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/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
66 #include <asm/sections.h>
67 #include <asm/tlbflush.h>
68 #include <asm/div64.h>
71 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
72 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #define MIN_PERCPU_PAGELIST_FRACTION (8)
75 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
76 DEFINE_PER_CPU(int, numa_node);
77 EXPORT_PER_CPU_SYMBOL(numa_node);
80 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
82 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
83 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
84 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
85 * defined in <linux/topology.h>.
87 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
88 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
89 int _node_numa_mem_[MAX_NUMNODES];
93 * Array of node states.
95 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
96 [N_POSSIBLE] = NODE_MASK_ALL,
97 [N_ONLINE] = { { [0] = 1UL } },
99 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_HIGHMEM
101 [N_HIGH_MEMORY] = { { [0] = 1UL } },
103 #ifdef CONFIG_MOVABLE_NODE
104 [N_MEMORY] = { { [0] = 1UL } },
106 [N_CPU] = { { [0] = 1UL } },
109 EXPORT_SYMBOL(node_states);
111 /* Protect totalram_pages and zone->managed_pages */
112 static DEFINE_SPINLOCK(managed_page_count_lock);
114 unsigned long totalram_pages __read_mostly;
115 unsigned long totalreserve_pages __read_mostly;
116 unsigned long totalcma_pages __read_mostly;
118 * When calculating the number of globally allowed dirty pages, there
119 * is a certain number of per-zone reserves that should not be
120 * considered dirtyable memory. This is the sum of those reserves
121 * over all existing zones that contribute dirtyable memory.
123 unsigned long dirty_balance_reserve __read_mostly;
125 int percpu_pagelist_fraction;
126 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
136 static inline int get_pcppage_migratetype(struct page *page)
141 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
143 page->index = migratetype;
146 #ifdef CONFIG_PM_SLEEP
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
156 static gfp_t saved_gfp_mask;
158 void pm_restore_gfp_mask(void)
160 WARN_ON(!mutex_is_locked(&pm_mutex));
161 if (saved_gfp_mask) {
162 gfp_allowed_mask = saved_gfp_mask;
167 void pm_restrict_gfp_mask(void)
169 WARN_ON(!mutex_is_locked(&pm_mutex));
170 WARN_ON(saved_gfp_mask);
171 saved_gfp_mask = gfp_allowed_mask;
172 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
175 bool pm_suspended_storage(void)
177 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
181 #endif /* CONFIG_PM_SLEEP */
183 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184 unsigned int pageblock_order __read_mostly;
187 static void __free_pages_ok(struct page *page, unsigned int order);
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
200 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
201 #ifdef CONFIG_ZONE_DMA
204 #ifdef CONFIG_ZONE_DMA32
207 #ifdef CONFIG_HIGHMEM
213 EXPORT_SYMBOL(totalram_pages);
215 static char * const zone_names[MAX_NR_ZONES] = {
216 #ifdef CONFIG_ZONE_DMA
219 #ifdef CONFIG_ZONE_DMA32
223 #ifdef CONFIG_HIGHMEM
227 #ifdef CONFIG_ZONE_DEVICE
232 static void free_compound_page(struct page *page);
233 compound_page_dtor * const compound_page_dtors[] = {
236 #ifdef CONFIG_HUGETLB_PAGE
241 int min_free_kbytes = 1024;
242 int user_min_free_kbytes = -1;
244 static unsigned long __meminitdata nr_kernel_pages;
245 static unsigned long __meminitdata nr_all_pages;
246 static unsigned long __meminitdata dma_reserve;
248 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
249 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
250 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
251 static unsigned long __initdata required_kernelcore;
252 static unsigned long __initdata required_movablecore;
253 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
255 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
257 EXPORT_SYMBOL(movable_zone);
258 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
261 int nr_node_ids __read_mostly = MAX_NUMNODES;
262 int nr_online_nodes __read_mostly = 1;
263 EXPORT_SYMBOL(nr_node_ids);
264 EXPORT_SYMBOL(nr_online_nodes);
267 int page_group_by_mobility_disabled __read_mostly;
269 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
270 static inline void reset_deferred_meminit(pg_data_t *pgdat)
272 pgdat->first_deferred_pfn = ULONG_MAX;
275 /* Returns true if the struct page for the pfn is uninitialised */
276 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
278 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
284 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
286 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
293 * Returns false when the remaining initialisation should be deferred until
294 * later in the boot cycle when it can be parallelised.
296 static inline bool update_defer_init(pg_data_t *pgdat,
297 unsigned long pfn, unsigned long zone_end,
298 unsigned long *nr_initialised)
300 /* Always populate low zones for address-contrained allocations */
301 if (zone_end < pgdat_end_pfn(pgdat))
304 /* Initialise at least 2G of the highest zone */
306 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
307 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
308 pgdat->first_deferred_pfn = pfn;
315 static inline void reset_deferred_meminit(pg_data_t *pgdat)
319 static inline bool early_page_uninitialised(unsigned long pfn)
324 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
329 static inline bool update_defer_init(pg_data_t *pgdat,
330 unsigned long pfn, unsigned long zone_end,
331 unsigned long *nr_initialised)
338 void set_pageblock_migratetype(struct page *page, int migratetype)
340 if (unlikely(page_group_by_mobility_disabled &&
341 migratetype < MIGRATE_PCPTYPES))
342 migratetype = MIGRATE_UNMOVABLE;
344 set_pageblock_flags_group(page, (unsigned long)migratetype,
345 PB_migrate, PB_migrate_end);
348 #ifdef CONFIG_DEBUG_VM
349 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
353 unsigned long pfn = page_to_pfn(page);
354 unsigned long sp, start_pfn;
357 seq = zone_span_seqbegin(zone);
358 start_pfn = zone->zone_start_pfn;
359 sp = zone->spanned_pages;
360 if (!zone_spans_pfn(zone, pfn))
362 } while (zone_span_seqretry(zone, seq));
365 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
366 pfn, zone_to_nid(zone), zone->name,
367 start_pfn, start_pfn + sp);
372 static int page_is_consistent(struct zone *zone, struct page *page)
374 if (!pfn_valid_within(page_to_pfn(page)))
376 if (zone != page_zone(page))
382 * Temporary debugging check for pages not lying within a given zone.
384 static int bad_range(struct zone *zone, struct page *page)
386 if (page_outside_zone_boundaries(zone, page))
388 if (!page_is_consistent(zone, page))
394 static inline int bad_range(struct zone *zone, struct page *page)
400 static void bad_page(struct page *page, const char *reason,
401 unsigned long bad_flags)
403 static unsigned long resume;
404 static unsigned long nr_shown;
405 static unsigned long nr_unshown;
407 /* Don't complain about poisoned pages */
408 if (PageHWPoison(page)) {
409 page_mapcount_reset(page); /* remove PageBuddy */
414 * Allow a burst of 60 reports, then keep quiet for that minute;
415 * or allow a steady drip of one report per second.
417 if (nr_shown == 60) {
418 if (time_before(jiffies, resume)) {
424 "BUG: Bad page state: %lu messages suppressed\n",
431 resume = jiffies + 60 * HZ;
433 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
434 current->comm, page_to_pfn(page));
435 dump_page_badflags(page, reason, bad_flags);
440 /* Leave bad fields for debug, except PageBuddy could make trouble */
441 page_mapcount_reset(page); /* remove PageBuddy */
442 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
446 * Higher-order pages are called "compound pages". They are structured thusly:
448 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
450 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
451 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
453 * The first tail page's ->compound_dtor holds the offset in array of compound
454 * page destructors. See compound_page_dtors.
456 * The first tail page's ->compound_order holds the order of allocation.
457 * This usage means that zero-order pages may not be compound.
460 static void free_compound_page(struct page *page)
462 __free_pages_ok(page, compound_order(page));
465 void prep_compound_page(struct page *page, unsigned int order)
468 int nr_pages = 1 << order;
470 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
471 set_compound_order(page, order);
473 for (i = 1; i < nr_pages; i++) {
474 struct page *p = page + i;
475 set_page_count(p, 0);
476 p->mapping = TAIL_MAPPING;
477 set_compound_head(p, page);
481 #ifdef CONFIG_DEBUG_PAGEALLOC
482 unsigned int _debug_guardpage_minorder;
483 bool _debug_pagealloc_enabled __read_mostly;
484 bool _debug_guardpage_enabled __read_mostly;
486 static int __init early_debug_pagealloc(char *buf)
491 if (strcmp(buf, "on") == 0)
492 _debug_pagealloc_enabled = true;
496 early_param("debug_pagealloc", early_debug_pagealloc);
498 static bool need_debug_guardpage(void)
500 /* If we don't use debug_pagealloc, we don't need guard page */
501 if (!debug_pagealloc_enabled())
507 static void init_debug_guardpage(void)
509 if (!debug_pagealloc_enabled())
512 _debug_guardpage_enabled = true;
515 struct page_ext_operations debug_guardpage_ops = {
516 .need = need_debug_guardpage,
517 .init = init_debug_guardpage,
520 static int __init debug_guardpage_minorder_setup(char *buf)
524 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
525 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
528 _debug_guardpage_minorder = res;
529 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
532 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
534 static inline void set_page_guard(struct zone *zone, struct page *page,
535 unsigned int order, int migratetype)
537 struct page_ext *page_ext;
539 if (!debug_guardpage_enabled())
542 page_ext = lookup_page_ext(page);
543 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
545 INIT_LIST_HEAD(&page->lru);
546 set_page_private(page, order);
547 /* Guard pages are not available for any usage */
548 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
551 static inline void clear_page_guard(struct zone *zone, struct page *page,
552 unsigned int order, int migratetype)
554 struct page_ext *page_ext;
556 if (!debug_guardpage_enabled())
559 page_ext = lookup_page_ext(page);
560 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
562 set_page_private(page, 0);
563 if (!is_migrate_isolate(migratetype))
564 __mod_zone_freepage_state(zone, (1 << order), migratetype);
567 struct page_ext_operations debug_guardpage_ops = { NULL, };
568 static inline void set_page_guard(struct zone *zone, struct page *page,
569 unsigned int order, int migratetype) {}
570 static inline void clear_page_guard(struct zone *zone, struct page *page,
571 unsigned int order, int migratetype) {}
574 static inline void set_page_order(struct page *page, unsigned int order)
576 set_page_private(page, order);
577 __SetPageBuddy(page);
580 static inline void rmv_page_order(struct page *page)
582 __ClearPageBuddy(page);
583 set_page_private(page, 0);
587 * This function checks whether a page is free && is the buddy
588 * we can do coalesce a page and its buddy if
589 * (a) the buddy is not in a hole &&
590 * (b) the buddy is in the buddy system &&
591 * (c) a page and its buddy have the same order &&
592 * (d) a page and its buddy are in the same zone.
594 * For recording whether a page is in the buddy system, we set ->_mapcount
595 * PAGE_BUDDY_MAPCOUNT_VALUE.
596 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
597 * serialized by zone->lock.
599 * For recording page's order, we use page_private(page).
601 static inline int page_is_buddy(struct page *page, struct page *buddy,
604 if (!pfn_valid_within(page_to_pfn(buddy)))
607 if (page_is_guard(buddy) && page_order(buddy) == order) {
608 if (page_zone_id(page) != page_zone_id(buddy))
611 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
616 if (PageBuddy(buddy) && page_order(buddy) == order) {
618 * zone check is done late to avoid uselessly
619 * calculating zone/node ids for pages that could
622 if (page_zone_id(page) != page_zone_id(buddy))
625 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
633 * Freeing function for a buddy system allocator.
635 * The concept of a buddy system is to maintain direct-mapped table
636 * (containing bit values) for memory blocks of various "orders".
637 * The bottom level table contains the map for the smallest allocatable
638 * units of memory (here, pages), and each level above it describes
639 * pairs of units from the levels below, hence, "buddies".
640 * At a high level, all that happens here is marking the table entry
641 * at the bottom level available, and propagating the changes upward
642 * as necessary, plus some accounting needed to play nicely with other
643 * parts of the VM system.
644 * At each level, we keep a list of pages, which are heads of continuous
645 * free pages of length of (1 << order) and marked with _mapcount
646 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
648 * So when we are allocating or freeing one, we can derive the state of the
649 * other. That is, if we allocate a small block, and both were
650 * free, the remainder of the region must be split into blocks.
651 * If a block is freed, and its buddy is also free, then this
652 * triggers coalescing into a block of larger size.
657 static inline void __free_one_page(struct page *page,
659 struct zone *zone, unsigned int order,
662 unsigned long page_idx;
663 unsigned long combined_idx;
664 unsigned long uninitialized_var(buddy_idx);
666 unsigned int max_order = MAX_ORDER;
668 VM_BUG_ON(!zone_is_initialized(zone));
669 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
671 VM_BUG_ON(migratetype == -1);
672 if (is_migrate_isolate(migratetype)) {
674 * We restrict max order of merging to prevent merge
675 * between freepages on isolate pageblock and normal
676 * pageblock. Without this, pageblock isolation
677 * could cause incorrect freepage accounting.
679 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
681 __mod_zone_freepage_state(zone, 1 << order, migratetype);
684 page_idx = pfn & ((1 << max_order) - 1);
686 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
687 VM_BUG_ON_PAGE(bad_range(zone, page), page);
689 while (order < max_order - 1) {
690 buddy_idx = __find_buddy_index(page_idx, order);
691 buddy = page + (buddy_idx - page_idx);
692 if (!page_is_buddy(page, buddy, order))
695 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
696 * merge with it and move up one order.
698 if (page_is_guard(buddy)) {
699 clear_page_guard(zone, buddy, order, migratetype);
701 list_del(&buddy->lru);
702 zone->free_area[order].nr_free--;
703 rmv_page_order(buddy);
705 combined_idx = buddy_idx & page_idx;
706 page = page + (combined_idx - page_idx);
707 page_idx = combined_idx;
710 set_page_order(page, order);
713 * If this is not the largest possible page, check if the buddy
714 * of the next-highest order is free. If it is, it's possible
715 * that pages are being freed that will coalesce soon. In case,
716 * that is happening, add the free page to the tail of the list
717 * so it's less likely to be used soon and more likely to be merged
718 * as a higher order page
720 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
721 struct page *higher_page, *higher_buddy;
722 combined_idx = buddy_idx & page_idx;
723 higher_page = page + (combined_idx - page_idx);
724 buddy_idx = __find_buddy_index(combined_idx, order + 1);
725 higher_buddy = higher_page + (buddy_idx - combined_idx);
726 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
727 list_add_tail(&page->lru,
728 &zone->free_area[order].free_list[migratetype]);
733 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
735 zone->free_area[order].nr_free++;
738 static inline int free_pages_check(struct page *page)
740 const char *bad_reason = NULL;
741 unsigned long bad_flags = 0;
743 if (unlikely(page_mapcount(page)))
744 bad_reason = "nonzero mapcount";
745 if (unlikely(page->mapping != NULL))
746 bad_reason = "non-NULL mapping";
747 if (unlikely(atomic_read(&page->_count) != 0))
748 bad_reason = "nonzero _count";
749 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
750 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
751 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
754 if (unlikely(page->mem_cgroup))
755 bad_reason = "page still charged to cgroup";
757 if (unlikely(bad_reason)) {
758 bad_page(page, bad_reason, bad_flags);
761 page_cpupid_reset_last(page);
762 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
763 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
768 * Frees a number of pages from the PCP lists
769 * Assumes all pages on list are in same zone, and of same order.
770 * count is the number of pages to free.
772 * If the zone was previously in an "all pages pinned" state then look to
773 * see if this freeing clears that state.
775 * And clear the zone's pages_scanned counter, to hold off the "all pages are
776 * pinned" detection logic.
778 static void free_pcppages_bulk(struct zone *zone, int count,
779 struct per_cpu_pages *pcp)
784 unsigned long nr_scanned;
786 spin_lock(&zone->lock);
787 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
789 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
793 struct list_head *list;
796 * Remove pages from lists in a round-robin fashion. A
797 * batch_free count is maintained that is incremented when an
798 * empty list is encountered. This is so more pages are freed
799 * off fuller lists instead of spinning excessively around empty
804 if (++migratetype == MIGRATE_PCPTYPES)
806 list = &pcp->lists[migratetype];
807 } while (list_empty(list));
809 /* This is the only non-empty list. Free them all. */
810 if (batch_free == MIGRATE_PCPTYPES)
811 batch_free = to_free;
814 int mt; /* migratetype of the to-be-freed page */
816 page = list_entry(list->prev, struct page, lru);
817 /* must delete as __free_one_page list manipulates */
818 list_del(&page->lru);
820 mt = get_pcppage_migratetype(page);
821 /* MIGRATE_ISOLATE page should not go to pcplists */
822 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
823 /* Pageblock could have been isolated meanwhile */
824 if (unlikely(has_isolate_pageblock(zone)))
825 mt = get_pageblock_migratetype(page);
827 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
828 trace_mm_page_pcpu_drain(page, 0, mt);
829 } while (--to_free && --batch_free && !list_empty(list));
831 spin_unlock(&zone->lock);
834 static void free_one_page(struct zone *zone,
835 struct page *page, unsigned long pfn,
839 unsigned long nr_scanned;
840 spin_lock(&zone->lock);
841 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
843 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
845 if (unlikely(has_isolate_pageblock(zone) ||
846 is_migrate_isolate(migratetype))) {
847 migratetype = get_pfnblock_migratetype(page, pfn);
849 __free_one_page(page, pfn, zone, order, migratetype);
850 spin_unlock(&zone->lock);
853 static int free_tail_pages_check(struct page *head_page, struct page *page)
858 * We rely page->lru.next never has bit 0 set, unless the page
859 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
861 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
863 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
867 if (page->mapping != TAIL_MAPPING) {
868 bad_page(page, "corrupted mapping in tail page", 0);
871 if (unlikely(!PageTail(page))) {
872 bad_page(page, "PageTail not set", 0);
875 if (unlikely(compound_head(page) != head_page)) {
876 bad_page(page, "compound_head not consistent", 0);
881 page->mapping = NULL;
882 clear_compound_head(page);
886 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
887 unsigned long zone, int nid)
889 set_page_links(page, zone, nid, pfn);
890 init_page_count(page);
891 page_mapcount_reset(page);
892 page_cpupid_reset_last(page);
894 INIT_LIST_HEAD(&page->lru);
895 #ifdef WANT_PAGE_VIRTUAL
896 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
897 if (!is_highmem_idx(zone))
898 set_page_address(page, __va(pfn << PAGE_SHIFT));
902 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
905 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
908 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
909 static void init_reserved_page(unsigned long pfn)
914 if (!early_page_uninitialised(pfn))
917 nid = early_pfn_to_nid(pfn);
918 pgdat = NODE_DATA(nid);
920 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
921 struct zone *zone = &pgdat->node_zones[zid];
923 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
926 __init_single_pfn(pfn, zid, nid);
929 static inline void init_reserved_page(unsigned long pfn)
932 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
935 * Initialised pages do not have PageReserved set. This function is
936 * called for each range allocated by the bootmem allocator and
937 * marks the pages PageReserved. The remaining valid pages are later
938 * sent to the buddy page allocator.
