2032 case ZONE_RECLAIM_NOSCAN:
2035 case ZONE_RECLAIM_FULL:
2036 /* scanned but unreclaimable */
2039 /* did we reclaim enough */
2040 if (zone_watermark_ok(zone, order, mark,
2041 classzone_idx, alloc_flags))
2045 * Failed to reclaim enough to meet watermark.
2046 * Only mark the zone full if checking the min
2047 * watermark or if we failed to reclaim just
2048 * 1<<order pages or else the page allocator
2049 * fastpath will prematurely mark zones full
2050 * when the watermark is between the low and
2053 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2054 ret == ZONE_RECLAIM_SOME)
2055 goto this_zone_full;
2062 page = buffered_rmqueue(preferred_zone, zone, order,
2063 gfp_mask, migratetype);
2067 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2068 zlc_mark_zone_full(zonelist, z);
2071 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2072 /* Disable zlc cache for second zonelist scan */
2079 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2080 * necessary to allocate the page. The expectation is
2081 * that the caller is taking steps that will free more
2082 * memory. The caller should avoid the page being used
2083 * for !PFMEMALLOC purposes.
2085 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2091 * Large machines with many possible nodes should not always dump per-node
2092 * meminfo in irq context.
2094 static inline bool should_suppress_show_mem(void)
2099 ret = in_interrupt();
2104 static DEFINE_RATELIMIT_STATE(nopage_rs,
2105 DEFAULT_RATELIMIT_INTERVAL,
2106 DEFAULT_RATELIMIT_BURST);
2108 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2110 unsigned int filter = SHOW_MEM_FILTER_NODES;
2112 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2113 debug_guardpage_minorder() > 0)
2117 * This documents exceptions given to allocations in certain
2118 * contexts that are allowed to allocate outside current's set
2121 if (!(gfp_mask & __GFP_NOMEMALLOC))
2122 if (test_thread_flag(TIF_MEMDIE) ||
2123 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2124 filter &= ~SHOW_MEM_FILTER_NODES;
2125 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2126 filter &= ~SHOW_MEM_FILTER_NODES;
2129 struct va_format vaf;
2132 va_start(args, fmt);
2137 pr_warn("%pV", &vaf);
2142 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2143 current->comm, order, gfp_mask);
2146 if (!should_suppress_show_mem())
2151 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2152 unsigned long did_some_progress,
2153 unsigned long pages_reclaimed)
2155 /* Do not loop if specifically requested */
2156 if (gfp_mask & __GFP_NORETRY)
2159 /* Always retry if specifically requested */
2160 if (gfp_mask & __GFP_NOFAIL)
2164 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2165 * making forward progress without invoking OOM. Suspend also disables
2166 * storage devices so kswapd will not help. Bail if we are suspending.
2168 if (!did_some_progress && pm_suspended_storage())
2172 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2173 * means __GFP_NOFAIL, but that may not be true in other
2176 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2180 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2181 * specified, then we retry until we no longer reclaim any pages
2182 * (above), or we've reclaimed an order of pages at least as
2183 * large as the allocation's order. In both cases, if the
2184 * allocation still fails, we stop retrying.
2186 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2192 static inline struct page *
2193 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2194 struct zonelist *zonelist, enum zone_type high_zoneidx,
2195 nodemask_t *nodemask, struct zone *preferred_zone,
2196 int classzone_idx, int migratetype)
2200 /* Acquire the OOM killer lock for the zones in zonelist */
2201 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2202 schedule_timeout_uninterruptible(1);
2207 * Go through the zonelist yet one more time, keep very high watermark
2208 * here, this is only to catch a parallel oom killing, we must fail if
2209 * we're still under heavy pressure.
2211 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2212 order, zonelist, high_zoneidx,
2213 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2214 preferred_zone, classzone_idx, migratetype);
2218 if (!(gfp_mask & __GFP_NOFAIL)) {
2219 /* The OOM killer will not help higher order allocs */
2220 if (order > PAGE_ALLOC_COSTLY_ORDER)
2222 /* The OOM killer does not needlessly kill tasks for lowmem */
2223 if (high_zoneidx < ZONE_NORMAL)
2226 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2227 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2228 * The caller should handle page allocation failure by itself if
2229 * it specifies __GFP_THISNODE.
2230 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2232 if (gfp_mask & __GFP_THISNODE)
2235 /* Exhausted what can be done so it's blamo time */
2236 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2239 clear_zonelist_oom(zonelist, gfp_mask);
2243 #ifdef CONFIG_COMPACTION
2244 /* Try memory compaction for high-order allocations before reclaim */
2245 static struct page *
2246 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2247 struct zonelist *zonelist, enum zone_type high_zoneidx,
2248 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2249 int classzone_idx, int migratetype, enum migrate_mode mode,
2250 bool *contended_compaction, bool *deferred_compaction,
2251 unsigned long *did_some_progress)
2256 if (compaction_deferred(preferred_zone, order)) {
2257 *deferred_compaction = true;
2261 current->flags |= PF_MEMALLOC;
2262 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2264 contended_compaction);
2265 current->flags &= ~PF_MEMALLOC;
2267 if (*did_some_progress != COMPACT_SKIPPED) {
2270 /* Page migration frees to the PCP lists but we want merging */
2271 drain_pages(get_cpu());
2274 page = get_page_from_freelist(gfp_mask, nodemask,
2275 order, zonelist, high_zoneidx,
2276 alloc_flags & ~ALLOC_NO_WATERMARKS,
2277 preferred_zone, classzone_idx, migratetype);
2279 preferred_zone->compact_blockskip_flush = false;
2280 compaction_defer_reset(preferred_zone, order, true);
2281 count_vm_event(COMPACTSUCCESS);
2286 * It's bad if compaction run occurs and fails.
2287 * The most likely reason is that pages exist,
2288 * but not enough to satisfy watermarks.
2290 count_vm_event(COMPACTFAIL);
2293 * As async compaction considers a subset of pageblocks, only
2294 * defer if the failure was a sync compaction failure.
2296 if (mode != MIGRATE_ASYNC)
2297 defer_compaction(preferred_zone, order);
2305 static inline struct page *
2306 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2307 struct zonelist *zonelist, enum zone_type high_zoneidx,
2308 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2309 int classzone_idx, int migratetype,
2310 enum migrate_mode mode, bool *contended_compaction,
2311 bool *deferred_compaction, unsigned long *did_some_progress)
2315 #endif /* CONFIG_COMPACTION */
2317 /* Perform direct synchronous page reclaim */
2319 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2320 nodemask_t *nodemask)
2322 struct reclaim_state reclaim_state;
2327 /* We now go into synchronous reclaim */
2328 cpuset_memory_pressure_bump();
2329 current->flags |= PF_MEMALLOC;
2330 lockdep_set_current_reclaim_state(gfp_mask);
2331 reclaim_state.reclaimed_slab = 0;
2332 current->reclaim_state = &reclaim_state;
2334 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2336 current->reclaim_state = NULL;
2337 lockdep_clear_current_reclaim_state();
2338 current->flags &= ~PF_MEMALLOC;
2345 /* The really slow allocator path where we enter direct reclaim */
2346 static inline struct page *
2347 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2348 struct zonelist *zonelist, enum zone_type high_zoneidx,
2349 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2350 int classzone_idx, int migratetype, unsigned long *did_some_progress)
2352 struct page *page = NULL;
2353 bool drained = false;
2355 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2357 if (unlikely(!(*did_some_progress)))
2360 /* After successful reclaim, reconsider all zones for allocation */
2361 if (IS_ENABLED(CONFIG_NUMA))
2362 zlc_clear_zones_full(zonelist);
2365 page = get_page_from_freelist(gfp_mask, nodemask, order,
2366 zonelist, high_zoneidx,
2367 alloc_flags & ~ALLOC_NO_WATERMARKS,
2368 preferred_zone, classzone_idx,
2372 * If an allocation failed after direct reclaim, it could be because
2373 * pages are pinned on the per-cpu lists. Drain them and try again
2375 if (!page && !drained) {
2385 * This is called in the allocator slow-path if the allocation request is of
2386 * sufficient urgency to ignore watermarks and take other desperate measures
2388 static inline struct page *
2389 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2390 struct zonelist *zonelist, enum zone_type high_zoneidx,
2391 nodemask_t *nodemask, struct zone *preferred_zone,
2392 int classzone_idx, int migratetype)
2397 page = get_page_from_freelist(gfp_mask, nodemask, order,
2398 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2399 preferred_zone, classzone_idx, migratetype);
2401 if (!page && gfp_mask & __GFP_NOFAIL)
2402 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2403 } while (!page && (gfp_mask & __GFP_NOFAIL));
2408 static void reset_alloc_batches(struct zonelist *zonelist,
2409 enum zone_type high_zoneidx,
2410 struct zone *preferred_zone)
2415 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2417 * Only reset the batches of zones that were actually
2418 * considered in the fairness pass, we don't want to
2419 * trash fairness information for zones that are not
2420 * actually part of this zonelist's round-robin cycle.
2422 if (!zone_local(preferred_zone, zone))
2424 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2425 high_wmark_pages(zone) - low_wmark_pages(zone) -
2426 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2430 static void wake_all_kswapds(unsigned int order,
2431 struct zonelist *zonelist,
2432 enum zone_type high_zoneidx,
2433 struct zone *preferred_zone)
2438 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2439 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2443 gfp_to_alloc_flags(gfp_t gfp_mask)
2445 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2446 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2448 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2449 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2452 * The caller may dip into page reserves a bit more if the caller
2453 * cannot run direct reclaim, or if the caller has realtime scheduling
2454 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2455 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2457 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2461 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2462 * if it can't schedule.
2464 if (!(gfp_mask & __GFP_NOMEMALLOC))
2465 alloc_flags |= ALLOC_HARDER;
2467 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2468 * comment for __cpuset_node_allowed_softwall().
2470 alloc_flags &= ~ALLOC_CPUSET;
2471 } else if (unlikely(rt_task(current)) && !in_interrupt())
2472 alloc_flags |= ALLOC_HARDER;
2474 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2475 if (gfp_mask & __GFP_MEMALLOC)
2476 alloc_flags |= ALLOC_NO_WATERMARKS;
2477 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2478 alloc_flags |= ALLOC_NO_WATERMARKS;
2479 else if (!in_interrupt() &&
2480 ((current->flags & PF_MEMALLOC) ||
2481 unlikely(test_thread_flag(TIF_MEMDIE))))
2482 alloc_flags |= ALLOC_NO_WATERMARKS;
2485 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2486 alloc_flags |= ALLOC_CMA;
2491 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2493 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2496 static inline struct page *
2497 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2498 struct zonelist *zonelist, enum zone_type high_zoneidx,
2499 nodemask_t *nodemask, struct zone *preferred_zone,
2500 int classzone_idx, int migratetype)
2502 const gfp_t wait = gfp_mask & __GFP_WAIT;
2503 struct page *page = NULL;
2505 unsigned long pages_reclaimed = 0;
2506 unsigned long did_some_progress;
2507 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2508 bool deferred_compaction = false;
2509 bool contended_compaction = false;
2512 * In the slowpath, we sanity check order to avoid ever trying to
2513 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2514 * be using allocators in order of preference for an area that is
2517 if (order >= MAX_ORDER) {
2518 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2523 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2524 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2525 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2526 * using a larger set of nodes after it has established that the
2527 * allowed per node queues are empty and that nodes are
2530 if (IS_ENABLED(CONFIG_NUMA) &&
2531 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2535 if (!(gfp_mask & __GFP_NO_KSWAPD))
2536 wake_all_kswapds(order, zonelist, high_zoneidx, preferred_zone);
2539 * OK, we're below the kswapd watermark and have kicked background
2540 * reclaim. Now things get more complex, so set up alloc_flags according
2541 * to how we want to proceed.
2543 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2546 * Find the true preferred zone if the allocation is unconstrained by
2549 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) {
2550 struct zoneref *preferred_zoneref;
2551 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2552 NULL, &preferred_zone);
2553 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2557 /* This is the last chance, in general, before the goto nopage. */
2558 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2559 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2560 preferred_zone, classzone_idx, migratetype);
2564 /* Allocate without watermarks if the context allows */
2565 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2567 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2568 * the allocation is high priority and these type of
2569 * allocations are system rather than user orientated
2571 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2573 page = __alloc_pages_high_priority(gfp_mask, order,
2574 zonelist, high_zoneidx, nodemask,
2575 preferred_zone, classzone_idx, migratetype);
2581 /* Atomic allocations - we can't balance anything */
2584 * All existing users of the deprecated __GFP_NOFAIL are
2585 * blockable, so warn of any new users that actually allow this
2586 * type of allocation to fail.
2588 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2592 /* Avoid recursion of direct reclaim */
2593 if (current->flags & PF_MEMALLOC)
2596 /* Avoid allocations with no watermarks from looping endlessly */
2597 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2601 * Try direct compaction. The first pass is asynchronous. Subsequent
2602 * attempts after direct reclaim are synchronous
2604 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2605 high_zoneidx, nodemask, alloc_flags,
2607 classzone_idx, migratetype,
2608 migration_mode, &contended_compaction,
2609 &deferred_compaction,
2610 &did_some_progress);
2615 * It can become very expensive to allocate transparent hugepages at
2616 * fault, so use asynchronous memory compaction for THP unless it is
2617 * khugepaged trying to collapse.
2619 if (!(gfp_mask & __GFP_NO_KSWAPD) || (current->flags & PF_KTHREAD))
2620 migration_mode = MIGRATE_SYNC_LIGHT;
2623 * If compaction is deferred for high-order allocations, it is because
2624 * sync compaction recently failed. In this is the case and the caller
2625 * requested a movable allocation that does not heavily disrupt the
2626 * system then fail the allocation instead of entering direct reclaim.