940 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
942 unsigned long start_pfn = PFN_DOWN(start);
943 unsigned long end_pfn = PFN_UP(end);
945 for (; start_pfn < end_pfn; start_pfn++) {
946 if (pfn_valid(start_pfn)) {
947 struct page *page = pfn_to_page(start_pfn);
949 init_reserved_page(start_pfn);
951 /* Avoid false-positive PageTail() */
952 INIT_LIST_HEAD(&page->lru);
954 SetPageReserved(page);
959 static bool free_pages_prepare(struct page *page, unsigned int order)
961 bool compound = PageCompound(page);
964 VM_BUG_ON_PAGE(PageTail(page), page);
965 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
967 trace_mm_page_free(page, order);
968 kmemcheck_free_shadow(page, order);
969 kasan_free_pages(page, order);
972 page->mapping = NULL;
973 bad += free_pages_check(page);
974 for (i = 1; i < (1 << order); i++) {
976 bad += free_tail_pages_check(page, page + i);
977 bad += free_pages_check(page + i);
982 reset_page_owner(page, order);
984 if (!PageHighMem(page)) {
985 debug_check_no_locks_freed(page_address(page),
987 debug_check_no_obj_freed(page_address(page),
990 arch_free_page(page, order);
991 kernel_map_pages(page, 1 << order, 0);
996 static void __free_pages_ok(struct page *page, unsigned int order)
1000 unsigned long pfn = page_to_pfn(page);
1002 if (!free_pages_prepare(page, order))
1005 migratetype = get_pfnblock_migratetype(page, pfn);
1006 local_irq_save(flags);
1007 __count_vm_events(PGFREE, 1 << order);
1008 free_one_page(page_zone(page), page, pfn, order, migratetype);
1009 local_irq_restore(flags);
1012 static void __init __free_pages_boot_core(struct page *page,
1013 unsigned long pfn, unsigned int order)
1015 unsigned int nr_pages = 1 << order;
1016 struct page *p = page;
1020 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1022 __ClearPageReserved(p);
1023 set_page_count(p, 0);
1025 __ClearPageReserved(p);
1026 set_page_count(p, 0);
1028 page_zone(page)->managed_pages += nr_pages;
1029 set_page_refcounted(page);
1030 __free_pages(page, order);
1033 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1034 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1036 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1038 int __meminit early_pfn_to_nid(unsigned long pfn)
1040 static DEFINE_SPINLOCK(early_pfn_lock);
1043 spin_lock(&early_pfn_lock);
1044 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1047 spin_unlock(&early_pfn_lock);
1053 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1054 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1055 struct mminit_pfnnid_cache *state)
1059 nid = __early_pfn_to_nid(pfn, state);
1060 if (nid >= 0 && nid != node)
1065 /* Only safe to use early in boot when initialisation is single-threaded */
1066 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1068 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1073 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1077 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1078 struct mminit_pfnnid_cache *state)
1085 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1088 if (early_page_uninitialised(pfn))
1090 return __free_pages_boot_core(page, pfn, order);
1093 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1094 static void __init deferred_free_range(struct page *page,
1095 unsigned long pfn, int nr_pages)
1102 /* Free a large naturally-aligned chunk if possible */
1103 if (nr_pages == MAX_ORDER_NR_PAGES &&
1104 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1105 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1106 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1110 for (i = 0; i < nr_pages; i++, page++, pfn++)
1111 __free_pages_boot_core(page, pfn, 0);
1114 /* Completion tracking for deferred_init_memmap() threads */
1115 static atomic_t pgdat_init_n_undone __initdata;
1116 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1118 static inline void __init pgdat_init_report_one_done(void)
1120 if (atomic_dec_and_test(&pgdat_init_n_undone))
1121 complete(&pgdat_init_all_done_comp);
1124 /* Initialise remaining memory on a node */
1125 static int __init deferred_init_memmap(void *data)
1127 pg_data_t *pgdat = data;
1128 int nid = pgdat->node_id;
1129 struct mminit_pfnnid_cache nid_init_state = { };
1130 unsigned long start = jiffies;
1131 unsigned long nr_pages = 0;
1132 unsigned long walk_start, walk_end;
1135 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1136 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1138 if (first_init_pfn == ULONG_MAX) {
1139 pgdat_init_report_one_done();
1143 /* Bind memory initialisation thread to a local node if possible */
1144 if (!cpumask_empty(cpumask))
1145 set_cpus_allowed_ptr(current, cpumask);
1147 /* Sanity check boundaries */
1148 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1149 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1150 pgdat->first_deferred_pfn = ULONG_MAX;
1152 /* Only the highest zone is deferred so find it */
1153 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1154 zone = pgdat->node_zones + zid;
1155 if (first_init_pfn < zone_end_pfn(zone))
1159 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1160 unsigned long pfn, end_pfn;
1161 struct page *page = NULL;
1162 struct page *free_base_page = NULL;
1163 unsigned long free_base_pfn = 0;
1166 end_pfn = min(walk_end, zone_end_pfn(zone));
1167 pfn = first_init_pfn;
1168 if (pfn < walk_start)
1170 if (pfn < zone->zone_start_pfn)
1171 pfn = zone->zone_start_pfn;
1173 for (; pfn < end_pfn; pfn++) {
1174 if (!pfn_valid_within(pfn))
1178 * Ensure pfn_valid is checked every
1179 * MAX_ORDER_NR_PAGES for memory holes
1181 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1182 if (!pfn_valid(pfn)) {
1188 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1193 /* Minimise pfn page lookups and scheduler checks */
1194 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1197 nr_pages += nr_to_free;
1198 deferred_free_range(free_base_page,
1199 free_base_pfn, nr_to_free);
1200 free_base_page = NULL;
1201 free_base_pfn = nr_to_free = 0;
1203 page = pfn_to_page(pfn);
1208 VM_BUG_ON(page_zone(page) != zone);
1212 __init_single_page(page, pfn, zid, nid);
1213 if (!free_base_page) {
1214 free_base_page = page;
1215 free_base_pfn = pfn;
1220 /* Where possible, batch up pages for a single free */
1223 /* Free the current block of pages to allocator */
1224 nr_pages += nr_to_free;
1225 deferred_free_range(free_base_page, free_base_pfn,
1227 free_base_page = NULL;
1228 free_base_pfn = nr_to_free = 0;
1231 first_init_pfn = max(end_pfn, first_init_pfn);
1234 /* Sanity check that the next zone really is unpopulated */
1235 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1237 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1238 jiffies_to_msecs(jiffies - start));
1240 pgdat_init_report_one_done();
1244 void __init page_alloc_init_late(void)
1248 /* There will be num_node_state(N_MEMORY) threads */
1249 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1250 for_each_node_state(nid, N_MEMORY) {
1251 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1254 /* Block until all are initialised */
1255 wait_for_completion(&pgdat_init_all_done_comp);
1257 /* Reinit limits that are based on free pages after the kernel is up */
1258 files_maxfiles_init();
1260 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1263 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1264 void __init init_cma_reserved_pageblock(struct page *page)
1266 unsigned i = pageblock_nr_pages;
1267 struct page *p = page;
1270 __ClearPageReserved(p);
1271 set_page_count(p, 0);
1274 set_pageblock_migratetype(page, MIGRATE_CMA);
1276 if (pageblock_order >= MAX_ORDER) {
1277 i = pageblock_nr_pages;
1280 set_page_refcounted(p);
1281 __free_pages(p, MAX_ORDER - 1);
1282 p += MAX_ORDER_NR_PAGES;
1283 } while (i -= MAX_ORDER_NR_PAGES);
1285 set_page_refcounted(page);
1286 __free_pages(page, pageblock_order);
1289 adjust_managed_page_count(page, pageblock_nr_pages);
1294 * The order of subdivision here is critical for the IO subsystem.
1295 * Please do not alter this order without good reasons and regression
1296 * testing. Specifically, as large blocks of memory are subdivided,
1297 * the order in which smaller blocks are delivered depends on the order
1298 * they're subdivided in this function. This is the primary factor
1299 * influencing the order in which pages are delivered to the IO
1300 * subsystem according to empirical testing, and this is also justified
1301 * by considering the behavior of a buddy system containing a single
1302 * large block of memory acted on by a series of small allocations.
1303 * This behavior is a critical factor in sglist merging's success.
1307 static inline void expand(struct zone *zone, struct page *page,
1308 int low, int high, struct free_area *area,
1311 unsigned long size = 1 << high;
1313 while (high > low) {
1317 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1319 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1320 debug_guardpage_enabled() &&
1321 high < debug_guardpage_minorder()) {
1323 * Mark as guard pages (or page), that will allow to
1324 * merge back to allocator when buddy will be freed.
1325 * Corresponding page table entries will not be touched,
1326 * pages will stay not present in virtual address space
1328 set_page_guard(zone, &page[size], high, migratetype);
1331 list_add(&page[size].lru, &area->free_list[migratetype]);
1333 set_page_order(&page[size], high);
1338 * This page is about to be returned from the page allocator
1340 static inline int check_new_page(struct page *page)
1342 const char *bad_reason = NULL;
1343 unsigned long bad_flags = 0;
1345 if (unlikely(page_mapcount(page)))
1346 bad_reason = "nonzero mapcount";
1347 if (unlikely(page->mapping != NULL))
1348 bad_reason = "non-NULL mapping";
1349 if (unlikely(atomic_read(&page->_count) != 0))
1350 bad_reason = "nonzero _count";
1351 if (unlikely(page->flags & __PG_HWPOISON)) {
1352 bad_reason = "HWPoisoned (hardware-corrupted)";
1353 bad_flags = __PG_HWPOISON;
1355 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1356 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1357 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1360 if (unlikely(page->mem_cgroup))
1361 bad_reason = "page still charged to cgroup";
1363 if (unlikely(bad_reason)) {
1364 bad_page(page, bad_reason, bad_flags);
1370 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1375 for (i = 0; i < (1 << order); i++) {
1376 struct page *p = page + i;
1377 if (unlikely(check_new_page(p)))
1381 set_page_private(page, 0);
1382 set_page_refcounted(page);
1384 arch_alloc_page(page, order);
1385 kernel_map_pages(page, 1 << order, 1);
1386 kasan_alloc_pages(page, order);
1388 if (gfp_flags & __GFP_ZERO)
1389 for (i = 0; i < (1 << order); i++)
1390 clear_highpage(page + i);
1392 if (order && (gfp_flags & __GFP_COMP))
1393 prep_compound_page(page, order);
1395 set_page_owner(page, order, gfp_flags);
1398 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1399 * allocate the page. The expectation is that the caller is taking
1400 * steps that will free more memory. The caller should avoid the page
1401 * being used for !PFMEMALLOC purposes.
1403 if (alloc_flags & ALLOC_NO_WATERMARKS)
1404 set_page_pfmemalloc(page);
1406 clear_page_pfmemalloc(page);
1412 * Go through the free lists for the given migratetype and remove
1413 * the smallest available page from the freelists
1416 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1419 unsigned int current_order;
1420 struct free_area *area;
1423 /* Find a page of the appropriate size in the preferred list */
1424 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1425 area = &(zone->free_area[current_order]);
1426 if (list_empty(&area->free_list[migratetype]))
1429 page = list_entry(area->free_list[migratetype].next,
1431 list_del(&page->lru);
1432 rmv_page_order(page);
1434 expand(zone, page, order, current_order, area, migratetype);
1435 set_pcppage_migratetype(page, migratetype);
1444 * This array describes the order lists are fallen back to when
1445 * the free lists for the desirable migrate type are depleted
1447 static int fallbacks[MIGRATE_TYPES][4] = {
1448 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1449 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1450 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1452 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1454 #ifdef CONFIG_MEMORY_ISOLATION
1455 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1460 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1463 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1466 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1467 unsigned int order) { return NULL; }
1471 * Move the free pages in a range to the free lists of the requested type.
1472 * Note that start_page and end_pages are not aligned on a pageblock
1473 * boundary. If alignment is required, use move_freepages_block()
1475 int move_freepages(struct zone *zone,
1476 struct page *start_page, struct page *end_page,
1481 int pages_moved = 0;
1483 #ifndef CONFIG_HOLES_IN_ZONE
1485 * page_zone is not safe to call in this context when
1486 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1487 * anyway as we check zone boundaries in move_freepages_block().
1488 * Remove at a later date when no bug reports exist related to
1489 * grouping pages by mobility
1491 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1494 for (page = start_page; page <= end_page;) {
1495 /* Make sure we are not inadvertently changing nodes */
1496 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1498 if (!pfn_valid_within(page_to_pfn(page))) {
1503 if (!PageBuddy(page)) {
1508 order = page_order(page);
1509 list_move(&page->lru,
1510 &zone->free_area[order].free_list[migratetype]);
1512 pages_moved += 1 << order;
1518 int move_freepages_block(struct zone *zone, struct page *page,
1521 unsigned long start_pfn, end_pfn;
1522 struct page *start_page, *end_page;
1524 start_pfn = page_to_pfn(page);
1525 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1526 start_page = pfn_to_page(start_pfn);
1527 end_page = start_page + pageblock_nr_pages - 1;
1528 end_pfn = start_pfn + pageblock_nr_pages - 1;
1530 /* Do not cross zone boundaries */
1531 if (!zone_spans_pfn(zone, start_pfn))
1533 if (!zone_spans_pfn(zone, end_pfn))
1536 return move_freepages(zone, start_page, end_page, migratetype);
1539 static void change_pageblock_range(struct page *pageblock_page,
1540 int start_order, int migratetype)
1542 int nr_pageblocks = 1 << (start_order - pageblock_order);
1544 while (nr_pageblocks--) {
1545 set_pageblock_migratetype(pageblock_page, migratetype);
1546 pageblock_page += pageblock_nr_pages;
1551 * When we are falling back to another migratetype during allocation, try to
1552 * steal extra free pages from the same pageblocks to satisfy further
1553 * allocations, instead of polluting multiple pageblocks.
1555 * If we are stealing a relatively large buddy page, it is likely there will
1556 * be more free pages in the pageblock, so try to steal them all. For
1557 * reclaimable and unmovable allocations, we steal regardless of page size,
1558 * as fragmentation caused by those allocations polluting movable pageblocks
1559 * is worse than movable allocations stealing from unmovable and reclaimable
1562 static bool can_steal_fallback(unsigned int order, int start_mt)
1565 * Leaving this order check is intended, although there is
1566 * relaxed order check in next check. The reason is that
1567 * we can actually steal whole pageblock if this condition met,
1568 * but, below check doesn't guarantee it and that is just heuristic
1569 * so could be changed anytime.
1571 if (order >= pageblock_order)
1574 if (order >= pageblock_order / 2 ||
1575 start_mt == MIGRATE_RECLAIMABLE ||
1576 start_mt == MIGRATE_UNMOVABLE ||
1577 page_group_by_mobility_disabled)
1584 * This function implements actual steal behaviour. If order is large enough,
1585 * we can steal whole pageblock. If not, we first move freepages in this
1586 * pageblock and check whether half of pages are moved or not. If half of
1587 * pages are moved, we can change migratetype of pageblock and permanently
1588 * use it's pages as requested migratetype in the future.
1590 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1593 unsigned int current_order = page_order(page);
1596 /* Take ownership for orders >= pageblock_order */
1597 if (current_order >= pageblock_order) {
1598 change_pageblock_range(page, current_order, start_type);
1602 pages = move_freepages_block(zone, page, start_type);
1604 /* Claim the whole block if over half of it is free */
1605 if (pages >= (1 << (pageblock_order-1)) ||
1606 page_group_by_mobility_disabled)
1607 set_pageblock_migratetype(page, start_type);
1611 * Check whether there is a suitable fallback freepage with requested order.
1612 * If only_stealable is true, this function returns fallback_mt only if
1613 * we can steal other freepages all together. This would help to reduce
1614 * fragmentation due to mixed migratetype pages in one pageblock.
1616 int find_suitable_fallback(struct free_area *area, unsigned int order,
1617 int migratetype, bool only_stealable, bool *can_steal)
1622 if (area->nr_free == 0)
1627 fallback_mt = fallbacks[migratetype][i];
1628 if (fallback_mt == MIGRATE_TYPES)
1631 if (list_empty(&area->free_list[fallback_mt]))
1634 if (can_steal_fallback(order, migratetype))
1637 if (!only_stealable)
1648 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1649 * there are no empty page blocks that contain a page with a suitable order
1651 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1652 unsigned int alloc_order)
1655 unsigned long max_managed, flags;
1658 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1659 * Check is race-prone but harmless.
1661 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1662 if (zone->nr_reserved_highatomic >= max_managed)
1665 spin_lock_irqsave(&zone->lock, flags);
1667 /* Recheck the nr_reserved_highatomic limit under the lock */
1668 if (zone->nr_reserved_highatomic >= max_managed)
1672 mt = get_pageblock_migratetype(page);
1673 if (mt != MIGRATE_HIGHATOMIC &&
1674 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1675 zone->nr_reserved_highatomic += pageblock_nr_pages;
1676 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1677 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1681 spin_unlock_irqrestore(&zone->lock, flags);
1685 * Used when an allocation is about to fail under memory pressure. This
1686 * potentially hurts the reliability of high-order allocations when under
1687 * intense memory pressure but failed atomic allocations should be easier
1688 * to recover from than an OOM.
1690 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1692 struct zonelist *zonelist = ac->zonelist;
1693 unsigned long flags;
1699 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1701 /* Preserve at least one pageblock */
1702 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1705 spin_lock_irqsave(&zone->lock, flags);
1706 for (order = 0; order < MAX_ORDER; order++) {
1707 struct free_area *area = &(zone->free_area[order]);
1709 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1712 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1716 * It should never happen but changes to locking could
1717 * inadvertently allow a per-cpu drain to add pages
1718 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1719 * and watch for underflows.
1721 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1722 zone->nr_reserved_highatomic);
1725 * Convert to ac->migratetype and avoid the normal
1726 * pageblock stealing heuristics. Minimally, the caller
1727 * is doing the work and needs the pages. More
1728 * importantly, if the block was always converted to
1729 * MIGRATE_UNMOVABLE or another type then the number
1730 * of pageblocks that cannot be completely freed
1733 set_pageblock_migratetype(page, ac->migratetype);
1734 move_freepages_block(zone, page, ac->migratetype);
1735 spin_unlock_irqrestore(&zone->lock, flags);
1738 spin_unlock_irqrestore(&zone->lock, flags);
1742 /* Remove an element from the buddy allocator from the fallback list */
1743 static inline struct page *
1744 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1746 struct free_area *area;
1747 unsigned int current_order;
1752 /* Find the largest possible block of pages in the other list */
1753 for (current_order = MAX_ORDER-1;
1754 current_order >= order && current_order <= MAX_ORDER-1;
1756 area = &(zone->free_area[current_order]);
1757 fallback_mt = find_suitable_fallback(area, current_order,
1758 start_migratetype, false, &can_steal);
1759 if (fallback_mt == -1)
1762 page = list_entry(area->free_list[fallback_mt].next,
1765 steal_suitable_fallback(zone, page, start_migratetype);
1767 /* Remove the page from the freelists */
1769 list_del(&page->lru);
1770 rmv_page_order(page);
1772 expand(zone, page, order, current_order, area,
1775 * The pcppage_migratetype may differ from pageblock's
1776 * migratetype depending on the decisions in
1777 * find_suitable_fallback(). This is OK as long as it does not
1778 * differ for MIGRATE_CMA pageblocks. Those can be used as
1779 * fallback only via special __rmqueue_cma_fallback() function
1781 set_pcppage_migratetype(page, start_migratetype);
1783 trace_mm_page_alloc_extfrag(page, order, current_order,
1784 start_migratetype, fallback_mt);
1793 * Do the hard work of removing an element from the buddy allocator.
1794 * Call me with the zone->lock already held.
1796 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1797 int migratetype, gfp_t gfp_flags)
1801 page = __rmqueue_smallest(zone, order, migratetype);
1802 if (unlikely(!page)) {
1803 if (migratetype == MIGRATE_MOVABLE)
1804 page = __rmqueue_cma_fallback(zone, order);
1807 page = __rmqueue_fallback(zone, order, migratetype);
1810 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1815 * Obtain a specified number of elements from the buddy allocator, all under
1816 * a single hold of the lock, for efficiency. Add them to the supplied list.
1817 * Returns the number of new pages which were placed at *list.
1819 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1820 unsigned long count, struct list_head *list,
1821 int migratetype, bool cold)
1825 spin_lock(&zone->lock);
1826 for (i = 0; i < count; ++i) {
1827 struct page *page = __rmqueue(zone, order, migratetype, 0);
1828 if (unlikely(page == NULL))
1832 * Split buddy pages returned by expand() are received here
1833 * in physical page order. The page is added to the callers and
1834 * list and the list head then moves forward. From the callers
1835 * perspective, the linked list is ordered by page number in
1836 * some conditions. This is useful for IO devices that can
1837 * merge IO requests if the physical pages are ordered
1841 list_add(&page->lru, list);
1843 list_add_tail(&page->lru, list);
1845 if (is_migrate_cma(get_pcppage_migratetype(page)))
1846 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1849 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1850 spin_unlock(&zone->lock);
1856 * Called from the vmstat counter updater to drain pagesets of this
1857 * currently executing processor on remote nodes after they have
1860 * Note that this function must be called with the thread pinned to
1861 * a single processor.
1863 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1865 unsigned long flags;
1866 int to_drain, batch;
1868 local_irq_save(flags);
1869 batch = READ_ONCE(pcp->batch);
1870 to_drain = min(pcp->count, batch);
1872 free_pcppages_bulk(zone, to_drain, pcp);
1873 pcp->count -= to_drain;
1875 local_irq_restore(flags);
1880 * Drain pcplists of the indicated processor and zone.
1882 * The processor must either be the current processor and the
1883 * thread pinned to the current processor or a processor that
1886 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1888 unsigned long flags;
1889 struct per_cpu_pageset *pset;
1890 struct per_cpu_pages *pcp;
1892 local_irq_save(flags);
1893 pset = per_cpu_ptr(zone->pageset, cpu);
1897 free_pcppages_bulk(zone, pcp->count, pcp);
1900 local_irq_restore(flags);
1904 * Drain pcplists of all zones on the indicated processor.
1906 * The processor must either be the current processor and the
1907 * thread pinned to the current processor or a processor that
1910 static void drain_pages(unsigned int cpu)
1914 for_each_populated_zone(zone) {
1915 drain_pages_zone(cpu, zone);
1920 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1922 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1923 * the single zone's pages.
1925 void drain_local_pages(struct zone *zone)
1927 int cpu = smp_processor_id();
1930 drain_pages_zone(cpu, zone);
1936 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1938 * When zone parameter is non-NULL, spill just the single zone's pages.
1940 * Note that this code is protected against sending an IPI to an offline
1941 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1942 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1943 * nothing keeps CPUs from showing up after we populated the cpumask and
1944 * before the call to on_each_cpu_mask().
1946 void drain_all_pages(struct zone *zone)
1951 * Allocate in the BSS so we wont require allocation in
1952 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1954 static cpumask_t cpus_with_pcps;
1957 * We don't care about racing with CPU hotplug event
1958 * as offline notification will cause the notified
1959 * cpu to drain that CPU pcps and on_each_cpu_mask
1960 * disables preemption as part of its processing
1962 for_each_online_cpu(cpu) {
1963 struct per_cpu_pageset *pcp;
1965 bool has_pcps = false;
1968 pcp = per_cpu_ptr(zone->pageset, cpu);
1972 for_each_populated_zone(z) {
1973 pcp = per_cpu_ptr(z->pageset, cpu);
1974 if (pcp->pcp.count) {
1982 cpumask_set_cpu(cpu, &cpus_with_pcps);
1984 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1986 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1990 #ifdef CONFIG_HIBERNATION
1992 void mark_free_pages(struct zone *zone)
1994 unsigned long pfn, max_zone_pfn;
1995 unsigned long flags;
1996 unsigned int order, t;
1997 struct list_head *curr;
1999 if (zone_is_empty(zone))
2002 spin_lock_irqsave(&zone->lock, flags);
2004 max_zone_pfn = zone_end_pfn(zone);
2005 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2006 if (pfn_valid(pfn)) {
2007 struct page *page = pfn_to_page(pfn);
2009 if (!swsusp_page_is_forbidden(page))
2010 swsusp_unset_page_free(page);
2013 for_each_migratetype_order(order, t) {
2014 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2017 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2018 for (i = 0; i < (1UL << order); i++)
2019 swsusp_set_page_free(pfn_to_page(pfn + i));
2022 spin_unlock_irqrestore(&zone->lock, flags);
2024 #endif /* CONFIG_PM */
2027 * Free a 0-order page
2028 * cold == true ? free a cold page : free a hot page
2030 void free_hot_cold_page(struct page *page, bool cold)
2032 struct zone *zone = page_zone(page);
2033 struct per_cpu_pages *pcp;
2034 unsigned long flags;
2035 unsigned long pfn = page_to_pfn(page);
2038 if (!free_pages_prepare(page, 0))
2041 migratetype = get_pfnblock_migratetype(page, pfn);
2042 set_pcppage_migratetype(page, migratetype);
2043 local_irq_save(flags);
2044 __count_vm_event(PGFREE);
2047 * We only track unmovable, reclaimable and movable on pcp lists.
2048 * Free ISOLATE pages back to the allocator because they are being
2049 * offlined but treat RESERVE as movable pages so we can get those
2050 * areas back if necessary. Otherwise, we may have to free
2051 * excessively into the page allocator
2053 if (migratetype >= MIGRATE_PCPTYPES) {
2054 if (unlikely(is_migrate_isolate(migratetype))) {
2055 free_one_page(zone, page, pfn, 0, migratetype);
2058 migratetype = MIGRATE_MOVABLE;
2061 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2063 list_add(&page->lru, &pcp->lists[migratetype]);
2065 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2067 if (pcp->count >= pcp->high) {
2068 unsigned long batch = READ_ONCE(pcp->batch);
2069 free_pcppages_bulk(zone, batch, pcp);
2070 pcp->count -= batch;
2074 local_irq_restore(flags);
2078 * Free a list of 0-order pages
2080 void free_hot_cold_page_list(struct list_head *list, bool cold)
2082 struct page *page, *next;
2084 list_for_each_entry_safe(page, next, list, lru) {
2085 trace_mm_page_free_batched(page, cold);
2086 free_hot_cold_page(page, cold);
2091 * split_page takes a non-compound higher-order page, and splits it into
2092 * n (1<<order) sub-pages: page[0..n]
2093 * Each sub-page must be freed individually.
2095 * Note: this is probably too low level an operation for use in drivers.
2096 * Please consult with lkml before using this in your driver.
2098 void split_page(struct page *page, unsigned int order)
2103 VM_BUG_ON_PAGE(PageCompound(page), page);
2104 VM_BUG_ON_PAGE(!page_count(page), page);
2106 #ifdef CONFIG_KMEMCHECK
2108 * Split shadow pages too, because free(page[0]) would
2109 * otherwise free the whole shadow.
2111 if (kmemcheck_page_is_tracked(page))
2112 split_page(virt_to_page(page[0].shadow), order);
2115 gfp_mask = get_page_owner_gfp(page);
2116 set_page_owner(page, 0, gfp_mask);
2117 for (i = 1; i < (1 << order); i++) {
2118 set_page_refcounted(page + i);
2119 set_page_owner(page + i, 0, gfp_mask);
2122 EXPORT_SYMBOL_GPL(split_page);
2124 int __isolate_free_page(struct page *page, unsigned int order)
2126 unsigned long watermark;
2130 BUG_ON(!PageBuddy(page));
2132 zone = page_zone(page);
2133 mt = get_pageblock_migratetype(page);
2135 if (!is_migrate_isolate(mt)) {
2136 /* Obey watermarks as if the page was being allocated */
2137 watermark = low_wmark_pages(zone) + (1 << order);
2138 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2141 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2144 /* Remove page from free list */
2145 list_del(&page->lru);
2146 zone->free_area[order].nr_free--;
2147 rmv_page_order(page);
2149 set_page_owner(page, order, __GFP_MOVABLE);
2151 /* Set the pageblock if the isolated page is at least a pageblock */
2152 if (order >= pageblock_order - 1) {
2153 struct page *endpage = page + (1 << order) - 1;
2154 for (; page < endpage; page += pageblock_nr_pages) {
2155 int mt = get_pageblock_migratetype(page);
2156 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2157 set_pageblock_migratetype(page,
2163 return 1UL << order;
2167 * Similar to split_page except the page is already free. As this is only
2168 * being used for migration, the migratetype of the block also changes.
2169 * As this is called with interrupts disabled, the caller is responsible
2170 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2173 * Note: this is probably too low level an operation for use in drivers.
2174 * Please consult with lkml before using this in your driver.
2176 int split_free_page(struct page *page)
2181 order = page_order(page);
2183 nr_pages = __isolate_free_page(page, order);
2187 /* Split into individual pages */
2188 set_page_refcounted(page);
2189 split_page(page, order);
2194 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2197 struct page *buffered_rmqueue(struct zone *preferred_zone,
2198 struct zone *zone, unsigned int order,
2199 gfp_t gfp_flags, int alloc_flags, int migratetype)
2201 unsigned long flags;
2203 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2205 if (likely(order == 0)) {
2206 struct per_cpu_pages *pcp;
2207 struct list_head *list;
2209 local_irq_save(flags);
2210 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2211 list = &pcp->lists[migratetype];
2212 if (list_empty(list)) {
2213 pcp->count += rmqueue_bulk(zone, 0,
2216 if (unlikely(list_empty(list)))
2221 page = list_entry(list->prev, struct page, lru);
2223 page = list_entry(list->next, struct page, lru);
2225 list_del(&page->lru);
2228 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2230 * __GFP_NOFAIL is not to be used in new code.
2232 * All __GFP_NOFAIL callers should be fixed so that they
2233 * properly detect and handle allocation failures.
2235 * We most definitely don't want callers attempting to
2236 * allocate greater than order-1 page units with
2239 WARN_ON_ONCE(order > 1);
2241 spin_lock_irqsave(&zone->lock, flags);
2244 if (alloc_flags & ALLOC_HARDER) {
2245 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2247 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2250 page = __rmqueue(zone, order, migratetype, gfp_flags);
2251 spin_unlock(&zone->lock);
2254 __mod_zone_freepage_state(zone, -(1 << order),
2255 get_pcppage_migratetype(page));
2258 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2259 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2260 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2261 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2263 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2264 zone_statistics(preferred_zone, zone, gfp_flags);
2265 local_irq_restore(flags);
2267 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2271 local_irq_restore(flags);
2275 #ifdef CONFIG_FAIL_PAGE_ALLOC
2278 struct fault_attr attr;
2280 u32 ignore_gfp_highmem;
2281 u32 ignore_gfp_reclaim;
2283 } fail_page_alloc = {
2284 .attr = FAULT_ATTR_INITIALIZER,
2285 .ignore_gfp_reclaim = 1,
2286 .ignore_gfp_highmem = 1,
2290 static int __init setup_fail_page_alloc(char *str)
2292 return setup_fault_attr(&fail_page_alloc.attr, str);
2294 __setup("fail_page_alloc=", setup_fail_page_alloc);
2296 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2298 if (order < fail_page_alloc.min_order)
2300 if (gfp_mask & __GFP_NOFAIL)
2302 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2304 if (fail_page_alloc.ignore_gfp_reclaim &&
2305 (gfp_mask & __GFP_DIRECT_RECLAIM))
2308 return should_fail(&fail_page_alloc.attr, 1 << order);
2311 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2313 static int __init fail_page_alloc_debugfs(void)
2315 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2318 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2319 &fail_page_alloc.attr);
2321 return PTR_ERR(dir);
2323 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2324 &fail_page_alloc.ignore_gfp_reclaim))
2326 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2327 &fail_page_alloc.ignore_gfp_highmem))
2329 if (!debugfs_create_u32("min-order", mode, dir,
2330 &fail_page_alloc.min_order))
2335 debugfs_remove_recursive(dir);
2340 late_initcall(fail_page_alloc_debugfs);
2342 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2344 #else /* CONFIG_FAIL_PAGE_ALLOC */
2346 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2351 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2354 * Return true if free base pages are above 'mark'. For high-order checks it
2355 * will return true of the order-0 watermark is reached and there is at least
2356 * one free page of a suitable size. Checking now avoids taking the zone lock
2357 * to check in the allocation paths if no pages are free.
2359 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2360 unsigned long mark, int classzone_idx, int alloc_flags,
2365 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2367 /* free_pages may go negative - that's OK */
2368 free_pages -= (1 << order) - 1;
2370 if (alloc_flags & ALLOC_HIGH)
2374 * If the caller does not have rights to ALLOC_HARDER then subtract
2375 * the high-atomic reserves. This will over-estimate the size of the
2376 * atomic reserve but it avoids a search.
2378 if (likely(!alloc_harder))
2379 free_pages -= z->nr_reserved_highatomic;
2384 /* If allocation can't use CMA areas don't use free CMA pages */
2385 if (!(alloc_flags & ALLOC_CMA))
2386 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2390 * Check watermarks for an order-0 allocation request. If these
2391 * are not met, then a high-order request also cannot go ahead
2392 * even if a suitable page happened to be free.
2394 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2397 /* If this is an order-0 request then the watermark is fine */
2401 /* For a high-order request, check at least one suitable page is free */
2402 for (o = order; o < MAX_ORDER; o++) {
2403 struct free_area *area = &z->free_area[o];
2412 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2413 if (!list_empty(&area->free_list[mt]))
2418 if ((alloc_flags & ALLOC_CMA) &&
2419 !list_empty(&area->free_list[MIGRATE_CMA])) {
2427 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2428 int classzone_idx, int alloc_flags)
2430 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2431 zone_page_state(z, NR_FREE_PAGES));
2434 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2435 unsigned long mark, int classzone_idx)
2437 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2439 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2440 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2442 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2447 static bool zone_local(struct zone *local_zone, struct zone *zone)
2449 return local_zone->node == zone->node;
2452 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2454 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2457 #else /* CONFIG_NUMA */
2458 static bool zone_local(struct zone *local_zone, struct zone *zone)
2463 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2467 #endif /* CONFIG_NUMA */
2469 static void reset_alloc_batches(struct zone *preferred_zone)
2471 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2474 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2475 high_wmark_pages(zone) - low_wmark_pages(zone) -
2476 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2477 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2478 } while (zone++ != preferred_zone);
2482 * get_page_from_freelist goes through the zonelist trying to allocate
2485 static struct page *
2486 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2487 const struct alloc_context *ac)
2489 struct zonelist *zonelist = ac->zonelist;
2491 struct page *page = NULL;
2493 int nr_fair_skipped = 0;
2494 bool zonelist_rescan;
2497 zonelist_rescan = false;
2500 * Scan zonelist, looking for a zone with enough free.
2501 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2503 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2507 if (cpusets_enabled() &&
2508 (alloc_flags & ALLOC_CPUSET) &&
2509 !cpuset_zone_allowed(zone, gfp_mask))
2512 * Distribute pages in proportion to the individual
2513 * zone size to ensure fair page aging. The zone a
2514 * page was allocated in should have no effect on the
2515 * time the page has in memory before being reclaimed.
2517 if (alloc_flags & ALLOC_FAIR) {
2518 if (!zone_local(ac->preferred_zone, zone))
2520 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2526 * When allocating a page cache page for writing, we
2527 * want to get it from a zone that is within its dirty
2528 * limit, such that no single zone holds more than its
2529 * proportional share of globally allowed dirty pages.
2530 * The dirty limits take into account the zone's
2531 * lowmem reserves and high watermark so that kswapd
2532 * should be able to balance it without having to
2533 * write pages from its LRU list.
2535 * This may look like it could increase pressure on
2536 * lower zones by failing allocations in higher zones
2537 * before they are full. But the pages that do spill
2538 * over are limited as the lower zones are protected
2539 * by this very same mechanism. It should not become
2540 * a practical burden to them.
2542 * XXX: For now, allow allocations to potentially
2543 * exceed the per-zone dirty limit in the slowpath
2544 * (spread_dirty_pages unset) before going into reclaim,
2545 * which is important when on a NUMA setup the allowed
2546 * zones are together not big enough to reach the
2547 * global limit. The proper fix for these situations
2548 * will require awareness of zones in the
2549 * dirty-throttling and the flusher threads.
2551 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2554 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2555 if (!zone_watermark_ok(zone, order, mark,
2556 ac->classzone_idx, alloc_flags)) {
2559 /* Checked here to keep the fast path fast */
2560 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2561 if (alloc_flags & ALLOC_NO_WATERMARKS)
2564 if (zone_reclaim_mode == 0 ||
2565 !zone_allows_reclaim(ac->preferred_zone, zone))
2568 ret = zone_reclaim(zone, gfp_mask, order);
2570 case ZONE_RECLAIM_NOSCAN:
2573 case ZONE_RECLAIM_FULL:
2574 /* scanned but unreclaimable */
2577 /* did we reclaim enough */
2578 if (zone_watermark_ok(zone, order, mark,
2579 ac->classzone_idx, alloc_flags))
2587 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2588 gfp_mask, alloc_flags, ac->migratetype);
2590 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2594 * If this is a high-order atomic allocation then check
2595 * if the pageblock should be reserved for the future
2597 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2598 reserve_highatomic_pageblock(page, zone, order);
2605 * The first pass makes sure allocations are spread fairly within the
2606 * local node. However, the local node might have free pages left
2607 * after the fairness batches are exhausted, and remote zones haven't
2608 * even been considered yet. Try once more without fairness, and
2609 * include remote zones now, before entering the slowpath and waking
2610 * kswapd: prefer spilling to a remote zone over swapping locally.
2612 if (alloc_flags & ALLOC_FAIR) {
2613 alloc_flags &= ~ALLOC_FAIR;
2614 if (nr_fair_skipped) {
2615 zonelist_rescan = true;
2616 reset_alloc_batches(ac->preferred_zone);
2618 if (nr_online_nodes > 1)
2619 zonelist_rescan = true;
2622 if (zonelist_rescan)
2629 * Large machines with many possible nodes should not always dump per-node
2630 * meminfo in irq context.
2632 static inline bool should_suppress_show_mem(void)
2637 ret = in_interrupt();
2642 static DEFINE_RATELIMIT_STATE(nopage_rs,
2643 DEFAULT_RATELIMIT_INTERVAL,
2644 DEFAULT_RATELIMIT_BURST);
2646 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2648 unsigned int filter = SHOW_MEM_FILTER_NODES;
2650 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2651 debug_guardpage_minorder() > 0)
2655 * This documents exceptions given to allocations in certain
2656 * contexts that are allowed to allocate outside current's set
2659 if (!(gfp_mask & __GFP_NOMEMALLOC))
2660 if (test_thread_flag(TIF_MEMDIE) ||
2661 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2662 filter &= ~SHOW_MEM_FILTER_NODES;
2663 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2664 filter &= ~SHOW_MEM_FILTER_NODES;
2667 struct va_format vaf;
2670 va_start(args, fmt);
2675 pr_warn("%pV", &vaf);
2680 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2681 current->comm, order, gfp_mask);
2684 if (!should_suppress_show_mem())
2688 static inline struct page *
2689 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2690 const struct alloc_context *ac, unsigned long *did_some_progress)
2692 struct oom_control oc = {
2693 .zonelist = ac->zonelist,
2694 .nodemask = ac->nodemask,
2695 .gfp_mask = gfp_mask,
2700 *did_some_progress = 0;
2703 * Acquire the oom lock. If that fails, somebody else is
2704 * making progress for us.
2706 if (!mutex_trylock(&oom_lock)) {
2707 *did_some_progress = 1;
2708 schedule_timeout_uninterruptible(1);
2713 * Go through the zonelist yet one more time, keep very high watermark
2714 * here, this is only to catch a parallel oom killing, we must fail if
2715 * we're still under heavy pressure.
2717 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2718 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2722 if (!(gfp_mask & __GFP_NOFAIL)) {
2723 /* Coredumps can quickly deplete all memory reserves */
2724 if (current->flags & PF_DUMPCORE)
2726 /* The OOM killer will not help higher order allocs */
2727 if (order > PAGE_ALLOC_COSTLY_ORDER)
2729 /* The OOM killer does not needlessly kill tasks for lowmem */
2730 if (ac->high_zoneidx < ZONE_NORMAL)
2732 /* The OOM killer does not compensate for IO-less reclaim */
2733 if (!(gfp_mask & __GFP_FS)) {
2735 * XXX: Page reclaim didn't yield anything,
2736 * and the OOM killer can't be invoked, but
2737 * keep looping as per tradition.