2628 if ((deferred_compaction || contended_compaction) &&
2629 (gfp_mask & __GFP_NO_KSWAPD))
2632 /* Try direct reclaim and then allocating */
2633 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2634 zonelist, high_zoneidx,
2636 alloc_flags, preferred_zone,
2637 classzone_idx, migratetype,
2638 &did_some_progress);
2643 * If we failed to make any progress reclaiming, then we are
2644 * running out of options and have to consider going OOM
2646 if (!did_some_progress) {
2647 if (oom_gfp_allowed(gfp_mask)) {
2648 if (oom_killer_disabled)
2650 /* Coredumps can quickly deplete all memory reserves */
2651 if ((current->flags & PF_DUMPCORE) &&
2652 !(gfp_mask & __GFP_NOFAIL))
2654 page = __alloc_pages_may_oom(gfp_mask, order,
2655 zonelist, high_zoneidx,
2656 nodemask, preferred_zone,
2657 classzone_idx, migratetype);
2661 if (!(gfp_mask & __GFP_NOFAIL)) {
2663 * The oom killer is not called for high-order
2664 * allocations that may fail, so if no progress
2665 * is being made, there are no other options and
2666 * retrying is unlikely to help.
2668 if (order > PAGE_ALLOC_COSTLY_ORDER)
2671 * The oom killer is not called for lowmem
2672 * allocations to prevent needlessly killing
2675 if (high_zoneidx < ZONE_NORMAL)
2683 /* Check if we should retry the allocation */
2684 pages_reclaimed += did_some_progress;
2685 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2687 /* Wait for some write requests to complete then retry */
2688 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2692 * High-order allocations do not necessarily loop after
2693 * direct reclaim and reclaim/compaction depends on compaction
2694 * being called after reclaim so call directly if necessary
2696 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist,
2697 high_zoneidx, nodemask, alloc_flags,
2699 classzone_idx, migratetype,
2700 migration_mode, &contended_compaction,
2701 &deferred_compaction,
2702 &did_some_progress);
2708 warn_alloc_failed(gfp_mask, order, NULL);
2711 if (kmemcheck_enabled)
2712 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2718 * This is the 'heart' of the zoned buddy allocator.
2721 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2722 struct zonelist *zonelist, nodemask_t *nodemask)
2724 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2725 struct zone *preferred_zone;
2726 struct zoneref *preferred_zoneref;
2727 struct page *page = NULL;
2728 int migratetype = allocflags_to_migratetype(gfp_mask);
2729 unsigned int cpuset_mems_cookie;
2730 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2733 gfp_mask &= gfp_allowed_mask;
2735 lockdep_trace_alloc(gfp_mask);
2737 might_sleep_if(gfp_mask & __GFP_WAIT);
2739 if (should_fail_alloc_page(gfp_mask, order))
2743 * Check the zones suitable for the gfp_mask contain at least one
2744 * valid zone. It's possible to have an empty zonelist as a result
2745 * of GFP_THISNODE and a memoryless node
2747 if (unlikely(!zonelist->_zonerefs->zone))
2751 cpuset_mems_cookie = read_mems_allowed_begin();
2753 /* The preferred zone is used for statistics later */
2754 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx,
2755 nodemask ? : &cpuset_current_mems_allowed,
2757 if (!preferred_zone)
2759 classzone_idx = zonelist_zone_idx(preferred_zoneref);
2762 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2763 alloc_flags |= ALLOC_CMA;
2766 /* First allocation attempt */
2767 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2768 zonelist, high_zoneidx, alloc_flags,
2769 preferred_zone, classzone_idx, migratetype);
2770 if (unlikely(!page)) {
2772 * The first pass makes sure allocations are spread
2773 * fairly within the local node. However, the local
2774 * node might have free pages left after the fairness
2775 * batches are exhausted, and remote zones haven't
2776 * even been considered yet. Try once more without
2777 * fairness, and include remote zones now, before
2778 * entering the slowpath and waking kswapd: prefer
2779 * spilling to a remote zone over swapping locally.
2781 if (alloc_flags & ALLOC_FAIR) {
2782 reset_alloc_batches(zonelist, high_zoneidx,
2784 alloc_flags &= ~ALLOC_FAIR;
2788 * Runtime PM, block IO and its error handling path
2789 * can deadlock because I/O on the device might not
2792 gfp_mask = memalloc_noio_flags(gfp_mask);
2793 page = __alloc_pages_slowpath(gfp_mask, order,
2794 zonelist, high_zoneidx, nodemask,
2795 preferred_zone, classzone_idx, migratetype);
2798 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2802 * When updating a task's mems_allowed, it is possible to race with
2803 * parallel threads in such a way that an allocation can fail while
2804 * the mask is being updated. If a page allocation is about to fail,
2805 * check if the cpuset changed during allocation and if so, retry.
2807 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2812 EXPORT_SYMBOL(__alloc_pages_nodemask);
2815 * Common helper functions.
2817 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2822 * __get_free_pages() returns a 32-bit address, which cannot represent
2825 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2827 page = alloc_pages(gfp_mask, order);
2830 return (unsigned long) page_address(page);
2832 EXPORT_SYMBOL(__get_free_pages);
2834 unsigned long get_zeroed_page(gfp_t gfp_mask)
2836 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2838 EXPORT_SYMBOL(get_zeroed_page);
2840 void __free_pages(struct page *page, unsigned int order)
2842 if (put_page_testzero(page)) {
2844 free_hot_cold_page(page, false);
2846 __free_pages_ok(page, order);
2850 EXPORT_SYMBOL(__free_pages);
2852 void free_pages(unsigned long addr, unsigned int order)
2855 VM_BUG_ON(!virt_addr_valid((void *)addr));
2856 __free_pages(virt_to_page((void *)addr), order);
2860 EXPORT_SYMBOL(free_pages);
2863 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2864 * of the current memory cgroup.
2866 * It should be used when the caller would like to use kmalloc, but since the
2867 * allocation is large, it has to fall back to the page allocator.
2869 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2872 struct mem_cgroup *memcg = NULL;
2874 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2876 page = alloc_pages(gfp_mask, order);
2877 memcg_kmem_commit_charge(page, memcg, order);
2881 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2884 struct mem_cgroup *memcg = NULL;
2886 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2888 page = alloc_pages_node(nid, gfp_mask, order);
2889 memcg_kmem_commit_charge(page, memcg, order);
2894 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2897 void __free_kmem_pages(struct page *page, unsigned int order)
2899 memcg_kmem_uncharge_pages(page, order);
2900 __free_pages(page, order);
2903 void free_kmem_pages(unsigned long addr, unsigned int order)
2906 VM_BUG_ON(!virt_addr_valid((void *)addr));
2907 __free_kmem_pages(virt_to_page((void *)addr), order);
2911 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2914 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2915 unsigned long used = addr + PAGE_ALIGN(size);
2917 split_page(virt_to_page((void *)addr), order);
2918 while (used < alloc_end) {
2923 return (void *)addr;
2927 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2928 * @size: the number of bytes to allocate
2929 * @gfp_mask: GFP flags for the allocation
2931 * This function is similar to alloc_pages(), except that it allocates the
2932 * minimum number of pages to satisfy the request. alloc_pages() can only
2933 * allocate memory in power-of-two pages.
2935 * This function is also limited by MAX_ORDER.
2937 * Memory allocated by this function must be released by free_pages_exact().
2939 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2941 unsigned int order = get_order(size);
2944 addr = __get_free_pages(gfp_mask, order);
2945 return make_alloc_exact(addr, order, size);
2947 EXPORT_SYMBOL(alloc_pages_exact);
2950 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2952 * @nid: the preferred node ID where memory should be allocated
2953 * @size: the number of bytes to allocate
2954 * @gfp_mask: GFP flags for the allocation
2956 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2958 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2961 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2963 unsigned order = get_order(size);
2964 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2967 return make_alloc_exact((unsigned long)page_address(p), order, size);
2971 * free_pages_exact - release memory allocated via alloc_pages_exact()
2972 * @virt: the value returned by alloc_pages_exact.
2973 * @size: size of allocation, same value as passed to alloc_pages_exact().
2975 * Release the memory allocated by a previous call to alloc_pages_exact.
2977 void free_pages_exact(void *virt, size_t size)
2979 unsigned long addr = (unsigned long)virt;
2980 unsigned long end = addr + PAGE_ALIGN(size);
2982 while (addr < end) {
2987 EXPORT_SYMBOL(free_pages_exact);
2990 * nr_free_zone_pages - count number of pages beyond high watermark
2991 * @offset: The zone index of the highest zone
2993 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2994 * high watermark within all zones at or below a given zone index. For each
2995 * zone, the number of pages is calculated as:
2996 * managed_pages - high_pages
2998 static unsigned long nr_free_zone_pages(int offset)
3003 /* Just pick one node, since fallback list is circular */
3004 unsigned long sum = 0;
3006 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3008 for_each_zone_zonelist(zone, z, zonelist, offset) {
3009 unsigned long size = zone->managed_pages;
3010 unsigned long high = high_wmark_pages(zone);
3019 * nr_free_buffer_pages - count number of pages beyond high watermark
3021 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3022 * watermark within ZONE_DMA and ZONE_NORMAL.
3024 unsigned long nr_free_buffer_pages(void)
3026 return nr_free_zone_pages(gfp_zone(GFP_USER));
3028 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3031 * nr_free_pagecache_pages - count number of pages beyond high watermark
3033 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3034 * high watermark within all zones.
3036 unsigned long nr_free_pagecache_pages(void)
3038 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3041 static inline void show_node(struct zone *zone)
3043 if (IS_ENABLED(CONFIG_NUMA))
3044 printk("Node %d ", zone_to_nid(zone));
3047 void si_meminfo(struct sysinfo *val)
3049 val->totalram = totalram_pages;
3050 val->sharedram = global_page_state(NR_SHMEM);
3051 val->freeram = global_page_state(NR_FREE_PAGES);
3052 val->bufferram = nr_blockdev_pages();
3053 val->totalhigh = totalhigh_pages;
3054 val->freehigh = nr_free_highpages();
3055 val->mem_unit = PAGE_SIZE;
3058 EXPORT_SYMBOL(si_meminfo);
3061 void si_meminfo_node(struct sysinfo *val, int nid)
3063 int zone_type; /* needs to be signed */
3064 unsigned long managed_pages = 0;
3065 pg_data_t *pgdat = NODE_DATA(nid);
3067 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3068 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3069 val->totalram = managed_pages;
3070 val->sharedram = node_page_state(nid, NR_SHMEM);
3071 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3072 #ifdef CONFIG_HIGHMEM
3073 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3074 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3080 val->mem_unit = PAGE_SIZE;
3085 * Determine whether the node should be displayed or not, depending on whether
3086 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3088 bool skip_free_areas_node(unsigned int flags, int nid)
3091 unsigned int cpuset_mems_cookie;
3093 if (!(flags & SHOW_MEM_FILTER_NODES))
3097 cpuset_mems_cookie = read_mems_allowed_begin();
3098 ret = !node_isset(nid, cpuset_current_mems_allowed);
3099 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3104 #define K(x) ((x) << (PAGE_SHIFT-10))
3106 static void show_migration_types(unsigned char type)
3108 static const char types[MIGRATE_TYPES] = {
3109 [MIGRATE_UNMOVABLE] = 'U',
3110 [MIGRATE_RECLAIMABLE] = 'E',
3111 [MIGRATE_MOVABLE] = 'M',
3112 [MIGRATE_RESERVE] = 'R',
3114 [MIGRATE_CMA] = 'C',
3116 #ifdef CONFIG_MEMORY_ISOLATION
3117 [MIGRATE_ISOLATE] = 'I',
3120 char tmp[MIGRATE_TYPES + 1];
3124 for (i = 0; i < MIGRATE_TYPES; i++) {
3125 if (type & (1 << i))
3130 printk("(%s) ", tmp);
3134 * Show free area list (used inside shift_scroll-lock stuff)
3135 * We also calculate the percentage fragmentation. We do this by counting the
3136 * memory on each free list with the exception of the first item on the list.
3137 * Suppresses nodes that are not allowed by current's cpuset if
3138 * SHOW_MEM_FILTER_NODES is passed.
3140 void show_free_areas(unsigned int filter)
3145 for_each_populated_zone(zone) {
3146 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3149 printk("%s per-cpu:\n", zone->name);
3151 for_each_online_cpu(cpu) {
3152 struct per_cpu_pageset *pageset;
3154 pageset = per_cpu_ptr(zone->pageset, cpu);
3156 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3157 cpu, pageset->pcp.high,
3158 pageset->pcp.batch, pageset->pcp.count);
3162 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3163 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3165 " dirty:%lu writeback:%lu unstable:%lu\n"
3166 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3167 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3169 global_page_state(NR_ACTIVE_ANON),
3170 global_page_state(NR_INACTIVE_ANON),
3171 global_page_state(NR_ISOLATED_ANON),
3172 global_page_state(NR_ACTIVE_FILE),
3173 global_page_state(NR_INACTIVE_FILE),
3174 global_page_state(NR_ISOLATED_FILE),
3175 global_page_state(NR_UNEVICTABLE),
3176 global_page_state(NR_FILE_DIRTY),
3177 global_page_state(NR_WRITEBACK),
3178 global_page_state(NR_UNSTABLE_NFS),
3179 global_page_state(NR_FREE_PAGES),
3180 global_page_state(NR_SLAB_RECLAIMABLE),
3181 global_page_state(NR_SLAB_UNRECLAIMABLE),
3182 global_page_state(NR_FILE_MAPPED),
3183 global_page_state(NR_SHMEM),
3184 global_page_state(NR_PAGETABLE),
3185 global_page_state(NR_BOUNCE),
3186 global_page_state(NR_FREE_CMA_PAGES));
3188 for_each_populated_zone(zone) {
3191 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3199 " active_anon:%lukB"
3200 " inactive_anon:%lukB"
3201 " active_file:%lukB"
3202 " inactive_file:%lukB"
3203 " unevictable:%lukB"
3204 " isolated(anon):%lukB"
3205 " isolated(file):%lukB"
3213 " slab_reclaimable:%lukB"
3214 " slab_unreclaimable:%lukB"
3215 " kernel_stack:%lukB"
3220 " writeback_tmp:%lukB"
3221 " pages_scanned:%lu"
3222 " all_unreclaimable? %s"
3225 K(zone_page_state(zone, NR_FREE_PAGES)),
3226 K(min_wmark_pages(zone)),
3227 K(low_wmark_pages(zone)),
3228 K(high_wmark_pages(zone)),
3229 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3230 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3231 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3232 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3233 K(zone_page_state(zone, NR_UNEVICTABLE)),
3234 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3235 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3236 K(zone->present_pages),
3237 K(zone->managed_pages),
3238 K(zone_page_state(zone, NR_MLOCK)),
3239 K(zone_page_state(zone, NR_FILE_DIRTY)),
3240 K(zone_page_state(zone, NR_WRITEBACK)),
3241 K(zone_page_state(zone, NR_FILE_MAPPED)),
3242 K(zone_page_state(zone, NR_SHMEM)),
3243 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3244 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3245 zone_page_state(zone, NR_KERNEL_STACK) *
3247 K(zone_page_state(zone, NR_PAGETABLE)),
3248 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3249 K(zone_page_state(zone, NR_BOUNCE)),
3250 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3251 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3252 zone->pages_scanned,
3253 (!zone_reclaimable(zone) ? "yes" : "no")
3255 printk("lowmem_reserve[]:");
3256 for (i = 0; i < MAX_NR_ZONES; i++)
3257 printk(" %lu", zone->lowmem_reserve[i]);
3261 for_each_populated_zone(zone) {
3262 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3263 unsigned char types[MAX_ORDER];
3265 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3268 printk("%s: ", zone->name);
3270 spin_lock_irqsave(&zone->lock, flags);
3271 for (order = 0; order < MAX_ORDER; order++) {
3272 struct free_area *area = &zone->free_area[order];
3275 nr[order] = area->nr_free;
3276 total += nr[order] << order;
3279 for (type = 0; type < MIGRATE_TYPES; type++) {
3280 if (!list_empty(&area->free_list[type]))
3281 types[order] |= 1 << type;
3284 spin_unlock_irqrestore(&zone->lock, flags);
3285 for (order = 0; order < MAX_ORDER; order++) {
3286 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3288 show_migration_types(types[order]);
3290 printk("= %lukB\n", K(total));
3293 hugetlb_show_meminfo();
3295 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3297 show_swap_cache_info();
3300 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3302 zoneref->zone = zone;
3303 zoneref->zone_idx = zone_idx(zone);
3307 * Builds allocation fallback zone lists.