2739 *did_some_progress = 1;
2742 if (pm_suspended_storage())
2744 /* The OOM killer may not free memory on a specific node */
2745 if (gfp_mask & __GFP_THISNODE)
2748 /* Exhausted what can be done so it's blamo time */
2749 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2750 *did_some_progress = 1;
2752 mutex_unlock(&oom_lock);
2756 #ifdef CONFIG_COMPACTION
2757 /* Try memory compaction for high-order allocations before reclaim */
2758 static struct page *
2759 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2760 int alloc_flags, const struct alloc_context *ac,
2761 enum migrate_mode mode, int *contended_compaction,
2762 bool *deferred_compaction)
2764 unsigned long compact_result;
2770 current->flags |= PF_MEMALLOC;
2771 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2772 mode, contended_compaction);
2773 current->flags &= ~PF_MEMALLOC;
2775 switch (compact_result) {
2776 case COMPACT_DEFERRED:
2777 *deferred_compaction = true;
2779 case COMPACT_SKIPPED:
2786 * At least in one zone compaction wasn't deferred or skipped, so let's
2787 * count a compaction stall
2789 count_vm_event(COMPACTSTALL);
2791 page = get_page_from_freelist(gfp_mask, order,
2792 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2795 struct zone *zone = page_zone(page);
2797 zone->compact_blockskip_flush = false;
2798 compaction_defer_reset(zone, order, true);
2799 count_vm_event(COMPACTSUCCESS);
2804 * It's bad if compaction run occurs and fails. The most likely reason
2805 * is that pages exist, but not enough to satisfy watermarks.
2807 count_vm_event(COMPACTFAIL);
2814 static inline struct page *
2815 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2816 int alloc_flags, const struct alloc_context *ac,
2817 enum migrate_mode mode, int *contended_compaction,
2818 bool *deferred_compaction)
2822 #endif /* CONFIG_COMPACTION */
2824 /* Perform direct synchronous page reclaim */
2826 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2827 const struct alloc_context *ac)
2829 struct reclaim_state reclaim_state;
2834 /* We now go into synchronous reclaim */
2835 cpuset_memory_pressure_bump();
2836 current->flags |= PF_MEMALLOC;
2837 lockdep_set_current_reclaim_state(gfp_mask);
2838 reclaim_state.reclaimed_slab = 0;
2839 current->reclaim_state = &reclaim_state;
2841 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2844 current->reclaim_state = NULL;
2845 lockdep_clear_current_reclaim_state();
2846 current->flags &= ~PF_MEMALLOC;
2853 /* The really slow allocator path where we enter direct reclaim */
2854 static inline struct page *
2855 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2856 int alloc_flags, const struct alloc_context *ac,
2857 unsigned long *did_some_progress)
2859 struct page *page = NULL;
2860 bool drained = false;
2862 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2863 if (unlikely(!(*did_some_progress)))
2867 page = get_page_from_freelist(gfp_mask, order,
2868 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2871 * If an allocation failed after direct reclaim, it could be because
2872 * pages are pinned on the per-cpu lists or in high alloc reserves.
2873 * Shrink them them and try again
2875 if (!page && !drained) {
2876 unreserve_highatomic_pageblock(ac);
2877 drain_all_pages(NULL);
2886 * This is called in the allocator slow-path if the allocation request is of
2887 * sufficient urgency to ignore watermarks and take other desperate measures
2889 static inline struct page *
2890 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2891 const struct alloc_context *ac)
2896 page = get_page_from_freelist(gfp_mask, order,
2897 ALLOC_NO_WATERMARKS, ac);
2899 if (!page && gfp_mask & __GFP_NOFAIL)
2900 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2902 } while (!page && (gfp_mask & __GFP_NOFAIL));
2907 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2912 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2913 ac->high_zoneidx, ac->nodemask)
2914 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2918 gfp_to_alloc_flags(gfp_t gfp_mask)
2920 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2922 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2923 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2926 * The caller may dip into page reserves a bit more if the caller
2927 * cannot run direct reclaim, or if the caller has realtime scheduling
2928 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2929 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2931 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2933 if (gfp_mask & __GFP_ATOMIC) {
2935 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2936 * if it can't schedule.
2938 if (!(gfp_mask & __GFP_NOMEMALLOC))
2939 alloc_flags |= ALLOC_HARDER;
2941 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2942 * comment for __cpuset_node_allowed().
2944 alloc_flags &= ~ALLOC_CPUSET;
2945 } else if (unlikely(rt_task(current)) && !in_interrupt())
2946 alloc_flags |= ALLOC_HARDER;
2948 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2949 if (gfp_mask & __GFP_MEMALLOC)
2950 alloc_flags |= ALLOC_NO_WATERMARKS;
2951 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2952 alloc_flags |= ALLOC_NO_WATERMARKS;
2953 else if (!in_interrupt() &&
2954 ((current->flags & PF_MEMALLOC) ||
2955 unlikely(test_thread_flag(TIF_MEMDIE))))
2956 alloc_flags |= ALLOC_NO_WATERMARKS;
2959 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2960 alloc_flags |= ALLOC_CMA;
2965 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2967 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2970 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2972 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2975 static inline struct page *
2976 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2977 struct alloc_context *ac)
2979 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2980 struct page *page = NULL;
2982 unsigned long pages_reclaimed = 0;
2983 unsigned long did_some_progress;
2984 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2985 bool deferred_compaction = false;
2986 int contended_compaction = COMPACT_CONTENDED_NONE;
2989 * In the slowpath, we sanity check order to avoid ever trying to
2990 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2991 * be using allocators in order of preference for an area that is
2994 if (order >= MAX_ORDER) {
2995 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3000 * We also sanity check to catch abuse of atomic reserves being used by
3001 * callers that are not in atomic context.
3003 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3004 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3005 gfp_mask &= ~__GFP_ATOMIC;
3008 * If this allocation cannot block and it is for a specific node, then
3009 * fail early. There's no need to wakeup kswapd or retry for a
3010 * speculative node-specific allocation.
3012 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3016 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3017 wake_all_kswapds(order, ac);
3020 * OK, we're below the kswapd watermark and have kicked background
3021 * reclaim. Now things get more complex, so set up alloc_flags according
3022 * to how we want to proceed.
3024 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3027 * Find the true preferred zone if the allocation is unconstrained by
3030 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3031 struct zoneref *preferred_zoneref;
3032 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3033 ac->high_zoneidx, NULL, &ac->preferred_zone);
3034 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3037 /* This is the last chance, in general, before the goto nopage. */
3038 page = get_page_from_freelist(gfp_mask, order,
3039 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3043 /* Allocate without watermarks if the context allows */
3044 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3046 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3047 * the allocation is high priority and these type of
3048 * allocations are system rather than user orientated
3050 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3052 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3059 /* Caller is not willing to reclaim, we can't balance anything */
3060 if (!can_direct_reclaim) {
3062 * All existing users of the deprecated __GFP_NOFAIL are
3063 * blockable, so warn of any new users that actually allow this
3064 * type of allocation to fail.
3066 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3070 /* Avoid recursion of direct reclaim */
3071 if (current->flags & PF_MEMALLOC)
3074 /* Avoid allocations with no watermarks from looping endlessly */
3075 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3079 * Try direct compaction. The first pass is asynchronous. Subsequent
3080 * attempts after direct reclaim are synchronous
3082 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3084 &contended_compaction,
3085 &deferred_compaction);
3089 /* Checks for THP-specific high-order allocations */
3090 if (is_thp_gfp_mask(gfp_mask)) {
3092 * If compaction is deferred for high-order allocations, it is
3093 * because sync compaction recently failed. If this is the case
3094 * and the caller requested a THP allocation, we do not want
3095 * to heavily disrupt the system, so we fail the allocation
3096 * instead of entering direct reclaim.
3098 if (deferred_compaction)
3102 * In all zones where compaction was attempted (and not
3103 * deferred or skipped), lock contention has been detected.
3104 * For THP allocation we do not want to disrupt the others
3105 * so we fallback to base pages instead.
3107 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3111 * If compaction was aborted due to need_resched(), we do not
3112 * want to further increase allocation latency, unless it is
3113 * khugepaged trying to collapse.
3115 if (contended_compaction == COMPACT_CONTENDED_SCHED
3116 && !(current->flags & PF_KTHREAD))
3121 * It can become very expensive to allocate transparent hugepages at
3122 * fault, so use asynchronous memory compaction for THP unless it is
3123 * khugepaged trying to collapse.
3125 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3126 migration_mode = MIGRATE_SYNC_LIGHT;
3128 /* Try direct reclaim and then allocating */
3129 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3130 &did_some_progress);
3134 /* Do not loop if specifically requested */
3135 if (gfp_mask & __GFP_NORETRY)
3138 /* Keep reclaiming pages as long as there is reasonable progress */
3139 pages_reclaimed += did_some_progress;
3140 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3141 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3142 /* Wait for some write requests to complete then retry */
3143 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3147 /* Reclaim has failed us, start killing things */
3148 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3152 /* Retry as long as the OOM killer is making progress */
3153 if (did_some_progress)
3158 * High-order allocations do not necessarily loop after
3159 * direct reclaim and reclaim/compaction depends on compaction
3160 * being called after reclaim so call directly if necessary
3162 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3164 &contended_compaction,
3165 &deferred_compaction);
3169 warn_alloc_failed(gfp_mask, order, NULL);
3175 * This is the 'heart' of the zoned buddy allocator.
3178 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3179 struct zonelist *zonelist, nodemask_t *nodemask)
3181 struct zoneref *preferred_zoneref;
3182 struct page *page = NULL;
3183 unsigned int cpuset_mems_cookie;
3184 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3185 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3186 struct alloc_context ac = {
3187 .high_zoneidx = gfp_zone(gfp_mask),
3188 .nodemask = nodemask,
3189 .migratetype = gfpflags_to_migratetype(gfp_mask),
3192 gfp_mask &= gfp_allowed_mask;
3194 lockdep_trace_alloc(gfp_mask);
3196 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3198 if (should_fail_alloc_page(gfp_mask, order))
3202 * Check the zones suitable for the gfp_mask contain at least one
3203 * valid zone. It's possible to have an empty zonelist as a result
3204 * of __GFP_THISNODE and a memoryless node
3206 if (unlikely(!zonelist->_zonerefs->zone))
3209 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3210 alloc_flags |= ALLOC_CMA;
3213 cpuset_mems_cookie = read_mems_allowed_begin();
3215 /* We set it here, as __alloc_pages_slowpath might have changed it */
3216 ac.zonelist = zonelist;
3218 /* Dirty zone balancing only done in the fast path */
3219 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3221 /* The preferred zone is used for statistics later */
3222 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3223 ac.nodemask ? : &cpuset_current_mems_allowed,
3224 &ac.preferred_zone);
3225 if (!ac.preferred_zone)
3227 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3229 /* First allocation attempt */
3230 alloc_mask = gfp_mask|__GFP_HARDWALL;
3231 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3232 if (unlikely(!page)) {
3234 * Runtime PM, block IO and its error handling path
3235 * can deadlock because I/O on the device might not
3238 alloc_mask = memalloc_noio_flags(gfp_mask);
3239 ac.spread_dirty_pages = false;
3241 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3244 if (kmemcheck_enabled && page)
3245 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3247 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3251 * When updating a task's mems_allowed, it is possible to race with
3252 * parallel threads in such a way that an allocation can fail while
3253 * the mask is being updated. If a page allocation is about to fail,
3254 * check if the cpuset changed during allocation and if so, retry.
3256 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3261 EXPORT_SYMBOL(__alloc_pages_nodemask);
3264 * Common helper functions.
3266 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3271 * __get_free_pages() returns a 32-bit address, which cannot represent
3274 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3276 page = alloc_pages(gfp_mask, order);
3279 return (unsigned long) page_address(page);
3281 EXPORT_SYMBOL(__get_free_pages);
3283 unsigned long get_zeroed_page(gfp_t gfp_mask)
3285 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3287 EXPORT_SYMBOL(get_zeroed_page);
3289 void __free_pages(struct page *page, unsigned int order)
3291 if (put_page_testzero(page)) {
3293 free_hot_cold_page(page, false);
3295 __free_pages_ok(page, order);
3299 EXPORT_SYMBOL(__free_pages);
3301 void free_pages(unsigned long addr, unsigned int order)
3304 VM_BUG_ON(!virt_addr_valid((void *)addr));
3305 __free_pages(virt_to_page((void *)addr), order);
3309 EXPORT_SYMBOL(free_pages);
3313 * An arbitrary-length arbitrary-offset area of memory which resides
3314 * within a 0 or higher order page. Multiple fragments within that page
3315 * are individually refcounted, in the page's reference counter.
3317 * The page_frag functions below provide a simple allocation framework for
3318 * page fragments. This is used by the network stack and network device
3319 * drivers to provide a backing region of memory for use as either an
3320 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3322 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3325 struct page *page = NULL;
3326 gfp_t gfp = gfp_mask;
3328 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3329 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3331 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3332 PAGE_FRAG_CACHE_MAX_ORDER);
3333 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3335 if (unlikely(!page))
3336 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3338 nc->va = page ? page_address(page) : NULL;
3343 void *__alloc_page_frag(struct page_frag_cache *nc,
3344 unsigned int fragsz, gfp_t gfp_mask)
3346 unsigned int size = PAGE_SIZE;
3350 if (unlikely(!nc->va)) {
3352 page = __page_frag_refill(nc, gfp_mask);
3356 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3357 /* if size can vary use size else just use PAGE_SIZE */
3360 /* Even if we own the page, we do not use atomic_set().
3361 * This would break get_page_unless_zero() users.
3363 atomic_add(size - 1, &page->_count);
3365 /* reset page count bias and offset to start of new frag */
3366 nc->pfmemalloc = page_is_pfmemalloc(page);
3367 nc->pagecnt_bias = size;
3371 offset = nc->offset - fragsz;
3372 if (unlikely(offset < 0)) {
3373 page = virt_to_page(nc->va);
3375 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3378 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3379 /* if size can vary use size else just use PAGE_SIZE */
3382 /* OK, page count is 0, we can safely set it */
3383 atomic_set(&page->_count, size);
3385 /* reset page count bias and offset to start of new frag */
3386 nc->pagecnt_bias = size;
3387 offset = size - fragsz;
3391 nc->offset = offset;
3393 return nc->va + offset;
3395 EXPORT_SYMBOL(__alloc_page_frag);
3398 * Frees a page fragment allocated out of either a compound or order 0 page.
3400 void __free_page_frag(void *addr)
3402 struct page *page = virt_to_head_page(addr);
3404 if (unlikely(put_page_testzero(page)))
3405 __free_pages_ok(page, compound_order(page));
3407 EXPORT_SYMBOL(__free_page_frag);
3410 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3411 * of the current memory cgroup.
3413 * It should be used when the caller would like to use kmalloc, but since the
3414 * allocation is large, it has to fall back to the page allocator.
3416 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3420 page = alloc_pages(gfp_mask, order);
3421 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3422 __free_pages(page, order);
3428 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3432 page = alloc_pages_node(nid, gfp_mask, order);
3433 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3434 __free_pages(page, order);
3441 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3444 void __free_kmem_pages(struct page *page, unsigned int order)
3446 memcg_kmem_uncharge(page, order);
3447 __free_pages(page, order);
3450 void free_kmem_pages(unsigned long addr, unsigned int order)
3453 VM_BUG_ON(!virt_addr_valid((void *)addr));
3454 __free_kmem_pages(virt_to_page((void *)addr), order);
3458 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3462 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3463 unsigned long used = addr + PAGE_ALIGN(size);
3465 split_page(virt_to_page((void *)addr), order);
3466 while (used < alloc_end) {
3471 return (void *)addr;
3475 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3476 * @size: the number of bytes to allocate
3477 * @gfp_mask: GFP flags for the allocation
3479 * This function is similar to alloc_pages(), except that it allocates the
3480 * minimum number of pages to satisfy the request. alloc_pages() can only
3481 * allocate memory in power-of-two pages.
3483 * This function is also limited by MAX_ORDER.
3485 * Memory allocated by this function must be released by free_pages_exact().
3487 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3489 unsigned int order = get_order(size);
3492 addr = __get_free_pages(gfp_mask, order);
3493 return make_alloc_exact(addr, order, size);
3495 EXPORT_SYMBOL(alloc_pages_exact);
3498 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3500 * @nid: the preferred node ID where memory should be allocated
3501 * @size: the number of bytes to allocate
3502 * @gfp_mask: GFP flags for the allocation
3504 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3507 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3509 unsigned int order = get_order(size);
3510 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3513 return make_alloc_exact((unsigned long)page_address(p), order, size);
3517 * free_pages_exact - release memory allocated via alloc_pages_exact()
3518 * @virt: the value returned by alloc_pages_exact.
3519 * @size: size of allocation, same value as passed to alloc_pages_exact().
3521 * Release the memory allocated by a previous call to alloc_pages_exact.
3523 void free_pages_exact(void *virt, size_t size)
3525 unsigned long addr = (unsigned long)virt;
3526 unsigned long end = addr + PAGE_ALIGN(size);
3528 while (addr < end) {
3533 EXPORT_SYMBOL(free_pages_exact);
3536 * nr_free_zone_pages - count number of pages beyond high watermark
3537 * @offset: The zone index of the highest zone
3539 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3540 * high watermark within all zones at or below a given zone index. For each
3541 * zone, the number of pages is calculated as:
3542 * managed_pages - high_pages
3544 static unsigned long nr_free_zone_pages(int offset)
3549 /* Just pick one node, since fallback list is circular */
3550 unsigned long sum = 0;
3552 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3554 for_each_zone_zonelist(zone, z, zonelist, offset) {
3555 unsigned long size = zone->managed_pages;
3556 unsigned long high = high_wmark_pages(zone);
3565 * nr_free_buffer_pages - count number of pages beyond high watermark
3567 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3568 * watermark within ZONE_DMA and ZONE_NORMAL.
3570 unsigned long nr_free_buffer_pages(void)
3572 return nr_free_zone_pages(gfp_zone(GFP_USER));
3574 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3577 * nr_free_pagecache_pages - count number of pages beyond high watermark
3579 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3580 * high watermark within all zones.
3582 unsigned long nr_free_pagecache_pages(void)
3584 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3587 static inline void show_node(struct zone *zone)
3589 if (IS_ENABLED(CONFIG_NUMA))
3590 printk("Node %d ", zone_to_nid(zone));
3593 void si_meminfo(struct sysinfo *val)
3595 val->totalram = totalram_pages;
3596 val->sharedram = global_page_state(NR_SHMEM);
3597 val->freeram = global_page_state(NR_FREE_PAGES);
3598 val->bufferram = nr_blockdev_pages();
3599 val->totalhigh = totalhigh_pages;
3600 val->freehigh = nr_free_highpages();
3601 val->mem_unit = PAGE_SIZE;
3604 EXPORT_SYMBOL(si_meminfo);
3607 void si_meminfo_node(struct sysinfo *val, int nid)
3609 int zone_type; /* needs to be signed */
3610 unsigned long managed_pages = 0;
3611 pg_data_t *pgdat = NODE_DATA(nid);
3613 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3614 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3615 val->totalram = managed_pages;
3616 val->sharedram = node_page_state(nid, NR_SHMEM);
3617 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3618 #ifdef CONFIG_HIGHMEM
3619 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3620 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3626 val->mem_unit = PAGE_SIZE;
3631 * Determine whether the node should be displayed or not, depending on whether
3632 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3634 bool skip_free_areas_node(unsigned int flags, int nid)
3637 unsigned int cpuset_mems_cookie;
3639 if (!(flags & SHOW_MEM_FILTER_NODES))
3643 cpuset_mems_cookie = read_mems_allowed_begin();
3644 ret = !node_isset(nid, cpuset_current_mems_allowed);
3645 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3650 #define K(x) ((x) << (PAGE_SHIFT-10))
3652 static void show_migration_types(unsigned char type)
3654 static const char types[MIGRATE_TYPES] = {
3655 [MIGRATE_UNMOVABLE] = 'U',
3656 [MIGRATE_RECLAIMABLE] = 'E',
3657 [MIGRATE_MOVABLE] = 'M',
3659 [MIGRATE_CMA] = 'C',
3661 #ifdef CONFIG_MEMORY_ISOLATION
3662 [MIGRATE_ISOLATE] = 'I',
3665 char tmp[MIGRATE_TYPES + 1];
3669 for (i = 0; i < MIGRATE_TYPES; i++) {
3670 if (type & (1 << i))
3675 printk("(%s) ", tmp);
3679 * Show free area list (used inside shift_scroll-lock stuff)
3680 * We also calculate the percentage fragmentation. We do this by counting the
3681 * memory on each free list with the exception of the first item on the list.
3684 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3687 void show_free_areas(unsigned int filter)
3689 unsigned long free_pcp = 0;
3693 for_each_populated_zone(zone) {
3694 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3697 for_each_online_cpu(cpu)
3698 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3701 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3702 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3703 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3704 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3705 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3706 " free:%lu free_pcp:%lu free_cma:%lu\n",
3707 global_page_state(NR_ACTIVE_ANON),
3708 global_page_state(NR_INACTIVE_ANON),
3709 global_page_state(NR_ISOLATED_ANON),
3710 global_page_state(NR_ACTIVE_FILE),
3711 global_page_state(NR_INACTIVE_FILE),
3712 global_page_state(NR_ISOLATED_FILE),
3713 global_page_state(NR_UNEVICTABLE),
3714 global_page_state(NR_FILE_DIRTY),
3715 global_page_state(NR_WRITEBACK),
3716 global_page_state(NR_UNSTABLE_NFS),
3717 global_page_state(NR_SLAB_RECLAIMABLE),
3718 global_page_state(NR_SLAB_UNRECLAIMABLE),
3719 global_page_state(NR_FILE_MAPPED),
3720 global_page_state(NR_SHMEM),
3721 global_page_state(NR_PAGETABLE),
3722 global_page_state(NR_BOUNCE),
3723 global_page_state(NR_FREE_PAGES),
3725 global_page_state(NR_FREE_CMA_PAGES));
3727 for_each_populated_zone(zone) {
3730 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3734 for_each_online_cpu(cpu)
3735 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3743 " active_anon:%lukB"
3744 " inactive_anon:%lukB"
3745 " active_file:%lukB"
3746 " inactive_file:%lukB"
3747 " unevictable:%lukB"
3748 " isolated(anon):%lukB"
3749 " isolated(file):%lukB"
3757 " slab_reclaimable:%lukB"
3758 " slab_unreclaimable:%lukB"
3759 " kernel_stack:%lukB"
3766 " writeback_tmp:%lukB"
3767 " pages_scanned:%lu"
3768 " all_unreclaimable? %s"
3771 K(zone_page_state(zone, NR_FREE_PAGES)),
3772 K(min_wmark_pages(zone)),
3773 K(low_wmark_pages(zone)),
3774 K(high_wmark_pages(zone)),
3775 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3776 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3777 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3778 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3779 K(zone_page_state(zone, NR_UNEVICTABLE)),
3780 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3781 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3782 K(zone->present_pages),
3783 K(zone->managed_pages),
3784 K(zone_page_state(zone, NR_MLOCK)),
3785 K(zone_page_state(zone, NR_FILE_DIRTY)),
3786 K(zone_page_state(zone, NR_WRITEBACK)),
3787 K(zone_page_state(zone, NR_FILE_MAPPED)),
3788 K(zone_page_state(zone, NR_SHMEM)),
3789 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3790 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3791 zone_page_state(zone, NR_KERNEL_STACK) *
3793 K(zone_page_state(zone, NR_PAGETABLE)),
3794 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3795 K(zone_page_state(zone, NR_BOUNCE)),
3797 K(this_cpu_read(zone->pageset->pcp.count)),
3798 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3799 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3800 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3801 (!zone_reclaimable(zone) ? "yes" : "no")
3803 printk("lowmem_reserve[]:");
3804 for (i = 0; i < MAX_NR_ZONES; i++)
3805 printk(" %ld", zone->lowmem_reserve[i]);
3809 for_each_populated_zone(zone) {
3811 unsigned long nr[MAX_ORDER], flags, total = 0;
3812 unsigned char types[MAX_ORDER];
3814 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3817 printk("%s: ", zone->name);
3819 spin_lock_irqsave(&zone->lock, flags);
3820 for (order = 0; order < MAX_ORDER; order++) {
3821 struct free_area *area = &zone->free_area[order];
3824 nr[order] = area->nr_free;
3825 total += nr[order] << order;
3828 for (type = 0; type < MIGRATE_TYPES; type++) {
3829 if (!list_empty(&area->free_list[type]))
3830 types[order] |= 1 << type;
3833 spin_unlock_irqrestore(&zone->lock, flags);
3834 for (order = 0; order < MAX_ORDER; order++) {
3835 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3837 show_migration_types(types[order]);
3839 printk("= %lukB\n", K(total));
3842 hugetlb_show_meminfo();
3844 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3846 show_swap_cache_info();
3849 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3851 zoneref->zone = zone;
3852 zoneref->zone_idx = zone_idx(zone);
3856 * Builds allocation fallback zone lists.
3858 * Add all populated zones of a node to the zonelist.
3860 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3864 enum zone_type zone_type = MAX_NR_ZONES;
3868 zone = pgdat->node_zones + zone_type;
3869 if (populated_zone(zone)) {
3870 zoneref_set_zone(zone,
3871 &zonelist->_zonerefs[nr_zones++]);
3872 check_highest_zone(zone_type);
3874 } while (zone_type);
3882 * 0 = automatic detection of better ordering.
3883 * 1 = order by ([node] distance, -zonetype)
3884 * 2 = order by (-zonetype, [node] distance)
3886 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3887 * the same zonelist. So only NUMA can configure this param.
3889 #define ZONELIST_ORDER_DEFAULT 0
3890 #define ZONELIST_ORDER_NODE 1
3891 #define ZONELIST_ORDER_ZONE 2
3893 /* zonelist order in the kernel.
3894 * set_zonelist_order() will set this to NODE or ZONE.
3896 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3897 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3901 /* The value user specified ....changed by config */
3902 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3903 /* string for sysctl */
3904 #define NUMA_ZONELIST_ORDER_LEN 16
3905 char numa_zonelist_order[16] = "default";
3908 * interface for configure zonelist ordering.
3909 * command line option "numa_zonelist_order"
3910 * = "[dD]efault - default, automatic configuration.
3911 * = "[nN]ode - order by node locality, then by zone within node
3912 * = "[zZ]one - order by zone, then by locality within zone
3915 static int __parse_numa_zonelist_order(char *s)
3917 if (*s == 'd' || *s == 'D') {
3918 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3919 } else if (*s == 'n' || *s == 'N') {
3920 user_zonelist_order = ZONELIST_ORDER_NODE;
3921 } else if (*s == 'z' || *s == 'Z') {
3922 user_zonelist_order = ZONELIST_ORDER_ZONE;
3925 "Ignoring invalid numa_zonelist_order value: "
3932 static __init int setup_numa_zonelist_order(char *s)
3939 ret = __parse_numa_zonelist_order(s);
3941 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3945 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3948 * sysctl handler for numa_zonelist_order
3950 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3951 void __user *buffer, size_t *length,
3954 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3956 static DEFINE_MUTEX(zl_order_mutex);
3958 mutex_lock(&zl_order_mutex);
3960 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3964 strcpy(saved_string, (char *)table->data);
3966 ret = proc_dostring(table, write, buffer, length, ppos);
3970 int oldval = user_zonelist_order;
3972 ret = __parse_numa_zonelist_order((char *)table->data);
3975 * bogus value. restore saved string
3977 strncpy((char *)table->data, saved_string,
3978 NUMA_ZONELIST_ORDER_LEN);
3979 user_zonelist_order = oldval;
3980 } else if (oldval != user_zonelist_order) {
3981 mutex_lock(&zonelists_mutex);
3982 build_all_zonelists(NULL, NULL);
3983 mutex_unlock(&zonelists_mutex);
3987 mutex_unlock(&zl_order_mutex);
3992 #define MAX_NODE_LOAD (nr_online_nodes)
3993 static int node_load[MAX_NUMNODES];
3996 * find_next_best_node - find the next node that should appear in a given node's fallback list
3997 * @node: node whose fallback list we're appending
3998 * @used_node_mask: nodemask_t of already used nodes
4000 * We use a number of factors to determine which is the next node that should
4001 * appear on a given node's fallback list. The node should not have appeared
4002 * already in @node's fallback list, and it should be the next closest node
4003 * according to the distance array (which contains arbitrary distance values
4004 * from each node to each node in the system), and should also prefer nodes
4005 * with no CPUs, since presumably they'll have very little allocation pressure
4006 * on them otherwise.
4007 * It returns -1 if no node is found.
4009 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4012 int min_val = INT_MAX;
4013 int best_node = NUMA_NO_NODE;
4014 const struct cpumask *tmp = cpumask_of_node(0);
4016 /* Use the local node if we haven't already */
4017 if (!node_isset(node, *used_node_mask)) {
4018 node_set(node, *used_node_mask);
4022 for_each_node_state(n, N_MEMORY) {
4024 /* Don't want a node to appear more than once */
4025 if (node_isset(n, *used_node_mask))
4028 /* Use the distance array to find the distance */
4029 val = node_distance(node, n);
4031 /* Penalize nodes under us ("prefer the next node") */
4034 /* Give preference to headless and unused nodes */
4035 tmp = cpumask_of_node(n);
4036 if (!cpumask_empty(tmp))
4037 val += PENALTY_FOR_NODE_WITH_CPUS;
4039 /* Slight preference for less loaded node */
4040 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4041 val += node_load[n];
4043 if (val < min_val) {
4050 node_set(best_node, *used_node_mask);
4057 * Build zonelists ordered by node and zones within node.
4058 * This results in maximum locality--normal zone overflows into local
4059 * DMA zone, if any--but risks exhausting DMA zone.
4061 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4064 struct zonelist *zonelist;
4066 zonelist = &pgdat->node_zonelists[0];
4067 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4069 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4070 zonelist->_zonerefs[j].zone = NULL;
4071 zonelist->_zonerefs[j].zone_idx = 0;
4075 * Build gfp_thisnode zonelists
4077 static void build_thisnode_zonelists(pg_data_t *pgdat)
4080 struct zonelist *zonelist;
4082 zonelist = &pgdat->node_zonelists[1];
4083 j = build_zonelists_node(pgdat, zonelist, 0);
4084 zonelist->_zonerefs[j].zone = NULL;
4085 zonelist->_zonerefs[j].zone_idx = 0;
4089 * Build zonelists ordered by zone and nodes within zones.
4090 * This results in conserving DMA zone[s] until all Normal memory is
4091 * exhausted, but results in overflowing to remote node while memory
4092 * may still exist in local DMA zone.
4094 static int node_order[MAX_NUMNODES];
4096 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4099 int zone_type; /* needs to be signed */
4101 struct zonelist *zonelist;
4103 zonelist = &pgdat->node_zonelists[0];
4105 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4106 for (j = 0; j < nr_nodes; j++) {
4107 node = node_order[j];
4108 z = &NODE_DATA(node)->node_zones[zone_type];
4109 if (populated_zone(z)) {
4111 &zonelist->_zonerefs[pos++]);
4112 check_highest_zone(zone_type);
4116 zonelist->_zonerefs[pos].zone = NULL;
4117 zonelist->_zonerefs[pos].zone_idx = 0;
4120 #if defined(CONFIG_64BIT)
4122 * Devices that require DMA32/DMA are relatively rare and do not justify a
4123 * penalty to every machine in case the specialised case applies. Default
4124 * to Node-ordering on 64-bit NUMA machines
4126 static int default_zonelist_order(void)
4128 return ZONELIST_ORDER_NODE;
4132 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4133 * by the kernel. If processes running on node 0 deplete the low memory zone
4134 * then reclaim will occur more frequency increasing stalls and potentially
4135 * be easier to OOM if a large percentage of the zone is under writeback or
4136 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4137 * Hence, default to zone ordering on 32-bit.
4139 static int default_zonelist_order(void)
4141 return ZONELIST_ORDER_ZONE;
4143 #endif /* CONFIG_64BIT */
4145 static void set_zonelist_order(void)
4147 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4148 current_zonelist_order = default_zonelist_order();
4150 current_zonelist_order = user_zonelist_order;
4153 static void build_zonelists(pg_data_t *pgdat)
4157 nodemask_t used_mask;
4158 int local_node, prev_node;
4159 struct zonelist *zonelist;
4160 unsigned int order = current_zonelist_order;
4162 /* initialize zonelists */
4163 for (i = 0; i < MAX_ZONELISTS; i++) {
4164 zonelist = pgdat->node_zonelists + i;
4165 zonelist->_zonerefs[0].zone = NULL;
4166 zonelist->_zonerefs[0].zone_idx = 0;
4169 /* NUMA-aware ordering of nodes */
4170 local_node = pgdat->node_id;
4171 load = nr_online_nodes;
4172 prev_node = local_node;
4173 nodes_clear(used_mask);
4175 memset(node_order, 0, sizeof(node_order));
4178 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4180 * We don't want to pressure a particular node.
4181 * So adding penalty to the first node in same
4182 * distance group to make it round-robin.
4184 if (node_distance(local_node, node) !=
4185 node_distance(local_node, prev_node))
4186 node_load[node] = load;
4190 if (order == ZONELIST_ORDER_NODE)
4191 build_zonelists_in_node_order(pgdat, node);
4193 node_order[j++] = node; /* remember order */
4196 if (order == ZONELIST_ORDER_ZONE) {
4197 /* calculate node order -- i.e., DMA last! */
4198 build_zonelists_in_zone_order(pgdat, j);
4201 build_thisnode_zonelists(pgdat);
4204 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4206 * Return node id of node used for "local" allocations.
4207 * I.e., first node id of first zone in arg node's generic zonelist.
4208 * Used for initializing percpu 'numa_mem', which is used primarily
4209 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4211 int local_memory_node(int node)
4215 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4216 gfp_zone(GFP_KERNEL),
4223 #else /* CONFIG_NUMA */
4225 static void set_zonelist_order(void)
4227 current_zonelist_order = ZONELIST_ORDER_ZONE;
4230 static void build_zonelists(pg_data_t *pgdat)
4232 int node, local_node;
4234 struct zonelist *zonelist;
4236 local_node = pgdat->node_id;
4238 zonelist = &pgdat->node_zonelists[0];
4239 j = build_zonelists_node(pgdat, zonelist, 0);
4242 * Now we build the zonelist so that it contains the zones
4243 * of all the other nodes.
4244 * We don't want to pressure a particular node, so when
4245 * building the zones for node N, we make sure that the
4246 * zones coming right after the local ones are those from
4247 * node N+1 (modulo N)
4249 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4250 if (!node_online(node))
4252 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4254 for (node = 0; node < local_node; node++) {
4255 if (!node_online(node))
4257 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4260 zonelist->_zonerefs[j].zone = NULL;
4261 zonelist->_zonerefs[j].zone_idx = 0;
4264 #endif /* CONFIG_NUMA */
4267 * Boot pageset table. One per cpu which is going to be used for all
4268 * zones and all nodes. The parameters will be set in such a way
4269 * that an item put on a list will immediately be handed over to
4270 * the buddy list. This is safe since pageset manipulation is done
4271 * with interrupts disabled.
4273 * The boot_pagesets must be kept even after bootup is complete for
4274 * unused processors and/or zones. They do play a role for bootstrapping
4275 * hotplugged processors.
4277 * zoneinfo_show() and maybe other functions do
4278 * not check if the processor is online before following the pageset pointer.
4279 * Other parts of the kernel may not check if the zone is available.
4281 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4282 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4283 static void setup_zone_pageset(struct zone *zone);
4286 * Global mutex to protect against size modification of zonelists
4287 * as well as to serialize pageset setup for the new populated zone.
4289 DEFINE_MUTEX(zonelists_mutex);
4291 /* return values int ....just for stop_machine() */
4292 static int __build_all_zonelists(void *data)
4296 pg_data_t *self = data;
4299 memset(node_load, 0, sizeof(node_load));
4302 if (self && !node_online(self->node_id)) {
4303 build_zonelists(self);
4306 for_each_online_node(nid) {
4307 pg_data_t *pgdat = NODE_DATA(nid);
4309 build_zonelists(pgdat);
4313 * Initialize the boot_pagesets that are going to be used
4314 * for bootstrapping processors. The real pagesets for
4315 * each zone will be allocated later when the per cpu
4316 * allocator is available.
4318 * boot_pagesets are used also for bootstrapping offline
4319 * cpus if the system is already booted because the pagesets
4320 * are needed to initialize allocators on a specific cpu too.
4321 * F.e. the percpu allocator needs the page allocator which
4322 * needs the percpu allocator in order to allocate its pagesets
4323 * (a chicken-egg dilemma).
4325 for_each_possible_cpu(cpu) {
4326 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4328 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4330 * We now know the "local memory node" for each node--
4331 * i.e., the node of the first zone in the generic zonelist.
4332 * Set up numa_mem percpu variable for all possible cpus
4333 * if associated node has been onlined.
4335 if (node_online(cpu_to_node(cpu)))
4336 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4338 set_cpu_numa_mem(cpu, NUMA_NO_NODE);
4345 static noinline void __init
4346 build_all_zonelists_init(void)
4348 __build_all_zonelists(NULL);
4349 mminit_verify_zonelist();
4350 cpuset_init_current_mems_allowed();
4354 * Called with zonelists_mutex held always
4355 * unless system_state == SYSTEM_BOOTING.
4357 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4358 * [we're only called with non-NULL zone through __meminit paths] and
4359 * (2) call of __init annotated helper build_all_zonelists_init
4360 * [protected by SYSTEM_BOOTING].
4362 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4364 set_zonelist_order();
4366 if (system_state == SYSTEM_BOOTING) {
4367 build_all_zonelists_init();
4369 #ifdef CONFIG_MEMORY_HOTPLUG
4371 setup_zone_pageset(zone);
4373 /* we have to stop all cpus to guarantee there is no user
4375 stop_machine(__build_all_zonelists, pgdat, NULL);
4376 /* cpuset refresh routine should be here */
4378 vm_total_pages = nr_free_pagecache_pages();
4380 * Disable grouping by mobility if the number of pages in the
4381 * system is too low to allow the mechanism to work. It would be
4382 * more accurate, but expensive to check per-zone. This check is
4383 * made on memory-hotadd so a system can start with mobility
4384 * disabled and enable it later
4386 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4387 page_group_by_mobility_disabled = 1;
4389 page_group_by_mobility_disabled = 0;
4391 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4392 "Total pages: %ld\n",
4394 zonelist_order_name[current_zonelist_order],
4395 page_group_by_mobility_disabled ? "off" : "on",
4398 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4403 * Helper functions to size the waitqueue hash table.
4404 * Essentially these want to choose hash table sizes sufficiently
4405 * large so that collisions trying to wait on pages are rare.
4406 * But in fact, the number of active page waitqueues on typical
4407 * systems is ridiculously low, less than 200. So this is even
4408 * conservative, even though it seems large.
4410 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4411 * waitqueues, i.e. the size of the waitq table given the number of pages.
4413 #define PAGES_PER_WAITQUEUE 256
4415 #ifndef CONFIG_MEMORY_HOTPLUG
4416 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4418 unsigned long size = 1;
4420 pages /= PAGES_PER_WAITQUEUE;
4422 while (size < pages)
4426 * Once we have dozens or even hundreds of threads sleeping
4427 * on IO we've got bigger problems than wait queue collision.
4428 * Limit the size of the wait table to a reasonable size.
4430 size = min(size, 4096UL);
4432 return max(size, 4UL);
4436 * A zone's size might be changed by hot-add, so it is not possible to determine
4437 * a suitable size for its wait_table. So we use the maximum size now.
4439 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4441 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4442 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4443 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4445 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4446 * or more by the traditional way. (See above). It equals:
4448 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4449 * ia64(16K page size) : = ( 8G + 4M)byte.
4450 * powerpc (64K page size) : = (32G +16M)byte.
4452 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4459 * This is an integer logarithm so that shifts can be used later
4460 * to extract the more random high bits from the multiplicative
4461 * hash function before the remainder is taken.