3309 * Add all populated zones of a node to the zonelist.
3311 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3315 enum zone_type zone_type = MAX_NR_ZONES;
3319 zone = pgdat->node_zones + zone_type;
3320 if (populated_zone(zone)) {
3321 zoneref_set_zone(zone,
3322 &zonelist->_zonerefs[nr_zones++]);
3323 check_highest_zone(zone_type);
3325 } while (zone_type);
3333 * 0 = automatic detection of better ordering.
3334 * 1 = order by ([node] distance, -zonetype)
3335 * 2 = order by (-zonetype, [node] distance)
3337 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3338 * the same zonelist. So only NUMA can configure this param.
3340 #define ZONELIST_ORDER_DEFAULT 0
3341 #define ZONELIST_ORDER_NODE 1
3342 #define ZONELIST_ORDER_ZONE 2
3344 /* zonelist order in the kernel.
3345 * set_zonelist_order() will set this to NODE or ZONE.
3347 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3348 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3352 /* The value user specified ....changed by config */
3353 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3354 /* string for sysctl */
3355 #define NUMA_ZONELIST_ORDER_LEN 16
3356 char numa_zonelist_order[16] = "default";
3359 * interface for configure zonelist ordering.
3360 * command line option "numa_zonelist_order"
3361 * = "[dD]efault - default, automatic configuration.
3362 * = "[nN]ode - order by node locality, then by zone within node
3363 * = "[zZ]one - order by zone, then by locality within zone
3366 static int __parse_numa_zonelist_order(char *s)
3368 if (*s == 'd' || *s == 'D') {
3369 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3370 } else if (*s == 'n' || *s == 'N') {
3371 user_zonelist_order = ZONELIST_ORDER_NODE;
3372 } else if (*s == 'z' || *s == 'Z') {
3373 user_zonelist_order = ZONELIST_ORDER_ZONE;
3376 "Ignoring invalid numa_zonelist_order value: "
3383 static __init int setup_numa_zonelist_order(char *s)
3390 ret = __parse_numa_zonelist_order(s);
3392 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3396 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3399 * sysctl handler for numa_zonelist_order
3401 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3402 void __user *buffer, size_t *length,
3405 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3407 static DEFINE_MUTEX(zl_order_mutex);
3409 mutex_lock(&zl_order_mutex);
3411 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3415 strcpy(saved_string, (char *)table->data);
3417 ret = proc_dostring(table, write, buffer, length, ppos);
3421 int oldval = user_zonelist_order;
3423 ret = __parse_numa_zonelist_order((char *)table->data);
3426 * bogus value. restore saved string
3428 strncpy((char *)table->data, saved_string,
3429 NUMA_ZONELIST_ORDER_LEN);
3430 user_zonelist_order = oldval;
3431 } else if (oldval != user_zonelist_order) {
3432 mutex_lock(&zonelists_mutex);
3433 build_all_zonelists(NULL, NULL);
3434 mutex_unlock(&zonelists_mutex);
3438 mutex_unlock(&zl_order_mutex);
3443 #define MAX_NODE_LOAD (nr_online_nodes)
3444 static int node_load[MAX_NUMNODES];
3447 * find_next_best_node - find the next node that should appear in a given node's fallback list
3448 * @node: node whose fallback list we're appending
3449 * @used_node_mask: nodemask_t of already used nodes
3451 * We use a number of factors to determine which is the next node that should
3452 * appear on a given node's fallback list. The node should not have appeared
3453 * already in @node's fallback list, and it should be the next closest node
3454 * according to the distance array (which contains arbitrary distance values
3455 * from each node to each node in the system), and should also prefer nodes
3456 * with no CPUs, since presumably they'll have very little allocation pressure
3457 * on them otherwise.
3458 * It returns -1 if no node is found.
3460 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3463 int min_val = INT_MAX;
3464 int best_node = NUMA_NO_NODE;
3465 const struct cpumask *tmp = cpumask_of_node(0);
3467 /* Use the local node if we haven't already */
3468 if (!node_isset(node, *used_node_mask)) {
3469 node_set(node, *used_node_mask);
3473 for_each_node_state(n, N_MEMORY) {
3475 /* Don't want a node to appear more than once */
3476 if (node_isset(n, *used_node_mask))
3479 /* Use the distance array to find the distance */
3480 val = node_distance(node, n);
3482 /* Penalize nodes under us ("prefer the next node") */
3485 /* Give preference to headless and unused nodes */
3486 tmp = cpumask_of_node(n);
3487 if (!cpumask_empty(tmp))
3488 val += PENALTY_FOR_NODE_WITH_CPUS;
3490 /* Slight preference for less loaded node */
3491 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3492 val += node_load[n];
3494 if (val < min_val) {
3501 node_set(best_node, *used_node_mask);
3508 * Build zonelists ordered by node and zones within node.
3509 * This results in maximum locality--normal zone overflows into local
3510 * DMA zone, if any--but risks exhausting DMA zone.
3512 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3515 struct zonelist *zonelist;
3517 zonelist = &pgdat->node_zonelists[0];
3518 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3520 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3521 zonelist->_zonerefs[j].zone = NULL;
3522 zonelist->_zonerefs[j].zone_idx = 0;
3526 * Build gfp_thisnode zonelists
3528 static void build_thisnode_zonelists(pg_data_t *pgdat)
3531 struct zonelist *zonelist;
3533 zonelist = &pgdat->node_zonelists[1];
3534 j = build_zonelists_node(pgdat, zonelist, 0);
3535 zonelist->_zonerefs[j].zone = NULL;
3536 zonelist->_zonerefs[j].zone_idx = 0;
3540 * Build zonelists ordered by zone and nodes within zones.
3541 * This results in conserving DMA zone[s] until all Normal memory is
3542 * exhausted, but results in overflowing to remote node while memory
3543 * may still exist in local DMA zone.
3545 static int node_order[MAX_NUMNODES];
3547 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3550 int zone_type; /* needs to be signed */
3552 struct zonelist *zonelist;
3554 zonelist = &pgdat->node_zonelists[0];
3556 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3557 for (j = 0; j < nr_nodes; j++) {
3558 node = node_order[j];
3559 z = &NODE_DATA(node)->node_zones[zone_type];
3560 if (populated_zone(z)) {
3562 &zonelist->_zonerefs[pos++]);
3563 check_highest_zone(zone_type);
3567 zonelist->_zonerefs[pos].zone = NULL;
3568 zonelist->_zonerefs[pos].zone_idx = 0;
3571 static int default_zonelist_order(void)
3574 unsigned long low_kmem_size, total_size;
3578 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3579 * If they are really small and used heavily, the system can fall
3580 * into OOM very easily.
3581 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3583 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3586 for_each_online_node(nid) {
3587 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3588 z = &NODE_DATA(nid)->node_zones[zone_type];
3589 if (populated_zone(z)) {
3590 if (zone_type < ZONE_NORMAL)
3591 low_kmem_size += z->managed_pages;
3592 total_size += z->managed_pages;
3593 } else if (zone_type == ZONE_NORMAL) {
3595 * If any node has only lowmem, then node order
3596 * is preferred to allow kernel allocations
3597 * locally; otherwise, they can easily infringe
3598 * on other nodes when there is an abundance of
3599 * lowmem available to allocate from.
3601 return ZONELIST_ORDER_NODE;
3605 if (!low_kmem_size || /* there are no DMA area. */
3606 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3607 return ZONELIST_ORDER_NODE;
3609 * look into each node's config.
3610 * If there is a node whose DMA/DMA32 memory is very big area on
3611 * local memory, NODE_ORDER may be suitable.
3613 average_size = total_size /
3614 (nodes_weight(node_states[N_MEMORY]) + 1);
3615 for_each_online_node(nid) {
3618 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3619 z = &NODE_DATA(nid)->node_zones[zone_type];
3620 if (populated_zone(z)) {
3621 if (zone_type < ZONE_NORMAL)
3622 low_kmem_size += z->present_pages;
3623 total_size += z->present_pages;
3626 if (low_kmem_size &&
3627 total_size > average_size && /* ignore small node */
3628 low_kmem_size > total_size * 70/100)
3629 return ZONELIST_ORDER_NODE;
3631 return ZONELIST_ORDER_ZONE;
3634 static void set_zonelist_order(void)
3636 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3637 current_zonelist_order = default_zonelist_order();
3639 current_zonelist_order = user_zonelist_order;
3642 static void build_zonelists(pg_data_t *pgdat)
3646 nodemask_t used_mask;
3647 int local_node, prev_node;
3648 struct zonelist *zonelist;
3649 int order = current_zonelist_order;
3651 /* initialize zonelists */
3652 for (i = 0; i < MAX_ZONELISTS; i++) {
3653 zonelist = pgdat->node_zonelists + i;
3654 zonelist->_zonerefs[0].zone = NULL;
3655 zonelist->_zonerefs[0].zone_idx = 0;
3658 /* NUMA-aware ordering of nodes */
3659 local_node = pgdat->node_id;
3660 load = nr_online_nodes;
3661 prev_node = local_node;
3662 nodes_clear(used_mask);
3664 memset(node_order, 0, sizeof(node_order));
3667 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3669 * We don't want to pressure a particular node.
3670 * So adding penalty to the first node in same
3671 * distance group to make it round-robin.
3673 if (node_distance(local_node, node) !=
3674 node_distance(local_node, prev_node))
3675 node_load[node] = load;
3679 if (order == ZONELIST_ORDER_NODE)
3680 build_zonelists_in_node_order(pgdat, node);
3682 node_order[j++] = node; /* remember order */
3685 if (order == ZONELIST_ORDER_ZONE) {
3686 /* calculate node order -- i.e., DMA last! */
3687 build_zonelists_in_zone_order(pgdat, j);
3690 build_thisnode_zonelists(pgdat);
3693 /* Construct the zonelist performance cache - see further mmzone.h */
3694 static void build_zonelist_cache(pg_data_t *pgdat)
3696 struct zonelist *zonelist;
3697 struct zonelist_cache *zlc;
3700 zonelist = &pgdat->node_zonelists[0];
3701 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3702 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3703 for (z = zonelist->_zonerefs; z->zone; z++)
3704 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3707 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3709 * Return node id of node used for "local" allocations.
3710 * I.e., first node id of first zone in arg node's generic zonelist.
3711 * Used for initializing percpu 'numa_mem', which is used primarily
3712 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3714 int local_memory_node(int node)
3718 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3719 gfp_zone(GFP_KERNEL),
3726 #else /* CONFIG_NUMA */
3728 static void set_zonelist_order(void)
3730 current_zonelist_order = ZONELIST_ORDER_ZONE;
3733 static void build_zonelists(pg_data_t *pgdat)
3735 int node, local_node;
3737 struct zonelist *zonelist;
3739 local_node = pgdat->node_id;
3741 zonelist = &pgdat->node_zonelists[0];
3742 j = build_zonelists_node(pgdat, zonelist, 0);
3745 * Now we build the zonelist so that it contains the zones
3746 * of all the other nodes.
3747 * We don't want to pressure a particular node, so when
3748 * building the zones for node N, we make sure that the
3749 * zones coming right after the local ones are those from
3750 * node N+1 (modulo N)
3752 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3753 if (!node_online(node))
3755 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3757 for (node = 0; node < local_node; node++) {
3758 if (!node_online(node))
3760 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3763 zonelist->_zonerefs[j].zone = NULL;
3764 zonelist->_zonerefs[j].zone_idx = 0;
3767 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3768 static void build_zonelist_cache(pg_data_t *pgdat)
3770 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3773 #endif /* CONFIG_NUMA */
3776 * Boot pageset table. One per cpu which is going to be used for all
3777 * zones and all nodes. The parameters will be set in such a way
3778 * that an item put on a list will immediately be handed over to
3779 * the buddy list. This is safe since pageset manipulation is done
3780 * with interrupts disabled.