4463 static inline unsigned long wait_table_bits(unsigned long size)
4469 * Initially all pages are reserved - free ones are freed
4470 * up by free_all_bootmem() once the early boot process is
4471 * done. Non-atomic initialization, single-pass.
4473 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4474 unsigned long start_pfn, enum memmap_context context)
4476 pg_data_t *pgdat = NODE_DATA(nid);
4477 unsigned long end_pfn = start_pfn + size;
4480 unsigned long nr_initialised = 0;
4482 if (highest_memmap_pfn < end_pfn - 1)
4483 highest_memmap_pfn = end_pfn - 1;
4485 z = &pgdat->node_zones[zone];
4486 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4488 * There can be holes in boot-time mem_map[]s
4489 * handed to this function. They do not
4490 * exist on hotplugged memory.
4492 if (context == MEMMAP_EARLY) {
4493 if (!early_pfn_valid(pfn))
4495 if (!early_pfn_in_nid(pfn, nid))
4497 if (!update_defer_init(pgdat, pfn, end_pfn,
4503 * Mark the block movable so that blocks are reserved for
4504 * movable at startup. This will force kernel allocations
4505 * to reserve their blocks rather than leaking throughout
4506 * the address space during boot when many long-lived
4507 * kernel allocations are made.
4509 * bitmap is created for zone's valid pfn range. but memmap
4510 * can be created for invalid pages (for alignment)
4511 * check here not to call set_pageblock_migratetype() against
4514 if (!(pfn & (pageblock_nr_pages - 1))) {
4515 struct page *page = pfn_to_page(pfn);
4517 __init_single_page(page, pfn, zone, nid);
4518 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4520 __init_single_pfn(pfn, zone, nid);
4525 static void __meminit zone_init_free_lists(struct zone *zone)
4527 unsigned int order, t;
4528 for_each_migratetype_order(order, t) {
4529 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4530 zone->free_area[order].nr_free = 0;
4534 #ifndef __HAVE_ARCH_MEMMAP_INIT
4535 #define memmap_init(size, nid, zone, start_pfn) \
4536 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4539 static int zone_batchsize(struct zone *zone)
4545 * The per-cpu-pages pools are set to around 1000th of the
4546 * size of the zone. But no more than 1/2 of a meg.
4548 * OK, so we don't know how big the cache is. So guess.
4550 batch = zone->managed_pages / 1024;
4551 if (batch * PAGE_SIZE > 512 * 1024)
4552 batch = (512 * 1024) / PAGE_SIZE;
4553 batch /= 4; /* We effectively *= 4 below */
4558 * Clamp the batch to a 2^n - 1 value. Having a power
4559 * of 2 value was found to be more likely to have
4560 * suboptimal cache aliasing properties in some cases.
4562 * For example if 2 tasks are alternately allocating
4563 * batches of pages, one task can end up with a lot
4564 * of pages of one half of the possible page colors
4565 * and the other with pages of the other colors.
4567 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4572 /* The deferral and batching of frees should be suppressed under NOMMU
4575 * The problem is that NOMMU needs to be able to allocate large chunks
4576 * of contiguous memory as there's no hardware page translation to
4577 * assemble apparent contiguous memory from discontiguous pages.
4579 * Queueing large contiguous runs of pages for batching, however,
4580 * causes the pages to actually be freed in smaller chunks. As there
4581 * can be a significant delay between the individual batches being
4582 * recycled, this leads to the once large chunks of space being
4583 * fragmented and becoming unavailable for high-order allocations.
4590 * pcp->high and pcp->batch values are related and dependent on one another:
4591 * ->batch must never be higher then ->high.
4592 * The following function updates them in a safe manner without read side
4595 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4596 * those fields changing asynchronously (acording the the above rule).
4598 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4599 * outside of boot time (or some other assurance that no concurrent updaters
4602 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4603 unsigned long batch)
4605 /* start with a fail safe value for batch */
4609 /* Update high, then batch, in order */
4616 /* a companion to pageset_set_high() */
4617 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4619 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4622 static void pageset_init(struct per_cpu_pageset *p)
4624 struct per_cpu_pages *pcp;
4627 memset(p, 0, sizeof(*p));
4631 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4632 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4635 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4638 pageset_set_batch(p, batch);
4642 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4643 * to the value high for the pageset p.
4645 static void pageset_set_high(struct per_cpu_pageset *p,
4648 unsigned long batch = max(1UL, high / 4);
4649 if ((high / 4) > (PAGE_SHIFT * 8))
4650 batch = PAGE_SHIFT * 8;
4652 pageset_update(&p->pcp, high, batch);
4655 static void pageset_set_high_and_batch(struct zone *zone,
4656 struct per_cpu_pageset *pcp)
4658 if (percpu_pagelist_fraction)
4659 pageset_set_high(pcp,
4660 (zone->managed_pages /
4661 percpu_pagelist_fraction));
4663 pageset_set_batch(pcp, zone_batchsize(zone));
4666 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4668 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4671 pageset_set_high_and_batch(zone, pcp);
4674 static void __meminit setup_zone_pageset(struct zone *zone)
4677 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4678 for_each_possible_cpu(cpu)
4679 zone_pageset_init(zone, cpu);
4683 * Allocate per cpu pagesets and initialize them.
4684 * Before this call only boot pagesets were available.
4686 void __init setup_per_cpu_pageset(void)
4690 for_each_populated_zone(zone)
4691 setup_zone_pageset(zone);
4694 static noinline __init_refok
4695 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4701 * The per-page waitqueue mechanism uses hashed waitqueues
4704 zone->wait_table_hash_nr_entries =
4705 wait_table_hash_nr_entries(zone_size_pages);
4706 zone->wait_table_bits =
4707 wait_table_bits(zone->wait_table_hash_nr_entries);
4708 alloc_size = zone->wait_table_hash_nr_entries
4709 * sizeof(wait_queue_head_t);
4711 if (!slab_is_available()) {
4712 zone->wait_table = (wait_queue_head_t *)
4713 memblock_virt_alloc_node_nopanic(
4714 alloc_size, zone->zone_pgdat->node_id);
4717 * This case means that a zone whose size was 0 gets new memory
4718 * via memory hot-add.
4719 * But it may be the case that a new node was hot-added. In
4720 * this case vmalloc() will not be able to use this new node's
4721 * memory - this wait_table must be initialized to use this new
4722 * node itself as well.
4723 * To use this new node's memory, further consideration will be
4726 zone->wait_table = vmalloc(alloc_size);
4728 if (!zone->wait_table)
4731 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4732 init_waitqueue_head(zone->wait_table + i);
4737 static __meminit void zone_pcp_init(struct zone *zone)
4740 * per cpu subsystem is not up at this point. The following code
4741 * relies on the ability of the linker to provide the
4742 * offset of a (static) per cpu variable into the per cpu area.
4744 zone->pageset = &boot_pageset;
4746 if (populated_zone(zone))
4747 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4748 zone->name, zone->present_pages,
4749 zone_batchsize(zone));
4752 int __meminit init_currently_empty_zone(struct zone *zone,
4753 unsigned long zone_start_pfn,
4756 struct pglist_data *pgdat = zone->zone_pgdat;
4758 ret = zone_wait_table_init(zone, size);
4761 pgdat->nr_zones = zone_idx(zone) + 1;
4763 zone->zone_start_pfn = zone_start_pfn;
4765 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4766 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4768 (unsigned long)zone_idx(zone),
4769 zone_start_pfn, (zone_start_pfn + size));
4771 zone_init_free_lists(zone);
4776 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4777 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4780 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4782 int __meminit __early_pfn_to_nid(unsigned long pfn,
4783 struct mminit_pfnnid_cache *state)
4785 unsigned long start_pfn, end_pfn;
4788 if (state->last_start <= pfn && pfn < state->last_end)
4789 return state->last_nid;
4791 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4793 state->last_start = start_pfn;
4794 state->last_end = end_pfn;
4795 state->last_nid = nid;
4800 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4803 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4804 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4805 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4807 * If an architecture guarantees that all ranges registered contain no holes
4808 * and may be freed, this this function may be used instead of calling
4809 * memblock_free_early_nid() manually.
4811 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4813 unsigned long start_pfn, end_pfn;
4816 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4817 start_pfn = min(start_pfn, max_low_pfn);
4818 end_pfn = min(end_pfn, max_low_pfn);
4820 if (start_pfn < end_pfn)
4821 memblock_free_early_nid(PFN_PHYS(start_pfn),
4822 (end_pfn - start_pfn) << PAGE_SHIFT,
4828 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4829 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4831 * If an architecture guarantees that all ranges registered contain no holes and may
4832 * be freed, this function may be used instead of calling memory_present() manually.
4834 void __init sparse_memory_present_with_active_regions(int nid)
4836 unsigned long start_pfn, end_pfn;
4839 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4840 memory_present(this_nid, start_pfn, end_pfn);
4844 * get_pfn_range_for_nid - Return the start and end page frames for a node
4845 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4846 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4847 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4849 * It returns the start and end page frame of a node based on information
4850 * provided by memblock_set_node(). If called for a node
4851 * with no available memory, a warning is printed and the start and end
4854 void __meminit get_pfn_range_for_nid(unsigned int nid,
4855 unsigned long *start_pfn, unsigned long *end_pfn)
4857 unsigned long this_start_pfn, this_end_pfn;
4863 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4864 *start_pfn = min(*start_pfn, this_start_pfn);
4865 *end_pfn = max(*end_pfn, this_end_pfn);
4868 if (*start_pfn == -1UL)
4873 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4874 * assumption is made that zones within a node are ordered in monotonic
4875 * increasing memory addresses so that the "highest" populated zone is used
4877 static void __init find_usable_zone_for_movable(void)
4880 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4881 if (zone_index == ZONE_MOVABLE)
4884 if (arch_zone_highest_possible_pfn[zone_index] >
4885 arch_zone_lowest_possible_pfn[zone_index])
4889 VM_BUG_ON(zone_index == -1);
4890 movable_zone = zone_index;
4894 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4895 * because it is sized independent of architecture. Unlike the other zones,
4896 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4897 * in each node depending on the size of each node and how evenly kernelcore
4898 * is distributed. This helper function adjusts the zone ranges
4899 * provided by the architecture for a given node by using the end of the
4900 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4901 * zones within a node are in order of monotonic increases memory addresses
4903 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4904 unsigned long zone_type,
4905 unsigned long node_start_pfn,
4906 unsigned long node_end_pfn,
4907 unsigned long *zone_start_pfn,
4908 unsigned long *zone_end_pfn)
4910 /* Only adjust if ZONE_MOVABLE is on this node */
4911 if (zone_movable_pfn[nid]) {
4912 /* Size ZONE_MOVABLE */
4913 if (zone_type == ZONE_MOVABLE) {
4914 *zone_start_pfn = zone_movable_pfn[nid];
4915 *zone_end_pfn = min(node_end_pfn,
4916 arch_zone_highest_possible_pfn[movable_zone]);
4918 /* Adjust for ZONE_MOVABLE starting within this range */
4919 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4920 *zone_end_pfn > zone_movable_pfn[nid]) {
4921 *zone_end_pfn = zone_movable_pfn[nid];
4923 /* Check if this whole range is within ZONE_MOVABLE */
4924 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4925 *zone_start_pfn = *zone_end_pfn;
4930 * Return the number of pages a zone spans in a node, including holes
4931 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4933 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4934 unsigned long zone_type,
4935 unsigned long node_start_pfn,
4936 unsigned long node_end_pfn,
4937 unsigned long *ignored)
4939 unsigned long zone_start_pfn, zone_end_pfn;
4941 /* When hotadd a new node from cpu_up(), the node should be empty */
4942 if (!node_start_pfn && !node_end_pfn)
4945 /* Get the start and end of the zone */
4946 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4947 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4948 adjust_zone_range_for_zone_movable(nid, zone_type,
4949 node_start_pfn, node_end_pfn,
4950 &zone_start_pfn, &zone_end_pfn);
4952 /* Check that this node has pages within the zone's required range */
4953 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4956 /* Move the zone boundaries inside the node if necessary */
4957 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4958 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4960 /* Return the spanned pages */
4961 return zone_end_pfn - zone_start_pfn;
4965 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4966 * then all holes in the requested range will be accounted for.
4968 unsigned long __meminit __absent_pages_in_range(int nid,
4969 unsigned long range_start_pfn,
4970 unsigned long range_end_pfn)
4972 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4973 unsigned long start_pfn, end_pfn;
4976 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4977 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4978 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4979 nr_absent -= end_pfn - start_pfn;
4985 * absent_pages_in_range - Return number of page frames in holes within a range
4986 * @start_pfn: The start PFN to start searching for holes
4987 * @end_pfn: The end PFN to stop searching for holes
4989 * It returns the number of pages frames in memory holes within a range.
4991 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4992 unsigned long end_pfn)
4994 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4997 /* Return the number of page frames in holes in a zone on a node */
4998 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4999 unsigned long zone_type,
5000 unsigned long node_start_pfn,
5001 unsigned long node_end_pfn,
5002 unsigned long *ignored)
5004 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5005 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5006 unsigned long zone_start_pfn, zone_end_pfn;
5008 /* When hotadd a new node from cpu_up(), the node should be empty */
5009 if (!node_start_pfn && !node_end_pfn)
5012 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5013 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5015 adjust_zone_range_for_zone_movable(nid, zone_type,
5016 node_start_pfn, node_end_pfn,
5017 &zone_start_pfn, &zone_end_pfn);
5018 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5021 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5022 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5023 unsigned long zone_type,
5024 unsigned long node_start_pfn,
5025 unsigned long node_end_pfn,
5026 unsigned long *zones_size)
5028 return zones_size[zone_type];
5031 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5032 unsigned long zone_type,
5033 unsigned long node_start_pfn,
5034 unsigned long node_end_pfn,
5035 unsigned long *zholes_size)
5040 return zholes_size[zone_type];
5043 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5045 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5046 unsigned long node_start_pfn,
5047 unsigned long node_end_pfn,
5048 unsigned long *zones_size,
5049 unsigned long *zholes_size)
5051 unsigned long realtotalpages = 0, totalpages = 0;
5054 for (i = 0; i < MAX_NR_ZONES; i++) {
5055 struct zone *zone = pgdat->node_zones + i;
5056 unsigned long size, real_size;
5058 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5062 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5063 node_start_pfn, node_end_pfn,
5065 zone->spanned_pages = size;
5066 zone->present_pages = real_size;
5069 realtotalpages += real_size;
5072 pgdat->node_spanned_pages = totalpages;
5073 pgdat->node_present_pages = realtotalpages;
5074 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5078 #ifndef CONFIG_SPARSEMEM
5080 * Calculate the size of the zone->blockflags rounded to an unsigned long
5081 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5082 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5083 * round what is now in bits to nearest long in bits, then return it in
5086 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5088 unsigned long usemapsize;
5090 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5091 usemapsize = roundup(zonesize, pageblock_nr_pages);
5092 usemapsize = usemapsize >> pageblock_order;
5093 usemapsize *= NR_PAGEBLOCK_BITS;
5094 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5096 return usemapsize / 8;
5099 static void __init setup_usemap(struct pglist_data *pgdat,
5101 unsigned long zone_start_pfn,
5102 unsigned long zonesize)
5104 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5105 zone->pageblock_flags = NULL;
5107 zone->pageblock_flags =
5108 memblock_virt_alloc_node_nopanic(usemapsize,
5112 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5113 unsigned long zone_start_pfn, unsigned long zonesize) {}
5114 #endif /* CONFIG_SPARSEMEM */
5116 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5118 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5119 void __paginginit set_pageblock_order(void)
5123 /* Check that pageblock_nr_pages has not already been setup */
5124 if (pageblock_order)
5127 if (HPAGE_SHIFT > PAGE_SHIFT)
5128 order = HUGETLB_PAGE_ORDER;
5130 order = MAX_ORDER - 1;
5133 * Assume the largest contiguous order of interest is a huge page.
5134 * This value may be variable depending on boot parameters on IA64 and
5137 pageblock_order = order;
5139 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5142 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5143 * is unused as pageblock_order is set at compile-time. See
5144 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5147 void __paginginit set_pageblock_order(void)
5151 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5153 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5154 unsigned long present_pages)
5156 unsigned long pages = spanned_pages;
5159 * Provide a more accurate estimation if there are holes within
5160 * the zone and SPARSEMEM is in use. If there are holes within the
5161 * zone, each populated memory region may cost us one or two extra
5162 * memmap pages due to alignment because memmap pages for each
5163 * populated regions may not naturally algined on page boundary.
5164 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5166 if (spanned_pages > present_pages + (present_pages >> 4) &&
5167 IS_ENABLED(CONFIG_SPARSEMEM))
5168 pages = present_pages;
5170 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5174 * Set up the zone data structures:
5175 * - mark all pages reserved
5176 * - mark all memory queues empty
5177 * - clear the memory bitmaps
5179 * NOTE: pgdat should get zeroed by caller.
5181 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5184 int nid = pgdat->node_id;
5185 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5188 pgdat_resize_init(pgdat);
5189 #ifdef CONFIG_NUMA_BALANCING
5190 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5191 pgdat->numabalancing_migrate_nr_pages = 0;
5192 pgdat->numabalancing_migrate_next_window = jiffies;
5194 init_waitqueue_head(&pgdat->kswapd_wait);
5195 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5196 pgdat_page_ext_init(pgdat);
5198 for (j = 0; j < MAX_NR_ZONES; j++) {
5199 struct zone *zone = pgdat->node_zones + j;
5200 unsigned long size, realsize, freesize, memmap_pages;
5202 size = zone->spanned_pages;
5203 realsize = freesize = zone->present_pages;
5206 * Adjust freesize so that it accounts for how much memory
5207 * is used by this zone for memmap. This affects the watermark
5208 * and per-cpu initialisations
5210 memmap_pages = calc_memmap_size(size, realsize);
5211 if (!is_highmem_idx(j)) {
5212 if (freesize >= memmap_pages) {
5213 freesize -= memmap_pages;
5216 " %s zone: %lu pages used for memmap\n",
5217 zone_names[j], memmap_pages);
5220 " %s zone: %lu pages exceeds freesize %lu\n",
5221 zone_names[j], memmap_pages, freesize);
5224 /* Account for reserved pages */
5225 if (j == 0 && freesize > dma_reserve) {
5226 freesize -= dma_reserve;
5227 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5228 zone_names[0], dma_reserve);
5231 if (!is_highmem_idx(j))
5232 nr_kernel_pages += freesize;
5233 /* Charge for highmem memmap if there are enough kernel pages */
5234 else if (nr_kernel_pages > memmap_pages * 2)
5235 nr_kernel_pages -= memmap_pages;
5236 nr_all_pages += freesize;
5239 * Set an approximate value for lowmem here, it will be adjusted
5240 * when the bootmem allocator frees pages into the buddy system.
5241 * And all highmem pages will be managed by the buddy system.
5243 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5246 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5248 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5250 zone->name = zone_names[j];
5251 spin_lock_init(&zone->lock);
5252 spin_lock_init(&zone->lru_lock);
5253 zone_seqlock_init(zone);
5254 zone->zone_pgdat = pgdat;
5255 zone_pcp_init(zone);
5257 /* For bootup, initialized properly in watermark setup */
5258 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5260 lruvec_init(&zone->lruvec);
5264 set_pageblock_order();
5265 setup_usemap(pgdat, zone, zone_start_pfn, size);
5266 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5268 memmap_init(size, nid, j, zone_start_pfn);
5269 zone_start_pfn += size;
5273 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5275 unsigned long __maybe_unused offset = 0;
5277 /* Skip empty nodes */
5278 if (!pgdat->node_spanned_pages)
5281 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5282 /* ia64 gets its own node_mem_map, before this, without bootmem */
5283 if (!pgdat->node_mem_map) {
5284 unsigned long size, start, end;
5288 * The zone's endpoints aren't required to be MAX_ORDER
5289 * aligned but the node_mem_map endpoints must be in order
5290 * for the buddy allocator to function correctly.