3782 * The boot_pagesets must be kept even after bootup is complete for
3783 * unused processors and/or zones. They do play a role for bootstrapping
3784 * hotplugged processors.
3786 * zoneinfo_show() and maybe other functions do
3787 * not check if the processor is online before following the pageset pointer.
3788 * Other parts of the kernel may not check if the zone is available.
3790 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3791 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3792 static void setup_zone_pageset(struct zone *zone);
3795 * Global mutex to protect against size modification of zonelists
3796 * as well as to serialize pageset setup for the new populated zone.
3798 DEFINE_MUTEX(zonelists_mutex);
3800 /* return values int ....just for stop_machine() */
3801 static int __build_all_zonelists(void *data)
3805 pg_data_t *self = data;
3808 memset(node_load, 0, sizeof(node_load));
3811 if (self && !node_online(self->node_id)) {
3812 build_zonelists(self);
3813 build_zonelist_cache(self);
3816 for_each_online_node(nid) {
3817 pg_data_t *pgdat = NODE_DATA(nid);
3819 build_zonelists(pgdat);
3820 build_zonelist_cache(pgdat);
3824 * Initialize the boot_pagesets that are going to be used
3825 * for bootstrapping processors. The real pagesets for
3826 * each zone will be allocated later when the per cpu
3827 * allocator is available.
3829 * boot_pagesets are used also for bootstrapping offline
3830 * cpus if the system is already booted because the pagesets
3831 * are needed to initialize allocators on a specific cpu too.
3832 * F.e. the percpu allocator needs the page allocator which
3833 * needs the percpu allocator in order to allocate its pagesets
3834 * (a chicken-egg dilemma).
3836 for_each_possible_cpu(cpu) {
3837 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3839 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3841 * We now know the "local memory node" for each node--
3842 * i.e., the node of the first zone in the generic zonelist.
3843 * Set up numa_mem percpu variable for on-line cpus. During
3844 * boot, only the boot cpu should be on-line; we'll init the
3845 * secondary cpus' numa_mem as they come on-line. During
3846 * node/memory hotplug, we'll fixup all on-line cpus.
3848 if (cpu_online(cpu))
3849 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3857 * Called with zonelists_mutex held always
3858 * unless system_state == SYSTEM_BOOTING.
3860 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3862 set_zonelist_order();
3864 if (system_state == SYSTEM_BOOTING) {
3865 __build_all_zonelists(NULL);
3866 mminit_verify_zonelist();
3867 cpuset_init_current_mems_allowed();
3869 #ifdef CONFIG_MEMORY_HOTPLUG
3871 setup_zone_pageset(zone);
3873 /* we have to stop all cpus to guarantee there is no user
3875 stop_machine(__build_all_zonelists, pgdat, NULL);
3876 /* cpuset refresh routine should be here */
3878 vm_total_pages = nr_free_pagecache_pages();
3880 * Disable grouping by mobility if the number of pages in the
3881 * system is too low to allow the mechanism to work. It would be
3882 * more accurate, but expensive to check per-zone. This check is
3883 * made on memory-hotadd so a system can start with mobility
3884 * disabled and enable it later
3886 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3887 page_group_by_mobility_disabled = 1;
3889 page_group_by_mobility_disabled = 0;
3891 printk("Built %i zonelists in %s order, mobility grouping %s. "
3892 "Total pages: %ld\n",
3894 zonelist_order_name[current_zonelist_order],
3895 page_group_by_mobility_disabled ? "off" : "on",
3898 printk("Policy zone: %s\n", zone_names[policy_zone]);
3903 * Helper functions to size the waitqueue hash table.
3904 * Essentially these want to choose hash table sizes sufficiently
3905 * large so that collisions trying to wait on pages are rare.
3906 * But in fact, the number of active page waitqueues on typical
3907 * systems is ridiculously low, less than 200. So this is even
3908 * conservative, even though it seems large.
3910 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3911 * waitqueues, i.e. the size of the waitq table given the number of pages.
3913 #define PAGES_PER_WAITQUEUE 256
3915 #ifndef CONFIG_MEMORY_HOTPLUG
3916 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3918 unsigned long size = 1;
3920 pages /= PAGES_PER_WAITQUEUE;
3922 while (size < pages)
3926 * Once we have dozens or even hundreds of threads sleeping
3927 * on IO we've got bigger problems than wait queue collision.
3928 * Limit the size of the wait table to a reasonable size.
3930 size = min(size, 4096UL);
3932 return max(size, 4UL);
3936 * A zone's size might be changed by hot-add, so it is not possible to determine
3937 * a suitable size for its wait_table. So we use the maximum size now.
3939 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3941 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3942 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3943 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3945 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3946 * or more by the traditional way. (See above). It equals:
3948 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3949 * ia64(16K page size) : = ( 8G + 4M)byte.
3950 * powerpc (64K page size) : = (32G +16M)byte.
3952 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3959 * This is an integer logarithm so that shifts can be used later
3960 * to extract the more random high bits from the multiplicative
3961 * hash function before the remainder is taken.
3963 static inline unsigned long wait_table_bits(unsigned long size)
3969 * Check if a pageblock contains reserved pages
3971 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3975 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3976 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3983 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3984 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3985 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3986 * higher will lead to a bigger reserve which will get freed as contiguous
3987 * blocks as reclaim kicks in
3989 static void setup_zone_migrate_reserve(struct zone *zone)
3991 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3993 unsigned long block_migratetype;
3998 * Get the start pfn, end pfn and the number of blocks to reserve
3999 * We have to be careful to be aligned to pageblock_nr_pages to
4000 * make sure that we always check pfn_valid for the first page in
4003 start_pfn = zone->zone_start_pfn;
4004 end_pfn = zone_end_pfn(zone);
4005 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4006 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4010 * Reserve blocks are generally in place to help high-order atomic
4011 * allocations that are short-lived. A min_free_kbytes value that
4012 * would result in more than 2 reserve blocks for atomic allocations
4013 * is assumed to be in place to help anti-fragmentation for the
4014 * future allocation of hugepages at runtime.
4016 reserve = min(2, reserve);
4017 old_reserve = zone->nr_migrate_reserve_block;
4019 /* When memory hot-add, we almost always need to do nothing */
4020 if (reserve == old_reserve)
4022 zone->nr_migrate_reserve_block = reserve;
4024 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4025 if (!pfn_valid(pfn))
4027 page = pfn_to_page(pfn);
4029 /* Watch out for overlapping nodes */
4030 if (page_to_nid(page) != zone_to_nid(zone))
4033 block_migratetype = get_pageblock_migratetype(page);
4035 /* Only test what is necessary when the reserves are not met */
4038 * Blocks with reserved pages will never free, skip
4041 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4042 if (pageblock_is_reserved(pfn, block_end_pfn))
4045 /* If this block is reserved, account for it */
4046 if (block_migratetype == MIGRATE_RESERVE) {
4051 /* Suitable for reserving if this block is movable */
4052 if (block_migratetype == MIGRATE_MOVABLE) {
4053 set_pageblock_migratetype(page,
4055 move_freepages_block(zone, page,
4060 } else if (!old_reserve) {
4062 * At boot time we don't need to scan the whole zone
4063 * for turning off MIGRATE_RESERVE.
4069 * If the reserve is met and this is a previous reserved block,
4072 if (block_migratetype == MIGRATE_RESERVE) {
4073 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4074 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4080 * Initially all pages are reserved - free ones are freed
4081 * up by free_all_bootmem() once the early boot process is
4082 * done. Non-atomic initialization, single-pass.
4084 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4085 unsigned long start_pfn, enum memmap_context context)
4088 unsigned long end_pfn = start_pfn + size;
4092 if (highest_memmap_pfn < end_pfn - 1)
4093 highest_memmap_pfn = end_pfn - 1;
4095 z = &NODE_DATA(nid)->node_zones[zone];
4096 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4098 * There can be holes in boot-time mem_map[]s
4099 * handed to this function. They do not
4100 * exist on hotplugged memory.
4102 if (context == MEMMAP_EARLY) {
4103 if (!early_pfn_valid(pfn))
4105 if (!early_pfn_in_nid(pfn, nid))
4108 page = pfn_to_page(pfn);
4109 set_page_links(page, zone, nid, pfn);
4110 mminit_verify_page_links(page, zone, nid, pfn);
4111 init_page_count(page);
4112 page_mapcount_reset(page);
4113 page_cpupid_reset_last(page);
4114 SetPageReserved(page);
4116 * Mark the block movable so that blocks are reserved for
4117 * movable at startup. This will force kernel allocations
4118 * to reserve their blocks rather than leaking throughout
4119 * the address space during boot when many long-lived
4120 * kernel allocations are made. Later some blocks near
4121 * the start are marked MIGRATE_RESERVE by
4122 * setup_zone_migrate_reserve()
4124 * bitmap is created for zone's valid pfn range. but memmap
4125 * can be created for invalid pages (for alignment)
4126 * check here not to call set_pageblock_migratetype() against
4129 if ((z->zone_start_pfn <= pfn)
4130 && (pfn < zone_end_pfn(z))
4131 && !(pfn & (pageblock_nr_pages - 1)))
4132 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4134 INIT_LIST_HEAD(&page->lru);
4135 #ifdef WANT_PAGE_VIRTUAL
4136 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4137 if (!is_highmem_idx(zone))
4138 set_page_address(page, __va(pfn << PAGE_SHIFT));
4143 static void __meminit zone_init_free_lists(struct zone *zone)
4145 unsigned int order, t;
4146 for_each_migratetype_order(order, t) {
4147 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4148 zone->free_area[order].nr_free = 0;
4152 #ifndef __HAVE_ARCH_MEMMAP_INIT
4153 #define memmap_init(size, nid, zone, start_pfn) \
4154 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4157 static int zone_batchsize(struct zone *zone)
4163 * The per-cpu-pages pools are set to around 1000th of the
4164 * size of the zone. But no more than 1/2 of a meg.
4166 * OK, so we don't know how big the cache is. So guess.
4168 batch = zone->managed_pages / 1024;
4169 if (batch * PAGE_SIZE > 512 * 1024)
4170 batch = (512 * 1024) / PAGE_SIZE;
4171 batch /= 4; /* We effectively *= 4 below */
4176 * Clamp the batch to a 2^n - 1 value. Having a power
4177 * of 2 value was found to be more likely to have
4178 * suboptimal cache aliasing properties in some cases.
4180 * For example if 2 tasks are alternately allocating
4181 * batches of pages, one task can end up with a lot
4182 * of pages of one half of the possible page colors
4183 * and the other with pages of the other colors.
4185 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4190 /* The deferral and batching of frees should be suppressed under NOMMU
4193 * The problem is that NOMMU needs to be able to allocate large chunks
4194 * of contiguous memory as there's no hardware page translation to
4195 * assemble apparent contiguous memory from discontiguous pages.
4197 * Queueing large contiguous runs of pages for batching, however,
4198 * causes the pages to actually be freed in smaller chunks. As there
4199 * can be a significant delay between the individual batches being
4200 * recycled, this leads to the once large chunks of space being
4201 * fragmented and becoming unavailable for high-order allocations.
4208 * pcp->high and pcp->batch values are related and dependent on one another:
4209 * ->batch must never be higher then ->high.
4210 * The following function updates them in a safe manner without read side
4213 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4214 * those fields changing asynchronously (acording the the above rule).
4216 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4217 * outside of boot time (or some other assurance that no concurrent updaters
4220 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4221 unsigned long batch)
4223 /* start with a fail safe value for batch */
4227 /* Update high, then batch, in order */
4234 /* a companion to pageset_set_high() */
4235 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4237 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4240 static void pageset_init(struct per_cpu_pageset *p)
4242 struct per_cpu_pages *pcp;
4245 memset(p, 0, sizeof(*p));
4249 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4250 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4253 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4256 pageset_set_batch(p, batch);
4260 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4261 * to the value high for the pageset p.
4263 static void pageset_set_high(struct per_cpu_pageset *p,
4266 unsigned long batch = max(1UL, high / 4);
4267 if ((high / 4) > (PAGE_SHIFT * 8))
4268 batch = PAGE_SHIFT * 8;
4270 pageset_update(&p->pcp, high, batch);
4273 static void pageset_set_high_and_batch(struct zone *zone,
4274 struct per_cpu_pageset *pcp)
4276 if (percpu_pagelist_fraction)
4277 pageset_set_high(pcp,
4278 (zone->managed_pages /
4279 percpu_pagelist_fraction));
4281 pageset_set_batch(pcp, zone_batchsize(zone));
4284 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4286 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4289 pageset_set_high_and_batch(zone, pcp);
4292 static void __meminit setup_zone_pageset(struct zone *zone)
4295 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4296 for_each_possible_cpu(cpu)
4297 zone_pageset_init(zone, cpu);
4301 * Allocate per cpu pagesets and initialize them.
4302 * Before this call only boot pagesets were available.
4304 void __init setup_per_cpu_pageset(void)
4308 for_each_populated_zone(zone)
4309 setup_zone_pageset(zone);
4312 static noinline __init_refok
4313 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4319 * The per-page waitqueue mechanism uses hashed waitqueues
4322 zone->wait_table_hash_nr_entries =
4323 wait_table_hash_nr_entries(zone_size_pages);
4324 zone->wait_table_bits =
4325 wait_table_bits(zone->wait_table_hash_nr_entries);
4326 alloc_size = zone->wait_table_hash_nr_entries
4327 * sizeof(wait_queue_head_t);
4329 if (!slab_is_available()) {
4330 zone->wait_table = (wait_queue_head_t *)
4331 memblock_virt_alloc_node_nopanic(
4332 alloc_size, zone->zone_pgdat->node_id);
4335 * This case means that a zone whose size was 0 gets new memory
4336 * via memory hot-add.
4337 * But it may be the case that a new node was hot-added. In
4338 * this case vmalloc() will not be able to use this new node's
4339 * memory - this wait_table must be initialized to use this new
4340 * node itself as well.
4341 * To use this new node's memory, further consideration will be
4344 zone->wait_table = vmalloc(alloc_size);
4346 if (!zone->wait_table)
4349 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4350 init_waitqueue_head(zone->wait_table + i);
4355 static __meminit void zone_pcp_init(struct zone *zone)
4358 * per cpu subsystem is not up at this point. The following code
4359 * relies on the ability of the linker to provide the
4360 * offset of a (static) per cpu variable into the per cpu area.
4362 zone->pageset = &boot_pageset;
4364 if (populated_zone(zone))
4365 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4366 zone->name, zone->present_pages,
4367 zone_batchsize(zone));
4370 int __meminit init_currently_empty_zone(struct zone *zone,
4371 unsigned long zone_start_pfn,
4373 enum memmap_context context)
4375 struct pglist_data *pgdat = zone->zone_pgdat;
4377 ret = zone_wait_table_init(zone, size);
4380 pgdat->nr_zones = zone_idx(zone) + 1;
4382 zone->zone_start_pfn = zone_start_pfn;
4384 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4385 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4387 (unsigned long)zone_idx(zone),
4388 zone_start_pfn, (zone_start_pfn + size));
4390 zone_init_free_lists(zone);
4395 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4396 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4398 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4400 int __meminit __early_pfn_to_nid(unsigned long pfn)
4402 unsigned long start_pfn, end_pfn;
4405 * NOTE: The following SMP-unsafe globals are only used early in boot
4406 * when the kernel is running single-threaded.
4408 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4409 static int __meminitdata last_nid;
4411 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4414 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4416 last_start_pfn = start_pfn;
4417 last_end_pfn = end_pfn;
4423 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4425 int __meminit early_pfn_to_nid(unsigned long pfn)
4429 nid = __early_pfn_to_nid(pfn);
4432 /* just returns 0 */
4436 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4437 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4441 nid = __early_pfn_to_nid(pfn);
4442 if (nid >= 0 && nid != node)
4449 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4450 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4451 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4453 * If an architecture guarantees that all ranges registered contain no holes
4454 * and may be freed, this this function may be used instead of calling
4455 * memblock_free_early_nid() manually.
4457 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4459 unsigned long start_pfn, end_pfn;
4462 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4463 start_pfn = min(start_pfn, max_low_pfn);
4464 end_pfn = min(end_pfn, max_low_pfn);
4466 if (start_pfn < end_pfn)
4467 memblock_free_early_nid(PFN_PHYS(start_pfn),
4468 (end_pfn - start_pfn) << PAGE_SHIFT,
4474 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4475 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4477 * If an architecture guarantees that all ranges registered contain no holes and may
4478 * be freed, this function may be used instead of calling memory_present() manually.
4480 void __init sparse_memory_present_with_active_regions(int nid)
4482 unsigned long start_pfn, end_pfn;
4485 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4486 memory_present(this_nid, start_pfn, end_pfn);
4490 * get_pfn_range_for_nid - Return the start and end page frames for a node
4491 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4492 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4493 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4495 * It returns the start and end page frame of a node based on information
4496 * provided by memblock_set_node(). If called for a node
4497 * with no available memory, a warning is printed and the start and end
4500 void __meminit get_pfn_range_for_nid(unsigned int nid,
4501 unsigned long *start_pfn, unsigned long *end_pfn)
4503 unsigned long this_start_pfn, this_end_pfn;
4509 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4510 *start_pfn = min(*start_pfn, this_start_pfn);
4511 *end_pfn = max(*end_pfn, this_end_pfn);
4514 if (*start_pfn == -1UL)
4519 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4520 * assumption is made that zones within a node are ordered in monotonic
4521 * increasing memory addresses so that the "highest" populated zone is used
4523 static void __init find_usable_zone_for_movable(void)
4526 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4527 if (zone_index == ZONE_MOVABLE)
4530 if (arch_zone_highest_possible_pfn[zone_index] >
4531 arch_zone_lowest_possible_pfn[zone_index])
4535 VM_BUG_ON(zone_index == -1);
4536 movable_zone = zone_index;
4540 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4541 * because it is sized independent of architecture. Unlike the other zones,
4542 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4543 * in each node depending on the size of each node and how evenly kernelcore
4544 * is distributed. This helper function adjusts the zone ranges
4545 * provided by the architecture for a given node by using the end of the
4546 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4547 * zones within a node are in order of monotonic increases memory addresses
4549 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4550 unsigned long zone_type,
4551 unsigned long node_start_pfn,
4552 unsigned long node_end_pfn,
4553 unsigned long *zone_start_pfn,
4554 unsigned long *zone_end_pfn)
4556 /* Only adjust if ZONE_MOVABLE is on this node */
4557 if (zone_movable_pfn[nid]) {
4558 /* Size ZONE_MOVABLE */
4559 if (zone_type == ZONE_MOVABLE) {
4560 *zone_start_pfn = zone_movable_pfn[nid];
4561 *zone_end_pfn = min(node_end_pfn,
4562 arch_zone_highest_possible_pfn[movable_zone]);
4564 /* Adjust for ZONE_MOVABLE starting within this range */
4565 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4566 *zone_end_pfn > zone_movable_pfn[nid]) {
4567 *zone_end_pfn = zone_movable_pfn[nid];
4569 /* Check if this whole range is within ZONE_MOVABLE */
4570 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4571 *zone_start_pfn = *zone_end_pfn;
4576 * Return the number of pages a zone spans in a node, including holes
4577 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4579 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4580 unsigned long zone_type,
4581 unsigned long node_start_pfn,
4582 unsigned long node_end_pfn,
4583 unsigned long *ignored)
4585 unsigned long zone_start_pfn, zone_end_pfn;
4587 /* Get the start and end of the zone */
4588 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4589 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4590 adjust_zone_range_for_zone_movable(nid, zone_type,
4591 node_start_pfn, node_end_pfn,
4592 &zone_start_pfn, &zone_end_pfn);
4594 /* Check that this node has pages within the zone's required range */
4595 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4598 /* Move the zone boundaries inside the node if necessary */
4599 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4600 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4602 /* Return the spanned pages */
4603 return zone_end_pfn - zone_start_pfn;
4607 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4608 * then all holes in the requested range will be accounted for.
4610 unsigned long __meminit __absent_pages_in_range(int nid,
4611 unsigned long range_start_pfn,
4612 unsigned long range_end_pfn)
4614 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4615 unsigned long start_pfn, end_pfn;
4618 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4619 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4620 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4621 nr_absent -= end_pfn - start_pfn;
4627 * absent_pages_in_range - Return number of page frames in holes within a range
4628 * @start_pfn: The start PFN to start searching for holes
4629 * @end_pfn: The end PFN to stop searching for holes
4631 * It returns the number of pages frames in memory holes within a range.
4633 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4634 unsigned long end_pfn)
4636 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4639 /* Return the number of page frames in holes in a zone on a node */
4640 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4641 unsigned long zone_type,
4642 unsigned long node_start_pfn,
4643 unsigned long node_end_pfn,
4644 unsigned long *ignored)
4646 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4647 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4648 unsigned long zone_start_pfn, zone_end_pfn;
4650 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4651 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4653 adjust_zone_range_for_zone_movable(nid, zone_type,
4654 node_start_pfn, node_end_pfn,
4655 &zone_start_pfn, &zone_end_pfn);
4656 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4659 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4660 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4661 unsigned long zone_type,
4662 unsigned long node_start_pfn,
4663 unsigned long node_end_pfn,
4664 unsigned long *zones_size)
4666 return zones_size[zone_type];
4669 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4670 unsigned long zone_type,
4671 unsigned long node_start_pfn,
4672 unsigned long node_end_pfn,
4673 unsigned long *zholes_size)
4678 return zholes_size[zone_type];
4681 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4683 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4684 unsigned long node_start_pfn,
4685 unsigned long node_end_pfn,
4686 unsigned long *zones_size,
4687 unsigned long *zholes_size)
4689 unsigned long realtotalpages, totalpages = 0;
4692 for (i = 0; i < MAX_NR_ZONES; i++)
4693 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4697 pgdat->node_spanned_pages = totalpages;
4699 realtotalpages = totalpages;
4700 for (i = 0; i < MAX_NR_ZONES; i++)
4702 zone_absent_pages_in_node(pgdat->node_id, i,
4703 node_start_pfn, node_end_pfn,
4705 pgdat->node_present_pages = realtotalpages;
4706 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4710 #ifndef CONFIG_SPARSEMEM
4712 * Calculate the size of the zone->blockflags rounded to an unsigned long
4713 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4714 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4715 * round what is now in bits to nearest long in bits, then return it in
4718 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4720 unsigned long usemapsize;
4722 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4723 usemapsize = roundup(zonesize, pageblock_nr_pages);
4724 usemapsize = usemapsize >> pageblock_order;
4725 usemapsize *= NR_PAGEBLOCK_BITS;
4726 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4728 return usemapsize / 8;
4731 static void __init setup_usemap(struct pglist_data *pgdat,
4733 unsigned long zone_start_pfn,
4734 unsigned long zonesize)
4736 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4737 zone->pageblock_flags = NULL;
4739 zone->pageblock_flags =
4740 memblock_virt_alloc_node_nopanic(usemapsize,
4744 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4745 unsigned long zone_start_pfn, unsigned long zonesize) {}
4746 #endif /* CONFIG_SPARSEMEM */
4748 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4750 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4751 void __paginginit set_pageblock_order(void)
4755 /* Check that pageblock_nr_pages has not already been setup */
4756 if (pageblock_order)
4759 if (HPAGE_SHIFT > PAGE_SHIFT)
4760 order = HUGETLB_PAGE_ORDER;
4762 order = MAX_ORDER - 1;
4765 * Assume the largest contiguous order of interest is a huge page.
4766 * This value may be variable depending on boot parameters on IA64 and
4769 pageblock_order = order;
4771 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4774 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4775 * is unused as pageblock_order is set at compile-time. See
4776 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4779 void __paginginit set_pageblock_order(void)
4783 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4785 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4786 unsigned long present_pages)
4788 unsigned long pages = spanned_pages;
4791 * Provide a more accurate estimation if there are holes within
4792 * the zone and SPARSEMEM is in use. If there are holes within the
4793 * zone, each populated memory region may cost us one or two extra
4794 * memmap pages due to alignment because memmap pages for each
4795 * populated regions may not naturally algined on page boundary.
4796 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4798 if (spanned_pages > present_pages + (present_pages >> 4) &&
4799 IS_ENABLED(CONFIG_SPARSEMEM))
4800 pages = present_pages;
4802 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4806 * Set up the zone data structures:
4807 * - mark all pages reserved
4808 * - mark all memory queues empty
4809 * - clear the memory bitmaps
4811 * NOTE: pgdat should get zeroed by caller.
4813 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4814 unsigned long node_start_pfn, unsigned long node_end_pfn,
4815 unsigned long *zones_size, unsigned long *zholes_size)
4818 int nid = pgdat->node_id;
4819 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4822 pgdat_resize_init(pgdat);
4823 #ifdef CONFIG_NUMA_BALANCING
4824 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4825 pgdat->numabalancing_migrate_nr_pages = 0;
4826 pgdat->numabalancing_migrate_next_window = jiffies;
4828 init_waitqueue_head(&pgdat->kswapd_wait);
4829 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4830 pgdat_page_cgroup_init(pgdat);
4832 for (j = 0; j < MAX_NR_ZONES; j++) {
4833 struct zone *zone = pgdat->node_zones + j;
4834 unsigned long size, realsize, freesize, memmap_pages;
4836 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4837 node_end_pfn, zones_size);
4838 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4844 * Adjust freesize so that it accounts for how much memory
4845 * is used by this zone for memmap. This affects the watermark
4846 * and per-cpu initialisations
4848 memmap_pages = calc_memmap_size(size, realsize);
4849 if (freesize >= memmap_pages) {
4850 freesize -= memmap_pages;
4853 " %s zone: %lu pages used for memmap\n",
4854 zone_names[j], memmap_pages);
4857 " %s zone: %lu pages exceeds freesize %lu\n",
4858 zone_names[j], memmap_pages, freesize);
4860 /* Account for reserved pages */
4861 if (j == 0 && freesize > dma_reserve) {
4862 freesize -= dma_reserve;
4863 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4864 zone_names[0], dma_reserve);
4867 if (!is_highmem_idx(j))
4868 nr_kernel_pages += freesize;
4869 /* Charge for highmem memmap if there are enough kernel pages */
4870 else if (nr_kernel_pages > memmap_pages * 2)
4871 nr_kernel_pages -= memmap_pages;
4872 nr_all_pages += freesize;
4874 zone->spanned_pages = size;
4875 zone->present_pages = realsize;
4877 * Set an approximate value for lowmem here, it will be adjusted
4878 * when the bootmem allocator frees pages into the buddy system.
4879 * And all highmem pages will be managed by the buddy system.
4881 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4884 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4886 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4888 zone->name = zone_names[j];
4889 spin_lock_init(&zone->lock);
4890 spin_lock_init(&zone->lru_lock);
4891 zone_seqlock_init(zone);
4892 zone->zone_pgdat = pgdat;
4893 zone_pcp_init(zone);
4895 /* For bootup, initialized properly in watermark setup */
4896 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4898 lruvec_init(&zone->lruvec);
4902 set_pageblock_order();
4903 setup_usemap(pgdat, zone, zone_start_pfn, size);
4904 ret = init_currently_empty_zone(zone, zone_start_pfn,
4905 size, MEMMAP_EARLY);
4907 memmap_init(size, nid, j, zone_start_pfn);
4908 zone_start_pfn += size;
4912 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4914 /* Skip empty nodes */
4915 if (!pgdat->node_spanned_pages)
4918 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4919 /* ia64 gets its own node_mem_map, before this, without bootmem */
4920 if (!pgdat->node_mem_map) {
4921 unsigned long size, start, end;
4925 * The zone's endpoints aren't required to be MAX_ORDER
4926 * aligned but the node_mem_map endpoints must be in order
4927 * for the buddy allocator to function correctly.