5292 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5293 offset = pgdat->node_start_pfn - start;
5294 end = pgdat_end_pfn(pgdat);
5295 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5296 size = (end - start) * sizeof(struct page);
5297 map = alloc_remap(pgdat->node_id, size);
5299 map = memblock_virt_alloc_node_nopanic(size,
5301 pgdat->node_mem_map = map + offset;
5303 #ifndef CONFIG_NEED_MULTIPLE_NODES
5305 * With no DISCONTIG, the global mem_map is just set as node 0's
5307 if (pgdat == NODE_DATA(0)) {
5308 mem_map = NODE_DATA(0)->node_mem_map;
5309 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5310 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5312 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5315 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5318 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5319 unsigned long node_start_pfn, unsigned long *zholes_size)
5321 pg_data_t *pgdat = NODE_DATA(nid);
5322 unsigned long start_pfn = 0;
5323 unsigned long end_pfn = 0;
5325 /* pg_data_t should be reset to zero when it's allocated */
5326 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5328 reset_deferred_meminit(pgdat);
5329 pgdat->node_id = nid;
5330 pgdat->node_start_pfn = node_start_pfn;
5331 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5332 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5333 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5334 (u64)start_pfn << PAGE_SHIFT,
5335 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5337 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5338 zones_size, zholes_size);
5340 alloc_node_mem_map(pgdat);
5341 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5342 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5343 nid, (unsigned long)pgdat,
5344 (unsigned long)pgdat->node_mem_map);
5347 free_area_init_core(pgdat);
5350 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5352 #if MAX_NUMNODES > 1
5354 * Figure out the number of possible node ids.
5356 void __init setup_nr_node_ids(void)
5358 unsigned int highest;
5360 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5361 nr_node_ids = highest + 1;
5366 * node_map_pfn_alignment - determine the maximum internode alignment
5368 * This function should be called after node map is populated and sorted.
5369 * It calculates the maximum power of two alignment which can distinguish
5372 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5373 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5374 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5375 * shifted, 1GiB is enough and this function will indicate so.
5377 * This is used to test whether pfn -> nid mapping of the chosen memory
5378 * model has fine enough granularity to avoid incorrect mapping for the
5379 * populated node map.
5381 * Returns the determined alignment in pfn's. 0 if there is no alignment
5382 * requirement (single node).
5384 unsigned long __init node_map_pfn_alignment(void)
5386 unsigned long accl_mask = 0, last_end = 0;
5387 unsigned long start, end, mask;
5391 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5392 if (!start || last_nid < 0 || last_nid == nid) {
5399 * Start with a mask granular enough to pin-point to the
5400 * start pfn and tick off bits one-by-one until it becomes
5401 * too coarse to separate the current node from the last.
5403 mask = ~((1 << __ffs(start)) - 1);
5404 while (mask && last_end <= (start & (mask << 1)))
5407 /* accumulate all internode masks */
5411 /* convert mask to number of pages */
5412 return ~accl_mask + 1;
5415 /* Find the lowest pfn for a node */
5416 static unsigned long __init find_min_pfn_for_node(int nid)
5418 unsigned long min_pfn = ULONG_MAX;
5419 unsigned long start_pfn;
5422 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5423 min_pfn = min(min_pfn, start_pfn);
5425 if (min_pfn == ULONG_MAX) {
5427 "Could not find start_pfn for node %d\n", nid);
5435 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5437 * It returns the minimum PFN based on information provided via
5438 * memblock_set_node().
5440 unsigned long __init find_min_pfn_with_active_regions(void)
5442 return find_min_pfn_for_node(MAX_NUMNODES);
5446 * early_calculate_totalpages()
5447 * Sum pages in active regions for movable zone.
5448 * Populate N_MEMORY for calculating usable_nodes.
5450 static unsigned long __init early_calculate_totalpages(void)
5452 unsigned long totalpages = 0;
5453 unsigned long start_pfn, end_pfn;
5456 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5457 unsigned long pages = end_pfn - start_pfn;
5459 totalpages += pages;
5461 node_set_state(nid, N_MEMORY);
5467 * Find the PFN the Movable zone begins in each node. Kernel memory
5468 * is spread evenly between nodes as long as the nodes have enough
5469 * memory. When they don't, some nodes will have more kernelcore than
5472 static void __init find_zone_movable_pfns_for_nodes(void)
5475 unsigned long usable_startpfn;
5476 unsigned long kernelcore_node, kernelcore_remaining;
5477 /* save the state before borrow the nodemask */
5478 nodemask_t saved_node_state = node_states[N_MEMORY];
5479 unsigned long totalpages = early_calculate_totalpages();
5480 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5481 struct memblock_region *r;
5483 /* Need to find movable_zone earlier when movable_node is specified. */
5484 find_usable_zone_for_movable();
5487 * If movable_node is specified, ignore kernelcore and movablecore
5490 if (movable_node_is_enabled()) {
5491 for_each_memblock(memory, r) {
5492 if (!memblock_is_hotpluggable(r))
5497 usable_startpfn = PFN_DOWN(r->base);
5498 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5499 min(usable_startpfn, zone_movable_pfn[nid]) :
5507 * If movablecore=nn[KMG] was specified, calculate what size of
5508 * kernelcore that corresponds so that memory usable for
5509 * any allocation type is evenly spread. If both kernelcore
5510 * and movablecore are specified, then the value of kernelcore
5511 * will be used for required_kernelcore if it's greater than
5512 * what movablecore would have allowed.
5514 if (required_movablecore) {
5515 unsigned long corepages;
5518 * Round-up so that ZONE_MOVABLE is at least as large as what
5519 * was requested by the user
5521 required_movablecore =
5522 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5523 required_movablecore = min(totalpages, required_movablecore);
5524 corepages = totalpages - required_movablecore;
5526 required_kernelcore = max(required_kernelcore, corepages);
5530 * If kernelcore was not specified or kernelcore size is larger
5531 * than totalpages, there is no ZONE_MOVABLE.
5533 if (!required_kernelcore || required_kernelcore >= totalpages)
5536 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5537 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5540 /* Spread kernelcore memory as evenly as possible throughout nodes */
5541 kernelcore_node = required_kernelcore / usable_nodes;
5542 for_each_node_state(nid, N_MEMORY) {
5543 unsigned long start_pfn, end_pfn;
5546 * Recalculate kernelcore_node if the division per node
5547 * now exceeds what is necessary to satisfy the requested
5548 * amount of memory for the kernel
5550 if (required_kernelcore < kernelcore_node)
5551 kernelcore_node = required_kernelcore / usable_nodes;
5554 * As the map is walked, we track how much memory is usable
5555 * by the kernel using kernelcore_remaining. When it is
5556 * 0, the rest of the node is usable by ZONE_MOVABLE
5558 kernelcore_remaining = kernelcore_node;
5560 /* Go through each range of PFNs within this node */
5561 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5562 unsigned long size_pages;
5564 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5565 if (start_pfn >= end_pfn)
5568 /* Account for what is only usable for kernelcore */
5569 if (start_pfn < usable_startpfn) {
5570 unsigned long kernel_pages;
5571 kernel_pages = min(end_pfn, usable_startpfn)
5574 kernelcore_remaining -= min(kernel_pages,
5575 kernelcore_remaining);
5576 required_kernelcore -= min(kernel_pages,
5577 required_kernelcore);
5579 /* Continue if range is now fully accounted */
5580 if (end_pfn <= usable_startpfn) {
5583 * Push zone_movable_pfn to the end so
5584 * that if we have to rebalance
5585 * kernelcore across nodes, we will
5586 * not double account here
5588 zone_movable_pfn[nid] = end_pfn;
5591 start_pfn = usable_startpfn;
5595 * The usable PFN range for ZONE_MOVABLE is from
5596 * start_pfn->end_pfn. Calculate size_pages as the
5597 * number of pages used as kernelcore
5599 size_pages = end_pfn - start_pfn;
5600 if (size_pages > kernelcore_remaining)
5601 size_pages = kernelcore_remaining;
5602 zone_movable_pfn[nid] = start_pfn + size_pages;
5605 * Some kernelcore has been met, update counts and
5606 * break if the kernelcore for this node has been
5609 required_kernelcore -= min(required_kernelcore,
5611 kernelcore_remaining -= size_pages;
5612 if (!kernelcore_remaining)
5618 * If there is still required_kernelcore, we do another pass with one
5619 * less node in the count. This will push zone_movable_pfn[nid] further
5620 * along on the nodes that still have memory until kernelcore is
5624 if (usable_nodes && required_kernelcore > usable_nodes)
5628 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5629 for (nid = 0; nid < MAX_NUMNODES; nid++)
5630 zone_movable_pfn[nid] =
5631 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5634 /* restore the node_state */
5635 node_states[N_MEMORY] = saved_node_state;
5638 /* Any regular or high memory on that node ? */
5639 static void check_for_memory(pg_data_t *pgdat, int nid)
5641 enum zone_type zone_type;
5643 if (N_MEMORY == N_NORMAL_MEMORY)
5646 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5647 struct zone *zone = &pgdat->node_zones[zone_type];
5648 if (populated_zone(zone)) {
5649 node_set_state(nid, N_HIGH_MEMORY);
5650 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5651 zone_type <= ZONE_NORMAL)
5652 node_set_state(nid, N_NORMAL_MEMORY);
5659 * free_area_init_nodes - Initialise all pg_data_t and zone data
5660 * @max_zone_pfn: an array of max PFNs for each zone
5662 * This will call free_area_init_node() for each active node in the system.
5663 * Using the page ranges provided by memblock_set_node(), the size of each
5664 * zone in each node and their holes is calculated. If the maximum PFN
5665 * between two adjacent zones match, it is assumed that the zone is empty.
5666 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5667 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5668 * starts where the previous one ended. For example, ZONE_DMA32 starts
5669 * at arch_max_dma_pfn.
5671 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5673 unsigned long start_pfn, end_pfn;
5676 /* Record where the zone boundaries are */
5677 memset(arch_zone_lowest_possible_pfn, 0,
5678 sizeof(arch_zone_lowest_possible_pfn));
5679 memset(arch_zone_highest_possible_pfn, 0,
5680 sizeof(arch_zone_highest_possible_pfn));
5681 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5682 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5683 for (i = 1; i < MAX_NR_ZONES; i++) {
5684 if (i == ZONE_MOVABLE)
5686 arch_zone_lowest_possible_pfn[i] =
5687 arch_zone_highest_possible_pfn[i-1];
5688 arch_zone_highest_possible_pfn[i] =
5689 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5691 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5692 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5694 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5695 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5696 find_zone_movable_pfns_for_nodes();
5698 /* Print out the zone ranges */
5699 pr_info("Zone ranges:\n");
5700 for (i = 0; i < MAX_NR_ZONES; i++) {
5701 if (i == ZONE_MOVABLE)
5703 pr_info(" %-8s ", zone_names[i]);
5704 if (arch_zone_lowest_possible_pfn[i] ==
5705 arch_zone_highest_possible_pfn[i])
5708 pr_cont("[mem %#018Lx-%#018Lx]\n",
5709 (u64)arch_zone_lowest_possible_pfn[i]
5711 ((u64)arch_zone_highest_possible_pfn[i]
5712 << PAGE_SHIFT) - 1);
5715 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5716 pr_info("Movable zone start for each node\n");
5717 for (i = 0; i < MAX_NUMNODES; i++) {
5718 if (zone_movable_pfn[i])
5719 pr_info(" Node %d: %#018Lx\n", i,
5720 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5723 /* Print out the early node map */
5724 pr_info("Early memory node ranges\n");
5725 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5726 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5727 (u64)start_pfn << PAGE_SHIFT,
5728 ((u64)end_pfn << PAGE_SHIFT) - 1);
5730 /* Initialise every node */
5731 mminit_verify_pageflags_layout();
5732 setup_nr_node_ids();
5733 for_each_online_node(nid) {
5734 pg_data_t *pgdat = NODE_DATA(nid);
5735 free_area_init_node(nid, NULL,
5736 find_min_pfn_for_node(nid), NULL);
5738 /* Any memory on that node */
5739 if (pgdat->node_present_pages)
5740 node_set_state(nid, N_MEMORY);
5741 check_for_memory(pgdat, nid);
5745 static int __init cmdline_parse_core(char *p, unsigned long *core)
5747 unsigned long long coremem;
5751 coremem = memparse(p, &p);
5752 *core = coremem >> PAGE_SHIFT;
5754 /* Paranoid check that UL is enough for the coremem value */
5755 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5761 * kernelcore=size sets the amount of memory for use for allocations that
5762 * cannot be reclaimed or migrated.
5764 static int __init cmdline_parse_kernelcore(char *p)
5766 return cmdline_parse_core(p, &required_kernelcore);
5770 * movablecore=size sets the amount of memory for use for allocations that
5771 * can be reclaimed or migrated.
5773 static int __init cmdline_parse_movablecore(char *p)
5775 return cmdline_parse_core(p, &required_movablecore);
5778 early_param("kernelcore", cmdline_parse_kernelcore);
5779 early_param("movablecore", cmdline_parse_movablecore);
5781 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5783 void adjust_managed_page_count(struct page *page, long count)
5785 spin_lock(&managed_page_count_lock);
5786 page_zone(page)->managed_pages += count;
5787 totalram_pages += count;
5788 #ifdef CONFIG_HIGHMEM
5789 if (PageHighMem(page))
5790 totalhigh_pages += count;
5792 spin_unlock(&managed_page_count_lock);
5794 EXPORT_SYMBOL(adjust_managed_page_count);
5796 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5799 unsigned long pages = 0;
5801 start = (void *)PAGE_ALIGN((unsigned long)start);
5802 end = (void *)((unsigned long)end & PAGE_MASK);
5803 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5804 if ((unsigned int)poison <= 0xFF)
5805 memset(pos, poison, PAGE_SIZE);
5806 free_reserved_page(virt_to_page(pos));
5810 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5811 s, pages << (PAGE_SHIFT - 10), start, end);
5815 EXPORT_SYMBOL(free_reserved_area);
5817 #ifdef CONFIG_HIGHMEM
5818 void free_highmem_page(struct page *page)
5820 __free_reserved_page(page);
5822 page_zone(page)->managed_pages++;
5828 void __init mem_init_print_info(const char *str)
5830 unsigned long physpages, codesize, datasize, rosize, bss_size;
5831 unsigned long init_code_size, init_data_size;
5833 physpages = get_num_physpages();
5834 codesize = _etext - _stext;
5835 datasize = _edata - _sdata;
5836 rosize = __end_rodata - __start_rodata;
5837 bss_size = __bss_stop - __bss_start;
5838 init_data_size = __init_end - __init_begin;
5839 init_code_size = _einittext - _sinittext;
5842 * Detect special cases and adjust section sizes accordingly:
5843 * 1) .init.* may be embedded into .data sections
5844 * 2) .init.text.* may be out of [__init_begin, __init_end],
5845 * please refer to arch/tile/kernel/vmlinux.lds.S.
5846 * 3) .rodata.* may be embedded into .text or .data sections.
5848 #define adj_init_size(start, end, size, pos, adj) \
5850 if (start <= pos && pos < end && size > adj) \
5854 adj_init_size(__init_begin, __init_end, init_data_size,
5855 _sinittext, init_code_size);
5856 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5857 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5858 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5859 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5861 #undef adj_init_size
5863 pr_info("Memory: %luK/%luK available "
5864 "(%luK kernel code, %luK rwdata, %luK rodata, "
5865 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5866 #ifdef CONFIG_HIGHMEM
5870 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5871 codesize >> 10, datasize >> 10, rosize >> 10,
5872 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5873 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5874 totalcma_pages << (PAGE_SHIFT-10),
5875 #ifdef CONFIG_HIGHMEM
5876 totalhigh_pages << (PAGE_SHIFT-10),
5878 str ? ", " : "", str ? str : "");
5882 * set_dma_reserve - set the specified number of pages reserved in the first zone
5883 * @new_dma_reserve: The number of pages to mark reserved
5885 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5886 * In the DMA zone, a significant percentage may be consumed by kernel image
5887 * and other unfreeable allocations which can skew the watermarks badly. This
5888 * function may optionally be used to account for unfreeable pages in the
5889 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5890 * smaller per-cpu batchsize.
5892 void __init set_dma_reserve(unsigned long new_dma_reserve)
5894 dma_reserve = new_dma_reserve;
5897 void __init free_area_init(unsigned long *zones_size)
5899 free_area_init_node(0, zones_size,
5900 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5903 static int page_alloc_cpu_notify(struct notifier_block *self,
5904 unsigned long action, void *hcpu)
5906 int cpu = (unsigned long)hcpu;
5908 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5909 lru_add_drain_cpu(cpu);
5913 * Spill the event counters of the dead processor
5914 * into the current processors event counters.
5915 * This artificially elevates the count of the current
5918 vm_events_fold_cpu(cpu);
5921 * Zero the differential counters of the dead processor
5922 * so that the vm statistics are consistent.
5924 * This is only okay since the processor is dead and cannot
5925 * race with what we are doing.
5927 cpu_vm_stats_fold(cpu);
5932 void __init page_alloc_init(void)
5934 hotcpu_notifier(page_alloc_cpu_notify, 0);
5938 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5939 * or min_free_kbytes changes.
5941 static void calculate_totalreserve_pages(void)
5943 struct pglist_data *pgdat;
5944 unsigned long reserve_pages = 0;
5945 enum zone_type i, j;
5947 for_each_online_pgdat(pgdat) {
5948 for (i = 0; i < MAX_NR_ZONES; i++) {
5949 struct zone *zone = pgdat->node_zones + i;
5952 /* Find valid and maximum lowmem_reserve in the zone */
5953 for (j = i; j < MAX_NR_ZONES; j++) {
5954 if (zone->lowmem_reserve[j] > max)
5955 max = zone->lowmem_reserve[j];
5958 /* we treat the high watermark as reserved pages. */
5959 max += high_wmark_pages(zone);
5961 if (max > zone->managed_pages)
5962 max = zone->managed_pages;
5963 reserve_pages += max;
5965 * Lowmem reserves are not available to
5966 * GFP_HIGHUSER page cache allocations and
5967 * kswapd tries to balance zones to their high
5968 * watermark. As a result, neither should be
5969 * regarded as dirtyable memory, to prevent a
5970 * situation where reclaim has to clean pages
5971 * in order to balance the zones.
5973 zone->dirty_balance_reserve = max;
5976 dirty_balance_reserve = reserve_pages;
5977 totalreserve_pages = reserve_pages;
5981 * setup_per_zone_lowmem_reserve - called whenever
5982 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5983 * has a correct pages reserved value, so an adequate number of
5984 * pages are left in the zone after a successful __alloc_pages().
5986 static void setup_per_zone_lowmem_reserve(void)
5988 struct pglist_data *pgdat;
5989 enum zone_type j, idx;
5991 for_each_online_pgdat(pgdat) {
5992 for (j = 0; j < MAX_NR_ZONES; j++) {
5993 struct zone *zone = pgdat->node_zones + j;
5994 unsigned long managed_pages = zone->managed_pages;
5996 zone->lowmem_reserve[j] = 0;
6000 struct zone *lower_zone;
6004 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6005 sysctl_lowmem_reserve_ratio[idx] = 1;
6007 lower_zone = pgdat->node_zones + idx;
6008 lower_zone->lowmem_reserve[j] = managed_pages /
6009 sysctl_lowmem_reserve_ratio[idx];
6010 managed_pages += lower_zone->managed_pages;
6015 /* update totalreserve_pages */
6016 calculate_totalreserve_pages();
6019 static void __setup_per_zone_wmarks(void)
6021 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6022 unsigned long lowmem_pages = 0;
6024 unsigned long flags;
6026 /* Calculate total number of !ZONE_HIGHMEM pages */
6027 for_each_zone(zone) {
6028 if (!is_highmem(zone))
6029 lowmem_pages += zone->managed_pages;
6032 for_each_zone(zone) {
6035 spin_lock_irqsave(&zone->lock, flags);
6036 tmp = (u64)pages_min * zone->managed_pages;
6037 do_div(tmp, lowmem_pages);
6038 if (is_highmem(zone)) {
6040 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6041 * need highmem pages, so cap pages_min to a small
6044 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6045 * deltas control asynch page reclaim, and so should
6046 * not be capped for highmem.