4929 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4930 end = pgdat_end_pfn(pgdat);
4931 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4932 size = (end - start) * sizeof(struct page);
4933 map = alloc_remap(pgdat->node_id, size);
4935 map = memblock_virt_alloc_node_nopanic(size,
4937 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4939 #ifndef CONFIG_NEED_MULTIPLE_NODES
4941 * With no DISCONTIG, the global mem_map is just set as node 0's
4943 if (pgdat == NODE_DATA(0)) {
4944 mem_map = NODE_DATA(0)->node_mem_map;
4945 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4946 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4947 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4948 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4951 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4954 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4955 unsigned long node_start_pfn, unsigned long *zholes_size)
4957 pg_data_t *pgdat = NODE_DATA(nid);
4958 unsigned long start_pfn = 0;
4959 unsigned long end_pfn = 0;
4961 /* pg_data_t should be reset to zero when it's allocated */
4962 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4964 pgdat->node_id = nid;
4965 pgdat->node_start_pfn = node_start_pfn;
4966 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4967 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4969 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4970 zones_size, zholes_size);
4972 alloc_node_mem_map(pgdat);
4973 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4974 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4975 nid, (unsigned long)pgdat,
4976 (unsigned long)pgdat->node_mem_map);
4979 free_area_init_core(pgdat, start_pfn, end_pfn,
4980 zones_size, zholes_size);
4983 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4985 #if MAX_NUMNODES > 1
4987 * Figure out the number of possible node ids.
4989 void __init setup_nr_node_ids(void)
4992 unsigned int highest = 0;
4994 for_each_node_mask(node, node_possible_map)
4996 nr_node_ids = highest + 1;
5001 * node_map_pfn_alignment - determine the maximum internode alignment
5003 * This function should be called after node map is populated and sorted.
5004 * It calculates the maximum power of two alignment which can distinguish
5007 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5008 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5009 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5010 * shifted, 1GiB is enough and this function will indicate so.
5012 * This is used to test whether pfn -> nid mapping of the chosen memory
5013 * model has fine enough granularity to avoid incorrect mapping for the
5014 * populated node map.
5016 * Returns the determined alignment in pfn's. 0 if there is no alignment
5017 * requirement (single node).
5019 unsigned long __init node_map_pfn_alignment(void)
5021 unsigned long accl_mask = 0, last_end = 0;
5022 unsigned long start, end, mask;
5026 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5027 if (!start || last_nid < 0 || last_nid == nid) {
5034 * Start with a mask granular enough to pin-point to the
5035 * start pfn and tick off bits one-by-one until it becomes
5036 * too coarse to separate the current node from the last.
5038 mask = ~((1 << __ffs(start)) - 1);
5039 while (mask && last_end <= (start & (mask << 1)))
5042 /* accumulate all internode masks */
5046 /* convert mask to number of pages */
5047 return ~accl_mask + 1;
5050 /* Find the lowest pfn for a node */
5051 static unsigned long __init find_min_pfn_for_node(int nid)
5053 unsigned long min_pfn = ULONG_MAX;
5054 unsigned long start_pfn;
5057 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5058 min_pfn = min(min_pfn, start_pfn);
5060 if (min_pfn == ULONG_MAX) {
5062 "Could not find start_pfn for node %d\n", nid);
5070 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5072 * It returns the minimum PFN based on information provided via
5073 * memblock_set_node().
5075 unsigned long __init find_min_pfn_with_active_regions(void)
5077 return find_min_pfn_for_node(MAX_NUMNODES);
5081 * early_calculate_totalpages()
5082 * Sum pages in active regions for movable zone.
5083 * Populate N_MEMORY for calculating usable_nodes.
5085 static unsigned long __init early_calculate_totalpages(void)
5087 unsigned long totalpages = 0;
5088 unsigned long start_pfn, end_pfn;
5091 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5092 unsigned long pages = end_pfn - start_pfn;
5094 totalpages += pages;
5096 node_set_state(nid, N_MEMORY);
5102 * Find the PFN the Movable zone begins in each node. Kernel memory
5103 * is spread evenly between nodes as long as the nodes have enough
5104 * memory. When they don't, some nodes will have more kernelcore than
5107 static void __init find_zone_movable_pfns_for_nodes(void)
5110 unsigned long usable_startpfn;
5111 unsigned long kernelcore_node, kernelcore_remaining;
5112 /* save the state before borrow the nodemask */
5113 nodemask_t saved_node_state = node_states[N_MEMORY];
5114 unsigned long totalpages = early_calculate_totalpages();
5115 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5116 struct memblock_region *r;
5118 /* Need to find movable_zone earlier when movable_node is specified. */
5119 find_usable_zone_for_movable();
5122 * If movable_node is specified, ignore kernelcore and movablecore
5125 if (movable_node_is_enabled()) {
5126 for_each_memblock(memory, r) {
5127 if (!memblock_is_hotpluggable(r))
5132 usable_startpfn = PFN_DOWN(r->base);
5133 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5134 min(usable_startpfn, zone_movable_pfn[nid]) :
5142 * If movablecore=nn[KMG] was specified, calculate what size of
5143 * kernelcore that corresponds so that memory usable for
5144 * any allocation type is evenly spread. If both kernelcore
5145 * and movablecore are specified, then the value of kernelcore
5146 * will be used for required_kernelcore if it's greater than
5147 * what movablecore would have allowed.
5149 if (required_movablecore) {
5150 unsigned long corepages;
5153 * Round-up so that ZONE_MOVABLE is at least as large as what
5154 * was requested by the user
5156 required_movablecore =
5157 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5158 corepages = totalpages - required_movablecore;
5160 required_kernelcore = max(required_kernelcore, corepages);
5163 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5164 if (!required_kernelcore)
5167 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5168 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5171 /* Spread kernelcore memory as evenly as possible throughout nodes */
5172 kernelcore_node = required_kernelcore / usable_nodes;
5173 for_each_node_state(nid, N_MEMORY) {
5174 unsigned long start_pfn, end_pfn;
5177 * Recalculate kernelcore_node if the division per node
5178 * now exceeds what is necessary to satisfy the requested
5179 * amount of memory for the kernel
5181 if (required_kernelcore < kernelcore_node)
5182 kernelcore_node = required_kernelcore / usable_nodes;
5185 * As the map is walked, we track how much memory is usable
5186 * by the kernel using kernelcore_remaining. When it is
5187 * 0, the rest of the node is usable by ZONE_MOVABLE
5189 kernelcore_remaining = kernelcore_node;
5191 /* Go through each range of PFNs within this node */
5192 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5193 unsigned long size_pages;
5195 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5196 if (start_pfn >= end_pfn)
5199 /* Account for what is only usable for kernelcore */
5200 if (start_pfn < usable_startpfn) {
5201 unsigned long kernel_pages;
5202 kernel_pages = min(end_pfn, usable_startpfn)
5205 kernelcore_remaining -= min(kernel_pages,
5206 kernelcore_remaining);
5207 required_kernelcore -= min(kernel_pages,
5208 required_kernelcore);
5210 /* Continue if range is now fully accounted */
5211 if (end_pfn <= usable_startpfn) {
5214 * Push zone_movable_pfn to the end so
5215 * that if we have to rebalance
5216 * kernelcore across nodes, we will
5217 * not double account here
5219 zone_movable_pfn[nid] = end_pfn;
5222 start_pfn = usable_startpfn;
5226 * The usable PFN range for ZONE_MOVABLE is from
5227 * start_pfn->end_pfn. Calculate size_pages as the
5228 * number of pages used as kernelcore
5230 size_pages = end_pfn - start_pfn;
5231 if (size_pages > kernelcore_remaining)
5232 size_pages = kernelcore_remaining;
5233 zone_movable_pfn[nid] = start_pfn + size_pages;
5236 * Some kernelcore has been met, update counts and
5237 * break if the kernelcore for this node has been
5240 required_kernelcore -= min(required_kernelcore,
5242 kernelcore_remaining -= size_pages;
5243 if (!kernelcore_remaining)
5249 * If there is still required_kernelcore, we do another pass with one
5250 * less node in the count. This will push zone_movable_pfn[nid] further
5251 * along on the nodes that still have memory until kernelcore is
5255 if (usable_nodes && required_kernelcore > usable_nodes)
5259 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5260 for (nid = 0; nid < MAX_NUMNODES; nid++)
5261 zone_movable_pfn[nid] =
5262 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5265 /* restore the node_state */
5266 node_states[N_MEMORY] = saved_node_state;
5269 /* Any regular or high memory on that node ? */
5270 static void check_for_memory(pg_data_t *pgdat, int nid)
5272 enum zone_type zone_type;
5274 if (N_MEMORY == N_NORMAL_MEMORY)
5277 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5278 struct zone *zone = &pgdat->node_zones[zone_type];
5279 if (populated_zone(zone)) {
5280 node_set_state(nid, N_HIGH_MEMORY);
5281 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5282 zone_type <= ZONE_NORMAL)
5283 node_set_state(nid, N_NORMAL_MEMORY);
5290 * free_area_init_nodes - Initialise all pg_data_t and zone data
5291 * @max_zone_pfn: an array of max PFNs for each zone
5293 * This will call free_area_init_node() for each active node in the system.
5294 * Using the page ranges provided by memblock_set_node(), the size of each
5295 * zone in each node and their holes is calculated. If the maximum PFN
5296 * between two adjacent zones match, it is assumed that the zone is empty.
5297 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5298 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5299 * starts where the previous one ended. For example, ZONE_DMA32 starts
5300 * at arch_max_dma_pfn.
5302 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5304 unsigned long start_pfn, end_pfn;
5307 /* Record where the zone boundaries are */
5308 memset(arch_zone_lowest_possible_pfn, 0,
5309 sizeof(arch_zone_lowest_possible_pfn));
5310 memset(arch_zone_highest_possible_pfn, 0,
5311 sizeof(arch_zone_highest_possible_pfn));
5312 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5313 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5314 for (i = 1; i < MAX_NR_ZONES; i++) {
5315 if (i == ZONE_MOVABLE)
5317 arch_zone_lowest_possible_pfn[i] =
5318 arch_zone_highest_possible_pfn[i-1];
5319 arch_zone_highest_possible_pfn[i] =
5320 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5322 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5323 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5325 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5326 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5327 find_zone_movable_pfns_for_nodes();
5329 /* Print out the zone ranges */
5330 printk("Zone ranges:\n");
5331 for (i = 0; i < MAX_NR_ZONES; i++) {
5332 if (i == ZONE_MOVABLE)
5334 printk(KERN_CONT " %-8s ", zone_names[i]);
5335 if (arch_zone_lowest_possible_pfn[i] ==
5336 arch_zone_highest_possible_pfn[i])
5337 printk(KERN_CONT "empty\n");
5339 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5340 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5341 (arch_zone_highest_possible_pfn[i]
5342 << PAGE_SHIFT) - 1);
5345 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5346 printk("Movable zone start for each node\n");
5347 for (i = 0; i < MAX_NUMNODES; i++) {
5348 if (zone_movable_pfn[i])
5349 printk(" Node %d: %#010lx\n", i,
5350 zone_movable_pfn[i] << PAGE_SHIFT);
5353 /* Print out the early node map */
5354 printk("Early memory node ranges\n");
5355 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5356 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5357 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5359 /* Initialise every node */
5360 mminit_verify_pageflags_layout();
5361 setup_nr_node_ids();
5362 for_each_online_node(nid) {
5363 pg_data_t *pgdat = NODE_DATA(nid);
5364 free_area_init_node(nid, NULL,
5365 find_min_pfn_for_node(nid), NULL);
5367 /* Any memory on that node */
5368 if (pgdat->node_present_pages)
5369 node_set_state(nid, N_MEMORY);
5370 check_for_memory(pgdat, nid);
5374 static int __init cmdline_parse_core(char *p, unsigned long *core)
5376 unsigned long long coremem;
5380 coremem = memparse(p, &p);
5381 *core = coremem >> PAGE_SHIFT;
5383 /* Paranoid check that UL is enough for the coremem value */
5384 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5390 * kernelcore=size sets the amount of memory for use for allocations that
5391 * cannot be reclaimed or migrated.
5393 static int __init cmdline_parse_kernelcore(char *p)
5395 return cmdline_parse_core(p, &required_kernelcore);
5399 * movablecore=size sets the amount of memory for use for allocations that
5400 * can be reclaimed or migrated.
5402 static int __init cmdline_parse_movablecore(char *p)
5404 return cmdline_parse_core(p, &required_movablecore);
5407 early_param("kernelcore", cmdline_parse_kernelcore);
5408 early_param("movablecore", cmdline_parse_movablecore);
5410 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5412 void adjust_managed_page_count(struct page *page, long count)
5414 spin_lock(&managed_page_count_lock);
5415 page_zone(page)->managed_pages += count;
5416 totalram_pages += count;
5417 #ifdef CONFIG_HIGHMEM
5418 if (PageHighMem(page))
5419 totalhigh_pages += count;
5421 spin_unlock(&managed_page_count_lock);
5423 EXPORT_SYMBOL(adjust_managed_page_count);
5425 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5428 unsigned long pages = 0;
5430 start = (void *)PAGE_ALIGN((unsigned long)start);
5431 end = (void *)((unsigned long)end & PAGE_MASK);
5432 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5433 if ((unsigned int)poison <= 0xFF)
5434 memset(pos, poison, PAGE_SIZE);
5435 free_reserved_page(virt_to_page(pos));
5439 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5440 s, pages << (PAGE_SHIFT - 10), start, end);
5444 EXPORT_SYMBOL(free_reserved_area);
5446 #ifdef CONFIG_HIGHMEM
5447 void free_highmem_page(struct page *page)
5449 __free_reserved_page(page);
5451 page_zone(page)->managed_pages++;
5457 void __init mem_init_print_info(const char *str)
5459 unsigned long physpages, codesize, datasize, rosize, bss_size;
5460 unsigned long init_code_size, init_data_size;
5462 physpages = get_num_physpages();
5463 codesize = _etext - _stext;
5464 datasize = _edata - _sdata;
5465 rosize = __end_rodata - __start_rodata;
5466 bss_size = __bss_stop - __bss_start;
5467 init_data_size = __init_end - __init_begin;
5468 init_code_size = _einittext - _sinittext;
5471 * Detect special cases and adjust section sizes accordingly:
5472 * 1) .init.* may be embedded into .data sections
5473 * 2) .init.text.* may be out of [__init_begin, __init_end],
5474 * please refer to arch/tile/kernel/vmlinux.lds.S.
5475 * 3) .rodata.* may be embedded into .text or .data sections.
5477 #define adj_init_size(start, end, size, pos, adj) \
5479 if (start <= pos && pos < end && size > adj) \
5483 adj_init_size(__init_begin, __init_end, init_data_size,
5484 _sinittext, init_code_size);
5485 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5486 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5487 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5488 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5490 #undef adj_init_size
5492 printk("Memory: %luK/%luK available "
5493 "(%luK kernel code, %luK rwdata, %luK rodata, "
5494 "%luK init, %luK bss, %luK reserved"
5495 #ifdef CONFIG_HIGHMEM
5499 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5500 codesize >> 10, datasize >> 10, rosize >> 10,
5501 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5502 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5503 #ifdef CONFIG_HIGHMEM
5504 totalhigh_pages << (PAGE_SHIFT-10),
5506 str ? ", " : "", str ? str : "");
5510 * set_dma_reserve - set the specified number of pages reserved in the first zone
5511 * @new_dma_reserve: The number of pages to mark reserved
5513 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5514 * In the DMA zone, a significant percentage may be consumed by kernel image
5515 * and other unfreeable allocations which can skew the watermarks badly. This
5516 * function may optionally be used to account for unfreeable pages in the
5517 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5518 * smaller per-cpu batchsize.
5520 void __init set_dma_reserve(unsigned long new_dma_reserve)
5522 dma_reserve = new_dma_reserve;
5525 void __init free_area_init(unsigned long *zones_size)
5527 free_area_init_node(0, zones_size,
5528 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5531 static int page_alloc_cpu_notify(struct notifier_block *self,
5532 unsigned long action, void *hcpu)
5534 int cpu = (unsigned long)hcpu;
5536 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5537 lru_add_drain_cpu(cpu);
5541 * Spill the event counters of the dead processor
5542 * into the current processors event counters.
5543 * This artificially elevates the count of the current
5546 vm_events_fold_cpu(cpu);
5549 * Zero the differential counters of the dead processor
5550 * so that the vm statistics are consistent.
5552 * This is only okay since the processor is dead and cannot
5553 * race with what we are doing.
5555 cpu_vm_stats_fold(cpu);
5560 void __init page_alloc_init(void)
5562 hotcpu_notifier(page_alloc_cpu_notify, 0);
5566 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5567 * or min_free_kbytes changes.
5569 static void calculate_totalreserve_pages(void)
5571 struct pglist_data *pgdat;
5572 unsigned long reserve_pages = 0;
5573 enum zone_type i, j;
5575 for_each_online_pgdat(pgdat) {
5576 for (i = 0; i < MAX_NR_ZONES; i++) {
5577 struct zone *zone = pgdat->node_zones + i;
5578 unsigned long max = 0;
5580 /* Find valid and maximum lowmem_reserve in the zone */
5581 for (j = i; j < MAX_NR_ZONES; j++) {
5582 if (zone->lowmem_reserve[j] > max)
5583 max = zone->lowmem_reserve[j];
5586 /* we treat the high watermark as reserved pages. */
5587 max += high_wmark_pages(zone);
5589 if (max > zone->managed_pages)
5590 max = zone->managed_pages;
5591 reserve_pages += max;
5593 * Lowmem reserves are not available to
5594 * GFP_HIGHUSER page cache allocations and
5595 * kswapd tries to balance zones to their high
5596 * watermark. As a result, neither should be
5597 * regarded as dirtyable memory, to prevent a
5598 * situation where reclaim has to clean pages
5599 * in order to balance the zones.
5601 zone->dirty_balance_reserve = max;
5604 dirty_balance_reserve = reserve_pages;
5605 totalreserve_pages = reserve_pages;
5609 * setup_per_zone_lowmem_reserve - called whenever
5610 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5611 * has a correct pages reserved value, so an adequate number of
5612 * pages are left in the zone after a successful __alloc_pages().
5614 static void setup_per_zone_lowmem_reserve(void)
5616 struct pglist_data *pgdat;
5617 enum zone_type j, idx;
5619 for_each_online_pgdat(pgdat) {
5620 for (j = 0; j < MAX_NR_ZONES; j++) {
5621 struct zone *zone = pgdat->node_zones + j;
5622 unsigned long managed_pages = zone->managed_pages;
5624 zone->lowmem_reserve[j] = 0;
5628 struct zone *lower_zone;
5632 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5633 sysctl_lowmem_reserve_ratio[idx] = 1;
5635 lower_zone = pgdat->node_zones + idx;
5636 lower_zone->lowmem_reserve[j] = managed_pages /
5637 sysctl_lowmem_reserve_ratio[idx];
5638 managed_pages += lower_zone->managed_pages;
5643 /* update totalreserve_pages */
5644 calculate_totalreserve_pages();
5647 static void __setup_per_zone_wmarks(void)
5649 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5650 unsigned long lowmem_pages = 0;
5652 unsigned long flags;
5654 /* Calculate total number of !ZONE_HIGHMEM pages */
5655 for_each_zone(zone) {
5656 if (!is_highmem(zone))
5657 lowmem_pages += zone->managed_pages;
5660 for_each_zone(zone) {
5663 spin_lock_irqsave(&zone->lock, flags);
5664 tmp = (u64)pages_min * zone->managed_pages;
5665 do_div(tmp, lowmem_pages);
5666 if (is_highmem(zone)) {
5668 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5669 * need highmem pages, so cap pages_min to a small
5672 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5673 * deltas controls asynch page reclaim, and so should
5674 * not be capped for highmem.
5676 unsigned long min_pages;
5678 min_pages = zone->managed_pages / 1024;
5679 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5680 zone->watermark[WMARK_MIN] = min_pages;
5683 * If it's a lowmem zone, reserve a number of pages
5684 * proportionate to the zone's size.
5686 zone->watermark[WMARK_MIN] = tmp;
5689 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5690 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5692 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5693 high_wmark_pages(zone) -
5694 low_wmark_pages(zone) -
5695 zone_page_state(zone, NR_ALLOC_BATCH));
5697 setup_zone_migrate_reserve(zone);
5698 spin_unlock_irqrestore(&zone->lock, flags);
5701 /* update totalreserve_pages */
5702 calculate_totalreserve_pages();
5706 * setup_per_zone_wmarks - called when min_free_kbytes changes
5707 * or when memory is hot-{added|removed}
5709 * Ensures that the watermark[min,low,high] values for each zone are set
5710 * correctly with respect to min_free_kbytes.
5712 void setup_per_zone_wmarks(void)
5714 mutex_lock(&zonelists_mutex);
5715 __setup_per_zone_wmarks();
5716 mutex_unlock(&zonelists_mutex);
5720 * The inactive anon list should be small enough that the VM never has to
5721 * do too much work, but large enough that each inactive page has a chance
5722 * to be referenced again before it is swapped out.
5724 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5725 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5726 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5727 * the anonymous pages are kept on the inactive list.
5730 * memory ratio inactive anon
5731 * -------------------------------------
5740 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5742 unsigned int gb, ratio;
5744 /* Zone size in gigabytes */
5745 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5747 ratio = int_sqrt(10 * gb);
5751 zone->inactive_ratio = ratio;
5754 static void __meminit setup_per_zone_inactive_ratio(void)
5759 calculate_zone_inactive_ratio(zone);
5763 * Initialise min_free_kbytes.
5765 * For small machines we want it small (128k min). For large machines
5766 * we want it large (64MB max). But it is not linear, because network
5767 * bandwidth does not increase linearly with machine size. We use
5769 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5770 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5786 int __meminit init_per_zone_wmark_min(void)
5788 unsigned long lowmem_kbytes;
5789 int new_min_free_kbytes;
5791 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5792 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5794 if (new_min_free_kbytes > user_min_free_kbytes) {
5795 min_free_kbytes = new_min_free_kbytes;
5796 if (min_free_kbytes < 128)
5797 min_free_kbytes = 128;
5798 if (min_free_kbytes > 65536)
5799 min_free_kbytes = 65536;
5801 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5802 new_min_free_kbytes, user_min_free_kbytes);
5804 setup_per_zone_wmarks();
5805 refresh_zone_stat_thresholds();
5806 setup_per_zone_lowmem_reserve();
5807 setup_per_zone_inactive_ratio();
5810 module_init(init_per_zone_wmark_min)
5813 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5814 * that we can call two helper functions whenever min_free_kbytes
5817 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5818 void __user *buffer, size_t *length, loff_t *ppos)
5822 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5827 user_min_free_kbytes = min_free_kbytes;
5828 setup_per_zone_wmarks();
5834 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5835 void __user *buffer, size_t *length, loff_t *ppos)
5840 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5845 zone->min_unmapped_pages = (zone->managed_pages *
5846 sysctl_min_unmapped_ratio) / 100;
5850 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5851 void __user *buffer, size_t *length, loff_t *ppos)
5856 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5861 zone->min_slab_pages = (zone->managed_pages *
5862 sysctl_min_slab_ratio) / 100;
5868 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5869 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5870 * whenever sysctl_lowmem_reserve_ratio changes.
5872 * The reserve ratio obviously has absolutely no relation with the
5873 * minimum watermarks. The lowmem reserve ratio can only make sense
5874 * if in function of the boot time zone sizes.
5876 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5877 void __user *buffer, size_t *length, loff_t *ppos)
5879 proc_dointvec_minmax(table, write, buffer, length, ppos);
5880 setup_per_zone_lowmem_reserve();
5885 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5886 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5887 * pagelist can have before it gets flushed back to buddy allocator.
5889 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5890 void __user *buffer, size_t *length, loff_t *ppos)
5893 int old_percpu_pagelist_fraction;
5896 mutex_lock(&pcp_batch_high_lock);
5897 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5899 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5900 if (!write || ret < 0)
5903 /* Sanity checking to avoid pcp imbalance */
5904 if (percpu_pagelist_fraction &&
5905 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
5906 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
5912 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
5915 for_each_populated_zone(zone) {
5918 for_each_possible_cpu(cpu)
5919 pageset_set_high_and_batch(zone,
5920 per_cpu_ptr(zone->pageset, cpu));
5923 mutex_unlock(&pcp_batch_high_lock);
5927 int hashdist = HASHDIST_DEFAULT;
5930 static int __init set_hashdist(char *str)
5934 hashdist = simple_strtoul(str, &str, 0);
5937 __setup("hashdist=", set_hashdist);
5941 * allocate a large system hash table from bootmem
5942 * - it is assumed that the hash table must contain an exact power-of-2
5943 * quantity of entries
5944 * - limit is the number of hash buckets, not the total allocation size
5946 void *__init alloc_large_system_hash(const char *tablename,
5947 unsigned long bucketsize,
5948 unsigned long numentries,
5951 unsigned int *_hash_shift,
5952 unsigned int *_hash_mask,
5953 unsigned long low_limit,
5954 unsigned long high_limit)
5956 unsigned long long max = high_limit;
5957 unsigned long log2qty, size;
5960 /* allow the kernel cmdline to have a say */
5962 /* round applicable memory size up to nearest megabyte */
5963 numentries = nr_kernel_pages;
5965 /* It isn't necessary when PAGE_SIZE >= 1MB */
5966 if (PAGE_SHIFT < 20)
5967 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5969 /* limit to 1 bucket per 2^scale bytes of low memory */
5970 if (scale > PAGE_SHIFT)
5971 numentries >>= (scale - PAGE_SHIFT);
5973 numentries <<= (PAGE_SHIFT - scale);
5975 /* Make sure we've got at least a 0-order allocation.. */
5976 if (unlikely(flags & HASH_SMALL)) {
5977 /* Makes no sense without HASH_EARLY */
5978 WARN_ON(!(flags & HASH_EARLY));
5979 if (!(numentries >> *_hash_shift)) {
5980 numentries = 1UL << *_hash_shift;
5981 BUG_ON(!numentries);
5983 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5984 numentries = PAGE_SIZE / bucketsize;
5986 numentries = roundup_pow_of_two(numentries);
5988 /* limit allocation size to 1/16 total memory by default */
5990 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5991 do_div(max, bucketsize);
5993 max = min(max, 0x80000000ULL);
5995 if (numentries < low_limit)
5996 numentries = low_limit;
5997 if (numentries > max)
6000 log2qty = ilog2(numentries);
6003 size = bucketsize << log2qty;
6004 if (flags & HASH_EARLY)
6005 table = memblock_virt_alloc_nopanic(size, 0);
6007 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6010 * If bucketsize is not a power-of-two, we may free
6011 * some pages at the end of hash table which
6012 * alloc_pages_exact() automatically does
6014 if (get_order(size) < MAX_ORDER) {
6015 table = alloc_pages_exact(size, GFP_ATOMIC);
6016 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6019 } while (!table && size > PAGE_SIZE && --log2qty);
6022 panic("Failed to allocate %s hash table\n", tablename);
6024 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6027 ilog2(size) - PAGE_SHIFT,
6031 *_hash_shift = log2qty;
6033 *_hash_mask = (1 << log2qty) - 1;
6038 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6039 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6042 #ifdef CONFIG_SPARSEMEM
6043 return __pfn_to_section(pfn)->pageblock_flags;
6045 return zone->pageblock_flags;
6046 #endif /* CONFIG_SPARSEMEM */
6049 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6051 #ifdef CONFIG_SPARSEMEM
6052 pfn &= (PAGES_PER_SECTION-1);
6053 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6055 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6056 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6057 #endif /* CONFIG_SPARSEMEM */
6061 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6062 * @page: The page within the block of interest
6063 * @pfn: The target page frame number
6064 * @end_bitidx: The last bit of interest to retrieve
6065 * @mask: mask of bits that the caller is interested in
6067 * Return: pageblock_bits flags
6069 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6070 unsigned long end_bitidx,
6074 unsigned long *bitmap;
6075 unsigned long bitidx, word_bitidx;
6078 zone = page_zone(page);
6079 bitmap = get_pageblock_bitmap(zone, pfn);
6080 bitidx = pfn_to_bitidx(zone, pfn);
6081 word_bitidx = bitidx / BITS_PER_LONG;
6082 bitidx &= (BITS_PER_LONG-1);
6084 word = bitmap[word_bitidx];
6085 bitidx += end_bitidx;
6086 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6090 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6091 * @page: The page within the block of interest
6092 * @flags: The flags to set
6093 * @pfn: The target page frame number
6094 * @end_bitidx: The last bit of interest
6095 * @mask: mask of bits that the caller is interested in
6097 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6099 unsigned long end_bitidx,
6103 unsigned long *bitmap;
6104 unsigned long bitidx, word_bitidx;
6105 unsigned long old_word, word;
6107 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6109 zone = page_zone(page);
6110 bitmap = get_pageblock_bitmap(zone, pfn);
6111 bitidx = pfn_to_bitidx(zone, pfn);
6112 word_bitidx = bitidx / BITS_PER_LONG;
6113 bitidx &= (BITS_PER_LONG-1);
6115 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6117 bitidx += end_bitidx;
6118 mask <<= (BITS_PER_LONG - bitidx - 1);
6119 flags <<= (BITS_PER_LONG - bitidx - 1);
6121 word = ACCESS_ONCE(bitmap[word_bitidx]);
6123 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6124 if (word == old_word)
6131 * This function checks whether pageblock includes unmovable pages or not.