6048 unsigned long min_pages;
6050 min_pages = zone->managed_pages / 1024;
6051 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6052 zone->watermark[WMARK_MIN] = min_pages;
6055 * If it's a lowmem zone, reserve a number of pages
6056 * proportionate to the zone's size.
6058 zone->watermark[WMARK_MIN] = tmp;
6061 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6062 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6064 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6065 high_wmark_pages(zone) - low_wmark_pages(zone) -
6066 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6068 spin_unlock_irqrestore(&zone->lock, flags);
6071 /* update totalreserve_pages */
6072 calculate_totalreserve_pages();
6076 * setup_per_zone_wmarks - called when min_free_kbytes changes
6077 * or when memory is hot-{added|removed}
6079 * Ensures that the watermark[min,low,high] values for each zone are set
6080 * correctly with respect to min_free_kbytes.
6082 void setup_per_zone_wmarks(void)
6084 mutex_lock(&zonelists_mutex);
6085 __setup_per_zone_wmarks();
6086 mutex_unlock(&zonelists_mutex);
6090 * The inactive anon list should be small enough that the VM never has to
6091 * do too much work, but large enough that each inactive page has a chance
6092 * to be referenced again before it is swapped out.
6094 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6095 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6096 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6097 * the anonymous pages are kept on the inactive list.
6100 * memory ratio inactive anon
6101 * -------------------------------------
6110 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6112 unsigned int gb, ratio;
6114 /* Zone size in gigabytes */
6115 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6117 ratio = int_sqrt(10 * gb);
6121 zone->inactive_ratio = ratio;
6124 static void __meminit setup_per_zone_inactive_ratio(void)
6129 calculate_zone_inactive_ratio(zone);
6133 * Initialise min_free_kbytes.
6135 * For small machines we want it small (128k min). For large machines
6136 * we want it large (64MB max). But it is not linear, because network
6137 * bandwidth does not increase linearly with machine size. We use
6139 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6140 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6156 int __meminit init_per_zone_wmark_min(void)
6158 unsigned long lowmem_kbytes;
6159 int new_min_free_kbytes;
6161 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6162 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6164 if (new_min_free_kbytes > user_min_free_kbytes) {
6165 min_free_kbytes = new_min_free_kbytes;
6166 if (min_free_kbytes < 128)
6167 min_free_kbytes = 128;
6168 if (min_free_kbytes > 65536)
6169 min_free_kbytes = 65536;
6171 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6172 new_min_free_kbytes, user_min_free_kbytes);
6174 setup_per_zone_wmarks();
6175 refresh_zone_stat_thresholds();
6176 setup_per_zone_lowmem_reserve();
6177 setup_per_zone_inactive_ratio();
6180 module_init(init_per_zone_wmark_min)
6183 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6184 * that we can call two helper functions whenever min_free_kbytes
6187 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6188 void __user *buffer, size_t *length, loff_t *ppos)
6192 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6197 user_min_free_kbytes = min_free_kbytes;
6198 setup_per_zone_wmarks();
6204 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6205 void __user *buffer, size_t *length, loff_t *ppos)
6210 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6215 zone->min_unmapped_pages = (zone->managed_pages *
6216 sysctl_min_unmapped_ratio) / 100;
6220 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6221 void __user *buffer, size_t *length, loff_t *ppos)
6226 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6231 zone->min_slab_pages = (zone->managed_pages *
6232 sysctl_min_slab_ratio) / 100;
6238 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6239 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6240 * whenever sysctl_lowmem_reserve_ratio changes.
6242 * The reserve ratio obviously has absolutely no relation with the
6243 * minimum watermarks. The lowmem reserve ratio can only make sense
6244 * if in function of the boot time zone sizes.
6246 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6247 void __user *buffer, size_t *length, loff_t *ppos)
6249 proc_dointvec_minmax(table, write, buffer, length, ppos);
6250 setup_per_zone_lowmem_reserve();
6255 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6256 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6257 * pagelist can have before it gets flushed back to buddy allocator.
6259 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6260 void __user *buffer, size_t *length, loff_t *ppos)
6263 int old_percpu_pagelist_fraction;
6266 mutex_lock(&pcp_batch_high_lock);
6267 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6269 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6270 if (!write || ret < 0)
6273 /* Sanity checking to avoid pcp imbalance */
6274 if (percpu_pagelist_fraction &&
6275 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6276 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6282 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6285 for_each_populated_zone(zone) {
6288 for_each_possible_cpu(cpu)
6289 pageset_set_high_and_batch(zone,
6290 per_cpu_ptr(zone->pageset, cpu));
6293 mutex_unlock(&pcp_batch_high_lock);
6298 int hashdist = HASHDIST_DEFAULT;
6300 static int __init set_hashdist(char *str)
6304 hashdist = simple_strtoul(str, &str, 0);
6307 __setup("hashdist=", set_hashdist);
6311 * allocate a large system hash table from bootmem
6312 * - it is assumed that the hash table must contain an exact power-of-2
6313 * quantity of entries
6314 * - limit is the number of hash buckets, not the total allocation size
6316 void *__init alloc_large_system_hash(const char *tablename,
6317 unsigned long bucketsize,
6318 unsigned long numentries,
6321 unsigned int *_hash_shift,
6322 unsigned int *_hash_mask,
6323 unsigned long low_limit,
6324 unsigned long high_limit)
6326 unsigned long long max = high_limit;
6327 unsigned long log2qty, size;
6330 /* allow the kernel cmdline to have a say */
6332 /* round applicable memory size up to nearest megabyte */
6333 numentries = nr_kernel_pages;
6335 /* It isn't necessary when PAGE_SIZE >= 1MB */
6336 if (PAGE_SHIFT < 20)
6337 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6339 /* limit to 1 bucket per 2^scale bytes of low memory */
6340 if (scale > PAGE_SHIFT)
6341 numentries >>= (scale - PAGE_SHIFT);
6343 numentries <<= (PAGE_SHIFT - scale);
6345 /* Make sure we've got at least a 0-order allocation.. */
6346 if (unlikely(flags & HASH_SMALL)) {
6347 /* Makes no sense without HASH_EARLY */
6348 WARN_ON(!(flags & HASH_EARLY));
6349 if (!(numentries >> *_hash_shift)) {
6350 numentries = 1UL << *_hash_shift;
6351 BUG_ON(!numentries);
6353 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6354 numentries = PAGE_SIZE / bucketsize;
6356 numentries = roundup_pow_of_two(numentries);
6358 /* limit allocation size to 1/16 total memory by default */
6360 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6361 do_div(max, bucketsize);
6363 max = min(max, 0x80000000ULL);
6365 if (numentries < low_limit)
6366 numentries = low_limit;
6367 if (numentries > max)
6370 log2qty = ilog2(numentries);
6373 size = bucketsize << log2qty;
6374 if (flags & HASH_EARLY)
6375 table = memblock_virt_alloc_nopanic(size, 0);
6377 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6380 * If bucketsize is not a power-of-two, we may free
6381 * some pages at the end of hash table which
6382 * alloc_pages_exact() automatically does
6384 if (get_order(size) < MAX_ORDER) {
6385 table = alloc_pages_exact(size, GFP_ATOMIC);
6386 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6389 } while (!table && size > PAGE_SIZE && --log2qty);
6392 panic("Failed to allocate %s hash table\n", tablename);
6394 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6397 ilog2(size) - PAGE_SHIFT,
6401 *_hash_shift = log2qty;
6403 *_hash_mask = (1 << log2qty) - 1;
6408 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6409 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6412 #ifdef CONFIG_SPARSEMEM
6413 return __pfn_to_section(pfn)->pageblock_flags;
6415 return zone->pageblock_flags;
6416 #endif /* CONFIG_SPARSEMEM */
6419 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6421 #ifdef CONFIG_SPARSEMEM
6422 pfn &= (PAGES_PER_SECTION-1);
6423 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6425 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6426 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6427 #endif /* CONFIG_SPARSEMEM */
6431 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6432 * @page: The page within the block of interest
6433 * @pfn: The target page frame number
6434 * @end_bitidx: The last bit of interest to retrieve
6435 * @mask: mask of bits that the caller is interested in
6437 * Return: pageblock_bits flags
6439 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6440 unsigned long end_bitidx,
6444 unsigned long *bitmap;
6445 unsigned long bitidx, word_bitidx;
6448 zone = page_zone(page);
6449 bitmap = get_pageblock_bitmap(zone, pfn);
6450 bitidx = pfn_to_bitidx(zone, pfn);
6451 word_bitidx = bitidx / BITS_PER_LONG;
6452 bitidx &= (BITS_PER_LONG-1);
6454 word = bitmap[word_bitidx];
6455 bitidx += end_bitidx;
6456 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6460 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6461 * @page: The page within the block of interest
6462 * @flags: The flags to set
6463 * @pfn: The target page frame number
6464 * @end_bitidx: The last bit of interest
6465 * @mask: mask of bits that the caller is interested in
6467 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6469 unsigned long end_bitidx,
6473 unsigned long *bitmap;
6474 unsigned long bitidx, word_bitidx;
6475 unsigned long old_word, word;
6477 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6479 zone = page_zone(page);
6480 bitmap = get_pageblock_bitmap(zone, pfn);
6481 bitidx = pfn_to_bitidx(zone, pfn);
6482 word_bitidx = bitidx / BITS_PER_LONG;
6483 bitidx &= (BITS_PER_LONG-1);
6485 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6487 bitidx += end_bitidx;
6488 mask <<= (BITS_PER_LONG - bitidx - 1);
6489 flags <<= (BITS_PER_LONG - bitidx - 1);
6491 word = READ_ONCE(bitmap[word_bitidx]);
6493 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6494 if (word == old_word)
6501 * This function checks whether pageblock includes unmovable pages or not.
6502 * If @count is not zero, it is okay to include less @count unmovable pages
6504 * PageLRU check without isolation or lru_lock could race so that
6505 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6506 * expect this function should be exact.
6508 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6509 bool skip_hwpoisoned_pages)
6511 unsigned long pfn, iter, found;
6515 * For avoiding noise data, lru_add_drain_all() should be called
6516 * If ZONE_MOVABLE, the zone never contains unmovable pages
6518 if (zone_idx(zone) == ZONE_MOVABLE)
6520 mt = get_pageblock_migratetype(page);
6521 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6524 pfn = page_to_pfn(page);
6525 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6526 unsigned long check = pfn + iter;
6528 if (!pfn_valid_within(check))
6531 page = pfn_to_page(check);
6534 * Hugepages are not in LRU lists, but they're movable.
6535 * We need not scan over tail pages bacause we don't
6536 * handle each tail page individually in migration.
6538 if (PageHuge(page)) {
6539 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6544 * We can't use page_count without pin a page
6545 * because another CPU can free compound page.
6546 * This check already skips compound tails of THP
6547 * because their page->_count is zero at all time.
6549 if (!atomic_read(&page->_count)) {
6550 if (PageBuddy(page))
6551 iter += (1 << page_order(page)) - 1;
6556 * The HWPoisoned page may be not in buddy system, and
6557 * page_count() is not 0.
6559 if (skip_hwpoisoned_pages && PageHWPoison(page))
6565 * If there are RECLAIMABLE pages, we need to check
6566 * it. But now, memory offline itself doesn't call
6567 * shrink_node_slabs() and it still to be fixed.
6570 * If the page is not RAM, page_count()should be 0.
6571 * we don't need more check. This is an _used_ not-movable page.
6573 * The problematic thing here is PG_reserved pages. PG_reserved
6574 * is set to both of a memory hole page and a _used_ kernel
6583 bool is_pageblock_removable_nolock(struct page *page)
6589 * We have to be careful here because we are iterating over memory
6590 * sections which are not zone aware so we might end up outside of
6591 * the zone but still within the section.
6592 * We have to take care about the node as well. If the node is offline
6593 * its NODE_DATA will be NULL - see page_zone.
6595 if (!node_online(page_to_nid(page)))
6598 zone = page_zone(page);
6599 pfn = page_to_pfn(page);
6600 if (!zone_spans_pfn(zone, pfn))
6603 return !has_unmovable_pages(zone, page, 0, true);
6608 static unsigned long pfn_max_align_down(unsigned long pfn)
6610 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6611 pageblock_nr_pages) - 1);
6614 static unsigned long pfn_max_align_up(unsigned long pfn)
6616 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6617 pageblock_nr_pages));
6620 /* [start, end) must belong to a single zone. */
6621 static int __alloc_contig_migrate_range(struct compact_control *cc,
6622 unsigned long start, unsigned long end)
6624 /* This function is based on compact_zone() from compaction.c. */
6625 unsigned long nr_reclaimed;
6626 unsigned long pfn = start;
6627 unsigned int tries = 0;
6632 while (pfn < end || !list_empty(&cc->migratepages)) {
6633 if (fatal_signal_pending(current)) {
6638 if (list_empty(&cc->migratepages)) {
6639 cc->nr_migratepages = 0;
6640 pfn = isolate_migratepages_range(cc, pfn, end);
6646 } else if (++tries == 5) {
6647 ret = ret < 0 ? ret : -EBUSY;
6651 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6653 cc->nr_migratepages -= nr_reclaimed;
6655 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6656 NULL, 0, cc->mode, MR_CMA);
6659 putback_movable_pages(&cc->migratepages);
6666 * alloc_contig_range() -- tries to allocate given range of pages
6667 * @start: start PFN to allocate
6668 * @end: one-past-the-last PFN to allocate
6669 * @migratetype: migratetype of the underlaying pageblocks (either
6670 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6671 * in range must have the same migratetype and it must
6672 * be either of the two.
6674 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6675 * aligned, however it's the caller's responsibility to guarantee that
6676 * we are the only thread that changes migrate type of pageblocks the
6679 * The PFN range must belong to a single zone.
6681 * Returns zero on success or negative error code. On success all
6682 * pages which PFN is in [start, end) are allocated for the caller and
6683 * need to be freed with free_contig_range().
6685 int alloc_contig_range(unsigned long start, unsigned long end,
6686 unsigned migratetype)
6688 unsigned long outer_start, outer_end;
6692 struct compact_control cc = {
6693 .nr_migratepages = 0,
6695 .zone = page_zone(pfn_to_page(start)),
6696 .mode = MIGRATE_SYNC,
6697 .ignore_skip_hint = true,
6699 INIT_LIST_HEAD(&cc.migratepages);
6702 * What we do here is we mark all pageblocks in range as
6703 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6704 * have different sizes, and due to the way page allocator
6705 * work, we align the range to biggest of the two pages so
6706 * that page allocator won't try to merge buddies from
6707 * different pageblocks and change MIGRATE_ISOLATE to some
6708 * other migration type.
6710 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6711 * migrate the pages from an unaligned range (ie. pages that
6712 * we are interested in). This will put all the pages in
6713 * range back to page allocator as MIGRATE_ISOLATE.
6715 * When this is done, we take the pages in range from page
6716 * allocator removing them from the buddy system. This way
6717 * page allocator will never consider using them.
6719 * This lets us mark the pageblocks back as
6720 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6721 * aligned range but not in the unaligned, original range are
6722 * put back to page allocator so that buddy can use them.
6725 ret = start_isolate_page_range(pfn_max_align_down(start),
6726 pfn_max_align_up(end), migratetype,
6731 ret = __alloc_contig_migrate_range(&cc, start, end);
6736 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6737 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6738 * more, all pages in [start, end) are free in page allocator.
6739 * What we are going to do is to allocate all pages from
6740 * [start, end) (that is remove them from page allocator).
6742 * The only problem is that pages at the beginning and at the
6743 * end of interesting range may be not aligned with pages that
6744 * page allocator holds, ie. they can be part of higher order
6745 * pages. Because of this, we reserve the bigger range and
6746 * once this is done free the pages we are not interested in.
6748 * We don't have to hold zone->lock here because the pages are
6749 * isolated thus they won't get removed from buddy.
6752 lru_add_drain_all();
6753 drain_all_pages(cc.zone);
6756 outer_start = start;
6757 while (!PageBuddy(pfn_to_page(outer_start))) {
6758 if (++order >= MAX_ORDER) {
6762 outer_start &= ~0UL << order;
6765 /* Make sure the range is really isolated. */
6766 if (test_pages_isolated(outer_start, end, false)) {
6767 pr_info("%s: [%lx, %lx) PFNs busy\n",
6768 __func__, outer_start, end);
6773 /* Grab isolated pages from freelists. */
6774 outer_end = isolate_freepages_range(&cc, outer_start, end);
6780 /* Free head and tail (if any) */
6781 if (start != outer_start)
6782 free_contig_range(outer_start, start - outer_start);
6783 if (end != outer_end)
6784 free_contig_range(end, outer_end - end);
6787 undo_isolate_page_range(pfn_max_align_down(start),
6788 pfn_max_align_up(end), migratetype);
6792 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6794 unsigned int count = 0;
6796 for (; nr_pages--; pfn++) {
6797 struct page *page = pfn_to_page(pfn);
6799 count += page_count(page) != 1;
6802 WARN(count != 0, "%d pages are still in use!\n", count);
6806 #ifdef CONFIG_MEMORY_HOTPLUG
6808 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6809 * page high values need to be recalulated.
6811 void __meminit zone_pcp_update(struct zone *zone)
6814 mutex_lock(&pcp_batch_high_lock);
6815 for_each_possible_cpu(cpu)
6816 pageset_set_high_and_batch(zone,
6817 per_cpu_ptr(zone->pageset, cpu));
6818 mutex_unlock(&pcp_batch_high_lock);
6822 void zone_pcp_reset(struct zone *zone)
6824 unsigned long flags;
6826 struct per_cpu_pageset *pset;
6828 /* avoid races with drain_pages() */
6829 local_irq_save(flags);
6830 if (zone->pageset != &boot_pageset) {
6831 for_each_online_cpu(cpu) {
6832 pset = per_cpu_ptr(zone->pageset, cpu);
6833 drain_zonestat(zone, pset);
6835 free_percpu(zone->pageset);
6836 zone->pageset = &boot_pageset;
6838 local_irq_restore(flags);
6841 #ifdef CONFIG_MEMORY_HOTREMOVE
6843 * All pages in the range must be isolated before calling this.
6846 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6850 unsigned int order, i;
6852 unsigned long flags;
6853 /* find the first valid pfn */
6854 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6859 zone = page_zone(pfn_to_page(pfn));
6860 spin_lock_irqsave(&zone->lock, flags);
6862 while (pfn < end_pfn) {
6863 if (!pfn_valid(pfn)) {
6867 page = pfn_to_page(pfn);
6869 * The HWPoisoned page may be not in buddy system, and
6870 * page_count() is not 0.
6872 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6874 SetPageReserved(page);
6878 BUG_ON(page_count(page));
6879 BUG_ON(!PageBuddy(page));
6880 order = page_order(page);
6881 #ifdef CONFIG_DEBUG_VM
6882 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6883 pfn, 1 << order, end_pfn);
6885 list_del(&page->lru);
6886 rmv_page_order(page);
6887 zone->free_area[order].nr_free--;
6888 for (i = 0; i < (1 << order); i++)
6889 SetPageReserved((page+i));
6890 pfn += (1 << order);
6892 spin_unlock_irqrestore(&zone->lock, flags);
6896 #ifdef CONFIG_MEMORY_FAILURE
6897 bool is_free_buddy_page(struct page *page)
6899 struct zone *zone = page_zone(page);
6900 unsigned long pfn = page_to_pfn(page);
6901 unsigned long flags;
6904 spin_lock_irqsave(&zone->lock, flags);
6905 for (order = 0; order < MAX_ORDER; order++) {
6906 struct page *page_head = page - (pfn & ((1 << order) - 1));
6908 if (PageBuddy(page_head) && page_order(page_head) >= order)
6911 spin_unlock_irqrestore(&zone->lock, flags);
6913 return order < MAX_ORDER;