6132 * If @count is not zero, it is okay to include less @count unmovable pages
6134 * PageLRU check without isolation or lru_lock could race so that
6135 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6136 * expect this function should be exact.
6138 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6139 bool skip_hwpoisoned_pages)
6141 unsigned long pfn, iter, found;
6145 * For avoiding noise data, lru_add_drain_all() should be called
6146 * If ZONE_MOVABLE, the zone never contains unmovable pages
6148 if (zone_idx(zone) == ZONE_MOVABLE)
6150 mt = get_pageblock_migratetype(page);
6151 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6154 pfn = page_to_pfn(page);
6155 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6156 unsigned long check = pfn + iter;
6158 if (!pfn_valid_within(check))
6161 page = pfn_to_page(check);
6164 * Hugepages are not in LRU lists, but they're movable.
6165 * We need not scan over tail pages bacause we don't
6166 * handle each tail page individually in migration.
6168 if (PageHuge(page)) {
6169 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6174 * We can't use page_count without pin a page
6175 * because another CPU can free compound page.
6176 * This check already skips compound tails of THP
6177 * because their page->_count is zero at all time.
6179 if (!atomic_read(&page->_count)) {
6180 if (PageBuddy(page))
6181 iter += (1 << page_order(page)) - 1;
6186 * The HWPoisoned page may be not in buddy system, and
6187 * page_count() is not 0.
6189 if (skip_hwpoisoned_pages && PageHWPoison(page))
6195 * If there are RECLAIMABLE pages, we need to check it.
6196 * But now, memory offline itself doesn't call shrink_slab()
6197 * and it still to be fixed.
6200 * If the page is not RAM, page_count()should be 0.
6201 * we don't need more check. This is an _used_ not-movable page.
6203 * The problematic thing here is PG_reserved pages. PG_reserved
6204 * is set to both of a memory hole page and a _used_ kernel
6213 bool is_pageblock_removable_nolock(struct page *page)
6219 * We have to be careful here because we are iterating over memory
6220 * sections which are not zone aware so we might end up outside of
6221 * the zone but still within the section.
6222 * We have to take care about the node as well. If the node is offline
6223 * its NODE_DATA will be NULL - see page_zone.
6225 if (!node_online(page_to_nid(page)))
6228 zone = page_zone(page);
6229 pfn = page_to_pfn(page);
6230 if (!zone_spans_pfn(zone, pfn))
6233 return !has_unmovable_pages(zone, page, 0, true);
6238 static unsigned long pfn_max_align_down(unsigned long pfn)
6240 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6241 pageblock_nr_pages) - 1);
6244 static unsigned long pfn_max_align_up(unsigned long pfn)
6246 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6247 pageblock_nr_pages));
6250 /* [start, end) must belong to a single zone. */
6251 static int __alloc_contig_migrate_range(struct compact_control *cc,
6252 unsigned long start, unsigned long end)
6254 /* This function is based on compact_zone() from compaction.c. */
6255 unsigned long nr_reclaimed;
6256 unsigned long pfn = start;
6257 unsigned int tries = 0;
6262 while (pfn < end || !list_empty(&cc->migratepages)) {
6263 if (fatal_signal_pending(current)) {
6268 if (list_empty(&cc->migratepages)) {
6269 cc->nr_migratepages = 0;
6270 pfn = isolate_migratepages_range(cc->zone, cc,
6277 } else if (++tries == 5) {
6278 ret = ret < 0 ? ret : -EBUSY;
6282 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6284 cc->nr_migratepages -= nr_reclaimed;
6286 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6287 NULL, 0, cc->mode, MR_CMA);
6290 putback_movable_pages(&cc->migratepages);
6297 * alloc_contig_range() -- tries to allocate given range of pages
6298 * @start: start PFN to allocate
6299 * @end: one-past-the-last PFN to allocate
6300 * @migratetype: migratetype of the underlaying pageblocks (either
6301 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6302 * in range must have the same migratetype and it must
6303 * be either of the two.
6305 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6306 * aligned, however it's the caller's responsibility to guarantee that
6307 * we are the only thread that changes migrate type of pageblocks the
6310 * The PFN range must belong to a single zone.
6312 * Returns zero on success or negative error code. On success all
6313 * pages which PFN is in [start, end) are allocated for the caller and
6314 * need to be freed with free_contig_range().
6316 int alloc_contig_range(unsigned long start, unsigned long end,
6317 unsigned migratetype)
6319 unsigned long outer_start, outer_end;
6322 struct compact_control cc = {
6323 .nr_migratepages = 0,
6325 .zone = page_zone(pfn_to_page(start)),
6326 .mode = MIGRATE_SYNC,
6327 .ignore_skip_hint = true,
6329 INIT_LIST_HEAD(&cc.migratepages);
6332 * What we do here is we mark all pageblocks in range as
6333 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6334 * have different sizes, and due to the way page allocator
6335 * work, we align the range to biggest of the two pages so
6336 * that page allocator won't try to merge buddies from
6337 * different pageblocks and change MIGRATE_ISOLATE to some
6338 * other migration type.
6340 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6341 * migrate the pages from an unaligned range (ie. pages that
6342 * we are interested in). This will put all the pages in
6343 * range back to page allocator as MIGRATE_ISOLATE.
6345 * When this is done, we take the pages in range from page
6346 * allocator removing them from the buddy system. This way
6347 * page allocator will never consider using them.
6349 * This lets us mark the pageblocks back as
6350 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6351 * aligned range but not in the unaligned, original range are
6352 * put back to page allocator so that buddy can use them.
6355 ret = start_isolate_page_range(pfn_max_align_down(start),
6356 pfn_max_align_up(end), migratetype,
6361 ret = __alloc_contig_migrate_range(&cc, start, end);
6366 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6367 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6368 * more, all pages in [start, end) are free in page allocator.
6369 * What we are going to do is to allocate all pages from
6370 * [start, end) (that is remove them from page allocator).
6372 * The only problem is that pages at the beginning and at the
6373 * end of interesting range may be not aligned with pages that
6374 * page allocator holds, ie. they can be part of higher order
6375 * pages. Because of this, we reserve the bigger range and
6376 * once this is done free the pages we are not interested in.
6378 * We don't have to hold zone->lock here because the pages are
6379 * isolated thus they won't get removed from buddy.
6382 lru_add_drain_all();
6386 outer_start = start;
6387 while (!PageBuddy(pfn_to_page(outer_start))) {
6388 if (++order >= MAX_ORDER) {
6392 outer_start &= ~0UL << order;
6395 /* Make sure the range is really isolated. */
6396 if (test_pages_isolated(outer_start, end, false)) {
6397 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6404 /* Grab isolated pages from freelists. */
6405 outer_end = isolate_freepages_range(&cc, outer_start, end);
6411 /* Free head and tail (if any) */
6412 if (start != outer_start)
6413 free_contig_range(outer_start, start - outer_start);
6414 if (end != outer_end)
6415 free_contig_range(end, outer_end - end);
6418 undo_isolate_page_range(pfn_max_align_down(start),
6419 pfn_max_align_up(end), migratetype);
6423 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6425 unsigned int count = 0;
6427 for (; nr_pages--; pfn++) {
6428 struct page *page = pfn_to_page(pfn);
6430 count += page_count(page) != 1;
6433 WARN(count != 0, "%d pages are still in use!\n", count);
6437 #ifdef CONFIG_MEMORY_HOTPLUG
6439 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6440 * page high values need to be recalulated.
6442 void __meminit zone_pcp_update(struct zone *zone)
6445 mutex_lock(&pcp_batch_high_lock);
6446 for_each_possible_cpu(cpu)
6447 pageset_set_high_and_batch(zone,
6448 per_cpu_ptr(zone->pageset, cpu));
6449 mutex_unlock(&pcp_batch_high_lock);
6453 void zone_pcp_reset(struct zone *zone)
6455 unsigned long flags;
6457 struct per_cpu_pageset *pset;
6459 /* avoid races with drain_pages() */
6460 local_irq_save(flags);
6461 if (zone->pageset != &boot_pageset) {
6462 for_each_online_cpu(cpu) {
6463 pset = per_cpu_ptr(zone->pageset, cpu);
6464 drain_zonestat(zone, pset);
6466 free_percpu(zone->pageset);
6467 zone->pageset = &boot_pageset;
6469 local_irq_restore(flags);
6472 #ifdef CONFIG_MEMORY_HOTREMOVE
6474 * All pages in the range must be isolated before calling this.
6477 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6481 unsigned int order, i;
6483 unsigned long flags;
6484 /* find the first valid pfn */
6485 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6490 zone = page_zone(pfn_to_page(pfn));
6491 spin_lock_irqsave(&zone->lock, flags);
6493 while (pfn < end_pfn) {
6494 if (!pfn_valid(pfn)) {
6498 page = pfn_to_page(pfn);
6500 * The HWPoisoned page may be not in buddy system, and
6501 * page_count() is not 0.
6503 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6505 SetPageReserved(page);
6509 BUG_ON(page_count(page));
6510 BUG_ON(!PageBuddy(page));
6511 order = page_order(page);
6512 #ifdef CONFIG_DEBUG_VM
6513 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6514 pfn, 1 << order, end_pfn);
6516 list_del(&page->lru);
6517 rmv_page_order(page);
6518 zone->free_area[order].nr_free--;
6519 for (i = 0; i < (1 << order); i++)
6520 SetPageReserved((page+i));
6521 pfn += (1 << order);
6523 spin_unlock_irqrestore(&zone->lock, flags);
6527 #ifdef CONFIG_MEMORY_FAILURE
6528 bool is_free_buddy_page(struct page *page)
6530 struct zone *zone = page_zone(page);
6531 unsigned long pfn = page_to_pfn(page);
6532 unsigned long flags;
6535 spin_lock_irqsave(&zone->lock, flags);
6536 for (order = 0; order < MAX_ORDER; order++) {
6537 struct page *page_head = page - (pfn & ((1 << order) - 1));
6539 if (PageBuddy(page_head) && page_order(page_head) >= order)
6542 spin_unlock_irqrestore(&zone->lock, flags);
6544 return order < MAX_ORDER;
6548 static const struct trace_print_flags pageflag_names[] = {
6549 {1UL << PG_locked, "locked" },
6550 {1UL << PG_error, "error" },
6551 {1UL << PG_referenced, "referenced" },
6552 {1UL << PG_uptodate, "uptodate" },
6553 {1UL << PG_dirty, "dirty" },
6554 {1UL << PG_lru, "lru" },
6555 {1UL << PG_active, "active" },
6556 {1UL << PG_slab, "slab" },
6557 {1UL << PG_owner_priv_1, "owner_priv_1" },
6558 {1UL << PG_arch_1, "arch_1" },
6559 {1UL << PG_reserved, "reserved" },
6560 {1UL << PG_private, "private" },
6561 {1UL << PG_private_2, "private_2" },
6562 {1UL << PG_writeback, "writeback" },
6563 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6564 {1UL << PG_head, "head" },
6565 {1UL << PG_tail, "tail" },
6567 {1UL << PG_compound, "compound" },
6569 {1UL << PG_swapcache, "swapcache" },
6570 {1UL << PG_mappedtodisk, "mappedtodisk" },
6571 {1UL << PG_reclaim, "reclaim" },
6572 {1UL << PG_swapbacked, "swapbacked" },
6573 {1UL << PG_unevictable, "unevictable" },
6575 {1UL << PG_mlocked, "mlocked" },
6577 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6578 {1UL << PG_uncached, "uncached" },
6580 #ifdef CONFIG_MEMORY_FAILURE
6581 {1UL << PG_hwpoison, "hwpoison" },
6583 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6584 {1UL << PG_compound_lock, "compound_lock" },
6588 static void dump_page_flags(unsigned long flags)
6590 const char *delim = "";
6594 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6596 printk(KERN_ALERT "page flags: %#lx(", flags);
6598 /* remove zone id */
6599 flags &= (1UL << NR_PAGEFLAGS) - 1;
6601 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6603 mask = pageflag_names[i].mask;
6604 if ((flags & mask) != mask)
6608 printk("%s%s", delim, pageflag_names[i].name);
6612 /* check for left over flags */
6614 printk("%s%#lx", delim, flags);
6619 void dump_page_badflags(struct page *page, const char *reason,
6620 unsigned long badflags)
6623 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6624 page, atomic_read(&page->_count), page_mapcount(page),
6625 page->mapping, page->index);
6626 dump_page_flags(page->flags);
6628 pr_alert("page dumped because: %s\n", reason);
6629 if (page->flags & badflags) {
6630 pr_alert("bad because of flags:\n");
6631 dump_page_flags(page->flags & badflags);
6633 mem_cgroup_print_bad_page(page);
6636 void dump_page(struct page *page, const char *reason)
6638 dump_page_badflags(page, reason, 0);
6640 EXPORT_SYMBOL(dump_page);